archives of civil and mechanical engineering 17 (2017) 609–622

Available online at www.sciencedirect.com

ScienceDirect

journal homepage: http://www.elsevier.com/locate/acme

Original Research Article

Detection of fatigue cracking in steel bridge girders:

A support vector machine approach

Hassene Hasni, Amir H. Alavi *, Pengcheng Jiao, Nizar Lajnef

Department of Civil and Environmental Engineering, Michigan State University, East Lansing, MI 48823, USA

article info abstract

Article history: This study presents an artificial intelligence approach for the detection of distortion-induced
Received 15 June 2016 fatigue cracking of steel bridge girders based on the data provided by self-powered wireless
Accepted 19 November 2016 sensors. The sensors have a series of memory gates that can cumulatively record the duration
Available online 31 January 2017 of the applied strain. The gates are activated as soon as the electrical charge generated by
piezoelectric strain transducer exceeds pre-defined thresholds. In the present study, the
Keywords: distribution of the sensor output has been characterized by a Gaussian cumulative density
Support vector machine function. For the analysis, extensive finite element simulations were carried out to obtain the
Fatigue cracking structural response of an existing highway steel bridge girder (I-96/M-52) in Webberville,
Steel bridges Michigan. Different damage states were defined by extending the lengths of the crack at the
Self-powered wireless sensors web gaps from 10 mm to 100 mm. Damage indicator features were extracted for different data
Damage detection acquisition nodes based on the sensor output distribution. Subsequently, support vector
machine (SVM) classifiers were developed to fuse the clustered features and identify multiple
damage states. The results indicate that the models have acceptable detection performance,
specifically for cracks larger than 10 mm. The best classification performance was obtained
using the information from a group of sensors located near the damage zone.
Published by Elsevier Sp. z o.o. on behalf of Politechnika Wrocławska.

The stiffeners connect the girder to the diaphragm. Over many

1. Introduction
Multi-girder steel bridges are widely used throughout the
highways in the United States. One of the main factors
affecting the performance of these structures is the applica-
tion of the repetitive loading over the steel-girder components.
These load-carrying components deform under the live (traffic
load) and the dead load of the structure. A typical steel girder
bridge is composed of three main parts: girders, diaphragms,
and stiffeners. The diaphragms are structural elements that
provide resistance to the transverse traffic and wind loading.
years, inspections conducted on steel-girder bridges revealed
that these structures are suffering from fatigue cracking under
cyclic loading [1]. More specifically, low resistance to fatigue
has been observed in structural members subjected to out-of-
plane distortion. The phenomenon of out-of-plane distortion
is impacted by a variety of factors such as thermal forces,
traffic flow, differential deflection of the adjacent beams, etc.
[2,3]. Fig. 1 displays the schematic formation of fatigue cracks
in a steel girder caused by out-of-plane distortion. Fig. 1(a)
displays an illustration of a steel bridge before deformation in
a perspective view, and Fig. 1(b) shows the side view of the

