Structural Damage Detection in a Truss Bridge Model

Using Fuzzy Clustering and Measured FRF Data

Reduced by Principal Component Projection

Ling Yu1,3,*, Jun-hua Zhu1,2 and Li-li Yu1,4

1MOE Key Lab of Disaster Forecast and Control in Engineering, Jinan University, Guangzhou 510632, China
2The 5th Electronics Research Institute of the Ministry of Industry and Information Technology, Guangzhou 510610, China
3Department of Mechanics and Civil Engineering, Jinan University, Guangzhou 510632, China
4School of Materials Science and Engineering, Hebei University of Technology, Tianjin 300130, China

Abstract: This study deals with vibration-based damage detection in a truss bridge
model and suggests a novel methodology based on fuzzy clustering and measured
frequency response function (FRF) data reduced by principal component projection. A
six-bay truss bridge model is designed and fabricated in laboratory, various connection
damages are simulated by loosening the end connecter bolts, and the environmental
effects are taken into account by changing in excitation force levels of a shaker. The
FRFs of the healthy and the damaged structure are used as initial data. The FRF data
normalization is performed for eliminating the effects caused by the environmental
and operational variability. Two data projection algorithms, namely principal
component analysis (PCA) and kernel principal component analysis (KPCA) are
adopted for data compression and the median values of principal components are
defined for damage feature extraction. The fuzzy c-means (FCM) clustering algorithm
is used to categorize these features for structural damage detection. The illustrated
results show that the proposed method can effectively identify the bridge damages
simulated by loosening the bolted joints of the truss bridge structure. It is sensitive to
the structural damage but it is non-sensitive to the effect of the environmental and
operational variations. This makes it quite generic and permits its potential
development for real and complex truss bridges in site.

Key words: structural health monitoring, truss bridge, structural damage detection, eigenspace projections, fuzzy
clustering.

1. INTRODUCTION detection methods have been widely developed over the

In the past few decades, structural health monitoring past years (Alvandi and Cremona 2006; Fan and Qiao
(SHM) has been becoming one of the most popular 2011) and regarded as one of the most attractive
research areas in the bridge engineering field (Carden technologies. The basic premise of the global damage
and Fanning 2004; Farrar and Worden 2007; Ou and Li detection methods that examine changes in the dynamic
2010; Yu and Xu 2011). Although SHM is inarguably a properties is that modal parameters, namely natural
broader research area, most of SHM applications have frequencies, mode shapes and modal damping, are a
been focusing on damage detection (Fan and Qiao function of the physical properties of the structure
2011). Detecting structural damage is an important but (mass, damping, stiffness and boundary conditions).
still challenging task because damage is an intrinsically However, from the viewpoint of bridge health
local phenomenon. As an alternative to the current local monitoring, dynamic properties of a bridge can be
inspection methods, global vibration-based damage significantly affected by changing environment

*Corresponding author. Email address: lyu1997@163.com; Fax: +86-20-8522-7932; Tel: +86-20-8522-0476/2012.

Advances in Structural Engineering Vol. 16 No. 1 2013 207

 Structural Damage Detection in a Truss Bridge Model Using Fuzzy Clustering and Measured FRF Data Reduced
by Principal Component Projection

