Download as pdf or txt
Download as pdf or txt
You are on page 1of 26

STRUCTURAL CONTROL AND HEALTH MONITORING

Struct. Control Health Monit. 2012; 19:161 186


Published online 5 August\ 201 0 in Wiley Online Library (wileyonlinelibrary.com). DOI: 10.1002/stc.4 12

Pattern recognition of structural behaviors based on learning


algorithms and symbolic data concepts

Alexandre Cury1 and Christian Cremona2,,y


1
Laboratoire Central des Ponts et Chaussees, 58 boulevard Lefebvre, Paris 75015, France
2
Commissariat General au Developpement Durable, Direction de la Recherche et de lInnovation,
92055 La Defense Cedex, France

SUMMARY

Learning algorithms have extensively been applied to classication and pattern recognition problems in the
past years. Some papers have addressed special attention to applications regarding damage assessment,
especially how these algorithms could be used to classify different structural conditions. Nevertheless, few
works present techniques in which vibration signatures can be directly used to provide insights about
possible modication processes. This paper proposes a novel approach in which the concept of Symbolic
Data Analysis (SDA) is introduced to manipulate not only vibration data (signals) but also modal
properties (natural frequencies and mode shapes). These quantities (transformed into symbolic data) are
combined to three well-known classication techniques: Bayesian Decision Trees, Neural Networks and
Support Vector Machines. The objective is to explore the efciency of this combined methodology. For this
purpose, several numerical simulations are rst performed for evaluating the probabilities of true detection
(or true classication) in the presence of different damage conditions. Several noise levels are also applied to
the data to attest the sensibility of each technique. Second, a set of experimental tests performed on a
railway bridge in France is used to emphasize advantages and drawbacks of the proposed approach. Results
show that the analysis combining the cited learning algorithms with the symbolic data concepts is efcient
enough to classify and discriminate structural modications with a high probability of true detection, either
considering vibration data or modal parameters. Copyright r 2010 John Wiley & Sons, Ltd.

Received 15 March 2010; Revised 14 June 2010; Accepted 15 June 2010

KEY WORDS: pattern recognition; damage assessment; learning algorithms; symbolic data; probability of
true detection

1. INTRODUCTION
In the past few years, numerous methods for damage assessment in connection with structural
health monitoring were proposed in the literature. In general, most of these methods consisted
in evaluating indexes or indicators which allowed detecting possible locations and even the
extent of damaged regions [1,2] with rather low probabilities of true detection (the probability of
true detection is the probability to detect a damage when a true damage is present in the
mechanical system). In structural health monitoring, the rst step is sometimes to determine
whether a structure presents an abnormal behavior or not. For this purpose, several studies have
been outlined using statistical tests and time series analysis confronting data obtained from
healthy and damaged conditions [35]. These approaches can be useful when a FEM model is

*Correspondence to: Christian Cremona, Commissariat General au Developpement Durable, Direction de la Recherche
et de lInnovation, 92055 La Defense Cedex, France.
y
E-mail: christian.cremona@developpement-durable.gouv.fr

Copyright r 2010 John Wiley & Sons, Ltd.


162 A. CURY AND C. CREMONA

not required and when structural responses can be obtained with a certain level of condence.
In addition, methods regarding pattern recognition and learning algorithms have been applied
to vibration data in the attempt to discriminate structural conditions [6,7,8]. These methods
proved to be effective when relatively small sets of data are used for the training and
testing phases. However, when extensive observations and a large number of features are
present, these techniques can be expensive and time consuming [9]. The latter has also evidenced
the problem of stocking data obtained from acquisition campaigns. Minor et al. [10]
have proposed a methodology to assemble information gathered from several accelerometers
to detect structural damaged. It has been shown that the robustness of their methods was
strictly related to the size of the analyzed database. Guo [11] has proposed a data fusion
scheme in which accelerometers were preselected and only their measurements were used for
further analyses. Results have shown that the information loss in this case might have
conducted to false alarms when it comes to damage detection. Feature extraction techniques
have also been applied to modal parameters. Trendalova and Heyleno [12] have
employed unsupervised learning classication algorithms applied to the natural frequencies
of a cantilevered beam. Several damage scenarios were simulated through laboratory
controlled tests. Results have shown that the proposed approach allowed detecting different
structural states of the beam. Its extension to more complex structures was still to be
veried.
Data acquisition campaigns of civil engineering structures can last from several minutes to
years. Dealing with large amounts of data is not an easy task and suitable tools are required to
correctly extract important features from them. As a matter of fact, dynamic measurements
gathered from instrumented structures can easily contain over several thousands of values
making a direct analysis process extensive and perhaps prohibitive. Several damage detection
methods exist in the literature based on the signature principles, but they usually fail when
making them practical. In this sense, despite the increasing power of computers, the necessary
computational effort to manipulate large data sets remains a problem.
Data mining [13] is the process of extracting hidden patterns from large amount of data. As
more data are gathered in monitoring, data mining may be an important tool to transform this
data into information. It is commonly used in a wide range of proling practices, such as
marketing, fraud detection and scientic discovery. However, although it can be used to uncover
hidden patterns in data that have been collected, obviously it can neither uncover patterns which
are not already present in the data, nor it can uncover patterns in data that have not been
collected. To deal with this issue, it is important to recall which types of data can be employed
and manipulated in data mining [14]:

 A single quantitative value, e.g. height(w) 5 3.5; where w is an individual;


 Interval-valued data. For instance: weight (w) 5 [20, 180] which means that the weight of
w varies in the interval [20, 180];
 Multi-valued categorical data. For example: price(w) 5 fhigh, average, lowg meaning that
the general price of a product w may be high, average or low;
 Modal multi-valued (a histogram function). For instance: height (w) 5 f[0, 1.20] (0.225);
[1.20, 1.50] (0.321); [1.50, 1.80] (0.335); [1.80, 2.10] (0.119)g meaning that 22.5% of the
population w has a height varying in the interval [0, 1.20], 32.1% vary within [1.20,1.50],
for 33.5% of the population, the height is comprised between [1.50, 1.80] and 11.9 within
[1.80, 2.10].

This richer type of data refers to symbolic data and it allows representing the variability and
uncertainty present in each variable. The development of new methods of data analysis suitable
for treating this type of data is called Symbolic Data Analysis (SDA). Most of the currently
developed techniques in SDA are extensions of classical statistical methods. In general, SDA
provides suitable tools for managing complex, aggregated, relational and higher-level data. The
methodological issues under development generalize the classical data analysis techniques, such
as visualization, factorial techniques, decision tree, discrimination, regression as well as

Copyright r 2010 John Wiley & Sons, Ltd. Struct. Control Health Monit. 2012; 19:161 186
DOI: 10.1002/stc
PATTERN RECOGNITION OF STRUCTURAL BEHAVIORS 163

classication and clustering methods. One of the advantages of the SDA is the fact that it
reduces the amount of data to analyze without loss of generality [15].
This paper proposes a novel approach which combines SDA with learning and classication
algorithms. The major advantage of such hybrid techniques is that enhanced information is used
since it can be histograms, intervals, etc. They can be applied to manipulate vibration data
(signals) as well as modal properties such as natural frequencies and mode shapes. When these
quantities are converted to symbolic data, this piece of information will be applied to three well-
known classication techniques applied to structural health-monitoring problems: Bayesian
Decision Trees (BDT), Neural Networks and Support Vector Machines (SVM). Although the
authors of this paper have already applied SDA in connection with classical clustering methods
such as hierarchy-divisive or dynamic cloud methods [16], it is here the rst time this hybrid
approach is used for classifying structural behavior and for identifying damage patterns in
connection with learning algorithms. For this purpose, in order to explore the effectiveness of
this methodology, two different studies are carried out. First, several numerical simulations
including four different damage scenarios are performed using a FEM model of a simply
supported beam. Moreover, to attest the sensibility of each classication method, three noise
levels are independently added to the simulated data. The goal is to discriminate each damage
scenario applied to the structure into four different groups (where each one characterizes a
structural condition). The idea is to observe whether the classication methods are robust
enough to discriminate damages in the presence of noise. The second study uses a set of
experimental tests carried out on a rail bridge. A dynamic monitoring of this railway bridge was
performed to characterize and quantify the effect of a strengthening procedure. In so doing,
several vibration measurements were obtained under three different conditions: before, during
and after strengthening. Similarly to the rst study, the idea is to use SDA-classication
algorithms to classify the different structural conditions. The objective of these two studies is to
appraise the probabilities of true detection (or true classication) and of misclassication when
detecting structural pattern changes.