* Corresponding author.
E-mail addresses: ah_alavi@hotmail.com, alavi@msu.edu (A.H. Alavi).
http://dx.doi.org/10.1016/j.acme.2016.11.005
1644-9665/Published by Elsevier Sp. z o.o. on behalf of Politechnika Wrocławska.
610 archives of civil and mechanical engineering 17 (2017) 609–622
Fig. 1 – Schematic illustration of distortion-induced fatigue cracking: (a) bridge before deformation in a perspective view, (b)
side view of the bridge in the initial stage, and (c) different types of fatigue cracks caused by out-of-plane distortion (D).
bridge in the initial stage. Fig. 1(c) schematically shows the
cracks caused by out-of-plane distortions. It can be seen that
the deformations of the girder web are caused by the
differential displacement (d) between the two girders, which
leads to the out-of-plane distortion (D). Such distortion
eventually causes fatigue cracks on the girders in the form
of horseshoe and horizontal cracks.
Therefore, fatigue cracks usually occur at the girder web
gap due to out-of-plane distortion. The distortion-induced
fatigue cracks may occur as horizontal or horseshoe cracks at
the top or bottom of the girder to stiffener connections (Fig. 1
(c)). More details on the forming mechanism of these cracks
can be found in [4,5]. Different models have been developed to
investigate the behavior of bridges [6], with particular focus on
the retrofitting approaches to deal with this common type of
structural damage [5,7]. However, the selection of an appro-
priate repair strategy is complicated and depends on many
factors. On the other hand, significant cost of maintenance and
retrofitting of stiffener-girder connections implies the neces-
sity of detecting the damage progression at early stages to
prevent severe damage to the bridge structures.
The main goal of this study is to develop a new artificial
intelligence (AI)-based approach for the detection of distortion-
induced fatigue cracking of steel bridges based on the data
provided by a newly developed self-powered wireless sensor.
The sensor is empowered using piezoelectric transducers
through harvesting energy from the mechanical loading
experienced by the structure. Different studies have been
conducted to investigate the energy harvested from ambient
excitation [8]. In order to calibrate the AI detection models,
different finite element (FE) models of steel girders with
complex geometry components were developed and the
structural response of the girder was subsequently obtained.
The fatigue life of the girder was determined based on the
J-integral concept and Paris Law [9,10]. Several damage states
were defined by extending the crack lengths. Different numbers
of sensing locations were defined to monitor the strain changes
due to damage progression. The sensing nodes were placed
around the connection between the webs and the stiffeners to
determine the optimal sensors configurations that maximize
the detection performance of fatigue cracking. Thereafter,
features representing the sensor output were extracted from
the strain data at the sensing nodes. The obtained features
were then fed into a support vector machine (SVM) classifier to
identify multiple damage states.
2. The self-powered wireless sensor
Wireless sensors are widely used as alternatives to the
traditional wired sensors for structural health monitoring
(SHM) [11,12]. A major limitation in the deployment of the
 archives of civil and mechanical engineering 17 (2017) 609–622
conventional wireless sensors is finding a reliable and cost-
effective power supply. Many of these sensors use batteries
that have a limited power lifetime and it needs to be replaced
regularly. A viable solution to this power dilemma is to
harvest energy from the ambient excitations [13,14]. In this
context, piezoelectric transducers are widely used to convert
environmental mechanical energy into an electrical energy
[15,16].
In general, piezo-based self-powering can be categorized
as: (1) harvesting electrical energy from the signals that are
different from the signal being monitored or (2) from the
signal being monitored [17]. Nearly all of the existing energy
harvesting studies are focused on the first approach [18,19].
Based on the second approach, a new class of self-powered
wireless sensors (SWS) has been recently developed and
tested by the authors [20–23]. This type of sensor uses
piezoelectric transducers to empower an array of ultra-low
power floating gate computational circuits. SWS has a series
of memory cells that cumulatively store the duration of
strain events when the amplitude of the input signal, coming
from the piezoelectric material, exceeds different thresholds.
In fact, the sensor acts as a non-volatile memory for data
storage, which optimize the need of the sensor in power. The
data could be retrieved offline without the need to power the
sensor. A radio frequency identification (RFID) scanner can
be used to periodically read the data stored on-board the
sensor [20,23].
One of the main advantages of this sensing system is the
fact that it is ‘‘response-based’’. All the effects due to
variations in load location, load magnitude, traffic wander,
environmental effects such as temperature and moisture,
material aging and degradation are aggregated in the strain
response recorded by the sensor over time. This feature
makes the sensor suitable for long-term SHM. Most of the
other existing solutions evaluate the conditions of the system
at a given instant. These methods present only a snapshot at
the time where the measurements are taken. Thus, the
obtained results are highly influenced by the environmental
conditions. Since the developed SWS records each and every
event at all time, it will aggregate all these short-term
fluctuations. Thus, if long-term shifts are observed in the
results, they are most probably correlated with condition
degradation. An illustrative example of the level crossing
cumulative time counting implemented by the gates with a
constant injection rate can be found in [21,22]. Despite several
advantages offered by this self-powered sensing technology,
interpretation of the compressed data generated by such
systems is a challenging task. With the purpose of solving
this, previous research of our group revealed the notable
efficiency of the SWS for continuous monitoring of infra-
structural systems [20]. In addition, previous studies by Alavi
et al. [21–24] on damage detection using the SWS did not take
into account the number of cycles to failure, which is crucial in
the case of fatigue failure.
The present study proposes an SVM approach for the
detection of fatigue cracking of steel bridges using the data
provided by the SWS. The emphasis was placed on the
distortion-induced fatigue cracks occurred at girder to stiffen-
er connections. The entire damage detection procedure can be
divided into the following major phases:
611

Numerical simulation of the targeted structure;

Extraction of strain data and generation of the distribution
histograms based on the SWS sensing mechanism;

Extraction of preliminary damage indicator features;

Fusion of data from a network of sensors to define more
informative damage indicators;

Fatigue damage classification using an SVM approach.
3. Numerical analysis of distortion-induced
fatigue cracking
3.1. Geometry, loading and boundary conditions of the
steel girder
For the numerical simulation carried out in this study, the
highway steel bridge (I-96/M-52) in Webberville, Michigan, U.S.
was selected. The steel girder under consideration was
modeled using ABAQUS Version 6.12. Fig. 2 shows the
geometry, loading and boundary conditions of the selected
structure. Fig. 2(a) presents the geometry of the structure. In
Fig. 2(a1), a 3-D model was created using SolidWorks. Fig. 2(a2),
(a3), and (a4) show the end view, top view, and side view of the
model respectively. It can be seen that the overall length of the
girder was 7.62 m and the three beams were spaced by 1.93 m.
The steel girders have the same cross section (W920  233) as
the I-96/M-52 bridge. A web gap length of 25.4 mm was used for
the simulations. Fig. 2(b) displays the boundary conditions of
the model. The steel girders were modeled with a simply
supported boundary conditions. The top flanges of the three
beams were restrained with respect to translation in the z
direction and rotation along x axis. The loading was applied in
the form of vertical displacements to the cut edges of the lower
flanges of the two exterior I-beams (Fig. 2(b)). The imposed
displacements were 5 mm vertical displacement at the left
outer girder and 15 mm at the right outer girder. More details of
the dimensions and material properties of the stiffener and
diaphragm plates are listed in Table 1. A linear elastic material
was selected for the analysis.
3.2. Numerical model
3.2.1. Shell element-based FE model
The structure was modeled using shell elements in ABAQUS.
Shell elements performs sufficient computation efficiency
since they formulate thin structures with much less elements
than 3-D solid elements. Quadrilateral shell elements with
reduced integration (S4R) were used in this study. The total
number of elements and nodes of the intact structure were
180,320 and 182,050, respectively. The simulation of the intact
structure took about 2 min. The meshed structure is shown in
Fig. 3(a) and an isometric drawing of the used S4R element is
shown in Fig. 3(b). A static analysis was selected for the
simulations. Besides, the analysis was done for small
deformations.
At the region of web gap, a finer mesh was used with an
element size smaller than 2 mm such that the stress
concentration zones can be accurately captured. The maxi-
mum stresses in the central girder were located around the
612 archives of civil and mechanical engineering 17 (2017) 609–622