temperatures (Peeters and Roeck 2001), and may be et al. 2008) may classify the features as well. Relatively,
amplitude-dependent, especially for long-span cable- fuzzy clustering can categorize the features into two
supported bridges (Zhang 2007). Environmental and groups, i.e., damaged and undamaged states without using
operational effects may induce significant changes on a specified threshold value (da Silva et al. 2008).
the dynamic characteristics of the structure and these Therefore, the fuzzy clustering algorithm is used for the
effects can mask the changes caused by structural recognition objective since the features developed here do
deterioration (Catbas and Akatan 2002; Sohn 1999). not exhibit the property of normal distributions.
Meanwhile, data measured from an actual bridge are In general, statistical pattern recognition based
inevitably contaminated with noise (Yin et al. 2007). methods seem to be suitable under the conditions where
These issues make it difficult to discriminate the a clear physical model of the structure of interest is not
changes in structural responses caused by damage from available. A drawback of this method, however, seems
those caused by changing environmental and to be the visualization of the outputs from the signal
operational conditions or those due to noise. processing, due to the high number of measurement
To address these issues, Farrar et al. (2000) posed the points. In this paper, an integrated procedure is
damage detection problem in the context of a statistical presented for structural damage detection, which
pattern recognition paradigm. The premise of the combines the data normalization, projection algorithm
statistical pattern recognition approach is that as the and fuzzy clustering techniques. The frequency
model is trained for the baseline model, new data response functions (FRFs) of both the healthy and the
coming from the damaged structure will likely be damaged structures are used as initial data, median
classified as outliers in the data. The core issue of values of the projections are considered as damage
statistical pattern recognition approach is to seek the features, and the fuzzy c-means (FCM) clustering
damage sensitive features and to determine the critical algorithm are used to categorize these features. Two
damage values (Fugate et al. 2001). projection algorithms, namely principal component
For damage sensitive features, Worden et al. (2000) analysis (PCA) and kernel principal component analysis
used transmissibility function as damage-sensitive (KPCA) are compared for better extraction of damage
features, whereas Sohn et al. (2001) used the coefficients features. The performance of the proposed method has
of the AR models as damage-sensitive features which been verified using a truss bridge model fabricated in
were analyzed by using X-bar control charts. Manson et laboratory. Further, some issues related are compared
al. (2003) also used similar methodologies to analyze and discussed as well.
data coming from different test specimens. In another
study, the direct use of measured raw frequency response 2. BACKGROUND OF THEORY
functions – FRFs for structural damage detection may In general, vibration-based damage detection method
represent, in most cases, a considerable advantage. The promises that damage will alter modal parameters of a
measured raw data should be compressed without losing structure, which can be obtained directly from one
much useful information before structural damage column of frequency response function (FRF) matrix. In
detection is conducted due to more redundant fact, it is sufficient to measure one column of FRFs for
information and less noise immunity. Principal structural damage detection.
component analysis (PCA) is a well-known tool for
feature extraction and dimensionality reduction. It has 2.1. FRF Data Normalization
been widely applied to the damage detection field (Zang Suppose hi ∈ Rn × 1 (i = 1, 2… p), denotes amplitudes of
and Imregun 2001; da Silva et al. 2008; Oh and Sohn FRF data with p spectral lines at all n measurement
2009). However, the PCA mostly are taken as data points. In order to eliminate the effects caused by
compression tool and rarely used to directly extract the environmental and operational variations from the
damage features of structures (Yu et al. 2010; Bellino et measured FRFs, the data standardization is necessary as
al. 2010; Li et al. 2011). follows
In order to distinguish the healthy and damaged
conditions of structures, it is very important to find a hi − h
ĥi = (1)
critical threshold value for structural damage detection. σ
There are many approaches to determine the critical
values, such as statistical process control methods (Fugate
et al. 2001), fuzzy clustering (da Silva et al. 2008),
sequential probability ratio test (Oh and Sohn 2009) and 1 p
so on. Other simple discriminated analysis (Trendafilova h= ∑ hi
p i =1
(2)

208 Advances in Structural Engineering Vol. 16 No. 1 2013

Ling Yu, Jun-hua Zhu and Li-li Yu

The eigenvectors are obtained from the normalized

2 1 p
σ = ∑ (hi − h )2 (3) FRF data in a reference state in this paper and the test
p − 1 i =1 points comes from the normalized FRFs data either in a
health or in a damage state of structures.
–
where h , σ 2 and ĥi are the mean, variance and
2.2.2. Kernel principal component analysis
standardized values of hi respectively.
(KPCA)
KPCA is an extension of PCA using techniques of
2.2. FRF Data Projection kernel methods (Schölkopf 1998). It also can be derived
The measured FRF data contain more raw information by extension of least-squares support vector machine
of structures, but it is impossible to use full-size FRF (Oh and Sohn 2009; Nguyen and Golinval 2010).
data in a damage detection process due to computation Firstly, the FRF data point are mapped into another
cost and effect of noise. Therefore, data reduction and feature space,
feature extraction technology are used here. Two
projection algorithms, namely principal component
hˆi → φ (hˆi ) (7)
analysis (PCA) and kernel principal component
analysis (KPCA) are adopted before extracting the
damage features. In this space, standard PCA is performed. The
covariance matrix C2 of φ (ĥi) is computed by
2.2.1. Principal component analysis (PCA)
p
PCA is a powerful technique for feature extraction and 1 1
C2 = ∑ φ (hˆi )φ (hˆi )T = Φ ΦT (8)
data dimensionality reduction. It is mathematically p − 1 i =1 p −1
(Jolliffe 2002) defined as an orthogonal transformation
of raw data to a new coordinate system such that the
which can be diagonalized with non-negative eigenvalues
greatest variance by any projection of the data comes to
satisfying
lie on the first coordinate (called the first principal
component), the second greatest variance on the second
C 2 V = VΛ (9)
coordinate and so on. Theoretically, PCA is an
optimum transform for given data in terms of least
square meaning. In fact, PCA attempts to extract the where Φ = [φ (ĥ1), φ (ĥ2), …,φ (ĥp)], Λ = diag. [λ1, λ2,
features from the original FRF data by a linear …, λn] is a diagonal matrix constructed from the
orthogonal projection. eigenvalues of C2, V = [v1, v2, …, vn] ∈ Rn × n is a matrix
The covariance matrix of ĥi can be computed by composed of the corresponding eigenvectors, each
column of V lies in the span of φ (ĥi)(i = 1, 2, …p) and
can be expressed as (Schölkopf 1998),
1 p ˆ ˆT
C1 = ∑ hi hi
p − 1 i =1
(4)
 p p p 
V =  ∑ ai1φ (hˆi ), ∑ ai 2φ (hˆi ),..., ∑ ainφ (hˆi )  = ΦA (10)
 i =1 i =1 i =1 
Then, the PCA can be performed by solving the
eigenvalue problem as following
where A ∈ Rp × n is a coefficient matrix, whose ij-th
element is aij. Substituting Eqns 8 and 10 into Eqn 9,
C1W = W ∑ multiplying both sides by φ (ĥi)T, and rewriting the
(5)
expression as another eigenvalue problem,
where Σ = diag.[s1, s2, …, sn], W = [w1, w2, …, wn], si 1
and wi are i-th eigenvalue and corresponding KA = AΛ
Λ
eigenvector, respectively. The k-th feature of the p −1 (11)
normalized test sample point ĥ, also called k-th principal
component, is the projection of ĥ on the k-th eigenvector K ij = φ (hˆi )T φ (hˆ j ) = κ (hˆi , hˆ j ) (12)
wk, which can be computed by

