Damage Detection of Structures Using Support Vector Machines

under Various Boundary Conditions

Marie Shimadaa, Akira Mitab, Maria. Q. Fenga
Department of Civil and Environmental Engineering, UC Irvine, Irvine, CA, USA 92697-2175
b
Department of System Design Engineering., Keio University,
3-14-1 Hiyoshi, Kohoku-ku, Yokohama 223-8522, Japan

ABSTRACT
Many Structural Health Monitoring (SHM) methods have been proposed for the purposes of reducing maintenance
costs and/or assuring performance of civil structures. The objective of this research is to propose a damage detection
system that can obtain the detailed damage information by use of the minimum number of sensors. The proposed
system minimizes the possibility of incorrect judgments. Modal frequencies of a structure are used for pattern
recognition in the proposed method. Changes in multiple natural frequencies can be correlated to the spatial information
of the location of damaged stories. Typically only two vibration sensors, one on the roof and the other on the ground,
detecting a single input and a single output for the structure are needed to determine modal frequencies. Out of many
pattern recognition tools, we propose to use the Support Vector Machine (SVM). This technique has been found
effective. Our previous studies demonstrated that the proposed damage detection method worked well for simple
models such as shear structures and bending structures. However, real buildings have various boundary conditions at
their supports. In this study, the SVM technique was applied to damage detection of structures with various boundary
conditions. The feature vectors for SVMs are generated based on the model of a structure. Then locations of structural
damage are detected by inputting the measured structural vibration data into the SVMs. From simulation, it was found
that the influence of the change in boundary conditions on the lower modes is larger. We performed experimental
studies on damage detection of power distribution poles that had overhead wires. We proposed a method for
determining the boundary conditions of the poles and verified this method based on measured vibration data. We
demonstrated the effectiveness of the proposed method in detecting damage in the poles.
Keywords: system identification, support vector machine, modal analysis, damage detection

1. INTRODUCTION
Structural Health Monitoring (SHM) systems are getting strong attention for maintaining proper performance of
building structures against hazards such as large earthquakes and strong winds. Many approaches for structural health
monitoring have been proposed for maintenance cost reduction or performance guarantee of civil and building
structures. However, installing an SHM system to a building is not an easy task as the conventional building has many
possible damage scenarios for such hazards. Detecting possible damages associated with each damage scenario by an
SHM system is very difficult and requires the prohibitive number of sensors. Thus the obstacle we often face when
installing an SHM system into a building is the trade-off relation between the number of sensors versus the accuracy of
the damage detection. The large number of sensors results in expensive costs for the system as well as enormous efforts
needed for wiring and designing a proper data acquisition system. Complicated and expensive SHM system is by no
means practical for most buildings. If we assume an occasion of applying the system to a tall building, it would be
unrealistic to say it is economically feasible to deploy several acceleration sensors to each floor for the monitoring
purpose. Thus, automatic systems and innovative diagnosis mechanisms are necessary for SHM. In line with this
scenario, the purpose of this research is to propose a system that can obtain the detailed damage information by
installing the minimum number of sensors. The system is optimized to minimize the possibility of incorrect judgments.
Many damage detection algorithms based on the modal properties of a structure such as modal frequencies, mode
shapes, curvature mode shapes and modal flexibilities have been studied for several decades. However, in most

Smart Structures and Materials 2006: Sensors and Smart Structures Technologies for Civil,
Mechanical, and Aerospace Systems, edited by Masayoshi Tomizuka, et al., Proc. of SPIE Vol. 6174,
61742K, (2006) 0277-786X/06/$15 doi: 10.1117/12.658956
Proc. of SPIE Vol. 6174 61742K-1
Downloaded From: http://proceedings.spiedigitallibrary.org/ on 10/06/2015 Terms of Use: http://spiedigitallibrary.org/ss/TermsOfUse.aspx

