See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/376255497

A dynamic stiffness-based two-step method for damage identification of

joints in hinged slab bridges using support vector machine

Article in Structures · December 2023

DOI: 10.1016/j.istruc.2023.105657

CITATIONS READS

0 40

7 authors, including:

Jiawang Zhan Chuang Wang

Beijing Jiaotong University Beijing Jiaotong University
28 PUBLICATIONS 355 CITATIONS 11 PUBLICATIONS 33 CITATIONS

SEE PROFILE SEE PROFILE

Zhang Fei Qikai Sun

Beijing Jiaotong University Beijing Jiaotong University
10 PUBLICATIONS 57 CITATIONS 20 PUBLICATIONS 82 CITATIONS

SEE PROFILE SEE PROFILE

All content following this page was uploaded by Chuang Wang on 21 December 2023.

The user has requested enhancement of the downloaded file.

 Structures 58 (2023) 105657

Contents lists available at ScienceDirect

Structures
journal homepage: www.elsevier.com/locate/structures

A dynamic stiffness-based two-step method for damage identification of

joints in hinged slab bridges using support vector machine
Jingqi Zhao, Jiawang Zhan *, Chuang Wang, Fei Zhang, Zhihang Wang, Qikai Sun, Xinxiang Xu
School of Civil Engineering, Beijing Jiaotong University, Beijing 100044, China

A R T I C L E I N F O A B S T R A C T

Keywords: As critical components of hinged slab bridges, joints are of great importance to the bearing capacity and
Damage identification serviceability of bridges. The existing model-based joint damage detection methods are based on precise finite
Hinge joints element models, and their identification accuracy is significantly affected by construction uncertainty. In this
Dynamic stiffness
paper, dynamic stiffness is introduced into joint damage detection and a two-step framework is proposed for the
Bayesian optimization
Support vector machine
damage localization and quantification of joints in hinged slab bridges. The dynamic stiffness-based data-driven
Field measurement method is suitable for the evaluation of bridges with incomplete information or bridges with serious damage.
Firstly, the second-order difference of dynamic stiffness (SDDS) calculated by slabs’ responses is used an indi­
cator to locate the damage. Secondly, Bayesian optimization is utilized to optimize the model hyperparameters of
the support vector machine (SVM), where the dynamic stiffness serves as the characteristic vector for training the
optimized SVM network. The applicability of this method was numerically validated using a grillage model of a
hinged slab bridge, taking into consideration factors such as noise interference, impact location, sensor place­
ment, and impact force characteristics. Additionally, on-site experiments were conducted on an in-service bridge
to verify the effectiveness of the proposed method. Analysis results demonstrate that the proposed two-step
method can properly evaluate joint damage with less dependence on the impact locations, force amplitude
and environmental factors, and thus can be used for rapid damage identification of in-service hinged slab bridges
and quantitative evaluation of repair effect.

1. Introduction Some studies evaluated the service conditions of joints in hinged slab
bridges by using statistical analysis, physical models, or vision-based
Hinge joints, as the important components of prefabricated hinged measurements, or conducting a series of load tests with large amounts
slab bridges, enable all slabs to function as an integrity and distribute the of sensors [5–8]. These methods cannot achieve rapid deployment
load uniformly. In the United States, 55% of hinged slab bridges utilize during on-site testing and their application in real bridges also faces
hinge joints and transverse ties to transfer load [1], which increases challenges. In recent years, the vibration-based damage identification
construction costs but effectively mitigates hinge joint cracking. In some methods have received significant attention due to their advantages of
countries, transverse posttensioning ties are not adopted in the design of simplicity, rapidity, non-destructiveness to the structure and no
hinge joints [2], so prefabricated hinged slab bridges are prone to a disruption of the traffic [9–11]. Dan et al. proposed a multi-beam me­
variety of damage. The hinge joints of slab bridges are particularly chanical model connected by distributed springs for prefabricated beam
crucial to the transverse collaborative working performance, as their bridges, and two indices based on transverse modal shapes for joint
failure will directly lead to single beam bearing phenomenon. Tradi­ damage identification and the transverse collaborative working perfor­
tional visual inspection based on the appearance of hinge joints is highly mance monitoring are constructed [12]. In a later study by Dan et al., the
subjective, time-consuming and laborious [3]. The detection of these lateral collaborative working capability of hinged slab bridge based on
joints is only possible when there are visible surface cracks, concrete the dynamic strain correlation coefficient index was assessed and the
spalling or corrosion [2,4]. It is urgent to explore a rapid and accurate effectiveness and accuracy of the method was proved through numerical
damage identification method for the hinge joints of prefabricated simulation analysis and real bridge monitoring data analysis [13]. Zhan
hinged slab bridges.Table 1. et al. proposed the coherence function of the response spectrum index

* Corresponding author.
E-mail address: [email protected] (J. Zhan).

https://doi.org/10.1016/j.istruc.2023.105657
Received 13 July 2023; Received in revised form 23 November 2023; Accepted 26 November 2023
Available online 5 December 2023
2352-0124/© 2023 Institution of Structural Engineers. Published by Elsevier Ltd. All rights reserved.
J. Zhao et al. Structures 58 (2023) 105657