It can be seen that Λ and A are the eigenvalues and

PCk (h) = wkT hˆ (6) corresponding eigenvectors matrix of K/(p−1)∈Rp × p, K

Advances in Structural Engineering Vol. 16 No. 1 2013 209

 Structural Damage Detection in a Truss Bridge Model Using Fuzzy Clustering and Measured FRF Data Reduced
by Principal Component Projection

called the kernel matrix, whose ij-th element is the

β ktest
inner-product kernel κ (ĥi, ĥj). In this paper, the γ kpca ( k ) = (18)
polynomial kernel κ (ĥi, ĥj) = (ĥTi ĥj + 1)d is utilized. Here β kref
d is the degree of polynomial, which is set to be 2. The
eigenvectors matrix V is normalized as I = VTV while where αk and βk denote the median values of the k-th
the coefficient matrix A is normalized and translated principal component obtained by PCA and KPCA
into ATA = Λ −1/(p−1). respectively. The superscripts in Eqns 17 and 18, i.e. test
The k-th nonlinear principal component of a test point and ref, denote the median values in the test and
can be extracted in the form reference states respectively. If no damage occurred, the
damage index (γpca, γkpca) theoretically equal to unity,
T otherwise, it should be greater or smaller than unity. In
 p 
KPCk (φ (hˆ)) = vkT φ (hˆ) =  ∑ aik φ (hˆi ) φ (hˆ) fact, the damage index cannot absolutely equal to unity
 i =1  even in a health state due to effect of environmental
p p
(13) factors. Structural damage detection can be performed
( )
= ∑ aik φ (hˆi ) φ (hˆ) = ∑ aikκ (hˆi , hˆ)
T
according to the whole features of damaged indexes.
i =1 i =1
2.4. Structural Damage Detection based on
In summary, the following steps were necessary to Fuzzy Clustering
compute the principal components (PCs): first, compute In the previous section, damage index has been defined,
the kernel matrix K defined by Eqn 12; second, compute but it is difficult to choose a threshold value that
its eigenvector matrix A and normalize it; third, characterizes damages. In order to perform the damage
compute projections of a test point onto the eigenvectors detection, fuzzy c-means clustering (FCM) algorithm,
by Eqn 13. In fact, the mapping φ does not need to be which was first presented by Bezdek (1981), and
computed explicitly since the whole procedures only recently applied to SHM problems by da Silva et al.
need kernel matrix K, which can be implicitly obtained. (2008), is employed to clarify the features and make a
Care must be taken into consideration regarding the fuzzy decision by using the membership of damage
fact that whether ĥi has zero-mean in its original space index in a cluster. This algorithm is an unsupervised
or not, it is not guaranteed to be centered in the feature classification algorithm which uses a certain objective
space. Since the centered data is required to perform an function described in Eqn 19 to determining the local
effective PCA, the centralized version of K can be minima iteratively.
computed by (Schölkopf 1998)
C N
K* = K − 1 p K − K1 p + 1 p K1 p (14) min J (C , m) = ∑ ∑ uijm dij2 (19)
i =1 j =1
where 1p denotes a p-by-p matrix for each element
taking value of 1/p. N
∑ uijm x j
2.3. Damage Features Extracted from j =1
Projection Data centeri = N
(20)
It is impossible to use all projection data obtained in the ∑ uijm
preceded section. Structural damage features should be j =1
extracted further from the PCs. Recently, the median
values of PCs are proposed in the reference (Yu et al.
dij2 = ( x j − centeri )T ( x j − centeri ) (21)
2010), which shows that they can reflect the structural
state. By definition,
−2
α k = median( PCk (hˆ)) (15) (dij m
) −1
uij = −2 (22)
C

β k = median( KPCk (hˆ)) (16)

∑ (d kj m
) −1
k =1

α ktest where C is the total number of clusters and N the total

γ pca ( k ) = (17) number of objects in calibration. uij is the membership
α kref function associated with the j-th object of the i-th