algorithms, identifying the precise location and magnitude of the damage is difficult. If not completely possible, the
accuracy and reliability is not sufficient. Alternative approaches based on identification of parametric models from the
time domain or frequency domain data are being extensively studied. They are promising tools to provide satisfactory
estimation of damage. However, they are not effective if the number of sensors is very limited1).
The purpose of this study is therefore to establish a new method to identify the location of damage using limited number
of sensors. Modal frequencies of a structure are used for pattern recognition in the proposed method. A natural
frequency change associated with a certain mode does not provide the spatial information of structural damage.
However, multiple natural frequency changes can provide the spatial information on the location of damaged stories.
Typically only two vibration sensors on the roof and the ground detecting a single input and a single output for a
structural system are needed to determine modal frequencies. The modal frequencies are used for forming feature
vectors for pattern recognition. The detection of the damage is based on the relation between the natural frequency
changes and the damage induced in a structure with the help of support vector machines (SVM). The SVM was recently
proposed as a new technique for pattern recognition and is effectively utilized in the system to achieve this purpose.
Based on the modal frequency changes, the damage location and its severity are identified by the SVM.
Our previous studies demonstrated that the proposed damage detection method worked well for simple models such as
shear structures2) and bending structures3). However, real buildings have various boundary conditions at their supports.
In this study, the SVM technique was applied for damage detection of structures with various boundary conditions. We
focused that how changes in boundary conditions of a structure influence the damage detection. Thus, we propose a
method to detect the damage location in structural part of structures with various boundary conditions and consider the
influences of boundary condition determined from prior information to diagnosis using SVM.

2. SVM-BASED DAMAGE DETECTION

2.1 Basis of support vector machine
The Support Vector Machine (SVM) is a mechanical learning system introduced by Vapnik and co-worker that uses a
hypothesis space of linear functions in a high dimensional feature space4), 5). The simplest model of SVM is the socalled Linear SVM (LSVM) shown in Fig 1. It works only for data which are linearly separable in the original feature
space, and hence cannot be applicable to many real-world problems. In the early 1990s, nonlinear classification in the
same procedure as LSVM became possible by introducing nonlinear functions called a Kernel function, without being
conscious of actual map space. The technique extended to nonlinear feature spaces is called Nonlinear SVM (NSVM)
shown in Fig 2. In what follows, we assume a training sample S consisting of N sets of vectors x R with
i = 1,2,L, N .
n

S = ((x1 , y1 ),L , (x N , y N ))

(1)

Each vector x belongs to either of two classes and thus is given a label y {1,1} . Then, the next step is to define a
boundary that divides two classes. The pair of (w, b) defines a hyperplane as
i

( w T x) + b = 0

(2)

This hyperplane is called separating hyperplane. It is noted that Eq (2) is not sufficient to define this separating
hyperplane uniquely. Many hyperplane are possible under the condition given by Eq (2). An example hyperplane is
shown in Fig 1. Among them, the optimum hyperplane is obtained as follows.
The optimal separating hyperplane (OSH) is defined by the hyperplane that divides S leaving all the points of the same
class on the same side while maximizing the margin which is the distance of the closest point of S (see in Fig 2). This
closest vector x is called the support vector.
i

Proc. of SPIE Vol. 6174 61742K-2

Downloaded From: http://proceedings.spiedigitallibrary.org/ on 10/06/2015 Terms of Use: http://spiedigitallibrary.org/ss/TermsOfUse.aspx

.
(wTx)+b<O

____________ a.' \
0

(wTx)+b>O

0
0

I (wTx)+b0}

Figure 1: An example of separating hyperplane

Figure 2: Optimal separating hyperplane maximizing distance

Hence, the OSH (w, b) can be determined by solving an optimization problem defined by
minimize
subject to

