Bridge Maintenance,

Safety, Management,
Life-Cycle Sustainability
and Innovations
Editors
Hiroshi Yokota and Dan M. Frangopol

International Association for

Bridge Maintenance and Safety
BRIDGE MAINTENANCE, SAFETY, MANAGEMENT, LIFE-CYCLE
SUSTAINABILITY AND INNOVATIONS
PROCEEDINGS OF THE TENTH INTERNATIONAL CONFERENCE ON BRIDGE MAINTENANCE,
SAFETY AND MANAGEMENT (IABMAS 2020), SAPPORO, JAPAN, 11-15 APRIL 2021

Bridge Maintenance, Safety,

Management, Life-Cycle
Sustainability and Innovations

Editors
Hiroshi Yokota
Faculty of Engineering, Hokkaido University, Sapporo, Japan

Dan M. Frangopol
Department of Civil and Environmental Engineering, ATLSS Engineering Research Center,
Lehigh University, Bethlehem, PA, USA
Front cover photograph: The Kintaikyo Bridge, located in Iwakuni City, Yamaguchi Prefec­
ture, makes five bold arches onto massive stone pillars as it crosses over the river. The bridge
length and width are 193.3 m and 5.0 m respectively. Construction of the bridge made of
wood without iron nails was completed in 1673 and it stood until 1950 when it was washed
away by a typhoon flood. Shortly after it collapsed, a precise reconstruction was determined
and completed in 1953. In the early 2000s, the bridge underwent its first renovations since its
reconstruction.

Back cover photograph: Aerial view of Sapporo City. Many bridges over the Toyohira River
can be found there.

CRC Press/Balkema is an imprint of the Taylor & Francis Group, an informa business

© 2021 Taylor & Francis Group, London, UK

Typeset by Integra Software Services Pvt. Ltd., Pondicherry, India

All rights reserved. No part of this publication or the information contained herein may be
reproduced, stored in a retrieval system, or transmitted in any form or by any means,
electronic, mechanical, by photocopying, recording or otherwise, without written prior
permission from the publisher.

Although all care is taken to ensure integrity and the quality of this publication and the
information herein, no responsibility is assumed by the publishers nor the author for any
damage to the property or persons as a result of operation or use of this publication and/or
the information contained herein.

Library of Congress Cataloging-in-Publication Data

Applied for
Published by: CRC Press/Balkema
Schipholweg 107C, 2316XC Leiden, The Netherlands
e-mail: Pub.NL@taylorandfrancis.com
www.routledge.com – www.taylorandfrancis.com

ISBN: 978-0-367-23278-8 (Hbk)

ISBN: 978-1-032-05524-4 (Pbk)
ISBN: 978-0-429-27911-9 (eBook)
DOI: 10.1201/9780429279119
https://doi.org/10.1201/9780429279119
 Bridge Maintenance, Safety, Management, Life-Cycle Sustainability and Innovations –
Yokota & Frangopol (eds)
© 2021 Taylor & Francis Group, London, ISBN 978-0-367-23278-8

Table of contents

Preface xliii

Sponsors xlv

Acknowledgements xlvii

IABMAS Executive Board xlix

Conference organization li

T.Y. LIN LECTURE

Durability assessment and re-design of massive concrete structures in sea-linking projects 3
Q.K. Su, K.F. Li, Q.W. Li & Z.H. Fan

KEYNOTE LECTURES
Digital transformation of bridges inspection, monitoring and maintenance processes 11
T.N. Bittencourt, M.M. Futai, A.P. da Conceição Neto & D.M. Ribeiro
How to design, construct, and maintain highly-durable concrete bridges in cold and local
regions 31
I. Iwaki
Development of the requirements to major infrastructure projects 39
E.S. Larsen, M.S. Lagergaard, A. Jørgensen, B. MacAulay & J. Laursen
Some lessons of more than 20 years of inspection, maintenance and rehabilitation of bridges
in Spain 51
F.M. Mato & I.P. Sánchez
Rising to the challenge of managing bridges in Australia 66
N.G. Powers
Rib-to-floor beam connections of orthotropic steel decks for bridge deck replacement
applications 76
R. Sause, Y. Chen, J. Saunders, I. Hodgson & J. Marks
Energy-efficient autonomous framework for monitoring railroad bridges in the USA using
wireless smart sensors 91
B.F. Spencer, T. Hoang & K.A. Mechitov
Why do they call Chongqing the bridge capital of China? 100
M.C. Tang
Applications of AI, BIM, and sensing to bridge maintenance 108
N. Yabuki

v
MINI-SYMPOSIA
MS01: Novel techniques regarding the assessment and monitoring of bridges
Organizers: A. Strauss & D.M. Frangopol
Overview on the prestress loss evaluation in concrete beams 117
M. Bonopera, K.C. Chang & Y.C. Ou
Crack identification and measurement of bridges by using CNN models 123
H. Qin, F. Huang & B. Cheng
Non-destructive detection of damages in concrete with thermal imaging 129
L. Mold, M. Auer, A. Strauss, M. Hoffmann & B. Täubling
Application of wireless sensor technology in load testing of large-span cable-supported
bridges 137
Z. Xu, X. Zhang & J. Wang
Effectiveness and durability of repair measures on corroding steel in concrete columns 144
F. Binder & S.L. Burtscher
Research on long-term health monitoring and operation evaluation system for long-span
self-anchored suspension bridge 151
X. Liu, X. Chen, J. Zhou & X. Li
Unmanned aerial vehicle (UAV)-enabled bridge inspection framework 158
B.J. Perry, Y. Guo, R. Atadero & J.W. van de Lindt
Structural parameter identification from image-based acquired information 166
M. Helmrich & G. Morgenthal
A machine learning approach to damage detection of bridges 173
R.C. George
Inverse response surface method for prestressed concrete bridge design 179
D. Lehký, M. Šomodíková, M. Lipowczan & D. Novák
Comparison of different low-cost sensors for structural health monitoring 186
S. Komarizadehasl, J. Turmo, B. Mobaraki & J.A. Lozano-Galant
Gaussian curvature as an indicator used for damage detection of bridge structures 192
T. Wu, L. Tang, C.Y. Jian, R.Y. Mao & Z.X. Zhou
Statistical modal analysis for bridges under ambient excitation by using improved random
decrement technique and wavelet transform 200
J. Liu & Q. Zhang
Contactless deformation detection for bridge monitoring: First application of Sentinel-1
radar data in Austria 208
A. Vorwagner, M. Schlögl, B. Widhalm, M. Avian, D. Prammer, P. Leopold & C. Honeger
Estimation of cable tension force based on digital image correlation 215
B. Yan, W. Chen, D. Li & J. Yu
Probabilistic analysis and safety formats approaches applied for Czech bridge structures
under the ATCZ190 SAFEBRIDGE project 221
M. Šomodíková, L. Novák, M. Lipowczan, M. Vyhlídal, J. Doležel, D. Lehký, D. Novák &
R. Pukl

vi
Monitoring systems for masonry tunnels 227
A. Strauss, H. Neuner, C. Harmening, C. Seywald, M. Österreicher & E. Pistone
Improved assessment of concrete bridges 234
M. Hauser, M. Rigler, E. Apostolidi, A. Strauss, T. Zimmermann & D.M. Frangopol

MS02: Recent trends in AI/IoT technologies for bridge maintenance and safety
Organizers: T. Kitahara, H. Furuta, P. Chen & M. Beer
A new reliability model of bridge fatigue based on mind evolutionary algorithm 245
J. Wu, B.F. Liu, K. Dong & M. Yang
Automated infrastructure inspection based on digital twins and machine learning 251
P. Furtner, E. Forstner & A. Karlusch
Development of the bridge inspection experience system with MR head-mounted display 256
Y. Baba, H. Emoto, S. Tanikawa, H. Nakamura & K. Kawamura
Bridge damage cropping-and-stitching segmentation using fully convolutional network based
on images from UAVs 264
J. Shi, J. Dang & R. Zuo
A crack detection method based on deep transfer learning 271
Y.G. Shen, Z.W. Yu & Z.L. Wen
Measuring traffic-induced loads and 3D bridge displacements with UAVs 279
B.J. Perry & Y. Guo
Tracking bridge condition over time using recurrent UAV-based inspection 286
B.J. Perry, Y. Guo, R. Atadero & J.W. van de Lindt
Estimating bridge characteristics with only situation characteristics using Bayesian networks 292
V. Panopoulos, A. Bougas, B. Garcia de Soto & B.T. Adey
Time series forecasting to jointly model bridge responses 299
O. Bahrami, R. Hou, W. Wang & J.P. Lynch
Mixed training of deep convolutional neural network for bridge deterioration detection with
UAV and inspection report sourced images 308
J. Dang & P. Chun
Fatigue stress spectra modeling of steel bridge decks using traffic monitoring data 313
N. Lu, Y. Liu & Y. Luo
Infrared image-analysis-based concrete inspection using machine learning 320
S. Hayashi, K. Kawanishi, T. Yamane, S. Izumi, I. Ujike & P. Chun
Reliability-based risk analysis for Maryland sign structures assessment 326
X. Liu, Y. Ye, C. Xu, Y. Zhu & C.C. Fu
A framework for addressing the uncertainty of factors influencing the overall deterioration
of existing concrete structures 333
P. Miao, Y. Zhang & H. Yokota
Applying fully convolutional neural networks for corrosion semantic segmentation for steel
bridges: The use of U-Net 341
S.-K. Chen, I-F. Huang & P.-H. Chen

vii
Application of virtual reality technology to cultivate skill for visual inspection of bridge 347
K. Ishibashi, H. Furuta, Y. Nomura, K. Nakatsu & K. Takahashi
Variability-based method for balancing structural optimization and reliability 353
K.J. Haas

MS03: Bridge health monitoring conjunction with smart cities

Organizers: A. Miyamoto, A. Yabe, P. Hradil & K. Koski
Development of a practical social big data collecting system for the bridge using by a large
vehicle 363
A. Yabe
Development of a remote monitoring system for road condition assessment and application 371
A. Miyamoto
Heavy vehicle-based bridge health monitoring system 381
K. Koski, L. Fülöp, T. Tirkkonen, A. Yabe & A. Miyamoto
Monitoring of a suspension bridge 387
L.L. Lai
State of play and challenges for the successful implementation of indirect structural health
monitoring (iSHM) for bridges 392
K. Gkoumas, F. Bono, M.C. Galassi, K. Gkoktsi & D. Tirelli
Monitoring stability of high-speed rail tracks: A feasibility study 399
A.A. Mosavi & D. Torres
Health monitoring of stress-laminated timber bridges 407
P. Hradil, S. Fortino, K. Koski & L. Fülöp

MS04: Bridge loading – Measurement and modelling

Organizers: C. Caprani, A. Nowak, E. O’Brien & X. Ruan
Calculation method of transverse load distribution in box girder bridge 419
Z.N. Yu, X.Z. Yang, Z.H. Yuan & X.L. Guo
Analysis of bridge-traffic system using agent-based cellular automaton traffic model 425
J. Wu, R.R. Liu, M. Yang & D. Kai
Multi-lane traffic load model of widening bridges considering lane load disparity 433
J.Y. Zhou, Z.X. Chen, X.F. Shi, X.J. Ye & J. Yi
Influence of monitoring duration on measured traffic action effects on road bridges 441
B. Sawicki & E. Brühwiler
Traffic load spectra for multi-span cable-stayed bridge 448
X.J. Wang & X. Ruan
Experimental analysis of the dynamic amplification factor under traffic load 455
D. Hekič, J. Kalin, A. Anžlin, M. Kreslin, A. Žnidarič & G. Turk
Live loads for assessment of bridges on heavy haul rail freight lines 462
M. Kabani & P. Moyo
Dynamic amplification of live loads on heavy-haul freight rail lines using monitoring data 469
M. Kabani & P. Moyo

viii
A time domain approach for reconstruction of moving loads acting on bridges from dynamic
response data 475
A. Firus, J. Schneider & R. Kemmler
Challenges in assessing and load rating of old railway viaducts for traction and braking forces 484
R. Salamy
Effect of local roughness damage and traffic flow on bridge dynamic responses 491
H. Ho & M. Nishio
Energy dissipating characteristics of Y-type shear connectors based on the number of ribs 499
D.Y. Kim, S.H. Kim, O. Han, T. Batbold & S.H.A. Shah
Measuring traffic load on Forth Road Suspension Bridge using Weigh-In-Motion and image
data 505
E.J. OBrien, E.A. Micu, A. Malekjafarian, P. Madden, E. Angus, M. Lydon & S.E. Taylor
Bridge Weigh-In-Motion using wireless accelerometers on a continuous girder bridge 511
T. Nagayama, S. Kato, D. Su & H. Wang
The influence of orthotropic steel bridge deck stiffness on the stress amplitude of roof 515
F.W. Wu, J. Dai, Z.D. Wu & Y.J. Wen
Stochastic traffic load models on road bridges for applying to finite element analysis 521
T. Kouta & C. Bucher
Development of LRFR provisions for emergency vehicles 528
B. Sivakumar, M. Ghosn & E. Senturk
Effect of operating temperature on the dynamic properties of a pultruded GFRP footbridge 535
J.W. Ngan, C.C. Caprani & S.H. Zhang
Use of structural health monitoring for assessing historical bridges under heavy loads 543
S. Zhang, C. Caprani, M.M. Melhem, A. Ng & N. Hodgins
Bridge safety assessment beyond deterministic methods: An Australian perspective 551
M.M. Melhem, C. Caprani, M.G. Stewart & A. Ng
Using structural reliability to decide on extreme loads accessing historical bridges 560
M.M. Melhem, C. Caprani, S. Zhang, A. Ng & N. Hodgins
Alternating iterative method for moving force identification 568
H.L. Liu, C. Li & L. Yu
Research on temperature load model of a modular cable-stayed bridge 575
J. Song, K. Hu, X.F. Shi & C.A. Yin

MS05: Prolonging the life of steel and steel composite bridges

Organizers: A. Pipinato, P. Collin & C. Rebelo
Experimental study on stability of compression-bending members strengthened under load 585
Q. Su, S. Wang, X. Jiang, L. Chen & W. Zhou
A methodology for assessment and retrofitting by TIG dressing of existing pre-fatigued
welded steel joints 592
A. Manai, F. Von Bock und Polach & J. Hedegård

ix
UHPFRC strengthening strategies on existing steel and steel-concrete bridges 598
A. Pipinato, R. Geier, S. Ivanov, C. Rebelo, P. Collin & R. Hallmark
Extending the fatigue life of existing truss bridges 606
A. Pipinato & C. Rebelo
Enhancement of riveted steel bridges by adding a trough in UHPFRC acting in composite action 611
E. Brühwiler
Effect of concrete cracks on the corrosion of stud shear connectors 619
X.Q. Xu, Z.W. Yu, D.Y. He, H.M. Tan & S.W. Zeng
Fatigue analysis of rib-to-deck double-sided welded joints in steel bridges 624
Y. Liu, F.H. Chen & N.W. Lu
Research and application of uncoated weathering steel bridge in China mainland 632
C.S. Wang, C.X. Tan, J.W. Zhang, L. Duan & X.L. Zhai
Shear behavior study for high strength I girders considering residual stress 640
C.S. Wang, H. Cao, L. Duan & Q. Wang
Analysis on the long-term performance of a curved composite girder bridge 648
S.G. Chen, C. Liang, Y.Q. Liu, B. Lei & C.J. Zhao
An innovative composite box girder bridge with webs composed of stiffened steel plates and
steel tubes 655
Y. Zhang, S.H. Wang, Y.Q. Liu, X.H. He, Y.Y Chen & Y. Zhang
Research on structural behavior of steel–concrete joint in hybrid girder under long-term
loads 664
X.P. Gao, Z.H. Huang, Y.Q. Liu & B. Ma
Study of mechanical behavior of a new precast steel-concrete composite pier cap 671
Q. Zhou, Y.J. Li, S.W. Liu & Y.Q. Liu

MS06: Bridge safety, maintenance and management under natural hazards and climate change
Organizers: Y. Dong & D.M. Frangopol
Risk assessment of bridge under hurricane with experimental and numerical method 681
D. Zhu & Y. Dong
Seismic intensity measure selection under multiple criteria and uncertainty 686
J. Qian & Y. Dong
Risk-based bridge scour management: A survey 693
M. Pregnolato, L.J. Prendergast, P.J. Vardanega, P.F. Giordano & M.P. Limongelli
Corrosion effect on shear behavior of locally ungrouted PT concrete beams 702
L. Wang, Z. Hu, Y. Ma & J. Zhang
Fatigue life prediction of corroded RC beams considering bond degradation 710
J. Zhang, Z. Guo, Y. Ma & L. Wang
Repair loss assessment of seismic-resistant rocking bridges 717
A.I. Giouvanidis & Y. Dong
Effectiveness of drilling-hole method on mixed model fatigue crack 726
Z.Y. YuanZhou, L. Fang, B.H. Ji & Z. Ye

x
Prediction of crack initiation position at rib-to-deck welds using effective notch stress
approach 731
Q.D. Wang, Z. Ye, Y. Yao & B.H. Ji
Research on lateral seismic-constraint systems for cable-stayed bridge 736
Y.Q. Xu
Study on probability distribution of HPS specimens subjected to spray corrosion 743
L. Xiao, J. Peng, L. Wang, J. Zhang & C.S. Cai
Earthquake disaster countermeasures for bridges with rocking piers 749
S. Saito, Y. Ito & R. Hara

MS07: Value of information in bridge monitoring and management

Organizers: S. Ghosh, S. Thöns, C. Caprani & M.G. Limongelli
Application of value of information theory in adaptive metamodeling for reliability
assessment 757
R. Teixeira, A. O’Connor & M. Nogal
Value of structural health monitoring for bridges subjected to severe loads 766
M.S. Khan, C. Caprani, S. Ghosh & J. Ghosh
Fine measurement and calculation of parallel strand in stay cables based on vibration
frequency method 774
J. Dong, Y. Zhao, N. Ernestine, S. Ma, X. Xiang & D. Liu
Application of bridge construction monitoring management system based on Internet 782
R. Zhu, X.F. Shi & X.X. Li
Decision theoretic approach for identification of optimal proof load with sparse resistance
information 789
M. Kapoor, J.D. Sørensen, S. Ghosh & S. Thöns

MS08: Research progress on submerged floating tunnels

Organizers: Y.Q. Xiang, H.K. Lee, B. Faggiano & L. Martinelli
Theoretical framework of life cycle design of the submerged floating tunnel 801
Y.Q. Xiang, B. Bai & Y. Zhao
Advanced concrete technology for submerged floating tunnels 809
H.Y. Kim, H.N. Yoon, H.M. Son, J.S. Kim, D.W. Jin, H.J. Kim, S. Park & H.K. Lee
A tube model with two tension legs to investigate the spatial dynamic response of submerged
floating tunnel 811
Z.P. Yi, D.H. Yan & Y.Y. Zeng
Dynamic behavior analysis of high performance fiber concrete submerged floating tunnel
tube under collision 816
Y. Yang, Y.L. He & Y.Q. Xiang
Design of new large structures; Norwegian experience of submerged floating tube bridges 823
A. Minoretti, X. Xiang & A. Rønnquist
Investigation of rotational deformation of the suction anchor installed in sand from initial to
ultimate state 829
J.S. Bae, Y.H. Jeong, S. Manandhar & D.S. Kim

xi
SFTs under seismic loading: Conceptual design and optimization tools 837
F. Foti, L. Martinelli & F. Perotti
Simulation of hydrodynamic loads for a submerged floating tunnel using a copula-based
model 844
G.A. Torres, O. Morales-Nápoles & S.N. Jonkman
Investigation of tensile force on mooring line for a submerged floating tunnel 853
G.-J. Kim & H.-G. Kwak
Target reliability for submerged floating tunnels 857
C.M.P. ‘t Hart, D.J. Peters, O. Morales-Nápoles & S.N. Jonkman
Experimental design on corrosion behavior of steel in submerged floating tunnels 864
J.C. Park, S.L. Cha & H.J. Jung
VFIFE based hybrid simulation for submerged floating tunnel research 868
Y. Fang, Y.F. Duan, Y.Q. Xiang & H. Lin

MS09: Structural health monitoring of bridges based on modern sensor technologies and novel
methodologies
Organizers: C. Kim, F. Zhang, N.F. Catbas & Á. Cunha
Numerical study of damage detection of a truss bridge using pseudo local flexibility method 877
T.Y. Hsu, M.C. Lu, S.Y. Shiao, K.C. Chang & C.W. Kim
IoT bridge components – specialized smart monitoring solutions to address user-specific
needs 882
M. Imam, P. Savioz & C. O’Suilleabhain
Damage detection in a real truss bridge using Hilbert-Huang Transform of transient
vibrations 890
R.M. Delgadillo & J.R. Casas
A study on the stress ratio around welding lines of ribs in orthotropic steel decks 899
S. Kakizaki, H. Onishi, S. Ubagami, K. Hoshikawa & A. Horiai
The impact vibration test using a portable FWD system 905
Y. Kimura, H. Onishi, D. Yaegashi, R. Ishikawa & A. Takahashi
Observations of the behavior of the classic steel railway bridge structure 911
W. Anigacz, D. Beben & J. Kwiatkowski
Construction monitoring and load test for a 3×35m continuous steel-concrete composite
girder bridge 919
S.G. Cao, H. Hong, P. Ye, H. Tian, H.H. Han, S.G. Cao & A.R. Chen
Bridge management systems - a review of the state of the art and recommendations for
future practice 926
H. Habeenzu, P.J. McGetrick, D. Hester & S.E. Taylor
Bayesian system identification of a reinforced concrete beam subject to temperature
variations based on static response data 934
P. Simon, R. Schneider & M. Baeßler
Viability assessment of a mixed steel-concrete bridge structure 942
M.C. Scutaru, N. Țăranu, D. Ungureanu, C.C. Comisu & G. Boacă

xii
A novel laser and video-based displacement transducer for structural monitoring of long-
span bridges and tall structures 949
M.A. Vicente, J. Mínguez, D.C. González, N. Brown & T. Schumacher
Embedded distributed optical fiber sensors for health monitoring of concrete bridges 956
M.F. Bado, J.R. Casas & G. Kaklauskas
A long-term monitoring system for maintenance and management of extradosed bridges 963
C.C. Chen, C.L. Jiang, B.H. Lee, Y.C. Sung & K.C. Chang
Tensile force monitoring for construction of FCM bridges using EEM sensors 969
J. Kim & H.S. Kim
Remote microtremor monitoring for scour assessment of railway bridge 973
C.W. Kim, Y. Yoshitome, S. Kitagawa, M. Shinoda, H. Yao & Y. Hamada
Time-domain modal identification of bridges based on uncertainty quantification 979
Y. Goi & C.W. Kim
On-site stress measurement for steel reinforcement using a portable X-ray diffraction system 987
Y. Oshima, T. Noda, T. Furuta, J. Tomiyama, Y. Suda, O. Aydan, T. Makino, T. Miyagi,
N. Kurokawa, K. Tanaka & M. Yamaguchi

MS10: Life-cycle performance assessment of existing bridges in an aggressive environment

Organizers: M. Akiyama, D.M. Frangopol & H. Matsuzaki
Performance of self-healing concrete applied to tunnel engineering 993
X.F. Wang, Y.J. Huang, W.L. Wang, J. Liu & F. Xing
Simulation method of PC member with corrosion crack and breaking of PC tendon 1001
H. Nakamura, Y. Watanabe, T.A. Badmayev, T. Miura & Y. Yamamoto
Probability distribution of ultimate strain for aging deteriorated rubber bearings by Bayesian
estimation 1008
J. Dang, A. Igarashi & K. Hayashi
Optimum bridge life-cycle management with updating based on inspected fatigue crack
under uncertainty 1014
S. Kim, B. Ge & D.M. Frangopol
Analysis of typical environmental effects on the surface of prestressed concrete members of
bridges 1021
Q.L. Xu
Applicability of 2D ultrasonic phased array nondestructive test for fatigue crack of
orthotropic steel deck 1028
H. Shirahata
Seismic performance assessment of bridges with deteriorated isolators 1036
H. Matsuzaki
Application of observational data in reliability estimation of aging RC bridge structures
considering spatial steel corrosion distribution 1044
S. Srivaranun, K. Masuda, S. Lim, M. Akiyama, D.M. Frangopol & O. Maruyama
Reliability assessment of RC bridge girders with non-uniform steel corrosion using
probabilistic analysis and finite element method 1050
M. Zhang, S. Lim, M. Akiyama & D.M. Frangopol

xiii
Effects of non-uniform steel corrosion on the structural behavior of RC beams 1057
S. Lim, M. Zhang & M. Akiyama
Life-cycle cost analysis for rebar type selection in RC bridges located in coastal regions 1063
M.A. Hasan, S. Lim, M. Akiyama & D.M. Frangopol
Study on resistivity characteristics of embedded cement-based sensor 1068
W. Wang, F. Chen, Z. Xu, X. Wang & J. Liu
Benefit-cost ratio analysis of retrofit strategies for bridges considering the resilience effect 1076
C. Chiu
Life-cycle performance assessment of existing bridges based on artificial neural networks 1081
S. Bianchi, C. Manni & F. Biondini
Model updating for bridge structures based on the Kriging meta-model enhanced with DE
algorithm and analytic hierarchy process 1089
L. Mei, X.Y. Xia, J.H. Chen & W.L. Wang
Time-dependent reliability of aging bridges exposed to imprecise deterioration information 1096
C. Wang

MS11: Design, construction, maintenance, and management to realize highly-durable concrete

structures under harsh environments
Organizers: I. Iwaki, T. Ishida & H. Yokota
Improving the fatigue resistance of existing reinforced concrete bridge deck using high
penetration bonding agent 1107
K. Kaba, Y. Nagata, T. Maeshima & I. Iwaki
Long term expansions and deformations of real scale RC deck on steel girders caused
by ASR 1113
Y. Takahashi, T. Maeshima, I. Iwaki & K. Maekawa
Development of an AI-aided hammering test system 1122
Y. Kubota, Y. Nozoe, S. Takatsu, Y. Nagata, K. Tsuno, M. Iwata, Y. Kasai, J. Ye,
T. Okuma & M. Murakawa
Practical improvement of deicing salt scaling resistance on RC bridge slab in the Tohoku
region of Japan 1127
M. Zhang, M. Aba, Y. Sakoi, Y. Tsukinaga, Y.H. Kuang, S. Wabiko & M. Kushita
Empirical research on performance evaluation of highly durable RC road bridge deck at
construction process 1134
N. Sakakibara, Y. Tanaka, I. Iwaki, Y. Koda, K. Sato & T. Ishida
Salt damage simulation on each part and member of concrete bridge superstructure 1143
J. Tomiyama, Y. Suda, T. Yamaguchi, Y. Kato & K. Arai
Numerical evaluation of remaining fatigue life of road bridge deck with data assimilation
approach 1152
Y. Tanaka, E. Fathalla & K. Maekawa
Evaluation of soundness of PC road bridge in severe chloride environment 1161
T. Iidoi, H. Ueda, Y. Koda & I. Iwaki

xiv
Overturning mechanisms and evaluation strategy of box girder bridges under extreme vehicle
load 1169
Z.J. Zhou, H.Y. Wu, X.F. Shi & H.Y. Ma
The difference of behaviors between RC deck and AFRP-RC deck in punching shear test 1175
S. Tashima, H. Onishi, M. Moriai, N. Amano & S. Matsubara
Repairing effect of latex modified rapid hardening concrete on RC road bridge decks
deteriorated due to ASR and fatigue 1181
R. Kishira, T. Maeshima, Y. Koda & I. Iwaki
Effect of fly ash in Southeast Asia on the properties of mortar 1189
T.T. Win, R. Wattanapornprom & W. Pansuk
Evaluating early age thermal cracking risk of RC slabs on girder bridges 1195
A.I. Zerin & A. Hosoda
Data driven maintenance cycle focusing on deterioration mechanism of road bridge RC decks 1204
T. Ishida, J. Fang, E. Fathalla & T. Furukawa
Time-dependent change of bending performance of aramid short-fiber reinforced concrete
exposed in water or air with different temperature 1211
A. Hokura, S. Miyazato, S. Okamura, D. Yoshimoto & H. Kurakata
Key technology on improving the durability of long-span bridges 1218
J.H. Zhan
A basic study on characteristics of RC beam using ESCON 1225
T. Nozawa, Y. Kobayashi, Y. Sonoda & H. Tamai
Experiments and structure performance of fabricated box channel 1232
Z.Q. Liu, K. Hu & X.F. Shi

MS12: Advances in experimental and computational simulation for extreme load performance
assessment of bridges
Organizers: J. Hashemi, R. Al-Mahaidi & A.S. Whittaker
Hydrodynamic demands on coastal bridges due to wave impact 1241
R. Nasouri, A. Shahriar, A. Majlesi, A. Matamoros, A. Montoya & F. Testik
Multi-hazard events for bridges: State of play and numerical modeling of chained impact
and fire scenarios 1249
F. Petrini, C. Rossi, K. Gkoumas & F. Bontempi
An integrated impact analysis for riverine bridges subjected to high river flows 1258
M. Pregnolato, P. Bates, A. Winter, A.D. Sem, D. Mascarenas & M.R. Motley
Numerical simulations of collapse tests on RC beams 1265
M. Domaneschi, G.P. Cimellaro, G.C. Marano, M. Morgese, C. Pellecchia & A.A. Khalil
Fragility analysis of FRP strengthened bridges under extreme wave-induced forces 1273
I.M.I. Qeshta, R. Gravina, S. Setunge & M.J. Hashemi
The real resistance of a 115-years old truss bridge 1281
P. Ryjáček, V. Stančík, J. Braun & P. Jašek

xv
MS13: Suitable maintenance for road bridge of municipality
Organizers: S. Miyazato, H. Matsuda & K. Kinoshita
Proposal of suitable maintenance procedure for municipalities managed road bridges with
early deteriorations 1287
S. Miyazato & D. Hanaoka
Discussion of maintenance strategies for a self-anchored suspension bridge’s main cable 1296
X.H. Luo, Q.E. Deng & J.M. Xiao
Study on strength properties of butt-welded joints with stainless and carbon steel 1301
S. Aramaki, T. Shimozato, M. Tai, H. Matsushita & Y. Shimura
A behavior brought by daily temperature changes in existing small bridge 1308
T. Sugawara, H. Onishi, K. Terashima & A. Takahashi
The factors for the deterioration of renewed coating systems on the steel members 1313
K. Okabe, H. Onishi, R. Yamazaki & Y. Shimamoto
Applicability of new inspection technologies and numerical analysis by 3D measurement for
damaged RCT bridge and cutting girder 1318
K. Yamaguchi, S. Toyofuku & H. Matsuda
Nondestructive high-sensitivity magnetic detection of corrosion in light pole bases 1325
T. Ishikawa, K. Tsukada & H. Furuta
Fundamental study on repair and life extension effect of reinforced concrete by surface
penetrants 1329
D. Hanaoka & S. Miyazato
Clarification of performance requirements for utilizing robotic technology for regular
inspection of steel bridges 1336
R. Hasuike, K. Kinoshita, H. Hatano, H. Morimoto & K. Rokugo

MS14: Bridge safety evaluation and risk assessment: Code requirements vs. practical
considerations
Organizers: D. Su, M. Liu & H. Nassif
Reliability assessment of prestressed bridges in Brazil using WIM data 1347
E. Portela, P. Lou, H.H. Nassif & T.N. Bittencourt
Risk-based decision tools for managing bridge assets 1355
G. Fiorillo & H. Nassif
Experimental study on acoustic emission characteristics of different bonding interfaces in
steel-concrete composite beams 1363
W. Lu, H. Si & D. Su
The joint effect of concrete strength and loss of longitudinal steel on the reliability
of existing RC columns 1369
L.C.R. Castro & S.M.C. Diniz
Dynamic effects of central green belt on a 3-span continuous concrete girder bridge 1377
H.Y. Wu, A. Wang, Y. Xia & L.M. Sun
Characterization of superload traffic for bridge safety evaluation 1381
D. Su, J. Jesso & Y. Wang

xvi
Scour stability evaluation of bridge pier considering fluid-solid interaction 1386
T.H. Yang, P.W. Chen, T.K. Lin & K.C. Chang

MS15: FRP strengthening of metallic structures

Organizers: A. Al-Mosawe, R. Al-Mahaidi, X. Zhao & E. Ghafoori
Improved bond behavior of CFRP-steel composites by novel film adhesive at room and
elevated temperatures 1395
L. Ke, C.X. Li, J. He, Z.Y. Chen, S. Dong, J.L. Zhou & Y. Jiao
State of the art review of the application of SMA in steel bridges 1403
T.P. Nguyen, A. Al-Mosawe & R. Al-Mahaidi
Strengthening steel bridges under fatigue loading: State-of-the-art review 1409
T.P. Nguyen, A. Al-Mosawe & R. Al-Mahaidi
Feasibility of novel SMA-CFRP retrofit system for fatigue strengthening of steel structures 1417
Z. Lv, G. Zhang, X. Jiang, X. Wu, X. Qiang & S. Dai
FE analysis of a steel bridge strengthened with CFRP laminates 1424
A. Al-Mosawe, R. Al-Mahaidi, D. Alwash, X.L. Zhao, A. Hosseini, M. Motavalli & E. Ghafoori
Quick strengthening technique of corroded steel pipe by thermosetting prepreg sheets of
biaxial carbon fiber 1431
M. Matsumura, T. Hara, K. Takemoto & K. Sugiura

MS16: Innovative methods in strengthening of concrete bridges

Organizers: R. Kalfat, R. Al-Mahaidi & A. Al-Mosawe
Flexure strengthening of concrete bridge girders with concavely curved soffit using near-
surface-mounted CFRP bars 1439
K. Al-Ghrery, R. Kalfat, A. Al-Mosawe, R. Al-Mahaidi & N. Oukaili
Flexural failure behavior of textile reinforced concrete 1444
B.S. Muoi, H. Ando, M. Kunieda, S.C. Lim & S. Terada
An FE study on the behavior of CFRP shear strengthened and anchored post-tensioned
beams 1453
R. Jumaah, R. Kalfat, R. Al-Mahaidi & K. Abdouka
Punching shear strengthening of bridge decks using L-CFRP laminates 1458
H. Saleh, K. Abdouka, R. Kalfat & R. Al-Mahaidi

MS17: Technical risk assessment in recent bridge accidents and disasters

Organizers: F. Collazos-Arias, D. García-Sánchez, M. Wahbhe, R. Boundouki, A.D. Orcesi
& M.A. Valenzuela
State of the art of Chilean bridge collapse. Proposal of actions in Chilean bridge engineering 1467
M.A. Valenzuela, H. Pinto, M. Marquez & J.L. Seguel
Rehabilitation of bridges after heavy vehicle impacts with the parapet 1474
F. Collazos-Arias & D. García-Sánchez
Repair of a 19th century masonry bridge with piles settlement 1479
F. Collazos-Arias, L.C. López-Fando & T. Echeveste

xvii
SHM as intervention and mitigation action in the context of H2020 RAGTIME case study–
Mondalavia Viaduct 1485
D. García-Sánchez, D. Zamora, J.C. Jimenez, F. Di Gennaro, F. Federizzi, C. Fuggini &
I. Tesfai
Historical records of events related to technical risks in bridges and viaducts 1492
D. García-Sánchez, N. Tárrago, S. Pérez, A. Herrera, M. Gutierrez, I. Robles, R. Socorro,
J.L. Burón & S. Lenart
Safety of existing infrastructures: The collapse of the Morandi bridge in Genoa 1499
M. Domaneschi, G.P. Cimellaro, F. Ansari & M. Morgese

MS18: New NDE and monitoring technique for evaluating discrete damage in unhealthy
structures
Organizers: J. Lynch, D. Ozevin, L. Cao & T. Attard
Non-destructive evaluation by permanent magnet type Magnetic Main Flux Method 1509
H. Itoi, S. Shiiki, T. Moriya, K. Tsukada & S.K. Lee
Non-destructive inspection of corroded steel bars in concrete structures 1516
K. Suzuki & H. Hirata
Simulation of Radio Frequency Inductive Testing (RFIT) for deep sub-surface defects in
concrete 1524
P.G. McDonald, Z. Abbasi & L. Cao
Hidden Markov models for sequential damage detection of bridges 1528
O. Bahrami, W. Wang & J.P. Lynch
Assessment and prioritization strategies for scour critical bridges 1535
J. Chen, G. Xu & M. Hunnemann

MS19: Recent trends in monitoring and analysis of wind effects on long-span bridges
Organizers: H. Wang, X. He & T. Tao
Monitoring of wind characteristics of tropical storms: A non-stationary perspective 1545
T.Y. Tao & H. Wang
Short-term wind speed forecasting of downburst based on improved VARX model 1551
P. Shi, H. Wang & T.Y. Tao
Inverse estimation wind-induced responses in bridges from acceleration and wind data 1556
Ø.W. Petersen & O. Øiseth
Exploitation of local wind measurement data in long-span bridge design 1564
A. Fenerci, T.M. Lystad, D.R.F. Castellon & O. Øiseth
Experimental determination of bridge deck aeroelastic derivatives by stochastic subspace
technique 1571
X.Q. Liu, L. Yan & X.H. He
Flutter derivatives identification of bridge decks from free decay tests using an improved
Artificial Bee Colony algorithm 1578
Y. Lin, Z. Feng, X. Hua & Z. Chen
Wind-resistance safety evaluation for derrick cranes under construction of a long-span
bridge: A case study 1584
Z.K. Zhang, X.Y. Kang, G.Y. Jing & Y. Xia

xviii
SPECIAL SESSIONS
SS01: Vibration-based monitoring and damage identification for bridges
Organizers: M.G. Limongelli, Y. Fujino & N.F. Catbas
Bridge scour identification based on time-frequency analysis of superstructures 1593
W. Xiong, C.S. Cai, X.T. Hou & X.Y. Gao
Localisation and quantification of stiffness loss based on the forced vibration of a beam
traversed by a quarter-car 1599
K. Feng, A. González & M. Casero
Accuracy of instantaneous frequencies predicted by the Hilbert-Huang transform for a
bridge subjected to a moving vehicle 1607
M. Casero, A. González & E. Covián
Bridge damage detection using acceleration influence line calibrated without access to
a pre-weighed vehicle 1615
E.J. OBrien, D.P. McCrum & M.A. Khan
SMU – an open-source MATLAB package for structural model updating 1621
Y. Otsuki, D. Li, X. Dong & Y. Wang
Operational modal analysis of light pole-viaduct system from video measurements using
phase-based motion magnification 1629
T.J. Saravanan, D.M. Siringoringo, Y. Fujino & S. Wangchuk
Structural model updating based on l2 and l∞ norm regularizations 1638
Z.W. Luo & L. Yu
Uncertainty quantification: Data assimilation of numerical model of the Arade river
cable-stayed bridge 1646
I.C. Santos, D.M. Frangopol, J.L.V. Brito & E.S. Caetano

SS02: Cracking simulation and measurement of concrete structures

Organizers: H. Nakamura & H. Naito
Embedded fiber-optic sensors in reinforced concrete elements of bridge structures 1657
M. Domaneschi, G.P. Cimellaro, F. Ansari, M. Morgese & D. Inaudi
Evaluation of internal cracks in RC beams using vibration testing and wave propagation analysis 1665
H. Naito & J.E. Bolander
Detection method of corrosion area of rebar and corrosion induced internal crack by using
electromagnetic wave radar 1672
Y. Tada, T. Miura & H. Nakamura
Effect of corrosion distribution along rebar length on concrete surface crack development 1680
D. Qiao, N. Kiyama, Z. Amalia, H. Nakamura & T. Miura

SS03: Steel bridge rehabilitation

Organizer: M. Sakano
Research on non-damage repair technology for fatigue crack of steel box girder bridge 1689
J. Zhou & Z. Xu

xix
Verification of effectiveness of fatigue retrofit for rib-to-crossbeam connections in
orthotropic deck bridges 1697
L.H. Ichinose, S. Kunitoshi & M. Sakano
Rapid rehabilitation of deteriorated beam ends with ultra-high performance concrete 1706
K.F. McMullen & A.E. Zaghi
Study on tightening torque of thread forming screws 1714
H. Suzuki
Fatigue strength improvement for weld root of sole plate by filling resin 1719
J. Tamari, T. Ishikawa, M. Hirohata & S. Tsutsumi
Development of retrofitting method for steel bridges under service 1727
S. Maeda, T. Kawahigashi & T. Kasugai
Analytical evaluation of fatigue retrofit for trough-rib to transverse-rib welded joints in
orthotropic steel decks 1732
A. Tanabe, R. Matsumoto, S. Kunitoshi & M. Sakano
Produce of root cracks in transverse fillet welds between sole plate and bottom flange 1740
T. Hirai, L.H. Ichinose & M. Sakano
Fatigue behavior of lower flanges in riveted girders removed from Amarube bridge 1746
Y. Shiraishi, M. Sakano & K. Matsumoto

SS04: Bridge design-maintenance and monitoring practices around the world

Organizers: N.M. Apaydin, A. Chen & N.F. Catbas
Research in bridge maintenance, safety and management: An overview and outlook for
Europe 1755
K. Gkoumas, F.L.M. dos Santos & F. Pekar
The design and application of a wheel-moving fatigue test machine 1762
H. Tian, X.L. Zheng, Z.M. You & A.R. Chen
A review of the durability of joints in concrete structures 1769
Z. Ai, S. Yang, F. Ruan, Y. Shen & G. Li
Long-term monitoring of the Humber Bridge Hessle anchorage chamber 1779
P.R.A. Fidler, P.J. Vardanega, N.A. Hoult & C.R. Middleton
Investigation of magnetic performance of bridge cables for damage detection 1787
Q.W. Zhang, Y.C. Ni & R.Y. Xin
Structure and maintenance design for Lanzhou Chaijiaxia Yellow River Bridge 1793
L. Zhou, J. Peng, Y. Wu, Z. Yin & X. Zhuang

SS05: Soil-steel composite bridges and culverts: Testing and numerical analysis
Organizers: D. Beben, J. Vaslestad, H. Sezen & T. Maleska
Behaviour of the soil-steel bridge with different soil cover height under seismic excitations 1801
T. Maleska & D. Beben
Testing of the longest span soil-steel bridge in Europe – new quality in measurements 1809
M. Miśkiewicz, B. Sobczyk & P. Tysiąc

xx
Analysis of soil-steel bridge with EPS geofoam under static loads 1816
J. Nowacka, D. Beben & T. Maleska
Monitoring of soil-steel structures during construction and exploitation 1824
Cz. Machelski, L. Janusz, A. Wadi & P. Tomala
Estimation of live load deflections for soil-steel composite bridges 1829
A. Wadi & R. Karoumi
Field testing of a corrugated steel culvert at a shallow cover depth 1837
Y. Liu, N.A. Hoult & I.D. Moore

SS06: Numerical simulation technique in life-cycle design and maintenance of bridge

Organizers: Z. Pan & Y. Dong
On the fatigue assessment based on probabilistic fracture mechanics for the rib-to-deck
welded joint of orthotropic steel decks 1847
B. Wang & A. Chen
Continuum damage mechanics-based corrosion-fatigue analysis of high-strength steel wires 1854
C. Cui, R. Ma & A. Chen
Reliability-informed probabilistic analysis of corroded RC structure 1862
H. Guo & Y. Dong
A review of numerical simulation technique in durability issues of cementitious materials 1868
Z. Pan & A. Chen
Performance-based method for the impact-resistant design of reinforced concrete bridge
members 1876
Y. Liu, R. Ma, A. Chen & X. Zhou
Seismic performance of deteriorating concrete bridges 1884
M. Domaneschi, J.R. Casas, A. De Gaetano & G.P. Cimellaro
The mechanical performance analysis of self-anchored suspension bridge with ultra-wide
and double-sided steel box girder under vehicle loads 1893
F.W. Wu, Y.J. Wen, J.F. Luo & J. Dai
Analysis of the service life of reinforced concrete bridges structures under the action of
chloride ions using computational software 1900
A. da C. Filgueira Filho, C.J.G. da Silva, F.C. Zarzar Júnior & R.A. Oliveira

SS07: Approaches to bridge management / Bridge management systems from around the world
Organizers: R.M. Ellis, P.D. Thompson & R. Hajdin
Implementation of bridge management in the Kingdom of Saudi Arabia 1909
R.M. Ellis, K. Galal & A. Almonbhi
Application of bridge maintenance and management system with BIM technology 1917
Z.H. Li & M. Dong
Approach for management of bridge structures in a heterogeneous railway 1923
L.F. Masini, M. Botelli, R. Queiroz, J. Junqueira, V. Barichello & J.F. Rodrigues
Experiences of the new management system of engineering structures in Finland 1930
J. Wuorenjuuri

xxi
Decision support framework for terrestrial transportation infrastructure – Resilience
approach 1935
N. Tanasić & R. Hajdin
Bayesian updating of deterioration models and forecasting capabilities of the bridge
management system in the Province of Prince Edward Island 1942
D.J. Evans & R.M. Ellis
Bridge decommissioning and its impact on bridge management 1947
B.A. Bektas & A.J.M. Albughdadi
Advanced asset management tool for bridges and culverts 1953
H.D. Tran, S. Setunge, E. Amiri, Y.C. Koay & M.I. Alam
Risk management of expressway bridge projects in Vietnam: Current status and future
researches 1962
L.D. Dao, A.D. Le & T.D. Nguyen
Census, diagnosis, supervision, securing and management of bridges 1966
M. Micheloni, M.La Monica & M. Mainetti

SS08: Standardization of quality specifications for highway bridges

Organizers: J.R. Casas & J.C. Matos
Proposed Bridge Management System and Quality Control plan in Chile 1977
M.A. Marquez, M.A. Valenzuela, L. Acuña & P. Valenzuela
Quality specifications for roadway bridges in Europe: Overview of COST action TU1406 1985
J.R. Casas & J.C. Matos
Performance indicators for European bridge management 1993
A. Strauss, F. Sattler, M. Hoffmann, E. Apostolidi, S. Fernandes & J. Matos
Inspection of roadway bridges: A comparison at the European level 2006
Z.I. Turksezer, P.F. Giordano, M.P. Limongelli & C. Iacovino

SS09: Sustainable marine structures

Organizers: M. Iwanami, E. Kato, Y. Kawabata & N.T. Trung
Application of life cycle cost analysis for a maintenance strategy of reinforcement concrete
pile-type wharf in Viet Nam 2017
N.D. Dinh & N.T. Trung
Nanoparticles for nanocomposite coatings for civil infrastructure systems 2024
X. Wang, Z. Lin, F. Tang, Q. Cao & X. Qi
Effects of marine growth on jacket structures for the Vietnamese continental shelf condition 2028
C.Q. Dinh & A.T. Bui
An estimation of total fatigue life of jack-up leg structures induced by wave loading 2034
C.D. Quang & C.D. Vu
Replaceable superstructure system in open-type wharf for reliable maintenance of port
concrete structures 2041
E. Kato, Y. Kawabata, T. Takahashi, Y. Izumida, E. Ooi, K. Kanemaru, K. Miyamaru,
M. Iwanami & H. Yokota

xxii
Research on shear performance of innovative demountable steel-concrete bolt connectors 2047
W. Xue, J. Chen & Y. Xin
Effect of stirrups on concrete crack propagation induced by steel bar corrosion 2052
X.Y. Xu, Y.X. Zhao & J. Xia
Research on overview of structure, evaluation and maintenance solutions for ports in Hai
Phong fairways, Viet Nam 2059
D.T. Nguyen & H.D. Nguyen
Effect of rebar corrosion in concrete on interaction zone under tensile load 2068
A. Kunawisarut, M. Iwanami, N. Chijiwa & K. Nakayama
Proposal of pile connection method for precast superstructure of port pier 2075
K. Ikeno, Y. Kawabata, E. Kato & M. Iwanami

SS10: Life-cycle redundancy, robustness and resilience of bridges and infrastructure networks
under multiple hazards
Organizers: F. Biondini & D.M. Frangopol
Resilience-based seismic risk assessment of aging bridge networks under climate change 2085
L. Capacci & F. Biondini
Risk-based life-cycle analysis of highway bridge networks 2094
G. Fiorillo & M. Ghosn
A procedure for estimating the risk and resilience of bridge networks under both seismic and
tsunami hazards 2101
H. Ishibashi, T. Kojima, M. Akiyama, D.M. Frangopol & S. Koshimura
Interdependency models for resilience analysis of transportation networks 2108
W. Sun, P. Bocchini & B.D. Davison
Exploring the connectivity reliability of large-scale bridge networks based on multilevel
k-way graph partition 2117
J. Wang & S. Li
Probabilistic model for resilience assessment of the Belgian road network 2120
L. Sgambi, T. Jacquin, N. Basso & E. Garavaglia
Bayes method for bridge robustness assessment based on monitoring data 2126
L. Zhang, L. Dong, Y. He, S. Cheng, G. Wang, B. Wang, K. Chen, H. Liu & M. Chen
Structural analysis and robustness assessment of cable-stayed bridges under cable corrosion 2131
L. Rossi & F. Biondini

SS11: Application of vehicle-bridge interaction to bridge maintenance and safety

Organizers: X. Kong, L. Deng, C.S. Cai & Y. Li
Foundation scour detection method based on dynamic responses of bridge due to vehicle braking 2143
T. Yang, Y. Li & X. Lin
Study of the influence of track and wagon defects on the safety of railway bridge structures 2151
J.F. Rodrigues, L.F. Masini, M. Botelli, J. Junqueira, V. Barichello, A. Sisdelli, A. Merheb &
L. Valente

xxiii
Dynamic stress analysis of steel bridge decks using coupled train-track-bridge system model 2158
H.L. Li & L. Zhang
The passenger comfort of human-bus-road coupled vibration due to different sitting positions 2165
J. Zhang, X. Kong & G. Wang
Influence of vehicle-bridge interaction on the accuracy of moving force identification 2173
Z. Chen, Z. Wang & L. Deng
Modal parameter identification for regular bridges using vehicle sensing technique:
Simulation and experiment 2182
X. Jian, L. Sun & Y. Xia
Multiple presence factor for live loads on road-rail bridges 2188
B.R. Dai, Q. Li & D.J. Wu

SS12: Deterioration modelling and structural reliability analysis of bridges

Organizers: M. Mahmoodian & C. Li
Statistical analysis of corrosion-induced area loss of steel bar in concrete 2199
F. Tang, L. Zhao, H. Cui, H.N. Li, H. Wang, X. Wang, Z. Lin & H. Pan
An integrated data-driven approach for deterioration modelling of flexural cracking in
concrete bridges 2205
E. Amiri, S. Setunge, M. Mahmoodian & H.D. Tran
Assessment of cracked concrete bridge girders 2213
H.D. Tran, S. Setunge & M. Mahmoodian
System reliability analysis of girders considering the spatial randomness of corrosion 2220
L. Li, C.Q. Li & M. Mahmoodian

SS13: Constructability difficulties of bridges and solutions

Organizers: U. Attanayake, H. Aktan & M. LaViolette
Effect of temporary bracing system on the overhang deflection in skewed bridges during deck
slab construction 2231
F. Hraib, L. Hui, R. Hindi & M. Vicente
Replacement of Fresh Pond Bridge 2237
A.P. Ranasinghe, M. Feteha & L.N. Weber
Eliminating deck overhang issues during deck placement on steel I-girder bridges 2240
A. Inceefe, U. Attanayake & H. Aktan
Systematic evaluation of steel I-girder bridge constructability 2249
A. Inceefe, U. Attanayake & H. Aktan

SS14: Bridge structural health monitoring for infrastructure management

Organizers: T. Yi & J. Li
Live-load strain evaluation of the concrete box-girder bridges using deep learning 2261
K. Yang, H.W. Zhao, Y.L. Ding & C.F. Wan
Analysis using modal information in SSI by dynamic observability method 2265
T. Peng, J.R. Casas & J. Turmo

xxiv
Fast bridge damage detection based on portable sensing equipment 2273
C.Y. Liu, Y. Xu & K.X. Wang
Deformation features of a long-span arch bridge based on long-term monitoring data 2279
G.D. Zhou & D.K. Liu
Study on the decrement factors in tension behavior of fractured stay cables with grout-filled
protective tubes 2285
K. Ono, K. Osada, M. Hattori & S. Nojima
Structural damage detection based on swarm intelligence technique and hybrid objective
function 2293
Z.H. Ding, J. Li & H. Hao
Damage localization of bridge based on moving sensing-filtering integrated system 2297
Z.H. Nie, Y.K. Xie, Z.F. Shen & H.W. Ma
Time-varying analysis of structural displacement of cable-stayed bridge subjected to thermal
effects 2301
D.H. Yang, T.H. Yi & H.N. Li
Spatial thermal effects for steel bridge girders subjected to time varying ambient temperature 2306
T.H. Yi & G.D. Zhou
A two-stage approach for structural damage detection using MSE and ALO-INM 2311
C.B. Chen, Z.W. Luo & L. Yu
Structural status pre-warning method for operational bridge utilizing single-class support
vector machine 2318
J.Y. Lei, L.F. Hu & Q.S. Xiao
Temperature effect extraction based on variational mode decomposition 2327
W. Lu, Y. Cui, J. Teng, W.H. Hu & Z.H. Li
Using deep learning technique for non-model based vibration response reconstruction 2332
G. Fan, J. Li & H. Hao

SS15: Monitoring techniques and their interpretation for the integrity assessment of bridges
Organizers: T. Shiotani, Y. Yang & E. Lantsoght
Monitoring structural responses during proof load testing of reinforced concrete bridges:
A review 2339
G.I. Zarate Garnica, F. Zhang, Y. Yang, C. van der Veen, E.O.L. Lantsoght, M. Naaktgeboren
& S.A.A.M. Fennis
A study on fabrications and vibration characteristics of steel finger joints simulating damage
stages 2347
H. Iwabuki, A. Yabe, S. Ono & S. Tanaka
Smart Bridge: The duraBASt test bridge equipped with RFID-based sensors 2353
C. Strangfeld, I. Hindersmann & E. Niederleithinger
The application of interferometric radar for measuring lateral vibration of bridges 2359
P. Olaszek
Monitoring techniques for deterioration of concrete bridges due to chloride attack 2367
K. Matsuyama, K. Nakatsui, T. Sonoda, T. Kouchi & M. Nakano

xxv
Improved SIBIE procedure with multi sensor array for 3D visualization of damage in RC
slab 2372
K. Hashimoto, T. Shiotani & M. Ohtsu
Monitoring method for the distribution of prestressing force by optical fiber 2379
K. Okubo, M. Imai, N. Sogabe, S. Yamanobe, M. Oikawa, S. Nakaue, K. Chikiri,
T. Kobayashi & J. Niwa
Evaluation of influence of prestressing on ASR-damaged concrete using NDT 2386
A. Sagradyan & N. Ogura
Acoustic Emission-based crack tracking for existing concrete structures: Influence of number
of load cycles and loading speed 2393
F. Zhang, Y. Yang & M.A.N. Hendriks
Evaluation of concrete strength and defects in concrete by elastic wave methods 2400
T. Watanabe, A. Nouchi, S. Fujimoto & C. Hashimoto
Reliability of a damaged RC slab structure using Model Code 2010 safety formats for NLFEA 2405
A. De Boer, E.O.L. Lantsoght & Y. Yang
A study on monitoring multi-scale concrete members with coda-wave interferometry using
embedded transducers 2413
C. Kevinly, F. Zhang, Y. Yang, D. Draganov & C. Weemstra

SS16: Application of UHPC for bridge structures

Organizers: M. Hosotani, C. Shi, T.B. Viet & H. Yokota
Experimental study on the shear performance of the stiffened web of UHPC box girder 2421
R.S. Pan, C.X. Li & X.D. Shao
The modification of UHPC composite deck at a bridge with orthotropic steel deck 2426
J.J. Shua
Long-term durability of railway structures using ultra high strength fiber reinforced
concrete 2432
T. Kawaguchi, K. Sakashita, H. Musha & H. Yoshimoto
Time-dependent strength gain of a nonproprietary ultrahigh-performance concrete 2439
M.P. Manning, T.S. Alahmari & B.D. Weldon
The study to stress regularities of UHPC prestressed T-beam’s upper plate 2448
Q.L. Xu
The features of various UFC bridges and its durability investigation in Japan 2454
H. Musha, T. Watanabe, O. Hashimoto, Y. Ishii & M. Ikeda
Roadmap for autogenous shrinkage control of UHPC 2463
L. Yang, C. Shi & Z. Wu
Experimental study on precast deck connection using ultra-high-performance concrete 2469
K. Sasaki

SS17: Innovations and advances of composite techniques in bridge maintenance and rehabilitation
Organizers: W. Lin & C. Xu
Application of a composite strengthening technique in steel bridge rehabilitation 2479
W. Lin, N. Taniguchi & T. Yoda

xxvi
Multi-parameters decision-making algorithms for project level bridge maintenance 2486
C.Z. Gui, J.Q. Lei, W.W. Lin, Y. Hou, Z.W. Huang, Z. Duan & Y.H. Zhang
Shear performance of demountable perforated steel-tube connector for accelerated assembly
bridge construction 2498
J. He, C. Li, G. Vasdravellis, E. Feidaki, S. Wang & Y. Liu
Effects of bonding between RC slab and steel upper flange on the elasto-plastic behavior of
steel-concrete composite girders 2506
C. Fang, K. Ono, T. Miyashita, W. Lin, M. Shirato, Y. Sato & H. Tachibana
Four-point bending test of composite girders for limit state design in Japan 2511
Y. Sato, H. Tachibana, T. Miyashita, K. Ono & M. Shirato

SS18: Advances in SHM of bridges and critical infrastructures

Organizers: F. Ansari, N.F. Catbas, G.P. Cimellaro, M. Domaneschi & R. Greco
Investigation of the structural system conversion under transverse wind load based on the
long-term monitoring lateral response of Sutong Bridge 2517
C.X. Ge
Ultrasonic coda wave monitoring of concrete structures: First experiences with large-scale
experiments and real structures 2525
E. Niederleithinger, T. Schumacher, X. Wang, N. Epple, S. Ahmed & P. Klikowicz
Ratio-based features for data-driven bridge monitoring and damage detection 2532
A. Döring, P. Waibel, J. Matthes, L. Bleszynski, O. Scherer, H.B. Keller, S. Keller, J. Müller
& O. Schneider
Artificial intelligent technology applications for bridge scour 2542
Y.B. Lin, X.Q. Liu, K.C. Chang, T.K. Lin, P.S. Yang & C.C.K. Shen

SS19: Sustainability considerations for bridge management

Organizers: K. Sakai, D. Choi & H. Yokota
New requirements to old railway structures challenges engineers in Finland 2549
J. Wuorenjuuri
Status and future of construction industry and road bridges in Japan from the viewpoint of
sustainability 2553
T. Matsuka, H. Yokota, N. Takeshita & K. Sakai
Evolution of bridge construction in Japan 2562
A. Kasuga

SS20: Mechanisms of internal swelling reactions and those effects on structural performances
Organizers: K. Yamada, Y. Takahashi, Y. Kawabata & J. Torrenti
Modeling the effect of fly ash on alkali-silica reaction in concrete considering the reduction
of alkali concentration in pore water 2571
R. Taguchi & Y. Takahashi
Performance based design and maintenance strategy with controlling ASR 2579
K. Yamada, T. Yamamoto, Y. Kawabata, Y. Sagawa, N. Ueda, Y. Kubo & S. Ogawa
Effect of aggregate mineralogical composition on DEF in concrete 2588
M. Malbois, F. Benboudjema, J.M. Torrenti, L. Divet & S. Lavaud

xxvii
Estimation of temperature in the French recommendations for the prevention of disorders
due to delayed ettringite formation 2593
J.M. Torrenti
Mechanisms of internal swelling reactions: Recent advances and future research needs 2599
Y. Kawabata, K. Yamada, S. Ogawa & Y. Sagawa
A method keeping total alkali content and maximizing water supply for concrete prism test
of potential expansion by ASR 2608
K. Yamada, Y. Kawabata, S. Ogawa & Y. Sagawa
Study of expansion and subsequent damage due to ASR and DEF 2613
T. Sriprasong, T. Okubo, N.R. Joshi & S. Asamoto
Poromechanical models for time-dependent mechanical performance of concrete with ASR 2619
Y. Takahashi
Modeling structural effects of DEF: Lessons learned from real-case studies 2628
J.F. Seignol
A study on ASR expansion behavior of concrete exposed to natural environment for 5 years:
Experimental and numerical approaches 2637
T. Kawakami, Y. Sagawa, Y. Kawabata, K. Yamada & S. Ogawa
An experimental discussion on bond strength reduction of ASR damaged concrete 2644
D. Yamamoto, H. Hamada & Y. Sagawa
Research on a new approach assessing ASR of concrete structures for nuclear facilities 2651
J. Eto, S. Ogawa, K. Shibuya, Y. Kawabata, G. Igarashi, A. Teramoto, I. Maruyama &
K. Yamada
Expansion behavior of cement pastes containing additives due to delayed ettringite
formation 2659
H. Takahashi, S. Ogawa, M. Shibata, M. Kuranaga, S. Watanabe, K. Mishiba & Y. Kawabata
Identification of iron sulfide minerals in aggregates by accelerated mortar bar test 2664
W. Saengsoy, L. Yongchaitrakul, P. Sinlapasertsakulwong & S. Tangtermsirikul

SS21: Analysis of fatigue members

Organizers: C. Fujiyama, T. Kisaku, T. Hanji & B. Suryanto
Local stress based fatigue assessment of welded joints in steel-concrete composite slab using
angle-shape shear connectors 2673
T. Hanji, K. Tateishi, M. Shimizu, K. Asano, T. Ishii, K. Kobayashi & D. Uchida
Numerical analysis of RC Gerber bridge girder subjected to fatigue loading 2682
A.I. Quadri & C. Fujiyama
Fatigue life prediction for concrete bridges using Bayesian network 2690
M. Yuan, Y. Liu, D. Yan & L. Huang
Correlation between the fiber content and orientation and the mechanical behaviour of fiber-
reinforced concrete subjected to static and cyclic three point bending test by the use of CT-
Scan technology 2697
D.C. González, A. Mena, J. Mínguez & M.A. Vicente
Influence of bond on the fatigue behavior of reinforced concrete beams without stirrups 2705
B. Suryanto & N. Chijiwa

xxviii
Fatigue crack growth arrester using gourd-shaped-insert-plate for steel bridge deck 2711
T. Murakami, Y. Yamashita, M. Nakatani & Y. Akizuki

SS22: Advances in NDE/NDT for the bridge assessment

Organizers: C.C. Fu, C. Xiang & Y. Gu
Damage identification of simply support beam based on strain energy information entropy 2719
C.S. Xiang, L.Y. Li, C. Dang & Y. Zhou
Damage detection of beam structure using influencing line based on information entropy
fusion 2728
Y. Zhou, S.K. Di, Z. Yuan & C.S. Xiang
Dynamic performance assessment on a MDTA overpass steel bridge with newly constructed
link slabs 2734
Y. Zhu, K. Hou, C.C. Fu & N. Li

SS23: Life-cycle analysis: Probabilistic modeling of the deterioration and recovery of bridges
and transportation infrastructure, and the optimal allocation of resources
Organizers: P. Gardoni, M. Sanchez Silva, M. Pandey & G. Jia
Modelling the interactions between defect mechanisms on metal bridges 2743
G. Calvert, L. Neves, J. Andrews & M. Hamer
Post-event regional seismic risk assessment via vector-IM based record updating 2752
A. Du & J.E. Padgett
A new infrastructure management software for the optimization of road investments 2759
M. Frizzarin & P. Franchetti
Age, state, and environment dependent non-homogeneous stochastic model for improved
bridge deterioration prediction 2766
M. Li & G. Jia
Dimension reduction and surrogate based approach for optimal seismic risk mitigation of
large-scale transportation network 2774
M. Li, Z. Wang & G. Jia
Bayesian updating the resistance estimate of existing aging bridges with service load history 2781
C. Wang & Q.W. Li
Time-varying fragility functions for bridges subject to main shock-aftershock sequences
including damage accumulation during the events and calibration based on available data 2787
L. Iannacone & P. Gardoni

SS24: Rehabilitation and strengthening of concrete and steel structures with UHPFRC
Organizers: T. Matsumoto, E. Brühwiler, A. Miyamoto, K. Rokugo, K. Maruyama &
M. Kunieda
Analytical investigation of structural behaviour of an RC void slab bridge improved with
UHPFRC 2797
T. Makita, H. Kitagawa, S. Kumagai & H. Tatematsu
Cracking resistance of UHPFRC for repair application 2805
M. Kunieda, K. Asai & K. Sasaki

xxix
Applications of UHPFRC for rehabilitation of bridges in severe winter climates 2810
Y. Kosaka, T. Imai, H. Mitamura & T. Matsumoto
Full-scale test for upgrading existing bridges using AFt-UHPFRC 2818
Y. Watanabe, S. Yanai, T. Makita & H. Kitagawa
Bridge enhancement by means of the UHPFRC Technology: Concepts and recent
applications 2827
E. Brühwiler
The seismic safety bearing system for existing concrete bridges with J-THIFCOM 2834
K. Ueda, T. Imai, H. Mitamura & Y. Kosaka
Development of UHPFRC overlay method for RC slabs focusing on interface treatment 2842
T. Watanabe, T. Kanou & M. Ishida
Influence of insufficient early-age strength of UHPFRC on rehabilitation of OSDs 2851
P.R. Deng, T. Matsumoto, H. Kaminishi & Y. Gouda
Finite element analysis on strengthening effect of the UHPFRC-steel composite deck 2858
C.H. Ma, P.R. Deng, T. Matsumoto, K. Ueda & H. Mitamura
Applications of UHPFRC to the rehabilitations and strengthenings to bridges 2865
T. Matsumoto, Y. Gouda & H. Mitamura
Wheel-load-running fatigue test of an UHPFRC-steel composite bridge deck 2871
D. Makino, Y. Gouda, H. Mitamura & S. Matsui

SS25: Protective surface treatment for durability enhancement and service life extension of
concrete bridges
Organizers: J. Dai, P. Zhang & F.H. Wittmann
A practical method for durability design of marine concrete structures treated with silane 2881
Y. Zeng, D.W. Zhang, J.G. Dai, M.S. Fang, W.L. Jin, S. Li & H. Wang
Influence of water repellent surface impregnation of cracked SHCC on steel corrosion 2888
P. Zhang, F.H. Wittmann & T.J. Zhao
Improvements of carbonation and chloride resistance of concrete treated with nano-modified
coatings 2892
C.H. Fan, G. Li & J.C. Zhou
Water penetration into strain hardening cementitious composites before and after water-
repellent surface impregnation 2898
F. Wittmann, P. Zhang & T. Zhao
Experimental study on fiber reinforced repair material for concrete structure 2903
S. Gao & L. Zhang

SS26: Advances in assessment, simulation, and retrofit of corroded bridges

Organizers: D. Lavorato, M.M. Kashani, N. Alexander, C. Nuti & B. Briseghella
Incremental Modal Pushover Analysis (IMPA) for bridges 2911
A.V. Bergami, D. Lavorato, C. Nuti & G. Fiorentino
Shear strength prediction of TCSWs with artificial pitting based on ANN 2920
X. Wei, Z.Y. Wen, L. Xiao & G. Li

xxx
Experimental investigation of corrosion damage on reinforced concrete beams to correlate
crack width and mass loss 2927
H. Nasser, C. Van Steen, R. Vrijdaghs, L. Vandewalle & E. Verstrynge
Test method of simulating erosion and fatigue load of prefabricated bridge joints 2935
J. Zhao, F. Li & Y. Fang
Residual flexural capacity of post-tensioned PC beams having ruptured tendons and the
effect of re-grouting 2944
H. Mutsuyoshi, I.S.K. Wijayawardane & T. Yokota

SS27: ICT and 3D technology for maintenance and management of large-scale structures
Organizers: S. Kanai, A. Takasu & H. Masuda
Robust extraction of steel materials of large structure from point clouds 2953
I. Yoshiuchi & H. Masuda
Deep learning approach to modeling bridge dynamics using cameras and sensors 2961
T. Kawakatsu, K. Aihara, A. Takasu & J. Adachi
A data management platform for efficient monitoring of infrastructures 2970
K. Aihara, A. Takasu, T. Kawakatsu, A. Kinoshita & J. Adachi
Automated correction of the bridge slab-bending effect using terrestrial laser scanning 2978
M. Nakagawa, K. Sasaki, S. Matsuda, H. Ito, Y. Yamaguchi & K. Kurita
Data modeling based on a 3D BIM standard and viewer system for the bridge inspections 2984
F. Tanaka, Y. Nakajima, E. Egusa & M. Onosato
Multi-view stereo reconstruction technique for weakly-textured surfaces 2992
K. Akutsu, S. Kanai, H. Date, Y. Niina & R. Honma

SS28: Risk control, smart operation and effective maintenance of long-span bridges: In
conjunction with TC-18, ACECC
Organizers: H. Kim & H. Katsuchi
Response of suspension bridge in yawed wind 3003
T. Hanai, M. Takeguchi & M. Hongo
Asymptotic formulas for vibration-based cable tension identification accounting for
uncertain boundary conditions 3008
X.L. Le, H. Katsuchi & H. Yamada
Aerodynamic properties of shape modified cables by illumination cables using Large Eddy
Simulation 3016
P.D. Tam, H. Katsuchi & H. Yamada
Seismic response of a cable-stayed bridge subjected to spatially varying orthogonal ground
motions 3024
T. Kim, O.-S. Kwon & J. Song
Loading test and development of structural health monitoring for extradosed bridges 3030
Y.C. Sung, H.H. Hung & C.K. Su

xxxi
SS29: Data informatics for SHM of bridges
Organizers: H. Kim & E. Caetano
Automated operational modal analysis of an end-supported pontoon bridge using
covariance-driven stochastic subspace identification and a density-based hierarchical
clustering algorithm 3041
K.A. Kvåle & O. Øiseth
Automated long-term damping estimation of the cable-stayed bridge using faulty data in
wireless sensor network 3049
S. Kim, B.F. Spencer & H.-K. Kim
Digital twin visualization of beam structure using strain-displacement relationship 3054
M. Han, S. Shin & J.H. Lee
Seismic monitoring of cable-stayed bridge using wireless sensor network 3060
D.M. Siringoringo, Y. Fujino, V. Mehta, Y. Kazui & M. Suzuki

SS30: Latest developments on the seismic response of integral abutment bridges

Organizers: B. Briseghella, S. Mitoulis, C. Nuti, A. Palermo & P.W. Yen
Pseudo-static test on mechanical behavior of pile with pre-hole filled by rubber particles 3071
B. Briseghella, R.H. Fu, J.Q. Xue, Y.B. Lin, F.Y. Huang & C. Nuti
Research on mechanical behaviors of multi-span jointless bridge with link slabs over piers
and abutments 3076
J.Q. Xue, J.H. Lin, B. Briseghella, F.Y. Huang & C. Nuti
Experimental study on SSI of flat buried approach slab in jointless bridge 3082
J.Q. Xue, Y.F. Tang, B. Briseghella, F.Y. Huang, B.C. Chen & C. Nuti
Physical model of an Integral Abutment bridge: Numerical and experimental analyses 3089
G. Fiorentino, D. Lavorato, C. Nuti, F. De Luca, C. Cengiz, A. Sextos, G. Mylonakis &
B. Briseghella
Ductility behavior of skewed integral abutment bridge 3095
Q.H. Zhao, Q.L. Weng & S. Dong
Critical seismic response analysis of skewed integral abutment bridges 3102
Q.H. Zhao, Q.W. Wang & S. Dong

SS31: Fatigue linked retrofitting, life extension and advanced assessment in metallic bridges
Organizers: K. Kinoshita, A. Nussbaumer, M. Chryssanthopoulos, K. Anami & S. Ono
Crack behaviour after High Frequency Mechanical Impact treatment in welded S355
structural steel 3113
H. Al-Karawi, M. Al-Emrani & J. Hedegård
Fatigue crack repair by TIG-remelting 3120
H. Al-Karawi, A. Manai, M. Al-Emrani, R.U. Franz von Bock, N. Friedrich & J. Hedegård
Experimental study on fatigue characteristic of high-strength bolted friction type joints using
corroded weathering steel 3128
K. Inoue, S. Ono & S. Tanaka
Fatigue analysis of riveted connections using the theory of critical distances 3135
B. Imam & H. Gorouhi

xxxii
Giving new life to fatigue life-expired critical details 3143
J.M. Bonnett, R.A. Percy, P.J. Robinson & K. Antoniou
Numerical investigation of rib-to-crossbeam joint in orthotropic steel decks 3151
H. Fang, N. Iqbal, G. Van Staen & H. De Backer
A literature review of pre-fatigued structures treated by TIG dressing 3159
A. Manai
Influence of grid blast on the fatigue strength improvement by peening 3165
Y. Banno, K. Kinoshita, T. Ishikawa & K. Anami

SS32: Performance-based earthquake engineering: Seismic resilience for bridges

Organizers: Q. Wu & Z. Sun
Seismic risk and resilience assessment of composite bridges using hybrid fragility 3177
Y. Liu, Z. Mei, D.G. Lu & F. Paolacci
Plastic and damage behaviour of a masonry arch viaduct to foreshocks and a mainshock
seismic sequence with a Barcelona Material Model 3184
I.J. Drygala & J.M. Dulinska
Dynamic response of a composite bridge to mining tremors from main mining regions in
Poland 3190
I.J. Drygala & J.M. Dulinska

SS33: Data-driven asset management – The Scandinavian way

Organizers: L.F. Pedersen, J.S. Jensen & P. Linneberg
From condition to performance assessment of bridges – The challenge 3201
P. Linneberg
Data-driven asset management by the Swedish Transport Administration 3209
F. Olsson & H. Pétursson
Data-driven asset management on the state roads of Denmark 3215
M. Ebbesen
Using the owners experience through big data & AI to improve new projects and asset
management 3220
L.F. Pedersen

SS34: Design, construction and evaluation of pedestrian bridges

Organizers: E. Hwang, W. Park & U. Starossek
FE modelling of the Streicker footbridge 3229
M.G. Mulas, B. Glisic, M. Domaneschi & F. Venuti
Assessment of the remaining service life of Curado walkway and the bridge over the River
Tanque, Pernambuco, Brazil 3237
C. Gomes, A. Vitório, A. Carneiro & R. Oliveira
Eccentric-wing flutter stabilizer – Effectiveness and costs 3244
U. Starossek

xxxiii
SS35: Monitoring strategies for enhancing transport infrastructure resilience
Organizers: S. Mitoulis, D. Achillopoulou, S. Argyroudis & V. Livina
Data driven condition assessment of railway infrastructure 3251
C. Hoelzl, V. Dertimanis, E. Chatzi, D. Winklehner, S. Züger & A. Oprandi
Laboratory investigation of digital image correlation techniques for structural assessment 3260
M. Domaneschi, G.P. Cimellaro, M. De Iuliis & G.C. Marano
Data-driven resilience assessment for transport infrastructure exposed to multiple hazards 3267
S.A. Argyroudis, D.V. Achillopoulou, V. Livina & S.A. Mitoulis
Resilient monitoring of the structural performance of reinforced concrete bridges using
guided waves 3275
D.V. Achillopoulou, S.A. Mitoulis & N.K. Stamataki
A risk-based taxonomy for bridges at risk of flooding 3282
M. Pregnolato, G. Gavriel & F.D. Lopane
Breakwater inspection system using airborne LiDAR 3286
T. Ueno, T. Emaru, A.A. Ravankar & Y. Kobayashi
SS36: Machine learning-based SHM data analysis and condition assessment for bridges
Organizers: Y. Bao, H. Sun & Z. Chen
Output-only structural modal identification methods based on neural network 3295
Y.Q. Bao, D.W. Liu & H. Li
Automated generation of FE mesh of concrete structures from 3D point cloud using
computer vision technology 3300
J. Shu, K. Zandi & W. Zhao
Seasonal effects on novelty detection using ANNs for SHM 3304
A.C. Neves, I. González, R. Karoumi & J. Leander
Deep convolutional neural network for the analysis of bridge element data 3313
G. Fiorillo & H. Nassif
Pixel-level damage detection for concrete spalling and rebar corrosion based on U-net
semantic segmentation 3319
Y. Xu, W.D. Qiao, Y.Q. Bao, H. Li & Y.F. Zhang
Structural damage detection based on SSA with CI and sensitivity analysis 3327
X.E. Wang, H.L. Liu & L. Yu

SS37: BRIDGE|50 research project: Experimental testing on a 50-year-old PC bridge

Organizers: F. Biondini, S. Manto, C. Beltrami & F. Tondolo
BRIDGE|50 research project: Residual structural performance of a 50-year-old bridge 3337
F. Biondini, S. Manto, C. Beltrami, F. Tondolo, M. Chiara, B. Salza, M. Tizzani, B. Chiaia,
A. Lencioni, L. Panseri & L. Quaranta
Corso Grosseto viaduct: Historical and technical overview 3345
P. Savino, M. Anghileri, M. Chiara, B. Salza & L. Quaranta
Dynamic tests and modal identification of Corso Grosseto viaduct decks before the
dismounting 3352
A. Quattrone, D. Sabia, F. Tondolo, L. Capacci, A. Lencioni & C. Legramandi

xxxiv
Bridge visual inspections: Experience of local authorities and the case study of the Corso
Grosseto viaduct 3358
C. Beltrami, S. Bianchi, M. Cervio, M. Anghileri, R. Felicetti, A. Quattrone, M. Chiara,
B. Salza & D. Masala
Deconstruction of the Corso Grosseto viaduct and setup of a testing site for full scale load
tests 3365
M. Anghileri, F. Biondini, G. Rosati, P. Savino, F. Tondolo, D. Sabia, S. Manto, M. Nivriera,
C. Trincianti, D. Ventura, G. Monti, C. Legramandi & C. Caruso

SS38: Electro-chemical corrosion protection to extend the service life of concrete bridges
Organizers: L. Haixue, R.G. Pillai & C. Xu
Life prediction of Nanjing Yangtze River Bridge double arch approach bridge 3373
C.F. Xu, D.L. Li, M.Y. Ma, S.Y. Cao & J. Guo
Galvanic protection of piles in a marine environment 3379
J.C. Ball
Experimental study on corrosion resistance of steel in simulated hole solution with corrosion
inhibitor and sacrificial anode protection method 3387
C.F. Xu, D.L. Li, H. Yan, J.P. Zhang, S.Y. Cao & Z.W. Lu

SS39: Bridge-vehicle interaction and their applications

Organizers: V. Pakrashi & A. O’Connor
Influence of quasi-static track deformations on the ride comfort of high speed trains 3397
Q. Li, Y. Wu & Q. Wu
Coupled vehicle-bridge-driver modelling of exposed long-span fjord crossings: An initial
study 3406
S. Reymert, A. Rönnquist, O. Øiseth, Ø. Petersen & L. Drugge
Identifying damage in a bridge by analysing rotation response to a moving load 3415
C. McGeown, F. Huseynov, E.J. O’Brien, V. Pakrashi, D. Hester & P. McGetrick

SS40: Design expectations and actual field performance of existing bridges

Organizers: C. Modena, P.G. Malerba & F. Bontempi
Seismic response analysis of long-span rigid frame bridge with high-rising piers subjected to
near-fault earthquake 3425
Y.P. Zeng, K.J. Chen, J. Dong & G.J. Yang
Three-dimensional seismic fragility analysis of long-span railway bridge components with
high pier 3431
J. Dong, G.J. Yang, Y.P. Zeng, L. Pang & L.W. Liu
Comfort performances improvements in existing footbridges by the Tuned-Mass Damper
Inerter (TMDI) 3437
M. De Angelis, F. Petrini & D. Pietrosanti

GENERAL SESSIONS
Wave-based depth evaluation of pile bents in a rail viaduct 3449
H. Wang, C.-H. Hu, S.-H. Hsieh, Y.-C. Tsai, H.-C. Tsai & C.-Y. Wang

xxxv
The development of bridge management system in Indonesia 3455
R.P. Pratama & R. Irawan
Conservation methods for prestressed concrete bridges and composite bridges with concrete
and steel 3463
H. Sakai
Research on seismic performance of suspension bridge with separated girder and tower 3472
L.P. Liu, J.Q. Li, Y.Y. Zhang & H.J. Zhu
Romanian Civil Engineer Anghel Saligny – An example for all civil engineers everywhere 3479
V. Popa
Bridge management: Data collection for bridge upgrades 3487
P.S. McCarten
Study on the influence of cast iron deck joint on asphalt pavement 3495
T. Tomoda, M. Nakano, A. Kiyokawa, M. Murayama, K. Kasahara & N. Fukuzumi
Sensor design and testing for bolt monitoring based on fiber optic technology 3502
H.Y. Wu & R.X. Zhao
New seismic retrofit methodology for wall-type RC bridge piers 3507
K. Kawamura, A. Hata, T. Shindoh, J. Sakamoto & M. Hosotani
Current state and problems of ASR-deteriorated and retrofitted bridges 3515
M. Tsuda, S. Ura, T. Minato & K. Torii
Improvement of inspection and maintenance gantries on the Honshu Shikoku Bridges 3523
K. Imai, K. Endo, T. Matsuo & S. Hirota
Stress reduction at drilled crack-arrest holes by adding splice plates 3531
S. Kiyokawa, K. Tateishi, T. Hanji & M. Shimizu
Mechanical properties of steel-concrete composite girder during fire 3540
Y. Takahashi, Y. Imagawa & O. Ohyama
Vibration-based anomaly detection method for structural health monitoring of in-service
structures under ambient vibration 3547
H.T. Khuyen, T. Mizutani, H. Uchibori & N. Nagamoto
Degradation factors of a metal spraying system for steel bridges 3555
T. Kondo, H. Hotta, H. Matsuno, T. Matsumoto & K. Yamamoto
An efficient approach towards climate change impact on seismic fragility of aging bridges 3565
M. Mortagi & J. Ghosh
Design method using the influence line for scissors-type bridge 3573
K. Chanthamanivong, I. Ario & Y. Hama
Maintenance of HPC pavement deterioration using GPR 3582
S.W. Hong, S.E. Noh, C.W. Lee, D.H. Kim & S.H. Park
Reinforced concrete highway bridge dynamic assessment method using data obtained by live
scale load testing 3588
I. Paeglite, A. Paeglitis & J. Smirnovs

xxxvi
Modular expansion joints for bridges subject to seismic damage potential 3594
S. Hoffmann, N. Meng & P. Savioz
OSIMAB – Online Safety Management System for Bridges 3601
A. Socher & P. Haardt
Research on reinforced concrete t-shaped columns subjected to combined loads 3606
Z. Yu, D. Shan & X. Zhou
Study on seismic behavior of circular steel pier subjected to three dimensional ground
acceleration 3612
K. Wada, M. Mori & H. Shirahata
Discussion on the excessive deflection in mid-spans of large-span prestressed concrete
continuous beam bridges 3618
Q.F. Gao, Q.L. Ma, K. Zhang & C.G. Liu
Stability of a large-span continuous rigid frame bridge with high pier 3625
Q.F. Gao, K. Zhang, Q.L. Ma, Y. Liu & J. Li
Statistical analysis on concrete compressive strength of core specimens extracted from
long-aged concrete beams 3631
S.D. Jo, S.H. Kwon, C.Y. Kim & Y.S. Park
Superior PU-based alternative to the asphaltic plug expansion joint – the “PA joint” in Japan 3637
K. Funabashi, M. Sakano & G. Gallai
Application of friction pendulum isolators to a landmark museum in the Middle East 3644
M. Imam, Z. Zhenghao, P. Savioz & C. Mendez-Galindo
Research on the informatization of bridge asset management and maintenance 3652
M. Dong & Z.H. Li
Fatigue behavior of bearing type bolted joint with pultruded CFRP composites 3657
M. Shimizu, K. Tateishi, T. Hanji & Y. Kitane
Overview of analysis and design requirements for Light Rail Transit 3664
A.P. Ranasinghe & E.G. Honarvar
Development of the image processing method for estimating axle load by use of AI 3668
R. Koshimizu & Y. Sato
Case study on stability assessment of bridges subjected to flood 3673
R. Suzuki & Y. Sato
A posteriori mechanical assessment of Risorgimento type arch bridges 3679
A. Menghini, E. Conti & P.G. Malerba
Viability of off-site inspections to determine bridge defect ratings 3688
D.T. Nepomuceno, P.J. Vardanega, T. Tryfonas, J. Bennetts, S.R. Denton, S. Collard-Jenkins,
C. Thackray, J. Green & M. DiNiro
Rehabilitation of the superstructure of a steel road bridge: Girder inspection, fatigue crack
repair, and repainting for corrosion protection 3696
O. Sanada, H. Imai, T. Yamagishi & M. Hanamoto
Study on the load distribution factor for fatigue evaluation of steel girder bridges 3701
R. Ishikawa, J. Murakoshi & Y. Kishi

xxxvii
Evaluation of post-weld treated steel welds subject to mechanical loading 3708
O.B. Friis, H.B. Blum & H.C. Yıldırım
Automated linking of 3D inspection data for damage analysis 3714
J. Taraben & G. Morgenthal
Investigation of fatigue damage in a RC deck slab due to moving loads through long-term
monitoring of a road viaduct 3721
I. Bayane & E. Brühwiler
Aerodynamic shape optimization of streamlined bridge deck using Space Mapping method 3728
S. Tinmitonde, X.H. He & L. Yan
Change of rigidity and fatigue life of RC beams without stirrups under repeated loading 3735
K. Shibanuma, M. Kurosawa, K. Takeda & Y. Sato
Safety evaluation of a small bridge subjected to compression restraint by landslide 3740
S. Imanishi, Y. Sato, R. Watanabe & Y. Tanaka
Reliability analysis of a post-tensioned bridge using NLFEA and random fields 3747
R.S. Díaz, S.S. Nova, L.M. Trautwein, L.C. de Almeida & M.C.A. Texeira da Silva
Analyzing the effects of shear deformations on the constrained observability method 3755
S. Emadi, J.A. Lozano-Galant & J. Turmo
Application of an objective methodology to analyze bridge performance in Brazil 3763
G.M.B. Warmling, M.L. Santos Filho, R.D. Machado & R. Pieralisi
Load testing of prestressed bridge girders from locally developed UHPC and HPC 3769
T.S. Alahmari, D.V. Jáuregui & B.D. Weldon
Impact of landslide on bridge superstructures 3776
M.A. Canales & J.A. Salazar
Short–medium span bridges based on full-scale experimental verifications for long life
bridges 3783
H.-J. Kim, Y.H. Sung, S.H. Kwon, C.Y. Kim & Y.S. Park
Deep learning with computer vision for conditional assessment of civil systems 3789
H. Pan, Z. Zhang, X. Wang, Z. Lin, Q. Cao, F. Tang & L. Wu
Assessment of different rotation prevention systems for bridge deck construction through
field monitoring 3793
L. Hui, F. Hraib, M. Vicente & R. Hindi
Structural effects of freeze–thaw depth on shear strength of an existing RC bridge pier 3799
T. Kanazawa, Y. Ushiwatari, M. Sakoh & R. Kawase
Estimation of the residual bearing capacity of corrosion damaged bridge beams using 3D
scanning and finite element analysis 3806
A. Hain, T. Zhang & A.E. Zaghi
Impact of climate change on multihazard vulnerability of highway bridges 3814
D.K. Devendiran & S. Banerjee
Efficiency of selected strengthening methods of masonry arch bridges 3820
T. Kamiński

xxxviii
The strength of concrete in historical bridges 3828
A.P. Fantilli & B. Chiaia
Dynamic analysis and vibration reduction control for bridge tower with MTLD under wave
action 3835
S. Chen
An experimental study on the steel girders with a horizontal stiffener in bending 3844
M. Shirato, K. Takahashi, M. Komeda & K. Ono
Experimental study on buckling strength of stub-column with box section made of SBHS700 3852
J. Nie, K. Ono, S. Okada & N. Takeshima
Looking to the future of bridge inspection and management in the UK 3858
J. Bennetts, S.R. Denton, G.T. Webb, D.T. Nepomuceno & P.J. Vardanega
Research on health monitoring parameters and system of small and medium-span bridges 3867
P. Lu, H. Qu, T. Wang & H. Wu
Atmospheric corrosion of weathering steel in Myanmar and its correlation on accelerated
test 3872
Wint Thandar, K. Sugiura, Y. Kitane & Y. Suzuki
Study of temperature effect on different colored steel specimens by solar radiation 3879
R. Sun, K. Chang & K. Sugiura
Fracture mechanics approach to predict the low cycle fatigue life of steel H-piles in integral
bridge 3884
M. Karalar & M. Dicleli
Effect of pile length on the low cycle fatigue performance of integral bridge steel H piles 3888
M. Karalar & M. Dicleli
Open source platforms for monitoring thermal parameters of structures 3892
B. Mobaraki, S. Komarizadehasl, F.J.C. Pascual & J.A.L. Galant
Optimum Span length for a PCI girder expressway bridge 3897
N. Amatya, N. Anwar & F.A. Najam
Experimental study on flexural behavior of UHPC-NC composite beams 3905
C. Liu & Q.X. Sun
Flexural performance of existing bridge footings under seismic loads 3912
Y. Yang, T. Masuda, E. Yoshida, S. Horiuchi & T. Kiriyama
A study about future investment cost of road bridge network in South Korea 3921
J. Lim, S. Park, J.H. Lee, Y. Choi, S. Jin, J. Cho & J. Kong
Guided wave based cable damage detection using magneto-strictive transducer 3925
X.D. Sui, Y.F. Duan, C.B. Yun & Z.F. Tang
Measurement method of coating thickness of slipped high strength bolted frictional joint
considering statistical characteristics 3931
T. Takai & Y. Matsumoto
An innovative concept for strengthening damaged RC beam using prestressed UHPC layer 3940
Y. Zhang, S. Huang, Y. Zhu & C. Zhang

xxxix
Crack sizing accuracy of a phased array ultrasonic scanner developed for inspection of
rib-to-deck welded joints in orthotropic steel bridge decks 3946
M. Hattori, T. Makita, K. Tateishi, T. Hanji, M. Shimizu & N. Yagi
Dynamic analysis of a Langer arch bridge with simulated earthquake ground motions of the
Nankai Trough earthquake 3955
T. Hung, S. Sthapit, K. Ono, S. Kataoka & K. Magoshi
Detectability of delamination in CFRP strengthened members by thermography 3959
J. Gu & S. Unjoh
New bridge seismic isolation design specifications of Turkey 3965
M. Dicleli, A.S. Milani & B. Kurtman
Classification of thermal induced strain cycles and study of associated fatigue damage in
integral bridge steel H-piles 3971
M. Dicleli & M. Karalar
Structural and geotechnical configuration of integral bridges to enhance their seismic
performance 3979
M. Dicleli & S. Erhan
Design of seismic isolated bridges in cold climates; a case study 3987
M. Dicleli & A.S. Milani
Using seismic restrainers with gap to reduce isolator displacements in seismic-isolated
bridges subjected to pulse-type ground motions 3996
M. Dicleli & B. Kurtman
Appropriate selection of isolator properties to enhance the seismic performance of seismic-
isolated bridges in near-fault zones 4002
M. Dicleli & M. Karalar
Possibilities of damage detection in reinforcement of retaining structures 4009
A. Vorwagner, M. Kwapiz, M.J. Rebhan & C. Honeger
Growing loads and aging bridges 4016
H. Raunio
Diagnostic study of Yelcho cable-stayed bridge according to Chilean maintenance program 4021
M.A. Márquez, N. Trotin, O.R. Ramos & M.A. Valenzuela
Load responses analysis of a pedestrian bridge using cellular automata method 4029
Z.R. Jin, X. Ruan & Y. Li
Development of structural deformation measurement system using multiple acceleration
sensors 4038
Y. Umekawa & H. Suganuma
Test method of pull-out behavior of post-installed anchor in cracked concrete 4046
R. Ishihara, M. Takahashi, M. Kunieda & Y. Arakawa
Application and dynamic performance analysis of tubular flange composite girder bridges 4055
X. Wang, Z. Zeng, Y. Zhu, L. Shen & J. Li
Application of u-shaped steel bellows damper as energy absorbers to bridges 4061
A. Arafat, T. Sumimura, M. Matsumura, H. Zui & K. Tanaka

xl
Management program on monitoring Chilean bridges during operational stage 4070
M.A. Márquez, R. Cabezas, F. Espinoza, L. Acuña, M.A. Valenzuela, F. Rodrigues & H. Pinto
Potential environmental impact of using ultra-high performance concrete in simple, two and
three-span continuous prestressed concrete bridges 4076
J. Márquez, D.V. Jauregui, B.D. Weldon & C.M. Newtson
Exploring the transfer of knowledge by Japanese engineers after participation in
infrastructure maintenance training programs 4084
M. Henry, K. Matsumoto, H. Yokota & K. Nagai
Estimation of axle loads by bridge-weigh-in-motion using strain measurement of transverse
stiffeners 4089
E. Yamaguchi, K. Tsuzuki & K. Oda
The Eiffel Bridge in Viana do Castelo 4095
P.J.S. Cruz
Service life considerations due to repairs during bridge construction 4102
V.S. Vemana, Z. Wu, S.D. Koduru & K.K. Vemana
Numerical simulation of dual time-dependent chloride diffusion in concrete with ANSYS 4108
X.W. Zhao, R.C. Xiao & B. Sun
Design of main span 938m cable-stayed bridge for Wuhan Qingshan Yangtze River Bridge 4117
G. Xu, Y. Zhang & H. Hu
Texas bridges deterioration rates 4122
J. Weissmann, A.J. Weissmann & A. Montoya
Restoration of damaged bridges due to the 2016 Kumamoto Earthquake based on lessons
learned from failure mode 4129
F. Miyahara, T. Imamura, H. Nishida & J. Hoshikuma
Corrosion evaluation of steel reinforcement embedded in steam cured high strength concrete
slab with electro-chemical technique 4137
N. Wang, A. Cahyadi & T. Sugiyama
A thickened-edge U-rib that can extend the life of orthotropic steel decks 4144
C.J. Wang, X.Y. Chen, L. Li & Q.G. Ma
Analysis of data for 6,978 bridges to inform a data strategy for predictive maintenance 4151
N.A. Stevens, M. Lydon, S.E. Taylor, G. Hamill, A.H. Marshall, K.E.J Campbell, T. Neeson
& A. O’Connor
Numerical simulation of chloride ingress effect on bridge structures 4159
Y. Li, Y. Wei & X. Ruan
Reliability-based life-cycle maintenance analysis of RC pile caps considering multiple
deterioration modes and social cost 4164
K. Wu, W. Jin, Z. Wang, D.Y. Yang & D.M. Frangopol

Author index 4173

xli
 Bridge Maintenance, Safety, Management, Life-Cycle Sustainability and Innovations –
Yokota & Frangopol (eds)
© 2021 Taylor & Francis Group, London, ISBN 978-0-367-23278-8

Preface

One of the biggest challenges facing bridge asset managers globally is to effectively,
efficiently and safely manage aging and deteriorating bridges while actively considering
sustainability issues, such as environmental changes and natural resource consump­
tions. The dramatic development of information-communication technologies is chan­
ging bridge engineers’ orientation for bridge design, construction and maintenance. In
this context, academics and practitioners are rising to the challenge with research and
practice focusing on innovative approaches to understand the problem and to imple­
ment solutions. IABMAS conferences bring together academic and technological devel­
opments in the fields of bridge maintenance, safety, risk, management, and life-cycle
sustainability, among others. The most recent developments in the field are expected to
be discussed at the 10th International Conference on Bridge Maintenance, Safety and
Management (IABMAS’20), held in Sapporo, Hokkaido, Japan, 11–15 April, 2021
(https://www.iabmas2020.org/).
The First (IABMAS’02), Second (IABMAS’04), Third (IABMAS’06), Fourth (IABMAS’08),
Fifth (IABMAS’10), Sixth (IABMAS’12), Seventh (IABMAS’14), Eighth (IABMAS’16) and
Ninth (IABMAS’18) International Conferences on Bridge Maintenance, Safety and Management
were held in Barcelona, Spain, July 14–17, 2002, Kyoto, Japan, October 18–22, 2004, Porto, Por­
tugal, July 16–19, 2006, Seoul, Korea, July 13–17, 2008, Philadelphia, USA, July 11–15, 2010,
Stresa, Lake Maggiore, Italy, July 8–12, 2012, Shanghai, China, July 7–11, 2014, Foz do Iguaçu,
Brazil, June 26–30, 2016, and Melbourne, Australia, July 9–13, 2018, respectively.
IABMAS 2020 is organized on behalf of the International Association for Bridge Mainten­
ance and Safety (IABMAS) under the auspices of Hokkaido University, Japan with the organ­
izational support of the IABMAS Japan Group. IABMAS encompasses all aspects of bridge
maintenance, safety and management. Specifically, it deals with: bridge repair and rehabilita­
tion issues; bridge management systems; needs of bridge owners; financial planning; whole life
costing and investment of the future; bridge-related safety and risk issues; and economic and
other implications. The objective of the Association is to promote international cooperation
in the fields of bridge maintenance, safety and management for the purpose of enhancing the
welfare of society (www.iabmas.org). The interest of the international bridge community in
the fields covered by IABMAS has been confirmed by the large response to the IABMAS
2020 call for papers. The Conference Secretariat received 854 abstracts, 561 of which were
selected for final publication as full papers and presentation at the Conference within mini­
symposia, special sessions, and general sessions.
Contributions presented at IABMAS 2020 deal with the state of the art as well as emerging
concepts and innovative applications related to the main aspects of maintenance, safety, man­
agement, life-cycle sustainability and technological innovations of bridges. Major topics
include: advanced bridge design, construction and maintenance approaches, safety, reliability
and risk evaluation, life-cycle management, life-cycle sustainability, standardization, analyt­
ical models, bridge management systems, service life prediction, maintenance and manage­
ment strategies, structural health monitoring, non-destructive testing and field testing, safety,
resilience, robustness and redundancy, durability enhancement, repair and rehabilitation,
fatigue and corrosion, extreme loads, and application of information and computer technol­
ogy and artificial intelligence for bridges, among others.

xliii
Bridge Maintenance, Safety, Management, Life-Cycle Sustainability and Innovations contains
571 contributions, comprising the T.Y. Lin Lecture and 9 Keynote Lectures, and 561 technical
papers from all around the world. This volume provides both an up-to-date overview of the
field of bridge engineering and significant contributions to the process of making more
rational decisions on bridge maintenance, safety, management, life-cycle sustainability and
innovations of bridges for the purpose of enhancing the welfare of society. The Editors hope
that these Proceedings will serve as a valuable reference to all concerned with bridge structure
and infrastructure systems, including engineers, researchers, academics and students from all
areas of bridge engineering.

Hiroshi Yokota and Dan M. Frangopol

Chairs, IABMAS 2020
October 2020

xliv
 Bridge Maintenance, Safety, Management, Life-Cycle Sustainability and Innovations –
Yokota & Frangopol (eds)
© 2021 Taylor & Francis Group, London, ISBN 978-0-367-23278-8

Sponsors

ORGANIZING ASSOCIATIONS

IABMAS, International Association for Bridge Mainten­

ance and Safety

IABMAS Japan

SUPPORTING ORGANISATIONS

ATLSS Engineering Research Center, Lehigh University,

Bethlehem, PA, USA

FHWA, The Federal Highway Administration

Hokkaido University

IALCCE, International Association for Life-Cycle Civil

Engineering

JCI, Japan Concrete Institute

JPCI, Japan Prestressed Concrete Institute

JSCE, Japan Society of Civil Engineers

xlv
JSSC, Japanese Society of Steel Construction

RCEAS, P.C. Rossin College of Engineering and Applied

Science, Lehigh University, Bethlehem, PA, USA

TRB, Transportation Research Board of the United States

IABMAS Australia

IABMAS Brazil

IABMAS Chile

IABMAS China

IABMAS Italy

IABMAS Korea

IABMAS Portugal (ASCP)

IABMAS Sri Lanka

IABMAS Turkey

IABMAS USA

xlvi
 Bridge Maintenance, Safety, Management, Life-Cycle Sustainability and Innovations –
Yokota & Frangopol (eds)
© 2021 Taylor & Francis Group, London, ISBN 978-0-367-23278-8

Acknowledgements

The Editors are extremely grateful to all the people who contributed to the scientific program
and organization of the IABMAS 2020 Conference. The Editors would like to express their
sincere thanks to all authors for their contributions, to the members of the International Sci­
entific Committee for their role in ensuring the highest scientific level of the Conference, to
the members of the Local Organizing Committee for their time and efforts dedicated to
making IABMAS 2020 a successful event, and to the National Advisory Committee for their
valuable advice to organizing members.
Moreover, the Editors wish to thank all organizations, institutions, and authorities that
offered their sponsorship. A special acknowledgement has to be given to Hokkaido University
for being the lead host and organizer of the conference. Also, special thanks go to the Inter­
national Association for Bridge Maintenance and Safety (IABMAS) and IABMAS Japan for
endorsing and supporting the conference organization.
The Editors are also extremely thankful to Yozo Fujino, Sreenivas Alampalli, Riadh Al-
Mahaidi, Tulio N. Bittencourt, Joan Ramon Casas, Airong Chen, Jens Sandager Jensen, Ho-
Kyung Kim, Pier Giorgio Malerba and Richard Sause, co-chairs of the International Scien­
tific Committee and Mitsuyoshi Akiyama and Ichiro Iwaki, co-chairs of the National Organ­
izing Committee.
Finally, the Editors wish to express their sincerest appreciation to Yoko Ishizu, Yuuki
Kawada, Masako Morita, Hisashi Miyauchi, Moe Nagata, and all the team at Japan Conven­
tion Services Inc., who professionally managed the Organizing Secretariat with outstanding
expertise, patience, energy and commitment which have been very important for the successful
organization of this Conference.

xlvii
 Bridge Maintenance, Safety, Management, Life-Cycle Sustainability and Innovations –
Yokota & Frangopol (eds)
© 2021 Taylor & Francis Group, London, ISBN 978-0-367-23278-8

IABMAS Executive Board

President of IABMAS:
Dan M. Frangopol USA

Vice-Presidents of IABMAS:
Pier G. Malerba Italy
Eiichi Watanabe Japan

Members:
Fabio Biondini Italy
Tulio N. Bittencourt Brazil
Eugen Brühwiler Switzerland
Airong Chen China
Paulo J. S. Cruz Portugal
Yozo Fujino Japan
Hitoshi Furuta Japan
Jens S. Jensen Denmark
Ho-Kyung Kim Korea
Hyun-Moo Koh Korea
Victor Popa Romania
Man-Chung Tang USA

Secretary General of IABMAS:

Joan R. Casas Spain

Secretary of Executive Committee of IABMAS:

Mitsuyoshi Akiyama Japan

xlix
 Bridge Maintenance, Safety, Management, Life-Cycle Sustainability and Innovations –
Yokota & Frangopol (eds)
© 2021 Taylor & Francis Group, London, ISBN 978-0-367-23278-8

Conference organization

CONFERENCE CHAIRS

Hiroshi Yokota Hokkaido University, Japan

Dan M. Frangopol Lehigh University, USA

INTERNATIONAL SCIENTIFIC COMMITTEE

Chair
Yozo Fujino Josai University, Japan

Co-Chairs
Sreenivas Alampalli New York State Department of Transportation, USA
Riadh Al-Mahaidi Swinburne University of Technology, Australia
Tulio N. Bittencourt University of Sao Paulo, Brazil
Joan R. Casas Technical University of Catalunya. UPC-Barcelona Tech,
Spain
Airong Chen Tongji University, China
Jens Sandager Jensen COWI A/S, Denmark
Ho-Kyung Kim Seoul National University, South Korea
Pier Giorgio Malerba Politecnico di Milano, Italy
Richard Sause Lehigh University, USA

Members
Mitsuyoshi Akiyama Waseda University, Japan
Haluk Aktan Wayne State University, USA
Alfredo H. S. Ang University of California, USA
Nurdan M. Apaydin General Directorate of State Highways, Turkey
André Beck University of São Paulo, Brazil
Jan Bien Wroclaw University of Science and Technology, Poland
Fabio Biondini Politecnico di Milano, Italy
Paolo Bocchini Lehigh University, USA
Eugen Brühwiler EPFL, Switzerland
Alp Caner Middle East Technical University, Turkey
Colin Caprani Monash University, Australia
Necati Catbas University of Central Florida, USA
Baochun Chen Fuzhou University, China
Marios Chryssanthopoulos University of Surrey, UK
Christian Cremona Laboratoire Central des Ponts et Chaussees, France
Paulo Cruz University of Minho, Portugal

li
Sofia Diniz Federal University of Minas Gerais, Brazil
You Dong The Hong Kong Polytechnic University, Hong Kong,
China
Bruce Ellingwood Colorado State University, USA
Reed Ellis Stantec Consulting Ltd., Canada
David De Leon Escobedo Autonomous University of Mexico State, Mexico
Michael Forde The University of Edinburgh, UK
Dan M. Frangopol Lehigh University, USA
Hitoshi Furuta Kansai University, Japan
Yaojun Ge Tongji University, China
Michel Ghosn City College of New York, USA
Nenad Gucunski Rutgers University, USA
Rade Hajdin Infrastructure Management Consultants GmbH,
Switzerland
Rosemarie Helmerich Federal Institute for Materials Research and Testing,
Germany
Riyadh Hindi Saint Louis University, USA
Naeem Hussain ARUP, Hong Kong, China
Kiyohiro Imai Honshu-Shikoku Bridge Expressway Co., Ltd., Japan
Ichiro Iwaki Nihon University, Japan
Bruce Johnson Oregon Department of Transportation, USA
Satoshi Kanai Hokkaido University, Japan
Hiroshi Katsuchi Yokohama National University, Japan
Sang-Hyo Kim Yonsei University, South Korea
Chan Ghee Koh National University of Singapore, Singapore
Jung Sik Kong Korea University, South Korea
Chad Kusko Lehigh University, USA
Hui Li Harbin Institute of Technology, China
Jerome Lynch University of Michigan, USA
Wenliang Lu Beijing Jiaotong University, China
Barney Martin MODJESKI and MASTERS Inc., USA
Takashi Matsumoto Hokkaido University, Japan
Hiroshi Matsuzaki National Defense Academy, Japan
Claudio Modena University of Padova, Italy
Hani Nassif Rutgers, The State Univ of New Jersey, USA
Andrzej Nowak Auburn University, USA
Eugene OBrien University College Dublin, Ireland
Kiyoshi Ono Waseda University, Japan
André D. Orcesi The French Institute of Science and Technology for
Transport, Development and Networks (IFSTTAR),
France
James Ellen Padgett Rice University, USA
Shamim Pakzad Lehigh University, USA
Alessandro Palermo University of Canterbury, New Zealand
Victor Popa The Romanian Academy of Technical Sciences, Romania
Nigel G. Powers Australian Road Research Board, Australia
Xin Ruan Tongji University, China
Samantha Sabatino University of Texas, Arlington, USA
M. Saiid Saiidi University of Nevada, USA
Masahiro Sakano Kansai University, Japan

lii
Soobong Shin Inha University, South Korea
Mohamed Soliman Oklahoma State University, USA
Billie F. Spencer Jr. University of Illinois at Urbana-Champaign, USA
Mark Stewart University of Newcastle, Australia
Alfred Strauss University of Natural Resources and Life Sciences, Austria
Kunitomo Sugiura Kyoto University, Japan
Luc Taerwe Ghent University, Belgium
Man-Chung Tang T. Y. Lin International, USA
Somnuk Tangtermsirikul SIIT, Thammasat University, Thailand
Kazuo Tateishi Nagoya University, Japan
Jin-Guang Teng Hong Kong Polytechnic University, Hong Kong, China
Paul D Thompson Consultant, Bellevue, USA
Yiannis Tsompanakis Technical University of Crete, Greece
José Turmo Universitat Politècnica de Catalunya, Spain
Matías A. Valenzuela Pontificia Universidad Catolica de Valparaiso, Chile
Pedro Vellasco State University of Rio de Janeiro, Brazil
Chunsheng Wang Chang'an University, China
Yiqiang Xiang Zhejiang University, China
Dong Xu Tongji University, China
You-Lin Xu The Hong Kong Polytechnic University, Hong Kong,
China
Eiki Yamaguchi Kyushu Institute of Technology, Japan
Takashi Yamaguchi Osaka City University, Japan
David Y. Yang Portland State University, USA
Hiroshi Yokota Hokkaido University, Japan
Yunfeng Zhang University of Maryland, USA
Jinsong Zhu Tianjin University, China
Aleš Žnidarič Slovenian National Building and Civil Engineering
Institute, Slovenia

NATIONAL ADVISORY COMMITTEE

Toshiro Hayashikawa Hokkaido University, Japan

Akio Kasuga Sumitomo Mitsui Construction Co., Ltd., Japan
Mitsuo Kawatani Kobe University, Japan
Takeshi Kitahara Kanto Gakuin University, Japan
Kyuichi Maruyama Nagaoka University of Technology, Japan
Ayaho Miyamoto Yamaguchi University, Japan
Hiroshi Mutsuyoshi Saitama University, Japan
Junichiro Niwa Tokyo Institute of Technology, Japan
Eiichi Watanabe Kyoto University, Japan
Zhishen Wu Ibaraki University, Japan

NATIONAL ORGANIZING COMMITTEE

Co-Chairs

Hiroshi Yokota Hokkaido University, Japan

Ichiro Iwaki Nihon University, Japan
Mitsuyoshi Akiyama Waseda University, Japan

liii
Members

Shingo Asamoto Saitama University, Japan

Pang-jo Chun The University of Tokyo, Japan
Shuichi Fujikura Utsunomiya University, Japan
Chikako Fujiyama Yokohama National University, Japan
Taku Hanai Honshu-Shikoku Bridge Expressway Co., Ltd., Japan
Atsushi Homma East Nippon Expressway Co., Ltd., Japan
Manabu Hosotani Taisei Corporation, Japan
Tetsuya Ishida The University of Tokyo, Japan
Hironori Ishii Yokogawa Bridge Holdings Corporation, Japan
Mitsuyasu Iwanami Tokyo Institute of Technology, Japan
Yuichiro Kawabata Port and Airport Research Institute, Japan
Koji Kinoshita Gifu University, Japan
Hiroshi Matsuda Nagasaki University, Japan
Takashi Matsumoto Hokkaido University, Japan
Koji Matsumoto Hokkaido University, Japan
Masahide Matsumura Kyoto University, Japan
Hiroshi Matsuzaki National Defense Academy, Japan
Tomohiro Miki Kobe University, Japan
Shinichi Miyazato Kanazawa Institute of Technology, Japan
Kohei Nagai The University of Tokyo, Japan
Tomonori Nagayama The University of Tokyo, Japan
Hikaru Nakamura Nagoya University, Japan
Itaru Nishizaki Public Works Research Institute, Japan
Osamu Ohyama Osaka Institute of Technology, Japan
Shuichi Ono Japan Construction Method and Machinery Research
Institute, Japan
Koji Osada Central Nippon Expressway Co., Ltd., Japan
Yasuhiko Sato Waseda University, Japan
Akira Shiratori Metropolitan Expressway Co., Ltd., Japan
Tomoo Tomoda Nippon Koei Co., Ltd., Japan
Kazuhiro Tsuno Kokushikan University, Japan
Hiroaki Tsuruta Kansai University, Japan
Ken Watanabe Railway Technical Research Institute, Japan

IABMAS 2020 SECRETARIAT

c/o Japan Convention Services, Inc.
Daido Seimei Kasumigaseki Bldg. 14F
1-4-2 Kasumigaseki, Chiyoda-ku, Tokyo 100-0013, Japan
Phone: (+81) 3 3508 1214

CONFERENCE WEBSITE
https://www.iabmas2020.org/

IABMAS WEBSITE
http://www.iabmas.org/

liv
T.Y. LIN LECTURE
 Bridge Maintenance, Safety, Management, Life-Cycle Sustainability and Innovations –
Yokota & Frangopol (eds)
© 2021 Taylor & Francis Group, London, ISBN 978-0-367-23278-8

Durability assessment and re-design of massive concrete structures in

sea-linking projects
Q.K. Su
Hong Kong-Zhuhai-Macao Bridge Authority, People’s Republic of China

K.F. Li & Q.W. Li

Tsinghua University, Beijing, People’s Republic of China

Z.H. Fan
CCCC Fourth Harbor Engineering Institute, Guangzhou, People’s Republic of China

ABSTRACT: This paper presents a method for the durability assessment and re-design of concrete structures in
sea-linking projects, such as the Hong Kong-Zhuhai-Macau Bridge (HZMB) project. In order to assess the durabil­
ity condition of the reinforced concrete structures associated with the chloride-induced corrosion of steel in con­
crete, data sets obtained from construction, managing, monitoring and exposure sites are used to build a durability
assessment model to predict when the depassivation of steel will happen, when necessary repair methods need to
be taken and to put forward an assessment method of the durability of the structures with a failure probability.
With different levels of failure probability obtained, different methods of strengthening and renovation can be
adopted to improve the durability of the concrete structures and make sure the target service life can be achieved.

1 INTRODUCTION redesign of the HZMB project is a systematical method

composed of the durability assessment of each
The Hong Kong-Zhuhai-Macau Bridge (HZMB) sea- reinforced concrete element, updated durability assess­
linking project is composed of massive transportation ments during service time and durability redesign
sub-projects including tunnels, artificial islands and based on the whole life cycle cost (WLCC) analysis.
bridges. The HZMB project with a designed service The durability assessment of each reinforced concrete
life of 120 years is currently the most complicated, element is to make sure that the minimum service time
advanced and sophisticated project in China. The 120 years can be achieved through the proper param­
exposure environment of the HZMB project is very eters and anti-corrosion methods of each element based
aggressive with an average temperature at 22.3-23.1 on the data and tests conducted. The updated durability
ºC, a high humidity and high salty sea water (chloride assessments during service time is to updating the dur­
ions 10700-17020 mg/L and sulfate ions SO42- 1140­ ability model with the data sets from the monitoring
2260 mg/L). It is a challenge for the HZMB project to and exposure tests from time to time during the service
withstand such environments together with the sea time of the HZMB project. The durability redesign
wind, salt fog, tides and wetting/drying cycles to reach based on the WLCC is to redesign the methods for
its target service life 120 years. Similar projects like enhancing the durability of the concrete structures at
HZMB project with different types of sub-projects and the beginning and during the service time through the
such aggressive environments have rarely been analysis of WLCC.
reported elsewhere in the world. The related durability
assessment and re-design methods of this type of mas­
sive projects are very limited. Therefore, establishing 2 DURABILITY ASSESSMENT
the proper scientific durability assessment and redesign
methods for providing the technology support of the 2.1 Background information of the durability
renovation needed for the massive sea-linking projects assessment
like HZMB project is essential and necessary, which
could thus provide good economic and social contribu­ Based on the initial durability analysis of the durability,
tions (Wang et al. 2012). The durability assessment and the critical durability parameter of the HZMB project is

DOI: 10.1201/9780429279119-1
https://doi.org/10.1201/9780429279119-1

3
the chloride-induced corrosion of steel in concrete. Since the Dcl could not decrease continuously during
Under normal condition, the steel in concrete does not the service time 120 years, the maximum t value can
corrode and this is because normal concrete usually has be set as 30 years as shown in Eq. 3.
a high PH and the steel forms a passivation film on the
surface contacting the concrete matrix. However, when
there is a critical content of the chloride ions which
ingress into the concrete and reached to the surface of
the steel, the PH in concrete will change and the depas­
The key parameters in this assessment model are
sivation of the steel will happen, followed with the cor­
Cs, C0, xd, Ccr, DCl° and n
rosion of steel. The local depassivation of the steel has
a potential difference with adjacent parts and
a corrosion cell is formed, and the pit corrosion of steel 2.3 Updated durability assessment model during
happens. Chloride ions accelerate this corrosion pro­ service time
cess and eventually cause the failure of the concrete The statistical properties of the parameters needed
element (DuraCrete 2000). for the model have been obtained during the prelim­
The penetration of chloride ions into concrete is inary stage of the durability design of the HZMB
a complicated process and could be classified as 3 project through the data obtained from long term
different mechanisms (permeation, sorption and dif­ exposure experiments and previous projects in simi­
fusion) due to their different driving forces. In real­ lar exposure environments. The probability-based
ity, the chloride ingress into concrete could be durability assessment model can be established
a combination of the 2 or 3 mechanisms, and it is through the statistical properties of each parameter,
also influenced by the physical and chemical reac­ which will be discussed as following.
tions between the chloride ions and concrete ingredi­
ents. Under the most conditions, diffusion is still the 2.3.1 Thickness of concrete cover xd
most important mechanism that controls the chloride The thickness of concrete cover determines the length
ions into concrete (FIB 2010). The diffusion process of the route that the chloride ions need to travel
of chloride ions in un-cracked concrete can be through in order to reach the surface of the steel in
described as the Fick’s second law (ACI 2000). concrete and it is a critical factor that affecting the
durability of reinforced concrete. Here, the thickness
2.2 Chloride ingress model for HZM project of concrete cover from an immerged immersed tunnel
of the HZMB project is discussed and two groups of
The durability assessment model for the HZMB pro­ data sets from 6386 interior walls and 6571 external
ject is in consistent with the initial durability design walls are analyzed. The distribution of two data
model based on the chloride diffusion model and the groups are shown in Figures 1(a) and (b).
ultimate durability state is the depassivation of the The distribution model of the concrete cover
steel in concrete as shown in Eq. 1 (CCCC & Tshin­ thickness of the HZMB project can be described as
ghua University 2011; DuraCrete 1998; DuraCrete Table 1 based on the onsite measurements.
2000; FIB 2006).
2.3.2 Surface chloride content Cs
The distribution model of the surface chloride con­
tent was obtained from the previous long term
exposure experiments in similar exposure environ­
ments with the consideration of different water-
binder ratios and different binder types. The results
are shown in Table 2.
where Ccr is the threshold chloride content to initi­
ate the steel corrosion (%), C0 is the initial chloride 2.3.3 Initial chloride content in concrete C0
content in concrete at the beginning (%), Cs is the sur­ The distribution model of the initial chloride content
face chloride content of the concrete (%), Dcl is the in concrete is shown in Table 3 based on the test
chloride diffusion coefficient of the concrete (m2/s), data from the raw materials during construction.
xd is the thickness of the concrete cover (m); t is the
exposure time (s). Dcl is a function of time and usu­ 2.3.4 Chloride threshold content Ccr
ally described as Eq. 2. The distribution of the chloride threshold content is
shown in Table 4, which is based on the statistical
model used in the design stage of HZM project.

2.3.5 Chloride diffusion coefficient Dcl

The chloride diffusion coefficient of the concrete
where n is a coefficient for the change of Dcl with was tested through RCM method on the samples
time; Dcl° is the diffusion coefficient of concrete at obtained from the onsite projects at the age of 28
the age of t0 = 28 days; t is the age of concrete. days and the results are shown in Table 5.

4
 Table 3. Distribution model of the initial chloride content
in concrete of HZMB project (uniform distribution).

Reinforced Minimum Maximum

concrete Distribution (%binder) (%binder)

Bridges Uniform 0.02 0.04

Tube tunnel Uniform 0.02 0.04

Table 4. The distribution model of the chloride threshold

content (%Binder).

Exposure
zones Distribution Parameters

Atmosphere Lognormal Mean = 0.85, SD = 0.13

Splash/ Beta Lower bound L = 0.45, upper
Tidal bound U = 1.25,
Immerged Beta Lower bound L = 1.0, upper
bound U = 3.5,

Table 5. The distribution model of the chloride diffusion

coefficients of concrete in HZMB project (10-12 m2/s)
(normal distribution).
Figure 1. Distribution of concrete cover thickness from Reinforced Elements/Concrete
interior walls (a) and exterior walls (b) of an immerged tube concrete grade Parameters
tunnel of HZMB project.
Bridges Bearing platforms/ Mean = 4.37,
C45 SD = 0.88
(28d)
Table 1. The distribution model of the thickness of the Pier/C50 Mean = 4.27,
concrete cover in HZMB project (normal distribution). SD = 0.76
(28d)
Reinforced SD Bridge deck/C60 Mean = 4.48,
concrete Construction Mean (mm) (mm)
SD = 0.92
(28d)
Bridges Cast-in-place Design value 5.3
(Minimum +10) Box girder/C55 Mean = 3.81,
SD = 0.16
Prefabricated Design value 3.5
(28d)
(Minimum +5)
Tube tunnel Prefabricated/C45 Mean = 4.71,
Immersed Prefabricated Design value 73 3.5
tunnel (interior) (Minimum +3) SD = 0.47
(28d)
Prefabricated Design value 93 3.9
(exterior) (Minimum +3) Cast-in-place/C45 Mean = 4.71,
SD = 0.47
Cast-in-place Design value 5.3
(28d)
(Minimum +10)

2.4 Durability assessment in consideration of

anti-corrosion methods
Table 2. The distribution model of the chloride surface con­ In order to enhance the durability of the reinforced
tent of the concrete in HZM project (lognormal distribution). concrete structures, different types of anti-corrosion
methods were adopted in different exposure zones,
Exposure Splash Tidal Immerged Atmosphere such as silane impregnation on concrete surface, use of
types zone zone zone zone
stainless steel and epoxy coating for steel. The results
Mean of A’ 16.47 11.57 13.65 5.99 of failure probability of different elements with the
SD of A’ 2.47 1.77 2.04 0.90 change of service time by considering the different
anti-corrosion methods are shown in Figure 2.

5
 2.5 Updated durability assessment during service
time
The properties and conditions of the reinforced con­
crete elements during service time could be different
from what is expected at the design stage or the ini­
tial stage. Therefore, it is necessary to update the
durability assessment model from time to time
during the service period based on the updated test
results and monitoring results on site. The data
sources for updating the durability assessment model
include routine onsite tests, monitoring tests and the
exposure experiments.
The exposure site was on the harbor of west artifi­
cial island and both of the reinforced concrete elem­
ents and steel structure were used for the exposure
tests. These elements and structures were prepared at
the same time with the construction of HZMB pro­
ject and the same batch of materials was used. The
Figure 2. Failure probability of the bridge elements in reinforced concrete elements included bearing plat­
splash/tidal zones. forms, piers, tubes and box girders of bridges and
tube tunnels. The mechanical and durability of these
elements under different loading conditions and with
In Figure 2, stainless steel and silane treat­ different anti-corrosion treatments were tested rou­
ment were used in all the elements except DB01 tinely and monitored. The exposure tests started in
non-navigable arch bridge and gas pipeline the year 2014 and a large amount of data has been
bridge. For the DB01 non-navigable arch bridge obtained during the last 5 years.
and gas pipeline bridge, double layers of epoxy Sensors were buried during the casting of con­
coating on the steel and silane treatment on the crete elements during the construction stage in order
concrete surface was used. It can be seen from to monitor the physical and durability properties of
the results that the combined treatments with the these elements. Totally, 86 sensors were under ser­
use of stainless steel and silane impregnation on vice in the bridge, piers, bearing platforms and tubes
concrete surface provided a very high reliability to monitor the chloride concentration in concrete,
(β > 5.0) and very low failure probability at the corrosion potential, polarization resistance and cor­
service age of 120 years, which indicates these rosion current of steel. These data are very useful for
structures might not need any repair work judging the current and future durability conditions
during service time. For the treatments of of these elements.
double layers of epoxy coating on the steel and
silane treatment on the concrete surface in the
DB01 non-navigable arch bridge and gas pipe­ 3 DURABILITY REDESIGN
line bridge, the reliability index β > 2.0 and the
failure probability is 1.5% at 120 years, which The target of durability redesign is to keep the reliabil­
indicates there is a possibility that additional ity of the reinforced concrete structures at the age of
repair work might need to be conducted during 120 years constantly above the preset level through
the service time. adjusting the renovation methods based on the updated
Meanwhile, the assessment results from the durability assessment results. The objects of durability
interior part of the bridge pier show that under redesign are the specific elements whose properties
normal working conditions this part belongs to and exposure environments are already known. The
the atmosphere zone, but this part could also durability redesign involves using the latest and
belong to the splash zone in consideration of the updated technologies to enhance and improve the dur­
leakage of water from rainfall and other parts of ability properties of the elements which failed to meet
the bridge. As a result, the assessment of the the specific standard with the increase of time. WLCC
interior part of the bridge pier in this paper is and other reasonable measures are considered and
based on the splash zone, and the resulted β < adopted during the durability redesign.
2.0 and the failure probability is close to 4%,
which indicates that this part should be moni­ 3.1 The critical levels for durability redesign
tored extensively during service time and any
precautious should be taken to prevent any leak­ In the WLCC analysis, the renovation times are
age of water into this part. related to how the durability decreased with time

6
and the acceptable durability levels, and the repair Table 6. Renovation measures and related costs.
cost is related to the specific measures adopted. The
durability assessment model is used to describe how Effective Cost Effect on
the durability changes with time. The acceptable dur­ time (Yuan/ service
ability levels can be described by the reliability Methods (years) m2) condition
index or the failure probability. There are 3 different Electro-chemical - 280 Slight
durability levels corresponding to 3 different failure removal of salt
probabilities： Surface crack-sealing - 50 Slight
Level -1: Threshold chloride content on the sur­ Silane impregnation 20 80-100 Slight
face of steel is reached to initiate the depassivation Polyured waterproofing 20 150 Slight
of the steel in concrete. The steel is not corroded at coating
this stage and there is no effect on the mechanical Surface epoxy coating 15 65-100 Slight
properties of the reinforced concrete elements. The Capillary crystalline - 80-100 Slight
corresponding failure probability is around 1-2% and waterproofing coating
preventative measures should be taken at this stage. Concrete cover replace­ - 300 Obvious
Level -2: Cracks appear on the concrete cover due ment + controlled per­
to the chloride-induced corrosion of steel in con­ meability formwork
crete. With the accumulation of the corrosion prod­ Rust removal and coat­ - 200 Obvious
ucts in the interface between the steel and concrete, ing on steel
the cracks are generated in concrete when the critical Replacement of corroded - 300 Obvious
stress of concrete is reached. At this stage, the mech­ steel
anical properties might not be affected significantly Cathodic protection of 20 900 Obvious
steel (sacrifice of anode)
but the performance of these cracked elements could
Cathodic protection of 50 2000­ Obvious
be affected due to the existence of cracks. The cor­
steel (external current) 3000
responding failure probability at this stage is around
5% and necessary repair measures with a certain
amount of cost should be taken.
Level -3: Severe corrosion of steel has happened summarizes the current mostly used technologies
and there is an obvious loss of steel effective cross and the related costs.
section which can significantly reduce the mechan­ The selection of the repair technologies during the
ical properties of the reinforced concrete elements. durability redesign of the concrete elements is based
The corresponding failure probability at this stage is on the durability assessment results, target durability
around 20% and mandatory renovation measures level and the WLCC analysis.
must be taken. The renovation measures could cost
a large amount of money and significantly affect the
normal working condition and service ability. 4 CONCLUSIONS

The durability assessment and redesign of the sea-

3.2 Whole Life Cycle Cost (WLCC) analysis linking projects, including the durability assessment
The model for WLCC analysis in this study can be of concrete elements, updated durability assessments
described as Eq. 4. The whole cost Ct of a concrete during service time and durability redesign based on
element is composed of construction cost C0, reno­ the whole life cycle cost (WLCC) analysis, are sys­
vation costs (direct cost Cri and indirect cost Cui) tematic, sophisticated and dynamic. Strict and high
and demolition cost Cd at the end of service (Cady standard quality control during the construction
1983). stage is required to minimize the initial defects in the
concrete elements. Through the effective and
updated durability assessment based on exposure
experiments, routine tests and the real-time monitor­
ing, together with the durability redesign and proper
management of the sea-linking projects, the target
service life of these projects can be achieved.

where, m is the required times of repair work, r is

REFERENCES
the discount rate of the cost, T is the service life
time. ACI 365 Committee. 2000. Life-365 Computer program
for predicting the service life and life-cycle costs of
reinforced concrete exposed to chloride, USA.
3.3 Renovation measures and costs
Cady P.D. 1983. Inflation and highway economic
The renovation measures are essential for keeping analysis. Journal of Transportation Engineering, 109
a high reliability of the concrete elements. Table 6 (5): 631–639.

7
CCCC Fourth Harbor Engineering Institute & Tsinghua Fédération International du Béton. 2006. FIB Model Code
University. 2011. Guideline for durability design of con­ for Service Life Design, Bulletin 34, Lausanne.
crete structure of HZMB. Fédération International du Béton. 2010. FIB Model
DuraCrete. 1998. Compliance Testing for Probabilistic Design Code 2010, First complete draft, Bulletin 55/56,
Purposes – Evaluation report, Contract BRPR-CT95-0132, Lausanne.
Project BE95-1347, Document BE95-1347/R7. Wang, S. et al. 2012. Study on durability of concrete struc­
DuraCrete. 2000. Probabilistic performance based durabil­ ture with 120-year service life of Hong Kong-Zhuhai-
ity design of concrete structures. Final Technical Report, Macao Bridge. Proceedings of the eighth national dur­
Contract BRPR-CT95-0132, Project BE95-1347, Docu­ ability of concrete structures.
ment BE95-1347/R17.

8
KEYNOTE LECTURES
 Bridge Maintenance, Safety, Management, Life-Cycle Sustainability and Innovations –
Yokota & Frangopol (eds)
© 2021 Taylor & Francis Group, London, ISBN 978-0-367-23278-8

Digital transformation of bridges inspection, monitoring and maintenance

processes
T.N. Bittencourt, M.M. Futai & A.P. da Conceição Neto
USP – University of São Paulo, São Paulo, Brazil

D.M. Ribeiro
Rio de Janeiro, Brazil

ABSTRACT: New communication and information systems and technologies – known as ICT (Information
and Communication Technologies) – have immense potential to aggregate new functionalities and services to
the management of infrastructure assets. This phenomenon, known as ‘Digital Transformation’, has influenced
the evolution of various sectors of our society, such as the emergence of ‘Industry 4.0ʹ. New wireless communi­
cation technologies, such as 5G networks (large capacity communication, high reliability, great coverage and
low consumption, for information processing and management – such as Cloud Computing, Edge Computing,
Big Data, Machine Learning and Artificial Intelligence – are considered as the enabling technologies of this
digital transformation, integrated with the concept of IoT(Internet of Things). High connectivity capacity and
intensive automation enable, for example, changes in inspection paradigms and asset maintenance by transfer­
ring product focus to service platforms (‘Everything’ as a Service – XaaS), bringing gains for efficiency, prod­
uctivity, comfort and operational safety, as well as cost reduction. Bridges constitute an important part of the
infrastructure and are subjected to damage caused by their continuous use over time. In addition to the effects of
loading (fatigue, impacts, overloads, etc.), they are subjected to degradation of materials and support conditions,
as well as exposure to adverse environmental conditions (storms, floods, gale, earthquakes, etc.). Scheduled
inspections to assess their structural conditions are essential to ensure their proper use within the established
safety limits. In other occasions, continuous or repeated monitoring of structural responses of bridges (displace­
ments, vibrations and rotations at critical points) may add important information for decision-making regarding
its maintenance, repair and reinforcement. The use of these data, together with techniques of structural reliability
for the treatment of the uncertainties, allows predictions about the structural behavior to be elaborated with the
consideration of different loading and degradation scenarios. The new ICTs can greatly contribute to the
improvement of maintenance capacity and, consequently, to the reliability of the assets and to the operational
availability of the system. Thus, the development of new predictive maintenance approaches, which make use
of the large amount of data available, can improve the efficiency of maintenance processes, producing more
accurate and reliable anticipated diagnostics. In this way, Digital Transformation can reduce maintenance costs
(avoiding unnecessary maintenance events) and improve system availability, reducing operational losses. The
use of Big Data Analytics, incorporating Artificial Intelligence and Machine Learning, are innovative solutions
that can be introduced. The adoption of Digital Twins, that incorporate all these tools, can lead to a reduction in
the total cost, allowing predictive and proactive maintenance. The concept of a Digital Twin for a railway bridge
will be illustrated in this paper.

1 INTRODUCTION and also consider structural deterioration (fatigue, cor­

rosion, e.g.), historical and social conditions, inspec-
Bridges are an essential and sensitive part of the infra- tion data to balance operational needs and budget
structure. Bridge integrity management is one of the constraints. Very often, managers and engineers do not
most important issues for infrastructure operators. have enough or appropriate data access to make these
Engineers in charge of making decisions about man- decisions in a quantitatively and data-oriented way.
agement priorities face a daunting challenge: they Therefore, they generally resort to qualitative judg­
must weigh many complex structural and load features, ments to balance the many priorities and decide on

DOI: 10.1201/9780429279119-2
https://doi.org/10.1201/9780429279119-2

11
a course of action. This is a problematic point for man­ Pontem perpetui mansurum in saecula mundi fecit
agers themselves and exposes a systematic risk to divina nobilis arte Lacer … (Permanent bridge will
operators and users. remain forever in the world Lacer (designer) made
An interesting way to go beyond qualitative judg­ by the famous art …)
ment and try to apply quantitative data-driven
approaches is to make use of new technologies that Caius Julius Lacer was apparently right in his
can capture and manage more data on the current state statement, but he completely underestimated deterior­
of assets. ation, hazards and the necessity of maintenance. In
Digital transformation, associated with innovative fact, the Alcantara Bridge had many partial destruc­
computational techniques, is rapidly paving the way in tion-reconstruction cycles along its long life motiv­
this direction, providing the following interesting fea­ ated by wars, invasions and, more recently by floods.
tures: (a) Collection, organization and analysis of The construction of a dam in the area led to a serious
existing information (design, tests, inspections. moni­ flood in 1967, and because of that the base of the col­
toring and other information available); (b) Traditional umns had to be fully repaired and recovered. Since
and noninvasive techniques for inspection with the use then, the bridge has received a lot of attention from
of new technologies. (c) More affordable SHM (Struc­ engineers and new technologies such as TLS (Terres­
tural Health Monitoring) systems and the Internet of trial Laser Scanning) and GPR (Ground-Penetrating
Things (IoT) to generate new monitoring and analysis Radar) have been applied (Figure 2). Through these
possibilities; (d) Implementation and application of techniques, the expected degradation of the stone
automatic methodologies based on performance indi­ blocks and the fillings has been confirmed and may
cators and metrics for detailed classification and diag­ lead to local failure. Recent studies using numerical
nosis of the bridge integrity, including the bridge deck, modeling and limit plastic analysis are then used to
piers, abutments, foundations and bearings (Ghosn access the safety and serviceability conditions of the
et al. 2016); (e) Development of Digital Twins for bridge continuously (Cortés et al. 2018).
bridges and the use of Extended Reality (Zhu et al. The endurance of the Alcantara Bridge is fantastic.
2019); (f) Integration of the collected information. But Lacer´s statement and pride of his beautiful, well
designed and constructed bridge was just a starting
1.1 An example of endurance: The Alcantara point to guarantee its longevity. Inspection, repair,
bridge - Spain retrofit and maintenance activities are essential engin­
eering procedures that are necessary to achieve
The roman bridge of Alcantara in Spain, also known
a considerable life span. In a certain way, all these
as Trajan’s Bridge at Alcantara (Figure 1), is a stone
activities have been applied to this bridge over its long
arch bridge built over the Tagus River between 104
life.
and 106 AD. It was commissioned by the Roman
emperor Trajan in 98 AD. The bridge has a total
length of 180m with 6 arches with spans between 13 1.2 Bridge maintenance modern challenges
and 28m. Despite its age, the bridge is still in oper­
The fundamentals for the design and construction of
ation. Because of recent bridge failures, it is an
roman bridges were established in the second half of
embarrassing confirmation that longevity can poten­
1st century BC by the roman architect Marcus Vitruvius
tially be achieved for bridge structures.
Pollio, known simply as Vitruvius, in his treatise “The
The idea of longevity was so strong and well-
Ten Books on Architecture”, based on the rules of har­
accepted during the construction period of the bridge
mony (Rowland & Howe 1999). The key point of the
that its designer Caius Julius Lacer decided to display
treatise is the Vitruvian Triad: firmitas, utilitas, venus­
a daring statement on plaques at the bridge tower:
tas (strength, utility, beauty). These ideas have been
recuperated by Leonardo da Vinci in the renaissance

Figure 2. Application of nondestructive techniques (Terres­

trial Laser Scanning (TLS) and Ground-Penetrating Radar
(GPR) (Cortés et al. 2018).
Figure 1. Alcantara bridge – Spain – 106AD.

12
and are still the fundamental core of structural and environmental changes or sudden deterioration due
architecture design. The ultimate and service limit state to natural or manmade hazards.
analyses, and the search for comfort and esthetics are The human population increased dramatically over
closely related to the Vitruvian Triad, but not enough to the last two centuries, from 1 billion habitants in 1800
face the new demands of a sustainable planet. to 7,6 billion in 2018. To accommodate the demands of
Durability and sustainability are important issues to its crescent population the development of the infra­
be addressed too. The concepts of structural robust­ structure has been necessary and this process ended up
ness and resilience are also necessary to achieve sus­ producing a gigantic increase in the CO2 emissions
tainability (Biondini & Frangopol 2016, Bocchini (Figure 3). As a direct consequence of the rise of the
et al. 2014), not only for the individual structural sys­ C02 emissions, an increase of the average temperature
tems but for their network as well. Structural robust­ of the planet has been observed (Climate Central
ness can be viewed as the ability of the system to 2019).
suffer an amount of damage not disproportionated Officially the atmospheric CO2 concentration
with respect to the causes of the damage itself reached an alarming 405.51 ppm in September of
(Saydam & Frangopol, 2011), while resilience is the 2018, and there is some nonconfirmed indication
ability of a system or a network of systems to absorb of a value of 415,26 ppm in 2019 (Agence France
or avoid damage without suffering complete failure. It Press 2019). Average global temperature data
is an important new objective of design, maintenance show an increase due to the increase of carbon
and repair of bridges and infrastructure in general, as dioxide concentration in the atmosphere of around
well as communities. Robustness and recovery of the 1°C in 2016 since 1880. Although it seems not to
transportation network depend strongly on the per­ be a considerable increase, there was a shift also
formance of its individual bridges. Bridges are the in the extreme events known as global warming
weakest links in any road transportation system effect (Figure 4). This shift brought more frequent
because of their vulnerability to hazards and deterior­ extreme level events of temperature and humidity.
ation (Akiyama & Frangopol 2018). The environment is becoming more aggressive in
Reliability-based performance indicators have been general and the deterioration processes, already
used recently to account for the uncertainty in resist­ underway, may be potentially accelerated, spe­
ance and loads. Reliability based performance indica­ cially carbonation and corrosion. Moreover,
tors can provide adequate information on the safety
of bridge components and individual bridges, but they
lack the ability to reflect the outcome of a failure
event in terms of economic losses. Therefore, risk
and risk-based performance indicators need to be
introduced to offer additional information on the per­
formance of structural systems under various hazards.
Risk allows combining the probability of component
or system failure with the consequence of the event.
Under the current level of social conscience and
knowledge about infrastructure, maintenance demands
Lacer´s statement sounds very naive for our present
society. The challenges to make our infrastructure
more reliable and resilient are considerable, but they
are fascinating. In this context, the introduction of
new technologies and digital transformation possibil­ Figure 3. Climate change and average temperature rise
ities will make the whole process more accessible and (Climate Central 2018).
comprehensive. Digital Transformation can reduce
maintenance costs (avoiding unnecessary maintenance
events), improve system availability, while reducing
operational losses. The use of Big Data Analytics
techniques, incorporating Artificial Intelligence and
Machine Learning, are innovative solutions that can
be introduced. The adoption of Digital Twins, that
incorporate all these tools, can lead to a reduction in
the total cost, allowing predictive and proactive main­
tenance. The future is there to be created.

1.3 Infrastructure demand and environmental

conditions
It is well known that civil infrastructure is subjected Figure 4. Change in average lead to change in extreme
to progressive deterioration due to aggressive events (Climate Central 2018).

13
climate change may have an important impact on with the following challenges: (a) potentially unsafe
natural hazards like for example floods, storms, bridges that are designed using old design codes; (b)
hurricanes, fires, etc. not only in terms of inten­ bridges with reduced resistance due to deterioration
sity but also in terms of frequency. or mechanical damage; (c) bridges exposed to nat­
Corrosion is the most important deterioration ural hazards that are not adequately considered or
problem of the infrastructure (ASCE Infrastructure not even considered in design; (d) increasing traffic
Report 2017), and it is directly affected by tempera­ volume and axle loads; (e) increased exposure of
ture, CO2 concentration and changes of humidity. To bridges to natural hazards due to climate change.
mitigate the corrosion associated problems in gen­ (Hajdin 2018)
eral, it is estimated that an amount of around US$ 2 Structural safety and serviceability are primary
to 3 trillion is spent annually in the world. In the US concerns in bridge design, as known and practiced
this cost is estimated to be US$ 1.1 trillion (around since the Roman times. In the current inspection pro­
US$ 9 dollar a day per capita). According to the cedure, the most important aspects of bridge service
ASCE Infrastructure Report of 2017, the overall
grade of the infrastructure in the US is D+, with
some improvements with respect to 2013. An
increase from 2.5% to 3.5% of U.S. Gross Domestic
Product (GDP) by 2025 of investments is necessary
to meet future needs and restore the competitiveness
of the country. These numbers demonstrate the
importance and the impact of the correct manage­
ment of the current and future infrastructure.
Fatigue is another important deterioration problem.
With the rise of the population, a strong demand in
terms of load cycles for transportation of people and
goods is applied to our infrastructure. The change in
the loading regime (natural or manmade) is an
important issue to be taken into account as well.

2 THE PROBLEM OF DETERIORATION AND

STRUCTURAL FAILURE IN BRIDGES

Recent bridge failures around the world are an evi­

dent demonstration of the need of a better under­
standing and new approaches to manage our
infrastructure. The sudden collapses of the I-35W
Mississippi River Bridge in Minneapolis and the
Morandi Bridge in Genova, in 2007 and 2018
respectively, brought a lot of attention to the problem
of bridge maintenance and future design provisions.
The losses in terms of human lives in these two
recent bridge failures have been irreparable, but the
direct and indirect economic and social problems
caused by the two collapses are enormous as well.
The conditions of bridges are not very different
around the globe (Fujino 2018, Oliveira et al. 2019).
The lack of a good Quality Control Plan for bridges
affects the everyday life of many people in different
countries. In Brazil, for example, there has been
a sequence of bridge failures in the recent years with
different causes, but the most evident one is poor
maintenance (Figure 5). These failures were not so
catastrophic as the ones observed in the US and
Italy, but they indicate a need of emergency and
urgent actions to avoid or anticipate these situations.
The sudden collapses also highlight the absence of
plans to mitigate the consequences of bridge failures,
especially for large and populated cities.
The bridge manager needs to plan and execute
timely intervention (maintenance) in order to cope Figure 5. Recent and repeated bridge failures in Brazil.

14
life are addressed indirectly by a qualitative measure, Ordinary maintenance work comprises interven­
defined by the condition state, or condition index, tion resulting from the existence of anomalies of
which is based upon observable damages recorded lesser importance, such as cleaning or protection of
during inspections. The condition state, defined by elements of a bridge. On the other hand, repair work
the condition index, is only a qualitative value loosely involves changes in the resistant capacity or struc­
correlated to safety, serviceability, and more recently tural operation, the replacement of resistant elements
durability. In the Brazilian Code, the condition or bearing appliances. They are characterized by
indexes are scaled from 1 to 5, where 1 is the worst greater complexity, the need for design and some­
value and 5 is the best (ABNT-NBR 9452 2016). times the use of heavy equipment.
In general, three types of maintenance schemes
are used: preventive maintenance (expected to delay
3 FUNDAMENTAL CONCEPTS the deteriorating process with no significant gain of
performance), corrective or essential maintenance
The digital transformation of the process of bridge (more expensive, but more effective in restoration of
maintenance requires a clear understanding of some performance), and predictive maintenance (designed
underlying concepts. Some of them are presented to help determine the condition of in-service equip­
briefly in the following subtopics. ment in order to estimate when maintenance should
be performed). This approach promises cost savings
over routine or time-based preventive maintenance
3.1 Bridge inspection
because tasks are performed only when needed. This
The term bridge inspection comprises a set of tech­ approach is heavily based on large amounts of data
nical and specialized procedures that include the col­ and suitable for the introduction of new technologies
lection of data necessary for the formulation of like machine-learning, for example.
a diagnosis and prognosis of the bridge behavior, in
order to maintain or re-establish the requirements of
3.3 Bridge management systems
structural safety, functionality, durability, etc. Bridge
inspection is directly connected with bridge Bridge Management Systems (BMS) can be defined
maintenance. as a process or a set of engineering and management
Depending on the scope and frequency, bridge functions that, if considered simultaneously, compre­
inspection may be classified as: (i) initial or registra­ hend the actions necessary to manage the resources
tion inspection (for new or retrofitted bridges), (ii) to keep a group of bridges in operation.
routine (basic inspection, carried out usually once or BMS terminology is often associated with the
twice a year), (iii) special (or detailed inspection car­ packages of software used for bridge networks man­
ried out at least every five years), and (iv) extraor­ agement. However, the software used for bridge
dinary (carried out by highly qualified experts management is simply BMS tools. In practice, the
according to technical needs after extraordinary or software subsidizes the bridge manager with the
accidental situations (e.g. after floods, earthquake, information necessary for decision making based on
blasts, hurricanes) to indicate rehabilitation or mod­ engineering criteria and knowledge of administrative
ernization needs) (Radomski 2001, ABNT-NBR engineering.
9452 2016). In several countries, the creation of large national
The determination of the optimum inspection databases began, with standardized and uniform
planning and timing is crucial for minimizing the records according to standards and procedures,
life cycle cost while maintaining the performance based on Condition or Performance Indexes. For
levels of the structure. management and support analyses in decision
making, current BMSs use a Life-Cycle Costs Ana­
lysis (LCCA) tool.
3.2 Bridge maintenance
LCCA’s main approach is to consider not only the
Maintenance or conservation of bridges comprises initial costs of bridges, such as projects, construction
actions or strategies that prevent, delay or reduce the and execution, but also long-term costs such as oper­
deterioration of bridges or their elements. It intends ation, maintenance, repairs and rehabilitation. The
to restore the functionality of existing bridges, main­ data resulting from inspections are usually the major
taining bridges in good condition and in order to entries into the analysis modules. These analysis
achieve or increase their service life. modules are divided into three: a cost module,
Proper maintenance requires good management a degradation module, and an optimization module.
procedures in which a network of bridges is cared
for from its conception to the end of its service life.
3.4 Bridge monitoring – Structural Health
The goal is to preserve the bridge functionality at
Monitoring (SHM)
least for the specified design life, along which the
entire structure, or part of it, is intended to be used Structural Health Monitoring (SHM) consists of
with the planned maintenance without the need of verifying the service structural behavior, based on
major repairs. relevant quantities measured by sensors of different
15
technologies. These provide meaningful information technologies, the generation of appropriated signals
regarding the structural conditions, identifying situ­ for the structural behavior diagnosis, signal transmis­
ations that require preventive actions, for instance, sion and processing, identification and interpretation
the decision of either blocking the access to the of occurrences, and implementation of an integrated
bridge/building or restricting its use (Wenzel, operational system to evaluate the structural health.
H. 2009). All these activities can benefit a lot from the new
Sensors collect various types of data, periodically technologies for digital transformation.
or continuously. Once acquired, they are analyzed SHM can be applied in a portable or fixed condi­
and stored for future analyses. The objective is to tion (usually online). A portable scheme is usually
have a better understanding of the structural behav­ applied in short-term controlled tests, like static or
ior, detecting damages in advance and, therefore, dynamic load tests or in evaluations based on non­
leading to a better planned and less expensive destructive or semi-destructive techniques.
management. The use of a short term SHM scheme has been
The integration of structural health monitoring applied to study the retrofit of old steel bridges com­
(SHM), inspection, and updating provide a powerful pany along the Vitoria-Minas Railway (Figure 6).
method to reduce uncertainty, calibrate, and improve Close observations of the responses from monitoring
structural assessment and performance prediction and modeling show that the global behavior of the
models. Life-cycle management is very efficient in structure can be captured very well by the calibrated
providing a practical predictive evaluation of cost, numerical model, but there are clear differences in
safety, and condition, but is not very effective in pro­ the strains in the connections and bearings (Figure
viding actual structural performance. On the other 7). These strain/stress fluctuations are very important
hand, SHM techniques effectively capture structural in the fatigue analysis of the bridge for the estima­
behavior and the demands on a structure but are not tion of its residual life (Ticona Melo et al. 2019).
as effective in translating this information to bridge Online long-term fixed monitoring systems are
managers’ actions. Consequently, it is very important applied usually in important or historical bridges
for the integration of SHM and updating techniques where a close look in the life-cycle process is needed.
in the life-cycle management framework. A system like this has been developed to study the
There are several reasons for the use of SHM: rehabilitation of the Dom Pedro II Bridge (Colombo
et al. 2016). This bridge was built in England in 1870
• Collect data to validate design hypotheses: and was assembled over the Paraguaçu River in 1885
a better understanding of the structural behavior. (Figure 8). Four isostatic puddled iron trusses of
• Search for valuable information to extend the 96 m in length compose the total structure of the
structure remaining service life: early detection of bridge. The deck was originally covered in wood in
damage.
order to allow the passing of not only rail cars but
• Improve structural safety through information
regular road traffic.
based on operating conditions. Therefore, it is
Dom Pedro II Bridge connects the cities of São
possible to detect, in advance, evidence of the
Felix and Cachoeira in the State of Bahia. The traffic
acceleration of deteriorating processes.
is open to road vehicles, pedestrians and trains. In
• Know, in real-time, the structural behavior corres­
2002 this bridge was declared a National Historic
ponding to all stages of constructive or repair
Landmark. The bridge has passed through some ren­
(retrofit) processes.
ovations and some major rehabilitation due to severe
• Improve planning on maintenance and structural
corrosion deterioration. The replacement of some
integrity.
parts of the structure has been extremely sensitive,
• Promote cost-efficiency.
causing some overloading in the remaining parts.
• Structures for which innovative solutions, mater­
Because of that, the whole operation had to be
ials and construction processes are utilized.
SHM is closely associated with and benefits a lot
from structural inspection, non-destructive and semi-
destructive testing, and numerical modeling. The use
of calibrated models constitute a powerful tool for
anticipating or testing imaginary scenarios of loading
or hazards. These systems are used commonly now­
adays to study the behavior of critical structures, for
their economic and social importance, structures
with significant uncertainties regarding geotechnical
conditions, seismic risk, environmental aggression,
weathering condition or vulnerability during con­
struction, existing structures with identified deficien­
cies or which conditions raise doubts. SHM requires Figure 6. Suaçuí river bridge - Warren truss with 41m
the application of sensors and supportive span.