2. SDA OVERVIEW
In this section, the methodology used to convert classical data to symbolic data is explained. In
general, data acquisition campaigns in civil engineering structures gather thousands of
accelerations values measured by several sensors. As a consequence, analyzing all of these
data (classical data) directly may usually be time consuming or even prohibitive. In this sense,
transforming this massive quantity of data into a compact but also rich descriptive type of data
(symbolic data) becomes an attractive approach. Let us consider, for instance, a signal X (which
is part of a dynamic test) containing 5000 acceleration values measured by one single sensor (see
Figure 1 on the left). There are several ways to transform classical data into symbolic data. This
signal can be represented by:
 a k-category histogram: X 5 f1(0.0025), 2(0.0721), 3(0.8546), 4(0.0626), y, k(0.0082)g;
 a min/max interval: X 5 [0.025; 0.025];
 an interquartile interval: X 5 [0.012; 0.015].

In this paper, the rst type of symbolic data is used in the SDA process. Figure 1 (on the
right) shows how a classical signal (one sensor) is converted to a symbolic representation of a
20-category histogram. In this case, all acceleration values are projected to the y-axis of
coordinates and a k-category histogram is constructed. In fact, this transformation can also be
applied to natural frequencies and mode shapes. When it comes to structural modal
identication, it is a common procedure to divide the acquired signals into several segments
and extract the modal parameters from these divisions. In so doing, several pairs of natural
frequencies and respective mode shapes are evaluated, which is called realizations. In this sense,
in order to convert the modal parameters into symbolic representations, the ensemble of
realizations is gathered and histogrammed into k-categories.

Copyright r 2010 John Wiley & Sons, Ltd. Struct. Control Health Monit. 2012; 19:161 186
DOI: 10.1002/stc
164 A. CURY AND C. CREMONA

0.1

0.08

0.06

0.04
Acceleration (m.s2)

0.02

-0.02

-0.04

-0.06

-0.08

-0.1
0 500 1000 1500 2000 2500 3000 3500 4000 4500 5000 0 100 200 300 400 500 600 700 800
Time Index Number of occurrences

Figure 1. Example of transforming classical signal to symbolic signal (20-category histogram).

Table I. Symbolic data table of dynamic tests represented by sensors, frequencies or mode shapes.
Representation Representation Representation
by signals by frequencies by mode shapes

Test Sensor 1 Sensor p Frequency 1 Frequency f Mode 1 Mode f


1 [1 (0.03), [1 (0.01),y, [1 (0.08),y, [1 (0.02),y, [1(0),y, [1 (0.02),y,
2 (0.05),y, 3 (0.07),y, 5 (0.35),y, y,7 (0.06), 4 (0.22),y, 8 (0.07),y,
.. 20 (0.04)]
.. 20 (0.03)]
.. 20 (0.02)]
.. y, 20(0)]
.. 20 (0.09)]
.. 20 .(0)]
. . . . . . ..
N [1 (0.012,y, [1 (0.05),y, [1 (0.002,y, [1 (0.04),y, [1 (0.012,y, [1 (0.05),y,
5 (0.02),y 10 (0.45),y 3 (0.41),y 8 (0.05),y, 7 (0.02),y 10 (0.05),y,
20 (0.009)] 20 (0.009)] 20 (0.001)] 20 (0.003)] 20 (0.002)] 20 (0.004)]

Table I provides a hypothetical example containing n dynamic tests. Columns 2 and 3 contain
a symbolic representation of signals, where each sensor (1yp) is described by a 20-category
histogram. In this table, values in parenthesis represent the ratio between the number of
occurrences within each category and the total number of occurrences. Similarly, columns 4 and
5 represent the dynamic tests, but now in terms of natural frequencies (1yf). For example, if the
rst frequency varies from 4.2 to 4.8 Hz, a histogram is created containing 20 categories (i.e.
4.24.23 Hz; 4.234.26 Hzy4.774.8 Hz). Finally, the same description can be used for mode
shapes (1yf) in which proportions are evaluated component wise as summarized in columns 6
and 7. An important remark is that the process of transforming classical data to symbolic data is
carried out almost instantly which does not prohibit or make difcult the use of this
methodology for a large ensemble of dynamic tests.

3. CLASSIFICATION METHODS
In this section, an overview of three classication methods based on the supervised learning is
presented. First, some concepts regarding BDT and its applicability are explained. Then, a brief
discussion about how Neural Networks are used in this study is exposed. Finally, a general idea
of SVM applied to classication problems is presented.

3.1. BDT
BDT form a decision procedure which is able to solve classication problems. The general idea
of this method is to classify a particular object into one of the classes (groups) previously dened
i.e. in the training set. For instance, let O fT1 ; T2 ; . . . ; Tn g be a set of n dynamic tests and C a
class variable containing values varying within f1; . . . ; mg where m is the number of classes. This

Copyright r 2010 John Wiley & Sons, Ltd. Struct. Control Health Monit. 2012; 19:161 186
DOI: 10.1002/stc
PATTERN RECOGNITION OF STRUCTURAL BEHAVIORS 165

Table II. Symbolic data table.


Test Sensor 1 y Sensor p Class
T1 [1 (0.03), 2 (0.05),y, k (0.04)] y [1 (0.01),y, 3 (0.07),y, k (0.03)] 1
T2 [1 (0.012,y,5 (0.02),y, k (0.009)] y [1 (0.05),y,10 (0.45),y, k (0.009)] 2
T~ [1 (0.032,y,10 (0.25),y, k (0.05)] y [1 (0.04,y,8 (0.39),y, k (0.002)] ?
Classication of a new test using BDTs.

Figure 2. Example of a Bayesian decision tree [15].

discriminant analysis tries to predict the unknown value C for a given test T~ according to its
p featuresz and a training set. The steps of this symbolic discrimination procedure are to
represent a given partition in the form of a BDT and to create a rule that is able to assign a new
test to one class of a prior partition.
Now, let us consider that each test Ti is described by two types of variables, taking into
account symbolic signalsy:
 p sensors described k-category histograms;
 a class variable C: this variable species the class of a test in the training set in the form of
a unique value (1, 2,y, m).

Table II provides an example containing three dynamic tests: T1 (belonging to class 1) and T2
(belonging to class 2) within the training set. The goal is to classify test T~ into one of these two
classes. In this example, all tests are described by two sensors (s1, s2).
In the framework of the recursive algorithm, each node division step is performed according
to a single variable (suitably chosen) and to yes/no answers to specic binary questions. The set
of binary questions that will be useful for the recursive partition procedure is based on the
Bayesian rule [15] (Figure 2). This rule can be enounced as follows. Suppose that for a set of
tests O, the test T~ is distributed with density functions:
fj T~ ; j 1; . . . ; m 1
Density functions are generally unknown and need to be estimated. One way to solve this
issue is to use the Kernel method [17] which can reasonably approximate density functions. The
Kernel estimator is dened as:
nj X  r;q 
1 X
p X k
1 T~  Tsr;q
fj T~ K ; j 1; . . . ; m 2
nj s1 r1 q1 h h

z
If tests are represented by signals, features are symbolic representations of sensors; if tests are represented by natural
frequencies or mode shapes, features are symbolic representations of each frequency or each mode shape, respectively.
y
For the sake of clarity and conciseness, all further explanations will be made taking into account manipulations with
signals. However, the same procedure can be done using natural frequencies or mode shapes.

Copyright r 2010 John Wiley & Sons, Ltd. Struct. Control Health Monit. 2012; 19:161 186
DOI: 10.1002/stc
166 A. CURY AND C. CREMONA

where nj is the number of tests within the class j, h40 is a smoothing parameter and K is called
kernel which is symmetric, continuous and can be evaluated by:
 r;q 
T~  Tsr;q
r;q
T~ r;q Ts 2
1
K p e 2h2 ; s 1; . . . ; nj ; r 1; . . . ; p; q 1; . . . ; k 3
h 2p
where Tsr;q is the qth category of the rth sensor corresponding to the sth test of a given class j. or
example, in Table II, if q 5 1, r 5 1, s 5 1 and j 5 1 (meaning the rst category of sensor 1
corresponding to the rst test of class 1, respectively), the value of T11;1 is 0.03. Similarly, if
q 5 10, r 5 p, s 5 1 and j 5 2 (meaning the tenth category of sensor p corresponding to the rst
test of class 2), the value of T1p;10 is 0.45.
By denition, the Bayesian rule assigns T~ to the class with the largest value of Pi  fj T~ ,
where Pj j 1; . . . ; m represent the prior probabilities of each class. For the previous example
introducing two classes, the following question is considered:
Is P1  f1 T~ 4P2  f2 T~ ? 4
If the rst term is greater than the second one, then the test T~ is considered to belong
to class 1; otherwise the test is assigned to class 2. Although this example contains only
two groups, this procedure can easily be extended to problems which have several classes.
In fact, if there are m possible classes, there have to be m1 questions to assign a new test to a
given class.
In addition, if prior probabilities are unknown, two choices are available to determine them:

 uniform prior probabilities based on the number of classes:


1
pj ; j 1; . . . ; m 5
m

 prior probabilities based on the proportions observed in the training set:


nj X m
pj with n nj 6
n j1

where n is the number of tests in the training set.