Fig. 2 – Selected steel structure: (a) technical drawings of the structure, including (a1) a perspective view (a2) an end view, (a3)
a top view, and (a4) a side view, and (b) assembly of the structure in ABAQUS with loading and boundary conditions.

Table 1 – Geometry and material properties of the girder.

Material Geometry (mm)

Overall model Stiffener Diaphragm Beams

Young's modulus (GPa) 200 Length 7620 Length 812 Length 1863 W920  223 (W36  150)
Spacing 1933 Width 140 Width 457
Poison's ratio 0.3 Web gap length 25.4 Thickness 12.7 Thickness 12.7

connection stiffener to web. As it seen in Fig. 4, the maximum
obtained principal strain of the intact structure was
1.103  103.
The stiffeners and diaphragms were meshed using an
element size of 25 mm  25 mm and the beams were meshed
using a 2 mm  2 mm element size around the stiffener to web
connection. The rest of the central girder was meshed using
2 mm  76 mm quadrilateral element. Different simulations
were performed to find the optimal mesh size that guarantees
the numerical convergence of the solution. The results of the
convergence analysis are shown in Table 2 and Fig. 5. It should
be noted that the strain values were obtained for an element of
 archives of civil and mechanical engineering 17 (2017) 609–622 613

Fig. 3 – Mesh details: (a) meshed structure, and (b) S4R element.

Fig. 4 – Maximum principal strain around the area stiffener to web connection of the central girder.

the central girder web at the connection stiffener-web. As

Table 2 – Results of the numerical convergence analysis. mentioned above, this element presents the highest deforma-
Element size h (mm) Maximum principal strain (104) tion of the central girder. Fig. 5 shows that the element size of
2 mm  2 mm is sufficient to obtain satisfactory results. In
200 1.74
particular, the error between the element sizes of
150 2.37
1.5 mm  1.5 mm and 2 mm  2 mm is only 0.18% (2 me
100 4.34
50 4.71 difference). Therefore, the 2 mm S4R elements provide
25 5.9 adequate accuracy of the numerical results.
10 8.29
5 9.19 3.2.2. Comparison between shell and 3D solid element-based
3 9.82 FE model
2 11.03
In order to validate the accuracy of the shell model developed
1.5 11.01
in the previous, a 3-D solid finite element model was
614 archives of civil and mechanical engineering 17 (2017) 609–622

12 Table 3 – A comparison of the shell and 3D FEM models.

Maximum Principale Strain(×10-4)

10
Shell 3D solid Error (%)
4
Maximum principal strain (10 ) 11.03 10.93 0.91
8 Computational time (min) 2 41

6
and the maximum principal strain at the connection between
the stiffener and the web.
4
Different simulations were conducted to obtain the optimal
element size that guarantees the numerical convergence. The
2
optimum was found to be 2 mm  2 mm  2 mm close to the
stiffener to web connection as shown in Fig. 6(a). The stiffeners
0
0.0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 were meshed with an element size of 20 mm  20 mm 
1/h (mm-1) 4 mm. The element size of all the other regions varied between
50 mm and 150 mm. A summary of the comparison between
Fig. 5 – Variation of the maximum principal strain as a the shell model (S4R element-based) and the 3D solid model
function of the element size. (C3D8R element-based) is illustrated by Table 3. The error of
the maximum principal strain occurred at the elements
located at the connection stiffener to web was found as
developed and simulated using ABAQUS. The FE model that 0.91%. However, the computational time of the 3D solid finite
has exact the same geometry, loading, and boundary condi- element model was up to 20.5 times higher than the shell
tions as the shell model. The entire steel girder was meshed model. Compared to the 3D solid model, the shell model
using eight-node brick elements with reduced integration delivered accurate results in a much less computational time,
(C3D8R). Fig. 6 shows the assembly of the structure, the refined and therefore, the shell model was used for the rest of the
portion of the central girder (at connection stiffener to web), simulations.