16
Figure 7. Comparison of results from monitoring and numerical analysis.

technologies have shown to be very helpful in data

manipulation and visualization, and alarms set up as
well.

3.5 Reliability
Structural reliability is defined as the probability that
a component or a system will adequately perform its
specified purpose considering a defined period and par­
Figure 8. Dom Pedro II bridge, a four span puddled iron ticular conditions. Either component or system reliabil­
truss bridge – Monitoring system using cloud computing. ity can be computed. A failure of a single component
or a combination of individual components may initi­
ate the failure of the system. Let R and S be the resist­
followed by monitoring (Figure 9). The complete res­
ance and the load effect, respectively, with the
toration and the replacement of the bearings are now
probability density functions (PDFs) fR and fS, charac­
being monitored by this long-term system. A lot of
terizing these respective random variables. The
data is generated as expected. Cloud Computing

Figure 9. Dom Pedro II bridge – Calibrated numerical model response compared to monitoring data.

17
probability that S will not exceed R, P(R > S), repre- expression for total risk can be obtained.
sents the reliability. The time-variant probability of A simplistic approach for calculating instantaneous
failure PF(t) can be expressed in terms of joint PDF of total risk R is (Ellingwood 2005):
the random variables R(t) and S(t), fR,S(t), as:

where Cm represents the consequences of failure,

P(Hi) describes the probability of occurrence of
And the reliability index can be expressed as: a hazard, PF|Hi(t) is the conditional failure probability
given the occurrence of a hazard, and n is the total
number of hazards considered within the analysis.

3.8 Robustness
where Φ-1(·) is the inverse of the standard normal
Structural robustness is understood as the ability of
cumulative distribution function (CDF). In addition to
a structure to withstand natural or manmade hazard
the evaluation of the probability of structural failure at
events like fire, explosions, impact or the consequences
a given point in time, characterizing an ultimate limit
of human error, without being damaged to an extent
state, the consideration of various functionality aspects
disproportionate to the original cause - as defined in
such as serviceability limit states is also possible.
EN 1991-1-7 of the Accidental Actions Eurocode (EN
1991-1-7 Eurocode 1 2006). Usually, three alternative
3.6 Life-cycle cost measures are applied, sometimes simultaneously, to
achieve structural robustness and reduce the risk of dis­
In the evaluation of bridge performance, a crucial per­
proportionate collapse: (a) reducing the possibility of
formance indicator is life-cycle cost. The proper alloca­
occurrence of accidental loading, (b) preventing the
tion of resources can be achieved by minimizing total
propagation of a possible initial failure and increasing
cost while keeping structural safety at a controled
redundancy, (c) designing the structure to withstand
level. The expected total cost CET during the lifetime
accidental loading (load resistance method). Of course,
of a bridge structure is commonly expressed as:
to address these issues more elaborated structural ana­
lysis and load records are necessary.

3.9 Resilience
where CT is the initial cost, CPM is the routine main­ Resilience is the ability to absorb or avoid damage
tenance cost, CINS is the expected cost of inspections, without suffering from complete failure. It is an object­
CREP is the expected cost of repair, and CF is expected ive of design, maintenance and restoration of bridges
failure cost. Optimum planning for life-cycle manage­ and infrastructure in general, as well as communities.
ment of infrastructure assets, including bridges, is Resilience considers the impact of infrastructure
achieved when cost and performance are balanced. damage, failure, and society recovery when subjected
to hazards with a low probability of occurrence and
potentially high consequences. Resilience is closed
3.7 Risk
related to sustainability (Bocchini et al. 2014). Sustain­
Risk is quantified by combining the probability of ability relates to current and future resource manage­
occurrence and the consequences of events generated ment and addresses the impacts of planning and
by hazards. The instantaneous total risk (R) of development on the economy, society, and the environ­
a structural system can be formulated as: ment. Resilience and adaptivity are necessary features
of long-term sustainable infrastructure.
The general approach to quantify the resilience of
an individual structure, a group of structures, or
a network of interrelated structures is to compute the
parameter RE(resilience) as the integration over time
of the functionality:

where κ(x) denotes the consequences associated

with events resulting from hazards and fX(x) is the
joint PDF of the random variables. The integral in
Eq. (4) is difficult to assess and often cannot be
solved. Under certain assumptions, a simpler

18
 in which Q(t) is the functionality, to is the occur­ directed acyclic graphical model, is a probabilistic
rence time of the extreme event, and tr is the investi­ graphical model that represents a set of variables and
gated time horizon. (Frangopol et al. 2016). their conditional dependencies via a directed acyclic
graph. Bayesian networks are ideal for taking an event
that occurred and predicting the likelihood that anyone
3.10 Bayesian update and Bayesian networks
of several possible known causes was the contributing
In the reliability analysis of bridges, the random factor. For example, a Bayesian network could repre­
variables of the degradation model can be obtained sent the probabilistic relationships between diseases
from historical data of similar bridges, located in and symptoms. Given symptoms, the network can be
similar aggressive environment regions. However, used to compute the probabilities of the presence of
this type of approximation can result in very differ­ various diseases.
ent predictions from those found in the real structure. Efficient algorithms can perform inference and
One way to minimize these differences is to obtain learn in Bayesian networks. Bayesian networks that
monitoring data from the structure itself, either model sequences of variables are called dynamic
through visual inspections, structural tests or through Bayesian networks. Generalizations of Bayesian net­
Structural Health Monitoring (SHM) techniques works that can represent and solve decision problems
(which may allow real-time monitoring of the under uncertainty are called influence diagrams.
Structure).
Bayes theorem can be used to update the variables
of the degradation model from the bridge monitoring 4 BRIDGE PERFORMANCE EVALUATION
data. During recent decades, the Bayesian approach
has found great acceptance in the scientific commu­ Under the current level of knowledge, it is much
nity and has been applied in several areas (Colombo, more reasonable to address safety, serviceability and
2016). durability directly in the inspection process, using the
Assuming that a degradation rate for a bridge information on bridge performance (performance
type, located in a given aggressive environment, is indicators) defined during the design and construc­
known and can be approximated depending on θ tion. There are several performance indicators that
variable. If no inspection has been carried out on the can be related to the possible occurrence of local and
bridge, the reliability of the bridge should be esti­ global failures, including system ductility, failure
mated by the θ variable. However, if an inspection is times, redundancy, robustness, and resilience (Bion­
carried out, and the degradation rate obtained dini & Frangopol 2016). These performance indica­
through inspection can be approximated by the vari­ tors need to be included into the practical design
able x, according to the Bayes Theorem, the prob­ codes of bridges (Biondini & Frangopol 2018; Ghosn
ability density function is πðθjxÞ or probability et al. 2016). The performance indicators change over
a posteriori, since the x event was observed, can be time due to structural deterioration and time-variant
written as: resistance. The development of a generalized frame­
work for assessing bridge life-cycle performance and
cost is necessary. This framework must emphasize
analysis, prediction, optimization, and decision
making under uncertainty. A life-cycle approach to
bridge engineering provides a rational basis for
making decisions regarding design, construction,
where p(x|θ) is the conditional probability density inspection, monitoring, maintenance, repair, rehabili­
function of x given θ, π(θ) is the function of distribu­ tation, replacement, and management of bridges
tion of probability a priori, θ is the vector of the vari­ under uncertainty by using multi-objective optimiza­
able a priori and x is the vector of the subsequent tion in order to balance performance and cost, having
observations. risk under control. (Frangopol et al. 2016)
Therefore, from the data obtained in the inspection Different maintenance scenarios have a direct
performed, the reliability of the bridge can be esti­ effect on the performance profile (Figure 10). The
mated. The probability density of the degradation rate probabilistic aspect of the performance prediction (the
πðθjxÞ, given that x was observed in the inspection, reliability index of the performance indicator e.g.) is
enables greater assertiveness in terms of the actual con­ illustrated by the probability density functions (PDFs)
servation status of the bridge. If a new inspection is of the initial performance index, deterioration initi­
carried out, the deterioration rate can be updated again ation, rate of deterioration, and service life with (i)
from the probability density function obtained from without maintenance (or do nothing scenario), (ii)
the first inspection, and so on if further inspections are with preventive maintenance (PM) only, and (iii) with
to be made. The successive repetition of this procedure both preventive and essential (or corrective) mainten­
will result in a better estimate since more and more ance (EM). Preventive maintenance includes actions
monitoring data are being incorporated into the model. of replacing small parts, repairing concrete surfaces,
Bayesian network, also known as belief network, injecting cracks, checking lubricants, and cleaning
decision network, Bayesian model or probabilistic and painting exposed parts of a deteriorating bridge.
19
 indicates clearly the current condition state of the
bridge under investigation, and also provides useful
data that can be used to update the performance pre­
diction models. Within the last decades, multiple
approaches have been proposed for updating the per­
formance prediction models based on inspection out­
comes and/or SHM data. The usual approaches apply
the Bayesian updating of model parameters. In the
updating process, information from inspection or
SHM is used to represent the likelihood function
which can be combined with the prior information on
model parameters to find their posterior distributions
(Frangopol & Soliman 2015). The use of dynamic
Bayesian networks has been recently applied in
fatigue reliability analysis to try to establish optimal
inspection/repair plans with the lowest expected life-
cycle cost (Yang & Frangopol 2018; Heng et al.
Figure 10. Probabilistic structural performance profiles 2019).
considering the effects of deterioration with preventive and There are in general three categories of life-
essential interventions (Frangopol et al. 2016). cycle management techniques applicable to civil
infrastructure: (i) Condition-based Markov deci­
In this sense, the preventive maintenance is expected sion processes, (ii) Reliability and Risk-based
to delay the deteriorating process with no significant management and (iii) Renewal Theory-based life-
gain of performance. However, this gain of perform­ cycle cost models (Yang & Frangopol 2019). The
ance may be important to increase the service life of first has been widely used in the recent years,
the bridge. As can be observed in Figure 10, essential constituting an important tool available in the
(or corrective) maintenance intends to restore per­ main Bridge Management Systems (BMS).
formance when a predefined threshold is reached. The second has been extensively studied in the
These corrective interventions usually are much more recent years. Civil infrastructure is exposed to
expensive, but they have a strong potential to improve both progressive (gradual) deterioration (corrosion,
the performance level to higher values. Typically, EM fatigue, etc.) due to aggressive environmental con­
actions include repair, strengthening and replacement ditions and sudden (shock) deterioration processes
of bridge components. They can be time consuming due to natural or manmade hazards (earthquakes,
and may the availability of the bridge. Figure 10(b) hurricanes, floods, fires, explosions). The capabil­
shows the cumulative maintenance cost as a function ity of the renewal-reward models to deal with pro­
of time for preventive and essential maintenance gressive and sudden and systematic replacement,
interventions and the cumulative effect on the total based on performance thresholds, makes it a very
cost of the bridge management. promising approach for the future.
Another important point to mention is that Future Bridge Management Systems will be able to
approaches for the life cycle management of bridges bridge performance indicators allowing systematic
involving reliability performance indicators usually assessment of safety, serviceability and durability
consider uncertainties associated with loads and based on visual inspection results, non-destructive test­
resistance but are not able to account for the conse­ ing and Structural Health Monitoring (SHM)
quences incurred from bridge failure. Risk-based (Figure 11).
indicators provide the means to combine the prob­
ability of structural failure with the consequences
associated with this event (Ellingwood 2005)
Considerable progress has been made in the topic
of bridge maintenance in recent years (Furuta et al.
2015; Bittencourt et al. 2016; Powers et al. 2018).
But still, some challenges remain in order to try to
take some advantage of new digital technologies, in
order to make the process more automatic and robust.
Establishing the best inspection, maintenance and
retrofit schedules requires a robust optimization pro­
cess which integrates the proper damage occurrence
and propagation models with the previous knowledge
about the safety and financial constraints. In this
regard, information obtained by manual or automated Figure 11. Quality control of bridges (COST TU1406
inspections, non-destructive inspection and SHM 2019).

20
 By including Bridge Digital Models (Hüthwohl advance the operation and maintenance phase to the
et al. 2018; Wikipedia 6D BIM, 2019), the next gener­ end of the bridge lifecycle (demolition and retrofit).
ations of BMS will become a key decision-support It is expected that the application of the Digital Twin
tool not only for maintenance planning but also for technology may offer benefits such as: downtime
managing heavy load permits and risk assessment due reduction, predictive analysis, preventive mainten­
to hazards. Another promising possibility is the devel­ ance, sustainability optimization and transparency.
opment of digital twins of bridges for maintenance The use of BIM or Digital Twins need to be closely
purposes. While Bridge Information Modeling (BIM) associated with the recent developments of bridge
is a process involving the generation and management inspection and maintenance planning (Figure 12).
of digital representations of physical and functional A clear organization of ideas and entities involved
characteristics of a bridge/building, applied in the in the process is highly necessary. An ontology for
design and construction phases and maintained bridge quality control, indicating the relationship
throughout its life cycle, a Digital Twin is a dynamic between the different entities has been recently pro­
virtual representation of an object or physical system posed by COST TU 1406 (COST TU1406. 2016,
throughout its lifecycle, using real-time data to enable 2017, 2019) (Figure 13). Quality Control, as
understanding, learning, and reasoning. A Digital intended, has two meanings: (i) verify, check or
Twin refers to a digital replica of physical assets, pro­ inspect and (ii) command, direct or rule. A quality
cesses and systems that can be used for various pur­ control plan defines the extent and the interval of
poses, like bridge maintenance e.g. (Dang et al. 2018). inspections or investigations and data necessary to
Digital Twins solve the practical problems, gener­ estimate key performance indicators (KPI) and fore­
ating a virtuous cycle with BIM and other technolo­ cast their future evolution.
gies, which are present, but in an integrated way Safety and serviceability are the natural choices
(Shim et al. 2017). Digital Twins use models, scans, of Key Performance Indicators (KPI) for existing
sensors, machine learning, data analytics, artificial bridges. They may also be combined with other indi­
intelligence, etc. It always focuses on offering an cators like durability, stability, costs and functional­
easy navigation and visualization environment, which ity, etc. In COST TU1406, the proposal for KPI
will allow for a better understanding of the built (qualitative, between the scale of 1-5) is defined as:
environment. BIM offers structured and consistent
i. Safety, Reliability and Security (S, R, S) ­
data management in a collaborative environment for
a combined KPI;
Digital Twins technology, which will use this basis to

Figure 12. Long-term maintenance planning (COST TU 1406 2019).

21
Figure 13. Quality control ontology (COST TU 1406
2019).

ii. Availability and Maintainability (A, M) ­ Figure 14. Spider net diagram for bridge assessment
a combined KPI; (COST TU 1406 2019).
iii. Economy $$ (i.e. Costs);
iv. Environment (E)
v. Health and Politics (H, P) - a combined KPI. information from the other dimensions of the Spider-
The current framework considers the following Net diagram.
KPI (Key Performance Indicators) with associated The quality control framework is designed to
definitions: have two stages - static and dynamic. The first one
comprises preparatory work, inspection tasks and
i. Reliability - the probability that bridge will be fit preliminary assessment of the KPIs. The second
for purpose during its service life. It is the com­ mode implies assessment of remaining service life,
plement to the probability of structural failure KPI development over time and finding an optimal
(i.e. safety), operational failure (i.e. serviceabil­ maintenance scenario for decision-making purposes.
ity) or any other failure mode. To predict future performance Monte Carlo simula­
ii. Availability - the proportion of time a system is in tion, degradation and maintenance models may be
a functioning condition. It is not reliability- used in conjunction with multi-objective optimiza­
related disruption of bridge users but originates tion techniques to search for the optimal mainten­
from planned maintenance interventions (e.g. ance schedules in terms of the minimization of
additional travel time due to an imposed traffic degradation and maintenance costs (Denysiuk et al.
regime on bridge). 2016).
iii. Safety - related to minimizing or eliminating the General frameworks for life cycle analysis and
harm to people during the service life of optimization considering uncertainty, generate
a bridge. The loss of life and limb due to struc­ a powerful tool for assessing bridge conditions in
tural failure is not included (see Reliability). a systematic way and should be incorporated in new
iv. Economy - related to minimizing long-term costs systems for bridge management, taking advantage of
and maintenance activities over the service life of the new digital technologies available.
a bridge. Herein the user costs incurred due to
detours and delays are not included.
v. Environment - associated with minimizing the 5 DIGITAL TRANSFORMATION
harm to environment during the service life of
a bridge. Digital Transformation is a process in which com­
The bridge performance is represented by panies make use of technology to improve perform­
a ‘Spider Net’ diagram (Figure 14). The larger the ance, increase range and ensure better results. It is
area in the diagram enclosed by the KPI values, the a structural change in organizations where technol­
better is the bridge performance and KPIs values in ogy has an essential role. At the present time, along
the green area are desirable. This format is supposed with the wave of Industry 4.0, the rapid development
to be applicable for a single or a group of bridges. and advances in digital technologies are rapidly
The key-performance indicators are correlated to transforming the traditional manufacturing industry.
established performance indicators associated to fail­ Digital representation provides the elements and
ure modes in the so-called vulnerable zones. In dynamics of how an Internet of Things device oper­
COST TU1406, all KPI are scaled from 1-5, where 1 ates and lives throughout its lifecycle.
is the best value and 5 is the worst. Notice that the Artificial intelligence, advanced prediction tools,
qualitative value of the KPI for the Safety, Reliability cloud and edge computing, big data analytics, new
and Security dimension may be associated with the generations of robots and sensors can be all inte­
usual Condition Index (for the Brazilian Code is just grated to create a new environment where the assets,
a question of inverting the scale). The difference lies represented by their digital twins, can interact,
in the fact that the KPI is obtained from a computed denominated the Internet of Things (Figure 15). It is
PI (Performance Indicator), in addition to the expected that all this process can bring simplicity to

22
 life of its corresponding twin”, as defined by Glaess­
gen & Stargel (2012). A Digital Twin consists of
three components: a physical part, a virtual part, and
the connections between the physical and virtual
part, and has three basic characteristics:
• Real-time reflection: both physical and virtual
parts exist in Digital Twin, the virtual part can
keep ultrahigh synchronization and fidelity to
reflect its physical counterpart.
• Interaction and convergence: it happens in both
Figure 15. New technologies applicable to bridge inspec­ physical part, virtual part and between physical
tion, monitoring and maintenance. and virtual part. Also, the interaction and conver­
gence between real-time data and historical data
can make the Digital Twin data more comprehen­
complex situations, like the maintenance of bridges, sible and useful.
following Leonardo Da Vinci’s statement: “Simpli­ • Self-evolution: a Digital Twin can collect and
city is the ultimate sophistication”. update data in real-time, the virtual part can be
continuously self-improved by comparing the vir­
tual part with the physical part.
5.1 Internet of things
The connection between the real and the virtual
The Internet of Things (IoT) refers to an ongoing worlds is established in real-time through data
trend of connecting all kinds of physical objects to obtained through sensors. A Digital Twin provides
the internet, especially ones that one might not interesting features in production and design, remote
expect. This can mean everything from common diagnostics and service. A Digital Twin relies on the
house objects like refrigerators and lightbulbs, to continuously accumulated data and real-time presen­
business assets like shipping labels and medical tation of the collected data to simultaneously update
devices, to unprecedented wearables, smart devices, and modify its physical counterpart. On the other
infrastructure assets and even smart cities. A typical hand, the manipulation and interpretation of the
IoT system, like a smart bridge, works by continu­ huge amount of collected data and information in an
ously sending, receiving, and analyzing data in intuitive manner remains a big challenge.
a feedback loop. Depending on the kind of IoT Extended reality (XR) is very helpful in this
system, analysis can be conducted either by humans aspect because it allows the user to visualize and
or artificial intelligence and machine learning (AI/ interact with Digital Twin data by integrating graph­
ML), in near real-time or over a longer period. In ics, audios and real-world objects. (Zhu et al. 2019).
order to predict the optimal time to inspect or main­ XR has been recently implemented in various areas
tain a bridge, the IoT system may connect to the of the industry from product design, to data manage­
Google Maps API (Application Programming Inter­ ment, assembly instructions, and maintenance.
face) data about real-time traffic patterns in a region, The Digital Twin of a bridge is a virtual and real-
as well as utilize long-term data for deterioration time representation of its structural system and com­
processes collected by different types of smart sen­ ponents. Therefore, it is capable to represent distinct
sors or robots. IoT data collected from every smart aspects of performance, environment, geometry evo­
sensor can be analyzed by bridge owners in larger- lution, anomalies detection, maintenance and condi­
scale optimization efforts. The Internet of Things tion states based on the continuously collected data,
(IoT) is a dynamic global information network con­ updates and changes from its physical counterpart.
sisting of Internet-connected objects, such as sen­ All this data evolves with time (Figure 16), enabling
sors, actuators, as well as other instruments and the Digital Twin to continuously evaluate its struc­
appliances that comprise the future Internet. tural integrity risk (Faridafshin et al. 2019).
(Figure 17).
Reliable assessment of the bridge requires well-
5.2 Digital twins
organized life-cycle information from design to
A digital twin is a digital replica of a living or unliv­ operation. 3D digital twin bridge models are natural
ing physical entity. By uniting the physical world candidates for the next generation of bridge mainten­
and the virtual world, the data is transmitted trans­ ance systems. 3D geometry models can be generated
parently, allowing the virtual entity to exist simultan­ by the combination of 3D scanning and wire-based
eously with the physical entity. parametric modeling. Damage anomalies are linked
More precisely, “A Digital Twin is an integrated to the elements of the bridge. Environmental condi­
multi-physics, multi-scale, probabilistic simulation tions including temperature, humidity, loading his­
of a product or system that uses the best available tory, non-destructive and monitoring data provide
physical models, sensor updates, etc., to mirror the essential information for future performance

23
 process from a remote workstation. (Dang et al.
2018, Hüthwohl & Brilakis 2018, Hüthwohl et al.
2018). All these features can be naturally incorpor­
ated in the Digital Twin of an existing bridge.

5.3 Extended reality (XR)

Extended reality (XR) is one of the critical technolo­
gies in Industry 4.0. It has been widely used in the
manufacturing environment. XR technology usually
has the following three characteristics: combination
of the real and virtual worlds, interaction in real-time,
and registration in the three dimensional space. It
also has at least six classes of potential applications,
including medical visualization, entertainment, manu­
facturing and maintenance, process planning, and
military aircraft navigation and targeting.
XR is usually divided into three classes: Virtual
Reality (VR), Augmented Reality (AR) and Mixed
Figure 16. Digital twin capabilities. Reality (MR). While, Virtual Reality (VR) is a 100%
digital content and can be enjoyed in a fully immer­
sive environment (Videogames, Simulators, etc.),
Augmented Reality (AR) superimposes digital con­
tent on the real world. In addition, Mixed Reality
(MR) is a digital overlay that allows interactive vir­
tual elements to integrate and interact with the real-
world environment.
XR, mainly in the AR and MR forms, allows the
user to see the physical world, with virtual objects
overlaid onto the physical world, rather than com­
pletely replacing it. It also enables the users to inter­
act with the physical world, where the information
by the virtual objects help them to conveyed perform
tasks in the physical world. It may include multiple
sensory modalities, such as visual, auditory, haptic,
somatosensory, and olfactory. The development of
Figure 17. Risk-based structural integrity management new XR features and new devices for visualization
(Faridafshin et al. 2019). (from Cell Phones to high-tech display devices, like
the Microsoft Hololens) is rapidly bringing new pos­
sibilities to extend the capabilities of the bridge
prediction of a structural member. The bridge digital inspector and bridge manager (Figure 18).
twin is then updated by each inspection or mainten­ The ability to visualize hidden parts of the struc­
ance record and by non-destructive or monitoring ture and associate the images with anomalies and the
data. The digital twin model allows big data analysis degree of degradation allows more precise diagnos­
and artificial intelligence applications for more reli­ tics of the structural integrity of the bridge. In add­
able prognosis of future performance of the bridge. ition, a sequence of more precise diagnostics will
Automated systems for bridge maintenance have
always been a difficult problem for bridge man­
agers. 3D digital BIM technologies are widely
used now for design and construction. They can
be associated naturally to the digital twin, where
more efficient ways to handle data is developed to
avoid all the complexity of the parametric models
and their interaction. Adapting machine vision
concepts, it is possible to detect the structural
anomalies in real-time. Inspection data can be gen­
erated by humans, by mobile devices, by drones,
or other wearable/XR devices. Post-processing of
the data to determine the extension of the anomal­
ies provides the support for the decision-making Figure 18. XR applications to structural integrity inspection
and maintenance.

24
lead to more reliable prognostics of the bridge per­ 6.1.1 Machine learning (ML)
formance, opening promising possibilities for pre­ Machine learning is an important area of artificial intel­
dictive maintenance. ligence that deals with the development of algorithms
to identify complex patterns from experimental data,
without assuming a pre-established equation as
6 UNDERLYING NEW TECHNOLOGIES a model, and can provide more accurate results, thus
enabling decision making intelligently, in real-time,
Digital Transformation has become a reality with the without human intervention. (Taffese & Sistonen
introduction of some underlying technologies. The 2017).
whole idea of Digital Transformation is to orchestrate Machine learning algorithms adopt various methods
these technologies to make the design, the construc­ from multiple fields, for example, pattern recognition,
tion and the maintenance processes of bridges more data mining, statistics, and signal processing. This
comprehensible and more natural to apply. allows machine learning to take advantage of the syn­
Among these emerging technologies, Artificial ergy of all these fields, which in turn leads to robust
Intelligence, Cloud and Edge Computing, Big Data, solutions that use different knowledge domains
Computer Vision and the new generation of robots and (Figure 19).
sensors are worth mentioning and presenting a brief Machine learning can be classified as supervised,
definition. unsupervised, and semi-supervised learning. In gen­
eral, machine learning algorithms to detect structural
damage are classified as supervised learning, as data
6.1 Artificial intelligence (AI)
on the conditions of structures and variables that can
In computer science, artificial intelligence (AI), influence such a condition, usually associated with
sometimes called machine intelligence, is the intelli­ degradation mechanisms, are needed. Machine learn­
gence demonstrated by machines, in contrast to the ing algorithms can be used in modeling degradation
natural intelligence displayed by humans. Artificial processes (corrosion, carbornatation, fatigue, etc.)
intelligence is often used to describe machines (or and are a powerful prediction tool for the prognostic
computers) that mimic “cognitive” functions that process of structural behavior required for predictive
humans associate with the human mind, such as maintenance of bridges and other structures.
“learning” and “problem-solving”. The traditional
problems (or goals) of AI research include reasoning,
6.2 Cloud and edge computing
knowledge representation, planning, learning, natural
language processing, perception and the ability to Cloud Computing is the on-demand availability of
move and manipulate objects. General intelligence is computer system resources, especially data storage
among the field’s long-term goals. Approaches and computing power, without direct active manage­
include statistical methods, computational intelli­ ment by the user. The term is used to describe data
gence, and traditional symbolic AI. Many tools are centers available to many users over the Internet.
used in AI, including versions of search and mathem­ Large clouds, predominant today, often have func­
atical optimization, artificial neural networks, and tions distributed over multiple locations from central
methods based on statistics, probability and econom­ servers. If the connection to the user is relatively
ics. The AI field draws upon computer science, infor­ close, it may be designated an edge server.
mation engineering, mathematics, psychology, In contrast, Edge Computing is computing that
linguistics, philosophy, and many other fields. AI is takes place at or near the physical location of either
proving to be an efficient alternative approach to the user or the source of the data. By placing com­
classical modeling techniques. AI refers to the branch puting services closer to these locations, users bene­
of computer science that develops machines and soft­ fit from faster, more reliable services while
ware with human-like intelligence. Compared to trad­
itional methods, AI has advantages to deal with
problems associated with uncertainties. In addition,
when testing is not possible, AI-based solutions offer
alternatives to determine design parameters, resulting
in significant savings in terms of human time and
experiments. AI makes the process of decision
making faster, decreases error rates, and increases
computational efficiency. Different AI techniques
such as machine learning (ML), pattern recognition
(PR), and deep learning (DL) have been recently
applied in structural engineering problems (Yan et al.
2019, Salehi & Burgueño 2018) and are establishing
themselves as a new class of intelligent methods for Figure 19. Machine learning algorithms. (Taffese & Sistonen
use in structural inspection and maintenance. 2017).

25
companies benefit from the flexibility of hybrid within the context of monitoring structures due to
cloud computing. Edge computing is one way that the possibility of contactless measurements, long
a company can use and distribute a common pool of distance, and multiple points measurements, associ­
resources across a large number of locations. ated with high precision, good cost-benefit and time
reduction.
Generally, the monitoring process complies with
6.3 Big data
the following requirements (Dong et al., 2018): (i)
Big data is a field that studies ways to analyze, sys­ capture of the structure image and determination of
tematically extract information from, or otherwise the scale. Usually, the process is performed from
deal with data sets that are too large or complex to a marker (element with known shapes and dimen­
be dealt with by traditional data-processing applica­ sions); (ii) identification of regions of interest in data
tion software. Data with many cases offer greater collection, through image processing; (iii) determin­
statistical power, while data with higher complexity ation of image characteristics that can be compared
(more attributes) may lead to a higher false discov­ frame by frame; (iv) introduction of corrections to
ery rate. Big data challenges include capturing data, the image, if necessary; (v) calculation of frame-to­
data storage, data analysis, search, sharing, transfer, frame offsets and conversion of measurements to
visualization, querying, updating, information priv­ millimeters by using a target as a reference.
acy and data source. Big data was originally associ­
ated with three key concepts: volume, variety, and
6.5 Robots and sensors
velocity. When handle big data is handled, one may
not sample, but simply observe and track what hap­ The inspection, monitoring and maintenance of
pens. As a consequence, big data usually includes bridges take a lot of advantage from the new gener­
data with sizes that exceed the capacity of traditional ation of robots and sensors.
software to process it within an acceptable time and An Unmanned Aerial Vehicles (UAVs), or
value. Big data analytics comprises the complex pro­ a drone, is essentially a flying robot that can be
cess of examining large and varied data sets to remotely controlled or fly autonomously through
uncover information such as hidden patterns and software-controlled flight plans in their embedded
unknown correlations that can help organizations systems. They work in conjunction with onboard
make managing decisions. sensors and GPS for positioning. The use of this
With the accumulation of information coming type of equipment is widely spread, but the possibil­
from inspections over the years, non-destructive test­ ities of new generations of onboard thermal and
ing and monitoring of bridges campaigns, the appli­ high-speed cameras open new promising
cation of big data analytics is crucial to detect and applications.
anticipate trends in deterioration and their correlation On the other hand, an underwater remotely oper­
to structural performance. (Kim & Queiroz 2017, ated vehicle (ROV) is a mobile robot designed for
Kobayashi & Kaito 2017). aquatic work environments. The remote control is
usually carried out through copper or fiber optic
cables. Usually, a human operator watches in
6.4 Computer vision
a display what the robot “sees” from a shore-based
Computer Vision Techniques, together with the station, boat or submarine bubble. This type of tech­
acquisition of images via portable cameras and nology is very important in the early detection of
Unmanned Aerial Vehicles (UAVs) offer promising scouring and underwater deterioration of foundation
solutions for the evaluation of bridges. The goal of elements.
such systems is the automatic and accurate conver­ As anticipated by Hola & Schabowics (2013), the
sion of images and videos into data that help in the area of non-destructive testing for buildings and
decision making of engineers and managers. These bridges has evolved a lot in the recent years with the
systems are an alternative to traditional monitoring development of equipment and methods based on
and inspection methods that can be laborious, time- acoustic, ultra-sound, electro-magnetic, radar and
consuming, high cost. They can be used in places thermal technologies, New augmented vibration and
that are difficult or dangerous to access allowing laser-based technologies can now be used to capture
a close contact with the structure. (Spencer et al. the dynamic signature of different types of structures
2019, Catbas et al. 2018) (Rothberg et al. 2017).
When the inspection is concerned, one of the Moreover, the evolution of GNSS (Global Navi­
main objectives of Computer Vision is the identifica­ gation Satellite System: GPS, GLONASS, Galileo,
tion and classification of pathological manifestations. Beidou e.g.) based sensors for Structural Health
Two ways to accomplish such a task are often Monitoring enables 3D positioning anywhere on the
applied: feature extraction or machine learning globe with no measurement drift. A GNSS station is
recognition. able nowadays to use satellite signals and its internal
Computer Vision can also be applied for displace­ clock to estimate position with accuracy up to a few
ment measurement using cameras and UAVs millimeters. In addition, the application of affordable
(Unmanned Aerial Vehicle). It has gained space and spread smartphone technologies (Alavi &
26
Buttlar 2019, Pravia & Braido 2015) is opening new
possibilities for inspection, anomalies detection and
monitoring of the structural behavior of bridges.

7 IMPLEMENTATION OF A BRIDGE DIGITAL

TWIN

A computational system needs to be developed to

implement a digital twin. Here, an evolution of the
system SIM (Structural Integrity Management) devel­
oped for oil and offshore naval structural applications Figure 21. Automatic anomalies detection.
is used to illustrate the potentialities of a digital twin
(Witte & Ribeiro 2012a). The new system is named
C4D and it is currently under development PhDsoft Images taken in human or automated inspections
(http://phdsoft.com/technology/c4d/). The implemen­ may be stored and classified using Deep Learning
tation of any digital twin requires the construction of (DL) algorithms. Some previous training and technical
a 3D model with the most efficient representation of expertise are necessary to implement the machine
geometry, material properties, internal connections learning scheme. Self-evolution of the system is
and boundary conditions, as well as ground localiza­ achieved because the: Digital Twin can collect and
tion information (Witte & Ribeiro 2012b). update data in real-time, and continuously self-
In the case of a bridge, the digital twin may improve its learning process. The digital twin can
incorporate and manage drawings, spread-sheets, handle real-time data acquisition provided by its net­
documents and technical reports, regarding the work of sensors. Through the information provided by
design and the construction of the bridge, as well as the sensors and the use of prediction tools, a real-time
regulatory provisions utilized, in an easy and com­ risk evaluation can be provided at the global or at the
prehensive way (Figure 20). All the data is stored element level. All the information is provided graphic­
and managed using Cloud Computing. ally to the users (Figure 22) in real-time, providing
The system is also capable of storing “as built” as important subsidies for the decision making and priori­
well as “as is” information, due to the changes of the tization processes involved in the bridge maintenance.
structure along its lifetime. Automatic tools for check­ Risk-based Predictive Maintenance (PdM) helps to
ing regulatory compliances and implementing best determine the optimal time and date when repairing
repairing practices are also available in the system. activities should be performed, offering cost savings
Another important feature of the system is pro­ over routine or preventive maintenance. PdM is fre­
vided by the module for automatic anomaly recogni­ quently used for constant condition monitoring and
tion and classification (Figure 21). equipment evaluation. The main goal of its associ­
ation with the digital twin is, by statistical process
control, to determine when future maintenance will
be more appropriate according to the structural condi­
tion or performance. A general view of the integration
of the technologies and information associated with
the bridge digital twin is provided in Figure 23.
The integration of data coming from design and
construction documentation, inspection, non­
destructive testing and monitoring data constitutes

Figure 20. Information and image handling using

cloud-computing. Figure 22. Digital twin capabilities.

27
Figure 23. Digital twin capabilities.

a powerful tool for decision support and may be very by VLI and VALE Railway Companies. The opin­
helpful in the management of maintenance of ions and conclusions presented in this paper are
a single or a network of bridges. those of the authors and do not necessarily reflect the
views of the sponsoring organizations. The inter­
action and the support provided by PhDsoft is
8 CONCLUSIONS deeply appreciated.

Continued inspection of bridges is essential for effi­

cient management of these assets. Structural integ­ REFERENCES
rity monitoring (SHM) and modern Non-destructive
Testing techniques are increasingly economically ABNT-NBR 9452. 2016. Inspeção de Pontes, Viadutos
e Passarelas de Concreto – Procedimento. Rio de
viable. Maintenance, repair and reinforcement of Janeiro, 2016.
bridges are essential to ensure its service life. Agence France Presse (AFP) 2019: https://www.afp.com/
The combination of modern inspection tech­ Accessed 14 November 2019].
niques, non-destructive testing and structural health Akiyama, M. & Frangopol, D.M. 2018. Life-cycle reliabil­
monitoring allows much more accurate diagnoses ity of bridges under independent and interacting haz­
and prognoses of the structural behavior of various ards. Maintenance, Safety, Risk, Management and Life-
types of structures, including bridges. Cycle Performance of Bridges. Powers, Frangopol, Al-
Digital transformation is an inevitable process and Mahaidi & Caprani (Eds) © 2018 Taylor & Francis
will affect the competitiveness of various sectors of Group, London, ISBN 978-1-138-73045-8.
Alavi, A.H. & Buttlar, W.G. 2019. An overview of smart-
industry. Bridge inspection, monitoring, maintenance
phone technology for citizen-centered, real-time and
and management processes can benefit greatly from scalable civil infrastructure monitoring. Future Gener­
introducing new technologies and digital transform­ ation Computer Systems. Volume 93, April 2019, Pages
ation in support to all the accumulated knowledge of 651–672. https://doi.org/10.1016/j.future.2018.10.059.
the field in recent years. ASCE (American Society of Civil Engineers) - Infrastruc­
ture Report 2017: https://www.infrastructurereportcard.
org/Accessed 14 November 2019].
ACKNOWLEDGEMENTS Biondini, F. & Frangopol, D.M. 2016. Life-cycle perform­
ance of deteriorating structural systems under uncer­
Authors would like to acknowledge CNPq (Brazilian tainty: Review. Journal of Structural Engineering 142
Ministry of Science and Technology Agency) and (9): F4016001.
Biondini, F. & Frangopol, D.M. 2018. Life-cycle perform­
CAPES (Higher Education Improvement Agency)
ance of civil structure and infrastructure systems:
for providing an important part of the financial sup­ survey. Journal of Structural Engineering 144(1):
port needed to develop this paper. The work 06017008.
described in this paper has been partially supported

28
Bittencourt, T.N., Beck, A.T. & Frangopol, D. 2016. Main­ Frangopol, D.M., Sabatino, S. & Dong, Y. 2016. Bridge
tenance, monitoring, safety, risk and resilience of bridges Life-Cycle Performance and Cost: Analysis, Prediction,
and bridge networks. 1.Ed. London: CRC Press - Taylor Optimization and Decision Making IABMAS 2016. CRC
and Francis. V. 1. 616p. ISBN: 978-1-138-02851-7. Press - Taylor and Francis. V. 1: 3–17. ISBN: 978­
Bocchini, P., Frangopol, D.M., Ummenhofer, T. & 1-138-02851-7.
Zinke, T. 2014. Resilience and sustainability of civil Fujino, Y. 2018. Bridge maintenance, renovation and man­
infrastructure: Toward a unified approach. Journal of agement - Research and Development of governmental
Infrastructure System 20(2): 04014004. program in Japan. IABMAS 2018, CRC: 2018. V.1, PP.
Catbas, N., Dong, C.Z., Celik, O., Khuc, T. 2018. A Vision 2–14.
for Vision-based Technologies for Bridge Health Moni­ Furuta, H., Frangopol, D.M. & Akiyana, M. 2015. Life-
toring. IABMAS 2018, CRC Press - Taylor and Francis, Cycle of Structural Systems: Design, Assessment, Main­
2018. V.1: 54–62. ISBN: 978-1-138-73045-8. tenance and Management. CRC Press - Taylor and Fran­
Climate Central: https://www.climatecentral.org/Accessed cis, 2015. V. 1. 433p. ISBN: 978-1-138-00120-6.
20 December 2019]. Ghosn, G., Duenas-Osorio, L., Frangopol, D.M.,
Colombo, A.B. et al. 2016. Development of a web interface McAllister, T.P., Bocchini, P., Manuel, L.,
for structural health monitoring data visualization and Ellingwood, B.R., Arangio, S., Bontempi, F., Shah, M.,
structural performance. Maintenance, Monitoring, Akiyama, M., Biondini, F., Hernandez, S. & Tsiatas, G.
Safety, Risk and Resilience of Bridges and Bridge net­ 2016. Performance indicators for structural systems and
works. London: CRC Press, 2016. v.1. p.227–227. infrastructure networks. Journal of Structural Engineer­
Colombo, A.B. 2016. Applications of Structural Health Moni­ ing 142(9): F4016003.
toring and Field Testing Techniques to Probabilistic Based Glaessgen, E. & Stargel, D. 2012. The Digital Twin Para­
Life-Cycle Evaluation of Reinforced Concrete Bridges, digm for Future NASA and U.S. Air Force Vehicles,
PhD thesis., University of São Paulo, Brasil, 2016. 53rd AIAA/ASME/ASCE/AHS/ASC Structures, Struc­
Cortés Pérez, J.P., Sanjosé Blasco, J.J., Atkinson, A.D.J. & tural Dynamics and Materials Conference, 20th AIAA/
del Río Pérez, L.M. 2018. Assessment of the Structural ASME/AHS Adaptive Structures Conference,14th AIAA,
Integrity of the Roman Bridge of Alcántara (Spain) 2012.
using TLS and GPR. Remote Sens., 10, 387. http://doi. Hajdin, R. 2018. Managing Existing Bridges – On the
org/10.3390/rs10030387 Brink of an Exciting Future. Maintenance, Safety, Risk,
COST TU1406. 2016. WG1 Technical Report, Perform­ Management and Life-Cycle Performance of Bridges.
ance Indicators for Roadway Bridges of COST Action CRC Press - Taylor and Francis, 2018. V.1, pages 70–
TU1406, 2016. 87. ISBN: 978-1-138-73045-8.
COST TU1406, 2017. WG2 Technical Report, Perform­ Heng, J., Zhenga, K., Kaewunruenb, S., Zhua, J. &
ance Goals for Roadway Bridges of COST Action TU Baniotopoulos, C. 2019. Dynamic Bayesian
1406, 2017. network-based system-level evaluation on fatigue reli­
COST TU1406. 2019. WG5 Technical Report, Drafting of ability of orthotropic steel decks. Engineering Failure
Guideline/Recommendatios of COST Action TU 1406, Analysis, Vol. 105, Pages 1212–1228. https://doi.org/
2019. 10.1016/j.engfailanal.2019.06.092
Dang, N.S., Kang, H.R., Lon, S. &, Shim, C.S. 2018. 3D Hola, J. & Schabowicz, K. 2013. State-of-the-art non­
Digital Twin models for bridge maintenance. 10th Inter­ destructive methods for diagnostic testing of building
national Conference on Short and Medium Span structures – anticipated development trends. Archives of
Bridges, Quebec City, Quebec, Canada, July 31 – Civil and Mechanical Engineering, Vol. 10, 3, pp. 5–18.
August 3, pages 73-1:9, 2018. Hüthwohl, P. & Brilakis, I. 2018. Detecting Healthy Con­
Denysiuk, R., Fernandes, J., Matos, J.C., Neves, L.C & crete Surfaces. Advanced Engineering Informatics, v.
Berardinelli, U. 2016. A Computational Framework for 37, 150-162, 2018.
Infrastructure Asset Maintenance Scheduling. Structural Hüthwohl, P., Ioannis Brilakis, I., Borrmann, A. &
Engineering International, 26:2, pages 94–102, http:// Sacks, R. 2018. Integrating RC Bridge Defect Informa­
doi.org/10.2749/101686616x14555428759046 tion into BIM Models, Journal of Computing in Civil
Dong, C.Z., Celik, O. & Catbas, F.N. 2018. Marker-free Engineering. 2018, 32(3):04018013. https://doi.org/
monitoring of the grandstand structures and modal iden­ 10.1061/(ASCE)CP.1943-5487.0000744
tification using computer vision methods. Structural Kobayashi, K. & Kaito, K. 2017 Big data-based deterior­
Health Monitoring, p.1–19. https://dx.doi.org/10.1177/ ation prediction models and infrastructure management:
1475921718806895 towards assetmetrics, Structure and Infrastructure
Ellingwood, B.R. 2005. Risk-informed condition assessment Engineering, 13(1), 84–93.
of civil infrastructure: state of practice and research issues. Kim, Y.J. & Queiroz, L.B. 2017. Big Data for condition
Structure and Infrastructure Engineering, 1(1),7–18. evaluation of constructed bridges. Engineering Struc­
https://doi.org/10.1080/15732470412331289341 tures, 141: 217–227.
EN 1991- 1-7Eurocode 1 2006 + A1:2014. Oliveira, C.B.L., Greco, M. & Bittencourt, T.N. 2019. Ana­
Faridafshin, F., Anvari, M. & Hellevig, N.C. 2019. Risk- lysis of the Brazilian federal bridge inventory. IBRACON
Based Approaches for Planning the Inspections of Structures and Materials Journal, v. 12, p.1–3, 2019.
Large-Scale Topside Structural Systems. Proceedings of https://doi.org/10.1590/s1983-41952019000100002.
the SPE Offshore Europe Conference and Exhibition. Powers, N., Frangopol, D.M., Al-Mahaidi, R & Caprani, C.
Society of Petroleum Engineers, UK, 2019. 2018. Maintenance, Safety, Risk, Management and Life-
Frangopol, D.M. & Soliman, M. 2015. Life-Cycle of Struc­ Cycle Performance of Bridges. 1.Ed. London: CRC
tural Systems: Recent Achievements and Future Direc­ Press - Taylor and Francis, 2018. V. 1. 550p. ISBN:
tions. IALCCE 2014. CRC Press - Taylor and Francis, 978-1-138-73045-8.
2015. V. 1, PP. 3–17. ISBN: 978-1-138-00120-6. Pravia, Z.M.C. & Braido, J.D. 2015. Measurements of
bridges’ vibration characteristics using a mobile phone,

29
 IBRACON Structures and Materials Journal, Volume 8, Engineering, v. 1: 1-21, 2019. https://doi.org/10.1080/
Number 5 (October 2015) p. 721–743. ISSN 1983-4195. 15732479.2019.1605394.
http://dx.doi.org/10.1590/S1983-41952015000500009. Wenzel, H. 2009. Health Monitoring of Bridges., Vienna:
Radomski, W. 2001. Bridge Rehabilitation. Warsaw: John Wiley & Sons.
Imperial College Press. Wikipedia 6D BIM, 2019. [Online] Available at: https://en.
Rothberg, S.J., et al. 2017. An International Review of wikipedia.org/wiki/6D_BIM [Accessed 20 December
Laser Doppler Vibrometry: Making Light Work of 2019].
Vibration Measurement. Optics and Laser in Engineer­ Witte, C.C. & Ribeiro, D.M. 2012-a. A statistical time
ing 99, pp. 11–22. dependent degradation curve analysis for marine structures.
Rowland, I. & Howe, T.N. 1999. Vitruvius. Ten Books on ASME 2012 - 31st International Conference on Ocean,
Architecture. Cambridge University Press, Cambridge Offshore and Arctic Engineering. OMAE2012: 1–5.
1999, ISBN 0-521-00292-3. Witte, C.C. & Ribeiro, D.M. 2012-b. Structural integrity
Salehi, H. & Burgueño, R. 2018. Emerging artificial intelli­ management: painting predictive control, SPE
gence methods in structural engineering, Engineering 155857-PP, SPE International Conference and Exhib­
Structures, Volume 171: 170–189, 15 September 2018. ition on Oilfield Corrosion held in Aberdeen, UK,
https://doi.org/10.1016/j.engstruct.2018.05.084. 28–29 May 2012: 1–5.
Saydam, D. & Frangopol, D.M. 2011. Time-dependent Yan, W., Denga, L., Zhang, F., Li, T. & Shaofan Li, S.
perfor-mance indicators of damages bridge superstructures. 2019. Probabilistic machine learning approach to bridge
Engi-neering Structures 33(9): 2458–2471. fatigue failure analysis due to vehicular overloading.
Shim, C.S., Kang, H.R., Dang, N.S. & Lee, D. 2017. Engineering Structures, Volume 193: 91–99,
Development o BIM-based bridge maintenance system 15 August 2019. https://doi.org/10.1016/j.
for cable-stayed bridges. Smart Structures and Systems, engstruct.2019.05.028.
20(6): 697–708 (2017). http://doi.org/10.12989/ Yang, D.F. & Frangopol, D.M. 2018. Probabilistic opti­
sss.2017.20.6.697. mization framework for inspection/repair planning of
Spencer Jr, B.F., Hoskere, V. & Narazaki, Y. 2019. fatigue-critical details using dynamic Bayesian
Advances in Computer Vision-Based Civil Infrastruc­ networks. Computers and Structures, 198: 40–50.
ture Inspection and Monitoring. Engineering, v5, n.2: https://doi.org/10.1016/j.compstruc.2018.01.006
199–222, 2019. http://doi.org/10.1016/j. Yang, D.F. & Frangopol, D.M. 2019. Life-cycle manage­
eng.2018.11.030. ment of deteriorating civil infrastructure considering
Taffese, W. Z. & Sistonen, E. 2017. Machine learning for resilience to lifetime hazards: A general approach based
durability and service-life assessment of reinforced con­ on renewal-reward processes. Reliability Engineering &
crete structures: Recent advances and future directions. System Safety, Vol. 183: 197–212. https://doi.org/
Automation in Construction 77: 1–14. 10.1016/j.ress.2018.11.016
Ticona Melo, L.R., Ribeiro, D., Calçada, R.B. & Zhu, Z., Liu, C. & Xu,X. 2019. Visualisation of the Digital
Bittencourt, T.N. 2019. Validation of a vertical Twin data in manufacturing by using Augmented
train-track-bridge dynamic interaction model based on Reality. Procedia CIRP vol. 81, pages 898–903. http://
limited experimental data. Structure and Infrastructure doi.org/10.1016/j.procir.2019.03.223

30
 Bridge Maintenance, Safety, Management, Life-Cycle Sustainability and Innovations –
Yokota & Frangopol (eds)
© 2021 Taylor & Francis Group, London, ISBN 978-0-367-23278-8

How to design, construct, and maintain highly-durable concrete bridges in

cold and local regions
I. Iwaki
College of Engineering, Nihon University, Koriyama, Japan

ABSTRACT: In Japan, reinforced concrete deck slabs of major road bridges in cold regions have suffered
from rapid aging due to the large-scale use of de-icing agent (NaCl). Additionally, a large number of minor
bridges in farming and other local regions are facing lack of financial and technological resources for mainten­
ance. In this paper, some methodologies to design, construct, and maintain highly-durable concrete bridges in
cold and local regions are introduced. Firstly, six full size RC deck slabs on steel main girders were reproduced
in the university campus. Varying the materials and mixture proportions, and construction methods for each
slab, the performance was evaluated. The test and evaluation results were effective to develop highly-durable
concrete deck slabs having excellent durability and crack resistance even under harsh chloride environment.
Then, the developed deck slab based on the multiple protection strategy has been applied to reconstruction
roads in the 2011 Tohoku earthquake-damaged regions. On the other hand, to maintain infrastructures by
a simple preventive method, the spirit of fushin, is a starting point for civil engineering in Japan, is deeply
ingrained in local regions. The concept of fushin is “the infrastructure desired by local residents should be built
and maintained by themselves.” The latter part of this paper covers a modern-day version of this traditional
maintenance method that has been introduced to local communities and put into practice.

1 INTRODUCTION of infrastructure in the Tohoku region and Fukushima

Prefecture, with a focus on road bridges.
In recent years, the simultaneous deterioration of the
societal infrastructure (hereafter “infrastructure”) that
was constructed intensively during Japan’s period of 2 CURRENT STATE OF BRIDGES IN THE
high-level economic growth has become a serious TOHOKU REGION AND RECOVERY FROM
social problem. In fact, the Sasago Tunnel accident THE GREAT EAST JAPAN EARTHQUAKE
occurred in December 2012, in which the tunnel ceil­
ing panels collapsed and 9 people died. The term The Great East Japan Earthquake and tsunami
“infrastructure” includes many different types of struc­ occurred on March 11, 2011. Damage to bridges result­
tures including bridges, tunnels, dams and so on. ing from the earthquake was kept to a minimum
Moreover, even structures of the same type, such as thanks to seismic retrofitting that had been conducted
road bridges, can have very different roles and degrees in the aftermath of the Great Hanshin-Awaji Earth­
of importance. Some bridges are on a main arterial quake of 1995. However, there was nothing that could
road in the Tokyo metropolitan area that is traversed be done about the unprecedented tsunami. As the
by 100,000 vehicles each day, and others bridges are centerpiece of the reconstruction effort, the Tohoku
used by only several dozen vehicles per day but are Regional Development Bureau of the Ministry of
indispensable to the local community. It would make Land, Infrastructure, Transport and Tourism began
no sense to use the same cookie-cutter methods to work on the construction of a highway (reconstruction
maintain both of these bridges, and surely it is a good road) through the regions of the Sanriku Coast that
thing if advanced technologies are used to prolong the were damaged by the tsunami, and three highways
service life of some bridges while for other bridges (reconstruction support roads) going from the inland
methods that will not take much out of the budget and part of the Tohoku region to the coast (See Figure 1).
will enable the bridge to be in service longer are used. These areas are located in a cold region, and large
This paper will discuss methods in which the author quantities of de-icing agent (NaCl) are expected to be
is involved that are aimed at extending the service life spread on the road in winter as a

DOI: 10.1201/9780429279119-3
https://doi.org/10.1201/9780429279119-3

31
 Figure 2. LOHAS bridge.

Table 1. Types of LOHAS bridge.

Type of deck w/c Air (%) Technology

(1) Standard 0.55 4.5

(2) Low quality 0.65 3.0
(3) High durability 0.45 6.0 Expansive agent
(4) Superior durability 0.45 6.0 Expansive aghent,
FA, and Long­
term curing
Figure 1. Reconstruction roads in Tohoku Region. (5) Curing measures 0.55 4.5 Long-term curing
(6) Machine finish 0.55 4.5 Machine finish

snow melting agent to prevent freezing. When

salt acts on concrete structures, it causes frost In summary, the 6 types were (1) a deck slab pro­
damage, chloride attack, and alkali-silica reaction duced in accordance with the current standards (stand­
(ASR); if this deterioration is compounded, it is ard), (2) a deck slab in which the micropore structure
expected that the structure will need to be replaced was inferior and the air content of the concrete was
after a few decades. Accordingly, it was decided to low compared with the standard (low quality), (3)
take measures to extend the service life of these a deck slab in which the micropore structure was
bridges. superior and the air content of the concrete was high
compared with the standard (high durability), (4)
a deck slab with concrete with finer micropore struc­
3 LOHAS BRIDGE PROJECT ture by adding fly ash to high durability concrete, and
in addition, wet/sealed curing was carried out for 3
3.1 Overview of the project months to utilize its pozzolanic reactivity (superior
durability), (5) a standard deck to which wet/sealed
The LOHAS Bridge Project was initiated to achieve curing was carried out continuously for 3 months
healthy and sustainable bridges. (LOHAS is the (curing measures), and (6) a standard deck in which
meaning of “lifestyles of health and sustainability.”) the top surface was finished with a machine (machine
To ensure a long service life for the concrete deck finish).
slabs of the reconstruction road and reconstruction
support road, which are exposed to harsh environ­
ments, actual size models of the bridge were con­ 3.2 Materials and mixes used in concrete
structed in the author’s laboratory. It was the first The materials used in the concrete were as follows.
time that this had ever been done on a university The water was groundwater, the cement was ordinary
campus in Japan (see Figure 2). Specifically, six Portland cement, the fly ash was type II ash from
types of concrete deck slabs constructed with differ­ Noshiro fossil-fired power plant, the expansive agent
ent materials, mixes and construction methods as was a low addition type, the fine aggregates were
shown in Table 1 were built on top of steel main gir­ crushed sand produced in Fukushima Prefecture, and
ders, and field measurements under an actual envir­ the coarse aggregates were crushed stone produced in
onment were conducted over a year; subsequently, Fukushima Prefecture. Also, other admixtures that
durability tests were conducted, in order to develop were used depending on the mix included an air
a concrete deck slab that could provide a long ser­ entraining (AE) water reducing agent (Ad), a high per­
vice life even in harsh environments. Finally, the formance AE water reducing agent (SP), an air entrain­
deck slab that had been developed was then con­ ing agent (AE-1), and an antifoaming agent (AE-2).
structed at the reconstruction site. The concrete was produced at a ready-mixed concrete

32
Table 2. Mixture proportions for each deck.

Unit content (kg/m3)

s/a
Type of deck w/c (%) W C EX FA S G Ad SP AE-1 AE-2

(1) Standard 0.55 46.2 172 313 834 997 1.2 0.0016
(5) Curing measures 0.55 46.2 172 313 834 997 1.2 0.0016
(6) Machine finish 0.55 46.2 172 313 834 997 1.2 0.0014
(2) Low quality 0.65 49.0 185 285 899 960 1.5 0.0020
(3) High durability 0.45 45.0 161 338 20 791 992 0.55 0.0020
(4) Superior 0.45 45.0 161 338 20 72 709 992 0.70 0.0100
durability

factory, and delivered. Table 2 shows the mixture pro­ ent from the other concretes. In other words, the air
portions for each deck. From the table, it can be seen permeability coefficient is small, and the electrical
that the mixes for (1), (5), and (6) were basically the resistance is high, which is an excellent result in terms
same, and in (1), (2), and (3) the water cement ratio of level 1 or level 2 concrete surface quality. It is con­
and the admixture content (AE water reducing agent, sidered that this result was due to the fact that wet/
air entraining agent, and antifoaming agent) were sealed curing was carried out continuously up to 91
varied to change the micropore structure and air con­ days, so the pozzolanic reaction of the fly ash pro­
tent of the concrete. Also, in (3) an expansive agent gressed, forming an extremely fine micropore structure,
was added to the cement in the proportion 20 kg/m3, in and the excess water in the micropores was consumed
order to compensate for shrinkage of the concrete. In by hydration.
addition, in (4) fly ash was added to the mix in (3) to Figure 4 shows the measured results for expan­
replace fine aggregate in the proportion of 20% of the sion and shrinkage strain in the concrete of each
total powder content. deck slab. From this figure, it can be seen that the
largest amount of shrinkage was obtained with the
low quality concrete, followed by the standard
3.3 Field measurement
and machine finish. Although the curing measures
Figure 3 shows air permeability test results measured concrete had almost no shrinkage up to 91 days, sub­
by the Torrent method (Torrent, 1992). The horizontal sequently shrinkage progressed, and at 365 days,
axis in the figure represents the air permeability coeffi­ there was no major difference from the standard con­
cient, the smaller the value the finer the micropore crete. With the high durability and superior durabil­
structure of the concrete, which means that the perme­ ity concrete, there was almost no shrinkage at 365
ability to gas is lower. On the other hand, the vertical days due to the effect of expansive agent.
axis represents the electrical resistance. Water easily
conducts electricity, so the lower the value the greater
3.4 Experimental results and discussion
the quantity of free water remaining within the fine
pores. The measurement positions were near the center After completion of measurements under the actual
of the top and bottom surfaces of the deck slabs. From environment for 1 year, cores were extracted from
the measurement results at 91 days, it can be seen that the deck slabs, and chloride ingress tests and freeze-
only the superior durability concrete was clearly differ­ thaw tests (scaling tests) were carried out.

Figure 3. Air permeability test results by Torrent method. Figure 4. Expansion and shrinkage behaviors for each deck.

33
 The chloride ingress tests were carried out using
cylindrical test specimens 150mm diameter and
80 mm height, by the method of exposing the test
surface to a 3% aqueous solution of NaCl for
a predetermined number of days with the concrete
casting surface as the test surface. The test environ­
ment had a temperature of 40C and a relative humid­
ity of 60%. Concentration control of the test solution
was carried out once per week. The chloride ingress
was carried out for 91 days from the start of the test.
Two cores were extracted from the test specimens
after the prescribed number of days had passed, and
powder material was extracted at 3 mm intervals
from the test surface. The quantity of chloride ion
was measured by the ion chromatography analysis
method. Figure 6. Freezing and thawing test results.
The freeze-thaw tests (scaling tests) were carried
out in accordance with ASTM C 6721. The shape of
the test specimens was 150 mm×80 mm, the same as superior durability concrete because the water
the test specimens used in the chloride ingress tests, cement ratio was low, fly ash was added, and in add­
with the concrete casting surface as the test surface. ition, wet/sealed curing was continued for 91 days,
The test solution was 3% NaCl aqueous solution, so the micropore structure was extremely fine. Also,
which was placed inside a wall made from water­ the results for high durability concrete depended on
proof tape. The item measured in the freeze-thaw the water cement ratio, but although the mix for the
test was the quantity of scaling per unit area every 5 curing measures concrete was unchanged from that
cycles. One cycle consisted of a thawing holding of the standard concrete, the chloride ingress resist­
time of 6 hours at a maximum thawing temperature ance was clearly superior, and it is considered that
of 20±3C, a temperature transition time of 1 hour, this clearly shows the effect of curing.
a freezing holding time of 16 hours at a minimum Figure 6 shows the results of the scaling tests car­
freezing temperature of -20±3C, and a temperature ried out under the chloride environment. Note that in
transition time of 1 hour. The freeze-thaw test was this test, the reference value was set at 0.3 kg/m2 for
carried out for 50 of these cycles. the scaling tests. This value can be considered to
Figure 5 shows the chloride ingress test results. indicate a state in which almost no scaling occurs.
From this figure, it can be seen that the concrete This figure shows that only the low quality concrete
with the best resistance to chloride ingress was the had a result greatly in excess of the reference value,
superior durability concrete. Next were the high dur­ but all the other cores had results less than the refer­
ability, curing measures, standard, and low quality ence value. From the results, it is difficult to say that
concretes, and the machine finish concrete had the the low quality concrete with a water cement ratio of
lowest result. It is considered that the chloride 65% and 3% air content has sufficient scaling resist­
ingress resistance results were extremely high for the ance under a chloride environment, but under other
conditions, it has a certain level of scaling resistance.
In particular, it has been showed that the superior
durability concrete with water cement ratios of 45%
and air content of about 6% have extremely high
scaling resistance too.

3.5 Application of results to reconstruction site

From the above research results, it can be seen that
the superior durability concrete has excellent
strength development, the quality of its surface layer
is very fine, it exhibits stable strain behavior, and its
chloride ingress resistance and scaling resistance
when de-icing agent has been sprinkled is also excel­
lent. Therefore, it has been decided to apply fly ash
concrete based on this mix to the construction of the
RC deck slab of Mukai-Sadanai Bridge on National
Route 283 Kamaishi Road which is being recon­
structed after the Great East Japan Earthquake. The
basic concept of this deck slab is multiple protection
Figure 5. Chloride ingress test results. strategy (Tanaka et al. 2017).

34
 In other words, road bridges in snowy cold bridges, and an RC deck slab using Type B blast fur­
regions on which a large quantity of de-icing agent nace cement has been constructed on 1 bridge
is sprinkled are subject to the combined actions of (Ishida et al. 2018). It is intended that this concept
chloride attack, frost damage, ASR, and fatigue. will be widely adopted not only in the Tohoku
Therefore, in order to provide resistance to these region, but in the whole country.
types of deterioration, the deterioration is reduced by
the multiple countermeasures shown in Figure 7, to
ensure high durability. It has been confirmed that the 4 PROMOTING BRIDGE PRESERVATION IN
required quality and performance can be obtained in COOPERATION WITH LOCAL RESIDENTS
the construction of this deck slab by the shipment of
concrete with stable quality, and proper construction 4.1 Outline of the project
on site. Figure 8 shows the concrete deck slab being
constructed. The reinforcement has been coated with As noted above, when constructing a new concrete
epoxy resin to protect it from corrosion. structure on a main arterial road managed by the
It was found that the initial construction cost of Ministry of Land, Infrastructure, Transport and Tour­
Mukai-Sadanai Bridge, on which this multiple pro­ ism or Expressway corporations, it is essential from
tection has been implemented, was about 26% the beginning to build something that is durable and
higher compared with a conventional deck slab. On will last. In the case of existing bridges, however,
the other hand, there have been examples in the the service life must be extended through mainten­
snowy and cold regions of Japan where after just 30 ance. Of these existing bridges, if the bridge is man­
years of service RC deck slabs that have been sprin­ aged by an expressway corporation, or the national/
kled with de-icing agent have been removed, and prefectural government, to a certain extent, it is pos­
replaced at about 4 times the initial cost. Based on sible to allocate money from the budget for mainten­
this fact, it has been judged that the specification ance and conduct maintenance by employing
adopted for Mukai-Sadanai Bridge can be suffi­ advanced technologies. However, in the case of
ciently accepted from the point of view of life cycle a bridge managed by a municipality that may not
cost. On this reconstruction road project as of have the financial or technical capability to do this,
March 2017, highly durable RC bridge deck slabs the situation is more serious. It is thought that
using fly ash have been constructed on a total of 4 a sufficiently long service life can be achieved for
these bridges by performing preventive maintenance
against deterioration. The author proposes simple
maintenance focusing on the action of water on the
bridge as a method to prevent bridge deterioration.
Water is either directly or indirectly involved in the
factors that cause deterioration in concrete struc­
tures, namely carbonation, chloride attack, frost
damage, ASR, chemical attack and fatigue. Accord­
ingly, it is thought that taking steps to prevent water
from needlessly acting on the concrete structures
that are in service will dramatically improve the dur­
ability of these concrete structures. Figure 9 shows
an example. This photo shows water containing de­
icing agent flowing out from a gap between a bridge
Figure 7. Multi protection strategy. girder and an abutment, with significant deterioration

Figure 8. Construction of highly-durable concrete deck. Figure 9. Deterioration due to action of water.

35
visible due to frost damage and chloride attack. In
contrast, the dry section nearby shows almost no
deterioration. Examples of methods that can be used
to prevent water from unnecessarily acting on the
concrete include removing deposits of soils and
sands on the concrete road surface, cleaning the
drainage basins, and revising the direction and
length of the drain pipes. Applying a simple paint
job to the bridge railings at regular intervals can also
protect the railing from rust and preserve its beauti­
ful appearance. These steps do not require special
technologies and will not require much of a budget,
so they are things that can be done easily even by
local residents. The author has encouraged that these
steps be done by local residents under the name Figure 11. Collaboration between local residents and
“bridge tooth-brushing” based on the concept of students.
fushin. Figure 10 shows the project scheme for
“bridge tooth-brushing” with the cooperation of gov­
ernment, academia, industry and the local commu­ The government also noted that it was essential to
nity. Under this scheme, the government would evaluate the daily condition of these bridges. How­
provide the materials and tools needed for “bridge ever, municipalities have their hands full just making
tooth-brushing” and local residents would perform sure that regular inspections of bridges are per­
the actual work, with industry providing technical formed, and they are unable to conduct day-to-day
guidance and academia (students) helping local resi­ management. For this reason, the author’s laboratory
dents perform the “bridge tooth-brushing” activities hit on the idea of getting local residents to perform
as an opportunity to contribute to the local commu­ daily inspections of the small-scale bridges that are
nity and learn about the actual situation in the com­ under municipal management, and prepared the
munity. Figure 11 shows local residents and students checklist shown in Figure 12 (Asano et al. 2016).
working together to paint the railing of a bridge in The checklist is printed in color on both sides and
Minamiaizu-machi. contains the inspection checklist on the front and
a supplementary catalogue of changes in status on
the back. The content of the checklist itself is based
4.2 Self-Maintenance model using bridge daily on the bridge inspection form prepared by Fukush­
inspection checklist ima Prefecture. However, out of concern for the
As part of the effort to strengthen inspections of safety of the local residents conducting the inspec­
infrastructure following the Sasago Tunnel accident tion, the content is based on the assumption that the
that occurred in December 2012, the government inspection will be conducted from the surface of the
made it obligatory that, starting from FY 2014, bridge only, and the inspection is limited to six
a proximal visual inspection would need to be con­ items: (1) Bridge railing, (2) Felloe guard, (3)
ducted once every 5 years for all of the roughly Lights, (4) Drainage basin, (5) Pavement, and (6)
700,000 bridges nationwide (Tamakoshi et al. 2015). Expansion joint equipment. Moreover, the items
comprise whether or not there is rusting or cracking
and, if the answer is yes, whether it is localized or
widespread. The bridge inspection catalogue on the
back comprises a catalogue of changes in status for
items (1) through (6), in order to enable residents to
consult the catalogue as they perform the inspection.
There are also “Emergency” spaces that enable the
person performing the inspection to immediately
notify bridge management personnel in the event
that damage requiring urgent action is discovered.
Finally, the checklist urges caution on the part of
residents, urging them to perform the inspection only
from the sidewalk and the shoulder, to ensure that
the inspection is performed by at least two persons,
to avoid performing the inspection when it is raining
or when it is dark and so on.
After the inspection has been conducted using the
checklist, the next issue is how to ensure visibility of
the inspection results. Accordingly, a “Bridge Map”
Figure 10. Project scheme for “bridge tooth-brushing”. shown in Figure 13 was prepared, in which the

36
 Figure 13. Bridge map reflecting inspection data.

bridge inspection results obtained from the checklist

could be checked on the Internet together with loca­
tion data, in order to encourage learning and increase
interest and attention, and with the aim of promoting
cleaning and further inspections.
The Bridge Map was designed with the idea that
it would be viewed by local residents and would
encourage those residents to perform the cleaning
activities that would serve as preventive maintenance
for the bridge, and to provide them with easily under­
standable information on bridge use that would allay
their concerns. The procedure is described below.
1) Convert into numerical values the inspection
results for five items that can be cleaned even by
local residents (“guardrail rusting” “soil and mud in
drainage basin” “weeds in drainage basin” “soil and
mud between curb and sidewalk” and “weeds
between curb and sidewalk”).
2) Convert the average value for each inspection
item into a value on a scale from 1 to 10 and make
this a “bridge tooth-brushing index = X.”
3) Depict high scores using cold colors (blue and
green) and low scores using warm colors (red and
orange), in order to make it possible to tell at a glance
which bridges are most in need of cleaning (tooth­
brushing). In addition, by clicking the pins on the map,
it is possible to display basic information about the
bridge, display the status of the bridge, or show photo­
graphs of various parts of the bridge. As in the case of
other types of maps, the Bridge Map can be viewed on
a smartphone or PC, and it is also possible to display
directions to the bridge and how long it would take to
get there.
In this way, a Simple Bridge Inspection Checklist
was created, initially with the aim of enabling local
Figure 12. Checklist for daily bridge inspection in residents to easily conduct inspections of bridges
Japanese. needing maintenance that are located in rural areas

37
 future generations should be “building highly-durable
structures and working together to protect them.”
Local infrastructure is literally the foundation that
supports the lives of people in local communities. If
it falls, there is no way that regional revitalization
can become a reality. On the other hand, local com­
munities have a strength that cities do not possess:
local cohesion. If the aforementioned activities are
carried out based on the cooperation of industry, aca­
demia, government and the local community, with
the watchword “working together to protect local
infrastructure,” both the infrastructure and the local
Figure 14. Self-maintenance bridge model. communities will survive. These activities will also
help to create independent communities and lead to
the opportunity to promote a new model for local
that have limited budgets and technological capabil­ communities both in Japan and abroad.
ities. As noted earlier, however, the checklist is now
being used for high school classes and for inspection
patrols by in-house engineers. It is possible that the ACKNOWLEDGEMENT
checklist will be used not only by Hirata Village
(population approximately 6,500) and Taiwa-cho This study was financially supported by Council for
(population approximately 30,000) but also in the city Science, Technology and Innovation, “Cross­
of Koriyama (population approximately 340,000), ministerial Strategic Innovation Promotion Program
a major urban area. Moreover, the checklist has been (SIP), Infrastructure Maintenance, Renovation, and
converted into a Bridge Map that clearly identifies the Management” granted by JST. The author also wishes
bridges that need preventive maintenance, resulting in to express their gratitude to Dr. Koda, Ms. Asano, and
the creation and implementation of a bridge “self­ the students in the author’s Laboratory, for their sup­
maintenance cycle” shown in Figure 14 that encour­ port in the experiments.
ages bridge surface cleaning and other preventive
maintenance and will lead to subsequent inspections
(Asano et al. 2019). As a result,
REFERENCES
requests to introduce this system are being Asano W., Koda Y. & Iwaki I. 2016. A proposal for the
received from various local residents, high schools introduction of a simple resident-led bridge inspection
and vocational high schools, universities, private method using a checklist. Proceedings of the Japan
companies and local governments. Concrete Institute, 38(2): 1573–1578. (in Japanese).
Asano W., Koda Y. & Iwaki I. 2019. Construction and
Implementation of Bridge Self-maintenance Model Led
5 CONCLUSIONS by Local Residents Using “Simple Bridge Inspection
Check Sheet” and “Bridge Map.” Journal of JSCE (in
print). (in Japanese).
Infrastructure construction in Japan still seems to Ishida, T., Pen, K., Tanaka, Y., Kashimura, K. & Iwaki, I.
take its cues from the country’s period of high-level 2018. Numerical Simulation of Early Abe Cracking of
economic growth 50 years ago, in which the Reinforced Concrete Bridge Decks with a full-3D Multi-
emphasis was on constructing things quickly and scale and Multi-Chemo-Physical Integrated Analysis.
inexpensively. From this point on, however, newly Applied Sciences, 8(3): 1–18.
constructed structures should design and construct Tanaka Y., Ishida T., Iwaki I. & Sato K. 2017. Multiple pro­
not to impose a burden in terms of maintenance con­ tection design for durable concrete bridge deck in cold
sidering their performance-based life cycle cost. regions. Journal of JSCE, 5(1): 68–77.
Tamakoshi, T., Shirato, M., Yokoi, Y., Mabuchi, T. &
In the case of existing structures, ordinary citizens
Yabu, M. 2015. Guidelines for Periodic Road Structure
will need to consider themselves to be participants Inspections. 2015 Annual Report of NILIM.
and work together with industry, academia and gov­ Torrent, R. J. 1992. A two-chamber vacuum cell for measur­
ernment to maintain these structures. The approach ing the coefficient of permeability to air of the concrete
to design, construct, and maintain infrastructure in cover on site. Materials and Structures, 25(6): 358–365.

38
 Bridge Maintenance, Safety, Management, Life-Cycle Sustainability and Innovations –
Yokota & Frangopol (eds)
© 2021 Taylor & Francis Group, London, ISBN 978-0-367-23278-8

Development of the requirements to major infrastructure projects

E.S. Larsen, M.S. Lagergaard, A. Jørgensen, B. MacAulay & J. Laursen
Danish Road Directorate, Copenhagen, Denmark

ABSTRACT: This presentation will describe how the Danish Road Directorate works with the develop­
ment and execution of construction of large infrastructure projects. The presentation will be divided into
two parts, respectively, the framework for reducing climate and environmental impact and the challenge of
norms and standards respectively; some illustrated by examples. In the planning phase, implementation of
how the project best can contribute to meeting society’s need for mobility must be implemented, which is
often regarded as a prerequisite for economic growth and human well-being. The planning must include the
latest knowledge on climate and environmental impact associated with the infrastructure life from cradle to
grave - and add knowledge that can form the basis for decisions on a focused and cost-effective climate
action. The planning also includes factors that determine how biodiversity can be maintained as well as
other considerations for the surrounding environment, including protected nature and habitat areas. The
presentation will include examples of the optimization of the design assumptions, considering the risks
associated with the geotechnical conditions and relevant load conditions such as ice and ship impact.
Examples will be taken from the new Roskilde Fjord Link and the coming road- and railway bridge cross­
ing Storstrømmen.

1 INTRODUCTION The tender process normally starts with a ‘draft’

version of the complete tender material being pub­
Developing the requirements to major construction lished, followed by an information meeting and site
projects is always a challenge referring to both quality, visit, which is held prior to the first round of dia­
time and economy. It is a balance between the men­ logue meetings. The rounds of dialogue meetings are
tioned issues combined with all the risks which might often held over the course of 6-8 months. The first
occur on these projects. This paper describes some of dialogue meeting can be a two-day meeting,
the overall challenges and how the Danish Road Dir­ the second can be a one-day meeting and the final
ectorate (DRD) try to manage these challenges while round only a half day or a written questionnaire.
also considering the demands from society. However, this depends highly on the complexity of
The paper is divided into two parts, respectively, the project.
The contractor can then reflect on assessments
1) the framework for reducing climate and environ­ and observations of the tender material in order to;
mental impacts and,
2) the challenge of norms and standards. • Be sure that the project can be within budget, i.e.
budget target is informed during the dialogue,
Each topic is followed by examples from two
• focus on “cost drivers” in order to avoid expen­
major bridge projects, the Roskilde Fjord Link open sive solutions during the contractor’s develop­
for traffic in September 2019 and the Storstrøms ment of his proposal,
Bridge expected to open for traffic in 2022/2023. • optimize the tender documents and requirements,
In more complex projects such as major bridge i.e. the Employer has great possibilities to accom­
structures, the competitive dialogue process is often modate the considerations and comments from
used, following a market dialogue with the potential the contractor within the framework of the law
contractors. The idea behind competitive dialogue is and general requirements,
to make the project clear and understandable for the • achieve a common understanding of the specifica­
contractors, while at the same time allowing the con­ tions and the contract conditions including the
tractors to point out their view on the project, most severe risks associated with the project, i.e.
requirements and risks.

DOI: 10.1201/9780429279119-4
https://doi.org/10.1201/9780429279119-4

39
 it should be clear who carries the risks – the Con­ are today – the so-called baseline – so that we have
tractor or the Employer. a forward-looking reference to how and how quickly
the work to achieve the various objectives progresses.
Following each round of dialogue meetings, the The DRD has also embraced the UN sustainable
tender material is updated, to include the changes development goals, as a state agency and authority in
which were incorporated as a result of the meetings. the road sector – but presumably as everyone else –
However, the final tender documents are only pub­ we start with testing, investigating - a little uncertain
lished at the end of the last round of dialogue meet­ and with great commitment. We started with
ings, these include significant changes which were
a screening of the 169 milestones under the 17 world
a direct result of the dialogue process. After this point targets and have identified where the DRD plays - or
no significant changes to the specifications is done. can play - a role. In fact, our diverse activities
address directly and indirectly, to a greater or lesser
extent, into almost all the world’s goals: from health,
2 CLIMATE AND ENVIRONMENT equality, responsibility for decent jobs and respon­
sible consumption and production, to climate adapta­
At the 2015 General Assembly, all 193 of the UN tion, nature conservation and, last but not least,
countries committed themselves to an unprecedented partnerships. As mentioned above, the work of the
ambitious agenda. An agenda aimed at improving DRD covers almost all the world global goals. How­
life on the planet for hundreds of millions of people ever, we have chosen to focus initially on a subset of
and bringing global development in a direction that the objectives, which include our current and future
recognizes and respects the fact that the finite efforts in the field of climate, including the reduction
resources of the earth cannot be exploited unilat­ of CO2, and various environmental measures, includ­
erally but must be shared and recycled.
ing the reduction and recycling of resources and the
The resolution is summed up in 17 global goals and
minimization of traffic impact on nature.
169 milestones for sustainable development (Figure 1)
One of the key tasks of the DRD is to help ensure
through efforts for human health and prosperity, our
a high level of mobility in Denmark. Mobility creates
land, for peace and security and our global cohesion.
value for people and society. Mobility is underpinned
“Leave no one behind” is the basic principle behind
by good, reliable and robust infrastructure, which is
the global goals to be reached in 2030 (UNDP).
also a prerequisite for more efficient and greener trans­
The global goals now constitute a common global
port modes on the roads. However, road transport also
framework under which each country must define its
leaves a significant footprint on our climate, in the
own efforts. The objectives indicate a common dir­
form of carbon emissions, and on our environment, in
ection so that everyone from the Danish parliament
terms of pollution, noise and pressures on nature and
to local businesses and organisations can set up and
biodiversity. We must reduce this negative footprint as
work towards achieving tangible, measurable results,
best we can.
all of which pull in the common desired direction.
The work of the DRD in this area is also import­
In Denmark, a cross-party political network in the
ant because, as the main road authority, we are help­
Danish parliament was already established in 2017.
ing to set standards and regulations in the sector and
They are working at a national level with strategies to
can thus affect much wider than just our own work.
support sustainable development. The most important
The DRD can contribute to the green transition,
task that has recently been launched is the definition
among other things. Through our planning, construc­
of indicators and the development of baselines for all
tion and operation of the state roads, and by support­
the world’s objectives in the Danish context. Indica­
ing greener mobility and transport. We can also
tors are measurable parameters that can represent the
contribute through increased efforts for climate and
goals one wants to achieve. Finally, one must calcu­
environmental behaviour and resource consumption
late a value for the indicators, which shows where we
in our internal organisation. The effort rests on two
legs: actual reduction of the negative impact on the
environment, and – where it is not feasible or too
expensive – compensatory interventions that outweigh
the negative consequences of our activities. In all
areas our work is supported by dialogue, cooperation
and sharing of the latest knowledge and data with the
research world, the transport industry, as well as other
sectors. Both on a national and international level.
Denmark has an ambition to reduce CO2 emis­
sions by 70 percent by 2030. In order to meet global
and national greenhouse gas reduction targets, in
particular CO2, all activities that leave a CO2 foot­
print must be assessed for efficiency or conversion.
Figure 1. The sustainable development goals from UNDP. In 2017 carbon emissions from traffic in Denmark
org. accounted for 35 percent or around 12 mio. ton CO2­
40
eq. of the total non-quota emissions (measured in calculations, which form the basis for decisions on
CO2 Equivalents, Aarhus University, Institute of a focused and cost-effective climate effort. The
Environmental Sciences) with passenger car traffic ambition being a totally circular resource economy.
as the largest contributor. Of the total emissions We increase the focus on environment and climate
from roads including construction, operation, main- through thorough analyses and cooperation:
tenance and traffic, the latter is by far the biggest
• By working with the environment, nature and
contributor. The emissions associated with the con­
neighbourhood in the planning phase in order to
struction and operation of the road network (which
minimize the environmental impact and the cli­
is the main responsibility and influence of the DRD)
mate footprint of our activities. We also analyse
are in relation to the order of magnitude less. We
how biodiversity can be strengthened at
believe, however, that we need to work on all fronts.
a national level.
When the DRD will reduce the total greenhouse gas
• We highlight the effects on the environment and
emissions in Denmark, it is obvious to see where
climate in transport analyses. This includes the
and how the road sector can contribute to the object­
climate footprint and derived costs, as well as the
ives – both in terms of emissions from traffic and
health-damaging costs of road traffic in terms of
from infrastructure.
air pollution and noise.
The DRD works to reduce the negative impacts of
• We aim at providing an overview of raw materials
infrastructure and traffic on the surrounding environ­
and surplus land and of supply and demand with
ment and nature. It covers the natural environment
the purpose of reducing resource consumption
through the planning and construction of wildlife
and replacing it with waste and residues from
crossings and other measures that protect animal and
production. We prepare an analysis of how to
plant species and enhance biodiversity, it includes
compensate for climate change by exploiting
responsible handling of contaminated soil and
unused land by establishing more forest as well
groundwater, and this includes cooperation with land­
as using solar or wind energy.
scapes and the protection of cultural historical values.
• We aim at choosing solutions and adapting the
In addition, the DRD works with noise reduction.
infrastructure to better exploit new technologies,
Noise from road traffic is a growing problem affect­
for example in vehicles and traffic management
ing thousands of people. Noise nuisance is
the second most important environmental cause of systems, as well as finding specific climate-
illness, as prolonged high noise exposure can be friendly solutions, for example through the choice
instrumental in including stress, cancer and cardio­ of materials and in production. We prepare ana­
vascular disease (WHO 2018). lyses of CO2 emissions for road users’ different
In Denmark, the indicative limit value for harmful driving patterns.
noise is 58 dB. A survey shows that over 700,000 • We aim at planning attractive cycle paths along
the national trunk network and collaborating with
homes in Denmark are exposed to noise from roads
that exceed this limit value, of which over 140,000 the municipal actors.
homes are highly exposed to noise i.e. the noise • We aim at already in the planning phase to
level is at least 10 dB higher than the limit value include support for climate-friendly fuels, e.g. at
(Danish Environmental Protection Agency 2013). service stations, for passenger cars and freight
The DRD is working with the climate and envir­ transport, and to contribute to good interaction
onmental footprint of roads and traffic within the between transport modes.
following overarching phases; Planning, construc­ • Already in the planning phase to include appro­
tion, operation and maintenance, and traffic man­ priate traffic information that can reduce driving
agement. The stakes are formulated in the diversions, queue-driving, etc. and assessment
following terms as current practice, although we are of traffic management as an alternative/
not there yet. It is a mix of “stepping-stones” of supplement.
already existing practises and action plans for the • In order to improve the opportunities for intercon­
work ahead of us. nection we endeavour to establish attractive car­
pooling places in strategic locations.

2.1 Planning phase

2.2 Construction, operation and maintenance
During the planning phase, we assess how the DRD phases
best contributes to meeting society’s mobility needs.
This initially involves examining whether accessibil­ In the construction, operation and maintenance
ity can be increased through, e.g. campaigns, traffic phases, we assess how best to contribute to redu­
management or new technology – before we plan cing the climate and environmental impact of road
physical improvements to existing or completely activities. We incorporate the latest technology and
new infrastructure. innovative solutions, as well as making demands on
We include the latest knowledge on climate and equipment and processes in order to obtain
environmental impacts associated with the life of the a focused and cost-effective climate and environ­
infrastructure from cradle to grave in analyses and mental effort.

41
 We increase the focus on the environment and cli- increased in signal-controlled junctions and at
mate in procurement and through requirements: start/stop. Use of the AI (Artificial Intelligence or
Machine Learning) in traffic light systems.
• We aim at placing higher demands on suppliers in • Planning road campaigns for more environmen­
the procurement provisions to ensure focus on the tally friendly driving habits.
use of sustainable resources, by ensuring lower • Better use of data, including display and sharing
emissions from construction sites and better with road users (passengers and freight transport)
waste management. Further, to develop a model and other stakeholders and actors developing ser­
for CO2 reduction requirements in procurement,
vices for road users.
using an Environmental Product Declarations • Providing support for a good network of refuel-
(EPD) concept for calculation of CO2 emissions ling and charging options for vehicles on climate-
from road projects. friendly fuel and for cooperation with other
• Also, through our procurement, to make demands industries.
on suppliers in relation to the use of machinery • Optimizing and infrastructure for bicycles and
and materials, when we operate and maintain the informing about it, as well as actively working to
roads and other associated facilities and equip­ promote the bicycle as a means of transport.
ment. Thus, increasing focus on digitization and • Establishing a good framework for experimental
machine control. projects that may become relevant for a climate
• We continuously implement more climate-friendly change agenda on roads.
alternatives, such as LED lighting along the roads.
• We work with the environment, nature and neigh­
bour relations, whether in the countryside, in and
2.4 Environmental measures – case: Roskilde
around cities or by sea. The action includes,
Fjord link at Frederikssund
amongst other things: To increase biodiversity
along roads and at service facilities. For all major construction projects, an environmental
• We continuously aim at mapping the CO2 foot­ impact assessment (EIA) is carried out. The purpose
print from our construction and maintenance pro­ of an EIA study is to describe and assess the impact
jects and subsequently implement new experience of a new construction project on the population, ani­
in the planning of future projects. mals, plants and cultural heritage etc. The study will
• We collect, validate and structure officially at the same time ensure that the plan is designed to
approved environmental documentation of prod­ minimize its effects on the environment and to iden­
ucts in connection with the procurement of mater­ tify, where appropriate, measures to reduce and/or
ials from suppliers. compensate for the adverse effects of the
• We engage in international collaborations to installation.
share, exchange and gather knowledge. In 2007-2010, the DRD carried out an EIA study
for a new fjord link at Frederikssund. In the follow­
ing, the measures for the Fjord Link, a new dual-lane
2.3 Traffic management phase carriageway south of Frederikssund between
Marbæk and Tørslev Hage will be described.
The DRD contributes to better mobility and accessi­ A new Fjord Link at Frederikssund was
bility by providing targeted and easily accessible a potential conflict with the designation of Roskilde
traffic information, by managing traffic and using Fjord as a Natura 2000 area. Therefore, in the con­
the latest technologies for optimum traffic flow. text of the EIA study, an assessment of the impact of
We aim at supporting the individual road operator the project was made, considering the preservation
to make it easier to choose a mode of transport with objectives of the Natura 2000 area Roskilde Fjord,
less climate impact, just as we, through cooperation
cf. Article 6 (3) of the Habitat Directive.
with other actors in the sector, aim at supporting the
interactions between different modes of transport. 2.4.1 Habitat directive article 6, section 3 and 4
We support making transport greener by: In general, the Habitats Directive aims to provide
• Providing traffic information that can reduce the basis for the identification and protection of
diversions, queue-driving and improve traffic a coherent European ecological network of special
flow, both concerning congestion and incidents. areas of preservation and protection under the name
Increased use of traffic management for speed “Natura 2000”. At the same time, the protection
harmonisation to prevent queueing. provisions of the Habitats Directive also cover the
• Improved traffic flow (low CO2 emissions) bird protection areas designated under the Birds
through better transport management and traffic Protection Directive (2009/147). In accordance
information. with the directives, several special protection areas
• Using traffic management and latest technologies have been designated in Denmark, including Natura
for optimum flow, both on road sections and in 2000 area No. 136 “Roskilde Fjord and Jægerspris
cities, this, among other things, being important Nordskov”, which includes the habitat areas H120
for freight transport, where fuel consumption is and H199, as well as the bird protection areas F105

42
and F107. It follows from Article 6 (3) of the Habi­ The European Commission’s guidelines on the
tats Directive that, in the context of inter alia: Con­ habitats Directive show that the use of article 6 (4)
struction projects shall carry out an assessment of of the Habitats Directive requires:
its implications for the site in view of the site’s con­
• To examine all feasible alternatives, as regards
servation objectives. If, based on the impact assess­
their impact on the conservation objectives of the
ment, it cannot be denied that the project will harm
site, the integrity of the area and its contribution
the integrity of the Natura 2000 site (site’s conser­
to the global coherence of the Natura 2000
vation objective), the project cannot be imple­
network,
mented unless the conditions set out in article 6 (4)
• that this examination should be carried out with­
of the directive to derogate from the protection of
out regard to costs, potential delay of the project,
the Natura 2000 area are met.
etc., therefore, other assessment criteria, such as
The assessments in the EIA has been carried out
economic criteria, cannot override the ecological
within the framework of article 6 (3) of the Habi­
criteria, and that,
tats Directive in accordance with previous Danish
• in accordance with the principle of subsidiarity, it
practice. According to this practice, in some situ­
is up to the competent national authorities to
ations, in order to avert damage within the meaning
assess the effects of alternative solutions in the
of the directive, within the same Natura 2000 area
area concerned.
and as an integral part of the project, compensatory
measures like designation of replacement nature Furthermore, the guidelines show that the overrid­
areas has been carried out which will fully com­ ing public interest which may justify the implemen­
pensate for the loss of the habitat types destroyed tation of a project, even though the effects on a site
by the project. However, in the European Court of have been assessed negatively, include public health,
Justice the judgement of 14th May 2014 in case regional economic development, environmental pro­
C-521/12, T. C. Briels and others, for the first time, tection and the pursuit of legitimate economic and
adopted a position on the demarcation between art­ social policy objectives. However, the essential
icle 6 (3) and (4) in relation to the application of interests of the public can only be significant if it is
such mitigation measures. The judgment from the of long-term. Short-term economic interests or other
European Court of Justice held, inter alia, that the interests that can only have benefits to society in the
mitigation measures provided for in a project to short term will not be enough to compensate for the
compensate for the negative effects of the project long-term conservation interests that the directive
in question on a Natura 2000 site cannot be con­ carries out.
sidered in the assessment of the impacts of the pro­
ject in question, as provided for in article 6 (3). 2.4.2 Impact assessments
Thus, the European Court of Justice held that miti­ The project for a new crossing at Frederikssund is in
gation measures which are not intended to avoid or the Natura 2000 area Nr. 136, which includes,
reduce the negative effects of the project on the respectively, Roskilde Fjord (Habitat area SAC120
natural habitat type, but subsequently to compen­ and bird Protection Area SPA105) and Jægerspris
sate for its adverse effect (e.g. by laying down nat­ large forest (Bird Protection Area SPA107).
ural substitutes), are compensatory measures
within the meaning of article 6 (4). It is therefore
a situation which must be dealt with under the
derogating procedure of the directive, which pre­
supposes that the conditions laid down in article 6
(4) are met.
Of article 6, paragraph 4 shows that “if a plan or
project, despite the fact that the effects on the site
are assessed negatively, must nevertheless be carried
out for imperative reasons of overriding public inter­
est, including those of a social or economic nature,
because no alternative solution exists, the Member
State shall take all compensatory measures necessary
to ensure that the global coherence of Natura 2000 is Figure 2. Map of Roskilde Fjord and the new crossing
(1:25.000).
protected. The Member State shall inform the Euro­
pean Commission of the compensatory measures
The green area illustrates the habitat area SAC120
taken. In the case of a site with a priority natural
and bird protection area SPA105. The yellow area
habitat type and/or a priority species, reference may
illustrates the placement of marsh land and chalk
be made only to considerations of human health and
grass lands also protected under Natura 2000.
public safety or to significant beneficial effects on
Roskilde Fjord belongs to the Maritime Atlantic
the environment or, after consulting the Commis­
region and is dominated by large marine areas with
sion, to other imperative reasons of overriding
large shallow bays and coves, as well as areas with
public interest.”
43
shallow water and sand dunes. These areas are desig­ Overall, the compensatory measures for the impact
nated as special areas of conservation. The areas are of habitats on the Natura 2000 area designation basis
covered with eel grass and mussels, which are also is planned from different strategies as follows:
characteristic of Roskilde Fjord.
In the area there are several natural habitats • Compensation within existing Natura 2000
which, by virtue of their size or rich flora, are of delimitation:
• The laying of natural substitutes on land not
regional or national importance. This is largely the
designated with a habitat type and not subject
case with the large continuous salt marshes areas
to nature measures for the management of the
along the fjord, as well as the priority natural habitat
Natura 2000 site.
type calciferous dry grassland and scrub.
• In principle, in order to take account of the
All terrestrial nature types belong to the continen­
uncertainties and development time of estab­
tal biogeographic region, while the marine part of
lishing new nature, the area ratio is 1:2.
the Natura 2000 area belongs to the marine Atlantic
• In order to support the development of natural
region.
nature, nature care is be carried out.
Roskilde Fjord is one of Denmark’s most import­
• For natural habitats that are developed very
ant breeding grounds for waterfowls. In the shallow
slowly, additional compensation (overcom­
fjord lies about 30 small islands and islets.
pensation) is carried out in the form of larger
Every year, on the islets in the fjord, breed 10­
area, extra care of existing areas, etc.
20,000 pairs of birds spread over 25-30 species, such
as the avocet, the tern and the artic tern. • Compensation outside existing Natura 2000
demarcation:
• Laying of substitute nature
2.5 Compensational measures for terrestrial • Nature care of areas with poor natural state to
nature achieve favourable status and involvement in
As part of the basis for the nature impact assess­ the Natura 2000 area
ments for the selected alignment, it is provided that
The compensatory measures must be measures
the permanent influences, including the permanent
which are beyond the nature-enhancing efforts planned
land involvement and a change in the state of certain
for the management of the Natura 2000 site. Among
areas e.g. shadow effect, can be compensated for.
other things, solutions were implemented to establish
Compensatory measures are carried out for perman­
compensatory measures in connection with the marsh
ent adverse effects of the following habitat types:
lands and the chalk grass land in connection with Ros­
• Marsh lands (land inclusion 0.37 ha) kilde Fjord.
• Chalk grass lands (land inclusion 0.16 ha) The implementation of compensatory measures is
planned in cooperation with the municipality of Fre­
Compensatory measures are defined in the Guide derikssund, but also other relevant municipalities
to Decree no 408 of 1 May 2007 as follows: was involved to find the most appropriate areas. In
• In cases where a derogation is necessary, all the addition, the National Nature Agency is involved to
necessary compensatory measures must be imple­ ensure optimal coherence with the national planning
mented in order to compensate for damage that in the Natura 2000 area, while ensuring that there is
could not be prevented or prevented by integrat­ no overlap between the activities of the authorities
ing remedial measures into the plan or project. and the planned compensatory measures.
Compensatory measures are thus different from For the two terrestrial habitat types, chalk grass
the defensive measures by having to “repair” the lands and marsh lands, where an impact from the
damage. project is to be compensated, the planning takes into
• Compensatory measures may be taken within the account the affected areas:
Natura 2000 area affected or in another Natura • Geographic location in Natura 2000 sites
2000 area. The essential thing is that the coher­ • Latest results of monitoring for species content
ence of the Natura 2000 network is maintained. and estimated natural state
• The compensation measure must normally be
completed before the damage resulting from the At Tørslev Hage, the natural state of both chalk
plan or project occurs. grass lands and marsh lands in 2010-2011 is rated
moderate (natural state 3, species index/Structure
Compensatory measures for marsh land, and index 4/2 and 2/4). On the east side of the fiord, the
chalk grass lands are carried out as a result of expos­ natural state of the affected marsh is rated as poor
ure in the form of land involvement, altered hydro­ (natural state 5, species index/Structure index 5/4).
logical conditions and altered species composition/ The establishment of marsh land was planned by
population of characteristic and shadow influences. first locating potentially suitable areas on maps
It focuses on creating a coherent relationship in the where there has previously been marsh, but where
Natura 2000 area, as well as on the use of pathways drainage and cultivation has displaced the habitat
for terrestrial habitats. type, or areas where the hydrological conditions due

44
to rising water levels in Roskilde Fjord have vegetation for chalk grass lands, it may be necessary to
changed. The most suitable area or areas was attempt the spreading of hay from existing grass lands
selected based on an assessment of: in the area to simulate the spread of seeds.
The monitoring program is carried out annually
• Probability of the area being able to develop during the first five years following the laying of the
a beach meadow of at least similar natural state, areas and thereafter at a larger interval until the com­
• the location in relation to existing beach meadows pensation areas have reached the planned state and
and hence the likelihood of spreading characteris­ may be designated as habitat type, according to
tic species to the new area,
which the areas are part of the overall management
• the time it will take before the area can be desig­ of the Natura 2000 area.
nated as the nature type marsh lands.
Land owner conditions for the selected areas was 2.7 Impact assessment and mitigation measures
investigated and the owners contacted for possible in relation to birds
cooperation and permission to inspect. The selected
area was inspected and recorded in accordance with Breeding birds which is part of the designation basis
DEVANO mapping method for habitats. In addition, for Roskilde Fjord as Community Bird Protection Area
the area’s hydrology and nutrient status/soil condi­ Nr. 105 (Roskilde Fjord, Kattinge Vig and Kattinge
tions are recorded. Based on this study, nature restor­ Sø), includes specified species that will potentially be
ation measures were planned and implemented, affected by the construction of a new Fjord link.
which for marsh land will primarily be the closure of Many existing and possible breeding sites lie to the
ditches and canals that drain the area. south, at Jyllinge Holme and Eskilsø more than 6 km
The initial planning of compensation areas for the away from the bridge site. To the north all of which are
chalk grass lands was carried out. In addition to existing or potential breeding sites are located at
identifying possible areas for the recovery of chalk a distance from 600 m to about 5 km away from the
grass lands inside and outside the Natura 2000, the bridge site.
area of Roskilde Fjord is also explored the possibil­ Outside the breeding season, a very large number
ity of improving the condition of existing grass lands of birds in Roskilde Fjord rooster and forage. The
that are not already subject to the duty of care in number varies according to the season and weather
Authorities. For example, it may be privately owned conditions, which is why there are large fluctuations
chalk grass lands outside the Natura 2000 area. in the number of birds.
As the development time for a new grass land is In the context of the EIA Studies and the develop­
very long, the assessment of the suitability of the ment of the bridge design prerequisites for the pro­
areas was specially sought for regions which, so to ject, there has been continuous adjustment and
speak, ‘have grown out of protection’ but not des­ adaptation of the project in order to provide add­
troyed by cultivation and accompanying fertilization. itional assurance that the Roskilde Fjord as a Natura
The planned nature recovery restoration measures 2000 area is not harmed, including the extent of dif­
were approved by the relevant authorities. ferent constructions, both temporary and permanent.
In order to avoid any harm from the project, sev­
eral measures have been taken to avoid the effects
2.6 Monitoring of compensatory measures on natural habitats and birds which may be affected
A monitoring program is carried out to check the com­ to varying degrees by a new fjord crossing.
pensatory measures after their implementation. The • Construction phase:
method of monitoring will follow the national method • The project respects the distance to the near­
for mapping terrestrial habitats (DCE TA-N03). est breeding area for e.g. more than
In the light of the monitoring, both the develop­ 300 m from the project area.
ment of the compensatory areas and the overall • Impacts from construction activities are as low
impact of the project on the coherence of the Natura as possible, e.g. in connection with ramming of
2000 area are evaluated. piles.
Monitoring and assessment of the development of • There is shielding of sites on land, thereby
the planned habitat in the compensation areas is used limiting the disruptive activities of the con­
to assess whether further measures are needed to struction activities.
improve the physical or biological conditions in the • Excavation is carried out within shields in
area. Thus, it may be appropriate to carry out/intensify a way that causes the least possible impact on
mowing or grazing in order to promote a light-open habitats, the feeding areas and food source for
and varied vegetation both on marsh land and on chalk the breeding birds.
grass lands as further action. The fight against particu­ • The aim is to make the construction activities
larly dominant/invasive species in the early succession on the fjord itself as short as possible and to
phase of the areas may also be necessary to ensure the carry out as many activities as possible out­
development of the natural habitat type envisaged. In side the usually most cold winter months of
order to encourage the development of a characteristic January – March.

45
• Operation phase: The size of the area temporarily to be used, for
• Design is carried out to limit collision for the quay and temporary dam during the construction
birds. phase was defined as a total of 2 hectares in order to
• Noise screens are not transparent, otherwise it limit the impact on feeding and resting areas.
will increase the risk of birds colliding with A requirement in the tender documents was
the structure. included to execute the bridge foundations in coffer­
• The lighting on the bridge and during periods dams. With these conditions for the foundation
of fog are minimized in order to avoid collision. works of the bridge, it was possible to excavate inde­
pendently of the time of year. And thereby avoid the
Monthly bird counts of roosting birds is carried
most important months for migrating birds.
out in the part of the bird protection area affected by
the construction works. The bird counts are com­
menced 2 years before the start of construction
3 TECHNICAL CONSIDERATIONS
work. Thus, eight counts are carried out annually in
the period September-April, i.e. during the period
During development of the technical specifications
when the number of roosting birds in Roskilde Fjord
for a major bridge a large number of aspects have to
is at its peak. The counting continue for a 2-5 years
be considered. The first thing is to write a project
period after the construction phase is ended.
specific Design Basis, as was previously done also
for the Great Belt Bridge and the Øresund Bridge in
2.8 Measures concerning the marine habitat types Denmark. This is done to compile the design
The bridge project will mean the inclusion of the requirements that are not explicitly stated in the
seabed, and hence make a permanent and temporary Eurocodes, design requirements which originate
impact of the natural habitat type ‘larger shallow from the Danish annexes to the Eurocode, require­
bays and coves. It was decided that the permanent ments from Danish Design Guides (Danish: Vejreg­
impact should not exceed an inclusion of 0.5 hec­ ler), as well as project specific requirements.
tares of the natural habitat type ‘larger shallow bays The Design Basis furthermore includes the results
and coves, corresponding to one half per one thou­ of special investigations carried out for the bridge or
sand (0,05 %) of the natural habitats of the Natura other similar major structures. In cases where such
2000 area. It is considered that an area involvement investigations caused a request for modification of
of this magnitude will not harm the integrity of the a standard clause, the relevant departure from Stand­
habitat in the Natura 2000 area. The 0.5 hectares ard, is enclosed in the Design Basis.
could carry the total of 15 bridge foundations. Also, through the development of the Design
In the tender documents for the design and build Basis, several specific design issues are investigated.
contractor, a few requirements were specified that Some of the exceptional loads, which is studied such
could not be changed. Thus, it was a requirement as ship impact and ice loads, as well as road and rail­
that only 15 piers could to be constructed in the fjord way loads, temperature loads and geotechnical con­
to reduce the permanent footprint on the Natura siderations. As compliance with these requirements
2000 area. Also, the permanent footprint of the 15 all are associated with significant costs, these are
foundations had to be less than 0.5 hectares in total. carefully considered and questioned prior to incorp­
During the construction phase, the excavation of orating in the Design Basis.
the seabed for the foundations took place within
a protective measures, sheet piling/cofferdams, in 3.1 Design basis and considerations – case:
which case there was practically no sediment spill­ Storstrøms bridge
age and no impact that could damage Roskilde
Fjord’s marine habitat types, which are nearby i.e. The Design Basis [4] developed for the New
bays and coves, waders and sand dunes, as well as Storstrøm Bridge and as for the Roskilde Fjord Link,
there was no impact on the bottom flora e.g. eel has already been tested through the competitive dia­
grass and the fauna on the sea bed. logue with the contractors – but has also proved to
Thus, with the cofferdam method for the covering be an excellent document to clarify the requirements.
of the bridge foundations, it was possible to excavate The following is limited to issues such as ice
independently of the time of year as there were no loads, ship impact and geotechnical conditions:
impact from sediment on the eel grass population.
Otherwise excavation should have been carried out 3.1.1 Ice load
during winter time to avoiding impacting on the The thickness of the ice to be considered, has
growth of eel grass. But this would have conflicted a significant impact on the design, and should be
with the migrating and breeding birds, that are pro­ considered in detail for the individual project. The
tected. But excavating now could take place outside first estimate took the observed temperature increase
of the period of the migrating birds. between 1900 and 2014 into account, and the ice
The excavated material is moved by barge out of thickness at the location of the New Storstrøm
the fjord to a dumping area in the Kattegat. Bridge, with 50 years return period, was assessed to
be 56 cm for the year 2015. Based on the Met Ocean
46
Report for Storstrømmen. However, the future tem­ 3.1.2 Ship impact
perature changes are expected to decrease the ice The new Storstrøm Bridge is a railway bridge, which
thickness at Storstrømmen during the service life of means it will have to obtain an APIS (authorisation
the bridge. A study was carried out in order to esti­ for placing into service), in accordance with the
mate the future ice thickness at Storstrømmen and European Commission Implementing Regulation
investigate the design ice thickness from a reliability (EU) No. 402/2013 of 30 April 2013 on the common
perspective (IABSE 2018). safety method for risk evaluation. This includes
Three scenarios were investigated: a systematic identification of all hazards of relevance
during the operation of the railway, including events
1) based on historical data collected from 1870 – like ship impact. In accordance with EU 402/2013,
2014 with no climate change expected, the risk acceptability of each hazard shall be evalu­
2) minimum - and ated by using the three risk acceptance principles (in
3) maximum expected change in global average sur­ the following order of preference) until an appropri­
face temperatures, respectively. ate principle is found for the individual hazard being
The conservative approach would be to base the evaluated:
design on the minimum annual reliability and not 1. The application of Codes of Practice
include the favourable effects of future climate 2. Comparison with similar systems (ref. system)
changes after 2022, but this comes at a significant 3. Explicit risk estimation
cost.
Application of the accumulated reliability It was considered very unclear how the Contrac­
approach requires some level of future risk mitiga­ tor’s Designer, would be able to document, to an
tion to avoid reduction in human safety. However, acceptable level, how the risk acceptability for haz­
since an ice cover of 34-52 cm does not come ards such as ship impact were appropriate. As simi­
unwarned, there will be plenty of time to take appro­ lar reference systems were hard to identify, and it
priate action if necessary. In the event of occurrence was considered that an explicit risk estimation
of the design ice load, if necessary, train speed approach could be challenging and time consuming
reductions can be enforced, or preventative measures to reach agreement on during the Detailed Design
can be put into place. Thus, it was decided to reduce phase, the preferred approach would be to apply
the design ice thickness to 49 cm in the Design Basis a Codes of Practice.
(DBS 01), for this specific bridge. However, it is not clearly defined in Eurocodes, if
the general requirements for structural safety which
are applicable to normally occurring loads, also
applies to accidental loads in case no level of accept­
able risk is given in national annex, ref EN 1991-1-7
§3.2 (DK/EN 1991). The reason behind this could
partly be the very rare occurrence of these accidental
loads, and partly also the very high costs associated
with the compliance of general requirements to
structural safety which are applied to other loads.
As the Danish Road Directorate and Railnet Den­
mark (Danish: Banedanmark), already had done
some work in relation to structural safety levels for
Figure 3. Table from Storstrøms broen, future ice thick­ existing bridges with regard to ship impact, it
ness, Design Criteria Background Note (IABSE Sympo­ seemed a good opportunity to develop common
sium Nantes 2018”New Cable Stayed Bridge Across guidelines for new bridges, as well as existing
Storstrømmen”).
bridges. A task force was established, to develop
a new National Annex.
The term accidental loads are for the Danish
National Annex to be understood as rare occur­
rences, such as ship collisions, vehicle impact and
other accidents which might affect a bridge’s struc­
tural safety, including explosions. However, fire and
seismic events are not included in the scope as they
are covered by separate codes. The methodology
adopted in the new National Annex, for the risk
acceptance criteria is described, as a choice between
two possible methods of risk evaluations:
Method I – is the use of the same requirements of
structural safety levels as used for normally occurring
loads, applied to accidental loads. Applying probabil­
Figure 4. DBS 01 – Table 6-8 Ice Properties (DBS 01).
ity of structural failure of 10-7 per for CC3 structures.
47
 Method II – is developed using a cost-benefit requirements for placements of the piers, which
approach, where the acceptable risk level is seen in could be reduced to 65 m span length with no
relation to the balance between the public gain by requirements on an upper bound.
the risk reducing measures and the costs associated Considering all the above-mentioned factors, it
with this. The public gain is measured as the reduc­ became apparent, that DRD needed to develop
tion of probability for structural collapse, combined a model that provided the necessary control of the
with the consequences of a collapse primarily in geotechnical basis to mitigate all the associated risks.
terms of user safety and cost of disruption and repair The project would in any case need to handle changes
of the new Storstrøm bridge. This principle is also on the geotechnical basis when the Contractor had fin­
known as a variant of the ALARP principle (As Low ished the supplementary geotechnical investigations
As Reasonably Practicable), where the risk is for the detailed design. The normal approach, with
reduced to the lowest reasonable level seen in rela­ bid evaluation and subsequent grading would not suf­
tion to the cost implications, associated with this. fice, no matter the evaluation criteria, as the price is
Application of Method II is subject to limits on so often governing. Basically, we could end up award­
the risk of loss of human lives. Represented by min­ ing the contract to a project with a very high geotech­
imum requirements of loss of human lives of 10-6 nical risk and possibility for claims.
and 10-5, for railway and road bridges respectively, Discussing the possible solutions, it quickly
as annual probabilities. became clear to the DRD, that we needed a model
It should also be noted that the application of that contained at least the following necessary
Method II can necessitate implementation of mitiga­ elements:
tion measures or monitoring plans throughout the 1. Would establish a clear and unambiguous basis
structures service life. The onerous will be on the for comparison
owner of the structure to monitor and actively follow 2. Created a firm framework for settling the eco­
up on the effectiveness of the implemented systems. nomic consequences
In general, it is expected that using Method I will lead 3. In any case would be fair to both Contractor as
to a more expensive design, than Method II, which is well as Employer
considered reasonable, as by using Method II a more
detailed analysis is carried out. An analysis which con­ A model as described below was developed.
siders the public risks by the loss of the bridges and
a detailed analysis of possible risk reducing measures. 3.2.1 Basis for comparison
For major structures, a zero-risk level, is imprac­ Establishing an unambiguous basis for comparison
ticable and in nearly all cases it is necessary to on a geotechnical project where the design is
accept a certain level of risk. In January 2016 a new unknown was initially no easy task. Under no cir­
Danish National Annex, DK/EN 1991-1-7 NA was cumstance should the model restrict the Contractor
published, together with a background note on risk in his preferred method.
acceptance criteria (Background note 2016). This The approach was ultimately dictated by the fact,
was a direct result of the work done for the Design that only a few geotechnical design parameters are
Basis (DBS 01) for the new Storstrøm Bridge. decisive for the solution. With this knowledge in hand,
DRD decided to Baseline the few governing geotech­
nical design parameters, leaving the rest up to the
3.2 Geotechnical risks Contractor. The geotechnical design basis (DBS 02)
The geotechnical investigations for the Storstrøms then preceded any information in the available geo­
bridge tender project was based on the illustrative technical interpretative reports (DBS02-A01, 2015).
design, which applied both direct foundation as well The parameters in the geotechnical design basis (DBS
as foundation on improved ground with inclusion 02) was referred to as Baseline 1 (for tender design).
piles. The Baseline 1 parameters included shear
Because of the geotechnical investigations being strength, pre-consolidation stress as well as param­
conducted simultaneously with the development of eters for settlement analysis, ending up with one
the illustrative design, and the presence of high volt­ unique profile for each section of the alignment.
age cables in the northern part of the alignment, the A description (ILD 05) of the applied approach that
geotechnical basis for tender ended up consisting of eventually lead to the Baseline 1 parameters was
far less information than originally planned. Of the included as information in the tender documents. This
47 pier positions including abutments, only approxi­ was done in order to make it clear for the Contractor,
mately half ended up with information as intended what to expect for the detailed design phase, and even­
for the tender, whereas the remainder had to be tually allow the Contractor to mitigate risk in the
based on information from adjacent piers or even tender phase.
information gathered up to +150 m from the pier The approach was repeated in the detailed design
positions. phase, including the results of the supplementary
The specifications included several degrees of geotechnical investigations. The outcome of this
freedom for the Contractor, where the most decisive analysis was a similar set of geotechnical design
for challenging the geotechnical basis was the parameters and was referred to as Baseline 2.

48
3.2.2 Settling the economic consequences than ± 2 % would result in additional or reduced
Analysing if Baseline 2 lead to a change compared payment.
to Baseline 1 was straight forward. Were the param­ Two possible types of change were defined, devi­
eters not completely identical, there was a change. ating soil conditions and unforeseen soil conditions,
However, settling the economic consequences are each being settled using the same unit prices.
not easy, as some changes could be insignificant Deviating soil conditions was defined as being the
whereas others in worst case could lead to result of the Contractor’s supplementary geotech­
a complete change in the foundation principle. nical investigations showing that the soil conditions
in Baseline 2 deviated from Baseline 1.
Unforeseen soil conditions were defined as being
significantly adverse soil conditions which were
identified during the Contractor’s permanent founda­
tion works.
A deviation soil condition could not also be
defined as an unforeseen soil condition. Hence, only
one financial claim may arise from the soil condi­
tions concerned.
With the applied model for basis of comparison
and method for settling the consequences, DRD
believes that we to the furthest extent possible suc­
Figure 5. Illustrative example of the Baseline comparison. ceeded in applying a risk mitigation tool, that proved
Notice that the undrained shear strength profile and fair to both Contractor as well as Employer.
pre-consolidation stress for Baseline 2 (red) has decreased
compared to Baseline 1 (grey). A change with economic
3.3 Geotechnical reflections so far and lessons
consequences must be settled.
learned
The “Storstrømsbroen” project is ongoing and will
The solution to the problem was decided on the be completed in 2022. By this time, it is still far too
basis of an analysis of what the actual consequences early to conclude if the applied risk mitigation man­
would be. Did the change eventually lead to agement tool has proved its worth through the whole
a smaller or larger foundation, did the gravel bed process, or if it would have been beneficial to
below need to have a smaller or wider extent, was include or correct elements. There are however
the foundation principle changed from direct to e.g. already some lessons learned and reflections, that are
piled foundation. These consequences were eventu­ worth keeping in mind if a similar model is applied
ally broken down into quantities and unit prices. on another project. The most important are:
DRD had 9 different types of quantities related to
foundation works (TAG), which was based on the 3.3.1 Positive
illustrative design. The Contractor was, for each The competitive dialogue proved to be an excellent
individual foundation, asked to fill out a detailed opportunity to ask the different Contractors for feed­
specification of quantities and unit prices for the back on the applied model, by this also making sure
foundation structure, allowing him to create new that the model was well received and understood.
items or remove items, if the Contractors design dif­ During the different dialogue stages, certain elem­
fered from the illustrative design. ents of the model were improved as per the request
When Baseline 2 was available, the Contractor or suggestion of the different Contractors. This
had to document the change in quantities by calcula­ strengthened the model, but also created a certain
tions, where the difference in quantities should be understanding between the parties.
determined as the difference in quantities calculated As Employer, you are in control of the governing
on the basis of Baseline 1, and the final quantities parameters for the geotechnical design, making it
calculated on the basis of Baseline 2. quite easy to include a certain amount of robustness
An important limitation to the disadvantage of the around your chosen tender baseline. Pending on
Contractor was, that the Contractor was per the your gut feeling, the budget security, possibility of
tender requirements not entitled to any further eco­ conflict etc. you can adjust your baseline parameters
nomic settlement, than the change in quantities to the more or less conservative side, if you find this
defined. Thus, increased complexity or more difficult to be beneficial or necessary.
conditions for the work, would not endorse further The Employer must develop a geotechnical Base­
compensation. line based on the results of Contractors supplemen­
As the calculations for documentation of change tary investigations. This means, that the Employer
would be a time-consuming effort for the Contractor must be fully confident that the extent and quality is
to produce, and time-consuming for the Employer to enough for the purpose. For this reason, the
review, at threshold was determined. Only changes Employer must write a thorough description of
that eventually lead to a change of payment more the requirements to the necessary level of the
49
supplementary investigations, and make sure with chalk grass lands and to avoid harm from the project
inspections that the requirements are fulfilled. DRD several measures were taken to avoid the effects on
sees this as a positive outcome of the approach, natural habitats and birds.
giving the Employer more control over the level of The article included examples of the optimization
geotechnical investigations and eventually the of the design assumptions, considering the risks
robustness of the design. associated with the geotechnical conditions and rele­
vant load conditions such as ice and ship impact.
3.3.2 Negative And the examples are taken from the new Roskilde
At certain stages in the design process, the Employer Fjord Link and the coming road- and railway bridge
is deeply involved in the project, and can ultimately crossing Storstrømmen. The article illustrates how
end up costing time delays maybe even with eco­ the process of competitive dialogue was utilized, as
nomic consequences. a tool for making projects clear and understandable
Independent third-party checkers may begin ques­ for the contractors, whilst at the same time allowing
tioning the model, if this is not satisfactorily contractors to put forward their view on the project.
explained. This is a huge risk for the Employer, as In Denmark we have a long tradition of learning
the consequence of the approach not being approved from previous major construction projects such as
for detailed design could be catastrophic, eliminating the Great Belt, Øresund Link and others, which the
the entire purpose. technical optimization benefit from. However, the
requirements to preserve climate and protect envir­
onment opens up for new types of challenges in
4 CLOSURE
large construction projects and it is essential that we
develop and find new solutions, so we can overcome
The above mentioned gives a small insight into how
these challenges in the future.
the Road Directorate works with the development
and execution of construction of large infrastructure
projects and the DRDs framework for reducing cli­
mate and environmental impact. A large number of REFERENCES
issues must be addressed and often opposing chal­ Background note – Accidental Loads for new Bridges. Risk
lenges have to be solved. Today, the most pressing Acceptance Criteria. January 2016.
tasks to be solved are the environmental and climate 93200-COW-REP-3-HYD-00042 Met-ocean report for
impacts and the challenge of fulfilling the norms and Storstrømmen.
standards respectively under considerable optimiza­ 93200-COW-NOT-4-DBS-00280 STORSTRØMSBROEN,
tion of the constructions for several reasons such as FUTURE ICE THICKNESS, Design Criteria Back­
costs, aesthetics, local conditions etc. ground Note.
In the planning phase, implementation of how the Danish Environmental Protection Agency. 2013. National
kortlægning af boliger belastet af vejstøj i 2012.
project best can contribute to meeting society’s need
Arbejdsrapport fra Miljøstyrelsen nr. 5. Report from the:
for mobility must be implemented, which is often https://www2.mst.dk/Udgiv/publikationer/2013/08/978­
regarded as a prerequisite for economic growth and 87-93026-45-2.pdf
human well-being. However, the planning must also DBS 01 – Design Basis - Bridge. Storstrømsbroen. Entre­
include the latest knowledge on climate and environ­ prise 93200.001.
mental impact associated with the infrastructure life DBS 02 – Design Basis – Geotechnical. Entreprise
from cradle to grave - and add knowledge that can 93200.001
form the basis for decisions on focused and cost- DBS 02-A01 – FUGRO. Interpretative Geotechnical
effective climate action. The planning shall also Report on Ground Investigations. September 2015.
DK/EN 1991-1-7NA National annex to Eurocode 1: Load
include factors that determine how biodiversity can
on construction works - Section 1-7: General loads ­
be maintained as well as other considerations for the Accidental load.
surrounding environment, including protected nature IABSE Symposium Nantes 2018. ”New Cable Stayed
and habitat areas. Bridge Across Storstrømmen”
The article illustrates how the DRD worked with ILD 05 – Reader’s Guide to geotechnical interpretative
nature, the environment and the surrounding area in report. Entreprise 93200.001
relation to the Roskilde Fjord Link in Frederikssund, TAG – Method of Measurement and Basis for Payment.
were there was a potential conflict with the designa­ 93200.001 Storstrømsbroen.
tion of Roskilde Fjord as a Natura 2000 area. The UNDP: UN Sustainable development goals: https://www.
undp.org/content/undp/en/home/sustainable-develop
DRD carried out an environmental impact assess­
ment-goals.html
ment (EIA), which described and assessed the WHO. 2018. Environmental Noise Guidelines for the Euro­
impact of a new construction project on the popula­ pean Region – Executive summary. http://www.euro.
tion, animals, plants and cultural heritage etc. who.int/en/health-topics/environment-and-health/noise/
A number of compensatory measures were imple­ publications/2018/environmental-noise-guidelines-for­
mented to ensure adverse effects on marsh lands and the-european-region-executive-summary-2018

50
 Bridge Maintenance, Safety, Management, Life-Cycle Sustainability and Innovations –
Yokota & Frangopol (eds)
© 2021 Taylor & Francis Group, London, ISBN 978-0-367-23278-8

Some lessons of more than 20 years of inspection, maintenance

and rehabilitation of bridges in Spain
F.M. Mato
Technical University of Madrid, Madrid, Spain
IDEAM SA. TYLIN INTERNATIONAL GROUP, Madrid, Spain

I.P. Sánchez
IDEAM SA. TYLIN INTERNATIONAL GROUP, Madrid, Spain

ABSTRACT: Spain has implemented a Management System for bridges belonging to the national road net­
work since 1998. Since 1999, IDEAM has carried out for the Spanish public administrations more than 8,000
principal inspections; 150 specialized inspections; 200 structural damage evaluation reports; 200 repair and
strengthening projects and 150 drafts of maintenance plans for structures.
In most cases, the pathologies detected derive from construction failures or deficiencies in the conception
of some details that could have been easily remedied or even anticipated, namely: drainage and waterproofing;
duct grouting in post-tensioned concrete girders; internal corrosion of weathering steel box girder bridges;
corbels; bearings; corrosion of tendons of stays belonging to the first cable-stayed bridges generation; prob­
lems arising from fatigue in steel bridges under increasing heavy traffic forces, etc.
This lecture focuses on the description of some relevant practical case studies related to the aforementioned
topics

1 INTRODUCTION • Primary or principal inspections. They are con­

ducted every 5 years -except for special situations-
Spain, like most developed countries, has implemented by specialized engineers using auxiliary means for
a Management System for bridges belonging to the them to be able to visually inspect all bridge elem­
national road network since 1998. The regional net­ ents, characterize and assess damages.
work is currently being implemented in some autono­ • Special inspections. They are specific of each struc­
mous communities. ture and provide an individual rating of pathologies
The Management System comprises the following diagnosis. Generally, they are complemented by
concepts: surveys, safety and functionality structural analysis,
damage identification maps and, depending on the
• Structures inventory case, the structure repair or strengthening.
• Periodic inspections
• General Maintenance actions, mostly preventive. IDEAM has performed many primary inspections
of bridges for the Spanish public administrations
Inspections are rated depending on the intensity, since 1999. We have inspected more than 8,000 struc­
technical specialization and frequency. As in most
tures over all these years and are a key collaborator
Management Systems, three inspection types may be
with the Ministry of Public Works in the development
distinguished:
of the ‘Guide for Bridge Inspections’ and the ‘Guide
• General or regular inspections. They are generally for drafting Maintenance Plans of Bridges’.
conducted on an annual basis, coinciding with Since 2002, we have also carried out more than:
general maintenance actions. General inspections, • 150 specialized inspections;
consisting of a visual verification of the structure
• 200 structural damage evaluation reports;
and its main components, are performed by non- • 200 repair and strengthening projects;
specialized staff and don’t require specific auxil­ • 150 drafts of maintenance plans for structures.
iary equipment.

DOI: 10.1201/9780429279119-5
https://doi.org/10.1201/9780429279119-5

51
 In very few cases, the pathologies detected were concerns of engineers involved in the design and con­
the consequence of failures in structural analysis or struction phases on the upgrade of conservation and
the design and dimensioning of a certain element. In maintenance conditions of this type of structures. As
contrast, they mostly derive from construction failures they are an important part of the heritage value of
or deficiencies in the conception of some details that developed countries’ infrastructures, they require con­
could have been easily remedied or even anticipated, tinuous investments in conservation, which are hardly
namely: sustainable if bridge designs aren’t upgraded from the
maintenance perspective (in Spain, there are around
• drainage and waterproofing;
18,000 km of conventional railway lines, approxi­
• duct grouting in post-tensioned concrete girders; mately. 4,500 km of high-speed lines, around
• internal corrosion of weathering steel box girder 170,000 km of roads and highways containing a total
bridges; of approximately 40,000 structures and is therefore
• corbels; estimated that the order of magnitude of the structural
• bearings. heritage is above 18,000-20,000 million euros).
Particular attention should be paid to other cases,
like the replacement of stays belonging to the first 2 DRAINAGE AND WATERPROOFING
cable-stayed bridges generation or problems arising
from fatigue in steel bridges under increasing heavy Based on our extensive experience of more than 20
traffic forces. years in bridge inspection, water is certainly the main
Due to space constraints, pathologies deriving cause of pathologies and needs for bridge strengthen­
from foundation scouring issues of piers and abut­ ing and repair.
ments located in riverbeds do not fall within the What’s striking to us is that this is broadly
scope of this article. Nonetheless, scouring is one of acknowledged by engineers and authorities involved,
the most important elements to be considered during and yet, no substantial improvement is being made
inspection (often very complex) and old bridge main­ in bridges constructed in the last 10 or 15 years. On
tenance due to the potential risk of brittle fracture. the contrary, errors in the design of drainage, dis­
charge and waterproofing details, as well as their
associated pathologies are continuously repeated.
Additionally, it is important to stress that risks
associated to water don’t affect steel bridges only.
Indeed, reinforced or prestressed concrete bridges are
risk-sensitive and, in many cases, repair processes of
those pathologies are more costly and complex.

Figure 1. Scouring of foundations.

Figure 2. Scouring of foundations.

The cases described below, most of them regarding

to bridges that were designed and built less than 25 or
30 years ago, aim to draw attention and focus the Figure 3. Water leakage through expansion joints.

52
 The origin of those pathologies is mainly the Water circulation through the structure may
following: cause:
• water tightness failures under expansion joints in • concrete carbonation processes, including PH
abutments or, mostly, between isostatic spans; loss, that may accelerate corrosion of
• drainage system failures and rainwater discharge reinforcements in permeable concrete due to
on the platform sides; low compactness;
• water tightness failure or deficient design of • surface concrete deterioration with cracking and
downspouts and sumps sewers; spalling due to expansion processes deriving
• waterproofing design or placement failures, or from reinforcement corrosion;
frequently, damages in the waterproofing system • chloride, sulphate or other salts corrosion pro­
during subsequent pavement replacement or cesses due to the concrete contamination through
strengthening operations. fissures, joints between concrete layers, etc.
(these processes are very intense when de-icing
salts are used);
• intense development of alkali-aggregate reac­
tion expansion processes with the presence of
aggregates susceptible to such chemical
attack;
• corrosion processes under the active reinforce­
ments pressure, referred to further on.

Figure 4. Missing elements in drains.

Figure 6. General corrosion due to concrete carbonation.

Serious pathologies recently detected on an

overpass spanning one of the main thoroughfares
into Madrid with an ADT of approximately
80,000 vehicles per day are described below. The
severity of the damages made it necessary to
completely demolish the deck and replace it for
a new precast deck, which will have a significant
impact on the accesses to Madrid for nearly 9
months.
The structure inspection was undertaken when
significant deviations from the deformed align­
ment with respect to the theoretical elevation
were detected. Such inspection revealed serious
alkali-aggregate reaction issues, having as
a consequence an important reduction of up to
60% of the deformation modulus of concrete
and intense expansion processes associated to
general longitudinal cracking processes at the
Figure 5. Water filtrations under the platform’s side bars. deck.

53
Figure 7a and b. Serious alterations at the alignment of the
overpass located at Colmenar & M40 junction (Madrid).

When the pavement layer was lifted, it was detected

that the subsequent replacement processes of the road
surface had virtually removed the waterproofing layer,
meaning that the platform water entered directly
through the cracks due to alkali-aggregate processes,
intensively polluting the concrete core. The lack of an
efficient drainage system stressed the runoff rainwater
of the structure’s lateral surfaces, showing significant
concrete degradations and intense corrosion processes.
Finally, the frequent use of de-icing salts in winter
general failures in duct-grouting processes had dra­
matically emphasized the damages. As a result, sys­
tematic fractures of the transversal reinforcement
were found under the platform’s longitudinal cracks,
as well as an advanced corrosion under stress in
post-tensioning cables, mostly unprotected due to Figure 8a and b. Cracking and general longitudinal fis­
inadequate grouting and corroded ducts. sures due to alkali-aggregate reaction expansion processes.
All this inevitably led to an immediate decommis­
sioning of the structure and a subsequent deck demo­ analysis and stability control theories (buckling) of
lition and replacement. compressed thin plates as it now appears in the
modern standards (AASHTO, Eurocodes, etc.).
Similarly, several accidents affecting concrete
3 DUCTS GROUTING IN POST-TENSIONED bridges took place in the second half of the 80s. The
CONCRETE GIRDERS Ynys-y-Gwas Bridge collapsed in 1985 in England
as a consequence of the fracture and sudden collapse
The accidents that took place during the construction of a post-tensioned concrete deck due to corrosion of
of steel box girder in developed countries like Ger­ tensioned tendons. De-icing salts were present and
many, Austria, Wales and Australia in the 70s had ducts had been inadequately grouted, which led to
a big impact on the analysis and design criteria of a delay of this type of decks by The Department of
structures and steel box girder bridges, as well as on Transport of the UK in 1992, a thorough special
construction controls of this type of structures. This inspection campaign and the drafting of new specifi­
translated into delays in the construction of steel cations for duct grouting, stability of the grouting
bridges and set the ground for the creation of the and control in most of the European and North
Merrison Committee, which completely transformed American Administrations (SETRA, FIB, etc.).

54
 Figure 11. Corrosion of ducts and post-tensioning cables
under the action of de-icing salts and a deficient duct
injection.

Figure 9. Systematic fractures of the transversal reinforce­

ment due to expansion and corrosion processes under
de-icing salts and runoff rainwater.

Figure 12. Photograph of the interior of the duct showing

the incomplete filling by the grouting.

Figure 10. Bottom of deck deterioration due to rainwater

drainage failure.

More than 160 bridges were inspected in

a campaign conducted jointly by the British
Highways Agency and the French LNPC, where
serious grouting problems were detected in nearly
15% of bridges. Similar campaigns were per­
formed with similar results in USA over the
2000s.
Figures 11 and 12, corresponding to the
demolished overpass at the highway in Madrid, Figure 13. Accident during the cantilever erection of the
Koblenz bridge (Germany) 1970.
show how the series of failures in the grouting
system of ducts, together with the described
water filtration issues with salts triggered the As a consequence, the Spanish Ministry of Public
collapse by corrosion under stress of an undeter­ Works has started an inspection campaign on the
mined number of corroded prestressing tendons, condition of ducts and post-tensioning cables of
which obliged the immediate demolition a number of structures built throughout the 90s,
of a crucial structure for the accesses into mainly those with waterproofing issues and, specific­
Madrid. ally, those subject to de-icing salts in winter.
55
 The results of these inspections, similarly to the
aforementioned French and English campaigns, show
a significant number of failures within the duct-
grouting systems, leading to the need for a precise
monitoring of affected structures and, in some cases, to
the implementation of repair measures contributing to
prevent the development of corrosion processes in
cables or to arrange external reinforcing post-
tensioning cables.
The figures below show some images illustrating
the importance of these pathologies, both in road Figure 17. Continuous water flow when loosen some
bridges (Figures 14 and 15) and railway bridges screws of anchors caps.
(Figures 16 and 17) built less than 30 years ago.

4 INTERNAL CORROSION OF WEATHERING

STEEL BOX GIRDER BRIDGES

Weathering steel has been largely used in Spain

since the 40s due to its excellent results and unques­
tionable cost and maintenance advantages in com­
parison to conventional painted carbon steel.
Our country’s weather conditions, characterized
by frequent damp and dry cycles are suitable for this
type of steel to develop its own self-protection sur­
face layers against corrosion processes.
The only exceptions provided in national and
Figure 14. Lack of corrosion protection due to a European standards (Eurocodes) for the use of this
not-conforming grouting of ducts. type of steel are the areas affected by industrial
environments or sites near the sea.
On the inner surfaces of box girders, depending
on the degree of atmospheric aggressiveness and
inner accessibility, standards establish different
internal protection requirements by means of paint,
special sealing measures or damp control in certain
inaccessibility cases.
However, maybe due to a tendency to rely exces­
sively on the anticorrosive properties of this type of
steel, in many cases, aspects such as the inner accessi­
bility, facilitating periodic inspections, and the design
of inner stiffening details avoiding water build-up and
ensuring the inspections’ regularity are often
Figure 15. Lack of corrosion protection due to a disregarded.
not-conforming grouting of ducts. Our extensive experience in the inspection of
this type of structures, quite common in Spain,
indicates that the inner state of conservation of
box girder composite bridges built less than 30
years ago is far below from the expected if the
required design and maintenance precautions had
been followed.
The failure of the waterproofing system, often
inexistent due to the aforementioned road surface
replacement processes, is systematically the origin
of most of pathologies.
Additionally, concrete slabs of composite bridge
(tensioned under a negative flexure and slightly com­
pressed at the mid-span due to the frequent construc­
tion processes with no formwork) are subject to
present passing fissures in their thickness unless spe­
cial care is taken to minimize shrinkage during hard­
Figure 16. Water leakage through anchor caps of upper ening. Such passing fissures are an easy ingress way
cables over piers on a box girder launched railway viaduct. fort water into the box girders.

56
 The advanced deterioration condition of the 5 CORBELS
bridge of Mengibar, over the Guadalquivir river, can
be observed on Figures 18 and 19. An overall failure It is important to emphasize the importance of
of the deck waterproofing system led to a water leak­ sharing the vast amount of information that may
ing that was accumulated due to the blockade of the be extracted from the results of hundreds of
steel stiffeners. This water built up between the bridge inspections in the Management Systems
bottom steel plate and the deck webs, producing an already implemented in many countries. It should
accelerated corrosion process with a net steel depth enable to avoid design failures of certain details
loss higher than the 5-10%. that, systematically, are subject to suffer patholo­
The poor ventilation in box girders, together with gies and accidents, i.e., corbels.
the high humidity coming from the riverbed, caused This type of bearings was greatly developed
frequent condensations inside the box girder, pre­ since the 50s in the design of precast concrete
venting the development of a self-protection layer bridges, as it enabled to increase main spans by
that required frequent damp and dry cycles that more than 60%, implementing isostatic solutions
never took place. (Gerber beams) without increasing the dimensions
and transport and assembly elements of conven­
tional precast elements.
By locating corbels at quarters or fifths of span,
almost coincident with the points of inflection of
continuous solutions, precast isostatic schemes of
a similar strength efficiency were achieved. Corbels
were also applied to steel bridges.
Problems deriving from the design of this types of
details are well-known:
• the non-ductility and weak character of its
strength behavior;
• its permanent exposure to leaked water through
common water tightness issues of expansion
joints;
• in concrete bridges, its high sensitivity to crack­
Figure 18. General corrosion at the bottom plates of the ing due to failures and execution tolerances, spe­
Mengibar bridge box girder. cifically of the corner reinforcement;
• in concrete bridges, the almost inevitable
development of concrete degradation processes
and the corresponding reinforcement corrosion
in an area constantly subject to moisture and
water leakage;
• in steel bridges, the high sensitivity to corrosion
and fatigue of the corner section, where there is
a high stress concentration;
• its exposure to dynamic traffic impacts in joints;

Figure 19a and b. Oxide thickness surface release due to Figure 20. Concrete degradation of spalled corbels with
the lack of weathering steels’ own self-protection layer. general corrosion at reinforcements.

57
 continuity between segments was achieved by means
of passing bars and pre-tensioning cables.

Figure 23. Precast, varying depth continuous viaduct at the

M50 to M501 connection in Madrid.
Figure 21. Concrete degradation of spalled corbels with
general corrosion at reinforcements. Notwithstanding the above, precast solutions with
corbels at quarters or fifths of span are still largely
used in many emerging countries despite their cost of
maintenance and frequent pathologies which, in many
cases, have caused serious accidents with fragile
breaking, such as the Chicago (steel) and Montreal
(concrete) bridges, involving 5 fatalities in 2006.

Figure 24. Failure by fatigue of a corbel in a steel bridge,

Chicago.

Figure 22. Corbel failure in a concrete bridge.

• the difficulty/inability of a design providing

access for a proper inspection, maintenance and
strengthening.
In Spain, this type of details is clearly obsolete
and many of them are rejected by Public Administra­
tions. Since the 90s, the precast technology of road
and railway bridges has undergone an important
technological development and Gerber solutions Figure 25. Concorde Overpass accident in Montreal due to
were replaced by permanent precast ones where the the concrete corbel failure.

58
 Figure 28. Arroyo de las Piedras viaduct (Spanish HSRL).
1210 meters between expansion joints at abutments and
piers of up to 95 meters height.
Figure 26. Detail of damages at the Montreal corbel before
collapse.
Going into details of the multiple pathologies of
bearing systems systematically encountered in the
6 BEARINGS multiple inspections carried out does not fall within
the scope of this article. For the past few years and
Bearing devices are one of the most susceptible given the economic and operational effects that they
elements to damages and pathologies leading to the may generate, damages in bearing devices are
replacement before the bridge life in service is over. a priority in congresses, technical notes and design
Generally, such damages don’t involve relevant and placement guidelines by Administrations and
structural safety risks, which could explain the rela­ Concession companies managing road and railway
tive lack of attention that is often given to the design communication networks.
of details and placements.
However, we needn’t downplay the increasing
influence of these factors on the structures’ proper
operation. A clear example of this are the high-speed
railway lines where the demanding alignment condi­
tions increasingly lead to construct long viaducts, in
some cases with high piers, where the proper oper­
ation of the different bearings is crucial to guarantee
the safety of railway traffic and the proper operation
of expansion joints, which sometimes accommodate
movements of up to 1000 mm.
The proper adjustment and dimensioning of long
sliding plates and uni-or multidirectional ends, as well
as the planning of suitable details to facilitate access
for inspection, maintenance and eventual replacement
of bearing devices are key conditions in order to avoid Figure 29. TPFE sheet removal of a POT-bearing.
dramatic affectations to communication routes that are
more and more important in developed countries.

Figure 27. Archidona viaduct (Spanish HSRL). 3150 Figure 30. TPFE sheet removal of a POT-bearing.
meters between expansion joints at abutments.

59
Figure 31. Excessive distortion at elastomeric bearings due Figure 34. Contreras viaduct: cable lifting from the car­
to errors in the estimation of imposed deformation. riageway in order to facilitate the bearing replacement, as it
had not been planned from pier caps (2014).
However, if we aim to highlight how complex
and costly the replacement of bearing devices out
of service due to a deficient design and place­ 7 FATIGUE PROBLEMS IN ROAD BRIDGES:
ment may be, as it was often the case. Until VIADUCT OVER THE CONTRERAS
now, as has often been the case, the projects RESERVOIR LOCATED AT THE
have not always considered the structure jacking MADRID-VALENCIA HIGHWAY
system in the design of pier caps and bearing
diaphragms, as it had occurred in the Viaduct Until recently, fatigue was mainly associated to
spanning the Contreras reservoir in the highway structures under intense vibration phenomena: indus­
Madrid-Valencia. The viaduct, with a traffic of trial, aeronautical, off-shore and railway bridges.
around 1,000 vehicles/day had to be taken out of Road bridge, except for some special cases, weren’t
service for several months in order to replace the regarded as susceptible to suffer fatigue problems
named bearing devices. that could affect their safety.
Spanish standards, as the SIA in Switzerland, pro­
vided a fatigue control under the action of a heavy
vehicle of 480 kN and a number of cycles depending
on the road ADT that, generally, provided satisfac­
tory results.
Due to a series of accidents occurred in Asia and
North America (Ontario, 2003; Minneapolis, 2007)
with a huge media impact and the confirmation that
the intensity of heavy vehicles in large highways and
European industrial areas exceeded by far the refer­
Figure 32. Contreras viaduct, 170-meter span composite ences that had served as a basis for traditional
truss. fatigue controls, between the years 80 and 90 it was
recommended to conduct several campaigns includ­
ing heavy vehicle traffic studies in different roads of
French (near Lyon) and German industrial areas.
Those studies evidenced the noticeable lack of cur­
rent controls and need for a new regulatory approach
on fatigue, both for actions (EC1 for actions on
bridges), exceeding up to 3 times the precedent ref­
erences (EC3-1.9 for steel structures), and construc­
tion controls (EN 1990).
In 2013, during a routine inspection, a complete
rupture of the welding connecting sections at the
bottom chord of a composite truss in the lateral span
of a viaduct located over the Contreras reservoir
(with a main span of 170 meters), at the Highway
A3 connecting Madrid-Valencia (Mediterranean
Figure 33. Contreras viaduct: bearing devices under out-of­
service conditions.
coast), was detected.

60
 and backing plates in order to determine the neces­
sary cleaning, repairing or strengthening operations.

Figure 35. Welding rupture at the truss traction chord.

Figure 37a and b. Sections opening for internal inspection

Figure 36. Replacement of the damaged element. and repair (root and backing plate areas).

The event, that could have had catastrophic conse­ The laboratory inspection of the damaged welding
quences given the strategic location of the highway identified the fatigue crack very clearly. The fissure
carrying high traffic intensities (approximately 100,000 started and developed due to the absence of backing
vehicles per day and traffic direction and a percentage plate, the deficient preparation of welding edges and
of heavy vehicles carrying an average load of 420 kN, inappropriate control of fit up between elements bor­
both in the order of magnitude already detected in the ders before starting welding.
aforementioned European campaigns), obliged to
immediately close the traffic of the A3 highway and
start an intensive inspection, survey and repair program
that lasted for nearly 2 years and a half and was per­
formed under the direction of the signatories to this
paper, including the following actions:
• preliminary full visual inspection of all truss
nodes, located under each highway’s carriageway
in order to provisionally reopen the viaduct for
a limited and light traffic while the detailed
inspection program continued;
• a subsequent detailed ultrasonic inspection and
survey campaign from the exterior of all trusses
welds, together with an accurate data collection,
geometric discontinuities, etc. that enables to iden­
tify and select the most susceptible welds to
develop fatigue cracks;
• opening closed sections in areas in which ultrasonic
tests detected a damage degree that obliged a direct
Figure 38a and b. Fatigue fracture from a notch due to lack
internal visual inspection, especially in the root area of backing.

61
 In another node open for inspection, a new fatigue steel and composite bridge’ under Eurocodes and
cracking process in similar conditions to the dam- following similar initiatives by SETRA, ECCS, etc.
aged welding was detected.

Figure 39. Fatigue crack initiation coming from a notch

due to lack of backing.
Figure 40. Fatigue design guides (SETRA, ECCS, Spanish
Ministry).
In both cases, there was an absence of a backing
plate in the inner face, necessary to ensure the proced­ 8 STAY CABLES CORROSION IN 1ST
ure of the non-accessible root area of V-weldings per­ GENERATION CABLE-STAYED BRIDGES:
formed from the outside. Presumably, misalignment CENTENARY BRIDGE IN SEVILLE
problems between adjacent elements hampered their
fitting, and the lack of technical experts specialized in Modern cable-stayed bridges date back to more than
fatigue on site during construction did not allow for the 60 years: Stromsund in Sweden (1953) and Nord­
identification of risks derived from this malpractice. brucke in Dusseldorf (1958). In Spain, the Rande
The complexity and high cost of the repair process bridge (Galicia) and Castejón (over the Ebro river)
and the confirmation of a potential accident with cata­ were constructed during the 70s.
strophic consequences set the path for the Adminis­ Steel cable protections in the first cable stayed
tration to undertake an urgent awareness and training bridges, until the 90s, were scarce: double barrier
campaign within Spanish engineering professionals in (PHD external duct and internal mortar grout). Since
the framework of the design and construction of steel the early days, there was a technical consensus between
structures with a fatigue risk, as well as to intensify experts on the necessity to improve the weakest elem­
teaching on this issues at Engineering Universities. ent of this typology, as the mortar cracking under dead
IDEAM was also commissioned by the Spanish loads and live loads reduced the protection of tendons
Ministry to draft a ‘Guide for the fatigue design of and caused dangerous damages due to friction.

62
 Figure 44. Alcoy stay-cable bridge (1987).

The most recent technological developments

focused on the following:
• strand insulation, avoiding contact with each
other in order to protect them from:
Figure 41. Observed damages in stay-cables with PHD - corrosion under tension;
ducts and mortar slurry.
- fatigue due to friction between strands;
- corrosion due to friction between strands;
• pay extra attention to the cables’ most vulnerable
areas: anchors, temporary cables, saddles, etc.;
• reducing aerodynamic vibrations: ducts com­
posed by spirals anti-wind/rain effects, deflectors,
anchor dampers, etc.;
• continuous monitoring of stay cables and possi­
bility to inspect and replace tendon by tendon.
In 2018, the unfortunate accident of the Polcevera
viaduct in Genoa intensified the media pressure on
public administrations for safety in this type of struc­
tures. Inspection campaigns on stay cables of the

Figure 42. Observed damages in stay-cables with PHD

ducts and mortar slurry.
Current technology is based on the use of multiple
protection barriers (3 to 5):
• triple barrier protection:
- external PHD duct (1)
- strands with individual duct (2) and self-
protected (grease or wax) (3) or
- galvanized/galfanized strands (2) and air humid­
ity control inside the duct (3)
• up to 4/5 barriers protection:
- external PHD duct (1)
- galvanized/galfanized strands (2), individually
self-protected with duct and grease (3 and 4)
- external PHD duct (1)
- possible internal filling of the duct with wax or
grease grouting between strands (perhaps exces­
sive) (5)

Figure 43. Multiple-barrier current technology for Figure 45. Generalized corrosion at the Alcoy bridge’s
stay-cables protection. broken cables (2016).

63
first-generation cable-stayed bridges, built before the As the Centenary bridge is a main infrastructure in
end of last century, evidenced the need to tackle the the city of Seville (with an ADT of approximately
replacement of such bridges’ stay cables for other 100,000 vehicles/day), the process of replacing stay
with technologies ensuring an increased durability. cables cannot be carried out with traffic closures, as it
In 2016, the sudden failure of one stay cable at the was the case at the Alcoy bridge. This obliged to
132-meter main span Alcoy bridge, put in service in develop a complex project that would enable to tackle
1987, led to a complete replacement of its stay-cable such replacement without posing any risk for users and
system. minimally affecting the traffic capacity of the roadway.
Subsequently, the condition of the Centenary
Bridge’s stay cables was inspected. The bridge,
inaugurated in 1992 in Seville on the occasion of
the Universal Exhibition, over the Guadalquivir
river and has a main span of 264 meters. The
inspection detected the start of the corrosion pro­
cesses development at stay cables. Although this
process did not affect the structural safety yet, it
was recommended to draft an urgent replacement
project of the stay cables, which was completed in
2019, allowing works to start in the second half
of 2020. The project and future construction sup­
port services has been developed by IDEAM, Figure 48. Traffic-jammed Centenary bridge (100,000
author of the original project that dates back 30 vehicles/day).
years, in JV with FHECOR.
Simultaneously, given the serious traffic capacity prob­
lems of the bridge, carrying 5 continuously jammed
lanes, a design to allow the widening of the current
roadway to 6 lanes was developed, for this purpose,
both cable-stayed planes were displaced sideways, as
well as their respective anchors at pylons and precast
deck by means of transversal steel beams.
Additionally, this proposal was designed in a way
that it is compatible with a subsequent lateral expansion
of the platform, encompassing two closed steel sections
external to the pylon on both sides of the deck, which
will allow for a future capacity of 10 lanes.
In our view, it is an extremely complex and high-
tech innovative action, and, to our knowledge, there
aren’t any similar precedents with similar circum­
stances requiring not to affect intense traffic flows
like this particular case.
The original transversal section and the expanded
Figure 46. Centenary cable-stayed bridge in Seville, 1992. sections after the replacement of the stay-cables are
included in Figures 49, 50 and 51.

Figure 47. Corrosion detected at some tendons of the Cen­

tenary Bridge, 2018.
Figure 49. Current cross-section (5 lanes).

64
 Figure 51. Cross-section of the future projected widening
Figure 50. Cross-section after the replacement of the (10 lanes).
stay-cables (6 lanes).