Table 1 dynamic stiffness (SDDS) index obtained from the impact responses of
Hyperparameter computing framework based on Bayesian optimization. all slabs and preliminarily evaluate the damage degree combined with
Input: number of initialization point n0 , surrogate model g(x), acquisition function the Bayesian optimization-support vector machine (BO-SVM) algorithm.
α(x), number of iterations epoch To verify the performance of the proposed method, a comprehensive
step1: Initialize the data points X∗ = (x1 , x2 , …, xn0 − 1 ) randomly, obtain their numerical analysis was conducted, including different damage cases and
corresponding function value f(X∗ ), and then obtain the initial point set {X∗ , f(X∗ )}
different excitation conditions. Finally, the field test was performed on a
and let n = n0 .
while n < epoch do case-study bridge to verify the feasibility of the method.
step2: Initialize the prior distribution of the surrogate model through Gaussian process
and build the surrogate model g(x). 2. Methodology
step3: Get the next test point xt to maximize the acquisition function α(x).
step4: Get the new sample point xt , obtain the result label f(xt ), and then add {xt , f(xt )
2.1. Damage localization using dynamic stiffness
} to the sampling point set. Update the surrogate model to obtain the posterior
distribution and increase n.
end 2.1.1. Dynamic stiffness calculation method
output: Get the global optimal hyperparameters including penalty parameter C and The transient response method is also known as the mechanical
kernel function K.
admittance method. The mechanical admittance spectrum Y is defined
as:
under vehicle load to locate and quantitatively analyze the joint damage V(f )
by combining with the frequency to construct an objective function and Y= (1)
F(f )
update the bridge model [14]. Zhang et al. presented an approach to
determining the damage location for hinge joints using mode shapes where V(f) and F(f) are the Fourier spectra of the velocity signal and
extracted from moving vehicle responses while taking into account the force signal respectively; f denotes the frequency. Fig. 1 illustrates the
effect of road irregularity and vehicle speed [4]. These methods rely on theoretical amplitude of admittance for a foundation tested using the
the finite element model and the detailed structural information of the transient response method [23]. In the low-frequency range f < f0
bridge to be tested in the field test, and cannot realize the large-scale (typically below 30 Hz for foundations), the ideal mobility response
detection of hinge joints and the detection of old bridges with incom­ varies linearly with the frequency due to the limited impact energy
plete original data. provided by the drop equipment. The slope k of the velocity admittance
The transient response method, a data-based non-destructive dy­ amplitude curve in the linear segment defines the compliance or flexi­
namic testing method, extracts the damage information from the bility of the area around the test point for a normalized force input [24].
structural frequency spectrum and has the advantages of rapidity and As shown in Fig. 1, k remains nearly constant within the mechanical
low cost. It has been widely used in evaluating the performance of bridge admittance’s linearity range before the first-order frequency.
substructures such as pile foundations. In previous studies, a static and
2πf 2π
dynamic contrast test have been commonly utilized to determine the DS = = (2)
static stiffness and estimate the quality, integrity, and bearing capacity |V(f )/F(f )| k
of the substructure [15,16]. However, vibration signals may be affected In engineering, it is more convenient to obtain the acceleration signal
by environmental noise, leading to inaccuracy of dynamic stiffness. It of a structure rather than the velocity signal. As a result, the acceleration
makes the application of dynamic stiffness limited to overall evaluation, response is often used to deduce the dynamic stiffness of the structure.
rather than quantitative evaluation of structural damage. For pre­ The relationship between the admittance function of velocity and ac­
fabricated assembled slab bridges, the dynamic stiffness of all slabs will celeration in frequency domain can be written as follows:
change with the joint damage condition. Therefore, instead of directly
using the dynamic stiffness value, this paper intends to locate the joint V(f )
=
1 A(f )
(3)
damage by establishing the relationship between the dynamic stiffness F(f ) j2πf F(f )
of the slabs on both sides of the joint.
where j denotes imaginary unit.
As an extensively employed technique in the supervised learning,
support vector machine has been applied to solving pattern recognition
2.1.2. SDDS index
problems, such as text processing, face recognition, and fault diagnosis.
Dynamic stiffness represents the deformation capacity of the struc­
Proposed by Vapnik [17], support vector machine is a data mining
ture under dynamic load. Prior studies concerning pile foundations have
method based on statistical learning theory and good at solving prob­
lems of small samples, nonlinearity, and highdimensional space. Some
studies compared the parallel processing ability, fault tolerance,
robustness, and learning association ability of SVM with other methods.
Finotti et al. [18] extracted original acceleration signals as model input
to identify structural changes. By comparing the results, it is found that
support vector machine has higher performance than neural network for
simply supported beams. Alves et al. [19] and Sharafi et al. [20]
compared multiple machine learning classification algorithms and
pointed out that support vector machine is more robust, higher in ac­
curacy and lower in error. Bulut et al. [21] and Pan et al. [22] considered
the influence of noise and environmental conditions on the classification
results and analyzed the advantages of SVM with different kernel
functions in identifying damage. It was proved that support vector ma­
chine has high accuracy even under noise interference and limited
operating conditions.
In this paper, the dynamic stiffness detection method is applied to the
bridge superstructure. A two-step method is proposed to accurately
locate the joint damage by directly using second-order difference of
Fig. 1. Theoretical mobility response spectrum.

