Inverse Problems, Design and Optimization Symposium

Miami, Florida, U.S.A., April 16-18, 2007

FINITE ELEMENT MODEL UPDATING OF A SUSPENSION BRIDGE

USING ANSYS SOFTWARE
Renata N. Merce, Graciela N. Doz and
Jos Lus Vital de Brito
Department of Civil Engineering
University of Braslia
Braslia, DF, Brazil

John H.G. Macdonald

Department of Civil Engineering
University of Bristol
Bristol, UK

merce@unb.br
Michael I. Friswell
Department of Aerospace Engineering
University of Bristol
Bristol, UK

ABSTRACT
The need to know the behaviour of as-built
bridges is becoming increasingly important.
Bridges are structures with large dimensions, long
service lives and, with advances in technology,
they continue to become more complex. This has
generated the need to understand their behaviour
more precisely.
Model updating is a tool able to improve the
representation of bridges. A Finite Element (FE)
model is updated to achieve better correlation
with experimentally observed dynamic properties.
This paper is concerned with updating an FE
model of the Clifton Suspension Bridge (CSB),
UK. The objective is to establish an updating
process for bridges using the tools available in the
commercial FE software ANSYS.
After updating, the model of the CSB yielded
modal parameters which were in significantly
closer agreement with the experimental values,
while preserving their physical meaning.
Nomenclature
Df - Difference between the experimental
and FE natural frequencies;
f
- Objective function;
FER - Frequency error;
MAC- Modal Assurance Criterion;
x
- Design variable (updating parameter);
~

i X - FE mode shape vector;

iX

- Experimental mode shape vector;

ai - FE eigenvalue of mode that has been

ei

paired with ei ;
- Experimental eigenvalue of mode i.

INTRODUCTION
Bridges are structures that due their large
dimensions, long service lives and structural
complexity need regular inspection and condition
assessment to guarantee their serviceability,
safety and reliability. Simulation of the physical
system of a bridge is a useful tool that can help
assess its structural condition.
Some research has been performed on
condition assessment of bridges using a
combination of numerical modelling and fullscale measurement to simulate the physical
system, such as by Teughels et al. [1] and
Garibaldi et al. [2] for the Z24 Bridge in
Switzerland, and by Jaishi and Ren [3] for the
Beichuan River Bridge in Xining City, China.
This
combination
provides
improved
understanding of the behaviour and properties of
the bridge and can help in regular inspections,
modification of the structure, extreme load
analysis, etc.
FE model updating is a procedure used to
minimise the difference between numerical and
experimental results. However, this process does
not only require satisfactory correlations between
the two sets of results, but also the need to
maintain physical significance of the updated
parameters. Updated parameters are the modified
parameters selected in the FE model with the aim

Inverse Problems, Design and Optimization Symposium

Miami, Florida, U.S.A., April 16-18, 2007

of correcting modelling errors. It has been used

for many years in the mechanical and aerospace
engineering sectors, but it is only recently that the
civil engineering community has begun to adopt
this technology.
The updating processes used in this paper
follows the flowchart shown in Figure 1. It can be
divided in two stages: manual and automatic
updating. The principle of both parts is the same;
changing the numerical model with the objective
of finding a numerical modal model closer to the
experimental modal model. The difference is only
in the method of changing the model, in one it is
performed manually, and in the other one by an
optimization method, in this case using the
commands in ANSYS [4].
Manual updating or tuning, after the initial FE
model is generated, is the first refinement of the
model. This stage principally involves making
changes to the form of the model (e.g.
characteristics represented, mesh configuration,
element types), as well as possibly improving
estimates of parameters describing the geometry
and materials, e.g. mass, area moments, elastic
modulus [5]. Only when the form of the model
represents the structure sufficiently accurately can
automatic updating be meaningfully applied.