65
 Bridge Maintenance, Safety, Management, Life-Cycle Sustainability and Innovations –
Yokota & Frangopol (eds)
© 2021 Taylor & Francis Group, London, ISBN 978-0-367-23278-8

Rising to the challenge of managing bridges in Australia

N.G. Powers
Australian Road Research Board, Melbourne, Australia

ABSTRACT: Australia’s freight task is growing at a rapid rate due to unprecedented population growth,
coupled with increased demand from our trading partners in Asia and rapid changes in technology, e-com­
merce and consumer behavior. A significant proportion of freight is transported using the road network and to
remain competitive this network needs to facilitate fast, efficient and effective access. This is a significant
challenge with an ageing bridge network, the regularly extreme climatic conditions and the freight industry
constantly pushing to increase access and allowable loads.
In rising to this challenge, Australia’s infrastructure managers are being innovative, using the latest stand­
ards and technology in building new bridges and assessing existing bridges and utilizing the latest technology
in managing existing bridge structures. In partnership with academic institutions and with funding from the
State and Federal Governments, there has been significant research and development with the monitoring,
modelling, assessment and strengthening of bridges.
This paper will outline what is currently being done to rise to this significant challenge and also recom­
mend future research and development to continue to meet this challenge.

1 INTRODUCTION which includes assessment, was updated in 2017

which has helped to progress the design of new
Australia’s freight task is growing at a rapid rate due bridges and refine the assessment of existing bridges.
to unprecedented population growth, coupled with In partnership with academic institutions and with
increased demand from our trading partners in Asia funding from the State and Federal Governments,
and rapid changes in technology, e-commerce and there has been significant research and development
consumer behavior. Overall, transport logistics was with the monitoring, modelling, assessment and
valued at over A$110 billion in 2010 and has been strengthening of bridges. There has also been signifi­
steadily increasing since. A significant proportion of cant change in the overall management of heavy
freight is transported using the road network and to vehicles on the road network at the national level
re-main competitive this network needs to facilitate which has contributed to meeting the challenges
fast, safe, efficient and effective access. posed by the freight industry.
This is a significant challenge with an ageing This paper will outline what is currently being
bridge network and many of the in-service bridges done to rise to this significant challenge and also rec­
built to outdated standards and for loads that are less ommend future research and development to con­
than the allowable loads on the road network. Aus­ tinue to meet this challenge.
tralia is also a land of temperature and weather
extremes with major climatic events quite frequent
which constantly tests the resilience of our bridges. 2 AUSTRALIA’S FREIGHT TASK
When you combine the above with a freight industry
constantly pushing to increase access and allowable The effective and efficient movement of freight
loads, infrastructure managers have a major chal­ across Australia is critical for the Australian econ­
lenge to rise to. omy. A significant proportion of freight is trans­
In rising to this challenge, Australia’s infrastruc­ ported on the Australian road network. This is due to
ture managers are being innovative, using the latest constraints on the rail network and the characteristics
standards and technology in building new bridges of the land not being condusive to using other trans­
and assessing existing bridges and utilizing the latest port forms like air and sea freight.
technology in managing existing bridge structures. It is estimated that the overall transport logistics
The Australian Standard for Bridge Design AS5100, industry contributed A$111.5 billion to Australia’s

DOI: 10.1201/9780429279119-6
https://doi.org/10.1201/9780429279119-6

66
GDP in 2010. Australia’s freight is expected to grow 2.3 E-commerce
by 80% from 2011 to 2031 (ACIL Allen Consulting,
As with many countries across the world, e-com­
2014) thus access across the road network is
merce is booming Australia wide. In 2018, Austra­
essential.
lians spent A$28.6 billion on online shopping and
this is expected to increase to A$35.2 billion by
2.1 Population growth 2021 (WebAlive, 2019).
The vast majority of this results in the delivery
Australia’s population is expected to grow by
of parcels across Australia via the road network.
10 million people over the next 30 years. This is
Due to the nature of these deliveries, they are gen­
a significant increase considering Australia’s esti­
erally small and light and thus does not create
mated resident population was 25,287,400 as of
a structural burden on our bridges but more
31 March 2019 (Australian Bureau of Statistics,
a capacity issue for our already congested urban
2019).
road networks.
This increase poses a significant challenge for
infrastructure managers. Not only will it result in
more vehicles being on the road network but it will 2.4 Road freight
also generate freight to provide the infrastructure
In 2017-2018 Australian freight vehicles moved an
needed to support such growth.
estimated 214,789 million tonne-kilometres of
freight across Australia’s road network (Australian
2.2 Trade Bureau of Statistics, 2019). This represents a 5.0%
increase since 2016.
Australia’s national freight is predominantly gener­
The majority of road freight transported crude
ated by our trading partners in Asia. 75% of Austra­
materials (30 per cent) such as stone, sand and
lia’s goods exports are going to Asia, in particular
gravel, followed by manufactured goods
the booming economies of China and south east
(12 per cent) and food and live animals
Asia (Infrastructure Australia, 2018). This is
(11 per cent).
expected to continue, if not increase over coming
Over 95% of Australia’s road freight is carried by
years.
vehicles hearier than 4.5 tonners with the B-Double
Similar to population growth, this also poses
(Figure 1) accounting for 40% of all frieght move­
a challenge to infrastructure managers. With
ment (ACIL Allen Consulting, 2014). B-Doubles
Australia’s reliance on the road network to
typically carry a load of approximately 68 tonnes.
transport freight, the increase in trade will
With current frieght industry initiatives across Aus­
increase the quantity of heavy vehicles across
tralia for more high productivity vehicles, such as
the network. With freight operators striving for
B-Triples with a mass of up to 85 tonne, this will
more efficient movement of freight, often result­
allow road freight to grow and be a cost effective
ing in larger and heavier vehicles, this will
form of transport. Such large and heavy vehicles
create further challenges for infrastructure man­
create challenges for the road network in terms of
agers to meet.
safety and bridge capacity.

Figure 1. Typical B-Double. Dimensions in mm. (Spark, 2013).

67
 With such a heavy reliance on road freight for • Approximately 23,000 are managed by the State
Australia’s economic prosperity, it is critical to and Territory Road Authorities,
ensure that our road infrastructure can safely • Approximately 30,000 are managed by Local
accommodate the movement of frieght across the Government Authorities, and
network • A small number are managed by private road
operators.

3 AUSTRALIAN BRIDGE NETWORK

3.3 Design eras in Australia
3.1 Background Before 1900, standards for bridge design and loading
Similar to the rest of the world, Australian bridges did not exist. Many were built to withstand the loads
have developed over recent centuries from timber from horse and cart and may have been tested using
and masonry structures prior to the 18th century, a 16 ton road roller. Some of these bridges are still in
progressing to steel structures towards the end of the service today.
19th century and the network is now dominated by In the early 1900’s, bridge design developed, and
steel and concrete structures. the loading gradually progressed. In the 1920’s and
Timber bridges have largely been removed from 1930’s a loading was specified for the class of road
the road network over the years to make way for with Class AA roads based on a 20 ton traction
more modern materials and structures and to carry engine, Class A roads 15 tons and Class B roads 10
the loads that are now on the road network. That tons.
said, timber structures are still very important on the From 1947, bridge design loading was made uni­
road network as timber is often the only material form across Australia with the adoption of AASHTO
readily accessible in the remote parts of Australia. and the design loading of H20 S16-44 truck
For example, the Queensland Department of Trans­ (approximately equivalent to 33 tonne semi-trailer).
port and Main Roads still has over 300 timber There was a construction boom across Australia
bridges in service across its network (Austroads, during this period with the current network made up
2018). of approximately 50% of bridges built to this stand­
This is also the case for masonry structures. ard (Caprani, 2018).
While some masonry structures remain in service In 1976 – 2004 this progressed to T44 (approxi­
such as Richmond Bridge in Tasmania (Figure 2) mately 47 tonne semi-trailer) with the current net­
many have been removed from service across the work consisting approximately 25% of bridges built
network due to condition and uncertainty of struc­ to this standard.
tural capacity. With the development of the Australian Standard
for Bridge Design, AS5100 – Bridge Design Code,
the design load progressed to SM1600 which is
3.2 Current asset manager information equivalent to 2 x 75 tonne semi-trailers, nose to tail.
In Australia there are approximately 53,000 road These are estimated to make up 10% of the current
bridges (Caprani, 2018). Of these: bridge network.

Figure 2. Richmond Bridge, Tasmania – Australia’s oldest masonry bridge (Noodle snacks, 2008).

68
4 NETWORK ACCESS the design standards were set locally and in more
recent times a set of national standards were set
4.1 Background by Austroads. This was then developed into an
Australian Standard in 2004 titled AS5100 –
In Australia, heavy vehicles are classed as a vehicle Bridge Design Code (Standards Australia, 2004)
that has a gross vehicle mass of aggregate trailer providing the broader industry with greater direc­
mass of more than over 4.5 tonne (National Heavy tion and guidance for the design and assessment
Vehicle Regulator, n.d.). of bridges.
Historically, access to the road network by heavy
vehicles was governed by the relevant State and Ter­
ritory Governments. There were inconsistencies in 5.2 Recent development of Australian bridge
the management of access across these Governments standards
and their relevant authorities which created signifi­ In 2011 the development of AS5100 was required to
cant frustration in the heavy vehicle industry, and in progress with the needs of the industry and the
particular with heavy vehicle operators who would development of technology. As explained by Pritch­
operate across numerous States and Terrorities. ard, 2017 the revision of AS5100 objective is to pro­
To ensure freight could be moved across Australia vide nationally acceptable requirements for:
effectively and efficiently to support the economy,
significant changes were required at a national level. • the design of road, rail, light rail, pedestrian and
cyclist bridges;
• the specific application of concrete, steel, timber
4.2 Development of heavy vehicle management and composite material construction, which
In 2003, the Australian Heads of Government agreed embody principles that may be applied to other
to establish the National Transport Commission materials in association with other relevant
(NTC) to “progress regulatory and operational standards;
reform for road, rail and intermodal transport in • the assessment of the load capacity of existing
order to deliver and sustain uniform or nationally bridges, and
consistent outcomes” (Commonwealth of Australia, • the strengthening and rehabilitation of existing
2003). The NTC is a national transport reform bridges.
agency that works with all Australian government to There was also a significant change in focus of
advance social and economic outcomes through an AS 5100, namely:
efficient, integrated and nationally consistent land
transport system (National Transport Commission, • AS5100.7 – Bridge Rating has been changed
2019). from Rating to Bridge assessment. This includes
In 2011, an agreement was reached between the a methodology for assessing the structural for
Commonwealth, State and Territory Governments to a range of vehicles including cranes for the first
establish a national system of regulation for all time.
heavy vehicles, with uniform laws administered by • AS5100.8 - Rehabilitation and strengthening of
a single national regulator. existing bridges is a new part of AS 5100. It
The National Heavy Vehicle Regulator (NHVR) gives a comprehensive focus to this topic and dis­
was established in 2013 with the vision of “A safe, cussed the process and materials for rehabilitation
efficient and productive heavy vehicle industry serv­ and strengthening of bridges. Importantly, this
ing the needs of Australia” (National Heavy Vehicle now covers Fibre Reinforced Polymers which is
Regulator, n.d.). NHVR plays a critical role in the the first time guidance has been provided locally
coordination of access across the Australian road on this topic in an Australian Standard.
network between heavy vehicle operators and the • AS5100.9 - Timber bridges is another new part. It
relevant road authorities. is future focused and discusses “engineered
The above two changes have been significant for timber” bridges and does not discuss the use of
the movement of freight across the Australian road traditional sawn timber logs and deck planks.
network. Since their introduction much progress has
been made regarding the standardisation of assess­
ments, both vehicles and road infrastructure, but also 5.3 Australian bridge standard scope and general
in other important areas such as safety. principles 2017

5.3.1 Overview of changes

5 STANDARDS While there were objectives set for the review as
explained in Section 5.2, the review of AS5100
5.1 Overview of bridge standards encompassed all aspects to ensure that standard as
kept current and has evolved with changing technol­
The standards for bridges in Australia have come ogy. Some of the more major changes included:
a long way over the last few decades. Originally

69
• Road bridge barriers performance level definition, guidance was given for the assessment for other
selection and design have been revised. This was loads so designers and assessors had to largely rely
in response to the increasing size and mass of on overseas standards.
vehicles operating on the road network, in par­ AS5100.7-2017 (Standards Australia, 2017)
ticular that of road freight. This is now more in superceeds the above standard and is now titled
line with AASHTO’s Manual for Assessing Bridge Assessment reflecting a marked change in the
Safety Hardware (American Association of State standard and the need for greater direction and guid­
Highway and Transportation Officials, 2015). ance to be provided. It allows for the assessment of
• Environmental impacts for their consideration in any vehicle and loading which is important consider­
design process. ing the diverse vehicles we have travelling across
• Collision protection from rail to bring in line with the road network such as cranes, and heavy load
international practice and address learnings from platforms.
recent natural disasters. The significant changes made were around meth­
• Design loads to reflect light rail loads and address odology, load rating vehicles and the loads them­
learnings from recent flood events regarding selves, determining structural capacity and adopting
urban debris. new methods and technologies like structural health
monitoring.
A positive change to assist bridge assessors is the
5.3.2 Climate change and natural disasters inclusion of an easy to follow process and flowchart
Australia is a land of weather extremes and can regu­ (Figure 3) which steps the assessor through under­
larly experience major natural disasters including taking a bridge assessment including the possibility
floods, fires and cyclones. In the last decade Austra­ of higher tier assessment.
lia has experienced numerous natural disasters and
with changing climatic conditions globally these dis­
asters are expected to continue.
Experience and learnings from these recent flood­
ing events identified that AS5100 debris loading did
not accurately reflect the modern world. Previously
the debris loading was based on rural debris
(ie. Large trees striking the bridge). What was
experienced during these events was urban debris
such as shipping containers, cars, pontoons and so
on. The load has increased from 2 tonnes for a log
impact to an estimation on the expected urban debris
to would likely be experienced such as pontoons or
shipping containers which can be significantly
larger.
Also, changes were made to provide guidance to
designers regarding the following areas:
• Attachment of services to the downstream of the
bridge to avoid direct impact from debris.
• Water and sewer mains should pass through
a bridge abutment rather than the spill through
abutment. If the abutment is scoured during the
flood, the main may be damaged or severed.
• Abutment protection to emphasise the importance
of ensuring scour of the spill through abutment
does not occur so the bridge is available for flood
response if required.

5.4 Bridge assessment changes

The accurate assessment of bridges across the road
network is critical to ensure safe and efficient access
for all vehicles, in particular heavy freight.
AS5100.7-2004 (Standards Australia, 2004) was
titled Bridge Rating and was focussed on rating
bridges and specifically calculating the load rating of Figure 3. AS5100.7:2017 Bridge Assessment Flowchart
a bridge based for a specific rating vehicle. This (Standards Australia, 2017).
approach was somewhat restrictive and minimal

70
6 DETERIORATION MODELLING on the experience of the engineer, this was
becoming a challenge due to a change across the
6.1 Background industry with engineers moving around frequently
and not developing the depth of technical and
Historically, infrastructure managers in Australia have local knowledge needed. There was also
operated independently with respect to how they a demand from the government and local commu­
manage their bridge resources. This has led to signifi­ nity for greater transparency around the decision
cant differences on the approach to asset management making for asset management of all road assets,
and the systems and processes that support this. Also, including bridges.
the data recorded for bridges and the way they are A major challenge to develop accurate and prac­
inspected, and condition recorded is somewhat tical deterioration models in Victoria is the data on
different. bridge condition. The capture of data on the condition
There is a good awareness of deterioration model- of bridges is governed by the VicRoads Road Struc­
ling for bridges across Australia but the development tures Inspection Manual (VicRoads, 2018). VicRoads
and implementation of working and practical models uses a Level 1 to 3 inspection system with a Level 1
in quite limited. inspection being routine maintenance based occurring
every 6 months, Level 2 being condition based occur­
6.2 A national approach? ring every 2-5 years and Level 3 being a detailed
bridge investigation occurring on a as needs basis.
In 2015 Austroads, the peak organization of Australa­ Level 2 bridge inspections rate the condition of
sian road transport and traffic agencies, commis­ a bridge’s components utilising a 1 to 4 scale (with
sioned a report into deterioration modelling titled Condition State 1 being good condition and 4 being
Improved Bridge Deterioration Models, Predictive advanced deterioration requiring replacement) and
Tools and Costs (Austroads, 2015). The report include a percentage for a rating of how much
looked into the state-of-the-art in deterioration model- a component is affected by the condition. For
ling with the overall aim of investigating a nationally example, a bridge beam could be rated at Condition
consistent approach to deterioration modelling. State 2 for 30% and Condition State 70%. Utilising
This report determined that the significant differ­ this data with such coarse condition states and
ences in approaches between the infrastructure man­ a range of condition values per element is quite chal­
agers across Australia creates significant challenges in lenging when developing a deterioration model.
delivering a national approach to deterioration model- Another challenge to developing the deterioration
ling. In the end it was seen as idealistic at that point in modelling is the recording of maintenance and other
time considering the constraints infrastructure man­ relevant data. There was not a single system that
agers have with systems, processes, procedures and recorded maintenance information, including the
budgets. resultant condition of a bridge following mainten­
A vision statement was developed to assist with ance, which made the development of accurate
the progression towards nationally consistent deterioration curves difficult. Also, the recording of
approach to deterioration modelling with the har­ other factors that contribute to deterioration includ­
monization of inspection practice, inventory infor­ ing traffic loading and volumes, exposure and so on
mation and bridge management approach. This will is sporadic across the road network.
assist to achieve this goal in future years. As described in the paper by Tran, Setunge,
Koay, & Luczak, 2018, the above posed many
6.3 A local approach challenges with developing a deterioration model.
Many aspects of data were missing such as year
With a national approach to deterioration modelling built, widening and abnormal condition informa­
proving difficult, VicRoads and RMIT University tion. In the end, through the compasion of models
began a collaboration to develop modelling to improve they were able to determine that Markov models is
the management of bridges in the state of Victoria, more suited to their particular application and was
Historically, the management of bridges within able to provide predictive curves for the various
Victoria has been reliant on the expertise of local conditions states for a open concrete girder.
engineers. Each year, an assessment of the condi­ Since this paper, this has been further applied to
tion of bridges based on Level 2 bridge inspec­ other bridge elements with an operating model being
tions and also monitoring inspections was made developed and hoped to be implemented in the
with the engineer responsible for the management coming years. It is important this work continues as
of bridges in a region making a decision based it will help to ensure that bridge assets are main­
on engineering judgement on where to invest tained effectively and efficiently and that funding is
maintenance and rehabilitation funding. While devoted to the bridges that are more critical for the
this approach can be quite effective, depending network and thus the movement of freight across it.

71
7 BRIDGE STRENGTHENING the Little River Bridge and contract manager for the
works, after investigation and discussion with sup­
7.1 Overview pliers, design consultants and researchers accepted
the proposal. FRP was applied to the soffit of each
Until recently, the strengthening of bridges across beam as part of the strengthening works.
Australia has been somewhat restricted to conven­ Following the installation of FRP the engineers in
tional techniques such as utilizing standard VicRoads were keen to test its success with adhesion
reinforced concrete overlays and steel plating. This and load testing. This uncovered some issues with
is not surprising considering the bridge assets being the tensile adhesion which resulted in the reapplica­
strengthened such as u-slab reinforced concrete tion of the FRP. Importantly this resulted in learnings
bridge decks and with the loads being transported. regarding specification, design, surface preparation,
With the loads on the road network increasing and quality control, ambient conditions and material stor­
doing so at a rapid pace there was the need to adopt age. These have now become very important for the
more modern and progressive technology to meet future use of FRP and have been widely adopted
this challenge. across the industry.
The load testing was conducted in conjunction with
7.2 Adoption of fibre reinforced polymer Monash University using 5 axle 60 tonne mobile
technology cranes. This was able to prove the effectiveness of the
treatment and instill confidence in its use into the
The adoption of Fibre Reinforced Polymer (FRP) future.
technology with the strengthening of bridges was
slow in Australia compared to the rest of the world. 7.2.2 West gate bridge strengthening
Strengthening of FRP has been common in North Following the trial at Little River Bridge and then
Amercia, Europe and Asia for more than 40 years subsquent trials on a number of other bridges across
but only recently has it been adopted in Australia the road network, confidence grew in the technology.
due to the collaboration between suppliers, academ­ This then paved the way for FRP to be used on
ics and infrastructure managers. Overall, this is a scale much larger, the strengthening of the West
a great example of the industry working together to Gate Bridge (Figure 5).
achieve outcomes to benefit the community. The West Gate Bridge is the primary east-west
connector within Melbourne across the Yarra River.
7.2.1 Little river bridge It is a steel box girder for the five main spans across
As described by Sakardy, 2017, thanks to a group of the river with concrete box girders for the approaches
materials experts and design professionals, the con­ to the main spans. It had four lanes in each direction
sideration and adoption of alternative technology for and carried up to 200,000 vehicles per day and sig­
strengthening of bridges was progressed and real­ nificant frieght to and from the Port of Melbourne.
ised. It wasn’t without its issues, but after persistence The bridge had become very congested and was
from all involved it paved the way for this technol­ a bottleneck in the Melbourne road network. Follow­
ogy to be used into the future. ing the consideration of numerous options to address
In the late 1990’s the opportunity arose to trial the congestion, it was decided to utilise the shoulders
FRP following the need to strengthen a bridge in the of the bridge to create five lanes of traffic in each
Geelong (Princes) Freeway with the upgrade of this direction. This required significant strengthening of
route. The Little River Bridge (Figure 4), built in the bridge for torsion and shear capacity and flexural
1919, and extended in 1955, is a reinforced concrete capacity of the deck.
bridge that has performed well in service but now To strengthen the concrete box girder, FRP was
needed to be strengthened or replaced to accommo­ nominated. As described by Sakardy, 2017, the ori­
date 68 tonne B-double vehicles now commonly ginal tender nominated over 100km of “off-the-shelf”
using this route. 1.4mm thick laminates. But after further refinement
While the orginal design called for steel plating, of the design and consideration of different FPR
the construction contractor looked for other solutions solutions a 100mm wide x 4mm thick and 260 GPa
and proposed FRP. VicRoads, the asset manager of modulus carbon fibre laminate was developed to

Figure 4. Little River Bridge 1919 (Sakardy, 2017).

72
Figure 5. West Gate Bridge, Melbourne (Sakardy, 2017).

strengthen the deck cantilevers. In combination with utilised steel plating as FRP could not provide the
other carbon fibre laminates, uni-directional fabric strengthening required.
and innovative anchoring techniques the concrete
box girder was able to be strengthened and the five
lanes of traffic in each direction achieved. 8 FUTURE PRIORITIES

7.2.3 Other strengthening examples With the freight task expected to increase over
With the technology being proven across the bridge coming years, it is important that the freight and
network and major applications like on the West Gate bridge industries continue to develop and innovate
Bridge, FRP is being used more widely and commonly to ensure the challenge is met and the Australian
to achieve great outcomes from the freight industry. economy remains healthy. The following are
A great example of this is its use to strengthen a number of initiatives that will contribute to
bridges for Higher Productivity Freight Vehicle ensuring this.
(HPFV) access across the network. With the freight To ensure the rapid adoption of new technologies
industry being demanding in the access required, pro­ the standards set by infrastructure managers must
grams have been put in place to assess and strengthen move towards performance based standards rather
bridges across the network to allow access. than the often prescriptive standards. This will also
HPFV’s include numerous vehicles from 79 tonne allow innovation across the industry and the benfits
A-Double to 113.5 tonne AB-Triples. Access to the that will bring.
network by this suite of vehicles is critical for the The vast majority of road assets, including
movement of various commodities from primary bridges, are built for today’s needs with little plan­
producers to Ports and interstate destinations. ning of allowance for the future in terms of function­
As described by Kimpton & Beavis, 2017, an ality and use. There is certainly measures in place to
assessment was undertaken on two key routes in the predict future volumes of traffic and future progress
north east of the state, being the Hume Freeway and in vehicle mass. With the rapid change in transporta­
Goulurn Valley Highway. This identified 17 bridges tion and the evolution of autonomous vehicles the
potentially dificient in load carry capacity for use of our road and bridge assets will change. This is
HPFV’s with 16 bridges requiring strengthening in of course difficult to predict so road authorities will
the end. This was determined using 75% SM1600 need to be able to quickly adapt and build infrastruc­
load case which would sufficiently cover all HPFV’s ture that can adapt with the changes in need and
and would not require individual assessment for functionality.
each vehicle configuration. The collaboration and coordination across State
The assessment determined numerous structures and Federal Governments, infrastructure managers,
were deficient where typically the abutment cross- academics and researchers has been at a reasonable
heads were deficient for shear capacity and positive level to date but many more benefits could be real­
bending and piers deficient for positive and negative ised. Greater collaboration and coordination between
bending. these groups and also private industry, including
Due to the confidence built up in FRP this was the frieght operators, could be very powerful and realise
prefered strengthening treatment. After further great potential with innovation in the way our road
assessment though only half used FRP and the rest network and road freight is managed. This will result

73
in more effective and efficient access across the road With the significant changes across the freight and
network. bridge industries over the last decade or two Austra­
While Australia is a large country in terms of land lia has been able to rise to the challenge of moving
size, it is actually quite small in terms of road net­ a huge freight task across and ageing road and
work, population, the freight industry and the engin­ bridge network.
eering industry that supports infrastructure managers This has been achieved, in part, through the
to manage the road network. Due to this, the devel­ effective combination of:
opment of uniform standards across Australia is
• national changes with the management of frieght,
needed as many companies and organisations oper­
• changes to Australian Standards for bridge design
ate on a national level. This will provide greater effi­
and assessment, and
ciency and thus be able to focus funding and energy
on the task at hand. • the collaboration and cooperation between gov­
ernment, infrastructure managers, academics and
The rapid development of digital engineering is
private industry.
a major challenge for the whole industry. infrastruc­
ture managers are struggling to keep up with the rest With such a heavy reliance on road freight for Aus­
of the industry and are struggling to realise the bene­ tralia’s economic prosperity, it is critical to continue to
fits of the technology, systems, data and information develop and innovate to ensure the challenge posed by
on offer. Frameworks need to be developed to pro­ the freight task is met into the future. Also important
vide the industry with highly prescriptive data and is the need for our road infrastructure to safely accom­
information requirements to enable the benefits to be modate the movement of freight across the network.
realised. This will then ensure the effective and effi­ Overall, the various infrastructure managers across
cient management of bridges and road assets Australia have risen to this challenge to date but they
throughout the lifecycle of the asset as there will be must strive for further innovations and development
one source of truth for all asset information and it to continue to meeting this challenge into the future.
will be easily accessible to ensure investigations and
decisions can be done with minimal additional
investigations. ACKNOWLEDGEMENTS
To ensure the efficient and effective use of fund­
ing and management of the bridge network, deterior­ I would like to acknowledge the following people
ation modelling and the lifecycle mangement of and groups for making this paper possible:
bridges must be progressed. It would be preferrable
for this to be done on a national level but if this • The IABMAS 2020 Committee for requesting me
proves to still be a challenge then infrastructure man­ to deliver a keynote paper and presentation at
agers must progress this locally to ensure their con­ their conference.
strained budgets get the best value for money and • IABMAS Australia Committee for their support
the bridge assets can transport road freight as and guidance in the development and delivery of
required. this paper.
Finally, while progression has been made in • IABMAS Australia members for their support
bridge assessment with the recent revision in 2017, and advice in the development of this paper.
further progress needs to be made with higher tier
assessments which are not just based on design
methodology. This will address the inherit conserva­
REFERENCES
tism in such an approach and enable greater under­
standing of a bridge’s structural capacity. This will ACIL Allen Consulting. (2014). The Economic Significance
allow greater access across the network and more of the Australian Logistics Industry. Australian Logistics
efficient design and construction of bridge strength­ Council.
ening to existing structures. Some progress towards American Association of State Highway and Transportation
this has been made in recent research projects by Officials. (2015). AASHTO Manual for Assessing Safety
Austroads but this needs to develop further and Hardware. Washington, USA: AASHTO.
infrastructure managers will need to adopt the final Australia Bureau of Statistics. (2019, March 20). Australian
Bureau of Statistics. Retrieved from Australian Bureau
outcomes.
of Statistics: https://www.abs.gov.au/AUSSTATS/abs@.
nsf/mediareleasesbyReleaseDate/6006BF6A6CC2F525
CA2574B20020D2AB?OpenDocument
9 CONCLUSION Australian Bureau of Statistics. (2019, September 19).
3101.0 - Australian Demographic Statistics, Mar 2019.
The case study illustrated by this paper shows the Retrieved from Australian Bureau of Statistics: https://
power of, and what can be achieved when, govern­ www.abs.gov.au/AUSSTATS/abs@.nsf/mf/3101.0
ment, infrastructure managers, academics and pri­ Austroads. (2015). Improved Bridge Deterioration Models
vate industry work together to solve the problems Predictive Tools and Costs. Sydney, Australia:
Austroads.
posed to society.

74
Austroads. (2018). Guide to Bridge Technology Part 1 ­ Pritchard, R. (2017). 2017 Revision of Australian Standard
Introduction and Bridge Performance. Sydney, Austra­ AS 5100 Part 1 - Scope and General Principles. Austroads
lia: Austroads. Bridge Conference 2017. Melbourne, Australia: Austroads.
Noodle snacks - Own work, C. B.-S. (n.d.). Richard Bridge Sakardy, A. (2017). The development of FRP strengthening
Tasmania. Retrieved from Wikipedia: https://commons. on the Victorian road network over the last 17 years.
wikimedia.org/w/index.php?curid=4955073 Austroads Bridge Conference 2017. Melbourne, Austra­
Caprani, C. (2018, August 22). Are Australian bridges safe, lia: Austroads.
and can we do it better? Retrieved from The Spark, I. (2013, June). A numerical control algorithm for
Conversation: https://theconversation.com/are-austra a B-double truck–trailer with steerable trailer wheels
lian-bridges-safe-and-can-we-do-better-101825 and active hitch angles. Part 2: Reversing. Retrieved
Commonwealth of Australia. (2003). Inter-Governmental from ResearchGate: https://www.researchgate.net/
Agreement for Regulatory and Operational Reform in figure/Typical-dimensions-of-a-B-double-truck-trailer­
Road, Rail and Intermodal Transport. vehicle-Lengths-are-in-mm_fig1_258177823
Infrastructure Australia. (2018, October 31). Australia’s Standards Australia. (2004). AS5100 - Bridge design.
Growing Freight Task: Challenges and Opportunities. Sydney, Australia: SAI Global.
Retrieved from Infrastructure Australia: https://www. Standards Australia. (2004). AS5100.7 Bridge Design ­
infrastructureaustralia.gov.au/listing/speech/australias­ Rating. Sydney, Australia: SAI Global.
growing-freight-task-challenges-and-opportunities Standards Australia. (2017). AS5100.7 - Bridge Assessment.
Kimpton, D., & Beavis, R. (2017). Strengthening of Sydney, Australia: SAI Global.
Bridges for Highway Productivity Freight Vehicles. Standards Australia. (2017). AS5100.1 - Scope and General
Austroads Bridge Conference 2017. Melbourne, Austra­ Principles. Sydney, Australia: SAI Global.
lia: Austroads. Standards Australia. (2017). AS5100.7:2017 - Bridge
National Heavy Vehicle Regulator. (n.d.). About the NHVR. Assessment. Sydney, Australia: SAI Global.
Retrieved from National Heavy Vehicle Regulator: Tran, H., Setunge, S., Koay, Y.-C., & Luczak, H. (2018).
https://www.nhvr.gov.au/about-us/who-we-are/about­ Predictive Modelling of the Deterioration of Australian
the-nhvr State Bridge Network. Maintenance, Safety, Risk, Man­
National Heavy Vehicle Regulator. (n.d.). What is a heavy agement and Life-Cycle Performance of Bridges
vehicle? Retrieved from National HEavy Vehicle Regu­ (IABMAS 2018) (pp. 1806–812). Melbourne, Australia:
lator: https://www.nhvr.gov.au/about-us/who-we-are/ Taylor & Francis Group.
what-is-a-heavy-vehicle VicRoads. (2018). Road Structures Inspection Manual.
National Transport Commission. (2019). National Trans­ Melbourne, Australia: VicRoads.
port Reform Implementation Monitoring Report ­ WebAlive. (2019, June 24). The State of Australia’s Ecom­
Report to the TIC. Melbourne, Australia: National merce in 2019. Retrieved from WebAlive: https://www.
Transport Commission. webalive.com.au/ecommerce-statistics-australia/

75
 Bridge Maintenance, Safety, Management, Life-Cycle Sustainability and Innovations –
Yokota & Frangopol (eds)
© 2021 Taylor & Francis Group, London, ISBN 978-0-367-23278-8

Rib-to-floor beam connections of orthotropic steel decks for bridge deck

replacement applications
R. Sause, Y. Chen, J. Saunders & I. Hodgson
Lehigh University, Bethlehem, Pennsylvania, USA

J. Marks
AECOM, Pittsburgh, Pennsylvania, USA

ABSTRACT: Orthotropic steel decks (OSDs) are used for new bridge decks and replacement decks of highway
bridges. An OSD includes various welded connections. This paper focuses on rib-to-floor beam (RFB) connections
with emphasis on their fatigue performance. The bridge superstructure affects the fatigue stresses in RFB connec­
tions. Independent OSD floor beams that are not supported or restrained by other transverse members in the
bridge are common for new bridge applications. Restrained OSD floor beams that are supported and restrained by
other transverse members in the bridge are common for deck replacement applications. When the OSD floor beam
is supported by an existing transverse floor truss (or beam), the fatigue stresses at the RFB connection from in-
plane loading are reduced. However, when the OSD floor beam is restrained by the existing floor truss, the fatigue
stresses at the RFB connection from out-of-plane loading are increased. Thus, for deck replacement applications,
a stress-relieving cut-out is often used in the OSD floor beam web to allow the ribs to rotate more freely under
out-of-plane loading. An “extended cut-out” RFB connection has been used extensively in deck replacement appli­
cations in the US. The extended cut-out RFB connection often exhibits good fatigue performance, but this type of
connection can be labor intensive to fabricate. This paper presents research on a new RFB connection for deck
replacement applications with a “slit” cut-out completely within the OSD floor beam web that allows the rib to
rotate, and is easier to fabricate than the extended cut-out RFB connection.

1 INTRODUCTION AND BACKGROUND One specific welded OSD connection, the rib-
to-floor beam (RFB) connection, can be labor
1.1 Orthotropic steel decks intensive to fabricate in order to achieve adequate
fatigue performance. The RFB connection is
The orthotropic steel deck (OSD) is a bridge deck
fatigue sensitive because it is subjected to com­
system consisting of a continuous steel deck plate sup­
plex in-plane and out-of-plane deformations when
ported by longitudinal open or closed ribs that pass
the OSD is subjected to traffic loading. One type
continuously through a cut-out in the transverse OSD
of fatigue stress response of the RFB connection
floor beams. These components are joined together
is from “in-plane loading” where traffic wheel
using welded connections. OSDs provide advantages
loads passing over the OSD floor beam produce
of rapid on-site construction, light weight, and long ser­
in-plane vertical shear forces (and shear stresses)
vice life compared to other bridge deck systems. OSDs
in the floor beam web, which distort the rib cut­
are often used for long-span or movable highway
out in the floor beam web, leading to local rib
bridges. An OSD is often used in deck replacement
wall distortion and corresponding stresses in the
applications because the OSD will be lighter than
RFB connection (Figure 1). In-plane loading also
a typical existing concrete bridge deck, resulting in less
induces stresses in the RFB connection from
dead load on the bridge after replacement.
transfer of shear stresses in the rib walls to the
While OSDs have advantages over other bridge
OSD floor beam, since the floor beam provides
deck systems, OSDs have experienced fatigue per­
the reactions for shear and torsion in the ribs.
formance issues. The components of the OSD (deck
Another type of fatigue stress response of the
plate, floor beam and rib) are joined via relatively
RFB connection is from “out-of-plane loading”
complex welded connections, and fatigue cracking
where traffic wheel loads within the span of the ribs
of these welded connections has been observed.

DOI: 10.1201/9780429279119-7
https://doi.org/10.1201/9780429279119-7

76
Figure 1. In-plane loading of RFB connection with distor­ Figure 3. Independent OSD floor beam.
tion of rib cut-out in OSD floor beam web and distortion of
rib wall from floor beam web shear.
In a deck replacement application, the design of the
OSD is often strongly influenced by the existing
between adjacent floor beams, produce rotation of bridge superstructure configuration. Compared to the
the ribs at the RFB connection, inducing out-of­ OSD for a new bridge application, the rib spans may
plane bending of the floor beam web. The floor be longer (influenced by the existing floor truss (or
beam web bending response to the rib rotation leads floor beam) spacing), and the depth of the OSD floor
to stresses at the RFB connection at the bottom of beam may be limited by the available distance from
the rib as shown in Figure 2. the top of the existing floor truss (or floor beam) to the
roadway. Also, in a deck replacement application,
1.2 Floor beam configuration some of the OSD floor beams are located on and are
attached to the top of the existing floor trusses (or floor
The configuration of the bridge superstructure system beams) of the bridge superstructure. These OSD floor
can significantly affect the fatigue stresses in RFB beams are “restrained” floor beams (see Figure 4) that
connections that develop under traffic loading. Inde­ are supported vertically and restrained out-of-plane by
pendent OSD floor beams, that are not supported by the other transverse members in the existing bridge
and are not restrained by other transverse members in superstructure.
the bridge, are common for an OSD designed for Due to the vertical support provided by the exist­
a new bridge (see Figure 3). Independent floor beams ing transverse members of the bridge, a restrained
are free to deform in-plane (i.e., are unsupported) floor beam is subjected to smaller in-plane floor beam
when traffic wheel loads pass over the floor beam, bending and floor beam shear, compared to an inde­
and are free to deform out-of-plane (i.e., are unre­ pendent floor beam. At the same time, however, the
strained) due to rib rotation when wheel loads are out-of-plane restraint of the OSD floor beam tends to
within the rib span between adjacent floor beams. increase the fatigue stresses at the RFB connection
from out-of-plane loading due to rib rotation. The
out-of-plane restraint of the OSD floor beam is from
attachment of the OSD floor beam flange to the larger
transverse members of the existing bridge superstruc­
ture. A schematic illustration comparing the out-of-

Figure 2. Out-of-plane loading of RFB connection with

out-of-plane bending of OSD floor beam web due to rib Figure 4. Restrained OSD floor beam located on top of
rotation. existing floor truss.

77
plane response of a restrained and an independent One configuration for the termination of the
floor beam is shown in Figure 5. The schematic illus­ extended cut-out on the rib wall is to utilize a small,
tration shows that the OSD floor beam bottom flange square-ended tab that terminates perpendicular to the
is restrained against translation and rotation in the rib wall. The connection is fillet welded and the fillet
out-of-plane direction of the OSD floor beam, result­ weld wraps around the tab on the rib wall (Figure 7).
ing in greater stress at the RFB connection. This extended cut-out RFB connection configuration
For deck replacement applications, the combination has several fatigue prone locations at the edges of
of the OSD floor beam bottom flange transverse the extended cut-out, and at the termination of the
restraint, the limited floor beam depth, and the extended cut-out on the rib wall, where stress con­
increased rib rotation from a long rib span all contrib­ centrations occur due to the change of geometry and/
ute to the resulting large fatigue stresses at the RFB or the presence of the fillet weld. Potential fatigue
connection from out-of-plane loading. This large stress cracks for the extended cut-out RFB connection are
motivates the use of a stress-relieving cut-out in the shown schematically in Figure 8. In some extended
floor beam web at the RFB connection, which allows cut-out RFB connections with wrap-around fillet
the rib to rotate more freely under out-of-plane load­ welds, internal bulkheads or stiffeners within the ribs
ing, resulting in reduced stress at the RFB connection. have been used to achieve adequate fatigue perform­
ance, but this internal stiffening complicates the fab­
rication process, therefore, internal stiffening was
1.3 RFB connection types
not considered in this research.
An “extended cut-out” RFB connection has been In order to decrease the stress at the termination
used extensively in deck replacement applications in of the extended cut-out on the rib wall, and to
the US. For the extended cut-out RFB connection, the increase the fatigue resistance at this location, an
closed rib passes through a cut-out in the OSD floor alternate termination configuration has been used in
beam web with the cut-out extended below the rib (as the US (e.g., Bronx Whitestone Bridge and Wil­
shown in Figure 6) to allow the rib to rotate more liamsburg Bridge). In the alternate configuration,
freely under out-of-plane loading. The extended cut­ shown in Figure 9, the weld is ground smooth (tan­
out in the OSD floor beam web terminates on the rib gent to the rib wall), where the extended cut-out in
wall as shown in Figure 6. The extended cut-out is the floor beam web terminates on the rib wall. When
the recommended RFB connection type in the the extended cut-out has the tangent termination on
AASHTO Bridge Design Specifications (AASHTO the rib wall, the RFB connection often includes
2018) and has been extensively used for deck replace­ a complete-joint penetration (CJP) weld at the
ment applications such as the Bronx Whitestone
Bridge and Williamsburg Bridge in New York City.

Figure 5. Comparison of out-of-plane loading of restrained

and independent OSD floor beams.

Figure 7. Fillet welded perpendicular termination of

Figure 6. Extended cut-out RFB connection. extended cut-out on rib wall (adapted from Kitner (2016)).

78
 transition to the rib wall, making this type of RFB
connection labor intensive to fabricate.
Research on a new type of RFB connection,
termed the “slit” RFB connection is presented in
this paper. For the slit RFB connection, the closed
rib passes through a matching cut-out in the OSD
floor beam web. An additional slit cut-out in the
floor beam web is added to the RFB connection
(see Figure 10) to allow the rib to rotate more
freely under out-of-plane loading. Unlike the
extended cut-out RFB connection, the additional
cut-out of the slit RFB connection is located
entirely within the floor beam web below the rib,
and this cut-out does not terminate on the rib wall.
Figure 8. Potential fatigue crack locations for extended The authors of the present paper were motivated to
cut-out RFB connection with fillet welded perpendicular study the slit RFB connection by a schematic pres­
termination. entation of this connection in (HNTB 2015).
Since the stress-relieving cut-out of the slit
RFB connection is located entirely within the
web of the floor beam and does not terminate on
the rib wall, the stress concentration and labor
intensive fabrication of a weld termination on the
rib wall are avoided. The fatigue prone locations
for the slit RFB connection are at the edges of
the slit cut-out, and at the weld toes of the fillet
welds between the rib wall and floor beam web.

Figure 10. Slit RFB connection.

Figure 9. Ground smooth (tangent) termination of extended

cut-out on rib wall (adapted from Kitner (2016)).

termination, transitioning to a fillet weld or partial-

joint (PJP) weld further up the rib wall, towards the
deck plate. This type of RFB connection weld was
developed from full-scale laboratory fatigue tests for
the Williamsburg Bridge deck replacement OSD
(Tsakopoulos & Fisher 2003). An extended cut-out
with this alternate type of termination has welds that
require joint preparation and non-destructive inspec­ Figure 11. Potential fatigue crack locations for slit RFB
tion, followed by grinding to achieve the smooth connection.

79
Potential fatigue cracks for the slit RFB connec­ Table 1. Cross section properties of longitudinal trusses
tion are shown schematically in Figure 11. in truss system.

Area Iz Iy
2 FINITE ELEMENT ANALYSIS APPROACH
mm2 mm4 mm4
2.1 Finite element analysis models
Top Chord 7550 1.28 × 108 1.80 × 107
To evaluate stresses for the extended cut-out RFB Diagonal 7550 1.28 × 108 1.80 × 107
connection type and the slit RFB connection type, Bottom Chord 7550 1.28 × 108 1.80 × 107
three dimensional linear elastic finite element ana­
lysis (FEA) of OSD models were performed. The Note: Iz is the moment of inertia in-plane of truss
FEA models used twenty-noded quadratic brick Iy is the moment of inertia out-of-plane of truss
elements with reduced integration. Most of the OSD
models included truss systems to support the OSD,
and the truss members were modeled with three­ Table 2. Cross section properties of transverse floor
noded quadratic beam elements. The stresses trusses in truss system.
reported in this paper for the RFB connections in the
OSD are from the FEA models with an average Area Iz Iy
mesh size of 25.4 mm within the OSD.
The FEA evaluated stresses for the extended cut­ mm2 mm4 mm4
out RFB connection under fatigue limit state loading,
and evaluated the slit RFB connection stresses under Top Chord 32,300 9.66 × 108 1.12 × 109
fatigue, strength, and service limit state loading. The Diagonal 7480 1.28 × 108 1.80 × 107
FEA was performed using a simple bridge with an Bottom Chord 9810 1.30 × 108 3.61 × 108
OSD shown in Figure 12. This OSD has five trans­
verse floor beams, two longitudinal plate diaphragms Note: Iz is the moment of inertia in-plane of truss
Iy is the moment of inertia out-of-plane of truss
which have full height stiffeners at the floor beam
locations, a 19.1 mm thick deck plate and eleven
U-shaped ribs. The ribs are 355 mm deep, spaced The truss system supporting the OSD is similar to
710 mm center-to-center, and have 7.9 mm thick the truss systems of long-span cable-supported
walls. The ribs are numbered from one to eleven, bridges that were studied in past research at Lehigh
with the south-most rib identified as Rib 1. The floor University on OSDs for deck replacement applica­
beam flanges are 12.7 mm thick and 305 mm wide tions. The plate diaphragms are supported by longitu­
and the floor beam webs are 12.7 mm thick. Two dinal trusses and the bottom flange of the five OSD
floor beam web depths were used in the FEA study: floor beams are supported by five transverse floor
876 mm for the extended cut-out RFB connection and trusses. These longitudinal trusses and transverse
673 mm for the slit RFB connection. Two center-to­ floor trusses comprise the truss system. The cross
center floor beam spacings were used for the FEA. section properties of the truss system are listed in
For studies of in-plane loading of the RFB connec­ Table 1 and Table 2. The depth of the longitudinal
tions, the floor beam spacing was 3580 mm, and for trusses is 2670 mm and the depth of the transverse
studies of out-of-plane loading, the floor beam spa­ floor trusses is 1450 mm. To provide boundary condi­
cing was 6100 mm, which is the maximum recom­ tions to the OSD FEA model, the three translational
mended floor beam spacing by (Connor et al. 2012). displacement degrees of freedom are restrained at the
The floor beams are numbered from one to five, with bottom of each longitudinal truss at each end, adja­
the east-most floor beam identified as Floor Beam 1. cent to Floor Beam 1 and adjacent to Floor Beam 5.

2.2 Loading
To evaluate the RFB connection stresses under
fatigue limit state loading, the OSD FEA model was
loaded with the AASHTO Fatigue I load combin­
ation factored tandem axle load for OSDs (AASHTO
2018), shown in Figure 13. This factored tandem
axle load for OSDs consists of four wheel loads of
93.1 kN spaced 1220 mm longitudinally and
1830 mm transversely. The wheel loads are distrib­
uted uniformly over a rectangular tire contact area
(i.e., the “wheel load patch”) with a width of
508 mm and a length of 254 mm, with the shorter
dimension parallel to the traffic direction. The
Figure 12. Overview of OSD model.
80
 To evaluate the fatigue stresses in the RFB con­
nections under a large in-plane loading condition,
Rib 1 was studied. Rib 1 is closest to the plate dia­
phragm and is therefore located in the region of the
floor beam with the greatest vertical shear force.
When the longitudinal center of the full tandem axle
load is centered on Floor Beam 3 with one half of
the full tandem (two of the wheel load patches) cen­
tered transversely between Rib 1 and Rib 2, as
shown in Figure 14, a large shear force develops in
the floor beam at Rib 1. This loading configuration is
Figure 13. Full tandem axle load for OSDs. denoted the “in-plane loading configuration”.
To evaluate the fatigue stresses in the RFB con­
nections under a large out-of-plane loading condition,
associated uniformly distributed pressure is approxi­ a condition that created a large rib rotation was iden­
mately 0.72 MPa. This load configuration, based on tified. In this loading configuration, the half tandem
the factored tandem axle load for the AASHTO load was centered transversely on Rib 6 at the mid­
Fatigue I load combination is termed the “full span of the floor beam. Rib 6 was selected since its
tandem load”. transverse location is farthest from the plate dia­
Another load configuration used to evaluate the phragms adjacent to Ribs 1 and 11, as shown in
RFB connection stresses under fatigue loading Figure 15, and therefore the rib rotation for Rib 6 is
includes only two wheel load patches spaced least affected by the longitudinal trusses. Six different
1220 mm apart longitudinally, as shown in the blue longitudinal positions of the half tandem were studied
box in Figure 13. This loading configuration is termed using the OSD FEA model to determine the position
the “half tandem load”. The half tandem load was that results in the largest rib rotation. The longitu­
used to focus the FEA results on the effects of the lon­ dinal center of the half tandem was moved from the
gitudinal position of the load patches and to eliminate center of Floor Beam 3, west, towards Floor Beam 4.
the floor beam shear force response to the other two As shown in Table 3, when the longitudinal center of
load patches that are included in the full tandem load. the half tandem is 1.8 m from Floor Beam 3, the
To evaluate the slit RFB connection stresses under absolute rib rotation from FEA is largest. This load­
strength limit state and service limit state loading, two ing configuration, with the half tandem centered
load configurations were applied to the OSD FEA transversely on Rib 6 and with longitudinal center
model. To study the strength limit state, the AASHTO located 1.8 m west from Floor Beam 3 is denoted the
Strength I load combination was applied. The Strength “half tandem out-of-plane loading configuration”.
I limit state load combination includes a factored tandem The full tandem axle load was also used to study
axle load consisting of four factored wheel loads of out-of-plane loading of the RFB connections, includ­
129 kN spaced 1220 mm longitudinally and 1830 mm ing the full effect of the AASHTO Fatigue I load
transversely, added to a factored uniform lane load of combination. The corresponding loading configur­
16.3 kN/m applied across a lane width of 3050 mm. ation has two wheel load patches centered trans­
To study the service limit state, the AASHTO Ser­ versely on Rib 6 while the other two wheel load
vice I load combination was applied. The Service II patches are transversely located 1830 mm to the
limit state load combination includes a factored north, as shown in Figure 16. The two wheel load
tandem axle load consisting of four factored wheel patches centered on Rib 6 induce a large rib rotation.
loads of 74 kN added to a lane load of 9.3 kN/m. The longitudinal center of the full tandem is
The self-weight of the OSD FEA model, with 1.8 m west from Floor Beam 3 to maximize the rib
a superimposed dead load corresponding to a 76 mm rotation. This loading configuration is denoted the
thick asphalt wearing surface, were also included in “full tandem out-of-plane loading configuration”.
the Strength I and Service II load combinations.
The stresses under fatigue loading for the extended
cut-out RFB connection type and the slit RFB connec­
tion type were investigated for loading conditions that
produce: (1) largely in-plane (in the plane of the OSD
floor beam web) response (denoted in-plane loading);
or (2) largely out-of-plane response (denoted out-of­
plane loading). In-plane loading produces in-plane
stress response, such as in-plane vertical shear forces
(and shear stresses) in the floor beam web and transfer
of shear stresses from the rib walls to the floor beam
through the RFB connection. Out-of-plane loading
produces out-of-plane response, including rib rotation Figure 14. Transverse position of full tandem load for Rib
and out-of-plane bending of the floor beam web. 1 in-plane loading configuration.

81
Figure 15. Transverse position of half tandem load for Rib
6 half tandem out-of-plane loading configuration.

Table 3. Rib rotation of Rib 6 at Floor Beam 3 for slit

RFB connection under half tandem out-of-plane loading
configuration.

Half tandem longitudinal center Absolute rib rotation

position from Floor Beam 3 from FEA Figure 17. Extended cut-out RFB beam connection with
dimensions.
M radian
0.0 0.00 × 10-4
0.6 6.09 × 10-4
stress along the edge of the extended cut-out from
1.2 1.08 × 10-3
FEA is shown in Figure 18. The largest tension princi­
1.8 1.26 × 10-3
pal stress on the south side of the extended cut-out is
2.4 1.25 × 10-3
50.3 MPa. The largest compression principal stress on
3.0 1.13 × 10-3
the north side of the extended cut-out is -157 MPa.
The magnitudes of these principal stresses are
below the 165 MPa constant-amplitude fatigue limit
(CAFL) for AASHTO Fatigue Category A (AASHTO
2018), which is the appropriate fatigue category for
this fatigue detail (base metal). The magnitudes of
these stresses relative to the CAFL suggests that
fatigue cracking at the edge of the extended cut-out is
unlikely. Throughout this paper, fatigue stresses from
FEA using the AASHTO Fatigue I load combination
are compared with the CAFL for the appropriate
fatigue category. When the fatigue stress is below the
CAFL, fatigue cracking is unlikely. Fatigue stress
Figure 16. Transverse position of full tandem load for Rib above the CAFL suggests that infinite fatigue life and
6 full tandem out-of-plane loading configuration. good fatigue performance are unlikely.
A plot of the variation of the stress normal to the
3 EXTENDED CUT-OUT RFB CONNECTION weld toe at the rib wall weld of the wrap-around
FEA RESULTS fillet weld is shown in Figure 19. The maximum ten­
sion stress normal to the weld toe is 147 MPa, and
The stresses for the extended cut-out RFB connection occurs symmetrically on either side of the floor beam
were evaluated for fatigue limit state loading only. web. This stress (normal to the weld toe) is much
Stresses for the extended cut-out RFB connection greater than the 68.9 MPa CAFL for AASHTO
under the strength limit state and service limit state Fatigue Category C, which is the appropriate fatigue
loading, were not evaluated. The dimensions of the category for this fatigue detail. Therefore, the
extended cut-out RFB connection that was studied are extended cut-out with a perpendicular termination
shown in Figure 17. and wrap-around fillet weld, as shown in Figure 19,
is unlikely to exhibit good fatigue performance.
A plot of the variation of the stress normal to the
3.1 FEA results for in-plane loading weld on the rib wall for the extended cut-out RFB
configuration connection with a ground smooth weld termination is
FEA of the extended cut-out RFB connection provided shown in Figure 20. The largest tension stress normal
the fatigue stresses for the connection of Rib 1 under to the weld on the rib wall is 91.1 MPa as shown in
the in-plane loading configuration described earlier. Figure 20. Since the weld is ground smooth with the
A plot of the variation of largest magnitude principal rib wall, this stress is compared to the 165 MPa

82
 wall for the ground smooth termination (91.1 MPa)
is smaller than the stress normal to the fillet weld toe
for the wrap-around fillet weld (147 MPa), and the
fatigue category is higher (Category A versus Cat­
egory C). Therefore, the ground smooth weld termin­
ation has been used often in the US, instead of the
wrap-around fillet weld, to provide good fatigue per­
formance of the extended cut-out RFB connection.

3.2 FEA results for half tandem out-of-plane

loading configuration
Figure 18. Variation of largest magnitude principal stress FEA of the extended cut-out RFB connection provided
along cut-out edge of extended cut-out RFB connection of the fatigue stresses for the connection of Rib 6 under
Rib 1 with perpendicular termination and wrap-around the half tandem out-of-plane loading configuration
fillet weld under in-plane loading configuration. described earlier. A plot of the largest magnitude prin­
cipal stress along the extended cut-out edge at the east
face and west face of the floor beam web for the half
tandem out-of-plane loading configuration is shown in
Figure 21. The difference in largest magnitude princi­
pal stresses between the east face and west face of the
floor beam web represent the portion of the stress that
is caused by out-of-plane bending of the floor beam
web due to rib rotation. The largest difference between
the east face principal stress value and west face prin­
cipal stress value is 40.7 MPa, which is relatively low
indicating that the fatigue stresses at the edge of the
extended cut-out are not strongly influenced by out-of­
plane loading (i.e., by out-of-plane bending of the floor
beam web due to rib rotation). These results show that
the extended cut-out provides stress-relief in the floor
beam web at the RFB connection, which allows the rib
Figure 19. Stress normal to wrap-around fillet weld toe on to rotate more freely under out-of-plane loading.
rib wall of north side of extended cut-out RFB connection Figure 22 shows the largest magnitude principal
of Rib 1 with perpendicular termination and wrap-around
fillet weld under in-plane loading configuration.
stress along the extended cut-out edge at the floor
beam web mid-surface under half tandem loading
with the center of the half tandem in two different
longitudinal positions. The mid-surface stress, which
is unaffected by out-of-plane bending of the web, is
the in-plane component of stress. In Figure 22, the lar­
gest compression principal stress along the extended

Figure 20. Stress normal to weld on rib wall of north

side of extended cut-out RFB connection of Rib 1 with
ground smooth weld termination under in-plane loading
configuration.

CAFL for AASHTO Fatigue Category A, which is

the appropriate fatigue category for this fatigue Figure 21. Variation of largest magnitude principal stress
detail. The comparison shows that fatigue cracking is at east face and west face of web along cut-out edge of
extended cut-out RFB connection of Rib 6 with perpen­
unlikely for the ground smooth weld termination.
dicular termination and wrap-around fillet weld under half
Note that the stress normal to the weld on the rib tandem out-of-plane loading configuration.

83
cut-out decreases in magnitude from -150 MPa to
-96.5 MPa as the center of the half tandem is moved
1.8 m away from Floor Beam 3. The decrease in cut­
out edge stress is from a decrease in the shear stresses
in the rib walls that are transferred to the OSD floor
beam through the RFB connection, since more of the
half tandem load is transferred to Floor Beam 4 when
the half tandem is 1.8 m from Floor Beam 3.
The largest compression principal stress along the
extended cut-out edge of -115 MPa on the west face
under out-of-plane loading (with the half tandem at
1.8 m from Floor Beam 3), shown in Figure 21, is
smaller in magnitude than the -150 MPa mid-surface Figure 23. Stress normal to wrap-around fillet weld toe on
stress for the 0 m longitudinal load position shown in rib wall of north side of extended cut-out RFB connection
Figure 22. These results show that for the extended of Rib 6 with perpendicular termination and wrap-around
cut-out RFB connection, the stress along the extended fillet weld under half tandem loading for two different lon­
cut-out edge is more sensitive to in-plane loading gitudinal positions.
than out-of-plane loading, due to the stress-relieving
cut-out in the floor beam web at the RFB connection. components of the slit geometry. The slit has three
Figure 23 shows the stress normal to the weld toe at different radii to ensure that the cut-out edge is
the rib wall weld of the wrap-around fillet weld under a continuous curve (and that the slope of this curve
half tandem loading with the center of the half tandem is continuous). The radii can be varied to reduce the
in two different longitudinal positions. The largest stress at the edge of the slit cut-out. There is an
stress normal to the weld toe decreases from 106 MPa upper portion of the slit (represented by Radius #1),
to 71.0 MPa as the center of the half tandem is moved a lower portion of the slit (represented Radius #2),
1.8 m away from Floor Beam 3. These results show and the keyhole portion of the slit (represented by
that for the extended cut-out RFB connection, the Radius #3). The remaining portion of floor beam
stress at the wrap-around fillet weld is more sensitive web between the top of the slit and the bottom of the
to in-plane loading than out-of-plane loading due to rib wall is represented by tTAB. The width of the slit
the stress-relieving cut-out in the floor beam web. is represented by tSLIT.
The two slit geometries studied in this paper are
shown in Table 4. The slit geometry termed Slit A,
4 SLIT RFB CONNECTION FEA RESULTS was used in the studies of in-plane loading. The slit
geometry termed Slit B was used in the studies of
The slit RFB connection was not studied previously out-of-plane loading.
by FEA, so an initial FEA study in Saunders et al.
(2019) focused on understanding the behavior of the
slit RFB connection and on determining favorable
slit geometries that provide smaller stresses. The two
slit geometries that were used predominately in the
FEA studies in Saunders et al. (2019) are discussed
here. A representative slit RFB connection is shown
in Figure 24 along with notation that represent the

Figure 24. Slit rib-to-floor beam connection with geometric

components labeled.

Table 4. Slit geometries used in slit RFB connection

FEA.

Radius Radius Radius

#1 #2 #3 tslit ttab
Figure 22. Variation of largest magnitude principal stress
at mid-surface of web along cut-out edge of extended mm mm mm mm mm
cut-out RFB connection of Rib 6 with perpendicular ter­
mination and wrap-around fillet weld under half tandem Slit A 165 191 12.7 25.4 19.1
loading for two different longitudinal positions. Slit B 152 191 25.4 25.4 19.1

84
4.1 FEA results for in-plane loading heat-affected zone (AASHTO 2018). This location on
configuration the slit edge of the slit RFB connection of Rib 1 is
expected to be under compression under dead load
FEA of the slit RFB connection provided the stresses
and under all live load conditions, and thus fatigue
for the connection of Rib 1 under the in-plane load­
cracking under this compression stress is not expected.
ing configuration. The largest magnitude principal
Figure 27 and Figure 28 show the stress normal to
stress along the top edge and bottom edge of the slit
the weld toe of the fillet weld of the slit RFB connec­
RFB connection (with Slit A) can be seen in Figure
tion on the rib wall and on the floor beam web,
25 and Figure 26, respectively. The largest tension
respectively. These stresses are well below the 68.9
principal stress is 107 MPa on the north side of the
MPa CAFL for AASHTO Fatigue Category C,
top slit edge, which is well below the 165 MPa
which suggests that fatigue cracking of the fillet
CAFL for AASHTO Fatigue Category A.
weld is unlikely. Compared to the extended cut-out
The largest magnitude compression principal stress
RFB connection, the slit RFB connection has signifi­
is -204 MPa on the north side of the bottom slit edge.
cantly smaller stress normal to the fillet weld toe.
Compression stress is considered to contribute to
The presence of the slit protects the weld from high
fatigue damage only if a tension stress develops at the
stress conditions and thus the weld has the potential
same location under a different fatigue load condition
for good fatigue performance.
(i.e., a stress reversal occurs) according to (AASHTO
2018). If the tensile component of the stress range
does not exceed the compressive stress due to the 4.2 FEA results for full tandem out-of-plane
unfactored dead loads, a net tensile stress does not loading configuration
occur, and therefore, the stress range is compression
FEA of the slit RFB connection provided the stresses
only and a fatigue crack will not propagate beyond the
for the connection of Rib 6 under the full tandem out-

Figure 25. Variation of largest magnitude principal stress

along top slit edge of slit RFB connection (with Slit A) of Figure 27. Variation in stress normal to weld toe on rib
Rib 1 under in-plane loading configuration. wall around rib bottom of slit RFB connection (with Slit A)
of Rib 1 under in-plane loading configuration.

Figure 26. Variation of largest magnitude principal stress Figure 28. Variation in stress normal to weld toe on floor
along bottom slit edge of slit RFB connection (with Slit beam web around rib bottom of slit RFB connection (with
A) of Rib 1 under in-plane loading configuration. Slit A) of Rib 1 under in-plane loading configuration.

85
of-plane loading configuration. The variation of the lar­ herein. On the rib wall the stress normal to the
gest magnitude principal stress along the slit edge at fillet weld toe is compressive. The largest magni­
the east face and west face of the floor beam web are tude compressive stress is -69.6 MPa on the east
shown for the top of the slit in Figure 29 and the side of the floor beam at the bottom of the rib.
bottom of the slit in Figure 30. The largest difference The magnitude of this stress is slightly above the
between the east face principal stress value and west 68.9 MPa CAFL for AASHTO Fatigue Category
face principal stress value on the north side of the C. As noted earlier, compression stress is con­
bottom slit edge is 46.2 MPa. This stress difference is sidered to contribute to fatigue damage only if
relatively low, indicating that the stresses at the slit tension stress develops at the same location under
edge are not strongly influenced by out-of-plane load­ a different fatigue load condition according to
ing (i.e., by out-of-plane bending of the floor beam (AASHTO 2018). The stress normal to the fillet
web due to rib rotation). These results show that the weld toe at the bottom of the rib is dominated by
slit provides stress-relief in the floor beam web at the primary (negative) bending of the rib, and tension
RFB connection, similar to the extended cut-out RFB stress is not expected at this location. On the
connection, which allows the rib to rotate more freely floor beam web, the stress normal to the fillet
under out-of-plane loading. weld toe is small at all locations, well below the
The stresses normal to the weld toe on the rib 68.9 MPa CAFL for AASHTO Fatigue Category
wall and on the floor beam web of the fillet weld C. Thus, fatigue cracking of the fillet weld of the
of the slit RFB connection of Rib 6 under the slit RFB connection of Rib 6 under the full
full tandem out-of-plane loading configuration tandem out-of-plane loading configuration is
were also determined from FEA (Saunders et al. unlikely.
2019), but for brevity, plots are not included
4.3 FEA results for strength limit state and service
limit state
For the OSD of the simple bridge shown in
Figure 12, the strength limit state and the service
limit state defined by the AASHTO Bridge
Design Specifications (AASHTO 2018) were
checked to determine if the dimensions of this
OSD are reasonable.
For the strength limit, according to (AASHTO
2018), the rib flexural resistance is 438 kN-m, the rib
shear resistance is 1380 kN, and the rib compression
force resistance is 3220 kN, which exceed the corres­
ponding demands of 81.9 kN-m, 142 kN, and 1237
kN, respectively. The OSD floor beam has a flexural
Figure 29. Variation of largest magnitude principal stress
resistance of 2860 kN-m, a shear resistance of 758 kN,
along top slit edge of slit RFB connection (with Slit B) of
Rib 6 on east and west faces of web under full tandem and compression force resistance of 1180 kN, which
out-of-plane loading configuration. exceed the corresponding demands of 52.4 kN-m, 143
kN, and 115 kN, respectively. The demands for the rib
and OSD floor beam were determined from the FEA
results by integrating the stresses over the effective
dimensions.
For the service limit, according to (AASHTO
2018), the deflection of the OSD under dead load
plus live load was checked against the limit of the
span divided by 300. This check was performed for
the larger center-to-center floor beam spacing (i.e.,
larger rib span) that was studied, 6.1 m. The max­
imum deflection from FEA is 7.4 mm, which is sig­
nificantly less than the span/300 value of 20.3 mm.
The live load deflection of the OSD was checked
against the limit of the span divided by 1000. The
maximum deflection from FEA is 4.6 mm, which is
less than the span/1000 value of 6.1 mm. The max­
imum relative deflection under live load between
Figure 30. Variation of largest magnitude principal stress adjacent ribs was checked. The maximum relative
along bottom slit edge of slit RFB connection (with Slit B)
of Rib 6 on east and west faces of web under full tandem
deflection from FEA is 1.5 mm, which is also less
out-of-plane loading configuration. than the limit of 2.5 mm.

86
4.4 FEA results for modified OSD model with
independent floor beams under in-plane
loading configuration
A modified OSD model for a simple bridge was
developed, in which the OSD floor beams are
independent. The independent OSD floor beams
are not supported by and are not restrained by
other transverse members in the bridge (i.e.,
which is common for an OSD designed for Figure 31. Overview of modified OSD model.
a new bridge). The three dimensional linear elas­
tic FEA model for this simple bridge with inde­
pendent OSD floor beams is shown in Figure 31.
This model was used to study the slit RFB con­
nection with an independent floor beam under in-
plane fatigue loading. The modified OSD model
has the same floor beam, deck plate, and rib
geometry as the OSD of the simple bridge sup­
ported by the truss system shown in Figure 12,
which was used in the previous studies. Two
I-shaped longitudinal plate girders (one on either
side of the OSD) were used instead of the plate
diaphragms along the longitudinal edges of the
OSD. To provide boundary conditions to the
modified OSD FEA model, the three translational Figure 32. Variation of largest magnitude principal stress
displacement degrees of freedom were restrained along top slit edge of slit RFB connection (with Slit A) of
at the bottom of each end of each longitudinal Rib 1 with independent floor beam under in-plane loading
plate girder. configuration.
Stresses from FEA for the slit RFB connection of
Rib 1 with an independent floor beam under in-plane
fatigue loading are shown Figure 32 and Figure 33,
which show the variation of largest magnitude prin­
cipal stress along the slit edge for the top of the slit
and the bottom of the slit, respectively. The largest
tension principal stress is 187 MPa on the south side
of the bottom slit edge. The largest (in magnitude)
compression principal stress is -283 MPa on the
north side of the bottom slit edge. These large princi­
pal stresses develop at the slit edge from the shear
force in the OSD floor beam. The shear force is lar­
gest in the region near Rib 1 which is adjacent to the
longitudinal plate girder that supports the floor
beam. The slit edge principal stresses are signifi­ Figure 33. Variation of largest magnitude principal stress
cantly larger for an independent floor beam com­ along bottom slit edge of slit RFB connection (with Slit A)
pared to a restrained floor beam (i.e., compare of Rib 1 with independent floor beam under in-plane loading
Figure 32 and Figure 33 with Figure 25 and Figure configuration.
26) under in-plane loading due to the larger shear
force in the independent floor beam.
5 SUMMARY OF FULL-SCALE
Increasing the depth or thickness of the floor beam LABORATORY TESTS FOR SLIT RFB
web will decrease the floor beam shear stresses and CONNECTION
the principal stresses along the slit edge. The largest
tension and compression principal stresses at the slit
To assess the stress response and the potential for
edge are above the 165 MPa CAFL for AASHTO good fatigue performance of the slit RFB connection
Fatigue Category A, and thus, without increasing the in deck replacement applications, full-scale laboratory
floor beam web depth or thickness to reduce these
tests were performed on two identical OSD test speci­
stresses, the slit edge of the slit RFB connection at mens with the slit RFB connection. Details of the test
Rib 1 is not expected to have infinite fatigue life. program, including test fixtures, test specimens, test
Therefore, unless the OSD floor beam web is
protocol, and test results are given in Saunders et al.
designed to control the stresses along the slit edges,
(2019). The tests and test results are summarized here.
the slit RFB connection would not be appropriate for An overview of the slit RFB connection specimen
OSD applications with independent floor beams. test setup is shown in Figure 34. The test specimens
87
were studied under cyclic in-plane loading with two 5.1 Fatigue test results under in-plane loading
different loading configurations, and under out-of­
Two different cyclic in-plane loading configurations
plane static loading. The test specimens include:
were used for fatigue testing of the slit RFB connec­
a 19.1 mm thick deck plate with an area of 3.1 m by
tion specimens (Figure 35). One loading configur­
1.8 m; four U-shaped 355 mm deep ribs with 7.9 mm
ation, used in the phase of testing denoted Phase A,
thick walls, and 710 mm center-to-center spacing;
used the factored tandem axle load for OSDs from
a floor beam with a 12.7 mm thick and 673 mm deep
(AASHTO 2018). In Phase A testing, four equal
web and a 19.1 mm thick, 305 mm wide bottom
wheel load patches were centered longitudinally on
flange; an I-shaped extension beam with dimensions
the floor beam web and located transversely to maxi­
similar to the floor beam; and an I-shaped edge plate
mize the in-plane loading of the slit RFB connec­
girder with a 14.2 mm thick, 978 mm deep web,
tions at Rib 1 and Rib 4. A second loading
a 19.1 mm thick, 305 mm (12 inches) wide bottom
configuration, used in the phase of testing denoted
flange, and full height stiffeners on the outer web
Phase B, used a load that was 25% larger than the
face at the two supports and midspan. The slit geom­
factored full tandem axle load for OSDs from
etry of the slit RFB connection of the test specimens
(AASHTO 2018). In Phase B testing, three wheel
was the same as Slit B listed in Table 4.
load patches were used. Two of the wheel load
The ribs were numbered from one to four, with
patches, each equal to 25% of the total applied load,
Rib 1 next to the edge plate girder as shown in Figure
were located on the deck plate of the test specimens.
34. The bottom flange of the test specimen floor beam
The third wheel load patch, equal to 50% of the total
was restrained by test fixtures that simulate the
applied load, was located on the extension beam.
restraint provided to the OSD floor beam by the truss
The longitudinal center of the load in Phase
system in the OSD model shown in Figure 12.
B testing was centered on the floor beam web and
The test specimens and test fixtures were assessed
the load was placed transversely to maximize the in-
using FEA prior to fabrication and installation in the
plane loading of the slit RFB connections at Rib 2
test setup. The FEA results and the processes used to
and Rib 3. The 25% greater load used in Phase
fabricate the test specimens and test fixtures are pre­
B testing was intended to generate stress ranges at
sented in detail in Saunders et al. (2019).
the edges of the slits at Rib 2 and Rib 3 that were
Strain gages were installed on the rib walls and
similar in magnitude to the stress ranges at the edges
floor beam webs of the slit RFB connection speci­
of the slits at Rib 1 and Rib 4 in Phase A testing.
mens. Strain gages were installed normal to the weld
During Phase A testing, a total of 2.6 million
toe, on the rib wall and the weld toe on the floor
cycles were applied to each of the two slit RFB spe­
beam web, of the fillet weld of the slit RFB connec­
cimens and no fatigue cracks were observed. At the
tions. Strain gages were also installed on the top
slit RFB connection fillet welds, the stresses at the
edge and bottom edge of the slit cut-out. Inclinom­
weld toe on the rib wall and the weld toe on the floor
eters were installed on the rib walls to measure rib
beam web were small in magnitude; the largest
rotations. Further details on the installation of instru­
stress range at either weld toe was 37.9 MPa. At the
mentation and on the data acquisition system can be
slit edge, the stress ranges that were entirely tension
found in Saunders et al. (2019).
were not large, but the stress ranges that were
The various loading configurations were applied
entirely compression were relatively large in magni­
to the slit RFB connection specimens using
tude. At Rib 1, the largest stress range was 123 MPa
hydraulic actuators. Details of the hydraulic actu­
in tension, and -218 MPa in compression.
ators and the hydraulic control system can be found
To assess the fatigue performance (i.e., the
in Saunders et al. (2019).
potential for fatigue cracking) of the slit RFB
A summary of the slit RFB connection specimen
results is given below. More detail and additional
laboratory test results can be found in Saunders et al.
(2019).

Figure 34. Overview of slit RFB connection specimen test Figure 35. Slit RFB connection specimen in-plane loading
setup. test.

88
connection specimens from the fatigue tests under Specifications (AASHTO 2018) do not require these
in-plane loading, these stress ranges from Phase locations to be assessed for fatigue, the 2.3 million
A testing are compared with the mean fatigue cycles applied to each of the two slit RFB specimens
resistance in the finite-life range of fatigue resist­ significantly exceeds the mean fatigue life of
ance (i.e., compared with the mean S-N curve). 1.7 million cycles for one test specimen. Due to
The finite-life fatigue resistance was used instead schedule constraints, it was not possible to apply
of the CAFL because the stress range magnitude in enough cycles to exceed the mean fatigue life of
compression at the slit edges of the test specimens 3.2 million cycles for the other test specimen. How­
was intended to exceed the 165 MPa CAFL for ever, the 2.3 million cycles exceeded the design
AASHTO Fatigue Category A, based on the design fatigue life of 1.2 million cycles based on the stress
of the test specimens and selected magnitude of range of -196 MPa, using AASHTO Fatigue Category
loading. The design finite-life fatigue resistance of A (AASHTO 2018). No fatigue cracking was
the AASHTO Bridge Design Specifications observed, and thus the slit RFB connection specimens
(AASHTO 2018) for each fatigue category was exhibited adequate fatigue resistance and potential for
established from the probability distribution for the good fatigue performance during Phase B of the in-
fatigue test data, and is a lower bound to the data plane loading tests.
(Keating & Fisher 1986). For a given AASHTO
fatigue category and stress range, the mean number
5.2 Test results under static out-of-plane loading
of cycles to failure (i.e., the mean finite-life fatigue
resistance) can be determined from the probability As observed from FEA of the slit RFB connection for
distribution used to establish the design fatigue out-of-plane loading, the stress demands from out-of­
resistance (Keating & Fisher 1986). plane loading are small for the fillet weld toe (on the
For the stress range of -218 MPa in compression rib wall and on the floor beam web) and for the slit
at the edge of the slit from the Phase A testing, with edges of the slit RFB connection. Since the stresses at
AASHTO Fatigue Category A as the appropriate these fatigue prone locations were not expected to be
fatigue category, the corresponding mean fatigue life large in the test specimens under out-of-plane loading,
is approximately 2.2 million cycles. Note that the the test specimens were tested under static out-of­
applied stress range at this location was entirely plane loading rather than under cyclic fatigue loading.
compression, and the dead load stress at this location In the static out-of-plane loading (Figure 36), the
is expected to be compressive, so the AASHTO slit RFB connection specimens were tested using
Bridge Design Specifications (AASHTO 2018) do hydraulic actuators to generate rib rotations up to
not require this location to be assessed for fatigue. 36% larger than the rib rotations observed from FEA
Nonetheless, the 2.6 million cycles applied to each of the OSD model (shown in Figure 12) under the
of the two slit RFB specimens exceed the mean half tandem out-of-plane loading configuration.
fatigue life of 2.2 million cycles, with no fatigue Under out-of-plane loading that produced these
cracks observed. Thus, the slit RFB connection spe­ large rib rotations, the largest stress ranges normal to
cimens exhibited adequate fatigue resistance and the weld toe on the floor beam web and on the rib
potential for good fatigue performance during Phase wall of the fillet weld were 6.9 MPa and 65.5 MPa,
A of the in-plane loading tests. respectively, which are less than the 68.9 MPa
During Phase B of testing, a total of 2.3 million
cycles were applied to each of the two slit RFB spe­
cimens and no fatigue cracks were observed. At the
slit RFB connection fillet welds, the stresses at the
weld toe on the rib wall and the weld toe on the
floor beam web were small in magnitude; the largest
stress range was 45.5 MPa. Similar to the results for
Phase A testing, for Phase B testing, the tension
stress ranges at the slit edge were not large, but the
compression stress ranges were relatively large in
magnitude. At Rib 2, the largest stress range was 117
MPa in tension, and -241 MPa in compression for
one test specimen, and 94 MPa in tension and -196
MPa in compression for the other test specimen.
These stress ranges from Phase B testing are com­
pared with the mean finite-life fatigue resistance. For
the stress range of -241 MPa in compression at the
edge of the slit, the corresponding mean fatigue life is
approximately 1.7 million cycles. For the stress range
of -196 MPa in compression at the edge of the slit, Figure 36. Cross section of slit RFB connection specimen
the corresponding mean fatigue life is 3.2 million out-of-plane loading test.
cycles. Although the AASHTO Bridge Design
89
CAFL for AASHTO Fatigue Category C. The largest ACKNOWLEDGMENTS
stress range observed on the slit edge was 85.8 MPa,
which is much less than the 165 MPa CAFL for The research reported in this paper is based upon
AASHTO Fatigue Category A. Although the rib work supported by the Federal Highway Administra­
rotations were much larger than those based on FEA tion under Cooperative Agreement No. DTFH61-11­
of the OSD model, the corresponding stress ranges H-00027. Any opinions, findings, and conclusions or
are unlikely to cause fatigue cracking. recommendations expressed in this paper are those
of the Authors and do not necessarily reflect the
view of the Federal Highway Administration.
6 CONCLUSIONS Support was also provided through the Pennsylva­
nia Infrastructure Technology Alliance from the
1. Based on the FEA results, the extended cut-out Pennsylvania Department of Community and Eco­
RFB connection with a wrap-around fillet weld nomic Development.
termination does not appear to have the potential
for good fatigue performance for U-shaped ribs
without internal bulkheads or stiffeners within the REFERENCES
ribs.
2. Based on the FEA results, the extended cut-out AASHTO. 2018. AASHTO LRFD Bridge Design Specifica­
RFB connection with a ground smooth termin­ tions (8th Edition ed.). Washington, DC: American Associ­
ation of State Highway and Transportation Officials.
ation has the potential for good fatigue perform­ Connor et al. 2012. FHWA Manual for Design, Construc­
ance, as also shown by previous research; tion, and Maintenance of Orthotropic Steel Deck
however the ground smooth termination of the Bridges. Washington, D.C.: US Department of Trans­
extended cut-out can be labor intensive to portation Federal Highway Administration.
fabricate. HNTB. 2015. Cost-Effective Orthotropic Bridge Decks
3. Based on FEA and full-scale laboratory test results, (An Evaluation of Optional Welding Processes).
the slit RFB connection demonstrated the potential New York: HNTB Corporation.
for good fatigue performance for deck replacement Keating, P. B. & Fisher, J. W. 1986. Review of Fatigue Tests
applications, where the OSD floor beams are and Design Criteria on Welded Details, Final Report.
NCHRP Report 286. Fritz Laboratory Report 488.1.
“restrained” (i.e., the OSD floor beams are sup­
Bethlehem PA: Fritz Laboratory, Lehigh University.
ported vertically by other transverse members in the Kitner, K. 2016. A Study of Manufacturable Rib-to-Floor
existing bridge superstructure). Beam Connections in Steel Orthotropic Bridge Decks.
4. For independent OSD floor beams that are not M.S. Thesis. Bethlehem PA: Lehigh University.
supported by other transverse members in the Saunders, J., Chen, Y., Marks, J., Hodgson, I. & Sause, R.
bridge (common for an OSD designed for a new 2019. Fatigue Resistant Rib-to-floor Beam Connections
bridge), FEA results (for the OSD floor beam that for Steel Orthotropic Bridge Decks. ATLSS Report No.
was studied) show that the slit RFB connection 19-02. Bethlehem, PA: ATLSS Engineering Research
has large stresses at the cut-out edge, and that Center, Lehigh University.
Tsakopoulos, P. A. & Fisher, J. W. 2003. Full-scale Fatigue
infinite fatigue life and good fatigue performance
Tests of Steel Orthotropic Decks for the Williamsburg
are unlikely; therefore, unless the OSD floor Bridge. Journal of Bridge Engineering, 8(5): 323–333.
beam web is designed to control the stresses at Tsakopoulos, P. & Fisher, J. 2005. Full-scale Fatigue Tests
the slit edges, the slit RFB connection would not of Steel Orthotropic Deck Panel for the
be appropriate for applications in OSDs with Bronx-Whitestone Bridge Rehabilitation. Bridge Struc­
independent floor beams. tures, 1(1), 55.

90
 Bridge Maintenance, Safety, Management, Life-Cycle Sustainability and Innovations –
Yokota & Frangopol (eds)
© 2021 Taylor & Francis Group, London, ISBN 978-0-367-23278-8

Energy-efficient autonomous framework for monitoring railroad bridges

in the USA using wireless smart sensors
B.F. Spencer, T. Hoang & K.A. Mechitov
University of Illinois, Urbana-Champaign, USA

ABSTRACT: One of the most critical components of the US transport system is railroads: providing means
of transportation for 48% of the nation’s total modal tonnage. Despite such important tasks, more than half of
the railroad bridges, the essential component of railroads to maintain flow of network, were built before 1920,
making them the most fragile components of the railroad system. Wired and wireless sensor systems have
been deployed, but none is designed specifically to address the challenges of railroad bridges monitoring,
including: 1) limited energy source for sensors; 2) short and random nature of train schedule; 3) unavailable
autonomous monitoring systems; and 4) difficult rapid decision-making process due to long data processing
time. This paper focuses on efforts to develop an autonomous schedule-based framework for monitoring rail­
road bridges using wireless smart sensor network (WSSN). This framework, which bases on WSSN platform
Xnode, makes use of multiple components, including hardware, software, and algorithms to fulfill the needs
for railroad bridge condition monitoring. To demonstrate the efficacy of this system, a full-scale monitoring
campaign has been conducted. With these improvements to overcome the challenges of monitoring railroad
bridges, this system is expected to become an important tool for railroad engineers and decision makers.

1 INTRODUCTION warning (Big Bayou Canot Bridge in Mobile, AL –

September 22nd, 1993 – Bridge collapsed after 20
Railroads are a critical component of minutes due to a bridge striking a bridge span to out
U. S. transportation and economy. On average, the of alignment (Specker 2016; Garner & Huff 1997;
railroad system carries 48% of the nation’s total Mayville et al. 1999); (ii) an monitoring system that
modal tonnage while emitting the least amount of can be remotely controlled to increase its monitoring
greenhouse gas compared to waterborne, truck, and events frequency based on potential hazard weather
air (2017 CFS Preliminary Data 2017; O’Rourke (Bridge in Kingman, AZ – August 9, 1997 – Train
et al. 2015). Since 1980, freight railroads have spent derailed, and bridge collapsed due to erosion and
more than $685 billion to maintain a reliable freight scouring of the early morning flash flood (Mayville
rail network. This investment is still growing, e.g., et al. 1999; Gurzinskil 2015); and (iii) a low-cost,
there was a $24.9 billion increase from 2017 to 2018. easy-to-deploy campaign type assessment tool that
U.S. freight shipments are also projected to rise from can provide a quantitative ratings to existing struc­
17.8 billion tons in 2017 to 25.5 billion tons in 2040 tures (structures show visual damage but still rated as
(Association of American Railroads 2019). Such safe by railroad companies (Gunderson 2015).
a rapid increment will put more stress on the struc­ To tackle these problems, on September 13, 2010,
tural integrity of the railroad systems. the FRA required mandatory management programs
Crucial components of the railroad system are the for all bridges. Under this new regulation, all rail­
more than 100,000 railroad bridges, of which more road bridges need to be structurally inspected and
than half were built before 1920 (American Railway rated at least annually. To date, visual inspection is
Engineer and Maintained-of-way Association 2003). the main approach (Federal Railroad Administration
In the past, there were multiple cases when railroads 2010). However, in many structures, especially com­
system and society were profoundly impact due to plex ones, in-depth understanding can only be effi­
railroad bridges in adverse conditions. In these inci­ ciently learned from quantitative structural health
dents, the preventative practice of using a monitoring monitoring (SHM). In fact, such SHM systems
system could be a very straight forward solution, already deployed using wired sensor systems on
such as (i) an impact detection system with timely many structures such as bridges (Tsing Ma Bridge,

DOI: 10.1201/9780429279119-8
https://doi.org/10.1201/9780429279119-8

91
Kap Shui Mun Bridge, Ting Kau Bridge (Ko et al. required, or where a power source is accessible.
1999; Wong 2004), Bill Emerson Memorial Bridge Due to the aforementioned limited energy issue,
(Caicedo et al. 2002)), buildings (Millikan Library this is not feasible in railroad bridge monitoring.
(Clinton et al. 2006), One Rincon Hill Tower 4) Challenging data-to-decision process: The main
(Huang et al. 2012), or other types of infrastructures goal of all monitoring systems is to provide
(Muffti 2003; Ni et al. 2009). However, the major actionable information regarding the structural
drawback is in the cost of cabling and installation, condition to the bridge engineers in a precise,
which could sum up to more than $22,000/sensor straightforward, and timely manner. If this goal is
(Farrar 2001) on the bridges, or more than $5,000/ approached in an edge-computing direction, it
channel for general tall buildings (Celebi 2002). can lift the burden of the energy usage for com­
To overcome the barrier of high cost, which munication while speed-up the decision-making
mainly originates from cable installation and main­ process significantly. This is not a trivial task as
tenance (Lynch 2006), researchers and engineers it demands to address both the theory and prac­
have switched to wireless sensors, which makes use tical aspects of the applications.
of radio communication. Since 1998, many gener­
ations of multiple platforms of the wireless sensor
have been designed and applied to monitor real
structures, but very few were applicable for monitor­ 2 RAILROAD BRIDGE MONITORING
ing railroad bridges and none specifically designed SYSTEM
for such purpose. Thus, even the most advanced and
widely accepted solution is not well-match for one The main advantages to adopt wireless smart sensors
of the most critical components of the infrastructure in SHM are due to 4 major features: 1) intelligence
and economy. Several issues are unique to railroad capabilities of the onboard microprocessor, which
can handle digital signal processing, Analog to Digi­
bridges:
tal Converter (ADC) or Frequency to code conver­
1) Limited energy: One of the most significant char­ sion, communication interface functions, and other
acteristics of wireless sensors is that they must condition-based decision-making functions, 2)
operate on a limited energy budget. The power small-size due to increasing use of Micro-
usually is provided in the form of a battery whose electromechanical System (MEMS) based sensing
capacity can be improved by increasing its size, components which have been decreasing in size with
which, however, retrains the types of applica­ time, 3) low cost due to mass production and multi­
tions. In addition to this limitation, battery-based purpose of MEMS-based parts, and 4) wireless to
energy source usually suffers from low efficiency make use of the existing protocol for radio-
in cold weather, which is very typical to the site frequency data transmissions (Spencer 2004).
condition of railroad bridges.
2) Unpredictable nature of train events: As one of
2.1 Xnode wireless smart sensor platform
the top bridge research priorities is to help rail­
road owners in their inventory management by To facilitate large-scale applications, Spencer et al.
measuring bridge responses under train loadings (2017), developed the next-generation sensor plat­
(Moreu & Lafave 2012). However, the train dur­ form – the Xnode (Figure 1) to provide a system
ation can vary from one to a few minutes and it capable of high-fidelity sensing, reliable communica­
can be either running hours ahead of or after the tion, and efficient power and data management. To
pre-defined schedule. A trivial monitoring enable such highly demanding tasks, the sensor hard­
system, if not designed for such unpredictable ware platform has multiple improvements from the
characteristics, cannot capture precise and com­ previous generation of Imote2. More details are
plete data when service-load is required for the shown in Table 1.
inspection process. Researchers developed the Xnode based on Free-
3) The high cost of human-in-the-loop: A smart and RTOS – a real-time operating system, capable of
long-running system needs to be able to keep high demanding industrial control system tasks. This
communicating with the owner or decision-
maker to be able to handle the information
promptly, as well as to update the system
remotely. Without this capability, regular check­
ing or software updating is difficult and costly as
personnel needs to be present periodically. One
of the simple practices to address such a matter is
to have an on-site minicomputer with remote
access functions (Jang et al. 2010; O’Connor
et al. 2017). This method only works economic­
ally and efficiently for large-scale deployments
Figure 1. Next-generation wireless smart sensor platform
where only one or very few computers are
Xnode.

92
Table 1. Comparison between the iMote2 with ISM-400 minimal extra energy. The low-cost, low-powered
sensor board and the Xnode (Fu et al. 2018). accelerometer features ultralow power, 3-axis
MEMS accelerometer consumes less than 2 μA at
iMote2 w/ISM-400 Xnode Smart a 100 Hz output data rate, and 270 nA when in
sensor board (Adler Sensor (Spencer motion-triggered wake-up mode (Analog Devices
et al. 2005) et al. 2017) 2016).
The hardware component that enables the sched­
Sensing 4 8
channels
uled rendezvous scheme is the real-time clock
Sample rate Up to 5 kHz 1-16kHz
DS3231m. This MEMS-based real-time clock has
A/D 16-bit 24-bit
a temperature compensated crystal oscillator for
resolution highly accurate timekeeping of ±5ppm (±0.432
Time synch +/-20us +/-10us Second/Day), and two Time-of-Day Alarms for the
error purpose of keeping track and trigger the Xnode
Acquisition periodic sensing periodic/trigger based on pre-defined tasks at low energy usage.
schemes sensing Figure 2 sums up the integration of these two com­
LOS range 300m >1km ponents to the existing Xnode radio board. More
Data rate 250kbps 250-1000kbps details regarding the hardware integration process
Transmission IEEE 802.15.4 IEEE 802.15.4 are available at Fu et al. 2018.
protocol
Clock speed 13-416MHz 12-204MHz
2.3 Near-real-time data retrieval and emergency
Volatile 32MB 32MB
memory
notification
Permanent N/A 128MB NAND A SHM system with access to the cellular and the
storage Flash + 4GB SD Internet could potentially accelerate the data retrieval
card and notification process in an emergency. This direc­
Operating TinyOS FreeRTOS tion is achievable for WSSN with an extension of
system a dedicated cellular modem.
Sensing ~190mA ~220mA In the selection process, due to the upcoming clos­
power
ing of 3G infrastructure in the U.S. (IEEE Communi­
Sleep power ~0.5mA ~0.5mA
cations Society 2019), only 4G LTE or Cat 1 cellular
Energy Solar panel Solar panel
harvesting
network is considered for their reasonable speed and
cost. The final option is to use the 4G LTE network

option addresses the issue of limited resources of

TinyOS operating systems, including resource allo­
cation, single-application focus, not real-time sup­
port, non-popular development language of nesC
(Levis et al. 2005). More details are available in the
literature (Fu et al. 2016; Fu et at. 2019).

2.2 On-demand event triggering and scheduled

rendezvous design and framework
The original WSS Xnode platform provides the
elemental functionalities for a SHM system. How­
ever, more hardware and software components are
still required to fulfill the aforementioned needs of
a railroad bridge monitoring system. These exten­
sions are designed to enable the functionalities of
event triggering (for train detection), and scheduled
rendezvous (for regular network tasks such as data
retrieval, condition query, etc.) while consuming
minimal energy. Both of the following additional
components are purposefully chosen and designed
such they can run independently from the main wire­
less sensor node, enabling low energy mode.
ADXL362 is selected as the hardware component
that fulfills all the essential characteristics that allow Figure 2. Final PCB board design and components
the Xnode to be able to follow an on-demand with connections.

93
as this is more suitable for initial design phases. The onboard microcontroller and Free-RTOS operating
selected component is Sierra Wireless HL7588 LTE­ system. Table 7 lists the approaches that are experi­
CAT4 modem (Sierra Wireless 2018). This module mented (with ‘X’ mark meaning the approach
communicates with the Xnode via Universal Asyn­ included the corresponding steps in its process), and
chronous Receiver/Transmitter serial port (UART) Table 8 shows the corresponding computation time.
and 2 GPIO pins for turning on and off. Due to its The final approach allows an improvement of
power demands and current spikes during transmit­ 100,000 faster for the processing time (and poten­
ting, HL7588 is powered separately tially down to no additional processing time at all if
In conjunction with this component, a cloud- approach nine is successfully validated).
based Amazon Web Services (AWS) data reposi­ As Table 2 and 3 show, the most valuable compo­
tory server is also set up to collect data directly nent added to the online, onboard displacement estima­
from the gateway as well as to access the data tion pipeline include: i) fast processing of data using
remotely. the built-in library CMSIS DSP for decimation and fil­
tering: reduces the computation time significantly by
reducing the size of the original raw data while main­
2.4 Online estimation of dynamic displacements as
taining the critical measurements in the lower fre­
a metric for structural health
quency range (Keil.com 2017); ii) data saving in
According to a survey done in 2011, railroad bridge binary form: ignores the conversion and saving step
structural engineers showed specific interest in col­ for human-readable format, resulting in a significant
lecting and use bridge displacement as a metric for saving of both computation time and memory space;
condition assessment (Moreu & LaFave 2012).
A SHM system capable of incorporating such needs
in its data acquisition and analysis pipeline and pro­ Table 2. Details of implementation approaches.
viding the engineers with such data in a timely
manner can help streamline the decision making sig­ Approach
nificantly, saving both time and cost while reducing
the risk of late action making due to long data pro­ 1 2 3 4 5 6 7 8 9*
cessing. The Xnode with a high performance embed­ Use External FFT X X X X
ded processor designed for demanding digital signal library
processing task is qualified to efficiently process the Decimate data X X X X X X X
acceleration data on-line to provide the engineer CMSIS DSP for X X X X X X X
with more insight into the structural performance. filtering
The algorithm for reference-free displacement esti­ CMSIS DSP for X X X X X X
mation (Gomez et al. 2018) is implemented for the the decimation
onboard computation of the Xnode. This algorithm is Precompute and X X X X X
save coefficients
an estimator of dynamic displacement from acceler­
Save raw data X X X X X
ation based on and Finite Impulse Response (FIR)
Save binary data X X X
filter, which is stable and can be applied directly in
On-line process- X X
time-domain. More details can be found in Gomez
ing using Free-
et al. (2018). RTOS
Although the algorithm was initially designed for On-line saving X
easy to implement on WSS, the implementation is using Free-RTOS
not straightforward, and multiple steps are required
to optimize the algorithm to make use of all the * Prototyped but not extensively tested

Table 3. Performance (in computation time – seconds) of implementation approaches.

Approach

1 2 3 4 5 6 7 8 9*

Read data** 38.05 38.05 38.05 38.05 38.04 0 0 0 0

Obtain coefficients 190.80 190.80 190.80 2.93 0.04 0.04 0 0 0
Filter *** 3085.65 396.68 12.08 2.59 2.59 2.59 2.59 0 0
Save data 38.86 8.57 38.86 8.57 8.58 8.57 0.03 0.03 0
Total time (seconds) 3353.35 634.1 279.78 52.14 49.24 11.2 2.61 0.03 0

* Prototyped but not extensively tested

** 10 minutes of data or 600,000 samples
*** Filter includes Decimation and Displacement Estimation

94
and iii) online-processing: utilizing the real-time cap­ Table 4. Autonomous monitoring framework descriptive
ability of the Free-RTOS operating system while not example.
interfering with the highest priority sampling task. The
improvement allows us to finish the whole process Train
30ms after sensing is done – or more than 100,000 Time position Gateway task Sensor nodes task
times faster than the first approach.
8:01 Not near Sleeping Sleeping
AM the bridge
2.5 Energy-efficient autonomous framework for 8:02 Approach Sleeping Wake up due to
monitoring railroad bridges AM the bridge ADXL362
Sense and record
With the addition of the wake-up sensor ADXL362, acceleration
the real-world clock DS3231m, and the 4G LTE- 8:04 Passed Sleeping Finish sensing
module, a framework is established. The framework AM the bridge due to ADSL362
provides the following critical functions of a railroad Back to sleeping
bridges SHM system without disrupting the effi­ 8:05 Not near Wake up due to Wake up due to
ciency of energy usage: AM the bridge DS3231m, DS3231m, listen
Request acceler­ for incoming
• Waking up of sensor nodes from deep sleep up­ ation data from all commands
on high acceleration input (i.e., train loading) sensor nodes
• Actionable information of dynamic displacement 8:06 Not near Collect data from Send data to the
computed online on the sensor nodes AM the bridge all sensor nodes gateway
• Minimal delay scheduled data retrieval and 8:07 Not near Done collecting Finishing sensing
upload to the cloud AM the bridge data, upload data data, back to the
• Notification directly sent to the bridge engineers to AWS server, sleep state
and decision-makers in case of emergency send the report to
bridge engineer
The basic framework is conveyed descriptively by via text message
using the following example shown in Table 4 and then come back to
visualized in Figure 3. sleep
… … … …
9:00 Not near Wake up due to Wake up due to
3 LABORATORY VALIDATION AM the bridge DS3231m, DS3231m, listen
Request sensor for incoming
To validate the performance of the whole system, nodes condition commands
a laboratory test is conducted. The displacement meas­ 9:00 Not near Collect sensor Send condition
urement from the field test is reproduced by Load and AM the bridge nodes’ conditions data
Boundary Condition Boxes (LBCBs), and dynamic 9:01 Not near Send report to Finish sending
AM the bridge bridge engineer data to the gate-
then come back to way, come back
sleep to sleep
… … … …
12:00 Not near Wake up due to Wake up due to
PM the bridge DS3231m, broad­ DS3231m, listen
cast real-world for incoming
clock synchroniza­ commands
tion to all the
sensor nodes
12:00 Not near Done broadcast- Adjust clock
PM the bridge ing, come back to based on the
sleep received message
then come back to
sleep

displacement is estimated through the accelerometer

attached to the LBCBs, as shown in Figure 4. The
LBCBs utilizes six independent actuators to allow con­
trol of the Cartesian coordinate. In the experiment,
only the vertical axis is controlled using 3-minute time
history data of field-measured the lateral displacement
Figure 3. Autonomous framework for monitoring of a timber trestle bridge in the CN Rockford line. As
railroad bridges. a means of comparison, a high-precision Krypton

95
Figure 4. Laboratory displacement tracking test setup.

Figure 6. Comparison between measured dynamic dis­

placement from Krypton camera and estimated dynamic
displacement from Xnode for the case of total vertical dis­
placement input from LBCBs.

Figure 5. Input displacement to the LBCBs.

K600 camera is used to track displacement based on

continuously capturing the position of LED tracking
markers.
In total, there are two types of tests done: input as
total displacement and as dynamics displacement Figure 7. Comparison between measured dynamic displace­
from the field test. Regarding the former type of test, ment from Krypton camera and estimated dynamic displace­
the input signal needed to be pre-processed by pass­ ment from Xnode for the case of dynamics vertical
ing through a low-pass filter to get dynamics dis­ displacement input from LBCBs.
placement before compared with estimated
dynamics displacement from Xnode. Each type of
input is scaled differently to fit into the range of
LBCBs. Figure 5 shows the total and dynamics dis­
placements input for a laboratory test. Table 5. Summary of errors from experimental tests.
The dynamic displacement is autonomously esti­
mated from acceleration measurement with the RMS Peak-to-peak
implementation from Approach 8 in the Xnode; it is Test error (%) error (%)
compared with displacement measured with the
Krypton camera system. Figure 6 and Figure 7 pre­ Total Displacement 1 3.86 1.31
sent the comparison of the results in time-domain. Total Displacement 2 3.66 5.17
Table 5 summarizes RMS and peak-to-peak error Dynamic Displacement 1 3.53 0.78
from the four tests. Good agreement in terms of Dynamic Displacement 2 3.32 1.01
measured and estimated data could be observed.

96
4 FIELD MONITORING of the tripod contributed by the ground and
wind) could be observed.
4.1 Bridge structures and deployments overview
Ten timber railroad bridges for freight trains in South­ 4.3 Railroad bridges monitoring result
ern Illinois, USA, were selected by careful consulting With the additional functionality of event-trigger, the
with the Federal Railroad Association and the Canad­ SHM system successfully operated for a 2-month
ian National Railway. Figure 8 and Figure 9 show an
example of one deployment setup with one base sta­
tion and three sensor nodes installed on one timber
railroad bridge. The sensors’ installation locations
were carefully chosen to collect more vibration data.

4.2 Field test validation of online displacement

estimation
A field displacement measuring using the com­
puter vision technique of checkerboard pattern
tracking was utilized to verify the performance of
the online dynamic displacement estimation
system. Figure 10 shows the test setup and the
comparison between the vision and reference-free
estimation of displacement is shown in Figure 11.
Note the disparity at the beginning of the record Figure 10. One Xnode sensor node and checkerboard pat­
is due to the late record button pressing. From tern for dispalcement tracking.
the 20th second toward the end of the record,
a good agreement (apart from the small vibration

Figure 11. Comparison between displacement estimated

Figure 8. Example of WSS SHM system instrumentation
online by the Xnode and measured by Camera.
on a timber railroad bridge.

Figure 9. Example of WSS SHM system instrumentation on a timber railroad bridge.

97
Figure 12. Sample lateral acceleration measurement and Figure 14. Battery voltage recorded during the
estimated displacement from a train event. deployment.

long period. By studying the change in the displace­ being able to operate for nearly two weeks (day 41
ment over time of the measurements from the two until day 53) with limited sunlight in the winter.
sensors, meaningful results for the bridge engineers To study the change in behavior over 1-year
could be obtained. Records of pier cap acceleration period, one more deployment has been done in late
and subsequently estimated displacement are com­ 2019 and more extensive analysis is being conducted
puted (Figure 12). Over the two months’ deploy­ and will be reported in the future research.
ment, the displacement is shown to be stable within
the two standard deviation range, signaling no sig­
nificant and noticeable change to the structural integ­ 5 CONCLUSIONS
rity (Figure 13).
Additionally, the WSS monitoring system proved In this paper, a WSSN system is proposed to over­
its capabilities to reliably capture the train crossing come the existing challenge of SHM of railroad
events with 944 data sets, corresponding to 419 bridges including: 1) limited energy; 2) short and
recorded train crossing events. Despite operating in random nature of train schedule; 3) un-available
a harsh environment of the two coldest month in Illi­ autonomous monitoring systems; and 4) difficult
nois, USA, the system maintained a level of 90% rapid decision-making process. The newly proposed
battery charged (3.82V on average and 4.0V at fully system with a wake-up sensor, a real-time clock, 4G
charged) because of the efficient use of energy. LTE network access integrated and smart onboard
Figure 14 shows the evidence of the sensor node data processing successfully proved its accuracy and
usefulness in assisting rapid inspection campaigns.
With all these improvements to overcome the chal­
lenges while monitoring railroad bridges, this system
is expected to be an important tool for decision
makers.

REFERENCES
2017 CFS Preliminary Data | Bureau of Transportation
Statistics. 2017. Retrieved 21 November 2019, from
https://www.bts.gov/surveys/commodity-flow-survey/
2017-cfs-preliminary-data
Adler R, Flanigan M, Huang J, et al. 2005. Intel mote 2: an
advanced platform for demanding sensor network
applications. Proceedings of the 3rd international con­
ference on embedded networked sensor systems, San
Diego, CA, 2–4 November, pp. 298–298. New York:
ACM.
American Railway Engineer and Maintained-of-way Asso­
Figure 13. Maximum lateral displacement is stable over the ciation. 2003. Practical Guide. Retrieved from https://
period of 60 days.

98
 www.arema.org/AREMA_MBRR/Store/Practical_Gui monitoring of a cable-stayed bridge using smart
de_TOC.aspx. sensor technology: deployment and evaluation. Smart
Analog Devices, 2016. ADXL362 Datasheet. Retrieved Structures and Systems, 6(5-6), 439–459.
from: http://www.analog.com/media/en/technical-docu Keil.com. 2017. CMSIS DSP Software Library. [online]
mentation/data-sheets/ADXL362.pdf. Available at: http://www.keil.com/pack/doc/CMSIS/
Association of American Railroads. 2019. A Short History DSP/html/index.html [Accessed 19 Sep. 2017].
of U.S. Freight Railroads. Retrieved from https://www. Ko, J., Ni, Y. & Chan, T. 1999. Dynamic Monitoring of
aar.org/wp-content/uploads/2018/05/AAR-Short-His Structural Health in Cable Supported Bridges, Smart
tory-American-Freight- Railroads.pdf. Structures and Materials 1999: Smart Systems for
Caicedo, J.M., Clayton, E., Dyke, S.J., Abe, M. & Bridges, Structures, and Highways, Proceedings of
Tokyo, J. 2002. Structural health monitoring for large SPIE, Vol. 3,671, pp. 161–172.
structures using ambient vibrations, Proc. of the ICANC­ Levis, P., Madden, S., Polastre, J., Szewczyk, R.,
EER Conference, Hong Kong, August, 15–20. Whitehouse, K., Woo, A., … & Culler, D. 2005. TinyOS:
Celebi M. 2002. Seismic Instrumentation of Buildings An operating system for sensor networks. Ambient intelli­
(With Emphasis on Federal Buildings), Technical gence (pp. 115–148). Springer, Berlin, Heidelberg.
Report No. 0-7460-68170, United States Geological Lynch, J.P. & Loh, K.J. 2006. A summary review of
Survey, Menlo Park, CA. wireless sensors and sensor networks for structural
Clinton, J.F., Bradford, S.C., Heaton, T.H. & Favela, J. health monitoring, Shock and Vibration Digest,38(2),
2006. The observed wander of the natural frequencies in 91–130.
a structure. Bulletin of the Seismological Society of Mayville, R., Rancatore, R. & Tegeler, L. 1999. Investiga­
America, 96(1),237–257. tion and simulation of lateral buckling in trains. Pro­
Farrar, C.R. 2001. Historical overview of structural health ceedings of the 1999 ASME/IEEE Joint Railroad
monitoring. Lecture Notes on Structural Health Moni­ Conference (Cat. No. 99CH36340) (pp. 88–93). IEEE.
toring using Statistical Pattern Recognition. Los Moreu, F. & LaFave, J.M. 2012. Current research topics:
Alamos Dynamics, Los Alamos, NM. Railroad bridges and structural engineering. Newmark
Federal Railroad Administration. 2010. Bridge Safety Structural Engineering Laboratory. University of Illi­
Standards, Final Rule. nois at Urbana-Champaign.
Fu, Y., Hoang, T., Mechitov, K., Kim, J. R., Zhang, D. & Mufti, A.A. 2003. Restoration and structural health moni­
Spencer, B.F. 2018. Sudden event monitoring of civil toring of Manitoba’s Golden Boy. Canadian Journal of
infrastructure using demand-based wireless smart Civil Engineering, 30(6),1123–1132.
sensors. Sensors, 18(12), 4480. Ni, Y.Q., Xia, Y., Liao, W.Y. & Ko, J.M. 2009. Technology
Fu, Y., Mechitov, K., Hoang, T., Kim, J.R., Lee, D.H. & innovation in developing the structural health monitor­
Spencer Jr, B.F. 2019. Development and full-scale valid­ ing system for Guangzhou New TV Tower. Structural
ation of high-fidelity data acquisition on a next- Control and Health Monitoring: The Official Journal of
generation wireless smart sensor platform. Advances in the International Association for Structural Control and
Structural Engineering, 1369433219866093. Monitoring and of the European Association for the
Fu, Y.G., Mechitov, K.A., Hoskere, V., Spencer, Jr., B.F. Control of Structures, 16(1), 73–98.
2016. Development of RTOS-based wireless SHM O’Connor, S.M., Zhang, Y., Lynch, J.P., Ettouney, M.M. &
system: benefits in applications, International Confer­ Jansson, P.O. 2017. Long-term performance assessment
ence on Smart Infrastructure and Construction. Cam­ of the Telegraph Road Bridge using a permanent wire­
bridge, UK, June 27- 29. less monitoring system and automated statistical process
Garner, A. C. & Huff, W. K. 1997. The wreck of Amtrak’s control analytics. Structure and infrastructure engineer­
Sunset Limited: News coverage of a mass transport dis­ ing, 13(5), 604–624.
aster. Disasters, 21(1), 4–19. O’Rourke, L., Read, K. & Johnson, E. 2015. U.S. Freight
Gomez, F., Park, J.W. & Spencer Jr, B.F. 2018. Reference- GHG Emissions by Consuming Industry Segment (Paper
free structural dynamic displacement estimation No. 13-4191). Retrieved from http://business.edf.org/files/
method. Structural Control and Health Monitoring, 25 2015/03/TRB_GHGemissions_poster_RELEASE.pdf.
(8), e2209. Sierra Wireless. 2018 AirPrime HL7588 Mini Card Acces­
Gunderson, D. 2015. Inspectors find dozens of rail bridges sory Board | 4G LTE Cat-4 | Sierra Wireless. Retrieved
need repair. Minnesota Public Radio, Retrieved from 1 November 2019, from https://www.sierrawireless.
http://www.mprnews.org/story/2015/07/23/inspectors­ com/products-and-solutions/embedded-solutions/prod
find-dozens-of-rail-bridges-need-repair. ucts/hl7588/.
Gurzinskil, J. Worst U.S. train crashes in recent history, Specker, L. 2016. Amtrak Sunset Limited disaster 23 years
2015 CBS News, Retrieved from https://www.cbsnews. later: 1993 Bayou Canot train wreck remembered. AL.
com/pictures/deadliest-train-crashes-in-recent-history/9/. com, Retrieved from http://www.al.com/news/mobile/
Huang, M., Shakal, A., Petersen, C., Celebi, M., Hooper, J. 2016/09/sunset_limited_disaster_a_retu.html.
& Klemencic, R. 2012. Strong motion instrumentation Spencer Jr, B.F., Park, J.W., Mechitov, K.A., Jo, H. &
of a 62-story concrete core residential building in San Agha, G. 2017. Next generation wireless smart sensors
Francisco. I Proceedings of SMIP12 Seminar on Utiliza­ toward sustainable civil infrastructure. Procedia Engin­
tion of Strong-Motion Data (pp. 81–96). eering, 171, 5–13.
IEEE Communications Society. 2019. AT&T to shut down Spencer Jr, B.F., Ruiz-Sandoval, M.E. & Kurata, N. 2004.
3G network in 2022; Verizon at end of 2019 – Technol­ Smart sensing technology: opportunities and challenges.
ogy Blog. Retrieved 1 November 2019, from https://tech Structural Control and Health Monitoring, 11(4),
blog.comsoc.org/2019/02/22/att-to-shut-down-3g-net 349–368.
work-in-2022-verizon-at-end-of-2019/. Wong, K.Y. 2004. “Instrumentation and health monitoring
Jang, S., Jo, H., Cho, S., Mechitov, K., Rice, J.A., of cable-supported bridges”, Structural Control and
Sim, S.H., … & Agha, G. 2010. Structural health Health Monitoring, 11(2), 91–124.

99
 Bridge Maintenance, Safety, Management, Life-Cycle Sustainability and Innovations –
Yokota & Frangopol (eds)
© 2021 Taylor & Francis Group, London, ISBN 978-0-367-23278-8

Why do they call Chongqing the bridge capital of China?

M.C. Tang
T.Y. Lin International China, Chongqing, China

ABSTRACT: Why do they call Chongqing the bridge capitol of China? Here are the main reasons: 1.
Bridge construction in Chongqing is most progressive; 2. Chongqing has most bridges in proportion to land
area; 3. Chongqing has the most variety of bridges; 4. Bridges in Chongqing are more accessible; 5. People in
Chongqing love their bridges; 6. Bridges in Chongqing are very beautiful!

1 INTRODUCTION there are other reasons why Chongqing may be rec­

ognized as the “Bridge Capitol of China”: 1. Bridge
Geographically, Chongqing is located slightly east in construction in Chongqing is constantly progressing;
the middle of China. The population in eastern 2. Chongqing has the most bridges in proportion to
China is much greater than in the west. Chongqing land mass; 3. Chongqing has the most variety of
has a population of about 32 million and a land area bridges; 4. Bridges in Chongqing are more access­
of about 83,000 square kilometers. By comparison, ible; 5. People in Chongqing love their bridges; 6.
the State of California has 425,000 square km of Bridges in Chongqing are very beautiful!
land and a population of 39 million. Bridge construction in Chongqing is constantly
The name Chongqing is pronounced as “Chong progressing. Many bridges in Chongqing were the
Ching.” In Chinese, Chong means “double”, Qing “first-of-their-kind” in the country and in the world.
means “festivity or celebration”. So, the name
implies that people in Chongqing are happy people!
Overall, 80% of Chongqing’s population are farm­
ers (before 1985). With so many farmers, the average
piece of land a farmer owned was very small. Farm­
ers were very poor. In order to lift them out of pov­
erty, it was necessary to encourage them to leave
their farms and move to urban areas where they
could take up industrial jobs. The aim was to encour­
age 60% of farmers to migrate to cities and have
20% of people to remain on their farms. The remain­
ing farmers could own four times the size of their
previous farmland and therefore, could live better.
To accommodate newcomers to the cities, urban
infrastructure was extensively upgraded and
expanded. Many new roads and bridges were built.
To date, urbanization is about 60% complete. More
roads and bridges are still needed.
This is especially true in Chongqing—a very hilly Figure 1. Map of China and location of Chongqing.
city. Chongqing is also bisected by the Yangtze River
and its many tributaries, such as the Jialing and Wu China missed the first two industrial revolu­
Rivers. Therefore, many bridges are needed connect tions of the world. Up until the 1980s, China’s
areas. In the last 30 years, Chongqing has built about steel production was very low. Almost all steel
16,000 bridges. Among them, 57 are major bridges used for bridge construction was imported at
in the metropolitan area alone, 24 bridges cross the that time. The bridge construction industry was
Yangtze River, and 33 bridges cross the Jialing basically non-existent. Even though China built
River. In addition to this great number of bridges the Qiantang River Bridge near Hangzhou in

DOI: 10.1201/9780429279119-9
https://doi.org/10.1201/9780429279119-9

100
1937 with steel imported from Germany, not
a single major bridge was built in China
between 1937 and 1950. The Yangtze River ran
across the entire China from west to the east
without a single bridge crossing it. During the
1950s, the former Soviet Union helped build the
first bridge over the Yangtze River in Wuhan.
The Soviet Union was to help build the second
bridge over the Yangtze River in Nanjing in the
1960s as well. But, for political reasons, Soviet Figure 2. The first Jialing River Bridge.
engineers left China before the start of construc­
tion. This forced Chinese engineers to do it
alone and they finally completed the Nanjing
Bridge in 1968. At the same time, they also
built the First Jialing River bridge in Chongqing
—an 88m span steel truss bridge. It may be
lauded as the first Chinese design and the first
long-span steel truss bridge built in China, with
all steel produced in China. And, this happened
in Chongqing.
The industrialization of China began in 1978,
after many years of political tumult, a period
when hardly any bridges were built. During this
entire time, Chongqing has been the most pro­
gressive city in China with respect to bridge
construction.
In 1981, the First Shibanpo Bridge was opened
to traffic. Its 174m main span was the longest Figure 3a. The First Shibanpo Bridge.
pre-stressed concrete girder span in China at the
time of its completion. We may compare this to
the first long-span prestressed concrete girder
bridge in the US, the 137m span Pine Valley
Creek Bridge in Southern California, completed
in 1975. However, the First Shibanpo Bridge has
prestressed concrete hammerhead piers. The
center piece of the bridge in each span is a drop-
in girder. This is an effective method that avoids
excessive long-term deflection.
The first long-span single-tower cable-stayed
bridge in China, the Shimen Bridge over the Jialing
River, was opened to traffic in 1988. This bridge has
a main span of 230m with cables located at the
middle strip and with roadways on both sides. This Figure 3b. The Second Shibanpo Bridge.
was an advanced design at the time. In the 1980s,
cable technology in China was not as advanced as in
the US. Large-diameter cables were still difficult to
produce in China. Accordingly, all the cables for the
Shimen Bridge were small cables and with two
cables at each anchorage point. However, because
the two cables were placed closely to each other,
they act as one cable.
By comparison, the first single-plane, cable-
stayed bridge in the US, the 366m span Sunshine
Skyway Bridge in Florida, USA was opened to traf­
fic in 1987. The Sunshine Skyway has two towers
and the Shimen Bridge has only one tower. There­
fore, technically, the span of the two bridges is com­
parable. Figure 4. Shimen Bridge.

101
 Later, the 440m span concrete cable-stayed bridge, Chongqing has been very progressive in arch bridge
the Lijiatau Bridge, was completed in 1998—another construction. The First Wujiang Bridge was completed
record span in China at the time. It has a simple slab in 1989. This 200m span concrete arch bridge was the
and edge girder cross section. This slab type girder first to use a horizontal swing construction technique.
was originally proposed by the German engineer Fin­ Arch bridges are very efficient with respect to con­
sterwalder and was first used for the Penang Bridge struction materials and quantities, but they are difficult
in Malaysia, when he was an advisor during the to build because the bridge is not stable until it is fully
bridge’s design stage. It is very efficient and simple. completed. Most old arch bridges were built using
The 440m span of Lijiatau is one of the longest spans local falsework. As the span gets longer, such false-
in the world for this bridge type. By comparison, cur­ work supports become very difficult and expensive.
rently the longest span concrete cable-stayed bridge Therefore, long-span arch bridges were not popular in
with a similar deck design in the US, the Dames the past.
Point Bridge in Jacksonville, has a main span of The Wujiang Bridge employed a completely new
396m. It was completed in 1989. construction method, without falsework. The idea is
very simple. An arch can be designed to have little to
no bending moment in the arch rib under dead load.
When we cut the arch at the middle point and separate
it into two halves and replace it with a horizontal sup­
port in the direction of the arch rib, the two half arches
are stable if the induced horizontal force remains in the
arch plane. Based on this, each half of the arch rib of
the First Wujian Bridge was built on forms set up on
the mountain slope by cutting out an arch shape there.
Half arches were poured in a position roughly perpen­
dicular to the bridge axis. After the concrete attained
enough strength, a tie rod was used to tie the end of
the half arch to a pivot point. The two halves of the
Figure 5a. First Wujiang Bridge. arch ribs were then rotated to align along the bridge
axis. A closure segment was poured to form the arch.
This proved to be a very efficient construction method
and the first in the world.
The world’s longest concrete arch bridge to date,
the 420m span Wanxian Bridge over the Yangtze
River was completed in 1997. At the time of its con­
struction, China’s bridge technology had just begun
to advance. The country still lacked heavy construc­
tion equipment. High-power highlines and lifters
were not yet available or too expensive to acquire.
So, a light steel truss arch skelett was constructed
first. Then the form for the three-cell box arch rib
was attached to the steel arch and the arch rib was
built segmentally from both ends. This construction
technique was very successful.
Figure 5b. First Wujiang Bridge.

Figure 6a. Wanxian Yangtze River Bridge, final bridge.

Figure 5c. First Wujiang Bridge.

102
 Table 1. Longest bridge spans in China.

Main Year
Bridge Type Name Span (m) Complete

Box Girder Shibanpo 330 2006

Beipanjiang 290 2013
Humen 270 1997
Arch Chaotianmen 552 2009
Lupu 550 2003
ZiGui 531 2019
Cable-Stayed Sutong 1,088 2008
Stonecutters 1,018 2009
Figure 6b. Wanxian Yangtze River Bridge during Edong 880 2007
construction. Suspension Yangsigang 1,700 2019
Xihoumen 1,650 2009
Runyang 1,490 2005

Table 2. Longest bridge spans in the world.

Bridge Main Year

Type Name Span (m) Country Complete

Girder Shibanpo 330 China 2006

Stolmasundet 301 Norway 1998
Costa e Silva 300 Brazil 1974
Arch Chaotianmen 552 China 2009
Lupu 550 China 2003
ZiGui 531 China 2019
Cable- Russky 1,104 Russia 2012
Stayed Sutong 1,088 China 2008
Figure 7. Chaotianmen Bridge, world’s largest arch span.
Stonecutters 1,018 China 2009
Suspension Akashi- 1,991 Japan 1998
Building technologically advanced bridge struc­ Kaikyo 1,700 China 2019
tures has been a tradition of Chongqing. This makes Yangsigang 1,650 China 2009
it a distinguished place and special to the country Xihoumen
and the world!

traffic in 2013. China built a 420m span self-anchored

2 CURRENT WORLD RECORDS suspension bridge over the Yellow River after that.
The 600m span Ergongyan Bridge is an-other giant
We can group all bridges into four categories: girder step forward in the world of bridge engineering.
bridges, arch bridges, cable-stayed bridge and suspen­ The Second Ergongyan Bridge is a new bridge
sion bridges. Chongqing has them all. Table 1 lists the located next to the First Ergongyan Bridge (opened
three longest span bridges in China and Table 2 lists to traffic in 2000). Chongqing specified that the new
the three longest span bridges in the world. In the bridge had to be a suspension bridge with the same
four categories of the world’s longest spans, Chong­ span length and tower height as the first suspension
qing holds two of the four world records: the world’s bridge. But the First Ergongyan Bridge is a true sus­
longest girder span bridge, the 330m span Shibanpo pension bridge with anchors at both ends. The new
Bridge (opened to traffic in 2006) and the world’s bridge did not have space for its cable anchors.
longest span arch bridge, the Chaotianmen Bridge Thus, it was designed as a self-anchored suspension
(completed in 2009). This is, by itself, a world record. bridge. The challenge in building a self-anchored
In the subgroups, the Wanzian Bridge is still the suspension bridge lies in the construction sequence;
longest concrete arch span in the world while the the main cables are anchored to the ends of the
newly completed. Second Ergonyan Bridge, with its bridge girder so the bridge girder must be completed
600m span, is, by far the longest span for a self- first before the main cable can be installed. Without
anchored suspension bridge. When the Eastern Span of the main cables to suspend the girder, the girder
San Francisco-Oakland Bay Bridge in the US was cannot support itself. Girders of some self-anchored
designed in 2000, it was the world’s longest self- suspension bridges were built using falsework. Such
anchored suspension bridge span. It was opened to is the case with the San Francisco-Oakland Bay

103
Bridge. Falsework in the river was not possible in transferred to the main cables. This is a rather com­
the Ergonyan Bridge due to navigational require­ plicated procedure. Nevertheless, construction of the
ments. Instead, the bridge was first built using cable- bridge was successful.
stays. Because for a cable-stayed bridge the towers
are almost twice as high as a suspension bridge, an
extension was added above each tower and the tem­
porary stay cables were erected to support the girder
temporarily. The main suspension cables were then
erected with their ends anchored to the ends of the
completed girder.

Figure 9. New Ergongyan Bridge construction stage.

Figure 8a. Old Ergongyan Bridge.

Figure 10a. Jiayue Bridge.

Chongqing has built many beautiful bridges that

fit well with their natural surroundings. The
Figure 8b. New Ergongyan Bridge. Caiyuanba Bridge is a two-level bridge with six lanes
on the upper deck and two monorail tracks on the
lower deck. Monorail fits well with the hilly Chong­
After the main cables were completely installed
qing landscape because it can achieve steeper slopes
and the suspenders attached to the main cables and
and accommodate smaller horizontal curves. Many
the girder, the weight of the girder was then

104
Chongqing bridges carry both rail and highway traf- 3 PEOPLE IN CHONGQING LOVE THEIR
fic. The City has also built several rail-only bridges. BRIDGES
The Jiayue Bridge over the Jialing River is
a partially cable-supported girder bridge. Its towers 35 years ago, before many bridges were built,
are much lower than a conventional cable-stayed people in Chongqing had to ride the ferries to
bridge, so it fits well with a landscape where the cross the rivers. This was time consuming and
bridge is very tall and connects the mountain tops on could be dangerous when the waters were rough.
both ends. For this reason, people in Chongqing appreciate
The First Fujing Bridge and the Lijia Bridge offer and love bridges. At each bridge opening there
Chinese architectural elements. Similarly, the are always big crowds.
Second Yangtze River Bridge at Chengsu, currently People walk on the bridges oftentimes as it is
under construction, features Chinese architecture. a typical pastime. One of the bridges, the Qianxi­
men Bridge, is especially well liked because one
can see the beautiful cityscape along the Jialing
and Yangtze Rivers. The City opened the bridge
exclusively to pedestrians during public holidays
in order to celebrate. Some even picnic on the
bridge. This makes it a special tourist attraction
of the City.

Figure 10b. First Fujiang Bridge.

Figure 10c. Lijia Garden Bridge.

Figure 11b. Dongsuimen Bridge and Qianximen Bridge

and people on the bridge during the holidays.

Figure 11a. Dongsuimen Bridge and Qianximen Bridge.

105
 of its completion. It has a span arrangement of 86.5
+4x138.0+156.0+174.0+104.5 meters. Due to traffic
increase a second bridge had to be built next to the
old bridge. The owner wanted a similar bridge as the
first one, so a girder bridge was mandated. But the
Waterway Department, based on hydraulic studies,
insisted that one of the main span piers of the new
bridge be eliminated. This is because the original
Shibanpo Bridge, completed in 1981 has a main
span of 174m, which was appropriate back in the
1980s, when river traffic was low. That would have
been too narrow for today’s navigation, especially if
a new pier were to be placed alongside the old pier,
which would cause a tunnel effect so the ships
would have difficulty to maneuver in the river. By
taking out one pier, the new bridge’s main span
became 330m, which is a world record span. For
a 330 span a cable-stayed bridge or an arch bridge
would have been more economical. However, due to
aesthetic reasons, the City insisted that the new
Figure 11c. Qianximen Bridge. bridge must be a girder bridge that matched the
existing bridge as much as possible.

4 CHONGQING BRIDGES ARE ACCESSIBLE

The width of the Yangtze River inside the metropol­

itan area is about 1.5 km. The width of the Jialing
River is about 1 km wide. This is a nice distance to
walk across. Most bridges in Chongqing have
accessible pedestrian paths. People walk between the
two river banks to work, to jog or walk just for fun.
This is different from Nanjing or Wuhan, where the
Yangtze River is much wider. In addition, the main
span, which is usually the most interesting part of
the bridge, is in the middle of the wider river, which
is far apart from both shores. If people wanted to
look at the main span from up close, they would
have to hire a boat. In this way, people are not able
to develop a feeling towards bridges. Because resi­
Figure 12. The 103-m-long center piece of steel box girder
dents are more familiar with the bridges in Chong­ for the Second Shibanpo Bridge being towed 1,000 km
qing, there is a special place in their hears for the from factory to bridge site.
bridges.

5 CHONGQING BRIDGES ARE BEAUTIFUL

As mentioned above, in Chongqing the city govern­

ment pays special attention to aesthetics of bridges.
The Second Ergongyan Bridge described above is
a good example. For a 600m span bridge, a cable-
stayed bridge is more economical than a suspension
bridge, and especially more economical than a self-
anchored suspension bridge. But the City insisted
that the Second Ergongyan Bridge be suspension
bridge to match the existing suspension bridge. The
two are parallel to one other, only a few meters
apart.
Another example is the twinning of the Shibanpo
Bridge. The first Shibanpo Bridge was open to traffic Figure 13a. Shibanpo Bridge before the middle portion
steel box was attached to the concrete cantilever on both
in 1981. The bridge was the longest span pre­
ends.
stressed concrete girder bridge in China at the time
106
 bridge is built alongside the existing bridge.
Another such bridge, the Lijiatou Bridge is now
under construc-tion. The new bridge must match
closely with the old bridge for appearance’s sake.
Cities like Chongwing are willing to pay more to
achieve their aesthetics.
The Twin River Bridges consist of the Qianxi­
men Bridge over the Jialing River and the Dong­
shuimen Bridge over the Yangtze River. These
two bridges are located at the tip of Chongqing.
Aesthetics played an important role in the selec­
tion of the bridge schemes. The location of the
bridges is such that they had to be beautiful by
themselves but also harmonious with each other.
Figure 13b. Shibanpo Bridge after the middle portion steel
The bridges must not be too heavy so as not to
box was attached to the concrete cantilever on both ends.
obstruct the view of the City. Each bridge carries
four lanes of city traffic on the upper deck and
There was a common belief in China at the time two tracks of rail traffic on the lower level.
that a prestressed concrete girder bridge should not During the proposal stage of consultant selection
have a span more than 300m because most long-span process, aesthetics was one of the most important
prestressed concrete girder bridges have excessive se-lection criteria.
deflection problems due to creep and shrinkage of con­
crete, a problem that still has no definite solution. The
longest span steel girder bridge to date is the 300m
span Rio Niteroi Bridge in Brazil, completed in 1974
while the longest span prestressed concrete bridge is
the 301m span Stolmasunde Bridge in Norway, com­
pleted in 1998. The middle portion of the Stolmasunde
Bridge was built using light-weight concrete, the light
weight aggregate of which was imported from the
United States. There was little experience with light­ Figure 14. Dongshuimen Bridge at the left and Qianximen
weight concrete in China at that time. Importing the Bridge at the right of the Yuzhong Peninsula.
aggregate from the US was economically prohibitive.
Finally, an innovative hybrid design using a 103m­
long steel box girder in the middle of the main span to 6 SUMMARY
lighten up the girder was conceived. The steel box was
fabricated in Wuhan, on the Yangtze River about Chongqing is a very beautiful city. The people
1,000 km downstream of Chongqing. The box girder are keenly aware and are very proud of the
was designed also as a barge so after closing both ends beauty of their City. They scrutinize the aesthet­
it was launched to the river. It was then towed ics of each bridge structure to make sure it fits
upstream to Chongqing and lifted in place. It was con­ well with the cityscape. Because the government
nected to the prestressed concrete portion of the bridge and the people are very concerned about the
to form a continuous structure. beauty of the City, all bridges in Chongqing are
Due to increased traffic in the City, many very beautiful.
older bridges are overloaded. Usually, a parallel

107
 Bridge Maintenance, Safety, Management, Life-Cycle Sustainability and Innovations –
Yokota & Frangopol (eds)
© 2021 Taylor & Francis Group, London, ISBN 978-0-367-23278-8

Applications of AI, BIM, and sensing to bridge maintenance

N. Yabuki
Osaka University, Suita, Japan

ABSTRACT: As a huge amount of structures built during the high economic growth period are approaching
the end of their life, the extension of the longevity of bridges on a limited budget has become a crucial issue.
Due to the retirement of a large number of maintenance engineers with expertise and lack of the young gener­
ation in this field, sustainability and knowledge transfer have become significant concerns. However, it is dif­
ficult to spend substantial money on infrastructure maintenance unlike construction. Therefore, advancing
Information and Communication Technology (ICT) is expected to lead to smart bridge maintenance. This
paper reviews the research on the applications of Artificial Intelligence (AI), Building Information Modeling
(BIM), and sensing to bridge maintenance, discuss their effects and challenges, and consider future directions.

1 INTRODUCTION have serious problems in storing related documents

and maintenance data, and thus, it is difficult for
1.1 Inspection and data management them to access necessary data once a severe accident
or disaster would occur.
The whole bridge maintenance system can be repre­
sented as a cycle of 1) inspection, 2) investigation/sens­
ing, 3) health assessment, 4) decision making, 5) 1.2 The objective of the paper
execution of construction and recording data/docu­
The number of aging infrastructures has been
ments and should be managed properly. Although
increasing rapidly. As a huge amount of struc­
much research has been done for sensing, health
tures built during the high economic growth
assessment, decision making, and construction/repair
period are approaching the end of their life, asset
methods, little attention had been paid to inspection
management, especially, the extension of the lon­
and recording data/documents until about
gevity of bridges, on a limited budget has
a decade ago.
become a crucial issue. Furthermore, due to the
On December 2, 2012, nearly 150 concrete ceiling
retirement of a large number of maintenance
panels inside the Sasago Tunnel on the Chuo
engineers with expertise and lack of the young
Expressway in Japan collapsed, killing 9 and injur­
generation in this field, sustainability and know­
ing 2 people. Most of the Japanese civil engineers
ledge transfer have become significant concerns.
were very shocked and realized that insufficient
However, it is difficult to obtain substantial
inspection and repair of infrastructures such as
budget on bridge maintenance unlike construc­
bridges and tunnels would cause serious conse­
tion. Therefore, nowadays, the applications of
quences that are fatal to humans. The Ministry of
advancing Information and Communication Tech­
Land, Infrastructure, Transport and Tourism (MLIT)
nology (ICT), which is making a big impact on
decided to mandate the inspection of all bridges and
both society and industry, are expected to lead to
tunnels in Japan at least once in every 5-year period
innovative efficiency and to the improved sophis­
by close visual contact and hammering test, starting
tication of the bridge maintenance.
from April 2014. Although all bridges and tunnels
This paper reviews the research on the applica­
had been inspected by March 2019, both bridge
tions of cutting-edge ICT such as Artificial Intelli­
owners and industry engineers have recognized that
gence (AI), Internet of Things (IoT), Building
the current inspection method is labor-intensive,
Information Modeling (BIM), Sensing to bridge
error-prone, dangerous, and it takes much time,
maintenance, particularly, inspection and data man­
effort, and cost. Moreover, all the inspection reports
agement, discuss their effects and challenges, and
are stored in the offices independently as PDF files
consider the future directions.
and cannot be queried. In fact, most bridge owners

DOI: 10.1201/9780429279119-10
https://doi.org/10.1201/9780429279119-10

108
2 AI FOR BRIDGE MAINTENANCE Internet of Things (IoT), etc. Now that such Big
Data will possibly become datasets needed for train­
2.1 AI in general ing, the timing is perfect for AI.
AI started by the so-called Dartmouth conference
organized by John McCarthy in 1956 and the first AI 2.2 Applications of AI to bridge engineering
boom, of which research was related to search and The data used for deep learning in bridge mainten­
reasoning, occurred in the 1960s. However, once it ance can be grouped into 3 categories, i.e., images,
was known that only toy problems could be solved sound (vibration), and texts. The main research in
by the AI then, the boom passed. each data source is as follows.
In the 1980s, many people were excited about the Active research can be seen in recognizing defects
anticipation of knowledge systems or expert sys­ such as cracks, free lime, exposure of reinforcing bars
tems, which started the second AI boom. AI of concrete structures (Figure 1), road cracks and pot­
researchers then thought that computers could think holes, rust and paint deterioration of steel structures
like human beings if knowledge is represented as from digital images. Furthermore, members such as
rules by a human language, which is a physical girders, piers, abutments, cables, etc., are detected
symbol system, and if it is driven by the inference from images, and the kinds and performance of con­
engine. The idea of separating the knowledge and struction workers and machinery are detected from
inference engine was revolutionary and a huge images for repair and refurbishment construction.
budget was allocated to AI research in developed As for sound, research on diagnosing concrete by
countries. However, in the early 1990s, people real­ the sound of hammering test and impact noise of
ized that working expert systems could be developed laser beams to concrete has been active. Research
only in very limited domains although their develop­ can be seen in finding loose bolts by hammering
ment cost, time, and labor were enormous. tests and accelerometers attached to steel structures.
The second AI boom plunged into winter and the AI research in bridge maintenance can also be
term, AI, became obsolete. seen in making rules or clustering conditions from
Then, neural networks, which were not regarded table data including character strings in ledgers and
as AI during the second AI boom, fortunately, were inspection reports.
applied to machine learning for recognition of vari­
ous things or patterns from images, sound, sensing
data, etc. In conventional machine learning based on
simple neural networks, a feature, which is an indi­
vidual measurable characteristic of a phenomenon
being observed, must be selected and threshold
values must be specified by researchers, while the
performance and accuracy rate were usually disap­
pointing. However, in deep learning, which is based
on Convolutional Neural Networks (CNN) with
many layers of neurons, features can be extracted
automatically without human specification and the
accuracy rates increased significantly. The Univer­
sity of Toronto team won the first prize in the Ima­
geNet Large Scale Visual Recognition Challenge
(ILSVRC) 2012 and achieved a breathtaking
decrease in error rates. In 2016, Google DeepMind’s
AlphaGo program defeated the world human Go
champion then and the third AI boom has begun.
Now, deep learning or CNN is regarded as almost
equivalent to AI. Although deep learning is an excel­
lent algorithm, a large amount of data is required as
the training dataset. If a small amount of data in
some specific narrow area is used for training, over-
fitting usually occurs.
When AI’s third boom began, the term “Big
Data” has begun to be paid attention. Big Data
means literally, data which is so big that conven­
tional database management systems cannot manage.
Big Data includes texts and images that the users of
Social Networking Services (SNS), transactional
data of online shipping, video data of security cam­ Figure 1. Cracks, free lime, and exposure of reinforcing
eras, dashboard cameras, data derived from the bars of concrete structures.

109
2.3 Challenges of AI applications should have more chances to “relearn” AI and its
applications.
To obtain high accuracy in deep learning, a huge
When talking about AI, many people have
amount of training data is required. For detecting
a concern that AI will steal humans’ jobs. Looking
special objects in a specialized domain such as
back into history, we realize that innovations, while
bridge inspection from images, less amount of data
increasing productivity revolutionarily, have stolen
would be required to obtain satisfactory results by
many people’s jobs, e.g., coachmen, running express
employing fine-tuning or transfer learning tech­
messengers in the Japanese “Edo” era, tracers (draft­
niques, borrowing excellent CNN models developed
spersons who traced drawings by ink), typists, etc.
by universities and research institutes. On the other
While AI will steal some humans jobs, AI will
hand, these techniques may not be applicable to
create new jobs just like the case of CAD that cre­
other problems and have limitations. Thus, a large
ated 2D CAD operators in place of tracers.
amount of data set is necessary to apply AI to bridge
maintenance. However, owners or managing organ­
izations of civil infrastructures including bridges are 2.4 Future prospective of AI applications
usually very reluctant to provide their data, whether
Some people say that the current AI boom is hitting
they are photo images or sensing data, to other
the peak and will be passing in the near future. How­
organizations, or even to other sections within the
ever, the author believes that the current AI boom
same organizations. Furthermore, it would take
will continue much longer and stronger than
much time and effort to make a clean dataset for
the second one because of the following reasons.
machine learning. Therefore, researchers who
The current AI can be applied to relatively simple
obtained the data from those reluctant organizations
and broad problem domains and once the necessary
in a painstaking way and who made their training
dataset is obtained, the training can be left to com­
dataset would not want to open it to others. In com­
puters without having any interviews with experts.
puter science, especially, in the AI research field,
This coincides with the prevail of IoT and the avail­
a new movement has been observed that those who
ability of Big Data.
openly provide large and high-quality training data
for AI researchers should be evaluated highly in an
academic sense. In bridge engineering, similar effort
3 BIM FOR BRIDGES
should be done and we should establish an open
environment and sharing rules for owners and
researchers to provide datasets. 3.1 BIM
Although expert systems of the second AI boom Currently, bridges are represented by 2D drawings
could explain why and how the answer was deducted except for some special purposes such as landscape
logically by the human language, current deep learn­ simulation, 3D Finite Element Analysis (FEA), etc.
ing systems of the third AI boom can only tell the Although 2D drawings can accurately represent any
algorithms and dataset used for learning but cannot 3D objects theoretically, it is very difficult to avoid
answer the questions of why and how by natural all mistakes in multiple 2D drawings. These mis­
human languages. Thus, there is a concern that takes could give significantly negative impacts on
people, especially experts such as engineers, may the lifecycle of artifacts. Furthermore, 2D drawings
have a suspicious feeling toward AI and may be cannot facilitate automatic data conversion and
reluctant to use such AI systems. In order to solve input/export between multiple software applications.
this problem, the accumulation of experiences of Therefore, the manufacturing industry, especially in
using AI systems many times successfully by con­ automobiles and airplanes, has been using 3D CAD
firming the answers provided from AI is correct extensively, adopting Computer Integrated Manufac­
would be important. Such accumulations will turing (CIM) which is an integration of CAD, Com­
improve the social cognition of AI. The current situ­ puter Aided Engineering (CAE), Computer Aided
ation is somewhat similar to the one in the dawn of Manufacturing (CAM), and Computer Aided Testing
computers when experts had a doubt in the results (CAT) to increase the efficiency significantly since
that the computers made. around 1980.
Data scientists, who are experts in AI as well as As for buildings, research on utilization of 3D
statistics, play a very important role in applying AI CAD to design and construction began in the late
to bridge maintenance. At the same time, bridge 1980s and the industry has begun to adopt 3D
engineers should collaborate with data scientists and models since around 2004 with the new termin­
should have, at least, minimum data science literacy, ology, Building Information Modeling (BIM). BIM
including basic knowledge about machine learning. is intended to integrate the whole lifecycle of build­
Therefore, educational programs in the department ings by sharing a single 3D model by all the players
of civil engineering or related fields in universities of the project with the methods called front loading
should incorporate more coursework related to data and concurrent engineering, including planning,
science, AI, computer science. Furthermore, civil architectural design, structural design, facility
engineers who already graduated from universities

110
design, construction, operation and maintenance, 3.3 International standardization of product
and demolition. models
In steel bridge manufacturing, 3D CAD technol­
If each country, company, or organization develops
ogy has already been utilized since the 1990s. For
its own product model for a domain, the interoper­
example, cambers are computed to design steel plate
ability of data among different software applications
cutting, bolt-holes, etc. and temporary assembly of
cannot be realized. Thus, product models are devel­
steel bridge members at the factory has been
oped following the international standard, ISO 10303,
replaced with the virtual assembly of 3D member
Industrial automation systems and integration - Prod­
objects in 3D CAD. However, 3D technology has
uct data representation and exchange, also known as
been used solely in bridge manufacturing at the fac­
ISO STEP, which stands for Standard for the
tory and 3D model data has not been shared with
Exchange of Product model data. In the domains of
design, construction, or operation and maintenance
machinery, vessels, and plants, product models were
phases. Recently, BIM has gradually been adopted in
developed and standardized following the ISO STEP
various infrastructure fields such as roads, bridges,
in the 1990s. But the product model development in
tunnels, harbors, railways, etc. and it is often referred
the building domain was late. Thus, in 1996, an inter­
to as BIM for infrastructure or InfraBIM.
national consortium named International Alliance for
Interoperability (IAI) was founded and it started to
3.2 Object-oriented product models develop a product model for buildings, called Indus­
try Foundation Classes (IFC). IAI changed its name
3D models created by using conventional CAD or
to buildingSMART International (bSI) later and IFC
Computer Graphics (CG) are mere sets of geometric
has become an international standard, ISO 16739 in
shapes such as points, lines, circles, curves, surfaces,
2013.
solids, etc. In the late 1960s, object-oriented technol­
The French research group started developing
ogy was born in computer science and in the 1980s,
a product model for bridges in the late 1990s and
3D CAD changed to object-oriented, where objects
released the first draft of IFC Bridge, which was an
in 3D environment are represented as objects or
extension of IFC to represent bridge members in
products such as wheels, disks, bolts, beams, col­
2002. Around the same time, another bridge model
umns, etc., which are meaningful to humans instead
was developed by Yabuki et al. by extending IFC.
of just geometry such as lines, circles, etc. This kind
Then, those two bridge models were merged into
of object-oriented model is called ‘product model.’
one as a new IFC Bridge between 2005 and 2007
Product models can be represented by two differ­
(Yabuki et al. 2006). However, fully-fledged devel­
ent types of objects: class and instance. A class is
opment of IFC Bridge had to wait until 2013 when
an abstract concept which is represented by
IFC for buildings became ISO standard. In 2013,
a general terminology such as “bridge,” which is
Infrastructure Room was founded in bSI to develop
a group of objects that can be called bridges. An
product models for infrastructures. They are IFC
instance is an actual, individual, existing object of
Bridge, IFC Road, IFC Rail, IFC Ports & Water­
the class such as Yokohama Bay Bridge. Each
ways, and IFC Tunnel. They will be complete and
instance can have its characteristic data called
integrated as IFC 5, which means IFC version 5, by
“attribute” such as bridge name, location, bridge
the end of 2020 and will be ISO 16739 soon after.
length, main material, etc., and attributes defined in
a class can be inherited to its instances. Classes can
be divided into more concrete classes as “bridge” 3.4 Bridge product model
can be classified as “concrete bridge,” “steel
Since the current IFC Bridge product model is being
bridge,” “masonry bridge,” etc. On the other hand,
developed as a fast-track project in bSI, it does not
classes can be divided into parts of classes as
cover all the bridge types and all members but just
“bridge” can be divided into “superstructure,” “sub­
typical ones. Therefore, the author and his group
structure,” and “bearing.” Classes can be repre­
members are developing a revised IFC Bridge prod­
sented as a network of relationships between
uct model that all types of bridges and all types of
related classes by this way. This network represent­
members would be covered. During the conceptual
ing the body of knowledge of some particular
modeling phase, an ontology was employed. Con­
domain such as bridges is called the product model.
ceptually, the developed bridge product model is div­
Thus, a product model is defined as a generic con-
ided into the following four main classes: 1) Bridge,
text-rich networked data model representing 3D
of which attributes include its name and bridge
geometric data, various attribute data, and relation­
type, 2) Bridge Part, which represents compositions
ships of all members of products such as automo­
of a bridge, such as superstructure, substructure,
biles, plants, buildings, bridges, etc. A data model
etc., 3) Member, which represents a member of
described as a generic product model is called
a bridge, such as girder, pier, cross girder, etc.,
schema, which is clearly differentiated from some
and 4) Member Part, which represents an element
particular, individual bridge data that can be
that makes a member together with other elements,
described as an instance file based on the bridge
such as web, flange, stiffener, etc. The conceptual
schema.
111
 structural health monitoring has usually been done for
a limited period with a limited number of sensors so
far. Now, thanks to Micro Electro-Mechanical Systems
(MEMS), progress of Information and Communication
Technology (ICT), especially Internet of Things (IoT)
and Low Power, Wide Area (LPWA) networks, sen­
sors have become very small, light, and wireless, bat­
teries could work for several years, and sensing data
can be transmitted to the data point for about 50km.
Thus, it is expected that many small sensors to be
installed for important bridges to monitor their behav­
ior for a long time.
However, if a very large number of sensors are
installed for a bridge, it will be difficult for engineers
to evaluate the sensing data by referring to 2D draw­
ings showing the location and direction of each
sensor. In order to solve this problem, first,
a context-rich 3D bridge product model such as
Industry Foundation Classes (IFC) should be devel­
oped. Then, a sensor (IoT) information model should
be develosec10_3_4ped. Finally, both of the data
models should be linked together so that the system
can automatically make complicated queries about
the relationship between the bridge and sensors, and
can infer the bridge behavior based on the sensing
data and condition of the bridge. As described in sec­
tion 3.4, a bridge product model is being developed
by us. We are also developing a sensor data model
and a linkage model between the bridge product
model and the sensor data model.

4.2 Previous research on sensor data modeling

The most well-known sensor data model is SensorML
(Sensor Modeling Language), which was firstly devel­
oped under the Sensor Web Enablement (SWE) activ­
ity of Open Geospatial Consortium (OGC). SensorML
is now a set of OGC approved standard models and an
Figure 2. Conceptual bridge product model developed by Extensible Markup Language (XML) encoding for
the author and his group. describing processes of sensors and measurement.
Another standard related to sensor data models is
Observations and Measurements, known as O&M,
which is an international standard defining
bridge product model is represented in the Unified a conceptual schema encoding for observations and
Modeling Language (UML) in Figure 2. Based on features such as time.
the conceptual model, we are modifying the current Installation of sensors to bridges is usually drawn
IFC Bridge product model, which is represented in on 2D drawings so that engineers must read them care­
the EXPRESS language of ISO 10303. fully to evaluate sensing data according to the sensor
types, locations, directions, installation method, char­
acteristics of the member which the sensor is attached,
4 LINKAGE BETWEEN BRIDGE PRODUCT bridge types, defects or deterioration of the bridge, etc.
MODEL AND SENSOR DATA MODEL However, if the number of sensors to be installed to
each bridge becomes very large, e.g., 1,000, it would
4.1 Sensing take a long time and effort for the engineers and
researchers to evaluate the behavior of the member or
Much research has been done to monitor the behavior
the bridge because the process of referencing the sens­
of infrastructures, especially bridges, with sensors to
ing data to the sensors and members that the sensors
detect anomalies. However, most sensors in the past
are attached cannot be automated at the moment.
were large and heavy, needed long and heavy cables,
Peng & Law (2004) proposed a reference data
alternate current electric power, and data processing
model for storing and querying a large amount of data
and storing units near each bridge. Therefore,
obtained at the experimental facilities of George
112
E. Brown, Jr. Network for Earthquake Engineering includes the location of the sensor is attached, to
Simulations (NEES) of the US. The NEES Reference which member is the sensor attached, which side of
Data Model was represented by both the relational the member, to which element of the member such as
data model and the object data model. Yabuki & upper flange, web, or lower flange, how the sensor is
Yoshida (2006) proposed a data model, EDgrid, for attached, e.g., screwed shut, glued, or welded, the dir­
storing and querying experimental data of E-Defense, ection of the sensor corresponding to the member
which is the largest 3D shake table facility in Miki, alignment and to the bridge alignment. The data model
Japan. Both data models have not been actually for representing the above information can be repre­
utilized. sented as a class named Sensor Installation as shown
Jeong et al. (2017) developed a database that can in Figure 4. This class is connected with the Member
integrate bridge information model and sensor data class in the bridge product model and Sensor Device
model based on NoSQL, known as Not only SQL. class in the sensor data model as shown in Figure 5.
They extended SensorML to define an attribute to
represent the sensor’s location of the member of the
bridge. They also employed cloud computing to pro­
vide bridge monitoring and management.