1 T
(w w ),
2
T
yi ((w xi ) + b) 1,

d (w ) =

(3)

i = 1,2,L, N

The resulting SVM is called Hard Margin SVM because no error is allowed. However, such SVMs can be used only for
a limited number of problems. In order to relax the situation, we allow a small number of misclassified feature vectors.
The previous optimization problem in Eq (3) is generalized by introducing N nonnegative variables = ( , , L , )
such that
1

1
(w w ) + C ,
2
y ((w x ) + b) 1 , i = 1,2, L , N

minimize d (w ) =
subject to

(4)

The SVM defined by Eq (4) is called Soft Margin SVM. The purpose of the term C , when the sum is for
i = 1,2, L, N is to keep under control the number of misclassified feature vectors. The parameter C is a regularization
parameter. The OSH tends to maximize the minimum distance 1 w for small C, and minimize the number of
misclassified vectors for large C.
So far, we described the case of LSVM. To allow flexible recognition, we introduce nonlinear transformation to a set of
original feature vectors x into a high-dimensional space by a mapping F : x a z as explained in Fig 2 so that
linear separation would become possible in the high-dimensional space there. However the direct computation of inner
products in a high-dimensional space is prohibitively time-consuming. Therefore, we are interested in cases where these
expensive calculations can be significantly reduced by using a kernel function which satisfies the Mercers theorem
such that
i

(F (x) F (x)) = K (x, x )

(5)

The typical kernel function, Gaussian kernel, is expressed in the form.

K (x, x ) =
i

exp( x x )
i

(6)

(Gaussian kernel)

Many kernel functions satisfying the condition given by Eq (5) have been proposed. For our problems, the accuracy of
the damage diagnosis using the Gaussian kernel was found satisfactory.

Input space

High-dimensional feature space

Figure 3: Concept of nonlinear SVM

Original space

Proc. of SPIE Vol. 6174 61742K-3

Downloaded From: http://proceedings.spiedigitallibrary.org/ on 10/06/2015 Terms of Use: http://spiedigitallibrary.org/ss/TermsOfUse.aspx

2.2 Proposed Algorithm

We proposed the method to detect the damage of structure under various conditions. In this proposed method, the
natural frequency changes induced from damage of structure are used as feature vectors and detect the location of
damage using classification technique of SVM.
In this method, first, the boundary condition for the target structure is determined from prior information. Then, the
model of the structure is developed with determined boundary condition. From this model, SVM is constructed using
the modal information. Feature vectors for construction of SVMs are generated based on the model of a structure. After
that, the modal information of the real structure is extracted using system Identification method. Finally, locations of
structural damage are detected by inputting the measured structural vibration data into the SVMs.
We verified the proposed method by simulations and experiments. Through simulations, the influence of the changes in
the boundary conditions to modal frequencies was studied using modal analysis. And we performed an experimental
study on damage detection of power distribution poles with overhead wires as structures under various boundary
conditions. The method to determine the boundary condition of the poles from prior information was verified based on
the measured data of vibration tests using system identification technique.

3. EFFECTS OF VARIOUS BOUNDARY CONDITIONS ON DAMAGE DETECTION

3.1 Simulation models under various boundary conditions

For the purpose of verifying the proposed method, we studied the influences of changes in boundary conditions to
modal frequencies using shear beam models and bending beam models. We use three kinds of boundary conditions at
each end of the beams such as clamped (C), free (F) and pined (P) according to the way of their supports.
The C-F model was assumed as the basic model for each model set. And we compared it with C-C model and C-P
model. In addition, for the bending model set, we include models with more complicated boundary conditions such as
models added a mass and a spring on the top.
3.2 Influence of the change in boundary conditions

CC1

From FE analysis, the 1st to 5th modal frequencies of each model were computed. Assuming the modal frequencies of
C-F model as a basic, we generated the frequency change rate for each mode. The results for C-C model ad C-P model
are shown in Fig 4 and the results for mass additional model and spring additional model are shown in Fig 5.
In case of comparing C-C and C-P model to C-F model, the modal frequencies decrease and the changes are larger on
the lower mode for shear beam models and bending beam models.
For mass additional model, the modal frequencies increase and for spring additional model, they decrease and the
change rate is larger on the lower mode. From these results, it was shown that the influence of the change in boundary
conditions on the lower mode is larger.