2
J. Zhao et al. Structures 58 (2023) 105657

focused on the amplitude-frequency and phase-frequency response 2.2. Damage quantification based on machine learning
characteristics of the system. For multi-sheet prefabricated slab bridges,
due to the effect of transverse connections, not only does the stiffness of 2.2.1. Classification model using support vector machine
a single slab decrease when the joint is damaged, but the stiffness and Support vector machine aims to maximize the margin between two
transverse collaborative working performance of the whole bridge will sets of data in a space, minimize misclassification, and construct a hy­
be adjusted and redistributed. perplane (classifier) with maximum margin. The support vectors, which
Fig. 2a shows the acceleration response of each slab when the side are the edge data points of the two sets of data, are utilized to classify
slab is impacted. By extracting a portion of the data in the red area and maximize the optimization function of the Euclidean norm of the
(Fig. 2b), it is evident that the response between slabs roughly follows a hyperplane direction vector. This maximizes the blank area on both
linear relationship which can be expressed as: sides of the hyperplane and establishes the decision boundaries between
the two classes [25]. In terms of nonlinear problems (linear insepara­
Ai (t) = mA1 (t) (4)
bility), the training data will be mapped to a higher-dimensional space
where Ai(t) denotes the acceleration response of the i-th slab (i = 1,…, through the kernel function to achieve linear separability.
N); N is the total number of slabs; m is a proportional coefficient. By It is assumed that the training set is {xi , yi }ni=1 , where xi is the input
performing Fourier transform on both sides of Eq. (4), the linear rela­ sample and yi ∈ { +1, − 1} is the classification label. The construction of
tionship of the response between slabs in frequency domain can be an optimal hyperplane to separate the samples can be achieved through
expressed as: the utilization of weight vectors w= (w1,w2,…,wn) and bias b, provided
∫∞ ∫∞ the dataset is linearly separable. The objective function of the maximum
Ai (ω) = F [Ai (t)] = Ai (t)e− iωt dt = mA1 (t)e− iωt dt = mA1 (ω) (5) margin classifier can be defined by the concept of function margin,
− ∞ − ∞ which is:
Therefore, the corresponding distribution of dynamic stiffness in the 1 T
transverse direction should exhibit an oblique line under the given force min w w s.t. yi (wxi + b) ≥ 1 (7)
2
spectrum. If the joint is damaged, the response transmission path
Assuming that the existence of outliers within the dataset results
changes, thus causing a substantial increase in dynamic stiffness dif­
from deviations caused by noise and that the impracticality of achieving
ferences between slabs on both sides of the damaged joint. At the two
a hyperplane capable of perfectly distinguishing all sample points is
inflection points on the dynamic stiffness curve shown in Fig. 3, the
recognized, the implementation of a dual learning method becomes
second-order difference of dynamic stiffness (SDDS) exhibits a positive
necessary. By incorporating slack variable ξi and penalty parameter C
value and a negative value, respectively. Based on the above analysis,
into the objective function to allow for misclassification, the constraints
the SDDS index is proposed and defined as follows:
can be relaxed. Thus, a new objective function can be defined as follows:
⃒ ⃒ ⃒ ⃒
SDDSj = ⃒DS ̃ j ⃒ − ⃒DS
̃ j+1 − DS ̃ j− 1 ⃒
̃ j − DS (6) ∑n
1 T
min w w+C ξi s.t. yi (wxi + b) ≥ 1 − ξi , ξi ≥ 0 (8)
2
where DS
̃ j denotes the dynamic stiffness of the j-th slab (j = 1,…, N-1) in i=1

the normalized dynamic stiffness vector. When j = 1, set DS ̃ 0 = 0. The If the dataset cannot be separated linearly in the low-dimensional
SDDS index reflects the response difference of adjacent slabs under space, a nonlinear transformation is required to map the input space χ
impact and represents the transverse collaborative working performance (discrete set) to the high-dimensional feature space H (Hilbert space),
of the bridge. If the SDDS index approaches or falls below 0, it indicates and the kernel function is then applied to calculating the inner product
that the difference between the two adjacent slabs is light, and the joint of the vectors after mapping to the high-dimensional space. The gener­
between two slabs is undamaged. Conversely, a value notably greater alization ability and classification accuracy of support vector machine
than 0 suggests ineffectual cooperation between adjacent slabs, and the depend on penalty parameter C and kernel function K. An appropriate
joint between the two slabs is damaged. kernel function can improve the computational efficiency of the model
in the high-dimensional space and the penalty parameter is used to
control the range of misclassification during model training. The rise of
the penalty level will reduce the number of support vectors but increase
the training time.

Fig. 2. Linear relationship of the response between slabs.

3
J. Zhao et al. Structures 58 (2023) 105657

Fig. 3. Dynamic stiffness of all slabs in the case of damaged joint 4#.

2.2.2. Determination of support vector machine hyperparameters by are not efficient for parameter optimization due to their lengthy
Bayesian optimization computation time. Genetic algorithm and particle swarm optimization
The two hyperparameters, penalty parameter C and kernel function algorithm belong to population optimization algorithms, which require
K, define the properties and training process of the model. Many users a sufficient number of initial sample points. These algorithms can be
blindly adjust these parameters based on the results obtained. When the time-consuming due to the high cost of calculating the objective func­
kernel function type cannot be determined through experience, the best tion and may not always result in optimal efficiency. Bayesian optimi­
option is to use an optimization algorithm. Common optimization al­ zation is regarded as the most advanced and classical approach to black-
gorithms include grid search, random search, genetic algorithm, particle box function estimation. The method can find a comparatively optimal
swarm optimization and Bayesian optimization. Grid search and random solution within a limited time frame, making it well-suited for
search methods aim to traverse all points in the large parameter space addressing the complex relationships among variables in optimization
and then return the optimal loss function value. However, these methods and for resolving the objective function, yet arduous to express in

Fig. 4. Scheme of the proposed methodology.

4
J. Zhao et al. Structures 58 (2023) 105657

mathematical formulas or programs. This technique is not susceptible to index is defined as the ratio of crack height to the whole height of the
overfitting and has a low likelihood of converging to a local optimal cross section as shown in Eq. (9). The damage index is defined as follows:
solution.
hc
ηj = (9)
h0
2.3. Damage identification process by the proposed method
where hc and h0 denote the crack height and cross section height,
The damage identification method for the joint based on dynamic respectively, and j denotes joint number. The section characteristics of
stiffness of bridge is proposed as shown in Fig. 4. The following three the virtual cross-beams represent the transverse stiffness of the hinge
steps are summarized to identify the location of joint damage and pre­ joint and the effective working height is h0-hc. The transverse bending
liminarily determine the degree of damage: stiffness of hinge joints can be described as follows:

(1) Apply impact load F(t) on the side slab of the hinged slab bridge Ebh0 3
D0 = EI0 = (10)
and collect the response A(t) at the mid-span of each slab. 12
(2) Calculate dynamic stiffness by the spectra of force signal F(f) and
Eb(h0 − hc )3
acceleration responses of the bridge A(f) after FFT. Dd = EId = (11)
(3) Propose a two-step damage identification method, including 12
damage localization and damage quantification by SDDS index where E is the elastic modulus of concrete; D0 and Dd denote the
and BO-SVM algorithm respectively. transverse bending stiffness of the undamaged and damaged joints,
respectively. The transverse bending stiffness reduction index ηd is
3. Numerical study calculated as follows:
( )/
3.1. Model establishment of hinged slab bridge D Ebh0 3 1 − η3 12
ηd = 1 − d = 1 − / = η3 (12)
D0 Ebh0 3 12
In this study, a grillage model of a hinged slab bridge composed of 8
slabs and 7 funnel-shaped hinge joints is built in the ANSYS software.
The cross-section of the bridge is shown in Fig. 5a. The bridge has a span 3.2. Dynamic stiffness calculation of slabs
of 8 m and a width and height of 8 m and 0.3 m, respectively. The
pavement layer is 0.15 m in thickness and slabs are made of C40 con­ Transient response method can release high energy to the structure
crete with an elastic modulus of 32500 MPa. The material data in the in a short time by applying a transient impact force, which generates a
linear elastic stage is used for subsequent calculation. The beams are broadband impact force spectrum that can improve the testing accuracy.
simulated as Euler beam elements and each slab is divided longitudi­ However, the method requires consideration of noise interference,
nally into 16 beam elements with a length of 0.5 m. Virtual cross-beams impact pulse energy, and interaction time. If the excitation force spec­
are used to simulate the transverse connection between slabs, with a trum is too wide, the average energy allocated at each frequency will
spacing of 0.5 m (Fig. 5b). The 0.15 m thick concrete pavement is decrease, and the signal-to-noise ratio of the collected signal will
considered as a part of bridge, i.e., the heights of the beams and the decrease as well.
transverse connections are increased from 0.3 m and 0.1m to 0.45m and In this paper, the impact force pulse has a width of 20 ms with a peak
0.25 m, respectively. of 105 N as shown in Fig. 6a, and the effective frequency range of
It is assumed that the hinge joints cracks do not change the mass and excitation is 0–200 Hz (Fig. 6b). The sampling frequency is 400 Hz, and
elastic modulus of concrete. To quantify the damage state, a damage the sampling time is 10 s

Fig. 5. Analyzed bridge: a) cross section of bridge, b) grillage model.

5
J. Zhao et al. Structures 58 (2023) 105657

Fig. 6. Impact force: a) time history, b) power spectrum.

Based on the structural and stress characteristics of the bridge su­

perstructure, the site layout for impact location and sensor placement is
proposed as shown in Fig. 7. In order to achieve a rapid layout, as a non-
disruptive bridge inspection technology, sensors are predominantly
positioned at the mid-span of each slab bottom. If a more detailed
evaluation is required, additional tests can also be conducted at 1/4 span
and 3/4 span. The excitation signal shown in Fig. 6 is applied to the No.
1 side slab and the vertical acceleration signal of 8 slabs within 10 s is
collected, as shown in Fig. 8.
The acceleration admittance of the bridge is obtained through the
Fourier transform of the acceleration signal and the impact force signal.
The velocity admittance spectrum in Fig. 9a is calculated using Eq. (3)
and the first three frequencies are 11.4 Hz, 19.4 Hz and 35.0 Hz,
respectively. The low-frequency range is defined as 0–5 Hz for this
bridge based on its fundamental frequency. The low-frequency velocity
admittance is shown in Fig. 9b.
Fig. 9b shows that the mobility of each slab changes nearly linearly
with frequency, which corresponds to the characteristics of an ideal
velocity admittance curve. The slope of the velocity admittance curve
for all slabs in the low frequency is determined through linear curve
fitting. Taking No. 2 slab as an example, the correlation coefficient R2 in
the frequency band of 0–5 Hz exceeds 0.99, indicating that the slope of Fig. 8. Time-domain response of the slabs.
the curve at 0–5 Hz is close to a constant, making this frequency range
appropriate (Fig. 10a). Frequencies such as 0–2 Hz or even 0–1 Hz are interference, resulting in a decrease in fitting effectiveness. Conse­
not selected due to limitations in the sampling frequency. In field quently, for the numerical model in this paper, the frequency range of
measurements, too few sampling points can amplify the impact of noise 0–5 Hz is thought as an empirical choice.
Based on the slope of the fitting curve, the approximate dynamic
stiffness for each slab is calculated, as shown in Fig. 10b. The transverse
distribution of dynamic stiffness exhibits an upward trend due to the
impact location of No.1 slab.

3.3. Damage localization by SDDS index

In engineering, cracks are likely to propagate throughout the longi­

tudinal section of the hinged slab bridge once the joint is damaged.
Therefore, the joint is simulated as a significant damage in this paper,
and the damage index η of all damaged joints is set to 80%. At this
damage level, the reduction index of transverse bending stiffness
ηd= 0.83 = 0.512 shows that the stiffness decreases to half of the initial
value. This section aims to examine the effectiveness of the SDDS index
in identifying joint damage by using single-damaged and double-
damaged scenarios as examples.
The single-damaged joint is simulated in the numerical model and
the SDDS of each joint under different damaged scenarios is shown in
Fig. 11. The joints with SDDS greater than 0 and less than 0 significantly
correspond to the two inflection points in the DS curve, respectively
(Fig. 3). It can be seen that the SDDS of the intact joint is close to 0 or less
Fig. 7. The site layout of vibration test. than 0, while the SDDS of the damaged joint is obviously greater than 0,

6
J. Zhao et al. Structures 58 (2023) 105657

Fig. 9. Velocity admittance curve: a) containing three-order natural frequencies, b) at low-frequency range.

Fig. 10. Velocity admittance curve: a) linear fitting of admittance in low frequency band, b) linear relationship of dynamic stiffness between slabs.

Fig. 11. SDDS index at different single-damage locations: a) 2#, b) 3#, c) 4#, d) 7#.