Fig. 6 – 3D finite element model (a) assembly and refined mesh portion, and (b) maximum principal strain around the
connection stiffener to web.
 archives of civil and mechanical engineering 17 (2017) 609–622
Fig. 7 – Mesh around the crack tip.
3.2.3. Fatigue analysis of the steel girder
In this paper, the contour integral method was used to
calculate the stress intensity factor. A predefined crack was
placed 25.4 mm (1 inch) from the upper flange of the central
girder, with an initial length of 10 mm. Damage states were
defined by increasing the crack length from 10 mm to 100 mm.
For each damage sate, the maximum principal strains and the
J-integral values were extracted to determine the number of
cycles needed to propagate the crack and to compute the
sensor output.
The optimal element size of the damaged structure was
0.2 mm around the crack tip vicinity and 1.8 mm in the contact
area between the stiffener and the web of the central girder.
Triangular elements were used to mesh the crack front. In
order to perform a contour integral analysis under ABAQUS,
three different entities should be defined: the crack front, the
crack tip and the crack extension direction. The crack front is
the forward part of the crack and it is useful to evaluate the first
contour integral (Fig. 7). The crack tip is a point (for 2-D parts) to
be selected from the assembly where the crack extension
direction is defined. Thereafter, the crack extension direction
is defined by selecting the points representing the start and the
end of the crack.
Triangular elements of type S3 were used to mesh the crack
front region and the remaining part is meshed using
quadrilateral elements (S4R) as indicated in Fig. 7. The
triangular elements are recommended to mesh the crack
front. However, this type of elements should not be used to
mesh the contour integral region. Quadrilateral elements
should be used instead [25].
A finer mesh around the crack tip was used to capture the
high stress field at the crack front region. The J-integral is
calculated based on elements surrounding the crack front.
More contours should be requested in the analysis to check the
accuracy of the results. In some cases, J-integral estimates
might vary from one contour to the other. A strong or gradual
variation of these estimates indicates an error in the contour
integral definition or the mesh is not small enough to
accurately calculate the J-integrals [25]. In linear elastic
615
problems, the first and second contours are usually inaccurate.
In the present study, 8 contours were requested to ensure the
accuracy of the J-integral calculations. The percent variation of
J-integral values between the last two contours varies between
1% and 3% depending on the damage state. Therefore, the
average of the last 2 contours was used for the analysis.
In addition, the element dimensions were 14 mm  14 mm
for the stiffeners and 18 mm  15 mm for the diaphragms.
Computational time took about 3 min for each case. The
maximum principal strains for two typical crack lengths are
presented in Fig. 8.
Thereafter, based on the J-integral estimates, the stress
intensity factor was calculated with respect to the crack
length. Then, Paris Law was used to estimate the number of
loading cycles to propagate the crack by Da based on the
following equation [9,10]:
Da
¼ cDKn (1)
DN
where c and n are material constants, a is the crack length, N is
the number of cycles and K is the stress intensity factor. These
parameters were taken 2.40  102 and 3.3, respectively [26].
As seen in Fig. 9, a 100 mm crack will occur after about six
million cycles. Accordingly, the life span of the girder was
divided into 6 different stages. These periods represent the
date/time of readings of the sensor output for the post-proces-
sing. Each stage consists of 1 million loading cycle.
4. Damage detection results and analysis
4.1. Damage detection mechanism
The level crossing cumulative time counting implemented by
the proposed SWS is schematically presented in Fig. 10. The
main information that can be extracted from the sensors is the
cumulative duration of strain events shown in Fig. 10. Based on
previous studies [21–23], the sensor output can be expressed by
the following Gaussian cumulative density function (CDF):
616 archives of civil and mechanical engineering 17 (2017) 609–622

Fig. 8 – The FE results for (a) a = 20 mm (maximum principal strain = 2753 me), and (b) a = 70 mm (maximum principal
strain = 3192 me).

  
120 1 gm
CDFGaussian ðgÞ ¼ 1 þ erf pffiffiffi (2)
2 s 2

100 in which m, s, a and g refer to the mean of the strain distribu-

tion, standard deviation with respect to load and frequency,
Crack length (mm)

the total cumulative time of the applied strain measured by

80
the entire gates and the gate number, respectively. For a better
representation, the CDF can be transformed to a Probability
60 Density Function (PDF) as indicated in Fig. 11.

For the analysis, 400 sensing nodes were placed in the

40
horizontal and vertical directions on the upper half of the
central girder around the stiffener to web contact area (20  20
20 with 20 mm spacing). These nodes represent the actual
piezoelectric transducers (PZTs) that are attached to the
0 structure. The average strain of the elements located inside
0 2 4 6 8 a 10 mm diameter circle centered at the sensing node locations
Number of cycles (× 106) was used for the analysis. Fig. 12 displays the locations of the
400 data acquisition nodes on the structure. The geometry
Fig. 9 – Number of cycles vs. crack length. information is summarized in Table 4. Ten strain levels were
defined for the girder as shown in Fig. 13.