Figure 1 - Flowchart of the updating process

After each modification of the model, modal

analysis is carried out and the results are paired
with the experimental modal model. The vibration
modes are compared using the Modal Assurance
Criterion (MAC), defined as:

{ }{ }
{ } { } { } { }

~
{ } and { } are the

[MAC X~X ]ii = ~

2
~X T X
i
i

T ~ X X T X
i i
i

(1)

X
X
where
FE and
i
i
experimental mode shape vectors respectively.
The value of MAC is between 0 and 1. A
value of 1 means that one mode shape vector is a
multiple of other, and a value of 0 means that the
two mode shapes are completely independent of
each other. Generally, values above 0.9 mean that
the modes are well correlated.
The correlation for the natural frequencies is
calculated using the frequency error (FER) that is:
ei
FERi = ai
(2)

ei

where ei is the ith experimental eigenvalue and

ai is the FE eigenvalue of the mode that has
been paired with it.
Manual tuning can be done as many times as
the analyst judges necessary, until a satisfactory
model is achieved, both representing the physical
structure appropriately and with modal
characteristics close to the experimental ones.
This model is the starting point for the automatic
updating, and is defined here as the MT model.
The automatic updating stage is an iterative
process based on the MT model where
quantitative parameters are updated using a
computational routine. It can be divided in two
parts: pre-processing, where the parameters for
updating are defined, and automatic updating
itself, where the computational routine is applied
and the parameters are iteratively refined. It is
important to highlight that the automatic updating
stage was developed in ANSYS.
This paper focuses on the automatic updating.
A practical FE model automatic updating
technique for bridges, using the optimization
process in the ANSYS commercial FE software,
is proposed. The procedure was tested on the
Clifton Suspension Bridge, an iron chain
suspension bridge designed by Brunel, spanning
the Avon Gorge and linking Clifton in Bristol to
Leigh Woods in North Somerset, UK.

Inverse Problems, Design and Optimization Symposium

Miami, Florida, U.S.A., April 16-18, 2007

CASE STUDY

Table 1 - Details of the MT model

Bridge description
The Clifton Suspension Bridge (CSB) has a
main span of 214 m between the centres of the
towers, which are 26 m high from road level to
the chain saddles [6]. The total width of the deck
is 9.45 m. The suspension system of the bridge
consists of three wrought iron chains on each side
of the roadway, arranged one above the other.
Suspension rods, at intervals of 2.44 m, are
attached successively to each of the three chains
in turn. The deck consists of two layers of timber
beams overlaid with highway asphalt [7]. Figures
2 and 3 show the elevation and cross section of
the CSB.

Leigh
Woods

Clifton

Figure 2 - Clifton Suspension Bridge [6]

Figure 3 - Cross section of the CSB

Manually updated model MT model
A brief description of manually updated
model (MT model) and its dynamic proprieties
are presented, since this is the initial FE model for
the automatic updating stage. The model was
designed in the ANSYS Parametric Design
Language (APDL).
The chains in the main span, where the
suspension rods are attached, were each modelled
as three separate chains, but the chains in the side
spans were each modelled as only one chain with
proportional proprieties. Modelling details of the
chains, suspension rods, deck and girders are
given in Table 1.
Static analysis was first carried out to achieve
the deformed equilibrium configuration of the
bridge due to dead loads, in which the structural
members are pre-stressed. After this, the dynamic
analysis was realized, using the Block Lanczos
method in ANSYS.

Chain
Suspender

Element
Type
(ANSYS)
BEAM4
BEAM4

Deck Roadway

SHELL41

Deck Footway

SHELL41

Parapet Girders

BEAM44

Cross Girders

BEAM44

Longitudinal Girder BEAM44

Details

80 elements
in each side
of the bridge
80 elements
in each side
of the bridge
81beam
(5 elements in
each beam)
80 elements
in each side
of the bridge

Number
of
elements
226
162

Modulus of Mass
Elasticity Density Material
(N/m) (Kg/m)
1.92E+11
7800 Iron
1.92E+11
7800 Iron

240 1.00E+10

1500 Timber

160 1.00E+10

800 Timber

160 1.92E+11

7800 Iron

405 1.92E+11

7800 Iron

160 1.92E+11

7800 Iron

In the frequency range from 0.22 to 3Hz, 106

modes of vibration were identified. The mode
shapes observed were lateral modes, vertical
modes, torsional modes, longitudinal modes,
coupled vibration modes and modes where the
deck and chains interact.
Considering that the identified experimental
modes were obtained from accelerometers placed
on the deck, for this paper only modes dominated
by deck motion were used in the analysis. Thus,
the natural frequencies for torsional, lateral and
vertical deck-dominated modes are presented in
Table 2.
Table 2 - Natural frequencies from numerical
modal analysis of the MT model
m ode
num ber