7
J. Zhao et al. Structures 58 (2023) 105657

so that the location of damaged joint can be identified. 3.4. Sensitivity analysis
For the identification of double-damaged joints locations, the dam­
age of adjacent joints (Fig. 12a, Fig. 12b), and non-adjacent joints 3.4.1. Effect of impact location and sensor placement
(Fig. 12c, Fig. 12d) are considered. Fig. 12 shows that the identification The impact location is a pivotal factor influencing the dynamic
rule for single-damaged joint is still applicable. In the cases of non- stiffness distribution of a bridge. The primary modes of vibration in the
adjacent damaged joints, the SDDS of the subsequent numbered joint first three orders for a simply-supported beam bridge are characterized
(the nearest joint to the impact location designated as joint 1#) obvi­ by longitudinal vertical bending and transverse torsion. When selecting
ously falls below 0, corresponding to another inflection point in the the excitation location, it is imperative to consider providing more en­
dynamic stiffness distribution diagram. When the SDDS value is small, it ergy to the bridge and exciting the first several vibration modes. To
can assist in identifying the damage location such as joint 2# in Fig. 12c pinpoint the optimal impact location, simulations are conducted for
and joint 1# in Fig. 12f. impacts at the intermediate slab and the side slab.
Fig. 13 shows the SDDS of joints at varying degrees of single damage. Fig. 14 illustrates the results of the SDDS indices obtained at various
For a certain damage, as the damage degree increases, so does the SDDS impact locations when joint 3# is damaged. The findings suggest that
value of the damaged joints. Based on the above analysis, the SDDS when the mid-span of No. 4 to No. 8 slabs is impacted, the SDDS indices
index is highly sensitive to transverse connection damage. This makes it become more pronounced, facilitating damage identification. However,
a valuable tool for identifying damage locations and qualitatively it is essential to note that the index at the undamaged joint is also
evaluating the degree of damage to hinge joints, particularly in the case notably large, potentially leading to misjudgment. In practical scenarios,
of older bridges with incomplete raw data or disaster-stricken bridges. determining the excitation point on the intermediate slab necessitates
The SDDS index can be used directly without the need for a finite the measurement of the beam’s dimensional information for precise
element model, which significantly reduces the time required for dam­ positioning, constituting a time-consuming process. In contrast, identi­
age identification. fying the mid-span of the side slab is a swifter and more straightforward
process. Moreover, the impact location set on the side slab proves more
effective in exciting both the longitudinal first-order vertical bending

Fig. 12. SDDS index at different double-damaged joints locations: a) 3#4#, b) 5#6#, c) 2#6#, d) 5#7#, e) 1#2#, f) 1#5#.

8
J. Zhao et al. Structures 58 (2023) 105657

3.4.2. Effect of impact force

During the field testing, it is challenging to maintain consistency in
the peak value of pulse force for each impact, often with fluctuations in
the force measurement results. To investigate the influence of pulse peak
on the identification of hinge joint damage, the pulse force peak dis­
cussed in the previous section is multiplied by coefficients of 0.9 and 1.1,
respectively, resulting in three forms of impact force. As shown in
Fig. 17, the differences between the identified damage indices in the
three scenarios are very slight (damaged joint 3#), indicating that pre­
cise control of the impact force is unnecessary, which can free the impact
operation.
Fig. 18 shows the effectiveness of dynamic stiffness as a traditional
indicator to identify damage under different impact scenarios. It is
evident that the robustness of dynamic stiffness, as mentioned in prior
studies, is not as high as the SDDS index since dynamic stiffness tends to
fluctuate significantly with variations in the impact conditions.
Although the impact location and the force amplitude less influence the
location result, it is suggested to control the impact position and force
magnitude whenever possible to avoid the possibility of incorrect
damage localization.

3.4.3. Effect of noise

In order to consider the effect of noise on the measured dynamic
response, Gaussian white noise is added to the acceleration response of
Fig. 13. SDDS index with different single-damage degree (damaged joint 4#).
each slab. The random noise is generated using the following expression:

frequency and the transverse first-order torsional frequency compared ̃ = A(t) + τ⋅std(A(t))⋅rand ∼ N(0, 1)
A(t) (13)
to the intermediate slab.
Expanding on this, the impact of changes in the impact point along
the longitudinal direction on the identification results is explored, as
depicted in Fig. 15. The SDDS indices at different positions exhibit
negligible differences, suggesting that the identification method is not
highly sensitive to the impact location in the longitudinal direction.
Consequently, the optimal excitation position is determined to be at the
middle of the side slab, ensuring the maximum transfer of vibrational
energy to the bridge.
Concerning the placement of sensors, hinged slab bridges generally
undergo maximum deflection and bending moment in the longitudinal
direction at the mid-span of the bridge. Sensors should be strategically
located at critical nodes corresponding to the vibration modes of the
beam. Fig. 16 presents the damage identification outcomes for various
sensor configurations. It is evident that the discrepancies in SDDS
indices among them are minimal. In conclusion, the outcomes of joint
damage identification exhibit insensitivity to both impact location and
sensor placement. Fig. 15. SDDS identified at different impact locations in the longitudi­
nal direction.

Fig. 14. SDDS identified at different impact locations in the transverse direction.

9
J. Zhao et al. Structures 58 (2023) 105657

Fig. 16. SDDS identified at different sensor placement.

Fig. 17. SDDS identified at different peak value of impact force.