210 Advances in Structural Engineering Vol. 16 No. 1 2013

Ling Yu, Jun-hua Zhu and Li-li Yu

cluster, which is updated by using Eqn 22 in each 2 4 6 8 10

T8 T9 T10 T11 T12 T13 T14

iteration step. The exponent m is a measurement of
fuzzy partition. centeri is the centroid of the i-th cluster, T1 T2 T3 T4 T5 T6 T7

xj is j-th object of data set to be clustered, which is set to 1 3 5 7 9

be γ here, dij denotes the distance between the j-th object T20
12
T21 T22
14
T23 T24

16
and the centroid of i-th cluster. Here, Euclidean distance T15 T16 T17 T18 T19
11 13 15
is defined as Eqn 21 (Matlab 2000).
Excitation Accelerometer and its direction

3. EXPERIMENTAL VERIFICATION
3.1. Truss Bridge Structure Figure 2. Layout of acceleration measurement and excitation location
After a lot of simulations have been carried out, the
new method proposed in the last section will be verified
by various experiments in this section. A deck-truss electro-dynamic shaker is vertically attached at the node
composite bridge structure is designed and fabricated in T23 using a stringer through a PCB ICP 208C02 force
laboratory as shown in Figure 1, which consists of a sensor with sensitivity of 50 mV/lbs.
bridge deck and a six-bay truss structure. The bridge The LMS Vibration Measurement and Modal
deck is a uniform Q235 steel plate (3100 mm × 450 mm Analysis system is used to record the acceleration
× 4.5 mm) stiffened with five hollow rectangular ribs response and excitation force signals with the SCMDAS
(30 mm × 15 mm × 1.9 mm) welded under the plate. Mobile Front-End Module SCM05-VB08. The LMS
The deck is put on the six laterally horizontal tubes of Test.lab software is used to analyze the sampled data.
the truss and connected through U-shaped bolts. The data is recorded at a sampling rate of 320 Hz. The
The six-bay 3D truss structure consists of fifty-four excitation band is 0–160 Hz. Total sample time period is
stainless steel tubes (φ22 × 1.5 mm thick) jointed together 99.2s, the length of sample points are 31744 for each
by twenty-four standard Zinc copper alloy ball nodes. acceleration response.
Each tube is fitted with a screwed end connector, which,
when tightened into the node, also clamps the tube by 3.3. Damage Cases
means of an internal compression fitting. All the In order to simulate structural connection damage, five
connection bolts are tightened with the same torsional damage cases are set up by loosening connection bolts
moment to avoid asymmetry or nonlinear effects caused between node and tube as shown in Table 1, in which
by man-made assembly errors. The length of all the the capital letters H and D represent healthy and damage
horizontal and vertical tube members between the centers states of the truss bridge respectively. Cases 01D and
of two adjacent nodes is exactly 500 mm, but the length 02D are for single connection damages of the bridge,
of all diagonal members is 707 mm after assembly. The cases 03D, 04D and 05D are for multiple damages with
whole deck-truss structure is simple supported at two increasing damage extent. It is easy to find that the
ends through two ball nodes at each end. damage extent is increased with changes in structural
states from 01D through 05D. Moreover, three cases are
3.2. Experimental Setup shown in Figure 3, which shows that the relative places
There are twenty-four ball nodes in the truss structure, of the end connecter are obviously different for each
in which sixteen nodes are selected to vertically mount case, especially for the pins in the pinhole.
PCB ICP 333B30 single axis accelerometers with
sensitivity of 100 mV/g as shown in Figure 2. An 3.4. Dynamic Characteristics of Bridge
Although the experimental modal analysis is not
required to extract frequency and corresponding modal
shape for the method proposed in this study, it is still
conducted to understand the effects of damage states
and environmental conditions on the dynamic
characteristics of the bridge.
According to the order of structural state in Table 1,
the first experiment is conducted in the healthy state
01H, i.e. the baseline state of bridge. All the sixteen
acceleration responses and one specific excitation force
signals are recorded simultaneously and the
corresponding sixteen FRFs are then calculated. The
Figure 1. Deck-truss composite bridge structure PolyMAX method in the modal modulus of LMS

Advances in Structural Engineering Vol. 16 No. 1 2013 211

 Structural Damage Detection in a Truss Bridge Model Using Fuzzy Clustering and Measured FRF Data Reduced
by Principal Component Projection

Table 1. Structural damage scenarios

Structural state Description

01H Healthy state 01
01D Loosening bolt at the end T2 of T2-T15 tube and then fastening by hand
02D Completely loosening bolt at the end T2 of T2-T15 tube
03D Half loosening bolts at two ends of T2-T15 tube
04D Completely loosening bolts at two ends of T2-T15 tube
05D Completely loosening bolts at two ends of T2-T15 tube, loosening bolt at the end T21 of
T20-T21 tube and then fastening by hand
02H Healthy state 02, recovered from the damage state 05D