As already mentioned, tests are classied into different nodes according to cut rules and
splitting criteria. To dene each split, the most discriminant feature (sensor, frequency or mode
shape) must be used, that is, the feature capable to optimally separate tests into different
classes. This optimal feature is called cut variable and it leads to the purest nodes of the
Bayesian tree. The purity of a node is actually the number of tests that are correctly classied in
it (by leave-one-out and/or bootstrap methods), where the classication is veried in relation to
the class variable C. This number can be obtained by the Bayesian rule computed for each
feature. The choice of the most discriminant cut variable will result in selecting the feature which
minimizes the impurity measure i.e. the number of misclassied tests. Finally, the cut variable
and its associated cut value will form the cut rule which will be used to properly classify
a new test [18].

3.2. Neural Networks


Neural Networks have proved themselves as procient classiers and are particularly well suited
for addressing non-linear problems. Given the non-linear nature of real-world phenomena, like
the classication of dynamic tests, Neural Networks are a potential approach for dealing with
this problem.
Commonly, Neural Networks are adjusted or trained, so that a particular input leads to a
specic target output (supervised learning). In this case, the network is adjusted based on a
comparison of the output and the target, until the network output matches the target. Such
situation is shown in Figure 3.

Copyright r 2010 John Wiley & Sons, Ltd. Struct. Control Health Monit. 2012; 19:161 186
DOI: 10.1002/stc
PATTERN RECOGNITION OF STRUCTURAL BEHAVIORS 167

Figure 3. Supervised learning scheme.

Figure 4. One-hidden-layer MLP (adapted from [20]).

Supervised learning is a machine-learning technique for deducing mapping functions from a


training data set consisted of inputoutput pairs. The goal is to predict output values of the
mapping function for any valid input after having seen a number of training examples (i.e. pairs
of inputs and target outputs). To achieve this, the network has to generalize from the training
data to unseen situations in a reasonable way.
Mapping functions are obtained through an optimization scheme based on the evaluation of
the mean-squared error (explained further). This scheme tries to minimize the average squared
error between the networks output and the target value over all the training data set pairs. In
this paper, a feed-forward multi-layer perceptron (MLP) Neural Network is used for classifying
dynamic tests. Multi-layer networks use a variety of learning techniques, the most popular being
back-propagation. In this case, the output values are compared with the correct answer to
compute the value of some predened error-function and the error is then fed back through the
network [19]. Using this information, the algorithm adjusts the weights of each connection to
reduce the value of the error function by some small amount. After repeating this process for a
sufciently large number of training cycles, the network will usually converge to some state
where the error of the calculations is negligible. At this point, it is said that the network is
trained. Figure 4 shows a one-hidden-layer MLP with n inputs, N hidden neurons and n
outputs which is used in the applications of this paper. The inputs are the dynamic tests Ti
(i 5 1,y,n) which are symbolic representations of signals, frequencies or mode shapes as
described in Section 2. They are dened by p  k matrices, where p is the number of sensors and
k is the number of categories. Similarly to the BDT method, outputs are set to represent the
target vectors (labels) corresponding to each class i.e. f1; 2; . . . ; mg. These outputs are
transformed into a binary notation according to the number of classes used.
Each input Ti is multiplied by adjustable weights denoted wil before being fed to the lth
neuron in the hidden/output layers, yielding:
!
XN
yj f wil Ti 1bi ; j 1; . . . ; n 7
i1

where bi represents the bias of each perceptron.


In order to adjust weights properly, a general method for non-linear optimization called
gradient descent is applied [20]. Briey, the derivative of the error function with respect to the

Copyright r 2010 John Wiley & Sons, Ltd. Struct. Control Health Monit. 2012; 19:161 186
DOI: 10.1002/stc
168 A. CURY AND C. CREMONA

network weights is calculated, and the weights are then changed such that the error decreases
(thus going downhill on the surface of the error function). Equation 8 shows the expression to
evaluate the updated weights of this network:
@J
wil t11 wil t  Z 8
@wil

where Z is the learning rate, t is the iteration step and J is mean error for a perceptron,
written as:

1 Xn
J dj  yj 2 9
2n j1

where dj represent the desired outputs (targets) and yj the observed outputs (evaluated by the
Neural Network). For the simulations presented in this paper a free Matlab toolboxr named
Netlab and developed by Bishop et al. [19] was used. This toolbox allows performing several
types of supervised multi-class classications.

3.3. SVM
SVM are a useful technique for data classication problems. As usual, the objective is to separate
two different classes by a function which is induced from available examples (training data set).
SVMs were rst suggested in the 1960s for classication and have recently become an area of
intense research due to developments in the techniques and theory coupled with extensions to
regression and also density estimation [21]. This technique came up from statistical learning
theory and is based on the structural risk minimization principle. Commonly,
SVMs are used for two-class classication problems (Figure 5). However, this can easily be
extended from two-class classications to m-class classication problems by constructing m
two-class classiers [23]. The geometrical interpretation of support vector classication
(SVC) is that the algorithm searches for an optimal separating surface, i.e. a hyperplane
equidistant from the two classes [24]. First, SVC is briey outlined for the linearly separable case.
Consequently, a notion of kernel functions is introduced to show how non-linear decision
surfaces can be constructed. Similarly to the Neural Networks classication, the inputs Ti
(i 5 1,y,n) are dened by symbolic representations of dynamic tests (p  k matrices, where p is
the number of sensors, frequencies or mode shapes and k is the number of categories) and the
outputs yi (i 5 1,y,n) represent the target binary vectors (labels) corresponding to each class
i.e. f1; 2; . . . ; mg.

Figure 5. Scheme representing support vectors and separation margins for the linear case (adapted from [22]).

Copyright r 2010 John Wiley & Sons, Ltd. Struct. Control Health Monit. 2012; 19:161 186
DOI: 10.1002/stc
PATTERN RECOGNITION OF STRUCTURAL BEHAVIORS 169

3.3.1. Linear case. Let us consider that the training data are linearly separable. Then, there is a
pair (w,b) such that:
wT Ti 1bX1 8Ti 2 O1
10
wT Ti 1bp  1 8Ti 2 O2
where O1 [ O2 O, w the weight vector, b he bias and T is the vector containing the ensemble
of tests. The classication of tests is made according to the following criteria:
Ifyi 1 ! wT Ti 1b411 8Ti 2 O1
T
11
Ifyi 1 ! w Ti 1bo  1 8Ti 2 O2
This equation can be combined to give:
yi wT Ti 1bX1 8Ti 2 O 12
The SVM searches for the simplest solution that classies the data correctly. The learning
problem can be formulated as: minimize jjwjj2 wT w subject to the constraints of linear
separability (shown in Equation 14). This is equivalent to maximize the distance, normal to the
hyperplane, between the two classes; this distance is called margin [21]. The optimization
problem is enounced as [25]:
1
Minimize Fw jjwjj2
w;b 2 13
 T 
subject to yi w Ti 1b X1; i 1; . . . ; n
The Lagrangian for this problem can be written as follows:
1 Xn
Lw; b; L jjwjj2  li yi wT Ti 1b  1 14
2 i1

where L l1 ; . . . ; ln T are the Lagrange multipliers. The solution to this problem is given by
maximizing L with respect to LiX0 and minimizing with respect to w,b. The optimal solution is
given by:
Xn
w li yi Ti 15
i1

As a matter of fact, the Lagrange multipliers are only non-zero when yi wT Ti 1b 1. In this
case, these vectors are called support vectors since they lie closest to the separating hyperplane.
The optimal of the bias is evaluated as:
b yi  wT Ti ; i 1; . . . ; n 16
Figure 5 shows a schematic bi-dimensional representation of tests as well as the margins
dening the classication criteria.