4.3 Sensor data model

The sensor data model which we are developing is
shown in Figure 3. The sensor data model consists of
five classes: 1) Network, which represents a network
of multiple sensors for a bridge, 2) Sensor Device,
which represents a generalized sensor device, com­
posed of the following three classes, 3) Sensor, which Figure 4. Sensor installation data model.
represents each sensor, 4) Sensing Device, which
includes data logger, amplifier, cable, computer,
etc., 5) Support Device, which includes light, survey
equipment, etc.

4.4 Sensor installation data model

The linkage between the bridge product model and
sensor data model can be achieved by describing and
recording how each sensor is installed to a member of
the bridge. The important installation information

Figure 5. Linkage between bridge product model and

sensor data model by sensor installation model.
Figure 3. Sensor data model.

113
5 CONCLUSION ACKNOWLEDGMENT

This paper, first, raised an issue that little attention This research has been partially done as
had been paid to inspection and data management in a commissioned research on “Research & Develop­
bridge maintenance until about a decade ago and ment and Standardization on Coordination between
then, pointed out that advanced ICT should be Heterogeneous Domains of 3D Infrastructure
employed to increase the productivity and quality of Models and IoT Sensor Information Models in Infra­
the maintenance works. The paper reviewed AI in structure Monitoring” from the Ministry of Inter­
general and AI applications to bridge inspection and national Affairs and Communications (MIC) under
summarized several issues to be considered when AI the public invitation for “Strategic Information and
is applied to actual practices. Next, the paper Communications R&D Promotion Program
described how BIM was born and has been growing (SCOPE), International standard acquisition type,
and the importance of utilization of object-oriented Japan-US cooperative research” in FY2018.
product models and their international standardiza­ The author would like to express his sincere grati­
tion. Although a bridge BIM model, IFC Bridge, has tude to Prof. Kincho H. Law, Stanford University and
been developed by bSI and will be a part of ISO Prof. Jerome Lynch, University of Michigan, Ann
16739 soon, due to the fast track project, it does not Arbor. The author extends his thanks to Dr. Takashi
cover all types of bridges and members. Thus, the Aruga of Conport, Co., Ltd., and colleagues from
author and his group are developing a complete Kansai Institute of Information Systems and OKI Elec­
bridge product model by extending the current IFC tric Industry Co., Ltd.
Bridge. Thirdly, the paper discussed the reason why
the sensor data model and bridge product model
should be linked and our models were explained. REFERENCES
By these novel technologies and models,
inspection and sensing data will be linked to 3D Jeong, S., Hou, R. Lynch, J. & Law, K.H. 2017. An Infor­
semantic-rich bridge product model and all the mation Modeling Framework for Bridge Monitoring.
Advances in Engineering Software, 114, pp. 11–31.
data will be integrated so that the user can make
Peng, J. & Law, K.H. 2004. Reference NEESgrid Data
various queries and extract necessary data from Model, Technical Report NEESgrid-2004-40.
a huge amount of bridge data. This can be Yabuki, N., Lebegue, E., Gual, J., Shitani, T. & L,. Z. 2006.
achieved by linking remote, heterogeneous data­ International collaboration for developing the bridge
bases or data servers with Application Program product model IFC-BRIDGE. In Proc. of the Joint
Interface (API) without making an integrated Intern. Conf. on Computing and Decision Making in
huge database. Once the linked data platform is Civil and Building Engineering, Montreal, Canada:
developed, AI researchers will be able to make 1927–1936.
very big training datasets for deep learning, Yabuki, N. & Yoshida, Y. 2006. A data model for storing
a large amount of sensor data. In Proc. of the First Asia-
which will facilitate the AI application to bridge
Pacific Workshop on Structural Health Monitoring,
maintenance significantly. Yokohama, Japan, Paper No. 36: 1–8.

114
 MINI-SYMPOSIA
MS01: Novel techniques regarding the assessment and monitoring of
bridges
Organizers: A. Strauss & D.M. Frangopol
 Bridge Maintenance, Safety, Management, Life-Cycle Sustainability and Innovations –
Yokota & Frangopol (eds)
© 2021 Taylor & Francis Group, London, ISBN 978-0-367-23278-8

Overview on the prestress loss evaluation in concrete beams

M. Bonopera
National Center for Research on Earthquake Engineering, Taipei, Taiwan

K.C. Chang & Y.C. Ou

National Taiwan University, Taipei, Taiwan

ABSTRACT: In the last few decades, prestressing techniques have been used to build very important struc­
tures, infrastructures, and bridges. Since the serviceability and the safety of prestressed concrete members
rely on the effective state of prestressing, development of tools and dynamic procedures capable of estimating
the effective prestress loss have been widely carried on. Amongst other techniques, static identification
using second-order deflections has proved to be an accurate and reliable method to evaluate prestress loss.
This paper represents an overview of some important research works on prestress loss evaluation in concrete
beams presented in the literature. The attention is focused mainly on static procedures, and a comparison with
dynamic procedures is elaborated. Comments and recommendations will be made at proper places, while con­
cluding remarks including future research directions will be presented at the end of the paper.

1 INTRODUCTION nondestructive methods are normally required.

Nevertheless, as far as the influence of prestress­
The first applications of prestressing techniques to ing on the dynamic properties of PC beams is
concrete structures go back to the first half of the concerned, the debate is still on going.
twentieth century. Today prestressing is largely This paper represents an overview of some
used for several applications, ranging from small important research works on prestressing loss evalu­
beams, such as railway sleepers, to more important ation in concrete beams. At first, experimental and
structures such as span bridges, long and light pre­ theoretical investigations are reviewed. The next
cast flooring and roofing elements for buildings. stage focuses on a static method, and a comparison
Both the serviceability and safety of Prestressed with dynamic procedures is elaborated. In fact,
Concrete (PC) structures rely on the effective state amongst other techniques, static identification by
of prestressing force (Breccolotti & Materazzi using second-order deflections has proved to be an
2015). Indeed, prestressing is mainly used to effective method to evaluate prestressing loss. Com­
reduce deflections and to partially counterbalance ments and recommendations are made at proper
the effect of dead and live loads in the case of places, whilst concluding remarks including future
bridges. Consequently, an extreme loss of pre­ research directions are illustrated at the end.
stressing may jeopardize the performance of PC
beams and indicate damage phenomena. For this
reason, tools and procedures capable of estimating 2 LITERATURE REVIEW
prestressing losses are highly desirable. Dynamic
methods were generally applied, also for prevent­ Researches were devoted to study the influence of
ing maintenance, repair, or replacement of prestress force on the dynamic behavior of beams.
a bridge. Therefore, the operating state of bridges An overview of the primary contributions on this
must be controlled through nondestructive argument is presented herein, including experimental
approaches. However, prestressing losses can be and theoretical investigations.
directly, simply, and accurately estimated if the
internal tendons of PC girders are instrumented by
2.1 Experimental investigations
using a load cell during construction (Saiidi et al.
1996, 1998). Although instrumentation of external Kerr (1976) executed several experiments on a steel
tendons is easy during the serviceability, cantilever beam with a supported barycentric tendon.

DOI: 10.1201/9780429279119-11
https://doi.org/10.1201/9780429279119-11

117
Based on the results, the researcher found that the declared that for lower values of prestressing force,
natural frequencies were not affected by the pre­ an increase of prestressing generates an increase in
stressing force. eigenfrequencies, especially for the fundamental fre­
Hop (1991) monitored the dynamic response of quency. Vice versa, for higher levels of prestressing,
a series of PC beams. Investigation focused on the the rate of increase of the eigenfrequencies tends to
influence of prestressing forces on frequency and decrease. Furthermore, changes in vibration frequen­
damping. The author discovered that applying an cies were observed to be higher for smaller eccentri­
increase of prestressing force, acting unevenly on the cities of the prestressing tendon. By considering the
beam, would increase the vibration frequency. In aforementioned investigations, it can generally be
many cases, it was measured that the application of claimed that prestressing force slightly affects the
further levels of prestressing increase, would result dynamic properties of concrete beams. Nevertheless,
in drop of vibration frequency. relevance of these modifications depends on many
Similar test results were gained by Saiidi et al. factors (cracking and nonlinearity of concrete, bond­
(1994) on a PC member with concentric tendon. The ing and eccentricity of tendon) that generate counter­
research proved an increase of the first eigenfre­ balancing effects making it difficult to identify
quency from 11.41 Hz for the case of null prestress­ a clear relation between dynamic properties and
ing, to 15.07 Hz for the maximum magnitude of magnitude of prestress force.
prestressing. Also, the authors observed that an
increase in prestressing seems to influence micro-
2.2 Theoretical investigations
crack closure and, consequently, increment the stiff­
ness and natural frequencies of PC beams. Several theoretical studies were conducted by
Miyamoto et al. (2000) tested the dynamic vibra­ addressing the effect of prestress force on the
tion of PC beams, strengthened with external ten­ dynamic characteristics of PC beams.
dons. According to their results, prestressing forces During the simulation of a moving force identifi­
applied to external tendons influence the frequency cation method that considered the effects of pre­
vibrations of girders. stressing, Chan & Yung (2000) discovered that the
Lu & Law (2006) tested a PC beam with natural frequencies of a PC bridge decrease with an
a straight concentric tendon. Two conditions were increase in prestressing force. This is notorious as
examined, i.e., with and without the prestress force the “compression–softening” effect, and generally
of 66.7 kN. The authors observed that the prestress­ occurs in Euler–Bernoulli beams and PC members
ing induced an increment in the first three eigenfre­ preserved against crack formation (Timoshenko &
quencies within a range of 0.4-2.1%. Gere 1961, Bazant & Cedolin 1991).
Xiong & Zhang (2009) tested 3 simply supported Kim et al. (2004) studied prestress loss predic­
PC girders with different configurations of external tions in PC girders based on measurement variations
tendons. The authors noted that the natural frequency in the natural frequencies. Comparison between the
initially increased with the increase of prestressing test results obtained by Saiidi et al. (1994) and the
force. Conversely, the natural frequency decreased previsions of their model validated their method.
after the cracks induced by the prestressing. Law & Lu (2005) analyzed the time-domain
Kim et al. (2010) tested a PC girder with many response of a PC beam under dynamic excitation. By
damage scenarios of prestressing loss. Starting from comparing the results of analytical simulations with
a state of absence of prestressing loss, the prestress the theoretical findings, the authors identified the
force was then gradually reduced to Zero. During prestressing force in the time domain by recording
this unloading procedure, vibration measurements displacements and strains. According to their find­
allowed to determine reductions of the first four ings, the vibration frequencies decrease as the pre­
eigenfrequencies up to values of 4.0-4.4% from the stressing increases.
initial stage to the final one. Hamed & Frostig (2006) developed a nonlinear
Jang et al. (2010) tested 6 PC beams with model for describing the behavior of PC beams with
a bonded tendon. By applying continuously an a bonded or an unbonded tendon. Based on the
increase of prestressing from 0 to 523 kN, the derived governing equations, the authors demon­
authors noted a progressive increase of the first strated that the prestress force does not affect the
eigenfrequency from 7.6 to 8.7 Hz. natural frequencies of PC members.
Noh et al. (2015) executed experiments on 3 PC Jaiswal (2008) pointed out that the increase of
beams with different configurations of tendons. The a PC beam’s stiffness (and frequency) depends on
researchers predicted that the natural frequency gener­ the eccentricity of the tendon, thus inducing greater
ally increases as tension force in the prestressing steel moment and stiffening effect along the member.
increases. Moreover, they observed that the frequen­ Limongelli et al. (2016) studied the detection of
cies of PC members are affected by other parameters, early warning signs of deterioration in a PC beam
such as beam camber, cable geometric stiffness, and due to prestressing loss. The investigators underlined
stiffness effect of the beam-tendon system. that the natural frequencies of PC members signifi­
The results of the most relevant works (Hop 1991, cantly vary only under the effects of crack initiation
Saiidi et al. 1994, Kim et al. 2010, Jang et al. 2010) or crack re-opening.
118
 Gan et al. (2019) validated the experiments 3 STATIC METHODS FOR PRESTRESS LOSS
made by Jang et al. (2010) and Noble et al. EVALUATIONS
(2016) by a finite element model, in which the
influence of prestressing on the natural frequen­ Amongst vibration-based techniques, static identifi­
cies was simulated by the existence of early-age cations by using second-order deflections have
shrinkage cracks inside the concrete. proved to be reliable methods when applied to beam
Bonopera et al. (2019b) proved that the funda­ elements (Tullini et al. 2012, Tullini 2013, Bonopera
mental frequency of uncracked PC girders with et al. 2018a). Subsequently, a series of experimental
a parabolic tendon is unaffected by the prestressing simulations were conducted on members belonging
force because the course of the “compression–soften­ to space frames and trusses (Turco 2013, Bonopera
ing” theory being cancelled out by the increase of et al. 2018b, c).
elastic modulus due to the concrete’s consolidation/ Bonopera et al. (2018d) instead verified the feasi­
hardening with time. bility of estimating prestress force in a PC beam spe­
Looking at the most relevant studies above men­ cimen by using vertical displacements. This
tioned (Hamed & Frostig 2006, Jaiswal 2008, approach only uses static parameters, thus, in con­
Limongelli et al. 2016), no significant agreement trast to dynamic procedures, does not require select­
between the effect of prestressing force and dynamic ing experimental data for use in the algorithms.
behavior of PC beams was observed. As a result, the
natural frequencies were declared unsuitable indica­ 3.1 The method proposed by Bonopera et al.
tors for prestress loss evaluations, as reported by (2018d)
Saiidi et al. (1994), Jaiswal (2008), Limongelli et al.
(2016) and Bonopera et al. (2019b). The method was before developed for axial load
detection in compressed steel members (Bonopera
et al. 2018a) and, consequently, employed for pres-
2.3 Dynamic methods for prestress loss tress force identification in PC beams (Bonopera
evaluations et al. 2018d). The reference model comprises
Kim et al. (2004), Law & Lu (2005), Lu & Law a simply supported Euler–Bernoulli beam of
(2006), Bu & Wang (2011), Xu & Sun (2011) and 250 mm in width, 400 mm in height, and length L of
Shi et al. (2014) presented vibration methods for 6.62 m, prestressed by a straight tendon, in which
prestress force identifications based on natural fre­ the prestressing force N is assumed as an external
quencies and dynamic responses of the PC beams compressive force eccentrically applied to the end
under investigation. Local vibration measurements constraints N e. The cross sectional second moment
along the members are required. of the area of the PC member’s section I is equal to
Law et al. (2008), Li et al. (2013) and Xiang 1.3333 × 109 mm4. Besides, a bending deflection v(1)
et al. (2016) performed numerical simulations by along the aforementioned beam’s model is properly
taking into account the vibration responses of approximated by multiplying the corresponding first-
PC girders to moving vehicular loads. In their order deflection by the “magnification factor” of
methods, prestress loss evaluations can be the second-order effects, i.e., based on the “compres­
obtained through vibration measures if the pres- sion–softening” theory (Timoshenko & Gere 1961,
tress force is assumed as an equivalent external Bazant & Cedolin 1991).
compressive force applied to the end constraints
(Tullini & Laudiero 2008, Tullini et al. 2012,
Maes et al. 2013, Rebecchi et al. 2013, Tullini
2013, Luong et al. 2017, Bonopera et al. 2018a,
b, c, d, e, Kernicky et al. 2018). As
a consequence, the natural frequencies of PC
girders tend to increase with a decrease in pres-
tress force according to the “compression–soften­
ing” theory.
Most of the dynamic techniques, cited therein, are
based on the modal characteristics of a beam. Once
the methods are applied, the modal characteristics
depend on the corresponding PC member’s stiffness
and, accordingly, become affected by the prestress­
ing force. In detail, such vibration-based identifica­ Figure 1. Reference model of the PC beam. (a) Deflection
tions require an accurate selection of the mode shape curve v(°) after the application of the eccentric prestress
to be utilized in the procedures. In fact, selecting the force N. (b) Deflection curve v(1) after the application of the
optimal PC beam’s frequency a priori is challenging, vertical load F to the deflection curve v(°). The dashed lines
and different frequencies yield varying degrees of represent the initial deflection curves (Bonopera et al.
accuracy in estimations. 2018d).

119
 Experiments on a PC beam specimen, having Table 2. Prestress force evaluations Na based on Eq. (8b)
the configuration above mentioned, were arranged (Bonopera et al. 2018d) for each test day.
in the laboratory of the National Center for
Research on Earthquake Engineering of Taipei, Test 2 - v4(1)
Taiwan (Bonopera et al. 2018d). First, the
deflected-shape measurements v(1) along the PC deflections at
beam’s length, obtained from 27 three-point the midspan
bending tests with different applied prestress
forces N, were examined to assess the accuracy Days of Eaver N F Na Δ
of the assumptions of the beam’s model. Second, concrete
curing (MPa) (kN) (kN) (kN) (%)
based on the “magnification factor” formula,
prestress force identifications were achieved by
426 34870 620 20.2 789 27.3
using two series of deflections v(1), i.e., those
620 22.6 857 38.2
measured at the quarter v2(1) and those measured
617 25.0 550 -10.9
at the midspan v4(1) of the PC member. Informa­
427 37618 724 20.1 732 1.1
tion regarding the flexural rigidity of such PC
721 22.6 761 5.5
member was also required. Specifically, an aver­
721 25.1 718 -0.4
age value of the static elastic modulus Eaver of
433 38791 820 20.2 823 0.4
the used concrete was obtained by compression
820 22.9 825 0.6
tests on a set of cylinders cast at the same time
820 25.1 870 6.1
of the PC beam.

3.2 Prestress loss evaluations through the method

proposed by Bonopera et al. (2018d)
Tables 1 and 2 list the prestress force identifications
Na obtained by using the deflections v2(1) and the for each day of execution of the experiments was
corresponding experimental values ψ = FL3/Eaver used as parameter in the identification process. The
I in Eq. (8a) (Test 1), as well as the deflections v4(1) corresponding first-order deflections were instead
and the corresponding parameters ψ in Eq. (8b) calculated by using Eqs. (4a) and (4b), similarly
(Test 2), respectively. Both equations are illustrated reported in Bonopera et al. (2018d). Tables 1 and 2
in Bonopera et al. (2018d). Specifically, the 9 test also show the related percentage errors Δ= (Na □
combinations represent the best prestress force Nψψ /N. In general, poor estimates Na were obtained
evaluations among the thrice test repetitions con­ when prestressing N equal to 617 and 620 kN were
ducted (Section 3.1). The static elastic modulus Eaver applied. Conversely, the combinations with pre­
stressing that induced second-order effects greater
than 6.5%, i.e. N ≧ 721 kN, provided excellent force
evaluations Na. In fact, in this last case, estimation
Table 1. Prestress force evaluations Na based on Eq. (8a) errors were lower than 6.1% (in absolute value).
(Bonopera et al. 2018d) for each test day.
Sensitivity analyses were performed for the
Test 1 - v2(1) estimations based on the aforementioned Eqs.
(8a) and (8b). The deflections v2(1) and v4(1) cal­
culated by Eqs. (3a) and (3b) (Bonopera et al.
deflections
at a quarter 2018d) and parameter ψ were modified to repro­
duce possible experimental errors. Specifically,
Days of Eaver N F Na Δ the values v2(1), v4(1), and ψ were alternatively
concrete multiplied by 0.99 and 1.01 to generate 14 com­
curing (MPa) (kN) (kN) (kN) (%) binations of simulated values for 9 different
assumed prestress forces N. The average value of
426 34870 620 20.2 777 25.3 the applied vertical loads Faver = 22.6 kN was
620 22.6 857 38.2 assumed in the manipulations. Figures 2 and 3
617 25.0 386 -37.4 depict a comparison between the worst estimated
427 37618 724 20.1 729 0.7 Na and assumed values N conducted by using
721 22.6 721 0.0 deflections v2(1) and v4(1), both of which yielded
721 25.1 715 -0.8 a constant error of approximately ±107 kN.
433 38791 820 20.2 846 3.2 Based on all results obtained, a favorable corres­
820 22.9 825 0.6 pondence between analytical N and experimental
820 25.1 824 0.5 force identifications Na was observed when mid­
span deflection v4(1) was assumed.

120
 assumption that PC girders are not generally affected
by crack formation, the procedure can instantaneously
estimate the existing prestressing force through the
measurement of second-order vertical deflections. The
estimation’s accuracy improves when the PC girder
under investigation is subjected to a high prestress
force and, also, when the midspan displacement is
assumed. Information regarding the flexural rigidity of
the PC girder are additionally necessary. Specifically,
an average value of the static elastic modulus must be
evaluated by compression tests on a set of concrete
cores drilled along its span. Besides, the static method
Figure 2. Prestress force evaluations based on Eq. (8a). does not require a direct measure of the tension force
Symbols + refer to the comparison between estimated Na in the tendon and, mostly, in contrast to dynamic pro­
and measured values N for all 27 test combinations. The cedures, does not require selecting experimental data
dashed lines with symbol × represent the sensitivity
for use in the algorithms. Future investigations are
analyses.
intended to be developed by using vehicle loading on
a concrete bridge, in which the constraint stiffness of
the spans must be evaluated for PC girders with
unknown boundary conditions. The fiber Bragg grating
differential settlement measurement sensors are an
effective tool for the measurements of the bridge verti­
cal deflections (Bonopera et al. 2019a).

ACKNOWLEDGMENTS

M.B. acknowledges the funding provided by the

Ministry of Science and Technology of Taiwan
(MOST 108-2811-E-492-001).

Figure 3. Prestress force evaluations based on Eq. (8b).

Symbols + refer to the comparison between estimated Na REFERENCES
and measured values N for all 27 test combinations. The
dashed lines with symbol × represent the sensitivity Bažant, Z.P. & Cedolin, L., 1991. Stability of Structures.
analyses. Oxford University Press, New York, USA.
Bonopera, M., Chang, K.C., Chen, C.C., Lee, Z.K. &
Tullini, N., 2018a. Axial load detection in compressed
steel beams using FBG–DSM sensors. Smart Structures
4 CONCLUDING REMARKS and Systems 21(1): 53–64.
Bonopera, M., Chang, K.C., Chen, C.C., Lin, T.K. &
In this paper, an overview of some important experi­ Tullini, N., 2018b. Compressive column load identifica­
mental and theoretical studies on prestress loss evalu­ tion in steel space frames using second–order deflec­
ation in concrete beams allowed to analyze several tion–based methods. International Journal of Structural
details on this topic. Some references state that pre­ Stability and Dynamics 18(7): 1–16, 1850092.
stressing force does not affect the vibration frequencies Bonopera, M., Chang, K.C., Chen, C.C., Lin, T.K. &
Tullini, N., 2018c. Bending tests for the structural safety
of PC beams. As a result, the natural frequencies were
assessment of space truss members. International Jour­
declared unsuitable indicators for prestress loss estima­ nal of Space Structures 33(3-4): 138–149.
tions. Conversely, most of the experimental works Bonopera, M., Chang, K.C., Chen, C.C., Sung, Y.C. &
found in the literature show a slight increase in the Tullini, N., 2018d. Feasibility study of prestress force
eigenfrequencies under the increase of prestressing prediction for concrete beams using second–order
force. This behavior is related to the concrete mechan­ deflections. International Journal of Structural Stability
ics and, particularly, is a main consequence of the and Dynamics 18(10): 1–19, 1850124.
effect of crack and microcrack closure along the PC Bonopera, M., Chang, K.C., Chen, C.C., Sung, Y.C. &
members. However, vibration-based identification Tullini, N., 2018e. Prestress force effect on fundamental
frequency and deflection shape of PCI beams. Structural
methods require an accurate selection of the mode
Engineering and Mechanics 67(3): 255–265.
shape because different frequencies yield varying Bonopera, M., Chang, K.C., Chen, C.C., Lee, Z.K.,
degrees of accuracy in prestress loss evaluations. Sung, Y.C. & Tullini, N., 2019a. Fiber Bragg grating–
Taking these aspects into account, the paper focuses differential settlement measurement system for bridge
then on the feasibility study of the static method pro­ displacement monitoring: Case study. Journal of Bridge
posed by Bonopera et al. (2018d). Based on the Engineering 24(10): 1–12, 05019011.

121
Bonopera, M., Chang, K.C., Chen, C.C., Sung, Y.C. & Luong, H.T.M., Zabel, V., Lorenz, W. & Rohrmann, R.G.,
Tullini, N., 2019b. Experimental study on the funda­ 2017. Non-destructive assessment of the axial stress
mental frequency of prestressed concrete bridge beams state in iron and steel truss structures by dynamic
with parabolic unbonded tendons. Journal of Sound and measurements. Procedia Engineering 199: 3380–3385.
Vibration 455: 150–160. Maes, K., Peeters, J., Reynders, E., Lombaert, G. & De
Breccolotti, M. & Materazzi, A., 2015. Prestress losses and Roeck, G., 2013. Identification of axial forces in beam
camber growth in wing-shaped structural members. PCI members by local vibration measurements. Journal of
Journal 60(1): 98–117. Sound and Vibration 332(21): 5417–5432.
Bu, J.Q. & Wang, H.Y., 2011. Effective pre–stress identifi­ Miyamoto, A., Tei, K., Nakamura, H. & Bull, J., 2000.
cation for a simply supported PRC beam bridge by BP Behavior of prestressed beam strengthened with external
neural network method. Journal of Vibration and Shock tendons. Journal of Structural Engineering 126(9):
30(12): 155–159. 1033–1044.
Chan, T. & Yung, T., 2000. A theoretical study of force Noble, D., Nogal, M., O’Connor, A. & Pakrashi, V., 2016.
identification using prestressed concrete bridges. Engin­ The effect of prestress force magnitude and eccentricity
eering Structures 22(11): 1529–1537. on the natural bending frequencies of uncracked pre­
Gan, B.Z., Chiew, S.P., Lu, Y. & Fung, T.C., 2019. The stressed concrete beams. Journal of Sound and Vibra­
effect of prestressing force on natural frequencies of tion 365: 22–44.
concrete beams - A numerical validation of existing Noh, M.H., Seong, T.R., Lee, J. & Park, K.S., 2015.
experiments by modelling shrinkage crack closure. Experimental investigation of dynamic behavior of pre­
Journal of Sound and Vibration 455: 20–31. stressed girders with internal tendons. International
Hamed, E. & Frostig, Y., 2006. Natural frequencies of Journal of Steel Structures 15(2): 401–414.
bonded and unbonded prestressed beams-prestress force Rebecchi, G., Tullini, N. & Laudiero, F., 2013. Estimate of
effects. Journal of Sound and Vibration 295(1-2): 28–39. the axial force in slender beams with unknown boundary
Hop, T., 1991. The effect of degree of prestressing and age conditions using one flexural mode shape. Journal of
of concrete beams on frequency and damping of their Sound and Vibration 332(18): 4122–4135.
free vibration. Materials and Structures 24(3): 210–220. Saiidi, M., Douglas, B. & Feng, S., 1994. Prestress force
Jaiswal, O.R., 2008. Effect of prestressing on the first flex­ effect on vibration frequency of concrete bridges. Jour­
ural natural frequency of beams. Structural Engineering nal of Structural Engineering 120(7): 2233–2241.
and Mechanics 28(5): 515–524. Saiidi, M., Shields, J., O’Connor, D. & Hutchens, E., 1996.
Jang, J., Lee, H., Hwang, K. & Song, Y., 2010. Variation of prestress forces in a prestressed concrete
A sensitivity analysis of the key parameters for the pre­ bridge during the first 30 months. PCI Journal 41(5):
diction of the prestress force on bonded tendons. 66–72.
Nuclear Engineering and Technology 42(3): 319–328. Saiidi, M., Hutchens, E. & Gardella, D., 1998. Bridge Pres-
Kernicky, T., Whelan, M. & Al-Shaer, E., 2018. Dynamic tress Losses in Dry Climate. Journal of Bridge Engin­
identification of axial force and boundary restraints in eering 3(3): 111–116.
tie rods and cables with uncertainty quantification using Shi, L., He, H. & Yan, W., 2014. Pre–stress force identifica­
Set Inversion Via Interval Analysis. Journal of Sound tion for externally pre–stressed concrete beam based on
and Vibration 423: 401–420. frequency equation and measured frequencies. Mathem­
Kerr, A.D., 1976. On the dynamic response of a prestressed atical Problems in Engineering, 840937.
beam. Journal of Sound and Vibration 49(4): 569–573. Timoshenko, S.P. & Gere, J.M., 1961. Theory of Elastic
Kim, J.T., Yun, C.B., Ryu, Y.S. & Cho, H.M., 2004. Identi­ Stability, McGraw-Hill, New York, USA.
fication of prestress-loss in PSC beams using modal Tullini, N. & Laudiero, F., 2008. Dynamic identification of
information. Structural Engineering and Mechanics 17 beam axial loads using one flexural mode shape. Journal
(3-4): 467–482. of Sound and Vibration 318(1-2): 131–147.
Kim, J.T., Park, J.H., Hong, D.S. & Park, W.S., 2010. Tullini, N., Rebecchi, G. & Laudiero, F., 2012. Bending
Hybrid health monitoring of prestressed concrete girder tests to estimate the axial force in tie–rods. Mechanics
bridges by sequential vibration-impedance approaches. Research Communications 44: 57–64.
Engineering Structures 32(1): 115–128. Tullini, N., 2013. Bending tests to estimate the axial force
Law, S. & Lu, Z., 2005. Time domain responses of in slender beams with unknown boundary conditions.
a prestressed beam and prestress identification. Journal Mechanics Research Communications 53: 15–23.
of Sound and Vibration 288(4-5): 1011–1025. Turco, E., 2013. Identification of axial forces on statically
Law, S.S., Wu, S.Q. & Shi, Z.Y., 2008. Moving load and indeterminate pin-jointed trusses by a nondestructive
pre–stress identification using wavelet based method. mechanical test. The Open Civil Engineering Journal 7:
Journal of Applied Mechanics Transactions ASME 75 50–57.
(2),021014. Xiang, Z., Chan, T.H., Thambiratnam, D.P. & Nguyen, T.,
Li, H., Lv, Z. & Liu, J., 2013. Assessment of pre–stress 2016. Synergic identification of prestress force and
force in bridges using structural dynamic responses moving load on prestressed concrete beam based on vir­
under moving vehicles. Mathematical Problems in tual distortion method. Smart Structures and Systems 17
Engineering, 435939. (6): 917–33.
Limongelli, M.P., Siegert, D., Merliot, E., Waeytens, J., Xiong, H.X. & Zhang, Y.T., 2009. Theoretical analysis of
Bourquin, F., Vidal, R., Le Corvec, V., Gueguen, I. & natural frequency of externally prestressed concrete
Cottineau, L.M., 2016. Damage detection in a post ten­ beam based on rigidity correction. Academic Journal of
sioned concrete beam – Experimental investigation. Xi’an Jiaotong University 21(1): 31–35.
Engineering Structures 128: 15–25. Xu, J. & Sun, Z., 2011. Pre–stress force identification for
Lu, Z. & Law, S., 2006. Identification of prestress force eccentrically pre–stressed concrete beam from beam
from measured structural responses. Mechanical Sys­ vibration response. Tech Science Press SL 5(2):
tems and Signal Processing 20(8): 2186–2199. 107–115.

122
 Bridge Maintenance, Safety, Management, Life-Cycle Sustainability and Innovations –
Yokota & Frangopol (eds)
© 2021 Taylor & Francis Group, London, ISBN 978-0-367-23278-8

Crack identification and measurement of bridges by using CNN models

H. Qin, F. Huang & B. Cheng
Shanghai Jiao Tong University, Shanghai, China

ABSTRACT: In recent years, with the development of artificial intelligence (AI) technology, automatic
equipment such as robots and drones has been introduced into the damage detection of bridges, where
advanced image processing technology is commonly required in order to recognize the damages efficiently.
This paper utilizes convolutional neural network (CNN) technology to identify the surface cracks in bridge
structures. A database consisting of 2120 images with or without crack was first established by photographing
practical bridge structures, in which 1696 images were used as training sets and 424 images were used as test
sets. A convolutional neural network was then constructed by employing four convolutional layers, four pool­
ing layers, and three fully connected layers. The CNN model was trained with backpropagation stochastic
gradient descent method on the basis of training data, and the accuracy of prediction was verified by the test
sets. In addition, the dimensions including length and width of the crack were also numerically calculated by
setting reference points on images. Finally, by further using image mosaic technologies, the panorama picture
of surface cracks was automatically generated from interval and continuous images. Results show that the
accuracy of presented crack recognition using CNN technology is obviously higher than traditional image
processing technology, and the efficiencies of bridge maintenance could be significantly improved.

Keywords: bridge, crack identification, image processing, convolutional neural network, image mosaic
technology

1 INTRODUCTION 2 APPLICATION OF CNN IN CRACK

DETECTION
During the use of engineering structures, cracks are
easily caused by loads, temperature differences, In recent years, with the rapid development of CNN
chemical corrosion and other environmental factors. technology and its excellent performance for image
The generation of cracks will reduce the bearing cap­ recognition, CNN has begun to be applied for the crack
acity of the structure, also accelerate the aging of the identification and detection in the field of civil
structure and give rise to security risks. Therefore, engineering.
the efficient technology of crack detection, which For road crack detection. Wang and Hu[1]introduced
would make the targeted maintenance and restor­ principal component analysis in the CNN model,
ation work more easy and accurate, is very important which can accurately classify longitudinal cracks, trans­
to ensure the normal use of the structure. Recently, verse cracks, and mesh cracks. Liu Honggong[2]used
the crack detection is mainly depended on the LeNet-5 model for classifying transverse cracks, longi­
manual work, whose accuracy is significantly tudinal cracks, and massive shapes. It shows that the
affected by the experience of the inspectors. In add­ CNN-based classification method performs better than
ition, the manual work is relatively time-consuming, the traditional manual one in crack detection.
efficient and sometime dangerous for inspectors. For the crack detection in house construction, Moon
As a kind of unsupervised machine learning and Kim[3]used convolutional neural networks to detect
method, CNN technology has fast recognition speed and identify cracks on the surface of buildings. The rec­
and high accuracy. It has been proved to be very ognition rate of crack images is 90%, and the recogni­
suitable for mass image recognition and multi-type tion rate of non-crack images is 92%, which meets the
damage recognition, which has the advantage that demand of the general detection. Cha[4] used the Faster
traditional image recognition technology can’t R-CNN method to simultaneously detect structural
match. concrete cracks, bolt corrosion and steel delamination,

DOI: 10.1201/9780429279119-12
https://doi.org/10.1201/9780429279119-12

123
which extended the scope of application of CNN in 4 IMAGE DATABASE
structural defect detection.
In the detection of metal cracks, since the crack on In this experiment, the quality of the collected images
the metal surface is slender and the contrast with the has a significant impact on the training effect of the
background color is not obvious, the detection effect of CNN model. In order to make the CNN mode widely
the conventional method is not ideal. Faghih-Roohi[5] applicable for crack identification, the collected images
used DCNN to detect orbital surface defects, and the should satisfy the variety of structure types, back­
recognition accuracy reached 92%. Chen and ground colors and lighting conditions. Therefore, it is
Jahanshahi[6] combined CNN with Bayesian method to recommended that the surface of structure for image
detect surface cracks in nuclear power facilities, and capturing should be painted in white, yellow, or blue,
the detection effect is better than traditional detection and the lighting condition of uniform illumination and
method. softness should be chosen. The picture selection is
The above engineering application not only shown in Figure 2. The pictures in this experiment are
proves that CNN has extensive practicability in the selected from self-photographing and online collection,
detection of engineering structures, but also shows of which the number is 128.
that CNN has strong scalability and can be combined It is also necessary to segment the collected ori­
with other traditional algorithms to improve its com­ ginal image before crating the image database. On the
puting speed and recognition accuracy. one hand, the larger the image database, the more
fully trained the CNN model. Dividing the original
image can greatly increase the number of pictures in
3 CRACK IDENTIFICATION PROCESS the database, which is beneficial to the later CNN
model training. On the other hand, the original images
The use of CNN to identify structural cracks mainly differ in size, which do not meet the input require­
includes the establishment of a database and classifi­ ments of the CNN model. Thus, the original images
cation construction. The establishment of the data­ need to be tailored to a uniform size. Herein, the ori­
base includes image collection, image processing ginal image is uniformly cut into 2120 small images
and database formation. In order to guarantee the with size of 128*128 pixels, as shown in Figure 3.
accuracy of crack identification with CNN under dif­ In the model training phase, the degree of model
ferent environment, it is necessary to collect crack training needs to be judged based on the cross-
images with different kinds and background colors entropy loss. In the traditional image recognition
in the image collection stage. After then, the original experiment, the pictures need to be marked before
image is processed by clipping and scaling, so that training. While for CCN-based method, this ineffi­
the database including training date, test data and cient marking process is skipped, the training pic­
verification data is built according to the label. The tures only need to be classified into the two types,
constructed CNN-based classification is optimized i.e., 960 crack images and 1160 crack-free images,
through the progress of training and test. and then were respectively stored into the corres­
After completing the crack recognition experiment, ponding folders. Before using the dataset to train the
that is, model construction and verification, the original model, the dataset is divided into three parts: training
image is processed by the image mosaic technique to set, verification set, and test set. Firstly, the CNN
generate the panorama, and the crack feature from the model is optimized through the training set, and then
image is extracted using OpenCV. Figure 1 shows the the performance of the model on the verification set
typical process of crack identification based on CNN. is improved by iterative according to the result of

Figure 1. The process of crack identifying based on CNN.

124
Figure 2. The selected images.

Figure 3. Processing of images.

cross-validation. Finally, the test set is used to evalu­ and classifies the images according to the feature
ate the performance of the model. In addition, the information. The feature information is extracted
proportion of the training set, the verification set, from an image through convolution and pooling oper­
and the test set will also affect the detection effect of ations. For the crack detection, the CNN model clas­
the CNN model. Generally, for a database with sifies the images in the database according to the
a small data size, ratio of training set, verification characteristics of the cracks with the standard of
set, and test set can be 6:2:2. “whether the cracks are contained in the image”,
thereby realizing the detection of cracks in the image.

5 EXPERIMENT OF CNN MODELS

5.1 Model construction
The CNN-based pictures recognition is essentially In the experiment of CNN identifying structural
the process that extracts the features from the pictures cracks, the construction of CNN model is the most

125
important process, and the selection of model struc­ backpropagation in reverse error sensitive terms.
ture directly affects the calculation speed and accur­ The back-propagation process of CNN model can be
acy. CNN model includes input layer, convolution summarized into the following steps:
layer, pooling layer and full connection layer. The (1) Initialize the parameters of the CNN model
input layer is located at the front end of the CNN and calculate δi,L of the output layer through forward
model. Through image decoding, the computer sum­ propagation;
marizes the pixel information such as the brightness (2) Calculate the output δi,l of the front layer
and gray of the image into a matrix that can be through δi,L:
called by the CNN model, which is called by the A) If the front layer is full connection layer, then:
CNN input layer. The convolution layer is composed
of multiple convolution kernels. By moving the con­
volution kernels on the image matrix, the feature
matrix of the original image can be obtained and the
image features can be extracted. Each convolution
kernel in the convolution layer represents a feature
selection. Usually, multiple convolution kernels are
superimposed to better reflect the feature information C) If the front layer is a pooling layer, then:
of the original image. The pooling layer operates on
an area of the input matrix and retains only one
value in that area. The full connection layer is
located at the end of the CNN model. The local fea­
tures extracted from each layer are assembled into (3) Update the weight parameter W and bias param­
a complete image through weight matrix. Generally, eter b of each layer:
two or more full connection layers are set to solve A) If the current layer is full connection layer,
nonlinear problems. The CNN model constructed in then:
this experiment is composed of four convolution
layers, four pooling layers and three fully connected
layers. The size of convolution kernel on the four
convolution layers is 5*5, the sampling method of
pooling layer adopts maximum sampling, and the
final output of the full connection layer is 2 types.
The structure of CNN model is shown in Figure 4.

5.2 Model training

The training process of CNN model is essentially the B) If the current layer is the convolution layer, then:
iterative optimization process of weight parameter
W and bias parameter b through the back propagation
algorithm. The traditional neural network adopts the
form of full connection. In the process of back
propagation, the error sensitive term of each layer
can be obtained by using chain partial derivative,
and then the weight and the gradient of the bias term
can be calculated, in turns the weight can be
updated. However, in the CNN model, when the fea­
ture graph after convolution operation is reversely
propagated to the previous layer, the parameters of (4) When the variation value of all weight param­
convolution kernel need to be updated. Since the eters W and bias parameter b is less than the constant
output of the pooling layer is the maximum value of of the stop iteration threshold ε, the iteration is
a sliding window and its partial derivative is 1, the finished;
pooling layer is different from the traditional (5) Output the trained CNN model.

Figure 4. Structure of CNN model.

126
5.3 Model test and analysis 6 CASE STUDY
The experimental environment was TensorFlow2.7,
In practical applications, the images were obtained
Intel Core CPU, 1.8ghz main frequency and 8GB
from an independently designed bridge inspection
memory. The original images in the experiment were
system, and the CNN model is used to identify the
taken and collected independently, and a total of
area where the crack exists. Due to the limitation of
2,120 images were processed, including 960 fracture
shooting distance and image definition, only the
images and 1,160 non-fracture images. According to
images in size of up to 15×10 cm can be acquired in
the ratio of 6:2:2, they were divided into training
a single time. When the image is captured, the dis­
sets, test sets and verification sets. The initial learn­
tance between the camera and the surface to be
ing rate of the CNN model was set as 0.0001. 16
detected remains fixed, which indicates the size of the
samples were used to update the weights once, and
acquired single image area is constant. However, in
100 iterations were carried out in total. The experi­
order to ensure that the detection information is not
mental results are shown in Figure 5. The accuracy
missed, there will be overlapped at the edges of adja­
of the CNN model on the test set and the verification
cent images. As shown in Figure 6, the overlapping
set is 98%, which can meet the accuracy require­
portion is clipped based on the position information
ments of crack identification.

Figure 5. Results of the experiment.

Figure 6. Cutting and stitching of original pictures.

127
 types under different background colors and differ­
ent lighting environments. The proposed approach
can effectively enhance the identification efficiency
and reduce the risk of miscarriages in current health
inspection systems, and can make the health inspec­
tion of structures more intelligent and automatic. In
the future, more efforts should be devoted to the
identification of other types of structural diseases
Figure 7. Measurement result of panoramic image. (e.g., corrosion, settlement, slip, and bolt-loosing) so
that the application scope of this approach could be
extended.
of the image, and then the image is stitched . In the
image mosaic process, the SIFT algorithm is also
used, which not only has the invariance of scale, rota­ REFERENCES
tion, affine, angle of view and illumination, but also 1 Wang X, Hu Z. Grid-based pavement crack analysis
maintains a good matching degree for the effects of using deep learning [C]//2007 4th International Confer­
target motion, occlusion, noise, etc. The image stitch­ ence on Transportation Information and Safety (ICTIS),
ing method of location information provides compara­ Banff, AB, 2017: 917–924.
tive verification. 2 Liu Honggong, Wang Xuejun, Li Bingying, et al.Detec­
After generating the panoramic image, the pano­ tion and recognition of bridge crack based in convolu­
ramic image is processed using OpenCV. Firstly, the tional neutral network[J]. Journal of Hwbei University
gray level and contrast of the image are adjusted, of Science and Technology, 2016, 37(05): 485–490.
3 Moon H, Kim J. Intelligent crack detecting algorithm on
and then the Canny edge detection method is used to the concrete crack image using neural network[C]//Pro­
extract the connected domain of the crack. After ceedings of the 28th ISARC, Seoul: 2011: 695–701.
finding all connected domains, extract the connected 4 Cha Y, Choi W, Suh G, et al. Autonomous structural
domain skeleton and measure the length and width visual inspection using region-based deep learning
of the crack. The measurement result is shown in for detecting multiple damage types[J]. Computer-
Figure 7. Aided Civil and Infrastructure Engineering. 2018, 33
(9): 731–747.
5 Faghih-Roohi S, Hajizadeh S, Núñez A, et al. Deep con­
7 CONCLUSIONS volutional neural networks for detection of rail surface
defects[C]//2016 International Joint Conference on
Neural Networks (IJCNN), Vancouver, BC: IEEE, 2016:
This paper presents a method of structural crack 2584–2589.
detection based on conventional neural network. By 6 Chen F, Jahanshahi M R. Deep learning-based crack
using the software of Tensor Flow, the CNN models detection using convolutional neural network and Naïve
were constructed and trained to successfully identify Bayes Data Fusion[J]. IEEE Transactions on Industrial
the cracks from the images of various structures Electronics. 2018, 65(5): 4392–4400.

128
 Bridge Maintenance, Safety, Management, Life-Cycle Sustainability and Innovations –
Yokota & Frangopol (eds)
© 2021 Taylor & Francis Group, London, ISBN 978-0-367-23278-8

Non-destructive detection of damages in concrete with thermal imaging

L. Mold, M. Auer, A. Strauss, M. Hoffmann & B. Täubling
Institute of Structural Engineering, University of Life Sciences Vienna, Austria

ABSTRACT: Surface recordings of concrete bridges with laserscan or digital imaging with the aid of
drones are used to detect superficial damages on bridges.This method is more time-saving than the conven­
tional, more complex procedures for visual inspections. In order to be able to also detect damages below the
surface by means of areal recordings, a promising approach - thermography - is investigated. A big advantage
of thermography is that the cameras can be applied on drones. Together with photogrammetric reconstruction
a rapid recording of the entire bridge can be carried out. This method makes it possible to detect damages
such as delamination, voids and moist areas in the concrete to a certain degree. Now the aim is to determine
the suitability of this method for the detection of the mentioned damages with regard to the detection accuracy
and limits. This paper is a summary and translation of an article submitted in a German Journal.

1 INTRODUCTION AND MOTIVATION examine the possible applications and limits of

damage detection on concrete structures using
Concrete structures, in particular, bridge structures, thermography, see also (Mold et al., 2019).
are subject to a great variety of stresses which can
impair their load capacity, serviceability and durabil­
ity. These influences, such as environmental impacts, 2 STATE OF THE ART
traffic loads, etc. usually lead to damage or defects in
the structure depending on the material properties. In In the field of condition determination of concrete
order to ensure safety and reliability over the planned structures by means of thermography, scientific stud­
lifetime of an engineering structure, regular inspec­ ies are already available. The investigations were all
tions and investigations are carried out (Strauss et al., done in the field of delaminations on bridge decks.
2016). The necessary tests to determine the condition By surveying the bridge carriageway with the aid
of the structure are usually associated with a costly of a drone to which a thermal camera was attached,
effort in terms of the applied investigation methods is was possible to detect layer delamination in the
and the required staff (Strauss et al., 2011). road surface in the thermograms (Omar and Nehdi,
Non-destructive methods offer the opportunity to 2017). The use of classification by means of
carry out effective and detailed examinations of k-means clustering enabled a condition mapping of
bridge structures. Thermography is a non-destructive the bridge carriageway in relation to delaminated
method that can be used for the detection of damages areas. The results of the evaluation could be valid­
on and under the concrete surface of structures. ated with the help of conventional test methods.
This paper is a summary and translation of an art­ With the help of concrete specimens with artifi­
icle submitted in a German Journal about thermal cially prepared delamination layers (Hiasa et al.,
imaging for non-destructive testing. 2017a), it was shown that the temperature gradient
Due to the possibility of using thermal cameras between intact and delaminated regions is greater,
with the help of drones, a comprehensive and effi­ the greater the underlying delamination is. Further­
cient investigation of the entire bridge surface can be more, it could be shown that by combining the
carried out. When assessing damages on structures, thermographic evaluation with a FE-modeling of the
it is important to know which type of damage can be specimen, the depth of the delamination layers can
unequivocally determined in terms of extent and be concluded. This FE-assisted approach has also
severity. (Hoffmann et al., 2018, 2019) been confirmed in further studies (Hiasa et al.,
In the course of this paper, the steps for determin­ 2017b).
ing the reliability of the thermography method are Investigations by (Hiasa et al., 2018) on the deter­
shown. The aim of the research project was to mination of the optimum time window for the

DOI: 10.1201/9780429279119-13
https://doi.org/10.1201/9780429279119-13

129
detection of delamination in concrete structures as the optimal threshold for the evaluation to show
under natural daily sunlight showed the morning as the outlines of the delaminated areas on the surface.
well as the period from early evening to midnight as
the preferred time window with the highest thermal
contrast between intact and delaminated sites. 3 PHYSICAL BASIS OF THERMOGRAPHY
According to the authors, the time points of the so-
called “crossover” have to be avoided, since no tem­ Everybody with a temperature above the absolute
perature differences can be detected here. Investiga­ zero emits thermally excited electromagnetic radi­
tions after sunset are recommended due to the lower ation. As an imaging method, Thermography allows
disturbance. (Watase et al., 2015) recommends meas­ the imaging of this heat radiation (Bernhard, 2014).
urements taken at about noon on sun-facing surfaces The measurement of temperatures can be carried
and midnight inspections for sun-exposed areas. out by different measuring principles. In the case of
Also (Janků et al., 2017) shows similar findings. thermal contact measurements (e.g., thermometers)
Using concrete test specimens with built-in thermodynamic equilibrium is established between
defects of polystyrene in different thicknesses and the object to be measured and the thermometer due
depths showed that even in the sun remote areas to heat exchange. As soon as an equilibrium is
where the heat flow is only excited by the daily tem­ reached, the kinetic temperature of the object can be
perature of the air, delamination can be detected up read (Bernhard, 2014). This kinetic temperature
to a depth of 5cm (Rocha et al., 2019). reflects the average internal kinetic energy of the
Tests performing thermography on bridge road­ object’s molecules (Lillesand et al., 2015). In add­
ways by means of vehicle-mounted IR cameras ition to this internal kinetic energy, objects emit
showed that at a driving speed of about 50 km/h energy as a function of their temperature. By means
defects can still be detected, whereby the quality of of this correlation, it is possible to deduce the energy
the data is significantly influenced by the integration level of an object in a non-contact manner via the
time and resolution of the used IR camera (Hiasa measurement of the emitted radiation energy, and
et al., 2016). thus subsequently to the kinetic body temperature
In a review (Milovanović and Pečur, 2016) (Lillesand et al., 2015).
summarized the state of knowledge on thermog­ Stefan and Boltzmann realized around 1880 that
raphy of reinforced concrete. Accordingly, it is the total radiation of a body is proportional to the
important for infrastructure operators not only to fourth power of its absolute temperature (doubling
know if there are any defects in bridge lanes but of the temperature = 16 times the radiation power).
also in which location or layer they are located. It If the total radiation of a body is detected by means
also refers to the difficult boundary conditions of of suitable sensors, via the Stefan-Boltzmann law,
thermography, as the temperature differences the radiation temperature can be deduced from ideal­
between intact sites and defects often show only ized radiation sources in which the entire incident
0.5 K and the depth of the defect affects the radiation energy is absorbed and re-emitted (black
timing and extent of thermal contrast (in the daily radiators). This relationship is used in infrared
cycle of natural solar radiation deeper delamin­ thermography (Lillesand et al., 2015).
ations become visible later than near-surface If one only wants to measure temperature differ­
delaminations). The authors recommend daily ences between same materials in the form of
temperature differences of at least 8 ° C on the a qualitative measurement, it is sufficient to record
examination days to ensure sufficient heat flow and map the radiant energy. However, in order to be
stimulation in the concrete structures. able to make quantitative statements, the kinetic
Regarding the limits of application of the detec­ body temperature must be deduced from the radi­
tion of delaminations on reinforced concrete bridges, ation energy - this requires further parameters and
(Rocha and Póvoas, 2017) summarized in a review the use of a calibrated radiometric thermal camera.
the recognition achievements of previous investiga­ In contrast to idealized black bodies, surfaces of
tions with respect to delamination size, depth and real bodies absorb only a fraction of the incident
time of inspection. This shows a very heterogeneous radiation - some of the incident energy is reflected or
result of the relationships between time and apparent transmitted through the body (see Figure 1). The
depth, were the maximum recognizable depths were emissivity of a body describes its ability to emit and
around 7cm. Subsequent investigations such as indicates how much radiation the real body emits in
(Auer, 2019) were able to detect layer detachments comparison to an ideal black emitter (Bernhard,
up to 10 cm under optimal boundary conditions. 2014). The emissivity depends mainly on the mater­
(Escobar-Wolf et al., 2018) showed that the com­ ial, surface condition and radiation direction and it
bination of IR inspection with simultaneous RGB changes over the spectral range.
exposures improves the detection rate of defects by Because real objects can achieve emissivities close
reducing the false positives. to 1 (= 100%), but not exactly 1, the total radiation
Investigations by (Sultan and Washer, 2017) on energy of a body does not only result from the
the optimal thermal contrast for the detection of internal energy of the body but from the totality of all
delamination give 0.6 - 0.8 K temperature difference absorbed, transmitted and reflected radiation sources.
130
 Thermography makes it possible to record and
display the heat radiation from surfaces. Different
temperature fields on the surface may indicate
damage to the structure. It is possible to detect
wet spots, cracks or spalling on the structure, but
also damages below the surface, such as delamin­
ation in the concrete. Delamination is the peeling
off of layers in material composites. In bridge
lanes, e.g. a detachment of the road surface from
the substructure leads to voids, in reinforced con­
crete delamination can be caused by the corrosion
Figure 1. Absorption, reflection and transmission of the
of the reinforcing bars. Subsequent, a visible
radiation acting on a body and emission of the inherent release of the concrete cover and unprotected
radiation, extracted from Mold et al., 2019. exposure of the reinforcement and, as a result,
corrosion can occur. Thus, delamination, when
detected, provides an indicator of the detection of
potentially corroded reinforcement. The resulting
Assuming that the objects to be examined are ther­ air space between the detached layers creates
mally opaque (transmission = 0) and are relying on the a kind of insulation that restricts heat flow. Due
Kirchhoff radiation law (degree of absorption = emis­ to the obstruction of the heat flow, the external
sivity), the total radiation results from the emitted radi­ heated surface, as shown in Figure 2, heats up
ation of the object and a reflected portion from other faster (heat accumulation) in the area of damage
surrounding radiation sources. When measuring the and cools faster in the case of heat flow from
radiation temperature of a body with an emissivity, for inside to outside than the surface of the intact
example, 0.8, the measured value consists of 80% of area. The temperature differences on the surface
the sought “true” kinetic body temperature and 20% of due to the disturbed heat flow are best detected
reflected radiation from the environment. By knowing at midday, when the building is heated the most
the emissivity and the mean temperature of the and at night when the building cools down again
reflected environment, the desired kinetic body tem­ (see Figure 2 a). This effect is used for bridge
perature can be calculated. deck damage detection (NEXCO-West USA, Inc.,
The principle of the thermography can be summar­ 2014), (Hiasa, 2016) and has been studied in the
ized as follows: Every opaque body with a temperature series of experiments as described in Chapter 5.
above absolute zero emits thermal radiation as If environmental conditions fit, delamination can
a function of its emissivity. This radiation can be also be detected on other components of the
recorded by a sensor without contact. The temperature bridge (Janků et al., 2017).
can be deduced from the intensity of the radiation with
knowledge of further parameters.
This very simplified representation of the conver­
sion becomes more difficult in practical application
by further influencing factors and possible measure­
ment errors, which must be considered or corrected
for a correct measurement.
The aim of the research project regarding the
damage detection with thermography was therefore
to determine the necessary environmental and basic
conditions for the most accurate and reliable meas­
urements possible.

4 DETECTION OF DAMAGES BY MEANS OF

THERMOGRAPHY

As described in (Mold et al., 2019), the prerequisite

for detecting subsurface thermal phenomena is
a temperature gradient in the object that results in
Figure 2. Principle for the detection of delamination with
heat transfer with any detectable anomalies. The heat thermography in reinforced concrete with a) temperature of
transfer can be generated from the outside to the the concrete surface during the course of the day depending
inside, usually active (IR emitters) or natural (diurnal on the intact and delaminated area, b) heat flow during
movement of the sun/air temperature), or from recording period A (noon), c) heat flow during recording
inside to outside, usually generated passively. period B (night), extracted from Mold et al., 2019.

131
5 THERMOGRAPHY TESTS

5.1 Motivation
The regular condition assessment of infrastructure con­
struction plays a crucial role in ensuring their safety,
reliability and durability. The focus is on the assess­
ment of damage and the possible consequences in
terms of the safety and reliability of the structure. Non­
destructive methods allow a detailed and non-invasive
detection of damages of the structure. When using
non-destructive methods, it is necessary to be able to
assess the reliability of the method in order to ensure
that the necessary safety level can be maintained.
Thermal detection depends on a number of factors
that can affect the detection rate. The probability of
detection depends on environmental factors such as
wind, solar radiation and moisture as well as on the size Figure 3. Illustration a specimen for the investigation of
and depth of the damage in the structure. The investiga­ voids in concrete by means of thermography with the asso­
tion of structures by means of thermography requires ciated thermographic images, extracted from Mold et al.,
a certain know-how with regard to the optimal use due 2019. a) Externally heated specimen with air-flowed voids
to environmental factors and the damage to be detected. b) Externally heated specimen with water-cooled voids.
In order to be able to identify the necessary conditions
for the most reliable detection of damage, the detectabil­
ity of voids, delamination and moisture in concrete known. This allowed a comparison of the results of
under different environmental conditions was examined the thermographic images with the actual location
in the research project. The limits of damage detection and extent of the damage.
in terms of their size and depth in the structure were As shown in Figure 3, the damaged area near the
determined in the course of laboratory tests and simula­ surface cools fastest. The additional cooling of the
tions. The experiment was divided into laboratory tests boreholes with cold water clearly shows the tem­
with test specimens for the detection of delamination perature differences in this area.
and voids, outdoor experiments for the detection of The data obtained from the test series were used to
delamination, investigations of damage detection on real calibrate numerical finite element models, which were
objects, numerical analyzes and the determination of the used to test further variants with regard to damage
probability of detection (Mold et al., 2019). dimensions and their depth for determining the prob­
ability of detection (see Chapter 6), (Mold et al., 2019).
5.2 Laboratory tests 5.2.2 Delamination in concrete slabs
The investigation of the delamination was carried
5.2.1 Heat transport in voids out on three different concrete slabs with the dimen­
Concrete samples with prefabricated holes (see sions 60/60/15 cm, which were prepared inside with
Figure 3) with 1.3 and 2.5 cm diameter were exam­ polystyrene layers with a thickness of 1 cm in differ­
ined at depths of 2 to 5 cm respectively 0 to 10 cm ent depths. In the first test run, the concrete slabs
overlap. The samples had a dimension of 40/10/ were actively heated and the photographs were taken
10 cm and a compressive strength class of C25/30. during the cooling phase.
For the thermographic images, a handheld camera Figure 4 Shows the built-in delamination layers in
(FLIR E60bx) and a drone camera (FLIR Vue Pro) the concrete slabs and the associated thermographic
were used. The specimens were tested in the labora­ images of the surface during cooling.
tory at an ambient temperature of 24 ° C. After the The heat source for the heating of the concrete
active heating of the test bodies, the cooling process slabs was limited in size, therefore the test speci­
could be recorded in predefined time steps with the mens were heated irregularly (strongest warming in
infrared cameras. The boreholes were further cooled the middle of the slab). For this reason, the tempera­
with water in order to determine the influence of the ture differences on the surface caused by the delam­
temperature in the boreholes on the damage detec­ inated areas are recognizable, but not clearly
tion. The heat flow in the concrete body was demarcated (Mold et al., 2019).
recorded by built-in Pt1000 temperature sensors.
Figure 3 shows the thermal images of one of the
investigated concrete specimens a) with air-flowed 5.3 Outdoor tests - delamination in slabs
voids and b) water-cooled voids. In the next test series, the slabs with delamination
Due to the artificial hollow spaces in the concrete, were exposed to natural solar radiation for a full day
the location and size of the damaged areas was and recorded every hour with the thermal camera.

132
 Figure 5. Thermographic images of a plate with delamin­
ation at different times of the day (position of the delamin­
ation layers of the recorded plate see Figure 4) (Mold et al.,
2019).

Figure 4. A) - c) illustration of the concrete slabs with the

incorporated delamination layers (yellow - at 2 cm depth,
orange - at 5 cm depth, brown – at 10 cm depth) and the
associated thermographic images during cooling, d) prepar­
ation of the concrete slabs with styrofoam installations,
extracted from Mold et al., 2019.

The uniform heating of the panels by solar radiation

and the strong temperature differences between day
and night (min 10 ° C at night and max 21 °
C during the day) provided clearly definable damage
patterns, which were most pronounced during the
greatest warming between 12:00 and 13:00 and
during the greatest cooling between 22:00 and
23:00. F
Figure 5 Shows recordings of plate b) (according to
Figure 4) at different times of day (Mold et al., 2019).
Figure 6. Thermographic images of a detail of the railway-
bridge Viaduktgasse at 10:00 and 12:00, extracted from
5.4 Thermography on real objects on site Mold et al., 2019.
In the last test program, thermographic images
were taken directly at the building. The details of
the railwaybridges Viaduktgasse and Italieners­ 6 NUMERICAL ANALYSIS FOR THE
chleife in Vienna were recorded at different times DETERMINATION OF THE PROBABILITY
of the day with different weather conditions. OF DETECTION
These recordings confirmed that the recording
time is important, but also that it is important to To assess the reliability of a system, the reliability of
make several consecutive observations during the used test and monitoring system and the prob­
the day so that different types of damages can be ability of detection of damage is an indispensable
detected. As shown in Figure 6, fouling and soil­ piece of information.
ing on the surface and also inhomogeneities A basic requirement for the evaluation of
below the surface could be detected at different a recognition probability is the knowledge about the
times of the day (Mold et al., 2019). true value of the defect. Determining the actual extent

133
of damage is the most difficult step in the assessment body has reached 50 ° C. At time a) the cooling
process. The development of reference samples plays begins in dependence of the defined temperature
a crucial role in assessing the reliability of nondes­ of the environment and in the voids. The distri­
tructive testing (Kessler and Kanzler, 2019). butions c), d), e) and f) illustrate the temperature
In order to be able to give a statistical statement curve according to c) 30 min, d) 60 min, e) 90
regarding the detection rates of voids in concrete by min and f) 120 min after the cooling start. On
means of thermography, finite element models with the basis of the peaks in the temperature curve,
further variations of the void size and depth in the test the temperature fluctuations in the damaged areas
specimen were simulated. The data and information become visible.
from the laboratory tests were used for the calibration The simulations were carried out with further
of the models. variations of the test specimens with different
The modeling was done in the program “GiD”, ambient and void temperatures and with varying
in which the heat flow in the concrete specimen diameter and varying depths of the voids. The
was numerically simulated. The simulation was temperature values from the simulation were used
carried out in several steps. In the first step, for the analysis of the probability of detection
a temperature of 50 °C was applied to the outer (POD) of voids.
surfaces of the model as own temperature of the Defective areas can only be detected if there is
concrete body. After the heat was distributed a temperature difference between the intact and the
throughout the whole concrete body, the ambient damaged area. Assuming, as shown in Figure 8,
temperature was defined as 25 °C and the tem­ a normal distribution of the surface temperatures of
perature in the void was set at 12 °C. Subse­ the two regions, the probability that the damage is
quently, the cooling process of the concrete body not recognized (referred to as PD) arises from the
was simulated. The temperature profile at the sur­ overlap of the two distributions. Thus, the detection
face was gathered from the models after specified probability (POD) was calculated as POD = 1-PD.
time steps. The temperature trend shows clear Figure 8 shows an example of the detection possibil­
changes in the area of the voids (see Figure 7). ities for point 5 on the finite element model from
The graph in Figure 7 shows the progression of Figure 7 at time c).
the surface temperature at several points on the The detection probability was calculated for
modelled specimen. The distribution b) in the different situations. The temperature of the con­
graph shows the temperature over the total length crete sample, the environment and the tempera­
(0 to 40 cm) of the front side of the concrete ture in the void was varied and calculated for
specimen at the time when the heating of the depths of 1,5 to 6 cm. The probability of detec­
tion is increased by the temperature difference
between the cavity and the concrete surface and
can reach up to 99% depending on the depth
(Mold et al., 2019).

Figure 8. Determination of the POD with the data from

Figure 7. Surface analyzes of the modeled concrete bodies Figure 7, a) shows the POD due to the temperature differ­
with borehole diameters of 2.5 cm, with a test body tem­ ence between the points c5 and c1, in b) the POD is lower
perature of 50 ° C; an ambient temperature of 25 ° C and due to the small temperature difference between c5 and c6,
a borehole temperature of 12 ° C, extracted from Mold extracted from Mold et al., 2019. a) High POD b) Low
et al., 2019. POD.

134
7 RECOMMENDED APPLICATION AND • Material inhomogeneities with a high to medium
INTEGRATION INTO THE CONDITION probability
ASSESSMENT • Built-in parts with a high to medium probability

The test results show that thermography together with Developments in measuring technology allow more
photogrammetric reconstruction are a promising efficient and detailed 3d reconstructions for BIM –
method for the non-destructive detection of visually models as well as an implementation of structural
unrecognizable damages. inspections and condition assessments. The reliabil­
ity analysis for damage detection by means of therm­
The application of thermography for the detection
of voids and delamination can be carried out well ography not only allows a more efficient evaluation
under appropriate environmental conditions. An essen­ of the condition but above all makes it possible to
tial factor for the possibility of damage detection is the ensure the safety of bridge structures.
temperature difference between the building and its
surroundings so that a heat flow can occur in the struc­ REFERENCES
ture. The most suitable times of day for the measure­
ment are those at which the building warms up or Auer, M., 2019. Thermografische Zustandserfassung von
cools down the strongest. In the tests carried out, it Betonstrukturen mit Hochleistungsdrohnen.
was the midday and the hours before midnight when Bernhard, F. (Ed.), 2014. Handbuch der Technischen Tem­
the damage was best recognized. In order to be able to peraturmessung. Springer Berlin Heidelberg.
make a clear statement on the extent of the damage, it Escobar-Wolf, R., et al., 2018. Unmanned Aerial Vehicle
is recommended to take several shots over the course (UAV)-Based Assessment of Concrete Bridge Deck
of the day, as the damage can be more accurately Delamination Using Thermal and Visible Camera Sen­
sors: A Preliminary Analysis. Research in Nondestruc­
determined based on the temperature changes. The tive Evaluation 29, 183–198.
measurements on the building show that properties Hiasa, S., 2016. Investigation of infrared thermography for
such as material inhomogeneities, soiling or growth on subsurface damage detection of concrete structures.
the building become more visible at different times of Hiasa, S., Birgul, R., Catbas, F.N., 2017a. Effect of Defect
the day. A major factor in the detection of damages is Size on Subsurface Defect Detectability and Defect
solar radiation, which enables the greatest temperature Depth Estimation for Concrete Structures by Infrared
differences between the environment and the structure. Thermography. Journal of Nondestructive Evaluation
The thermography is suitable for selective detail 36.
Hiasa, S., et al., 2018. Experimental and numerical studies
shots as well as for area-wide imaging of the entire
for suitable infrared thermography implementation on
structure. It makes sense to take thermographic concrete bridge decks. Measurement 121, 144–159.
images of the bridge during regular inspections. The Hiasa, S., Birgul, R., Necati Catbas, F., 2017b. A data pro­
reference images of earlier images make it easier to cessing methodology for infrared thermography images
detect changes in the heat flow in the structure and of concrete bridges. Computers & Structures 190,
thus damage (Mold et al., 2019). 205–218.
Hiasa, S., et al., 2016. Monitoring concrete bridge decks
using infrared thermography with high speed vehicles.
8 CONCLUSION Structural Monitoring and Maintenance 3, 277–296.
Hoffmann, M., 2018. Lebenszykluskosten der
Straßeninfrastruktur. Habilitationsschrift an der Tu
The analysis of the probability of detection of dam­ Wien.
ages by thermography provided deep insights into Hoffmann, M., et al., 2019. Forschungsprojekt RISKMON -
the possibilities for non-destructive and contactless Anlageninspektion und RISK-MONitoring mit Hochleis­
detection of damages on bridge structures. In the tungsdrohnen (UAS/UAV).
course of the experiments it was found that repeated Janků, M., Březina, I., Grošek, J., 2017. Use of Infrared
observations during one day (also during one year) Thermography to Detect Defects on Concrete Bridges.
under certain environmental conditions allow Procedia Engineering, Structural and Physical Aspects
a reliable detection of the following damages (see of Construction Engineering 190, 62–69.
Kessler, S., Kanzler, D., 2019. Basic requirements for POD
also (Hoffmann et al., 2019)):
using NDT in Civil Engineering.
• Delaminations in concrete structures at a depth Lillesand, T.M., Kiefer, R.W., Chipman, J.W., 2015.
of 5 cm. Remote sensing and image interpretation, Seventh edi­
• Delaminations in concrete structures at a depth of tion. ed. John Wiley & Sons, Inc, Hoboken, N.J.
Mehlhorn, G., Curbach, M., 2014. Handbuch Brücken -
10 cm with a medium probability
Entwerfen, Konstruieren, Berechnen, Bauen und Erhal­
• Voids with diameter 2cm, depth 2cm with ten, 3. Auflage ISBN 978-3-658-03339–2. ed. Springer,
a medium probability Wiesbaden.
• Voids with diameter 2cm, depth 4cm with a low Milovanović, B., Pečur, I.B., 2016. Review of Active IR
probability Thermography for Detection and Characterization of
• Cracks (especially under moist conditions) with Defects in Reinforced Concrete.
a high probability Mold, L. et al., 2019. Thermografie zur Erfassung von
Schäden and Brückenbauwerken. Bautechnik.

135
NEXCO-West USA, Inc., 2014. Comparison of Infrared Langzeitbewertung der integralen Marktwasserbrücke
Cameras for Concrete Bridge Deck Scanning: - Vol.2 S33.24]. Beton- und Stahlbetonbau, 106 (4), pp.
Field Test at Haymarket Bridge. 231–240.
Omar, T., Nehdi, M.L., 2017. Remote sensing of concrete Strauss, A., Vidovic, A., Zambon, I., Dengg, F., Matos, J.
bridge decks using unmanned aerial vehicle infrared C., 2016. Performance indicators for roadway bridges.
thermography. Automation in Construction 83, Maintenance, Monitoring, Safety, Risk and Resilience
360–371. of Bridges and Bridge Networks - Proceedings of the
Rocha, J.H.A., Póvoas, Y.V., 2017. Infrared thermography 8th International Conference on Bridge Maintenance,
as a non-destructive test for the inspection of reinforced Safety and Management, IABMAS 2016, pp. 965–970.
concrete bridges: A review of the state of the art. Sultan, A.A., Washer, G., 2017. A pixel-by-pixel reliability
Rocha, J.H.A., Póvoas, Y.V., Santos, C.F., 2019. Detection analysis of infrared thermography (IRT) for the detec­
of Delaminations in Sunlight-Unexposed Concrete tion of subsurface delamination. NDT & E International
Elements of Bridges Using Infrared Thermography. 92, 177–186.
Journal of Nondestructive Evaluation 38. Watase, A., Birgul, R., Hiasa, S., Matsumoto, M.,
Strauss, A., Wendner, R., Bergmeister, K., Reiterer, M., Mitani, K., Catbas, F.N., 2015. Practical identification
Horvatits, J., 2011. Monitoring and influence lines of favorable time windows for infrared thermography
based performance indicators [Modellkorrekturfaktoren for concrete bridge evaluation. Construction and Build­
als “performance Indikatoren” für die ing Materials 101, 1016–1030.

136
 Bridge Maintenance, Safety, Management, Life-Cycle Sustainability and Innovations –
Yokota & Frangopol (eds)
© 2021 Taylor & Francis Group, London, ISBN 978-0-367-23278-8

Application of wireless sensor technology in load testing of large-span

cable-supported bridges
Z. Xu & X. Zhang
CCCC HIGHWAY CONSULTANTS CO., LTD

J. Wang
Jiangsu Sutong Bridge Co., Ltd

ABSTRACT: Load testing is the most direct way to verify the actual bridge condition which includes static
load test and dynamic load test. Local or overall structural damage can lead to changes in the sectional stress,
deformation or vibration. For the load testing of large-span bridges, it required many monitoring items and
had strong correlation between different parameters. According to the traditional sampling method, it is diffi­
cult to achieve all parameters effective and reliable in a short time. Through the integration of wireless sensors
and online monitoring systems, the load testing of Sutong Bridge during the operation period realizes the real-
time and remote monitoring. It has outstanding advantages in improving the timeliness and ensuring the com­
pleteness of data monitoring. It can be used as a reference scheme for structural load tests of large-span
bridges with similar time requirements and numerous monitoring parameters.

1 INTRODUCTION project was completed and opened to traffic in 2008.

The main bridge is a 1088-m double-tower double-
Bridge load testing is the most direct and effective plane cable-stayed bridge, which is a landmark pro­
way to evaluate the structural performance and ject for the development of China’s bridge industry.
working state. The static and dynamic characteristics Since the completion, a large number of routine
of the structure or component are tested by applying inspections and evaluations have been carried out for
loads, normally focused on such items as structural the bridge, and the performance of the bridge struc­
deformation and section stress. Usually, the deform­ ture is in good condition. Limited by the heavy traf­
ation and the resistance are tested by the photoelec­ fic over the Yangtze River, it has not been able to
tric method; while the stress is tested by the carry out the load testing under fully traffic-closed
vibrating wire strain method. These methods can conditions for last ten years of operation. According
meet the requirements of load tests in small-and­ to other similar engineering practices in China, it is
medium span bridges. For large-span cable- now in the infancy of the disease germination, and it
supported bridges, however, the complexity of the is urgent to carry out a load test for the bridge.
structure and the diversity of the components Based on the above considerations, the load testing
increase the monitoring items, in which the correl­ under full traffic-closed state of the bridge was car­
ation between different parameters is strong. On the ried out in due course. In accordance with traffic
other hand, the large-span cable-supported bridge at plan approval, the total closed time of this load test­
operation belongs to the key project on traffic; and ing was from 0:00AM to 5:00AM. It carried out
affected by heavy traffic pressure, it is difficult to a total of seven inspection conditions. If using the
close the traffic for a long time. According to the traditional data collection method, it is difficult to
traditional sampling method, it is hard to achieve all complete the whole test tasks. Therefore, the issue
parameters effective and reliable in a short time, and how to carry out the load testing efficiently was an
only through the large investment of personnel or important subject that needs to be solved.
equipment or the abandonment of certain non-key During the test course, wireless sensor technology
items as a compromise. is adopted to improve the efficiency and coverage of
Sutong Bridge across Yangtze River is the world’s monitoring data acquisition, which achieves good
first cable-stayed bridge over a kilometer span. The application results.

DOI: 10.1201/9780429279119-14
https://doi.org/10.1201/9780429279119-14

137
2 OVERVIEW OF WIRELESS MONITORING through wireless, and then collect in-site data to
SYSTEM a relay node or a gateway. After that, upload data to
an online monitoring cloud through a carrier network
With the development and research of intelligent (2G/3G/4G).
sensing devices and information software, the inte­ The data transmission adopts a point-to­
gration of wireless sensor devices and transmission multipoint star topology or improved star network
systems can save manpower and other investment, topology, and performs signal transmission-reception
which is convenient for personnel to put the main control and data concentration in a point-to­
energy into the loading. In this once-a-decade load multipoint manner.
testing, it adopts wireless monitoring and transmis­ In order to realize the long-distance transmission
sion technology which improves the timeliness of of data and the expansion of system network scale
monitoring and ensures the completeness of data on large-span bridges, the data transmission module
monitoring compared with other load testing adopts an improved star network topology. Each
methods. The architecture of the wireless monitoring node in the network is connected to the relay node in
system during the load testing is as follows: the effective network by point-to-point, and then the
The wireless monitoring system communicates the relay node forwards the data or command informa­
acquisition sensor and system platform through the tion to the control center or the destination node, i.e.,
wireless transmission part. By using the existing com­ the gateway. Either the relay node or the gateway is
munication means, it realizes sensor distribution, trans­ executed a centralized communication control strat­
mission automation and data centralization. The egy, and the relationship of the data topology is as
system architecture consists of three levels, namely the shown in Figure 1.
acquisition part, the transmission part, and the process­
ing part. The system scheme is composed as shown in
Table 1. 3 SETTINGS AND COMPARISON OF
The most important function of wireless sensors MONITORING ITEMS
is the acquisition, transition, treatment and storage of
the information in the physical world. Based on this From the perspective of the structural behavior, the
function, the wireless sensor node mainly includes load testing of large-scale cable-supported bridges
a sensor module, a processor module, a wireless has the following characteristics compared to con­
communication module and an energy supply ventional-type bridges:
module. In terms of concerned items in the load test­
1) Variable monitoring items and strong correlation
ing, the data acquisition module can be divided into
between different parameters
three types, namely the load action, the structural
Cable-supported bridges have the characteristics
response and the environmental parameters, which
of complex structural forms and various compo­
covers key parameters such as acceleration, displace­
nents. The behavior is characterized by statically
ment, dip angle, temperature & humidity, wind
indeterminate and structural nonlinear. Therefore,
speed and direction.
the inspection of structural behavior should be
The data transmission module adopts a wireless
made comprehensive consideration from the per­
transmission system, which can achieve real-time
spective of multi-parameter correlation analysis.
data transmission. It can set up a network flexibly
The monitoring items are far more than the trad­
itional bridge structures, and different parameters
need be obtained synchronously. The conven­
Table 1. Scheme of short-term intelligent monitoring tional data collection methods are limited by
system. space, difficult to ensure synchronization. With
the wireless monitoring system, the monitoring
Data trans- Data pro- command can be released, and simultaneous sam­
Data processing mission cessing pling of each position parameter can be realized.
environment wind speed and mobile oper- on-line
monitoring wind direction ator service monitoring
temperature and (2G/3G/4G) system
moisture relay node
rainfall
structure stress
response cable force
deflection Expert
…… evaluation
inclination system
Load traffic volume
monitoring vehicle load
Figure 1. Relationship of the data transfer topology.

138
2) Long-term data stability period Table 2. Comparison of monitoring methods.
Cable-supported bridges are equipped with special
performance damping members for seismic or Monitoring wireless sensor Conventional
vibration suppression requirements. Because the items monitoring methods
large span of the bridge and the strong correlation
of the parameters, the response of the loaded struc­ Tower dis- GPS, wirelessly trans- total station
ture requires a certain buffering time, and the placement mit, Remote
monitoring
period required for data stabilization exceeds that
Beam dis- Laser displacement Manual measure-
of a conventional structural bridge. It is necessary placement meter + Wireless ment, such as ruler
to continuously monitor the structural response transmission system
within a certain time to determine whether the Wireless transmission,
data is stable. With the wireless monitoring remote monitoring
system, continuous coverage in the time dimen­ Crack Crack gauge + Wire- Manual measure-
sion can be achieved, avoiding the misjudgment monitoring less transmission ment, such as crack
caused by the test data at a single moment. system width gauge
3) A large demand for loading vehicles and a risk of Wireless transmission,
overload in non-loading locations remote monitoring
In the loading process of large-scale cable- temperature Automatic recorder Manual
supported bridges, more loading vehicles need to & humidity measurement
be arranged to achieve the loading efficiency.
When the monitoring cross-section efficiency is Wireless sensor acquisition mode is using the
achieved, it may cause the results of the non- same strain acquisition sensor as conventional
monitoring section out of control. Structural acquisition at the sensor level. Then, the relay
response needs to be monitored over a certain node is added, collecting the data at each point.
spatial range to avoid damage to non-monitored The relay node sends a packet, and the data is
section locations. With the wireless monitoring uploaded to the online monitoring cloud platform
system, full coverage in the spatial dimension by means of the gateway and the operator net­
can be achieved, avoiding over-limit damage in work (2G/3G/4G) to realize real-time data
non-monitoring locations. transmission.
4) Complex structural responses affected by the 2) Strain monitoring over the whole span
external environment When subjected to the equivalent critical section
The external environment, such as temperature load, other non-control section positions may be
and ambient wind, will have a certain impact on loaded force exceeding the design requirements.
the structure. For instance, due to the difference in Therefore, the overall strain of the full span bridge
heat conduction between the concrete and the needs to be monitored to avoid overloading risk.
steel, the structural temperature of the concrete Conventional acquisition modes generally do not
cable tower and the steel box girder will be incon­ pay attention to it.
sistent under the same external environmental con­ In wireless sensor acquisition mode, distributed
ditions. It must be considered the influence of the fiber strain monitoring is a new type of fiber sens­
above factors when correcting the temperature. ing method that can be used to monitor the strain
Load test monitoring mode with the wireless sen­ on the surface of a structure. It is generally com­
sors is adopted in the test process. In the overall posed of a demodulator and a sensing fiber
process, it is monitoring the key section strain, the (optical cable). The demodulator emits two paths
full span strain, the girder deflection, the girder of pump light and probe light (Stokes light),
inclination, the pylon deviation, the segment dis­ which are respectively injected into the sensing
placement, the cable force, crack development, fiber from both ends of the fiber.
temperature and humidity, et.al. When the frequency difference between the two-
The wireless sensor monitoring methods are com­ way light falls within the Brillouin spectrum,
pared with the conventional data collection method, a stimulated Brillouin effect is generated in the
as shown in Table 2. fiber. The energy is transferred from the pump
light to the probe light, so that the probe light
1) Strain monitoring at critical sections gains. The time-domain distribution of the back­
The behavior of the critical section is equiva­ scattered probe light reflects the Brillouin gain
lently tested by applying the load, and the vari­ of each point at the fiber. When obtained the
ation of sectional force can be characterized by Brillouin frequency shift value of each point at
the strain. the fiber, the strain of each point of the fiber can
Conventional acquisition modes are taking be measured.
a resistive or vibrating wire strain sensor, with 3) Girder deflection
special reading equipment, data acquisition and The behavior of the critical section is tested by
feedback from the artificial field. applying the load, among which the deflection is

139
 a symmetrically arranged differential pressure
static level. At the same time, the difference
between the two sides is compared and the trans­
verse deflection of the girder is analyzed.
5) Cable force
Based on the principle of tension string, the cable
force has a functional relationship with the nat­
ural frequency of the cable. By measuring the
vibration acceleration signal of the cable, the
acceleration signal is spectrally transformed and
the fundamental frequency is extracted. Then, the
Figure 2. Principle of distributed optical fiber strain cable force can be calculated.
monitoring.
In conventional acquisition mode, the cable force
test of a single cable is obtained by using a cable
the most reliable evaluation index reflecting the force measuring device.
overall structural behavior. In wireless sensor acquisition mode, high-
In the conventional acquisition mode, the means precision vibration acquisition chip is integrated
includes level, total station or GPS-RTK technol­ with wireless transmission, data processing and
ogy. With the advancement of measurement tech­ energy module. After collecting the data col­
nology, the measurement robots have liberated lected at each point, upload the data to the oper­
some manpower, but it still need to be equipped ator network (2G/3G/4G) via gateway and carrier
with certain personnel for auxiliary measurement network. Monitor the cloud platform online to
at field. realize real-time data transmission.
In wireless sensor acquisition mode, the monitor­ Other wireless monitoring methods are shown in
ing sensor of the girder deflection adopts the dif­ Table 2.
ferential pressure static level. Using the principle The comparative analysis of the operational phase
of the connected pipe and the physical law of the load test and the completed phase load test need to
liquid transfer pressure, the reference point and be carried out under the same load level, to deter­
the monitoring point are connected to each other, mine whether the actual performance of the structure
and the pressure changes between the monitoring is attenuated. Therefore, the main loading conditions
point and the reference point is measured in real are as consistent as possible with the completed
time. Then, the settlement of the monitoring bridge stage, so as to facilitate the comparative ana­
point relative to the reference point can be calcu­ lysis of the data. The static load test arrangement is
lated. Installing a liquid pressure sensor on the shown in Figure 3, and the scheme comparison is
bottom of the monitor, the liquid level height shown in Table 3.
value is calculated indirectly by measuring the The main working conditions of this test (the
liquid pressure. mid-span deflection loading condition of the 300m
The liquid level height value is calculated indir­ side span along Nantong, the mid-span deflection
ectly by measuring the liquid pressure by install­ symmetrical/eccentric loading condition of the main
ing a liquid pressure sensor on the bottom of the span) are the same as the completed bridge stage.
monitor. There are many sensing technologies for The number, position and overall weight of the load­
liquid pressure sensors, such as strain gauge pres­ ing vehicles are as consistent as possible. The
sure sensors and silicon resistance pressure sen­ arrangement of dynamic measuring point is consist­
sors. The collected data at each point is also ent with the completed bridge test, and the consist­
summarized by the relay node. The relay node ency of the above working conditions is the basis of
sends the packet, and the data is uploaded to the the subsequent comparison.
online monitoring cloud platform by means of Compared with the completed bridge test, this test
the gateway and the operator network (2G/3G/ has obvious advantages in terms of timeliness of
4G) to realize real-time data transmission. data, spatial continuity and investment in monitoring
4) Transverse inclination of the girder process. This is mainly reflected in the following
Transverse inclination of the girder can character­ aspects:
ize structural symmetry performance.
In conventional acquisition mode, the leveling
device is adopted and the relative height differ­
ence is tested by laterally arranging the measur­
ing points. Associated with the absolute value of
the girder deflection, the absolute inclination can
be calculated.
In wireless sensor acquisition mode, the deflec­
tion of the girder is monitored by means of Figure 3. Layout of static load test.

140
Table 3. Comparison of static load test condition.

efficiency
load condition Vehicle load Items LoadValue coefficient Notes

Condition1: Negative bending moment at the top of 40 Vehicles Moment -121401 0.75 Consistent
the auxiliary pier along Nantong side (kN.m)
Condition 2: Mid-span bending moment at the 300m 26 Vehicles Moment 69958 0.86 Inconsistent
side span along Nantong direction (14Vehicles) (kN.m)
Condition 3: Mid-span deflection at the 300m side 40 Vehicles Deflection 0.292 0.86 Consistent
span along Nantong direction (m)
Condition 4: Mid-span bending moment at the middle 24 Vehicles Moment 101008 0.88 Inconsistent
span (16Vehicles) (kN.m)
Condition 5: Mid-span deflection at the middle span 52 Vehicles Deflection 1.488 0.87 Consistent
(m)
Condition 6: Mid-span bending moment at the side 26 Vehicles Moment 66443 0.82 Inconsistent
span along Suzhou direction (26 Vehicles) (kN.m)
Condition 6: Mid-span deflection at the middle span 39 Vehicles Deflection 1.486 0.87 Consistent
under eccentric load (m)
Notes: (1) The load value is the actual load value determined by the actual loaded vehicle weighing and loading position.
(2) ↓ indicates that the vehicle is stayed at the loading position of the previous working condition, at which the influence
line is small in this condition. (3) Monitoring items include cable force, steel box girder stress, defect development, deform­
ation (girder, cable tower, beam end displacement, etc.).

1) It is monitoring the whole process of the test. On bridge can be monitored by distributed monitoring
the one hand, all the positions where the sensors equipment.
are provided can be continuously monitored to 5) The monitoring process needs less staff, and
ensure all the structure positions are in achieves remote visualization of all monitoring
a controllable state during the non-controlled load­ data by means of the cloud monitoring platform.
ing conditions. On the other hand, the response Only a small number of data monitoring person­
during the loading process is monitored through nel are on site, and the staff focuses on loading
high-frequency monitoring to evaluate the data organization scheduling and data evaluation.
stabilization time. Routine tests are monitoring
results at specific times, which are greatly influ­
enced by the experience of loading control.
4 TEST RESULT ANALYSIS AND
2) The timeliness of loading is improved. The traffic
APPLICATION
closure approval time of this test is 5 hours and
the actual loading process time is only 4.5 hours.
The results obtained from the tests are analyzed to
In this period, 7 loading conditions are completed;
judge changes in structural performance states.
that is to say, the average loading takes less than
40 minutes per each working condition. The com­
pleted bridge test took a total of 48 hours and 4.1 Deflection and strain analysis at critical
completed 15 loading conditions. Its average time section
was 75 minutes per each condition, which was
almost twice that of the current loading method. Based on the static parameters of the structure (static
3) The monitoring process achieves simultaneous displacement, strain, etc.), the static load test results
monitoring of various parameters. Through the are evaluated in accordance with the specifications.
remote control command, the communication The results of the two load tests before and after are
control function is applied between the relay compared to analyze the performance changes after the
node and the gateway, ensuring the synchroniza­ bridge operation. Because the wheelbase and the
tion of each parameter monitoring. Conventional vehicle weight in the test is not strictly consistent with
tests are carried out independently of each moni­ that in the completed bridge test, the maximum calcu­
toring project and synchronization cannot be lated effect value (Ss) of the force or the displacement
fully guaranteed. at the critical section is inconsistent in two load tests;
4) It has high coverage of the monitoring range. thereby, a calibration coefficient is used as
Besides the critical section locations in the conven­ a comparison index for the comparable load condi­
tional test monitoring, the condition of the full-span tions. The test results illustrate that the calibration

141
 coefficient of the main measuring points in each load 4.3 Analysis of structural response lags
condition is basically less than 1. Compared with the
Based on the high-frequency sampling method, the
completed bridge test, the calibration coefficient of this
beam end displacement and the pylon deflection are
test can be basically the same, and the relative devi­
continuously monitored to analyze the component per­
ation does not exceed 10%, indicating that the structure
formance and structural response. The results show
does not appear obvious functional degradation.
that the structural response is normal and basically
consistent with the theoretical calculation. There are
no obvious restrictions on the constrain system such as
4.2 Distribution of the overall structural behavior expansion joints and bearings. At the same time, the
Based on the differential pressure static level and dis­ time history curve shows that the deformation response
tributed optical fiber arranged along the span, the dis­ slightly lags when the expansion joint is pulled out in
tribution of the overall structural behavior is the small displacement stroke. After the strain and
analyzed. As shown in Figure 4, the results show that deflection test data is stabilized for 15 minutes, the dis­
the actual force distribution trend is in good agree­ placement of the beam end still appears a trend of
ment with the theoretical calculation. The extreme increases, which shows a curve with inclined growth.
position and absolute value are consistent with the When the expansion joint is compressed in the large
calculation results, and the overall performance of the displacement stroke, the lag effect of deformation
structure is consistent with the design expectation. response is not obvious. After unloading, the beam end

Figure 6. Continuous displacement of expansion joint (unit:

Figure 4. Deflection distribution of the whole span in Con­ cm).
dition 5 (unit:cm).

displacement recovers rapidly, and the curve appears

as a nearly vertical descending curve.

4.4 Analysis of monitoring data under free flow

In order to verify whether the free-flow simulation
test can be adopted when the traffic is not closed,
after the load testing the existing monitoring equip­
ment is used to continuously monitor the overall
strain and deformation of the structure under normal
traffic flow for two weeks.
During the monitoring period, the measured strain
and deflection envelope curves are shown in the
Figure 7 and Figure 8.
The results show that the response trends of the
measured structure at each point under free flow are
consistent with the theoretical calculations, and are
basically within the envelope range allowed by the
theoretical calculations, indicating a well-performed
Figure 5. Stress distribution of the whole span in Condition macroscopic mechanical behavior during the moni­
5 (unit:kpa). toring. After continuous monitoring within a certain

142
 response and loading process, the structural
response process can be obtained by using the
high-frequency real-time and continuous moni­
toring. Then, the abnormal conditions during the
test are identified and the performance of key
members is analyzed.
3) It reduces on-site labor costs and simplifies the
loading and coordination process. By the means
of wireless transmission and remote monitoring,
Figure 7. Strain envelope curve of the girder during moni­ remote command can substitute for manual on-
toring period (units με in vertical axis, m in horizontal axis). site acquisition process, which reduces the per­
sonnel investment in the field. The remote moni­
toring can be conducted by 1~2 people collecting
data records. It simplifies the coordination pro­
cess of on-site loading and reduces the one-to­
many instruction delivery process in the field.
Compared with the conventional test methods,
the load testing mode of the large cable-supported
bridge based on the wireless sensor technology
improves the timeliness of work and the complete­
ness of the data, and solves the problem of short
Figure 8. Deflection envelope curve of the girder during loading time in the load testing of large-span
monitoring period (units cm in vertical axis, m in horizontal bridges during the operation period. The model
axis). can be used as a reference case for other similar
bridge tests.
For small-and-medium conventional bridges, the
test conditions are relatively simple, and the above-
period, it can be used as one of the methods for mentioned wireless sensor monitoring methods have
evaluating the macroscopic mechanical state of limitations due to the inability to arrange it quickly.
a bridge without closed traffic conditions.

REFERENCES
5 CONCLUSIONS
[1] Ministry of Transport of the People’s Republic of
In the load testing process of the Sutong Bridge at China. Specification for Inspection and Evaluation of
operation, a load test monitoring system based on wire­ Load-bearing Capacity of Highway Bridge[S]. Beijing:
less sensor technology was formed to solve the load China Communication Press, 2011.
[2] Ministry of Transport of the People’s Republic of
inspection task during the short-term closure period.
China. Specifications for Load Testing of Highway
The system has the following features and advantages: Bridges[S]. Beijing: China Communication Press,
1) It can improve the timeliness of monitoring and the 2015.
completeness of data monitoring. Through remote [3] SONG Yifan. Load testing and structural evaluation of
highway Bridges[M]. Beijing: China Communication
monitoring and distributed measuring point layout, Press, 2002.
the synchronization of each monitoring parameter [4] CUI Fei, YUAN Wancheng, SHI Jiajun. Damage
can be achieved, and the coverage monitoring of detection of structures based on static response[J].
the full span control section position and the non- Journal of Tongji University, 2000, 28(1):5–8.
control section position can be ensured. [5] CAI Jing, WU Zhishen, LI Zhaoxia. Statistical
2) Real-time, continuous monitoring can be approach to evaluating the status of an existing struc­
achieved to analyze and judge structural response ture using static test data[J]. Engineering Mechanics,
effectively. In views of the lag effect of structural 2004, 21(6): 76–80.

143
 Bridge Maintenance, Safety, Management, Life-Cycle Sustainability and Innovations –
Yokota & Frangopol (eds)
© 2021 Taylor & Francis Group, London, ISBN 978-0-367-23278-8

Effectiveness and durability of repair measures on corroding steel in

concrete columns
F. Binder
ASFINAG Service GmbH, Vienna, Austria
Institute of Structural Engineering, University of Natural Resources and Life Sciences, Vienna. Austria

S.L. Burtscher
burtscher consulting GmbH, Vienna, Austria

ABSTRACT: The 5,000 bridges which are maintained by ASFiNAG (Austrian highway operator) are its
main asset. Despite all maintenance efforts, these structures are still subjected to degradation, thus, leading to
high expenses. Decisions for these maintenance actions are usually driven by the actual condition of the struc­
ture or deterministic (conservative) maintenance cycles proposed by guidelines. To ensure that a high road net
availability can be provided at moderate costs, the effectiveness of possible maintenance methods, as well as
their maintenance intervals must be assessed. The aim of this paper is to determine condition and degradation
(condition indices) by using objective indicators, like humidity, electrical resistivity and corrosion potential.
For this purpose, long-term measurements from an overpass situated in the vicinity of Vienna where different
maintenance measures (Hydrophobic treatment, Coating) were applied. The monitoring data allows to con­
tinuously determine the degradation, as well as the effectiveness of repair measures. Thus, a direct statement
can be made regarding the effectiveness of the maintenance actions and thereafter important conclusions can
be made for future repair measures and repair intervals.

1 INTRODUCTION detection of all influences leading to the defects.

The first step should therefore always be the deter­
The chloride load from de-icing salts is one of the mination of the condition by a detailed visual
primary causes of corrosion in reinforced concrete inspection of the components and investigations by
structures. The processes involved in corrosion non-destructive material testing. Currently, inspec­
inside reinforced concrete structures are complex tions are mainly conducted visually at regular inter­
and significantly influenced by environmental con­ vals. On visual inspection, the eye perceives
ditions. Especially for bridges the corrosion which damages, which are marked in appropriate plans for
is induced by chlorides due to de-icing salt, repre­ damage assessment. Thus, further areas can be
sents a danger that must be considered carefully. determined, on which the building inspection is to
The aim of this paper is to present the relevant cir­ be complemented by appropriate additional measur­
cumstances for degradation on corroding reinforce­ ing points. Nowadays concrete tests which can
ment treated with different repair measures, in either be non- or nearly non-destructive can be
order to evaluate them and thus, to provide state­ listed as follows (Küchler, 2013), (Hillemeier and
ments on the effectiveness and their durability. This Taffe, 2012) (Schickert et al., 1991):
allows for a simple prognosis regarding the pro­
gress of degradation. • Determination of the carbonation depth of the
concrete
• Measurement of the concrete cover of the
2 CONDITION ASSESSMENT reinforcement
• Determination of the compressive strength of
Prerequisite for a successful and durable rehabilita­ concrete
tion of reinforced concrete structures is the know­ • Determination of the chloride penetration depth
ledge of the mechanisms of damage and a reliable of the concrete
• Potential field mapping

DOI: 10.1201/9780429279119-15
https://doi.org/10.1201/9780429279119-15

144
• Concrete resistivity mapping tion 4), can be classified with respect to damage or
• Crack recording including the measurement of potential damage due to corrosion. Threshold and
crack widths limits for classification of damages were gathered from
standards, guidelines and literature ((RVS 13.03.11,
Based on this information, a proposal with pos­ 2011), (Baumann et al., 2014), (B4706, 2003),
sible repair strategies should be elaborated. In add­ (EN1504, 2014), (Polder et al., 2000) (Andrade and
ition, reports of any risks can be given as well as the Alonso, 1996)) and are given in Table 1.
emerging costs of recommended measures can be If a measured value exceeds the threshold docu­
determined.
mented in the literature, the index jumps to the next
higher level. This leads to a step-wise indication of
damage over time (abrupt changes in condition),
3 THE INDEX METHOD which is physically not accurate and prevents a good
foresight of necessary maintenances. Therefore, regres­
3.1 General sion procedures (linear or exponential) were used to
The evolution of the durability and damage deter­ gain smooth and gradually changing functions. The fit­
mines a cost-effective maintenance, while the safety ting parameters and the function which are from now
level acts as the baseline for strengthening or repair on used are given in Table 2. Hence, the “Corrosion
methods at – in most cases – considerable higher Damage Index” generated on this way is a gradual and
costs. For this reason, the need for adequate and more suitable function for the representation of the
objective tools for assessing structures has become condition considering main influences.
a subject of crucial economic interest. Several efforts
for quantifying these effects have been undertaken 3.3 Corrosion damage index
(Andrade and Martinez, 2009), (CONTECVET,
2008), (R. Wendner, 2008), (Schneck, 2012), (Furuya The corrosion damage index (CDI) is obtained by
et al., 2011). However, the outcome of a structural first transferring measured values to Corrosion indi­
assessment must contain at least the following points: cator levels. In Table 1 the corrosion indicator levels
are given for certain thresholds of the assessment
• Identification of the damage mechanism values. The levels for the corrosion indicators were
• Extent of damage compiled and extracted from several sources. Next,
• Dynamic of the degradation progress. the data was fitted to a function which describes the
In (CONTECVET, 2008), two procedures are behavior best (see Figure 1a-d). The functions can
described to establish these three aspects of the be used to determine the gradual corrosion indicator
assessment in case of concrete structures affected by level for each assessment method (not individual
corrosion. In this paper a simplified procedure, that investigations). Secondly, the applied corrosions
excludes the factor for exposure, is applied and indicator levels were averaged, which results in the
described in the following. overall assessment called the “Corrosion Damage
Index” (see Eq. 1).

3.2 Indicators for corrosion

Many (nearly all) measurements in the condition
survey, used in the paper (see also discussion in sec-

Table 1. Indicator for corrosion and their indices.

Corrosion indicator level

1 2 3 4
Corrosion Indicator Negligible Low Moderate High
Carbonation depth XCO2 [mm] 0% C ≤ 75% C >75% C ≥C
Chloride content XCl­ [M%] ≤0,2 >0,2≤0,6 >0,6≤1,0 >1,0
Crack width due corr. w [mm] no <0,3 >0,3 Spalling
Concrete resistivity ρ [Ωm] >1000 500-1000 100-500 <100
Corrosion potential1 Ecorr [mV] > -200 -200 ­ -350 - -500 < -500
-350
Corrosion current icorr [μA/cm²] <0,1 0,1-0,5 0,5-1 >1

1 using Cu/CuSO4 reference electrode

C means thickness of the concrete cover

145
Table 2. Fitting parameters, function and coefficient of determination for each Corrosion indicator.

Indicator corrosion Akr. Function a b R²

b
Carbonation depth XCO2 ax 1.0123 0.2788 0.9783
Chloride content XCl­ a þ bx 2.8998 1.1878 0.9772
Crack width due corr. w axb 4.0006 0.1866 0.9621
Concrete resistivity ρ aebx 3.7676 -0.0010 0.9777
Corrosion potential Ecorr a þ bx -0.0153 -2.6353 0.9803
Corrosion current icorr axb 3.8655 0.2934 0.9963

Figure 1a. Linear function for mapping the chloride content to its corrosion indicator level. Figure 1b Power function for
mapping the corrosion rate to its corrosion indicator level. Figure 1c Exponential function for mapping the concrete resistivity
to its corrosion indicator level. Figure 1d Linear function for mapping the corrosion potential to its corrosion indicator level.

The procedure in Equation 1, together with the 3.4 Corrosion damage index derived from
levels for the individual corrosion indicators, lead to monitoring data
the corrosion damage index, which can subdivided
Monitoring data from an overpass was compiled into
in the following four classes:
the models. Since installation of the monitoring
• 1 = negligible or no corrosion, system, trends could be observed, that indicate
• 2 = low corrosion, a gradual corrosion damage. The monitoring data
• 3 = medium corrosion, used for the investigation were
• 4 = high corrosion.

146
• Electrolytic resistance, ant when it comes to classification of the measured
• Potential of the reinforcement, potentials and is already considered in Table 1.
• corrosion current.
The monitoring data is now available for 7.5 4.4 Galvanic currents of macro cells
years. For this time range the chloride content and Nowadays, many techniques for the assessment of
the carbonation were assumed to be constant. Thus, steel corrosion in concrete are available. Measure­
the progress and the variation of the indicators can ments of galvanic currents were used for obtaining
be mapped across all recorded indicators.
information about the corrosion rates due to macro-
cell action under defined conditions (Raupach, 1996).
The method has the advantage of providing direct
4 MEASURED PARAMETERS indication of electrochemical activity in the system
without the need for sophisticated instrumentation.
4.1 Temperature In a macro element, anode and cathode are spa­
Temperature sensors are usually electronic compo­ tially separated. The corrosion current can be meas­
nents that provide an electrical signal to measure ured directly by electrically separating the anode and
temperature. In case of the installed monitoring the cathode using a Zero Resistance Ampere meter
system, the component temperature is measured by (ZRA). The measurement of the macro element cur­
several sensors (multi-ring electrode and reference rent can be used for continuous monitoring of the
electrode). In addition, the air temperature is also corrosion progress in reinforced concrete structures.
recorded so that any differences between air and For this purpose, individual reinforcement sections
component temperature can be recorded and, thus, can be separated from the remaining reinforcement
data verified. (Broomfield et al., 2002), (Chrisp et al., 2002),
(Schiegg et al., 2007) or a short reinforcement bar
(approx. 10 cm long) can be embedded in mortar
4.2 Electrolytic resistivity with a high chloride content. Thus, the macro-cell is
The concrete resistance can be measured easily with anodically active and an anodic current is flowing.
the aid of a multi-ring electrode (MRE). The multi- Actively corroding reinforcements will show posi­
ring electrode is used to measure the electrolytic resist­ tive currents, while passive reinforcements will show
ance in concrete as a function of depth. It consists of negative currents (Alonso et al., 1998). The galvanic
eight stainless steel rings arranged at an axial distance activity can be expected to be dependent on the ambi­
of 5 mm from each other and a temperature sensor ent temperature. In this paper, however, the objective
(measurement at the level of the innermost ring). was to investigate, how the galvanic activity of
The measurement data on reinforcement level was a rebar-assembly would respond to applied repair
used to compile the corrosion indicator for the electro­ measures as a monitoring tool to quantify the
lytic resistivity. For the investigated overpass the mean effectiveness.
value of the concrete cover is about 30 mm. Therefore,
an average value of those three rings, which were
close to the rebar was taken to calculate a more realis­ 5 REPAIR MEASURES
tic moisture behaviour, as well as conductivity of the
environment surrounding the reinforcement. The determination of the effectiveness and the dur­
ability of the repair measures presented in this paper
is done on a real structure, where the environmental
4.3 Corrosion potential impact has also been monitored on columns at differ­
The principle of measuring the electrical reinforce­ ent heights.
ment potential is based on the different electrode The structure is an overpass which was planned in
potentials of the corrosive and the passive reinforce­ 1980, has a length of about 100 m and shows a tunnel-
ment. Potential differences of up to several 100 mV like characteristic. The beams are supported by the
can be expected (Elsener et al., 2003). abutment and the columns are situated between the dir­
To measure a potential, a reference electrode with ection lanes of the motorway. In 1994, concrete spall­
a known and constant potential is required. The man­ ing was patched and the 21 columns were treated with
ganese dioxide reference electrodes (ERE 20) fulfil an epoxy-based coating up to a height of 2.0 m. The
these requirements and so they were used in this substructure, especially the columns, showed severe
monitoring system. Usually manganese dioxide ref­ damage of the concrete structure at the time of investi­
erence electrodes have a potential which is 87 mV gation in 2010, such as spalling and cracks over
more positive (MnO2 +445 mV NHE) than the 0.3 mm width as well as traces of rust in the contact
copper/copper sulphate electrode (Cu/CuSO4 +318 and splash zones. Regions of the substructure were
mV NHE) (Krebs, 2019). This difference is import­ repaired in 2014 by patching and additionally protected
by hydrophobic treatment, others by coating.

147
6 PERFOMANCE OF REPAIR MEASURES

To obtain and measure the effectiveness of the repair

measure, the change of the Corrosion Damage Index
was determined before and after the intervention. In
order to obtain valid statements about the effective­
ness, all values prior to the intervention can be com­
pared with the values taken shortly (one year) after
the repair measure. Different statistical techniques, i.e.
Boxplot and Distribution plots, are used to analyse the
effect of each treatment on each column. In addition,
evaluations were made that summarize all columns
treated with hydrophobic agents, for example.
To obtain an objective value for the durability it is
considered to observe the gradient of the index over
time of a repair measure.

7 RESULTS AND DISCUSSION

Figures 2a and 2b show the course of the monitored

sensor data over the complete monitoring period
including compiled CDI in two exemplary columns,
which are treated with hydrophobic agents and coat­
ing, respectively. Figure 2a shows the measured
values of a column whose concrete surface was
sealed with a coating after intervention, whereas
Figure 2b shows the results from a hydrophobic
treatment. The time of the intervention is highlighted
in the figures as a grey shadowed area. Since inter­
vention took place a significant improvement in the Figure 2a. Sensor data over complete monitoring period
corrosion damage indicators occurred. including compiled CDI for column # 09. Figure 2b. Sensor
Hydrophobic treatment achieved on average data over complete monitoring period including compiled
CDI for column # 10.
a reduction of the corrosion damage index by 0.69.
However, coating reduced the CDI by 0.42. (See
Figure 3 in general and 4a 4b for individual struc­
tural components, respectively.)
However, both treatments led to a significant
reduction of the CDI. This is clearly visualized in
Figures 5a and 5b. After the interventions (repair
measures) took place, the variances of the observed
corrosion rates became smaller. This effect is clearly
shown in Figure 3.
This phenomena was also observed on corrosion
potential (Broomfield, 1992) and can be explained as Figure 3. Boxplots for measuring the effectiveness of the
a more controlled and less active degradation pro­ applied repair measures before and after intervention.
cess. The treatments obviously had a corrosion rate
controlling influence which is especially shown by
the statistical analyses of the measurements. An add­
itional observation was, that the corrosion current
density is increasing and decreasing with
temperature.
The gradient calculated as a ratio to assess the
durability of the repair measures, shows a decreasing
trend after the intervention. Nevertheless, four years
after intervention the gradient turns to rise, which
can be interpreted as an increasing degradation. It
seems, that the initiation of corrosion of the repaired
areas was passed. This observation is totally inde­ Figure 4a. Histogram of CDI values pre and post interven­
pendent of the reached CDI after repair. For better tion of column #09. Figure 4b. Histogram of CDI values pre
understanding and interpreting the course of the and post intervention of column #10.

148
 The CDI fluctuates considerable due to the envir­
onmental impacts like humidity and temperature.
However, this apparently high scatter diminishes
when the yearly average were chosen. The CDI also
shows that independent of the CDI-level before
rehabilitation the improvement is considerable.
The data also shows that the corrosion is depend­
ent on several influences. On the one hand there are
the environmental influences, but on the other hand
there are initially measured corrosion intensity, as
well as the electrical resistance. Further research has
to be done to establish a solid performance index
that can objectively assess the effectiveness of repair
measures.

REFERENCES
ALONSO, C., ANDRADE, C., NÓVOA, X. R.,
Figure 5a. Course of the CDI for each corrosion indicator IZQUIERDO, M. & PÉREZ, M. C. 1998. Effect of pro­
and gradient for column #09. Figure 5b. Course of the CDI tective oxide scales in the macrogalvanic behaviour of
for each corrosion indicator and gradient for column #10. concrete reinforcements. Corrosion Science, 40,
1379–1389.
ANDRADE, C. & ALONSO, C. 1996. Corrosion rate
monitoring in the laboratory and on-site. Construction
averaged CDI was plotted against the gradient (see and Building Materials, 10, 315–328.
Figure 5a and 5b). ANDRADE, C. & MARTINEZ, I. 2009. Use of indices to
assess the performance of existing and repaired concrete
structures. Construction and Building Materials, 23,
3012–3019.
8 CONCLUSION B4706, ÖN. 2003. Concrete buildings - repair works,
reconstruction and strengthening. Ausg.: 2003-08-01,
The paper deals with corrosion conditions and ed.: Wien: Österr. Normungsinst.
degradation behaviour of a real concrete structure BAUMANN, J., ÖBV, AUSTRIAN SOCIETY FOR
before and after repair efforts. Measurements of CONSTRUCTION, T. 2014. Erhaltung und Instandset­
an in-situ monitoring system were used to compile zung von Bauten aus Beton und Stahlbeton: Richtlinie,
corrosion indicators. Literature data was chosen to Wien - Book, ÖBV.
BROOMFIELD, J. P. 1992. Field Survey of Cathodic Pro­
classify the individual indicators relevant for the
tection on North American Bridges. MATERIALS PER­
corrosion activity and long-term corrosion behav­ FORMANCE, 31, 28.
iour. Three different types of sensors were BROOMFIELD, J. P., DAVIES, K. & HLADKY, K. 2002.
installed directly on corroding spots in the struc­ The use of permanent corrosion monitoring in new and
ture. The three corrosion indicators are corrosion existing reinforced concrete structures. Cement and
potential, electrolytic resistivity and corrosion rate. Concrete Composites, 24, 27–34.
However, installation of sensors for determining CHRISP, T. M., MCCARTER, W. J., STARRS, G.,
the corrosion rate offers an adequate possibility to BASHEER, P. A. M. & BLEWETT, J. 2002. Depth-
measure the corrosion activity of repaired compo­ related variation in conductivity to study cover-zone
concrete during wetting and drying. Cement and Con­
nents. The developed assessment model - the
crete Composites, 24, 415–426.
introduced corrosion damage index (CDI) – allows ELSENER, B., ANDRADE, C., GULIKERS, J.,
on one hand to assess the structures condition, but POLDER, R. & RAUPACH, M. 2003. Half-cell poten­
also to derive statements about the performance of tial measurements—Potential mapping on reinforced
maintenance measures. Despite the wisely chosen concrete structures. Materials and Structures, 36,
monitoring point on a hot-spot in the structure, it 461–471.
is hard to determine a holistic behaviour of the FURUYA, K., YOKOTA, H., HASHIMOTO, K. &
degradation of the treatment and of the structural HANADA, S. 2011. Reliability of Deterioration Predic­
component. Both repair measures, coating applica­ tion Based on Markov Model for Mooring Facilities.
J. JSCE, Ser. F4 Journal of Japan Society of Civil
tion and hydrophobic treatment were reducing the
Engineers, Ser. F4 (Construction and Management), 67,
corrosion damage indicator significantly. The evo­ I_159-I_168.
lution of the corrosion potential also indicated an HILLEMEIER, B. & TAFFE, A. 2012. Aktuelle Regel­
ennoblement of the reinforcement. The efficiency werke der Bauwerksdiagnostik. Bauphysik-Kalender
of the intervention for the column with hydropho­ 2012, 57–101.
bic treatment was more pronounced as for the RODRIGEZ, J. & ANDRADE, C. 2008. CONTECVET
coated column. A validated Users Manual fo r assessing the residual

149
 service life of concrete structures. In: practical consequences. Construction and Building
INSTITUTE, G. A. T. (ed.) Manual for assessing corro­ Materials, 10, 329–338.
sion-affected concrete structures. Madrid, Spain. RAUPACH, M. & BÜTTNER, T. F. 2014. Concrete repair
KREBS, T. 2019. Permanent reference electrodes - Manga­ to EN 1504: diagnosis, design, principles, and practice,
nese dioxide (MnO2). In: GMBH, G. G. C. P. (ed.). Boca Raton, CRC Press.
45307 Essen. SCHICKERT, G., KRAUSE, M. & WIGGENHAUSER, H.
KÜCHLER, M. 2013. Instandsetzung von Betontragwerken. 1991. Studie zur Anwendung zerstörungsfreier Prüfver­
Beton-Kalender 2013. fahren bei Ingenieurbauwerken: ZfPBau-Kompendium,
POLDER, R., ANDRADE, C., ELSENER, B., Bremerhaven, Wirtschaftsverl. NW, Verl. für Neue Wiss.
VENNESLAND, Ø., GULIKERS, J., WEIDERT, R. & SCHIEGG, Y., AUDERGON, L., BÖHNI, H. &
RAUPACH, M. 2000. Test methods for on site measure­ ELSENER, B. 2007. 11. - On-line monitoring of corro­
ment of resistivity of concrete. Materials and Structures, sion in reinforced concrete structures. Elsevier Ltd.
33, 603–611. SCHNECK, U. Non-destructive corrosion surveys:
R. WENDNER, S. H., A. STRAUSS, K. BERGMEISTER Methods and opportunities. 3rd International Sympo­
& R. GEIER 2008. Inspection of bridges in Austria— sium on Life-Cycle Civil Engineering, IALCCE 2012,
Practice and outlook. Life-cycle civil engineering, 2012 2012 Vienna. 559–566.
703–708. VERKEHR, Ö. F. S.-S.-. 2011. Überwachung, Kontrolle
RAUPACH, M. 1996. Chloride-induced macrocell corro­ und Prüfung von Kunstbauten. RVS 13.
sion of steel in concrete—theoretical background and 03.11Straßenbrücken. Vienna.

150
 Bridge Maintenance, Safety, Management, Life-Cycle Sustainability and Innovations –
Yokota & Frangopol (eds)
© 2021 Taylor & Francis Group, London, ISBN 978-0-367-23278-8

Research on long-term health monitoring and operation evaluation system

for long-span self-anchored suspension bridge
Xueshan Liu
Tongji University, Shanghai, China

Xiaohu Chen
T.Y. Lin International Engineering Consulting (China) Co., Ltd, Chongqing, China

Jianting Zhou & Xiaogang Li

Chongqing Jiaotong University, Chongqing, China

ABSTRACT: Self-anchored suspension bridges have been increasingly popular in the engineering and
academic field, but there is only few researches on their health monitoring and evaluation system in
operation stage. In the case of Egongyan No. 2 Yangtze River Bridge in Chongqing, the bridge with
largest span of its kind, a set of intelligent health monitoring system was established in the paper ­
based on cloud platform, intellectualization and internet of things technologies. It forms a monitoring
network by collecting data, deploying real-time load, displacement, stress-strain, acceleration and other
mechanical sensors, and transmitting sensor information to the cloud platform for analysis and process­
ing. The system realizes real-time monitoring, early warning, analysis, evaluation, safety warning and
other functions to ensure bridge safety. The research on long-term health monitoring and operation
evaluation system is of great significance for infrastructure and urban traffic safety, it also provides
a reference for similar bridges.

1 INTRODUCTION and bridge construction technologies constantly get

mature, this bridge type will become more competi-
A suspension bridge is a flexible suspension com- tive in bridge construction and witness significant
posite bridge structure with the main cable as the development.
main load-bearing component, and its main compo- Bridge health monitoring technologies have been
nents include main cable, main tower, saddle, sling, applied on a number of important bridges for long-
and stiffening girder. Suspension bridges can be div- term safety of bridge in operation stage, and have
ided into anchored suspension bridges and self- realized the monitoring and evaluation of bridge
anchored suspension bridges. For self-anchored sus- health [1], especially that the application in the large-
pension bridges, the anchoring position of the main scaled anchored suspension bridges of Runyang
cable is transferred to the stiffening girder end from Bridge [2] and Ma’anshan Yangtze River Bridge [3]
the anchor, which can not only reserve the attractive has played a role of vital importance in disaster pre­
appearance of a suspension bridge but also save vention and reduction for long-span suspension
anchors with huge construction volume. This bridges. For the development and construction of
method is practical to build suspension bridges long-span self-anchored suspension bridges, Liu Jian
under adverse landform or geological conditions and [4], et al. carried out bridge health monitoring design
has greatly enriched the application scope of suspen- for Foshan Pingsheng Bridge, and Li Qiuyuan [5],
sion bridges. However, self-anchored suspension et al. carried out bridge health monitoring design for
bridges are more complicated than traditional Yinchuan Binhe Yellow River Bridge. It can be seen
anchored suspension bridges in terms of structure from practices in these bridges that the existing bridge
stress, construction control, and health monitoring. health monitoring technologies have three basic char-
As design concepts get consistently completed, acteristics, namely, overall perception, transmission,
engineering materials get innovated continuously, and intelligent processing.

DOI: 10.1201/9780429279119-16
https://doi.org/10.1201/9780429279119-16

151
 Chongqing Egongyan No. 2 Yangtze River 2 STRUCTURE & STRESS
Bridge, also called Chongqing Egongyan Rail CHARACTERISTICS OF LONG-SPAN SELF­
Bridge, is located about 45m from the upstream of ANCHORED SUSPENSION BRIDGE
the Egongyan Yangtze River Bridge as shown in
Figure 1, The bridge applies the bridge scheme laid Structure characteristics of self-anchored suspension
out corresponding to the existing bridge as shown in bridges determine the stress characteristics of the
Figure 2. Since the current bridge is an anchored sus­ bridge. Seen from appearance, the sole difference
pension bridge with larger anchor, a self-anchored between self-anchored suspension bridges and
suspension bridge scheme is applied on the new anchored suspension bridges lies in anchoring mode.
bridge to avoid influence on structure of the old Therefore, self-anchored suspension bridges have
bridge. The span combination of Chongqing Egong­ stress characteristics similar to anchored suspension
yan No. 2 Yangtze River Bridge is 50+210+600+210 bridges. However, the movement of the main cable
+50=1,120m, with 5 spans in total. The 600m main anchoring position to the girder end gains its stress
span of the bridge will be the largest self-anchored characteristics of cable-stayed bridges.
suspension bridge in the world after its completion. As a cable composite structure, self-anchored sus­
The bridge deck is 22m wide. The main beam is of pension bridges have the following stress characteris­
steel beam-concrete beam composite structure, and tics: (1) The main cable is the main load-bearing
both the main pier and the main tower are of component of a long-span self-anchored suspension
reinforced concrete structure. bridge, and cannot be replaced. The main cable sees
Chongqing Egongyan No. 2 Yangtze River Bridge not only elastic deformation but also changes of geo­
has features of large span and complex stress [6], metrical shape under load. Therefore, it is necessary
[7]. Meanwhile, the bridge applies many new tech­ to consider internal force change caused by a change
nologies and new materials, which does not only of geometrical shape of the main cable while analyz­
bring great difficulties to bridge construction but ing structures. (2) The main cable of a self-anchored
propose stricter requirements for future operation suspension bridge is directly anchored to the stiffen­
and maintenance management. The bridge needs to ing girder end. Stiffening girder forms a self-
deal with railway vehicles on daily basis after open balancing system under the tremendous axial pressure
to traffic. Maintenance and repair must be completed conveyed by the main cable and has to have enough
within six hours at night. Therefore, it is quite neces­ section area to resist the huge axial pressure and
sary to conduct preventive health monitoring and enough flexural stiffness to ensure its stability. (3)
operation evaluation to the bridge. The anchoring system of the main cable at the stiffen­
ing girder end constitutes the key and difficult point
of the self-anchored suspension bridge structure, and
the shift of the huge cable force to stiffening girder
from the structure makes it capable of adequately
anchoring the main cable force. Meanwhile, in the
case of anchoring a tremendous cable force to the
girder end, it is necessary to realize beam splitting
within a limited space to ensure anchoring reliability
and safety of the main beam. (4) Generally,
a deadweight of the side-span main beam of a long-
span self-anchored suspension bridge is not enough to
provide upward force pulling the main cable, so it is
necessary to set tension-compression bearings or con­
duct ballasting to the main beam. These structural
measures make the stress state of side span more
Figure 1. Chongqing Egongyan No. 2 Yangtze River
complicated. (5) The top of the main tower bears the
bridge. concentrated force of the saddle, bringing enormous
axial pressure to the cable bent tower. Meanwhile, the
main tower is the most important landscape control
element of a self-anchored suspension bridge, which
generally takes landscape as the design target, besides
meeting the basic stress performances.

3 OVERALL TECHNICAL ROUTE FOR

HEALTH MONITORING

The general aim of the bridge health monitoring

Figure 2. Elevation of Chongqing Egongyan No. 2 Yangtze system is to monitor and evaluate bridge conditions,
River bridge. replace partial manual operations, comprehensively

152
improve information fusion and utilization and pro­ Combining structural forms of self-anchored sus­
mote the conversion of bridge repair mode to fore­ pension bridge, and taking environment factor, load,
seeable repair, centered on the application of bridge structural response, dynamic characteristics and others
operation and maintenance. Characteristics of the as the main monitoring categories, specific monitoring
system are as follows: (1) Complementary mechan­ contents include: (1) Environment factor monitoring:
ism of automatic real-time monitoring and manual wind velocity, wind direction, temperature, humidity,
detection: conduct complementary analysis of the precipitation, etc.; (2) Load monitoring: wind load,
long-term real-time structural response monitoring ship collision load, temperature load, etc.; (3) Struc­
information and the regular manual detection infor­ tural response monitoring: spatial deformation of main
mation, to improve the accuracy of bridge safety beam, foundation settlement, tower top deviation,
analysis; (2) Comprehensive evaluation and analysis anchor beam displacement, strain-stress of key sec­
of a number of reliable and effective bridge safety tion, main cable strand force, sling force, etc.; (4)
evaluation theories and technologies: the positioning Dynamic characteristic monitoring: for Egongyan Rail
of the long-term health monitoring system in the Bridge, the dual-tower, dual-cable self-anchored sus­
operation stage is the structural state evaluation and pension bridge with the largest span in the world, the
safety evaluation guided by structural maintenance. dynamic characteristics of the structure of the rail
In order to improve the accuracy of bridge safety bridge in operation are relatively remarkable due to
evaluation, three kinds of structural safety evaluation larger excitation of train load and larger train struc­
systems based on time-variant reliability, envelope tural plane. Therefore, in terms of monitoring of
theory, and statistical analysis are proposed; (3) dynamic characteristics, plenty of response monitoring
Modular design: reduce system complexity to facili­ sensors for acceleration and other mechanical powers
tate system maintenance and upgrading; (4) Human­ are added. The general layout of bridge health moni­
ized design: include system operation, human toring points is shown in Figure 3.
machine interface and comfort evaluation; (5) According to bridge calculation report, the layout
Redundancy design: ensure system reliability, and principle of monitoring points is: (1) Based on reason­
meet the needs of improvement, expansion and per­ able bridge monitoring parameter analysis principle:
fection in the future; (6) Open design: Meet needs of confirm the parameters which should be and may be
system upgrading and updating. monitored, and stress of main beam control section,
main tower deviation, sling and main cable strand
force and environment factor. Propose reasonable
4 HEALTH MONITORING CONTENTS monitoring points based on a comprehensive analysis
of monitoring parameters necessary for the bridge. (2)
The basic content of the health monitoring system is Principle of scientific analysis of bridge internal force
to send out early warning signals through long-term and deflection: monitoring points of the bridge moni­
real-time monitoring and evaluation of bridge struc­ toring system should be laid out at the point with the
ture state under special weather and traffic condition worst internal force and deflection obtained from ana­
or in case of serious abnormal operation conditions, lysis in the whole bridge operation period, namely, the
providing a basis to make decisions on bridge main­ position where the bridge is most likely to fail. (3)
tenance, repair and management. Principle of point distribution based on bridge defect

Figure 3. General layout of bridge health monitoring points.

153
Figure 4. Layout of different sensors.

investigation: during the operation of Chongqing

Egongyan No. 2 Yangtze River Bridge, monitoring
points are laid out in positions where defects likely
occur to the bridge based on investigation into
common bridge defects and previous engineering
experience, aimed to realize safety warning and fore­
casting by monitoring internal force and deformation
sate in positions where bridge defects easily occur. (4)
Principle of geometric size coverage: appropriately
consider the principle of geometric size coverage based
on the above monitoring parameter principle, internal Figure 5. Demonstration of sensor installation.
force analysis principle, and defect investigation prin­
ciple. Therefore, the above analysis is conducted under
the same construction conditions of the bridge. How­
ever, there exist more substantial differences since the
construction conditions of all bridge parts are different
actually. Therefore, monitoring points should cover the
bridge as much as possible in the geometric size range
to “capture” the worst positions of the bridge. (5) Sym­
metry principle: use the symmetry of the structure to
reduce the layout of the sensors. The layout of different
sensors are illustrated in Figure 4. There are 144 smart
string strain sensors, 7 Leica GNSS (Global Navigation
Satellite System) receivers, 12 laser joint meters, 26
hollow rope meters, and 30 acceleration sensors on the
whole bridge. See the sensor installation example in
Figure 5 and the list of sensors in Figure 6. To obtain
the service state of the main cable and sling more
accurately, the cables will be monitored with the rope
meter in a long time, while the cable force dynamic
meter will be used for regular detection with the fre­
Figure 6. Part of equipment used for monitoring.
quency method.

154
5 FUNCTION DESIGN OF HEALTH state monitoring and early warning, system state
MONITORING SYSTEM monitoring and early warning, and system remote
control. (4) Module 4: The structural health evalu­
A relatively independent intelligent health monitoring ation system is a high-performance computer system,
system is adopted in Chongqing Egongyan No. 2 which is the core of the whole system. It is mainly
Yangtze River Bridge. Through the non-destructive used to analyze the original monitoring data, evalu­
monitoring of the physical and mechanical properties ate, diagnose, and predict the structure state. (5)
of the bridge, the system can monitor the whole Module 5: The structural health data management
behavior of the bridge in real-time, integrate it into system consists of a high-performance computer
the cloud platform to diagnose the position and extent system and a relational data system, and it is mainly
of the damage of the bridge, and analyze the service used for access to monitoring data and analysis
condition, reliability, durability and bearing capacity results. (6) Module 6: The inspection and mainten­
of the bridge, so that the bridge can trigger the early ance system consists of a toolbox, and it is used to
warning signal in case of emergency or abnormal inspect and maintain sensors, data acquisition units,
severe service condition. The system can provide data transmission networks, display devices, and etc.
basis and guidance for repair, maintenance, manage­
ment decision-making of the bridge, which plays an
essential role in bridge health and rail transit safety. 6 OPERATION STATE EVALUATION
The bridge health monitoring system adopts
modular design, which can reduce the complexity of Chongqing Egongyan No. 2 Yangtze River Bridge
the system, facilitate system maintenance and focuses on the monitoring of bridge displacement,
upgrading, and each module can work independently, dynamic response and beam end reactor device,
so the failure of one module will not affect the regu­ which can reflect the safety and adaptability of vehicle
lar operation of other modules, reflecting the soft­ operation. Through the development of automatic ana­
ware development idea of “high integration and high lysis software, the original data can be purified and
resiliency”; The system is mainly composed of the problem of massive data processing and analysis
sensor subsystem, data acquisition and transmission can be solved. The monitoring system is based on the
subsystem, data processing subsystem, data storage application of management and maintenance person­
and management subsystem, analysis and early warn­ nel, and it can judge the state through comprehensive
ing subsystem, inspection and maintenance subsys­ warning and evaluation. The monitoring system can
tem. Optical fiber is used for communication judge the structure state comprehensively by incorpor­
between modules to ensure uninterrupted long- ating the results of manual patrol inspection.
distance data transmission. Figure 7 is the structure During the construction of Chongqing Egongyan
diagram of the health monitoring of Egongyan No. 2 No. 2 Yangtze River Bridge, the smart string strain
Yangtze River Bridge. (1) Module 1: The sensor sensors will be arranged in the main beam, which
system consists of various sensors used to test the can be connected to the acquisition module to test
physical parameters of the structure and its surround­ the stress under the dead load and live load response.
ing environment. (2) Module 2: Data acquisition and Before the sling is tensioned, the hollow rope meter
transmission system consisted of two kinds of equip­ should be installed, and the cable force test should
ment, namely, data acquisition unit and data trans­ be conducted during the construction and operation.
mission network. They are used for signal After the completion of the structural construction,
acquisition, processing, caching and transmission. (3) the Leica GNSS deformation monitoring system
Module 3: Data processing and control system is should be installed in the middle of the main span of
a high-performance computer system, which is used the main beam, with the reference point set at the
for routine data processing and analysis, structural abutment position, which mainly monitors the longi­
tudinal, transverse and vertical deformation of the
key sections of the bridge deck as well as the dis­
placement of the main tower; laser joint meters
should be installed at anchor beam and cable saddle
to measure displacement changes.
The monitoring information of the bridge should
be collected in real-time with 02:00 a.m. as the
benchmark time before the railway train passes
every day. The main beam stress, main cable force
and suspender cable force should be automatically
collected once every 10 min., and the main beam
and main tower deformation should be automatically
collected once every 10 sec. After the calculation
and analysis by the program, they should be included
Figure 7. Structure diagram of the health monitoring of
Egongyan No. 2 Yangtze River bridge. in the evaluation system to evaluate the rigidity,

155
strength and cable force and make maintenance deci­
sions and vehicle traffic permits.
The positioning of the long-term health monitor­
ing system in the operation stage is the structural
state evaluation and safety evaluation guided by
structural maintenance. To improve the accuracy of
the bridge safety evaluation, three kinds of structural
safety evaluation systems based on time-variant reli­
ability, envelope theory, and statistical analysis are
proposed. The automatic computer evaluation of the
long-term health monitoring system can be divided
into three situations: (1) Normal state monitoring of Figure 8. The interface of Egongyan No. 2 Yangtze River
the structure: when the monitoring value is less than Bridge’s Health Monitoring System.
75% of the allowable value of normal use state, it
belongs to the normal state monitoring stage of the
structure. In this stage, the main work is to under­
regular manual detection information should be
stand the working mechanism of the structure, iden­
made for the management and maintenance of
tify the working mode of the structure and its
Chongqing Egongyan No. 2 Yangtze River Bridge in
statistical analysis, and the analysis of the original
operation period, so as to realize the complementary
monitoring data is mainly routine data processing
mechanism of automatic real-time monitoring and
and analysis. (2) Critical state monitoring of the
manual detection, and improve the accuracy of
structure: when the monitoring value is greater than
bridge safety analysis.
75% of the allowable value of the normal use state
To improve the efficiency and friendliness of use,
but less than the allowable value of the normal use
a visual interface as demonstrated in Figure 8, is devel­
state, it belongs to the critical state monitoring stage
oped in this system to analyze and evaluate the trans­
of the structure. In this stage, the main work is to
mitted data, display the state of the bridge on-site in
evaluate the structural state, check the deformation
real-time through the back-end server, and intuitively
and stress overrun of other unmonitored compo­
feed it back to the staff. According to the state of the
nents, combined with the monitoring data and the
automatic bridge, the bridge can be divided into three
structural finite element model. If any overrun is
states by this system: normal state, early warning state,
found, it should instruct the structural maintenance
and degradation state. Based on the state provided by
personnel to check the problem parts/components in
this system, combined with the results of manual
detail. (3) Evaluation of structural degradation state:
evaluation, this system can take management actions
when the monitoring value is greater than the allow­
such as vehicle speed reduction, traffic restriction or
able value of normal use state, it belongs to the stage
bridge maintenance and reinforcement.
of structural degradation evaluation. In this stage,
the main work is to identify structural damage and
evaluate damage state, and systematically remind to 7 CONCLUSION
check whether there is any damage to the structure
and the adverse effect of damage on the overall per­ According to the stress characteristics of the long-
formance of the structure combined with the moni­ span self-anchored suspension bridge, this paper
toring data and structural finite element model. takes Chongqing Egongyan No. 2 Yangtze River
It is an important part of the health monitoring Bridge as an example, systematically and compre­
system based on the test data of the system to evalu­ hensively studies the design method of the health
ate the operation state of the bridge structure reason­ monitoring system of the long-span self-anchored
ably. Due to the limitations of the scale of the health suspension bridge, and the main conclusions are as
monitoring system, sensor layout and information follows: (1) The design of the health monitoring
transmission, it is not complete to evaluate Chong­ system for Chongqing Egongyan No. 2 Yangtze
qing Egongyan No. 2 Yangtze River Bridge only River Bridge is systematically described, including
based on the data collected by this health monitoring the overall layout, the design of each subsystem, and
system. At the same time, due to the lack of know­ the specific implementation of each functional
ledge about and experience in the response of cable module. (2) The monitoring items of the sensor sub­
type track bridge structure under the complex envir­ system and the general layout of the whole bridge
onment and trainload, it is difficult to give an accur­ sensor monitoring points are introduced in detail. In
ate and effective early warning mode, while the the general layout plan, the monitoring method of
method of manual inspection has some limitations. cable force is different from the previous monitoring
Therefore, the traditional structural inspection and system, that is, two kinds of equipment, rope meter
maintenance measures and advanced health monitor­ and cable force dynamic meter, are used for the
ing system should be organically integrated, and the monitoring points layout, making full use of the
complementary analysis of the long-term real-time advantages of the two types of equipment. (3) The
structural response monitoring information and the structure of the early warning and safety evaluation
156
subsystem introduces the main contents of special span rail bridge state intelligent perception cloud plat­
safety evaluation in detail. Combined with the exist­ form” (csccs20180601113) [8] of Chongqing Science
ing evaluation methods and the characteristics of the and Technology Commission. Thanks.
monitoring system, it proposes the evaluation
methods and indicators suitable for Chongqing
Egongyan No. 2 Yangtze River Bridge. (4) This REFERENCES
system adopts the way of visual interface, which can
analyze and evaluate the transmitted data, display Zhou Jianting, Zhou Zhixiang, Xu Mou, et al. 2006.
the bridge state in real-time through the back-end Research on the Remote Health Monitoring and Evalu­
server, and realize the functions of mass data storage ation System of Large Bridges. The Thesis Album of
55th Anniversary of Chongqing Jiaotong University.
and automatic early warning. Practical application
Ou Qingbao, Shen Gang, et al. 2015. Research on Structural
shows that this system can quickly and stably obtain Health Monitoring Strategy of Runyang Bridge. Municipal
bridge health monitoring data, improve work effi­ Engineering Technology Supplement. January 2015.
ciency, and reduce monitoring costs. Zheng Weifeng, Zheng Gang, et al. 2018. Research on
Structural Health Monitoring System of Ma’anshan
Yangtze River Highway Bridge. Journal of Highway
ACKNOWLEDGMENT and Transportation Research and Development 5.
Liu Jian, Liu Ding, et al. 2011. Optimized Layout of Accel­
Based on the principle of “synchronous design, syn­ eration Sensor Monitoring Points of Long-term Health
chronous construction and synchronous operation”, Monitoring System of Pingsheng Bridge. Journal of
the main construction and health monitoring system China & Foreign Highway. June 2011.
Li Qiuyuan, Wang Jian, et al. 2015. Design of Automatic
of Chongqing Egongyan No. 2 Yangtze River Bridge
Sensor Test System for Health Monitoring of Main
have been gradually realized. The main structure of Bridge of Yinchuan Binhe Yellow River Bridge. Low
the bridge was completed in December 2018. The Carbon World. November 2015.
bridge is expected to be officially opened in Decem­ Shen Ruili, Cheng Xin, et al. Study on Static Ultimate
ber 2019, and the health monitoring system will be Bearing Capacity and Safety Evaluation Method of
operated synchronously. In the next stage, the system Self-anchored Suspension Bridges. Journal of the China
will be further optimized to equip better applicability, Railway Society. November 2017.
safety and stability according to the long-term oper­ Duan Li, Li Yuansong, et al. Analysis of Dynamic Charac­
ation requirements of the system. It is hoped that this teristics of Egongyan Bridge. Modern Transportation
Technology. August 2019.
system can provide reference for the design and oper­
Chen Xiaohu. Key Technology and Application Demon­
ation of similar long-span self-anchored suspension stration Implementation Scheme of Intelligent Percep­
bridges. This system and this paper are supported by tion “Cloud Platform” Construction of Long-span Track
the scientific research project “key technology and Bridge across Two Rivers. Chongqing Science and
application demonstration of the construction of long- Technology Commission. March 2018.

157
 Bridge Maintenance, Safety, Management, Life-Cycle Sustainability and Innovations –
Yokota & Frangopol (eds)
© 2021 Taylor & Francis Group, London, ISBN 978-0-367-23278-8

Unmanned aerial vehicle (UAV)-enabled bridge inspection framework

B.J. Perry, Y. Guo, R. Atadero & J.W. van de Lindt
Department of Civil and Environmental Engineering, Colorado State University, Fort Collins, Colorado, USA

ABSTRACT: Bridge inspections of medium or large scale are often cumbersome, expensive, and time-
consuming. With a large number of bridges in the United States that require at least a bi-yearly inspection,
there is a need to improve bridge inspection techniques to save time and reduce costs. Unmanned Aerial
Vehicles (UAVs) have made tremendous advancements in recent years to allow for better data collection with
enhanced sensors, more controllability with precise global-positioning-systems and inertial measurement units,
and increased safety with omnidirectional sensors to avoid collisions while becoming more affordable. In cur­
rent bridge inspection practice, UAVs have been used as “eyes-in-the-sky,” simply assisting inspectors to view
bridges or other structures from different vantage points with the inspectors still taking measurements and
making decisions with the traditional techniques. However, to take full advantage of the UAV’s capabilities
and allow for the UAV to perform and quantitate inspections automatically to create a more streamlined work-
flow, there is a need for more robust data processing of the information attained by the UAV. A streamlined
decision-making support framework is proposed that uniquely integrates UAV-based field inspection, auto­
mated damage identification, and establishment of an element-wise As-Built Bridge Information Model (AB-
BrIM) for the damage documentation. In this framework, a UAV platform with optical sensors first collects the
data. Next, an automated damage detection algorithm that highlights cracks and spalling is developed to
quickly extract quantitative information (i.e. type, size, amount, and location). Finally, a 3-D point cloud is
created with photogrammetry and then segmented into identified structural elements (e.g. beam, girders, deck,
etc.) to serve as a base for the AB-BrIM. The identified damage information is automatically linked to each
element. The resulting AB-BrIM with 3-D visualization of element-wise, quantitative damage information
offers a transparent condition evaluation and thus can greatly ease the planning of repair/maintenance.

1 INTRODUCTION ability (Ryan, Mann, Chill, & Ott 2012, Omar &
Nehdi 2017, Moore, Phares, Graybeal, Rolander, &
To measure, track, and manage the aging and deterior­ Washer 2001). For instance, Moore et al. demonstrated
ation of the bridge infrastructure, bridges in the United the variance and subjectivity of routine bi-yearly
States are inspected at least once every two years. With inspections. They had multiple crews perform a routine
over half a million bridges in the United States, the inspection on the same structures. It was found that
routine inspection creates a burden for state and federal 68% of the ratings showed �1 point variation and
officials (Ryan, Mann, Chill, & Ott 2012). In addition 27% of ratings had �2 points variation; while the
to the burden of a large number of inspections, human- remaining 5% of ratings had a variation of greater
based bridge inspections also tend to be more subject­ than �2 (Moore, Phares, Graybeal, Rolander, &
ive and costly. The current bridge inspection practices Washer 2001). This indicates an inconsistency in rating
in the United States are standardized by The American data. It is also worth noting that this inconsistent rating
Association of State Highway Transportation Officials’ is even more susceptible to structurally deficient
(AASHTO) Manual for Bridge Evaluation and the bridges, where accurate data is even more critical for
Federal Highway Administration’s (FHWA) Bridge safe operations and planning.
Inspection Reference Manual (Ryan, Mann, Chill, & Other systemic downsides are present with current
Ott 2012, AASHTO 2018). These conventional tech­ techniques regarding safety, time, and costs. To gain
niques typically incorporate visual and sounding access to the underside of these large-scale bridges
methods to qualitatively rate each bridge’s components, during in-depth inspections, roping rigs, scaffolding,
i.e. deck, girders, and abutments. These qualitative rat­ and/or “snooper” trucks are used. These systems
ings tend to be relatively subjective due to their reli­ give great accessibility to the underside of the bridge
ance on human inspectors’ experience and hearing and bring the components “within arm’s reach” to

DOI: 10.1201/9780429279119-17
https://doi.org/10.1201/9780429279119-17

158
the inspector, which is currently required for certain of Transportation has completed a three-phase pro­
types of inspections by AASHTO and FHWA in the ject on utilizing UAVs to assist inspectors on a wide
United States. However, extensive training and variety of bridges in Minnesota in the United States.
safety procedures are required. The use of this equip­ They created photogrammetric models with the use
ment during inspection not only puts the inspectors of Pix4D to log and organize the data collected from
in a dangerous position, due to the inherent risks of the UAVs. They also demonstrated the potential for
the systems and the proximity to vehicular traffic, cost savings of about 60% using a UAV system as
but often requires two-weeks or more for set-up, opposed to traditional techniques (Wells & Lovelace
inspection, and tear-down of the equipment. Add­ 2018). The photogrammetric model created greatly
itional office time is required for the final report cre­ aided inspectors and supplied a wealth of data that
ation. With this long turnaround time, fewer bridges would otherwise not be possible. Despite the benefits
can be inspected per season by an inspector. Also, of safety, cost-effectiveness, and valuable inspection
during this time of on-site inspection, highway lanes, data, these studies did not provide inspectors with an
railways, and/or boat routes may be altered or efficient data processing system that merges UAV-
closed, negatively impacting traffic flow. Lastly, the based inspection data into current practice. To pro­
monetary cost of renting/buying/operating these sys­ vide this streamlined decision-making support,
tems is considerable while the deadweight loss of a holistic inspection system that integrates UAV-
not having the best data to make the most informed based field inspection with element-wise damage
and logical decision is common. These shortcomings identification, documentation, and visualization
of the current practice highlight the need to develop incorporated in an As-Built Bridge Information
a more consistent, efficient, and cost-effective deci­ Model (AB-BrIM) is needed.
sion-making support system while maintaining In response to this need and the gap in research,
a high level of safety for inspectors. Such a system this paper proposes a unique UAV-based bridge
would provide bridge owners and decision-makers inspection framework, which encompasses the entire
with better data to plan accordingly with the limited inspection process from field survey to final AB-
resources available. Recently, Unmanned Aerial BrIM creation. The proposed framework has two
Vehicles (UAVs) based remote sensing technology unique features that are expected to promote future
has emerged as a promising tool to aid in the deci­ technology transfer into current practice: 1) to ensure
sion-making process of transportation assets seamless merging of the new technology with current
management. inspection practice, the data analysis module of UAV-
Recently, UAVs have greatly improved and based inspection system will be constructed to pro­
become more user-friendly and affordable. Multi- vide element-wise damage assessment that is consist­
rotor UAVs can maintain a stable position without ent with current inspection practices within AASHTO
much movement or vibration even in moderate wind and FHWA regulations; and 2) to enable the integra­
conditions and non-preferred environmental condi­ tion of the new inspection system with governmental
tions (Hernandez 2016). Advances in visual naviga­ agencies’ current GIS-based data management
tion systems have allowed UAVs to hover in place system, a geo-referenced AB-BrIM model will be
without the need for a strong Global-Positioning- developed. This will allow convenient and compre­
System (GPS) signal by using optical, infra-red, and/ hensive documentation of condition assessment
or ultrasonic sensors (Zingg, Scaramuzza,Weiss, & results with geo-referenced and element-wise damage
Siegwart 2010, Gageik, Strohmeier, & Montenegro information. The proposed framework proves to be
2013). The price and size of a complete system have safer and more efficient, cost-effective, and consistent
reduced considerably, making the technology more than the current techniques. Inspectors can fly the
accessible for government and research institutions. UAV system at a safe distance from traffic without
Moreover, the camera technology has continuously additional equipment to inspect the underside of the
evolved, with improved image quality, sensor size, bridge. On-site time could be significantly reduced to
resolution, resulting in a better ground sampling a day or less, saving on time, traffic interruptions, and
density (GSD). equipment costs. The use of data analysis algorithms
Several studies have been developed incorporat­ enables quantification of damage, and thus can pro­
ing UAVs to inspect bridges. Hernandez et al. devel­ vide objective results that are reproducible and con­
oped a customized UAV system with tethered data sistent throughout inspections.
relay and power to photograph bridges (Hernandez This UAV-enabled bridge inspection framework is
2016). Gillins et al. also deployed a UAV to inspect comprised of three main modules which allow for
bridges; however, these two studies only used the semi-autonomous data processing, organization, and
obtained photographs as an aid to the inspectors, presentation for bridge inspectors and decision-
providing “eyes in the sky” without locating, quanti­ makers. The proposed framework is illustrated in
fying, or measuring damage to assess bridge condi­ Figure 1. A UAV platform carrying visual sensors
tion (Gillins 2016). Furthermore, Duque et al. used will hover around the structure collecting data
a UAV system to photograph two bridges and create (images) through flight missions. Post-flight, the
a robust defect detection algorithm (Duque, Seo, & images will pass through the data processing modules
Wacker 2018). Similarly, the Minnesota Department to extract the pertinent information. With the collected
159
 Table 1. Comparison of UAV Platforms.

Platform Platform Platform

1 2 3

Built-in
Obstacle None 5­ Omni-
Directions Directional
Detection
Resolution (MP) 12.4 20 12
Sensor (inch) 1/2.3 1 1/2.3
GSD (cm/pixel) 0.214 0.137 0.175
Flight Time (min) 32 30 27
Wingspan (feet) 5.47 1.15 1.05
Figure 1. Flow Chart of Proposed Framework. Price (USD) $5,898 $1,499 $1,249

Platform 1: DJI Matrice 600 Pro with Zenmuse X3 Camera

photos and videos, a 3-D point cloud model will be Platform 2: DJI Phantom 4 Pro V2.0
Platform 3: DJI Mavic 2 Zoom
created from the 2-D image set. This 3-D model will
serve as the base of the AB-BrIM. From the 3D point
cloud, each bridge element (e.g. beam, girders, deck,
etc.) will be identified and highlighted using unsuper­ sensor payload, and price and provides an idea of the
vised clustering techniques. Next, the images will trade-off and benefit of each system. No platform
pass through automated damage/defect detection algo­ proved to be more effective over another; rather,
rithms and surface and/or subsurface defects (e.g. the needs of the inspection would dictate which
cracks, spalling and scaling of concrete) will be iden­ platform(s) should be selected.
tified. The damage detection algorithms will quickly Two test sites were selected to demonstrate the
provide quantitative damage information. The proposed framework. Both bridges were unique and
extracted damage information (i.e. type, amount, and provided different challenges and demonstrated the
location) can be readily mapped to each of the identi­ performance of the proposed inspection framework
fied bridge elements in the AB-BrIM. With the bridge in the field. The “Palmer Bridge” shown in Figure 2
model completely digitized, it can be uploaded to the spans a spillway to a dammed-up creek. It had
cloud and accessed remotely. The resulting AB-BrIM a continuous flow of high-speed water under the
with element-wise damage information and structure with a clearance of about 1-meter. Due to
3-D visualization capability will make problem areas a UAV’s wide wingspan and created thrust, it was
obvious and provide quantitative measures to allow unable to safely fly under the bridge; therefore, pic­
bridge managers/owners to efficiently evaluate exist­ tures of only the top and sides of the bridge were
ing structures while keeping inspectors safe from traf­ taken. The “Oxbow Bridge” in Figure 2 was a newly
fic and on the ground. acquired bridge donated to the city. There was
In the subsequent sections, each module of the a large clearance under the structure; therefore, the
proposed system will be illustrated through a case UAVs were able to safely fly under the structure to
study of two Colorado bridges. First, how the UAV collect upward-angled imagery. Both structures pro­
system collected the data will be introduced. Then, vided a real-world implementation of the proposed
the methodology of each module of the data process­ framework and demonstrated the potential of future
ing tools (i.e. 3-D model creation, defect detection, UAV-enabled bridge inspections. Both flights were
and AB-BrIM establishment) and the corresponding conducted safely and within local and federal regula­
results will be elaborated. Lastly, the concluding tions and were considered successful. The images
remarks and future work will be discussed. and data collected from the “Palmer Bridge” and
“Oxbow Bridge” will be used and analyzed in subse­
quent sections to demonstrate the potential of UAV-
2 DATA COLLECTION enabled bridge inspections.
One major concern for bridge inspections is a loss
Three UAV platforms were tested in this study. All of GPS signal under the structure. However, due to
three platforms showed effectiveness within bridge the built-in obstacle detection sensor in the DJI
inspection and were used in many previous research Phantom 4 Pro V2.0 and the DJI Mavic 2 Zoom, the
studies (Mulakala 2018, Seo, Duque, &Wacker 2018, two UAVs performed well under the structure in the
Gillins, Gillins, & Parrish 2016, Bang, Kim, & Kim GPS-deprived environment and maintained its
2017, Greenwood, Lynch, & Zekkos 2019, Doraf­ course to remain stable during flight. A second con­
shan, Maguire, Hoffer, & Coopmans 2017, Wells & cern in particular for bridge inspection is the per­
Lovelace 2018). The systems were compared on sev­ formance of upward-facing photography under
eral metrics shown in Table 1 including camera, size, a bridge. None of the tested platforms supported

160
Figure 2. North side of the palmer bridge and its corresponding 3-D point cloud.

upward facing sensors; however, they were able to structure-from-motion was used (Moulon, Mon­
angle upwards at least 25� . Although this was not asse, & Marlet 2012). Structure-from-motion finds
ideal for upward imagery, it did provide enough key points and their associated descriptors, in this
upwards angle to capture the underside of the struc­ case through Scale-Invariant Feature Transform
ture and create a 3D point cloud model. (SIFT) to match images (Lowe 1999). Knowing
the focal length, the optical center of the camera,
and these key point matches, the camera poses are
3 METHODOLOGY OF DATA PROCESSING estimated to project pixels into 3-D space creating
TOOLS the 3-D point cloud. The camera pose will also be
used in a later section to map the defects onto the
Following the framework in Figure 1, each module 3-D point cloud. The 3-D point clouds for the two
of the data processing tools in the proposed system test bridges are shown in Figure 2 and Figure 3.
(i.e. 3-D model creation, defect detection, and AB- To create the 3-D point cloud, a certain amount of
BrIM establishment) will be discussed in detail in overlap between the images was needed to ensure
the order of the process. that there were enough matching key points between
the set of images. Chiu et al. and Shahnaz et al.
recommended 50 and 60-percent overlap, respect­
3.1 3-D model creation
ively (Chiu, Ong, Kuen, & Courtney 2017, Shahnaz
With 500+ images obtained from each test site, 2010). When flying the two flights manually, the
a 3-D model could be created using photogram­ overlap was approximately 95%. With future imple­
metry, which is a process of taking a 2-D library mentation, autonomous flight could ensure more effi­
of photographs and creating a 3-D point cloud cient data collection with a reduced overlap rate.
(Shahnaz 2010, Wu 2013, Schoenberger & Frahm Once the 3-D point cloud is created, it can be seg­
2016). This formed the base for the AB-BrIM and mented into the bridge’s elements (i.e. deck, girder,
helped to visualize the geometric information and abutments). Current AASHTO and FHWA regula­
surface condition of the structure. Although com­ tions require a rating on the bridge’s components;
mercial, licensed software exists which provides therefore, this segmentation step would help facili­
a simple workflow for the point cloud creation, tate future implementation. The presented frame­
open-sourced software was used and attained simi­ work proposes a semi-autonomous segmentation
lar results. Meshroom with AliceVision, an open- procedure that uses meta-data of the point cloud, sur­
source, photogrammetry package that is based on face normals, and two clustering techniques,

Figure 3. North side of the oxbow bridge and its corresponding 3-D point cloud.