rate of the bridge frequency in levels of 3V, 5V, 7V, 9V

with respect to ones in the level of 1V is shown in
Figure 4, where V9:V1 indicates the frequencies in the
level of 9V with respect to ones in the level of 1V,
others can be identified in the similar way.
It can be seen from Figure 4 that the structural
frequencies decrease with increase of environmental
01H 01D effect, i.e. the excitation force level. But the change rate is
obvious different for each vibration mode. The highest
change is corresponding to the fourth mode. The sixth
mode is next to that. The highest change rate is −3.28%,
which occurs in the 9V level with respect to the 1V level
of excitation force. The modal testing of Alamosa Canyon
Bridge in New Mexico, USA (Doebling and Farrar 1997)
showed that the first mode frequency of the bridge varied
02D
approximately 5% during the 24 h cycle. Therefore, the
Figure 3. Layouts of end connecter under various cases environmental effect simulating through various
excitation force level is appropriate here. This simulation
way can approximately represent the practical effect of the
Test.Lab software is selected to obtain the modal environmental conditions of the bridge.
properties, namely, frequency, modal shape and
damping in each case. For the following each damage 3.4.2. Effects of various damage
case, the same way is repeated accordingly for the When the excitation force level is 1V, the first ten
corresponding modal properties. frequencies of bridge in the healthy state 01H and the
3.4.1. Effects of environmental conditions 0
Damage detection is based on the premise that damage
in the structure will cause changes in the measured −0.5
vibration data. However, many existing methods neglect
the important effect of changing environmental and −1
Frequency rate (%)

operational conditions on the underlying structure. For

−1.5
in-service structures, the variability in dynamic
properties can be a result of time-varying environmental −2
and operational conditions. In this study, the
environmental conditions are simulated by various −2.5
excitation force levels of the shaker, which includes five −3.28 V3:V1
V5:V1
−3
levels for each structural state, i.e. 1V, 3V, 5V, 7V and V7:V1
V9:V1
9V respectively. In the healthy state 01H, the first ten −3.5
1 2 3 4 5 6 7 8 9 10
frequencies of the truss bridge are obtained. They are
Modal number
19.436, 26.005, 41.771, 62.264, 68.386, 99.84, 115.312,
121.629, 134.97, 144.193 Hz respectively. The change Figure 4. Frequency rate in various environmental conditions

212 Advances in Structural Engineering Vol. 16 No. 1 2013

Ling Yu, Jun-hua Zhu and Li-li Yu

Table 2. Comparison on frequency change rate under various cases

Frequency (Hz) Frequency change rate

Order 01H 01D 02D 03D 04D 05D 02H
1 19.436 −0.376 −2.192 −0.931 −0.931 −0.962 0.010
2 26.005 −0.077 0.031 −0.169 −0.162 −0.138 0.300
3 41.771 −0.730 −2.854 −2.952 −2.775 −2.825 0.429
4 62.264 −0.011 −0.338 −0.450 −0.441 1.649 2.213
5 68.386 −0.181 −0.146 −0.608 −0.554 −0.412 0.465
6 99.840 −0.318 0.128 −0.795 −0.726 −0.775 2.102
7 115.312 −0.014 0.100 −0.004 −0.016 0.251 0.120
8 121.629 −0.129 0.131 −0.278 −0.266 −0.013 0.548
9 134.970 −0.095 0.162 −0.281 −0.265 −0.021 0.705
10 144.193 0.006 −0.029 −0.044 −0.035 −0.078 0.014

change rate of first ten frequencies in damage states with 8

f1
respect to ones in state 01H are listed in Table 2

Cumulative frequency rate (%)

7 f2
respectively. 6
f3
f4
It can be seen that most of damages cause the f5
5
decrease of frequencies but some frequencies increase f6
as well, especially for ones in the healthy state 02H, 4 f7
f8
which is recovered from the damage state 05D to the 3 f9
healthy state 02H. Where the changes in the fourth and 2
f10
sixth frequencies are the most significant, which reach
1
to 2.213% and 2.102% respectively. The cumulative
0
rate of frequency change is shown in Figure 5. It can be 01D 02D 03D 04D 05D 02H
found that the highest cumulative rate occurs in the Cases
damage case 05D and the lowest one is in the case 01D.
Figure 5. Cumulative rate of frequency change in various cases
However, the cumulative rate in the healthy case 02H is
also very high, which almost close to the highest one.
This makes the structural damage detection more for each structural state. However, only one measured
complicate. It is very difficult to identify the structural data at the healthy state 01H is choose as the reference
damage if the structural frequencies are used only. sample in the mentioned procedure, the case number is
Therefore, the new structural damage detection 1. All the other data are set to be test sample under any
proposed in this study is needed. structural states as listed in Table 3, the case numbers
are defined from 2 to 41. Here, the reference sample
3.5. Extraction of Damage Features data and the test ones are different measurement FRF
Some previous results showed that it is beneficial to data although both of them are in healthy state 01H.
weak damage detection if median values of all principal
components are extracted (Yu et al. 2010; Zhu 2011). 3.6.2. FRF Data Matrix
But for the data projection algorithm used here, any one Based on the measured time history data of 16
measured data sample can be choose as the reference acceleration response and one excited force in each
sample, other measured data samples in various structural state, the sample points are set to be 4096, the
environmental states can be used as the test ones. FRF average times are four for reducing the
Extracting and comparing the distribution of damage environmental effects of measurement noise, in which the
features respectively in the reference and the test states, calculation sample points are 1024 in time domain and
the structural damage can be detected after the fuzzy c- the spectrum lines are 512 in frequency domain
means clustering analysis is performed. respectively. The FRF data can be first resembled into a
matrix H (512 × 16) for the reference sample, in which
3.6. Structural Damage Detection 512 is the spectrum lines and 16 the acceleration
3.6.1. Measured reference and test data samples measurement numbers. The FRF data under various test
In the various environmental conditions, five excitation states in Table 3 can be then resembled into matrixes H0,
force levels, such as 1V, 3V, 5V, 7V and 9V are set up H1, H2, H3, H4, H5 and H6 respectively in a similar way.