3.3.2. Non-linear case. A linear classier may not be the most suitable hypothesis for the two
classes. The SVM can be used to learn non-linear decision functions by rst mapping the data to
some higher dimensional feature space and constructing a separating hyperplane in this space
(Figure 6). Hence, the concept of kernel functions must be introduced [21]. Denoting the
mapping to feature space by:
X !H
17
T 7!fT
the mapping of a given set of tests Ti,Tj is made as follows:
KTi ; Tj  fTi T fTj 18
These functions allow constructing an optimal separating hyperplane in the feature space
without explicitly performing calculations in this space. Now, it is necessary to choose an
appropriate kernel function easy to compute and capable of adjusting a reasonable discriminant

Copyright r 2010 John Wiley & Sons, Ltd. Struct. Control Health Monit. 2012; 19:161 186
DOI: 10.1002/stc
170 A. CURY AND C. CREMONA

Figure 6. Scheme representing the separation hyperplane for the non-linear case (adapted from [26]).

hyperplane. In this paper, the Gaussian radial basis function is used:


KTi ; Tj expgjjTi  Tj jj2 19
For the simulations presented in this paper regarding SVMs, a free Matlabr toolbox
developed by Chang and Lin [27] was used. This toolbox allows performing two-class as well as
multi-class classications.

4. NUMERICAL SIMULATIONS
Numerical simulations presented in this paper correspond to dynamic tests performed on a
simply supported steel beam. This beam is discretized by means of the FEM into 200 elements
(two node-linear elements with three degrees of freedom per node) and a random excitation is
located at node 178 (at 0.69 m from the right support). Dynamic responses are measured at 11
equidistant points of the beam during 50 s with a sampling time of 0.001 s. Figure 7 shows a
schematic view of the beam, the position of excitation as well as the measurement points used.
The material properties of the beam can be summarized as:
 Youngs modulus (E): 210 GPa.
 Linear mass density (r): 7850 kg m3.
 Cross-sectional area (S): 2.81  103 m2.
 Moment of area (I): 1.0845  108 m4.

As previously mentioned, one of the objectives of this paper is to simulate four damage levels
and to use the clustering methods from Section 3 to discriminate each structural condition, i.e.
from the undamaged state (D0) to the most severe damaged state (D3). The rst damage
scenario (D1) consists of a 5% reduction of the Youngs Modulus in elements 100, 101 and 102
(mid-span of the beam). The second damage scenario (D2) includes an additional 5% reduction
of the Youngs Modulus in elements 1, 2 and 3 (left support of the beam). Finally, the third
damage scenario (D3) consists of an additional 5% reduction of the Youngs Modulus in
elements 150, 151 and 152 (between the mid-span and the right support of the beam) and a 10%
reduction on the elements at mid-span (Figure 8). In addition, three different noise levels (1, 2
and 5%) are applied to the signals obtained from the simulated dynamic data. For each noise,
the percentage of noise was multiplied to the standard deviation value of each measuring
channel and was randomly added to other component of the current channel. In order to have a
signicant amount of simulated tests, for each damage state and noise level 10 dynamic tests are
performed. A total of 160 tests are simulated, since there are 4 damage scenarios (D0, D1, D2
and D3), 4 noise levels (N0, N1, N2 and N3) and 10 dynamic tests for each couple state/noise
(Di,Nj). D0 and N0, respectively, correspond to the absence of damage and noise.
Initially, the hybrid SDA-classication algorithms are applied to a set of 40 tests for each
noise level. Such an analysis helps to assess the inuence of the different noise levels. The set of
80 D0 and D3 tests containing all noise levels is then used. The objective is to discriminate the
two extreme damage levels no matter the incidence of noise levels. At last, all the 160 tests are

Copyright r 2010 John Wiley & Sons, Ltd. Struct. Control Health Monit. 2012; 19:161 186
DOI: 10.1002/stc
PATTERN RECOGNITION OF STRUCTURAL BEHAVIORS 171

Figure 7. Schematic view of the test beam and excitation point.

Figure 8. Schematic view of the three damage scenarios applied to the beam.

considered as a whole. The idea is to observe the capability of each SDA-classication algorithm
for discriminating all the damage states with all the noise levels concomitantly.
It is important to stand out again that not only the signals but also the natural frequencies
and the mode shapes are transformed into symbolic data and applied to the clustering methods.
Modal parameters are extracted from the response measurements by applying the vectorial
Random Decrement method combined to the Ibrahim Time Domain method [28]. For each
dynamic test (for a given noise level and damage scenario), 80 realizations are performed,
meaning 80 estimated natural frequencies and respective mode shapes per test. This strategy
allows representing both modal parameters into symbolic histograms. Randomness in the
modal properties is issued from inter-test and intra-test variations.

4.1. Classical analysis


Before detailing the results obtained from SDA-classication algorithms, it is important to
present a simple study of the identied natural frequencies and mode shapes. Figures 9 and 10
show a comparison between the rst ve mode shapes considering the undamaged state (D0)
and the damaged state (D3) without noise. It becomes that they are slightly affected by damage
in the absence of noise. When noise is included, these minor changes are mostly disappearing.
Table III presents the mean values and the standard deviations for each frequency, each damage
scenario and each noise level. It can be stated that the changes in the frequency are only
signicant for the largest ones. Figure 11 shows the evolution of the different frequencies
according to the damage level. The noise level has notably a large effect on the frequency
pattern. This effect can be emphasized by analyzing the condence intervals around the mean
frequencies (71 standard deviation). Figure 12 shows that the uncertainties due to the tests and
to the noises for the different damage states may be superposed; it is therefore very difcult to
discriminate if a change is issued from noise or from damage. This is particularly the case for D1
and D2 levels for which noise affects the identication in an opposite manner than the damage.

Copyright r 2010 John Wiley & Sons, Ltd. Struct. Control Health Monit. 2012; 19:161 186
DOI: 10.1002/stc
172 A. CURY AND C. CREMONA

1.5
D0

Amplitude
1 D3

0.5

0
0 20 40 60 80 100 120 140 160 180 200
Nodes

1
2
Amplitude

Amplitude
0 0

-2
-1
0 50 100 150 200 0 50 100 150 200
Nodes Nodes

2 1
Amplitude

Amplitude
0 0

-1
-2
0 50 100 150 200 0 50 100 150 200
Nodes Nodes
Figure 9. Comparison between mode shapes (undamaged and third damage state, no noise).

1.5
D0
Amplitude

1 D3

0.5

0
0 20 40 60 80 100 120 140 160 180 200
Nodes

1
2
Amplitude

Amplitude

0 0

-2
-1
0 50 100 150 200 0 50 100 150 200
Nodes Nodes

2 1
Amplitude

Amplitude

0 0

-1
-2
0 50 100 150 200 0 50 100 150 200
Nodes Nodes
Figure 10. Comparison between mode shapes (undamaged and third damage state, noise level 3).

Classical statistical tests based on condence intervals will not be sensitive enough to
discriminate the different damage states in the presence of noise. It is therefore essential to use
enhanced analyses in order to get a better discrimination [7,29].

Copyright r 2010 John Wiley & Sons, Ltd. Struct. Control Health Monit. 2012; 19:161 186
DOI: 10.1002/stc
PATTERN RECOGNITION OF STRUCTURAL BEHAVIORS 173

Table III. Probabilities of true classication using signals (30% training, 10% validation and 60% testing).
BDT Neural Networks SVM

Noise level N0 N1 N2 N3 N0 N1 N2 N3 N0 N1 N2 N3
Best (%) 83 79 75 75 100 100 96 92 100 100 100 100
Average (%) 61 58 59 44 78 76 74 78 73 71 78 74
Worst (%) 38 25 33 11 21 25 48 55 17 48 21 38

0.6
1st frequency
0.58

D0 D1 D2 D3

2.35
2st frequency
2.3
No Noise (N0)
2.25
D0 D1 D2 Noise level (N1) D3
Noise level (N2)
5.4
Noise level (N3)
Hz

5.2
3rd frequency
5
D0 D1 D2 D3

9.26
9.25
4th frequency
9.24
D0 D1 D2 D3

15
5th frequency
14.5
14
D0 D1 D2 D3
Damage level
Figure 11. Comparison between damage states and noise levels for the rst ve frequencies.

4.2. Hybrid SDA-classication algorithms


In this section, the hybrid SDA-classication algorithms are applied to classify dynamic tests
according to four different structural states (D0, D1, D2 and D3). First, simulations are
performed for analyzing the inuence of each noise level. Different sizes of the training,
validation and testing data sets are used. The idea is to perform a benchmarking study to
observe how the training and testing data sets inuence the classication. For all classication
methods, signals, natural frequencies and mode shapes are converted to 20-category histograms
and used as inputs. For the NN method, a one-hidden layer feed forward network is created
with 20 neurons in the hidden layer. Moreover, for SVMs simulations, a RBF kernel is used
with calibrated parameters (C and g) obtained though cross-validation. The architecture of the
NN and SVMs kernel are kept the same throughout the following simulations described in this
paper.
When using symbolic signals, the rst simulations use only 30% of the set of dynamic tests
for training, 10% for validationz and 60% for testing. This means that the 40 tests are

z
For the BDT method there, is no validation. In this case, the training data set included the percentages of the validation
data set.