m ode
num ber
2
4
18
19
22
27
43
48
53
55
56
62
63
65
69
73
82
89
96
106

m ode

num ber
1
9
31
42
45

Vertical m odes
Natural
num ber of nodes
Frequency along the bridge
deck
(Hz)
0,27834
1
0,40273
2
0,58813
2
0,60952
3
0,66184
3
0,80998
4
1,0539
5
1,2872
6
1,3491
6
1,3625
6
1,3653
6
1,5876
0
1,6185
7
1,6211
7
1,6959
7
1,9868
8
2,1666
8
2,3949
9
2,8173
10
2,9971
10
Lateral m odes
Natural
num ber of nodes
along the bridge
Frequency
deck (excluding
(Hz)
the ends)
0,22829
0
0,50095
1
0,89359
2
1,0268
1
1,2482
2

3
6
7
10
17
20
21
23
26
29
36
44
47
50
52
54
57
58
59
68
70
78
79
80
81
92
95
103
104
105

Torsional m odes
Natural
num ber of nodes
Frequency along the bridge
(Hz)
deck

0.34609
0.46473
0.47771
0.5034
0.5858
0.61098
0.65712
0.72635
0.78842
0.84847
0.96836
1.2385
1.2805
1.3264
1.346
1.3597
1.3937
1.5686
1.5708
1.6785
1.8663
2.0318
2.1241
2.1399
2.1538
2.4348
2.6315
2.9487
2.9816
2.9873

1
2
2
2
2
2
3
3
3
4
4
5
6
6
6
5
6
6
6
6
7
7
7
7
6
8
8
9
9
8

Inverse Problems, Design and Optimization Symposium

Miami, Florida, U.S.A., April 16-18, 2007

Ambient vibration testing and modal

parameter identification
A brief description of the ambient vibration
tests and modal parameter identification are
presented, but details are provided elsewhere [8].
The modal parameters of the bridge deck were
identified accurately by analysing data from a
continuous 6-day period. The modal parameters
were estimated from the response measurements
using an ambient vibration (output only) analysis
technique. A set of three accelerometers was used
as a reference set at one cross-section of the deck
and another set was used sequentially at 7 other
cross-sections to build up a picture of the mode
shapes. Two of the accelerometers measured the
vertical motion in each side of the deck bridge
from which the pure vertical and torsional
components of motion could be determined, while
the third measured lateral acceleration at the same
cross section.
The natural frequencies and damping ratios
were calculated from the Power Spectral
Densities (PSD) using the Iterative Windowed
Curve-fitting Method (IWCM), developed to
extract modal parameters from such ambient
vibration data with maximum accuracy [9].
In the frequency range from 0.2 to 3 Hz,
twelve predominantly vertical modes, eleven
predominantly torsional modes and four
predominantly lateral modes were identified, as
shown in Table 3.
Table 3 - Natural frequencies from the
experimental modal analysis [8]
mode
number
1
2
3
4
5
6
7
8
9
10
11
12

mode

number
1
2
3
4

Vertical modes
Natural number of nodes
Frequency along the bridge
(Hz)
deck
0,293
1
0,424
2
0,657
3
0,821
4
0,9
4
1,146
5
1,383
6
1,653
0
1,755
7
2,094
8
2,476
9
2,894
10
Lateral modes
Natural number of nodes
along the bridge
Frequency
deck (excluding
(Hz)
the ends)
0,24
0
0,524
1
0,746
0
0
0,965

Automatic updating
The automatic updating stage was entirely
developed in ANSYS, using its optimization
tools. However, it is important to relate the FE
model updating to optimization design.
The ANSYS optimization routines employ
three types of variables that characterize the
design process: design variables, state variables,
and the objective function.
The independent variables in an optimization
analysis are the design variables. The vector of
design variables is indicated by:
x = x1 x2 x3 K xn
(3)
and they are subject to n constraints with upper
and lower limits, that define what is commonly
called the feasible design space. It is
xi xi xi
i = 1,2,3,K, n
(4)
where n is the number of design variables. As can
be seen in Figure 4, the design variables are
related to the updating parameters. They are
defined using the gradient tool in ANSYS. The
feasible design space is defined using the random
analysis tool.