In constructing the database, the damage degree of each joint ranges

where A(t) and A(t)
̃ denote acceleration response signals before and
from 5% to 95%. For every 5% change in the damage degree, corre­
after adding noise, respectively; τ denotes the noise level in the response
sponding numerical simulation results of dynamic stiffness are recorded
signal (0 ≤ τ ≤ 1); rand~N(0,1) denotes a group of Gaussian vectors
in the database, which entails 19 damage conditions considered for each
with a mean value of 0 and variance of 1; std denotes the standard de­
joint. Assuming that joint 4# is damaged, to enhance the number and
viation of the corresponding response data.
diversity of samples in the database and improve the robustness of the
In order to simulate the fluctuating noise interference that sensors
machine learning model, dynamic stiffness samples from various impact
experience during field measurements, it is postulated that all signals are
locations, different peak values of impact forces, and varying noise
subjected to random noise levels, including the scenarios of no noise,
levels are included. Consequently, a total of 342 sets of dynamic stiffness
2%, and 5%. Fig. 19 illustrates that damage localization remains unaf­
data are obtained as training and test sets (6 impact force models×3
fected (damage to joint 3#), while the amplitude of the identified
noise levels×19 structural conditions), as shown in Fig. 20.
indices is influenced by noise.
3.5.2. Data training and testing
3.5. Damage quantification by BO-SVM classifier The classification learners and built-in functions in MATLAB R2021a
are utilized to develop the BO-SVM algorithm and statistical indicators
3.5.1. Establishment of dynamic stiffness database in this section. All statistical data derive from the dynamic stiffness
After determining the damage location in the preceding section, the calculation results of the numerical model. The feature matrix X shown
BO-SVM algorithm is adopted to quantitatively identify the damage in Eq. (14) is a [342 × 8] matrix.
degree. The condition of hinge joints can be classified into three grades
according to damage index η: slight damage, moderate damage and se­
vere damage, as summarized in Table 2.

10
J. Zhao et al. Structures 58 (2023) 105657

is iterated 15 times with 5-fold cross-validation. A one-against-one

approach is adopted as the multi-class method, resulting in the con­
struction of 3 classifiers, and the unknown samples are classified into the
class with the maximum classification function value.
The classification accuracy (CA) and the area under curve (AUC) are
used to evaluate the data training and prediction performance of the BO-
SVM method, as shown in Table 3. As the number of iterations increases
during the hyperparameter optimization process, the minimum classi­
fication error decreases, as illustrated in Fig. 21. The results reveal that
the optimal kernel function K is a linear kernel function, and the penalty
parameter C is 178.597. The value of AUC is 1 (Fig. 22), indicating that
the detection method is highly reliable, and the learning effect of the
classifier is very satisfactory.
The confusion matrices of training set and test set are shown in
Fig. 23. The values on the diagonal of the confusion matrix indicate the
number of correct predictions, while non-diagonal elements represent
incorrect predictions. This method establishes a mapping relationship
between damage degree and dynamic stiffness distribution. The CA of
the test set demonstrates that the BO-SVM method is highly effective in
predicting damage degree, confirming the validity of the hyper­
Fig. 18. Dynamic stiffness identified at different impact conditions. parameter tuning conducted in the previous stage. The test set accounts
for noise interference, variation in impact location and differences in
⎡
x1
⎤ ⎡
T1 DS1 T1 DS2 ⋯ T1 DS8
⎤ impact force peaks, but the accuracy of the prediction results has not
⎢ x2 ⎥ ⎢ T2 DS1 T2 DS2 ⋯ T2 DS8 ⎥ been significantly impacted, indicating the strong robustness of machine
⎢
X=⎣ ⎥ = ⎢ ⎥ (14) learning model in field measurements.
⋮ ⎦ ⎣ ⋮ ⋮ ⋱ ⋮ ⎦
x342 T342 DS1 T342 DS2 ⋯ T342 DS8
4. Field verification
The row corresponds to the test number, and the column corresponds
to the slab number. T denotes test case and the element T2 DS1 refers to
4.1. Case study bridge
the dynamic stiffness at No.1 slab in the second test case. The dataset D
added with the damage grade label y can be expressed as:
The proposed method was experimentally verified on a solid hinged
⎡ ⎤
x1 y1 slab bridge composed of two spans of simply-supported beams with a
⎢ x2 y2 ⎥
D=⎢ ⎣ ⋮
⎥ (15)
⋮ ⎦
x342 y342 Table 2
The division of hinge joints condition grades.
The dataset comprises 342 samples, including 8 characteristic vari­
Grade division Joint state Damage index
ables and 1 categorical variable (3 damage grades). Single-damaged
samples and double-damaged samples of hinge joints are trained sepa­ Slight damage The joint is damaged slightly, and no 0 < η ≤ 30%
rately, of which four-fifths are used as training sets and one-fifth as test additional treatment is required.
Moderate The joint is in a damaged state and needs to be 30% < η ≤ 60%
sets. The support vector machine algorithm is employed with Bayesian damage repaired.
optimization as the optimizer. The acquisition function used to deter­ Severe The joint has been severely damaged and η > 60%
mine the next sampling point is set as expected improvement. The model damage needs to be replaced.

Fig. 19. SDDS identified at different noise level.

11
J. Zhao et al. Structures 58 (2023) 105657

Fig. 20. Dynamic stiffness database.

Table 3
Data training and prediction performance of BO-SVM classifier.
Sample CA (%) AUC

Training set 274 99.3 1.00

Test set 68 100 1.00

Fig. 22. ROC curve and AUC value of training set.