Fig. 10 – Level crossing cumulative time counting implemented by SWS.

 archives of civil and mechanical engineering 17 (2017) 609–622 617

Fig. 11 – Sensor output distribution (a) CDF fit, and (b) transformed PDF.

Fig. 12 – Locations of the sensing nodes.

4.2. Information fusion where xi, horizontal distance of the sensing node from the
sensing region center; yi, vertical distance of the sensing node
In order to evaluate the information provided by the sensing from the sensing region center; m, mean of the PDF distribu-
nodes, the m and s values for all of the 400 sensors for different tion; s, standard deviation of the PDF distribution; Zi, functions
damage scenarios were calculated. Then, a data fusion that include the information collected by the whole network of
approach was proposed for the detection of damage progres- sensors. These functions are defined as follows:
sion in the girder. The basic idea was to use original parameters
from the distribution (m and s), sensor location, and new 8 p p
features that include the effect of different groups of sensors. >
> s s av
>
>
p
Zs i ¼ i p
>
>
Subsequently, the damage state (D) was defined as follows: >
>
s STD
>
>
>
>
>
>
p
m mav
p
>
> Zpmi ¼ i p
D ¼ function of ðm; s; xi ; yi ; Zi Þ (3) >
< m STD
(4)
>
> p p
>
> s i mav
>
> Z p
¼
>
> sm p
>
i
> s av
Table 4 – Geometry of the sensing region, crack, and >
>
>
> p p
sensors. >
> m mav
>
: Zpmsi ¼ i p
Geometry (mm) s av
p p p p
Sensing region (length  height  thickness) 400  400  12.7 where mav ; mSTD ; s av ; s STD are, respectively, the average of m,
Crack length 10; . . .; 100 standard deviation of m, average of s and the standard devia-
|fflfflfflfflfflfflfflffl{zfflfflfflfflfflfflfflffl}10
Distance between the centers of 20 tion of s of all of the sensors for a damage state p. The subscript
neighboring sensors i is the number of the sensor (the data acquisition node). The
Diameter of data acquisition nodes 10 new defined features were inspired from the z-score function
618 archives of civil and mechanical engineering 17 (2017) 609–622
600
Strain Threshold Level (με)
500
400
300
200
100
0 safety of soil nailed slopes [38], prediction of soil liquefaction
1 2 3 4 5 6 7 8
Gate Number
Fig. 13 – Strain threshold levels.
[21,22]. As indicated by the equations above, the average and
standard deviation of m and s include the effect of the whole
group of sensors at the sensing location i.
4.3. Damage detection using support vector machines
A damage detection process can be treated as a pattern
recognition problem [21]. The solution is to use a classifier
which can classify structures either as damaged or healthy. To
this aim, an AI-based data fusion system was proposed for
damage detection (Fig. 14). The AI techniques are considered
as alternatives to existing traditional methods for tackling real
world problems. They determine the model structure by
automatically learning from data. AI has different well-known
branches such as artificial neural network (ANN), fuzzy
inference system (FIS), adaptive neuro-fuzzy system (ANFIS),
and support vector machines (SVMs). In the last two decades,
the AI methods have been widely used for tackling problems in
civil engineering domains such as structural engineering,
hydraulic engineering, geotechnical engineering, earthquake
engineering, etc. In this context, some well-established studies
are: calculation tensile strength and yield strength of dual
phase steels [27], modeling true stress of dynamic strain aging
Fig. 14 – Data fusion flowchart.
regime of austenitic stainless steel [28], non-destructive
investigation of corrosion current density in steel reinforced
concrete [29], prediction of the capacity of CCFT short columns
[30], prediction of the pull-off adhesion of the concrete layers
in floors [31,32], modeling of shear strength of RC deep beams
[33], structural assessment and damage identification, predic-
tion of the scour below submerged pipeline and scour depth
downstream of sills [34,35], prediction of settlement of shallow
foundations [36], maximum dry density and unconfined
compressive strength of cement stabilized soil [37], factor of
9 10 susceptibility [39], etc.
However, the proposed AI-based data fusion framework
consists of the following main stages:

Structural simulation;

Information fusion in which features that are expected to
characterize different properties of structures are extracted
from a network of sensors; and

Fusion of the clustered features. The data fusion is
performed using the AI classifier.
Using the fitted m and s parameters obtained of a Gaussian
distribution, different damage indicator features were
obtained for a specified number of sensors. The defined
features simultaneously fuse the information provided by
array of scattered sensors. The damage indicator vectors were
then used for the calibration of the classifier. Subsequently, a
validation phase was performed to check the damage
detection performance of the classifier.
Among different AI techniques, SVMs have been widely
used in the field of damage detection and structural
identification [40,41]. Some of the main advantages of the
SVM method are as follows [42]:

It has regularization parameters in order to avoid over
fitting;

It uses the kernel trick which makes the user able to design
the kernel via engineering approach;

The SVMs optimization techniques are based on a convex
optimization to avoid local minima problems. SVM training
always find a global minimum;

They provide an alternative solution when linear decisions
hyper-plans are not sufficient to separate the classes by
mapping the input data into a feature space resulting in a
nonlinear classifier;