Advances in Structural Engineering Vol. 16 No. 1 2013 213

 Structural Damage Detection in a Truss Bridge Model Using Fuzzy Clustering and Measured FRF Data Reduced
by Principal Component Projection

Table 3. Test samples

Structural state Cases Excitation level /V Structural state Cases Excitation level /V
01H 2 1 03D 22 1
3 1 23 3
4 3 24 5
5 3 25 7
6 5 26 9
7 5 04D 27 1
8 7 28 3
9 7 29 5
10 9 30 7
11 9 31 9
01D 12 1 05D 32 1
13 3 33 3
14 5 34 5
15 7 35 7
16 9 36 9
02D 17 1 02H 37 1
18 3 38 3
19 5 39 5
20 7 40 7
21 9 41 9

3.6.3. FRF data normalization case number 11. Others can be recognized similarly in
In order to reduce the effects of the environmental and the following section. It can be seen that there is a
operational variation, the FRF matrix data are first little shift for the peak values of two FRF curves and
normalized according to the procedure in section 2.1. there are variations in some frequency points in the
Figure 6 is the comparison of the FRF before and after upper Figure 6(a). After data normalization, the lower
normalization, where the legend “case11: 01H-9V- Figure 6(b) shows that there is a good agreement with
11” is for the case 11 in Table 3, i.e. the structural each other in the lower frequency band, which shows
state 01H under the excitation force level 9V with that the data normalization has a good effect of signal
processing.

3.6.4. FRF data projection

(a) Original FRFs curve at location T2
The PCA and KPCA algorithms are used for FRF data
102
projection as procedures in section 2.2. Under the
reference state case 1 and the test states of case 19 and
Amplitude

100
case 40, the distribution of median values of principal
10−2 components is shown in Figure 7. It is easy to find that
Case1: 01H-1V-1
Case11: 01H-9V-11 they are relative close to each other for healthy states of
10−4
0 20 40 60 80 100 120 140 160 case 1 and case 40, but they are obvious different from
Frequency (Hz) the damage state case 19 either from Figure 7(a) by PCA
(b) Normalized FRFs curve at location T2
or Figure 7(b) by KPCA.
102
3.6.5. Extraction of damage features
Amplitude

100 The non-dimensional structural damage features are

then extracted following the procedures in section 2.3,
10−2 Case1: 01H-1V-1 and the comparison on distribution of structural
Case11: 01H-9V-11
10−4 damage features are shown in Figure 8 for cases 1, 20
0 20 40 60 80 100 120 140 160
and 39 respectively. It can be seen that the distribution
Frequency (Hz)
of damage features are agreement with each other for
Figure 6. Comparison on FRF curve before and after data both cases 39 and 1 although the former is recovered
normalization after more than 25 damage cases. However, the

214 Advances in Structural Engineering Vol. 16 No. 1 2013

Ling Yu, Jun-hua Zhu and Li-li Yu

distribution of damage feature for case 20 is great

different from ones for cases 1 and 39 either in upper
Figure 8(a) by PCA or in lower Figure 8(b) by KPCA.
This is beneficial for the membership of damage
feature in a clustering analysis.

3.6.6. Structural damage classification

After the fuzzy c-means (FCM) clustering algorithm
is performed, Figures 9 and 10 give the extracted
Case 1: 01H-1V-1 membership of damage features by using PCA and
Case 40: 02H-7V-40 KPCA respectively. It can be seen from Figure 9 that
Case 19: 02D-5V-19
both the healthy states 01H and 02H can be effectively
classified, the damage state 02D completely identified
as well, and the damage cases 26, 30, 31 and 32
effectively located, but most of the other damage
cases cannot be correctly identified and the damage
state 01D cannot be identified completely, this is
Case 1: 01H-1V-1
Case 40: 02H-7V-40 because the damage extent is very weak. Moreover, it
Case 19: 02D-5V-19 can be found from Figure 10 that damage detection
results due to KPCA are better than ones due to PCA.
All the 15 cases in both the healthy states 01H and
02H and most of the damage cases in damage states
can be correctly located except for the wrong
diagnosis on cases 17, 22 and 27, and damage
state 01D.