Copyright r 2010 John Wiley & Sons, Ltd. Struct. Control Health Monit. 2012; 19:161 186
DOI: 10.1002/stc
174 A. CURY AND C. CREMONA

D3,N3 D3,N3
D2,N3 D2,N3
D1,N3 D1,N3
D0,N3 D0,N3
D3,N2 D3,N2
D2,N2 D2,N2
D1,N2 D1,N2
D0,N2 D0,N2
D3,N1 D3,N1
D2,N1 D2,N1
D1,N1 D1,N1
D0,N1 D0,N1
D3,N0 D3,N0
D2,N0 D2,N0
D1,N0 D1,N0
D0,N0 D0,N0

0.55 0.56 0.57 0.58 0.59 0.6 2.29 2.295 2.3 2.305 2.31 2.315 2.32 2.325 2.33
Frequency [Hz] Frequency [Hz]

First frequency Second frequency

D3,N3 D3,N3
D2,N3 D2,N3
D1,N3 D1,N3
D0,N3 D0,N3
D3,N2 D3,N2
D2,N2 D2,N2
D1,N2 D1,N2
D0,N2 D0,N2
D3,N1 D3,N1
D2,N1 D2,N1
D1,N1 D1,N1
D0,N1 D0,N1
D3,N0 D3,N0
D2,N0 D2,N0
D1,N0 D1,N0
D0,N0 D0,N0

5.12 5.14 5.16 5.18 5.2 5.22 5.24 9.23 9.235 9.24 9.245 9.25 9.255 9.26 9.265 9.27 9.275
Frequency [Hz] Frequency [Hz]

Third frequency Fourth frequency

D3,N3
D2,N3
D1,N3
D0,N3
D3,N2
D2,N2
D1,N2
D0,N2
D3,N1
D2,N1
D1,N1
D0,N1
D3,N0
D2,N0
D1,N0
D0,N0

14.2 14.25 14.3 14.35 14.4 14.45 14.5 14.55


Frequency [Hz]

Fifth frequency
Figure 12. Condence intervals (mean71 standard deviation) on the identied frequencies.

distributed over three groups: 12 tests for the training data set, 4 tests for the validation data set,
24 tests for the testing data set. Since this distribution can be randomly performed, 10 000
simulations (or groups of training, validation and testing data sets) are generated. Results are
summarized in Table III, showing the best and the worst classication ratios for each simulation
as well as the average value of the probability of true detection. Considering the rst column of
this table, for example, it is possible to see that the BDT had its best performance when it
classied 83% of tests correctly. In average, it classies 61% of tests. Its worst performance

Copyright r 2010 John Wiley & Sons, Ltd. Struct. Control Health Monit. 2012; 19:161 186
DOI: 10.1002/stc
PATTERN RECOGNITION OF STRUCTURAL BEHAVIORS 175

occurred when it classied only 38% of tests correctly. As expected, when the noise levels
increase, the probability of true detection decreases. In general, the best and the average ratios
decrease, whereas the worst score does not change signicantly. SDA-SVM is the only method
capable to provide a 100% probability of detection for all the simulations.
The training data set is now increased to 40% (per noise level) and the testing data set is
reduced to 50%. Now, the probabilities of true detection are improved signicantly (average and
worst ratiosTable IV). Finally, when 20 tests are used for training, very good results are
obtained for all the methods (Table V). It can be observed that the SDA-SVM is achieving better
results than the other two methods. This is might be due to the fact that SDA-SVM could better
adjust separation thresholds for each damage state. Also, The SDA-BDT is providing slightly
worse probabilities of detection compared with the SDA-NN. This could be explained by the fact
that the SDA-BDT is not true learning procedure. New tests are affected to groups according to
the logical questions. If the four groups (corresponding to the damage levels) do not provide a
good representation of each structural state, the classication may be compromised.
Symbolic natural frequencies are now used as inputs to the SDA-classication methods.
Results are presented in Tables VIVIII. Even when the training data set is small, the SDA-NN

Table IV. Probabilities of true classication using signals (40% training, 10% validation and 50% testing).
BDT Neural Networks SVM

Noise level N0 N1 N2 N3 N0 N1 N2 N3 N0 N1 N2 N3
Best (%) 80 80 80 75 100 100 95 95 100 100 100 100
Average (%) 66 55 56 48 78 76 78 74 85 81 80 84
Worst (%) 35 30 35 20 50 50 50 55 63 56 56 63

Table V. Probabilities of true classication using signals (50% training, 10% validation and 40% testing).
BDT Neural Networks SVM

Noise level N0 N1 N2 N3 N0 N1 N2 N3 N0 N1 N2 N3
Best (%) 83 83 83 75 100 100 100 100 100 100 100 100
Average (%) 68 58 58 50 80 77 73 80 85 83 81 77
Worst (%) 38 42 33 19 50 44 50 56 66 58 50 58

Table VI. Probabilities of true classication using frequencies (30% training, 10% validation and 60% testing).
BDT Neural Networks SVM

Noise level N0 N1 N2 N3 N0 N1 N2 N3 N0 N1 N2 N3
Best (%) 100 100 100 100 100 100 100 100 100 100 100 100
Average (%) 100 100 100 99 97 97 90 90 96 97 98 92
Worst (%) 100 100 100 96 58 50 58 58 55 55 55 55

Table VII. Probabilities of true classication using frequencies (40% training, 10% validation and 50%
testing).
BDT Neural Networks SVM

Noise level N0 N1 N2 N3 N0 N1 N2 N3 N0 N1 N2 N3
Best (%) 100 100 100 100 100 100 100 100 100 100 100 100
Average (%) 100 100 100 100 99 95 95 94 99 99 98 86
Worst (%) 100 100 100 100 65 65 65 50 55 55 55 40

Copyright r 2010 John Wiley & Sons, Ltd. Struct. Control Health Monit. 2012; 19:161 186
DOI: 10.1002/stc
176 A. CURY AND C. CREMONA

Table VIII. Probabilities of true classication using frequencies (50% training, 10% validation and 40% testing).
BDT Neural Networks SVM

Noise level N0 N1 N2 N3 N0 N1 N2 N3 N0 N1 N2 N3
Best (%) 100 100 100 100 100 100 100 100 100 100 100 100
Average (%) 100 100 100 100 99 99 95 94 99 99 97 95
Worst (%) 100 100 100 100 88 81 81 58 82 81 82 80

Table IX. Probabilities of true classication using mode shapes (30% training, 10% validation and 60%
testing).
BDT Neural Networks SVM

Noise level N0 N1 N2 N3 N0 N1 N2 N3 N0 N1 N2 N3
Best (%) 80 68 58 58 100 96 92 92 92 92 92 92
Average (%) 57 33 33 35 73 71 55 57 75 72 67 62
Worst (%) 33 13 8 8 33 30 21 21 42 33 21 21

Table X. Probabilities of true classication using mode shapes (40% training, 10% validation and 50%
testing).
BDT Neural Networks SVM

Noise level N0 N1 N2 N3 N0 N1 N2 N3 N0 N1 N2 N3
Best (%) 75 70 60 60 100 100 95 90 100 95 95 90
Average (%) 56 33 35 33 82 77 57 61 64 75 71 59
Worst (%) 30 10 15 5 40 35 40 25 35 35 35 25

and the SDA-SVM are able to classify the four damage levels correctly, no matter the incidence
of noise. Although the best probabilities of detection do not change signicantly when
the training data set increases, average and worst values are nevertheless improved considerably.
Similar analyses are performed using the mode shapes converted to symbolic data
(Tables IXXI). In general, classications based on mode shapes provide better results
compared with those obtained with signals, but poorer classications compared with natural
frequencies. This can be explained by the fact that mode shapes may be more sensitive to noise
than frequencies.
In this second part, the 80 D0 and D3 tests with all the noise levels are used. The objective is
to discriminate the two damage levels no matter the incidence of noise. Results obtained when
signals are used are summarized in Table XII. When the training data set comprises 30% of
tests, only the SDA-SVM is able to achieve 100% of correct classication. The SDA-NN does
not reproduce similar results than those obtained for the separate analysis of noise levels. The
SDA-BDT method is not able to get probabilities of detection higher than 90%. This shows that
although damage levels are very different, the presence of noise can mislead the classication.
A total of 10 000 simulations have been used for this study. When natural frequencies are used,
results are improved (Table XIII). The SDA-NN and the SDA-SVM produce 100% of correct
classication based on 10 000 simulations. For the SDA-BDT, very reasonable results are also
achieved. For mode shapes (Table XIV), the SDA-NN and the SDA-SVM highlight a better
performance compared with the SDA-BDT. Results are slightly worse than those obtained with
frequencies.
Finally, all the 160 tests are used. The rst simulations are performed using symbolic signals
and only 30% of the data set for training. It can be observed that results obtained are not good
enough. In this case, only the SDA-SVM is able to discriminate each damage level properly
(Table XV). It seems that this method is more robust than the others to classify structural