Figure 4 - Design variables and updating

parameters
Now, the aim is to minimize
f = f (x )

(5)

subject to
g i (x ) gi

(i = 1,2,3,K, m1 )
wi wi (x ) wi (i = 1,2,3,K, m2 )
hi hi (x ) (i = 1,2,3,K, m3 )

Torsional modes
mode
Natural number of nodes
Frequency along the bridge
number
deck
(Hz)

1
2
3
4
5
6
7
8
9
10
11

0.356
0.498
0.759
0.846
0.97
1.069
1.367
1.593
1.808
2.043
2.559

1
2
3
3
4
4
5
6
0
7
8

(6)

where f is the objective function, gi (x ) , wi (x )

and hi (x ) are the state variables containing the
design, with underbars and overbars representing
lower and upper bounds respectively. m1, m2 and
m3 are the number of state variables constraints
with various upper and lower limit values.
Equations 5 and 6 represent a constrained
minimization problem whose aim is the
minimization of the objective function f under
the constraints imposed by Equation 6.
The state variables are dependent on the
design variables vector, x, (i.e. the updating
parameters) and the parameters of the

Inverse Problems, Design and Optimization Symposium

Miami, Florida, U.S.A., April 16-18, 2007

experimental modal model. In practice the state

variables are the natural frequencies and mode
shapes or some function of them.
Pre-processing for automatic updating
In the pre-processing the state variables,
updating parameters, the objective function and
their constraints are defined.
Selecting updating parameters, setting up the
feasible design space, the state variables and the
objective function are crucial steps in model
updating, since these are the variables that define
the optimization problem.
State variables
The state variables for the CSB model were
defined as the differences between the
experimental and numerical frequencies ( Df ).
Df (i ) = ai ei , i = (1,2,3...n f )
(7)
n f = n fv + n ft + n fl

where n fv , n ft and n fl are the number of

vertical, torsional and lateral frequencies from
the experimental model, respectively. A total of
27 state variables were defined.
The lower and upper bounds for the state
variables are important values. It defines the error
limit that can occur in the optimization process
for the state variables. The lower and upper
bounds were chosen as 0 and 0.2Hz respectively.
Updating parameters
The updating parameters are the variables that
are modified in the updating process. As a general
rule, only uncertain parameters in the model that
the responses are sensitive to should be selected,
otherwise the updating process will be illconditioned since insufficient information is
available to estimate the parameters accurately.
Sensitivity studies examine the impact of the
variation of parameters on the dynamic and static
proprieties of the bridge. They are carried out to
determine the most sensitive parameters for FE
updating.
Sensitivity analysis computes the sensitivity
coefficient S j which is defined as:
z = S jx
Sj =

z z m z j
=
x
xxj

(8)
(9)

where the numerator represents the changes in the

measured output and the denominator the
perturbation in the parameters.
The sensitivity matrix can be computed for
material and geometric properties of the FE
models elements. Calculation of the eigenvalue
sensitivity may not be easy when the study is
carried out using commercial software whose
program code is not available and system
matrices cannot be extracted easily. In that case,
finite difference approximation is one of the
alternatives for calculation of the sensitivity. In
this approach, the eigenvalue sensitivity matrix is
approximated using the forward difference of the
function with respect to each parameter
considered.
f r
f (x + xi e ) f r (x )
= r
xi
xi

xi =

D
xi xi
100

(10)
(11)

Where x is the updating parameter, e is the

vector with 1 in its ith component and 0 for all
other components. D is the forward difference
step size (in percent), taken to be 0.2 in this study,
and xi , xi are the upper and lower limits for the
parameter (design variable).
In this paper the gradient tool in ANSYS was
used for the selection of the updating parameters.
This tool computes a gradient at a point in the
design space using the finite difference
approximation, as explained above. It was used
on the CSB model to study the sensitivities of the
objective function and the state variables in
relation to the candidate updating parameters.
The candidate updating parameters were the
mass density and modulus of elasticity of the
different elements. A total of 16 parameters were
chosen. Examples of the gradient results can be
seen in Figure 5 for the seventh and eighth
vertical modes.
A total of 28 plots were analysed, 27 were
related to the state variables and one to the
objective function. The updating parameters
chosen, based on the gradient analysis, were the
density of the side chains, centre chains and cross
girders, and the modulus of elasticity of the side
chains and centre chains.