Other AI methods suffers from local minima, however the
solution given by an SVM is always unique and global. Unlike
ANN that they are based on empirical minimizations, the
SVMs are less prone to overfitting because they use
structural minimization.
The main objective of SVM is to find an optimal hyperplane
that maximize the margin between the nearest points of
different classes as indicated in Fig. 15. SVM uses the kernel
method to map the input feature to a higher dimensional
space where the data becomes separable and then find an
optimal hyperplane that maximize the margin between
classes. This concept is called the 'kernel trick' as it illustrated
in Fig. 16.
 archives of civil and mechanical engineering 17 (2017) 609–622 619

formulation of the SVM method is defined by the flowing

maximization problem:

Xm
1X m
maxa 2 Rm WðaÞ ¼ ai  a a y y kðxi ; xj Þ (5)
i¼1
2 i;j¼1 i j i j

subject to the constraints of:

8
< 0m< ai < C
X and i ¼ 1 to m (6)
: ai yi ¼ 0
i¼1

Fig. 15 – Margin and separating hyperplane illustration for

the SVM method.
The SVM classification process can be defined as follows:
Find a decision function f that maps inputs x into labels y,
where (x,y) are the input-output pairs. By identifying the
function f, a new label xn could be classified as yn based on the
sign of the discriminant function f(xn). The mathematical
The set (xi,yi) are the input pairs (data, class label), m is the
number of training samples, and the a is a representative
vector of the relative importance of training sample i in the
classification of a new sample. C is the regularization
parameter (C > 0) and k is the kernel function. Finally, in
order to classify a new data xn, the discriminant function f is
defined as follows:
X
f ðxn Þ ¼ ai yi kðxn ; xi Þ þ b (7)
i¼1:m

2 Class 1 5
Class 2
1.5 Class 1
4 Class 2
1 Seperating
hyperplane
3
z = x 2 + y2

0.5

0 2
y

-0.5
1

-1
0
-1.5 5
r = 1.2
0
-2 4
0 2
-2 -1 0 1 2 -5 -4 -2
y
x x

Fig. 16 – The mapping process using the transformation function: z ¼ x2 þ y2 : (a) non-linearly separable 2D data, and (b)
linearly separable 3D data.

Fig. 17 – The SVM performance on the: (a) testing data, and (b) validation data.
620 archives of civil and mechanical engineering 17 (2017) 609–622

Fig. 18 – Set of sensors with the highest detection rate.

Typical SVM kernels are: linear kernel, polynomial, radial Number of data correctly classified
basis function (RBF) and sigmoid [40,41]. For the current study, Detection ¼ (9)
Total number of data
the RBF kernel was used. The RBF kernel function is given by
the following expression: The total number of data points was 2400 (400 sensing
nodes  6 damage states). The labels to be classified were 6
! damage states (D1, . . ., D6) each represented the girder
jjxyjj2 condition after 1 million cycles. The data was divided into
kðx; yÞ ¼ exp  (8)
2d2 three subsets: 70% for training, 15% for testing and 15% for
validation. The SVM algorithm was run for different sensor
where d is the RBF parameter to be specified along with the
configurations by adding the number of sensors from 1 (single
SVM parameter C. Detection rate was used to evaluate the
sensor) to 400 (the entire sensor network). The optimal
classifier performance as follow:

Fig. 19 – Confusion matrixes for the best sensor configuration.