161
 owners and decision-makers. Additionally, damage
rating tends to be the most subjective section of the
existing bridge inspection process, often resulting in
inconsistent data (Moore, Phares, Graybeal, Rolan­
der, & Washer 2001). Currently, these ratings mostly
come from qualitative category ratings and not expli­
citly delineated on the rating scale. Detecting and
measuring these defects in concrete can be challen­
ging due to the low luminance and the thinness of
the cracks. In the literature, there are several studies
Figure 4. Fully segmented 3-D point cloud of the palmer on crack detection for concrete surfaces using
bridge. mainly morphological transforms (Jahanshahi &
Masri 2012, Talab, Huang, Xi, & Haiming 2016).
These systems work well in controlled environ­
ments, but because of the required fine-tuning of
several parameters, they lose generality in broader
applications. Recently, Convolutional Neural Nets
(CNNs) have made tremendous progress in image
classification and recognition (Zhang, Donahue, Gir­
shick, & Darrell 2014, Zhou, Park, & Koltun 2018,
Girshick, Donahue, Darrell, & Malik 2014, Li,
Wong, Lo, & Xin 2018, Dorafshan, Thomas, &
Maguire 2018, Hoang, Nguyen, & Tran 2018), thus
have become another alternative for damage/defect
Figure 5. Fully segmented 3-D point cloud of the oxbow
bridge.
identification. CNNs are able to robustly detect and
identify features in a given image; however, they
require a significant amount of training data to be
successful. Using CNNs to identify defects of
Gaussian Mixture Model (GMM) and Agglomera­ bridges shows promise in the future but is not pos­
tive clustering. The GMM works best for distribu­ sible without first labeling a large set of images in
tions and groups with different, irregular, or oblong a training and validation set. Gathering this training
shapes such as those found in a point cloud of data would negate the purpose of creating an auto­
a bridge. Agglomerative Clustering, however, works matic defect detection algorithm. Therefore, a new
best when the clusters can be circumscribed in defect detection algorithm optimized for the defects
a sphere. This technique creates a dendrogram of the typically found in concrete bridges which does not
points, in this case, based on the Euclidean distances, require a significant amount of training data is pro­
to separate the data. posed. This algorithm does require optimizing many
To begin the segmentation process, GMM cluster­ parameters, but, more importantly, demonstrates
ing is performed over the entire 3-D point cloud. how a defect detection technique could be imple­
After the GMM clustering is complete, the surface mented in the proposed framework. In a future
normals are used to identify the points associated implementation of the framework, this module can
with the deck from the other points. Once the deck is be readily updated to include a more robust and
identified, labeled, and segmented out, GMM clus­ effective defect detection algorithm which could
tering is again performed on the remaining include a CNN or other machine technique when
3-D points. This time, the point associated with the large labeled data sets become available.
abutments and/or piers can be identified and labeled. The proposed algorithm using the open-source,
The remaining points can be grouped using Agglom­ OpenCV library in a Python script (Bradski 2000). By
erative Clustering. Once the points are in the initial first performing minor color correction of increasing
classification, manual fine-tuning can be performed the contrast and reducing the brightness, OpenCV’s
to correctly classify all the 3-D points of the bridge Non-Local Means Denoising algorithm and
model. The results of the 3-D point registration pro­ a Gaussian Blur were performed. Next, a Top Hat
cess described are shown in Figure 2 and 3 for the Transform was performed on the image using a 3x3
two bridges. Each colored group of points repre­ square structural element as recommend by (Sankaras­
sented an element of the bridge. rinivasan, Balasubramanian, Karthik, Chandrasekar, &
Gupta 2015). This created a greyscale image where
the white portions are regions of greater contrast com­
3.2 Defect detection
pared to adjacent pixels. Then, the image was thre­
One important aspect of the framework is the defect sholded to create a simple black and white image.
detection, as the defect/damage information is crit­ A Canny Edge Detector was used to outline the white
ical in the decision-making process for bridge regions on the image, which represent the potential

162
defects. A minimum area of the outline was set, so
that the largest contours remained. The area/length
and location information were extracted from the cre­
ated contours. The outputs from the proposed defect
detection algorithm are shown in Figures 6, 7, and 8.
The proposed defect detection algorithm can
detect, highlight, and measure cracks in the structure.
The location and area/length of the crack can also be
extracted and used as a quantitative measure for
rating the structure. This would provide more con­
sistent and objective data for bridge owners and
decision-makers. Figure 9. A 2-D Image with identified defects from the
Once the damage is identified in the images taken UAV image mapped onto the 3-D point cloud.
by the UAV, it can be readily mapped onto the
3-D point cloud using the camera poses calculated
during the 3-D point cloud creation. This allows for
3-D location information to be associated with the 3.3 AB-BrIM establishment
identified defect. Figure 9 represents an image with With the 3-D model established and the damage
the asphalt crack identified. It is then projected onto information acquired, a complete AB-BrIM can be
the 3-D point cloud using the calculated camera pose. built. Three candidate BIM software suites were con­
The red cube of the point cloud is the camera location sidered to build the AB-BrIM, i.e. LEAP Bridge,
when the image was taken (i.e. the camera pose). Tekla Structures, and Revit (Bently 2017, Trimble
2018, Autodesk 2018). For this particular applica­
tion, a robust BIM software package that could
handle a point cloud and input customizable, user-
defined parameters was needed. Based on
a comparison study by McGuire et al., the Revit
software suite, designed by Autodesk, Inc., was
chosen for this need, as it was able to associate any
user-defined information (i.e. damage information,
dimensions, images) to any specific element for easy
query and access (McGuire, Atadero, Clevenger, &
Ozbek 2016). Additionally, the Revit software can
automatically generate reports of these user-defined
Figure 6. Identified cracks on palmer bridge girder.
inputs. With a few clicks in the user interface, all the
damage information for each element (i.e. girder,
deck, guardrail, etc.) can be listed and quantified in
formatted reports.
First, using the ground-truth data from GPS
points, the 3-D point cloud was scaled to the real-
world dimension, and georeferenced. Each structural
element in the point cloud was identified and labeled.
Next, the concept of “damage cubes,” used by
McGuire et al., was adapted here for documentation
of each detected defect. A damage cube is a user-
Figure 7. Identified cracks on palmer bridge asphalt deck. defined object which can be placed on the structural
elements and contain all the user-defined information
for the specified location. The damage cubes are
highlighted by the blue color shown in Figure 10.
In Figure 10, each color of the point cloud repre­
sents an element (e.g. red is the road deck, cyan is the
south girder, etc.). The damage cube is placed at the
damaged location on the corresponding bridge elem­
ent. The damage type, (i.e. cracking, spalling, delam­
ination), element location (i.e. deck, girder, beam,
etc.), specific georeferenced location (i.e. x, y, z/eleva­
tion), area/length, remarks/comments, the images
from the UAV, and any other associated information
Figure 8. Identified crack on oxbow bridge girder. deemed necessary are readily linked to the damage

163
 Ongoing research will focus on reducing human
intervention, to achieve maximum autonomy from
the UAV-based data collection to the final AB-BrIM
with damage information mapped. A module for
analyzing thermal imagery to detect subsurface
delamination of the bridge deck can be readily
incorporated in the current framework. In addition,
the integration of autonomous UAV data collection
and BIM-enabled documentation of condition
assessment results allows for frequent and repeated
inspections, as well as consistent results presenta­
Figure 10. Final deliverable: an AB-BrIM with each tion; therefore, tracking of changes of the structure
bridge element segmented and highlighted within the over time becomes much easier in a long-term
3-D point cloud and damage cubes placed where damage implementation. The temporal information is valu­
was located which storage damage information. able for conducting life-cycle analysis, or deterior­
ation modeling.

cube. Selecting a damage cube on the user interface ACKNOWLEDGEMENT

will display the properties box with all the associated
information, as shown in Figure 10. Selecting the The work presented in this paper was conducted
image name in the properties box will display the ori­ with support from Colorado State University and
ginal image taken from the UAV and the image with the Mountain-Plains Consortium, a University
the detected defects. This visualization of the geore­ Transportation Center funded by the U.S. Depart­
ferenced damage cubes and the concise information ment of Transportation (FASTACT Grant No.
presented in the automatically generated reports from 69A3551747108). The contents of this paper reflect
Revit will help bridge owners and decision-makers the views of the authors, who are responsible for
visualize the damage and make more informed deci­ the facts and accuracy of the information presented.
sions about planning and rehabilitation projects based Additionally, the authors would like to acknow­
on quantitative and consistent data. ledge the Colorado State University’s Drone Center
for providing the UAVs and support.
4 CONCLUSIONS
REFERENCES
The shortcomings of current bridge inspection prac­
tices demonstrate a need for an efficient bridge AASHTO (2018). The Manual for Bridge Evaluation,
inspection system integrating field inspection and Third Edition, 2017 (3rd ed.). Washington, D.C.: Ameri­
documentation of damage with quantitative measures can Association of State Highway and Transportation
on a more cost-effective and safer platform. Utilizing Officials.
Autodesk (2018). Revit.
UAVs and leveraging current advances in photogram­
Bang, S., H. Kim, & H. Kim (2017). UAV-based automatic
metry and machine learning, a novel UAV-based generation of high-resolution panorama at
bridge inspection system is proposed. This system a construction site with a focus on preprocessing for
creates a point cloud from UAV-collected images and image stitching. Automation in Construction 84
automatically segments each component of the bridge (August), 70–80.
(i.e. girders, guardrail, deck) using unsupervised Bently (2017). LEAP Bridge.
machine learning. Each component is uploaded and Bradski, G. (2000). The OpenCV Library.
georeferenced in Revit. Automated damage detection Chiu, W. K., W. H. Ong, T. Kuen, & F. Courtney (2017).
algorithms are developed to extract quantitative Large Structures Monitoring Using Unmanned Aerial
damage information from images. Next, damage Vehicles. Procedia Engineering 188(i), 415–423.
Dorafshan, S., M. Maguire, N. V. Hoffer, & C. Coopmans
cubes are placed at the damaged location on the com­ (2017, 6). Challenges in bridge inspection using small
ponents to document and visualize the component- unmanned aerial systems: Results and lessons learned.
wise damage information. The system, however, is In 2017 International Conference on Unmanned Aircraft
not able to automatically label the segmented point Systems (ICUAS), pp. 1722–1730. IEEE.
cloud and required a user to provide this information Dorafshan, S., R. J. Thomas, & M. Maguire (2018). Com­
manually. The proposed bridge inspection system not parison of deep convolutional neural networks and edge
only provides bridge owners and managers with con­ detectors for image-based crack detection in concrete.
venient damage information query and visualization Construction and Building Materials 186, 1031–1045.
Duque, L., J. Seo, & J. Wacker (2018). Bridge Deterior­
to better support the decision-making process but also
ation Quantification Protocol Using UAV. Journal of
offers a seamless transition from current practices to Bridge Engineering 23(10), 04018080.
the implementation of a UAV-enabled bridge inspec­ Gageik, N., M. Strohmeier, & S. Montenegro (2013). An
tion system. autonomous UAV with an optical flow sensor for

164
 positioning and navigation. International Journal of estimation. Lecture Notes in Computer Science 7727
Advanced Robotic Systems 10, 1–9. (PART 4), 1–14.
Gillins, M. N. (2016). Cost-Effective Bridge Safety Inspec­ Mulakala, J. (2018). Measurement Accuracy of the DJI
tions Using Unmanned Aerial Vehicles (Uavs). Tech­ Phantom 4 RTK & Photogrammetry. Technical report,
nical report, United States Department of DroneDeploy, San Francisco, CA.
Transportation, Washington, D.C. Omar, T. & M. L. Nehdi (2017). Remote sensing of con­
Gillins, M. N., D. T. Gillins, & C. Parrish (2016). Cost- crete bridge decks using unmanned aerial vehicle infra­
Effective Bridge Safety Inspections Using Unmanned red thermography. Automation in Construction 83
Aircraft Systems (UAS). In Geotechnical and Structural (April), 360–371.
Engineering Congress 2016, Phoenix, AZ. Ryan, T. W., E. Mann, Z. M. Chill, & B. T. Ott (2012).
Girshick, R., J. Donahue, T. Darrell, & J. Malik (2014). Rich Bridge Inspector’s Reference Manual. Technical report,
Feature Hierarchies for Accurate Object Detection and FHWA, Washington, D.C.
Semantic Segmentation. In 2014 IEEE Conference on Sankarasrinivasan, S., E. Balasubramanian, K. Karthik,
Computer Vision and Pattern Recognition, pp. 580–587. U. Chandrasekar, & R. Gupta (2015). Health Monitoring
Greenwood, W. W., J. P. Lynch, & D. Zekkos (2019). of Civil Structures with Integrated UAV and Image Pro­
Applications of UAVs in Civil Infrastructure. Journal of cessing System. Procedia Computer Science 54,
Infrastructure Systems 25(2), 04019002. 508–515.
Hernandez, I. (2016). Overcoming the Challenges of Using Schoenberger, J. L. & J.-M. Frahm (2016). Structure-from-
Unmanned Aircraft for Bridge Inspections. Ph. D. Motion Revisited. In Conference on Computer Vision
thesis, University of Missouri-Kansas City. and Pattern Recognition (CVPR).
Hoang, N.-D., Q. L. Nguyen, & V. D. Tran (2018). Auto­ Seo, J., L. Duque, & J. P. Wacker (2018). Field Application
matic recognition of asphalt pavement cracks using of UAS-Based Bridge Inspection. Transportation
metaheuristic optimized edge detection algorithms and Research Record 2672(12), 72–81.
convolution neural network. Automation in Construction Shahnaz, S. (2010). Gravel Road Condition Monitoring
94(June), 203–213. Using Unmanned Aerial Vehicle (UAV) Technology. Ph.
Jahanshahi, M. R. & S. F. Masri (2012). Adaptive D. thesis, South Dakota State University.
vision-based crack detection using 3D scene reconstruc­ Talab, A. M. A., Z. Huang, F. Xi, & L. Haiming (2016).
tion for condition assessment of structures. Automation Detection crack in image using Otsu method and mul­
in Construction 22, 567–576. tiple filtering in image processing techniques. Optik 127
Li, J., H. C. Wong, S. L. Lo, & Y. Xin (2018). Multiple (3), 1030–1033.
Object Detection by a Deformable Part-Based Model Trimble (2018). Tekla Structures.
and an R-CNN. IEEE Signal Processing Letters 25(2), Wells, J. & B. Lovelace (2018). Improving the Quality of
288–292. Bridge Inspections Using Unmanned Aircraft Systems
Lowe, D. (1999). Object recognition from local (UAS). Technical Report July, Minnesota Department of
scale-invariant features. In Proceedings of the Seventh Transportation, St. Paul, MN.
IEEE International Conference on Computer Vision, Wu, C. (2013). Towards linear-time incremental structure
Kerkyra, Greece, pp. 1–8. IEEE. from motion. In Proceedings - 2013 International Con­
McGuire, B., R. Atadero, C. Clevenger, & M. Ozbek ference on 3D Vision, 3DV 2013, pp. 127–134.
(2016). Bridge Information Modeling for Inspection and Zhang, N., J. Donahue, R. Girshick, & T. Darrell (2014).
Evaluation. Journal of Bridge Engineering 21(4), Part-Based R-CNNs for Fine-Grained Category
04015076. Detection. ECCV 2014 1, 834–849.
Moore, M., B. Phares, B. Graybeal, D. Rolander, & Zhou, Q.-y., J. Park, & V. Koltun (2018). Open3D:
G. Washer (2001). Reliability of Visual Inspection for A Modern Library for 3D Data Processing.
Highway Bridges. Technical Report FHWA-RD-01-020, Zingg, S., D. Scaramuzza, S. Weiss, & R. Siegwart (2010).
FHWA, Atlanta, GA. MAV navigation through indoor corridors using optical
Moulon, P., P. Monasse, & R. Marlet (2012). Adaptive flow. In Proceedings - IEEE International Conference
Structure from Motion with a contrario model on Robotics and Automation, pp. 3361–3368.

165
 Bridge Maintenance, Safety, Management, Life-Cycle Sustainability and Innovations –
Yokota & Frangopol (eds)
© 2021 Taylor & Francis Group, London, ISBN 978-0-367-23278-8

Structural parameter identification from image-based acquired information

M. Helmrich & G. Morgenthal
Chair of Modelling and Simulation of Structures Institute of Structural Engineering Bauhaus-Universität Weimar,
Weimar, Germany

ABSTRACT: To ensure the structural safety and serviceability is a top priority for large infrastructures. To
determine the load bearing capacity for the current state and to predict the structural behaviour for the future
life span, regular inspections and evaluations of the structural condition are required. Recently, many different
approaches using modern digital technologies were developed to support the engineer with the acquisition and
interpretation of structural information. This article focuses on the evaluation of visual data acquired by using
small unmanned aircraft systems (UAS) equipped with high-quality cameras. The images obtained are used
amongst others to reconstruct a dense 3D point cloud, which contains local (e.g. cross sectional shape, spalling
areas) as well as global (e.g. bending line with respect to a reference condition or to earlier observation) geo­
metric information. Furthermore, surface anomalies (e.g. cracks), automatically detected by image analysis
algorithms, are mapped to the 3D structure. This paper presents a methodology, which processes and combines
the acquired data from consecutive UAS-based inspections to quantify identifiable parameters usable for
a validation or calibration of a numerical model. This parameters cover for example cross sectional values,
displacements and curvature at each position of the element axis as well as magnitudes of load and effective
bending stiffness. The introduced procedure systematically shows on different structural levels and based on
the number of data acquisitions and available a priori information about the structure which parameters are
identifiable and how they are obtained. The quantification is performed by a combination of mathematical opti­
mization strategies with linked cross sectional and system analysis. Case studies on a laboratory experiment
and a semi-integral bridge are used to show the functionality and potential of the proposed method.

1 INTRODUCTION engineer with the acquisition (Gastineau et al. 2009,

Koch et al. 2016), processing (Mohan and Poobal
Bridges are exposed to many different types of 2018, Li et al. 2019) and interpretation (Ellenberg
impacts, such as temperature, natural hazards or et al. 2016, Braun et al. 2015) of structural informa­
increasing traffic, while they are at the same time tion. Focus of this article are three dimensional (3D)
key elements of the infrastructure system. Therefore, point clouds (PtCls) used for structural assessment of
it is important to ensure their serviceability and (semi-) integral bridges. Those PtCls are the result of
structural safety at any point of their life time. By data acquisitions using either terrestrial laser scanners
knowing the structural condition, critical areas of the or unmanned aircraft systems (UAS’s) equipped with
structure can be determined and the future structural high-quality cameras and contain local (e.g. cross
behaviour can be predicted. Assessing the structural sectional shape) as well as global (e.g. bending line
condition is currently still performed manually by with respect to a reference state) geometric informa­
trained engineers, which follow the national guide­ tion. Previous studies (e.gTang et al. 2007, Haller­
lines such as (DIN Deutsches Institut f ¨ur Normung mann et al. 2018) have already exemplarily shown
e.V. 1999) in Germany or (Hartle 2002) in the that this type of geometrical data can be used in the
United States. Due to the large dimensions of process of visual, structural inspections of structures.
bridges, this procedure is time consuming, costly Taking these results up, this research work aims to
and sometimes dangerous for the operating engineer. tailor the existing algorithms and combine them with
Furthermore, the result of the assessment depends new approaches with the objective of structural con­
highly on the expertise of the inspector and has dition evaluation based on geometric data provided
therefore a subjective component. by PtCls.
Recently, many different approaches using modern In order to use the provided geometric data in the
digital technologies were developed to support the context of a framework for an automated assessment

DOI: 10.1201/9780429279119-18
https://doi.org/10.1201/9780429279119-18

166
(Morgenthal et al. 2019), it is compulsory to process
and interpret the data from a mechanical point of view.
Three different methods of analysing and processing
the geometric data for further calibration and updating
of a finite element model (FE-model) are proposed in
this article: analysis on cross sectional, element and
system level respectively. All three processing steps
lead to a different type of information, which provide
in combination an advice for the structural condition
assessment. Additionally to the level on which the ana­
lysis is performed, more significant information are
obtained, if a reference state is known. For some appli­
cations, the planned geometry is already enough for
deriving higher order information. However, if geo­
metric data of a previous acquisition is available, it is Figure 2. Side view of Scherkonde Viaduct with investi­
possible to extract the exact deformation happened in gated pier (highlighted), its FE-model and a snapshot of the
between these two states. For each introduced process­ reconstructed PtCl.
ing step, the difference of the results with respect to
the type of the reference state will be discussed.
The presented results in the following sections of the UAS-based data acquisitions (Aug. year 1 and
refer to a laboratory experiment and to a semi- Apr. year 2) consists of up to 3 × 108 points.
integral railway bridge in Germany. The experiment
consists of a single load Fi ¼ 2:5kN · i 8i 2 ½0; 11]
applied to a simply supported, reinforced concrete 2 ANALYSIS ON CROSS SECTIONAL LEVEL
beam (see Figure 1). The deformed shape of each
load step were recorded by seven synchronised cam­ In order to represent the structural behaviour of struc­
eras, such that 12 PtCls (1 × 108 points respectively) tures, which are numerically approximated by one
representing the top and bottom as well as one side dimensional (1D) elements, as realistic as possible, the
of the beam could be reconstructed. For further cross sectional values are essential. The common way
information about the test setup and first results, the to obtain them is by using the dimensions and shapes
reader is referred to Taraben et al. (2018). from the plans. However, this neglects uncertainties of
As second validation example, a temperature sen­ the construction process as well as changes during the
sitive pier of Scherkonde Viaduct is chosen (see lifetime, e.g. due to spalling. By using the geometric
Figure 2). For a detailed description of the bridge data of the PtCl, the exact shape at each point, i.e. at
and first results, the reader is referred to Morgenthal the Gaussian integration points of the corresponding
et al.(2019). Each PtCl reconstructed from the images FE-model, can be integrated into the calculation
procedure.

2.1 Processing data of a single PtCl

There exists plenty research in extraction the
cross sectional shape. Since cross sections of
bridges, especially of the superstructures, can be
rather complex (e.g. non-convex shapes, girders
with more than one web), algorithms for example
proposed by Kresslein et al. (2018) do not work
for all types of cross sections. Problems of such
3D skeletonisation algorithms could occur if the
superstructure consists of more than one web,
which would lead to separate cross sections for
each web. On the other hand side, these algo­
rithms do not need any previous segmentation of
the PtCl. Other methods, for example presented
by Gyetvai et al. (2018), uses also a cross sec­
tional approach, but require a manual processing.
However, the presented procedure in this article
follows the steps of Morgenthal et al. (2019) and
requires a pre-segmentation of the PtCl based on
Figure 1. Test setup of laboratory experiment with the main structural components, e.g. superstruc­
expected crack spacing based on a smeared crack model, its ture or piers. This segmentation does not neces­
FE-model and a snapshot of the reconstructed PtCl. sarily be performed manually, if there exists
167
a geo-referenced 3D model, e.g. the geo-referenced
volumetric representation of an FE-model. The PtCl
of each structural component is then sliced perpen­
dicular to its main axis. Additionally to the points
where the shape shall be extracted, the thickness of
each slice needs to be set. Due to the high density
of the used PtCls (≈70pts surface area), a thickness
of Δt ¼ 1 is sufficient for both validation examples.
The unsorted PtCl of each slice is then treated as
weighted alpha shape (Edelsbrunner 1992), such
that 1) the outer shape is recovered, 2) the points
are sorted along the outer shape, 3) the PtCl is
iteratively segmented based on the angle between
the normal vectors, 4) intersections of fitted planes
(excluding outliers) are calculated and 5) the final
polygon is obtained after applying simplification
algorithms (e.g Douglas and Peucker 1973). This Figure 5. Cross sectional values of investigated Scher­
procedure is schematically depicted in Figure 3. konde Pier extracted from the PtCl, exemplarily shown for
By applying the described procedure to the investi­ the cross sectional area. Left: absolute value, Right: relative
gated pier of the Scherkonde Viaduct, the exact geom­ difference to initially assumed area in the FE-model.
etry (see Figure 4) and with this also the relevant

cross sectional values are obtained, exemplarily shown

for one of the data acquisitions in Figure 5 (left).

2.2 Processing data with respect to a reference

state
Having the exact cross sectional shape of one state of
the structure and taking the shape of a previously
known reference state into account, further values can
be derived from the new acquired PtCl. On hand
side, the shapes can be compared to each other, such
that changes in the geometry and with this in the
Figure 3. Schematic procedure of cross section extraction. cross sectional values are detectable (e.g. due to spall­
a: outer shape in blue with order of points in grey (step 1 ing). For the validation case of Scherkonde Viaduct,
and 2); b: interval boundaries of iteratively segmentation no geometrical changes between the data acquisition
(from dark blue to light blue to green) of sorted points campaigns are detected, but, as presented in Figure 4
(step 3); c: found cross section polygon in light blue and and Figure 5 Values (right), a difference with respect
simplified cross section polygon in dark blue (step 4 and 5). to the initially modelled geometry in the FE-model.
Furthermore, deformation with respect to the ref­
erence state of the cross section itself are detectable,
including both solid body movements uy ; uz ; ’x as
well as distortions μy ; μz ; γyz . However, for bridge
engineering applications only the displacement of
the centre point along the local coordinate axes and
the rotation about the element axis are of interest and
distortions of the cross section are assumed to be
neglectable small. Potential horizontal and vertical
displacements uy;j and uz;j of a cross section corner
point j are therefore the result of a combination of
the three solid body movements:

Figure 4. Extracted cross section (blue and black; dark

blue represents the simplified polygon) for one slice of
Scherkonde Viaduct pier (grey) with initially modelled rect­ where yj and zj are the local coordinates of the corner
angular geometry of the FE-model (orange) in local coord­ point j with respect to the centre of gravity. Since for
inates of the cross section. further analysis, e.g. a model calibration, the
168
 conditions, the cross sectional shape at the same
position can have different number of corner
points (see also Figure 4), e.g. due to spallings or
due to small reconstructions inaccuracies. To
assign pairs of corresponding corner points, each
cross section polygon is transformed firstly to its
individual local coordinate system. Then, the pairs
are found by the smallest difference between poly­
gon points, assuming again that for civil engineer­
ing applications distortions can be set to zero.
Finally, the centre point displacement and rotation
is calculated with Eq. 2 but using only corner
point pairs assigned in the previous step.
Figure 6. Example of results of the beam experiment (slice
thickness: Δx ¼ 0:5mm) at x ¼ 0:3m for load steps i ¼ 0
(black, reference state) and i ¼ 11 (grey). 3 ANALYSIS ON ELEMENT LEVEL

3.1 Processing data of a single PtCl

movement of the centre point is needed, the calculated Combining the results of the cross sectional analysis,
differences of the corner points are splitted into their namely connecting the centre of gravity of each
contributions with respect to the centre. Taking still slice, the absolute position of the pier axis, e.g. in
the assumption of neglectable small distortions into longitudinal direction of the bridge (see Figure 7
account, already two corner points are sufficient to (left)), is obtained.
calculate the centre point displacement and rotation, Nonetheless, this information cannot directly be
as for the laboratory beam experiment exemplarily integrated in an FE-model, since it would neglect the
demonstrated in Figure 6. If in any case distortions stress state corresponding to this deformation. Even
are expected, more corner points need to be detected though, those stresses can be considered in an FE-
in the PtCls and Eq. 1 has to be slightly modified. model as residual stresses, it is not feasible to
In most cases, however, more than two cross sec­ assume that the complete stress state can be recorded
tion corners are captured and are thus part of the for bridge structures. For this reason, the absolute
reconstructed PtCl. Even if the cross sectional shape position for the different data acquisition are taken
does not change between the data acquisition cam­ into account as intermediate results and provide
paigns, problems due to small inaccuracies of the 3D valuable information by comparison with a reference
reconstruction can occur when determining uy ; uz state (see section 3.2).
and ’x with more than two cross section corner
points. There exists several possibilities to overcome
this problem, e.g. taking distortions μy ; μz ; γyz for the 3.2 Processing data with respect to a reference
calculation into account, but ignore their values for state
the further analysis. For the work of this article the In contrast to the data processing on cross sectional
displacement and rotation of the cross section centre level (sec:section 2.2), the comparison with a perfectly
point is found by an optimization task:

Two additional, optional processing steps to find

the optimal solution of Eq. 2 need to be mentioned:
1. If the set of corner points has more than one sym­
metry, multiple possible solutions exist, because
different rotational angles ’x 2 ½_π; π] lead to
the same result. Since for civil engineering appli­
cations relatively small displacements and rota­ Figure 7. Processed results of PtCl of investigated Scher­
konde pier. Left: absolute position of pier axis in longitu­
tions are expected, the solution for further dinal bridge direction (black: first data acquisition,
analysis is chosen by the minimal absolute value grey: second data acquisition), Middle: difference of the
of all possible ’x ‘s. pier axes in longitudinal direction (black: raw values, blue:
2. Applying the described method to PtCl recon­ curvature-based fitted curve), Right: corresponding curva­
structed from data acquired under realistic field ture of fitted curve.

169
built or rather planned geometry, is often misleading,
since it would neglect inaccuracies during the con­
struction process. The processed data of a previously
acquired and processed PtCl should be preferred for
this reason at any time, such that certain parameters of
an FE-model can be calibrated more accurately. Aim
of these model calibrations is mostly to determine the
current structural load bearing behaviour, such as the
effective as-is bending stiffness (see section 4.1). If
Figure 8. Results of curvature extraction for the laboratory
only the difference of the displacement is taken into beam experiment. Top: orthophoto of reconstructed PtCl
account, several and also unrealistic stiffness distribu­ with highlighted cracks, Bottom: extracted difference of
tions are determined, since the sensitivity of the dis­ curvature between load step i ¼ 10 and load step i ¼ 0 (ref­
placement regarding the bending stiffness is quite low. erence state).
More feasible results are obtained using a highly sen­
sitive model output with respect to the bending stiff­
ness, i.e. the curvature, as objective value in a model either manually or by applying automated image ana­
calibrating process. lysis processing steps. Including for each crack
Calculating the change of curvature between two a length of influence and an additional curvature
structural states from the difference of the centre axis change in the optimization task (Eq. 3) and modify­
displacement cannot be done straight forward (e.g. by ing the regularization term for only the curvature of
using finite differences), since the data is noisy as the parts outside of the crack influence lengths, the
depicted for the investigated pier of Scherkonde Via­ curvature is obtained as shown in Figure 8.
duct in Figure 7 (middle). Chartrand (2011) presented
a method for extracting derivatives from noisy data.
This method is enhanced for the purpose of this 4 ANALYSIS ON SYSTEM LEVEL
research work, the extraction of the second derivative.
Considering small intervals, the curvature κy can be Up to this point, the presented processing steps lead
approximated by a piecewise linear function, such that either to displacements of surface points or to
the integral of the second derivative κy00 with respect to a complete set of displacement values, i.e. in local
the spatial variable z over the whole element length L coordinate notation: uy , uz , ’x , ’y , ’z , κy and κz .
tends to be zero. This assumption is represented in the Only a displacement along the element axis ux
regularization term of the final optimization task: cannot be detected. However, this set allows a very
precise validation or calibration of certain param­
eters of an FE-model.

4.1 Processing data of a single PtCl

For the same reason as discussed in section 3.1 the
data of one PtCl is not sufficient for deriving any fur­
where ω is an optional weighting factor for scaling ther structural information, since the stress state is
the regularization term and ^f ð zÞ are the transformed not known. Hence, a reference state is obligatory.
function values of f ðzÞ ¼ Δux;glob ðzÞ:
4.2 Processing data with respect to a reference
state
As presented in the previous sections, both scenarios
The functionality of this procedure is demon­ for reference states, i.e. the planned geometry or an
strated with the data of the investigated pier in earlier acquired and processed PtCl, are applicable.
Figure 7. The blue curve, which has a good visual However, for a model calibration or validation only
agreement with the extracted data from the PtCl, is an as-is geometry as reference state leads to feasible
obtained by integrating the curvature results, since effects of deviations during the con­
r0 twice by using struction process are cancelled out. For this paper,
f 0 ¼ Δux;glob ð0Þ and f00 ¼ Δux;glob ð0Þ as integra­
tion constants. In the same way the rotation is the investigated pier is chosen to show exemplarily
obtained by integrating the curvature only one time. the model calibration process. The parameters for
Furthermore, this procedure can also be used to the optimization are divided into two groups:
extract the curvature from data sets, which have local 1. Load: For both data acquisitions no temperature
discontinuities, e.g. due to cracks. Advantage of the data was captured, such that the a constant tem­
data type used in this paper are the visual information perature change in the superstructure ΔTsup as
provided by the images, such that the positions of well as a linear temperature gradient in the pier in
cracks along the longitudinal axis can be determined longitudinal bridge direction ΔTl_r ðzÞ are
170
 unknown. Further loads did not change between significantly by refining the discretization. The cor­
the data acquisition campaigns. responding constant temperature change of the super­
2. Stiffness: In fact, the second moment of area Iy is structure is ΔTsup ¼ _3:3K for all interval numbers.
calculated by using the method described in section While the qualitative distributions of the results as
2.1, the effective bending stiffness EIy;eff ðzÞ, how­ well as the magnitudes of the temperature changes
ever, cannot be derived directly. During the second are feasible, the loss of stiffness in the lower half of
data acquisition campaign, small cracks (crack the pier is greater than expected considering that
width wcrack � 0:15mm) at the surface of the only small cracks on the surface were detected.
lower half of the pier were recorded manually. More data needs to be acquired to either confirm this
Even though, cracks are detected on the surface, values or to help to find a better assumption for scal­
they were not detected by the curvature-based ing the results of the optimization process.
approach of section 3.2. This leads already to the
conclusion, that the influence of a single crack is
relatively small. But since this type of damage can 5 CONCLUSIONS AND DISCUSSION
reduce in sum the effective bending stiffness sig­
nificantly, its value is set to unknown. Regular inspections are essential for ensuring the
structural safety and serviceability of bridges. The
Both load and stiffness parameters for the pier are proposed methods of this paper showed on two case
discretized as piecewise constant functions with studies that UAS-based acquired images are
5=10=15 intervals each. The calibration of the FE- a powerful data set for obtaining information struc­
model is then done by a simple sum of least squared tural parameters and help therefore operating engin­
errors problem: eers with the evaluation of the structural condition.
The reconstructed, geo-referenced PtCls are used to
extract the as-is geometry on cross section level as
well as on element level. Based on the exact shape
cross sectional values and the absolute axis position
are determined. By combining then processed data
of multiple PtCl, a complete set of displacement
where n is the number of nodes in the pier in the values (in local coordinate notation: uy , uz , ’x , ’y ,
FE-model. Since both impact side and resistance side ’z , κy , κz , excluding ux ) at each position is calcu­
are unknown and therefore varied during the opti­ lated with the tailored algorithms introduced in this
mization process, the final result needs to be scaled. article. Furthermore, a model calibration on system
Any arbitrary assumption is sufficient. Here, the level using the set of displacement values leads to
maximum effective bending stiffness is set to be information about both load changes and changes of
equal with the initially assumed on, such that the the effective bending stiffness.
ratio EIy;eff ð zÞ=EIy;ref ðzÞ is equal to one. The scaled Up to this point, the method can be applied to any
results for the pier are presented in Figure 9. Already dense PtCl, for example also to a PtCl acquired with
with 10 intervals the results do not change a terrestial laser scan. However, image-based acquired
data provide additional information. In this paper,
crack positions in longitudinal direction were extracted
and used to enhance the introduced curvature extrac­
tion method. Hence, more research needs to be done to
combine and ameliorate the presented methods with
visual information.
Since this article shows the potential and func­
tionality of the introduced method, it can be taken as
basis for future work. Those should cover on one
hand side the acquisition and processing of more
than two data sets of a structure (as already dis­
cussed in section 4.2), such that uncertainties of the
PtCl reconstruction process can be taken into
Figure 9. Results of model calibration on system level for account. Furthermore, the load, i.e. the temperature,
different discretization of load and stiffness parameters should be measured and a coupled thermal and
(5=10=15 intervals correspond to dark blue, light blue and mechanical analysis can then be included into the
green respectively). Left to right: 1) difference of pier axes model calibration process.
in longitudinal direction (grey, see also Figure 7) with
bending line of calibrated FE-model (coloured), 2)
extracted curvature of the PtCls of the pier (grey, see also
Figure 7) with calculated curvature of calibrated FE-model ACKNOWLEDGEMENTS
(coloured), 3) change of temperature gradient in longitu­
dinal bridge direction, 4) ratio of effective bending stiffness The authors would like to acknowledge the support
with respect to initially assumed bending stiffness. of the Federal Ministry of Education and Research

171
of Germany to the funding of the project AISTEC monitoring of bridges. In Maintenance, Safety, Risk,
(No. 13N14657). Management and Life-Cycle Performance of Bridges:
Proceedings of the Ninth International Conference on
Bridge Maintenance, Safety and Management (IABMAS
REFERENCES 2018), 9-13 July 2018, Melbourne, Australia, pp. 157.
CRC Press.
Braun, A., S. Tuttas, A. Borrmann, & U. Stilla (2015, jan). Hartle, R. A. (2002). Bridge Inspector’s Reference Manual
A concept for automated construction progress monitor­ (BIRM). US Department of Transportation, Federal
ing using bim-based geometric constraints and photo­ Highway Administration.
grammetric point clouds. Journal of Information Koch, C., S. G. Paal, A. Rashidi, Z. Zhu, M. König, &
Technology in Construction, Special issue ECPPM I. Brilakis (2016). Achievements and challenges in
2014-10th European Conference on Product and Pro­ machine vision-based inspection of large concrete
cess Modelling 20, 68–79. structures. Advances in Structural Engineering 17(3),
Chartrand, R. (2011). Numerical differentiation of noisy, 303–318.
nonsmooth data. ISRN Applied Mathematics 2011. Kresslein, J., P. Haghighi, J. Park, S. Ramnath,
DIN Deutsches Institut für Normung e.V. (1999). Bauwerk­ A. Sutradhar, & J. J. Shah (2018). Automated
sprüfung nach DIN 1076 - Bedeutung, Organisation, cross-sectional shape recovery of 3d branching struc­
Kosten; Dokumentation 1999. Berlin: Beuth Verlag. tures from point cloud. Journal of Computational
Douglas, D. H. & T. K. Peucker (1973). Algorithms for the Design and Engineering 5(3), 368–378.
reduction of the number of points required to represent Li, S., X. Zhao, & G. Zhou (2019). Automatic pixel-level
a digitized line or its caricature. Cartographica: the multiple damage detection of concrete structure using
international journal for geographic information and fully convolutional network. Computer-Aided Civil and
geovisualization 10(2), 112–122. Infrastructure Engineering, 616–634.
Edelsbrunner, H. (1992). Weighted alpha shapes, Volume Mohan, A. & S. Poobal (2018). Crack detection using
92. University of Illinois at Urbana-Champaign, Depart­ image processing: A critical review and analysis. Alex­
ment of Computer Science. andria Engineering Journal.
Ellenberg, A., A. Kontsos, F. Moon, & I. Bartoli (2016). Morgenthal, G., N. Hallermann, J. Kersten, J. Taraben,
Bridge related damage quantification using unmanned P. Debus, M. Helmrich, & V. Rodehorst (2019). Frame­
aerial vehicle imagery. Structural Control and Health work for automated uas-based structural condition
Monitoring 23, 1168–1179. assessment of bridges. Automation in Construction 97,
Gastineau, A., T. Johnson, & A. Schultz (2009). Bridge 77–95.
Health Monitoring and Inspections - A Survey of Tang, P., B. Akinci, & J. H. Garrett (2007). Laser scanning
Methods: Final Report. Minnesota Department of for bridge inspection and management. In IABSE Sym­
Transportation. posium Report, Volume 93, pp. 17–24. International
Gyetvai, N., L. Truong-Hong, & D. F. Laefer (2018). Laser Association for Bridge and Structural Engineering.
scan-based structural assessment of wrought iron Taraben, J., N. Hallermann, J. Kersten, G. Morgenthal, &
bridges: Guinness bridge, ireland. Proceedings of the V. Rodehorst (2018). Case study for the integration of
Institution of Civil Engineers-Engineering History and geometrical analyses for structural condition assessment
Heritage 171(2), 76–89. in building information models. In IOP Conference
Hallermann, N., J. Taraben, & G. Morgenthal (2018). Bim Series: Materials Science and Engineering, Volume
related workflow for an image-based deformation 365, pp. 022–054. IOP Publishing.

172
 Bridge Maintenance, Safety, Management, Life-Cycle Sustainability and Innovations –
Yokota & Frangopol (eds)
© 2021 Taylor & Francis Group, London, ISBN 978-0-367-23278-8

A machine learning approach to damage detection of bridges

R.C. George
Hiroshima University, Japan

ABSTRACT: This study proposes a method for damage detection on bridges using its operational vibrations
under a moving vehicle. The method uses signal energy of acceleration responses as the damage sensitive
feature. Outlier analysis using one class support vector machines are used to classify the data and detect the
presence of damage in the system. Factors influencing the accuracy of damage identification are investigated.
It is observed that using adequate training data and enough number of sensors, a reasonable estimate on the
presence of damage can be obtained. The method can handle the operational variability of the vehicular traffic
such as mass and velocity of the vehicle. The method is illustrated using simple numerical simulations.

1 INTRODUCTION understanding the type of damage and prognosis of

the structure, as well as the location and quantity of
1.1 Bridge health monitoring damage. Data based methods rely mostly on feature
identification from the vibration data. These methods
A bridge directly and indirectly benefits different
are faster and less complex but are used only to get
communities. While commuters depend on the
information about the presence, location and quan­
bridge to travel from one place to another, others
tity of the damage. The type of damage or the prog­
benefit from the goods transported through the
nosis would be difficult with these methods. Farrar
bridge. Collapse of a bridge pose direct threat to the
& Worden (2013) discussed the data-based methods
safety of the commuters and vehicles passing over it
and machine learning methods used in SHM in
and cause huge economic loss to the government.
detail.
Shut down of a bridge cause inconvenience to all its
beneficiaries. However, replacement of bridges as
soon as they near the end of their planned service 1.2 Machine learning techniques
life is not always feasible due to the economic
Artificial Intelligence (AI) is probably the most
burden. This is especially true in the current times as
popular recent terminology that is both used and
the number of old bridges are increasing annually.
abused. While many definitions of AI do exist, the
For safe and sustainable infrastructure, it becomes
widely accepted one is that of it being both science
necessary to monitor the health of bridges periodic­
and engineering that makes computers behave in
ally. Damages need to be identified as early as pos­
ways that require human intelligence (Wladawsky-
sible and maintenance measures should be taken in
Berger, 2018). This suggests that it is a balancing act
order to prevent the growth of these damage and to
between engineering through required technology
prolong the safe use of bridges.
and sciences through the required theoretical inter­
Vibration based, structural health monitoring
est. While the theoretical aspects give it the unre­
(SHM) methods suggests many methods to identify
stricted feel, while the technological aspects provide
the presence, location, quantity and type of damage
with its restrictions (Coppin, 2004). A lot of what is
of the structure(Brewick & Smyth, 2013; Doebling
being called as AI today is actually Machine Learn­
et.al, 1996;Farrar & Worden, 2013; Gul & Catbas,
ing (ML), which is an algorithmic field that com­
2009; Sohn et al, 2003). These methods are mainly
bines statistics, computer science, operations
of two types: model based and data based. Model
research, etc. to design algorithms that process large
based methods are complex due to the need to create
amounts of data to assist the decision maker (Lipton,
an analytical or mathematical model of the structure
2017). Most ML techniques can be broadly classified
to be monitored. The accuracy of the results depend
into five (Coppin, 2004; Sutton & Barto, 2014; Rus­
on the accuracy of the model. However, if the model
sell & Norvig, 2017), viz., (i) Supervised learning
is accurate enough then these methods can help in
that utilizes labelled data where if the data is

DOI: 10.1201/9780429279119-19
https://doi.org/10.1201/9780429279119-19

173
categorical, then classification algorithms are used measured every time a vehicle moves over a bridge.
and regression models are used for continuous data, The signal energy of the response is calculated from
(ii) Unsupervised learning that utilizes unlabelled each sensor response, as given by Equation 1
data, where attempts to make sense of the data by (George et.al.,2017a),
putting into clusters (clustering) or reduction of
dimensions, (iii) Semi-supervised learning that util­
izes a mix of both labelled and unlabelled data and
can combine benefits of the earlier two approaches,
(iv) Reinforcement learning that utilizes feedback or
reinforcement signals that focuses on learning what
to do or how to map situations to actions, and (v) where n is the number of timestep and xi is the
Recommendation engine that focuses on finding pat­ acceleration at ith timestep.
terns in the historic data to provide accurate and The energy values corresponding to each sensor is
meaningful recommendations. stored as a data base. The acceleration time histories
Most learning models work on identifying the and thus the signal energies depend on the mass and
right set of features to make the learning process the speed of the vehicles moving over the bridge.
easy. However, identifying the “right set of features” However, the effect due to variabilities in mass and
is considered quite difficult (Ventura, 2009). Vapnik velocity gets reflected in all sensor readings. In order
(1995) developed the Support Vector Machines to cancel this effect, the signal energy values in each
(SVM) methodology that is widely used for classifi­ observation are scaled by dividing with the min­
cation and regression problems, which utilizes imum signal energy value among the sensors in that
a linear/non-linear mapping function that can trans­ observation. This makes the minimum scaled energy
form input data into feature space data so that the in each observation to be 1. Large data set is col­
problem becomes linearly separable. The optimal lected when the bridge is at ‘healthy’ state. When the
separating hyperplane is then automatically dis­ current state of the bridge needs to be evaluated,
covered by SVM. Also, SVM belongs to the super­ energy values from one observation is compared
vised machine learning class of algorithms. against the healthy data (training dataset) using out-
SVMs have been successfully used in condition lier analysis based on one-class SVM.
assessment of rotating machineries. One class SVM
based outlier analysis can be used for detecting 2.2 Damage identification
anomalies and damages (Zhang et al, 2009; Mahade­
van & Shah, 2009). The training data of the bridge signal energies are of
the form of a matrix where the rows correspond to
observations and columns corresponds to different
1.3 Present study sensors. The signal energies at each observation is
This study explores the use of support vector machines divided by the minimum of the signal energies in
(SVM) to find damage in a bridge using acceleration that observation to cancel the effect of variability in
responses under variable vehicle traffic. George et.al. vehicle masses and velocities. This data is used to
(2017) used signal energy of the acceleration model an SVM classifier with one class (in this case,
responses as damage sensitive feature and damage of healthy). If any prior information regarding the pos­
railway bridge models were determined using Mahala­ sible outliers in the healthy data such as those due to
nobis distance-based outlier analysis. The present high noise content is available, then those data
method uses outlier analysis using one-class SVM to points may be removed from the data set.
find the changes in acceleration signal energies due to To test the current state of the bridge, the class of
damage in highway bridge. The method is illustrated the data from an observation is predicted using this
through numerical simulation using simple finite elem­ SVM model. The SVM algorithm calculates the pos­
ent model of a simply supported bridge under moving terior probability or score of the test data to be in the
load due to a single vehicle. same class as the training data. Based on the value
The data collected continuously, while the bridge of this score, the data is classified as the same class
is operational, is used to check for damage in the of the training data or an outlier. The observations
structure. In the present study only one vehicle is with negative score values are identified as outliers.
assumed to be passing over the bridge at a time. This The damage modifies the acceleration signals near
assumption is aimed to be relaxed in future studies. the damage location on the bridge. This makes the
data from a damaged bridge different from those of
the training data. This difference is captured and
2 METHODOLOGY data from the damaged bridge gets identified as an
outlier. The effect of damage is high near the
2.1 Data collection and management damage location. The signal energy from sensors far
away from damage locations are not affected by the
Accelerometers are placed along the bridge at fixed damage in the bridge. So adequate number of sen­
intervals, and the acceleration responses are sors are needed to capture the damage correctly.

174
3 NUMERICAL SIMULATION

A single span simply supported bridge of length 50m

is idealized as a simply supported beam, and the
responses under a moving load are simulated using
finite element analysis. The mass per unit length of
the bridge, M, is taken to be 12.75 × 103 kg/m and the
flexural rigidity to be 1.015 × 103 kgm3/s2. Damage in
the bridge is simulated by reducing the stiffness of the
elements at the intended damage location. Figure 1
shows the simplified model of the bridge.
Operational variability caused by the differences
in masses and velocities are simulated by assuming
a uniform distribution. Masses of vehicles are
assumed to follow uniform distribution with min­
imum and maximum values as 5% and 10% of M.
Velocity of vehicles are assumed to follow uniform
distribution between 15 to 20 m/s. Different sets of
masses and velocities are considered for training and
testing. Figure 2 shows the mass and velocity distri­
bution used for generating the training data.
Finite element analysis of the beam is carried out
using eight Euler Bernoulli beam elements per span.
Vehicle is modeled as a moving load. Rayleigh damp­
ing is assumed, with 2% modal damping in the first
two modes. Acceleration responses at the vertical
DOFs at the nodes are used as sensor data. The true
acceleration responses are computed using the New­
mark-β method. Eight elements are considered over the
span. Figure 3 shows the sensor locations on the
bridge.
Bridge with no assumed/known damages is con­
sidered as the healthy bridge, and its responses form
the training data set or baseline dataset. Damages are
simulated by reducing the stiffness of the element by
0% (healthy state itself), 5%, 10%, 15% and 20%. Five Figure 2. Vehicle mass and velocity distribution in training
different locations of damage are considered. Damage data.
scenarios A to D corresponds to damage in element 1
to 4 respectively. Damage scenario E corresponds to
damage in element 2 and 5. Table 1 below shows the
summary of training data size, sensor locations,
damage locations and damage intensities used in the
study.
To find the influence of sensor density in damage
detection different numbers of sensors are assumed to Figure 3. Sensor locations at the beam.
be placed on the beam. In scenario 1, a single sensor
(S4), is assumed to be placed at the middle of the
span. In scenario 2, three sensor locations are available
(S2, S4 and S6). In scenario 3, sensor locations are The acceleration response at the nodes corres­
each of the nodes other than the supports. i.e. seven ponding to the sensor locations are evaluated using
sensors S1 to S7. finite element analysis. Table 2 shows the sensor
locations used in the three cases.
Real vibration records tend to be contaminated with
noise from multiple sources. To account for this, 10%
Gaussian noise is added to each of the simulated
accelerograms. Signal energies at each of the sensors
are calculated and the training data set is formed using
the healthy bridge responses. The signal energy values
are scaled using the minimum value in each observa­
tion, except when only one sensor is available. To find
Figure 1. The bridge model. the influence on the size of the training data in

175
Table 1. Damage locations and intensities for various damaged states are well separated and it can be very
damage scenarios. easily classified using a machine learning based
algorithm.
Damage Scenario A B C D E The data sets are separated depending on the
amount of damage present in the bridge. As the
Elements damaged 1 2 3 4 2.5 damage increases, the data sets become further apart.
Damage intensity at each 0 5% 10% 15% 20%
location
4 RESULTS AND DISCUSSIONS

Tables 3 to 5 shows the results of outlier analysis

Table 2. Number and locations of sensors used in differ­ using one class SVM classifier using training data
ent scenarios. sets of size 100, 300 and 500 respectively for each
of the sensor scenarios.
Sensor scenario 1 2 3 Tables 3, 4 and 5 summarize the performance of
this method in various damage scenarios, with different
Number of sensors 1 3 7
training data sizes and using different sensors available
Sensors locations S4 S2, S4, S6 S1 to S7
on the bridge. In each case, 100 test data were tested
for outliers. Number of correctly classified data points
out of this 100 are reported in the tables.
damage identification, training data base with number From the Tables 3-5, it can be observed that, the
of observations equal to 100, 300, and 500 are used. healthy data (with damage intensity 0%) is classified
Observations in training data are labelled healthy. accurately when largest data set (i.e. 500) in the
One-class SVM models are trained using each of training data. When the training data set is smaller,
these training datasets. A testing dataset consisting as in the case here with 100 or 300, the information
of 100 observations are created for each of the is not enough for the SVM model to learn from these
damage intensities (0, 5, 10, 15 and 20%). Each limited data. This leads to classifying healthy data as
observation in the testing dataset is compared with damaged causing false alarms.
the training data set by predicting the class of the
observation using the trained SVM model. Depend­
ing on the score or posterior probability values from
the SVM algorithm, each of the observations (data Table 3. The number of test data points correctly classi­
points) can be classified as healthy or not. Data fied when training data set of size 100 is used.
points with negative score are classified as not
healthy. In the present study outlier indicates that the Sensor Scenario 1
possibility of damage.
Figure 4 shows a scatter plot of scaled signal Damage Damage Scenario
energy value pairs from S2 and S4 for 500 healthy intensity A B C D E
data set and 500 damaged data set. As it can be
observed in the figure the data points in healthy and 0% 100 98 100 100 100
5% 0 0 0 2 1
10% 0 7 27 96 76
15% 0 12 99 100 99
20% 0 21 100 100 100
Sensor Scenario 2
Damage Damage Scenario
intensity A B C D E
0% 100 96 100 99 99
5% 0 79 80 6 8
10% 1 100 100 88 80
15% 10 100 100 100 100
20% 27 100 100 100 100
Sensor Scenario 3
Damage Damage Scenario
intensity A B C D E
0% 100 100 94 100 100
5% 5 93 93 11 66
10% 44 100 100 40 100
15% 88 100 100 90 100
Figure 4. scaled signal energies from S2 versus those of S4 20% 100 100 100 100 100
for healthy and 10% damage states.

176
Table 4. The number of test data points correctly classi­ Table 5. The number of test data points correctly classi­
fied when training data set of size 300 is used. fied when training data set of size 500 is used.

Sensor Scenario 1 Sensor Scenario 1

Damage Damage Scenario Damage Damage Scenario

intensity A B C D E intensity A B C D E

0% 100 99 100 100 100 0% 100 100 100 100 100

5% 0 0 0 2 0 5% 0 0 0 2 0
10% 0 2 15 88 42 10% 0 2 7 92 31
15% 0 3 89 100 98 15% 0 2 78 100 96
20% 0 14 100 100 100 20% 0 13 100 100 100
Sensor Scenario 2 Sensor Scenario 2
Damage Damage Scenario Damage Damage Scenario
intensity A B C D E intensity A B C D E
0% 100 99 100 99 100 0% 100 100 100 100 99
5% 0 66 75 6 6 5% 0 64 75 5 8
10% 1 100 100 89 77 10% 1 100 100 59 81
15% 9 100 100 100 100 15% 8 100 100 100 100
20% 14 100 100 100 100 20% 12 100 100 100 100
Sensor Scenario 3 Sensor Scenario 3
Damage Damage Scenario Damage Damage Scenario
intensity A B C D E intensity A B C D E
0% 100 100 100 100 100 0% 100 100 100 100 100
5% 0 93 84 3 70 5% 3 87 86 4 68
10% 3 100 100 25 100 10% 23 100 100 25 100
15% 21 100 100 79 100 15% 80 100 100 77 100
20% 79 100 100 100 100 20% 100 100 100 100 100

Due to the same reason, the tendency to classify When multiple sensors are used, in sensor scen­
the new data point as an outlier is more when the arios 2 and 3, damaged data is correctly classified
training data set is small. This will lead to more for damage scenario B and C for intensities greater
damaged data identified correctly. However, this is than 5%. At 5% damage intensity, 64 to 86 times the
due to the lack of enough training data, and this damage is identified correctly when training data of
should not be misjudged as the efficiency in damage size 500 is used. This is a reliable prediction because
detection. The adequate size of the training data will all the healthy test data are identified correctly. This
lead to accurate prediction of healthy as well as dam­ indicates that the increased number of damage detec­
aged test data. tion is not due the smaller training data set. 500 can
Sensor scenario 1 consisted of only one sensor. be considered as a good enough size for training data
The damage scenario D is identified correctly when set to detect damage in the structure when damage is
damage intensity is 15% and 20% when training at elements 3 and 4. It should also be noted that, as
data size is 500. The 75 damaged data points are the number of sensors are increasing from 3 to 7, the
identified correctly when the damage intensity is accuracy of the prediction also increases. Damage at
10%. The results are inaccurate for 5% damage these locations affect a greater number of sensor
cases. Damage scenario C also gets predicted with responses due to the positioning of the sensors.
reasonable accuracy for damage intensities of 15 and However, damage scenario D is not predicted with
20%. Damage location at both these damage scen­ the same accuracy as that of B and C for smaller
arios are closer to the sensor S4, than other damage intensities of damage. Further investigations are
locations. The effect of damage is reflected in the needed to understand the reasons for this. Lastly in
dynamic responses closer to damage location. As the case of multiple damages, the sensor scenario 3
Damage scenario D is closest to the sensor S4, that yields the best results. As the damage is located at
is identified more precisely than Damage scenario multiple places, a greater number of sensors provide
C. A single sensor is inadequate to find the damage better prediction of damage.
in other locations farther from it. In such cases, It can be said in a nutshell that, if large data set
a damaged data set gets wrongly classified as healthy (size 500) is used for training data and adequate
data. Adequate number of sensors are needed to clas­ number of sensors are used, then damage can be pre­
sify the data correctly and detect damage. dicted with acceptable accuracy, when the damage

177
intensity is 15% or more. For smaller intensities of Doebling, S.W., Farrar, C.R., Prime, M.B. & Shevitz, D.W.
damage, the accuracy of prediction depends on the 1996. Damage identification and health monitoring of
number of sensors, position of damage with respect structural and mechanical systems from changes in their
to the damage, the training data size, etc. vibration characteristics: A Literature Review, Technical
Report LA-13070-MS, Los Alamos National Laboratory,
Los Alamos, New Mexico 87545.
5 CONCLUSIONS Farrar, C. R. & Worden, K. 2013. Structural health moni­
toring: A machine learning perspective. Wiley, Chiches­
ter, U.K.
Based on the numerical simulation of the simple George R.C., Posey J., Gupta A., Mukhopadhyay S.,
bridges it can be observed that, the signal energy of Mishra S.K. 2017a. Damage Detection in Railway
the acceleration response gets influenced by the Bridges Under Moving Train Load. In: Barthorpe R.,
damage in the bridge. The accuracy of prediction Platz R., Lopez I., Moaveni B., Papadimitriou C. (eds)
using one class SVM based outlier analysis depends Model Validation and Uncertainty Quantification,
on the size of the training data set. The larger the Volume 3. Conference Proceedings of the Society for
training data set, the better. Accuracy of prediction Experimental Mechanics Series. Springer, Cham.
Gul, M. & Catbas, F. N. 2009. Statistical pattern recogni­
also depends on the number of sensors available for
tion for structural health monitoring using time series
data collection. Inadequate number of sensors leads modeling: Theory and experimental verifications. Mech.
to inaccurate predictions. This also signifies the Syst. Sig. Process. 23(7): 2192–2204.
number of sensors near the damage locations. If Lipton, Z.C. 2017. The AI misinformation epidemic;
there are many sensors far away from the damage Approximately Correct: Technical and Social perspec­
sensors, those sensor data will not have any contri­ tives on Machine Learning (March).
bution towards the damage identification. When Mahadevan, S. & Shah, S.L. 2009. Fault detection and
adequate number of data points are used in the train­ diagnosis in process data using One-Class Support
ing data set, the accuracy in predicting both the Vector Machines. Journal of Process Control, 19:
healthy and damaged data increases i.e. false posi­ 1627–1639.
Russell, S.J. & Norvig, P. 2016. Artificial Intelligence:
tives and false negatives get avoided. The method A Modern Approach (3rd Ed), Pearson Education.
can also handle operational variability of the vehicu­ Sohn, H., Farrar, C. R., Hemez, F. M., Shunk, D. D.,
lar traffic such as mass and velocity of vehicles. The Stinemates, D. W., Nadler, B. R. & Czarnecki, J. J.
efficiency of this method can be improved in the 2003. A review of structural health monitoring litera­
future studies by considering different kernel tricks ture: 1996–2001. Los Alamos National Laboratory, Los
to transform the data and using larger training data Alamos, New Mexico.
set. Localization of damage needs to be attempted by Sutton, R.S. & Barto, A.G. 2014. Reinforcement Learning:
using smaller groups of sensors along the beam. An Introduction. MIT Press.
Vapnik, V.N. 1995. The Nature of Statistical Learning
Theory. Springer.
Ventura, D. 2009. SVM Example – lecture notes.
REFERENCES Wladawsky-Berger, I. 2018. What Machine Learning can
and cannot do. CIO Journal (July).
Brewick, P. T. & Smyth, A. W. 2013. An investigation of
Zhang, Y., Meratnia, N. & Havinga, P. 2009. Adaptive and
the effects of traffic induced local dynamics on global
Online One-Class Support Vector Machine-Based Out-
damping estimates using operational modal analysis,
lier Detectio Techniques for Wireless Sensor Networks,
Mech. Syst. Signal Pr. 41: 433–453.
2009 International Conference on Advanced Informa­
Coppin, B. 2004. Artificial Intelligence Illuminated. Jones
tion Networking and Applications Workshops, Bradford,
and Bartlett.
990–995.

178
 Bridge Maintenance, Safety, Management, Life-Cycle Sustainability and Innovations –
Yokota & Frangopol (eds)
© 2021 Taylor & Francis Group, London, ISBN 978-0-367-23278-8

Inverse response surface method for prestressed concrete bridge design

D. Lehký, M. Šomodíková, M. Lipowczan & D. Novák
Brno University of Technology, Brno, Czech Republic

ABSTRACT: The paper describes application of the inverse response surface method when performing reli­
ability-based design of selected bridge parameters. Employed adaptive procedure combines response surface
method with an artificial neural network-based inverse reliability method. The first method serves for approxi­
mation of computationally demanding limit state function based on finite element method model. The second
method provides design parameters for prescribed reliability indicators which correspond to analyzed limit
states. Since the initial set of design parameters is not known in advance both methods must be efficiently
combined to obtain sufficiently accurate results. The proposed method was employed for determination of
selected uncertain material and prestress parameters of aging concrete bridge made of post-tensioned precast
segments. Presented analysis is part of lifetime and reliability assessment of the aging bridge subjected to
gradual deterioration which brings increasing level of uncertainty to its parameters.

1 INTRODUCTION determination of optimal design parameters of a single-

span post-tensioned concrete bridge made of precast
Aging bridges are regularly inspected to ensure the sus­ MPD girders. The bridge inspection pointed to
tainability of transport infrastructure. In particular, their a significant degree of deterioration of some parts of
mechanical properties and the degree of deterioration the structure. These were mainly concrete of transverse
are investigated, as well as the associated level of reli­ joints and some of prestressing tendons. Since these
ability, load-bearing capacity and residual lifetime. are important parts of the structural system, it was
Bridge engineers, administrators or owners are often necessary to identify their minimum values to maintain
faced with the question of what the minimum require­ the required load-bearing capacity and reliability.
ments for selected bridge properties – so called design The paper presents the process of determination of
parameters – are with respect to uncertainties while bridge design parameters by performing inverse reli­
ensuring the required level of reliability. When per­ ability analysis with the help of inverse response sur­
forming reliability-based design optimization the reli­ face method. First, the analyzed bridge is introduced
ability level related to a particular limit state needs to including brief information on applied load, analyzed
be quantified via reliability indicators such as failure limit states, results of diagnostics etc. Then, the basic
probability or reliability index. Since calculating these aspects of forward and inverse response surface
indicators for complex structural systems is usually methods are presented. Application section describes
a time-consuming task, the utilization of approximation the adaptive procedure when determining selected
methods is an appropriate solution. A popular approxi­ bridge parameters using two types of surrogate models.
mation method is the response surface method (RSM, Finally, obtained results are compared and discussed.
Myers 1971, Bucher 2009), where the limit state func­
tion is approximated using a suitable surrogate model.
Construction of a response surface requires all vari­ 2 POST-TENSIONED CONCRETE BRIDGE
ables of stochastic model to be known in advance.
However, during the structural design, there are design A single-span post-tensioned concrete bridge built in
parameters which are subject of reliability-based 1957, crossing a single-track railway, is situated on the
design optimization procedure and thus not known at main road near the village Uherský Ostroh in the
the start of the process. For such cases, an adaptive Czech Republic. Based on the diagnostic survey from
inverse response surface procedure is proposed. The 2007, the bridge is made of twelve precast post-
procedure is based on a coupling of the adaptive tensioned concrete MPD3 (outer) and MPD4 (inter­
response surface method and the inverse reliability ana­ mediate) type girders, which were used from 1955 for
lysis. The proposed method was applied for construction of slab bridges up to a clear span of

DOI: 10.1201/9780429279119-20
https://doi.org/10.1201/9780429279119-20

179
18 m. Each of MPD girders was composed of six seg­ Table 1. Reduced stochastic model.
ments that are connected to each other by the trans­
verse joints, see Figure 1. Varibale Distribution Mean CoV
In order to model the structural response,
a simplified numerical finite element method (FEM) ft (MPa) Weibull min. ? 0.35
model was created in ATENA 2D software (Červenka Gf (N/m) Weibull min. 47.82 0.25
et al. 2012). Here, data on the geometry and material P1 (MN) Normal ? 0.09
parameters assessed within the bridge diagnostic g1,n (kN/m) Normal –65.55 0.05
survey were used. The following load cases were
modelled: dead load of the structure, longitudinal pre­
stressing, secondary dead load and traffic load corres­
ponding to normal loading class according to Czech (CoV). The mean values of tensile strength and pre­
technical standard ČSN 73 6222 (2013). stressing force were considered as uncertain design
Stochastic model containing parameters of con­ parameters (see question marks in Table 1) with the
crete and reinforcement was created to cover uncer­ aim of finding their critical values using response sur­
tainties in material properties. The secondary dead face methods mentioned in the following section.
load was also randomized. Values according to
recommendations of Joint Committee of Structural
Safety (JCSS 2015) and Technical standards were 3 FORWARD AND INVERSE RESPONSE
used and updated based on the material parameters SURFACE METHODS
testing within the diagnostic survey. Statistical correl­
ation between some of material parameters was also 3.1 Polynomial-based response surface method
taken into account. Details on the complex stochastic Response surface modeling condenses complex
model containing the total number of 18 random vari­ engineering simulations into surrogate models that
ables can be found in (Lehký et al. 2016). make it possible to evaluate new designs in seconds
Sensitivity analysis of input variables was per­ compared to evaluation of the original model. Within
formed to capture the most significant variables of the the context of reliability assessment, the original limit
model. Only a few of them proved to be significantly state function (LSF, structural response simulated
related to the structural response, hence the stochastic using nonlinear FEM model in this case) is approxi­
model was reduced; see Table 1, where ft and Gf stand mated using a suitable function mostly of polynomial
for tensile strength and fracture energy of transverse type. The reliability indicators (failure probability or
joints concrete, respectively, P1 represents the pre­ reliability index) are then calculated by utilization of
stressing force in the bottom tendon, and g1,n is the sec­ classical simulation techniques using the approxi­
ondary dead load. All the parameters of the reduced mated function as a surrogate model.
stochastic model are defined by their probability dens­ When using the polynomial-based RSM, a second
ity functions, mean values and coefficients of variation order polynomial is most often employed. Approxi­
mated LSF, , of the vector of input random vari­
ables X is defined in the form of:

where Xi, i = 1, …, n are the input basic variables and

parameters a, bi, ci are the unknown regression coeffi­
cients of the approximation function. A polynomial
function suggested in a simplified form can be found
in (Bucher & Bourgund 1990):

with the same notation as in Eq. (1). It is seen that

the Eq. (2) does not contain mixed terms Xi∙Xj.
The regression coefficients in Eqs. (1)–(2) are
obtained by performing a series of numerical “experi­
ments” with input variables selected according to some
“experimental design” based on the mean square error
method. In this case, numerical experiments mean
Figure 1. Longitudinal and transversal section of the bridge. a series of structural FEM analyses where individual

180
realizations of input random variables are sampled
using Latin hypercube sampling (LHS) method. Based
on the presence of mixed terms in the polynomial
approximation function, 1 + n + n(n + 1)/2 or 2n + 1
analyses need to be performed. Although the RSM pro­
vides sufficiently accurate results in reliability applica­
tions, the number of experiments (calculations of
actual LSF) is the most limiting factor, especially in
cases of time-consuming FEM analyses. As obvious,
the number of the required evaluations of the original
LSF significantly increases with the increase in
number of variables, which is often the case in struc­
tural analysis. Therefore, the reduced stochastic models
taking into account the most significant input variables
is recommended to use when applying this method.

3.2 ANN-based response surface method

As an alternative to polynomial type response Figure 2. Flow chart of the small-sample artificial neural
function an artificial neural network-based (ANN) network-based response surface method.
response surface can be used to approximate an
original LSF. An ANN is a mathematical model
inspired by biological neural networks. The ANN
is a signal processing system composed of simple simulations. This number depends on several fac­
processing elements, called artificial neurons, tors such as the complexity of the original function,
which are interconnected by direct links the number of input variables, their statistical cor­
(weighted connections). The aim of such system relation, quality of sample set etc.
is to perform parallel distributed processing in
order to solve a desired computational task. An
3.3 Inverse response surface method
advantage of ANN is its capability to adapt itself
by changing its connection strengths or structure. The RS is an alternative to the real LSF. However, in
Quite a number of neural network types are contrast to the forward approach, when designing
available; in the proposed method a feed-forward the structure, the function values that are used to
multi-layer network type is used. In this type of construct the RS are not available until the desired
network, the artificial neurons are organized into design variables are determined. Therefore, an
different layers (the input layer, the hidden inverse response surface method (IRSM) is pro­
layers, and the output layer) and input signal posed. It is based on a coupling of the adaptive RSM
only moves in the forward direction, i.e. the data of Bucher & Bourgund (1990) and the ANN-based
goes from the input nodes through the hidden inverse reliability method of Lehký & Novák
nodes (if any) to the output nodes. If the output (2012). The method proposed in this paper utilizes
vector of the whole network is required, vectors ANN and LHS methods which make it robust, effi­
have to be calculated layer by layer from the cient and therefore feasible for solving time-
input layer to the output layer of the network. consuming problems such as structural design.
The process of ANN-based response surface The identification of design parameters that
method application is divided into two main phases enable the achievement of the desired reliability
(see Figure 2): (1) the approximation phase, where described by reliability indicators related to particu­
the original LSF is approximated by a suitable lar limit states can be categorized as inverse reliabil­
ANN using a series of numerical experiments, and ity analysis. The parameters to be identified are
(2) the reliability calculation phase, where the deterministic or random design parameters related to
approximated LSF is used instead of the original the structure itself, the acting load or the surrounding
one in combination with classical simulation or environment. The known (in this case desired)
approximation methods (e.g. Monte Carlo) for the response is the safety level described by reliability
calculation of reliability indicators. For more indicators. The functional relationship between
details on ANN-RSM method see Lehký & Šomo­ design parameters and reliability indicators can take
díková (2017). Note that number of “experimental the form of an analytical formulation or a stochastic
samples” needed for adjusting the parameters of an nonlinear FEM model as it is the case in this paper.
ANN-RSM doesn’t directly depend on number of A corner stone of the ANN-based inverse reliability
input random variables like it is in the case of method is an ANN, which is used as a surrogate
a polynomial RSM, see formulas in section 3.1. model of unknown inverse function describing rela­
Thanks to its ability to generalize, the ANN is effi­ tion between the design parameters and correspond­
cient at fitting the LSF with only a small number of ing reliability indicators:
181
 are depicted for a case with just one random variable
X1 and a single design parameter.
1. In the first step of the IRSM, with the initial
where P ¼ d [ r is the vector of all design parameters values for the design parameters, the initial RS is
(deterministic and random ones), I = β (or I = pf) is the constructed using foregoing RSM (see light green
vector of reliability indicators. dash-dotted line in Figure 3 top). Based on this
The efficiency of the inverse method is emphasized approximate response surface the ANN-based
by utilization of the small-sample LHS simulation inverse reliability analysis is carried out and a new
method used for the stochastic preparation of the estimate of design parameters is obtained as well
training set utilized in training the ANN. For that pur­ as the design point. A response surface with a new
pose, the design parameters P (e.g. mean values or estimate of design parameters is depicted by dark
standard deviations of selected random variables) are green dash-dotted line in Figure 3 top.
considered as random variables with a scatter reflect­ 2. In the second step, the new anchor point is calcu­
ing the physical range of design values. Subsequently, lated from the design point using Eq. (4). It
the calculation of reliability is performed using appro­ serves together with the previously obtained
priate simulation or approximation method and reli­ design parameters for the response surface update
ability indicators I are obtained. Once the ANN has (see light blue dotted line in Figure 3 middle).
been trained, it represents an approximation conse­ 3. Based on this updated RS the ANN-based inverse
quently utilized in a following way: To provide the reliability analysis is carried out again to seek the
best possible set of design parameters corresponding new design parameters and the design point. An
to prescribed reliability. See Lehký & Novák (2012) updated RS together with an updated estimate of
for more complex explanation of the method. design parameter is depicted by dark blue dotted
An iterative scheme to update the response sur­ line in Figure 3 bottom.
face and, at the same time, to accomplish the inverse
reliability analysis is proposed in the following This process is repeated until the convergence is
steps. See also Figure 3 where all individual steps achieved at the design parameter with acceptable tol­
erance. In the second step of the IRSM procedure
the new anchor point XM is calculated from the
design point XD, for the surface ~gðXÞ ¼ 0 based on
the assumption of uncorrelated Gaussian variables:

where X is the mean vector. Obtained anchor point

is used as a new center point when preparing a new
set of samples for constructing an updated response
surface. The strategy is to position the new samples
reasonably close to the failure domain and thus
improve accuracy of approximation in the area
which is decisive for the calculation of reliability
indicators.

4 DETERMINATION OF THE BRIDGE

PARAMETERS

According to diagnostic survey, the average value of

concrete compressive strength of joints between
bridge precast segments was about 40 MPa but it
was classified only as C6/7.5 strength class due to
high variability in measurements probably caused by
bridge spatial deterioration. This also brings uncer­
tainty into the actual loss of prestressing. Its value
was roughly estimated according to code specifica­
tions as 17 % which corresponds to the value of pre­
stressing force P1 = 14.20 MN, including actual loss
Figure 3. Individual steps of inverse response surface of prestressing. Since the tensile strength of trans­
updating process. verse joints and the bridge prestressing has

182
a significant effect on the bridge load-bearing cap­ approximations. ANN utilized in inverse reliability
acity, the mean values of both were considered as analyses (see Figure 5) consisted of five nonlinear
uncertain design parameters with the aim to find neurons in a hidden layer and two linear output
their critical values corresponding to desired reliabil­ neurons corresponding to two design parameters ft,mean
ity level and load-bearing capacity. and P1,mean. There were two inputs to the network cor­
Two limit states were taken into account – service­ responding to reliability indices of two analyzed limit
ability limit state of decompression (SLSD) and ser­ states. In order to create the training set, the reliability
viceability limit state of crack initiation (SLSC). The calculations using 1 million Monte Carlo simulations
both limit states have implicit form – structural resist­ have been performed with 50 random samples of
ance is calculated using the nonlinear FE model, load design parameter. After ANN training, the ANN is
action is considered as a deterministic variable placed ready to provide the best design parameters related to
according to the normal loading class scheme. Target the initial response surfaces. This is performed by
reliability indices were considered as β1 = 0 for SLSD, means of a network simulation using target reliability
and β2 = 1.3 for SLSC, respectively. According to indices as inputs.
diagnostic survey and needs of bridge administrator With updated design parameters, updated
desired load-bearing capacity related to normal loading response surfaces have been constructed for the next
class was considered as 25 t. iteration. Here, the stochastic model has been
In order to construct the response surface and to changed with respect to the updated design param­
perform ANN-based inverse reliability analysis the eters and the new anchor points calculated according
design parameters ft,mean and P1,mean have been to Eq. (4), i.e. random sampling was performed in
treated as uniformly distributed random variables a region closer to the design points. Standard devi­
with the initial ranges according to Table 2. Both, ations of the design parameters have been reduced to
a two-degree polynomial response surfaces without half of the original values in order to speed up the
the mixed terms according to Eq. (2) and an ANN- process and improve its convergence.
based response surfaces, respectively, have been Tables 3 and 4 show the values of design param­
used to substitute the original LSFs corresponding eters and reliability indices during iteration process
to both analyzed limit states. Utilized ANNs (see for both response surface types. Here, two iterations
Figure 4) consisted of three nonlinear neurons in
a hidden layer and a linear output neuron corres­
ponding to the safety margin of particular limit
state. There were four inputs to the network corres­
ponding to four input random variables. In order to
calculate unknown coefficients of the response sur­
faces, 50 evaluations of the original LSFs has been
carried out with random samples of input param­
eters generated by LHS method. Note that the same
set of samples has been used for both types of
response surfaces.
Based on the constructed response surfaces the
inverse reliability analyses have been carried out
for both a polynomial as well as ANN-based
Figure 5. A schematic view of an ANN utilized in inverse
reliability analyses.

Table 2. Randomization of the design parameters.

Design parameter Distribution min max

Table 3. Results of iterative process for polynomial
ft,mean Uniform 1.8 3.0 response surface – values of design parameters and corres­
ponding reliability indices.
P1,mean Uniform 12.0 18.0
Identification

Design
parameters Iteration 1 Iteration 2 Target values

ft,mean 3.247 2.799 –

P1,mean 14.436 14.420 –
β1 0.041 –0.008 0.000
Figure 4. A schematic view of an ANN utilized for β2 1.293 1.309 1.300
response surfaces.

183
Table 4. Results of iterative process for ANN-based The number of iterations of the inverse response
response surface – values of design parameters and corres­ surface method depends on the shape of the original
ponding reliability indices. function, the accuracy of the replacement model
and the initial estimation of the design parameters.
Identification Multiple iterations easily confirm the convergence
of the iterative process. Nevertheless, in many prac­
Design tical cases, two iterations are sufficient to obtain
parameters Iteration 1 Iteration 2 Target values a reasonably accurate solution while maintaining an
acceptable computational burden. It is important to
ft,mean 2.916 2.277 – realize that when solving large structures by non­
P1,mean 14.538 14.544 _ linear finite element method, one numerical simula­
β1 0.076 0.005 0.000 tion can last several hours. In such cases, it is
β2 1.384 1.310 1.300 necessary to consider whether the addition of
another iteration, which means performing tens of
simulations, is worth a slight refinement of the
results.
have been enough for reaching acceptable accuracy.
An important aspect of the obtained solution is
Let’s note that reliability indices were calculated by
also the type of the utilized approximation function.
ten million Monte Carlo simulations of response sur­
Two types, namely polynomial function and an artifi­
faces. When comparing both response surface types
cial neural network model, were employed for
one can see a very good agreement when identifying
response function approximations of analyzed
prestressing force. Accuracy for tensile strength of
bridge. The suitability of both models was con­
concrete is slightly lower. This can be explained by
firmed, especially for the resulting values of pre­
the dominant influence of prestressing on response
stressing force which is a significantly more
in both SLSD and SLSC, respectively, while tensile
dominant parameter for both limit states.
strength is important just for SLSC.
Results show that the required mean value of con­
crete tensile strength in transverse joints corresponds ACKKNOWLEDGMENT
to concrete strength class C20/25 or C25/30 depend­
ing on response surface type. It matches to concrete
The authors give thanks for the support provided
type used for transverse joints during bridge con­
struction and as was also confirmed by findings of from the Czech Science Foundation (GAČR) project
diagnostic survey. Note that requirement for reliabil­ RESUS No. 18-13212S.
ity index β2 = 1.3 in case of SLSC is relatively strict
too. For lower values, an even lower demand for
concrete strength would be obtained. REFERENCES
Resulting request to the value of prestressing force
Bucher, C.G., 2009. Computational analysis of randomness
is almost the same as the one estimated according to in structural mechanics. CRC Press/ Balkema, Leiden.
code specifications where loss of prestressing was Bucher, C.G. & Bourgund, U., 1990. A fast and efficient
considered as 17 %. Identified value of prestressing response surface approach for structural reliability
force indicates loss of prestressing equal to 15 %. problems. Structural Safety, 7(1): 57–66.
From both results we can conclude that requirement Červenka, V., Jendele, L. & Červenka, J. 2012. ATENA
for normal load-bearing capacity Vn = 25 t is adequate Program Documentation – Part 1: Theory. Prague: Cer­
and the desired level of safety would be met. venka Consulting.
ČSN 73 6222. 2013. Load bearing capacity of road
bridges. Prague: Czech Office for Standards, Metrology
and Testing. (in Czech).
5 CONCLUSIONS Joint Committee on Structural Safety. Probabilistic Model
Code [online]. Last updated 1. 7. 2015. Available from:
The paper describes application of the inverse http://www.jcss.byg.dtu.dk/Publications/Probabilistic_
response surface method when performing reliabil­ Model_Code.aspx.
ity-based design of selected bridge parameters. It is Lehký, D. & Novák, D., 2012. Solving inverse structural
highly recommended to use proposed adaptive pro­ reliability problem using artificial neural networks and
cedure when solving the inverse reliability problem, small-sample simulation. Advances in Structural Engin­
with the help of the surrogate model. Otherwise, the eering, 15(11): 1911–1920.
identified values of design parameters may not be Lehký, D., Šomodíková, M. & Lipowczan, M. 2016.
Determination of uncertain design parameters of
precise enough, as shown by the results of the first post-tensioned composite bridge. In Bittencourt, T. N.,
iterations in the process of structural analysis. The Frangopol, D. M. & Beck A. T. (Eds.), Maintenance,
reason is the inaccurate first response surface Monitoring, Safety, Risk and Resilience of Bridges and
approximation, which is based on the initial estima­ Bridge Networks – Proceedings of the 8th International
tion of the design parameters. Conference on Bridge Maintenance, Safety and

184
 Management (IABMAS 2016), Foz do Iguaçu, Brazil, artificial neural network-based response surface
26–30 June 2016. London, UK: Taylor & Francis method. Neural Computing and Applications, 28(6):
Group. 1249–1263.
Lehký, D. & Šomodíková, M. 2017. Reliability calcula- Myers, R.H., 1971. Response surface methodology. Allyn
tion of time-consuming problems using a small-sample and Bacon, Inc., Boston.

185
 Bridge Maintenance, Safety, Management, Life-Cycle Sustainability and Innovations –
Yokota & Frangopol (eds)
© 2021 Taylor & Francis Group, London, ISBN 978-0-367-23278-8

Comparison of different low-cost sensors for structural health monitoring

S. Komarizadehasl & J. Turmo
Universitat Politècnica de Catalunya, Barcelona, Spain

B. Mobaraki & J.A. Lozano-galant

Univesidad de Castilla-La Mancha, Ciudad Real, Spain

ABSTRACT: Recently, the need for Structural Health Monitoring (SHM) and inspection is growing as
a result of the increasing age of the structural stock worldwide. SHM includes bringing into use a sensor or an
array of sensors into an engineered structure. This paper evaluates the practical aspects to consider when
choosing different low-cost sensors for using in a SHM application. For this reason, a microcontroller (Ardu­
ino) with attached sensors (Ultrasonic, Laser, Gyroscope, Accelerometer, Temperature-humidity, SD-module)
has been used throughout different controlled ambient situations and experiments. During the sensor installa­
tion and data acquisition, some issues related to the Arduino codes and position of the sensors arise. In this
article solutions for tackling faced problems have been presented for achieving an efficient way of using these
electronic devices. Their data have been collected to assess and compare their tolerances as well as advantages
and disadvantages of their use, cost among them. The information from this article could be used either to
choose the appropriate sensors either for static or for dynamic Structural System Identification (SSI).

1 INTRODUCTION 2.1 Arduino uno

Arduino is an open-source electronics platform based
Lately, the demand for adopting sensors in infrastruc­
on easy-to-use hardware and software. Arduino Uno
tures and buildings is increasing. This need is due to
is a microcontroller board based on the ATmega328P.
the monitoring and inspection of the health state of
It has 14 digital input/output pins, 6 analog inputs to
those structures. Consequently, more economic ways
measure and convert the voltage to a digital value. It
of SHM applications are getting huge attention. In this
facilitates numerous interfaces to communicate with
article, several models of low-cost sensors have been
other microcontrollers and computers such as Inter-
introduced. Furthermore, the features of each one of
Integrated Circuit (I2C), Serial Peripheral Interface
them have been explained. Each sensor could be
(SPI) and Universal asynchronous receiver/transmitter
applied in a particular circumstance. Every sensor has
(UARTA)(Pathak, Kumar, Sharma, & Kumar, 2019).
benefits as well as weaknesses in various conditions.
Figure 1 is the schematic of the Arduino uno circuit.
To be able to work with them, firstly Arduino Uno
(which is the main programmable logic controller
(PLC) in this project) has been introduced. Secondly, 2.2 Ultrasonic senso
diverse types of sensors along with their specifications
Ultrasonic ranging module HC - SR04 provides 2cm
have been introduced. Each type of sensor may use
- 400cm non-contact measurement function which
a different way to send its data. Thirdly, different ways
on Figure 2 its schematic and on Table 1 its charac­
of communicating with this PLC will be presented.
teristic has been presented. The ranging accuracy
Finally, the results of these sensors have been
can reach to 3mm. The modules have transmitters,
illustrated.
receiver and processing circuit. The basic concept of
work is: firstly, using I\O trigger for at least 10uS
(microsecond) high level signal. Secondly, The
2 STATE OF THE ART
Module automatically sends frequency of 40 kHz
and detect whether there is a pulse signal back.
In this section the sensors and a microcontroller that
Finally, the range can be calculated through the time
has been used in the project will be reviewed along
interval between sending trigger signal and receiving
with their technical descriptions.

DOI: 10.1201/9780429279119-21
https://doi.org/10.1201/9780429279119-21

186
 Figure 3. Schematic of a DHT22, digital temperature and
humidity sensor (Patil et al., 2019).

Figure 1. Schematic of the Arduino Uno (Pathak et al.,

2019). Table 2. Technical specification of DHT22 (Patil et al.,
2019) (Liu, n.d.).

Characteristic Description

Working Voltage 3.3-6V DC

Working Current 0.3mA
Sensing element Polymer capacitor
Operating range Humidity: 0-100%RH
Temperature: 40~80Celsius
Accuracy Humidity: +-2%RH
Temperature: <+0.5Celsius
Figure 2. Schematic of an ultrasonic ranging module
HC-SR04 (Kamal & Hemel, 2019). Resolution or sensitivity Humidity: 0.1%RH
Temperature: 0.1Celsius
Repeatability Humidity: +-1%RH;
Temperature: +-0.2Celsius
Humidity hysteresis +-0.3%RH
Table 1. Technical specification of the ultrasonic sensor Long-term Stability +-0.5%RH/year
(Kamal & Hemel, 2019). Sensing period Average: 2s
Interchangeability fully interchangeable
Characteristic Description
Dimensions 14*18*5.5mm
Working Voltage DC 5 V
Working Current 15mA
Working Frequency 20Hz
Max Range 4m 2.3 The VL53L0X
Min Range 2cm The VL53L0X, which has been shown on Figure 4
Measuring Angle 15 degree is a new generation Time-of-Flight (ToF) laser-
Trigger Input Signal 10uS TTL pulse ranging module housed in the smallest package on
Echo Output Signal Input TTL lever signal and the the market today, providing accurate distance meas­
range in proportion urement whatever the target reflectance unlike con­
Dimension 45*20*15mm ventional technologies. The technical specification
of this sensor has been given on Table 3(Adafruit,
2016).
It can measure absolute distances up to 2m, setting
echo signal. Test distance = (time × velocity of a new benchmark in ranging performance levels, open­
sound (340M/S)) (Kamal & Hemel, 2019). ing the door to various new applications. The
VL53L0X integrates a leading-edge SPAD array
2.2.1 Humidity sensor (Single Photon Avalanche Diodes) and embeds
As it has been written on section 2.2, for measuring ST’s second generation FlightSenseTM patented tech­
the distance using that sensor the speed of sound is nology. The VL53L0X’s 940 nm VCSEL emitter
needed. The sound travels in different speeds in dif­
ferent temperatures and humidity. The schematic of
this sensor has been shown on Figure 3. Moreover,
its characteristic has been illustrated on Table 2.
DHT22 has already been calibrated during produc­
tion process and provides accurate information(Patil,
Khedkar, & Jadhav, 2019).
Figure 4. Schematic of a VL53L0X (Adafruit, 2016).

187
Table 3. Technical specification of VL53L0X (Adafruit, Table 4. Technical specification of VL53L1X (Adafruit,
2016). 2018).

Characteristic Description Characteristic Description

Working Voltage 2.6-5.5V DC Working Voltage 2.6-5.5V DC

Working Current 10mA up to 40mA Working Current 10mA up to 40mA
Working Frequency 50Hz Working Frequency 1-50 Hz max sampling rate
Out-put format (I²C) 16-bit distance reading 1-Short: up to (most immune to interference
(in millimeters) ~130 cm from ambient light)
Distance measuring range Up to 2m with minimum 2-Medium: up to 2- 30 Hz max sampling rate
range of 3cm ~300 cm in the dark
Weight without header pins 0.5 g 3-Long: up to 400 cm 3- 30 Hz max sampling rate
Dimension 13*18*2mm in the dark
Out-put format (I²C) 16-bit distance reading (in
millimeters)
Distance measuring Up to 2m with minimum range
range of 4cm
(Vertical Cavity Surface-Emitting Laser), is totally Weight without 0.5 g
header pins
invisible to the human eye, coupled with internal phys­
Dimension 13*18*2mm
ical infrared filters, it enables longer ranging distances,
higher immunity to ambient light, and better robust­
ness to cover glass optical crosstalk (Adafruit, 2016).
device and incorporates the latest InvenSense design
2.4 The VL53L1X innovations, enabling dramatically reduced chip size
The VL53L1X, which has been shown on Figure 5 and power consumption, while at the same time
improving performance and cost. The MPU-9250
together with its specification on Table 4 is a state-of
-the-art, Time-of-Flight (ToF), laser-ranging sensor, MotionTracking device sets a new benchmark for
enhancing the ST FlightSense™ product family. 9-axis performance with power consumption only
It is the fastest miniature ToF sensor on the 9.3μA and a size that is 44% smaller than the com­
market with accurate ranging up to 4 m and fast pany’s first-generation device. Gyro noise perform­
ranging frequency up to 50 Hz Housed in ance is 3 times better, and compass full-scale range
a miniature and reflowable package, it integrates is over 4 times better than competitive offerings
a SPAD receiving array, a 940 nm invisible (Mobaraki & Vaghefi, 2016).
Class1 laser emitter, physical infrared filters, and The MPU-9250 is a System in Package (SiP)
optics to achieve the best ranging performance that combines two chips: the MPU-6500, which
in various ambient lighting conditions with contains a 3-axis gyroscope, a 3-axis accelerom­
eter, and an onboard Digital Motion Processor™
a range of cover window options. Unlike con­
ventional IR sensors, the VL53L1X uses ST’s (DMP™) capable of processing complex Motion-
latest generation ToF technology which allows Fusion algorithms; and the AK8963, the market
absolute distance measurement whatever the leading 3-axis digital compass. Improvements
target color and reflectance. It is also possible to include supporting the accelerometer low power
program the size of the ROI on the receiving mode with as little as 6.4μA of and it provides
array, allowing the sensor FoV to be reduced improved compass data resolution of 16-bits
(Adafruit, 2018). (0.15 μT per LSB) (InvenSense, 2014). On
Figure 6 schematic of this sensor can be seen
(Vaghefi & Mobaraki, 2018).
2.5 MPU 9250
The MPU-9250, delivered in a 3x3x1mm QFN pack­
age, is the world’s smallest 9-axis MotionTracking

Figure 6. Schematic of accelerometer sensor(InvenSense,

Figure 5. Schematic of a VL53L0X (Adafruit, 2018). 2014).

188
3 COMMUNICATION WAYS platform, were a white paper, a black paper, a clear
and transparent plastic cover and some thin tissues.
While many sensors use digital and analog ports for
uploading the measured data to the microcontroller,
4.1 Analysis
some sensors use the inter-integrated circuit (I2C)
protocol. This is a protocol that allows multiple On the Table 5 standard deviations driven from the
“slave” digital integrated circuits (Sensors) to com­ performed tests has been illustrated. On Table 5 ran­
municate with one or more “master” chips (Ardu­ ging result of same experiment under different cir­
ino). Like the Serial Peripheral Interface (SPI) which cumstances has been shown.
is only intended for short-distance communications On this table, the last column has been created to
within a single device [1]. The ultrasonic sensor and provide filtered data from the first laser sensor due to
DHT22 have been connected to the Arduino’s digital the extreme environmental light and heat test. The
ports. The laser ones and the accelerometer had to be filter has deleted the ranging outputs equal to 8190.
connected to I2C port (SCL, SDA) on the board. When this sensor is not able to read or the measure­
Since both of the laser sensors had the same board ment distance is more than its capacity it declares this
addresses, introducing them to the Arduino raised number.
a problem. For solving this issue, the X shut pin of It should be mentioned that the ultrasonic sensor
these 2 sensors has been used to change their circuit which was the chipset sensor and the easiest one to
address. install, had shown better performances compare to the
All the different types of ranging circuits have laser ones. On down side, this sensor needs 5v inter­
been connected and glued together, so data from all action digital ports and needs at least 4 volts for its full
3 of them would be measured almost simultaneously functionality. The only problem with this sensor could
on the static experiment. be its data providing speed. Although the laser has
A different circuit had to be made for the acceler­ a faster rate (50Hz data production), this sensor has
ometer, in order to perform the dynamic experiment. a frequency of only 20Hz. In other words, this sensor
Finally, 2 different codes were written on the Ardu­ can provide up to 20 data each second. The biggest
ino platform and uploaded to the board via a USB problem with the ultrasonic sensors would be their
cable. For getting the main characteristics of these sen­ dependence on the ambient temperature and humidity.
sors, a few dynamic and static tests have been car­ Since the speed of sound changes from an environment
ried on. to another. This sensor needs the accurate speed of
sound for its calculations. The proposition of this
paper would be using the ultrasonic sensor with a laser
4 STATIC EXPERIMENT sensor if there was the probability of a changing tem­
perature or of an extreme ambient light. Using the first
The device was tested against for the purpose of laser sensor or the second one is due to what range and
same measurement against different materials. In the circumstances the experiment may experience.
Figure 7, 2 tests with and without extreme ambient
light has been done for getting the distance from the
big book. For the one with the light bulb, the tem­ 5 DYNAMIC EXPERIMENT
perature sensor has been moved a bit far from the
source of the light and heat. The reason was that the For testing this sensor and its reliability, an experi­
excruciating heat coming from the light bulb would ment has been implemented. With a dynamic jack,
not harm the sensor. The other tested objects on this a sinus signal has been programed and the vibrations
had been saved by the accelerometer. On the Figure 8
a picture of the jack can be seen. This jack can shake
its bottom plate as was programmed. The instructions

Table 5. Results of ranging experiments.

Sensors Ultra Laser1 Laser2

Experiments
a thick book 0.61 2.5 1.5
a white paper 1.87 2.67 1.48
a black paper 1 7.18 1.87
a tissue 352 4.66 1.62
a transparent plastic cover 0.7 5.46 3.12
Extreme ambient light 3.23 3607 21.94
Extreme ambient light* _ 39.86 _
Figure 7. The schematic of static experiment.

189
 had to be glued to the bottom plate of the jack for
getting accurate information. The second problem
was that the written python code could save only
120 data per second while the sensor was reporting
more than 300 data per second. Although by using
a serial port commercial software on the computer it
could have been possible to save data with the same
speed of their production. Since here getting the
accurate time of capture was vital, it was obligatory
to use python to attach the provided data with their
corresponding time. To tackle this problem, the
speed of data capture had to be dialed down, so the
python could get and save them. To be on the safe
side the speed of capture had been set on 84Hz.
The other unexpected issue that this project faced
was that, though this sensor had been calibrated in
the company, it had a constant number added to all
provided data which from now here it would be
named as the white noise. As on the Figure 9 it has
been illustrated, the averaged data is around
-50milig while they had to fluctuate around zero.
This -50milig was considered as the white noise of
Figure 8. The hydraulic jack subjected to the experiment. this sensor. In order to measure this correctly, the
average of 10000 set of data in a vibration free test
has been calculated. For this sensor the white noise
to the hydraulic jack was to make a wave with a fixed had been calculated as -49.8535 milig.
frequency 5 hertz (5 complete waves in one second). By removing this amount from the provided
The movement of the jack was to go up to 0.1 milli­ accelerations, the values had been pulled up where
meter up and -0.1 millimeter down from its null axis they needed to be. The data were clearer and more
to make a sinus wave. With a very simple two time
differential, the acceleration equation could be calcu­
lated. A picture of the used dynamic jack can be seen
on Figure 8.

In the above equation y is the displacement in

time t, d is the maximum allowed movement of
the jack in each cycle, ω is the angular frequency
and f is the set frequency which equal to 5Hz
and ’ is the phase constant. On the equ.3, accel­ Figure 9. Acceleration, time diagram with white noise.
eration has been calculated from the equ.1 and
equ.2. this was done by getting the second order
derivative of the equ.1. By putting all the data in
the equ.3 the maximum acceleration was calcu­
lated as 10.4352 g*10-3 m/s2.

5.1 Analysis
The very first faced problem in this experiment was
that the sensor could not record data or if recorded, Figure 10. Acceleration, time diagram without the white
the data were messy. It was deducted that the sensor noise.

190
understandable when this white noise had -1-R directed by José Turmo and BIA2017-86811-C2
been removed. This improvement can be seen on -2-R, directed by Jose Antonio Lozano-Galant. All
Figure 10. Different accelerometer boards had been these projects are funded with FEDER funds. Authors
tested and it was concluded that each circuit has its are also indebted to the Secretaria d’ Universitats
own different amount of white noise that must be i Recerca de la Generalitat de Catalunya for the fund­
dealt with. It should have been assured that this ing provided through Agaur (2017 SGR 1481). It is
white noise is constant. Moreover, the circumstances also to be noted that funding for this research has been
and time cannot change it. For this reason, the jack provided for MR. SEYEDMILAD KOMARIZADE­
was programed with many more different frequen­ HASL by Agencia Estatal de Investigación del Minis­
cies and displacements. In all of those experiments terio de Ciencia Innovación y Universidades grant and
the white noise stayed the same. the Fondo Social Europeo grant (PRE2018-083238).

6 CONCLUSIONS ORCID

In this section, the overall result deduction of the Seyedmilad Komarizadehasl:

https://orcid.org/0000-0002-9010–2611
given experiments has been briefly presented.
Behnam Mobaraki:
https://orcid.org/0000-0002-2924-643X
6.1 Static experiments Jose Antonio Lozano Galant:
http://orcid.org/0000-0003-0741–0566
Notwithstanding that the laser sensors did not have Jose Turmo:
as good results as the ultrasonic one, they can be https://orcid.org/0000-0001-5001–2438
useful as well. They are smaller, they are lighter and
faster and have noise-free technology (no noise can
enter from the wires). Moreover, they work inde­ REFERENCES
pendent of the temperature of their testing situation.
Adafruit. (2016). Adafruit VL53L0X Time of Flight
Best results would only appear if a ultrasonic sensor
Micro-LIDAR Distance Sensor Breakout. Retrieved from
(attached to its temperature and humidity sensor) be https://www.st.com/resource/en/datasheet/vl53l0x.pdf
used alongside of a Laser sensor. They can cover the Adafruit. (2018). STMicroelectronics, VL53L1X - A new
downsides of each other and provide accurate, useful generation, long distance ranging Time-of-Flight sensor
set of data. The selection between Laser type one or based on ST’s FlightSenseTM technology. Retrieved from
two depends on the circumstances of the experiment. https://www.st.com/resource/en/datasheet/vl53l1x.pdf
If there is enough budget, the Laser type 2 provides InvenSense, T. (2014). MPU-9250, Nine-Axis (Gyro+ Accel­
way better set of data and is less sensitive to ambient erometer+ Compass) MEMS MotionTrackingTM Device.
light. Kamal, A., & Hemel, S. (2019). Comparison of Linear Dis­
placement Measurements Between A Mems Accelerom­
eter and Hc-Sr04 Low-Cost Ultrasonic Sensor.
6.2 Dynamic experiment Ieeexplore.Ieee.Org. Retrieved from https://ieeexplore.
ieee.org/abstract/document/8934569/
As on the Figure 10 it is visible the sinus wave con­ Liu, T. (n.d.). Digital-output relative humidity & temperature
ducted from the accelerometer is quite close with the sensor/module DHT22 (DHT22 also named as AM2302)
expected behavior. As it is observable the sinus Capacitive-type humidity and temperature module/sensor.
wave is fluctuating about 10.5 milig from its aver­ Retrieved from http://www.datasheet4u.com
age. As it was calculated on the last section the Mobaraki, B., & Vaghefi, M. (2016). Effect of the Soil
graph should have had a 10.453 milig fluctuation. It Type on the Dynamic Response of a Tunnel under
is quite notable to see that they have worked out Surface Detonation. Original Russian Text c B.
Mobaraki, M. Vaghefi. Published in Fizika Goreniya
almost the same. i Vzryva, 52(3), 119–127. https://doi.org/10.1134/
On the Figure 10 the filtered data from the 5Hz S0010508216030175
experiment has been shown. As it can be seen, the Pathak, B., Kumar, P., Sharma, A., & Kumar, R. (2019).
result is not so accurate, for the fluctuation has other Earthquake Detector using Arduino Uno. Retrieved
unexpected data or noises as well. In the future from http://122.252.232.85:8080/jspui /handle/
works, filters must be applied to delete the unwanted 123456789/22911
data and ambient noises which may have entered in Patil, T., Khedkar, S., & Jadhav, M. (2019). Internet
to this experiment unwantedly. of Things (IoT) based Warehouse Monitoring and
Control Interface Implementation. Retrieved from
https://www.academia.edu/download/60437773/
IRJET-V6I678520190830-121843-gk14fp.pdf
ACKNOWLEDGEMENTS Vaghefi, M., & Mobaraki, B. (2018). INVESTIGATION
OF DYNAMIC BEHAVIOR OF CONCRETE BRIDG
The authors are indebted to the Spanish Ministry of E COLUMNS UNDER SURFACE BLASTING. Jour­
Economy and Competitiveness for the funding pro­ nal.Sharif.Ir. Retrieved from http://journal.sharif.ir/
vided through the research project BIA2017-86811-C2 journals/sjce/?_action=export⌋bibtex⌉0,1339

191
 Bridge Maintenance, Safety, Management, Life-Cycle Sustainability and Innovations –
Yokota & Frangopol (eds)
© 2021 Taylor & Francis Group, London, ISBN 978-0-367-23278-8

Gaussian curvature as an indicator used for damage detection of bridge

structures
T. Wu, L. Tang, C.Y. Jian & R.Y. Mao
State Key Laboratory of Mountain Bridge and Tunnel Engineering, Chongqing Jiaotong University, Chongqing,
China

Z.X. Zhou
College of Civil and Transportation Engineering, Shenzhen University, Guangdong, China

ABSTRACT: Aiming at the shortcomings of the current small and medium-sized bridges in the daily safety
inspections, a method for damage detection of structures using Gaussian curvature was proposed. According to the
bending theory of the elastic thin plate, the qualitative relationship between the structural stiffness distribution and
the bridge deck bending degree, Gaussian curvature, is obtained and verified by numerical examples and a model
polymethyl methacrylate (PMMA) bridge. A simple supported T-beam model was established in the finite element,
and different heights, positions, numbers of cracks were set on the web of the main beam to simulate the structural
stiffness damage. The analysis results show that the stiffness distribution of the structure can be clearly seen
through the Gaussian curvature map of the bridge deck, and the cracks in the structure can be accurately identified
by the discontinuities in Gaussian curvature map. Finally, a loading test was carried out on a PMMA bridge with
crack in the laboratory, and the crack was successfully detected by the Gaussian curvature method. The research
results in this paper can provide a reference for the damage identification of structures with Gaussian curvature.

Keywords: bridge health monitoring, stiffness distribution, damage detection, Gaussian curvature

1 INTRODUCTION residual force error and deflection error into the identi­
fication problem of structural damage, and studied the
Bridge structures are susceptible to deterioration and iterative algorithm for solving the optimization prob­
damage during their service lifetime. If an initial lem. Because the measured information is not very
damage in these structures is not detected timely, it much, and the influence of noise is added to the meas­
can lead to increased damage and even catastrophic urement, a non-linear equation which is not very accur­
failure. Therefore, Structural Health Monitoring ate is established by using the algorithm. Cui et al.
(SHM) is essential to ensure the safety of the bridge (2000,2003) solved these problems by gradient
over its service life. method, Gauss-Newton (G-N) method and Monte-
It can be known from structural mechanics that Carlo method. Zhang & Fan (1998) studied an algo­
when the external load is constant, if the stiffness of rithm based on model modification of structural stiff­
the structure changes, it will inevitably be reflected in ness, and then derived another method, which
the d/eformation. Damage localization based on struc­ combines vibration modal parameters and static dis­
tural characteristic deflection shapes or their spatial placement measurements. On the basis of static meas­
derivatives has been widely investigated due to their urement data, Cai et al. (2004) studied a probabilistic
high sensitivity to local damage. Wang et al. (2001) analysis method, damage identification and state evalu­
proposed a structural damage identification algorithm ation of existing structures. By using finite element
using static test data and natural frequency variation: method, the structures were discretized into analytical
firstly, the damage index is used to locate the damage models expressed by basic parameters and matrices.
of the structure, then the iterative calculation is used to By analyzing the gap between the models and the solid
solve the nonlinear optimization problem, and the structures, the G-N method was used to deduce the
damage degree is estimated. Banan et al. (1994) trans­ results. Then two methods of parameter identification
formed the optimization problem of minimizing

DOI: 10.1201/9780429279119-22
https://doi.org/10.1201/9780429279119-22

192
are derived. According to the relationship between Where, is the bending stiffness of the
bending moment and curvature in material mechanics. plate, E is the elastic modulus, v is the material Pois­
3D laser scanning technology and close-range son’s ratio, and h is the thickness of the plate. Bring
photogrammetry technology can obtain bridge geom­ Eq. (2) into Eq. (3), then:
etry data more efficiently and comprehensively (Gikas
& Vassilis, 2012; Khuc et al., 2018; Matsue et al.,
2019; Teza et al., 2009; Vaghefi et al., 2015; Yang
et al., 2014). However, at present, its application at
home and abroad is limited to obtaining structural sur­
face morphology images, and the purpose of visual
inspection or deformation measurement of the structure
is achieved by analyzing the surface images. There is Curvatures and torsion expressed by stiffness can
little research on the inversion of structural stiffness be obtained by solving eq. (4).
reduction through the deformation shape feature. In
this paper, the bridge structure is regarded as elastic
bending thin plate. The relationship between the stiff­
ness distribution of the bridge and the bending degree
of the bridge deck, that is, the Gaussian curvature, was
explored. The method of detecting the damage of the
bridge by Gaussian curvature was proposed and was
verified by a numerical example and a PMMA model
bridge in the laboratory. 2.2 Sensitivity analysis of Gaussian curvature to
structural damage
2 SENSITIVITY OF GAUSSIAN CURVATURE Suppose the surface w ¼ wðx; yÞ have continuous
TO STRUCTURAL STIFFNESS DAMAGE second-order partial derivatives. Let E, F and G as
the first basic quantities, L, M and N as the second
2.1 Relation between Gaussian curvature basic quantities. The first and second basic forms of
of deflection and bending stiffness the surface are:
For shell structures, the three-dimensional problem
can be reduced to a two-dimensional problem by
introducing the Kirchhoff’s straight-line hypothesis.
If w is the deflection of the plate, that is, the dis­
placement surface of the plate along the Z-axis, the
deflection equation of the plate is (Han, 2002):
The normal curvature of any point on the surface is:

The curvature of a point on the surface is

the second derivative of the deflection equation:

Suppose k1 ¼ðkn Þmax , k2 ¼ðkn Þmin , the formula for

calculating the principal curvatures are as follows:

Where, kx is the curvature in the X-axis direction;

ky is the curvature in the Y-axis direction; kxy is the
twist rate of the XY plane.
Where,

193
It can be proved that the first order differential can 3 SIMULATION EXAMPLES
be neglected when surface w ¼ wðx; yÞ is approxi­
mated to a thin plate with minimal bending degree. 3.1 Gaussian curvature of intact structure
Let p ¼ 0 and q ¼ 0, the formula for calculating the
principal curvatures can be obtained from eq. (10) A finite element model of a simply supported T-beam
(Li et al., 2003): bridge with a span of 40 m is simulated to verify the
proposed method. They represent the simplest and
common structural forms of bridges. The bridge con­
sists of five T-beams. Each T-beam is 1.8 m high,
2.0 m wide and the thickness of the web is 0.2 m. Two
end diaphragms and three middle diaphragms are
arranged vertically. The material properties of the beam
Where, kx ¼ 1r , ky ¼ 1t , kxy ¼ 1s are as follows: Young’s modulus E = 3.45×104 MPa,
From the calculation formula of Gaussian Poisson ratio μ=0.2, density ρ=2500 kg/m3. Figure 1.
curvature: shows the sketch of the simply supported T-beam
bridge model.
The structural damage of the model bridge is
simulated by setting artificial cracks on the web of
the main beam. The main beams of the model are
By introducing Eq. (5) and Eq. (10) into Eq. (11), divided into 20 equal-length segments. The seed size
a formula for calculating the Gaussian curvature of the whole model is 0.01 m×0.01 m. The Gaussian
expressed by the stiffness of an elastic thin plate can curvature variation law of bridge deck under dead
be obtained. load is analyzed first.
After ABAQUS is used to calculate and extract
the deformation of bridge deck under dead load, the
Gaussian curvature of each point is estimated by
MATLAB. Then a surface is fitted by interpolation
(Figure 2). Figure 2 (a) shows the Gaussian curva­
ture map of the bridge deck of the intact bridge. The
dotted lines in the Y direction indicate the location
of the diaphragms and the dotted lines in the
X direction indicate the location of the main girder
webs. As webs and diaphragms reinforce the stiff­
ness of the deck, the Gaussian curvatures will
It can be seen from Eq. (12) that when the decrease at the location of webs and diaphragms, as
bending stiffness decreases, the Gaussian curva­ is stated above. Figure 2(b) is a three-dimensional
ture will increase relatively, and when the bend­ map of the Gaussian curvatures. Where the webs and
ing stiffness increases, the Gaussian curvature diaphragms are located, the surface is depressed
will decrease relatively. Next, a numerical downward, and the Gaussian curvature value
example of a simply supported T-beam bridge is decreases. the positions of the web and diaphragm
presented to demonstrate the damage identifica­ can be clearly seen through the Gaussian curvature
tion method using Gaussian curvature as the map. Through the distribution of Gaussian curvature
index of structural stiffness change. of bridge deck, the stiffness distribution of the

Figure 1. Sketch of the simply supported T-beam bridge model/cm.

194
Figure 2. Gaussian curvature cloud map of the bridge deck without damage.

bridge can be “perspective”, which is the basis of Figure 4~Figure 7 show several Gaussian curva­
damage detection in the following sections. ture maps of different damage scenarios. The rela­
tionship between structural damage and Gaussian
curvature can be analyzed intuitively.
3.2 Damage detection with Gaussian curvature
From Figure 4~Figure 5, it can be seen that when
From the previous analysis, it can be seen that when one crack takes place in the structure, the Gaussian
the structure is damaged, the Gaussian curvature at curvature map at the cracked position suffers local
the damaged position will suffer a local discontinu­ discontinuities. Because the crack reduces the local
ity, and the more serious the damage is, the greater stiffness of the structure in the corresponding area,
the peak value of Gaussian curvature is. In this sec­ then the Gaussian curvature value in this region
tion, some cracks are set in the web of main beams increases. By comparing the Gaussian curvature
as local stiffness reduction according to Table 1. maps of different damage scenarios, it can be seen
that when the damage severity is relatively small
(Figure 4), the increase of Gaussian curvature caused
Table 1. Different damage scenarios on simply supported by damage is in the same order of magnitude as the
beam. decrease of Gaussian curvature caused by webs and
diaphragms, both of which can be seen from the map
Damage Number Heights of clearly. With the increase of damage severity, the
scenario of cracks Position of crack crack/cm
Gaussian curvature caused by damage increases sig­
1 1 15m of 3rd beam 30
nificantly, and the discontinuities of Gaussian curva­
2 1 15m of 3rd beam 40
ture at other locations are no longer obvious
3 1 15m of 3rd beam 50
(Figure 5).
4 1 15m of 3rd beam 60
As can be seen from Figure 8, with the increase of
5 1 15m of 3rd beam 70
crack height, the peak value of Gaussian curvature
6~25 1 1m, 2m, 3m, ……, 50 increases exponentially. Figure 9 depicts the relation­
19m, 20m of 1st beam ship between the maximum Gaussian curvature values
26~45 1 1m, 2m, 3m, ……, 50 and the longitudinal positions of the crack in 1st beam
19m, 20m of 2nd beam to 3rd beam. As the crack position approaches the
1m, 2m, 3m, ……, middle span, the maximum value of Gaussian curva­
46~65 1 50 ture increases gradually, but it decreases slightly as the
19m, 20m of 3rd beam
66 3 11m, 15m, 19m of 2nd 50 crack approaches the diaphragms (10m and 20m),
beam because the diaphragms increase the stiffness of bridge
67 5 11m, 13m, 15m, 17m, 50 deck, and then reduce the influence of cracking on
19m of 2nd beam structural stiffness. It also can be seen that the

195
maximum Gaussian curvature value is approximately 4.2 Damage detection of the beam using Gaussian
equal when the crack is located in the middle beam curvature
(2nd beam and 3rd beam) and is obviously larger than
that when the crack is located in the side beam (1st
beam).
Figure 6~Figure 7 are Gaussian curvature maps of
multiple cracks (scenario 66 and scenario 67). When
cracks take place in the structure, the Gaussian curva­
ture map at the cracked position suffer local discon­
tinuities, and the number of which is the same as the
number of cracks. The maximum value of Gaussian
curvature for multiple cracks is slightly smaller than
that for single crack. Cracks will reduce the stiffness of
the structure in the area around the cracks, so the gradi­
ent of the reduction of stiffness by multiple cracks will Figure 3. The original point cloud from P50.
tend to be flat, which is manifested in the reduction of
the maximum value of Gaussian curvature.
Figure 3 shows the original point cloud of the whole
bridge. In CYCLONE software, the redundant point
4 EXPERIMENTAL VALIDATIONS cloud was deleted, leaving only the point cloud of

4.1 Description of the laboratory experiment

The applicability of the proposed method for
damage detection on a real structure was studied on
an experimental beam in the laboratory. The beam
with a span length of 1.60 m and the width is 0.80m.
The bridge consists of five T-beams, which
numbered1st~5th from front to back. Each T-beam is
100 mm high, 160 mm wide and the thickness of the
web is 20mm. Two end diaphragms and three middle
diaphragms are arranged vertically. The material
used were PMMA.
In order to increase the reflectivity of laser on Figure 4. Bridge surface Gaussian curvature of scenario 1.
the bridge deck, a layer of grinding paper was
pasted on the bridge deck. As the density of PMMA
is small, so the deflections caused by the self-
weight is small, too. In order to increase the deflec­
tions, a uniform load was added to the webs of each
main beam. To do this, first, a row of hooks was
glued on the web. Then the weights were hanged on
the hooks.An artificial small crack was arranged on
the web of the 5th T-beam, which is 700 mm from
the right end of the 5th beam. The longitudinal
width of the crack along the bridge is 5 mm, and
the height is 30 mm.
In this experiment, Terrestrial Laser Scanner
(TLS) Leica ScanStation P50 (P50) was selected to Figure 5. Bridge surface Gaussian curvature scenario 5.
obtain the bridge deck’s deflection data. P50 is an
innovative 3D laser scanner launched by Leica
Switzerland. The operation is simple and remote
operation control can be realized. The advantages of
TLS have already been discussed with respect to
bridge inspection/monitoring. Within several
research projects, TLS methods were applied for
deflection measurement or bridge monitoring under
dynamic excitation.
Because of the small scanning range in this
experiment, one-station scanning is enough without
any transfer station and target setting. This produced
very high precisions for the deflection. Figure 6. Gaussian curvature map of scenario 66.

196
the bridge deck. is the Projection of original point
cloud on XOY plane. Since the origin of the point
cloud data is at the scanner location, in order to
facilitate the later data processing and unify the
point cloud coordinates before and after the damage,
it is considered to first transform all the point clouds
so that the origin of the transformed coordinate
system is located at the end of the bridge deck, the
long side of the bridge deck coincides with the
X axis, and the short side coincides with the Y axis.

Figure 10. The crack on the web.

Figure 7. Gaussian curvature map of scenario 67.

Figure 11. Data acquisition with leica scanstation P50.

Figure 8. Relationship between the crack heights and the

maximum Gaussian curvature. Figure 12. Projection of original point cloud on XOY plane.

Figure 13. Projection of point cloud on XOY plane after

coordinate transformation.
In essence, point cloud data is a scattered array of
numerous points, and the XY coordinates of each
point before and after the damage can not correspond
Figure 9. Relationship between the crack heights and the
one by one, so it is necessary to interpolate point
maximum Gaussian curvature. cloud data to form a surface (Figure 14). After the

197
 explored. By setting different severities, different loca­
tions and different numbers of cracks in the finite elem­
ent model, the relationship between damage properties
and Gaussian curvature of deflection under dead load
is studied. Finally, a loading test was carried out on
a PMMA bridge with crack in the laboratory, and the
crack was successfully detected by the Gaussian curva­
ture method. The following conclusions can be drawn.

Figure 14. Surface formed by point cloud interpolation. 1. The Gaussian curvature of deflections under load
has a significant relationship with the stiffness dis­
tribution of the structure. In case of local structural
damage, the continuous stiffness surface will pre­
sent a disruption. Therefore, discontinuity is also
shown on the Gaussian curvature map, each stiff­
ness change leads to discontinuities in the course
of the Gaussian curvature map. In the position
where the local stiffness of the structure is larger
(webs and diaphragms), the Gaussian curvature
map sinks downward.
2. The peak point of Gaussian curvature clearly indi­
cates the damage location of bridges: when the
main beam web cracks, the Gaussian curvature
Figure 15. The Gaussian curvature of deflection map of bridge deck will suffer discontinuities
difference.
(increase). The coordinates and number of max­
imum values indicate the location and number of
formation of the surface, in order to detect the cracks. The maximum value of Gaussian curva­
damage of the structure using the curvature character­ ture increases exponentially with the increase of
istics, it is necessary to resample the deflection value. crack height. And the maximum value of Gauss­
Then the deflection of several points of the bridge ian curvature of multiple cracks is slightly smaller
deck before and after the damage at a certain interval than that of a single crack.
were extracted. The deflection value of each point
after damage was subtracted from the deflection value The Gaussian curvature can be used as the
before damage to get the deflection difference. Then index of the stiffness change of bridge structures.
the Gaussian curvature of deflection difference was The stiffness change caused by constructional
calculated and used for damage detection. details such as webs and diaphragms and the
In the following, the results of the damage detec­ stiffness reduction caused by damages can be
tion of the model bridge will be discussed by using seen through the Gaussian curvature map. This
the Gaussian curvature method. The Gaussian curva­ method is able to detect and localize all kind of
ture values from the deflection data were calculated stiffness reduction due to local damage of
using MATLAB and the results were presented in a bridge structure without the need of a reference
Figure 15. measurement of the structure in undamaged
The deflection surface of the damaged bridge deck condition.
under uniform load is shown in Figure 14, which
showed no clear localization of the cracked area. ACKNOWLEDGEMENT
However, the localization of the crack pattern area
was successful for the uniform load using the Gauss­ This paper is supported by the National Natural Sci­
ian curvature map. In Figure 15, a discontinuity of the ence Foundation of China (No.51708068 and
Gaussian curvature map located exactly the position No.51778094) and the Chongqing Jiaotong Univer­
of the crack. The maximum value of the Gaussian sity Graduate Education Innovation Fund Project
curvature is about 1.5×10-6. At other locations in the (No.2019S0148).
map, the Gaussian curvature does not change much. We declare that we have no conflict of interest.

5 CONCLUSIONS REFERENCES
Banan, M. R., Banan, M. R., & Hjelmstad, K. D. 1994. Par­
In this paper, the relationship between the damage of
ameter estimation of structures from static response. ii:
bridges and the Gaussian curvature of deflection under numerical simulation studies. Journal of Structural
dead load is deduced by using the theory of bending of Engineering, 120(11), 3259–3283.
elastic thin plate, and the sensitivity of Gaussian curva­ B. Güldür Erkal, & N. M. Apaydın. 2017. Bridge surface
ture to the damage of structural local stiffness is damage detection application with a laser-based

198
 software prototype. ISPRS - International Archives of Matsue, K., Matsumoto, T., Kakuma, K., Nishi, H.,
the Photogrammetry, Remote Sensing and Spatial Infor­ Awaludin, A., & Matsumoto, T., et al. 2019. Image
mation Sciences, XLII-4/W6, 55–57. analysis on the flexural behaviors of rc beams with
Cui F, Yuan W., & Shi J. 2000. Damage Detection of Struc­ artificial damage and the effects of repair. MATEC
tures Based on Static Response. Journal of Tongji Uni­ Web of Conferences, 258.
versity: Natural Science Edition. 28(1):5–8. Teza G, Galgaro A, & Moro F. 2009. Contactless recognition
Cui F, Yuan W., & Shi J. 2003. Bridge Structural Damage of concrete surface damage from laser scanning and curva­
Identification Based on Static Test. Bridge construction. ture computation. NDT&E International. 42(4):240–249.
2003(2):4–7. Vaghefi, Khatereh, et al. 2015. Combined Imaging Tech­
Cai J., Wu Z., & Li Z. 2004. Statistical Approach to nologies for Concrete Bridge Deck Condition
Evaluating the Status of an Existing Structure Using Assessment. Journal of Performance of Constructed
Static Test Data. ENGINEERING MECHANICS. 21 Facilities. 29(4):1–8.
(6):76–83. Wang, X., Hu, N., Fukunaga, H., & Yao, Z. H. 2001. Struc­
Gikas, & Vassilis. 2012. Three-dimensional laser scanning tural damage identification using static test data and
for geometry documentation and construction manage­ changes in frequency. Engineering Structures, 23(6),
ment of highway tunnels during excavation. Sensors, 12 610–621.
(12), 11249–11270. Yang H, Xu X, & Neumann I. 2014. The Benefit of 3D
Han Q. 2002. Higher plate theory. Laser Scanning Technology in the Generation and Cali­
Khuc, Tung, Catbas, F., & Necati. 2018. Structural identifi­ bration of FEM Models for Health Assessment of Con­
cation using computer vision-based bridge health crete Structures. Sensors. 14(11):21889.
monitoring. Journal of Structural Engineering. Zhang Q., & Fan L. 1998. Damage Detection for Bridge
Li Z., Zeng Z., & Luo W. 2003. A new approach for Structures Based on Dynamic and Static Measurements.
predicting fractures using principal curvature. Pet­ Journal of Tongji University: Natural Science Edition.
roleum Exploration and Development. 30(6):83–85. 1998(5):528–532.

199
 Bridge Maintenance, Safety, Management, Life-Cycle Sustainability and Innovations –
Yokota & Frangopol (eds)
© 2021 Taylor & Francis Group, London, ISBN 978-0-367-23278-8

Statistical modal analysis for bridges under ambient excitation by using

improved random decrement technique and wavelet transform
Jianfeng Liu
Department of Bridge Engineering, Tongji University, Shanghai, China

Qiwei Zhang
State Key Laboratory for Disaster Reduction in Civil Engineering, Tongji University, Shanghai, China

ABSTRACT: A new statistical identification method for modal parameters of bridges using structural
ambient vibration responses is presented. The method is composed of improved random decrement
technique, wavelet transform and bootstrap method. The improved random decrement technique can be
used to deal with zero-mean non-stationary ambient vibration signals directly. The wavelet transform is
applied to decouple dense frequency and low damping system because of its advantages in time-
frequency domain. The bootstrap procedure is employed to evaluate and decrease the uncertainty of the
identification results. The reliability of the proposed method has been validated by a four-degree-of­
freedom numerical example. Then, modal parameters of bridges under ambient excitation are estimated
using the presented method. The identification accuracy of the method is higher than the traditional
time-domain method and the wavelet transform only method, especially for structural damping ratios.
The precise identified results can be used to be the foundation of structural health monitoring during
the operational phase.

1 INTRODUCTION domain method, time domain method and time-

frequency domain method. As a new mathemat­
With the increasing of long-span bridges and the ical tool developed in recent years, wavelet
frequent occurrence of bridge accidents, the transform has been widely used in signal pro­
safety and durability of structures have gradually cessing due to its advantages in the field of
become the focus of attention. Bridge health time-frequency domain analysis. Staszewski
monitoring has been widely used. Modal param­ et al. summarized three methods for identifying
eter identification is an important part of bridge structural damping ratios based on Morlet wave­
health monitoring. It is the precondition for finite let reconstruction on asymptotic signals [1].
element model modification, structural damage Kijewski et al. analyzed the frequency resolution
identification and condition assessment. The and boundary effect encountered in wavelet
modal parameter identification under environmen­ transform for parameter identification, and pro­
tal excitation is the process of identifying the nat­ posed the method for reducing the effect of
ural frequency, damping ratio and mode shape of boundary effects by adding data at both ends of
the bridge through the measured structural the boundary and improving the frequency reso­
dynamic response under ambient excitations such lution by selecting the center frequency of
as wind load, vehicle load, and ground pulsation. Morlet wavelet [2]. Lardies et al. used the trad­
Because the entire process does not require exter­ itional random decrement technique to convert
nal excitation equipment and does not affect the environmental excitation response of a multi­
normal traffic, it has become the main means of degree-of-freedom system into an impulse
bridge modal parameter identification, especially response, then used the wavelet transform to
for long-term bridge monitoring and condition estimate the natural frequency and damping ratio
assessment. of the system[3]. Min Zhihua et al. [4] proposed
Traditional modal parameter identification a modal parameter identification method combin­
techniques are mainly composed of frequency ing wavelet transform and singular value

DOI: 10.1201/9780429279119-23
https://doi.org/10.1201/9780429279119-23

200
decomposition, which showed that the identifica­ (t) denotes the r-th component of the modal displace­
tion accuracy was higher than the direct wavelet ment response q(t). In Eq. (2), qr(t) can be further
transform method. Liu Jianfeng et al. proposed expressed by combining the homogenous solution
a statistical identification method of structural qrh(t) and particular solution qrp(t) in the following
modal parameters based on improved random equation:
decrement technique and wavelet transform
(IRDT-Bootstrap-WT) [5]. Theoretical derivation
and numerical verification show that the identifi­
cation accuracy of the method is higher than
direct wavelet transform and traditional time-
domain method, especially for damping ratios.
At the same time, the method also has good
anti-noise capability.
In this paper, based on the theoretical and numer­ Where hr ðt _ τÞ is the unit impulse response func­
ical verification[5], the IRDT-Bootstrap-WT method tion for a viscous-damped system.
is further used to identify the structural modal
parameters under environmental excitation. The field
feasibility of the method is studied by the on-site
vibration tests of a truss composite arch bridge.

2 A NEW MODAL PARAMETER The modal velocity q(t) can be obtained by differen­
IDENTIFICATION METHOD tiating q(t)

In this section, a new statistical method for modal par­

ameter identification based on the improved random
decrement technique and wavelet transform (IRDT­
Bootstrap-WT) is presented, which is proper to the
structures with low frequencies, dense modes and low
damping ratios. The improved random decrement tech­
nique extracted responses from the non-stationary sig­
nals directly. Then data are expanded by Bootstrap
method. Wavelet transform is used to decouple the
system with dense modes and low damping ratios. The
final identification results are in the form of bootstrap
mean value and approximate percentile confidence
Eq. (6) can be derived from Eq. (5).
intervals. The main principles of the method is derived
as follows:
The standard matrix equations of motion of
a discrete linear system subjected to a nonstationary
zero-mean random excitation f(t) can be expressed as

Substituting t into t þ τ, Eq. (3) will be written as

where M, C, and K are the structure’s global mass,
damping, and stiffness matrices respectively, Intro­
duce the transformation as follows:

Where � denotes the complex modal matrix, q(t)

denotes a vector of the modal displacement response, Inserting Eq. (3) and Eq. (6) into Eq. (7), Eq. (8) can
’r denotes the vector of the r-th mode shape, and qr be obtained

201
 method most often supplies a single estimate of the
modal parameter for each time history, and the
repeated measurements for the real structure are also
restricted. No enough information for appropriate stat­
istical description can be used to consider and evalu­
ate the impacts of uncertainties derived from the
noise, measurement errors and so on inherent on iden­
tification accuracies. Due to these limitations,
The traditional random decrement technique divides a simple but effective bootstrap scheme is proposed in
the response into N time segments, each with the conjunction with the wavelet transform to consider
same time length. All of these segments have the the influences of uncertainty on dynamic parameters.
same trigger value. Through the ensemble averaging The bootstrap theory pioneered by Efron was
of the N sample segments, the following time func­ originally introduced to evaluate the statistical
tion can be obtained. accuracy by calculating the confidence intervals
of the random variables with unknown probabil­
ity distribution and limited data or samples. The
typical bootstrap method can bedescribed as
follows.
(1) Suppose a random independent sample
where δðτÞ denotes the randomdec signature X ¼ ðx1 ; x1 ; _ _ _; xn Þ drawn from an
obtained from the ensemble average of the unknown identical distribution F(x). Its statistic
N segments.
inference of interest is . A bootstrap
Inserting Eq.(8) into Eq. (9),
sample X � ¼ ðx1 � ; x1 � ; _ _ _; xn � Þ can be
gained by sampling with replacement from the
original data randomly. Its corresponding esti­
mation is .
(2) Through replicating the Bootstrap extraction
B times, B numbers of Bootstrap ensemble
ðθ1 � ; θ2 � ; _ _ _; θn � Þ are produced.
(3) From the histogram of the B numbers of Boot­
strap ensemble, the probability density function

Eq. (10) contains the free decay vibration response can be determined and the standard deviation
of the values ðθ1 � ; θ2 � ; _ _ _; θn � Þis the esti­
and force
mate of the standard error of s(x). The equa­
vibration response listed in the flowing Eq. (11). tion is,

Where,

The improved random decrement technique can be

applied to the zero-mean non-stationary environmen­
tal vibration response directly. Force vibration com­
The bootstrap percentile method is used to define the
ponents can be removed by performing the random
confidence interval. The main principle is the follow­
decrement averaging. The improved method avoids
ing equation,
the distortion caused by the error involved in the
approximate quasi-stationary response obtained
through curve-fitting technique.
Another main technique involved in the new
method is the Bootstrap-based uncertainty estimation
technique. The reason considering this statistical Where θ is the estimation of θ�B . So the lower and
method is that the traditional wavelet transform upper bound of the 1 _ α confidence interval of θ�B is

202
represented by 1 _ α=2 percentile of the Bootstrap 3 ON-SITE VIBRATION TESTS
ensemble,
The calculated span of the arch bridge is 105m, and
the bridge width is 9m. The temperature during the
tests was 20°C, and the wind speed was between
2 m/s and 6 m/s. The test system included some
large mass piezoelectric acceleration sensors,
If the number of B is adequate large, the Bootstrap a multi-channel signal amplifier, a multi-channel
histogram will become normal shaped. So an accur­ signal acquisition instrument and a computer. The
ate confidence interval can be obtained. on-site vibration tests mainly included two parts:
The wavelet transform is a linear transform. It (1) Vibration tests under environmental excitations.
decomposes signal x(t) by zooming in frequency The field tests collected the acceleration
domain and panning in the time domain using the responses of each measuring point under environ­
mother wavelet g(t). The main principle can be writ­ mental excitation (natural wind, ground pulsa­
ten as follows, tion, etc.). A total of seven vertical measuring
points (V1~V7) and seven horizontal measuring
points (H1~H7) were arranged along the bridge.
The tests used measuring points V3 and H3 as
the reference points. Arrangment of the sensors is
shown in Figure 2.
Where * denotes the conjugate function. Wavelet (2) Free vibration tests when a heavy vehicle crossed
coefficients W(a,t) is the function of scale a and the bridge. The tests collected the acceleration
time t. responses of each measuring point when a 300kN
Complex Morlet wavelet is selected as the truck drived across the bridge at different speeds
mother wavelet for modal parameter identification (30km/h and 40km/h), especially the free vibra­
due to its capabilities in time-frequency localiza­ tion responses after the vehicle left the bridge.
tion. It can decouple the dense modes better by A total of 2 vertical bridge measuring points
importing the frequency band coefficient . Sub­ (FV1, FV2) and 2 horizontal bridge measuring
ject to the paper space limitation, details can be points (FH1, FH2) were set along the bridge.
obtained from the paper presented by Kijewski Arrangment of the sensors is shown in Figure 3.
et al. The sampling frequency of the tests were 51.2Hz.
Figure 1 Shows the entire scheme of the new par­ The vibration tests under environmental excitation
ameter identification method IRDT-Bootstrap-WT. was mainly used to compare the natural frequencies,
damping ratios of the former modes of the bridge

Figure 2. Arrangement of the sensors under the environ­

mental excitation.

Figure 1. Entire scheme of the parameter identification Figure 3. Arrangement of the sensors when the heavy
method IRDT-Bootstrap-WT. truck driving across the bridge.

203
obtained by different parameter identification transform on each segment of data. Firstly, the fre-
methods. The free vibration tests with a heavy quencies of the main vertical and horizontal modes of
vehicle crossing the bridge was mainly used to the bridge are concentrated between 0 and 7 Hz
accurately identify the damping ratios. which are analyzed by a simple peak-picking method.
Then the parameters of the vertical and hori-
zontal responses are determined approximately
4 IDENTIFICATION RESULTS AND between [4.1, 26] and [3.4, 14.8] respectively. fb is the
COMPARISON bandwidth parameter, fc is the central wavelet fre-
quency. According to the minimum Shannon wavelet
The method proposed in this paper is used to identify entropy criterion, when fb ¼ 16; fc ¼ 2:5, the wavelet
the frequency and damping ratio of the arch bridge entropy E of the vertical and horizontal responses are
under environmental excitation. Figure 4 shows the relatively small. Figure 5 and Figure 6 are time-
typical time history of the acceleration of the meas­ frequency plots obtained from the wavelet transform
uring points. The first step of the whole identification of the vertical and horizontal random decrement
scheme is to extract the random decrement signals signals.
from the acceleration responses using the improved The corresponding transform results (phase angle
random decrement technique (IRDT). The acceler­ ’ðtÞ and logarithm of amplitude lnðBðtÞÞ) of the first
ation signals of V3 and H3 are respectively used as two modes are displayed in Figure 7 and Figure 8.
reference channels for the vertical and horizontal It can be seen from the figures that the boundary
measurement points. effect of the low-order signals is relatively obvious,
Using the Bootstrap method, the resulting random and some pieces in the logarithmic graph are signifi­
decrement signals between each measuring point and cantly curved. They are mainly caused by the
the reference point is extended to 1000 segments
(selecting the parameter B equals 1000). So the full
bridge has a total of 7000 vertical and 7000 horizon­
tal random decrement signals.
Subsequently, the complex Morlet wavelet is
selected as the base wavelet to perform wavelet

Figure 5. Time-frequency plot from the wavelet transform

of the vertical random decrement signal δ V3 V1 .

Figure 4. Typical time history of the acceleration: (a)Verti- Figure 6. Time-frequency plot from the wavelet transform
cal, (b) Horizontal. of the horizontal random decrement signal δ H3 H1 .

204
Figure 7. First vertical mode: (a)the semi-logarithm of Figure 8. First horizontal mode: (a)the semi-logarithm of
amplitude lnðBðtÞÞ, (b) the phase angle ’ðtÞ. amplitude lnðBðtÞÞ , (b) the phase angle ’ðtÞ.

signals’ quality that is prone to be influenced by the error of the three methods is 5.19%, which
test noise. However, the results have the characteris­ should be mainly caused by the calculation error
tics of approximately linearity and stability within of the theoretical model. The theoretical model
a certain period of time. Using these periods as the can be modified according to the identification
analysis length can improve the identification accur­ results.
acy, and all modes of signals have the same feature. In order to verify the accuracy of the identification
The histograms of the bootstrap estimation results results of the damping ratio under environmental
for modal parameters using the responses from the excitation, the free vibration tests were performed
measuring point V1 and the reference point V3 are dis­ after the heavy vehicle passed the bridge. Figure 10
played in Figure 9. The histogram results are similar to shows the typical acceleration response during the
the numerical simulation [5]. The identification results whole process of the heavy truck crossing the bridge
of the frequency of each order are concentrated. The with different constant speeds.
results of the damping ratio of each order are relatively Firstly, a 10th-order Butterworth band-pass filter is
discrete, but they are accord with the normal used to obtain the main frequency components from
distribution. the acceleration responses when the heavy vehicle
Table 1 lists the identification results of the crossed the bridge (Figure 10). By comparing the
former vertical and horizontal frequencies of the amplitude and phase angle between each measurement
arch bridge obtained by three methods. The point, the results of the filtered signals show that the
results show that the natural frequencies identified phase angle relationship between the filter components
by the IRDT-Bootstrap-WT method, the WT-Only at each measurement point is completely consistent
method and the traditional RDT-ITD method are with the corresponding mode shape. Then, curve fit­
consistent with each other. The maximum identifi­ ting is performed on the decrement envelopes of the
cation error between different methods is within filtered signals of the main frequency components.
0.3%, which indicates that the three methods Finally, the damping ratio of each mode is calculated
have the same identification accuracy. Compared by the logarithmic decay rate method. Since the
with the theoretical frequency, the maximum damping ratio is related to the vibration intensity, and

205
Figure 9. Histogram of bootstrap estimation results on the natural frequency (top) and damping ratio (bottom) for: (a) first
vertical mode, (b) second vertical mode, (c) third vertical mode, (d) first horizontal mode.

Table 1. Natural frequencies identified with different methods.

Natural frequency (Hz) Error (%)

Modea FEA IRDT-Bootstrap-WTb WT-Only RDT-ITD IRDT-Bootstrap-WT WT-Only RDT-ITD

SL1 1.816 1.843 1.847 1.844 1.49 1.71 1.54

UV1 3.034 3.248 3.250 3.251 3.64 3.70 3.73
UL1 3.997 4.163 4.166 4.165 4.15 4.23 4.20
SV1 5.023 5.389 5.388 5.386 5.19 5.17 5.13
UV2 6.054 6.301 6.295 6.318 4.08 3.98 4.36

a - The mode shapes are identified using the RDT-ITD method. The MAC between the identification and theoretical mode
is greater than 0.95.
SV -Symmetrical vertical bending. UV -Unsymmetrical vertical bending.
SL -Symmetrical horizontal bending. UL -Unsymmetrical horizontal bending.
b -average frequency of the bootstrap distribution.

damping ratio of each order estimated from the loga­

rithmic decrement rate in Table 2 is an average value.
Table 2 lists the identification results of the
former vertical and horizontal damping ratios of
the arch bridge obtained by four methods. It can
be seen that the maximum identification error of
the damping ratio of each order is 8.76%. The
damping ratios identified by the and calculated
by the logarithmic decay rate method are basic­
ally consistent with each other. The maximum
error of the damping ratio of each order identi­
fied by WT-Only and traditional RDT-ITD
methods reach 52.76% and 27.87%, respectively.
For the identification of damping ratios, the
Figure 10. Vertical time history of the acceleration when IRDT-Bootstrap-WT method proposed in this
the heavy truck was passing the bridge. paper is obviously superior to the WT-Only
method and the traditional RDT-ITD method. It
is more fitful to identify the damping ratios
the selection of different start times of free decrement of the structures under the environmental
will cause different exponential regression, so the excitation.

206
Table 2. Damping ratios identified with different methods.

Damping ratio (%) Error (%)

Mode Shapea LDR IRDT-Bootstrap-WTb WT-Only RDT-ITD IRDT-Bootstrap-WT WT-Only RDT-ITD

SL1 1.94 1.77 2.33 1.67 -8.76 20.10 -13.92

UV1 0.96 1.00 1.08 0.83 4.17 12.50 -13.54
UL1 –c 0.77 0.99 0.92 – – –
SV1 0.61 0.56 0.93 0.78 -8.20 52.46 27.87
UV2 0.62 0.57 0.75 0.77 -8.06 20.97 24.19

a - The same with Table 1.

b - average damping ratio of the bootstrap distribution.
c - The vibration intensity of this mode is too small to fulfil the filter and curve fitting

Figure 11. Damping ratio identification using the curve fitting and the logarithmic decrement rate method.

5 CONCLUSIONS Kijeswski T, Kareem A. Wavelet Transform for System

Identification in Civil Engineering [J]. Computer Aided
The results of structural modal parameter identification Civi l and Infrastructure Engineering, 2003, 18:
under environmental excitation show that the proposed 339–355.
IRDT-Bootstrap-WT method has high identification J Lardies, M N Ta, M Berthillier. Modal parameter estima­
accuracy. The improved random decrement method tion based on the wavelet transform of output data[J].
directly analyzes the non-stationary signals that Archive of Applied Mechanics, 2004, 73: 718–733.
Min Zhihua, Sun Limin, Sun Zhi, Dan danhui. Modal
removes the parameter identification error caused by parameters identification method based on wavelet
the approximately stationary assumption. Non- transform and singular value decomposition. Journal of
stationarity is the characteristic of most signals under Tongji University. 2009, 37(4): 460–465.
environmental excitation. Then bootstrap method Liu Jianfeng, Li Yuanbing, Zhang Qiwei. Statistical Identi­
expands the analysis data, uses bootstrap means and fication for Modal Parameter Based on Improved
confidence intervals to reduce the identification uncer­ Random Decrement Technique and Wavelet Transform.
tainty of modal parameters, especially the damping Journal of Tongji University. 2015(10): 1447–1453.
ratio. According to the Table 2, for the identification of Chiang DY, Lin CS. Identification of modal parameters
damping ratio, the proposed method in this paper is from nonstationary ambient vibration data using correl­
ation technique. AIAA J 2008;46(11):2752–2759.
obviously superior to the traditional identification B. Efron, R. Tibushirani, An Introduction to the Bootstrap,
methods. It has higher applicability in the bridge Chapman & Hall, London, 1993.
health monitoring. Luo Guangkun, Zhang lingmi. Study on identification of
modal parameters based on Morlet wavelet
transformation. Journal of vibration and shock. 2007,
REFERENCES 26(7), 135–138.
Y.T. Sheen, C.K. Huang, Constructing a wavelet-based
Staszewski W J. Identification of Damping in Mod Systems envelope function for vibration signal analysis,
Using Time-Scale Decomposition [J]. Journal of Sound Mechanical System and Signal Processing 18 (2004)
and Vibration, 1997, 203(2): 283–305. 119–126.

207
 Bridge Maintenance, Safety, Management, Life-Cycle Sustainability and Innovations –
Yokota & Frangopol (eds)
© 2021 Taylor & Francis Group, London, ISBN 978-0-367-23278-8

Contactless deformation detection for bridge monitoring: First application

of Sentinel-1 radar data in Austria
A. Vorwagner, D. Prammer & P. Leopold
AIT Austrian Institute of Technology GmbH, Vienna, Austria

M. Schlögl, B. Widhalm & M. Avian

Zentralanstalt für Meteorologie und Geodynamik, Vienna, Austria

C. Honeger
ASFINAG Service GmbH, Vienna, Austria

ABSTRACT: Persistent Scatterer Interferometry (PSI) is a powerful remote sensing method for measuring
and monitoring surface displacements on the Earth’s surface. With the launches of Sentinel-1A and 1B,
C-band SAR data with a spatial resolution of 5×20 meters became freely available. This study investigates
the applicability of freely available Sentinel-1 radar data for supporting assessment of infrastructure assets
within the ASFINAG road network. While the absence of consistently reflecting objects remains a limiting
factor in open areas, measuring points are well detected at prominent features and corresponding deform­
ation rates can be derived. At the same time, certain challenges related to selection of reference points,
phase unwrapping, topographic phase correction and atmospheric correction still need to be overcome. The
application with freely available Sentinel-1 data is tested for the first time on a prestressed concrete bridge
in the alpine region.

1 INTRODUCTION investigations are usually carried out as a supplement

to the visual inspection. Non-destructive methods
Maintenance and upgrading of existing infrastructure could help to understand the degradation process of
is becoming increasingly important. Particularly in buildings and materials. For example, investigations
times of shortened resources and space for new con­ on-site are needed to determine the carbonation and
structions, the need for efficient maintenance of the the chloride content in concrete structures. Some of
functionality of buildings exceeds the construction of these tests require samples to be taken from the sup­
new infrastructure facilities. Impacts such as changes porting structure. Therefore, a disadvantage, in addition
in the load situation or minor repair work underline to the relatively high workload, is the local character
the necessity of continuous structural assessment. of the results, as the material can only be tested select­
There is a need to evaluate the load-bearing capacity ively. Non-destructive methods such as potential field
of these existing structures in order to be able to make measurement or ultrasound test require contact
a statement about their structural safety and life time between the measuring device and the concrete surface
prediction and future maintenance work. and examine the building material in the immediate
In Austria the condition of bridges is determined vicinity of the measuring device. A comprehensive
periodically by visual inspections, whereby the applied investigation of the structure is therefore correspond­
regulations are defined in (RVS 2011). An essential ingly time-consuming and therefore the application is
advantage of this inspection is the use of competent, not possible network-wide.
experienced personnel with historical knowledge of To enable continuous, network-wide observation
the object including local conditions. Disadvantages of important transport infrastructure there is a need
are, however, the lower objectivity/repeatability, the for other strategies and measurement/surveying
impossibility of inspecting inaccessible areas and the methods. In addition to local measurements of
lack of quantifiability of the damage. In deep material integrity, methods based on relevant global

DOI: 10.1201/9780429279119-24
https://doi.org/10.1201/9780429279119-24

208
 in condition assessment, in particular for applica­
tions of continuous monitoring, is constantly being
further developed.
The permanent observation of deformation condi­
tion of the structures is potentially a promising indi­
cator, but not easy to perform for permanent
monitoring. Methods have been developed which try
to determine the deformation from measurements
from inclination sensors (Helmi et al. 2015) or form
acceleration sensors (Hester et al. 2017). The most
common permanent system is a flexible tube level. It
measures the deformation directly and uses the prin­
Figure 1. Example of a road network section A2 in Austria.
Illustrated points are points of possible interest. ciple of a hydrostatic settlement measurement by
©ASFINAG. sensors applied to the structures. This involves deter­
mining the vertical settlement of the sensors in rela­
tion to a reference sensor. The method is suitable for
continuous monitoring of geotechnical or deflection
structure parameters could provide a network-wide problems, but needs permanent sensor maintenance
observation. These methods should identify abnor­ (e.g. changes of liquids….) and linked tubes between
mal characteristics over time of the most important the sensors.
structures. The most promising techniques operate In recent years, more and more new methods have
contactless and should provide changes of deform­ been tested and applied for a wide range of applica­
ation in e.g. a sub cm range. This paper presents tions, which do not use individual sensors but record
a prove of concept by application of contactless the measured values linearly or area-wide on the
method measuring deformation changes based on supporting structure. Non-contact methods are par­
radar measurements from space on Austrian motor­ ticularly in demand, as there is no need to install sen­
ways Figure 1 Differential satellite radar interferom­ sors on the supporting structure. Possible methods
etry is a promising technology which is using freely are laser or radar based.
available Sentinel-1 data. The first application for For example, the just finished research project SIBS
bridge monitoring in the Alpine region is presented safety assessment of retaining walls, tests various pos­
here. sibilities of deformation measurements and assessment
on real objects. The measurement is carried out using
tow high-speed laser mounted on a vehicle (1 Laser
2 METOHDS FOR CONTACTLESS with 1 mio pix/sec) and operates at a speed of up to
DETECTION OF DEFORMATION 100 km/h. The tests were performed under real traffic
flow. The achieved precision of the inclination was
Deformation based methods measure structural reac­ 0.03 ° with a 95% probability (for smooth surface).
tion, caused by temperature deformation, mechanical The developed identification concept uses this method
system changes, or changes in loading. Changes in for a general identification of problem structures which
structural safety are thus indirectly deduced by should be permanent monitored in detail (Vöbu 2019).
measuring the variation in mechanical parameters
such as inclination, deformation or deformation line.
2.2 Persistant scatterer interferometry
2.1 Exiting methods The Persistent Scatterer Interferometry (PSI) method
is a powerful analysis method that can determine
The most common test is the measurement of the deformations on the Earth’s surface with potentially
bridge levelling, i.e. the structural reaction to con­ very high accuracy. In detail, PSI represents
stant loads (dead loads) or changes of deflection a differential, interferometric SAR (Synthetic Aper­
caused e.g. by material behavior, settlement, or ture Radar) method in which the properties of wave
changed loading. This is usually done on a case-by­ interference are used to derive measured quantities.
case basis during bridge inspection, although no gen­ With the help of PSI - by comparing the phase values
eral procedure is currently defined for the use of this of corresponding pixels of different acquisition times
data in a reassessment. - distance differences in the range of a fraction of the
Loading tests are routinely carried out during the wavelength λ can be derived. The principal function
acceptance test of new railway bridges. This is demonstrated in Figure 2 and Figure 3. Basically,
approach was extended, e.g. in an EU - project no reflector onsite is needed. In the PSI - method,
(Samaris 2006), where a bridge-weigh-in-Motion those pixels are examined which remain coherent
(BWIM) measuring system was used to record the over several interferograms seen in Figure 2, as the
structural reaction of road bridges under traffic strongest reflection over a series of images. For this
loads. The bridge model and structural analysis were purpose, a larger number of SAR images (usually
adapted with this data. The use of measurement data more than 20) is processed, whereby a time series of
209
 of landslide boundaries and activity (Barra et al.
2016).