frequency of 1024 Hz. Through time-domain and frequency-domain

analysis of the measured data, the first and second dominant fre­
quencies of the bridge are 17 Hz and 23.75 Hz, respectively, as shown in
Fig. 26.
Fig. 21. Minimum classification error in hyperparameters optimiza­ The grillage model of the bridge to be tested was built using ANSYS
tion process. software, and the same numerical simulation as above was conducted.
On this basis, steel plates were added at the bottom of each slab. The
length of 8 m. The transverse direction of the bridge is composed of eight curb stones were simulated by beam elements, and virtual cross-beams
solid slabs, each with a width of 1 m and a height of 0.3 m. The pave­ with a height of 0.25 m were established to simulate the connection
ment thickness is 0.15 m, and the slabs and hinge joints are made of C40 between the curb and the side slab. When impact force acts on No. 1
concrete. The hinge joints of the bridge were severely damaged in 2015 slab, the acceleration response and spectrum of No. 1 slab are shown in
due to prolonged lack of maintenance. In 2017, the bridge was repaired Fig. 26. The first and second dominant frequencies of the numerical
by recasting of concrete, grouting materials in the hinge joints, model are 17 Hz and 24 Hz respectively, which are close to the
rebuilding the concrete pavement, and installing 1 cm thick steel plates measured frequencies. It confirms the precision and reliability of the
at the bottom of the slabs (Fig. 24). calculation results of the finite element model.
The test arrangement is shown in Fig. 25. Rubber hammer was used
to impact the mid-span of No. 1 slab of the case-study bridge. The force 4.2. Identification of joint damage
sensor was mounted on the hammer head, while eight acceleration
sensors were installed at the midspan of the bridge bottom. The data In order to determine the damage location and degree of the joints,
acquisition system sampled the dynamic response of the eight slabs at a the impact force signal and acceleration response signals of slabs were
used to calculate SDDS indices of joints, as illustrated in Fig. 27a. The

12
J. Zhao et al. Structures 58 (2023) 105657

Fig. 23. Confusion matrix: a) training set, b) testing set.

Fig. 24. Case study bridge.

Fig. 25. Field measurements of tested bridge.

13
J. Zhao et al. Structures 58 (2023) 105657

Fig. 26. The measured results and numerical simulation results of acceleration responses: a) time histories, and b) spectra.

SDDS indices at joints 4#, 6#, and 7# are significantly greater than 0, suffered from severe damage with conspicuous longitudinal cracks
indicating that these hinge joints are damaged. Although the SDDS (Fig. 29c). It is proved that the damage localization results using SDDS
indices at joints 1# and 2# are slightly greater than 0, there is no indices and damage quantification results based on machine learning
negative value evident at joint 3#, indicating that joints 1# and 2# are agree well with the visual inspection results.
not damaged. Fig. 27b displays the degree of damage to three hinge
joints identified by BO-SVM classifier. Since the damage of joints 4#, 5. Conclusions
6#, and 7# are all judged to be slight, it can be concluded from Table 2
that no additional treatment is required. It is speculated that the In this paper, a two-step method based on dynamic stiffness is pro­
cracking of the bridge deck axis and the reduction of the transverse posed to locate and quantify the damage of joints in prefabricated
connection height caused by the half-side construction of the pavement hinged slab bridges. The SDDS index is defined using dynamic stiffness,
layer may have caused damage to joint 4#. and the dynamic stiffness database as the data input for support vector
In order to verify the robustness of the proposed method, various machine classifier model is built and optimized using the Bayesian
impact locations and different peak values of impact force were exam­ optimization algorithm. The effectiveness of the proposed method is
ined to determine their effect on the results. Specifically, three distinct verified through numerical simulations using several assumed damage
locations were selected for assessment: L/2 and L/4 of No.1 slab, as well cases and field assessment of an in-service bridge, considering the
as L/2 of No.4 slab. Additionally, there was no control over the peak impact location, sensor placement, peak value of impact force, mea­
values of the impact forces, resulting in three unique impact pulse forms. surement noise that may influence the damage identification. The main
As depicted in Fig. 28, there may be slight variations in the measured conclusions are as follows:
SDDS results due to noise interference. Despite this variability, the
damage indices obtained using the proposed method are nearly identical 1. The proposed SDDS index is capable of accurately identifying the
across all test cases, successfully identifying the location of joint dam­ location of damage joints, providing numerical evidence of the effi­
age. It is concluded that the proposed method is capable of accurately ciency of the method. When the joint is damaged, the dynamic
identifying joint damage in the field test. stiffness of the adjacent slabs on both sides of the joint changes
Upon conducting a visual inspection on site, it was observed that significantly. The SDDS index of the damaged joint is obviously
joints 4#, 6# and 7# have incurred varying degrees of damage. A lon­ greater than 0 and that of the subsequent numbered joint is below 0.
gitudinal crack traversing the entire bridge has appeared at joint 4# on When the hinge joint is intact, the SDDS index approaches 0.
the bridge deck (Fig. 29a), while its corresponding bottom remains 2. The BO-SVM classifier with dynamic stiffness as the feature vector
unaffected. Concrete shedding and holes have occurred in multiple areas can quantify the damage degree. Moreover, the method is robust to
at joint 6# (Fig. 29b). Joint 7#, located underneath the bridge, has uncertain factors in field measurements, such as noise interference,

Fig. 27. The identification results in the field test: a) damage localization, b) damage degree.

14
J. Zhao et al. Structures 58 (2023) 105657

Fig. 28. Damage indices identified at different impact conditions in the field test.

Fig. 29. Results of on-site visual inspection.