 archives of civil and mechanical engineering 17 (2017) 609–622
parameters d and C were found through an extensive searching
algorithm.
Fig. 17 displays the results for the testing and validation
sets for different combination of the input parameters. The
results indicate that the new Z features have good perfor-
mance for both of the testing and validation sets. Using only
the individual sensing features m and s results in a very low
performance (around 20–33%). An interesting observation
from Fig. 17 is that both curves present the highest values
for a number of sensors below 30. This indicates that the
damage is located in the upper area of the web where the first
30 sensors are located. The best detection rate for the training,
testing and validation was obtained using Sensing nodes 4–16
as shown in Fig. 18.
The optimal value of the d and C parameters were 0.9, 100,
respectively. The best performance of the classifier was:
- Training = 83%
- Testing = 82%
- Validation = 85%
Fig. 19 displays the confusion matrixes for the best sensor
configuration. As seen in Fig. 19, class 2 is misclassified as class
1 in all cases. This indicates that there is not important
variation in strain amplitude between 10 mm and 20 mm crack
lengths. Therefore, the classifier cannot differentiate between
the two classes. However, for the other damage states, the
detection performance is satisfactory.
5. Conclusions
Many older steel girder bridges are affected by the distortion-
induced fatigue cracking. In this study, an SVM-based
procedure has been proposed for the detection of fatigue
cracking in steel girders based on the data of the SWS. The
sensor operates by harvesting the energy from the host
structure. It can record the cumulative durations of the applied
strain signal at predefined threshold levels. However, the
sensor output is compressed into a single histogram. This
makes the interpretation of the data very challenging.
Moreover, detection of the fatigue cracking stages is a
complicated task due to notable changes of the strain patterns
during crack propagation. In order to evaluate the perfor-
mance of the sensor, a numerical study was carried out on a
typical girder. The damage detection process was divided into
different phases: First, the data was collected from the FE
strains and the features were extracted to define individual
damage indicators (PDF parameters of m and s). Thereafter, a
data fusion model was defined based on the previously
extracted features and the sensors group effect. Finally, the
new features were inputted to SVM classifiers. The results
indicate that the classification performance was increased
using the data fusion model. It was observed that the SVM
models can accurately classify most of the damage stages,
specifically for cracks larger than 10 mm. Besides, tracking
the performance of the SVM models gives an insight into the
damage location. Although the FE simulation results
still remain the standard for most of the elasticity problems,
verification of the proposed approach with laboratory or
621
full-scale experiments would be an interesting topic for future
study.
Acknowledgment
The presented work is supported by a research grant from the
Federal Highway Administration (FHWA) (DTFH61-13-C-
00015).
Funding
The U.S. Federal Highway Administration (FHWA) (DTFH61-
13-C-00015).
references
[1] J.W. Fisher, Fatigue and Fracture in Steel Bridges: Case
Studies, John Wiley & Sons. Inc., 1984.
[2] Y. Zhao, W.M.K. Roddis, Fatigue Prone Steel Bridge Details:
Investigation and Recommended Repairs, K-TRAN: KU-99-2,
Final Report, Kansas Department of Transportation, Topeka,
KS, 2004.
[3] D. Juntunen, Study of Michigan's Link Plate and Pin
Assembly. Michigan Department of Transportation (MDOT).
Research Report No. R-1358, 1998.
[4] J.W. Fisher, D.R. Mertz, Hundreds of bridges-thousands of
cracks, civil engineering, ASCE April (1985) 64–76, Civil
Engineering, ASCE 55 (4) (1985) 64–67.
[5] M.A. Elewa, Influence of Secondary Components on the
Serviceability of Steel Girder Highway Bridges, Ph. D.
Dissertation, Michigan State University, East Lansing, MI,
2004.
[6] R.J. Dexter, J.M. Ocel, Manual for Repair and Retrofit of Fatigue
Cracks in Steel Bridges, Report: FHWA-IF-13-020, Federal
Highway Administration (FHWA), Minnesota, 2013.
[7] J.M. Stallings, T.E. Cousins, J.W. Tedesco, Fatigue of
Diaphragm-Girder Connections. RP 930-307, The Alabama
Department of Transportation, Montgomery, Alabama, 1996.
[8] P. Jiao, W. Borchani, H. Hasni, A.H. Alavi, N. Lajnef, Post-
buckling response of non-uniform cross-section bilaterally
constrained beams, Mechanics Research Communication 78
(Part A) (2016) 42–50. , http://dx.doi.org/10.1016/j.mechrescom.
2016.09.012.
[9] M.H. El Haddad, N.E. Dowling, T.H. Topper, K.N. Smith, J-
integral applications for short fatigue cracks at notches,
International Journal of Fracture 16 (1) (1980) 15–30.
[10] N. Pugona, M. Ciavarellab, P. Cornettia, A. Carpinteria, A
generalized Paris Law for fatigue crack growth, Journal of the
Mechanics and Physics of Solids 54 (2006) 1333–1349.
[11] J.P. Lynch, K.J. Loh, A summary review of wireless sensors
and sensor networks for structural health monitoring, Shock
and Vibration Digest 38 (2006) 91–128.
[12] H. Salehi, S. Das, S. Chakrabartty, S. Biswas, R. Burgueno,
Structural assessment and damage identification algorithms
using binary data, in: ASME 2015 Conference on Smart
Materials, Adaptive Structures and Intelligent Systems,
2015, 1–10.