Copyright r 2010 John Wiley & Sons, Ltd. Struct. Control Health Monit. 2012; 19:161 186
DOI: 10.1002/stc
PATTERN RECOGNITION OF STRUCTURAL BEHAVIORS 177

Table XI. Probabilities of true classication using mode shapes (50% training, 10% validation and 40% testing).
BDT Neural Networks SVM

Noise level N0 N1 N2 N3 N0 N1 N2 N3 N0 N1 N2 N3
Best (%) 94 91 81 63 100 100 94 94 94 94 94 93
Average (%) 57 35 33 33 87 81 58 63 65 75 69 56
Worst (%) 25 20 10 8 31 44 13 19 19 38 31 13

Table XII. Probabilities of true classication using signals for damage levels D0 and D3.
30% Tr., 10% V, 60% T 40% Tr., 10% V, 50% T 50% Tr., 10% V, 40% T

Method BDT NN SVM BDT NN SVM BDT NN SVM


Best (%) 71 71 100 73 75 100 75 78 100
Average (%) 54 52 76 55 52 78 56 55 80
Worst (%) 33 28 40 35 30 37 28 31 42

Table XIII. Probabilities of true classication using frequencies for damage levels D0 and D3.
30% Tr., 10% V, 60% T 40% Tr., 10% V, 50% T 50% Tr., 10% V, 40% T

Method BDT NN SVM BDT NN SVM BDT NN SVM


Best (%) 100 100 100 100 100 100 100 100 100
Average (%) 99 100 100 100 100 100 100 100 100
Worst (%) 98 100 100 100 100 100 100 100 100

Table XIV. Probabilities of true classication using mode shapes for damage levels D0 and D3.
30% Tr., 10% V, 60% T 40% Tr., 10% V, 50% T 50% Tr., 10% V, 40% T

Method BDT NN SVM BDT NN SVM BDT NN SVM


Best (%) 92 100 100 93 100 100 97 100 100
Average (%) 75 96 97 75 96 99 76 97 100
Worst (%) 52 53 95 50 55 97 50 73 100

Table XV. Probabilities of true classication using signals for all damage and noise levels.
30% Tr., 10% V, 60% T 40% Tr., 10% V, 50% T 50% Tr., 10% V, 40% T

Method BDT NN SVM BDT NN SVM BDT NN SVM


Best (%) 80 60 100 82 65 100 84 65 100
Average (%) 75 25 65 76 33 68 76 35 71
Worst (%) 12 13 25 25 15 20 15 16 22

conditions even when noise is added to the data. For modal parameters, results are largely
improved. Tables XVI and XVII present the probabilities of true detection when natural
frequencies and mode shapes are used. The SDA-NN and the SDA-SMV systematically achieve
100% for the best ratios. The SDA-NN results are better for natural frequencies, whereas the
SDA-SVM gives better scores for mode shapes. The SDA-BDT provides uniform results for
each damage state.

Copyright r 2010 John Wiley & Sons, Ltd. Struct. Control Health Monit. 2012; 19:161 186
DOI: 10.1002/stc
178 A. CURY AND C. CREMONA

Table XVI. Probabilities of true classication using frequencies for all damage and noise levels.
30% Tr., 10% V, 60% T 40% Tr., 10% V, 50% T 50% Tr., 10% V, 40% T

Method BDT NN SVM BDT NN SVM BDT NN SVM


Best (%) 98 100 98 98 100 98 98 100 100
Average (%) 89 98 87 91 99 87 91 99 88
Worst (%) 71 88 71 75 85 69 78 94 73

Table XVII. Probabilities of true classication using mode shapes for all damage and noise levels.
30% Tr., 10% V, 60% T 40% Tr., 10% V, 50% T 50% Tr., 10% V, 40% T

Method BDT NN SVM BDT NN SVM BDT NN SVM


Best (%) 80 92 100 80 94 100 82 95 100
Average (%) 76 78 97 77 81 97 79 83 97
Worst (%) 35 22 82 47 39 84 49 25 90

5. EXPERIMENTAL APPLICATION
The studied bridge is an embedded steel bridge located on the South-East high speed track in
France at the kilometric point 0751317; it crosses the secondary road 939 between the towns of
Sens and Soucy in the Yonne County. A dynamic monitoring of this railway bridge was
performed to characterize and quantify the effect of a strengthening procedure (Figures 13 and
14). This strengthening procedure consists of tightening special bearings to shift the rst natural
frequency (around 5 Hz) from the excitation frequency due to train crossings (around 45 Hz).
The risk of resonance is thus increased, especially when ballast recharging occurs. Three sets of
dynamic tests were performed: before strengthening (15 tests represented by the letter A), during
(13 tests represented by the letter R) and after (13 tests represented by the letter B). The 13 tests
performed during strengthening were divided into four phases, as summarized in Table XVIII.
The transducers are stuck and mechanically maintained by various mounting systems which
allow the adjustment of the sensor orientation and dismounting. The instrumentation comprises
eight vertical accelerometers and two horizontal accelerometers (longitudinal and transversal)
under the bridge deck and two temperature gauges. The sampling frequency was xed at
4096 Hz. The signal analysis due to train crossings highlights a good repeatability [30].

5.1. Classical analysis


Modal parameters are extracted from the response measurements by applying the vectorial
Random Decrement method combined to the Ibrahim Time Domain method [28]. Table XIX
summarizes the variation of the identied natural frequencies for each strengthening stages. It
should be noted that the numerical variation between these stages is quite small, thus
classication methods need to be sensitive enough to detect those differences. Figures 1518
show the histograms obtained for the rst four natural frequencies comparing both before and
after states of the bridge. Dashed vertical lines represent the mean values and the limits of a
95% condence interval of the identied frequencies. Although the difference between these two
states is reasonably clear for the rst frequency, this cannot be stated for the others. By
analyzing the histograms, it is really difcult to distinguish these two states, since the values are
superposed. It is therefore essential to use subsequent analyses in order to get a better
discrimination. Figure 19 gives a comparison between the rst four modes before and after
strengthening. For this gure, the measurements on the eight measurement points located on the
two bridge sides (upstream and downstream) are extrapolated to 129 points spread on three
lines. The extrapolation has been performed by piecewise cubic interpolations. It comes that
both modes 2 and 3 are the most sensitive modes to structural modication according to the
MAC index. It is interesting to note that the strengthening (although made on both sides of the

Copyright r 2010 John Wiley & Sons, Ltd. Struct. Control Health Monit. 2012; 19:161 186
DOI: 10.1002/stc
PATTERN RECOGNITION OF STRUCTURAL BEHAVIORS 179

Figure 13. View of the bridge during the strengthening process.

Figure 14. Strengthening system and procedure.

Table XVIII. Description of tests.


State Strengthening phase Tests
Before TGV1A, TGV2A,y, TGV15A
During 1 TGV1R, TGV2R, TGV3R
During 2 TGV4R, TGV5R, TGV6R
During 3 TGV7R, TGV8R, TGV9R
During 4 TGV10R,y, TGV13R
After TGV1B, TGV2B,y, TGV13B

Table XIX. First 4 frequencies for each strengthening stages.


Frequency Before (Hz) During (Hz) After (Hz)
1 5.84 6.27 6.44
2 8.61 8.74 8.78
3 13.09 13.2 13.31
4 16.95 17.21 17.34

bridge) induces a loss of symmetry on the exural and torsional modes. The strengthening
decreases the modal amplitudes of the rst torsional mode and does not affect considerably the
bending modes. The second mode which is a coupled exural/lateral mode is largely modied
especially on one side of the bridge.

Copyright r 2010 John Wiley & Sons, Ltd. Struct. Control Health Monit. 2012; 19:161 186
DOI: 10.1002/stc
180 A. CURY AND C. CREMONA

Figure 15. Comparison between before and after states for the rst frequency.

Figure 16. Comparison between before and after states for the second frequency.