0.2

CD

Mode iNombe

(2) Shear model

Figure 4: Compared models and Frequency change rate

z"
C

(1) Compared models

(IC

(c) C-P model

05

0.4

(b) C-C model

Freqeocy Chaoge

0.l 5

Rate

(a) Basic model (C-F)

(3) Bending model

Proc. of SPIE Vol. 6174 61742K-4

Downloaded From: http://proceedings.spiedigitallibrary.org/ on 10/06/2015 Terms of Use: http://spiedigitallibrary.org/ss/TermsOfUse.aspx

Spring value [kgfm]

Mass ratio
0.0

009

ij 008

05

15

Ct
cr
007
Ct

0.4

*5

006

10

-c 005
Ct

0.3

C-)

4 20

0 04

003
C 002

02

Ct

0.1

4- 200

500

w 001

1000

50
-100

Mode number

Mode number

(1) Mass additional model

(2) Spring additional model
Figure 5: Frequency change rate of mass additional model and spring additional model

4. PERFORMANCE OF DAMAGE DETECTION METHOD

4.1 Simulation model of power distribution poles

In this study, we performed the analytical and experimental verification using power distribution poles as an example of
simple structures under various boundary conditions. The boundary conditions of power distribution poles are
determined by many aspects such as overhead wires, branch line, switches and transformers. And for the poles with
overhead wires, the weight and number of wires, distance between poles and the sag of wire were considered.
From the design drawing, we develop the model represents a pole with overhead wires. The simulation model is shown
in Fig 6. The model is a 1-dimensional bending FE model divided into 20 elements. The section of the model is circular
cavity and the outside and inner diameters of the support point become smaller along its height. The wire model for
along the wire direction includes the effects of masses and springs and for wire the orthogonal direction only includes
the effects of springs.
K M

K M

(a) Wire direction model

K
(b) Wire orthogonal direction model
(2) Wire model
(1) Pole model
Figure 6: Simulation model of power distribution pole with overhead wires

4.2 Vibration tests of power distribution poles

We conducted a series of vibration tests for power distribution poles for verifying the model developed from the design
drawing. We hit a pole without overhead wire and a pole with overhead wire by the impact hammer. The hammering
and measurement of acceleration were conducted in 2 directions of wire direction and wire orthogonal direction. System
identification techniques were used for modal analysis using measured input and output data. Using ARX models and
ARMAX models, the transfer functions were obtained as shown in Fig 8. In this vibration tests, we could identify the
1st to 4th modal frequencies of the system. We compared the analytical results of model developed in Section 4.1 and
measured results as shown in Table 1. From these results, the developed analytical model was updated.

Proc. of SPIE Vol. 6174 61742K-5

Downloaded From: http://proceedings.spiedigitallibrary.org/ on 10/06/2015 Terms of Use: http://spiedigitallibrary.org/ss/TermsOfUse.aspx

10

Z0

Fre quenv3?Hz)

40

50

Wire orthogonal direction

I&

10

20

30

40

Fq,y (HZ)

10

20
Fqe,y

40

50

Wire direction
(1) Pole without overhead wire
Figure 7: Power distribution pole

(2) Pole with overhead wire

Figure 8: Transfer function of poles from vibration tests

Table 1: Modal frequency [Hz] of vibration analysis and FE analysis

Mode
1
2
3
4

Vibration test analysis

with wire
without wire
wire
orthogonal
direction
direction
2.0470
5.2611
9.4435
12.1212
13.9169
24.0789
20.6730
23.9655
45.5349
42.8965