Inverse Problems, Design and Optimization Symposium

Miami, Florida, U.S.A., April 16-18, 2007
DF - 7 Vertical Mode
4.00E-03

2.00E-03

0.00E+00

-2.00E-03

-4.00E-03

-6.00E-03

-8.00E-03

-1.00E-02

-1.20E-02

DF - 8 Vertical Mode
1.50E-03

Figure 6 was generated from the results of the

analysis for the centre span chain density. The
feasible design space is the space highlighted by
the ellipse, which covers the values of the
updating parameter that gives the lower values for
the objective function. As can be seen that the
bounds defined for the updating parameter are
approximately 7000 and 9000 Kg/m for lower
and upper limits, respectively. The limits found
for all the updating parameters are defined in
Table 4.

1.00E-03

Density - Chain (center span)

5.00E-04

0.00E+00

-5.00E-04

10000
9000
8000

-1.00E-03

-1.50E-03

-2.00E-03

7000
6000
5000
4000

-2.50E-03

3000
2000
1000
0

0.00E+00

2.00E-01

4.00E-01

6.00E-01

Objecive Function

Figure 6 - Updating parameter vs. objective

function centre span chain density
Table 4 - Lower and upper limits for the
updating parameters
Lower Upper
limits limits
7000
9000
7000
9000
3000
5000

Density - Chain (center span)

Density - Chain (side)
Density - Cross Girder
Module of elasticity - Chain
(center span)
1.0E+11 3.0E+11
Module of elasticity - Chain
(side)
1.0E+11 3.0E+11

Figure 5 - Sensitivity analysis for the 7th and 8th

vertical modes
Defining the feasible design space
In order to guarantee the physical significance
of the updated parameter values, lower and upper
bounds for the updating parameters should be
applied [10]. The choice to find the constraints for
the updating parameters has been done by random
analysis in ANSYS.
Random analysis is undertaken to define the
feasible design space. The analysis shows how the
objective function changes for different values of
the updating parameters. One analysis is
performed for each updating parameter. In this
case 5 random analyses were carried out.

un.
Kg/m
Kg/m
Kg/m
N/m
N/m

Objective function
The objective function in the updating process
reflects the difference between the numerical
predictions and the real behaviour of a structure.
In this paper it was formulated in terms of the
discrepancies between the FE and experimental
eigenvalues and mode shapes and was given by:
2

m
ei
(12)
+ i M i
f ( x ) = i ai

i =1
i
=
1
ei

where i is the weighting factor for the ith

eigenvalue, i is the weighting factor for the ith
mode shape, and M i is the mode shape related
function proposed by Moller and Friberg [11]:
m

1 MACi
M i = f (MACi ) =
MACi

(13)

Inverse Problems, Design and Optimization Symposium

Miami, Florida, U.S.A., April 16-18, 2007

where MAC is the Modal Assurance Criterion.

The choices of i and i were made by trialand-error approaches, and they were defined as
0.8 and 0.2 for i and i respectively.
Optimization
For the automatic updating itself, the MT
model, defined in the manual updating stage, was
updated based on the experimental modal model.
The method used in this work was the first order
optimization method, available in the design
optimization tool in ANSYS. This method makes
use of derivative information, that is, gradients of
the dependent variables with respect to the design
variables. It is highly accurate and works well for
problems having dependent variables that vary
widely over a large range of design space.
However, the method can be computationally
intense [4].
After automatic updating, the correlation
between the updated model and the experimental
model, using the MAC and frequency error (FER)
criteria, were calculated to define the accuracy of
the updated model. The results found for this
correlation are shown in Tables 5, 6 and 7 for the
vertical, torsional and lateral modes, respectively.
Table 5 - MACs and FERs for the vertical modes
(updated model)
Vertical mode comparison ( experimental vs.MTmodel modes)
1V
2V
3V
4V
5V
6V
FE mode number
2
4
19
27
27
43
MAC
0,9924 0,99805 0,99242 0,92011 0,95935 0,99088
FER [%]
4,12% 3,94% 5,96% 0,33% 8,48% 6,27%
7V
8V
9V
10 V
11 V
12 V
FE mode number
56
63
68
73
89
96
MAC
0,9695 0,96422 0,92356 0,99265 0,99529 0,99268
FER [%]
0,63% 0,36% 1,79% 2,60% 0,45% 0,31%