Figure 17. Comparison between before and after states for the third frequency.

Copyright r 2010 John Wiley & Sons, Ltd. Struct. Control Health Monit. 2012; 19:161 186
DOI: 10.1002/stc
PATTERN RECOGNITION OF STRUCTURAL BEHAVIORS 181

Figure 18. Comparison between before and after states for the fourth frequency.

5.2. Symbolic analysis


In this section, experimental vibration data (signals) and experimental modal parameters
(natural frequencies and mode shapes) are transformed into symbolic quantities and used as
inputs to the classication methods. Briey, the same procedure outline in Section 4.2 to convert
classical data to symbolic data is adopted. Also, the same NN architecture is used as well as the
calibration of parameters for the SVM method. To attest and verify the robustness of the
proposed methodology, four different studies are carried out.
Four analyses are performed in order to assess the efciency and the reliability of the SDA-
classication algorithms for detecting structural changes. For the rst study, the training data
set contains 28 dynamic tests (15 corresponding to the state before and 13 corresponding to the
state after). The state before represents class 1 and the state after, class 2. The testing data set
contains the 13 dynamic tests corresponding to the state during. The idea is to classify those
tests according to one of the two previous classes. Since the strengthening was a gradual process,
it is interesting to see how the techniques affect the different tests during to the states before
and after. For the second study, the training data set is reduced to 10 dynamic tests (ve
randomly chosen tests corresponding to the state before and ve randomly chosen tests
corresponding to the state after). Similarly, the testing data set comprises ve random tests
from the state during. The third study is intended to validate the classication methods. In this
case, the training data set contains 10 tests (the same used in the second simulation), but now the
testing data set is composed of ve different random tests chosen among both before and
after states. The idea is to observe if the classication methods are able to conrm if a before
test is indeed affected to class 1 as well as an after test to class 2. At last, the fourth study is
carried out considering all of the 41 dynamic tests and the three structural states (before,
during and after). Now, the objective is to simulate a benchmarking procedure to see how
these methods will respond to different training and testing groups.

5.2.1. Results/study 1. Table XX summarizes the results given by the SDA-BDT method applied
to signals, natural frequencies and mode shapes. Tests TGV1A to TGV15A correspond to class 1
and tests TGV1B to TGV13B are related to class 2. The decision tree is created and the during
tests will be assigned to one of those classes according to the Bayesian rule described in
Section 3.1. Results are quite reasonable: 70% of tests are correctly classied when using the signals
as well as the frequencies and the mode shapes. One interesting observation lies on the fact that
the test TGV6R is systematically misclassied. This may be due to the fact that this test is close to
the transition between phases 2 and 3. Table XXI summarizes the classications obtained when the
SDA-NN is used. Again, reasonable results are achieved, although the classication obtained with
the signals is worse (62%). However, results with the modal parameters are improved: 92% for the

Copyright r 2010 John Wiley & Sons, Ltd. Struct. Control Health Monit. 2012; 19:161 186
DOI: 10.1002/stc
182 A. CURY AND C. CREMONA

MAC(% ) : 99.82 99.9 99.93

0
0 43 86 129

0 MAC(% ) : 63.64 75.03 82.05

-1
0 43 86 129

0 MAC(% ) : 98.23 38.18 99.95

-1
0 43 86 129

0 MAC(% ) : 99.34 99.61 99.84

-1
0 43 86 129
Line 1 Line 2 Line 3
Figure 19. Comparison between interpolated mode shapes (before and after strengthening).

frequencies and 85% for mode shapes. Once more, test TGV6R is misclassied. Finally,
Table XXII summarizes the results obtained when the SDA-SVM is employed. The classication
based on the signals is better than those obtained with the other methods (77%). However, the
probabilities of true classication for the modal parameters are decreased in contrast to the Neural
Networks (77% for the frequencies and 70% for the mode shapes). This might be explained by the
fact that the SDA-SVM is not sensitive enough to discriminate these conditions.

5.2.2. Results/study 2. As described in Section 5.2, the training data set contains ve dynamic
tests randomly chosen from the state before (TGV1A, TGV2A, TGV5A and TGV13A) and
ve from the state after (TGV4B, TGV6B, TGV7B, TGV9B and TGV11B). The testing data
set now comprises only ve tests pertaining to the state during. Table XXIII summarizes the
results obtained by the SDA-BDT method applied to the signals, natural frequencies and mode
shapes. As it can be observed, only test TGV5R is misclassied (for signals and mode shapes).
For the frequencies, all tests are correctly classied. Table XXIV summarizes the classications

Copyright r 2010 John Wiley & Sons, Ltd. Struct. Control Health Monit. 2012; 19:161 186
DOI: 10.1002/stc
PATTERN RECOGNITION OF STRUCTURAL BEHAVIORS 183

Table XX. Classication obtained by the BDT method.


Detected class

Test True class Signals Frequencies Mode shapes


TGV1R 1 1 1 1
TGV2R 1 1 1 1
TGV3R 1 1 2 2
TGV4R 1 1 1 2
TGV5R 1 1 2 1
TGV6R 1 2 2 2
TGV7R 2 1 2 2
TGV8R 2 1 2 2
TGV9R 2 1 1 1
TGV10R 2 2 2 2
TGV11R 2 2 2 2
TGV12R 2 2 2 2
TGV13R 2 2 2 2

Table XXI. Classication obtained by the NN method.


Detected class

Test True class Signals Frequencies Mode shapes


TGV1R 1 1 1 1
TGV2R 1 1 1 1
TGV3R 1 2 1 1
TGV4R 1 2 1 1
TGV5R 1 2 1 1
TGV6R 1 2 2 2
TGV7R 2 2 2 1
TGV8R 2 2 2 2
TGV9R 2 1 2 2
TGV10R 2 2 2 2
TGV11R 2 2 2 2
TGV12R 2 2 2 2
TGV13R 2 2 2 2

Table XXII. Classication obtained by the SVM method.


Detected class

Test True class Signals Frequencies Mode shapes


TGV1R 1 1 1 1
TGV2R 1 1 1 1
TGV3R 1 1 1 1
TGV4R 1 1 1 1
TGV5R 1 1 1 2
TGV6R 1 2 2 2
TGV7R 2 2 1 1
TGV8R 2 2 1 1
TGV9R 2 1 2 2
TGV10R 2 2 2 2
TGV11R 2 2 2 2
TGV12R 2 2 2 2
TGV13R 2 2 2 2

obtained when SDA-NN is employed. Now, tests TGV4R and TGV5R are misclassied (for
signals) and again, TGV5R for the frequencies. All tests are correctly classied when mode
shapes are used as inputs. Finally, Table XXV summarizes the results obtained when the

Copyright r 2010 John Wiley & Sons, Ltd. Struct. Control Health Monit. 2012; 19:161 186
DOI: 10.1002/stc
184 A. CURY AND C. CREMONA

Table XXIII. Classication obtained by the BDT method.


Detected class

Test True class Signals Frequencies Mode shapes


TGV3R 1 1 1 1
TGV4R 1 1 1 1
TGV5R 1 2 1 2
TGV9R 2 2 2 2
TGV10R 2 2 2 2

Table XXIV. Classication obtained by the NN method.


Detected class

Test True class Signals Frequencies Mode shapes


TGV3R 1 1 1 1
TGV4R 1 1 1 1
TGV5R 1 2 2 1
TGV9R 2 2 2 2
TGV10R 2 2 2 2

Table XXV. Classication obtained by the SVM method.


Detected class

Test True class Signals Frequencies Mode shapes


TGV3R 1 1 1 1
TGV4R 1 1 1 1
TGV5R 1 2 1 1
TGV9R 2 1 1 2
TGV10R 2 2 2 2

Table XXVI. Probabilities of true classication using signals for all dynamic tests.
30% Tr., 10% V, 60% T 40% Tr., 10% V, 50% T 50% Tr., 10% V, 40% T

Method BDT NN SVM BDT NN SVM BDT NN SVM


Best (%) 90 100 100 94 100 100 94 100 100
Average (%) 64 83 67 65 89 80 64 93 85
Worst (%) 21 17 24 30 26 25 38 40 43

SDA-SVM is used. Test TGV5R is here correctly classied by the mode shapes; however, test
TGV9R is incorrectly classed by signals and frequencies.

5.2.3. Results/study 3. The training data set of this simulation is identical to the second study.
The testing data set now comprises the tests TGV8A, TGV10A, TGV5B, TGV7B and TGV13B.
For this study, all the methods provide 100% of correct classication considering either signals
or modal parameters. For the sake of conciseness, tables exposing these results are omitted. This
shows that all methods are capable to learn from an actual structural condition and then classify
similar but unseen behaviors according to one of the dened classes.