FE analysis
pole model with wire
wire
orthogonal
direction
direction
5.9317
5.6453
12.0395
13.9236
20.6310
26.9759
34.3360
48.1206

pole model
2.2653
9.9053
24.7999
46.9818

4.3 Performance of SVM for damage detection

4.3.1 Construction of feature vectors

The simulations representing damage scenarios using developed models for power distribution poles were performed.
The i-th vector of natural frequency change associated with the i-th element damaged pattern for a pole described in Fig
6 is given by
T

,
, L,
p = [ p , p ,L, p ] =
,

1i

2i

1i

2i

Ni

Ni

(i = 1,2, L, N )

(7)

where is the natural frequency of r-th mode for the model without damage and is the change of the r-th
natural frequency due to the i-th element damaged pattern. The stiffness of the damaged element is varied from 90% to
50% of the original stiffness. The 1st to 4th natural frequency change rates were used for feature vectors. Slight
modification is made for the vector defined by Eq (7) to remove harmless components. The feature vector is, therefore,
defined in Eq (8). This modification corresponds to removing components associated with the stiffness reduction for all
stories with the same reduction ratio.
r

ri

q = p min [ p ], p min [ p ],L , p min [ p ]

1i

all j

ji

2i

all j

ji

Ni

all j

ji

Proc. of SPIE Vol. 6174 61742K-6

Downloaded From: http://proceedings.spiedigitallibrary.org/ on 10/06/2015 Terms of Use: http://spiedigitallibrary.org/ss/TermsOfUse.aspx

(8)

4.3.2 SVMs for damage detection

Considering a 20-element pole, training feature vectors were generated. Assuming five levels of stiffness reduction,
10%, 20%, 30%, 40% and 50% reduction of the stiffness for each element, the corresponding five feature vectors were
generated for every elements. For the undamaged structure, a feature vector consisting of zero elements was used. The
number of feature vectors used for training is, therefore, 101 ( = 5 20 + 1 ). As we use a two class separation algorithm,
we need 21 SVMs defined by no damage (SVM0) and i-th element damage (SVMi) for 20 elements. From Fig 9, we
could observe that the vectors show distinctive patterns depending on which element is damaged. This fact indicates
that recognition of the damage location may be possible from the natural frequency change vectors.
The parameters characterizing support vector machines are the order of Gaussian kernel and the regularization
parameter C appeared in Eq (7). The correctness is the probability of the training feature vectors that are classified
correctly by the separating hyperplane. The l-o-o (leave-one-out bounds) represents the probability of feature vectors
that do not exist in a margin. Thus, the discernment accuracy is better as l-o-o is closer to 100%. The parameters
and C were chosen to maximize the l-o-o while keeping the correctness to be 100%.
The performance tests were conducted assuming the stiffness reduction 30% reduction of the stiffness for each element.
The i-th feature vector is the i-th element damaged data. In total, 20 feature vectors were prepared for the performance
test. In Fig 10, outputs from each SVM are plotted. The positive output from SVMn indicates that the tested feature
vector is classified to n-th class. If the negative output is observed, then the tested feature vector should be classified to
other classes. We could detect the damage of each element correctly by constructed SVM. We could identify the
damaged element by using SVM.
io_3

a 0.021

004

0
002

005

004

>

002

>

004

' 005

002

>

In- 0

In_0

00
004

0.04

0.04

002
0

1ii
IIt

-0i
1>
0
I-J
>
C0

0.04

i:-J
_1
I

2 005
>
In-

004

iga

004

oo2L

:000402!;
Mode Number

Mode Number

Figure 9: Feature Vectors for damage detection of power distribution poles

with 2wires of Sag 1%, 0.35kgf/m in wire direction
(Stiffness reduction: 10%, 20%, 30%, 40% and 50%)

Proc. of SPIE Vol. 6174 61742K-7

Downloaded From: http://proceedings.spiedigitallibrary.org/ on 10/06/2015 Terms of Use: http://spiedigitallibrary.org/ss/TermsOfUse.aspx

0 .................