2.3 Failure detection

For failure deduction model’s additional investiga­
tions are needed. The system identification is done
by a comparison measurement data with
a computational mechanical structure model (e.g.
finite element Model). By using model update algo­
rithms, the most likely damage can be determined by
comparing measurement data with the calculated
data (Ralbovsky 2014). Also, only data-based
methods (Erdenebat 2018) have been used for
assessment of abnormal changes. Both mentioned
methods can be based on deformation data, but the
applications are rare so far. They have one thing in
Figure 2. Phase simulation of a persistent scatterer pixel common: the accuracy of the measuring method
(blue) based on (Hooper et al. 2007). must be sufficient that disturbances due to environ­
mental influences on the system behavior (tempera­
ture deformations) can be excluded.

3 SENTIENL 1 MISSION

Since 2014, ESA’s Copernicus program has been

operational with the radar satellites Sentinel 1A
(launched in April 2014) and 1B (launched in
April 2016) seen in Figure 3, providing free satellite
data for deformation and other analyses.
The Sentinel-1 mission has a high temporal over­
flight (or repetition frequency) of now 6 days
achieved by the satellite pair Sentinel 1A and 1B.
Figure 3. Sentinel-1 derived displacement measurement is This opens new possibilities in the field of deform­
taken in the Line of Sight (LOS) with wavelength λ= ation monitoring where displacements are measured
5,5 cm, form the orbit (red line) at altitude of 693 km based
in Line of Sight direction (LOS, blue line Figure 3).
on (© ESA 2019).
The Sentinel mission is still under development, in
the next years (from approx. 2022) further satellites
will be brought into orbit (e.g. Sentinel 1C and 1D).
the deformations as well as the mean offset rate can The Sentinel-1 satellites 1A and 1B overfly the
be calculated for the Persistent Scatterer (PS). earth on a synchronous (approximate) polar orbit at an
Explicit research -using the noncommercial altitude of 693 km. The repetition cycle per satellite is
high-resolution satellite TerraSAR-X (TSX) oper­ 12 days, which - as already mentioned - results in
ated by the German Aerospace Center- tested the a repetition cycle of 6 days per location on the Earth’s
methods on serval structures and buildings. For surface due to the arrangement of both satellites. Points
example, a current project is observing hydro- on the Earth’s surface are flown over in two orbital dir­
power dams using the highest resolution mode, ections, one from south to north (ascending orbit) and
the so-called Staring SpotLight with a spatial one from north to south (descending orbit).
resolution of approximately 1.0 m x 0.25 m and To measure deformation on infrastructure Persist­
a footprint coverage of about 5 km x 3 km. They ent Scatterer Interferometry (PSI) will be obtained
could demonstrate that the sub-cm accuracy is using multi-temporal Sentinel-1 images. APSI
possible (Lange 2019). Dam infrastructures have approach is used, which uses two key aspects of the
also been surveyed using Sentinel-1 data, delin­ Sentinel-1 data. The first one is the increased inter­
eating sub-cm/yr displacement rates (Ruiz- ferometric coherence, which is mainly due to the
Armenteros et al. 2018). Applying PSI over short revisiting cycle of these data. This approach
a larger area, Sentinel-1 data has successfully exploits temporally consecutive interferograms to
been applied to detect subsidence areas in fully take advantage of its high coherence.
a metropolitan area (Delgado Blasco et al. 2019, The second aspect is the reduced orbital tube of Sen­
Barra et al. 2016). A further example where Sen­ tinel-1, which decreases the importance of the so-
tinel-1 data has been shown applicable is land­ called residual topographic phase component. In this
slide detection, allowing to update existing maps approach, this phase component is neglected.
210
4 CASE STUDY APPLIED ON A CONCRETE
BRIDGE IN AUSTRIA

To test the feasibility of freely available Sentinel-1

data a case study was performed on a bridge within
the Austrians road network operated by ASFINAG.
The observed bridge object is crossing a valley in
a height of 130 m. It has a length of 632 m and is
one of the largest prestressed concrete bridges in Aus­
tria. The concrete piers (2 double columns and one
single column) are fixed on the structure, the highest Figure 5. Identified PS- points on the observed bridge.
pier is 75 m high. The main span is 250 m. All 4 ©ASFINAG.
lanes (2 directions) are on the same bridge section.

4.1 Sentinel data form the test side In Figure 5 a plan view of the bridge including
Along the point of interest, a larger number of PS the identified PS- points are plotted. The bridge and
could be identified for the period 2016-01-01 to location of piers are highlighted green. For the 2 dif­
2017-12-31 (approx. 60 scenes). The ascending orbit ferent orbits PS were found. They are marked red for
73 and the descending orbit 124 of the satellite Sen­ the ascending (south north) and blue for the descend­
tinel-1A serve as data sources. In Figure 4 the ing (north south) orbit. Some PS points lie nearby
master scene of the observed area and the ellipsoid the structure, they were reflected from some striking
corrected master image in ascending orbital direction points on the ground or from the roof of buildings.
is seen. The observed object is marked red. The LOS in the ascending orbit is inclined by the
angle 32.5° and in descending orbit by the angle
43.8° with respect to the vertical. The pre-evaluation
of the data with respect to the start of the recording
and the representation of the LOS show both uplift
and subsidence along the LOS in a range from -5 to
+7 mm/year, but cannot yet be used directly for
statements on structural deformations.
The measuring points of the two orbits were
recorded at a time interval of 6 days and at different
times of the day. These time differences can have
a considerable influence on the deformation analysis
in combination with deformations from environmen­
tal and temperature deformations.
Further the identified points are distributed
unevenly over the structure. Thus, the individual
orbits absorb either the eastern or the western end of
the bridge well, but there are minor overlaps of both,
so calculating the vertical deflection will be difficult.
For further analysis PS-points of significant points of
the bridge were examined in detail. On the bridge,
points were chosen directly above piers (2) and (3)
and in the middle of the bridge (M) (Figure 6 up). In
addition, reference points next to the structure on
a nearby road section (K) were analyzed, (seen in
Figure 6 down). The reference points are located
nearby the bridge and should have similar atmos­
pheric situations as the points located on the bridge.
But the influence of the temperature deformation on
the ground is subordinate compared to the bridge.

4.2 Data analysis and first results of preliminary

analysis
The time variation curves for the reference points is
Figure 4. The bridge (marked red) in the master scene shown for the ascending orbit in Figure 7 below.
(top) and the ellipsoid corrected master image in ascending A negative sign means a deformation in LOS direc­
orbital direction (below). tion away from the satellite. In addition, the air

211
 environmental influences. For these deformations,
both an annual cycle (depending on the average actual
building temperature) and a daily cycle (depending on
the ambient temperature as well as the actual solar
radiation, weather conditions such as cooling by driv­
ing rain) will be visible. Basically, the accuracy can
be significantly improved by using compensation
models.
A measured value is available per orbit or viewing
direction in an interval of 12 days. As a simplified
approach for the partial compensation of the daily vari­
ations the consideration of the measured values in the
early morning hours is used, because on the one hand
the disturbing influence of the solar radiation is smal­
lest here, and experience shows that the building tem­
perature corresponds better with the daily minimum
Figure 6. Detailing points, M in bridge mid span, 2 and 3 than with the daily maximum. Although no satisfactory
are located over piers (top), and reference points K (below)
compensation can be achieved with this approach,
©ASFINAG.
improvement tendencies should still be possible.
The measurements of the descending orbit in
Figure 8 shows scattering. Again, and again outliers
can be detected whose cause is probably related to
the phase unwrapping and atmospheric failures.
However, a yearly cycle is assumed for pier 2. The
deformation of Pier 3 shows jumps and no real cor­
relation to the temperature.
Under the given conditions, a first approach has
now been taken to reduce the local atmospheric
effects. It will be assumed that the variation in the ref­
erence points is only caused from the atmosphere
effects and can therefore compensate these effects by
simple subtracting. In fact, a slight improvement of
Figure 7. Reference point K- deformation (mm) over time the descending orbit was achieved with this simple
in LOS direction data from the desending orbit. method seen in Figure 9. The results for Pier 2
improves noticeably after 07/2016. In between,

temperature data, evaluated from a vera flex reana­

lyze (resolution of 250x250 m in a timestep of 1h)
are plotted in grey. The base air temperature was
taken at 5:00 am, which should also approximate the
temperature of the structure.
As expected, the time variation curve of the
points hardly deviates within an orbit or the viewing
direction. In addition, the control points do not have
an annual or daily cycle. Only long-term settlements
or trends could assume this deformation over the ref­
erence period. When looking closely at the points
recorded, deviations of up to ±10 mm can be
observed. These can either be traced back to prob­
lems with the recognition of the point or very prob­
ably to existing atmospheric effects or problems
related from incorrectly resolved phase jumps (so
called phase unwrapping).
The situation on the bridge is much more complex.
In addition to structural deformations, environmental
deformations are also present and can falsify the
measurements. This environmental deformation is
determined by the building temperature, which Figure 8. Uncorrected deformation (mm) over time in LOS
depends on the location under consideration (tempera­ direction, points of Pier 2 (top)- Pier 3 (below), descending
ture fields) and the temporal course of the orbit (5:00AM).

212
 In addition, there is an atmospheric error due to not
or not exactly known refractions of the radar beam
when passing through different air packets. Although
this fact is partly considered in the simplified atmos­
pheric correction in the data processing, it has to be
assumed that these methods provide good results espe­
cially in lowlands. In hilly and mountainous regions,
however, this influence is greater than in lowlands and
compensation is much more difficult. A first pragmatic
approach to improving the results is shown in
Figure 9.
Furthermore, there may be inaccuracies in the rec­
ognition of measuring points if, for example,
a similar point is measured in the immediate vicinity.
The actual measuring point has changed its position
in the measuring period or errors in the phase shift
(phase unwrapping) may occur. In order to prevent
this, either striking, well reflecting building compo­
nents or the use of specially attached reflectors
Figure 9. First attempts for correction of deformation (corner reflectors) are of decisive advantage.
(mm) over time in LOS direction, points of Pier 2 (top)­ In the case of engineering structures, the deform­
Pier 3 (below), descending orbit (5:00AM). ations from environmental influences have to be
taken into account. The compensation of different
deformation states caused by disturbing influences of
the structure between the individual satellite images
however, leaps are still visible whose cause is sus­ is essential for the accuracy of the method. Causes
pected to be a phase error. Especially for Pier 3, of the mentioned influences are:
a correlation to the periodic temperature changes is
noticeable. • Weather-related deformations from temperature,
In the annual cycle of air temperatures at 5:00 UTC solar radiation, wind….
(Local Time 6:00 respective 7:00 in summer time), • Mechanical influences (traffic, other loads) and
a difference in temperature of ΔT= 30°C was detected. corresponding deformation
The longitudinal expansion of Pier 3 and thus the cor­ These influences can lead to both horizontal and
responding change in height at these points can be cal­ vertical displacements of recorded measuring points.
culated to a difference Δl= 27 mm using simple linear However, from the measurement data itself with
temperature extension. This estimation illustrates the only one observation direction (LOS) it is not clear
order of the magnitude, the actual temperature deform­ in which direction the observed point has shifted,
ation over the course of the year is not known. In add­ since only the change of the measurement distance
ition, the temperature changes act as a temperature between the object and satellite is known.
field with various temperature over the entire structure
and the pier and bridge will not only move vertically
but spatially in all three axis directions. 5 CONCLUSIONS
A full temperature compensation could only be
done by involving distributed onsite temperature Synthetic Aperture Radar method of the Persistent
sensor data in combination with a detailed model Scatterer Interferometry based on freely available
(e.g. FEM model), taking also into account the tem­ Sentinel-1A satellite data for deformation measure­
perature fields distribution over the structure and non- ment on objects is basically well applicable. The first
stationary temperature effects and the real mechanical application on an alpine bridge proved, that InSAR
behavior such as stiffness and boundaries. This was works with Sentinel 1 data despite the coarser spatial
not performed in this first stage of investigations. resolution and is identifying movements in LOS in
an unexpected good quality.
4.3 Interpretation and discussion However, the results are still subject to a certain
uncertainty, which can be further eliminated with
The first characteristic in determining the deform­ future research work. In order to derive reliable
ation according to the SAR method is that without statements for the structural safety, deformation
additional measures it is not possible to determine accuracies in the mm range are necessary. The Euro­
whether this PSI point lies on the structure or equip­ pean Space Agency (ESA) is currently preparing
ment (noise barrier, railing…). The location of additional Sentinel satellites (1C to 1D), which will
points is strongly dependent on the reflection proper­ potentially further shorten overflight times and
ties of the radar waves of the structure. improve data and reaction times.

213
 Netwide deformation monitoring should help to ESA 2019 https://sentinel.esa.int/called on 1. Of
monitor the condition of bridges and retaining walls. April 2019.
The first main challenge in extracting useful informa­ Helmi, K., Taylor, T., Zarafshan, A. & Ansari, F. 2015:
tion from this data is the compensation of environmen­ Reference free method for real time monitoring of
tal impacts. Structural temperature is a major cause of bridge deflections, Engineering Structures, 103,
uncertainties in deformation-based condition identifica­ 116–124. DOI: 10.1016/j.engstruct.2015.09.002
tion, therefore its effect must be compensated. While Hester, D., Brownjohn, J., Bocianc, M. & Xu, Y. (2017):
Low cost bridge load test: Calculating bridge displace­
conventional monitoring uses installed temperature ment from acceleration for load assessment calculations,
sensors, netwide monitoring requires a new reliable Engineering Structures 143, 358–374. DOI: 10.1016/j.
method to determine structural temperatures. engstruct.2017.04.021
The first analysis of the tested bridge have also Hooper, A., P. Segall, and H. Zebker 2007: Persistent scat­
shown that compensation measures helps to improve terer interferometric synthetic aperture radar for crustal
the quality. To handle the problem with point recog­ deformation analysis, with application to Volcán
nitions corner reflectors mounted on the structure Alcedo, Galápagos. Journal of Geophysical Research,
could help to improve the data and will be investi­ 112, B07407.
Lange, O., Walter D, Anderssohn J. 2019 Satellite Based
gated on the ongoing research.
Longterm Deformation Monitoring on Dams and its
Surroundings. 5 th Conference on Smart Monitoring,
ACKNOWLEDGEMENTS Assment and Rehabilitation of Civil Structures, Pots­
dam. 27 – 29 August 2019 Potsdam, Germany.
Ralbovsky, M., Santos, J., Kwapisz, M., Dallinger, S., &
This work was funded by ASFINAG Service GmbH, Catarino, J.M. 2014: Damage detection based on struc­
Vienna, Austria. The authors would like to thank for tural response to temperature changes and model
the support and very good cooperation. updating. Proceedings of European Workshop on Struc­
tural Health Monitoring, 427–434, Nantes, France.
Ruiz-Armenteros, A., Lazecky, M., Hlaváčová, I.,
REFERENCES Bakoň, M., Delgado, J.M., Sousa, J.J., Lamas-
Fernández, F., Marchamalo, M., Caro-Cuenca, M.,
Barra, A. Monserrat, O. Mazzonti, P., Esposito, C., Papco, J., Perissin, D. 2018: Deformation monitoring of
Crosetto, M., Scarascia Mugnozza, G. 2016: First dam infrastructures via spaceborne MT-InSAR. The
insights on the potential of Sentinel-1 for landslides case of La Viñuela (Málaga, southern Spain). Procedia
detection. Geomatics, Natural Hazards and Risk. Computer Science, 138, 346–353.
DOI:10.1080/19485805.2016.1171258 RVS 13. 03.11, 2011:Qualitätssicherung bauliche Erhaltung-
Delgado Blasco, J.M., Foumelis, M., Stewart, C., Überwachung, Kontrolle und Prüfung von Kunstbauten
Hooper, A. 2019: Measuring Urban Subsidence in the Straßenbrücken, Österreichische Forschungsgesellschaft
Rome Metropolitan Area (Italy) with Sentinel-1 Straße, Schiene, Verkehr, 2011, Vienna.
SNAP-StaMPS Persistent Scatterer Interferometry. SAMRIS, Sustainable and Advanced Materials for Road
Remote Sensing, 11(2), 129, DOI:10.3390/rs11020129 InfraStructure, project ID: GRD2/30228/2000, 2006:
Erdenebat, D., Waldmann, D., Scherbaum, F. & Teferle N. Deliverable D30: “Guidance for the optimal assessment
2018: The Deformation Area Difference (DAD) method of highway structures. EU- FP5-Projekt.
for condition assessment of reinforced structures, Engin­ VÖBU and SIBS Project partners, 2019. Final Report on
eering Structures, 155, 315–329. DOI: 10.1016/j. SIBS – Saftey asssessment of exiting retaioningwalls;
engstruct.2017.11.034 FFG Project #861163; VÖBU; Vienna.

214
 Bridge Maintenance, Safety, Management, Life-Cycle Sustainability and Innovations –
Yokota & Frangopol (eds)
© 2021 Taylor & Francis Group, London, ISBN 978-0-367-23278-8

Estimation of cable tension force based on digital image correlation

Banfu Yan, Wenbing Chen, Derui Li & Jiangyong Yu
College of Civil Engineering, Hunan University, Yuelushan, Changsha, Hunan, P.R. China; Key Laboratory for
Wind and bridge Engineering of Hunan Province, Changsha, Hunan, China

ABSTRACT: This paper proposed an approach transforms the cable force estimation problem from the
common procedure of constructing and solving the equation of motion of the cable to a simpler problem of
finding the modal nodes of its mode shapes. Digital image correlation is employed to capture the dynamic dis­
placement of the point of interested on the surface of the measured cable under ambient vibration by matching
the corresponding image sub-regions before and after the deformation. The frequency domain decomposition
(FDD) approach is then utilized for modal identification from the measured displacements of the cable. Accord­
ing to the obtained modal identification results, the cable segment between any two modal nodes is selected to
establish the equivalent tensioned beam model with two pinned ends, from which the cable force can be esti­
mated. The laboratory test proved that the integrated approach has good applicability in practical engineering.

1 INTRODUCTION external dampers installed at the cable ends near to

the deck and pylon divides the cable into two or
The vibration-based method is by far the most reli­ three segments. Many researchers investigated the
able and effective tool to estimate the tension forces vibration characteristics of the cable-damper systems
within stay cables or hanger cables for cable stayed from the viewpoint of damper design rather than
bridge or through-type arch bridge due to its simpli­ cable force estimation. Ceballos & Prato (2008) sim­
city, convenience and economy. plified the intermediate supports near the anchors as
The vibration method based on the string vibration deviators having no own rotational stiffness and
theory is used to estimate cable forces from the meas­ restraining the transversal displacement only. Geier
ured natural frequency. For an inclined cable with et al. (2005) proposed a method which considers the
fixed boundary conditions, several parameters such as boundary conditions by calculating the free vibration
the flexural rigidity (EI), sag extensibility ( ) (Irvine length. The length is determined by measuring the
1981) and end conditions should be considered for distance between the nodes of high vibration modes
force determination. The cable force is determined by close to the anchor construction and multiplying this
solving a transcendental equation. Zui et al. (1996) section length by the mode order. Yan et al. (2014,
introduced the dimensionless parameters 2019) proposed a promising solution to tension force
, , and (where T = identification of a cable with arbitrary boundary con­
cable force, and are the theoretical ditions by transforming the common vibration-model
values of nth natural frequencies for taut-string and based method to construct an equivalent segmental
beam structure, respectively), and proposed practical model of the cable with pinned ends according to the
approximate formulas, independent of the length and measured mode shape. Cheng et al. (2015, 2016)
internal force of the cable, to estimate the cable force proposed a similar concept of combining the mode
by an iterative method. Mehrabi & Tabatabai (1998) shape ratios with the modal frequencies was recently
considered these dimensionless parameters and introduced to develop a convenient and accurate
employed a finite difference approach to provide method for the determination of stay cable forces.
a more accurate identification of the cable force from Digital image correlation (DIC) has become one of
the cable frequency. Fang & Wang (2012), Ren et al the most popular target-free displacement measure­
(2005), and Ham & Nghia (2011) presented empirical ment techniques in the field of non-contact optical
explicit expression for the tension force of the cable measurement for experimental solid mechanics
with fixed end conditions. and civil engineering due to its advantages of accuracy
For the existing in-service bridge, the introduction in sub-pixel level, long-distance multi-point non-
of intermediate supports such as the internal or contact test, simple equipment, and easy operation

DOI: 10.1201/9780429279119-25
https://doi.org/10.1201/9780429279119-25

215
(Pan et.al. 2007, Sun & He 2009). At present, the inte­ and . As shown in Figure 1., this transform­
ger-pixel registration algorithms of fast normalized ation indicates that the original model of a tensioned
cross correlation (FNCC) and Fourier transform cross cable with complex end conditions always can be
correlation (FTCC), and the more efficient sub-pixel represented in the form of an equivalent model of
registration algorithm of inverse-compositional Gauss– a tensioned beam with arbitrary boundary conditions.
Newton (IC-GN) have been considered as standard Note that it is difficult to directly identify the ten­
solutions for accurate identification of displacements sion force from the equivalent model of the cable
(Pan 2018). with unknown frequency-dependent dynamic rota­
In the paper, we extend Yan et al.’s method (2014, tion restraints. As shown in Figure 1, we extract the
2019) to the tension force estimation of a laboratory effective length between any two zero-amplitude
cable model from the measured mode shape using DIC points of i and j of the nth mode shape. This corres­
techniques. DIC is employed to capture digital images ponds to the equivalent model as the calculation
of the surface of the measured cable structure under length of the equivalent segmental model of
ambient vibration, and then the dynamic dis-placement apinned-pinned cable with the properties T, m, and
of the region of interested (ROI) on the surface feature EI, and the natural frequency . Obviously,
of the measured cable is obtained. The frequency the equivalent segmental model of the cable in the
domain decomposition (FDD) approach is then utilized form of tensioned beam model is independent of the
for modal shape identification from the measured dis­ complex boundary conditions of the original model.
placements of the cable. The cable segment between The tension force of a cable with pinned end condi­
two stagnation points is selected to establish the tions is presented as
equivalent tensioned beam model of the original cable
structure, from which the cable force can be estimated.
The laboratory test is conducted to validate the applic­
ability of the integrated approach in practical
engineering.
where and n0 is generally determined by the
roundness of .
2 METHODOLOGY OVERVIEW Thus, the tension force identification problem of
a cable with complex boundary conditions is trans­
formed to a simplified problem of identifying the
2.1 Determination of modal nodes of mode shape
mode shape of the original model and finding its
The original model of the cable with multiple inter­ zero-amplitude points as the cable ends of the
mediate elastic supports （k ¼ 1; 2; …; p） and equivalent segmental model.
rotation restraints （k ¼ 1; 2; …; p） is shown in
Figure 1. The cable under tension T has the length
2.2 Displacement extraction: DIC Algorithm
L1;p , the mass per unit length m, and the flexural
Intergrating FTCC And IC-GN
rigidity EI. For the nth mode, the natural frequency
of the cable is ωn . Obviously, it is very difficult to A vision sensor system composed of high-speed video
construct a model-based solution to the identification camera and computer processing system is employed
of the cable tension due to the complexity of the to capture the image sequences including the selected
boundary conditions. Generally, the original model of region of interest (ROI) with existing surface feature.
the cable with complexity boundary conditions can The DIC technique integrating integer-pixel registra­
be transformed to an equivalent model with elastic tion algorithm and subpixel registration algorithm is
support stiffness of and , and then used to extract the displacement time history of
frequency-dependent dynamic rotation restraints of the target points in ROI. In the study, the integer-pixel
image registration algorithm FTCC is employed to
provide an initial guess of the deformed images close
to the true value. Then, a novel subpixel registration
algorithm IC-GN with higher registration accuracy
and efficiency is used to track the target points from
the image sequence. The DIC technique based on
these two algorithms can reach acceptable sub-pixel
resolution accuracy in displacement time history.

2.2.1 FTCC
Figure 2 displays the flowchart of integer-pixel
registration algorithm FTCC. The algorithm utilizes
the overlapping region between the reference subset
tðx0 ; y0 Þ and the current deformed subset gðx; yÞ, and
Figure 1. Original model, equivalent model and equivalent performs the inverse Fourier transform on the cross-
segmental model of the tensioned cable. power spectrum of the region in the frequency

216
 The IC-GN employs the initial guess determined
by FTCC to start the iteration until ZNSSD con­
verges to the minimum value of RZNSSD . Then we
obtain the sub-pixel level displacement results
between the deformed subset and the reference
subset. More detailed introductions about IC-GN is
shown in Reference (Pan 2018). Compared to the
conventional FA-NR algorithm, the IC-GN algo­
rithm avoiding the calculation of Hessian matrix
exhibits high performances in accuracy, calculation
efficiency, and noise robustness.

3 LABORATORY VERIFICATION TEST

3.1 Experimental setup

To verify the effectiveness and accuracy of the pro­
posed approach integrating DIC and mode shape-
aided techniques in estimation of cable tension force,
we conduct an laboratory test on tension force esti­
mation of an inclined cable model using DIC-based
vision system. As shown in Figure 3, the analyzed
Figure 2. Flowchart of FTCC. cable with a length of 2.8 m and an inclination angle
of 31� is equally distributed with a mass of 0.05kg
every 5cm in longitudinal direction. One end of the
domain to obtain its impulse response function. The cable model is hinged to the steel beam which repre­
peak position of the function is detected as the two- sents the slab deck, and the other end is connected to
dimensional displacement of the image. the mass (weight: m) through a rubber sheet with
The 2D inverse Fourier transform is conducted on a size of 5cm � 5cm and a pulley located at the steel
the cross-power spectrum to attain the impulse column representing the pylon. As such, the tension
response function whose peak value is exactly the force T ¼ m · g (g ¼ 9:8m=s2 is the acceleration of
relative shifting of the corresponding subset in x and gravity) is applied to the cable, and the boundary
y directions, respectively. By tracking the coordinate condition at the pylon end appears to be complex
of the impulse response function, we can obtain the due to the introduction of the rubber sheet. The nom­
integer-pixel displacement ðu0 ; v0 Þwhich is the so- inal diameter and mass per unit length of the cable
called initial guess for the further sub-pixel registra­ model are 4mm and 0.098kg/m, respectively. The
tion algorithm. The original data in Figure 2 is taken nominal cross section and moment of inertia of the
from the same position of the first and tenth pictures cable are 12.57 mm2 and mm2 , respectively.
in the open source data “sample12” of the Society As shown in Figure 3, 52 metal blocks are uni­
for Experimental Mechanics (SEM), and the subset formly distributed along the cable. The metal blocks
size is 64X64 pixels. The DIC initial value calculated as the existing surface features on the cable are used
by FTCC is ð_1; _9Þ, which is completely identical as the targets. The non-contact DIC system is then
to that using the traditional FNCC algorithm.

2.2.2 IC-GN
The IC-GN subpixel registration algorithm proposed
by Pan et al.[13] is popular and strongly recommended
as a standard algorithm in the DIC-based displacement
field measurement. The algorithm employs the initial
guess of deformation via FTCC to initialize the itera­
tive calculation. To attain an effective and robust
subset matching, an objective function related to cor­
relation criterion and shape function is then defined
and optimized to obtain the desired deformation
vector. In the study, the bicubic spline interpolation
function and the common adopted correlation criterion
termed as zero-mean normalized sum of squared dif­
ference (ZNSSD) are recommended in the IC-GN
algorithm. Figure 3. Experimental test setup for an inclined cable with
distributed mass.

217
employed to measure the dynamic displacement of
the target point by tracking the ROI covering the
metal block. A consumer-grade digital single lens
mirrorless camera is employed to provide high quality
video recording. The camera has a 20.3MP digital
live CMOS sensor and the ability to record 4K video
at up to 60 fps, 10-bit 4:2:2 internal recording with
select settings, and full HD video (1080P) at up to
180 fps. To ensure the resolution of the target points
and to capture sufficient cable vibration components,
the camera is applied to record the image sequences
of the cable with a pixel resolution of 1920 � 1080 at
the frame rate of 180fps.

3.2 Identification result

An impact hammer is used to excite the cable every 25 Figure 5. FDD-based mode shape identification of the
to 30 seconds, and the digital camera is employed to cable model.
record the image sequence of the cable with
a resolution of pixels at the frame rate of 180fps. The
dynamic displacement time histories of the targets can
be obtained by DIC-based target tracking techniques A comparison between the calculated and identi­
integrating FTCC and IC-GN algorithms. The further fied mode shapes of the cable model using FDD is
FDD-based output-only modal identification approach presented in Figure 5. Note that the global modal
is then employed to identify the mode shape of the shape of the cable is derived from the combination
cable from the measured dynamic displacement time of the segmental mode of the four cable segments
histories. via the reference points. The identified second, third,
Figures 4(a) and (b) show the typical time histor­ fourth, and fifth-order mode shapes fitted by the
ies and power spectrums for the displacements of sinusoidal functions are respectively very close to
target point 36, respectively, when the tension force the calculated theoretical modes. Furthermore, it is
320N is applied to the cable. It is seen that the first easy to find the locations of the zero-amplitude
five-order natural frequencies of the cables can be points such as A and B for the third mode, A, B, and
clearly identified. Figure 4(c) plots the singular C for the fourth mode, and A, B, C, and D for the
values of the power spectral density function matrix fifth mode from the measured mode shape. Accord­
of the displacement responses in the frequency ing to the mode shape-aided cable tension force esti­
domain. It is observed that at least five-order reson­ mation techniques proposed in section 2.3, the
ance frequencies can be clearly identified from the segmental cable length between any two zero-
peaks of the spectral lines. amplitude points of the mode shape can be con­
sidered as the effective length of the equivalent seg­
mental model of a pinned-pinned cable. The
difficulty in estimating the tension force of a cable
with complex end conditions is thus simplified.
We conducted one loading cases of applying
320N to the cable. Table 1. presents the comparative
results of the estimated cable forces. The effective
lengths of the segmental cable for the third, fourth,
and fifth mode, the natural frequencies for the related
mode order, and the relative error between the theor­
etical and measured cable forces in different cases
are considered to clarify the effectiveness of the pro­
posed identification method. It is observed that: (1)
In these three cases, there is a little difference of
about 0.01 m or 0.02 m among the effective lengths
with respect to the identical mode order and segmen­
tal cable; (2) The relative errors between measured
and theoretical value are in the range of ±5%, which
is accepted for the common tension force estimation
Figure 4. (a) Typical time history and (b) power spec-trum of the cable. For the fifth mode, the relative error
of one target point 32, and (c) singular values of the power attains a maximum value of 4.03% when the seg­
spectrum density matrix of the dis-placement responses of mental cable between the zero-amplitude points of
target point numbers from 24 to 40.
B and C is considered; (3) The relative errors for the

218
Table 1. Comparisons of the theoretical and measured cable forces in different cases.

Effective Natural Measured cable Theoretical cable

length frequency force force
Mode Segmental Relative error
order cable m Hz N N (%)

3rd A-B 0.92 9.668 324.86 320 1.52

A-B 0.69 321.91 320 0.60
4th B-C 0.69 12.832 321.91 320 0.60
A-C 1.38 321.91 320 0.60
A-B 0.56 321.31 320 0.41
B-C 0.57 332.88 320 4.03
C-D 0.56 321.31 320 0.41
5th A-C 1.13 15.796 327.07 320 2.21
B-D 1.13 327.07 320 2.21
A-D 1.69 325.14 320 1.61

higher mode (for instance: the fifth mode) appears to convenience of the mode shape identification. (2)
be larger than those for the lower mode. One of the For the stiff stayed cable or hanger cable, it is diffi­
reasons is that for the higher modes the uncertainty- cult to extract the third-order mode using the DIC-
induced target tracking error in DIC techniques based vision system due to its weak vibration in the
results in the degradation of identification accuracies cases of ambient excitation and even artificial shake.
of the mode shape. Hence, it is necessary to extract the effective length
In summary, the technique integrating vison­ of the cable from the first or second-order modes via
based target tracking with FDD-based modal identi­ DIC system.
fication is capable of obtaining the zero-amplitude
points of the modal shape of the cable, from which
the equivalent segmental model with pinned-pinned ACKNOWLEDGEMENTS
ends is employed to estimate the cable force.
Although the measured cable force matches closely The authors gratefully acknowledge supports from
with the theoretical value, the yielding relative error the National Natural Science Fund of China (Grant
may be caused by the reasons: (1) the inherent error Number: 51578227) and the Major Program of Sci­
of the proposed mode shape-aided method; (2) the ence and Technology of Hunan Province (Grant
distortion of the target points along the cable during Number: 2017SK1010).
manual impact hammer test may introduce identifi­
cation errors on the mode shape and the position of
the zero-amplitude point, and (3) uncertainty- REFERENCES
induced vison-based target tracking error.
Ceballons MA, Prato CA. 2008. Determination of the axial
force on stay cables accounting for their bending stiff­
ness and rotational end restraints by free vibration tests.
4 CONCLUSIONS Journal of sound and vibration 317(1): 127–141.
Chen C. C., Wu W. H., Tseng H.Z., Chen C.H., Lai G.
The study proposed a promising solution based on 2015. Application of digital photogrammetry techniques
mode shape-aided estimation of effective length via in identifying the mode shape ratios of stay cables with
non-contact DIC and FDD techniques to measure the multiple camcorders. Measurement 75: 134–146.
tension force of the hanger cable with unknown Chen C. C., Wu W. H., Leu M. R., Lai G. 2016. Tension
determination of stay cable or external tendon with com­
boundary conditions. The identification accuracy of
plicated constraints using multiple vibration
the integrated methodology is initially confirmed measurements. Measurement 86: 182–195.
through a laboratory verification test on tension Fang Z., Wang J.Q. 2012. Practical formula for cable ten­
force estimation of an inclined cable model. sion estimation by vibration method. Journal of bridge
In practical application of the proposed engineering 17(1): 161–164.
approaches, several technical difficulties are still Geier R, DeRoeck G, Petz J. 2005. Cable force determin­
required to be addressed: (1) For the common stayed ation for the Danube Channel Bridge in Vienna. Struc­
cable and hanger cable, the identification of the pos­ tural engineering international 15(3): 181–185.
ition of the surface feature along the cable is not an Irvine H.M, 1981. Cable structures, The MIT Press, Cam­
bridge, Massachusetts, and London, England.
easy work. One of the solutions is to mark some
Mehrabi AB, Tabatabai H. 1998. Unified finite difference
target panels along the cable with known spacing formulation for free vibration of cables. Journal of
length during the construction stage for the structural engineering 124(11): 1313–1322.

219
Nam H, Nghia NT. 2011. Estimation of cable tension using Sun W., He X.Y. 2009. Experimental studies on application
measured natural frequencies. Procedia engineering 14 of digital image correlation in measuring field of civil
(2011): 1510–1517. engineering. Journal of Nanjing university of aeronaut­
Pan B. 2018. Digital image correlation for surface deform­ ics & astronautics 41(2): 271–275.
ation measurement: historical developments, recent Yan B.F., Yu J.Y, Soliman M, 2014. Estimation of cable
advances and future goals. Measurement science and tension force independent of complex boundary
technology 29: 082001. conditions. Journal of engineering mechanics 141(1):
Pan B, Xie H.M., Li Y,j. 2007. Three-dimensional digital 06014015.
image correlation method for shape and deformation Yan B.F., Chen W, Yu J, Jiang X, 2019. Mode shape-aided
measurement of an object surface. Journal of experi­ tension force estimation of cable with arbitrary bound­
mental mechanics 22(6): 556–567. ary conditions. Journal of sound and vibration 440
Ren W.X, Chen G. Hu W.H. 2005. Empirical formulas (2019): 315–331.
to determine cable tension using fundamental Zui H, Shinke T, Namita Y. 1996. Practical Formulas for
frequency. Structural engineering and mechanics 20 Estimation of Cable Tension by Vibration Method. Jour­
(3): 363–380. nal of structural engineering 122(6): 651–656.

220
 Bridge Maintenance, Safety, Management, Life-Cycle Sustainability and Innovations –
Yokota & Frangopol (eds)
© 2021 Taylor & Francis Group, London, ISBN 978-0-367-23278-8

Probabilistic analysis and safety formats approaches applied for Czech

bridge structures under the ATCZ190 SAFEBRIDGE project
M. Šomodíková, L. Novák, M. Lipowczan, M. Vyhlídal, J. Doležel, D. Lehký, D. Novák &
R. Pukl
Faculty of Civil Engineering, Brno University of Technology, Brno, Czech Republic

1 INTRODUCTION capacity of selected bridge/structural component.

The process of selection of suitable bridge structures
Project ‘ATCZ190 SAFEBRIDGE – Advanced ana­ for case studies is also briefly presented.
lysis of existing reinforced and pre-stressed concrete
bridges: Nonlinearity, reliability, safety formats, life­
time aspects’, awarded by European Regional Devel­ 2 PROBABILISTIC ANALYSIS AND SAFETY
opment Fund within the European Union program FORMATS APPROACHES
Interreg Austria–Czech Republic, aims at the design of
advanced procedure of numerical assessment of bridge There are several methods for determination of
structures based on reliability theory (on the basis of a design value of response in normative documents or
EN 1990) and make this available for more engineer­ scientific papers, which are briefly described in fol­
ing offices as well as road/railway infrastructure oper­ lowing sections. So called safety formats are more or
ators in particular regions of Austria and the Czech less commonly used in practice due to the possibility
Republic. of their usage without complex reliability knowledge.
Road system in Vienna, Lower Austria and Mor­
avia regions faces great challenges: aging structures,
2.1 Partial safety factors method
increasing traffic load, effect of climate changes, qual­
ity requirements, etc. In order to maintain a reliable According to EN 1990 (CEN 2002), nonlinear FEM
road network, new and innovative approaches must analysis is performed with design values of input
be pursued, especially in case of bridge structures. random variables, Xd , resulting in design value of
The detailed stochastic nonlinear modeling of struc­ a resistance, R~ d . It means, partial safety factors
tural response with respect to its deterioration can pro­ method is a deterministic method for the assessment
vide relevant information and complex assessment of of design value of structural resistance and only one
bridge structures; hence it has a significant impact on FEM simulation is needed in this case. Design values
effective utilization of a bridge management budget. of material characteristics can be obtained from
Nonlinear modeling based on finite element method laboratory experiments or directly from tables
(FEM) is used more frequently for design and analysis included in EN.
of structures nowadays. In the last decade, a reliability For lognormally distributed random variables,
analysis is also more frequently used in practice. The input design values of material characteristics are
combination of both procedures is a strong and effect­ determined as:
ive tool for realistic modeling of new as well as exist­
ing structures. However, utilization of fully
probabilistic approach in combination with nonlinear
FEM analysis is highly time consuming. Hence struc­
tural designers and engineers are more often interested where μR and vR are the mean value and coeffi­
in semi-probabilistic methods as a suitable tool for cient of variation (CoV) of resistance, αR is
determination of a design value of structural response. a sensitivity factor and βn is the target reliability
The aim of this paper is to introduce strategies index from normative documents. Statistical param­
and possibilities of safety formats approaches and eters obtained by laboratory experiments are needed
advanced reliability methods for nonlinear FEM ana­ in this case. Additional reduction of a resistance
lysis of existing structures. Several approaches are due to model uncertainty is included in EN and
applied to determine the design value of ultimate

DOI: 10.1201/9780429279119-26
https://doi.org/10.1201/9780429279119-26

221
therefore, a safety factor ¼ 1:06 is applied. The where γs ¼ 1:15 and γc ¼ 1:5.
final design value of structural resistance is then:

2.3 Global safety factor methods – ECoV

Lognormal distribution for response variable
R is assumed in proposal by Červenka (2013),
thus coefficient of variation vf can be esti­
In case of the normative coefficients of individual mated as:
material input variables, γM , which contain also the
model uncertainty, no additional reduction by
a safety factor is needed. In this case, the final
design value of structural resistance is as follows:

Note that just two simulations of nonlinear FEM

analysis are needed in this approach: the mean
response Rm ¼ Rðf cm ; f ym ; anom ; …Þ with mean
The design values of input variables using partial values of input random variables, and Rk using char­
safety factors method are extremely low, especially in acteristic values (5% percentile). The global resist­
case of normative values. It may lead to unrealistic ance safety factor is then calculated as:
redistribution of internal forces and different failure
mode of a structure. Therefore, nonlinear FEM ana­
lysis using mean values of input variables (resulting
in mean response, Rm ) in combination with global
safety factor, γR , is recommended. The design value Described concept was adopted in the fib Model
of resistance is then calculated as: Code 2010 (fib Federation internationale du beton
2013) and design value Rd was later decreased by
¼ 1:06 to take model uncertainties into account:

2.2 Global safety factor methods – EN 1992-2

There are two main alternative methods how to Improved ECoV method, where vR is extended by
define Rd according to global safety factor approach: variability of model and geometrical uncertainties, is
deterministic (only one simulation is needed) EN proposed by Schlune et al. (2012):
1992-2 method and semi-probabilistic estimation of
a coefficient of variation (ECoV) method.
In the global safety factor concept according to
EN 1992-2 (CEN 2005), the design value of resist­
ance is estimated as:
where vg and vm represents coefficient of variation
of geometrical and model uncertainties. The coeffi­
cient of variation of material vf , if material param­
eters are not correlated, can be calculated as:

with the global safety factor of resistance

γR ¼ 1:27, including model uncertainty. The mean
value of yield strength of steel reinforcement, fym , and
reduced mean value of concrete compressive strength,
~ , is defined based on their characteristic values as
follows: where the response of construction RΔXi is deter­
mined by nonlinear FEM analysis using reduced
mean values of material variables by ΔXi and σXi is
standard deviation of i-th variable. If the lognormal
distribution of material variables is assumed, the
reduced values of a variable Xi can be calculated as:

222
 and EXprojekt s.r.o. The strategic partners provided
the necessary information to support the selection of
the most interesting cases. The characteristic structural
where Xmi is the mean value of i-th material charac­ systems of bridges, their numbers and location within
teristic and step size parameter . Note the Czech road and railway network were discussed.
that this approach requires N þ 1 simulations of non­ The most representative bridges for Safebridge
linear FEM analysis, whereN is the number of material project case studies were selected based on statistical
random variables. The extension of the method for cor­ data of ŘSD ČR (2018) and Rotter et al. (2018) about
related material variables can be found in (Schlune existing bridges in the Czech Republic; see also para­
et al. 2012). This technique is generally Taylor series graphs below. As the strategic partners involved in
expansion with one-sided differencing, thus it is pos­ the project are the main national operators for road
sible to assumed higher order of Taylor series in order and railway bridges, they have direct access on data
to obtain more accurate results. about the full existing bridge network. The bridge
selection process was made conceptually so that
common types of bridges were analyzed, ensuring the
2.4 Monte Carlo type simulation methods objectivity and supporting the development of
Besides safety formats mentioned above, various a general and standardized procedure for advanced
advanced methods can be used to assess the design numerical analysis and reliability assessment of exist­
value of response Rd . Eq. 1 can be used for this pur­ ing bridges. A structural type, material, size and age
pose. In this case, it is necessary to estimate statis­ of bridges were established as the most important
tical moments of response function R. bridge characteristics. Structural condition (SC) and
The most accurate, on the other hand most time the frequency of individual bridge types within the
consuming approach, is an approach based on Monte program region were also taken into account.
Carlo (MC) type simulation. The main idea of MC In the Czech Republic, almost all railway bridges
techniques is pseudo-random sampling and statistical (with the exception of railway bridges on siding
analysis of performed deterministic simulations. For rails) are operated by the Railway Infrastructure
efficient estimation of statistical moments, stratified Administration, state organization (SŽDC). On 31.
sampling technique called Latin Hypercube Sampling 12. 2017, SŽDC managed 6,751 bridges. Direct
(LHS) was developed (McKay et al. 1979). In com­ management is provided by 7 regional offices –
parison with crude MC, the number of needed simula­ Bridge and Tunnel Management (BTM) – and
tions is drastically reduced (from thousands–millions Regional Directorates (RD). Data and inspection
in case of MC to tens–hundreds in case of LHS). results are stored in the Bridge Management System.
LHS is not dependent on size of stochastic model, Regarding road bridges, bridges on motorways and
thus it is recommended for extensive stochastic 1st class roads are managed by the Road and Motor­
models; see e.g. FReET software (Novák et al. 2014). way Directorate of the Czech Republic (ŘSD ČR),
Further, MC type simulation techniques are able bridges on 2nd and 3rd class roads are owned by
to take correlation of input variables into account. regions. Bridges on local roads are owned by towns
Several methods was developed for this purpose, and municipalities, or they have private owners. On
e.g. generalized Nataf transformation (Lebrun & 1. 7. 2018, the Road and Motorway Directorate, the
Dutfoy 2009) or optimization techniques such as Road Data Bank Department and the National Trans­
simulated annealing (Vořechovský & Novák 2009). port Information Centre (NDIC) registered a total of
17,545 bridges on motorways and 1st , 2nd and 3rd
class roads. The number of bridges on local roads in
3 BRIDGE SELECTION PROCESS, towns and municipalities is roughly estimated as
STATISTICAL DATA 18–20 thousands.
In case of road bridges, a large number of bridges
In the framework of the project, five existing bridges with classifications of 5–7 (5 – wrong; 6 – bad; 7 –
within the program region of the Czech Republic – i.e. emergency state of a bridge) can be found on Czech
regions of South Moravia, South Bohemia and Vyso­ motorways and roads in South Moravian region; see
čina – were carefully selected to be studied and mod­ Figure 1. The load-bearing capacity of such bridges is
eled with nonlinear finite element method on significantly reduced in most cases. Hence, advanced
deterministic and stochastic levels. The selection of analysis is increasingly supported. Regarding age,
bridges was performed based on general information most bridges on motorways and 1st class roads were
on the road and railway network in the Czech Republic built between 1971 and 1980; bridges on 2nd and 3rd
and discussions with strategic partners involved in the class roads were built mostly in 1950s and 1960s (see
project. These are: Railway Infrastructure Administra­ Figure 2). Reinforced concrete and prestressed precast
tion, state organization (SŽDC), Road and Motorways concrete are the most common materials of bridges on
Directorate of the Czech Republic (ŘSD ČR), Motor­ local roads in cities of Prague, Brno and Plzeň as dis­
ways Administration and Maintenance of the South played in Figure 3. Structural condition SC 2 (means
Moravian Region (SÚS JMK), Brněnské komunikace that an object requires repairs beyond normal mainten­
a.s., and engineering offices Dopravoprojekt Brno a.s. ance), age between 80 and 89 years and span length up
223
 Figure 4. Number of railway bridges with respect to the
Figure 1. Total number of bridges on motorways and structural condition (SC) according to RD-BTM.
roads considering structural condition and region.

Figure 5. Number of horizontal superstructures considering

span and age.

1. The bridge on 1st class road over a spillway near

the village of Pasohlávky (South Moravian
region) – a four-span slab bridge with the super­
Figure 2. Number of bridges on motorways and 1st class structure formed in the longitudinal direction
roads (above) and on 2nd and 3rd class roads (below) within from prefabricated prestressed beams I-73 115/
the program region considering year of construction. 125 of length 27 m; 8 beams are located in the
cross-section. Built in 1979, SC 6.
2. The Fryčajova bridge on 2nd class road across the
River Svitava (Brno, South Moravian region) –
a slab bridge with the superstructure formed in the
longitudinal direction from prefabricated pre­
stressed beams of the I-MPD 6 type; 12 beams are
located in the cross-section. Built in 1955, SC 6.
3. The Bělohorská bridge on 2nd class road across
tram lines (Brno, South Moravian region) –
a two-span slab bridge with the superstructure
formed in the longitudinal direction from prefab­
ricated prestressed I-73 beams of length 27 m; 13
beams are located in the cross-section. Built in
1982, SC 6.
Figure 3. Number of bridges on local roads (Prague,
4. The Veslařská bridge on local road across the
Brno, Plzeň) with classification of 5–7 considering River Svratka (Brno, South Moravian region) –
material. a three-span slab bridge with the superstructure
formed in the longitudinal direction from prefab­
ricated prestressed KA-61 beams; 16 beams are
to 10 m are the most common characteristics of rail­ located in the cross-section. Built in 1966, SC 5.
way bridges in the Czech Republic; see Figures 4–5. 5. The Horní Cerekev bridge within railway sec­
For five case studies, four road bridges and one tion Veselí n. L.–Jihlava (Vysočina region) –
railway bridge were selected as follows: A one-span slab bridge made of two

224
 superstructures formed by cast-in steel beams; Non-linear finite element model was created in
27 beams per superstructure. Built in 1887, software ATENA and it consists of 13,000 elements
reconstructed in 1936, SC 2. of hexahedra type in the major part of volume and tri­
angular ‘PRISM’ elements in the blue-colored parts
Available documentation of selected bridges was of cross-section (see Figure 6). Concrete is described
analyzed in terms of: (i) administrative data; (ii) by a non-linear mathematical model according to Čer­
design and construction specifications; (iii) oper­ venka & Papanikolaou (2008) based on the theory of
ational characteristics; (iv) schematic diagrams and fracture mechanics. Reinforcement and tendons are
photos; (v) material; (vi) load; and (vii) defects and represented by discrete 1D elements with geometry
damages. Data from standard and main bridge according to original documentation. The ultimate
inspections and diagnostic surveys, original design limit state is represented by the load applied in the
documentation and construction diary, including last step of the analysis representing structural failure
inspection and probative material tests, are analyzed. (peak of load-deflection diagram). Bending failure
A loading test was also available for some of the mode and locations of cracks can be seen in Figure 7.
bridges. In next steps, all the structures are going to Variability of concrete material characteristics is
be modeled with nonlinear FEM tools and recalcu­ summarized in Table 1 and it is assumed according
lated on deterministic and stochastic levels using to recommendations of JCSS (JCSS 2001) –
actual codes and semi-probabilistic approach. Partial Young’s modulus E, laboratory experiments of
results are presented in following Section 4. material obtained from original documentation of the
bridge, such as compressive strength of concrete fc –
4 CASE STUDY and experiences from large testing campaign (Novák
et al. 2015, Strauss et al. 2017) – tensile strength of
The very first case study (selected bridge No. 4 in sec­ concrete ft and fracture energy Gf . Mean values of
tion 3) is focused on comparison of ultimate limit state E; ft and Gf were determined according to formulas
design values of resistance determined by normative implemented in the fib Model Code 2010 (fib Feder­
methods and semi-probabilistic approach. In the first ation internationale du beton 2013). Note that PDF
step of the study, the structure is represented by one means probability density function.
girder: simply supported prefabricated prestressed For correct comparison of all selected methods,
bridge girder KA-61 failing in bending and loaded by there is an assumption of uncorrelated variables.
a half of special six-axle truck according to Eurocode Note that, there are model uncertainties in partial
national annex. The KA-61 girder is 19.98 m long; safety factors tabular method usingγM taken from
a rectangular cross-section (see Figure 6) is Eurocode and in EN 1992-2 method, thus it was
0.85 m high and 0.98 m wide. The girder is prestressed necessary to reduce design values determined by
by 15 tendons (4 parabolic and 11 straight). Material partial safety factors calibrated by experiments and
of reinforcement is 10 400B and concrete class is of advanced semi-probabilistic methods by γRd ¼ 1:06.
C35/45 according to bridge documentation. Determined results are depicted in Figure 8. Ref­
erence statistical moments of resistance R were
obtained by LHS method with 30 simulations: the
mean value is μLHS ¼ 692:7 kN and standard devi­
ation is σLHS ¼ 40:2 kN. The ultimate limit state
design values were determined for reliability index
β ¼ 3:8 (which leads to target failure probability
pf ¼ 0:0012) and under the assumption of Lognor­
mal distribution of resistance. Determined results are
depicted in Figure 8.
Although design value estimated by ECoV method
is very close to the result of LHS, it significantly under­
Figure 6. KA-61 girder cross-section. estimates standard deviation (σECoV ¼ 21:5 kN) and
the mean value (μECoV ¼ 652:8 kN). On the other

Table 1. Stochastic parameters of concrete material

model.

Parameter Mean CoV [%] PDF Unit

fc 59 9 Lognormal [MPa]
ft 4.6 16 Lognormal [MPa]
E 38.8 15 Lognormal [GPa]
Figure 7. Location of cracks during failure of KA-61
Gf 152 22 Lognormal [J·m2]
girder.

225
 awarded by the European Regional Development Fund
within the European Union program Interreg Austria–
Czech Republic. Moreover, theoretical background of
the Czech Science Foundation (GAČR) project No.
20-01781S is acknowledged.

REFERENCES
CEN (2002). EN 1990 – Eurocode: basis of structural
Figure 8. Comparison of ultimate limit state design values
design. Brussels: European Comitee for Standardization.
Rd determined by safety formats and advanced reliability
CEN (2005). EN 1992 – Eurocode: design of concrete struc­
methods.
tures. Brussels: European Comitee for Standardization.
Červenka, V. (2013). Reliability-based non-linear analysis
according to fib model code 2010. Structural Concrete
Journal fib. 14:19–28.
hand, ECoV by Schlune slightly overestimates stand­ fib Federation internationale du beton (2013). fib Model
ard deviation σECoVSchlune ¼ 44:8 kN, which in combin­ Code for Concrete Structures 2010. Ernst & Sohn.
ation with the mean value of μECoVSchlune ¼ 652:8 kN JCSS (2001). JCSS Probabilistic Model Code. Joint Com­
(the same as in case of standard ECoV) leads to lower mittee on Structural Safety. ISBN 978-3-909386-79-6.
design value of resistance. Nevertheless, the prob­ Lebrun, R. & A. Dutfoy (2009). A generalization of the
lem is in the estimation of mean value by a single Nataf transformation to distributions with elliptical
copula. Probabilistic Engineering Mechanics. 24:172–8.
simulation with mean values of material character­
doi:10.1016/j.probengmech.2008.05.001.
istics Rm ¼ Rðfcm ; fym ; anom ; …Þ, which does not McKay, M. D., R. J. Beckman, & W. J. Conover (1979).
lead to correct estimation in this case. Interesting Comparison of three methods for selecting values of
results are obtained by normative methods: partial input variables in the analysis of output from
safety factors (PSF) method highly overestimates a computer code. Technometrics, 239–245.
the design value of resistance (even after calibration Novák, D., M. Vořechovský, & B. Teplý (2014). FReET:
of safety factors by experiments). On the other Software for the statistical and reliability analysis of engin­
hand, EN 1992-2 method leads to very accurate eering problems and FReET-D: Degradation module.
result close to the reference value obtained by LHS. Advances in Engineering Software. 72:179–192.
Novák, D., L. Řoutil, L. Novák, O. Slowik, A. Strauss,
Further work will be focused on updating the sto­
& B. Krug (2015). Database of fracture-mechanical
chastic model by diagnostic survey, FEM modeling concrete parameters and its implementation into reli­
of the whole bridge and mathematical modeling of ability software freet. Proc. of the 13th International
defects and deterioration. Probabilistic Workshop (IPW 2015) (ISBN 978-981­
09-7963-8).
Rotter, T., M. Teichman, V. Škoch, & J. Volek (2018). Stav
5 CONCLUSIONS mostů v České republice [condition of bridges in the
czech republic]. In Proceedings of 23rd International
In the framework of the European Project Symposium Bridges 2018 – Appendix, pp. 1–21. (in
Czech).
‘ATCZ190 SAFEBRIDGE’, five existing bridges
ŘSD ČR (2018). Přehledy z informačního systému
within the Czech regions of South Moravia, o silniční a dálniční síti ČR – stav k 1. 7. 2018 [Over­
South Bohemia and Vysočina were selected to view of the information system on the road and motor­
be studied and modeled with NLFEM on deter­ way network in the Czech Republic – data on 1. 7.
ministic and stochastic levels. The statistical 2018]. https://www.rsd.cz/wps/portal/web/rsd/Silnicni­
data provided by the main national operators for databanka.
road and railway bridges were used for this pur­ Schlune, H., K. Gylltoft, & M. Plos (2012). Safety formats
pose. Strategies of safety formats approaches for non-linear analysis of concrete structures. Magazine
and advanced reliability methods for nonlinear of Concrete Research. 64:563–74.
Strauss, A., B. Krug, O. Slowik, & D. Novák (2017). Com­
FEM analysis of structures were also presented.
bined shear and flexure performance of prestressing con­
As the very first case study, normative methods crete T-shaped beams: Experiment and deterministic
and semi-probabilistic approach were applied to modeling. Structural Concrete. doi: 10.1002/
determine the design value of ultimate capacity suco.201700079.
of KA-61 girder. Results according to individual Červenka, J. & V. K. Papanikolaou (2008). Three dimen­
methods were discussed. sional combined fracture-plastic material model for
concrete. International Journal of Plasticity,
2192–2220.
ACKNOWLEDGMENT Vořechovský, M. & D. Novák (2009). Correlation control
in small sample Monte Carlo type simulations
The authors would like to acknowledge the financial I: A Simulated Annealing approach. Probabilistic
support of the ‘ATCZ190 SAFEBRIDGE’ project, Engineering Mechanics. ISSN 0266–8920.

226
 Bridge Maintenance, Safety, Management, Life-Cycle Sustainability and Innovations –
Yokota & Frangopol (eds)
© 2021 Taylor & Francis Group, London, ISBN 978-0-367-23278-8

Monitoring systems for masonry tunnels

Alfred Strauss
Department of Civil Engineering and Natural Hazards, University of Natural Resources and Life Sciences, Vienna,
Austria

Hans Neuner & Corinna Harmening

Department of Geodesy and Geoinformation, TU Wien, Vienna, Austria

Christian Seywald
Streckenmanagement und Anlagenentwicklung, Fachbereich Bautechnik, Tunnelbau Austrian Federal Railways,
Salzburg, Austria

Michael Österreicher
iC consulenten ZT GmbH, Vienna, Austria

Elisabetta Pistone
Vienna Consulting Engineers ZT GmbH, Vienna, Austria

ABSTRACT: The paper presents a review of testing methods and a classification of strategies and tools in
terms of technologies and techniques applied to the monitoring of tunnels during rehabilitation measures. There
is a need of an assessment tool to correctly use classic and innovative monitoring systems while surveying exist­
ing tunnel system before and after tunneling rehabilitation measures. A uniform classification of monitoring
methods and their suitability is presented in this contribution with respect to standards and the successive imple­
mentation steps. The whole analysis is based on European international research projects as well as on the
record of many years of practical experience and implementation of monitoring management systems. The form
of the monitoring system evaluation below is on masonry tunnels that are in need of rehabilitation.

1 INTRODUCTION load-bearing capacity of the structure can occur due

to masonry removing, thus making monitoring during
In Austria, the ÖBB-Infra AG, as infrastructure com­ construction indispensable.
pany for rail transportation, operates about 150 tunnel The concept of a monitoring system for perform­
structures that are more than 100 years old and the ance evaluation applied to new and existing tunnel
majority consists of masonry tunnels. The masonry cross sections, especially during a construction pro­
can be very different in nature and can include various ject, requires the differentiation among the following
types, ranging from sandstone or brick masonry to scale levels:
natural stone and quarry stone brickwork. Aging
effects as well as effects by the operation life or the • system: overall cross section of the tunnel.
environment can trigger damages, especially in the • Subsystem: ridge, calotte, bench, elm, sole, lining
area of the vault and of the abutments. These typically segments.
can manifest in the form of cracks (longitudinal, • Material: soil, rock, concrete, spayed concrete,
transverse, oblique or reticular), spalling, shuttering masonry.
as well as holes and wet spots. In order to repair these As a wide and general consideration, existing
damages, renovation measures of various types have tunnel systems are governed by several defect types
to be carried out (for example, conventional vault while degradation mechanisms are associated with
renewal, tunnel-in-tunnels, vault renewal by full- maintenance measures. Regarding monitoring pol­
construction or partial renewal). Depending on the icies, they can be load dependent or load independent
depth and the profile width, potential reduction in the and a distinction regarding the temporal perspective

DOI: 10.1201/9780429279119-27
https://doi.org/10.1201/9780429279119-27

227
should be made between short term, long term and during maintenance work, a variety of criteria are
construction phase. In particular, in these three phases defined, which can be classified into four categories.
in the immediate area of construction, as well as in (1) The first one characterizes the systems’ perform­
areas before the construction/or intervention, or in the ance parameters. It includes the quality criterion
area after the intervention semi-destructive testing accuracy, the measuring range, the distortion’s
(SDT) methods should be provided, if appropriate. dimension – the object area as well as the geometric
Technical monitoring systems are mainly carried dimensions – and the time-related characteristics,
out during construction phases and are mainly used which are relevant for the sensors’ integration into
for quality assurance and guideline estimation of the multi-sensor-systems as well as for the acquisition of
overuse of remaining tunnel cross section or tunnel­ dynamic processes. (2) The second category contains
ing material, which is left during the intervention. parameters, which characterize the sensors’ use in
The scope is the exact estimation of the tunnel con­ the field. These include the dimension, the assembly
dition in areas that cannot be inspected differently. and the reaction with respect to environmental influ­
The short and long-term monitoring concepts and ences. In a subsequent level, the latter can be subdiv­
the associated sensor provisioning and data transmis­ ided into atmospheric and operating conditions, the
sion have to be planned accordingly in areas after influences due to the site situation as well as the
completion of interventions. These monitoring sys­ measuring object’s properties. (3) The third category
tems allow a structural performance evaluation for serves as a general assessment of the sensor systems
the performed interventions and construction meas­ and comprises the parameters costs, detectability of
ures, whereby the monitoring effects are depending different types of damage (e.g. spalling, cavities, wet
on the type of monitoring system. areas, cracks) and the general suitability with regard
The fact that three different levels can be distin­ to the monitoring of tunnels in general and of tunnels
guished for the application of monitoring systems, in use in particular. The table’s rows represent the
gives access to different information, but it also sensors’ classification according to the type of
requires the appropriate choice of monitoring sys­ acquired measures. Primarily, these are the directly
tems. However, these systems and associated sensors measured linear movements. In the further categories
will also vary depending on various situations, e.g.: (angular movements, velocity, acceleration, strain
and crack measurement), intermediate steps are
• tunnels with innovative structural and/or material
required during the analysis in order to infer the
solutions,
occurred movement.
• very important tunnel structures that represent
Among the systems for measuring linear displace­
strategic components of transportation system,
ments the total station, which belongs to the category
• damaged structures awaiting rehabilitation or
of geodetic sensors, has to be highlighted. The total
replacement,
station is a classical measuring instrument for per­
• selected reference tunnel structures, representa­
forming 3D convergence measurements, whose use is
tive of national tunnels assets, monitored for cre­
recommended for performing reference measurements
ation and verification of degradation models.
when evaluating alternative measuring systems. Lev­
Thus, when designing a monitoring system for elling is difficult to automate, whereas in photogram­
existing tunnel systems, it is necessary to further metric procedures the dependence on light conditions
characterize the frequency of measurements, the has a negative effect. From a technical point of view,
range of measurements, the type of monitored pro­ terrestrial laser scanner would have been very suitable
cesses and the data transmission. for the monitoring. However, as no sophisticated
Furthermore, in existing tunnels, both SDT and monitoring software exists, and because of their great
Nondestructive Testing (NDT) methods are used. For expense, they are not the first choice.
a detailed analysis, lists and specific applications, we A compromise between the space-continuous 3D­
refer the reader to Strauss et al. (2020). The main dif­ acquisition of the measuring object and the pointwise
ference between NDT and SDT methods, is that in the acquisition by means of total stations are profile scan­
case of SDT, either material samples have to be taken ner from the automotive sector. They are also more
out or minor breaches of structural integrity have to be economically advantageous than 3D laser scanners,
performed. In relation to the monitoring of loads and which is why it is worth testing them. Mechanical
environmental effects for existing tunnels or tunnels sensor systems like strain gauges with invar wires are
during maintenance work, a clear distinction in the not pursued due to their low level of automation and
classifications of bridge structures has to specified. their significant rising into the operating area. Induct­
ive sensors are characterized by their low standard
deviation, a high level of automation and small dimen­
2 ANALYSIS OF SENSOR SYSTEMS FOR OLD sion. Thus, they are suitable for the assigned task.
MASONRY TUNNELS Capacitive sensors have similar properties. However,
due to their sensitivity with respect to vibrations, they
By analyzing the sensor systems, which are worth to do not have an explicit added value compared to
be considered for a monitoring of masonry tunnels inductive sensors. A classification matrix concerning

228
monitoring systems that can be used during mainten­ because of their level of accuracy and their stability
ance work on old masonry tunnels is given in characteristics. For inclinometers, there exists
Figure 1. a variety of solutions. Some of them are very well
A similar situation occurs in case of the eddy cur­ suited for the monitoring of masonry tunnels, among
rent sensors, which also have similar properties as the other things because of their high accuracy and high
inductive sensors, but are affected by alternating level of automation, small dimensions as well as
fields and vibrations. Solutions in the field of fibre their favourable stability characteristics (especially
optics are technically very suitable because they are servo solutions). When choosing an appropriate
very precise, but on the other hand, they are also sensor, its temperature stability as well as its damp­
expensive. Radar techniques and hydraulic/hydrostatic ing properties are of further interest; the latter espe­
measurement methods have – similar as the mechan­ cially with regard to the sensors’ vulnerability to
ical sensors – a considerable impact on the tunnel’s vibrations. In principle, these systems are signifi­
operation and, thus, are not suited. Laser techniques cantly cheaper than fibre-optic systems. Although
are, on principle, suitable for the tunnel monitoring. the latter measures with high precision, the angular
However, they do not stand out from profile scan­ change is only indirectly acquired by a strain meas­
ners – which acquire a variety of measurement points urement. Thus, in this study, they are seen as an
as an added value – and/or inductive sensors – which alternative to the inclinometers with regard to higher
provide higher accuracies. Thus, they are not taken measuring frequencies and a lower degree of influ­
into account for a closer inspection. ence due to vibrations.
Figure 2 gives an overview of the scope of applica­ The suitability of the different monitoring systems
tion for various monitoring systems with a particular at varying surrounding conditions is presented in
focus on their application limits. Figure 3.
In the class of the angular displacements measur­ Four sensor types were investigated with regard to
ing sensors, the linear movements are either derived the detection of strains. (1) Electrical resistance gauges
from the acquired angular change and the subsequent have proven to be a good choice, as they score particu­
computation of the bending line or from the concat­ larly for their comparatively simple assembly, low
enation of similarity transformations. Micro-electro­ price and high accuracy. In case higher measuring fre­
mechanical systems (MEMS) as well as fibre optic quencies are required, (2) fibre-optic sensors, which
gyros are not suitable for a long-time monitoring allow a high-accurate and high-frequent strain detec­
tion, are a well suited alternative. However, they
require the largest financial expense of the four sensors
under investigation. (3) Vibrating wire sensors are
characterized by a comparatively large measuring
range. However, as they are very sensitive with regard
to vibrations, and can be disturbed by passing trains as
well as the construction works, they are only suitable
to a limited extent. As a fourth option, (4) MEMS sen­
sors were investigated. However, they were excluded
because of their drift and their comparatively low
accuracy.
The four investigated sensor types for vibration vel­
ocity and acceleration measurement (piezoelectric sen­
Figure 1. Classification sheet of the surveyed monitoring sors, capacitive sensors, inertial sensors and inductive
systems for the maintenance/intervention work of old sensors) are all in the same price segment and, further­
masonry tunnels – first cutout (Strauss and Neuner 2019). more, have similar properties in the remaining

Figure 2. Classification sheet of the surveyed monitoring Figure 3. Classification sheet of the surveyed monitoring
systems for the maintenance/intervention work of old systems for the maintenance/intervention work of old
masonry tunnels – first cutout (Strauss and Neuner 2019). masonry tunnels – third cutout (Strauss and Neuner 2019).

229
assessment categories. A general disadvantage of this conditions of existing tunnels or tunnels that are
sensor group is the increase of the noise level due to under construction:
the temporal integration of the measured values, which
• material level: strain sensors, fiber optic sensors,
is necessary to obtain the linear displacement. A dis­
transducers,
tinction has to be made with regard to the measuring
• component or component level: fiber-optic sen­
frequency, which is the smallest for the inertial sensors.
sors, displacement transducers, inclination
An advantage of the inductive sensors with respect to
transducers,
the other three sensor types is their insensitivity regard­
• system level or tunnel cross section: fiber optic
ing vibrations and pressure waves caused by passing
sensors, displacement transducers, tilt transducers.
trains. Piezoelectric sensors, capacitive sensors and
inertial sensors are therefore only suitable to a limited
extent for use during running train traffic.
When evaluating the investigated sensor types
with regard to the detection of cracks, the same
exclusion criteria comes into account: the vibrating
wire sensors are limited suitable due to their high
susceptibility to vibrations. Inductive and fibre-optic
sensors offer the best conditions in a metrological
sense in order to detect cracks, with the former being
the cheaper alternative.

3 PROPOSED SENSOR SYSTEMS FOR OLD Figure 4. Tunnel section and possible sensor systems
MASONRY TUNNELS (Strauss and Neuner 2019).

For the analysis of the appropriate monitoring systems

capable of assessing the performance of the tunnel For the suitability evaluation of monitoring sys­
cross-section immediately before and after the inter­ tems concerning sensor applicability during recon­
vention area, two observation lengths were defined struction works in tunnels under operation, the
with respect to the longitudinal axis of the tunnel: following criteria have to be considered:
• observation length of 10.00 m before and after • workload of instrumentation, A.
the tunnel construction site, • Flexibility and multiple use in assembly, F.
• observation length of 100.00 m after the tunnel • Data transmission and associated cabling, D.
construction site. • Uncertainties or influences deriving from interfer­
ences, such as contact lines, M.
The monitoring task of controlling limit states in: • Costs of sensor technology, installation, operation
• ridge, calotte, bench, elm, sole and lining per m², K.
segments, Sensor System for corridor of smaller than 10.00 m
• soil or rock, concrete or sprayed concrete,
masonry structures,
• total cross-section of tunnel tube,
can be divided into the following three monitoring
tasks:
• monitoring at material level (deformations,
strains, stresses),
• monitoring at component level (deformations,
inclinations, crack pattern developments, etc.),
• monitoring at system level, e.g., the tunnel cross-
section (deformations, inclinations, convergence
measurements, etc.).
A purposeful and robust monitoring system will
be a combination of these three levels. Depending Figure 5. Tunnel section and sensor systems for a design
length of 10.00 m (Strauss and Neuner 2019).
then on local circumstances and specifications of the
instrumentation degree relative to the three levels,
a diversification of the monitoring system will result The aforementioned suitability evaluation, when
(see Fehler! Verweisquelle konnte nicht gefunden applied to an observation assessment length of
werden.). The following systems are suitable as 10.00 m, leads to the following considerations on the
a first approximation, taking into account boundary basis of literature search, expert survey and data

230
product sheets, whereby color green means “appro­
priate”, color red means “not appropriate” and color
black means “neutral”:
• material level: strain sensors (A, F, D, M, K), fiber
optic sensors (A, F, D, M, K), position transducer
(A, F, D, M, K).
• Component or component level: fiber optic sen­
sors (A, F, D, M, K), displacement transducers
(A, F, D, M, K), inclinometers (A, F, D, M, K).
• System level or tunnel cross-section: fiber optic
sensors (A, F, D, M, K), displacement transducers
(A, F, D, M, K), inclinometers (A, F, D, M, K).
Sensor System for corridor of larger than 100.00 m
Figure 7. Cost analysis of the monitoring systems for
maintenance and repair work on old masonry tunnels
(Strauss and Neuner 2019).

investigations as a function of the observation evalu­

ation lengths for four different sensor types, namely,
inclination sensors, strain gauges, displacement sen­
sors and fiber-optic sensors.

5 DISCUSSION

The following recommendations stem from the pre­

vious discussions. In particular, the following solu­
tions result to be the best possible ones in terms of
Figure 6. Tunnel section and sensor systems for a design
attainable performance and cost:
length of 100.00 m (Strauss and Neuner 2019).
• the use of displacement transducers (vibration lat­
eral sensors), inclination transducers in combin­
The aforementioned suitability evaluation, when ation with strain sensors, in case of an evaluation
applied to an observation assessment length of length of <100 m.
100.00 m, leads to the following considerations on • The use of fiber optic sensors (Bragg systems) in
the basis of literature search, expert survey and data combination with convergence measuring belts or
product sheets: the use of Brillouin in case of an evaluation
length of > 100 m.
• material level: strain sensors (A, F, D, M, K), fiber • The use of fiber optic sensors (Bragg systems) in
optic sensors (A, F, D, M, K), position transducer combination with convergence measuring belts
(A, F, D, M, K). (possibly in combination with displacement trans­
• Component or component level: fiber optic sen­ ducers), for an extensive monitoring, in case of
sors (A, F, D, M, K), displacement transducers an evaluation length of > 100.
(A, F, D, M, K) inclinometers (A, F, D, M, K).
• System level or tunnel cross-section: fiber optic
sensors (A, F, D, M, K), displacement transducers
(A, F, D, M, K), inclinometers (A, F, D, M, K). 6 SENSOR SYSTEMS BEYOND FOR OLD
MASONRY TUNNELS

The following monitoring systems are listed in

4 COST ANALYSIS OF PROPOSED SENSOR accordance with the current state of the art:
SYSTEMS FOR OLD MASONRY TUNNELS
• material level: thermographic camera systems,
Cost analysis of the monitoring and sensors systems thermocouple systems, hyperspectral systems,
supporting the refurbishment measures in masonry voltage optics.
tunnels. This process comprises the measuring • Component or component level: thermographic
system, the data transmission, the data acquisition, camera systems, hyperspectral systems, impact
the data processing, the interpretation of data, the vehicle systems, echo methods, distributed fiber
decision and maintenance processes. Figure 7 shows Brillouin sensors, quasi strain gage networks sys­
the monitoring costs derived from the surveys and tems, voltage, optics.

231
 In particular, the topic is contextualized through
an introduction regarding the monitoring in tunnels
during rehabilitation measures and related topics
followed by the analysis of sensor systems for the
maintenance work in the case of old masonry
tunnels.
The research and survey related with this project
show for example:
• the taxonomy of monitoring plays a very import­
ant role for the monitoring tasks in tunneling and
Figure 8(a). Innovative monitoring concepts for mainten­ allows a very good overview of the damage
ance and repair work on old masonry tunnels (Strauss and mechanisms, the possible detection systems and
Neuner 2019). monitoring tasks,
• In addition to defect classification, knowledge of
the mechanisms of degradation, the choice of
monitoring strategies and the use of different
monitoring technologies is of high importance,
and includes also the implementation and oper­
ation of the monitoring system,
• It could be shown that the monitoring of the inter­
locking of several disciplines, such as NDT of the
SDT, from geodetic measuring methods to infor­
mation processing is effective
• The monitoring tasks and strategies as well as the
choice of sensor are very dependent on the exter­
nal boundary conditions, in this case the site con­
ditions during a tunnel rehabilitation. The
procedure for choosing an optimized monitoring
concept for the repair of existing natural stone or
masonry tunnels was presented.
In addition, the following recommendations stem
from the discussions about tunnel rehabilitation. In
particular, the following solutions result to be the
best possible ones in terms of attainable performance
and cost:
• the use of displacement transducers (vibration lat­
eral sensors), inclination transducers in combin­
Figure 8(b). Innovative monitoring concepts for mainten­ ation with strain sensors, in case of an evaluation
ance and repair work on old masonry tunnels (Strauss and length of <100 m.
Neuner 2019). • The use of fiber optic sensors (Bragg systems) in
combination with convergence measuring belts or
• System level or tunnel cross-section: thermo- the use of Brillouin in case of an evaluation
graphic systems, impact vehicle systems. length of > 100 m.

In addition, Figure 8 shows further innovative The use of fiber optic sensors (Bragg systems) in
combination with convergence measuring belts (pos­
monitoring concepts that might be of high potential
sibly in combination with displacement transducers),
for maintenance and repair work on old masonry
for an extensive monitoring, in case of an evaluation
tunnels, too, and will be investigated in more detail
in next research activities. length of > 100.

REFERENCES
7 CONCLUSIONS
Mottola, L., Picco, G.P., Ceriotti, M., Gunǎ, Ş. and
The paper presents a review of testing methods Murphy, A.L. (2010) Not all wireless sensor networks
and a classification of strategies and tools in are created equal: A comparative study on tunnels.
terms of technologies and techniques applied to ACM Transactions on Sensor Networks (TOSN), 7
the monitoring of tunnels during rehabilitation (2), p.15. (see also http://triton.disi.unitn.it/laboratory.
measures. html)

232
Strauss, A., Wendner, R., Bergmeister, K., Reiterer, M., of Bridges and Bridge Networks - Proceedings of the
Horvatits, J. (2011) Monitoring and influence lines 8th International Conference on Bridge Maintenance,
based performance indicators. Beton- und Stahlbeton­ Safety and Management, IABMAS 2016, pp. 965–970.
bau, 106 (4), pp. 231–240. Strauss, A., Neuner, H.,B., (2019) Machbarkeitsstudie und
Strauss, A., Vidovic, A., Zambon, I., Dengg, F., Matos, J.C. Monitoringsysteme für Tunnelinstandsetzungsmaßnah­
(2016) Performance indicators for roadway bridges. men – Report Institut für Konstruktiver Ingenieurbau
Maintenance, Monitoring, Safety, Risk and Resilience (IKI), BOKU-Universität für Bodenkultur.

233
 Bridge Maintenance, Safety, Management, Life-Cycle Sustainability and Innovations –
Yokota & Frangopol (eds)
© 2021 Taylor & Francis Group, London, ISBN 978-0-367-23278-8

Improved assessment of concrete bridges

M. Hauser, M. Rigler, E. Apostolidi & A. Strauss
University of Natural Resources and Life Sciences, Vienna, Austria

T. Zimmermann
City of Vienna, Building Inspection MA37, Vienna, Austria

D.M. Frangopol
Department of Civil and Environmental Engineering, Engineering Research Center for Advanced Technology for
Large Structural Systems (ATLSS Center), Lehigh University, Bethlehem, PA, USA

ABSTRACT: The assessment and retrofitting of existing bridge structures is an important task for structural
engineers. A new approach to bridge evaluation was developed, involving a multi-level strategy for structural
analysis. These analysis concepts for verifying the required safety margins provide engineers with tools to
more accurately simulate the structural behaviour of bridges, as needed to predict the failure mode. This paper
describes the basic procedures and demonstrates the applicability and the potential benefits of an increasingly
advanced and accurate analysis for bridge assessment using nonlinear FE analysis and including model uncer­
tainties, of example case studies.

1 INTRODUCTION bridge assessment, it is necessary to include aspects of

structural safety. The main objective of this paper is to
In recent decades, the need for bridge assessment and describe these aspects through a process of connecting
repair, through reinforcement replacement and con­ safety concepts to different levels of the strategy, as
crete retrofitting has been increased. The reason for well as to demonstrate the multi-level approach in
evaluation measures is the aging bridge stock in practice.
Europe and the rising of heavy vehicle traffic loads.
In practice, evaluation is usually performed through
standard approaches that mainly involve linear elastic 2 BRIDGE ASSESSMENT PROCEDURE
analysis (Level 1 according to Figure 1). Advanced
methods, such as nonlinear analysis or probabilistic 2.1 Multi-level strategy for structural analysis
assessment, are rarely used. At the same time, large-
scale tests on concrete bridges have shown that the In Bagge et al. (2018), a multi-stage strategy was
current technical codes are often very conservative. described and investigated to systematically opti­
Difficulties in predicting the load-bearing capacity are mize the structural analysis of the bridge superstruc­
associated especially with the shear forces. Often, ture (see Figure 1). If an initial assessment does not
instead of a detailed consideration of the particularly satisfy the required safety standards, the accuracy of
safety - critical components, such as expansion joints, the structural ultimate limit state (ULS) assessment
supports, etc., they are only simulated with simple can be systematically improved by increasing the
and consequently overly conservative models. There­ level of accuracy. Based on the methods of calcula­
fore, it is important to develop robust and practical tion and the material models to be used, multiple
methods for detailed analysis. A multi-stage strategy stages of analysis can be defined.
for successively improving the structural assessment The first level of structural analysis is the sim­
of concrete bridges in load-bearing condition (ULS) plest. Using simple calculation methods, the internal
was described in Bagge et al. (2018). forces are determined and compared with the mater­
The multi-level strategy describes how the detailing ial resistance models used. This is done for different
in structural analysis can be systematically improved, internal forces such as bending moments, shear
but in order to make it complete and applicable to forces, normal forces and their combinations. In the

DOI: 10.1201/9780429279119-28
https://doi.org/10.1201/9780429279119-28

234
Figure 1. Multi-level strategy for structural analysis of bridges (adapted from Bagge et al. 2018).

following three approximation steps, the internal

forces are recorded step by step by structural ana­
lysis using nonlinear FE analysis. Therefore, no
explicit review at the highest level is required. or often by Equation (2):
Level 1 is called the initial stage and should
always be the basis for the bridge evaluation, to find
critical areas and potential failure modes. Levels 2 ­
4 are referred to as extended stages. With these
stages a more detailed evaluation is possible.
In conclusion, the first structural assessment is where Z = safety margin, R = resistance, S =
important in order to be able to concentrate on the action, G = continuous load, P = preload, Q = vari­
critical points and thus improve the quality of further able load.
analysis. If R = S then the system is exactly at the limit
state. This may concern the serviceability limit state
or fatigue limit state. At R < S then the system fails
2.2 Safety verification and at R > S the system is in a safe state. The safety
To guarantee a suitable level of safety, for increased level is defined with a certain probability of failure.
loads or decreased durability, it is necessary to carry Through the variables R and S, the failure mode
out a multi-level structural analysis. Starting from refers to the total structural resistance or action,
the simplest level (Level 1) of static analysis, a more respectively, but, in the advanced levels, also to the
detailed examination of single parts or entire struc­ local resistances and actions. The verification of the
tures can be performed. The deeper levels are associ­ safety level according to a probability-based concept
ated with different, complex approaches (Casas & is carried out according to Model Code 2010 (2013).
Wisniewski 2013; fib 2013; O’Brien et al. 2005; Depending on the degree of accuracy, the following
Strauss et al. 2011, Strauss et al. 2016, Suda et al. procedures can be applied:
2009, Zimmermann et al. 2012). The structure’s 1. Partial safety factor (PSF)
robustness, however, is not considered in this 2. Global resistance safety factor (GRSF)
approach, mainly due to the increased computational 3. Full probabilistic (FP) method
effort among other requirements. For this purpose,
Anitori et al. (2013) developed a method to calculate Level 1 analysis calculates the structural system and
and assess the robustness and the associated struc­ the corresponding internal forces elastically or plastic­
tural safety. The safety level is a key factor in the ally. These are compared with elastic or plastic resist­
assessment of reinforced concrete bridges and, ances, determining the cross-sectional or system-wide
above all, it should take into account the associated load-bearing capacity (see Figure 2). For level 1 ana­
economic, social and environmental risks (Melchers lysis the PSF and the FP method have proven them­
2001; Stewart et al. 2001; Sykora et al. 2017). selves and are often used. As Figure 1 shows, the next
The general condition for safety is given by Equa­ levels (2-4) are used to model and analyze the struc­
tion (1): tural behavior, especially the type of failure, but also
the deformation and load-bearing capacity more

235
 Resistance Safety Factor has been proposed, which
allows to more accurately consider the specific situ­
ation of any structure, and the potential interplay
between various factors contributing to the resistance
or actions, while still maintaining the expected level
of safety.
Even more realistic results can be achieved using
a fully probabilistic approach. Instead of working
with fractile values and applying safety factors at all,
the relevant input data for the calculations can be
described in the form of mean values and statistical
distributions. Out of these, stochastic values can be
obtained to be used in a series of calculations, result­
Figure 2. Framework for improving the safety verification ing in a statistical distribution of the generated
in the course of a multi - level strategy. results, which can be used directly to obtain the
probability of the structure’s failure. Thereby, the
modelling uncertainties regarding material proper­
ties, geometrical imperfections, etc. can be directly
precisely. In order to determine a realistic load capacity
and realistically depicted in the analysis.
and failure mode, it is necessary to determine the
Another important method of improving the
material behavior as precisely as possible. The way the
assessment is the use of data directly obtained from
material laws are modelled should reflect their real
the structure, like degradation, defects, site specific
properties as closely as possible, to reap the most bene­
load - and boundary conditions or material quality.
fits from the increased complexity of the calculations.
In – situ testing and monitoring is invaluable for an
The way the assessments and calculations are made
accurate depiction of the structure’s condition. The
can differ significantly between the levels. For this
use of the highly detailed approaches to structural
reason, these analyses should be verified based on at
analysis described above is limited without ground­
least two security formats, as required in the Model
ing in on – site, structure specific information.
Code 2010 (2013).
In conclusion, according to the considerations
above, the following guidelines can be given:
2.3 Successively improved assessment
• The initial assessment can be performed accord­
There are two different methods of increasing the ing to the relevant technical standards and design
quality and accuracy of the assessment described codes, or, if available, using data gathered from
above. On the one hand, the model used can be the structure’s documentation, or the structure
improved to be increasingly close to reality, suc­ itself. A fully probabilistic analysis however,
cessively reducing the amount of simplifications should be based on material data obtained from
and idealizations, at the cost of computational the structure itself, if at all possible. (O’brien et al
time, as well as time spent creating the model 2005) The use of such in – situ collected data is
itself. Going from handbook methods to increas­ generally recommended at al levels of assess­
ingly detailed FE – models and using elaborate ment. (Bagge et al 2018)
FE – solvers that account for the various effects • Partial safety factors can be used at any level of
of non-linearity, this can dramatically improve the analysis, especially for the actions, even when the
achievable level of safety. Due to the increased resistance is determined fully probabilistically or
effort required for the analysis, the use of smaller, using a general resistance safety factor. Site spe­
local models of the structure’s critical areas is cific loads should be included in a fully probabil­
often preferable (fib 2013). These local models istic assessment or can be used as an intermediate
can be used after the critical points of a structure step before proceeding to full probabilistic from
have been identified by using a (less demanding) a lower level analysis. (O’brien et al 2005)
lower assessment level. • The structure’s actual condition, geometry and
On the other hand, the underlying calculations defects should always be taken into consideration,
assumptions, especially the material parameters on regardless of the level of analysis.
the resistance side, and the actions can be improved. • Assumptions made for the structural analysis
The semi – probabilistic safety concept of the Euro- must be validated at all levels of assessment by
code for example used 5% - fractile values for comparison to the structure’s actual behavior.
material strengths, and 95% - fractile values for the This becomes especially important with increas­
actions, in combination with partial safety factors, ingly detailed modelling of the structure.
which decrease the resistances and increase the
loads, respectively. Instead of the rather rigid par­ If, as a result of the assessment, the structure is
tial – safety factor system, the use of a Global found to be unable to satisfy the stated requirements

236
of the ultimate limit state, improvements are in
order. In the course of these improvements, special
focus should be placed on aspects, where the assess­
ment has shown significant uncertainties. Mitigating
or removing these uncertainties can improve the
accuracy of the results of future assessments.

3 BRIDGE CASE STUDIES

Figure 4. Joint of the pre-stressed cable with the sealed
In the framework of the current research, several
tensioning heads.
characteristic bridges were selected that serve as case
studies for the application of the presented multi-level
procedure. In the current paper, two case studies are
presented with emphasis on the second one, where
the FE analysis results are presented and evaluated.

3.1 Road bridge B41.09-TÜ Schützberg

The bridge was built in 1979 and was designed as
a pre-stressed multi-span girder system (nine spans).
The distance at the end span and in the regular span
area is approximately 28.5 m and 39.0 m, respect­
ively. The span distances at the end spans are
approx. 28.5 m. The cross-section was selected to be
a box girder with a total upper chord width of
13 m. A general view and specific details of the
bridge are presented in Figure 3 to Figure 5, along
with some initial FE simulations. The FE analysis is
under process.
Figure 5. View of the box girder in an end field: (a) photo­
3.2 Railway bridge case study B12-1165 Paasdorf graph from the inside and (b) FE model (software SCIA).

3.2.1 General data

The bridge was constructed in 1981 and is
a reinforced single span concrete bridge over a road
with a clear span of 10 m and a width of 5.5 m. The
slab has a thickness of 85 cm and a skew angle of

Figure 6. General view of B12-1165 Paasdorf bridge.

46°. The bridge rests on two elastomer bearings per

side and a steel horizontal force bearing. A general
view of the bridge is presented in Figure 6.

3.2.2 Initial structural assessment

As described above, the initial linear calculation was
performed in SOFiSTiK using a relatively simple,
shell-element-based FE model. This model type is rea­
sonably close to engineering practice and, therefore,
can serve to illustrate the potential benefits of a more
elaborate assessment. The material properties derived
from the original plans and they were further adapted,
Figure 3. General view of B41.09-TÜ Schützberg bridge: following (ONR24008 2014), to fit a recalculation
(a) photograph and (b) FE model (software SCIA). according to current standards (Figure 7 and Figure 8).

237
 Figure 9. Stresses XX at the top of the slab, resulting from
the linear calculation (software SOFiSTiK).

Figure 7. Stress-strain law used for concrete in the linear

FE - calculation (output of the software SOFiSTiK).

Figure 10. Maximum principal compressive stress result­

ing from the linear calculation (software SOFiSTiK).

structure, since the relevant in – situ testing data was

not available.

3.2.3 Nonlinear assessment using a more detailed

Figure 8. Stress-strain law for reinforcement steel used in the model
linear FE- calculation (output of the software SOFiSTiK). Using the Software ATENA (2016a), a detailed 3D
nonlinear model was constructed, including the exist­
ing reinforcement. The nonlinear calculation accounts
For the actions, as well as for the resistances, for the effects of cracking in the concrete, compressive
partial safety factors were applied according to hardening, as well as other nonlinear phenomena like
the current standards (Table 1). The linear calcu­ redistribution of forces, plastic deformation of the
lation also cannot account for the effects of the reinforcement, tension stiffening, etc. The higher level
available reinforcement in the structure. All this of detail can provide a better understanding of the
results in a relatively conservative estimation of structure’s behavior and possible risks of failure. This
the structure’s capabilities. For comparison pur­ additional detailed assessment can also account for
poses between the results of linear and nonlinear nonlinear phenomena that increase the structure’s
assessment, some representative stresses occurring resistance and could potentially allow for a lower
in the structure were selected and are presented safety margin, moving the calculation results closer to
in Figure 9 and Figure 10. reality, while still maintaining the same overall level of
The simpler nonlinear calculation requires consider­ safety.
ably less computational effort than a more detailed The level 3 analysis is based on the least favorable
nonlinear analysis. However, linear analysis is well load case derived from the level 1 calculation. To stay
suited to find the most critical aspects or areas of congruent with the level 1 assessment, the same partial
a structure, where a detailed, nonlinear calculation can safety factors were applied to the resistance and the
be performed. A calculation for the suggested level 1b actions (Table 1). Comparing the results yields a basis
assessment (see Figure 2) was not possible for the to estimate potential benefits of the nonlinear analysis.
For the comparison, the maximum occurring com­
pressive stresses in the concrete, as well as the ten­
Table 1. Partial safety factors.
sile stresses in the reinforcement were used.
Additionally, the ATENA model provides a better
Resistance representation of the exact distribution of forces
throughout the model, not only in the concrete, but
γc (concrete) 1.5 also in the reinforcement, and gives a prediction of
γy (reinforcement steel) 1.15 where cracking in the concrete or plastic deform­
ation in the reinforcement steel is to be expected.
Actions The expected cracks from the nonlinear model can
γQ (leading live load) 1.45 also be compared to any cracking occurring in the
γG (self – weight and dead load) 1.35 real structure to validate the calculation or make pre­
dictions regarding structural damages.

238
Figure 11. Layout of the reinforcement in the slab bridge
(software ATENA).

Figure 13. Crack width [m] in the concrete predicted

The model consists of hexahedral volume elem­ through non-linear analysis (Software ATENA).
ents and includes the existing reinforcement, pre­
sented in Figure 11.
Comparing, e.g., the compressive stresses acting areas determined by the initial structural assessment,
in the concrete at the top-center part of the slab, the which in this case were the stresses in the slab’s center.
nonlinear analysis returns an average of about It can be shown that even when using partial safety
-10 MPa (Figure 12), well below the value of -12.7 factors instead of a full probabilistic analysis, the
to -12.9 MPa resulting from the linear calculation. increased detail in the calculation alone can improve
The maximum compressive and tensile stresses the level of safety. Combining the higher level of FE –
occurring in the reinforcement bars range from calculation with GRSF or FP assessments (see
-75.5 MPa to +82.1 MPa, way below the design Figure 2), could improve the expected results even
resistance of 434.8 MPa. This illustrates the further.
increased safety gained by the more complex
assessment, even while using the same partial
safety factors. Consequently, the loading could 4 MATERIAL MODELS USED FOR THE
potentially be substantially increased, while main­ RECALCULATIONS
taining the same level of safety. Especially, since
the semi-probabilistic Eurocode standard of using 4.1 Adaptation of the outdated material data
the 5% - fractile values with partial safety factors provided by the structure’s original
could safely be replaced by more realistic assump­ documentation
tions for actions and resistances (see section 2), to The original design was performed in 1980 based on
mobilize even more strength reserves of the struc­ the standards of that time. In regard to concrete, the
ture. The maximum crack width predicted by the obsolete ÖNORM B4200 specifies a resistance class
model (Figure 13) is 0.19 mm, which is comfort­ based on the compressive strength of cubes with a side
ably below the 0.3 mm crack width permitted by length of 200 mm, whereas calculations following the
(CEN 2004) for frequent load combinations. Eurocode are based on fck, which represents the 5% ­
fractile of the compressive strength of cylinders of
3.2.4 Summary of the enhanced structural analysis 150 mm length. Following ONR 24008, the outdated
For the more accurate load bearing capacity definition strength class “B 400” can be translated to fck = 26.4
and the better understanding of the structural behavior, kN/mm2, and a mean strength (fcm) of 34.4 kN/mm2.
the bridge was investigated at Level 3 of the multi­ From this, the additional material parameters needed
level strategy, presented in Figure 1. For comparison, for the calculations can be derived, based on the Euro-
an improved verification of the structural safety has code (CEN, 2004) and they are summarized in Table 2.
been made based on the PSF method according to the
proposed framework illustrated in Figure 2, in con­
junction with the enhanced nonlinear FE – calculation. 4.2 Constitutive material model used for the
The analysis – results were focused on the critical nonlinear FE - model
The nonlinear analysis is performed using the
CC3DNonLin-Cementitious2 material model. For
detailed information refer to (Červenka et al. 2013).
For the use of this constitutive model, additionally
required material parameters can be derived from the
basic parameters given in Figure 14. For the sake of
practicality, the parameters were adapted following
the suggestions given in the software’s documenta­
tion (ATENA 2016a).
The material model combines constitutive models
Figure 12. Stress in XX direction [MPa] in the concrete and for tensile and compressive behavior, using the Ran­
in the reinforcement (cross section in the slab mid span). kine failure criterion for modelling the response to

239
Table 2. Basic material properties of “B400” concrete stiffening are expected to be non – negligible. The
according to Eurocode 2. value of cts was chosen to be 0.4, in accordance with
the CEB-FIP Model Code 1990. (CEB-FIP 1993)
Characteristic fck 26.4 [MPa] The concrete material’s post – cracking
compressive strength behaviour is also accounted for in detail, e.g.
Mean compressive strength fcm 34.4 [MPa] regarding opening and closing of cracks and
Mean tensile strength fctm 2.63 [MPa] remaining shear resistance.
Mean modulus of elasticity Ecm 31900 [MPa]

5 CONCLUSIONS

The application of the proposed concept of

a multi-level assessment of engineering struc­
tures (Bagge et al 2018) has demonstrated that
by utilizing increasingly detailed methods of
structural analysis, a considerable amount of
safety can be gained. The closer the assessment
can come to the structure-mechanics happening
in the real structure, the more efficiently the
strength – reserves contained in the structure can
be accessed, while maintaining the expected
level of safety. This increased degree of accur­
acy can be achieved by improving the quality of
the calculation models applied in the assessment,
as well as depicting the actions and resistances
used for the calculations more true to reality, by
employing probabilistic methods. Furthermore,
the aspects of the structure, where the highest
Figure 14. Depiction of tension stiffening in the uncertainties exist can be identified in the pro­
CC3DNonLin-Cementitious2 material model (ATENA cess. As a consequence of a detailed assessment,
2016a). the measures taken to improve the structure can
more efficiently target the most critical areas, as
well as the greatest uncertainties found. Thus,
tensile stress. Once the tensile strength of the mater­ improvement measures can be applied more
ial is exceeded, the fracturing strain and the resulting accurately, but also uncertainties can be reduced,
crack opening are calculated. The material’s behav­ to make future recalculations more effective.
ior under compression is depicted using a plasticity As a result, multi-level assessments could contrib­
model based on Menétrey – William failure surface. ute significantly to the future longevity of structures.
The combination of these two models is accom­ In the light of the large amount of aging structures
plished by simultaneously solving equations (3) and worldwide, formulating a reasonably simple method,
(4), each representing one of the material behaviors, that could be applied by engineers in practice is
by way of iteration. desirable.

ACKNOWLEDGEMENTS

This work was carried out with the financial support

to be solved for of the research project “ATCZ190-SAFEBRIDGE”
in the framework of the EU program INTERREG
Austria – Czech Republic. The support is gratefully
acknowledged.

to be solved for REFERENCES

The CC3DNonLin-Cementitious2 can also account
Anitori, G., Casas, J. R., & Ghosn, M. 2013. “Redundancy
for the effect of tension stiffening. This is achieved and robustness in the design and evaluation of bridges:
by limiting the amount of tension softening as much European and north american perspectives.” Journal of
as possible. A factor cts is specified, tensile stress Bridge Engineering, 18(12), 1241–1251.
during tension softening cannot drop below ATENA. 2016a. “ATENA program documentation - Part 1:
The analysed railway Paasdorf bridge is heavily Theory.” V. Červenka, L. Jendele, and J. Červenka, eds.,
reinforced, consequently the effects of tension Červenka Consulting, Prague, Czech Republic, 312.

240
Austrian Standards Institute. 2014 „ONR 24008:2014 Stewart, M. G., Rosowsky, D. V., & Val, D. V. 2001. “Reli­
Bewertung der Tragfähigkeit bestehender Eisenbahn­ ability-based bridge assessment using risk-ranking deci­
und Straßenbrücken“. sion analysis.” Structural Safety, 23(4), 397–405.
Bagge, N., Nilimaa, J., & Elfgren, L. 2017a. “In-situ Strauss, A., Vidovic, A., Zambon, I., Dengg, F., Matos, J.C.
methods to determine residual prestress forces in 2016. Performance indicators for roadway bridges.
concrete bridges.” Engineering Structures, 135, Maintenance, Monitoring, Safety, Risk and Resilience
41–52. of Bridges and Bridge Networks - Proceedings of the
Bagge, N., Plos, M., & Popescu, C. 2018. “A multi-level 8th International Conference on Bridge Maintenance,
strategy for successively improved structural analysis of Safety and Management, IABMAS 2016, pp. 965–970.
existing concrete bridges: examination using Strauss, A., Wendner, R., Bergmeister, K., Reiterer, M.,
a prestressed concrete bridge tested to failure.” Journal Horvatits, J. 2011. “Monitoring and influence lines
of Structure and Infrastructure Engineering, 27–53. based performance indicators, Modellkorrekturfaktoren
Bagge, N., Popescu, C., & Elfgren, L. 2017b. Failure tests als “performance Indikatoren” für die Langzeitbewer­
on concrete bridges: Have we learnt the lessons? Struc­ tung der integralen Marktwasserbrücke S33.24, (2011)
ture and Infrastructure Engineering, 1–28. Beton- und Stahlbetonbau, 106 (4), pp. 231-240. DOI:
Casas, J. R., & Wisniewski, D. 2013. “Safety requirements 10.1002/best.201100003
and probabilistic models of resistance in the assessment Suda, J., Strauss, A., Rudolf-Miklau, F., Hubl, J. 2009.
of existing railway bridges.” Structure and Infrastructure Safety assessment of barrier structures. Structure and
Engineering, 9(6), 529–545. Infrastructure Engineering, 5 (4), pp. 311–324. DOI:
CEB-FIP. 1993. CEB-FIP MODEL CODE 1990. Commit­ 10.1080/15732470701189498
tee Euro-International du Béton. Sykora, M., Diamantidis, D., Holicky, M., & Jung, K.
CEN. 2004. Eurocode 2: Design of concrete structures ­ 2017. “Target reliability for existing structures consider­
Part 1-1: General rules and rules for buildings. fib­ ing economic and societal aspects.” Structure and Infra­
fédération internationale du Béton. 2013. fib Model structure Engineering, 13(1), 181–194.
Code for Concrete Structures 2010, 1st ed. Ernst & Zimmermann, T., Strauss, A., Bergmeister, K. 2012. Struc­
Sohn, Berlin. tural behavior of low-and normal-strength interface
Melchers, R. E. 2001. “Assessment of existing structures ­ mortar of masonry. Materials and Structures/Materiaux
Approaches and research needs.” Journal of Structural et Constructions, 45 (6), pp. 829–839.
Engineering, 127(4), 406–411. Zimmermann, T., Strauss, A., Lehký, D., Novák, D., &
O’Brien, E. J., Znidaric, A., Brady, K., González, A., & Keršner, Z. 2014. “Stochastic fracture-mechanical char­
O’Connor, A. 2005. “Procedures for the assessment of acteristics of concrete based on experiments and inverse
highway structures.” Proceedings of the Institution of analysis.” Construction and Building Materials, 73,
Civil Engineers - Transport, 158(1), 17–25. 535–543.

241
MS02: Recent trends in AI/IoT technologies for bridge maintenance
and safety
Organizers: T. Kitahara, H. Furuta, P. Chen & M. Beer
 Bridge Maintenance, Safety, Management, Life-Cycle Sustainability and Innovations –
Yokota & Frangopol (eds)
© 2021 Taylor & Francis Group, London, ISBN 978-0-367-23278-8

A new reliability model of bridge fatigue based on mind evolutionary

algorithm
J. Wu, B.F. Liu, K. Dong & M. Yang
Highway College, Chang’an University, Xi’an, Shaanxi, P.R.China

ABSTRACT: Fatigue is an important safety indicator of bridge member under long-term random live
loads. This paper focuses on the bridge member’s fatigue assessment caused by traffic and wind loads.
The numerical model of the fatigue analysis of the bridge-traffic-wind system is implicit, which includes
several time-progressive sub-modules. A new fatigue reliability assessment model is constructed based
on Mind Evolutionary Algorithm (MEA). The Back Propagation (BP) neutral network optimized by
MEA can be used to generate the limit state function by rationally considering all the time-progressive
random variables. Particle Swarm Optimization (PSO) is then applied to improve the traditional Monte
Carlo method, which can help choose samples with high efficiency. Through applying the new reliability-
computation model to a typical cable-stayed bridge under variable traffic and wind loads, it is found the
new method can compute the fatigue reliability index with higher efficiency compared to tradition model.

1 INTRODUCTION polynomial to express the nonlinear surface of the

whole random space (Faravelli 1989, Yang 2012).
Long-span bridges’ vibration is remarkable under wind The back propagation (BP) neutral network can
and traffic flow loadings since they are kind of slender express the nonlinear relationship between the input
structures with small stiffness. The long-term vibration and output parameters; the problem of BP neutral
will intrigue the fatigue problem of bridge members, network is the slow speed of convergence. Generally,
which will lead to serious damage or even collapse. researchers use genetic algorithm (GA) to improve
A rational reliability fatigue model of long-span the efficiency of BP neutral network. The crossover
bridges is desired to estimate the health state of and mutation used in GA may help find the optimum
bridges. gene or damage the original gene, and there is no
The bridge-wind-traffic system is a complicated method to assess its result. In 1998, Sun et al. pro­
coupling system; the fatigue calculation is an implicit posed mind evolutionary algorithm (MEA) which
problem and involves a large amount of calculation has stronger ability in searching global optimal solu­
(Wu 2010). Thus, it is hard to find an appropriate reli­ tion than GA. This method has been applied in fields
ability model for the fatigue assessment for long-span of transportation engineering and information engin­
bridges under wind and traffic flow. The commonly- eering, but has seldom been used in the nonlinear
used reliability methods in bridge engineering include structural analysis for bridges.
the first order reliability method (FORM), Monte This paper introduces a new reliability model based
Carlo method (MCM) and response surface method on mind evolutionary algorithm for long-span bridges’
(RSM) (Melchers 1999). The FORM has the advan­ fatigue problem. In the first part, the principle of the
tages of simplicity, but it is not suitable for implicit reliability model is presented. Subsequently, the model
problem. The MCM compute the reliability based on is applied to a typical long-span bridge. Finally, the
thousands of computations for random sampling; efficiency and rationality of the model is discussed.
which poses a prohibitively high workload for com­
plex system (Xiao 2002). The RSM reconstruct the
limit state function by substituting the implicate func­ 2 RELIABILITY MODEL BASED ON MIND
tion by polynomial; this is a method with high effi­ EVOLUTIONARY ALGORITHM
ciency to estimate the reliability index. However, for
large and complicated structure, the limit state function The whole framework of the reliability model is
is highly nonlinear, and it is hard to find a proper shown in Figure 1. The whole framework is divided

DOI: 10.1201/9780429279119-29
https://doi.org/10.1201/9780429279119-29

245
 2.2 MEA-BP neutral network
Back propagation (BP) neutral network is
a multi-layer feed-forward neutral network with
good self-taught and derivation ability to deal
with the nonlinear mapping problem. In the pro­
cess of network transmission, the information of
input data is passed from the input layer to the
hidden layer and finally reaches the output layer.
If the output layer results are not expected, the
error information will propagate backward. The
structure of BP neutral network is shown in
Figure 2. Xi (1, 2, …, n) and Y represent input
data and output results, respectively; ωij and ωjk
represent weights and thresholds between differ­
ent layers. For bridges, the basic parameters
and load conditions can be input to the network
Figure 1. Framework of reliability model based on mind to approximate the limit state function.
evolutionary algorithm.
In the process of information transmission,
the initial weights and thresholds of the network
into three parts. Firstly, the common scenarios of are determined by the researcher or random
wind and traffic flow are considered in the bridge­ data, and the fitting results and efficiency of the
wind-traffic coupling system and then construct the neutral network will be greatly affected by the
basic database. Secondly, the mind evolutionary initial value. Mind evolutionary algorithm
algorithm (MEA) is used to optimize the BP- (MEA) can help the neutral network get better
neutral network for obtaining the nonlinear rela­ initial parameters and optimize the function fit­
tionship between the input data and fatigue damage, ting ability.
and derived the approximated limit state function. The basic elements of MEA are shown in
Thirdly, the reliability index is computed through Figure 3 (Sun et al. 1998, 2000). The billboard is
optimized Monte Carlo method by particle swarm composed of global and local billboards. The
optimization algorithm. Each step is introduced in global billboard is to post the score of each sub­
the following sections. group and the local one is for individual’s score in
each sub-group. The score is the evaluation of the
individual’s action by environment which is used
2.1 Limit state function of fatigue damage to find the optimal individual and sub-group.
The bridge-wind-traffic coupling analysis can refer During the evolution process, all the individuals
to Wu (2010). The fatigue damage of bridge for each generation construct a group. Each group
member is computed based on linear cumulative is divided into sub-groups called group i. Sub­
damage theory (Miner 1945) as expressed by groups can be classified into superior group and
Eqn. (1): temporary group, as illustrated in Figure 4, the
optimization process of MEA-BP is as follows
(Sun et al. 1998):

where Ni = fatigue life of the structure; ni = loading

cycle number.
Through computation, when the cumulative
damage index of the structure is larger than 1, the
structure can be defined as fatigue failure; otherwise
the structure has a residual life. So an implicit limit
state function can be obtained based on Eqn. (2) and
above-mentioned theories:

where ΣDFy = cumulative fatigue damage index of

members over ny years. Figure 2. BP neutral network structure.

246
 In the process of MEA optimization, the group
size (SG) is calculated through Eqn. (3).

where popsize = population size; supsize = superior

group size; tempsize = temporary group size. The
superior group records the superior information while
the temporary group restored the competition process.
The number of input neurons S1 is determined by
the number of random variables considered in this
paper; the number of neurons in the hidden layer S2 is
obtained by empirical Eqn. (4); the number of output
Figure 3. Basic elements of MEA.
neurons S3 is determined by the limit state function.

where α is a constant between 1 and 10.

2.3 PSO-Monte Carlo sampling method

After obtaining the limit state function of bridge mem­
bers through MEA-BP neutral network, the PSO-
Monte Carlo method can be used to solve the failure
probability. Particle Swarm Optimization (PSO) was
first proposed by Kennedy & Eberhart (1995). PSO is
a global random search algorithm based on swarm
intelligence and the basic idea of PSO comes from the
study of predation behavior of birds. In this algorithm,
the potential solution for each optimization problem is
a particle in the search space. Each particle correspond­
ing to a fitting value is determined by the objective
function; the speed of particles determines their direc­
tion and distance.
Figure 4. Optimization process of MEA-BP neutral network. In the optimization process, the initialization is to
assign random values to the particle position and vel­
ocity. Each particle follows the existent optimal par­
ticle and searches in the solution space until
1) In the whole search space, numerous individuals satisfying the optimal requirements.
are randomly generated. Then, the optimal indi­
viduals and temporal individuals are found
through the score or adaptability. 3 CASE STUDY
2) New individuals are generated by taking the
optimal individuals and temporal individuals in
the last step as centers. Subsequently, the opti­ 3.1 Basic parameters of bridge, wind and traffic flow
mal groups and temporal groups are identified; This paper takes a four-lane cable-stayed bridge with
3) Apply similartaxis operation in each sub-group to main span of 372.5 m as an example to analyze the
make the subgroup become mature. Similartaxis is fatigue damage reliability of the bottom members in
the term of the competitive process for individual the mid-span, as shown in Figure 5.
who becomes the winner. Then, the score of the
best individual is taken as the score of the sub­
group;
4) When sub-groups are mature, dissimilation oper­
ation is conducted among sub-groups. Dissimila­
tion refers to the searching process of new point
of subgroup that competed for the winner.
Finally, compute the final score of the optimal
individual in the whole search space. Figure 5. Cable-stayed bridge (unit: m).

247
 Three traffic conditions- free flow (9 veh/km/ln), Table 2. Parameters of MEA-BP neutral network.
medium flow (20 veh/km/ln) and busy flow (32 veh/
km/ln)- are selected to represent all the traffic scen­ Parameter Value Note
arios during the bridge lifetime. In general, the min­
istry of transport uses growth factors (gf) to predict Popsize 200 Population size
future traffic and gf can be obtained by the following Supsize 5 Superior group size
expression (Wu 2010): Tempsize 5 Temporary group size
SG 20 Group size
S1 12 Number of neurons in the input layer
S2 10 Number of neurons in the hidden layer
S3 1 Number of neurons in the output layer
Iteration 10 Number of iterations

where N is the total number of years -1; xlast is

the annual average daily traffic of the last year;
xfirst is the annual average daily traffic of the
first year.
the fatigue damage of the bridge from the 25th to the
Eight wind speeds are grouped based on the
75th year of service time, and substitutes these data
Beaufort wind scale: W1 (2 m/s), W2 (5 m/s), W3
into the MEA-BP neutral network for training.
(8 m/s), W4 (10 m/s), W5 (15 m/s), W6 (20 m/s),
In order to choose a better training network, the
W7 (25 m/s), W8 (32 m/s).
network training is carried out several times
The statistical properties of each random vari­
every year to ensure that the error is stable, and then
able are shown in Table 1. Assume that all random
select the network structure with the minimum error
variables obey normal distribution, the standard
as the limit state function of the bridge structure in
deviation of mechanical properties of the bridge is
the corresponding year. Figure 6 shows the error
assumed to be 10% (E, I) or 20% (ξ, damping
results of one year, with a total of 50 training times.
ratio) of their mean values. The mean (μ) and
It shows that the error of BP neutral network opti­
standard deviation (σ) of annual average daily traf­
mized by MEA is obviously smaller than that of un­
fic growth rate (gf) are assumed to be 0.05 and
optimized BP neutral network, and the error of
0.015. The average of the probability PWj of any
MEA-BP neutral network is stable between 0
particular wind speed is obtained from the local
and 0.1.
wind rose diagram, and the standard deviation is
assumed to be 10% of the mean value (Wu 2010).
3.3 Fatigue reliability assessment
3.2 MEA-BP neutral network analysis The PSO-Monte Carlo sampling methods is applied
to calculate the fatigue reliability. Firstly, the PSO
The basic parameters of MEA-BP neutral network is
algorithm is applied and the optimized value is
shown in Table 2. After determining the initial param­
shown in Figure 7. It can be seen from the coordin­
eters of MEA-BP neutral network, this paper studies
ate system on the left of Figure 7, as the number of

Table 1. Statistical properties of random variables.

Mean Standard
Random variable Sign Unit
value deviation

Bridge girder ξ / 0.005 0.001

E GPa 200 20
I m4 3.086 0.309
Traffic growth
gf / 0.05 0.015
factor
Occurrence PW1 (2 m/s) % 60.33 6.033
probability of PW2 (5 m/s) % 30.56 3.056
wind speed PWj PW3 (8 m/s) % 6.41 0.641
PW4 (10 m/
% 1.44 0.144
s)W4
PW5 (15 m/s) % 0.08 0.008
PW6 (20 m/s) % 0 0
PW7 (25 m/s) % 0.26 0.026 Figure 6. Fitting errors comparison between MEA-BP and
PW8 (32 m/s) % 0.92 0.092 BP neutral networks.

248
Figure 7. Optimal individual fitting value.

iterations increases, the individual fitting value are fitted by polynomials, as shown by the solid and
decreases, and the individual fitting value tends to dash lines in Figure 8. The failure probability value
be stable. It can be seen from the coordinate system of fatigue damage shows a quadratic increasing
on the right of Figure 7, all parameters tend to be trend, and the reliability index show a quadratic
the same point, and the coordinate of the point is decreasing trend.
the initial checking point obtained by the algorithm. As shown in Figure 8, the failure probability
Secondly, the checking point is substituted into the value of the bridge during the service period
Monte Carlo sampling method, obtaining the failure under traffic and wind is relatively large. This is
probability and reliability index as shown in Figure probably because the lack of reports on bridge
8. The failure probability increases slowly from the maintenance and repair during the design life, so
20th year to the 35th year, but after the 35th year, the this paper assumes that the bridge has not been
failure probability increases sharply, and the corres­ maintained during its service, which directly
ponding reliability decreases rapidly. The scatter leads to a sharp increase in the failure probabil­
plots of failure probability and reliability indexes ity value.

Figure 8. Failure probability and reliability index of bridge members.

249
4 CONCLUSIONS REFERENCES
This paper introduced a new reliability analytical Faravelli, L. 1989. Response-surface Approach for Reli­
model for bridge fatigue analysis under wind and ability Analysis. Journal of Engineering Mechanics, 115
traffic load. The main feature of the new model is to (12): 2763–2781.
apply the mind evolutionary algorithm (MEA) to Kennedy, J. & Eberhart, R. 1995. Particle Swarm
Optimization. IEEE Int’l Conf. on Neutral Networks,
optimize the BP neutral network for obtaining an
Perth, Australia, 1942–1948.
optimum result for approximate limit state function. Melchers, R. E. 1999. Structural Reliability: Analysis and
By comparing the MEA-BP neutral network and Prediction, John Wiley & Sons, New York, NY.
traditional BP neutral network, it is found that the fit­ Miner, M. A. 1945. Cumulative damage in fatigue. Journal
ting error of MEA-BP neutral network is small and of Applied Mechanics, 12(3): A159–A164.
stable. This study shows the great potential of apply­ Sun, C., Sun, Y. & Wei, L. 1998. Mind-evolution-based
ing MEA-BP neutral network in the reliability ana­ machine learning: framework and the implementation of
lysis for complicated dynamic analysis of bridge. optimization. Proceedings of IEEE International Con­
ference on Intelligent Engineering Systems: 355–359.
Sun, C., Yan, S. & Xie, K. 2000. Mind-evolution-based
machine learning and applications. Proceedings of the
ACKNOWLEDGEMENT 3rd World Congress on Intelligent Control and Automa­
tion (Cat. No. 00EX393), 1: 112–117.
The research is supported by funds provided by the Wu, J. 2010. Reliability lifetime performance analysis of
Fundamental Research Funds for the Central Univer­ long-span bridges. Dissertation, Fort Collins, USA:
Colorado State University.
sities, CHD 300102219220) and China Postdoctoral
Xiao, G. 2002. A Monte Carlo method for obtaining reliabil­
Science Foundation under Project 2014M552395. ity and availability confidence limits of complex main­
Opinions, findings, and conclusions expressed are tenance system. Acta Armamentarii, 23(2): 215–218.
Yang, X. F. 2012. Investigation on response surface
those of the writers, and do not necessarily represent
methods with high efficiency for structural reliability
the views of the sponsors. analysis. Guangxi, China: Guangxi University.

250
 Bridge Maintenance, Safety, Management, Life-Cycle Sustainability and Innovations –
Yokota & Frangopol (eds)
© 2021 Taylor & Francis Group, London, ISBN 978-0-367-23278-8

Automated infrastructure inspection based on digital twins and machine

learning
P. Furtner
VCE Vienna Consulting Engineers ZT GmbH, Vienna, Austria

E. Forstner & A. Karlusch

Palfinger Structural Inspection GmbH, Vienna, Austria

ABSTRACT: One of the key challenges in our modern society is the provision of safe transport infrastruc­
ture. Infrastructure managers are subject to regulations requiring major infrastructures to be periodically
checked for damage before it becomes a safety hazard. In the standard structural inspection, professionally
qualified civil engineers travel to the object to be inspected on site. Notes, sketches and photos are prepared
for the subsequent report. The inspectors are introduced to non-directly accessible locations with special,
heavy inspection equipment. During such a test the object is not or only partially usable, which leads to inter­
ruptions, delays, traffic jam and thus considerable non-availability costs. In recent years, approaches of
a drone-based structural inspection are increasingly noticeable. These are mainly limited to a visual inspection
of the created optical images. By using new technologies, a more objective and faster structural inspection
can be carried out at a lower cost.

1 INTRODUCTION structure, As-Built-BIM-models, structure progress

control, maintenance planning, etc.
Engineering structures (= civil engineering structures)
have to meet requirements with regard to stability, traf­
fic safety and durability. Civil engineering structures 2 USE FOR THE INFRASTRUCTURE
are to be subjected to periodic inspection at as even MANAGER
intervals as possible. In this regard, there are laws and
regulations in all industrialized countries. For example, In the course of developing a concept for a pilot pro­
in Germany these are the DIN 1076 and the RI-EBW- ject, the project participants made fundamental con­
Prüf; and in Austria the RVS series 13.03.XX and the siderations regarding the use of UAVs for the
document “06.01.02 Maintenance//Maintenance plan” inspection process. The expected advantages of this
of ÖBB Infrastruktur AG. technology are summarized below.
This paper presents the concept and implementation
of a pilot project on the use of innovative technologies
2.1 Advantages
in the inspection of structures. Service-based support
on the basis of unmanned aerial vehicles (UAVs,
“drones”) and other platforms as well as digital image 2.1.1 Data collection
processing technologies, in particular artificial intelli­ • Fast, secure, and complete capture of objects
gence methods, should be provided for inspecting civil • Efficient inspection of hard-to-access structures
engineering structures. The aim is to ensure objectified (such as large viaducts)
and comprehensible damage detection while minimiz­ • No/minimal operational restrictions during the
ing cost-intensive processes (road closures, track inspection (at least if there is no need to fly dir­
locks, underbridge inspection units, etc.). ectly above the traffic area)
As a by-product of the application of these tech­ • Operational restrictions, if required, can be sig­
nologies and techniques, a digital twin of the particu­ nificantly shortened by saving time
lar structure is created, which can be also used for • Comprehensive recording and exact localization of
other applications e.g. digitization of the existing structure condition data

DOI: 10.1201/9780429279119-30
https://doi.org/10.1201/9780429279119-30

251
2.1.2 Data evaluation captured information. An additional compression of
location-related structural condition data can be
• Automation-based, objectified and comprehen­ carried out by determined precisely located infor­
sible classification of structure damage mation by means of further chemical-physical sen­
• Location-independent visual assessment of the sors/analyses.
structure condition by experts based on the aerial
survey data 2.2.3 Data visualization and management
• Geo-referencing of the collected data and pos­ The end product accessible for the client via web-
ition-related evaluation options via temporal
GIS – meaning software independent – can be
change of damaged areas/anomalies etc. integrated as a module into existing structural
• Safe and semi-automated detection and quantifi­ management databases by infrastructure operators
cation of condition changes and damage through via links, or used as a stand-alone module. An
periodic structural inspection authorization system ensures the control of the
• Objective and repeatable damage detection information availability per structure and thus the
data confidentiality.
2.1.3 Data visualization and documentation The start page of the web-GIS allows the client
to choose the desired structure, similar to the func­
• Working basis for client, independent test engin­ tionality of Google Maps. By clicking on the struc­
eers and asset accounting ture displayed as an object located on the map, its
• Software independent web-GIS mapping and three-dimensional photorealistic representation is
BIM data delivery opened and can be zoomed and rotated in three
• Structured archiving and retrieval options of dimensions as desired.
existing structure data and structure condition Recorded or automatically determined information
data such as photos and defects can be queried by filter
• Interface for the integration of planning data functions (time filter, layer) or by clicking on the cor­
• Use of the three-dimensional structure representa­ responding position on the 3D object from the under­
tion as a planning basis for necessary renovation lying database and evaluated by the expert/
or maintenance measures commented/categorized (→ tagged). The correspond­
• Easy creation of accurate structure plans from the ing information is also stored in the underlying data­
results base and can be retrieved in a structured manner using
• Unrestricted (or client-definable) group of users the filter functions.
• Data represent the basis for lifecycle management

3 PROJECT GOALS FOR A PILOT PROJECT

2.2 Innovation and further development
A pilot project was carried out in the course of
2.2.1 Data acquisition a joint research project with ÖBB. The main object­
The recording of structures with no or minimal ive of this pilot project is to evaluate the extent to
impact on the current use is carried out by optical which the inspection process at ÖBB can be sup­
and multispectral sensors in an accuracy sufficient ported by the use of UAVs on the basis of a real
for clear assessments and other planning tasks. bridge object, the Falkensteinbrücke. In the course
Obtained information is combined to form a three- of the project the demonstration of the data acquisi­
dimensional digital image of the structure using tion takes place by means of UAV.
photogrammetric methods. The three-dimensional
model forms the basis for further planning tasks.
The information obtained during data collection is 4 DESCRIPTION OF THE TECHNIQUE AND
better – exactly geo-referenced. All information on METHODS USED IN THE PILOT PROJECT
the condition of the structure can thus be clearly and
continuously assigned to concrete positions. 4.1 The drone
A development assessment of the temporal course of
the structure condition is thus possible. A state-of-the-art multicopter is used as the central
carrier platform. The selection of the aircraft (pay­
2.2.2 Data evaluation load) is based on the selected on-board sensors (cam­
Image editing and analysis tools are used to auto­ eras). The flight control of the drone takes place
matically determine structure condition informa­ (semi-)automatically or manually dependent on the
tion. In addition to visually detectable structural situation. Compliance with the trajectory according
damage (e.g. discoloration, cracks, geometric to the flight plan is primarily achieved by using the
anomalies that allow inferences on water ingress, GPS-RTK and distance sensors. Especially in
moss growth, concrete spalling, sulphate rubbing, restricted areas (e.g. bridge underside, between pil­
etc.), physical material condition data can also be lars) a stable flight condition is thus ensured with
evaluated and interpreted by means of multispectral additional pilot monitoring.

252
4.2 The flight planning and aerial survey 4.5 The 3D visualization
In order to create an accurate 3D model based on On the basis of the data from the aerial survey (point
photos, it is necessary to obtain photos containing as data and aerial photos), with the help of structure
much information as possible. That is why an accur­ condition plans and information about the bridge
ate flight planning plays a key role. object from the client, 3D models for the structure
The flight planning of the bridge aerial survey can visualization are created. The visualizations com­
be done in two ways – with a previously carried out prise a) the creation of a photorealistic model from
3D flight planning or with an on-site flight planning the aerial survey data and b) the creation of a 3D
in the field. The choice of the method depends on approximation model of the bridge as a BIM model.
the available data which differ with regard to the The provision of the visualizations takes place
workload. For a 3D flight planning an existing 3D software independent via web-based viewers. The
model is necessary. This allows a semi-automation visualization is retrieved via web-GIS, which serves
of the aerial survey and the identification of the crit­ to locate the project base data. Access to the web-
ical points as well as the take-off and landing places GIS is also software independent via web browser.
in advance in the office. If a 3D model is not yet The web-GIS has an entitlement control that allows
available, proceed as follows: first, a manual coarse sharing data paying attention to different hierarchical
flight is carried out, then the 3D flight planning is levels of an organization. The data can be available
created and finally, the actual measurement flight is to a broad user base by providing them via web ser­
completed. In this case, the flight planning must take vices. The user is able to use all functionalities of
place in the field. Only like that it is still possible to the web-GIS and the 3D viewer without special soft­
identify the critical points, as well as the take-off ware. In addition to the visualization of the test
and landing place before the flight. object, the web-GIS can also be used for structured
data storage. Besides the localization of the informa­
tion (photos of damage, test reports…), information
4.3 The sensors
can also be filed in a structured way and easily
The central on-board sensor of a drone is a high- retrieved by the allocation of tags and with the help
resolution system camera for recording in the RGB of a time filter.
range. The fixed focal length lens with the highest
possible luminous intensity is calibrated before and
4.6 Detection of concrete damage
after use. The RGB camera reproduces everything
that the human eye sees. The heterogeneous (natural) The detection of concrete damage takes place on the
light conditions must be taken into consideration, basis of the onboard sensors images. For this pur­
meaning the use of a photo flash or video light and pose, software for image evaluation and pattern rec­
the corresponding implementation in the context of ognition is used. The individual pixels of a photo are
the development work are aimed for. Direct sun and divided into different classes due to their absorption
therefore high-contrast shadows can be avoided by and reflection properties. Areas with known spectral
planning the flight during a day with slight cloud properties are marked in advance and then used as
cover. a reference. Using machine-learning algorithms, this
process can be automated as the number of pro­
cessed recordings increases.
4.4 The image processing and evaluation software
The detected damages include e.g. discoloration,
A 3D model of the bridge is created based on the spalling, exposed reinforcing steel, moss, cracks,
recorded images. For this purpose, software for deformations, rust marks, etc. In addition to the
robust photogrammetric evaluation is used. Using detection, the exact location of the respective
algorithms of the bundle adjustment, a precise damage is determined. Cracks in the concrete are
mutual orientation of the images takes place. This classified according to their length and width. The
process is supported by the use of pre-geodetically selection of the camera is influenced by the smallest
measured control points. In addition, the correct spa­ crack width to be detected.
tial positioning of the object in a superordinate refer­
ence system is effected by these control points. The
4.7 Concrete moisture determination
output of the photogrammetric evaluation are 3D­
data for the visualization as well as exactly located Concrete moisture is detectable in the near infrared
and corrected images as basis for the detection of range at a spectral wavelength of about 970 nm. Cur­
concrete damage. rent own research results with the involvement of
The detection of structure damage in the indi­ Graz University of Technology promise more mean­
vidual images takes place by using methods of ingful values at approx. 1420 nm. Further develop­
artificial intelligence. For this purpose, fully auto­ ment work concerning camera, illumination and
mated software based on a so-called neural net­ evaluation is planned: first in the laboratory, then in
work is used. natural test environment and subsequently on real

253
test objects based on a metrological investigation by triggered manually by the pilot. The recording posi­
a drone. tions were chosen in such a way that all surfaces are
The determination of concrete moisture with ther­ detected as frontally and diagonally as possible. The
mal cameras is singular not promising, but as an add­ overlap of the pictures is about 80%.
itional sensor, the use of such systems is conceivable. For the subsequent geo-referencing of the 3D
model, approximately 40 control points on the
bridge were geodetically measured in the M31 Land
4.8 Chloride content determination
Survey System.
Chloride as a cause of damage on concrete structures
can be determined on the surface by means of
remote sensing methods. First experiments with 6 DATA EVALUATION AND VISUALIZATION
Graz University of Technology show a spectral
imaging of chloride in the range of approx. 1420 6.1 3D modeling
nm – similar to the spectral range of the moisture
content. An imaging and measurement by means of To create a photorealistic 3D model of the aerial
multispectral sensors is thus possible. photographs, a photogrammetric method – Dense
Image Matching – is used. The operating principle of
the algorithm searches for equal points on several (at
5 THE PILOT PROJECT least two) images and calculates their coordinates.
FALKENSTEINBRÜCKE The result of this process is a point cloud that can
later serve as a basis for a BIM or 3D mesh model.
The Falkensteinbrücke near Obervellach, with Of the 3599 photos taken, 2919 were selected for
a length of 396 m, is the longest bridge in the ÖBB modelling and ultimately used. Blurred, too dark or
Tauern Railway and spans an eastern side valley of repetitive photos were not used. The first step in
the Möll Valley via second arches. modelling is the orientation of the photos. To speed
For the aerial survey of the Falkensteinbrücke, the up this process, the GNSS coordinates of the photos
northern half of the southern arch was selected; espe­ of the drones were used. The results of this step pro­
cially the support plate, pillars and bow have been vide a “Sparse Point Cloud”, improved coordinates
inspected. This corresponds to about a quarter of the and orientation of the photos by calibrating the
entire bridge. camera.
Within the project area, the bridge deck, the arch In the last step, a dense point cloud was created.
and the pillars 10-13 were in focus (see Figure 1). With a total of 2919 photos, this process is a major
The flight took place on 27th of November 2017 hardware challenge.
in good weather with a hexacopter by the company
Skyability GmbH. The control of the drone was 6.2 Image analysis
done manually by the pilot from the ground, assisted
by an observer, as well as by a FPV data link. The As already mentioned, the image analysis and the
distance between bridge and drone averaged 3 to damage detection on taken pictures are done using
5 m. The sensors could optionally be mounted on the artificial intelligence methods. In the pilot project we
lower or upper side of the drone to pivot and tilt. used neural networks for this purpose. The further
The sensor used was a Sony Alpha 7RII with a 42 developed software tools are FALCO and IRIS,
MP full frame. The camera was mounted partly whereby FALCO is used for the training of the
above, partly below the drone. The camera was neural network and IRIS for the analysis of a large
number of images with the help of a classifier cre­
ated in FALCO.
All images are examined for different types of
damage, such as cracks, spalling, moisture penetra­
tion and chloride content. The detected damage is
marked on the images (by means of polyline or poly­
gon in a case of a surface damage) and then written
to a database. The damage type and the image coord­
inates of the tag are now saved.
In the next step, the damage entries in the data­
base are geo-referenced with special software.

6.3 Visualization
The so-called web-GIS software was developed for
the visualization of the evaluation results. This
allows the representation of the 3D model with all
Figure 1. Pillars in the recording area. detected damage to the model (see Figure 2).

254
 using internet access, a web browser and login data
only.

6.4 Summary
In this contribution the potential of new technologies
in the field of inspection and condition testing of civil
engineering structures is being demonstrated by
means of a pilot project. Especially with large objects,
such as the mentioned Falkensteinbrücke, where
a complete, close-at-hand examination with conven­
tional access technology is not possible or only at
very high costs, these technologies can already pro­
Figure 2. Web-GIS visualization. vide valuable support. At the same time, the unavail­
able costs due to traffic restrictions can be minimized.
An additional added value can result in the future
The representation of the damage in the web-GIS
through the use of additional sensors (multi-spectral
takes place dynamically – all data is loaded into
cameras, laser scanners etc.) on the carrier platform.
a 3D viewer and a table directly from the database,
Initial tests show potential for surface moisture
and the changes made by the evaluation team are
measurement or determination of chemical surface
also saved there. Every defect that has a defined ID
condition (e.g. chloride content).
is also marked with it.
In the web-GIS there are two views with the
data representation – a main view and a detail REFERENCES
view. The main view consists of a 3D viewer and
a damage list, in which all damages are shown and Morgenthal, G., Hallermann N. & Achterlik M. 2015.
described. For the assessment, a detailed view was Unbemannte Fluggeräte zur Zustandsermittlung von
created in which each individual defect is shown Bauwerken, In Forschungsinitiative Zukunft Bau,
with the associated photos and the description. F 2971, Fraunhofer IRB Publishing House; ISBN 978­
The results presented in the web-GIS (visualiza­ 3-8167-9589-6.
tion, tabular list of damages and classification TÜV Rheinland on behalf of the Federal Highway Research
Institute in Germany. 2014. Research paper FE-No.
details) will be checked, interpreted and evaluated
15.0602/2014/LRB: Unterstützung der Bauwerksprüfung
in the next step by an experienced civil engineer. durch innovativ digitale Bildauswertung – Pilotstudie.
The software also has a report generator which VCE & Angst-Group (UTC) for ÖBB Infrastruktur AG.
supports the editor in the report generation. 2016. Concept for an ÖBB pilot project: Unterstützung
The functionality of the web-GIS software allows der Bauwerksprüfung mit UAV - Pilotprojekt Falken­
the evaluation team the access to visualize and process stein, Leistungsbeschreibung.

255
 Bridge Maintenance, Safety, Management, Life-Cycle Sustainability and Innovations –
Yokota & Frangopol (eds)
© 2021 Taylor & Francis Group, London, ISBN 978-0-367-23278-8

Development of the bridge inspection experience system with MR

head-mounted display
Y. Baba, H. Emoto & S. Tanikawa
Fukushima KOSEN (National Institute of Technology, Fukushima College), Fukushima, Japan

H. Nakamura & K. Kawamura

Yamaguchi University, Yamaguchi, Japan

ABSTRACT: In Japan, bridges were constructed in a period of high economic growth, and almost all bridges
have recently exceeded 50 years old. These many bridges exhibit increasing damage conditions, such as cracks or
spalling due to aging. With this, the amount of inspection work has increased. In this study, for the education for
young engineers performing inspection work, we developed a bridge inspection experience system with a MR
head-mounted display that can simulate inspection work in a VR space and utilize the information for future main­
tenance management plans. After developing this system, we surveyed students and faculty by questionnaire to
gain objective opinions and find points of improvement. In this paper, we describe how we developed this system,
how to use it, and, moreover, its usefulness, as clarified by the questionnaire.

1 INTRODUCTION With this background, seminars are currently held

everywhere to train inspectors in bridge inspection.
Much of civil infrastructure, like bridge infrastructure, In most seminars, onsite training is held in the field,
was constructed in the period of high economic and part of a collapsed bridge is used as an educa­
growth, and most bridges have now exceeded 50 years tional material (Tamada, K. 2016, Tamada, K.,
in age. Thus, older bridges accrue damage conditions Tamura, T. and Shimada, T. 2016). It is necessary for
based on age. Because of this situation, in 2014 the young inspectors to learn to inspect bridges using
Ministry of Land, Infrastructure, Transport and Tour­ real damaged structure and to collect data on old
ism revised part of the Road Law Enforcement Rules structures as training. However, damaged bridges
to establish specific standards for road maintenance should be repaired and strengthened for long-term
and repair based on the provisions of Article 35-2, use, and it is impossible to eternally retain damaged
Paragraph 2 of the Road Law Enforcement Order. The bridges. Keeping collapsed structures for training
road administrator is obliged to inspect to bridges and requires a vast area and lots of money. Furthermore,
tunnels closely and visually once every five years safety and lane regulation is required for this type of
(MLIT 2014). In addition, to properly evaluate the on-site collapsed bridge inspection training.
structural health of each part of a bridge, it is necessary Recently Mixed Reality (MR) (Tachi, S., Satou,
for the inspection engineer to know the structure and M. and Hirose, M. 2011, Milgram, P. and Kishino,
the features of the materials. In other words, legislation F. 1994) has become popular to resolve these issues.
regarding the maintenance and renewal of social infra­ MR is mixed reality, the general technology for mer­
structure is progressing. On the other hand, the local ging the real and virtual worlds. By using this MR
government for road administrators do not have technology, we can make a 3D-VR model of a real
enough inspectors to staff experts on bridge mainten­ bridge and import it into a VR world, where we pro­
ance and renovation. Local governments have pose the bridge inspection work will take place. In
a responsibility for maintenance and renovation as road this way, it becomes possible to use the expert know­
administrators, even with staff restrictions. Thus, gov­ ledge for bridge inspection work safely and easily,
ernments need to quickly educate inspectors with pro­ without considering weather and time. Furthermore, it
fessional knowledge for bridge inspection work and is easy to express the VR-Model on MR, instead of
improve inspection work efficiency (Tanaka, T., Mat­ using an old bridge that may break down and need
suda, H. and etc. 2017)2). replacing. In this study, we develop a bridge

DOI: 10.1201/9780429279119-31
https://doi.org/10.1201/9780429279119-31

256
inspection support system using MR technique for described above, it is possible to confirm the damage
concrete bridges. conditions at the time of inspection at any time using
the database. As a feature, it is possible to share
knowledge of the bridge’s damage conditions and
2 HOW TO USE THE MR TECHNIQUE IN compare with past conditions.
BRIDGE INSPECTION Finally, we describe the training support offered for
inspection techniques. First, the primary system users
Here, we describe how to use the MR technique for are young engineers who do not learn bridge inspection
bridge inspection. First, MR is mixed reality, the gen­ technology. In general, young engineers learn about
eral technology for merging the real virtual worlds. As bridge damage conditions through workshops based on
shown in Figure 1 (Milgram, P. and Kishino, F. 1994), bridge engineering, structural mechanics and material
when the virtual experience is stronger than reality it is engineering. In addition, they inspect and diagnose
called VR: Virtual Reality; and when reality is stronger bridge conditions under another professional, and they
than the virtual experience, it is called AR: Augmented improve the inspection quality by gaining personal
Reality. This system uses a VR-Model based on real experience. However, due to the rapid increase in
bridge and hammering test sounds in a VR space. inspection work because of aging infrastructure and the
Therefore, in this research, MR is expressed as a real decrease in the maintenance budget because of the
space on a virtual space. reduction of public works expenses, the education of
Figure 2 Illustrates the MR technique viewpoint for young engineers is not sufficiently conducted. In add­
bridge inspection, which includes an inspection data­ ition to the budget, on-site education cannot be easily
base, time series database, supporting discussion of implemented due to the limitations of weather, time,
inspection results, and bridge inspection training usage permission, traffic blockage, and safety. There­
support. fore, education support for bridge inspections can be
For use as an inspection database, a 3D-VR model provided without going to the site by using MR tech­
with bridge damage conditions is created as a texture, nique to gain inspection experience in the virtual space.
and the results of visual inspection are also used. After Furthermore, the cost of inspection training is reduced,
this, the condition of the bridge at the time of inspec­ without restrictions.
tion is confirmed at any time by referencing the detail
in the database. In addition, the technique is useful for
evaluating the progress of deterioration based on 3 OUTLINE OF THE SYSTEM
a time-series database created by storing 3D-VR model
data, along with the results of visual inspection. One of the objectives of this system is training
Next, we discuss the inspection result support a bridge inspector indoors, instead of at the bridge site.
function. The 3D-VR model does not change once To install the simulated experience indoors, we apply
made and has a high reproducibility. Therefore, as Virtual Reality (VR) using the MR-HMD (Mixed
Reality – Head-Mounted Display), as shown in Figure
3. With this, it is possible experience a bridge inspec­
tion in immersive virtual space. With MR-HMD, both
eyes are covered to display stereoscopic images. The
left and right MR-HMD are slightly angled, making it
possible to experience an immersive video experience
using perspective. In addition, an installed motion
sensor on the MR-HMD provides a mobile viewpoint
that moves based on movement of the user’s head. In
this way, it is possible to enter the VR space.
Figure 1. Outline of MR.

Figure 2. Point of utility of MR technique in bridge

inspections. Figure 3. Example of a MR-HMD and motion controller.

257
Furthermore, the system is applied to a 3D-VR model
based on a real bridge with a hammering sound, based
on the sounds made by real bridges. Figure 4 explains
the system flow.

3.1 System configuration and the development

environment
We explain how to “Set up a development environ­
ment (Unity)”, as shown in Figure 4. This system con­
sists of the “VR data”, “Hardware” and “Software”,
shown in Figure 5. “VR-data” is included in the 3D­
VR model of real bridges, the 3D-VR model for
hammer sounds, and the sound data of a hammering
test. The “Hardware” requires the minimum PC speci­
fication shown in Figure 5 to enable “Windows Mixed
Reality” made by Microsoft Co. Ltd. The “MR­
HMD” is the “Dell Visor with Controllers vrp100”
made by Dell. Finally, the “Software” that is the core
of the development environment uses the gaming Figure 5. System components.
engine “Unity” made by Unity Co. Ltd. The 3D- VR
model software is Metasequoia made by Tetraface Inc.
be adapted for VR content made by Unity. For this,
the “Virtual Reality Supported” function on Unity is
3.2 The MR development environment on Unity enabled, and a library file of “Virtual Reality SDKs”
Here, we describe the development environment, on Unity is setup. With these settings, Unity will
shown in no. 2 of Figure 4. First, Unity is required to support VR and automatically recognize the headset.
Since this system is developed as content for
a “SteamVR”, “OpenVR” is selected as the SDK.

3.3 Discussion of the versatility of a MR Head-

mounted display platform
Here, we describe the MR Head-mounted display
shown in no. 3 in Figure 4. A MR-HMD enables the
use of MR content. MR-HMDs include various types
of headsets, such as the “Dell Visor with Controllers
vrp100” used in this system; the HTC VIVE, devel­
oped by HTC Corporation; and the Oculus, developed
by Facebook Technologies, LLC. At the same time,
the shape of the controller and the type and number
of buttons differ for each headset, so when develop­
ing MR content, it is necessary to create a system
that is compatible with each headset and controller.
Therefore, we are developing this system using
“Steam VR” content to create a common control
headset and controller. This “Steam VR” is made by
the “Valve Corporation”, which is a game distribution
platform developer. It is a distribution platform for
computer games compatible with Windows MR.
Because of this, it can be used with any hardware. In
this study, “Steam VR” starts under Windows MR, as
shown in Figure 6, and the system operates within it.
VR content development is installed as a “Steam
VR” plugin to implement headset and controller use.
The “Steam VR” plugin is an “Asset” in Unity and
is made by “Valve Corporation”. This plugin includes
various libraries to enable easy operation, such as the
controller, loading a 3D-Model and tracking automat­
ically in the VR space. Furthermore, the “Steam VR”
Figure 4. Flowchart of system development. plugin handles controller input in an abstract unit

258
Figure 6. Platform components.

called Action. The “Steam VR” plugin can control

all controllers to set up each button before defining
an “Action” such as “holding something” or “open­
ing the menu”, as shown in Figure 7.

3.4 Making and installing a 3D-VR model

Here, we describe the 3D-VR model and its installa­
tion, as shown in no. 4 in Figure 4. A 3D-VR model
in our system is made by the flowchart process shown
in Figure 8 and using Metasequoia (Tetraface Inc.
2019). First, according “Procedure A” in Figure 8, the
system takes a photo of each bridge material, and
photos are adjusted by size, shape, brightness using
a photo editor. Then, each edited photo is composited
for each material. These composite photos become the
texture of the 3D-VR model. Next, according to “Pro­
cedure B” in Figure 8, the system creates the frame­
work for the 3D-VR model by Metasequoia, and this
framework is set to a texture. Finally, using the above
procedure, the 3D-VR model is completed. Using this
procedure, the authors created a 3D-VR model of
a real bridge in Yamaguchi Prefecture (Emoto, H.,
Miyamaoto, A. and Takahashi, J. 2016). Figure 8. Procedure of a making 3D-VR model.

Figure 7. Method of setting each action.

259
 After making the 3D-VR model, to keep the rela­ HMD as shown in Figure 9 and puts the laptop PC
tionship between the polygon data and the bridge tex­ that operates the bridge inspection experience
ture image data of in the Unity system, a 3D-VR system into a backpack. In this state, the user per­
model by Metasequoia is imported. The polygon data forms the bridge inspection experience. At this time,
is the shape of an object. Therefore, a 3D-VR model is since the user cannot confirm the external state,
serialized by the extension type of FBX (.fbx) (Auto­ space is necessary for safety.
desk Inc. 2019). This file format can smoothly
exchange 3D data developed by “Kaydara filmbox”.
4.2 Reproduction of the visual inspection and
Currently, the FBX format is licensed by Autodesk.
hammering test
Keeping the relationship of the polygon data and image
data, the Unity system can import a 3D-VR model. We created a crack gauge using Metasequoia to
measure the crack width of concrete and introduced
this tool into Unity in a translucent state. In addition,
4 SYSTEM FUNCTIONS we created the hammer used during the hammering
test and the choke used for choking the peeling part
4.1 The movement method when a user experiences of concrete in Unity.
inspection work The experience of the hammering test is like the
actual hammering test. A user shakes the motion-
The experience of inspection work targeted by this controller as a hammer and hits the bridge to mark the
system is the visual inspection and hammering test damaged area using a hitting sound. On the system, we
of a concrete bridge. A user approaches the 3D-VR set the entire structure to a normal hammering sound,
model by moving themselves in the VR space and and we place some flat plates with abnormal hammer­
experiences the visual inspection and hammering ing sounds, having 0.2 mm in thickness, 0.3 mm away
test. We implemented a movement method using from the structure surface. The hammering sounds in
absolute coordinates of the VR space and relative this system use hammering sounds from the normal
coordinates of the HMD in this system. and abnormal portions recorded at Amanogawa Bridge
First, the movement using absolute coordinates in Iwaki City, Fukushima Prefecture. In addition, we
assumes movement by a bridge inspection car. For made it possible to visualize the abnormal sound areas
this reason, the movement is based on the xyz axis arbitrarily and to make a comparison between the test
of the VR space, and the movement direction result after the hammering test and the correct answer.
doesn’t change from the preset direction, even if
a user moves the viewpoint.
Second, the movement using relative coordinate 4.3 View of the 3D-VR model
assumes that the direction of the scene towards which The 3D-VR model in this system is made from real
a user wants to move is in front of the user. For this bridge photos. Thus, the 3D-VR model’s texture has
reason, the xy axis is based on the direction of the a real damage appearance. Therefore, a user can check
HMD. Therefore, the front of a user is always the for­ for damage to the bridge surface, such as cracking and
ward direction of movement, and they can move freely peeling of concrete, rebar exposure and rebar corrosion.
back and forth, left and right. However, movement in Consequently, a user can confirm what kind of damage
the height direction is accomplished using a movement is likely to appear on a bridge before the inspection
method based on absolute coordinates to avoid confu­ work, and the user is expected to be able to perform the
sion for users. inspection work quickly. Figure 10 shows the 3D-VR
Moreover, moving in VR space by moving in real model viewed through MR-HMD.
space is also possible. In this case, a user wears an Conventionally, the created 3D model has been
browsed using 3D model viewer software on a PC
screen (Nishida, K., Emoto, H., Takahashi, J. and

Figure 9. Experiencing this system in a backpack. Figure 10. View of 3D-VR model in this system.

260
Miyamoto, A. 2013). However, most of the software
need operations to rotate and enlarge/reduce the bridge
model to browse any part of 3D model. Those needs
require user skill. By contrast, in this system, the view­
point moves according to the direction of the user’s
face, and the position of the user on the MR changes as
the real user moves. Hence, a user can easily operate
the system using their own sense in space to look the
3D-VR models as if they were at a bridge site. In add­
ition, it is easier to grasp the position of the bridge,
since the 3D model is not rotated. Therefore, depending
on the resolution of the actual bridge image used to
create the 3D model, even small cracks can likely be
confirmed on MR.