For vertical modes, the maximum frequency

error obtained was 8.48% for mode 5V. The mean
value of the FERs among all the mode shapes was
2.936%. The mean MAC for all the modes was
0.974.
Table 6 - MACs and FERs for the torsional
modes (updated model)
Torsional mode comparison ( experimental vs.MTmodel modes)
1T
2T
3T
4T
5T
6T
FE mode number
3
7
23
26
32
32
MAC
0,9909 0,97462 0,98676 0,95173 0,91679 0,92955
FER [%]
1,40% 2,04% 2,53% 3,59% 2,31% 7,16%
7T
8T
9T
10 T
11 T
FE mode number
53
59
69
76
90
MAC
0,971 0,84249 0,9374 0,90471 0,98459
FER [%]
1,38% 1,80% 3,48% 2,17% 2,10%

For the torsional modes, the maximum

frequency error was 7.16% for mode 6T. The
mean FER was 2.72% and the mean MAC 0.945.
Table 7 - MACs and FERs for the lateral modes
(updated model)
Lateral mode comparison
( experimental vs.MTmodel modes)
1L
2L
3L
4L
FE mode number
1
13
24
33
MAC
0,7734 0,97623 0,94108 0,85067
AFE [%]
2,41% 1,64% 0,83% 3,41%

For the lateral modes, the maximum frequency

error was 3.41% for mode 4L. The mean FER
was 2.07% and the mean MAC 0.885.
The natural frequencies of the experimental,
updated FE and MT models are shown in Tables
8, 9 and 10, for the vertical, torsional and lateral
modes, respectively.
Table 8 - Frequencies of the vertical modes
Vertical Modes
Natural
Natural
frequency
Natural
frequency
(Hz)
frequency (Hz)
(Hz)
(updated
Mode No. (experimental) model- UM) FER UM (%) (MT model)

1
2
3
4
5
6
7
8
9
10
11
12

0.293
0.424
0.657
0.821
0.9
1.146
1.383
1.653
1.755
2.094
2.476
2.894

0.28094
0.4073
0.61785
0.82371
0.82371
1.0741
1.3917
1.659
1.7236
2.0396
2.4649
2.9031

4.12%
3.94%
5.96%
-0.33%
8.48%
6.27%
-0.63%
-0.36%
1.79%
2.60%
0.45%
-0.31%

0.27834
0.40273
0.60952
0.80998
0.80998
1.0539
1.3625
1.5876
1.6959
1.9868
2.3949
2.8173

FER MT
(%)

5.00%
5.02%
7.23%
1.34%
10.00%
8.04%
1.48%
3.96%
3.37%
5.12%
3.28%
2.65%

Table 9 - Frequencies of the torsional modes

Torsional modes
Natural
Natural
frequency
Natural
frequency
Mode
(Hz)
frequency (Hz)
(Hz)
(updated
No.
(experimental) model- UM) FER UM (%) (MT model)

1
2
3
4
5
6
7
8
9
10
11

0.356
0.498
0.759
0.846
0.97
1.069
1.367
1.593
1.808
2.043
2.559

0.35101
0.48783
0.73983
0.87645
0.99243
0.99243
1.3859
1.6216
1.7451
2.0874
2.5054

1.40%
2.04%
2.53%
-3.60%
-2.31%
7.16%
-1.38%
-1.80%
3.48%
-2.17%
2.09%

0.34609
0.5034
0.72635
0.72635
0.96836
0.96836
1.3597
1.5708
1.6785
2.0318
2.6315

FER MT
(%)

2.78%
-1.08%
4.30%
14.14%
0.17%
9.41%
0.53%
1.39%
7.16%
0.55%
-2.83%

Inverse Problems, Design and Optimization Symposium

Miami, Florida, U.S.A., April 16-18, 2007

Table 10 - Frequencies of the lateral modes

Lateral modes
Natural
Natural
frequency
frequency
Natural
Mode
(Hz)
(Hz)
frequency (Hz)
(updated
No.
(experimental) model- UM) FER UM (%) (MT model)