PCA
Membership in each cluster

0.8

0.6
01H 01D 02D 03D 04D 05D 02H
Figure 7. Distribution of median values under reference and test 0.4
Undamaged
states Damaged
0.2

0
100 0 5 10 15 20 25 30 35 40
Cases
10−2
γpca

Figure 9. Membership of damage features extracted by PCA

Case 1: 01H-1V-1
Case 20: 02H-7V-20
10−4 Case 39: 02D-5V-39
2 4 6 8 10 12 14 16 KPCA
Membership in each cluster

Number of PCs 1

0.8

0.6
100 01H 01D 02D 03D 04D 05D 02H
γkpca

0.4
Case 1: 01H-1V-1 Undamaged
Case 20: 02H-7V-20 0.2 Damaged
10−2 Case 39: 02D-5V-39
0
50 100 150 200 250 300 0 5 10 15 20 25 30 35 40
Number of PCs Cases

Figure 8. Distribution of structural damage sensitive features Figure 10. Membership of damage features extracted by KPCA

Advances in Structural Engineering Vol. 16 No. 1 2013 215

 Structural Damage Detection in a Truss Bridge Model Using Fuzzy Clustering and Measured FRF Data Reduced
by Principal Component Projection

4. CONCLUSIONS environmental and operational variations

An integrated procedure is proposed for structural approaches 3.28% and the frequency change
damage detection in this study. It includes data rate due to damage varies 2.2% in this study. All
selection, data normalization, projection, damage the mentioned above show that the proposed
feature extraction, and fuzzy clustering analysis for method is sensitive to the structural damage but
structural damage assessment. The frequency response it is non-sensitive to the effect of the
functions (FRFs) of both the healthy and the damaged environmental and operational variations.
structures are used as initial data, median values of the
projections are considered as damage features, and the ACKNOWLEDGEMENTS
fuzzy c-means (FCM) clustering algorithm is used to This research is jointly supported by National Natural
categorize these features. The performance of the Science Foundation of China (50978123, 51278226 and
proposed method has been verified using a truss bridge 11032005), Guangdong Natural Science Foundation
model fabricated in laboratory. Two projection (10151063201000022) and Fundamental Research
algorithms, namely principal component analysis (PCA) Funds for the Central Universities (21611512).
and kernel principal component analysis (KPCA) are
compared for better extraction of damage features. The REFERENCES
illustrated results show that: Alvandi, A. and Cremona, C. (2006). “Assessment of vibration-
(1) The environmental effect simulating through based damage identification techniques”, Journal of Sound and
various excitation force level of a shaker can Vibration, Vol. 292, No. 1–2, pp. 179–202.
approximately represent the practical effect of Bellino, A., Fasana, A. and Garibaldi, L. (2010). “PCA-based
the environmental conditions of the bridge. detection of damage in time-varying systems”, Mechanical
(2) It is very difficult to identify the structural System and Signal Processing, Vol. 24, No. 7, pp. 2250–2260.
damage if the structural frequencies are used Bezdek, J.C. (1981). Pattern Recognition with Fuzzy Objective
only because the cumulative rate of frequency Function Algorithm, Plenum Press, NY, USA.
changes in the healthy and damage states of Carden, P.E. and Fanning, P.A. (2004). “Vibration based condition
structure are very close to each other. Therefore, monitoring: a review”, Structural Health Monitoring, Vol. 3,
the new method should be proposed for No. 4, pp. 355–377.
structural damage detection. Catbas, F.N. and Akatan, A.E. (2002). “Condition and damage
(3) The data normalization is necessary because it assessment: issues and some promising indices”, Journal of
can effectively eliminate the effects of Structural Engineering, ASCE, Vol. 128, No. 8, pp. 1026–1036.
environmental and operational variations, da Silva, S., Dias Júnior, M., Lopes, J.V. and Brennan, M.J. (2008).
particularly in the lower frequency band. “Structural damage detection by fuzzy clustering”, Mechanical
(4) The distributions of median values of principal System and Signal Processing, Vol. 22, No. 7, pp. 1636–1649.
components in the damage states are obvious Doebling, S.W. and Farrar, C.R. (1997). “Using statistical analysis to
different from ones in the healthy state of the enhance modal-based damage identification”, Proceedings of
truss bridge. The distributions of damage DAMAS 97: Structural Damage Assessment using Advanced Signal
features in healthy states are agreement with Processing Procedures, University of Sheffield, UK, pp. 199–210.
each other although some healthy state is Fan, W. and Qiao, P.Z. (2011). “Vibration-based damage
recovered after more than 25 damage cases. identification methods: a review and comparative study”,
However, they are great different from any of Structural Health Monitoring, Vol. 10, No. 1, pp. 83–111.
ones from any damage states. Farrar, C.R., Duffey, T.A., Doebling, S.W. and Nix, D.A. (2000). “A
(5) The proposed method in this study can statistical pattern recognition paradigm for vibration-based
effectively identify the bridge damages structural health monitoring”, Proceedings of the 2nd
simulated by loosening bolted joints of the truss International Workshop on Structural Health Monitoring,
bridge structure. The damage detection results Stanford, CA, USA.
due to KPCA are better than ones due to PCA. Farrar, C.R. and Worden, K. (2007). “An introduction to structural
All the healthy states in the different time stage health monitoring”, Philosophical Transactions of the Royal
and four out of five damage states can be well Society A, Vol. 365, pp. 303–315.
detected. The damage features extracted either Fugate, M.L., Sohn, H. and Farrar, C.R. (2001). “Vibration-based
by PCA or by KPCA cannot be effectively used damage detection using statistical process control”, Mechanical
to identify the weak damage state 01D which System and Signal Processing, Vol.15, No. 4, pp.707–721.
causes the structural frequency change rates less Jolliffe, I.T. (2002). Principal Component Analysis, Springer, NY,
than −0.736%. However, the effect of the USA.