[13] S. Das, H. Salehi, Y. Shi, S. Chakrabartty, R. Burgueno, S. Biswas,
Towards packet-less ultrasonic sensor networks for energy-
harvesting structures, Computer Communications (2017),
http://dx.doi.org/10.1016/j.comcom.2016.11.001 (in press).
[14] W. Borchani, K. Aono, N. Lajnef, S. Chakrabartty, Monitoring
of post-operative bone healing using smart trauma-fixation
device with integrated self-powered Piezo-floating-gate
622 archives of civil and mechanical engineering 17 (2017) 609–622
sensors, IEEE Transactions on Biomedical Engineering 63 (7)
(2016) 1463–1472.
[15] N. Elvin, A. Elvin, D.H. Choi, A self-powered damage detection
sensor, Journal of Strain Analysis 38 (2) (2003) 115–124.
[16] N. Lajnef, M. Rhimi, K. Chatti, L. Mhamdi, Toward an
integrated smart sensing system and data interpretation
techniques for pavement fatigue monitoring, Computer-
Aided Civil and Infrastructure Engineering 26 (2011) 513–523.
[17] C. Huang, N. Lajnef, S. Chakrabartty, Self-calibration and
characterization of self-powered floating-gate usage
monitors with single electron per second operational limit,
IEEE Transactions on Biomedical Circuits and Systems 57
(2010) 556–567.
[18] C. Alippi, C. Galperti, An adaptive system for optimal solar
energy harvesting in wireless sensor network nodes, IEEE
Transactions on Circuits and Systems 55 (6) (2008) 1742–1750.
[19] B.C. Yen, J.H. Lang, A variable-capacitance vibration-to-
electric energy harvester, IEEE Transactions on Circuits and
Systems 53 (2) (2008) 288–295.
[20] N. Lajnef, K. Chatti, S. Chakrabartty, M. Rhimi, P. Sarkar,
Smart Pavement Monitoring System, Report: FHWA-HRT-12-
072, Federal Highway Administration (FHWA), Washington,
DC, 2013.
[21] A.H. Alavi, H. Hasni, N. Lajnef, K. Chatti, F. Faridazar, An
intelligent structural damage detection approach based on
self-powered wireless sensor data, Automation in
Construction 62 (2016) 24–44.
[22] A.H. Alavi, H. Hasni, N. Lajnef, K. Chatti, F. Faridazar, Damage
detection using self-powered wireless sensor data: an
evolutionary approach, Measurement 82 (2016) 254–283.
[23] A.H. Alavi, H. Hasni, N. Lajnef, K. Chatti, Continuous health
monitoring of pavement systems using smart sensing
technology, Construction and Building Materials 114 (2016)
719–736.
[24] A.H. Alavi, H. Hasni, N. Lajnef, K. Chatti, Damage growth
detection in steel plates: numerical and experimental
studies, Engineering Structures 128 (2016) 124–138.
[25] ABAQUS Analysis User's Manual 6.11.
[26] P.J.G. Schreurs, Fracture Mechanics, Technical Report,
Eindhoven University of Technology, Netherland, 2012.
[27] S. Krajewski, J. Nowacki, Dual-phase steels microstructure
and properties consideration based on artificial intelligence
techniques, Archives of Civil and Mechanical Engineering 14
(2) (2014) 278–286.
[28] A. Garg, K. Tai, A.K. Gupta, A modified multi-gene genetic
programming approach for modelling true stress of dynamic
strain aging regime of austenitic stainless steel 304,
Meccanica 49 (5) (2014) 1193–1209.
[29] L. Sadowski, Non-destructive investigation of corrosion
current density in steel reinforced concrete by artificial
neural networks, Archives of Civil and Mechanical
Engineering 13 (2013) 104–111.
[30] M. Ahmadi, H. Naderpour, A. Kheyroddin, Utilization of
artificial neural networks to prediction of the capacity of
CCFT short columns subject to short term axial load, Archives
of Civil and Mechanical Engineering 14 (3) (2014) 510–517.
[31] L. Sadowski, J. Hoła, Neural prediction of the pull-off
adhesion of the concrete layers in floors on the basis of
nondestructive tests, Procedia Engineering 57 (2013) 986–995.
[32] Ł. Sadowski, Non-destructive evaluation of the pull-off
adhesion of concrete floor layers using RBF neural network,
Journal of Civil Engineering and Management 19 (4) (2013)
550–560.
[33] A.H. Gandomi, G.J. Yun, A.H. Alavi, An evolutionary approach
for modeling of shear strength of RC deep beams, Materials
and Structures 46 (12) (2013) 2109–2119.
[34] H.M. Azamathulla, Gene expression programming for
prediction of scour depth downstream of sills, Journal of
Hydrology 460–461 (2012) 156–159.
[35] H.M. Azamathulla, A. Guven, Y.K. Demir, Linear genetic
programming to scour below submerged pipeline, Ocean
Engineering 38 (8–9) (2011) 995–1000.
[36] P. Samui, Support vector machine applied to settlement of
shallow foundations on cohesionless soils, Computers and
Geotechnics 35 (3) (2008) 419–427.
[37] S.K. Das, P. Samui, A.K. Sabat, Application of artificial
intelligence to maximum dry density and unconfined
compressive strength of cement stabilized soil, Geotechnical
and Geological Engineering 29 (3) (2011) 329–342.
[38] A. Garg, A. Garg, K. Tai, S. Sreedeep, An integrated SRM-
multi-gene genetic programming approach for prediction of
factor of safety of 3-D soil nailed slopes, Engineering
Applications of Artificial Intelligence 30 (2014) 30–40.
[39] P. Samui, T.G. Sitharam, Machine learning modelling for
predicting soil liquefaction susceptibility, Natural Hazards
and Earth System Sciences 11 (2011) 1–9.
[40] K. Worden, A.J. Lane, Damage identification using support
vector machines, Smart Materials and Structures 10 (2001)
540–547.
[41] S.B. Satpal, A. Guha, S. Banerjee, Damage identification in
aluminum beams using support vector machine: numerical
and experimental studies, Structural Control and Health
Monitoring 23 (3) (2016) 446–457.
[42] C.J.C. Burges, A tutorial on support vector machines for
pattern recognition, Data Mining and Knowledge Discovery 2
(2) (1998) 121–167.