1
2
3
4

0.24
0.524
0.746
0.965

0.23421
0.53261
0.73983
0.99795

2.41%
-1.64%
0.83%
-3.41%

FER MT
(%)

0.22829
0.51441
0.76607
0.9715

4.88%
1.83%
-2.69%
-0.67%

Figure 7 shows the changes of the objective

function during the iteration process of the
automatic updating. The final values of the
updated parameters can be seen in Table 11.
Objective Function vs. Iterations
0.07
0.065
0.06
0.055
0.05
0.045
1

13 19 25 31 37 43 49 55 61 67 73 79 85 91 97 103 109 115 121 127

Iterations

Figure 7 - Objective Function vs. Iteration

Table 11 - Updated Parameters
Updated
Param eters
D en sity - C h ain (cen ter spa n )
D en sity - C h ain (side)
D en sity - C ross G irder
M odule of elasticity - C h ain
(cen ter span )
M odule of elasticity - C h ain
(side)

which made this research possible and the Clifton

Suspension Bridge Trust for supporting this work.
The assistance of the late Dr J.W. Smith of the
University of Bristol is greatly acknowledged.

un .

7016.1 K g/m
7101.9 K g/m
4522.7 K g/m
1.9 E +11 N /m
2.1 E +11 N /m

CONCLUSIONS
The study of the Clifton Suspension Bridge
has shown that automatic updating of FE models
using ANSYS software is feasible. It can thus
improve prediction of the behaviour of bridge
structures and it could be used as a tool in a
bridge assessment process.
The improvement in the natural frequencies
was significant, the mean value of the frequency
error for the 27 identified modes decreased to
2.7%.
The selection of updating parameters made by
the sensitivity tool gave satisfactory results.
Considering the 8th vertical mode, which showed
high sensitivity to the updating parameters, good
improvement in the frequency error was achieved,
from 3.96% to 0.36%.
ACKNOWLEDGEMENTS
The authors gratefully acknowledge the
financial support of CAPES and CNPq (Brazil),

REFERENCES
1. A. Teughels, G.D. Roeck and J.A.K.
Suykens, Global optimization by coupled local
minimizers and its application to FE model
updating, Computers and Structures, 81, 2337
2351 (2003)
2. L. Garibaldi, S. Marchesiello and E.
Bonisoli, Identification and up-dating over the
Z24 Benchmark, Mechanical Systems and Signal
Processing, 17(1), 153161 (2003)
3. B. Jaishi and W. Ren, Structural Finite
Element model updating using ambient vibration
test results, ASCE Journal of Structural
Engineering, 131(4), 617-628 (2005)
4. ANSYS, User manual, Revision 5.6.
Swanson Analysis System (1999)
5. W.E. Daniell and J.H.G. Macdonald,
Improved finite element modelling of a cablestayed bridge through systematic manual tuning,
Engineering Structures, 29(3), 358-371 (2007)
6. W.H. Barlow, Description of the Clifton
Suspension Bridge, Minutes Proc. ICE, 26, 243253 (1867). Reprinted Proc. ICE: Bridge
Engineering, 156(BE1), 5-10 (2003)
7. W.T. Yeung and J.W. Smith, Damage
detection in bridges using neural networks for
pattern recognition of vibration signatures,
Engineering Structures, 27, 685698 (2005)
8. J.H.G. Macdonald, Dynamic behaviour of
the Clifton Suspension Bridge: Modal behaviour
in light wind and response to crowd loading,
BEELAB report CSB703/REP/1 (2003)
9. J.H.G. Macdonald, Identification of the
dynamic behaviour of a cable-stayed bridge from
full-scale testing during and after construction,
PhD thesis, University of Bristol (2000)
10. J.M.W. Brownjohn, Pin-Qi Xia, Hong Hao
and Yong Xia, Civil structure condition
assessment by FE model updating: methodology
and case studies, Finite Elements in Analysis and
Design, 37, 761-775 (2001)
11. P.W. Moller and O. Friberg, Updating
large finite element models in structural
dynamics, AIAA Journal, 36(10), 1861-1868
(1998)