216 Advances in Structural Engineering Vol. 16 No. 1 2013

Ling Yu, Jun-hua Zhu and Li-li Yu

Li, J.C., Dackermann, U., Xu, Y.L. and Samali, B. (2011). “Damage Sohn, H., Farrar, C.R. and Hunter, N.F. (2001). “Structural health
identification in civil engineering structures utilizing PCA- monitoring using statistical pattern recognition techniques”,
compressed residual frequency response functions and neural Journal of Dynamic Systems, Measurement, and Control, Vol. 123,
network ensembles”, Structural Control & Health Monitoring, No. 4, pp. 706–711.
Vol. 18, No. 2, pp. 207–226. Trendafilova, I., Cartmell, M.P. and Ostachowicz, W. (2008).
Manson, G., Worden, K. and Allman, D. (2003). “Experimental “Vibration-based damage detection in an aircraft wing scaled
validation of a structural health monitoring methodology. Part II: model using principal component analysis and pattern
Novelty detection on a GNAT aircraft”, Journal of Sound and recognition”, Journal of Sound and Vibration, Vol. 313, No. 3,
Vibration, Vol. 259, No. 2, pp. 345–363. pp. 560–566.
MATLAB (2000). Toolbox User’s Guide, The Math Works, Inc., Worden, K., Manson, G. and Fieller, N.R.J. (2000). “Damage
Natick, MA, USA. detection using outlier analysis”, Journal of Sound and Vibration,
Nguyen, V.H. and Golinval, J.C. (2010). “Fault detection based on Vol. 229, No. 3, pp. 647–667.
kernel principal component analysis”, Engineering Structures, Yin, T., Zhu, H.P. and Yu, L. (2007). “Noise analysis for sensitivity-
Vol. 32, No. 11, pp. 3683–3691. based structural damage detection”, Applied Mathematics and
Oh, C.K. and Sohn, H. (2009). “Damage diagnosis under Mechanics, Vol. 28, No. 6, pp. 741–750.
environmental and operational variations using unsupervised Yu, L. and Xu, P. (2011). “Structural health monitoring based on
support vector machine”, Journal of Sound and Vibration, Vol. 325, continuous ACO method”, Microelectronics Reliability, Vol. 51,
No. 1–2, pp. 224–239. No. 2, pp. 270–278.
Ou, J.P. and Li, H. (2010). “Structural health monitoring in mainland Yu, L., Zhu, J.H. and Chen, L.J. (2010). “Parametric study on PCA-
China: review and future trends”, Structural Health Monitoring, based algorithm for structural health monitoring”, Proceedings of
Vol. 9, No. 3, pp. 219–231. IEEE Prognostics and Health Management Conference, Macau
Peeters, B. and Roeck, G.D. (2001). “One-year monitoring of the Z24- University, Macau, China, January.
Bridge: environmental effects versus damage events”, Earthquake Zang, C. and Imregun, M. (2001). “Structural damage detection
Engineering and Structural Dynamics, Vol. 30, No. 2, pp. 49–71. using artificial neural networks and measured FRF data reduced
Schölkopf, B., Smola, A. and Müller, K.R. (1998). “Nonlinear via principal component projection”, Journal of Sound and
component analysis as a kernel eigenvalue problem”, Neural Vibration, Vol. 242, No. 5, pp. 813–827.
Computation, Vol. 10, No. 5, pp. 1299–1319. Zhang, Q.W. (2007). “Statistical damage identification for bridges
Sohn, H., Dzwonczyk, M., Straser, E.G., Kiremidjian, A.S., Law, K.H. using ambient vibration data”, Computers and Structures, Vol. 85,
and Meng, T. (1999). “An experimental study of temperature effect No. 7–8, pp. 476–485.
on modal parameters of the Alamosa Canyon Bridge”, Earthquake Zhu, J.H. (2011). Structural Damage Detection under Environmental
Engineering and Structural Dynamics, Vol. 28, No. 8, pp. 879–897. Variability, PhD Thesis, Jinan University, Guangzhou, China.

Advances in Structural Engineering Vol. 16 No. 1 2013 217

Copyright of Advances in Structural Engineering is the property of Multi-Science Publishing Co Ltd and its
content may not be copied or emailed to multiple sites or posted to a listserv without the copyright holder's
express written permission. However, users may print, download, or email articles for individual use.