4.4 Check the crack width using a crack gauge

In this system, a user can measure the crack width of
concrete using a crack gauge on MR, as shown in Figure 12. View of the crack gauge in this system.
Figure 11. The user can measure an actual crack
width by creating a 3D-VR model to scale, because
the crack gauge was created to actual scale. How­
ever, at this stage 0.2mm cracks (MLIT 2019), those abnormal parts, as shown in Figure 13. Therefore, this
regarded as small cracks in reinforced concrete method is expected to prove highly effective in sup­
structures, are not yet resolved, as the 3D-VR model pressing variation in inspection results.
in this system was not created for close visual
inspection and has low resolution. Nevertheless, the 4.6 Reduce outdoor work
crack gauge in this system is slightly lower in reso­
lution than the original image, but it can read cracks When carrying out on-site training for inspection work
up to 0.2mm, as shown in Figure 12. on actual bridges, it is necessary to obtain permission
from the road administrator. In addition, it is necessary
to prepare scaffolding, inspection tools, safety manage­
4.5 Hammering test on the 3D-model ment, and have good weather conditions. In contrast,
The hammering test, which is often performed at this system can ignore these preparations and condi­
a bridge inspection, is a diagnostic method based on tions because it was developed for indoor use.
the engineer’s senses. For this reason, there is no
standard for judging abnormal locations for the ham­ 4.7 Storage, reproducibility, and follow-up of past
mering test, and variations in test results occur easily. damage condition data
However, the variations in test results can be sup­
pressed by experiencing the hammering test with Some facilities have preserved damaged parts of old
a same sample in advance. Therefore, this system per­ structures as teaching materials and specimens (Han-
forms a hammering test using the same sample in VR shin Expressway Company Ltd. 2019). However, pre­
space. In addition, a user can compare the inspection serving damaged parts of old structures is not easy
results and the correct answer with the hammering because it requires a large amount land and carries
test in this system, because the user can mark the
abnormal part of the bridge and visualize preset

Figure 13. Method of chalking, hammering test and view­

Figure 11. Measuring crack width on MR. ing the damage condition point in this system.

261
high maintenance costs. Moreover, preserving an 5.2 Results and discussion
entire damaged structure is even more difficult. In
Figure 14 Shows the summary results of the question­
contrast, damage data can be saved and reproduced at
naire. The average by question is (a) 4.6 points, (b) 4.8
any time without land and maintenance costs by creat­
points, (c) 4.7 points, (d) 3.4 points, and (e) 4.8 points.
ing 3D-VR data of structures to be dismantled due to
First, the immersive experience of the MR is con­
aging. In addition, it is possible to observe the pro­
firmed based on question (a). Therefore, the system
gress of damage due to aging of existing structures by
can reproduce the experience of a bridge inspection.
creating 3D-VR models at regular intervals.
Next, evaluation of question (b) is high, which
confirms that virtual space is better recognized with
5 VERIFICATION OF RESULTS AND HMD than a 3D model viewed on a PC display. It
DISCUSSION BY QUESTIONNAIRE can be used to locate bridge material easily.
Next, the opinion of question (c) is divided, with
For verification and improvement of the system func­ an average value lower than other questions. Among
tions, we provided a user questionnaire to four teach­ others, points of improvement provided by the free
ing staff in our Fukushima KOSEN and 14 young description include that the system is “difficult to
engineering students. In this session, we describe the operate when moving.” and “the resolution of the
questionnaire’s content and evaluate its results. 3D-VR model needs to increase”.
Regarding movement in the MR, the absolute
coordinates used at the time of the experience
5.1 Outline of the questionnaire assumes the experience utilizes a bridge inspection
Here, we explain the questionnaire flowchart. Before car. In this case, the direction of movement by the
taking the questionnaire, the tester must experience controller with respect to the bridge is predeter­
using this system. The contents of the questionnaire mined. So, it is impossible to recognize the direction
are below: of movement when the direction of the tester moves.
At this point, it will adopt relative coordinates from
(a) Do you feel an immersive reality? the viewpoint of the experiencer. With this, the oper­
(b) Compared with the conventional 3D model ation is made easier if the front of the user is always
viewing methods (viewing on a PC display), is in the forward direction. Next, because the system is
this system easier to understand the positional set to achieve the highest image quality for reso­
relationship of each part of the bridge? lution of the 3D-VR model, there is a problem with
(c) At this time, how is this system as a substitution the method of creating the 3D-VR model.
for a visual inspection of a real bridge? Question (d) scored higher points than question
(d) In the future, advanced UAV surveying and (c), indicating that when the work of making the 3D­
image processing is expected. Then 3D-VR VR model is reduced, this system is expected to be
models can efficiently make high quality used for bridge inspection work.
images. How would this system substitute for Finally, question (e) also scored high points. This
a visual inspection of a real bridge? confirms its usefulness in learning bridge inspection
(e) How useful is the hammering test experience in work based on the hammering test experience. How­
this system? ever, this system is difficult to operate. Many com­
Each question has 5 levels. Level 1 is bad and ments such as, “it is difficult to understand the
level 5 is good. Other questions allow for free operation method”, and “it takes time to get used to the
description/impressions of the system and sugges­ operation” were provided in the free description. There­
tions for improvement. fore, operation improvement and a manual are needed.

Figure 14. Method of setting each action.

262
6 SUMMARY in Kyushu Yamaguchi area, Journal of JSCE, Con­
struction and Management F4, Vol.73, No.4,
The results of this study are summarized below: pp.112–119, 2017.
Tamada, K. 2016. Tihousousei to gaku -tiikiwo haguku­
I. We developed a bridge inspection experience mudobokuno [ti]to[hito] – tiikino infurawo mamoru
system using MR-HMD and proposed a method dobokugizyutusyawoikuseiseyo maidurukousen syakai­
for using it as an educational material for learn­ kibanmentenancekyouikusenta-notyousen, Civil Engin­
ing bridge inspections in future maintenance eering, Vol.101, No.12, pp.26–27.
management plans. Tamada, K., Tamura, T. and Shimada, T. 2016. Active
II. The system is expected to be useful for support­ learning wo kizikutoshita kyouryou mentenance gizyut­
ing the education of young engineers. Further­ susya ikusei karikyuramunokaihatsu, JSCE 2016
Annual Meeting, CS1-006, pp.11–12.
more, it makes experiencing bridge inspection
Tachi, S., Satou, M. and Hirose, M. 2011. ba­
work safely indoors possible. tyaruriariteligaku, The Virtual Reality Society of
III. As a result of using the 3D-VR model created Japan.
from real bridge data on MR, a higher immersive Milgram, P. and Kishino, F. 1994. A Taxonomy of Mixed
feeling is produced. Reality Visual Displays, IEICE TRANS. INF. & SYST.,
IV. Damage conditions of “crack” and “spalling” are Vol. E77-D, No.12, pp.1321–1329.
taken from real-world photos. A “hammering Tetraface Inc. 2019, Metasequoia4, <http://www.metaseq.
test sound” is also recorded in the real world. net/jp/>, (accessed:2019.10.8).
Therefore, the system is effective for learning Emoto, H., Miyamaoto, A. and Takahashi, J. 2016. Devel­
opment of VR-based inspection support system for
bridge inspection work.
crack and spalling detection in deteriorating bridges,
V. The survey questionnaire demonstrates this Memoirs of the faculty of engineering, Yamaguchi uni­
system’s effectiveness and clearly defines its versity, Vol.66, No.2, pp.21–34.
problems. Autodesk Inc. 2019. FBX Adaptable file format for 3D ani­
mation software, <https://www.autodesk.com/products/
In the future, we plan to improve system operabil­ fbx/overview>, (accessed 2019.10.8).
ity, simplify the 3D-VR model creation method, Nishida, K., Emoto, H., Takahashi, J. and Miyamoto, A.
improve the model resolution, and provide feedback 2013. Added detection of cracks and spalling in bridge
for engineers who conduct daily inspection work. visual inspection support system using VR technology,
Proceedings of the symposium on civil engineering
informatics, Vol.38, pp.139–132.
REFERENCES Ministry of Land, Infrastructure, Transport and Tour­
ism (MLIT) 2019. Kyouryouteikitenkenyouryou
Ministry of Land, Infrastructure, Transport and Tourism furoku - 1 sonsyouhyoukakizyun, pp.10, <http://
(MLIT) 2014. About regulations of ministerial ordin­ www.mlit.go.jp/road/ir/ir-council/pdf/yobo3_1_6.
ances and notifications regarding maintenance and pdf>, (accessed 2019.1.15).
repairs. <http://www.mlit.go.jp/common/001042579. Hanshin Expressway Company Ltd. 2019. Shinsai shiryou
pdf>, (accessed 2019.04.14). hokanko – hanshin awaji daishinsai deno hisaikeiken
Tanaka, T., Matsuda, H., Makizumi T. and wo kousei ni kataritsugutameni -, <https://www.han
Takahashi, K. 2017. Survey on infrastructure main­ shin-exp.co.jp/company/torikumi/kouken/shinsaishiryou
tenance and management work of local governments kan.html>, (accessed:2019.10.8).

263
 Bridge Maintenance, Safety, Management, Life-Cycle Sustainability and Innovations –
Yokota & Frangopol (eds)
© 2021 Taylor & Francis Group, London, ISBN 978-0-367-23278-8

Bridge damage cropping-and-stitching segmentation using fully

convolutional network based on images from UAVs
Jiyuan Shi, Ji Dang & Rongzhi Zuo
Department of Civil and Environmental Engineering, Saitama University, Saitama, Japan

ABSTRACT: This research aims to provide a cropping-and-stitching segmentation-based method to find cor­
rosions on steel bridge from a great number of images. In this study, a data set is generated by manually label­
ing damages pixel by pixel, on the images of steel with different levels of corrosion. The data set is used to
train a Fully Convolutional Network (FCN) model for detecting location of corrosion on images. Then a new
data set with cropped images is built by cropping the images from previous data set and then used to train
a new FCN model. By the new model, the location of damage can be shown in a stitched image to visualize
the damage and its distribution. Two FCN models are compared with each other to find the one with better
performance. Besides, one 4k image taken by UAV is tested by the FCN model with better performance.

1 INTRODUCTION with higher accuracy to detect the damages from the

images or videos taken by UAVs. Deep Learning-
During the period from 1954 to 1973, economy in based image classification have been conducted to
Japan grew rapidly, with many infrastructures such as detect cracks (Cha, Choi, & Byüköztürk 2017). How­
roads and bridges being constructed. However, many ever, most of methods were proposed for laboratory
such infrastructures are over 50 years (Ministry, of photographs with fix camera to object distance and
Land, Infrastructure, Transport, & Tourism. 2016) and angle to recognize the damage correctly. And image
nowadays are facing many problems such as ageing segmentation method is also successful in detecting
deterioration. Thus, many structures are expected to be object from aerospace images of several cities (Iglovi­
inspected in an effective method in the next few kov & Shvets 2018) to identify building seismic
decades. damage. It could detect structural damage robustly
The traditional method, manual inspection, is from real world images in more universal sources.
expensive in human cost, rental cost of special However, the traditional segmentation method,
inspection vehicles, and transportation constrain time. uses the data set composed of compressed full images
New methods such as UAV-based bridge inspection with small size. Since the images from UAVs provide
have been studied to fetch high quality images and high resolution images and the process of image com­
videos quickly. Besides, by these methods, deterior­ pression will lose part of feature information from the
ation detection is available for some structure members original images, the traditional segmentation method
where are difficult to access by traditional manual is not suitable for UAV images detection.
inspection. These methods will obviously fasten the In this study, a new damage segmentation method
process to diagnose the status of the infrastructure. based on Fully Convolutional Network (FCN) is pro­
And to evaluate the status of inspected structures, it is posed. Large amounts of images are collected and
necessary to mark the damages on collected images. manually marked to generate the training data to sat­
However, this operation also cost large amount of isfy the need of training an FCN model. Then previ­
human processing hours. ous and new methods are used to prepare two
To address this issue, a lot of methods based on different data set for training. The two FCN model
image and video processing to detect damage on struc­ trained from different data sets are compared to get
tures have been proposed. With the application of mul­ a better one. And finally, the behavior of the better
tiple layer structured Convolutional Neural networks model is improved and then used to test its capability
(Deep Learning), it is possible to develop approaches of corrosion detection on the images taken by UAVs.

DOI: 10.1201/9780429279119-32
https://doi.org/10.1201/9780429279119-32

264
2 STRUCTURE OF FCN layers connected with the rectified linear unit
(ReLU) layer and maxpooling layer, while up-
In general, to detect the location of the objects from sampling layer is consisted of upsampling layer and
images, the FCN architecture consists of two parts convolutional layer, which is followed by an auxil­
(Figure 1). One part using 5 down-sampling blocks iary layer named batch normalization (BN) layer.
is based on VGG16 structure (Simonyan & Zisser­
man 2014) to capture context. The other part based
on U-Net (Jonathan, Evan, & Trevor 2017) is con­
sists of 5 up-sampling blocks and using a symmetric
expanding path to make precise localization enable.
This section explains both of two blocks and intro­
duces the details or backgrounds of the layers used
in the study.
Table 1 presents the detail of each layer. The
input size of the images is 224�224 pixels with
three channels (Red, Green and Blue). Down-
sampling is combined from convolutional (Conv) Figure 1. Overall architecture.

Table 1. The detailed specifications of FCN.

Layer Type Pad Kernel Size Stride Output Size Note

1 Input - - - 224�224�3 __________

2 Conv+ReLU 1 3�3�64 1 224�224�64 Down-sampling
3 Conv+ReLU 1 3�3�64 1 224�224�64 block 1
4 MaxPool 0 2�2 2 112�112�64 __________
5 Conv+ReLU 1 3�3�128 1 112�112�128 Down-sampling
6 Conv+ReLU 1 3�3�128 1 112�112�128 block 2
7 MaxPool 0 2�2 2 56�56�128 __________
8 Conv+ReLU 1 3�3�256 1 56�56�256
9 Conv+ReLU 1 3�3�256 1 56�56�256 Down-sampling
10 Conv+ReLU 1 3�3�256 1 56�56�256 block 3
11 MaxPool 0 2�2 2 28�28�256 __________
12 Conv+ReLU 1 3�3�512 1 28�28�512
13 Conv+ReLU 1 3�3�512 1 28�28�512 Down-sampling
14 Conv+ReLU 1 3�3�512 1 28�28�512 block 4
15 MaxPool 0 2�2 2 14�14�512 __________
16 Conv+ReLU 1 3�3�512 1 14�14�512
17 Conv+ReLU 1 3�3�512 1 14�14�512 Down-sampling
18 Conv+ReLU 1 3�3�512 1 14�14�512 block 5
19 MaxPool 0 2�2 2 7�7�512 __________
20 Upsampling - - - 14�14�512
21 Concatenate - - - 14�14�1024 Up-sampling
22 Conv+BN 1 3�3�512 2 14�14�512 block 1
23 Conv+BN 1 3�3�256 2 14�14�256 __________
24 Upsampling - - - 28�28�256
25 Concatenate - - - 28�28�512 Up-sampling
26 Conv+BN 1 3�3�256 2 28�28�256 block 2
27 Conv+BN 1 3�3�128 2 28�28�128 __________
28 Upsampling - - - 56�56�128
29 Concatenate - - - 56�56�256 Up-sampling
30 Conv+BN 1 3�3�128 2 56�56�128 block 3
31 Conv+BN 1 3�3�64 2 56�56�64 __________
32 Upsampling - - - 112�112�64 Up-sampling
33 Concatenate - - - 112�112�128 block 4
34 Conv+BN 1 3�3�64 2 112�112�64 __________
35 Upsampling - - - 224�224�64 Up-sampling
36 Conv+BN 1 3�3�3 2 224�224�3 block 5
37 Output - - - 224�224�3

265
During the process of up-sampling block, some
layers are copied and then concatenated with the
output of convolutional layer. After the data calcu­
lated through a down-sampling block, the width and
height will become half. The up-sampling block will
double the width and height of the input.
Convolutional is adopted in both down- and up-
sampling block. One convolutional layer uses a set
of kernels with learnable weight to perform the con­
Figure 3. MaxPooling layer example.
volution operation. Each kernel slides on the input
array with a specific step size defined as stride and
the convolution implemented by this process. The
multiplications are done between the element from
kernel and the element from subarray of input to get
a receptive field. Then the multiplied values are
summed with bias added to get a value in the output
array. To maintain the output size is equal to the
input size, adopt zero-padding (Pad) for the input
array. This process is shown as Figure 2. The output
size of convolutional layer is calculated by the
formula: Figure 4. Up-sampling layer example.

Up-sampling layer is widely used in the up-

sampling blocks. It increases the width and height
from the input array. Any element in the input array
will be repeated in two directions (Figure 4). Every
where O = output size, I = input size, P = pad time after up-sampling, resolution will become 2�2
number, K = kernel size and S = stride. The depth of of the input array. Since the number of Up-Sampling
output is same as the depth of convolutional layer. layers and MaxPooling layers are same, the input
Another key aspect in FCN is Maxpooling layer, and output size of images are same.
which reduces the spatial size of the input array and
often defined as downsampling. It uses a subarray to
get the max values. Figure 3 shows the MaxPooling 3 BUILDING DATABASE
layer with a stride of two. And the output size of the
array is calculated by the formula: Here used 200 high-resolution images full of corro­
sions collected from different steel bridges under dif­
ferent illuminance conditions and at different
shooting angles. Then the corrosion parts in these
images are manually marked by MATLAB 2019a to
make the annotations database. The corrosion part is
where O = output size, I = input size, P = pooling thought as category B, while the undamaged part
size and S = stride size. In this study, P = 2 and S = and the background is marked as category A. After
2. Thus, once the input array goes pass the Maxpool­ that, images and annotations were divided randomly
ing layer, the width and height become half. into two parts, 80% of images is for training and the
other part is for testing. The data set is called data
set A. To compare the difference results of different
method from preparing the data set, we copied both
training and testing data from data set A and cropped
the images and the annotations into 224�224 pixels
to make the data set B. Table 2 shows the detail of
training and testing data set. The process of training

Table 2. The proportion of training and testing data set.

Training Testing

Data set A 160 40

(Full images)
Data set B 9728 2719
Figure 2. Convolutional layer example. (Cropped images)

266
 where JðA; BÞ is the accuracy of detection com­
pared from annotations and the predictions from the
model. A is called ground truth and means the set of
pixels from annotation data set and B is the set of
pixels predicted by the FCN model.
The accuracy of method of using full images is
about 82.79%, while the accuracy of the other
method is 88.68%. To see and compare the results of
two methods in an easier approach, after testing the
Figure 5. Process of two data sets.
cropped images, we stitch the 224 � 224 images
into a full image. The results are shown in Figure 6.
the two data sets is on the computer with CPU i7­ In the Figure 6 I-(c)and I-(d), the pixels of corro­
6700 with memory 16 GB. The training and predict­ sion are predicted, but the numbers of pixels are
ing process for the pixels of corrosion in the images quite different. The same phenomenon can be found
is shown as Figure 5. in Figure II-(c), II-(d) and III-(c), III-(d). The main
reason is full images will make the system focuses
on the relationship between corrosion part and the
4 EVALUATION OF TRAINING MODEL undamaged part instead of being only focused on the
pixels of damage and the pixels adjacent to them.
After training the model, testing data set is used to The model trained from cropped images is too sensi­
evaluate the detection ability of the model to show tive to damage and coatings and easily predict coat­
the location of the corrosion in the images. Here use ing degradation as the category of corrosion even
Jaccard index (Intersection Over Union) as evalu­ though such class of damage haven’t been marked in
ation metric. The result is intersection over union for the original images. What’s more, it is because of the
similar measure between to different sets A and B, size limitation of cropped images that some concrete
and the formula is defined as following: pixels are not distinguished from steel and then pre­
dicted as corrosion. Conversely, the other model
would not so sensitive to damages that even ignore
some damage pixels but would confuse concrete and
steel.

Figure 6. Examples of prediction results: (a) Original image, (b) Ground truth, (c) Prediction result from full images,
(d) Prediction result from cropped Images.

267
5 ACCURACY IMPROVEMENT batch size of training the model was set very small
and the processing of training cost too much time.
5.1 Enhance data set Then the new data set was trained on computer with
GPU, Nvidia GeForce GTX 1080, with a reasonable
To improve the accuracy of results, 100 more images training hyperparameter. The accuracy is about 94.7%
are added to enhance the data set. Most of these and improved about 6% compared with the results
images are with slight corrosion or even no damage, before (shown in Figure 7). Some results are shown as
because the previous outcomes showed that most Figure 8. From I, II and III in Figure 8 we could find
errors are wrongly predictions of no corrosion that the capability of prediction of red paintings obvi­
pixels, such as concrete and red painting. Table 3 ously improved. And some errors such as prediction
shows the detail of the data set after enhancing it. concrete as corrosion also reduced, from I and IV.

5.2 Change hyperparameter

Training hyperparameter is also very significant to the
accuracy of results. Because the limitation of CPU, the

Table 3. The proportion of data set after enhancing.

Training Testing

Data set A 240 60

(Full images)
Data set B 13939 3408
(Cropped images)
Figure 7. Results of training.

Figure 8. Examples of Prediction results: (a) Original image, (b) Ground truth, (c) Before improvement, (d) After
improvement.

268
Figure 9. Prediction results on images taken by UAV: (a) Original image, (b) Before improvement, (c) After improvement.

269
New model also showed better performance of detec­ taken by UAV is used to test the capability of predic­
tion corrosion, compared with the old model. tion corrosion.
In the future, more images with corrosions of dif­
ferent damage levels would be collected to enhance
5.3 Test on images taken by UAV
accuracy. And then the variety of damage to be
One 4k image (4096x2160) of steel bridge sliced detected will be enriched. Future more, UAVs will
from videos taken by UAV in Kawataki is used to be used to collect more images and videos to test the
test the capability of corrosion detection. Figure 9 capability of the FCN model.
shows the results. (a) is Original image while (b) and
(c) are the prediction results before and after accur­
acy improvement. Compared (b) and (c), (b) predict ACKNOWLEDGEMENTS
too many pixels of handrail as corrosion while (c)
does not. Even though both of two models predict The images used in the study to build the data set are
some trees as corrosion because color of these pixels supported by S.Matsunaga, Kyodo Engineering Con­
is similar as color of corrosion, (c) also performs sultant Co., Ltd, Fukuoka Branch, Japan.
better than (b).
REFERENCES
6 CONCLUSIONS Cha, Y.-J., W. Choi, & O. Büyüköztürk (2017). Deep
learning-based crack damage detection using convolu­
A damage detection method based on FCN is pro­ tional neural networks. Computer-Aided Civil and Infra­
posed to identification and position the corrosion structure Engineering 32(5), 361–378.
from the images of bridges. And two different data Iglovikov, V. & A. Shvets (2018). Ternausnet: U-net with
set in two different method are created to train the vgg11 encoder pre-trained on imagenet for image
model and then compare the outcomes from these segmentation. arXiv preprint arXiv:1801.05746.
two data set. And the accuracies come up to 82.8% Jonathan, L., S. Evan, & D. Trevor (2017). Fully convolu­
and 88.7%. It can generally reflect the location of the tional networks for semantic segmentation. IEEE 39,
640–651.
damage and to some extent show the size of the
Ministry, of Land, Infrastructure, Transport, & Tourism.
damage. (2016). White paper on present state and future of social
Then, a new data set is made and the hyperpara­ capital aging infrastructure maintenance information,
meter of training is changed to improve the accuracy. 2016 (in japanese). Technical report.
The accuracy of new data set is 94.7% and the new Simonyan, K. & A. Zisserman (2014). Very deep convolu­
model shows a better performance than the one tional networks for large-scale image recognition. arXiv
before accuracy improvement. Then one image preprint arXiv:1409.1556.

270
 Bridge Maintenance, Safety, Management, Life-Cycle Sustainability and Innovations –
Yokota & Frangopol (eds)
© 2021 Taylor & Francis Group, London, ISBN 978-0-367-23278-8

A crack detection method based on deep transfer learning

Y.G. Shen, Z.W. Yu & Z.L. Wen
College of Civil Engineering and Architecture, Zhejiang University, Hangzhou, China

ABSTRACT: The crack recognition method based on deep learning requires a huge amount of data and it is
difficult to collect adequate data manually, which leads to insufficient training and poor recognition. To solve
this problem, a crack recognition method based on deep transfer learning (DTL) was proposed. An unmanned
aerial vehicle (UAV) was used to quickly collect and product crack data set, then convolution layers and the
fully connected layer of visual geometry group-16 (VGG-16) were trained on ImageNet data set and crack
data set, respectively. Finally, a DTL model was built by connecting two parts with freezing and fine-tuning
some parameters. Experimental results showed that the method significantly reduced image requirements, and
improved the precision of crack detection. 95.9% of the recognition precision was obtained on the validation
set, and 0.983 AP value was obtained on the test set.

1 INTRODUCTION derivative of luminance image (Nishikawa et al.

2012). However, the detection accuracy of IPTs is
The crack is the most common threat to a concrete greatly affected by the image shooting conditions
bridge. It is indispensable to detect the crack width and noise. The rise of machine learning and deep
and propagation on the bridge surface. The com­ learning (LeCun et al. 2015) has brought new ideas
monly used detection method is manual detection, and methods to crack detection. Based on computer
which has high work intensity and low efficiency. vision technology, this method has the advantages of
The more advanced non-destructive testing methods long distance, non-contact, high precision and high
such as ultrasonic method (Sinha et al. 2003), ther­ efficiency, and has become a research hotspot of
mal imaging method (Omar & Nehdi 2017), wireless bridge diseases inspection. A deep structure of com­
strain measurement and electromagnetic-acoustic puter vision-based convolutional neural networks
emission sensor detection method are expensive. (CNN) was proposed using deep learning, which is
With the development of computer technology, a method for detecting concrete cracks without cal­
image processing techniques (IPTs) based on com­ culating defect features, and its data set size is
puter vision is used in crack detection. Four IPTs 40,000, which is cut from 277 crack images with
methods (FHT, FFT, Sobel and canny) were used to 4928 × 3264 pixel resolution (Cha et al. 2017).
compare the results of concrete crack recognition, A deep learning framework called NB-CNN was
and the results show that FHT was superior to the proposed to detect crack information of each video
other three methods (Abdel-Qader et al. 2003). frame, and a new data fusion scheme was proposed
A region growing technique based on support vector to aggregate the information from each frame of
machine (SVM) and similarity standard was pro­ each video to improve the overall performance and
posed to extract road features (Song & Civco 2004). robustness of the system. Its data set size is 300,000,
An image-based penetration model was proposed to which is obtained by data augmentation of 5326
extract the texture of continuous cracks by referring images with 120 × 120 pixel resolution (Chen &
to the connectivity of brightness and the shape of Jahanshahi 2018). The potential of deep learning for
penetration area (Yamaguchi et al. 2008). The appli­ pavement crack detection based on CNN was dem­
cation of morphological multi-level watershed seg­ onstrated using inexpensive smartphones to the col­
mentation in the detection of defects in the lection of crack image data sets. Its data set size is
radiographic images of welds was discussed (Ala­ 1,000,000, which is cut from 500 crack images with
knanda et al. 2009). Based on genetic programming 3264 × 2448 pixel resolution (Zhang et al. 2016).
(GP), an image filter was developed to detect cracks A CNN was used to detect pavement cracks, and
in concrete structures, and the width of cracks then principal component analysis (PCA) was used
detected was quantified based on the spatial to classify the detected pavement cracks. Its data set

DOI: 10.1201/9780429279119-33
https://doi.org/10.1201/9780429279119-33

271
consists of 30,000 images of 32 × 32 and 64 × 64 potentially advanced features from the data set. In
pixel resolution, which is cut from the original crack reality, the data in most areas is limited, resulting in
image of 3264 × 2448 pixel resolution (Zhang et al. an insufficient amount of training data, and data col­
2016). It can be seen from the above scholars’ lection is very time consuming and laborious. Also,
researches that deep learning requires a large amount it is very difficult to build a large-scale, high-quality
of labeled image data, and the training set, the valid­ labeled data set.
ation set and the test set are required to conform to Transfer learning (TL) (Pan & Yang 2010) is an
the same data distribution so that a good detection important tool for solving the problem of insufficient
effect can be obtained. But there is no recognized training in machine learning. It solves the problem
crack database in the field of concrete crack identifi­ of only a small amount or even no labeled sample
cation, which leads to limited sources of crack data, data in the target domain by transferring knowledge
and it takes a lot of manpower and material from the existing source domain to the target
resources to make a massive data set. Therefore, the domain. DTL is a combination of deep learning and
solution to the data set problem is a core step in the transfer learning. By using transfer learning, the
field of crack identification. deep neural network can effectively acquire know­
This paper proposed a crack detection method ledge and only need relatively few data to obtain
based on DTL. The rapid acquisition of images and better recognition effect.
fabrication of the crack data set using a UAV avoids In visual classification problems, especially in
the inconvenience of manual method, and the DTL deep CNN, a lot of research has been done using
method is used to transfer the pre-trained convolu­ DTL. A DTL-based pre-classification method was
tion layer weights as feature extractors, which the adopted to detect and distinguish common cracks
training data is no longer required to have the same and closed cracks in the same framework. The pro­
data distribution as the test data. In this paper, the posed T-DCNN method is far superior to CrackIT,
ImageNet project, which is a large visualization CrackForest and traditional Canny edge detection
database for visual object recognition, was used for method (Zhang et al. 2018). A method was proposed
training of the CNN. Compared with the crack data based on VGG Net and transfer learning to detect
set, ImageNet (Krizhevsky et al. 2017) not only has and identify the component types, spalling states,
a large amount of data but also a very rich variety, damage extents and damage types of 2000 structural
and its image classification data set ILSVRC2011 images, and the application of DTL in image-based
contains 1,229,413 images in training set, 50,000 structural damage identification was revealed (Gao
images in the validation set, and 100,000 images in & Mosalam 2018). Using a pre-trained deep learning
the test set. Although there are not a lot of crack model and TL technique could achieve to detect
images in ImageNet, the DTL model focuses on the crack damage in images of civil infrastructure. The
extraction of general features. Common features proposed method had a precision of up to 90% with­
such as edges and textures extracted by CNN can be out any enhancement and pretreatment (Gopalakrish­
used in all image classification and recognition nan et al. 2018). The above research results show
tasks. Using ImageNet as the source data set, a wide that DTL has great potential and can reduce the
variety of image common features, such as edge, tex­ demand for data of deep learning models without
ture, local detail and color, can be extracted to trans­ reducing the recognition precision or even
fer to the crack detection problem. improving.
The remainder of this paper is organized as fol­
lows. The basic components of the proposed method
2.2 Fine-tuning
are outlined in Section 2. The experimental process
and network structure are introduced in Section 3. Fine-tuning is the most common technique for transfer
The experimental results are presented in Section 4 learning. It transfers the weight parameters trained on
and followed by conclusions in Section 5. the source domain (usually with a large data set) to the
target domain (usually with a small data set) as the ini­
tialization weight, and then retrains the network to
2 OVERVIEW OF PROPOSED APPROACH achieve very good results. Therefore, the amount of
data required to obtain satisfactory performance by
2.1 Deep transfer learning fine-tuning is relatively greatly reduced. Features
extracted from the first few layers of CNN are
Deep learning uses convolutional neural networks as common to many tasks. However, the feature learning
feature extractors to learn the potential low-level and of the convolutional layer at the bottom of the CNN
high-level features from massive data, which makes focuses on advanced features (i.e. semantic features),
deep learning surpass traditional machine learning. which are specific to different data sets. As shown in
But at the same time, data dependence has become Figure 1a, convolutional layers of the pre-trained
one of the most serious problems in deep learning. source model are copied into the target model for fine-
Compared to traditional machine learning methods, tuning, then initialize the output layer of the target
deep learning relies heavily on a large amount of model and trained to obtain a model for the target task.
training data, because only then can it train to extract As shown in Figure 1b, fine-tuning can be divided into
272
the following four ways according to the size of the (3) When the target data set is small and the similar­
target data set and the similarity with the source ity with the source data set is high, fine-tune the
data set: full connection layer of the pre-training model;
(4) When the target data set is large and the similar­
ity with the source data set is high, fine-tune the
(1) When the target data set is small and the similar­ whole pre-training model.
ity with the source data set is low, fine-tune the
top convolution layer of the pre-training model;
(2) When the target data set is large and the similar­
2.3 3VGG-16 net
ity with the source data set is low, train the
whole neural network model; VGG-16 (Simonyan & Zisserman 2014) is
a CNN developed by the Visual Geometry
Group. It won the ILSVR (ImageNet) 2014
championship and is a very representative clas­
sical CNN structure. As shown in Figure 2, the
VGG-16 is composed of five convolutional
blocks and one fully connected classifier, each of
the blocks is composed of a plurality of identi­
cally convolutional layers with a padding number
of one and a window shape of 3 × 3 followed by
a maximum pooling layer with a stride of two
and a window shape of 2 × 2, contains a total of
thirteen convolutional layers, five pooling layers,
and three fully connected layers. The convolu­
tional layers and the fully connected layer contain
the feature parameters extracted by the neural
network, so they are also called weight layers,
and the total number is 13 + 3 = 16, which is
the source of 16 in VGG-16.
Compared with the traditional CNN, VGG-16
is characterized by stacking multiple small convo­
lution kernels in one block and then using max­
imum pooling operations, which can extract local
features more carefully and reduce the number of
parameters. For example, by stacking three 3 × 3
convolution layers instead of using a single 7 × 7
convolutional layer, three nonlinear functions can
be combined instead of just one, making the
decision function more discriminative and charac­
terization. At the same time, the parameter
amount is reduced by (7 × 7 - 3 × 3)/7 × 7 =
81.6%, and the receptive field remains
unchanged. The increase in the number of convo­
lution kernels also makes the weighting of the
convolution kernels more regular, reducing the
risk of overfitting and increasing the effectiveness
Figure 1. Overview of fine-tuning. of different convolution kernels.

Figure 2. Visualization of VGG-16 structure.

273
3 CRACK DETECTION BASED ON DEEP 3.2 Deep neural network structure
TRANSFER LEARNING
The model used in this paper contains thirteen con­
The DTL method was used to transfer the feature volutional layers, five down-sampling layers, i.e.
extraction layers (convolution layers and pooling pooling layers, and two fully connected layers. The
layers) of the VGG-16 model which were pre­ number of convolution filters in the 1-2th layer is 64,
trained on ImageNet to the classification task of in the 3-4th layers is 128, in the 5-7th layers is 256,
crack data set. The transferred feature layers of in the 8-13th layers is 512. The 15th layer is the
DTL covered complex and diverse features, so as to dropout layer, and the corresponding dropout rate is
improve the generalization ability of the crack 0.5. The 14th and 16th layers are fully connected
detection model. Although ImageNet and the target layers, the corresponding number of neurons are 256
crack data set have some differences, there are and 1. The size of all convolution filters is 3 × 3, and
some common general underlying abstract features the total model parameters are 16,812,353.
at the feature level, so CNN-based DTL can be As shown in Figure 4. Among them, the left side is
performed. the DTL crack detection model, which is mainly com­
posed of three parts: 1) freezing layer. It consists of
convolution blocks 1# to 4#, which are transferred
3.1 Experiment process from VGG-16 and loaded with parameters that have
Figure 3 shows the experimental process consisted been trained by ImageNet, which are immutable; 2)
of four stages: fine-tuning layer. Consisting of the 5# convolution
block, parameters of this part will be fine-tuned with
the training of the model; 3) classifier layer. The fully
(1) A UAV with an upward camera was used to col­ connected layer of this part comes from the pre-trained
lect the crack images. In the image pre­ VGG-16 fully connected layer, and parameters of this
processing stage, all the images were resized part are also constantly fine-tuned as the model is
according to the corresponding categories and trained.
unified into three channels. The image reso­
lution was adjusted to 150 ×150 and data
enhancement (including random rotation, cut­
ting, and flipping) and image normalization
were carried out;
(2) Pre-trained a fully connected layer using
VGG-16 on the crack data set. The specific
method is to take the crack data set as input,
and exported the convolution layer in the
trained VGG-16 on ImageNet as a feature
extractor, then trained a fully connected layer
for the crack data set;
(3) Transferred the feature layers above the fully
connected layer in VGG-16 trained on Ima­
geNet, and connecting with the pre-trained fully
connected layer to obtain a DTL model for the
crack identification task;
(4) Froze partial parameters, fine-tuning and train­
ing to predict other images.

Figure 3. Experiment process. Figure 4. Structure of DTL model based on VGG-16.

274
4 EXPERIMENTAL RESULTS eye to distinguish smaller targets, which will
greatly increase the difficulty of labeling. As
For concrete crack detection, it mainly relies on shown in Figure 6, the database is made up of
a classifier that can distinguish crack images from a wide range of crack types, the thickness of the
crack-free images. For DTL, the determination of fine- crack, the direction of development, the position
tuning convolutional layers and the selection of hyper- and clarity of the map are all different, and the
parameters (such as learning rate and loss rate) are non-crack images contain stains, scratches, stencil
tedious because there is no precise guidance on the marks, artificial handwriting or watermarks, etc.,
optimization of these parameters. The optimal fine- which provides a guarantee for training a crack
tuning layers of the model must be explored through detector with robustness and stability.
trial and error, corrected by the identification error of
the validation set, and finally tested with the test set.
4.2 Result analysis and comparative evaluation
All work done in this paper is performed on
a workstation with four GPUs, CPU: Intel Xeon E5­ In order to explore the transfer effect of the
2620v4 8 Core/2.1 GHz/20M, RAM: 128 GB and model and the optimal transfer layers, this paper
GPU: GeForce GTX 1080 Ti. set up six comparative experiments, among which
there were five recognition methods based on
DTL: (1) freezing the 1# convolution block, fine-
4.1 Data preparation
tuning the 2-5# convolution block and the fully
The total number of original images obtained in this connected layer; (2) freezing the 1-2# convolution
paper was 142, and the pixel resolution was 4608 × block, fine-tuning the 3-5# convolution block and
3456. Compared with the conventional deep learning the fully connected layer; (3) freezing the 1-3#
method described above, the number of images is convolution block, fine-tuning the 4-5# convolu­
greatly reduced. The number of images required for tion block and the fully connected layer; (4)
the DTL model is only one-tenth or even one percent freezing the 1-4# convolution block, fine-tuning
of the conventional method. These images were taken the 5# convolution block and the fully connected
by a UAV (model: DJI Jingwei M210RTK, lens ZEN­ layer, that is, the original model adopted in this
MUSE X4S) on a long-standing bridge, as shown in paper; (5) freezing the 1-5# convolutional block
Figure 5, and the illumination intensity and shooting and only fine-tuning the fully connected layer,
distance of each image were different. Of the 142 ori­ that is, transfer learning. There was one type of
ginal images, 120 were used for training, validation, recognition method based on deep learning, that
and test, and 22 were used for second test. The 120 is, retraining the VGG-16 model. For the con­
original images were cropped to an image of 150 × venience of description, according to the number
150 pixel resolution, and each image was labeled as of convolution block freezes, the above models
a crack or non-crack image for building a database of are referred to as DTL-1, DTL-2, DTL-3, DTL-4,
training, validation, and test. The total number of TL, and VGG-16, respectively.
training, validation, and test images prepared in the Figure 7 Shows the precision curve comparison
database was 8000, 2000, and 2000, respectively, and of DTL-1, DTL-2, DTL-3, DTL-4, TL and VGG­
images were randomly selected from the database to 16 on the training set and validation set. Figure
generate a training set, a validation set, and a test set. 7a shows the precision changes of six methods
The reason for choosing a relatively small crop based on the training set, among which, all DTL
size is determined by the size of the crack, and methods and VGG-16 methods tend to be stable
a CNN trained on a small image can scan any after the rapid increase of precision in the first
image larger than the design size. However, if 100 epochs, all DTL methods are almost stable at
a smaller image is selected, the network may 100%, VGG-16 methods are stable at about 98%,
extract more elongated features such as scratches, while TL methods in 400 epochs, from 80% to
handwriting, etc., increasing the difficulty of clas­
sification. Meanwhile, it is not easy for the naked

Figure 5. Images collection by a UAV. Figure 6. Diagram of some samples in the data set.

275
 In order to quantitatively evaluate the pros and
cons of the generalization ability of the model after
the final training, this paper evaluated the six indica­
tors by using precision (P), recall (R), average preci­
sion (AP), F1-Score value (F1), receiver operating
characteristic (ROC) curve and area under curve
(AUC), as shown in Figure 8. For the binary classifi­
cation problem, the confusion matrix can classify the
target sample according to the combination of its
real category and the model prediction category into
true positive (TP), false negative (FN), false positive
(FP), and true negative (TN). The specific classifica­
tion results are shown in Table 1.
According to the four indicators of the confusion
matrix, we can calculate the following five metrics:

Figure 7. Comparison of training and validation results.

where TPR is true positive rate, and FPR is false
positive rate.
90%, still have a certain upward trend. Figure 7b Using six models to identify cracks in the image
shows the precision variation of six methods of the test set and the quantization results are shown
based on the validation set, among which, the in Figure 8. The PRC curve is shown in Figure 8a,
precision of all DTL methods and TL methods the F1-PR curve is shown in Figure 8b, the ROC
increased rapidly in the first 20 epochs, and then curve is shown in Figure 8c, and the other quantita­
basically stabilized, DTL-1 and DTL-3 methods tive comparison results are shown in Figure 8d. It
stabilized between 95% and 96%, DTL-2 can be seen from Figure 8d that as the number of
methods stabilized between 96% and 97%, DTL­ frozen convolution blocks increases, the AP value,
4 methods stabilized between 91% and 92%, TL F1-Score value and AUC value both show a trend of
methods nearly tended between 87% and 88%, increasing first and then decreasing, and DTL-2
while the precision of VGG-16 method increased obtains the highest AP value, F1-Score value and
rapidly in the first 50 epochs, and then tended to AUC value, which are 0.983, 0.970 and 0.986,
be stable, but there is still a weak downward respectively. This is because, in the CNN, the fea­
trend, and the precision fluctuated greatly, and tures extracted by the underlying convolutional
finally stabilized between 92% and 93%. blocks are mainly general features, while the features
Figure 7 shows that all DTL methods have signifi­
cantly better learning ability on a small crack data set
than VGG-16 and TL methods, mainly because the
Table 1. Confusion matrix of classification results.
VGG-16 model has a great dependence on data and
requires a large amount of data for feature learning. True background Prediction results
However, this paper used a small data set, and the
VGG-16 method was easy to fall into over-fitting,
crack non-crack
which leads to the reduction of the generalization abil­
ity of the model, resulting in lower precision of the val­ Crack TP FN
idation set. Among all DTL methods, the DTL-2 Non-crack FP TN
method was obviously superior to other DTL methods.

276
 extracted by the top-level convolutional blocks are
the semantic features of the target task. It can be
seen from Figure 9a and b that after the fine-tuning
of the 3# block of DTL-2, the first layer of the 4#
block extracts more features than DTL-3 without
fine-tuning the 3# block. Figure 9b also confirms this
phenomenon. The generalization ability of the TL
model is far worse than the VGG-16 model and the
four DTL models, indicating that if the source data
differs greatly from the target data, the recognition
effect of the direct transfer will be far less than that
of the model trained on the target data and DTL
model after freezing and fine-tuning. Therefore, the
best freezing and fine-tuning model obtained on the
crack data set of this paper is DTL-2, which freezes
1# and 2# convolutional block, fine-tunes the 3#, 4#
and 5# convolutional block and the fully connected
layer.
After determining the optimal freezing and fine-
tuning layers, the DTL-2 model was used to test the
crack images of other structures. The test results are
shown in Figure 10. The results indicate that the
DTL model also has strong generalization ability
and robustness, which can be applied to practical
engineering tasks.

Figure 9. Visualization of features extracted by convolu­

tion layer.

Figure 8. Comparison of various index values on test set. Figure 10. Prediction results of other images.

277
5 CONCLUSIONS Gao, Y.Q. & Mosalam, K.M. 2018. Deep transfer learning
for image-based structural damage recognition. Com-
When the amount of crack data is insufficient, the puter-Aided Civil & Infrastructure Engineering 33(9):
experimental results obtained by deep learning are 748–768.
often poor. TL is applied to deep learning and good Gopalakrishnan, K., Gholami, H., Vidyadharan, A.,
results are obtained, which shows that DTL can be Choudhary A., & Agrawal A. 2018. Crack damage
used to solve the problem of data shortage and has detection in unmanned aerial vehicle images of civil
infrastructure using pre-trained deep learning model.
strong potential. This paper introduced DTL into the International Journal for Traffic & Transport Engineer­
crack identification by transferring the feature extractor ing 8(1):1–14.
from VGG-16 pre-trained on ImageNet and a fully Krizhevsky, A., Sutskever, I., & Hinton, G. 2012. ImageNet
connected layer pre-trained on crack data set, thus classification with deep convolutional neural networks.
a DTL model for crack detection task was constructed. Communications of the Acm 60, 84–90.
From the experiment results, the following conclusions Lecun, Y., Bengio, Y., & Hinton, G. 2015. Deep learning.
can be obtained: Nature 521(7553): 436–444.
Nishikawa, T., Yoshida, J., Sugiyama, T., & Fujino, Y.
2012. Concrete crack detection by multiple sequential
(1) Compared with the traditional deep learning image filtering. Computer-Aided Civil & Infrastructure
method, DTL only needs one-tenth or even Engineering 27(1): 29–47.
one percent of the data. Omar, T. & Nehdi, M.L. 2017. Remote sensing of con­
(2) Among all the DTL models in this paper, the crete bridge decks using unmanned aerial vehicle
infrared thermography. Automation in Construction
DTL-2 model which only freezes the 1# and 2#
83: 360–371.
convolution blocks and fine-tunes the remaining Pan, S.J. & Yang, Q. 2010. A survey on transfer learning.
weight layer performs best, and it achieves IEEE Transactions on Knowledge & Data Engineering
95.9% recognition precision on the validation set, 22(10): 1345–1359.
an AP value of 0.983 on the test set, an F1-score Simonyan, K. & Zisserman, A. 2014. Very deep convolu­
value of 0.970 and an AUC value of 0.986. tional networks for large-scale image recognition. Com­
(3) Compared with the deep learning, the evaluation puter Science.
indexes of the DTL model are greatly improved, Sinha, S.K., Schokker, A.J., Iyer, S.R., & Council, L. S.
with strong robustness and generalization ability. 2003. Non-contact ultrasonic imaging of post-tensioned
bridges to investigate corrosion and void status. Pro­
ceedings of the IEEE Sensors 2003, 487–492.
Song, M.J. & Civco, D. 2004. Road extraction using SVM
REFERENCES and image segmentation. Photogrammetric Engineering
& Remote Sensing 70(12): 1365–1371.
Abdel-Qader, I., Abudayyeh, O., & Kelly, M.E. 2003. Ana­ Wang, X.L. & Hu, Z.Z. 2017. Grid-based pavement crack
lysis of edge-detection techniques for crack identifica­ analysis using deep learning. International Conference
tion in bridges. Journal of Computing in Civil on Transportation Information & Safety (ICTIS)
Engineering 17(4): 255–263. 917–924.
Alaknanda, Anand, R.S., & Kumar, P. 2009. Flaw detection Yamaguchi, T., Nakamura, S., Saegusa R., & Hashimoto S.
in radiographic weldment images using morphological 2008. Image-based crack detection for real concrete
watershed segmentation technique. NDT & surfaces. IEEJ Transactions on Electrical & Electronic
E International 42(1): 2–8. Engineering 3(1): 128–135.
Cha, Y.J., Choi, W., & Büyüköztürk, O. 2017. Deep Zhang, K.G., Cheng, H. D., & Zhang, B.Y. 2018. Unified
learning-based crack damage detection using convolu­ approach to pavement crack and sealed crack detection
tional neural networks. Computer-Aided Civil & Infra­ using preclassification based on transfer learning. Jour­
structure Engineering 32(5): 361–378. nal of Computing in Civil Engineering 32(2): 04018001.
Chen, F.C. & Jahanshahi, M.R. 2018. NB-CNN: Deep learn­ Zhang, L., Yang, F., Zhang, Y.M., & Zhu, Y. 2016. Road
ing-based crack detection using convolutional neural net­ crack detection using deep convolutional neural
work and Naïve Bayes data fusion. IEEE Transactions on network. IEEE International Conference on Image Pro­
Industrial Electronics 65(5): 4392–4400. cessing (ICIP 2016).

278
 Bridge Maintenance, Safety, Management, Life-Cycle Sustainability and Innovations –
Yokota & Frangopol (eds)
© 2021 Taylor & Francis Group, London, ISBN 978-0-367-23278-8

Measuring traffic-induced loads and 3D bridge displacements with UAVs

B.J. Perry & Y. Guo
Department of Civil and Environmental Engineering, Colorado State University, Fort Collins, Colorado, USA

ABSTRACT: Fully understanding the bridge performance under traffic loadings is critical for improving
bridge design, condition assessment, and load rating. Modern structural health monitoring (SHM) has enabled
measurements of the traffic loads and dynamic bridge response to help enhance the knowledge on the mech­
anism of vehicle-bridge interactions; however, challenges still exist for accurately measuring the moving traf­
fic loads and synchronizing the loads with the corresponding traffic-induced response. Recently, with the
tremendous advancements in unmanned aerial vehicles (UAVs) technology (including better camera perform­
ance and longer flight times), UAVs can offer unique advantages to hover at specified heights and key loca­
tions and access difficult to reach areas and critical angles while providing relatively stable and high-quality
imagery. By leveraging the recent advantages in UAV technologies, image computation, and camera vision-
based SHM, this study proposes a UAV-based SHM system to track vehicular loading and measure the dis­
placement of the bridge at the same time. In the system, multiple UAVs will hover adjacent to a bridge and
communicate with each other to take videos of the moving traffic and structure response from different angles
simultaneously. Then, an object identification algorithm in image computation will be developed to identify
and track moving vehicles. Digital Image Correlation (DIC) will be employed to quantify the three-
dimensional dynamic displacement of the bridge. The feasibility of using videos from multiple UAVs in
object identification and displacement measurement will be investigated. In the future, the data measured by
the proposed UAV-based SHM system can be fused with data from traditional weigh-in-motion systems to
allow precise estimation of moving traffic loadings while measuring corresponding structural response, thus
providing valuable data for accurate modeling and assessing bridge performance under traffic loads.

1 INTRODUCTION management of the wires if a wired system or syn­

chronization of the sensors if a wireless system is
Understanding the dynamics of bridges is critical for adopted. For instance, Linderman et al. attached
evaluating long-term structural performance and deci­ twenty-six accelerometers on the structure to measure
sion-making regarding the operation of the structure. the necessary dynamic response of I-35W Saint
However, there still exist two challenges when study­ Anthony Falls Bridge in Minneapolis, Minnesota (Lin­
ing the dynamics of bridges. Firstly, the constantly derman 2019). Although the accelerometer sensors
changing and fluctuating dynamic loads on the struc­ provided a wealth of information that is not typically
ture are difficult to measure and thus are often treated available for large-scale bridges, the scale of the sen­
as unknowns. Secondly, the measurement of dynamic sory array may impose logistical problems for the
displacement is more difficult compared to the acceler­ bridge owners and researchers. Several of the sensors
ation. The total displacements cannot be fully lost communication due to a number of reasons. More­
recovered from accelerometers of inertial measurement over, once the sensors were attached, they were not
units due to the inability to solve for the constants of easily able to be relocated if a different, critical location
integration when double integrating the accelerations. was found. Furthermore, the entire undertaking did not
Overcoming these two challenges would help engin­ measure the induced dynamic loads of the structure.
eers, researchers, and city planners to better understand Recently, Unmanned Aerial Vehicles (UAVs) are
the performance of the structure and make the most being deployed to assist bridge inspectors and man­
informed decisions about maintenance and operations. agers. UAVs have the potential to provide information
Traditionally, accelerometers have been deployed to from various difficult-to-access locations of a bridge
measure the dynamic response of bridges. However, on a faster, more cost-effective, and safer platform
these systems require a large number of sensors, care­ when compared to traditional techniques. Some studies
ful pre-planning of the sensor placement, and have already explored the possibility of utilizing UAVs

DOI: 10.1201/9780429279119-34
https://doi.org/10.1201/9780429279119-34

279
as “eyes-in-the-sky” to assist inspectors in the field Intel’s RealSense senor and Microsoft’s Kinects
(Gillins, Gillins, & Parrish 2016, Wells & Lovelace sensor implement a virtual speckle pattern onto an
2018). These studies investigate the viability of using object to measure the depth from the camera to an
UAVs equipped with optical sensors to assist inspectors object. It incorporates two infrared cameras and
and show the possibility for future implementation for an infrared laser projector. The laser projects the
bridge inspection. As the size and capabilities of vari­ speckle pattern on a subject. Next, the two infrared
ous sensors improve, a more diverse application of cameras on each side of the laser projector capture the
UAV sensors become possible. This study proposes reflection of the speckle pattern. The two infrared
equipping a UAV with various optical and infrared sen­ cameras are rigidly attached to the laser source, and,
sors to capture both the dynamic loading and structural with the camera matrix known, can track each speckle
response of bridges. Currently, there are a few sensors and calculate the depth of the object from the sensor.
that can potentially be used to robustly collect this Utilizing the virtual speckle pattern technique built
information. This paper examines the use of an Intel® into the Intel’s or Microsoft’s sensors systems have
RealSense D35 depth sensor in conjunction with typ­ been successfully applied in human-computer inter­
ical optical cameras attached to a UAV. The RealSense actions and medical fields. Nguyen et al. used two
sensor was initially developed for human-computer Microsoft’s Kinects sensors and the virtual speckle pat­
interactions in 3-D environments. Applying this sensor tern to measure the displacements in a variety of
to civil infrastructure, it can tackle the measurements objects and was able to measure the breath rate of
of the displacement of a bridge. To provide the live a chest and the pulse from a subject’s neck with a high
load information, a truck-tracking procedure based on accuracy (Nguyen & Choi 2018). Aoki et al. created
computer vision is developed which allows a UAV to a virtual grid using a green laser to project a pattern on
follow a truck from a weigh station to the bridge a subject’s chest to measure with high accuracy the
location. heartbeat of a subject (Aoki, Suzuki, & Shiga 2018).
There is some research on using UAVs to track the Within these two studies, the placement and movement
displacements of a bridge. These UAVs systems rely of the sensors were highly controlled and measured. If
on an optical sensor on a UAV and a painted speckle these systems were implemented on a UAV where
pattern on a bridge to measure dynamic displacement there is random vibration and/or movement due to the
(Kalaitzakis, Kattil, Vitzilaios, Rizos, & Sutton 2019, instability of the UAV hovering, there could be signifi­
Catt, Fick, Hoskins, Praski, & Baqersad 2019). Baqer­ cant error induced in the measurements. Therefore,
sad et al. mounted two cameras with known distances there needs to be a system developed that can either
apart and known camera parameters to measure the overcome or compensate the vibrations and/or move­
deformations of a deformable board. The proposed ments of the projector or sensors during the tests.
algorithm processes the speckled pattern placed on the This paper presents a proof-of-concept study of
board to accurately measure the precise a framework to 1) measure the 3-D displacement of
2-D deformation and movement of the speckles. More­ a structure using an optical sensor for 2-D planar
over, (Kalaitzakis, Kattil, Vitzilaios, Rizos, & Sutton movement and an Intel® RealSense D35 Sensor for
2019) implemented a similar set up to measure strains 1-D depth movement, as well as 2) tracks the
of a concrete beam during a four-point loading test. imposed loading with a truck-tracking procedure
Lastly, Instron Testing Equipment has a commercial with UAVs. The paper will be organized to first
software to measure strains during tensile tests using examine the overall framework and real-world
similar techniques of measuring the displacements of implementation. Next, three sections are presented to
a speckle pattern on a specimen (Instron 2017). How­ introduce the algorithms of the truck-tracking,
ever, these three systems require an applied speckle 2-D planar movement, and 1-D depth movement of
pattern to be manually drawn on the specimen prior to an object and demonstrate the implementation of the
the tests. Implementing this type of system in the real- proposed three algorithms, respectively.
world might be difficult due to the large-scale of the
structures and the workforce required. Additionally, the
systems measure only the 2-D planar deformation of 2 OVERVIEW OF PROPOSED FRAMEWORK
the object; the plane that is perpendicular to the AND ITS REAL-WORLD IMPLEMENTATION
camera. It is not able to measure the full,
3-D movement of an object. Recently, several products The three modules of the framework will be presented
have made an appearance on the market place which through a proof-of-concept study. The first module is
can measure displacements using a virtual speckle pat­ “truck-tracking” which follows a truck to measure the
tern instead of the real speckle patterns used in the pre­ dynamic loading. The second module is tracking the
vious studies. However, these systems have 2-D planar movements of an object using a typical
a limitation of only being able to measure 1-D, i.e., the RGB/optical sensor on a UAV. The third module is to
distance from the sensor to the object (depth). Consid­ measure the 1-D depth movement of an object using
ering the advantages and shortcomings of the afore­ the RealSense sensor. The accuracy of the implemen­
mentioned techniques, this study proposes to integrate tation of the three modules will be examined.
the 2-D planar measurement and 1-D depth measure­ First, to measure the dynamic loading of the struc­
ment to measure all 3-D dynamic displacements. ture to build an input-output model, it is proposed to
280
implement a “truck-tracking” system. A UAV is
synced to the weigh-in-motion sensors currently
installed in the United State’s interstate infrastructure.
Originally developed and used to ensure compliance
with state and federal regulations, the information can
be collected and used in the framework. The weight of
a large truck provides the critical live loads of
a structure, which is often the most relevant load
regarding the operation of the infrastructure. Once the
weight is collected, a UAV tracking system is used to
follow the weighed truck from the weighing area to the
bridge and follow the truck as it passes over the bridge.
Next, the feasibility of measuring the 3-D Figure 1. UAV Tracking a truck at 90-meters Above
dynamic response of a bridge will be presented. The Ground Level.
3-D measurement is accomplished by two modules.
The first module uses RGB/optical sensors on
a UAV to track and measure 2-D planar dynamic dis­ and can be implemented in real-time. Connecting with
placement perpendicular to the camera. The second a weigh-in-motion system, the weight of a truck is
module uses the RealSense sensor to measure depth associated with the position of the vehicle through the
from the sensor to a region of interest. A UAV tracking algorithm. The truck’s load will be known as
equipped with a set of RealSense sensors will hover it crosses over the structure.
adjacent to a bridge to measure the displacement of To track and follow a truck, the algorithm and tech­
the structure through tracking the movement of pro­ nique are fairly robust and well studied. The technique
jected the virtual speckle pattern. Integrating the was demonstrated with a UAV flying about 90-meters
2-D planar motion with the 1-D depth motion infor­ above ground level and followed a truck on a road as
mation, the total 3-D dynamic displacement of the shown in Figure 1. The red box is the identified truck
bridge is known. These three modules together will with the weight of the vehicle displayed on top.
enable simultaneous measurement of both dynamic
loading and displacement on a bridge, which will
allow in-depth analysis of the structural dynamic 4 2-D PLANAR DISPLACEMENT
performance and enhance the accuracy of the condi­ MEASUREMENT MODULE
tion evaluation of bridges. The three proposed mod­
ules will be discussed in the following sections in With the truck-tracking algorithm implemented to
detail with results from case studies presented. measure the dynamic traffic load, the UAV sensors can
be used to measure the displacement of the bridge sim­
ultaneously. Two techniques are integrated to find the
3 TRUCK-TRACKING MODULE dynamic 3-D displacement of a bridge. The first algo­
rithm is an implementation of a 2-D measurement tech­
To provide the live loading of a bridge to allow for nique onto an optical sensor attached to a UAV
a full input-output analysis, a truck is tracked from the (Kalaitzakis, Kattil, Vitzilaios, Rizos, & Sutton 2019,
weigh-in-motion station until it passes over the bridge. Catt, Fick, Hoskins, Praski, & Baqersad 2019). Meas­
For this implementation, it is assumed that the weigh- uring the 2-D planar motion perpendicular to the
in-motion stations are in close proximity to a bridge of sensor requires tracking of key-points within an image
interest. However, as a bridge becomes more remote, over time to calculate the movement. This type of tech­
the live load truck tracking may not be possible due to nique has shown potential using stationary cameras and
the energy required to track a truck over a longer dis­ sensors; however, when applying these techniques to
tance. Tracking of objects is currently a trivial task in the camera attached to UAVs, a problem arises. There
computer vision. In this implementation, a “Minimum are movement and instability with the UAV as it
Output Sum of Squared Error” (MOSSE) filter is hovers. Although the UAV may remain stable in the
used. The MOSSE filter is created by a pixel-wise div­ air, there is slight drifting and movement of the sensor
ision of an image with only a Gauss point placed at and error is imposed. A correction of this error must be
the known location of the object and the entire frame applied to get the movement of the actual object. In the
from a video in the frequency domain (Bolme, Bever­ following, the algorithm to track 2-D planar motion
idge, Draper, & Lui 2010). Once the filter is obtained, and the treatment of correcting the error due to the
it is multiplied by the next frame to produce a new UAV movement are introduced.
Gauss point of where the object is. The filter is then The proposed 2-D displacement tracking tech­
updated and the process is continued. It requires only nique uses the scale-invariant feature transform
the initial position of the object being tracked and (SIFT) algorithm to identify key-points and their
requires the object to move continuously from frame associated descriptors within a region of interest
to frame. It has proved to be a highly effective, fast, (ROI)(Lowe 1999). SIFT key-points are irrelevant of
and accurate technique to track an object in a video scale and orientation and prove to work well in

281
a variety of applications. SIFT key-points track the from the movement of the object and the true move­
subpixel locations of key-points within an image. ment of the object is found.
The number of SIFT key-points within an ROI
depends on the size, resolution, and texture of the
4.1 Case study
object. Smoother objects such as concrete tend to not
reveal many key-points in an image, thus they may To test this algorithm, the dynamic displacement meas­
introduce difficulties for the following motion track­ urement of a transmission line was studied. Although
ing. Therefore, to overcome this problem, an artifi­ the methodology was originally developed for bridge
cial pattern (i.e. the speckle pattern) is implemented application, the transmission line test was used as
to provide areas of interest for the SIFT algorithm to a proof of concept because of the available experimen­
identify more key-points. Performing this SIFT ana­ tal opportunity. Another research group from Colorado
lysis on a frame-by-frame basis of video collected State University was conducting tests on the vibration
by a UAV allows tracking on a subpixel level of the of a transmission line under blast-induced excitation to
area of interests. Next, using a brute force technique, study the mechanism of how the transmission line
the key-points and its related descriptors are matched dynamically reacted to the blast. Therefore, it provided
with the initial frame for a point of reference. The a good experimental opportunity for the authors to test
matches are then sorted and the top five matches are the developed UAV-based dynamic displacement
used for the analysis. Since the selected ROI is rela­ measurement algorithm. Since both transmission lines
tively small, averaging the top five matches provide and bridges are line-like structures, the feasibility
a measurement of the location of the ROI. Compar­ tested on a transmission line implies the potential
ing this average with the initial frame identifies the of the method for bridge application. An elevation
movement of the ROI from frame to frame. At first view of the experimental setup is shown in Figure 2.
appearance, it may seem best to match the same Figure 3 is the experimental setup taken from the UAV
identified key-points from frame to frame; however, in flight before the blast was set off.
during the implementation, it was found that aver­
aging produced more precise results. This is prob­
ably due to the variation in the accuracy of the
matches. For instance, for different frames, different
key-points may be better and more precisely
matched, and, therefore, the related movement is
better estimated by taking the average.
As discussed earlier, the movement of the UAV
during flight introduces an error on the 2-D motion
measurement results. Therefore, it is necessary to cor­
rect or compensate for the UAV’s movement. To over­
come the movement of the UAV during flight, SIFT
key-points of the background are also found and
tracked from frame to frame. Assuming the movement
of the background is due only to the movement of the
UAV, the background movement is subtracted out Figure 2. Elevation View of Blast Test Set-Up.

Figure 3. Plan View of Experimental Set-up of Using a UAV to Measure Dynamics of Transmission Line.

282
 Since the transmission line was a relatively smooth, 5 1-D DEPTH MEASUREMENT MODULE
and narrow surface, few SIFT key-points were pro­
duced within a small ROI. To provide better SIFT An Intel RealSense Depth Camera was used. This
points identification, an artificial pattern (i.e. the camera was initially developed to measure and track
speckle pattern) was manually imposed on the trans­ hand motions to interact with computers within vir­
mission line. The pattern was a random black and tual environments with virtual reality head-mounted
white polka dot pattern which provided high contrast displays. The camera casts an infrared laser that pro­
for identification of SIFT key-points. The dot pattern jects a virtual speckle pattern onto an object. In add­
was printed on paper and wrapped around the transmis­ ition to an RGB camera, two infrared cameras are
sion line. Since the paper was thin, light-weight, and attached to the sensor which records the reflection of
matched the shape of the transmission line, it was the laser pattern. The camera’s “camera matrix”
assumed that any error imposed by the artificial speckle which consists of the intrinsic and extrinsic informa­
pattern by stiffening or restricting the movement of the tion is known and constant. Knowing the “camera
wire was negligible. The UAV was flown about matrix,” the speckle pattern of the infrared laser is
6-meters above the transmission line and about tracked and measured. Using Intel’s RealSense SDK
7.5-meters above ground level. Using a DJI Mavic 2 2.0, an algorithm was developed which measured the
Zoom with the 2X optical zoom provided a ground- distance from an object to the sensor and the power
sampling distance of 0.13-cm/pixel. Next, the same spectral density of the obtained time history of dis­
random speckle polka dot pattern was secured to the placements was estimated using a Fast Fourier
ground below the transmission line. This was used as Transform. After adjusting the camera’s parameters,
a reference to correct the movement of the UAV as pre­ a small ROI is selected which is small enough to be
viously mentioned. considered as a point on the object. In this way, the
The movement of the transmission line and the dynamic displacement of the object in the depth dir­
movement of the stationary ROI were measured and ection from the sensor to ROI is measured.
subtracted to get the true movement of the transmission
line. Since it is not feasible to install accelerometer or
5.1 Case study
GPS sensors on the thin power line to do a comparison
study, the ground-truth data is currently not available; To test the proposed method, a simple, small scale
however, the measurement results appeared consistent experiment was set up. Using a custom-built shake-
with what was observed, which indicates the initial suc­ table which will produce known, and controlled dis­
cess of this method. The resulting, estimated dynamic placements, a prismatic, rigid structure was secured
displacement of the transmission line captured by the onto it. The excitation of the shake table was varied
UAV is shown in Figure 4. The power spectral density between 0.50 - 2.00-Hz. The cameras were set up per­
estimated using a Fast Fourier Transform is also pendicular to the rigid structure at a distance approxi­
shown, where the dominant frequency is 2.80-Hz. mately 0.5-meters away in a defined x- and y-direction.
The cameras recorded the two infrared cameras and the
depth was calculated. The max frame-rate of the
camera is 30-fps. According to the classic Nyquist-
Shannon sampling theorem, the sampling frequency
needs to be at least twice the maximum frequency of
the vibration to faithfully record the signal. So the sam­
pling frequency of 30-fps is large enough to capture
vibrations up to 15-Hz. For the tested systems, the max­
imum frequency of vibration is less than 2-Hz; there­
fore, the camera records at a sufficient rate to provide
enough sampling points to measure the frequency of
vibration. These measured vibration frequencies from
the RealSense sensor were compared to an accelerom­
eter secured onto the shake-table. The results for two
tests with different excitation frequencies are shown in
Figure 5. The results for all five of the tested frequen­
cies are compared with those obtained from the accel­
erometer in Table 1.
From the comparison results listed in Table 1,
the percent difference from the accelerometer and
the RealSense camera are less than 2%. This dem­
onstrates the efficacy of the proposed algorithm to
measure the frequency of the vibration. One should
note that with accelerometers, the total displace­
Figure 4. 2-D Planar Displacement of the Transmission ments cannot be fully recovered due to the inability
Line recorded by a UAV. to solve for the constants of integration when
283
 support decision-making for maintenance and
repairs of existing infrastructure by bridge owners
and managers.

7 FUTURE WORK

The proposed framework presented case studies for

each module. The entire framework needs to be
tested and studied on a bridge with all three modules
connected and the performance needs to be studied
and further evaluated.
The truck tracking module needs to be tested in an
environment with multiple vehicles present. The
MOSSE filter is quite robust when tracking one
object. But often time, multiple trucks can appear
very similar from above. The efficacy of the MOSSE
filter, in this case, needs to be tested. Moreover, as the
distance from a weigh-in-motion station to the bridge
Figure 5. Displacement Time History and Frequency Com­ of interest increases, the feasibility of tracking the
ponent obtained from the RealSense Sensor. truck needs to be evaluated considering the limitation
of UAV flight distance. The optimal distance needs to
be studied to see what type of range is achievable for
the proposed system.
Table 1. A Comparison between Displacement and To measure the planar movement of an object, an
Acceleration Measurements. artificial pattern must be implemented on an object.
When applying the technique in the real environ­
Ground Truth Frequency from Percent
Test Frequency RealSense Sensor Different
ment, the implementation of this speckle pattern
could become an arduous task negating the benefit
1 0.49 0.50 0.98% of the algorithm. Therefore, to provide a better
2 1.48 1.50 1.68% implementation in the future, other alternatives need
3 1.99 2.00 0.55% to be studied. Concrete and steel are more or less
4 1.76 1.76 -0.13% smooth surfaces that do not produce many SIFT
5 2.23 2.25 1.13% key-points. One possibility is to utilize multiple laser
projection sensors installed on multiple UAVs to
measure depth in different directions to provide the
3-D displacement measurements (i.e. measuring
with one artificial laser projection in each x, y, and
double integrating the accelerations, thus a large z-directions). This potential alternative will be inves­
error may be induced. However, the RealSense tigated. Some limitations with the RealSense
sensor measures the displacement directly. This is Camera need to be considered in the future investi­
an advantage of using the RealSense sensor over gation. Since the RealSense sensor was originally
accelerometers. developed for humans interacting with computers,
the working distance is rated to 10-meters. However,
through testing, it was shown that above a working
6 SUMMARY distance of about 3-meters the error of the measure­
ment becomes significant. Additionally, the infrared
In measuring the dynamics of bridges, the proposed laser and sensors do not work as effectively outside
framework and case studies show potential in meas­ in sunlight; the infrared speckle pattern becomes
uring the dynamic traffic loading and 3-D structural washed out with the radiation from the sun in out­
response. First, by using a MOSSE filter, a truck is door environments. However, working in shady
tracked from a weigh-in-motion sensor on the Inter­ environments outside of direct sunlight, the infrared
state Highway in the United States to the test bridge laser is still captured by the infrared camera. There­
which provides the real-time measurement of load­ fore, the promise of working under a bridge is still
ing of the structure. Next, 2-D planar movement viable. Moreover, either a more powerful or different
and 1-D depth movement measurement techniques laser type may perform better in outdoor environ­
are integrated to find the true 3-D displacement of ments. The idea behind the projection of an artificial
the structure. With both dynamic loading and laser-pattern does show promise to provide a robust
response known, the performance of the structural technique to measure depth. Therefore, the most
system can be better understood. This would allow appropriate laser type still needs to be studied and
for better condition assessment and ultimately examined.
284
ACKNOWLEDGEMENT Structural Health Monitoring of Structures Using Digital
Image Correlation (DIC). Structural Health Monitoring,
The work presented in this paper was conducted with Photogrammetry & DIC 6(February), 9.
support from Colorado State University and the Moun­ Gillins, M. N., D. T. Gillins, & C. Parrish (2016). Cost-
tain-Plains Consortium, a University Transportation Effective Bridge Safety Inspections Using Unmanned
Center funded by the U.S. Department of Transporta­ Aircraft Systems (UAS). In Geotechnical and Structural
tion. The contents of this paper reflect the views of the Engineering Congress 2016, Phoenix, AZ.
authors, who are responsible for the facts and accuracy Instron (2017). DIC Replay — 2D Digital Image Correl­
ation Software.
of the information presented. (FASTACT Grant No. Kalaitzakis, M., S. R. Kattil, N. Vitzilaios, D. Rizos, &
69A3551747108). Additionally, the authors would like M. Sutton (2019). Dynamic Structural Health Monitor­
to acknowledge the UAVs and assistance provided by ing using a DIC-enabled drone. In 2019 International
the Colorado State University’s Drone Center. Also, Conference on Unmanned Aircraft Systems (ICUAS),
the authors appreciate the experimental testing oppor­ Atlanta, GA, pp. 1–6. IEEE Xplore.
tunities provided by the Colorado State University’s Linderman, L. (2019). Displacement Monitoring of I-35W
researcher Dr. Paul Heyliger and his two graduate stu­ Bridge with Current Vibration-Based System. Technical
dents Lubna Al Ani and Mohammed Alkharisi. Report January, Minnesota Department of Transporta­
tion, Minneapolis, MN.
Lowe, D. (1999). Object recognition from local
REFERENCES scale-invariant features. In Proceedings of the Seventh
IEEE International Conference on Computer Vision,
Aoki, H., A. Suzuki, & T. Shiga (2018). Study on Kerkyra, Greece, pp. 1–8. IEEE.
Non-contact Heart Beat Measuremtn Method by Using Nguyen, C. H. P. & Y. Choi (2018). Comparison of point
Dpeth Sensor. World Congress on Medical Physics and cloud data and 3D CAD data for on-site dimensional
Biomedical Engineering 68(1), 341–345. inspection of industrial plant piping systems. Automa­
Bolme, D., R. Beveridge, B. Draper, & Y. M. Lui (2010). tion in Construction 91(February), 44–52.
Visual Object Tracking using Adaptive Correlation Filters. Wells, J. & B. Lovelace (2018). Improving the Quality of
Technical report, Colorado State University, Fort Collins. Bridge Inspections Using Unmanned Aircraft Systems
Catt, S., B. Fick, M. Hoskins, J. Praski, & J. Baqersad (UAS). Technical Report July, Minnesota Department of
(2019). Development of a Semi-autonomous Drone for Transportation, St. Paul, MN.

285
 Bridge Maintenance, Safety, Management, Life-Cycle Sustainability and Innovations –
Yokota & Frangopol (eds)
© 2021 Taylor & Francis Group, London, ISBN 978-0-367-23278-8

Tracking bridge condition over time using recurrent UAV-based inspection

B.J. Perry, Y. Guo, R. Atadero & J.W. van de Lindt
Department of Civil and Environmental Engineering, Colorado State University, Fort Collins, Colorado, USA

ABSTRACT: With bridge owners and managers tasked with making major maintenance/repair decisions with
inadequate funding and resources, there is a need to provide quantifiable metrics on the overall health of a bridge
that can help the decision-makers prioritize bridge maintenance/repair projects. One important metric indicating
the health of a structure is the changing rate of the structure condition over time. The condition rating from human
visual inspection in the current bridge inspection practice is not adequate for evaluating the changing rate, because
it is subjective in nature and severity and extent of defects is not rigorously delineated on the rating scale.
Recently, with advanced sensor and flight performance, as well as more affordable prices, unmanned aerial
vehicles (UAVs) have become a popular tool in infrastructure inspection practice. The enhanced controllability
with precise global-positioning-systems and inertial measurement units and increased safety with omni-directional
collision avoidance sensors has enabled UAVs to navigate in the otherwise inaccessible key spaces for bridge
inspection. In this study, a methodology is presented using UAVs to conduct recurrent inspections of a bridge
through flight missions. With reliable flight control applications coming into the market, the exact flight trajector­
ies and camera locations can be near-perfectly replicated during each flight. This provides a unique opportunity to
capture comparable images for the same area over time. Then a defect detection and quantification algorithm is
developed to quantify the progression of the defect over time using a set of images for a similar location. The
feasibility and efficacy of the proposed method is tested in an experimental case study of a split cylinder test.

1 INTRODUCTION researchers (Gillins, Gillins, & Parrish 2016, Seo,

Duque, & Wacker 2018, Wells & Lovelace 2018,
Traditional, human-based inspection techniques of Perry, Guo, Atadero, & W. van de Lindt 2019). UAV
bridges in the United States typically assign a qualita­ based remote sensing has shown promise to aid inspec­
tive rating to each component (i.e.deck, girders, and tors by easily accessing difficult-to-reach areas of
abutments), which is not explicitly delineated on a bridge while providing a safer platform as opposed
a rating scale. This qualitative rating leads to inspec­ to “snooper trucks” or roping rigs currently used (Her­
tions that tend to be more subjective. The final rating nandez 2016). With UAVs, inspectors are able to
of a bridge has been shown to be correlated with the easily fly the UAV around a bridge structure and col­
inspector’s experience or, in the case of chain-drag lect hundreds to thousands of images to later post-
tests, hearing ability (Moore, Phares, Graybeal, Rolan­ process. Also, utilizing UAVs alleviate the variation
der, & Washer 2001). Consequently, current bridge from inspectors and allows for more effective data col­
inspections often have more subjective results which lection. The larger challenge of implementing UAVs
can lead to a larger variance from inspector to for bridge inspection, however, is being able to extract
inspector. This variation of the results does not provide useful information quickly and easily. Manually
the most thorough information for inspectors and deci­ extracting pertinent information from big data collected
sion-makers. The incomplete data then contributes to by UAVs would require significantly more time and
a deadweight-cost with the inability to make the most workforce negating the benefits that a UAV-enabled
informed decisions and planning. In addition, under inspection would provide. Wells et al. developed
current inspection practices, large equipment is utilized a framework for a 3-D bridge model with geo­
for the difficult-to-reach areas, which is inherently dan­ referenced, localized images (Wells & Lovelace 2018).
gerous and time-consuming. A full inspection framework that takes the collected
The shortcomings of current bridge inspection prac­ images from a UAV and created a 3-D bridge informa­
tices have lead researchers to develop more effective tion model with element-wise damage information was
assessments of bridges. Therefore, UAV-enabled also created which highlights the potential of future
bridge inspection has been a focus for many implementation (Perry, Guo, Atadero, & W. van de

DOI: 10.1201/9780429279119-35
https://doi.org/10.1201/9780429279119-35

286
Lindt 2019). However, if these inspection frameworks combination with a variety of edge detectors to iden­
were to be transitioned from research into practice, tify cracks in concrete (Dorafshan, Thomas, &
a module that identifies the growth of the cracks over­ Maguire 2018, Hoang, Nguyen, & Tran 2018). Both
time needs to be developed. This would prove benefi­ studies were successful with a classification accuracy
cial to bridge owners or managers as cracks that are of 86% and 92%, respectively. However, both used
stagnant are often handled differently than cracks that large amounts of labeled training data. For example,
grow over time. The growth of a crack often indicates Dorafshan et al. used 3,420 images (319 images with
more significant problems in a structure. Moreover, if cracks and 3,101 images without defects) and Hoang
a bridge inspection is performed every two years with et al. used 400 images, 320 of which were labeled and
a UAV, there would be no need for additional data col­ used for training. Using CNNs on the images from the
lection to easily and objectively measure the growth of UAV flight is not possible without first labeling a large
a crack. set of images in a training and validation set. Gather­
This paper presents a computer vision-based tech­ ing this training data would negate the purpose of cre­
nique to identify and quantify the growth of cracks ating an automatic defect detection algorithm. Until
using images collected over time. First, a defect more robust data and/or algorithms are available,
detection algorithm that identifies the cracks will be CNNs are not the best option for defect detection.
discussed. Next, the proposed method to measure the Therefore, in this paper, morphological transforms that
change of a crack over time will be presented. require tuning several parameters, but perform well in
Lastly, a demonstration of the proposed method will a controlled environment will be implemented.
be discussed using data from a split tensile test of A new defect detection algorithm optimized for
concrete cylinders in the lab. the cracks found in images collected from the lab
experiment was developed. The open-source
OpenCV library was implemented in python to
2 CRACK DETECTION detect and highlight cracks (Bradski 2000). First,
pre-processing of the image was performed. Minor
Currently, the ratings of bridges in the United States color correction was done; the contrast was
mostly come from qualitative categorical ratings. increased while the brightness was reduced. Non-
Adding quantitative defect information to an inspec­ Local means denoising and a Gaussian blur were
tion report would help to alleviate some of the subject­ passed over the image. Next, the morphological
ivity within current practice; however, detecting and transform was executed. The top hat transformation
measuring cracks in concrete from images can be was performed on the image using a 3x3 square
challenging due to the low luminance and the thinness structural element as recommend by Sankarasrini­
of the cracks. In the literature, there are several studies vasan et al. (Sankarasrinivasan, Balasubramanian,
on crack detection for concrete surfaces that focus Karthik, Chandrasekar, & Gupta 2015). This cre­
mainly on using morphological transforms (Jahan­ ated a greyscale image where the white portions are
shahi & Masri 2012, Sankarasrinivasan, Balasubrama­ regions of greater contrast compared to adjacent
nian, Karthik, Chandrasekar, & Gupta 2015, Talab, pixels. The image was thresholded to create
Huang, Xi, & Haiming 2016). These systems perform a simple black and white image. A Canny Edge
well in controlled environments. For example, Jahan­ Detector was used to outline the white regions on
shahi et al. and Sankarasrinivasan et al. mainly the image, which represent the potential defects.
focused on identifying cracks present on exterior/inter­ Lastly, a minimum area of the outline was set, so
ior facades that typically have longer cracks. Talab that only the largest contours remained. This elim­
et al. focused on the formation of cracks in a lab set­ inated some outliers of smaller anomalies within
ting. Because the methods proposed in these studies the image. In this way, the developed technique
were tailored to their specific applications (i.e. exterior consistently detects and highlights the cracks
concrete facades and laboratory concrete beam tests), within an image.
they worked well in their respective environments but
lost generality in broader applications.
Recently, machine learning has made tremendous 3 GROWTH DETECTION
progress in image classification and recognition in
images and, thus, has become another alternative for With a defect detection algorithm developed and opti­
damage/defect identification (Zhang, Donahue, Gir­ mized for the typical images collected, the crack
shick, & Darrell 2014, Zhou, Park, & Koltun 2018, growth can be measured. The images collected at dif­
Girshick, Donahue, Darrell, & Malik 2014, Li,Wong, ferent times may have geometric distortions due to
Lo, & Xin 2018, Dorafshan, Thomas, & Maguire different camera angles and scales. Therefore, before
2018, Hoang, Nguyen, & Tran 2018). The most robust the identification of the crack growth is performed,
techniques rely on convolutional neural nets (CNNs) the initial image and the current image must be
which is a category of machine learning to identify aligned to be readily compared. To perform this align­
cracks; however, CNNs require a significant amount of ment, a scale-invariant feature transform (SIFT) algo­
training data to be successful. Dorafshan et al. and rithm is utilized. SIFT key-points and their related
Hoang et al. studied implementing CNNs in descriptors are identified and matched across the two
287
images. Once the pixel locations of the key-points are develop within the specimen. To capture this crack
found, an affine transformation, which is a linear growth and simulate real-world conditions, a DJI
mapping method typically used to correct for geomet­ Mavic 2 Pro and a DJI Mavic 2 Zoom were set up
ric distortions or deformations that occur with non- adjacent to the cylinders and used in a stationary pos­
ideal camera angles, is performed which warps the ition; the UAVs were unable to hover within the con­
current image to align the shape to the initial image. fined space of the testing set-up. The experimental
After the alignment, the crack detection algorithm setup is shown in Figure 1, where the two UAVs are
involving the morphological transform as described pointed in the direction of the universal testing
above is performed; however, the algorithm is stopped machine’s compression plates onto which the concrete
after the top hat transformation with the 3x3 square cylinder will be placed. Two UAVs were used to
structural element. This transform identifies the regions simulate two flights with multiple camera angles and
of high and low contrast and creates a grey-scale different but similar camera technologies. The split
image. To find the growth of the crack, the grey-scale cylinder concrete test provided slow, controlled crack
morphological images are subtracted from one another growth in the concrete cylinder that could easily be
to find the remaining transforms (i.e. the difference recorded and pictured with a typical 60-fps camera.
between the two images). Next, a Gaussian blur is con­ This provided the data to demonstrate the proposed
voluted over the images. The difference between the method to identify the growth of cracks over time.
two images identifies the growth of cracks. Once the Video of the test was taken and frames from the video
difference is found, the thresholding and outlining of were extracted to simulate the repeated flights of
the difference image is performed as described above a UAV around a bridge over time. The results will be
to highlight the areas of crack growth. discussed in the following.

4 DATA COLLECTION 5 RESULTS

The collection of images of actual crack growth on With the experimental data collected from the split
bridges is difficult to obtain due to the lack of data cylinder testing, the experimental results are exam­
sets and the slow growth of cracks. Therefore, to ined. Since the number of test cylinders was small,
obtain the data to test and develop the algorithm, lab only two result data sets are available. The proposed
experiments were performed. Split cylinder tests were method is validated using these two data sets. The
performed and recorded with two UAVs. The lab first test used the same camera at the same angle.
experiment followed ASTM C496. In a split cylinder Beginning with the test, Figure 2 and Figure 3 are
test, a concrete cylinder is laid on its side and the two images collected by the UAV over time.
a compressive force is applied uniformly over the two They show growth of about 10 cracks on the side of
points of contact with a universal testing machine. the cylinder. After performing the affine transform
This compressive force creates tension in the center and the morphological transformation, and finding
of the cylinder which causes a crack to gradually the difference as explained above, Figure 4 shows the

Figure 1. Set Up of Experimental Data Collection.

288
 Next, the second test is conducted. In this
test, the initial image was still obtained from the
first camera, while the second camera was used
to collect the image with the progressed cracks
over time. Using images from two cameras is to
simulate data collected from two different UAV
surveys. The proposed method was then applied
to identify the growth of cracks over time using
images taken from different camera angles. Fig­
ures 5 and 6 show the initial image and the cur­
rent image. Again, to align the two images taken
at differing angles, key-points were identified
and matched and an affine transform was per­
formed. Then, the morphological transformation
was performed as described to find the differ­
ence. Figure 7 shows the results of the identified
Figure 2. Time 1 Image of First Test. crack growth.

Figure 3. Time 2 Image of First Test with Growth of

Cracks Shown.

Figure 5. Time 1 Image of Second Test.

Figure 4. The Growth of the Cracks Highlighted over

Time.

results. It is seen that the proposed method is able to

identify and highlight the changes of the cracks Figure 6. Time 2 Image of Second Test with Growth of
between the two images taken at different times. Cracks Shown.

289
 as UAV navigation develops, there is a prospect
of creating a fully autonomous flight system for
UAVs. Autonomous flight not only has the benefit
of more efficient surveys, but it could near-
perfectly replicate flight paths. With a replicated
flight path from inspection to inspection, the col­
lected images would be more similar and the pro­
posed method would be better and more robust
due to better alignment of the images. Consumer-
level UAV flight plan applications currently do
not provide the level of detail needed to suffi­
ciently plan for the complex flight paths required
for bridge surveying. Once the more sophisticated
control applications that allow elevation controls
become available in the future, the performance
Figure 7. The Growth of the Cracks Highlighted over of the proposed method can be further tested on
Time. images collected from the fully autonomous
flights. In addition, with an increase in dimen­
sional data (i.e. displaying a 3-D structure over
time), there needs to be a user-friendly manner to
6 CONCLUSIONS
display the information which helps inspectors
and decision-makers easily understand and inter­
With more automated assessments of infrastructure
pret the results.
using remote sensing technologies such as UAVs
coming into research and eventually practice, big
data can be collected in a very short time period. ACKNOWLEDGEMENT
Thus, more efficient data processing tools that can
expedite the extraction of information from the data The work presented in this paper was conducted
are very critical to reduce the burden of the other­ with support from Colorado State University (CSU)
wise manual data analysis. In this particular applica­ and the Mountain-Plains Consortium, a University
tion for monitoring the changes of structural Transportation Center funded by the U.S. Depart­
condition, there is a need to develop a method which ment of Transportation (FASTACT Grant No.
automatically detects the growth of concrete cracks 69A3551747108). The contents of this paper reflect
over time from images collected at differing angles the views of the authors, who are responsible for the
and scales. The proposed method addresses this need facts and accuracy of the information presented.
and provides not only a more streamlined transition Additionally, the authors would like to acknowledge
from research to real-world application but also the assistance and UAVs provided by the CSU’s
a more robust, quantitative information for bridge Drone Center and experimental support provided by
inspectors and decision-makers to use for future CSU’s lab manager Mr. Todd Atadero.
bridge maintenance and repairs plans. The proposed
algorithm aligned two images over time using SIFT
and affine transformation and then implemented REFERENCES
a morphological transformation over the two time-
related images. The transformed images at the initial Bradski, G. (2000). The OpenCV Library.
time point were then subtracted from its counterpart Dorafshan, S., R. J. Thomas, & M. Maguire (2018). Com­
at a later time point and the result was assessed to parison of deep convolutional neural networks and edge
find the growth of cracks over time. The proposed detectors for image-based crack detection in concrete.
Construction and Building Materials 186, 1031–1045.
method provided satisfactory results highlighting the Gillins, M. N., D. T. Gillins, & C. Parrish (2016). Cost-
growth of the cracks over time. Effective Bridge Safety Inspections Using Unmanned
Aircraft Systems (UAS). In Geotechnical and Structural
Engineering Congress 2016, Phoenix, AZ.
7 FUTURE WORK Girshick, R., J. Donahue, T. Darrell, & J. Malik (2014).
Rich Feature Hierarchies for Accurate Object Detection
The proposed method for identifying growth of and Semantic Segmentation. In 2014 IEEE Conference
cracks over time was presented with the promise on Computer Vision and Pattern Recognition, pp.
of future implementation. However, the method 580–587.
Hernandez, I. (2016). Overcoming the Challenges of Using
was not demonstrated on a real bridge. There Unmanned Aircraft for Bridge Inspections. Ph.
could be additional challenges involved when D. thesis, University of Missouri-Kansas City.
transitioning from lab tests to real-world imple­ Hoang, N.-D., Q. L. Nguyen, & V. D. Tran (2018). Auto­
mentation. For instance, the crack growth may be matic recognition of asphalt pavement cracks using
less dramatic than seen in the lab tests. Moreover, metaheuristic optimized edge detection algorithms and

290
 convolution neural network. Automation in Construction Sankarasrinivasan, S., E. Balasubramanian, K. Karthik,
94(June), 203–213. U. Chandrasekar, & R. Gupta (2015). Health Monitoring
Jahanshahi, M. R. & S. F. Masri (2012). Adaptive of Civil Structures with Integrated UAV and Image Pro­
vision-based crack detection using 3D scene reconstruc­ cessing System. Procedia Computer Science 54, 508–515.
tion for condition assessment of structures. Automation Seo, J., L. Duque, & J. P. Wacker (2018). Field Application
in Construction 22, 567–576. of UAS-Based Bridge Inspection. Transportation
Li, J., H. C. Wong, S. L. Lo, & Y. Xin (2018). Multiple Research Record 2672(12), 72–81.
Object Detection by a Deformable Part-Based Model Talab, A. M. A., Z. Huang, F. Xi, & L. Haiming (2016).
and an R-CNN. IEEE Signal Processing Letters 25(2), Detection crack in image using Otsu method and mul­
288–292. tiple filtering in image processing techniques. Optik 127
Moore, M., B. Phares, B. Graybeal, D. Rolander, & (3), 1030–1033.
G. Washer (2001). Reliability of Visual Inspection for Wells, J. & B. Lovelace (2018). Improving the Quality of
Highway Bridges. Technical Report FHWA-RD-01-020, Bridge Inspections Using Unmanned Aircraft Systems
FHWA, Atlanta, GA. (UAS). Technical Report July, Minnesota Department of
Perry, B., Y. Guo, R. Atadero, & J. W. van de Lindt (2019). Transportation, St. Paul, MN.
Unmanned Aerial Vehicle (UAV) Enabled Building Zhang, N., J. Donahue, R. Girshick, & T. Darrell (2014).
Information Modeling for Bridge Inspection. In 9th Part-Based R-CNNs for Fine-Grained Category
International Conference on Structural Health Monitor­ Detection. ECCV 2014 1, 834–849.
ing of Intelligent Infrastructure, St. Lois, MO, pp. 1–6. Zhou, Q.-y., J. Park, & V. Koltun (2018). Open3D:
Sensors. A Modern Library for 3D Data Processing.

291
 Bridge Maintenance, Safety, Management, Life-Cycle Sustainability and Innovations –
Yokota & Frangopol (eds)
© 2021 Taylor & Francis Group, London, ISBN 978-0-367-23278-8

Estimating bridge characteristics with only situation characteristics using

Bayesian networks
V. Panopoulos
Rhomberg Bahntechnik AG, Zurich, Switzerland

A. Bougas
Atkins Danmark A/S, Esbjerg, Denmark

B. Garcia de Soto
New York University, Abu Dhabi, United Arab Emirates

B.T. Adey
Swiss Federal Institute of Technology, Zurich, Switzerland

ABSTRACT: When deciding where a new road or rail line should be built, it is useful to have some idea as
to the characteristics of the bridges to be built. One way to do this is to have engineers develop preliminary
designs for the bridges that would likely be included in each route. The developed preliminary design, then,
stands as a basis for cost estimation and the detailed design of the structure. This would, of course, require
a significant amount of time and effort. Another way to do this would be to exploit existing data on the types
of bridges built in different situations using Bayesian networks. In this article, it is shown how this can be
done using the information in existing databases to estimate bridge characteristics, knowing only the values of
a number of situational characteristics, e.g., span to be covered, what the bridge carries, and what the bridge
traverses.

1 INTRODUCTION 2 METHODOLOGY

Although considerable research has been con­ 2.1 Data pre-processing

ducted to improve estimations in civil engineering
in general (Deublein et al. 2015, García de Soto The dataset used consisted of entries for single
et al. 2014), and bridges specifically, no one has bridges. The dataset consisted of 1,793 instances with
yet developed and tested a methodology for pre­ 17 different variables of interest, both categorical and
dicting future bridge characteristics based only on numerical. Not all variables, however, could be used.
situational information. Based on an overview of The variables removed were done so because they
literature, Bayesian networks, with their ability to 1. played at best only a minor role in the bridge
represent relationships between variables (Khoda­ design process, e.g. bridge identity, bridge cross­
karami and Abdi 2014), and their potential to ing angle, slope, transverse inclination, curvature,
model the causal relationships between the design and the minimum superstructure height, as their
parameters and design characteristics in bridge difference with the respective maximum height
design (Matthews 2008), seem, however, particu­ was –if not negligible– smaller than 40 cm in
larly suitable. With this in mind, the research 90% of the cases.
presented in this paper was focused on the use of 2. did not have a sufficient number of entries, e.g.
Bayesian networks to estimate bridge characteris­ the superstructure construction process for which
tics using an existing bridge date base and only 50% of the bridges had values.
knowing only the values of a number of charac­ 3. were strongly correlated with other variables, e.g.
teristics, such as the span to be covered, the the static system of the bridge which was strongly
functionality of the bridge, and what the bridge correlated with the bridge type.
traverses.

DOI: 10.1201/9780429279119-36
https://doi.org/10.1201/9780429279119-36

292
4. were adequately represented through grouped Table 1. Variables used for model development and their
variables, e.g., bridges with a cross sections discrete states.
denoted as “Single-celled hollow box”, “Single­
celled hollow box, walkable”, “Single-celled Variable name Related question Discrete states
hollow box, crawlable”, “Single-celled hollow
box, non-walkable, non-crawlable” and “Super­ Operation How is the bridge Autobahn
structure as a hollow box”, were all considered to used? Federal Highway
Other street
be “Hollow”. Cycling path
After removing the unnecessary variables, the District road
Hartemink’s discretization algorithm (Hartemink State road
2001) principled was used to convert the continuous Underneath What is below the Highway
variables X ¼ Xi ; i ¼ 1; …; N into N discrete vari­ bridge? Typical Street
ables, each with k2 levels, preserving as much River
Small Intervention
mutual information between variables as possible.
Train
This resulted in the variables shown in Table 1.
Bridge_length How long is the [1.6;4.28]
A Cramer’s V association diagram (Cramér bridge (m)? (4.28;8.26]
1999) (Figure 1) and an unidirectional correlation (8.26;14.5]
graph, which connects the pairs of variables that (14.5;45.5]
have a cor-relation of at least 0.40 (Figure 2), were (45.5;1.02e+03]
used to identify the possible correlations between Pier_height How high are the [0;1]
the variables (Table 1). Figure 2 is an undirected piers (m)? (1;3]
network where the edges between the nodes are (3;4.6]
a direct result from the correlation between the vari­ (4.6;6.5]
ables of the dataset. (6.5;78.6]
This information was used as a guide to identify­ Span_No How many spans are 1 span
ing causal, and not only correlative, relations that there? 2 spans
exist between the variables. For example, the varia­ > 2 spans
ble”Underneath” was connected with variable”­ Bridge_type What type of bridge Beam bridge
is it? Arch bridge
Span_no”, which complies with engineering
Frame bridge
intuition and knowledge, as the number of bridge Continuous Girder
spans is influenced by what was underneath the Max_span_pier How long is the max­ [0.7;3.4]
bridge, i.e., a road, a river, etc. In other words, imum bridge span (3.4;7.78]
the”Underneath” variable has a cause-effect relation (m)? (7.78;13]
with”Span_no” variable. The variable”Underneath” (13;22.9]
was not directly related to the bridge cross-section, (22.9;171]
a fact that can be logically interpreted, as “Cross_­ Cross_section What type of cross- Single girder
section” variable was more influenced by the varia­ section does it have? Hollow
bles”Bridge_type” and”Max_span_pier”. Other
Multiple girders
2.2 Model formation Sup_height How high it the (0;0.5]
bridge super­ (0.5;0.76]
The model formulation consisted of two parts, 1) the structure (m)? (0.76;0.9]
model structure, and 2) the parameters. To determine (0.9;1.35]
the optimal model structure, expert opinion coupled (1.35;19.8]
with consideration of the design process was used.
The structure was then compared with the results of
the model structure determined using seven structure
learning algorithms.
The exact interactions between the model variables
in the Bayesian network are depicted in Figure 3, internal variables, i.e. “Bridge Type”, “Underneath”,
where the arcs denote the direct causal influences “Operation”, “Span No”, “Cross section”, and “Sup
between the nodes. The nodes represent variables height”.
external to the bridge design process, i.e. “Operation”, Although other models are possible, this one
“Underneath”, dependent solely on the situational exploits the most significant conditional dependen­
variables, i.e. “Bridge length”, and “Max span pier”, cies between the available variables, i.e. the depend­
dependent on both internal and external variables, encies illustrated in Figure 2 closely resemble those
i.e., “Span No”, and “Pier height”, and dependent on in Figure 3. To learn the structure of the network

293
 Table 2. Structure-learning algorithms used.

No. Name Reference Type

1 Grow-Shrink (Margaritis 2003) Constraint

based
2 Incremental (Tsamardinos et al.
Association 2003)
3 Fast Incremental (Yaramakala and
Association Margaritis 2005)
4 Interleaved Incre- (Tsamardinos et al.
mental Association 2003)
5 Hill-Climbing (Bouckaert 1995) Score
based
6 Tabu-search (Bouckaert 1995)
7 Max-Min Hill- (Tsamardinos et al. hybrid
Climbing 2006)

Figure 1. Dependencies between the variables of the data­

Figure 4 shows the networks learned using con­
set, expressed through pairwise Cramer’s V association of straint-based algorithms. Despite the validity and the
the categorical variables of the processed data. robustness of these algorithms, the obtained results
are far from satisfactory and are in contrast to the
interdependencies based on the unidirectional correl­
ation graph in Figure 2, as well as, the expert model
in Figure 3. These four learned structures show that
there are variables that are independent of all the
others in the network and, therefore, should be elim­
inated. In other words, there are nodes in all cases in
Figure 4 that are not connected with the others.
Figure 5 shows the learned networks based on the
score-based algorithms. The obtained networks
match better to the graph structure of the expert
model, including interdependencies that are mostly
reasonable. For example, bridge length plays an
important role in the estimation of almost all other
variables in both these networks. Differences, how­
ever, can still be observed. For example, the vari­
Figure 2. Undirected correlation graph, connecting the ables “Operation” and “Underneath” are considered
pairs of variables having a correlation of at least 0.40. here to be dependent with one another in both
graphs, a relationship which in the real world does
not make sense.
Figure 6 shows the learned network developed
using the Max-min climbing hybrid algorithm. It can
be seen that the variables “Pier_height”, “Operation”
and “Underneath” are considered as independent
from all the other variables and are excluded from
the learned network graph.
In addition it was tried to create a de-noised con­
sensus network produced from multiple single net­
works. It can result in a better predictive
performance than choosing a single, high–scoring
Figure 3. Expert model. network (Claeskens and Hjort 2008) (Figure 7).
The structure of the network was learned by repeat­
ing Hill-Climbing structure learning several times (i.e.,
only based on the available data, four constraint- bootstrapping) with different initializations. In this
based algorithms, two score-based algorithms and way, a larger number of networks could be explored in
a hybrid algorithm were used (Table 2). The different an effort to reduce the effect of locally optimal (but
network structures were generated using the globally suboptimal) networks on learning and subse­
“bnlearn” R-package (Scutari 2010). quent inference. The averaged network was created
using the arcs present in at least a specified ratio of the
294
 Figure 5. Learned networks through the score-based learn­
ing algorithms.

Figure 6. Learned network through the hybrid learning

algorithm: Max – min Hill – Climbing.

Figure 4. Learned networks through the constraint-based

learning algorithms.

networks. This ratio measures the strength of each arc

and provides the means to establish its significance
given a threshold. The critical threshold was computed
as a minimization problem of the Euclidian norm (L1
norm) (Nagarajan et al. 2013) (Eq. 1). Figure 7. Model averaging network.

295
 avoid overfitting the model on the data set and helped
estimate the parameters with no data. For example, in
Figure 8, where is no data indicating a continuous
girder bridge with a length between 1.6 m and 4.28 m,
Where is the cumulative distribution func­ a uniform distribution is assumed instead of assigning
tion of the computed arc strengths (empirical CDF) posterior probabilities of 0 to the cross section states.
and is the cumulative distribution function
of the ideal network (ideal CDF), in which the t frac­
2.3 Model inference
tion of the elements of equal to zero and the rest
equal to one. In order to learn the model, approximate inference
The critical threshold is a measure of the fraction was performed using both logic sampling and likeli­
of non-significant edges. At the same time, t provides hood weighting, and their accuracy and required
a threshold for separating the elements of . The solution sample size were compared. Figure 9 shows
identification of significant edges can be thought of the estimated probability of the bridge type being
either as a least absolute deviations estimation or an a “Continuous Girder”, given that the number of
L1 approximation (Eq. 2). For the bridge network, spans is greater than two, using increasing sample
the value calculated for the significance threshold is sizes (particles), where 20 simulations were run for
0.444. It can be seen that most of the interdependen­ each of sample size (ranging from 5; 000 to 100; 000
cies are the same as in the expert network. particles in increments of 5; 000).
It can be seen that both inference methods con­
verged to roughly the same accuracy with large
sample sizes (large number of particles). Likelihood-
weighting, however, converged more quickly than
logic sampling. The most accurate estimate was
In conclusion, not all of the networks developed found with a sample size of only 5; 000. Addition­
using learning algorithms can be used for predicting ally, logic sampling exhibited relatively large dis­
bridge characteristics given the existing dataset. The crepancies with variations in the sample size,
constraint-based algorithms resulted in networks that having, and their mean value initially deviates from
excluded many of the selected variables, so they the true value of the conditioned event. The likeli­
cannot be used to predict bridge characteristics. The hood weighting algorithm was, therefore, used for
learned networks based on the score-based algo­ all the subsequent calculations.
rithms, as well as the model averaging network gen­ An example using the variables bridge type
erated by averaging a score-based algorithm resulted (“Bridge_type”), cross-section (“Cross_section”)
in structures similar to that of the expert network. and superstructure height (“Sup_height”), which are
The expert network, however, has the additional most often subjected to change during bridge design
benefit that it is easily interpretable. and dimensioning, is given in Table 3, which shows
Bayesian estimation was used for parameter learn­
ing assuming a uniform prior of the probabilities.
Hyperparameter a of the prior was chosen equal to 1.
An example of the conditional probabilities for the
expert network for the cross-section variable is
shown in Figure 8.
It was assumed that there was one additional
imaginary observation for each state. This helped

Figure 8. Conditional probability of the “Cross_section”

variable calculated through Bayesian estimation for the Figure 9. Probability of Bridge_type == “Continuous_Gir­
Expert network. According to this, the parents of the der” conditioned on evidence: Span_No = “> 2 spans” (a)
“Cross_section” are “Bridge_type” and “Bridge_length”. using logic sampling, and (b) using likelihood weighting.

296
Table 3. Best sampling results of the Expert network for Table 4. Posterior classification error for the variables
Maximum span between piers = (13; 22.9) [m], Bridge “Bridge_type”, “Cross_section”, “Sup_height” among the
length = (14.5; 45.5] [m], Number of spans = 1, Maximum generated networks.
pier height = (4.6; 6.5] [m].
Network Bridge_type Cross_section Sup_height
Cross- Superstructure
ID Bridge type section height [m] Probability Proposed 0.216879 0.1351889 0.3532140
Hill – Climbing 0.2259775 0.1338635 0.3611663
1 Continuous Multiple (0.9; 1.35] 41.68% Tabu Search 0.2226640 0.1345262 0.361425
Girder girders Averaged 0.2319417 0.1338635 0.3664679
2 Beam Single (0.9; 1.35] 14.28%
bridge girder
3 Beam Single (0.76; 0.9] 7.17%
bridge girder
4 Arch bridge Other (0; 0.5] 3.53% Additionally, the prediction error from the differ­
5 Frame Multiple (1.35; 19.8] 3.42% ent networks was compared to a naïve predictive
bridge girders model, which uses only the database as a predicting
tool. The naïve predictive model has no graphical
representation of the variables and merely classifies
the variable of interest utilizing only the database by
the five most likely combinations for the specific set the elimination of all the non-matching cases,
of evidence and their probability of occurrence. according to Figure 10. For example, the “Cross_­
From Table 3, it can be seen that if the maximum section” variable was predicted according to the
span between the piers is between 13 and 22.9 m, level, which yielded the highest probability in the
the bridge length is between 14.5 and 45.5 m, the database for a specific combination of the levels of
number of spans is 1 and the maximum pier height is all the possible evidence.
between 4.6 and 6.5 m, the most likely bridge is
a continuous girder bridge with a multiple girder
cross-section and a superstructure height between
0.9 and 1.35 m with a probability of 41.68%. If the
information in the network were sparse, then a small
correlation between the variables would be observed. The naïve model was compared with the afore­
This would result in almost random transitions from mentioned ones using 10-fold cross-validation to
one variable to another, and therefore in possible determine the prediction error for the “cross_section”
design combinations that would have the same prob­ variable. The naïve model resulted in a prediction
ability. This is not the case in the expert model, as error for the “cross_section” variable equal to 0.335.
the assumptions of correlation and causation are sat­ This means that if one would try predicting the type
isfied. Thus, the obtained results from this Maximum of the cross-section only by using the data set and
a posteriori query are meaningful. eliminating all the instances for which the evidence
does not match exactly, then one would have on
average 33.5% wrong classifications. This error rate
2.4 Model validation is significantly higher than the result yielded from
To validate the expert model, a 10-fold cross- most of the graph structures. Specifically, the predict­
validation was performed. The data was split into 10 ive capability of this model is 67.5% accurate for the
equally sized random sets and the learning algo­ variable “cross_section”,whereas all the other Bayes­
rithms were trained on a subset of 90% of the ian network models showed an increased accuracy
instances and then tested on the remaining 10% up to 86.5% for the same variable of interest.
(≈150 instances). 10 iterations were performed.
Table 4 shows the mean posterior classification
error for predicting the variables “Bridge_type”,
“Cross_section” and “Sup_height” among the best
generated networks (Tabu search, Hill-Climbing,
Model averaging) by deploying 10-fold cross-
validation, and the expert network. It can be seen that
the expert model has the lowest classification error
for predicting the variables “Bridge_type” and
“Sup_height”, whereas its classification error regard­
ing the “Cross_section” variable is the lowest among
these network structures. This means that the expert
model is not only sensible in structure, but also gen­ Figure 10. 10-fold cross-validation of learning algorithms.
erates not only similar, but superior results in two of The Figure depicts the Log-loss values for each network
the three cases when compared to the other networks. structure.

297
3 CONCLUSION Deublein, M., M. Schubert, B. T. Adey, & B. García de
Soto (2015). A bayesian network model to predict acci­
In this study, we presented how it is possible to use dents on swiss highways. Infrastructure Asset Manage­
prior knowledge (i.e., stored data) to estimate the ment 2(4), 145–158.
design aspects of future bridges (e.g., bridge type, García Soto, B., B. T. Adey, & D. Fernando (2014).
deck, number of pies, general dimensions, etc.) using A process for the development and evaluation of prelim­
Bayesian Networks. The methodology was tested by inary construction material quantity estimation models
using backward elimination regression and neural
using the data within an existing bridge database, pre­ networks. Journal of Cost Analysis and Parametrics 7
processing the entries and building different discrete (3), 180–218.
Bayesian Networks. Various net-works, both from Hartemink, A. J. (2001). Principled computational methods
structure learning algorithms and from engineering for the validation discovery of genetic regulatory net­
experience, were investigated and compared with works. Ph. D. thesis, Massachusetts Institute of
respect to unbiased criteria. As long as the structure Technology.
was set and the parameters of the structure were Khodakarami, V. & A. Abdi (2014). Project cost risk ana­
learned, a general inference could be conducted for lysis: A bayesian networks approach for modeling
dependencies between cost items. International Journal
any attribute and observed variables. Despite its simpli­
of Project Management 32(7), 1233–1245.
city, it was shown that the expert model provided rela­ Margaritis, D. (2003). Learning bayesian network model
tively accurate results that comply with engineering structure from data. Technical report, Carnegie-
knowledge. Notwithstanding their dissimilarities, the Mellon Univ Pittsburgh Pa School of Computer
network structure representing the causal relationships Science.
behaves very well, out-performing in two of the three Matthews, P. C. (2008). A bayesian support tool for mor­
cases, the learned non-causal structures by a better pre­ phological design. Advanced Engineering Informatics
diction error. Finally, the prediction capability of the 22(2), 236–253.
network reached an accuracy of up to 86.5% for the Nagarajan, R., M. Scutari, & S. Lèbre (2013). Bayesian
single variable regarding the cross section type of the networks in r. Springer 122, 125–127.
Scutari, M. (2010). Learning bayesian networks with the
bridge, whereas naïve search models resulted in 66.5% bnlearn R package. Journal of Statistical Software 35
accuracy. (3), 1–22.
Tsamardinos, I., C. F. Aliferis, A. R. Statnikov, &
E. Statnikov (2003). Algorithms for large scale markov
REFERENCES blanket discovery. In FLAIRS conference, Volume 2, pp.
376–380.
Bouckaert, R. R. (1995). Bayesian belief networks: from
Tsamardinos, I., L. E. Brown, & C. F. Aliferis (2006). The
construction to inference. Ph. D. thesis.
max-min hill-climbing bayesian network structure learn­
Claeskens, G. & N. L. Hjort (2008). Model selection and
ing algorithm. Machine learning 65(1), 31–78.
model averaging. Technical report, Cambridge Univer­
Yaramakala, S. & D. Margaritis (2005). Speculative
sity Press.
markov blanket discovery for optimal feature selection.
Cramér, H. (1999). Mathematical methods of statistics,
In Fifth IEEE International Conference on Data Mining
Volume 9. Princeton university press.
(ICDM’05), pp. 4–pp. IEEE.

298
 Bridge Maintenance, Safety, Management, Life-Cycle Sustainability and Innovations –
Yokota & Frangopol (eds)
© 2021 Taylor & Francis Group, London, ISBN 978-0-367-23278-8

Time series forecasting to jointly model bridge responses

O. Bahrami, R. Hou, W. Wang & J.P. Lynch
Department of Civil and Environmental Engineering, University of Michigan, Ann Arbor, USA

ABSTRACT: Structural health monitoring (SHM) strategies often rely on data acquired from a single
bridge seeing a broad range of environmental and operational loads over time. However, the emergence of
wireless connectivity now allows multiple bridges in the same highway corridor to be monitored and their
response to the same truck linked. In this work, time series forecast models are explored as a tool for jointly
modeling the response of bridges on the same corridor and monitored under identical load conditions. More
specifically, two types of models are built for the task at hand. First, an encoder-decoder architecture with two
different cell types, namely gated recurrent unit (GRU) and long short-term memory (LSTM) are explored.
The second type is the autoregressive with exogenous inputs (ARX) model. To evaluate the performance of
the forecast models in taking the output of one bridge to predict the response of another under the same truck
load, finite element models are built for two real-world bridges and a simulation dataset containing 2,100
pairs of bridge responses to the same truck load created. The two encoder-decoder models provided accurate
prediction capabilities of bridge response compared to the ARX baseline approach.

1 INTRODUCTION can be identified from inaccurate. Hence, this model

enables joint health monitoring among multiple
Recent advances in deep learning are providing the bridges located on the same corridor (for which there
civil engineering community with powerful and is a high likelihood that the same truck loads are
valuable tools to solve unique problems emerging experienced). Another application of joint modeling
within the field. Applications of such techniques of bridges would be enabled by connected vehicles
range from the use of computer vision for infrastruc­ technology. The response of an upstream bridge to
ture inspection (Spencer Jr et al. 2019) to damage a truck contains information on the characteristics of
detection to structural health monitoring using meas­ the truck (i.e., weight). The joint model between the
ured structural responses(Salehi & Burgueno 2018). two bridges can be employed to predict truck impacts
This paper builds upon a recent successful applica­ on downstream bridges so that the speed of the truck
tion of deep learning in civil engineering of computer approaching the second bridge can be set to minimize
vision used to track trucks in a corridor with bridges the response and consumed life of that bridge (and
monitored (Hou et al. 2019). In the work research, thus maximize the life span of the bridge). This
the traffic within a highway corridor is monitored via requires the models to be exclusive to each bridge
cameras. A deep learning technique is used to auto­ pair and take into account the unique complexities
matically identify trucks traveling on a bridge in the inherent to the bridges and the truck (e.g., suspension
corridor and then re-identify the same truck as it tra­ system). It is almost impossible to do this with phys­
verses another bridge in the same corridor. Simultan­ ical models of the bridges. Hence data-driven models
eously, the structural response of the bridges to are an attractively scalable solution.
moving trucks are measured and synchronized to the In this work, the use of time series forecast models
same truck load (as confirmed by imagery). are explored for the aforementioned task of predicting
In this work a data-driven model is built that bridge responses. Two types of models, namely
enables forecasting the response of one bridge to sequence to sequence (Seq2Seq) models and ARX
a truck using that of another bridge whose response models are explored. Seq2Seq models are originated
corresponds to the same truck. Such a model can be from the field of neural machine translation (Sutskever
used for multiples purposes such as structural et al. 2014) but have been proven to be effective in
damage detection. For example, by building an accur­ many other applications such as time series forecasting
ate model that is capable of forecasting the response (Wilms et al. 2018, Salinas et al. 2019) and anomaly
of one bridge using that of another, structural damage detection (Malhotra et al. 2016). Moreover, ARX

DOI: 10.1201/9780429279119-37
https://doi.org/10.1201/9780429279119-37

299
models use a linear equation to predict the output of 2019). The corridor has a set of traffic cameras
a system using previous outputs and previous and the installed along the corridor including on two bridges,
current inputs (Ljung 1999) . Such a time series struc­ namely the Telegraph Road Bridge (TRB) and the
ture is suitable for the purpose of this work and pro­ Newburgh Road Bridge (NRB) and of a weigh-in­
vides a linear twin for the non-linear Seq2Seq models. motion system (WIMS) system. Trucks are identified
The remainder of this paper is organized as fol­ on the upstream TRB bridge to trigger reidentifica­
lows: the next section is devoted to the description tion of the same truck on the downstream NRB
of the dataset used within this work. In Section 3, bridge and weigh-in-motion (WIM) station. The TRB
potential Seq2Seq models are described and the and NRB both have extensive SHM systems installed
results are presented in Section 4. This paper con­ to collect bridge responses to trucks. Figure 1 shows
cludes with a summary of the results and the highway corridor and the monitoring system.
a description of future work. In this study, finite element (FE) models are cre­
ated for both TRB and NRB in CSiBridge (Com­
puters and Structures 2011) according to the
2 DESCRIPTION OF THE DATASET dimensions and properties of the bridges to evaluate
the feasibility of the proposed Seq2Seq model.
The subject of this study is the I-275 corridor previ­ Numerical simulations are carried out for both
ously studied by (Hou et al. 2019) and (Jeong et al. bridges using the same truck to imitate the scenario

Figure 1. I-275 Highway corridor and encoder-decoder architecture.

300
of monitoring a given truck passing through the
highway corridor loading the TRB and NRB.
For TRB, a total of 9,052 shell elements are used
to model the reinforced concrete slab (1,750) and the
steel girders (7,302). Lateral bracings and rebars are
modeled using 454 brace elements. The pin-and­
hanger mechanism and joints are modeled by link
elements. The concrete has a Young’s modulus of
3,600 ksi with a Poisson’s ratio of 0.17, the steel
used for girders and rebars has a Young’s modulus
of 29,000 ksi with a Poisson’s ratio of 0.3 and the
link plate steel had a Young’s modulus of 30,000 ksi.
The dominant modal frequency of TRB is 2.17 Hz.
For NRB, a total of 12,139 shell elements are
used to model the reinforced slab (3,840) and steel
girders (8,299), and a total of 1,020 brace elements
are used to model the lateral bracings. The concrete
has a Young’s modulus of 3,155 ksi with a Poisson’s
ratio of 0.2. The girder steel has a Young’s modulus
of 27,000 ksi and a Poisson’s ratio of 0.28. The
Young’s modulus for the rebar steel is 29,000 ksi. Figure 2. The CSiBridge model for: (a) TRB and (b) NRB.
Unlike the TRB, NRB has two dominant modal fre­
quencies: 4.0Hz and 5.8Hz. The model properties
for both bridges have been fine tuned by model lane on NRB (the “slow” lane on TRB is actually an
updating, thus, their modal frequencies closely on-ramp lane).
match those observed in the actual bridges. The simulation has a time step of 0.01 seconds
For the truck load properties, the WIM station (sampling frequency is 100 Hz). A sample simulation
located on the same corridor measures nine truck attri­ for TRB is shown in Figure 3. As can be seen, the
butes including: vehicle class based on Federal High­ strain response contains not only the forced response
way Administration (FHWA) classification, vehicle triggered by truck but also the free vibration response
speed, number of axles, axle weights and spacings, after the truck is off the bridge. Thus, each bridge
and the passing lane. Considering vehicle classes 1 to response can be divided into two parts: (1) forced
3 represents light weight vehicles, whose loads on the response whose duration and amplitude depend on
bridges are not of major interest. In addition, according truck speed and weights, and (2) free vibration
to the local speed limit (70 mph for cars, and 60 mph response after the truck is off. To capture enough
for trucks), measurements with abnormal speeds are structural information of the bridges, the simulation
ignored. Thus, the measured WIM dataset is pruned to duration is set to 5 and 3 seconds for TRB and NRB,
only consider trucks with a vehicle class between 4 to respectively. This ensures at least four cycles of free
13, and vehicle speeds between 40 to 85 mph. vibration are recorded for each bridge.
The sampled WIM data is used for generating
TRB and NRB responses using the CSiBridge model.
The speed of each truck is perturbed using a random
variable between -5 and 5 mph so that simulations of
the same truck on the two bridges vary slightly
making the response data sets more realistic.
The simulated responses of both bridges are the
strain values at the mid-span of the central girders.
The description of the whole procedure is shown in
Figure 1. Additionally, the detailed FE models for
these two bridges are shown in Figure 2.
Note that in this study, measurements (strains) are
taken at only one point on both bridges. Trucks are
passed over each bridge using a particular lane specific
to each bridge. In the real-world, trucks aren’t
restricted to this one lane combination. There are 9 dif­
ferent lane combinations that the truck may travel over
considering each bridge has three lanes. In this paper,
only one combination of lanes is used to validate the
capability of the proposed Seq2Seq model. Based on
field observations, the most common lane is: trucks Figure 3. Sample TRB strain response to truck weighing
pass through the middle lane on TRB and the slow 62.9kips crossing in the middle lane at 64 mph.

301
 A total of 2,100 truck events (pairs) are simulated For more description of these cell types, interested
on both bridges. 1,500 pairs of these simulations are readers are referred to Chung et al. (2014) and Mal­
used for training the models, 150 pairs for validation, hotra et al. (2016) for details.
and 450 pairs for testing. The RNN cells apply a combination of linear
matrix multiplications and non-linear transform­
ations. The equations describing the GRU are given
3 TIME SERIES FORECAST MODELS as (Paszke et al. 2017, Chung et al. 2014):

3.1 Seq2Seq model

The backbone of a Seq2Seq model is a recurrent
neural network (RNN). An RNN can be viewed as
a unit cell that is rolled over itself for a set number of
times. At each time step, the cell takes an input x½i],
and the hidden state from the previous step h½i _ 1]
and returns the current hidden state h½i] (Figure 4). The
network aims at capturing sequential information
within the data presented to it through the use of
hidden states from the previous time steps. The main
computation of the RNN is the unit cell F where �E
denotes the cell for the encoder and �D for the where tanh is the hyperbolic tangent function, rep­
decoder. The RNN architecture is used to model an resents the entrywise product, and σ represents the
encoder whose design is optimized to output a low sigmoid function defined as .
dimension context vector using input time series x[i]; Additionally, the index i represents the time index to
the decoder uses the context vector to feed another show how the current hidden state is related to both
RNN that outputs another time series y[i]. In this the current input and previous hidden state.x½i] is the
study, x and y correspond to responses of the two input at time i. The terms r½i], z½i], and n½i] are termed
bridges to the same truck. the reset, update, and new gates, respectively. Finally,
Two different cell types are used in this study: the W ‘s and B ‘s are the weights and biases of the
namely gated recurrent unit (GRU) and long short- linear transformations. Figure 4b shows the GRU cell.
term memory (LSTM) cells. The brief descriptions Similar, the equations of LSTM cell are expressed
of these two cell types are presented in this section. as follows (Paszke et al. 2017, Graves et al. 2013):

Figure 4. (a) Seq2Seq model schematics. From the Encoder’s side, the context vector is passed to the decoder. The decoder
uses it’s own predictions in the previous time step as the input to the current time step, (b) GRU cell, and (c) LSTM cell.

302
 where W and bda are the weight matrix and bias
vector of this layer, x½i] and y½i] are the input and
output at time i.
There are a handful of non-linear transformations
commonly used with neural networks, however, the
non-linearity function should have a range of
ð_∞; ∞Þ. The tanhshrink activation function is one
of the few that has such characteristics. The weight
matrix W da , bias vector bda, along with the internal
weights and biases of the cells are optimized in the
training stage so as to minimize the mean squared
where h and c are called the hidden state and cell state. error (MSE) between the NRB response and that pre­
Moreover, j, f , g, and o are referred to as the input, dicted by the Seq2Seq model when the correspond­
forget, cell and output gates, respectively. It is empha­ ing TRB response is fed as the input to the encoder.
sized that for the LSTM cell, at each time step, two set
of vectors are received from the previous time step,
3.2 Training model
namely the hidden and cell states (h and c). And the
LSTM cell is shown in Figure 4c. The Seq2Seq model is implemented in PyTorch
In summary, as shown in Figure 4a, the function­ (Paszke et al. 2017). This subsection contains tech­
ality of this cell unit in the encoder RNN can be nical details of the implementation and training the
expressed by: model using that software tool.
A key feature within the Seq2Seq model is the
dimension of the context vector in the last hidden
state of the encoder. A set of four different hidden
dimensions, namely H ^ ¼ 10; 30; 50; 70g for both
where x½i], h½i], and h½i _ 1] are the input, hidden cell types are iterated on and the one that results in
state at time i, and the hidden state at time i _ 1, highest performance is selected.
respectively. For the GRU cell, this hidden state is To train the model, a stochastic gradient-based
basically the h vector in Equation (1). While for the optimizer named Adam (Kingma & Ba 2014) is.
LSTM cell, this hidden state is a concatenated vector Similar to other stochastic gradient-based
of h and c vectors in Equation (2). approaches, at each iteration, Adam takes a subset of
The Seq2Seq model used here consists of two dif­ the training dataset (called a mini-batch) and
ferent RNNs: one is the encoder �E and the second approximates the gradient of the MSE with respect
is the decoder �D (this structure is also referred to as to the weights. The optimizer then takes a step using
an encoder-decoder architecture). The encoder takes the gradient at that iteration. Consequently, the size
the input time series and condenses the information of the mini-batches, referred to as the batch-size, is
into a fixed-length vector which is referred to as the another important training feature (so called hyper-
context vector. The context vector is the last hidden parameter). Four different batch sizes, namely
M^ ¼ f10; 20; 50; 100g are used and the one with
state of the encoder but serves as the input on the
decoder’s side. On the decoder side, the encoder’s best performance is selected.
last hidden state (context vector) is fed in with the Epoch is another hyperparameter to be set. It
decoder then predicting the output using that context refers to the number of times the entire set (consist­
vector. For each time step, the decoder recurrent ing of multiple batches) is used to update the neural
cell’s output is used as the input to the cell at the network. Here the dataset in training is 1500 obser­
next time step. This is in contrast to the encoder vations divided into batches (with network weights
where the actual time series values are used at each updated after each batch). During an epoch the train­
time step ing will repeat the learning process on 15 batches
The dimensions of the hidden states of the when batch size is 100, 30 times when batch size is
encoder and the decoder have to be the same. Mean­ 50, and so on. The number of epochs used during
while, the dimensions of the hidden state of the training will be described shortly.
decoder could be different than that of the output. Another important hyperparameter set prior to train­
Hence, a fully connected layer is used on the decoder ing is the learning rate for the Adam optimizer. Adjust­
side of the model. This layer acts as follows: let hd ½i] ing the learning rate throughout training is critically
be the decoder’s hidden state at time step i; then, this important in achieving an optimal set of network
hidden state is multiplied by a weight matrix W da weights.. Initially, the learning rate must be high to
such the product is a scalar; the scalar is then passed allow large enough steps towards the optimal weights
to a “tanhshrink” function defined as follows: but these steps must be gradually reduced to prevent

303
overshooting a local minimum of the cost function. To where p is the number of lags on the output (poles),
adjust the learning rate, a built-in feature of PyTorch q is the number of lags on the input (zeros) and k is the
named ReduceLROnPlateau is used. This function offset between the input and output. The set of param-
keeps track of the MSE on the validation dataset used eters are the weights in
after each epoch to test the Seq2Seq model. If this
value doesn’t decrease for 10 epochs, it reduces the the ARX model. Here, ψ½η is the target (NRB’s)
learning rate to a half of its previous value. It should response and y0 ½η is the ARX model forecasting NRB
be noted that it is up to the user to decide on the learn­ response at the ηth time step, respectively.
ing rate and the number of epochs to wait before redu­ This model can be expressed in the form of multi-
cing the rate. In this paper, the initial learning rate is plication of a matrix with the vector of weights as
selected as 0.02. follows. For each time step η, the right-hand side of
Overall, a maximum of 2,000 epochs are used for Equation (5) can be expressed as the inner product
training the Seq2Seq model. However, it was realized of a row vector
that beyond a certain point, the model overfitted the and a column
training data where the error on the validation dataset vector being Let Q be
would not decrease (and may increase). Thus, to speed a matrix whose rows are Qn vectors and let y0 be
up the process of search for optimal hyperparameters, column vector whose elements are y0 ½η. Then the
training was terminated if the validation error did not ARX model can be expressed as a matrix multiplica-
decrease for 70 epochs. For each value of hidden tion in the following form:
dimension and batch size, the set of weights that
resulted in the minimum MSE on the validation dataset
were stored. Training errors (MSE) for the encoder-
decoder network model is shown in Figure 5. Add­
itionally, a section of each plot has been magnified to To find the set of optimal weights, , for
depict the divergence of the training and validation a selected set of model orders and offset,
errors. a regularized least-squares approach is taken (Golub
et al. 1999). Let ψ be a column vector containing all
3.3 ARX Model target variables constructed similar to y0 . Then:

To validate the capacity of the proposed encoder-

decoder architecture, a linear ARX model is
employed as a baseline model. The ARX model can
be expressed as follows:

where λ is the regularization parameter and rep­

resents the l-2 norm. The solution to Equation (7) is
given by:

Figure 5. Training error (MSE) for the proposed encoder-decoder methods: (a) GRU unit cell, and (b) LSTM unit cell.

304
 response time series data used is microstrain (με)
and hence the RMSEs have a unit of με as well.
As shown in Table 1, for the encoder-decoder
model with GRU cell units, it was found that using
where representing the transpose of matrix, and I a batch size (M^ ) of 100 with a hidden dimension (H)^
being the identity matrix with the appropriate size. of 30 resulted in the lowest validation error. In that
Equation (8) is solved using Python’s Numpy pack­ situation, the validation RMSE is 1.03 με. Moreover,
age (Van Der Walt et al. 2011) for the encoder-decoder model with LSTM cell
For training, the true values of the target sequence units, it was found that using a batch size of 50 with
are used in the ψ vector (ψ ¼ y0 ). However, in the a hidden dimension of 70 resulted in the lowest val­
validation and testing stages, only the first p values idation error of 0.95 με. While, for the ARX model,
match the targets. For the remaining time steps, the the optimal hyper-parameters were p=50, q=50,
forecast of the ARX model is used. This implemen­ k=70, and λ=0.5, which resulted in a validation
tation is similar to that of the Seq2Seq model when RMSE of 1.42 με.
the model forecast of the NRB response at previous By selecting the model with the lowest error on
time steps are used as opposed to the actual values the validation dataset, the issue of over-fitting can be
of the target sequence. avoided. Note that the total number of training
To find the optimal set of hyperparameters, a grid epochs are different for the two cell types and this is
search similar to that of the Seq2Seq model was con­ due to terminating the training procedure when the
ducted. The values for p and q were from the set validation error did not decrease for 70 epochs.
^ ¼ f5; 10; · · · ; 50g, the values for k from
P During the testing stage, the test dataset contain­
^ ¼ f0; 10; · · · ; 100g, and the λ parameters were
K
^ ¼ f0:01; 0:1; 0:5; 1; 2; 5; 10; 100g. For each ing 450 observations was used. We emphasize that
from A the training process was fully independent of this
combination of hyperparameters, the ARX model testing dataset. The resulting RMSEs and the infer­
was trained by using Equation (8) and the trained ence time for each model are listed in Table 2. The
model was used on the validation dataset. The com­ computations were carried out on a Nvidia GTX
bination of hyper-parameters resulting in the lowest 1070 GPU.
validation error were used in the model. Furthermore, for two sample observations in the
In summary, the greatest two differences between test dataset, the model predictions along with the
the ARX and Seq2Seq models are: (1) the first p actual observations on NRB are shown in Figure 6. It
terms of the predicted results in the ARX model can be seen that the GRU cell type results is the best
match exactly that of the target values, whereas the performance on the test dataset. The performance of
Seq2Seq values are initiated with zeros and only the the encoder-decoder model with LSTM cells is close
model forecasts are used from the very beginning, to that of an encoder-decoder model with the GRU.
and (2) the encoder observes the entire input Both encoder-decoder models outperform the ARX
sequence and then passes the context vector to the model in both accuracy and elapsed inference time.
decoder, whereas in the ARX model, at each time
step, only a subset of the input observations are
used. Despite the differences, the ARX model is one
of the only time series model structures that has
Table 2. Test results of the Seq2Seq and ARX modes.
strong similarities with Seq2Seq model.
Average Inference
Method RMSE (με)
Time (ms)
4 RESULTS AND DISCUSSION
Encoder-Decoder: GRU 0.9 0.89
Root mean squared error (RMSE) is a common Encoder-Decoder: LSTM 1.03 1.90
metric for reporting the errors of time series models ARX 1.26 5.14
which will be used in this paper. The units of the

Table 1. Minimum RMSE of the GRU and LSTM cells during training with different batch sizes and hidden dimensions.

RMSE of GRUðμεÞ RMSE of LSTMðμεÞ

^
M
^
H 10 20 50 100 10 20 50 100

10 8.61 8.58 8.61 8.89 8.61 2.33 8.61 8.89

30 1.79 1.48 1.35 1.03 1.17 1.75 1.26 1.08
50 2.27 3.12 1.39 1.92 3.41 1.07 1.15 1.02
70 5.25 2.72 1.57 1.50 2.42 1.15 0.95 1.06

305
Figure 6. Predicted responses for two observations: (a) Sample 1, the RMSE values are 0.28 με, 0.27 με and 0.66 με for the
GRU and LSTM of encoder-decode methods and ARX models respectively, and (b) Sample 2, the RMSE values are
2.90με, 3.26με and 1.16με for the GRU and LSTM of encoder-decode methods and ARX models respectively.

Despite the ARX model seemingly being sim­ In future studies, we will seek to implement these
pler, the elapsed time for this model is much longer models on the experimental dataset captured from
than that of the encoder-decoder models. This the actual bridges. In such data sets, a lower signal­
should be attributed to the fact that the Seq2Seq to-noise ratio (SNR) is expected which may chal­
models takes batches of data (for GRU a batch of lenge the Seq2Seq models. Future work will explore
size 100 and for the LSTM a batch of size 50) how hyperparameters affect the models in the face of
whereas the ARX only takes one observation at sensor noise. Furthermore, we seek to investigate the
a time. In addition, we emphasize the code used for capability of the models in damage detection using
implementing the Seq2Seq models takes advantage a synthetic dataset and compare it with those of the
of the built-in features of PyTorch. The code used state-of-the-art methods.
for implementing the ARX model is written by the
authors without further optimization. It may affect
the elapsed time but should not change the fact that ACKNOWLEDGEMENTS
ARX model requires more time than the other two
models. The authors acknowledge the support provided by
the National Science Foundation under grant
#1831347. Additional support was provided by the
5 CONCLUSION University of Michigan Institute for Data Science
(MIDAS).
Time series forecast models were used in this study
for jointly modeling two bridges in a highway corri­
dor. A synthetic dataset was generated using the CSi- REFERENCES
Bridge FE platform to simulate the response of the Chung, J., Gulcehre, C., Cho, K. & Bengio, Y. 2014.
bridges to the same moving loads observed in the Empirical evaluation of gated recurrent neural networks
field. The load properties were sampled from the on sequence modeling. arXiv preprint arXiv:1412.3555.
actual measurements captured by a WIMS station. Computers and Structures, Inc. 2011. CSI analysis refer­
Slight perturbations were added to the speed of the ence manual for SAP2000, ETABS, SAFE, and CSi-
load to make the dataset more realistic. Bridge [Online]. CSI Berkeley, CA. Available: http://
For the encoder-decoder model, two types of docs.csiamerica.com/manuals/etabs/Analysis%20Refer
recurrent cell units, namely GRU and LSTM, were ence.pdf.
used to predict the response on the NRB and they Golub, G. H., Hansen, P. C. & O’Leary, D. P. 1999. Tikho­
nov regularization and total least squares. SIAM Journal
both showed equally competitive results. An ARX
on Matrix Analysis and Applications, 21(1): 185–194.
model was also proposed as a linear variant of the Graves, A., Mohamed, A. & Hinton, G. 2013. Speech rec­
Seq2Seq models for the task at hand. Both encoder- ognition with deep recurrent neural networks. IEEE
decoder models outperformed the ARX model in International Conference on Acoustics, Speech and
terms of both the error and inference time required Signal Processing, Vancouver, Canada, 26-31 May,
for the model. 2013, (6): 6645–6649.

306
Hou, R., Jeong, S., Law, K. H. & Lynch, J. P. 2019. Reiden­ Salehi, H. & Burgueno, R. 2018. Emerging artificial intelli­
tification of trucks in highway corridors using convolu­ gence methods in structural engineering. Engineering
tional neural networks to link truck weights to bridge Structures, 171(18): 170–189.
responses. Sensors and Smart Structures Technologies Salinas, D., Flunkert, V., Gasthaus, J. & Januschowski, T.
for Civil, Mechanical, and Aerospace Systems, Denver, 2019. DeepAR: Probabilistic forecasting with autore­
Colorado, 3-7 March, 2019, (10970P):1–14. gressive recurrent networks. International Journal of
Jeong, S., Ferguson, M. & Law, K. H. 2019. Sensor data Forecasting, 7(1): 1–11.
reconstruction and anomaly detection using bidirectional Spencer Jr, B. F., Hoskere, V. & Narazaki, Y. 2019.
recurrent neural network. Sensors and Smart Structures Advances in computer vision-based civil infrastructure
Technologies for Civil, Mechanical, and Aerospace Sys­ inspection and monitoring. Engineering, 5(19):
tems, Denver, Colorado, 3-7 March, 2019, 199–222.
(10970N):1–11. Sutskever, I., Vinyals, O. & Le, Q. 2014. Sequence to
Kingma, D. P. & Ba, J. 2014. Adam: A method for stochas­ sequence learning with neural networks. arXiv preprint
tic optimization. arXiv preprint arXiv:1412.6980. arXiv:1409.3215.
Ljung, L. 1999. System identification. Springer: 1–19. Van Der Walt, S., Colbert, S. C. & Varoquaux, G. 2011.
Malhotra, P., Ramakrishnan, A., Anand, G., Vig, L., The NumPy array: a structure for efficient numerical
Agarwal, P. & Shroff, G. 2016. LSTM-based computation. Computing in Science & Engineering, 13
encoder-decoder for multi-sensor anomaly detection. (2): 22–30.
arXiv preprint arXiv:1607.00148. Wilms, H., Cupelli, M. & Monti, A. 2018. Combining
Paszke, A., Gross, S., Chintala, S., Chanan, G., Yang, E., auto-regression with exogenous variables in
DeVito, Z., Lin, Z., Desmaison, A., Antiga, L. & sequence-to-sequence recurrent neural networks for
Lerer, A. 2017. Automatic differentiation in pytorch short-term load forecasting. IEEE 16th International
[Online]. Available: https://openreview.net/pdf? Conference on Industrial Informatics, Porto, Portugal,
id=BJJsrmfCZ. 18-20 July, 2018, (18):673–679.

307
 Bridge Maintenance, Safety, Management, Life-Cycle Sustainability and Innovations –
Yokota & Frangopol (eds)
© 2021 Taylor & Francis Group, London, ISBN 978-0-367-23278-8

Mixed training of deep convolutional neural network for bridge

deterioration detection with UAV and inspection report sourced images
J. Dang
Saitama University, Saitama, Japan

P. Chun
University of Tokyo, Tokyo, Japan

ABSTRACT: Instead of manual detailed bridge inspection, UAV based bridge patrol can roughly scanning
the surface of the bridge and bring back high-resolution image data in very low cost. To use those data
smartly, the application of Convolutional Neural Network (Deep Net) based image processing can found
deteriorations such as cracks, palling, corrosions, leaking water effectively. Efforts have been conducted to
extract labeled damage photos from past Inspection reports and to find out effective way to train CNN model
for damage detection. Although it can achieve accuracy about 90% after data argumentation using this data
base, its ability of recognize damage from real word UAV image is low. In this study, mixed trainings of CNN
model with both UAV sourced image data and inspection report sourced data were conducted to reinforce the
machines performance when it’s seeing the background and no-damage structural members, which are less in
the inspection reports but resourceful in real UAV scanning images. UAV videos acquired from a few real
bridge patrols were used as sources. The 4K images was sliced from video and split to small samples. Each
sample was then label manually to classes including background, a few types of damage, and a few types of
undamaged structural surfaces. Training with original inspection report sourced data, UAV sourced data, and
mixed data were conducted and compared in accuracy of damage recognition in UAV image.

1 INSTRUCTIONS However, considering the amount of UAV

images, the work of picking up damages from those
For bridge structures, the fast reopen after large earth­ data by human eyes need a lot of proccing working
quake is one of the essential issues for seismic resili­ hours. The missing of damage location, individual
ence of urban area. The emergency structure deviation in standard of damage, yielding in long
investigation checks the safety status and prevents the hours operation fatigue will be new challenges.
secondary disaster such as collapse of severely dam­ Thus, an Image Processing based automatic damage
aged bridges due to traffic load or after shock. How­ detection technic are expected.
ever, the emergency investigation after large With the successful application of Convolutional
earthquake is dangerous, difficult, and time consuming. Neural Network based image processing algorithms,
For long span large bridge, the inspection ladder the possibility of combine these two technologies to
path may have already been damaged and engineers a revolutionary bridge inspection approach becomes
may risk their lives to climbing on the high place to more and more clear. A lot of efforts are also put
check the potential damage in bearings when strong into developing damage detection machine using
after shock may happen in any time. In recent years, deep learning with different approaches.
the personal UAVs for photography grows very fast. In this study, bridge damage photo reported in
If it is used properly, the inspection using UAV recent bridge inspection con-ducted in Japan are used
would be far more effective as for the difficult to to train some Convolutional Neural Network models,
access part can be easily scanned by the UAV. and the accuracy of the damage detection machines are
Resent years, a lot of UAVs have been developed for evaluated. Some technics to train a high accuracy
bridge inspection. F. Moreu et. al even tried to using model with some data based are also tested and the
UAV to monitor the vibration caused deflection in results are also reported. Then the feasibility of bridge
railway bridges. inspection using UAV is investigated by a series of

DOI: 10.1201/9780429279119-38
https://doi.org/10.1201/9780429279119-38

308
bridge site verifications. Finally, some pictures taken number of extracted images are 485, 82, 55, 55,
by UAV was tried to be diagnosed by pretrained CNN and 36.
model. In this study, as a verification of the usefulness of
Deep Learning for damaged images, image classifica­
tion using deep learning was performed based on the
2 CONVOLUTIONAL NEURAL NETWORK images collected by past manual inspection reports of
BASED DAMAGE DETECTION bridges.
Here, a well-known model, GoogLeNet which
In Japan, inspections using robots are being encour­ won the competition of large scale image recognition
aged by the new development strategy of govern­ (ILSVRC) in 2015.
ment grand plan. UAV is considered as one of the GoogLeNet has a structure in which modules
most important tools to improve the inspection effi­ called Inception are stacked. In Inception, convolu­
ciency. However, the images collected by the UAV tion is performed in parallel, and each obtained fea­
are in large size and number. ture map is combined.
It will cost a lot of working hours for reviewing and This model is suggested and pretrained in Caffe
labelling the damage in videos frame by frame manu­ framework. Part of structure of this model is shown in
ally. By using Convolutional Neural Network (CNN) Figure 3, residual part can be found in reference. The
based Deep Learning (DL)methods, the damage detec­ input size is 256 * 256 pixels, and the output size is 5
tion can be processed in low cost and fast processing. classes.
The basic concept of UAV and DL based Inspection Table 1 shows the results of classifying verifica­
can be presented in Figure 1. The UAV, Smartphones tion data using a network trained using those data.
and record high resolution image data instead of oper­ 80% of the extracted damaged image is used as
ator and inspection vehicles-based inspection which training data, 10% is used as valuation, and 10% is
are commonly companied with suspension of the trans­ saved as test data.
portation function partly or fully of the bridge falling The classification accuracy of ‘Corrosion’
with traffics issue. Then the data can be real-timely exceeds 99%, indicating that the ‘cracking’ is also
processed or processed later, to recognize the damage roughly classified correctly. However, the precision
and evaluate their influence to whole structure. of ‘Exposed Reinforced Bar’, ‘Leakage, Ponding
In Japan, due to legislative amendment in 2014, Water’ is less than 20%.
periodic inspection was obligated. As a result, many Figure 4 shows an example of classification
inspection reports were prepared. In this research, results for ‘Leakage, Ponding Water’ images.
damaged images were extract-ed from a certain In the figure, water leakage can be seen in the
bridge inspection report and a learning database was middle part of the image, but the concrete in the lower
constructed. Many images were taken of the damage right part of the image is discolored to reddish brown,
from a distance in the report. which seems to have been judged as ‘Corrossion’.
Therefore, only images that can confirm damage in Such a tendency was also seen in other misrecognized
images were selected. Also, as a Structural Damage images. Images with such features are likely to be erro­
Detection data-base, as shown in Figure 2, only the neously recognized and there is room for improvement.
damage pictures that satisfies the following two condi­
tions as picked up and labelled due to the reports:
3 LIMITATION OF DATA BASE AND DATA
(1) Damage can be judged by visual inspection. ARGUMENTATION
(2) The number of images in the report is more
than 30. The accuracy in some damages detection is still low as
Therefore, damage types handled in research are the data size is small, the number of training pictures is
“Corrosion”, ‘Crack’, ‘Exposed reinforced bar’, ‘Leak­
age, Flee Lime’, ‘Leakage, Ponding Water’, and the

Figure 2. Damage detection data base from inspection

Figure 1. UAV and DL based inspection reports. reports.

309
 Figure 4. Example case of classification result.

classification, a few image process methods, commonly

called as data argumentation, can be used to improve
this situation of using small size of training data. For
data argumentation it can flip, rotate, or change contrast
of the original data pictures to generate new pictures,
as can be seen in Figure 5.
After find more pictures from the original reports
total 2410 damage pictures were prepared for the 5
types of damage, 1101 for Corrosion, 430 for Leach­
ing Flee Lime, 260 for cracks, 432 Reinforcing Bar
Exposed, 187 for Leakage. By data argumentation,
the total size of the data base is grown to 18,144.
The accuracy for some groups is improved greatly,
though, it is still very easy to be wrong to tell the dif­
ference between the damage of concrete deterior­
ations. Furtherly, for those pictures only contain the
damage area in part of the whole image, such as the
one shown in Figure 6, the damage area can be cut
out from the original pic and resized to the training
image size.
Thus, only the damage feature are shown in the
training data, so that the machine can understand
this part contains the feature of their class.

Figure 3. CNN model (inception).

not large enough. As this is a comment limitation for

structural engineers, as data for real damage is very
rare and they are not available soon enough. For image

Table 1. Classification result of verification data.

Type Correct/images Accuracy

Corrosin 120/121 99.2%

Leakage, Flee Lime 10/20 50.0%
Exposed Reinforced Bar 2/14 14.3%
Crack 11/14 78.6%
Leakage, Ponding Water 0/9 0.0%
Figure 5. Data argumentation based on image processing.

310
 it is not trained to find the different from health con­
crete and damaged one. The trained model didn’t have
enough training information to distinguish damage
structure from health steel, clean concrete surface,
background such as trees, grass, river waters.

5 MIXED TRAINING DATA BASE USING UAV

Figure 6. Damage area focused data refinement. AND INSPECTION IMAGE

The feature of pictures from inspection is there are

many damaged pictures but with very few back­
ground nor health structures. High resolution image
data had been recorded by UAV based bridge inspec­
tion for a group bridges in s preliminary study.
Images from real world bridge inspection are differ­
ent with those picked up pictures for damage report.
They contain a lot of information of background and
health bridge structure surface. However, there are
very few parts containing with features of structural
damage, especially those severe damages.
To improve the performance and accuracy for
Figure 7. Training result for data refinement and data CNN model in real world picture, mixed training
argumentation.
data from both inspection reports and UAV is pro­
posed in this study. The data size of Inspection
reports sourced database, UAV Image sourced data­
Those refined data based is prepared and com­ base, UAV + Inspection Mixed database, Mixed data
bined with data argumentation using image pro­ with Argumention are shown in Figure 9.
cessing. The training result of those combination After training with Mixed data base the improve­
can be seen in Figure 7. Data set with refinement ment can be observed on Figure 10. The model
increased the accuracy to about 80%, showing
the great advantage of refining the data.

4 DAMAGE DETECTION TEST UAV IMAGES

Using the trained model to detect the damage in a UAV

4K photo shows a lot of miss recognition for Corrosion
Class, as can be seen in Figure 8. Some of the back­
ground contents, such as buildings and trees was also
been recognized as corrosion. This implies that more
real word UAV pictures should be used in training the
Damage Detection. The training was conducted by
using Inspection Report Pictures. Most of pictures are Figure 9. Data size of mixed training data base.
about damaged concrete or steel structures. This model
was well trained to class the type of damage. Howere,

Figure 10. Damage detection on UAV image by reinforced

Figure 8. Application of CNN on UAV image. data trained model (Corrosion: Blue, health steel: Green,
backGround: Pink, concrete: White, leaking water: Black).

311
correctly recognized the corrosion area (Blue), and REFERENCES
also realized where is the area belong to healthy
steal (Green). Though, part of back ground is recog­ Hallermann, N., Morgenthal, G. 2014. Visual inspection
nized steel, though, it should be improved in the strategies for large bridges using unmanned aerial
future by increase the pictures taken by UAV in vehicles (UAV), 7th International Conference on Bridge
urban area, so that, building can be learned by the Maintenance Safety and Management(IABMAS),
pp.7–11.
model.
Haruta, D., Shrestha, A., Dang, J. 2016. A Preliminary
Study on Bridge Inspection Using UAV (Drone) Aerial
Photography, Proceeding of Japan Association for
6 CONCLUSION
Earthquake Engineering, Kouchi, Vol.16, P3-33,
pp.01–08.
Bridge damage photo reported in recent bridge Krizhevsky, A., Sutskever I., Hinton G. 2012. ImageNet
inspection conducted in Japan are used to train some classification with deep convolutional neural networks.
Convolutional Neural Network models, and the Advances in neural information processing systems
accuracy of the damage detection machines are 2012.
evaluated. The classification accuracy of ‘Corrosion’ Ministry of Land. 2013. Infrastructure, Transport and Tour­
exceeds 99%, indicating that the ‘cracking’ is also ism (MLIT) of Japan Report: Current status of road
roughly classified correctly. However, the accuracy structures (bridges), MLIT.
Ministry of Land. 2014. Infrastructure, Transport and Tour­
for UAV images was about only 60% as it failed to
ism (MLIT) of Japan: Routine inspection guideline for
distinguish the different of damage with back­ road bridges, MLIT.
grounds where are lack in inspection report pictures. Reagan, D., Sabato, A., Niezrecki, C. 2017. Feasibility of
To overcome this issue, Mixed training method using Digital Image Correlation for Unmanned Aerial
contain UAV picture and inspection report with Vehicle Structural Health Monitoring of Bridges, Struc­
some ration is proposed. The proposed Mixed train­ tural Health Monitoring.
ing performed better in test on the same UAV images Szegedy, C., Liu, W. Yangqing Jia, et al. 2015. Going
recognized the corrosion area accurately. Deeper with Convolutions, CVPR2015.