5.2.4. Results/study 4. The 41 dynamic tests are considered, representing the three structural
states. In total, 10 000 simulations for different training data sets are carried out. The idea is to
observe the capability of each SDA-classication algorithm for discriminating all the structural
into three groups. First, results based on the symbolic signals are summarized in Table XXVI.

Copyright r 2010 John Wiley & Sons, Ltd. Struct. Control Health Monit. 2012; 19:161 186
DOI: 10.1002/stc
PATTERN RECOGNITION OF STRUCTURAL BEHAVIORS 185

Table XXVII. Probabilities of true classication using frequencies for all dynamic tests.
30% Tr., 10% V, 60% T 40% Tr., 10% V, 50% T 50% Tr., 10% V, 40% T

Method BDT NN SVM BDT NN SVM BDT NN SVM


Best (%) 80 100 88 83 100 95 89 100 97
Average (%) 57 90 53 66 93 56 75 95 54
Worst (%) 12 20 16 25 33 15 31 50 17

Table XXVIII. Probabilities of true classication using mode shapes for all dynamic tests.
30% Tr., 10% V, 60% T 40% Tr., 10% V, 50% T 50% Tr., 10% V, 40% T

Method BDT NN SVM BDT NN SVM BDT NN SVM


Best (%) 77 85 90 81 81 85 80 85 90
Average (%) 33 38 34 31 40 50 35 41 55
Worst (%) 15 18 23 22 20 25 15 23 30

SDA-NN and SDA-SVM present slight better results compared with SDA-BDT. In general,
when the number of tests in the training data set increases, the probability of true classication
also increases. Tables XXVII and XXVIII present the percentages of correct classication when
natural frequencies and mode shapes are used. Overall, SDA-NN produces better results
compared with the other methods. The probabilities using natural frequencies are considerably
better than those using mode shapes.

6. CONCLUSIONS
In this paper, a novel approach based on the coupling of SDA and well-known classication
methods such as BDT, Neural Networks and SVM has been introduced to discriminate different
structural behaviors. For this purpose, raw information (acceleration measurements) as well as
processed information (modal data) has been used for feature extraction.
First, several numerical simulations containing different structural states (damage scenarii)
and different noise levels have been studied. The objectives were (i) to use the classication
methods to discriminate the different structural states and (ii) to analyze the sensitivity of the
proposed algorithms with respect to the presence of noise. First, investigations concerning
classical analyses showed to be not sensitive enough to point out the presence of structural
modications. In so doing, the proposed hybrid approach was employed and very reasonable
results were obtained. Several simulations were carried out, taking into account different sizes of
training and testing data sets. In general, SDA-NN and SDA-SVM methods are providing
better results compared with the SDA-BDT method. This can be explained by the fact that this
latter method is not really a learning procedure.
An experimental application containing three structural states of a railway bridge was also
studied. The objective was to use the previous techniques to classify the different bridge conditions.
To do so, four different studies were performed, containing distinct data sets. Results show that the
hybrid SDA-classication methods are efcient to classify and to discriminate structural
modications either considering the vibration data or considering the modal parameters.

REFERENCES

1. Yan YJ, Cheng L, Wu ZY, Yam LH. Development in vibration-based structural damage detection technique.
Mechanical Systems and Signal Processing 2007; 21:21982211.
2. Doebling SW, Farrar CR, Prime MB, Shevitz DW. Damage identication and health monitoring on structural and
mechanical systems form changes in their vibration characteristics: a literature review. Report LA-13070-MS, Los
Alamos National Laboratory, 1996.

Copyright r 2010 John Wiley & Sons, Ltd. Struct. Control Health Monit. 2012; 19:161 186
DOI: 10.1002/stc
186 A. CURY AND C. CREMONA

3. Zhang QW. Statistical damage identication for bridges using ambient vibration data. Computers and Structures
2007; 85:476485.
4. Lei Y, Kiremidjian AS, Nair KK, Lynch JP, Law KH, Kenny TW, Carryer Ed, Kottapalli A. Statistical damage
detection using time series analysis on a structural health monitoring benchmark problem. Proceedings of the 9th
International Conference on Applications of Statistics and Probability in Civil, San Francisco, U.S.A., 2003.
5. Iwasaki A, Todoroki A, Shimamura Y, Kobayashi H. An unsupervised statistical damage detection method for
structural health monitoring. Smart Material Structures 2004; 13:8085.
6. Yeung WT, Smith JW. Damage detection in bridges using neural networks for pattern recognition of vibration
signatures. Engineering Structures 2005; 27:685698.
7. Reda Taha MM, Lucero J. Damage identication for structural health monitoring using fuzzy pattern recognition.
Engineering Structures 2005; 27:17741783.
8. Fang X, Luo H, Tang J. Structural damage detection using neural network with learning rate improvement.
Computers and Structures 2005; 83:21502161.
9. Chun X, Weilian Q, Dongmei T. An application of data fusion technology in structural health monitoring and
damage identication. Proceedings of the Smart Sensor Technology and Measurement Systems, Belgium, 2005;
451461.
10. Minor C, Johnson K, Rose-Pehrsson S, Owrutsky J, Wales S, Steinhust D, Gottuk D. Data fusion with a
multisensor system for damage control and situational awareness. Proceedings of the IEEE Conference on Advanced
Video and Signal Based Surveillance, U.K., 2007; 10551068.
11. Guo H. Structural damage detection using information fusion technique. Mechanical Systems and Signal
Processing 2006; 20:11731188.
12. Trendalova I, Heylen W. Categorisation and pattern recognition methods for damage localisation from vibration
measurements. Mechanical Systems and Signal Processing 2003; 17:825836.
13. Hastie T, Tibshirani R, Firedman J. The Elements of Statistical Learning: Data Mining, Inference, and Prediction
(2nd edn). Springer Series in Statistics, Springer: New York, 2009.
14. Billard N, Diday E (eds). Symbolic Data Analysis. Wiley: Chichester, 2006.
15. Diday E, Noirhomme M (eds). Symbolic Data Analysis and the SODAS Software. Wiley: Chichester, 2008.
16. Cury A, Cremona C, Diday E. Application of symbolic data analysis for structural modication assessment.
Engineering Structures 2010; 32:762775.
17. Silverman B. Using kernel density estimates to investigate multimodality. Journal of the Royal Statistical Society
1986; 43:9799.
18. Rasson JP, Lallemand P, Adans S. Bayesian decision trees. Symbolic Data Analysis and the SODAS Software.
Diday E, Noirhomme-Fraiture (eds). Wiley, 2008.
19. Bishop C. Neural Networks for Pattern Recognition (1st edn). Oxford University Press: Oxford, 1995.
20. Principe J, Euliano N, Lefebvre C (eds). Neural and Adaptive Systems. Wiley: London, 2000.
21. Burbridge R, Buxton B. An introduction to support vector machines for data mining. 2002.
22. Varma M, Babu BR. More generality in efcient multiple kernel learning. Proceedings of the International
Conference on Machine Learning, Montreal, Canada, 2009; 10651072.
23. Hsu CW, Lin C-J. A comparison of methods for multi-class support vector machines. IEEE Transactions on Neural
Networks 2002; 13:415425.
24. Takahashi N, Nishi T. Global convergence of decomposition learning methods for support vector machines. IEEE
Transactions on Neural Networks 2006; 17:13621369.
25. Cristianini N, Shawe-Taylor J (eds). An Introduction to Support Vector Machines. University Press: Cambridge,
2000.
26. Kaundal R. A SVM-based server for rice blast prediction dedicated to the farming community, 2005. Available
from: http://www.imtech.res.in/raghav/rbpred/algorithm.html.
27. Chang CC, Lin CJ. LIBSVM: a library for support vector machines, 2001. Available from: http://www.csie.
ntu.edu.tw/ cjlin/libsvm.
28. Alvandi A, Cremona C. Assessment of vibration-based damage identication techniques. Journal of Sound and
Vibration 2006; 292:179202.
29. Farrar CR, Duffey TA, Doebling SW, Nix DA. Statistical pattern recognition paradigm for vibration-based
structural health monitoring. Proceedings of the 2nd International Workshop on Structural Health Monitoring,
Stanford, CA, 2000 September 810; 764773.
30. Cremona C. Dynamic monitoring applied to the detection of structural modications. A high speed railway bridge
study. Progress in Structural Engineering and Materials 2004; 3:147161.

Copyright r 2010 John Wiley & Sons, Ltd. Struct. Control Health Monit. 2012; 19:161 186
DOI: 10.1002/stc

You might also like