I
!0 ....;.............. 0
I

In

20

In

00

Ic

00

0 .................
20

0 .................
ID

29

29

IS

20

!_ -

IS

20

2 29

IS

20

0
29

ID

0
I

10

10

0
10

20
2

10

10

20
2

Ic

In

29
2

I
I

10

Data Number

70

70

10

IS

0
0
0
I
0

29
2

29
00

20

10

Ic

00

IS

20

lb

20

10

Ic

00

10

20

2?
7(1

IS

00

2
2

20

20

0 .................

2
2

Data Number

Figure 10: Outputs of 20 SVMs

4.3.3 Allowable error margin in decision of boundary conditions for damage detection
To verify the effectiveness of proposed method, how the error in decision of boundary conditions influences the
accuracy of damage detection was studied. In the case associated with an error in the spring constants of a wire in the
training models and testing models, we compared the number of misclassified data for the error in wire direction and
wire orthogonal direction. It was shown that 10 % error is allowed in wire orthogonal direction, that is, it kept the
status of 100 % correctness classification if the error is controlled below that level. And in wire direction, the allowable
error is lower like 1%.

5. CONCLUSION
The objective of this research is to propose a damage detection system that can obtain the detailed damage information
by use of the minimum number of sensors and SVM. Modal frequencies of a structure are used for pattern recognition
in the proposed method to minimize the possibility of incorrect judgments. Modal frequencies can be determined using
only two vibration sensors typically. In this study, the SVM technique was applied for damage detection of structures
with various boundary conditions. The feature vectors for construction of SVMs are generated based on the model of a
structure. Then locations of structural damage are detected by inputting the measured structural vibration data into the
SVMs. From simulation, it was found that the influence of the change in boundary conditions on the lower modes is
larger. We performed an experimental study on damage detection of power distribution poles with overhead wires. We
proposed a method for determining the boundary conditions of the poles and verified this method based on measured
vibration data. We demonstrated the effectiveness of the proposed method in detecting damage in the poles through
simulation. In terms of allowable error in decision of boundary conditions, it was shown that plus-minus 10% is allowed
in wire orthogonal direction, and in wire direction the error margin is low.

Proc. of SPIE Vol. 6174 61742K-8

Downloaded From: http://proceedings.spiedigitallibrary.org/ on 10/06/2015 Terms of Use: http://spiedigitallibrary.org/ss/TermsOfUse.aspx

6. ACKNOWLEDGEMENT
Mr. T. Tsutsumi and Mr. Y. Kobayashi at Tokyo Electric Power Company provided the insightful suggestions. The
vibration tests of power distribution poles were conducted at the research facility of Tokyo Electric Power Company.
Deep appreciation is owed for their help to make the vibration tests possible. The authors also thank Mr. R. Kudo,
undergraduate student at Keio University, for kind help for the vibration tests.

REFERENCES
1.
2.
3.
4.
5.

A. Mita, Structural Dynamics for Health Monitoring, SANKEISHA, Japan, 2003

A. Mita and H. Hagiwara, Quantitative Damage Diagnosis of Shear Structures Using Support Vector
Machine, KSCE Journal of Civil Engineering, Vol.7, No.6, p.683-689, November 2003
M. Shimada, A. Mita, Damage Assessment of Bending Structures Using Support Vector Machines, Smart
Structures and Materials 2005: Sensors and Smart Structures Technologies for Civil, Mechanical and
Aerospace, San Diego, California USA, March 2005, 923-930
Vapnik and V.N., The Nature of Statistical Learning Theory, Springer, 1995
Nello Christianini and Jhon Shawe-Taylor, An Introduction to Support Vector Machines, 2000

Proc. of SPIE Vol. 6174 61742K-9

Downloaded From: http://proceedings.spiedigitallibrary.org/ on 10/06/2015 Terms of Use: http://spiedigitallibrary.org/ss/TermsOfUse.aspx