312
 Bridge Maintenance, Safety, Management, Life-Cycle Sustainability and Innovations –
Yokota & Frangopol (eds)
© 2021 Taylor & Francis Group, London, ISBN 978-0-367-23278-8

Fatigue stress spectra modeling of steel bridge decks using traffic

monitoring data
N. Lu
Changsha University of Science & Technology, Changsha, China

Y. Liu & Y. Luo

Hunan University of Technology, Zhuzhou, China

ABSTRACT: Orthotropic bridge decks are fatigue critical components since they directly suffer from cyclic
traffic loads. This study utilized a stochastic fatigue truck load model to simulate the fatigue stress spectra and
to evaluate fatigue reliability of orthotropic steel bridge decks. A computational framework associated with
deep learning technique was presented to deal with the uncertainty-induced computational complexity. In the
deep learning approach, initially, several uniformly designed training samples was generated accounting for
traffic load parameters including vehicle type, axle weight, etc. Subsequently, these training samples were in­
putted to the finite element model of the orthotropic steel bridge deck to simulate the equivalent stress range
under each truck load as the output samples. Finally, the training and output samples were connected to the
deep learning machine with a Gaussian Kernel function, and the accuracy of the learning machine was
checked in a parametric study. Based on the deep learning technology, the fatigue stress spectra of a prototype
bridge under actual traffic load was simulated It is concluded that the machine learning approach provides an
intelligent and efficient approach for probabilistic fatigue analysis of orthotropic bridge decks.

1 INTRODUCTION the welded joints under the bridge deck, where the
strain sensors are difficult to be placed (Gokanakonda
In general, a steel bridge is designed with enough et al., 2016). On this regard, the numerical simulation
fatigue resistance against the cyclic vehicle load (Sim approach can captures the fatigue stress characteris­
et al., 2009). However, recent field investigations on tics of a bridge under monitored truck loads. Mean­
several collapsed steel bridges (Lu et al. 2018; Lalthla­ while, the fast developing computer technology and
muana and Talukdar, 2013) indicated that the fatigue traffic weigh-in-motion (WIM) technology contribute
damage induced by accidental overloaded trucks con­ to the efficiently and accurately simulating the fatigue
tributed to the bridge failures. The fast- growing traffic damage of steel bridges under truck load.
volumes and loads will become a safety hazard for the Site-specific weigh-in-motion(WIM) measurements
fatigue safety of steel bridges especially in developing are big data that can be used for statistical analysis of
countries. Uncertainties in traffic flows add another traffic loading (OBrien and Enright, 2012). Numerous
challenge to accurately evaluate the fatigue damage traffic-load models have been developed based on site-
accumulation, where the probability model of fatigue specific WIM data. For in- stance, Wang et al., (2005)
damage accumulation mostly depends on the site­ utilized WIM measurements in Florida to develop
specific traffic loading. Therefore, integrating the actual a developed a live-load spectrum by combining static
traffic information into the fatigue reliability evaluation responses with estimated impact factors in
of existing steel bridges is of great importance, which a 3-dimensional nonlinear truck model. Zhao and
can provide a more actual evaluation result and Tabatabai (2012) developed a 5-axle single-unit truck
a theoretical basis for transportation management. model based on the WIM records in Wisconsin to sup­
Numerous research efforts have been concentrated plement the permit vehicle model. OBrien et al.,
to implementing structural health monitoring (SHM) (2010) estimated the characteristic maximum dynamic
data to evaluate fatigue damage accumulation in the load effects of short to medium span bridges with
fatigue-critical component joints of steel bridges (Ye extensive WIM measurements collected at 5 European
et al., 2012; Deng et al., 2015). However, the most sites. Marques et al., (2016) implemented WIM meas­
fatigue-critical component for steel bridge decks are urements of an old railway bridge in Portugal to

DOI: 10.1201/9780429279119-39
https://doi.org/10.1201/9780429279119-39

313
simulate the axle load, axle spacing and the velocity of pavement. When a moving vehicle crosses the
the train. In addition to the traffic load modeling, the WIM equipment, the sensors records the axle
WIM measurements in connection with the finite elem­ weight, the driving speed, the vehicle type and the
ent technology have also been used for fatigue reliabil­ time. These data can be used for a range of tasks,
ity evaluation of steel bridges. In this regard, Guo such as transportation management and overloaded
et al., (2012) evaluated the fatigue reliability of an truck monitoring. In the present study, those WIM
orthotropic steel bridge deck using a probabilistic finite measurements will be used as a statistical database
element approach. In the Guo’s truck load model, the for stochastic traffic flow simulation.
axle weights were fitted by lognormal distribution This study chooses a WIM system of a highway
functions that ignored the bimodal or tri-modal charac­ bridge in China as a prototype. The WIM system has
ters of the truck load. Subsequently, Guo and Chen been working since 2012, and information of the
(2013) demonstrated the effectiveness of integrating of WIM system can be found in Lu et al. (2019). For
the site-specific measurement and the finite-element the purpose of simulating stochastic fatigue truck
(FE) model for fatigue reliability assessment of in- load models, these data were filtered to remove
service steel bridges. Ye et al., (2015) conducted invalid data, such as the cars with GVW less than
a sensitive study on the influence of the bridge finite 30kN. According to the transportation control in
element model on the estimated stress under vehicle china (MOCAT, 2004), the maximum GVW for
loads. An advanced probabilistic fatigue stress analysis 6-axle trucks is 550kN. However, the maximum
approach is proposed by Zhang et al., (2016) using truck overloading rate more than 200%.
WIM measurements. The entire monitored vehicles were classified
As elaborated above, implementation of WIM into 6 categories according to the vehicle configur­
measurements is an effective way to simulate the sto­ ation characteristics. Figure 1 shows vehicle type
chastic traffic loading, which can be subsequently util­ pro- portions, driving lane proportions and the aver­
ized for fatigue damage evaluation of existing steel age daily truck traffic (ADTT), where V1 is light
bridges. However, the balance of computational effi­ trucks, V2 to V6 are the 2-axle to 6-axle trucks,
ciency and accuracy is still bottlenecking that limiting respectively, and AWij is the jth axle weight of the
the developing of existing fatigue reliability assessment ith vehicle type. It is observed that about 60% of
approaches. In Zhang’s computational framework, the the vehicles are 2-axle trucks and light cars, and
calculation of thousands of stress histories with respect most of heavy trucks were driving in the slow lane.
to the daily truck volume is obviously a time- These phenomena of the traffic information will
consuming problem. On the other side, since the actual contribute to the accurate fatigue damage evalu­
traffic loads are random in nature, the design fatigue ation of a steel bridge
truck load model cannot be utilized for probability ana­ As highlighted in opening literatures, the vehicle
lysis. The relative re- search efforts with respect to sto­ axle weight is the most sensitive parameter leading
chastic traffic load have focused on vehicle-bridge to the fatigue damage accumulation in steel bridges.
interaction and impact factor analysis (Zhou and Chen, Therefore, each axle weight of each type of vehicles
2015). However, implementing of stochastic fatigue was investigated. Taking V6 as an example, the axle
truck load model for fatigue reliability assessment of weight of the 6th axle indicated as AW64, are col­
steel bridges is still insufficient. lected, and the histogram and probability density
This study aims to develop an efficient framework functions (PDFs) are shown in Figure 2.
for implementing traffic WM measurements for fatigue In Figure 2, the Gaussian mixture model (GMM)
reliability evaluation of steel bridge decks. Initially, is superposition of a several normal distributions.
long-term monitored traffic data of a highway is util­ The GMM has the capability of capturing the over­
ized to simulate stochastic traffic load model, and the load behaviors of trucks, where the first and
traffic growth ratio is also included to consider the the second peaks in the PDF denotes the normal and
future traffic. Subsequently, a meta-model approxi­ overloaded trucks. The mathematical expression of
mated by neural network was utilized to substitute the the GMM can be written as (Xia et al., 2012).
traditional finite-element simulation. The proposed
computational framework was subsequently utilized to
evaluate fatigue reliability of the welded joints in
a steel bridge deck. Parametric studies are conducted
to provide suggestions to the traffic management based
on the predicted fatigue reliability index.

2 STOCHASTIC TRUCK LOAD MODELING

BASED ON WIM DATA
where, f ðy; a; θÞ is a predictive mixture density,
f ðyjθi Þ is a normal distributed PDF with respect to
2.1 WIM measurements θi, and ai, μi and σi are the weight, the mean value
In general, a WIM equipment has number of scales and the standard deviation of the ith normal distrib­
and pressure sensors that is embedded in the road uted functions, respectively.

314
 Figure 2. Histograms and PDFs of the 6-axle: (a) GVW;
and (b) axle weight.

Figure 1. Probability densities: (a) vehicle type; (b) traffic

lane; (c) ADTT.

The relationship between the GVW and each axle

weights were also investigated. Linear functions Figure 3. Fitted functions between axle weights and GVWs
were approximated as shown Figure 3 to represent of the 6-axle truck.
the axle weight proportion. In Figure 3, x is the
GVW of a vehicle, and y is an axle weight. It is contrary, the axle weight and vehicle configuration
assumed that the individual axle weight in the are chosen for modeling the stochastic truck load.
tandem and tri-dem axles is equal. The goodness of Based on the above illustration, a stochastic truck
fit for AW61, AW62, and AW66 in Figure 3 are 0.72, load model with was built considering the param­
0.83, and 0.91, respectively. eters of vehicle types, vehicle weights, and driving
lanes. The vehicle samples were generated using the
2.2 Stochastic truck load modeling Monte Carlo simulation method. In order to consider
the traffic growth in the bridge service period,
In a stochastic traffic flow, there are number of the a linear growth factor of the average daily truck traf­
random variables, including vehicle types, vehicle fic (ADTT) is assumed as 0.5%. A linear growth
weights, driving speeds and vehicle spacings. This factor of average daily traffic volume (ADTT) is
study considers the vehicle parameters with signifi­ assumed with a upper threshold value of 0.5%. On
cant influence on the fatigue stress of the bridge this basis, the current traffic and further traffic in the
deck. For a general load influence line of a welded 100th year are shown in Figure 4, where x-axis is the
joint in a steel bridge deck (Ji et al., 2013), the arrival time of individual truck, and the y-axis is the
effective load length is very short, usually between vehicle weight.
two diaphragms. Therefore, the vehicle spacing, It is observed from Figure 4 that the vehicles are
which is quite larger than the effective load length, random both in time domain and spatial domain.
can be ignored in the stochastic traffic model. On the Under the condition of a large number of vehicles,

315
Figure 4. Simulated stochastic truck load models.

these samples can reflect the probabilistic informa­

tion of the actual traffic. Therefore, the stochastic
traffic load model provides a foundation for utilizing
actual traffic data for the probabilistic analysis of
traffic load effects on a bridge.

3 COMPUTATIONAL FRAMEWORK FOR Figure 5. Flowchart of the proposed computational frame­

STRESS SPECTRUM SIMULATION work.

In general, the traditional approach for calculating

the truck-induced fatigue damage accumulation
includes three steps (Zhang and Cai 2011; Wang
et al. 2013). First, simulate the stress history of
a fatigue-critical point under a moving truck load.
Second, convert the history into stress blocks by util­
izing the rain-flow method. Finically, the individual
fatigue damage is evaluated by the S-N curve and
accumulated by the Miner’s rule. However, this
approach is inappropriate for the stochastic truck
load, because of the time-consuming problem caused Figure 6. Diagram of a RBF neural network.
by numerous repeating computer runs. Therefore,
the computational framework is presented for utiliz­ between the axle weights and the equivalent stress
ing the stochastic truckload model for probabilistic ranges.
modeling of the fatigue stress range. In Figure 6, the input parameters in the present
Since the time-consuming problem origins from study indicate the axle weights for a vehicle type.
a large number of stress histories by repeating finite Since there are 6 types of vehicle configuration, a total
element analysis, a meta-model can be utilized to number of 6 meta-models are necessary for the entire
substitute the finite element model. In the present trucks. First of all, the upper and lower bound of the
study, a meta-model approximated by neural net­ GVW should be determined, and then generate several
works is used to substitute the time-consuming finite uniformly distributed samples in the defined region via
element simulation procedures. The entire proced­ a uniform design (UD) approach. Subsequently, con­
ures are summarized in Figure 5. The critical steps duct the finite element analysis to compute the stress
in the flowchart include the neural network-based histories under individual truck passage, and then con­
response function approximation and GMM-based vert the stress histories into stress blocks via the ran-
probabilistic modeling of the stress range. flow method. Finally, the variant-amplitude stress
The first critical step is approximating response sur­ blocks are converted to constant-amplitude stress
face functions. Since each truck passage will pose sev­ blocks, which are used to approximate the response
eral stress blocks due to the multi-axle and dynamic function. Influence of the number of the training sam­
effects, the response function between the ax- le load ples on the prediction of the stress range will be dis­
and the equivalent stress ranges is complex. The inte­ cussed in the cased study.
gration of the uniform design (UD) approach (Liu The second critical step is the probabilistic model­
et al., 2016) and radial basis function (RBF) neural ing based on GMM. With the approximated response
networks shown in Figure 6, which is commonly used function, the probabilistic modeling can be carried out
for structural reliability evaluation, is utilized herein as efficiently. The purpose of the probabilistic modeling
a meta-model to approximate the implicit function is to establish a probability model of the equivalent
316
stress range. However, impacted by the probability the beams and pylons were simulated by Beam188
distribution of random variable in the stochastic traffic elements, and the cables and hangers were simulated
flow, the PDF of the equivalent stress range is com­ by Link180 elements. In the local model, all of the
plex. Thus, the PDF of the stress range may not be components were simulated by Shell63 elements.
well fitted by a single Gaussian or lognormal distribu­ The longitudinal, transverse and vertical length of
tion function. In order to deal with this problem, the local model is 12.8 m, 14 m and 3 m respectively.
a GMM is utilized herein to capture the characters of The deck and U-ribs were meshed by quadrilateral
the bimodal or tri-modal distribution of the PDFs. The elements, while the longitudinal stiffening plates, the
GMM provides an effective connection between the diaphragm plats and the web plates were meshed by
PDFs of GVWs and the equivalent stress range. triangular elements.

4 CASE STUDY 4.2 Finite-element simulation

A steel box-girder of a suspension bridge is utilized The local model is in the mid-span point of the
herein as a prototype to demonstrate the application global model. In order to observe the truck-induced
of the computational framework for the fatigue reli­ stress history, stress influence lines of the two
ability assessment. Influences of the parameters in welded joints in the global model were computed.
the stochastic truck load model on the fatigue reli­ The effective stress history is in the range of
ability index are discussed. a segmental girder. It is observed from static stress
influence lines that the effective stress influence line
is mostly confined to the region of two diaphragm
4.1 Bridge details plates. This has demonstrated the significance of the
Nanxi Suspension Bridge is a long-span high- way axle spacing and the vehicle configuration in the sto­
bridge in Sichuan, China. The orthotropic deck of chastic truck load model.
the steel box-girder directly supports vehicular loads Note that the pavement has not been considered
and contributes to the bridge structural overall load- in the finite element model. The axle load was simu­
bearing behavior. Dimensions of the bridge are lated by a vertical uniformly distributed load that ex­
shown in Figure 7. The WIM measurements in the tends to the bridge deck with 45°. For instance, the
present study were collected from the SHM system thickness of pavement is 6.7cm, and the load area of
(Liu et al., 2015) of this bridge. the back wheel is 60cm×20cm, then the revised load
A finite element model of this bridge was simu­ area is 73.4cm×33.4cm. Under the truck load of the
lated by a commercial finite-element program AN­ 6-axle trucks with maximum and minimum GVW,
SYS as shown in Figure 7(c). In the global model, the stress histories of the rib-to-deck joint were com­
puted as shown in Figure 8.
As observed from Figure 8, there are 6 peaks cor­
responding to the 6 axles of the truck. Each variant-
amplitude stress cycle is posed by an axle load. For
a 2-axle truck, 20 training data were designed as uni­
formly distributed samples. With 20 times of com­
puter runs each equivalent stress ranges were
obtained. The axle weight and the equivalent stress
range are input and output data for of the response
function. The “newrbe” function in MATLAB was
used to train the RBF neural network, where the
number of neurons equaled to the number of input
samples, and the kernel function is the Gaussian

Figure 7. Detailed dimensions of the bridge: (a) a half Figure 8. Stress-time histories of the rib-to-deck joint
cross- section; (b) U-rib; (c) finite element model. under a standard 6-axle truck load.

317
 traffic including averagely 2 thousands trucks needs
22 hours. The computational effort for a 100-day
traffic analysis is unimaginable without the meta-
model. With 6 types of trained neural networks, the
100-day stochastic traffic loads were translated into
stress ranges.

4.3 Probabilistic modeling

The PDF of the stress range under the 100-day
stochastic truck loading is fitted by the GMM,
and the daily number of stress cycles is fitted by
Figure 9. Response surface of the stress range under the the Gaussian distribution. Note that the number
2-axle truck load.
of stress cycles equals to the total number of
vehicle axles. Histograms and PDFs of the stress
range and the number of cycles of the rib-to­
function. The response function of the approximated
deck joint under the stochastic fatigue truck load
neural network is shown in Figure 9.
in 100 days are shown in Figure 11.
It can be seen from Figure 9 that the training
It can be found from Figure 11 that there are
samples have filled the space uniformly, and the
3peaks in the probability density of the stress
nonlinear response surface fits the samples well.
range that is well fitted by the GMM compared
It is acknowledged that the accurate of the meta-
to the nor- mal distribution model. In addition,
model is affected by the number of training sam­
the low- amplitude stress cycles has a higher
ples, which will also determine the computational
probability den- sity. The probability density of
effort. Therefore, influence of the number of
the number of daily cycles is approximated with
training samples on the accuracy of the meta-
normal distribution. As elaborated above, the
model was analyzed. The accuracy of the meta-
probability model of the stress range and the
model was reflected by the roots mean square
number of cycles established above provide
error (RMSE) of 100 random samples following
a basis for the subsequent fatigue reliability
uniform distribution. The roots-mean-square-error
evaluation.
of the approximated neural network of the 2-axle
truck (V2) and the 6-axle truck (V6) is shown in
Figure 10.
As observed from Figure 10, the 6-axle truck
needs more training samples to ensure the accuracy
of the meta-model. The 6-axle truck model includes
3 independent axle loads variables: the front axle
load, the tandem axle and the tridem axle. Under the
accuracy criterion of 0.1 MPa, the 2-axle and 6-axle
truck model needs 20 and 40 samples respectively.
Therefore, the 6 types of truck load model need 180
training samples. The entire finite-element computa­
tional effort for a core-7 computer is roughly 2
hours. However, without the meta-model, one-day

Figure 11. Histograms and PDFs of the rib-to-deck joint

Figure 10. Influence of the number of training samples on under the stochastic truck load: (a) stress range, and (b)
the RMSE of stress ranges. number of daily cycles.

318
5 CONCLUSIONS moni- toring data. Sesnors 19, 5056. (doi:10.3390/
s19225056).
A stochastic truck load model was developed based Deng, Y., Liu, Y., Feng, D. M., Li, A.Q. 2015. Investigation
on site-specific WIM measurements for fatigue reli­ of fatigue performance of welded details in long-span
ability evaluation of orthotropic steel bridge decks. steel bridges using long - term monitoring strain data.
The time-consuming problem of the finite- element Structural Control and Health Monitoring 22(11):
based fatigue stress simulation was solved by utiliz­ 1343–1358.
Ye, X.W., Ni, Y.Q., Wong, K.Y., and Ko, J.M. 2012. Statis­
ing a meta-model approximated by neural networks. tical analysis of stress spectra for fatigue life assessment
The effectiveness of the stochastic truck load model of steel bridges with structural health monitoring data.
in probabilistic modeling and fatigue re- liability Engineering Structures 45: 166–176.
assessment was demonstrated by the case study of Gokanakonda, S., Ghantasala, M. K., and Kujawski, D.
a steel box-girder bridge. 2016. Fatigue sensor for structural health monitoring:
The computational effort of the fatigue stress ana­ Design, fabrication and experimental testing of
lysis in the finite-element model is greatly reduced a prototype sensor. Structural Control and Health Moni­
by utilizing the meta-model. However, the accurate toring 23(2): 237–251.
Guo, T., Frangopol, D. M., and Chen, Y. 2012. Fatigue
of the prediction mostly depends on the number of
reliability assessment of steel bridge details integrat­
the training samples with uniform design scheme. ing weigh-in-motion data and probabilistic finite
Approximately 180 training samples is enough for element analysis. Computers & Structures 112:
the meta-models of the 6 types of vehicles. 245–257.
Future efforts are needed to improve the stochas­ Zhou, Y., and Chen, S. 2015. Dynamic Simulation of a
tic fatigue truck load model by considering the long- span bridge-traffic system subjected to combined
vehicle spacing parameter. The neural network serviceand extreme loads. Journal of Structural Engin­
approach can be replaced by a advanced approach to eering, 141(9). (DOI: 10.1061/(ASCE)ST.1943­
make the computational framework more efficient. 541X.0001188).
In addition, the vehicle-bridge interaction and the Ji, B., Liu, R., Chen, C., Maeno, H., and Chen, X. 2013.
Evaluation on root-deck fatigue of orthotropic steel
degradation of road surface roughness condition will bridge deck. Journal of Constructional Steel Research,
be considered in the future work. 90: 174–183.
Zhang, W., and Cai, C. S. 2011. Fatigue reliability assess­
ment for existing bridges considering vehicle speed and
REFERENCES road surface conditions. Journal of Bridge Engineering
17(3): 443–453.
Sim, H.B., & Uang, C.M. 2012. Stress analyses and para­ Wang, D., Zhang, D., Wang, S., and Ge, S. 2013. Finite
metric study on full-scale fatigue tests of rib-to-deck element analysis of hoisting rope and fretting wear evo­
welded joints in steel orthotropic decks. Journal of lution and fatigue life estimation of steel wires. Engin­
Bridge Engineering 17(5): 765–773. eering Failure Analysis 27: 173–193.
Lu, N., Liu Y., Deng Y. 2019. Fatigue reliability evaluation Liu, Y., Lu, N., and Yin, X. 2016. A hybrid method for
of orthotropic steel bridge decks based on site-specific structural system reliability-based design optimization
weigh-in-motion measurements. International Journal and its application to trusses. Quality and Reliability
of Steel Structures 19(1): 181–192. Engineering International, 32(2): 595–608.
Lalthlamuana, R. and Talukdar, S. 2013. Rating of steel Liu, Y., Deng, Y., and Cai, C. S. 2015. Deflection monitor­
bridges considering fatigue and corrosion. Structural ing and assessment for a suspension bridge using
Engineering and Mechanics 47(5): 643–660. a connected pipe system: a case study in China. Struc­
Lu, N., Ma, Y., Liu, Y. 2019. Evaluating probabilistic traffic tural Control and Health Monitoring 22(12):
load effects on large bridges using long-term traffic 1408–1425.

319
 Bridge Maintenance, Safety, Management, Life-Cycle Sustainability and Innovations –
Yokota & Frangopol (eds)
© 2021 Taylor & Francis Group, London, ISBN 978-0-367-23278-8

Infrared image-analysis-based concrete inspection using machine learning

S. Hayashi & K. Kawanishi
West Nippon Expressway Engineering Shikoku Company Limited, Takamatsu, Japan

T. Yamane, S. Izumi & I. Ujike

Ehime University, Matsuyama, Japan

P. Chun
The University of Tokyo, Tokyo, Japan

ABSTRACT: In recent years, due to the aging of concrete structures, deterioration such as flaking and
delamination has occurred. Since such damage leads to spalling and may cause concrete pieces to fall, thus
potentially injuring pedestrians or damaging vehicle passing below. Therefore, it is necessary to detect and
repair structures quickly. In Japanese highway companies, hammering inspection is undertaken once every 5
years to detect structural damage. However, to conduct the hammering test, it is necessary to hammer the
whole surface of the bridge, and arranging a bridge inspection vehicle is also necessary. Consequently, it is
a cost- and labor-intensive undertaking, accounting for half of the highway bridge maintenance costs. In add­
ition, traffic control is also required, which causes congestion and increases the risk of traffic accidents. This
study aimed to solve these problems by utilizing infrared thermography. Infrared thermography is
a nondestructive inspection technique used to detect regions of flaking and delamination by photographing the
temperature inhomogeneity of concrete structure surfaces using an infrared camera. The cost and labor are
drastically lower than that in the hammering test. Traffic regulation is also unnecessary, because the infrared
thermography can be measured nondestructively from a distance. However, the damage detection accuracy
based on the thermal image has been limited. This study reports the improvement in the detection perform­
ance by using deep learning, which ensures sufficient accuracy of the infrared method for practical use.

1 INTRODUCTION delamination based on the passive infrared thermog­

raphy method, which would not limit the target
Aging of concrete structures has been escalating bridges and should be able to inspect the target
recently, and several falling accidents of concrete frag­ remotely and entirely, in our project.
ments caused by fractures of PC reinforcement rod and The infrared thermography method has already
PC stranded wire in the prestressed concrete or by cor­ been applied in the detection of floating of building
rosion of reinforcing steel rods have been reported. On tiles, but currently, it has not been applied to the
April 3, 2009, at Tooridani bridge in Shikokuchuo, for inspection of floating and delamination of the con­
instance, a fragment of the cover concrete with 1m crete structures in practical use (Kimura et al. 2019;
long, 100mm wide and 260mm thick with the weight Plesu et al. 2012). It is because its detection rate is
of 6kg fell due to corrosion of reinforcing steel rod significantly lower than that of the conventional
placed in the water drainage unit. Such flaking inci­ hammering tests. To enhance the detection rate, we
dents must be prevented as much as possible, for it have examined infrared cameras by comparing their
could cause secondary disasters if the flaked concrete performances first. As a result, the type of a shorter
hit passers-by. range in the measurement wavelength region with
Several nondestructive inspection (NDI) tech­ a cooling type camera would be the most suitable
niques are available for the task (Oh et al. 2012; for the task. Also, we developed method that ana­
Cheng and Tian, 2012). We have decided to establish lyzes obtained thermal images to detect damage
a detection technique of areas with floating and thoroughly. Specifically, it enables the machine to

DOI: 10.1201/9780429279119-40
https://doi.org/10.1201/9780429279119-40

320
detect damages through machine learning trained 2.1.1 Building model structures and photographing
with features focused on anomalous forms in the conditions
areas of temperature alteration as well as the texture We built cubic concrete model structures. Each
within the target areas. model has a void with the size of 100mm x 100mm
(t=10mm) in a different depth from the surface;
20mm, 30mm, 40mm, and 60mm, respectively
2 EQUIPMENT AND METHOD OF (Figure 1). We set the models under the shade
INSPECTION beneath a bridge in Takamatsu Expressway to avoid
the influence of daytime sunshine and photographed
2.1 Equipment them at night (23 o’clock) under the weather condi­
tion with more than 10℃ of the day temperature
ASTM D 4788 (Standard Test Method for Detecting difference.
Delaminations in Bridge Decks Using Infrared
Thermography) requires the following conditions in 2.1.2 Thermal images of each infrared camera
thermal imaging (ASTM, 2007). We used three cameras with different measurement
1. Targets in constant contact with water, ice, or wavelength and NTED in photographing model
snow should not be applicable for shooting; they structures. We set two cameras of different NTED in
must be dried at least for 24 hours. line and photographed model structures simultan­
2. The condition with over 25 km/h wind velocity eously. Photographing with the camera of the min­
should not be applicable for shooting. imum thermal sensitivity below 0.06℃ could not
3. Thermal imaging at night should be taken in fine detect the void. The thermal images of the camera
weather. with minimum thermal sensitivity below 0.025℃
showed successful detection of the void up to the
It also requires specific weather criteria for shoot­ depth of 40mm. We also learned that even the
ing as shown in the following table. camera with below 0.025℃ minimum thermal sensi­
Selecting the right infrared camera is also tivity could not detect the void in 60mm deep, due to
critical to achieving highly accurate detection signal-to-noise ratio (SNR) of the camera. The result
results. Some of the essential specifications for suggests that the camera with minimum thermal sen­
an infrared camera should be its pixel resolution, sitivity below 0.025℃ is applicable for inspection of
detecting element, measurement wavelength, the upper structure but not capable for inspection of
noise equivalent temperature difference (NETD), the lower structure with a much thicker cover.
or frame rates, in which NETD is the most
important. The infrared camera candidates we
evaluated for our investigation were roughly 2.2 Thermal environment suitable for infrared
classified into two groups regarding NETD; thermography method
groups of 0.02℃ and 0.06℃ NETD. To deter­ The infrared thermography method requires ther­
mine the selection criteria for the thermal mal flow to generate and be maintained for sev­
camera, we need to confirm the temperature dif­ eral hours inside the target object. Usually,
ference in the target floating/delamination areas. Equation 1 should calculate the required time for
Considering the fact that the typical cover of the the operation. However, it takes more time than
bridge’s upper structure is 40mm, selecting the that derived from Equation 1 before the change
camera capable of detecting cracks and voids begins to appear on the surface temperature of
40mm below the surface is essential. We built floating/delamination areas because of heat trans­
model structures to investigate thermal cameras’ mission from areas surrounding the target. In our
detection capability and weather conditions to experiment, the temperature change appeared in
select the one capable of detecting the cracks the condition with 2℃ difference continuously
and voids 40mm below. for an hour between object and air temperatures.

Table 1. The weather criteria for thermal imaging.

The weather condition of two hours

before the test start Judgement

Fair Possible
Fair with occasional clouds Possible
Cloudy with occasional fine weather Possible
Cloudy with temporary fine weather Impossible
Cloudy Impossible
Rain Impossible
Figure 1. Concrete model structures and locations of voids.

321
 Figure 3. Thermal image example of RC hollow bridge
with temperature difference.

Figure 2. Thermal images of concrete model structure. (a)

Thermal sensitivity below 0.06℃, (b) Thermal sensitivity
below 0.025℃.

where Figure 4. Concept of thermal change in the damaged area.

d: Heat transmission depth (m)
λ: Heat conductivity (W/mk)
t:Time of apperance (sec) change amount but also the size of the damage. In
ρ:Density (kg/m3) reality, however, the surface temperature of the
c:Specific heat (J/kg K) target concrete structure has a significant tempera­
ture gradient due to the difference in component
thickness, which makes it harder to detect damages
3 EXAMINATION OF THERMAL IMAGE than in stable temperature conditions.
ANALYSIS METHOD To eliminate the influence of temperature gradient
in concrete structures in our research, we calculate the
moving average of temperature distribution and sub­
3.1 Necessity and outline of the thermal image tract the target pixel temperature. It enables damage
analysis method detection in concrete structures under the influence of
Some of the issues in the infrared thermography temperature gradient; the detection result should be
measurement are the risks of overlooking or false free from the temperature gradient in the target
detection of floating and delamination by inspection structures.
workers. We have constructed a floating and delam­ After the abovementioned preprocessing, we have
ination auto-detection system based on machine constructed the system through machine learning to
learning. detect floating/delamination from thermal images.
However, it requires the preprocessing of the Machine learning requires training data, for which
target images for the analysis. Since the component we collected through infrared thermography meas­
thickness of concrete structures should differ, exter­ urement, as well as the full-scale hammering tests on
nal influences like solar radiation would cause tem­ RC, steel and box-girder bridges ranging from five
perature differences in components according to to 38 years old. As for the infrared thermography,
their thickness (Figure 3). we conducted the measurement at night on the day
Figure 4 Is a model to show the singular tempera­ we confirm with a diurnal range over 10℃.
ture distribution in different thermal conditions, Detecting singular temperature distribution in the
namely, detection in uniform temperature and detec­ infrared thermography does not always mean the
tion with a temperature gradient. When the target detection of floating/delamination. There are, for
structure is in uniform temperature, it is easy to instance, objects that affect the temperature distribu­
detect the damaged area not only its temperature tion of the target such as free limes piled up by

322
rainwater infiltration or roughness of the target sur­ 3.2.6 Free limes
face itself. The following Table 2 shows the list of Free limes adhered to the concrete surface appear as
factors that cause temperature irregularity, each of singular temperature irregularities in thermal detec­
which we will discuss in the following sections. tion. The difference in reflectance or thermal conduct­
ivity between free limes and healthy concrete surface
or the existence of gaps in between free limes and
3.2 Factors of temperature irregularity
concrete surface would be the cause of the detection.
3.2.1 Floating area 3.2.7 Color irregularity
The floating area, in this case, is the area in which The color irregularity causes a singular temperature
abnormal noise is evident in the hammering test, but irregularity detection in the thermal image.
the concrete surface is still intact. Comparing with
the thermal image of the healthy area where
a foreign object attached, the temperature distribu­ 3.3 Full-scale hammering test results summary
tion image on the floating area is not distinct. Figure 5 shows thermal images organized according
to characteristics in detection shapes. We collated
3.2.2 Delamination area organized images with hammering test results to con­
The delaminating area, in this case, is the area in firm significant characteristics in the shape of areas
which concrete flakes should fall when the hammer­ with singular temperature irregularity in the following
ing tests are conducted. Compared to the floating five types; 1) delamination, 2) floating, 3) slag, 4) for­
area thermal image, the image of this area distinct­ eign substances, and 5) healthy area. Table 3 shows
ively shows a singular temperature distribution. each characteristic. We determined that classification
by machine learning of these five types is possible if
3.2.3 Adhered slag we can calculate the value expressing the shape of the
The delaminating area, in this case, is the area in temperature irregularity area (shape features).
Slags remaining on the joint form marks after the ini­
tial concrete placement are mostly 2 to 5mm thick;
thus, they are detected distinctively as thin linear sin­ 3.4 Discrimination index using geometrical
gular temperature distributions that run along with features in thermal images
the joint form marks in infrared thermography. In this research, we have examined the image filtering
process that ternarizes the thermal images. Based on
3.2.4 Foreign substances the emphasized index, we set the threshold values of
When foreign substances like wood chips are mixed ternarized red, yellow, and blue as follows; over 0.11
in the concrete of the covering components, they are as red, over 0.08 and less than 0.11 as yellow, and
detected as temperature irregularities in infrared over 0.04 and less than 0.08 as blue, respectively.
thermography. It is impossible to judge from the
visual image whether the foreign object mixed
within because the mortar plastered on top conceals
the wood chip.

3.2.5 Repair marks

The standard repair procedure of the floating/delam­
ination of concrete due to the steel reinforcement
corrosion is to remove the floating/delamination
areas and fill the damaged sections with materials
like shrinkage-compensating mortar. If the materials
used for repair have different thermal conductivity
from the concrete, they appear as singular tempera­
ture irregularities.

Table 2. Factors of temperature irregularity.

Floating area

Delamination area
Adhered slag
Foreign substances
Repair marks
Free limes
Color irregularity Figure 5. Relationship between hammering test results and
image processing.

323
Table 3. Relationship between hammering test results and In addition to the above mentioned values, we use
detection image shape characteristics. the co-occurrence matrix (Partio et al. 2002; Chun
et al. 2013; Pathak and Barooah 2013), which is an
Unique temperature image texture analysis method capable of quantify­
Sounding result note
area shape characteristic ing contrast changes in images. The co-occurrence
matrix firstly derives the matrix that uses the P-value
The red area is off the center Knock of the target contrast in a specific position δ = (r, θ)
delamination
of the whole area down
away from the point i should be the contrast j, Pδ =
Abnormal sound The red area is at the center
floating (i, j), as its element (hereinafter referred to as sto­
only of the whole area
Square shape, high red False
chastic matrix) to calculate several features by the
foreign substances matrix as shown in Figure 7. Values of the stochastic
occupancy positive
The occupancy rate of red is matrix represent frequency to each sample image,
high and the shape is square, but practically we normalized them so that all num­
False bers should be 1.
slag(t=2～5mm) but the periphery is more
positive
complicated than foreign We calculated following 14 features in total by
substances the co-occurrence matrix mentioned above;
healthy area
High yellow area occupancy
False Angular second moment, Contrast, Correlation, Sum
(Reflection) positive of square variance, Inverse difference moment, Sum
average, Sum variance, Sum entropy, Entropy, Dif­
ference variance, Difference entropy, Information
Considering the results in the previous section, the measure of correlation 1, Information measure of
positional relation among each area of red, yellow, correlation 2, Maximal correlation coefficient. We
and blue is significant. When the red area appears also set the position value δ as r = 1, 2 and the angle
near the center of the blue area, the detected image is value θ as 0, 45, 90, 135, 180, 225, and 270. As for
possibly a floating area with abnormal sound alone. the contrast, we set it in 32 levels of gray, and we
On the other hand, when the red area appears at the generated the co-occurrence matrix.
location apart from the center of the blue area, the
detected area could be a delamination area. Thus, dis­
tances among the center of each area should be calcu­ 3.5 Deep neural network
lated as the feature for the discrimination index. We used hammering test results as the training data
Shapes of red, yellow, and blue areas are also sig­ for the deep neural network (LeCun et al. 2015;
nificant. In the thermal images of false detection, it Simões et al. 2012; Sainath et al. 2013) developed in
could be a floating area when its edge is smooth. We, this research based on shape features obtained from
therefore, examined shape features of red, yellow, and analyzed images in the previous section, ternarized
blue areas. The number of pixels represents circumfer­ areas, and their positional relations, and those
ence and dimension of the area. We calculated occupa­ obtained by co-occurrence matrix. The data amount
tion rate, degree of shape complexity, and circularity was 2,353 cases (hammering test results from 2008
level of each shape by the following equations, setting to 2010).
circumference as L, area as S, respectively. In this study, a deep neural network with one
S input layer, six hidden layers, and one output layer
Occupationrate ðOÞ ¼ was developed. The number of nodes in hidden
RðheightÞ � RðwidthÞ
layers are 453, 500, 600, 400, 300, and 100. The
L ð2Þ
Degreeofshapecomplexity ðCÞ ¼
S
4πA
Circularitylevel ðCLÞ ¼ 2
L

Figure 6. Concept of shape feature calculation of ternarized

thermal images. Figure 7. Co-occurrence matrix sample.

324
dropout layer is also sandwiched between these used for the training. The result of this study seems
layers. In this paper, the number of layers of the to enable to evaluate floating/delamination remotely,
hidden layer for high accuracy is investigated by and it is considered that this can contribute to drastic
grid search. It is probably not the optimum value, efficiency improvement of present inspection which
but it is almost the optimum value and there is no depends on hammering sound.
problem in practical use.

REFERENCES
4 RESULTS AND CONCLUSION
ASTM. 2007. Standard test method for detecting delamin­
ations in bridge decks using infrared thermography.
4.1 Results ASTM D 4788.
Table 4 shows the accuracy of the deep neural net­ Cheng, L., & Tian, G. Y. 2012. Comparison of nondestruc­
tive testing methods on detection of delaminations in
work model. The overall accuracy was 88.7 %
composites. Journal of sensors, 2012.
(2,086/2,353). In addition, 800 data which were not Chun, P., Funatani, K., Furukawa, S., & Ohga, M. 2013.
used for the learning were prepared, and the analysis Grade classification of corrosion damage on the surface
by this method was carried out. As the result, it was of weathering steel members by digital image
proven that the correct answer was put in the predic­ processing, In Proceedings of the Thirteenth East Asia-
tion ranking of 2 within the data over 97.3% Pacific Conference on Structural Engineering and Con­
(778/800). struction (EASEC-13) (pp. G–4). The Thirteenth East
Asia-Pacific Conference on Structural Engineering and
Construction (EASEC-13).
Table 4. Analysis results of training dataset. Kimura, T., Zhang, T., & Fukuda, H. 2019. A Proposal for
the Development of a Building Management System for
Extending the Lifespan of Housing Complexes in Japan.
Sustainability, 11(20), 5622.
LeCun, Y., Bengio, Y., & Hinton, G. 2015. Deep learning.
nature, 521(7553), 436–444.
Oh, T., Kee, S. H., Arndt, R. W., Popovics, J. S., & Zhu, J.
2012. Comparison of NDT methods for assessment of
a concrete bridge deck. Journal of Engineering Mechan­
ics, 139(3), 305–314.
Partio, M., Cramariuc, B., Gabbouj, M., & Visa, A. 2002.
Rock texture retrieval using gray level co-occurrence
matrix. In Proc. of 5th Nordic Signal Processing Sympo­
sium (Vol. 75).
Pathak, B., & Barooah, D. (2013). Texture analysis based
4.2 Conclusion on the gray-level co-occurrence matrix considering pos­
sible orientations. International Journal of Advanced
We have confirmed that the image filtering process Research in Electrical, Electronics and Instrumentation
on thermal images taken with infrared cameras Engineering, 2(9), 4206–4212.
should remove the temperature gradient generated in Plesu, R., Teodoriu, G., & Taranu, G. 2012. Infrared therm­
the structure to improve the damage detection rate. ography applications for building investigation. Buleti­
In addition, the image filtering process should be nul Institutului Politehnic Din Lasi. Sectia Constructii,
sufficient to deal with the influence of temperature Arhitectura, 58(1), 157.
differences around bridge appendages. Sainath, T. N., Mohamed, A. R., Kingsbury, B., &
In addition, the construction of deep neural net­ Ramabhadran, B. 2013. Deep convolutional neural net­
works for LVCSR. In 2013 IEEE international confer­
work model which evaluated existence and kind of
ence on acoustics, speech and signal processing (pp.
damage from the temperature change was also suc­ 8614–8618). IEEE.
ceeded. As an input of deep neural network model, Simões, N., Simões, I., Tadeu, A., & Serra, C. 2012. Evalu­
14 feature quantities obtained from co-occurrence ation of adhesive bonding of ceramic tiles using active
matrix, etc. were used. As the result, the accuracy of thermography. In Proceedings of 11th Quantitative
88.7% was realized. In addition, the high accuracy InfraRed Thermography conference, paper QIRT2012­
was obtained even in the data set which was not 362, Naples (Italy).

325
 Bridge Maintenance, Safety, Management, Life-Cycle Sustainability and Innovations –
Yokota & Frangopol (eds)
© 2021 Taylor & Francis Group, London, ISBN 978-0-367-23278-8

Reliability-based risk analysis for Maryland sign structures assessment

X. Liu, Y. Ye, C. Xu, Y. Zhu & C.C. Fu*
Department of Civil and Environmental Engineering, The Bridge Engineering Software and Technology (BEST)
Center, University of Maryland, College Park, MD, USA

ABSTRACT: Performance assessment is important for existing infrastructures since it can effectively
help conduct maintenance and keep structures in good condition. In this study, a deterministic risk
ranking approach and a reliability analysis approach are developed for the assessment of sign structure
for the state of Maryland, USA. The performance assessment of road-side infrastructures relates to the
effects of structural age, current and future traffic volumes in average annual daily traffic volume
(AADT), number of anchor bolts, structural member combined stress ratio (CSR) and fatigue stress.
By analyzing the database of sign structures provided by the Maryland State Highway Administration
(MDSHA), in which the inventory data and performing structural condition inspections are included,
an assessment of sign structures is conducted with additional sampling fatigue analyses based on the
AASHTO Specification. Eventually, ranking of all existing sign structures is generated in order for pri­
oritizing the structural replacement. In the existing sign structures, 736 out of 2451 fatigue galloping-
sensitive structures, belonging to five major categories, are focused on. The results of this study will
provide a basis for future inspection and maintenance of all infrastructures. Some risk assessment equa­
tions are derived in the analysis. Furthermore, the assessment model may be updated by the data from
future construction and inspection.

1 BACKGROUD INTRODUCTION In this study, a deterministic approach, which

takes relevant factors including fatigue into con­
Road-side infrastructures are part of the highway sideration, is developed to rank all Maryland
plan, and they play a significant role in the sign structures to estimate their intended per­
traffic management system. Their safety is formance. Then the risk ranking can be used to
essential for maintaining traffic flow and traffic prioritize the maintenance or replacement of
safety. signs. Moreover, the failure probability of the
As sign structures installed along highways, whole sign system is calculated to help estimate
they are repeatedly subjected to wind loads. the budget for system upgrading.
Design of sign structures by AASHTO considers
both wind in the extreme limit state as well as
the fatigue limit state. Due to these cyclical load­ 2 DATABASE OF SIGN STRUCTURES IN
ings, with the highly flexible and low damping MARYLAND
properties, the structures are highly susceptible to
wind-induced fatigue loading (Hosch, 2009). Maryland has 2451 sign structural in total to
Also, there are documented failures of these sys­ date, and they are categorized into 19 types by
tems due to fatigue in the United States (Zu, SHA, as shown in Table 1 and Figure 1. The
2013). Thus, the recent AASHTO ASD and MDSHA supplied Maryland SHA Traffic Struc­
LRFD Specifications (2015) have incorporated ture Inventory Inspection Management (TSIIM)
fatigue limit states, e.g. galloping, natural-wind database, and excel format file can be extracted
gust and truck-induced gust, into the evaluation, and downloaded for the study. There are five
where galloping is considered most critical in major categories of sign structures considered
fatigue design. Currently, there is no assessment fatigue galloping sensitive: the cantilever (CN1­
system that takes fatigue into account to evaluate 3), butterfly (CN4-5 may include 6), span-
these sign structures. cantilevered (CN4-5 may include 6), double

DOI: 10.1201/9780429279119-41
https://doi.org/10.1201/9780429279119-41

326
Table 1. MDSHA sign structure type and their quantity.

Type Chord No. Number Ratio

CM1 Tri 7 0.29%

CM2 Tri 0 0.00%
CM3 Box 0 0.00%
CM4 Tri 7 0.29%
CM5 Box 23 0.94%
CM6 Box 47 1.92%
CM7 Tri 11 0.45%
CN1 Single 74 3.02%
CN2 Double 631 25.74% Figure 2. Numbers and types of structures influenced by
CN3 Box 616 25.13% the three types of fatigue loads.
CN4 Double 5 0.20%
CN5 Box 14 0.57%
CN6 Pedestal 12 0.49%
OH1 Single 18 0.73%
OH2 Double 8 0.33%
OH3 Tri 27 1.10%
OH4 Tri 275 11.22%
OH5 Tri 72 2.94%
OH6 Box 604 24.64%
Total 2451 1

Figure 3. Arc-GIST application program with the

galloping-influenced structure information (736 in total).

3 RANKING SYSTEM

3.1 Ranking approach

Sign structure performance, parallel phrased to
bridge performance, encompasses how sign struc­
tures function and behave under factors of loads,
high winds, or ices. Sign structure design, AADT,
age, and maintenance history also play roles in per­
formance (Hooks & Frangopol, 2013). Performance
is usually associated with certain set of standards, in
our case, AASHTO Standard (ASD) or LRFD Speci­
fications are adopted, and performance can be meas­
Figure 1. MDSHA sign structure type in pie chart. ured against those standards.
In this study, a semi-qualitative method is used to
assess the performance of sign structures in the State
(CM4-5, may include CM6) and the overhead of Maryland. Four relevant factors, including struc­
sign structure (OH1-6). ture analysis result, number of anchor bolts, AADT
As mentioned, based on the team’s research, gal­ and sign structure age, are used to evaluate each
loping is the most critical fatigue load of the three structure. When evaluating a structure based on one
types (galloping, natural-wind gust and truck- factor, a risk rank is given according to its condition
induced gust) specified by AASHTO (2015). Thus, data (which can be obtained from database), ranging
out of the five major categories, box type can be from 1 to 5. This risk ranking system is similar to the
waived for both cantilever (CN1-6), cantilever over­ Pontis-based health index system used for US
hang (CM1-3) and overhead (OH1-4) types. Gallop­ bridges. Risk rank 1, or Condition State (CS) 1 in
ing-influenced sign structures are CN1, CN2, CN4, Pontis, means the least seriousness while risk rank 5
OH1 and OH2, shown as bar 1 in Figure 2. means the most seriousness. To obtain the risk rank
Five out of 19 types for a total 736 galloping- in the structural analysis factor, the TSIIM database
influenced structures are left for consideration, whose was studied and grouped and analyzed by Sign
location are shown in Figure 3 in the Arc-GIS. Arc- Bridge Analysis and Evaluation System (SABRE)
GIS is a special mapping program that the application software as a preliminary calculation. Then, the rank
is developed to show all sign structures on base maps. in each factor is assigned to a certain weight to obtain

327
Table 2. Relevant factors and their weights. Table 3. Risk rank of structural analysis.

Factor Weight Fatigue

Max
Span Post stress limit
CSR
CSR 50% (ksi)
Structure. 40% CSR Fatigue
Fatigue 50% Dia Thk Risk Risk
ft Pole Galloping
(in) (in) Rank Rank
Bolt - 30% 16 0.313 1.009 11.94 5 4
AADT - 20% 18 0.438 0.529 6.8 2 2
Age - 10% 25 20 0.500 0.371 4.86 1 1
24 0.438 0.288 3.91 1 1
25 0.500 0.228 3.15 1 1
the final rank score. Table 2 shows the relevant fac­ 18 0.438 0.594 8.27 3 3
tors that combine to affect sign structure performance 20 0.500 0.499 6.73 2 2
30
and their weights, respectively. 24 0.438 0.451 6.5 2 2
25 0.500 0.408 6.18 1 1
-Structural analysis 16 0.313 1.072 9.78 5 4
Out of all 627 cantilever double chord type (CN2) 18 0.438 0.809 9.64 4 4
structures, 68 of them are with square-tube post, and 20 0.500 0.752 9.84 3 4
559 are of round-tube CN2 structures. As the main 35
24 0.438 0.616 8.52 3 3
part of 736 galloping-influenced structures, they are 25 0.500 0.569 7.74 2 2
studied as a pioneer group. 26 0.531 0.761 8.56 3 3
Inspired by the Sign/luminaire Standard Drawing 18 0.500 0.776 8.72 4 3
MD 803.08 released in 2002, 559 round-tube CN2 20 0.500 1.156 18.95 5 5
structures are summarized into 20 typical samples, 24 0.438 0.991 13.46 4 5
according to the distribution of span lengths and post 50
25 0.500 0.999 14.32 5 5
sizes. Then, the 20 samples are modeled and analyzed 26 0.531 0.846 12.27 4 5
by SABRE and the combined stress ratio (CSR) of
posts and arms as well as their fatigue stress ranges
of galloping, natural-wind gust and truck-induced
gust are obtained. Since the sectional data of arm are Table 4. Risk rank of the number of anchored bolt in
not available in the database, the CSR of the arm is CN2.
omitted in the structural ranking. The fatigue stress
range due to galloping is considered as the most crit­ No. of bolt Risk Rank
ical one and used to rank the fatigue risk.
Next, sort the results from low to high and assign BOLT 4 5
risk rank from 1 to 5 accordingly, as shown in Table 3. 6 3
In this way, any existing sign structure of this 8 1
group can be classified as one of the 20 samples, and
its risk rank in structural analysis can be found by
checking this table. Developed in this study, Maryland AADT is avail­
able as a GIS feature layer as shown in Figure 4, and
-Number of bolts CN2 structures with AADT data is available in Arc-
Four bolt anchor rod arrangements cause a stress GIS as shown in Figure 5.
concentration in the pole, which increase 40% stress Table 5 shows the risk ranking of AADT, which
in the pole when bending is considered (Ginal & takes the Risk Category by Traffic Volume table in
Scott, 2004). More anchor bolts result less stress AASHTO and the distribution of AADT in Maryland
concentration. For the eight bolt anchor rod arrange­ into consideration.
ments, it has a much smoother stress distribution at
the base of the pole and the bending stresses in the -Age
anchor rods are lessened. The risk ranking of the The age of a sign structure serves as a basic meas­
number of bolts is shown in Table 4. ure of the aggregate service provided during its ser­
vice. Also, age data can be meaningful in terms of
-Average annual daily traffic volume (AADT) understanding the sign structure materials, design
Annual daily traffic volume (AADT) of the route standards and specifications, and construction pro­
where a sign structure situates is important in the cesses used in the construction. The overall perform­
analysis of a structure. Moreover, based on the ance of sign structures of a certain age might be
AASHTO LRFD development for wind, AADT is expected to be poorer than that of sign structures
considered in the risk analysis. It relates to the truck- built after more recent advancements. The risk rank
induced gust the number of people under risk. of age of structure is shown in Table 6.

328
Figure 4. Maryland AADT Traffic Volume Map.

Figure 6. The proportion of the structure under different

rank score.

updated UT test shows a structure is failing, then the

risk rank of it will be overridden to the highest one
no matter what rank it has before.

3.2 Ranking results

With the developed chart (Table 3), all round-tube
CN2 with effective data are ranked. Most of the signs
are in ranks 3 and 4, and 5.7% are in rank 5, which is
Figure 5. CN2 with AADT data in Arc-GIS. the most critical part need to be concerned. The pro­
portion of sign structures in 1-5 rank score is shown in
Figure 6.
Table 5. Risk rank of AADT.
4 RELIABILITY CALCULATION
AADT Risk Rank
0 -100 1 The reliability of an engineering system can be
100 - 1000 2 defined as the ability to fulfill its design purposes
AADT
1,000 – 10,000 3 defined as performance requirements for certain
10,000 – 50,000 4 time period and environmental conditions (Ayyub,
50,000+ 5 2014). An analytical performance-based reliability
assessment is performed to calculate the reliabil­
ity of the sign structure system. The Monte Carlo
simulation, as a widely used method in reliability
analysis, is used in this study.
Table 6. Risk rank of age. In the Monte Carlo simulation, samples of the
basic correlated variables are randomly drawn
Years in Service Risk Rank according to their corresponding probabilistic
characteristics and fed into the performance func­
0-10 1
tion, which in this study is expressed as Eq. 1
AGE 10-20 2
and Eq. 2 for CSR and fatigue load induced
20-30 3
stress range σf , respectively:
30-40 4
50+ 5

-UT test
An ultrasonic test inspection (UT test) records The failure event happens while it satisfies
from the year 2011 to 2019 are provided. In the UT CSR41:0 or σf 47:0ksi(49MPa). Therefore, the
test, the condition of anchor bolts was checked and probability of failure and reliability could be
the structures with missing bolts or cracking bolts expressed as:
were defined fail. Thus, the result of UT test is used
as an override of the ranking results that if the

329
 and 10.7m5S �15.2m) follow the lognormal distribu­
tion as shown in Figure 7.
The mean value and standard deviation for each
variable could be summarized based on the data of
559 samples, as listed in Table 7.

Where Nf is the number of simulation cycles for 4.3 Monte-Carlo Simulation

which the failure event happens in total N cycles of
After the determination of performance function as
a simulation. According to the law of large numbers,
well as the probability of distribution for variables.
the probability of failure approaches the true value
Quantities of groups of variables are drawn based
when N approaches infinity.
on the assumed distribution. Then, based on the
failure numbers Nf according to Eq. 1 and Eq. 2.
4.1 Performance Function There is:
In this study, span length, height, sign area and the
moment of inertia are selected as the variable in the
Monte Carlo simulation.
The most possible used CSR and fatigue load
induced stress range are calculated based on the
code Structural Supports for Highway Signs, Lumi­ R is the reliability of the whole sign structure
naries, and Traffic Signals. Thus, the equations can system.
be written as:

Where I is the moment of inertia; S is the span

length; L is the height of the post; A is the area of the
sign panel. Among the four variables, S, L, A is inde­
pendent random variables based on the 571 samples,
while I is the conditional random variables related to
the span length range (grouped by S �25ft,
25ft5S �30ft, 30ft5S �35ft, 35ft5S �50ft or
S �7.6m, 7.6m5S �9.1m, 9.1m5S �10.7m,
10.7m5S �15.2m). fa , fb , fv are the axial, bending
and shear stresses based on random variables;
MG ; MWG ; MTG are the moment ranges induced by gal­
loping, wind gust and truck gust fatigue loads on
random variables I; S; L; A.

4.2 Probability distribution of variables

The probability of variables is based on the bar
charts, which are generated from the database. Mean
value and standard deviation based on 559 samples
of the cantilever planar sign structures.
The probability density function (PDF) of the vari­
ables are generated through MATLAB, and their fitted
distribution curves are obtained. The span length S,
height L and part of the moment of inertia I (when
S �25ft and 25ft5S �30ft or S �7.6m and
7.6m5S �9.1m) are assumed to follow the normal
distribution, while the sign area A and I (when
305S �35ft and 355S �50ft or 9.1m5S �10.7m

330
 5 CONCLUSIONS

The deterministic ranking approach and reliability

analysis can be expanded to all existing MD sign
structures. The CN2 structures with square section
and other types of sign structures can be ranked by
a similar method. Furthermore, the assessment
model can be updated by the data of the newly-built
structures, updated AADT, and future inspection
data like UT test.
The reliability analysis provides a macro
evaluation of the whole sign structure system.
The failure probability shows the proportion of
sign structures with relatively high risk, which is
supposed to be consistent with the proportion of
structures ranked in high risk in the risk ranking
system.
The results of the assessment could help infra­
structure maintenance and management departments.
With the risk-based evaluations of the sign struc­
tures, the relevant organizations can do the following
(Hooks & Frangopol, 2013):
• Evaluate serviceability and durability.
• Evaluate traffic safety with the risk of sign struc­
ture system.
• Prioritize resource allocation within the sign
structure system.
• Improve system reliability, redundancy, and
accountability.
• Evaluate organization-wide policies and programs
such as the split between maintenance and capital
funds.

REFERENCES

Figure 7. PDF distributions and fitted curves of variables. AASHTO, 2015. LRFD Specifications for Structural Sup­
(a) Span length S PDF and fitted curve, (b) Height L PDF ports for Highway Signs, Luminaires, and Traffic
and fitted curve, (c) Sign area A PDF and fitted curve, (d) I Signals.
(S �25ft) PDF and fitted curve, (e) I (25ft5S �30ft) PDF Ayyub, B.M., 2014. Risk analysis in engineering and eco­
and fitted curve, (f) I (30ft5S �35ft) PDF and fitted curve, nomics. Chapman and Hall/CRC.
(g) I (35ft5S �50ft) PDF and fitted curve. Fu, C.C., 2019, Sign Bridge Analysis and Evaluation System
(SABRE) software, the Bridge Engineering Software and
Technology (BEST) center, University of Maryland, Col­
lege Park, MD (http://best.umd.edu/sabre/)
Ginal, Scott, et al., 2004. Structural analysis of sign bridge
structures and luminaire supports. Wisconsin Highway
Table 7. Probability distribution of variables. Research Program.
Hooks, J.M. and Frangopol, D.M., 2013. LTBP bridge per­
Standard formance primer (No. FHWA-HRT-13-051). United
Variable Units Type Mean Deviation States. Federal Highway Administration. Office of Infra­
structure Research and Development.
S ft Normal 29.58 7.95 Hosch, I.E., 2009. Design of highway overhead
L ft Normal 20.26 1.90 cantilever-type sign support structures for fatigue loads
Vol. 71. No. 01.
A ft2 Lognormal 178.11 81.44
Kacin, J.A., 2009. Fatigue life estimation of a highway sign
I(S �25ft) in4 Lognormal 1065.50 967.70 structure (Doctoral dissertation, University of
I(25ft5S �30ft) in4 Lognormal 1498.05 908.92 Pittsburgh).
I(30ft5S �35ft) in4 Normal 1938.90 792.81 Puckett, J.A., Garlich, M.G., Nowak, A.A. and Barker, M.,
I(35ft5S �50ft) in4 Normal 2208.37 722.64 2014. Development and Calibration of AASHTO LRFD
Specifications for Structural Supports for Highway

331
 Signs, Luminaires, and Traffic Signals (No. Project Zonta, D., Zandonini, R. and Bortot, F., 2007. A
10–80). reliability-based bridge management concept. Structures
Stewart, M.G., Rosowsky, D.V. and Val, D.V., 2001. Reli- & Infrastructure Engineering, 3(3), pp.215–235.
ability-based bridge assessment using risk-ranking deci- Zu, G., 2013. Calibration of Fatigue Design Wind Pressure
sion analysis. Structural Safety, 23(4), pp.397–405. for Sign, Luminaire, and Traffic Signal Support.

332
 Bridge Maintenance, Safety, Management, Life-Cycle Sustainability and Innovations –
Yokota & Frangopol (eds)
© 2021 Taylor & Francis Group, London, ISBN 978-0-367-23278-8

A framework for addressing the uncertainty of factors influencing the

overall deterioration of existing concrete structures
P. Miao & Y. Zhang
Graduate School of Engineering, Hokkaido University, Sapporo, Japan

H. Yokota
Faculty of Engineering, Hokkaido University, Sapporo, Japan

ABSTRACT: The deterioration of concrete structures is influenced by various factors. However, neither the
interactions among those factors nor their impacts are explicitly determined. Nowadays, deterioration assessment
for concrete structures mainly relies on periodic inspections and on data in maintenance databases that are collected
during maintenance. However, superficial analyses of databases are insufficient when maintenance strategies are
formulated and/or proper intervention work is implemented. In addition, the factors that affect deterioration show
different characteristics and those differences should be taken into account. Usually, the inspection result is indi­
cated by an overall deterioration grade for the concrete structure. However, correlations between potential factors
influencing deterioration and the deterioration itself remain unknown. This paper proposes a framework for evalu­
ating the impacts of potential influencing factors on deterioration of concrete structures. A neural network was
combined with the Shapley value method to predict deterioration grades, and the factors affecting deterioration
were qualitatively and quantitatively calculated. Moreover, the “black box problem” of a neural network was
avoided effectively through the adoption of that framework, enabling the uncertainty of these factors to be
addressed. In practice, the framework can help to clarify factors that promote or suppress bridge deterioration and
can assist in the development of corresponding maintenance strategies.

1 INTRODUCTION processing, to predict deterioration and to quantita­

tively and qualitatively determine the factors that
Concrete structures that are in use for many years affect that deterioration.
suffer varying degrees of deterioration. Various fac­ An artificial neural network is used to infer the rela­
tors affect this deterioration, and many studies have tionships between multivariate independent variables
been conducted to evaluate the impacts of these fac­ and dependent variables in a large database (Rumelhart
tors on deterioration (Li & Zheng, 2005). However, et al. 1986). The noteworthy feature of a neural net­
given the complexity of concrete structures and their work is its ability to learn. In the process of establish­
environments, neither the combined impact nor the ing a neural network model, the network processes the
individual impacts of these factors have been expli­ input information, compares the output with the ideal
citly determined. Fortunately, engineers inspect and response, and calculates the errors; then it modifies the
assess the conditions of structures and record these weights of connections according to the errors (Figure
results at regular intervals. Over time, numerous 1). The attributes of structures can be used to predict
databases with large numbers of maintenance items deterioration grades by adjusting the weighting of con­
are obtained. However, the application of mainten­ nections. In other words, a neural network can establish
ance data to intervention work and to the develop­ the optimal relationship between independent variables
ment of strategies for inspection activities remains and dependent variables by altering the connection
insufficient. Therefore, there is untapped potential to weights between the processing units according to cer­
utilize maintenance databases for deterioration pre­ tain learning algorithms. Finally, an output is obtained
diction and for the development of inspection strat­ whose maximum probability corresponds to the inputs.
egies. For a thoughtful analysis of these data, an In maintenance engineering, it can help us to establish
inspection database was subjected to new data pro­ a relationship between the structural and environmental
cessing methods, including neural network properties and the overall deterioration of a structure.

DOI: 10.1201/9780429279119-42
https://doi.org/10.1201/9780429279119-42

333
 2 PROPOSED FRAMEWORK

Maintenance is essential to keep structures sound and

serviceable. Nowadays, condition assessment of exist­
ing structures is carried out mainly through periodic
inspections, such as every five years in Japan (Minis­
try of Land Infrastructure and Transport, 2014), every
two years in US (Ryan et al. 2006), and every 1 to 3
years in South Korea (Park & Yi, 2019). Based on
the assessment results, necessary interventions or fur­
ther inspection activities are scheduled. In addition,
data processing methods, such as those using a neural
network, have been applied to inspection (Kuo et al.
1992), evaluation (Hodhod & Ahmed, 2013), and
Figure 1. Structure of a neural network used for prediction. retrofit (Mikami et al. 1994). Inspired by these stud­
ies, we tried to integrate an inspection database and
a neural network toward predicting deterioration to
Because relationships trained by a neural net­ existing structures, based on which intervention work
work reflect the overall behavior of outputs and and inspection strategies can be made.
inputs, it is difficult to understand the relationship As mentioned earlier, relationships between sev­
between all the inputs and outputs for a specific eral inputs and one output can be established by
prediction; therefore, the neural network becomes a neural network in order to make predictions. How­
a “black box” (Kushida et al. 1997). The black ever, Kushida et al. (1997) and Miao et al. (2019)
box nature of the neural network makes it diffi­ noted that neural networks can give rise to black box
cult to understand the contribution of a specific problems. Although a neural network model can pre­
input (factor) to the output. In other words, even dict deterioration grades under given environmental
though a neural network approach can provide and structural properties for a concrete structure, the
prediction and can show better performance prediction model provides only one prediction value
(Hodhod et al. 2013), its application in actual without any explanation because of the black box
engineering is limited because its predictions are problem, which makes it difficult for engineers to
difficult to interpret. This lack of interpretability trust the prediction and to understand what factors
has thus far limited the application of neural net­ affect the prediction.
work methods in the practical maintenance of Understanding what factors drive a prediction is
structures. essential for determining targeted intervention work
Therefore, in this research, we want to solve the in the maintenance of a structure. To give the predic­
interpretability problem of a neural network model tion model interpretability, a typical consideration is
in order to explain how a particular prediction is to limit the complexity of the model (Caruana et al.
made for a concrete structure. To formulate 2015). For this reason, engineers avoid using com­
a deterioration prediction model with the best per­ plex but more accurate models (e.g., neural net­
formance and to make a convincing explanation for works) in favor of using simpler but less accurate
what factors most contribute to the deterioration, the interpretable models (e.g., linear models).
database must consist of high-quality data. There­ However, it is unfortunate to sacrifice the advan­
fore, before an inspection database is used, its data tages of neural network models only to gain interpret-
must be checked carefully. ability. Therefore, we want to preserve the predictive
This paper introduces the development and veri­ ability of a neural network model while solving its
fication of a proposed framework that can be used black box problem, so as to provide both predictions
to predict the risk of concrete structure deterioration and explanations of how a specific prediction value is
and to state a quantitative and qualitative interpret­ derived. A framework was proposed using an inspec­
ation for the impacts of potential factors on con­ tion database and a neural network to predict the over­
crete structure deterioration. First, an inspection all deterioration of a bridge structure, and using the
database and a neural network were used to estab­ Shapley value method to account for the factors that
lish a prediction model whereby the relationships contribute to deterioration. The framework consists of
between potential influencing factors and the over­ two parts: a predictive part and an explanatory part
all deterioration of concrete structures were estab­ (Figure 2).
lished. To make the prediction model more intuitive
and easier to understand, the Shapley value method
was used to calculate the importance of each factor 2.1 Predictive part
towards enabling the predictions to be explained. To predict deterioration, the predictive part is applied
For some factors, the estimated values were com­ to an inspection database towards establishing a neural
pared to the measured values, and the estimated network model. Since the database contains various
impacts were evaluated. maintenance data, data with high quality should be
334
 results will be more valuable. To interpret the predic­
tions, all the factors that may cause deterioration and
the relative importance of each factor for a concrete
structure should be provided, based on which corres­
ponding intervention work or further inspection activ­
ities can be scheduled.
This is an important step towards applying neural
networks to the practical maintenance of concrete
structures, because the explanatory part converts
abstract and inexplicable predictions into the relative
importance of factors, which are easy to understand
and act on. A proper interpretation method needs to
fulfill the purpose mentioned above, and the Shapley
value method was chosen because if all factors
acting together affect the deterioration, then it is an
excellent way of calculating the relative importance
of each factor (Littlechild & Owen, 1973).
We briefly introduce the Shapley value method.
We start by assuming that there are n potential
impact factors. The set of all factors is represented
by N ¼ f1; 2; . . . ng, and any subset (permutation) S
of N is a coalition. The characteristic function v
maps all subsets into variables used in actual main­
tenance, which in our case is the overall deterior­
ation of the concrete structure. The function v has
the following meaning: If S is a permutation (coali­
tion) of all factors, then vðSÞ, the value of
coalition S, describes the total impact of all factors
in coalition S (Littlechild & Owen, 1973). The
importance ’i ðvÞ of factoriis determined as follows:

Figure 2. Framework for interpretable predictions.

extracted and classified into two categories: the struc­

ture and environment properties that may influence the
deterioration of concrete structures, and the grades of
deterioration. The attributes of structures and environ­
ments are then applied as input values, and the struc­
tural deterioration is used as an output to train the Where S � N=fig indicates all sets S of N not
neural network. After the prediction model with the including factor i; |S| is the number of factors in
best performance is obtained, it can be used for deteri­ coalition S; ðn _ jSj _ 1Þ!jSj!=n! is the probability
oration prediction. By continuously integrating new of S in N. Equation (1) can be interpreted as follows.
inspection data and updating the prediction model, that There are many coalitions (permutations) for set N.
model will move closer to an actual situation. Details Each factor’s impact value vðS [figÞ _ vðSÞ in
about using an inspection database to build a neural permutation S was first computed, and then the aver­
network model can be found in Miao et al. (2019). age impact value of each factor was calculated
according to the probability of S in N. The sum of
the impacts for all factors is the final deterioration
2.2 Explanatory part ’ðvÞ of the concrete structure. Therefore, the total
To overcome the black box problem of the neural net­ impact of all factors is as follows:
work, an explanatory model for interpreting the predic­
tions was designed. However, conveying information
to engineers in a manner commonly used in engineer­
ing remains a crucial challenge, such as telling the
engineer what factors affect the deterioration. The
The Shapley value method can calculate the com­
grade of deterioration can be predicted by the predic­
bined importance of all factors, regardless of the com­
tion model given in Section 2.1. Nevertheless, if clear
plexity of the prediction model. Therefore, it can
information is given about what factors, such as tem­
explain any prediction model (Littlechild & Owen,
perature, chloride exposure, snowfall, and load, con­
1973). When a feature is chosen, the calculated value
tribute to specific deterioration, then the prediction

335
of importance for a particular factor represents the 3.1 Data preparation
difference between the output of the prediction model
Features that may affect degradation are regarded as
when we observe the feature (such as a bridge width)
inputs in the prediction model. These features are
and when we do not observe the feature (such as
extracted or derived from a database of bridge
when the bridge width is unknown). Therefore, the
inspection results collected over a long time span,
estimated importance is related to the magnitude of
such as bridge length, bridge width, and years in ser­
difference in the prediction values when a feature is
vice, but temperature and carbon dioxide concentra­
considered versus when it is not considered (Lund­
tion were derived from bridge location information.
berg et al. 2018). The value of importance of a given
The potential deterioration factors include bridge
feature does not imply a causal relationship; there­
geometry factors, environmental factors, and loading
fore, it does not represent a complete diagnosis of
conditions, among others (Table 1).
deterioration for a concrete structure. However, the
Because it is difficult to count the daily traffic
importance of features enables an engineer to better
volume exactly, the traffic volume is classified into
formulate an intervention or maintenance strategy by
the ten categories listed in Table 2.
knowing which attributes of the concrete structure
The output is the overall deterioration grade of
promote or suppress the current deterioration pre­
the bridge. Based on Inspection Guidelines (Ministry
dicted by the prediction model.
of Land Infrastructure and Transport of Japan,
Based on the above theory, we present the specific
2014), the deterioration grades were categorized as
steps to achieve the interpretation of the predictions,
follows: sound (Grade I), preventive maintenance
which is the main focus of this paper. For given
required (Grade II), prompt action required (Grade
structural and environmental properties, prediction
III), and emergency action required (Grade IV).
grades of deterioration can be obtained for any con­
However, no structures with Grade IV deterioration
crete structure (Section 2.1), with some predictions
were found in the database, because intervention
being successful (i.e., the same grade as that evalu­
work had been promptly done on those structures.
ated by the engineer) and others being unsuccessful.
All bridges with successful predictions were extracted
to form a new database. In computing the importance 3.2 Prediction model establishment
of each feature, the new database, the prediction
Using the attributes in Table 1 as potential influen­
model, and the explainer (Shapley value method) are
cing factors and three deterioration grades as predic­
used (Figure 2). The database and prediction model
tion results, a neural network prediction model was
are required because we have established the relation­
constructed. The neural network has an input layer
ships between inputs and outputs through the neural
of nine units, a hidden layer of nine units, and an
network. The calculation of estimated importance
uses the measured value in the database to identify
the relationship found by the neural network and to
Table 1. Possible influencing factors.
identify the corresponding impacts.
The previous section proposed a framework for Attributes (units)
combining a neural network and the Shapley value
method in order to provide a reasonable interpretation Bridge geometry Length (m), width (m)
of the predictions of the neural network model. The factors
framework allows the neural network to predict con­ Environment Elevation (m), Snowfall (cm), Yearly
crete structural deterioration while still providing an factors highest and lowest temperatures (°C),
intuitive explanation of the factors that promote or sup­ Carbon dioxide concentration (ppm)
press such deterioration, towards enabling concrete Loading Traffic volume* (vehicles/day)
structure maintenance and policy development. The condition
framework implementation process will be elaborated Other Years in service (years)
for a specific case. * Statistics on traffic volume include large and small vehicles
(Ministry of Land Infrastructure and Transport, 2010).

3 CASE STUDY

A bridge inspection database was employed to Table 2. Traffic grade classification.

verify the effectiveness of the framework men­
Traffic volume Traffic volume
tioned earlier. It was hope that the validation results Grade (Vehicles/day) Grade (Vehicles/day)
would reveal deficiencies in the framework, based
on which the framework would be optimized fur­ 1 <5000 6 25000r30000
ther. The database includes 96 bridges in various 2 5000r10000 7 30000r35000
locations around Sapporo. Below, the entire imple­ 3 10000r15000 8 35000r40000
mentation process will be interpreted from three 4 15000r20000 9 40000r45000
aspects: data preparation, prediction model estab­ 5 20000r25000 10 >45000
lishment, and feature importance estimation.
336
output layer of three units (Figure 1). According to importance of years in service for bridge B is 0.844,
the prediction model, if the structural and environ­ the recorded years in service is 2; therefore, the
mental characteristics of a specific concrete bridge number 2 is labeled. In the figure, the magnitude of
are known, the overall deterioration grade of the each feature indicates the impact of importance. The
bridge can be predicted. features with positive values are those that increase
the deterioration risk, and the features with negative
values are those that decrease the deterioration risk.
3.3 Feature importance estimation
For both bridges A and B, traffic volume, years in
Using the prediction model described in Section 3.2, service, snowfall, lowest temperature, and carbon
successful predictions were made for 93 of the 96 dioxide concentration promoted deterioration, and
bridges in the database (Table 3). Therefore, a new elevation suppressed deterioration, as shown in
database of 93 bridges was set up. Next the Shapley Figure 3. Other features, such as bridge length, bridge
value method was applied to calculate the value of width, and highest temperature, show no consistent
importance for each feature from the new database. promotion or suppression. Specifically, deterioration
was promoted for bridge A and suppressed for bridge
3.3.1 Estimated importance of each feature for B with increases in bridge length and highest tem­
a given structure perature. Bridge width had the opposite effect.
The results of the Shapley value method represent The magnitude of each feature also differs between
the relative impact of each factor, which explains bridges A and B, especially for years in service and
how the prediction model gives a particular deterior­ bridge length. The lengths of bridges A and B are
ation grade. To demonstrate the framework’s inter­ 5.25m and 28.8m respectively. Up to inspection,
pretability for the factors that result in deterioration, bridge B has an in-service period of 2 years, and
bridge A and bridge B were randomly selected. For bridge A has an in-service period of 8 years. How­
bridges A and B, the deterioration is Grade I, and the ever, the results show that the feature of years in ser­
calculated importance is shown in Figure 3. To show vice has less impact on bridge A than on bridge
the correlation between the estimated importance B. This is because the Shapley value method balances
and the measured environmental/structural value of all factors in the calculation of each factor’s import­
bridge factors, the actual importance is marked on ance. In this condition, bridge B is longer than bridge
the side of the bar chart. For example, the estimated A; therefore, the estimated value of importance for
bridge length for bridge B is negative considering the
balance between years in service and length. The phe­
Table 3. Performance of the prediction model. nomenon means that even though the deterioration
Predicted grade grades of bridge A and bridge B are Grade I, the cal­
culated relative impact differ for each factor because
Actual grade 1 2 3
they are exposed to different environments.
1 30 1 0 Then, bridge C with Grade II deterioration was
2 0 27 1 selected from the database and its attributes were com­
3 0 1 36 pared to those of bridge B as shown in Figure 4. The
Accuracy 96%

Figure 4. Comparison of influencing factors for bridges

Figure 3. Estimated value of importance for bridges A and B. B and C.

337
figure does not show the estimated importance for fea­ increase in traffic volume, years in service, snowfall,
tures because bridges B and C have different deterior­ lowest temperature, and carbon dioxide concentra­
ation grades. Bridges B and C have almost the same tion (Figure 5). The difference is that deterioration
attributes, except for bridge length and elevation. How­ decreases with increase in bridge width. Elevation
ever, the deterioration grade of bridge C indicates and length are found to not relate to deterioration.
more severe deterioration than that of bridge B. The Of all the features that promote deterioration,
situation is contrary to that observed in Figure 3, years in service is the most significant for each
where deterioration was suppressed with increases in grade. It was also found that traffic volume and
bridge length and elevation. Therefore, it is still snowfall are two other essential features in promoting
unknown whether bridge length and elevation promote deterioration. By comparing the differences in the
or suppress deterioration. The representation in Figures distribution of estimated importance for each feature
3 and 4 shows that for a single bridge, some features between different grades, it can be found that the
show the consistent promotion of deterioration, average measured environmental/structural value
whereas the effects of other features are unknown. Of (labeled value) and the estimated importance value
the features that promote deterioration, the one with are almost the same for carbon dioxide, lowest tem­
the greatest influence is years in service. Other fea­ perature, and highest temperature. Moreover, the
tures, such as traffic volume, carbon dioxide concentra­ impacts of the three features are relatively small. For
tion, snowfall, and lowest temperature have relatively snowfall, the average measured environmental/struc­
small impacts. tural value and the estimated importance are also
nearly identical, and its effect is relatively significant.
3.3.2 Estimated importance of each grade Years in service and traffic volume are two fea­
To compare differences between grades, the average tures that accelerate deterioration but that show
importance of each factor for its respective grade was mutually different tendencies. For Grades II and
computed, and the results are shown in Figure 5. Fea­ III, the greater is the years in service and the
tures with positive values are those that increase the greater is the traffic volume, the greater is the
deterioration risk, and features with negative values estimated importance. For Grades I and II, even
are those that suppress deterioration. Average values though the years in service are almost the same
for bridge attributes are labeled on the side of the bar and even though Grade II has slightly more years
chart. in service, the impact of years in service is less
Unlike the explanations shown in Figure 3, which for Grade II than for Grade I. With respect to the
are specific to individual bridges, Figure 5 shows the traffic volume, deterioration decreases with
average estimated importance for bridges with the increase in traffic volume, which is also abnor­
same deterioration grade. However, the estimated mal. Considering the estimated importance of traf­
value of importance for each grade shows the same fic volume and years in service, no definite
tendency as in Figure 3: Deterioration increases with relationship between deterioration and years in
service can be found because interventions had
been performed or some potential factors might
not be included in our model.
Similarly, the importance of each factor averaged
for all bridges was calculated (Figure 6). The

Figure 5. Average importance of each feature, broken

down by grade (Grades I, II and III). Figure 6. Estimated value of importance for all bridges.

338
average importance for different grades tends to be
the same for Figures 3 and 5. Of all the features,
deterioration significantly increased with increases
in three factors: years in service, traffic volume, and
snowfall. In general, there is a reasonable agreement
between the estimated importance of each feature
and factors already known to affect deterioration,
which means that the explanatory part can identify
features influencing deterioration. However, the spe­
cific importance of features may produce essential
differences. Since the Shapley value considers the
contribution of all features in the calculation, the
results differ from the actual situation, in which
some factors may be particularly significant and the
influence of other factors may be ignored. Moreover,
maintenance work is done for some bridges to slow
or prevent further deterioration, which is difficult to
incorporate into the calculation.
In practice, the framework can provide predicted
information to engineers when the deterioration risk
for a concrete structure is high and can inform engin­
eers of the features that drive the risk. This informed
risk prediction enables the engineer to schedule
appropriate actions to avoid further deterioration. In
addition, the explanatory model can provide infor­
mation on the relative importance, although the
accuracy of these values remains to be verified.

3.3.3 Comparison between estimated importance

and measured importance
Although six features promote deterioration (Figure
6), three of these features were chosen based on the
average estimated importance of the features. Figure
7 illustrates the relationship between the estimated
importance and the corresponding bridge attribute
value. Figures 6 and 7 show the variations in all the
measured values of bridge features and the estimated
Figure 7. A feature’s importance with respect to
values of importance for each feature. The blue his­
deterioration.
tograms in each plot indicate the distribution of
values for a feature in the successful prediction data­
base. The orange lines represent the estimated values deterioration predictions and quantitative insights
of importance for a given feature. into the prediction model and can give a more spe­
Unfortunately, among the three main features, no cific value for features that promote deterioration.
significant pattern is found between bridge attributes However, the accuracy of estimated importance still
and estimated importance. This may be due to the needs improvement.
lack of some data samples. For example, only a few
bridges with heavy traffic volume are included in the
database. In addition, differences in the design, con­ 4 CONCLUSIONS
struction, and maintenance of different bridges could
also result in the irregularities between environmen­ A framework was designed to integrate a damage
tal/structural attributes and estimated importance, prediction model with an interpretable method. The
such as the feature of years in service having no rela­ explanatory part of the framework offers insights
tionship to deterioration. It is also possible that the into the prediction model and provides the exact
current database lacks generalizability because it impact value of each feature that may affect the
includes only bridges in Sapporo. Of course, consid­ deterioration of a concrete structure. From the valid­
ering the complexity of the environment, it is chal­ ation results of an inspection database, the following
lenging to simulate reality by using only these conclusion can be drawn.
features. The results show that there is no absolute (1) The combination of predictive ability and
relationship between the three features and deterior­ interpretability can effectively avoid the black box
ation. The representative features illustrate that the problem and can maintain the predictive advantages
framework can provide qualitative explanations for of the neural network model.

339
 (2) From the estimated importance of features that diffusivity in high performance concrete. HBRC Jour­
may influence deterioration, it was found that the nal, 9(1): 15–21.
features of years in service, traffic volume, snowfall, Ichiishi, T. 2014. Game theory for economic analysis.
lowest temperature, and carbon dioxide concentra­ Elsevier.
tion promote deterioration, while the impacts of Kuo, S.S., Davidson, T.E. & Fiji, L.M. 1992. Development
other features remain unknown. In addition, the rela­ of Computer Automated Bridge Inspection Process.
tive importance of each feature that has the potential Computing in Civil Engineering and Geographic Infor­
mation Systems Symposium, 794–801. Texas, United
to contribute to bridge deterioration can be obtained. States.
In practice, the framework allows engineers to make Kushida, M., Miyamoto, A. & Kinoshita, K. 1997. Devel­
possible predictions while also understanding how opment of concrete bridge rating prototype expert
those predictions are made. system with machine learning. Journal of Computing in
Although our exercise at developing a framework Civil Engineering, 11(4): 238–247.
to predict and interpret deterioration is promising, it Li, C.Q. & Zheng, J.J. 2005. Propagation of reinforcement
should still be regarded as an initial attempt. In this corrosion in concrete and its effects on structural
first attempt, the types of concrete bridge structures deterioration. Magazine of Concrete Research, 57(5):
261–271.
were not categorized to assess deterioration predic­
Littlechild, S.C. & Owen, G. 1973. A simple expression for
tions. For this reason, the explanations of features the Shapley value in a special case. Management Sci­
that contribute to deterioration had to be somewhat ence, 20(3): 370–372.
generic. Additionally, several other factors are asso­ Lundberg, S.M., Nair, B., Vavilala, M.S., Horibe, M.,
ciated with deterioration, but are not directly con­ Eisses, M.J., Adams, T. & Lee, S.I. 2018. Explainable
sidered, such as deicing salt application, design machine-learning predictions for the prevention of hyp­
strength, and intervention records. Because the oxaemia during surgery. Nature Biomedical Engineer­
values of these features are not fully captured in the ing, 2(10), 749.
inspection databases nor are they obtainable in other McClelland, J.L., Rumelhart, D.E. & PDP Research Group.
ways, the impacts of these features were not evalu­ 1987. Parallel distributed processing (Vol. 2). Cam­
bridge, MA: MIT Press.
ated. Also, in assessing a feature’s importance, the Miao, P., Yokota, H., Zhang, Y. & Song, W. 2019. Predic­
Shapley value method considers all possible influen­ tion-based maintenance of concrete structures by using
cing factors rather than focusing on several major an artificial neural network. The 3rd ACF Symposium,
factors. Therefore, the estimated importance bal­ Sapporo, Japan, 10-11 September, 2019.
ances all attributes of the bridge, which will affect Mikami, I., Tanaka, S. & Kurachi, A. 1994. Expert system
the accuracy of the assessment for some features. with learning ability for retrofitting steel bridges. Jour­
For improved interpretations of features that contrib­ nal of Computing in Civil Engineering, 8(1): 88–102.
ute to deterioration, future attempts could focus on Ministry of Land Infrastructure and Transport. 2010.
the deterioration of a specific type of concrete struc­ Guidelines for General Traffic Survey. Viewed October
2019. <www.mlit.go.jp/road/census/h22-1/data/
ture. Another future improvement will be the inte­ kasyorep.pdf>. (In Japanese).
gration of more features and the inclusion of more Ministry of Land Infrastructure and Transport. 2014.
bridges distributed in different places. Guidelines for Regular Inspection of Road Bridges. (In
Japanese).
Park, H. & Yi, W. 2019. Proposed Improvements to the
REFERENCES Safety Inspection System: An Analysis of the Current
Status of Building Disaster Accidents and Safety Inspec­
Caruana, R., Lou, Y., Gehrke, J., Koch, P., Sturm, M. & tion Systems. Journal of the Korean Society of Hazard
Elhadad, N. 2015. Intelligible models for healthcare: Mitigation, 19(5): 11–21.
Predicting pneumonia risk and hospital 30-day Rumelhart, D.E., Hinton, G.E. & Williams, R.J. 1988.
readmission. Proceedings of the 21th ACM SIGKDD Learning representations by back-propagating errors.
International Conference on Knowledge Discovery and Cognitive modeling, 5(3): 1.
Data Mining, 1721–1730. ACM. Ryan, T.W., Hartle, R.A. & Mann, J.E. 2006. Bridge
Hodhod, O.A. & Ahmed, H.I. 2013. Developing an artifi­ inspector’s reference manual. Report No. FHWA NHI:
cial neural network model to evaluate chloride 03-001.

340
 Bridge Maintenance, Safety, Management, Life-Cycle Sustainability and Innovations –
Yokota & Frangopol (eds)
© 2021 Taylor & Francis Group, London, ISBN 978-0-367-23278-8

Applying fully convolutional neural networks for corrosion semantic

segmentation for steel bridges: The use of U-Net
S.-K. Chen, I-F. Huang & P.-H. Chen
National Taiwan University, Taipei, Taiwan

ABSTRACT: As the weather in Taiwan is warm and humid most of the time, steel bridges get rusted easily.
Nowadays, bridges are significant infrastructure in most countries, and, thus, it is crucial to come up with an effect­
ive corrosion detection method for steel bridge inspection, so as to maintain the health of steel bridges and reduce
the lifecycle costs of them at the same time. Browsing past research efforts, there were a number of image process­
ing techniques (IPTs) proposed for quick and effective rust image recognition. A crucial issue on rust recognition
is to distinguish real rust corrosion spots or areas from noises or patterns that look like rust. Also, the types of rust
and different rust colors would affect the accuracy of rust recognition. In view of the above issues, a fully convolu­
tional neural network, namely U-Net, will be explored in this paper to develop an image semantic segmentation
model, which will be able to deal with a wide range of rust image recognition.

1 INTRODUCTION technology in the field of civil engineering enhances

the feasibility of rust image identification.
Taiwan’s geographical location is located at the junc­ In the past, most of the algorithms proposed by
tion of the plate, the terrain is violent, resulting in the institute were “color” or “gradient” as the basis
diverse and complex terrain, in terms of transporta­ of the edge detection method, the instruction period
tion often rely on bridges as a transportation connec­ is fast, but the type of rust image that it can handle is
tion, thus showing the importance of bridge projects more limited, and it is impossible to determine
for Taiwan. In 2016, the total number of bridges in whether the observed image is rusty, distance from
Taiwan exceeds 28,000, and the bridge age of more the real application still needs to be overcome. To
than 30 years accounted for more than 50%, and this end, this study hopes to build an image recogni­
located in the subtropical and geographical environ­ tion model that can be applied to most types of rust
ment surrounded by the sea, high temperature, high through the fully connected convolutional neural net­
humidity climate makes steel structure rust easy, in work in deep learning technology, and provide
addition, the current domestic steel structure con­ a fast, objective and practical detection aid.
struction site and after assembly of corrosion preven­ As the rust images used as training data in this
tion measures are not strict. The specifications and study are self-photographed and collected, the
technical manuals make it easy to ignore the risks of number is about 200, the size is 4032 x 3024, the
severe corrosion environments and failure to prop­ film material for steel factory steel and steel structure
erly protect the environment, resulting in severe cor­ rust, covering a limited variety of rust, and most of
rosion of structural objects affecting overall safety the color is different from the rust color, coupled
and use. with human cost considerations, can not having
In recent years, the improvement of hardware a large amount of data in the previous location
computing power and the maturity of data analysis ready, the following limitations are found in this
technology prompting machine learning, deep learn­ study in the above context:
ing and other technologies to flourish, in the field of
1. There are parts of the image that have a deeper
image recognition is also benefiting from this. In
background and will be misjudged by the model.
recent years, the accuracy of identification more
2. If the coating or structure of the background
practicality because of the technique of deep learn­
color is similar to the rust color, it will also be
ing greatly improved, so this study hopes to intro­
misjudged by the model
duce is now widely. The applied in-depth learning

DOI: 10.1201/9780429279119-43
https://doi.org/10.1201/9780429279119-43

341
 The reason may be the training image set, the lack
of the above-mentioned situation of the photo to
model learning, deep learning is a training set, labels
to learn the prediction of the pulse of the technology,
its performance is quite dependent on the number of
training sets and the scope of the situation. Due to
human and time constraints, it is not possible to train
families in all situations in this study, but the most
valuable thing about this technology is that as the
training data accumulates, the trained models will
become more and more realistic and more accurate.
Figure 2. Labelbox label schematics.

2 SYSTEM DEVELOPMENT

2.1 Training image data collection

In this study, the rust shooting content of steel struc­
ture covers different coating color, different coating
weathering state, different types of rust before and
after the two shots a total of 213 rust images, the
size of 4032 x 3024, the expectation of a variety of
large-size rust images can make the identification
model closer to the real application. Figure 3. Original Rust And Shielding.

2.2 Data labels and processing

The filming of rust image material is about 213,
whether it should be classified as rust, painting off or
because the method of selection of this study is a full
cracking, etc, in order to data label consistency and
convolution neural network, the training data needs
convenience, in this study category, the marking pro­
to be manually labeled to input the full convolution
cess is considered rust if not normal coating part.
neural network model for training. To do this, select
After the data is labeled, the shield of the image
the “labelbox” image labeling tool for the marking
labeling can be produced by “labelbox” (Figure 3), the
of the data is shown in Figure 1.
original rust image and the shielded image produced
The use of “labelbox” allows you to label a picture
by labelbox will become a set of training data for the
in a hand-drawn manner (Figure 2), which varies in
model, and the model will be trained, meaning that the
time for each image footage, and looks at the rust dis­
model is told that the expected result after entering the
tribution and area of the image, ranging from fifteen
original image should be to shield the image.
minutes to a small time. Due to time and human limita­
We decided to do 200 rust images of general
tions, the results of this study are only marked with
degree of labeling, as training data, and 13 rust
213 images obtained from the filming, but if the train­
images to do more accurate labeling as a test set (an
ing data can be expanded to cover more types of rust
average of about 1 to 2 hours), as data material for
and various situations (ex: uneven light and shadow,
this study, while the data also as far as possible to
blurring, foreign objects), the trained image recognition
cover different coating colors, textures, photo
model will be more robust. In the marking process,
shadow states, to let the mold more perfect type.
because often encountered images do not know

3 DATA ANALYSIS AND DISCUSSION

3.1 Introduction to rust sample diagram

The rust images used in this study are all shot on the
ground (Figure 4), and the filming process also covers
as many as possible the coating of different back­
ground colors, rust, noise, texture and so on, expect
the trained model to be closer to the real application.
In the process of collecting data, in order to give
the model better generalization capability, it was spe­
cifically filmed for rust images with different under­
tones, as shown in Figure 5.
Figure 1. Labelbox.

342
Figure 4. Take a rust map on the ground.

Figure 6. Accuracy curve (top) and error curve (bottom).

the identification of the model, so that it can truly

identify “rust”, instead of just grabbing edges from
the image.
Figure 5. Data collection according to the coating color.

3.3 Research limit

3.2 Identification effect discussion The generality of this research identification model
can be gradually enhanced with the accumulation of
This study used the U-net architecture in the FCN data, but at present, some images can not be effect­
concept and trained 200 rust images and shielding as ively identified, as shown in several categories, pre­
training data input models, and tested 13 data labeled sumably due to the lack of image material (painting
with more fine. images similar to rust color and insufficient images
The pointer of the model evaluation can be containing depth of field). Figure 8 below identifies
viewed by observing the change of the pointer the parts of the model that have yet to be overcome
number to see the training effectiveness of the for this study:
model, and the following Figure 6 shows the change
curve of the training set image after several epochs
in the course of model training, accuracy (Accuracy) 3.4 Data set accuracy comparison
and error (Loss). (epoch: represents all samples in In this study, in order to explore the effect of training
a training set over a single session) set data label accuracy on the performance of the
Figure 7 below shows 13 original test data and model, the same number of data sets were specially
identification results, and the white part is the corro­ combined with high-precision labeling of different pro­
sion area of machine identification (original figure portions to observe: mainly divided into three groups:
on the left and model identification on the right):
It is worth mentioning that in the process of col­ 1. A total of 213 sheets, supplemented by 10 high-
lecting training data, this study has specially filmed precision label data, the rest are medium-
some non-rust or complete coating and so on rust- precision labels.
free images, in the marking process, that blank is not 2. A total of 213 sheets, supplemented by 20 high-
marked (rust-free area area). The purpose is to precision label data, the rest are medium-
expose the model to “counterexamples”, strengthen precision labels.

343
Figure 7. Research identification results display.

344
 Figure 9. Training Set Pointer Change Graph.
Figure 8. Research limit.

with the proportion of high-precision data sets and

3. A total of 213 sheets, supplemented by 30 high-
better, this part is expected, high-precision labeling can
precision label data, the rest are medium-
bring more accurate information for the model, but the
precision labels.
increase is not as large as expected, so if you want to
The Accuracy and Loss numbers vary with epoch improve the performance of the model in the future,
as follows: the author believes that the continuous increase in data
From the above Table 1 and Figure 9, we can see set the quantity would be a more cost-effective
that the performance of the model does indeed increase approach.

Table 1 . Margin settings for A4 size paper and letter size paper.

The first group:

Epoch 1 2 3 4 5 6 7 8
Accuracy 0.8817 0.9047 0.9205 0.9282 0.9351 0.9417 0.9436 0.9448
Loss 0.3547 0.2451 0.2111 0.1988 0.1835 0.1717 0.1685 0.1634
The second group:
Epoch 1 2 3 4 5 6 7 8
Accuracy 0.8860 0.9120 0.9259 0.9325 0.9393 0.9454 0.9475 0.9482
Loss 0.2744 0.2275 0.2045 0.1957 0.1736 0.1584 0.1522 0.1534
The third group:
Epoch 1 2 3 4 5 6 7 8
Accuracy 0.8806 0.8968 0.9134 0.9237 0.9325 0.9448 0.9478 0.9533

Loss 0.2759 0.2342 0.2130 0.1973 0.1828 0.1625 0.1539 0.1438

345
4 CONCLUSIONS Recognition Approach”, Computer-Aided Civil and
Infrastructure Engineering, Vol.17, pp.307–319, 2002.
With the improvement of data storage and collection [5] Po-Han Chen, Yuh-Chin Chang, and Luh-Maan
technology, data analysis related technology is Chang, “Application of Multiresolution Pattern Clas­
booming, and various industries are actively introdu­ sification to Steel Bridge Coating Assessment”, Jour­
cing data processing technology. The field of image nal of Computing in Civil Engineering, Vol.16(4),
recognition has also been greatly affected, with the pp.244–251, 2002.
[6] Po-Han Chen, Ya-Ching Yang, Chi-Yang Lei, and
technology of neural network is mentioned again, Luh-Maan Chang, “Automated Bridge Coating Defect
the technology of convolutional neural network Recognition Using Adaptive Ellipse Approach”,
matures, and the technique of applying deep learning Automation in Construction, Vol.18, pp.632–643,
has brought many breakthroughs to image 2009.
recognition. [7] Sangwook Lee, Luh-Maan Chang, and
This study introduces the purpose of a full convo­ Miroslaw Skibniew ski, “Automated Recognition of
lutional neural network in deep learning: Surface Defects Using Digital Color Image
Processing”, Automation in Construction, Vol.15,
1. In the hope that we can improve the practicality pp.540–549, 2006.
of the rust image recognition algorithm, in the [8] Sangwook Lee, “Digital Image Processing Methods
previous algorithm, there are many restrictions for Bridge Coating Management and Their
on rust image, too much noise, light and shadow Limitations”, Journal of Civil Engineering and Archi­
changes, coating color, coating pattern will affect tecture, Vol.1 (38), pp.39–47, 2011.
[9] Qing-Yuan He, and Chuan-Jiu Han, “Image Thresh­
the quality of identification, in addition, the types
olding Segmentation with Otsu Based on Particle
of rust can be identified are also limited. Swarm Optimization Algorithm”, Journal of Guilin
2. Using the identification method of this study,the University of Electronic Technology, May 2006.
above problems can be overcome as long asthe [10] R. Medina-Carnicer, A. Carmona-Poyato,
amount of training data is sufficient and Rafael Muñoz-Salinas, and Francisco José Madrid-
theimages covered are rich. Cuevas, “Determining Hysteresis Thresholds for Edge
Detection by Combining the Advantages and Disad­
vantages of Thresholding Methods”, IEEE Transac­
tions on Image Processing, Vol.19 (1), pp.165–173,
REFERENCES 2010.
[11] Raman Maini, and Himanshu Aggarwal, “Study and
[1] Heng-Kuang Shen, “Automatic Color Image Recogni­ Comparison of Various Image Edge Detection
tion for Steel Bridge Rust Defects Assessment”, Techniques”, International Journal of Image Process­
Department of Civil Engineering College of Engineer­ ing, Vol.3, pp.1–11, 2009.
ing National Taiwan University Doctoral Dissertation, [12] G.T. Shrivakshan, and C. Chandrasekar,
2013. “A Comparison of Various Edge Detection Tech­
[2] Heng-Kuang Shen, Po-Han Chen, and Luh-Maan niques Used in Image Processing”, International Jour­
Chang, “Automated Steel Bridge Coating Rust Defect nal of Computer Science Issues, Vol.9 (5),
Recognition Method Based on Color and Texture pp.272–276, 2012.
Feature”, Automation in Construction, Vol.4, [13] Li Bin, and Mehdi Samiei Yeganeh, “Comparison for
pp.338–356, 2013. Image Edge Detection Algorithms”, IOSR Journal of
[3] Heng-Kuang Shen, Po-Han Chen, and Luh-Maan Computer Engineering, Vol.2, pp.1–4, 2012.
Chang, “Support-Vector-Machine-Based Method for [14] Young-Jin Cha, “Deep Learning-Based Crack
Automated Steel Bridge Rust Assessment”, Automa­ Damage Detection Using Convolutional Neural Net­
tion in Construction, Vol.23, pp.9–19, 2012. works”, 2017.
[4] Po-Han Chen, and Luh-Maan Chang, “Intelligent [15] Will Nash, “Quantity beats quality for semantic seg­
Steel Bridge Coating Assessment Using Neuro-Fuzzy mentation of corrosion in images”, 2018.

346
 Bridge Maintenance, Safety, Management, Life-Cycle Sustainability and Innovations –
Yokota & Frangopol (eds)
© 2021 Taylor & Francis Group, London, ISBN 978-0-367-23278-8

Application of virtual reality technology to cultivate skill for visual

inspection of bridge
K. Ishibashi & H. Furuta
Kansai University, Osaka, Japan

Y. Nomura
Ritsumeikan University, Shiga, Japan

K. Nakatsu
Osaka Jonan Women’s Junior College, Osaka, Japan

K. Takahashi
Kagawa University, Kagawa, Japan

ABSTRACT: The purpose of this study is to propose a system to support cultivating the skill of visual
inspection of engineers in order to sustain the safety of bridge. Various researchers have attempted to auto­
mate bridge inspection by applying technologies of artificial intelligent and robot. However, those studies
have yet to achieve practical automation of inspection because large part of skill of visual inspection are black
box. This study attempts to analyze black box of skill for visual inspection of bridge through an application of
bridge visual inspection using virtual reality (VR) technology that can record eye movements of wearer. This
study develops a prototype of VR application of cultivation for skill of visual inspection by using bridge’s
photos shot by 360-degree camera. In addition, based on eye tracking data collected from the application, the
usefulness of proposal for cultivation of skill for visual inspection is examined.

1 INTRODUCTION (Doebling, Farrar, and Prime 1998) and robots to col­

lect data on visual inspection of bridge (Nishimura
In Japan, it is necessary to implement countermeasures et al. 2012). The implementation of these technologies
for sustaining the safety of bridges. This problem requires huge cost. Therefore, there is a limitation to
comes from the shortage of engineers for bridge man­ the number of bridges that can be implemented the
agement. The skill on bridge management, especially automation system under the circumstances with short­
visual inspection, has been cultivated by engineer’s age of budget. On the other hand, several studies have
experiences. Thus, it is difficult to inherit the skill of addressed the development of system for damage
experienced engineers to the next generation. More­ diagnosis using AI (artificial intelligence) technologies
over, the number of junior engineers has reduced due (Yokoyama and Matsumoto 2017). However, because
to the budget cut for bridge management. In the bride data on damage diagnosis by experienced engineers
management, visual inspection of bridges is essential has not been accumulated enough, it is difficult for the
for early detection of damage and implementation of system to have the practical accuracy.
appropriate measures. Therefore, it is necessary to auto­ In the cultivation of bridge management engineers,
mate the bridge management and improve the cultiva­ a system for skill acquisition is black box based on
tion of engineers. experiences, as described above. It is necessary for
The automation of bridge management has been young engineers to acquire skills through works in
realized by using ICT (information and communica­ practice due to the black box. As a result, the cultiva­
tion technology). However, it is difficult to implement tion of engineers has needed to be spent for a long
the automation of bridge management to all bridges time. On the other hand, damage of bridges progresses
because its development requires enormous cost. with every moment. Therefore, it is necessary to
Existing studies on the automation technologies have improve the efficiency of cultivation of engineers.
developed an all times monitoring system with sensors Existing studies have addressed implementation of

DOI: 10.1201/9780429279119-44
https://doi.org/10.1201/9780429279119-44

347
e-learning (Ebner and Holzinger 2002) and VR (virtual applying damage diagnosis with AI to practical use. In
reality) (Vora et al. 2002). However, a system of skill addition, diagnostic skill of experienced engineers has
acquisition has been black box yet. Thus, additional been cultivated by their experiences. Therefore, the
improvement has been required. practicality of diagnosis system is not clear because
This study attempts to propose a system to support the system of its skill has black box.
the cultivation of engineers for visual inspection of
bridges. Specifically, this study examines a system for
2.2 Cultivation of engineers
collecting data on visual attention of engineers on
visual inspection by making educational materials for Other measure to solve the shortage of bridge man­
VR with eye tracking. Existing studies have demon­ agement engineers is to establish a framework
strated that an educational material with VR had high enough for cultivating engineers. It is very important
learning effect (Bailenson 2018). On the other hand, it to cultivate young engineers and to inherit skill of
is difficult to implement data collection using eye experts to them. In Japan, many experienced engin­
tracking device in a field when engineers work visual eers have retired from work in parallel with the cut
inspection. Therefore, eye tracking with VR is in workforce due to the shortage of budget. Under
expected to be useful for collecting data on visual present circumstances, it is difficult to allocate
inspection. By analyzing black box of skills on visual enough time to the cultivation of engineers. More­
inspection from accumulated eye tracking data, the over, young engineers miss out on a chance to inherit
proposal aims to gain findings beneficial for the culti­ experts’ skill because the cut in workforce imposes
vation of engineers and the application of AI. This various works on them. Studies on e-learning (Ebner
paper develops a prototype of proposed system and and Holzinger 2002) and educational materials with
examines its applicability. VR (Vora et al. 2002) have been addressed in order
to solve this problem. E-learning is considered as
a measure to improve the efficiency of engineer’s
2 RELATED WORKS cultivation. The system of skill on bridge manage­
ment, especially visual inspection, consists of engin­
2.1 Automation of bridge management eers’ experiences as described above. Thus,
researchers have examined the use of educational
In Japan, the shortage of engineers for bridge material with relative effective media such as photo
management has been a serious problem. Auto­ and movie. VR has been used for educational mater­
mation of bridge management by using ICT has ial to improve learning effect (Bailenson 2018). In
been attempted to solve this problem. For the bridge management, it is difficult to practice on-
example, health monitoring system has been the-job training in a field. Therefore, educational
developed to monitor bridge’s state from sensors materials with VR are expected to be more effective
attached at the bridge (Doebling, Farrar, and for the cultivation of engineers than conventional
Prime 1998). The health monitoring system is media.
useful for automatically finding abnormal state of One reason of problems on the cultivation of engin­
bridge. However, the installation of health moni­ eers is unclear system of skill. That is, term spent on
toring system requires huge cost due to retrofit of the cultivation tends to become long because an engin­
sensors into bridge and application of power. eer should depend on work experience to acquire
Thus, this approach cannot apply to all bridges. essential skills. However, Japan has already faced
In recent years, several studies have addressed problems from the shortage of engineers. Therefore, it
implementation of robot technologies such as is necessary to implement immediate countermeasures.
UAV (uninhabited airborne vehicle) in order to
automate visual inspection (Nishimura et al.
2012). However, there are many problems for 3 PROPOSED SYSTEM
practical use of robot such as its control and pre­
vention of its drop. As an approach from compu­ The area of bridge management has problems on both
tation, application of AI to damage diagnosis has the development of automation technologies and the
been addressed (Nakatsu et al. 2012). cultivation of engineers as described in chapter 2.
One reason of problem on the implementation of A common factor between these problems is black box
automation technology to bridge management is the of bridge management skill. Clarification of black box
shortage of budget. That is, it is necessary to imple­ on bridge management skill is effective not only for
ment efficient countermeasures with limited budget. the systemization of engineer’s cultivation but also for
As described above, health monitoring system cannot improving accuracy of damage diagnosis system using
be applied to all bridges. Use of bridge management AI. Under circumstances with budget cut, interaction
robot requires enormous cost under present circum­ between man and machine is important for sustaining
stances. Damage diagnosis system using AI technology bridge management. This study proposes a learning
can be applied with relative low cost because it uses support system using VR with eye tracking in order to
collected data. However, it is necessary to accumulate analyze black box of bridge management skill and to
a huge amount of data to obtain accuracy enough for solve problems on the cultivation of engineers.
348
Figure 1. Outline of proposed system.

3.1 Outline of system collection of data on visual inspection and damage

diagnosis. Accumulation of data has been known to
Figure 1 shows the outline of proposition of this study.
be effective for improving accuracy of AI. In add­
This study applies VR headset with eye tracking
ition, analysis of accumulated data is expected to
to examining educational material on bridge manage­
clarify factors that have strong relation to decision-
ment, especially visual inspection. VR headsets with
making of experienced engineers. In this way, the
eye tracking have been developed in recent years.
proposed system attempts to support realization of
Fove 0 and Vive Pro Eye are devices available for
sustainable bridge management through the inter­
everyone. These devices are expected to enable engin­
action between human and machine.
eers to experience realistic learning on skills of visual
inspection in virtual space. Furthermore, accumulated 3.2 Learning contents using VR
data on eye tracking during the learning can be applied
to analyze factors related to acquisition of skill. The existing study used a structure made from 3DCG
Recently, wearable eye tracking devices such as Tobii (three dimensions computer graphic) as a learning con­
Pro Glassless 2, Pupil Core and etc. have been devel­ tent using VR (Vora et al. 2002). On the other hand,
oped. However, data collection with wearing eye track­ the proposed system attempts to use photos and
ing device in a field have dangerous for engineers. movies captured by 360-degree camera as a learning
Therefore, a system that can collect eye tracking data content. Recently, 360-degree cameras such as RICHO
in virtual space enable us to observe detailed data on theta have enabled us to produce VR contents easily.
visual inspection. VR contents made from 3DCG are useful for learning
The proposed system uses educational materials under various situations. On the other hand, photos
available for VR not only to the cultivation of engin­ and movies captured by 360-degree camera is expected
eers but also to the data collection on visual inspection to support engineers to obtain experiences similar to
of bridge. Especially, eye tracking data observed from on-the-job training. In addition, those contents are
experienced engineers is expected to be useful for ana­ available with smartphone. Therefore, using contents
lyzing black box of bridge management skills. More­ of both 3DCG and photos and movies captured by
over, by recording visual inspection process of 360-degree, the proposed system is expected to pro­
experienced engineer, young engineers can experi­ vide educational materials effective for learning.
ence it.
In the cultivation of engineers, the proposed edu­ 3.3 Analysis of expertise by using eye tracking
cational materials is useful as the same as conven­
tional medias with high learning effects. Eye tracking data is expected to be useful for analyzing
Furthermore, visualized eye tracking data collected black box between expertise of experienced engineer
from visual inspection of experienced engineers can and decision-making. In the marketing area, various
be available for a educational content. In this way, studies have been addressed based on the theory of
a system using VR with eye tracking utilizes col­ visual marketing (Wedel and Pieters 2008). For
lected data into next education. examples, existing studies investigated effects of adver­
In a support system using AI technologies, accu­ tisement design (Pieters, Wedel, and Batra 2010) and
mulated eye tracking data can be applied to analysis store shelf layout (Chandon et al. 2009) on consumer’s
and machine learning. Application of proposed visual attention and purchase behavior. In addition,
system to the cultivation of engineers promotes effects of consumer’s individual attributions have been
studies (Goldberg, Probart, and Zak 1999; Pieters and

349
Figure 2. Example of framework for analysis using eye tracking (revised from Figure 6. 1 in p. 39 of Wedel and Pieters (2008)).

Wedel 2007). This study applies a framework shown in 4.1 Learning content
Figure 2 to analyses based on the theory of visual
As a prototype of learning content, this study
marketing.
took the photo shown in Figure 3 by using 360­
As shown in Figure 2, this proposal examines an
degree camera. Captured structure is a bridge in
effect of expertise as human factors (top-down fac­
Osaka, Japan. The photo was taken under the
tors). That is, this study investigates effects of
bridge. As shown in Figure 3, a user can view
expertise on visual attention such as viewing time,
the photo from angle corresponding to the direc­
and then we analyze indirect effects of expertise on
tion of his/her face by using VR headset or
decision-making mediated by visual attention
smartphone.
(downstream effect). In addition, effects of the
A VR content is created by just taking photo with
number of vertical clacks and their width in visual
360-degree camera. Moreover, 360-degree camera
inspection are investigated. These have been con­
can record sounds in video mode. Therefore, learn­
sidered to have strong relation to damage of bridge.
ing contents can be created easily by taking photos
Based on thisframework, this study aims to analyze
of bridge with 360-degree camera in an actual visual
black box of visual inspection skills.
inspection. On the other hand, a content using 3DCG
enables engineers to experience visual inspection
under various situations as the same manner as exist­
4 DEVELOPMENT OF PROTOTYPE SYSTEM
ing studies (Vora et al. 2002).
The data collection on learning with VR contents
This study develops a prototype of proposed
enables us to investigate differences between experi­
system, and examines its usefulness for the culti­
enced and young engineers. In addition, the useful­
vation of bridge management engineers. Firstly,
ness of manual for visual inspection can be verified
we create a learning content using photo captured
through comparing learners that previously read it
by 360-degree camera. Next, this study examines
with others that did not read. In this way, the pro­
the applicability of data collected from a test by
posed system uses VR contents as educational
constructing a data collection system using VR
materials for cultivating engineers.
with eye tracking.

Figure 3. Example of learning content.

350
Table 1. Example of collected eye tracking data.

4.2 Eye tracking data on visual inspection analyzing black box of skills on bridge management.
A prototype of proposed system can be applied to ana­
VR with eye tracking can collect data as shown in
lysis of relationship between viewing time of certain
Table 1 (Ishibashi 2018). As shown in Figure 3, data
object and decision-making. Furthermore, eye tracking
collected by using photo and movie contains coord­
data can be applied to development of model which
inate of gaze point (LookedX, LookedY and Loo­
measures effects of engineer’s expertise and visual
kedZ) at regular intervals. The system of existing
stimuli of bridge. Therefore, future works attempt to
study (Ishibashi 2018) records data at intervals of 30
construct a model on visual inspection by using data
fps by using Fove 0. On the other hand, data col­
collected from the proposed system.
lected by a content made from 3DCG can observe
coordinate of participant’s position (PosX, PosY and
PosZ) in the virtual space. We calculate the summa­ ACKNOWLEDGEMENT
tion of viewing time on certain object (gaze dur­
ation) from collected eye tracking data, and then we This work was supported by The Research Institute
measure eye movements of engineers. for Socionetwork Strategies, Kansai University.
Data on decision-making in the framework shown
in Figure 2 is collected by answers in the learning or
questionnaires. Moreover, data on human factors is REFERENCES
generated from database on profile of engineer. Data
on visual stimuli is defined when creationof learning Bailenson, Jeremy. 2018. Experience on Demand: What
content. In this way, the proposed system constructs Virtual Reality Is, How It Works, and What It Can Do.
dataset for analysis. W. W. Norton & Company.
Chandon, Pierre, J. Wesley Hutchinson, Eric
T Bradlow, and Scott H Young. 2009. “Does
In-Store Marketing Work? Effects of the Number
5 CONCLUSIONS and Position of Shelf Facings on Brand Attention
and Evaluation at the Point of Purchase.” Journal
This study proposed a system for cultivation of of Marketing 73(6): 1–17. http://journals.ama.org/
engineers using VR with eye tracking in order to doi/abs/10.1509/jmkg.73.6.1.
support sustainable bridge management. Educational Doebling, Scott W., Charles R. Farrar, and Michael
material using VR is a media effective for cultivating B. Prime. 1998. “A Summary Review of
engineers. This proposal investigates factors related Vibration-Based Damage Identification Methods.”
to decision-making such as damage diagnosis Shock and Vibration Digest 30(2): 91–105.
Ebner, Martin, and Andreas Holzinger. 2002. “E-Learning
through collecting data of eye movements in visual
in Civil Engineering: The Experience Applied to
inspection in virtual space. Furthermore, by examin­ a Lecture Course in Structural Concrete.” Scientific
ing effects of engineer’s expertise on visual attention Journal of Applied Information Technology 1(1): 1–9.
and decision-making, we attempt to analyze black Goldberg, Joseph H., Claudia K. Probart, and Robert
box of skills on bridge management. In this paper, E. Zak. 1999. “Visual Search of Food Nutrition Labels.”
the applicability of this proposal was examined by Human Factors 41(3): 425–37.
constructing a prototype system. Ishibashi, Ken. 2018. “Availability of Virtual Reality with
A VR content can be created by using 360-degree Eye-Tracking for Survey Experiment.” In Proceedings
camera in recent years. In addition, consumers have of The 22nd Pacific Asia Conference on Information
Systems (PACIS 2018), 99–102.
become available reasonable VR devices. Educa­
Nakatsu, Koichiro et al. 2012. “Multiple Feature Selectoin
tional materials using VR on visual inspection of for Pattern Recognition Using ID3 Ensemble System.”
bridge are expected to support implementation of In Proceedings - 2012 6th International Conference on
practical learning that cannot be held frequently. Genetic and Evolutionary Computing, ICGEC 2012,
Moreover, visualization of process on visual inspec­ IEEE, 145–48.
tion of experienced engineer is effective for cultivat­ Nishimura, Shozo, Kenji Hara, Keisuke Kimoto, and
ing young engineers. Hiroshi Matsuda. 2012. “The Measurement and
Eye tracking data collected from the cultivation of Draw Damaged Plans at Gunkan-Island by Using
engineers and experiments is expected to be useful for 3D Laser Scanner and Digital Camera.” Journal of

351
 Japan society of photogrammetry and remote sens­ Vora, Jeenal et al. 2002. “Using Virtual Reality Technology
ing 51(1): 46–53. for Aircraft Visual Inspection Training: Presence and
Pieters, Rik, and Michel Wedel. 2007. “Goal Control of Comparison Studies.” Applied Ergonomics 33(6):
Attention to Advertising: The Yarbus Implication.” 559–70.
Journal of Consumer Research 34(2): 224–33. https:// Wedel, Michel, and Rik Pieters. 2008. Eye Tracking for
academic.oup.com/jcr/article-lookup/doi/10.1086/ Visual Marketing. Foundations and Trends.
519150. Yokoyama, Suguru, and Takashi Matsumoto. 2017.
Pieters, Rik, Michel Wedel, and Rajeev Batra. 2010. “Development of an Automatic Detector of Concrete
“The Stopping Power of Advertising: Measures and Surface Detereorations Using Deep Learning and
Effects of Visual Complexity.” Journal of Market­ Implementation of Web System.” Proceedings of the
ing 74(5): 48–60. http://journals.ama.org/doi/abs/ 20th Symposium on Applied Mechanics 73(2):
10.1509/jmkg.74.5.48. 781–89.

352
 Bridge Maintenance, Safety, Management, Life-Cycle Sustainability and Innovations –
Yokota & Frangopol (eds)
© 2021 Taylor & Francis Group, London, ISBN 978-0-367-23278-8

Variability-based method for balancing structural optimization and

reliability
K.J. Haas
Hinman Consulting Engineers, San Francisco, California, USA

ABSTRACT: Modern bridge designs are expected to operate under ever increasing lifecycle durations.
Increased lifecycles subject bridges to evolving and increasingly variable operational demands. The nature and
severity of environmental changes, security-related risks, and a host of other demands may dramatically vary over
a bridge’s lifecycle, all the while material degradation from fatigue, creep, and corrosion mechanisms may act to
degrade the bridge’s performance. The combinations of evolving factors lead to high potential for variability in
a bridge structure’s expected performance. Further challenging bridge designers, weight optimization efforts,
which are critical in achieving cost-effective designs, often come at a cost to reliability in the face of variability.
The need to explicitly address bridge reliability under variable conditions, though sensitivity analysis and/or uncer­
tainty quantification, is greatly increased by today’s lifespan expectations. Many sources of uncertainty under
extreme events such as blast cannot be readily quantified to utilize propagation of uncertainty. This paper demon­
strates a new variability-based method for addressing these challenges while simultaneously assessing the struc­
tural design’s optimization and system reliability. Machine learning and stability indicators are utilized, and a new
indicator called the Instability Index is proposed to predict relative system stability for a given loci of input vari­
ables. Both the structural optimization and reliability objectives are then balanced to find an optimal design that
best achieves both objectives. The utility of the proposed method is demonstrated through its application to
a realistic bridge design scenario that is subject to poorly defined sources of variability. The proposed methods are
applied in the contexts of structural system reliability under undisturbed (operational) conditions and under struc­
tural failure scenarios (element removal).

1 INTRODUCTION rod diameter, beam depth) given a fixed structural

configuration. Shape optimization identifies nodal
1.1 Optimization coordinates as variables, which are then iterated
through any number of algorithmic optimization
Structural weight optimization efforts continue to gain
techniques to identify the optimal nodal coordinates.
prevalence in the engineering community due to
In this case, the structural sizes are fixed, and the
increasingly available computing power. Large-scale
configuration is variable. Topology optimization
systems can be readily optimized to satisfy all relevant
seeks to identify optimal designs through an iterative
building code requirements with minimum material
approach of treating each meshed area as comprised
quantities through a wide range of techniques. How­
of either structural material or a void.
ever, the effects of weight optimization must be bal­
Though powerful in determining optimal struc­
anced against the potential negative effects on system
tural designs, over optimization can introduce
robustness and reliability. Industry standards currently
troublesome hazards. Overly optimized structures
exist to incorporate predictable and/or measured uncer­
often lack redundancy and/or opportunity for
tainty into final optimized designs, such as ASME VV­
adequate load redistribution after local damage is
10. However, the challenges associated with account­
incurred. Under the current state of research devel­
ing for undefined or unknown sources of uncertainty
opment, this characteristic of optimized structures is
are largely unaddressed by most industry standards.
relatively inherent. Therefore, deterministic designs
Optimization techniques typically fall into one of
that do not account for uncertainties and/or vari­
three categories – (1) Sizing Optimization, (2) Shape
ations are insufficient (Choi, 2017).
Optimization, or (3) Topology Optimization (Srivas­
The optimization of material and structural con­
tava, 2017). Sizing optimization generally describes
figuration often comes at the expense of robustness
the optimization of discrete component sizes (e.g.

DOI: 10.1201/9780429279119-45
https://doi.org/10.1201/9780429279119-45

353
and by extension, the overall system reliability (Mar­
janishvili, 2016). For the purpose of this paper,
robustness and reliability are characterized as fol­
lows (Haas, 2019):
• Robustness – Insensitivity of performance relative
to perturbations in the system’s loci of conditions
• Reliability – Sufficient performance of a system
while subject to uncertainties and variability over
its lifecycle
This paper introduces a method for tracking the
Figure 1. Proximity consequences on a typical cusp
relative reliability of the structural design over the surface.
course of optimization iterations, which does not
demand explicit quantification of uncertainty
sources. The resulting tabulated optimization and
reliability metrics can be compared and/or combined
to select the structural design that best combines
structural optimization and reliability.

2 METHOD OVERVIEW

2.1 Reliability evaluation - variability-based

reliability method
The variability-based method assesses system reli­
ability through a combination of gradient and prox­
imity-based metrics. These two metrics provide
indications of local and dynamic stability.
Figure 2. Variability-based reliability method steps.
• Gradient Field – L2 Norm indicates the local sta­
bility of the system. Higher L2 Norms indicate
that small perturbations in assumptions will lead
to disproportionate changes in system response Mitigation of overtraining through regularization
(e.g. peak displacement, energy absorption is critical in this application so as to not falsely
properties). predict sharp manifold changes (i.e. large gradi­
• Euclidean Proximity – Indication of Euclidean ents) where they do not exist.
proximity between loci of system variables of A properly trained emulator can be used
lower to higher L2 Norm classifications. Lower through the remaining steps to expeditiously
proximity indicates dynamic instability, i.e. that determine trial solutions corresponding to varied
minor perturbations in assumptions may lead the system variables. Sufficient data collection may
system down an unrecoverable path toward not allows be practical for conventionally trained
unstable regions via gradient flow. This is depicted and regularized machine learning models. Uncon­
on a typical cusp catastrophe surface in Figure 1. ventional, but powerful, tools such as persistent
The local and dynamic stability metrics are calcu­ homology have been demonstrated to limit the
lated through the process depicted in Figure 2 and quantity of training data required to limit over-
described through the remainder of this section. training. A detailed description of persistent hom­
ology and its applications to machine learning
• Step 1 – Establish Topological Relationships model regularization is beyond the scope of this
Relationships between system input variables and paper, but can add to the practicality of machine
the resulting structural performance must be learning based emulators
established to later calculate gradient characteris­ • Step 2 – Gradient Field Computation
tics. This can be done analytically or through The emulator from Step 1 allows for rapid calcu­
measured data. Alternatively, machine learning lation of the gradient field. The gradient field is
tools (e.g. regression models, artificial neural net­ determined by initiation of perturbations in each
works) may be employed where sufficient analyt­ individual system variable dimension and calcu­
ical or testing data is not available and/or lating the resulting change in system perform­
impractical. If trained models are used, the train­ ance. The gradient field and L2 Norm are
ing data collection must be dense enough to cap­ calculated with equations (1) and (2),
ture any sharp localized manifold features. respectively.

354
• Step 3 – L2 Norm Grouping
Loci of input variables (design variables and
assumptions) are grouped in accordance with
their L2 Norms. K-means clustering allows for
rapid and optimum clustering of similar gradi­
ents. K-means clustering seeks to group points
(loci of variables) based on the Euclidean dis­
tance of their corresponding gradient to
a predetermined number of variably located Figure 4. Relative proximity visualization.
gradient centroid nodes. Figure 3 depicts
a typical example of K-means clustering for
three dimensional systems. Although the similar to the concepts that underpin catastrophe
method cannot be readily visualized for higher theory, a subset of singularity theory. In catastro­
dimensions, its utility lies in its application to phe theory, stability can be considered relative to
both low or high-dimensional systems. Given the relative distance away from the degenerate
a continuous and differentiable manifold sur­ points of an underlying potential function
face, increasing the number of assumed cluster (Poston, 1978). These degenerative points repre­
centroids will tend to consolidate adjacent sent locations where minor perturbations would
regions, thus reflecting more connected data. lead the solution toward unrecoverable and severe
• Step 4 – Proximity Evaluation changes.
The proximity of a given point to areas of lower • Step 5 – Stability Metrics
local stability (higher gradients) is determined to The gradient and proximity metrics, equations (4)
assess dynamic stability. The Euclidean distance and (5) respectively, are produced from the ana­
between regions of lower stability can be viewed lytical data generated through the previous steps.
through the use of the gradient-based clustered Both are normalized to lie within the range [0,1].
groups determined in Step 3. Sudden changes in Gradient and proximity metrics are combined
gradient regions can be visualized, as depicted into a single indicator called the, “Instability
from a typical example in Figure 4. Sudden Index”, in equation (6). Higher values of the
jumps in gradient represent critical areas, where Instability Index indicate lower stability, i.e.
relative proximity to those jumps may provide lesser reliability.
indication of dynamic stability. The calculation of
relative proximity is performed with equation (3),
between point, “j”, and cluster, “k”.

This consideration of proximity to unstable

regions as a metric for dynamic stability is

2.2 Design optimization

Realtime tracking and postprocessing of the reli­
ability versus optimization progress is made more
straightforward with an optimization technique
that straddles the definitions of the typical opti­
mization categories. Artificial Bee Colony (ABC)
optimization is used due to its flexibility. ABC
Figure 3. K-means clustering example. falls into an alternate optimization category ­

355
swarm intelligence-based algorithms. ABC seeks 3 EXAMPLE APPLICATION
optimum solutions for unconstrained and con­
strained problems through mimicking the foraging 3.1 Example scenario
patterns of honeybees in their quest for nectar
(Abu-Mouti, 2012). The previously described procedures are demon­
ABC mimics honeybee foraging behavior through strated to the simple bridge structure depicted in
the use of the following: Figure 5, assuming AASHTO HS20-44 truck
loading.
• Employed Bees - Bees that seek solutions (nectar
sources) • P1 = P2 = 143.34-kN (32-kips)
• Observer Bees – Bees that observe the quality of • P3 = 35.6-kN (8-kips)
the solutions (nectar sources) found by employed • Element Width = 152-mm (6-in)
bees. Upon switching roles with employed bees, • (5) Variables:
the previous observations inform the direction • Dc1 – Bottom Chord Depth
and distance to travel to find the next potential • Dc2 – Top Chord Depth
solutions. • Dw1 – Diagonal Web Depth
• Scout Bees – Bees that seek solutions randomly, • Dw2 – Vertical Web Depth
without being informed by experience. • H – Truss Height

This process allows for rapid convergence to • Douglas Fir-Larch Structural Select
local minima, even in high-dimensional solution • Typical 4.88-m (16-ft) Bay Length
spaces. Additionally, ABC is useful for noncon­ The goal of the sample scenario is to optimize
vex optimization problems due to the presence of the (5) preselected variables relative to two goals:
scout bees to effectively leap to alternate areas of (1) displacement of Span/240 and (2) minimize
the solution space. The quality of a given solu­ weight. A robust and reliable bridge design is
tion is expressed through its, “fitness”. Fitness is also sought so reliability metrics will be evalu­
generally defined by a function bounded by the ated in concert with performance and weight
range, [0,1], with higher values corresponding to optimization.
greater optimization. Similar to how the Instabil­
ity Index represents a metric for assessing reli­
ability, the fitness value represents a metric for 3.2 ABC bridge design optimization
assessing the degree of optimization. The element widths and truss heights are classified
as the design variables to achieve optimization of the
2.3 Combined Reliability & Optimization displacement performance objective and minimal
structural weight. The ABC optimization utilized an
One can simultaneously address optimization and effective colony size of 16, with an even split of 8
reliability through consideration of both metrics bees acting as employed and observer bees at
described in previous sections. Plotting each a given time. The optimization algorithm was set to
metric simultaneously on a vertical axis against run over 1,000 iterations. The cost function incorpor­
number of optimization iterations on the horizon­ ates both displacement performance and weight, and
tal axis will tend to demonstrate an expected con­ is filtered through a sigmoid function to maintain
sequence. As optimization proceeds, the a similar scale for combination. These relationships
reliability (invariance to perturbations in assump­ are expressed in equations (8) through (10). The
tions) will tend to decrease. However, simultan­ resulting fitness through progressing iterations, as
eous tracking of the reliability and optimization expressed with equation (11), is depicted in Figure 6.
metrics informs the point of optimization that
jumps in instability occur, thus informing the
degree of optimization that can be utilized before
significant reliability is sacrificed.
The reliability and optimization metrics can also
be combined to form an overall decision-making
variable as shown in equation (7), with I and
F representing the Instability Indicator and optimiza­
tion fitness, respectively, for each optimization iter­
ation, i. WI and WF are weighting factors to
emphasize either reliability or optimization in the
final design.

Figure 5. Example bridge structure.

356
 • Gradient Field Computation
The trained polynomial regression model was
used to calculate the expected performance across
perturbations in each of the (5) applicable vari­
able dimensions. The resulting L2-Norm of the
gradient field at each iteration step is plotted in
Figure 7.
• Proximity Evaluation
The L2-Norms produced from the gradient
field calculations were grouped together with
K-Means clustering. Two significant features
are discovered from viewing Figure 8. Sudden
Figure 6. ABC optimization progression. and discontinuous jumps occur after cluster 6
and cluster 17. These jumps indicate
a deterioration of local stability directly adja­
cent to the points in cluster groups 6 and 17.
Equation (3) is used to calculate the Euclidean
distance between each point (loci of system
variables) and points within adjacent higher
L2-Norm cluster groups.
• Stability Metrics
Equations (4) through (6) are used to compute the
local, dynamic, and combined stability indicators
3.3 Variability-based reliability method
plotted in Figure 9. As expected, the instability is
(Operational conditions)
observed to quickly increase as optimization
• Establish Topological Relationships
A modified polynomial regression model was
used as an emulator to expeditiously perform the
large number of required analyses. The relatively
complex behavior a multi-degree of freedom
system proves challenging to simple regression
models. However, the relatively complex behav­
ior was able to be efficiently accounted for
through the randomized multivariate polynomial
permutation technique provided in equations (12)
through (15). A unique set of n random perturba­
tions were generated. The permutations involve
k selections from possible exponents of range
[0,m]. Larger numbers of n permutations only
nominally increases the computational runtime Figure 7. L2-Norm computation.
but provides a significant increase in predictive
power. Additionally, the use of polynomial
regression allows application of the normal equa­
tion, thus negating the need for gradient descent
training. n = 2,500 permutations were selected for
this example problem. R2 values for the training
and test data were 0.98 and 0.96, respectively.

Figure 8. K-Means clustering groups.

357
 3.4 Element removal evaluation (Post-local failure
conditions)
The concepts previously presented throughout this
paper can be applied similarly to assess dispropor­
tionate response potential due to element removal.
Disproportionate collapse potential is often evalu­
ated through simulating structural response after loss
of a single element. The same truss scenario from
the previous exercises is augmented as follows:
• Analysis is performed to determine the change in
Figure 9. Stability indicators.
total system internal strain energy between the
unperturbed state and the perturbed state from
removal of a single element.
• A separate analysis is performed for removal of
efforts proceed. After an initial sharp increase in each element (21 in total), for each optimization
instability (decrease in reliability), there are sev­ iteration (1,000). 21,000 (21 elements x 1,000
eral oscillations that occur. Equation (7) is iterations) total cases are evaluated.
expressed in Figure 10, assuming equal weighting • The instability indicator metric is taken to be the
of instability and optimization metrics. Critical arithmetic mean of the changes in internal strain
decision points indicated in Figure 9 are summar­ energy across each element removal case for
ized in Table 1. Point 3 appears to be the most a given iteration. Higher values indicate less stabil­
advantageous design to balance optimization and ity, i.e. larger changes in response due to element
reliability objectives, as there is a pronounced dip removal.
in instability, and it is located close to the end of
the optimization routine. Conversely, Point 2 Change in internal strain energy is used as
should be avoided due to its spike in instability, a surrogate for a measure of disproportionate system
thus indicating a non-robust design. response to initial damage (element removal). Other
metrics and/or progressive iterative analysis can also
be employed to increase analysis fidelity.
Figure 11 depicts the resulting strain energy stabil­
ity metric throughout the optimization progression.
Table 1. Solution results (Operating conditions).
Similar to the decision variable proposed in Section
Point Dc1 Dc2 Dw1 Dw2 H G P 2.3, the optimization and stability metrics are com­
bined in equation (7). Equation (7) can be used as
1 1.8 10.1 15.4 8.3 72.3 0.46 0.13 a relative indicator to score the best design solution
2 4.7 10.4 3.1 0.8 83.5 0.86 1.0 region to balance optimization and reliability object­
3 1.6 11.3 3.9 0.8 74.0 0.41 0.99 ives. Figure 12 depicts the application of Equation (7)
assuming equal weights between instability and opti­
mization. The critical decision points indicated in
1Variable dimensions are in cm. Figure 12 are summarized in Table 2. It is apparent

Figure 10. Operational conditions decision variable. Figure 11. Total strain energy change.

358
 critical features. Alternative tools, such as persist­
ent homology or autoencoder neural networks,
may prove valuable in uncovering critical features
of data sets so as to more expeditiously extract
relevant data from trial solutions to assess their
relative stability.

5 NOMENCLATURE

Figure 12. Element removal scenario decision variable. !

xn Principal Component Unit Vector
D Decision Variable
G Gradient-Based Stability Indicator
Hθ Hypothesis Prediction
Table 2. Solution results (Operating conditions)1. I Composite Stability Indicator
J Optimization Cost Function
Point Dc1 Dc2 Dw1 Dw2 H G P P Proximity-Based Stability Indicator
1 1.8 10.1 15.4 8.3 72.3 0.46 0.14 Pð X Þ Performance Function
Pm
k Permutations - k selections from m
2 3.9 10.4 3.2 0.8 83.5 0.49 0.88
Tθ Regression Coefficient Vector
3 1.6 8.8 6.0 0.8 69.0 0.58 1.0 WF Decision Weight – Optimization Fitness
WI Decision Weight – Instability
1. Variable dimensions are in cm.
X System Feature Space
rf Gradient Vector
that Point 3 is an advantageous design due to its rela­
tively low instability indicator and location near the jrf j L2-Norm
Dt Peak Displacement at Given Step
end of the optimization routine. Point 2 occurs at Dlim Target Displacement Performance Limit
a large spike in instability its corresponding non- DXjk Euclidean Distance - Point j to Cluster k
robust design should be avoided.
� Hadamard Power Operator

4 CONCLUSIONS

Structural design optimization is a valuable tool in

a world of limited resources and competitive building REFERENCES
markets. However, the use of common optimization
methods challenge designers to address potential Abu-Mouti, F.S., El-Hawary, M.E. 2012. Overview of Arti­
reductions in system robustness and reliability. The ficial Bee Colony (ABC) algorithms and its
methods presented in this paper have been demon­ applications. Systems Conference (SysCon), pp. 1–6.
ASME. 2006. V&V 10 – Guide for Verification and Valid­
strated to expeditiously evaluate optimization pro­ ation in Computational Solid Mechanics. New York,
gress and its effect on design solution stability. The NY: The American Society of Mechanical Engineers.
resulting decision metrics empower a designer to Choi, S., Grahndhi, R., Canfield, R. 2007. Reliability-
select a structural design that emphasizes both Based Structural Design. London: Springer-Verlag.
weight/performance optimization and reliability. Haas, K. 2019. Prediction of Structural Reliability Through
A simple truss bridge structure was used to an Alternative Variability-Based Methodology. Las
depict the proposed method’s application. How­ Vegas, NV: ASME 2019 Verification and Validation
ever, the real power behind the proposed methods Symposium.
lie in their applications to large and complex sys­ Marjanishvili, S., Katz, B. 2014. Uncertainties in Structural
Response Evaluation. Las Vegas, NV: ASME Verifica­
tems. The variability-based method is designed to tion & Validation Symposium.
be flexible so as to simultaneously incorporate Poston, T., Stewart, I. 1978. Catastrophe Theory and its
multiple load types/patterns and multiple perform­ Applications. Chelmsford, MA: Courier Corporation.
ance criteria (e.g. serviceability, strength, fatigue). Srivastava, P.K., Simant, Shukla, S. 2017. Structural Opti­
These methods are expected to be improved as mization Methods – A General Review. International
a result of current research into alternate means Journal of Innovative Research in Science, Engineering
of establishing topological relationships between and Technology, Vol. 6, Special Issue 9.

359
MS03: Bridge health monitoring conjunction with smart cities

Organizers: A. Miyamoto, A. Yabe, P. Hradil & K. Koski

 Bridge Maintenance, Safety, Management, Life-Cycle Sustainability and Innovations –
Yokota & Frangopol (eds)
© 2021 Taylor & Francis Group, London, ISBN 978-0-367-23278-8

Development of a practical social big data collecting system for the bridge
using by a large vehicle
Akito Yabe
Kozo Keikaku Engineering Inc., Tokyo, Japan

ABSTRACT: The “smart city concept” is expanding worldwide, which aims to administrate efficiently
social management by collecting, sharing and analyzing social big data generated from various activities. But
the application of the social big data for infrastructures such as the bridge is less than the other fields. So there
in this paper, the information collecting using daily operated large vehicles such as route buses is focused
on, an analysis method and application range was summarized for social big data collection and analysis
using traffic data of large-sized vehicles such as route buses, etc. for maintenance and management of bridges.
And in doing so, the results of demonstration experiments are introduced at Yamaguchi Prefecture, Kuma­
moto Prefecture, and using a crane truck as an another large vehicle type.

1 INTRODUCTION than “Transportation/mobility” and “Tourism and

regional revitalization” fields.
Recently grappling with newly social management is In Japan a toward to realize smart city the grap­
expanding by effective using of an advanced informa­ pling with many MaaS (Mobility as a Service) pro­
tion and communication technology. Especially the jects are going. Maas is the one of concept for using
monitoring data for the environment and made by any transportation and mobility seamlessly without
variously social activities are analyzed and shared for private cars. On the Maas, various parts of transporta­
administrating efficiently social management as social tion and mobility data shall be shared and applied for
big data. They are called “Social Big Data” and ana­ all users and all administrators. And those data will be
lyzed for deciding more smart social action. The one used keep sound transportation and mobility networks.
of clear idea is “Smart City Concept”. This concept It is able to infer that such as the situation increased
will make reduce an evil effect of the administrative the number of Transportation/mobility field project.
division and support effects on various parts of citizen Figure 2 is shown the ratio of composited fields
life. And the concept is expanding all over the world. with “Infrastructure maintenance” field. “Transporta­
In Japan August 2019,” the interim report con­ tion/mobility” projects were 91% in the relationship
cerning to realize smart city” was published. In the project with “Infrastructure maintenance”.
report, the concept named “society 5.0” was defined From the above, On the grappling with realization
and due to realize variously activities are grappled of the smart city concept the projects of composited
in the many local cities. In the “Smart city Public- fields with “Infrastructure maintenance” is high rela­
Private partnership platform” until November 2019 tion with “Transportation/mobility” project. But
105 projects are introduced linked by 11 fields. “Transportation/mobility” projects with “Infrastruc­
Concretely 11 fields are “Transportation/mobility”, ture maintenance” project is less.
“Energy”, “Disaster prevention”, “Infrastructure Now, In the case of the bridge maintenance these
maintenance”, “Tourism and regional revitaliza­ reason will be inferred that
tion”, “Health/Medical”,“Agriculture, forestry and
fisheries”,“Environment”, “Security and watching”,
“logistics”, “City planning and maintenance”. (1) Many bridges and load surface maintenance
Figure 1 is shown the ratio of smart city pro­ data are updated in 5 years by inspection. After
accident or disaster these data can’t use to grasp
jects each fields. “Transportation/mobility” and
“Tourism and regional revitalization” fields were the bridge condition.
included about 80% project. On the other hand (2) There are case that road surface and Bump
“Infrastructure maintenance” field was 30% less deterioration progress rapid than inspection

DOI: 10.1201/9780429279119-46
https://doi.org/10.1201/9780429279119-46

363
 Figure 3. Concept of bus monitoring system for short- and
medium-span bridges.

Figure 1. The ratio of smart city projects each fields.

Figure 4. The composition of data collecting system by bus.

Figure 2. The ratio of composited fields with “Infrastruc­

ture maintenance” field.

interval. So there these data collecting interval

is needed more short.
(3) If many sensors will be installed all bridges for
monitoring their condition, huge number of
sensor and amount of budget will be needed.
In this research the one of solution is proposed to
solve the above problems by a system of using Figure 5. A summary of the principle for extracting CD.
heavy vehicle like a city bus. The city bus is con­
nected between cities and operate on the important condition index, namely “Characteristic Deflection”
road in them. This system is collecting bridge and (shortened to CD) which is relatively free from the
road surface deterioration data on the root. And they influence of dynamic disturbances due to such factors
are also a kind of a practical social big data. as the unevenness of the road surface, and a structural
Figure 3 is shown the concept image of city bus anomaly parameter. Figure 5 is shown a summary of
monitoring system for short- and medium-span the principle for extracting CD. This system is used
bridges. Global Positioning System (GPS). This system is pin­
pointing bus’s position by GPS sensor data and
detecting ups or downs of a road. If the bus is over
2 THE DATA COLLECTING SYSTEM BY BUS the target bridge, this system automatically measures
acceleration data with position data. It is expected
2.1 The principle of detecting deterioration that this automatic data collection system will enable
to create low cost bridge management system.
Figure 4 is shown the composition of data collecting When a vehicle passes over a bridge, its mechan­
system by bus. An acceleration sensor is installed on ics model can be expressed as a dynamic interaction
the rear axis. This system has using an original bridge between the equation of motion of the bridge,

364
expressed by Eq. (1), and the equation of motion of bridge is proportional to the axis vibration of the
the vehicle, expressed by Eq. (2). The method of bus, and changes in Ab due to bridge anomalies are
formulating a bridge structure model and a vehicle reflected proportionately in changes in axis vibration
structure model as separate equations of motion and As of the bus.
expressing their interaction at joints with input and It is come to synchronize between the axis and the
output vectors is called the “substructure method”. bridge.
On the other hand, the road surface condition
index can be calculated by the acceleration data with
position information. A variance of acceleration in
the certain area has relation with road surface condi­
tion. If the road surface was flatly then a variance of
acceleration is small. And if it was rough then
a variance of acceleration is big (Figure 6).
In the case of investigating the road surface condi­
where the parameters are as follow tion by the acceleration data with position informa­
Mm,Cm,Km,Cs,Ks: mass/damping/stiffness matrix tion, “Area Block” is defined on the map. “Area
for bridge and for vehicle, Block” is made by position information log of the
u€m u_ m um ; u€s u_ s usf : responses for acceleration/vel­ vehicle. “Area Block” is composed longitude, lati­
ocity/displacement vector for bridge and for vehicle, tude and radius (R) like Figure 7. When the vehicle
usg: input forced displacement vector for crossing on the some “Area Block”, the acceleration
vehicle, and data with position information was extracting and
Rm Rs: support reaction vectors for bridge and calculating a variance of acceleration. After calculat­
vehicle. ing a variance of acceleration at all “Area Block”,
In a Mechanics problem in which the difference variances of acceleration are compared each other
method approximation holds true, the equation of and putting label of road surface condition on them.
motion is shown Eq. (3) by using the proportionality This system is able to collect not only the bridge
constant P, which is dependent on physical quantities condition data but also the road surface condition.
that remain unchanged in the system, such as time,
stiffness, damping and mass, and the state constant 2.2 The method of detecting bridge deterioration
Co (known) before time t obtained by a Taylor
expansion. This means that the system’s response to Figure 8 is shown the process of getting “Character­
the input vector is allocated proportionately depend­ istic Deflection”. In order to extract exactly position
ing on the system-dependent constants. when the bus is passing on the bridge, this system
need measuring acceleration one time on the center
of bridge. And the Synchronization between the bus
rear axis and the bridge has to confirmed by measur­
where the parameters are ing same time the bus rear axis and the bridge
M,C,K: mass/damping/stiffness matrix acceleration.
u(t): displacement vector,
Co: known value before time t obtained by Taylor
expansion, etc.,
F(t): input vector to vehicle system
P: proportionality constant dependent on system.
In here substructure segmentation of the bus is
assumed separating body and axis-bridge. Because the
spring of the bus is soft enough for the assumption. In
doing so, the axis–bridge vibration transmitted from Figure 6. The relation between road surface condition and
the bus body. The response of the axis and the bridge a variance of acceleration.
to the input from the bus body is allocated proportion­
ately depending on the physical constants of the
system. Let Ab represent the response vector of the
bridge, and As is the axis response vector of the bus.
Then, we have matrix P expressed as shown below:

This means that if Eq. (4) holds true in a Mechanics

problem in which the difference method approxima­ Figure 7. An image of “Area Block”.
tion holds true, then the vibration behavior of the

365
 The first process is extracting the bridge acceler­
ation data when the bus was passing on the bridge.
Because the extracting data is roughly positon
(Figure 9 (a)), more exactly position is detected by
wavelet analysis. Figure 9 is shown while some
vehicles had been passing on the bridge, around
12Hz vibration appeared. In this data at first around
12Hz moving time length is 7seconds, next is
9seconds. Already known passing time of the bus
was 7seconds. So It was estimated that the bus had
been passing at red the circle time on the Figure 9.
Figure 10 is shown comparing the accelerations
between the bridge and the bus acceleration. Expand­
ing that time at passing on the bridge. Red is the bus.
Blue is the bridge. Estimated synchronized time
between the bridge and the bus. And getting the

Figure 10. Compare the accelerations between the bridge

and the bus.
Figure 8. The process of getting “Characteristic Deflection”.

Figure 9. One of sample bridge vibration passing over bus.

366
Figure 11. The sample of plotted “Area Block” on the map
in a day operating.
Figure 14. Shingondai Bridge.
vibration pattern of the rear axis acceleration at passing
on the bridge. On and after, the other measured rear
axis acceleration at passing on the bridge is extracting
by the vibration pattern and calculating “CD”.

3 THE CASE OF YAMAGUCHI PREFECTURE

Figure 11 is shown the sample of plotted “Area

Block” on the map in a day operating at Ube in Yama­
guchi Prefecture by Ube city bus. Three bridges on the
bus route are extracting their condition information
Figure 15. Comparison of bridge and rear axis vibration at
where are “Jase Bridge”, “Shiratsuchi Daini Bridge”, Shingondai Bridge.
“Shingondai Bridge” (Figure 12,13,14). Figure 15 is
shown synchronization between the bus rear axis and
the bridge at Shingondai Bridge. Other bridges were able to calculate “CD”. In the case of Shingondai
confirmed synchronization too. So these bridges are Bridge, to model deterioration of a PC bridge in the
simulation (Figure 16), Table 1 was applied by assum­
ing that deterioration can be expressed as a decrease in
the geometrical moment of inertia related to
a decrease in prestressing force (Deterioration Phase 1
and Deterioration Phase 2). The limite “CD” value of
Shingondai Bridge by simulation and monitored “CD”
is plotted such as Figure 17.

4 THE CASE OF KUMAMOTO PREFECTURE

Figure 18 is shown the sample of plotted “Area

Block” on the map in a day operating in Kumamoto
Prefecture. The red point is the bigger 20% of the
variance of acceleration. It is means that the road
Figure 12. Jase Bridge (five-span).
surface in the red point area is rougher than other
area. And the red point is suggesting that these areas
have been damaged heavy action by vehicles.

Table 1. Assuming for Shingondai Bridge deterioration.

Decrease in Geomentrical moment of

Soundness of prestressing inertia ratio relative to
bridge force decrease of 0%

Sound 0% 1
Deterioration 50% 0.52
Phase 1
Deterioration 90% 0.35
Phase 2
Figure 13. Shiratsuchi Daini Bridge (two-span).

367
Figure 16. One of sample bridge vibration passing over bus.

Figure 17. Histories of CD at Shingondai Bridge.

368
Figure 18. The sample of plotted “Area Block” on the map Figure 20. Comparison of bridge and rear axis vibration at
in a day operating. Bridge.

Figure 19. Umamizu Bridge.

Figure 21. Ougi bridge and Crane truck.
A bridge on the bus route is extracting their con­
dition information where is “Umamizu Bridge” bridge and the rear axis of the crane truck acceler­
(Figure 19). Figure 20 is shown synchronization be­ ation. Expanding that time at passing on the bridge.
tween the bus rear axis and the bridge at Umamizu Orange is the crane truck. Blue is the bridge.
Bridge. So this bridge is able to calculate “CD”. This In case of this test, the rear axis of the crane truck
bridge is considering deterioration condition and acceleration was not synchronized with the bridge.
limite value of “CD”. Because enough data is short, So this crane truck is difficult to use collecting infor­
data is collecting now and in future. mation of bridge condition by “CD”.
Figure 23 is shown comparison of structural
models and vibration form by action from tire
5 A TEST OF USING A CRANE TRUCK between the bus and the crane truck. The bus rear
axis is moving by action separating with body. On
A test of using a crane truck had been carried out on the other hands the crane truck is moving by action
“Ougi brige” in 2017 at Kawasaki in Kanagawa Pre­ with body. The reason of difference is inferred that
fecture. Figure 21 is shown the crane truck was pass­ (1) The bus spring is soft, because the bus has to
ing on the “Ogi Bridge”. Figure 22 is shown keep comfortable to passengers. So the body
comparing the accelerations between the center of motion is separated from the rear axis.

Figure 22. Comparison of bridge and crane truck rear axis vibration Bridge.

369