different impact positions, and changes in the peak value of impact [2] Zhang J, Yi T, Qu C, Li H. Determining orders of modes sensitive to hinge joint
damage in assembled hollow slab bridges. J Bridge Eng 2022;27(3). https://doi.
force. The BO-SVM classifier is highly effective in practical mea­
org/10.1061/(ASCE)BE.1943-5592.0001834.
surements, making it well-suited for evaluating the repair effect of [3] Rehman SKU, Ibrahim Z, Memon SA, Jameel M. Nondestructive test methods for
the repaired bridge and detecting joint damage in hinged slab bridge concrete bridges: A review. Constr Build Mater 2016;107:58–86. https://doi.org/
in service regularly. This allows for ongoing monitoring of crack 10.1016/j.conbuildmat.2015.12.011.
[4] Zhang J, Yi T, Qu C, Li H. Detecting hinge joint damage in hollow slab bridges
propagation at each joint, enabling timely detection and intervention using mode shapes extracted from vehicle response. J Perform Constr Fac 2022;36
as needed. (1). https://doi.org/10.1061/(ASCE)CF.1943-5509.0001694.
3. The proposed method is data-driven, free of finite element model and [5] Abedin M, Basalo FJDC, Kiani N, Mehrabi AB, Nanni A. Bridge load testing and
damage evaluation using model updating method. Eng Struct 2022;252. https://
less depends on precise control of the hammer excitation with doi.org/10.1016/j.engstruct.2021.113648.
convenient operation, significantly reducing operation time. It is [6] Hu H, Wang J, Dong C, Chen J, Wang T. A hybrid method for damage detection and
suitable for detecting damage in old bridges with incomplete raw condition assessment of hinge joints in hollow slab bridges using physical models
and vision-based measurements. Mech Syst Signal Pr 2023;183. https://doi.org/
data and rapidly evaluating disaster-stricken bridges. 10.1016/j.ymssp.2022.109631.
[7] Tan G, Kong Q, Wang L, Wang X, Liu H. Reliability evaluation of hinged slab bridge
considering hinge joints damage and member failure credibility. Appl Sci 2020;10
Declaration of Competing Interest (14). https://doi.org/10.3390/app10144824.
[8] Liu H, Wang X, Tan G, He X, Luo G. System reliability evaluation of prefabricated
RC hollow slab bridges considering hinge joint damage based on modified AHP.
The authors declare that they have no known competing financial Appl Sci 2019;9(22). https://doi.org/10.3390/app9224841.
interests or personal relationships that could have appeared to influence [9] An Y, Chatzi E, Sim S, Laflamme S, Blachowski B, Ou J. Recent progress and future
the work reported in this paper. trends on damage identification methods for bridge structures. Struct Control Hlth
2019;26(10). https://doi.org/10.1002/stc.2416.
[10] Kong X, Cai C, Hu J. The state-of-the-art on framework of vibration-based
Acknowledgment structural damage identification for decision making. Appl Sci-Basel 2017;7(5).
https://doi.org/10.3390/app7050497.
[11] Roy K. Structural damage identification using mode shape slope and curvature.
This study is sponsored by the National Natural Science Foundation J Eng Mech 2017;143(9). https://doi.org/10.1109/MERCon.2019.8818910.
of China (Grant No. 52178100). [12] Dan D, Xu Z, Zhang K, Yan X. Monitoring index of transverse collaborative working
performance of assembled beam bridges based on transverse modal shape. Int J
Struct Stab Dy 2019;19(8). https://doi.org/10.1142/S021945541950086X.
References [13] Dan D, Zhao Y, Wen X, Jia P. Evaluation of lateral cooperative working
performance of assembled beam bridge based on the index of strain correlation
[1] Hussein HH, Sargand SM, Khoury I, Al-Jhayyish AK. Environment-induced coefficient. Adv Struct Eng 2019;22(5):1062–72. https://doi.org/10.1177/
behavior of transverse tie bars in adjacent prestressed box-girder bridges with 1369433218804924.
partial depth shear keys. J Perform Constr Fac 2017;31(5). https://doi.org/ [14] Zhan J, Zhang F, Siahkouhi M, Kong X, Xia H. A damage identification method for
10.1061/(ASCE)CF.1943-5509.0001068. connections of adjacent box-beam bridges using vehicle - bridge interaction

15
J. Zhao et al. Structures 58 (2023) 105657

analysis and model updating. Eng Struct 2021;228. https://doi.org/10.1016/j. [21] Bulut A., Singh A.K., Shin P., Fountain T., Jasso H., Yan L.J., et al. Real-time
engstruct.2020.111551. nondestructive structural health monitoring using support vector machines and
[15] Sun P., Wang Y., Niu Y., Liu H., Tan G. Damage Identification Method of Simply wavelets. In: Conference on Advanced Sensor Technologies for Nondestructive
Supported Beam Bridge with Multiple Girders. In: 2nd International Conference on Evaluation and Structural Health Monitoring. 2005;5770:180–9. https://doi.org/
Industrial Design and Mechanics Power. 2013;437:377–81. https://doi.org/ 10.1117/12.597685.
10.4028/www.scientific.net/AMM.437.377. [22] Pan H, Azimi M, Yan F, Lin Z. Time-frequency-based data-driven structural
[16] Ma M, Liu J, Ke Z, Gao Y. Bearing capacity estimation of bridge piles using the diagnosis and damage detection for cable-stayed bridges. J Bridge Eng 2018;23(6).
impulse transient response method. Shock Vib 2016. https://doi.org/10.1155/ https://doi.org/10.1061/(ASCE)BE.1943-5592.0001199.
2016/4187026. [23] Lo KF, Ni SH, Huang YH. Non-destructive test for pile beneath bridge in the time,
[17] Cherkassky V. The nature of statistical learning theory. IEEE Trans Neural Netw frequency, and time-frequency domains using transient loading. Nonlinear Dyn
1997;8(6):1564. https://doi.org/10.1109/TNN.1997.641482. 2010;62:349–60.
[18] Finotti RP, Cury AA, Barbosa FDS. An SHM approach using machine learning and [24] Ma M, Liu J, Zhang Z. Evaluating vertical conditions of bridge substructures of
statistical indicators extracted from raw dynamic measurements. Lat Am J Solids heavy-haul railway lines based on dynamic stiffness and pier vibration response.
Struct 2019;16(2):165. https://doi.org/10.1590/1679-78254942. Eng Struct 2021;235. https://doi.org/10.1016/j.engstruct.2021.112037.
[19] Alves V, Cury A, Roitman N, Magluta C, Cremona C. Structural modification [25] Liu H., Jiao Y., Gong Y., Bi H., Sun Y. Damage identification for simply-supported
assessment using supervised learning methods applied to vibration data. Eng Struct bridge based on SVM optimized by PSO (PSO-SVM). In: International Conference
2015;99:439–48. https://doi.org/10.1016/j.engstruct.2015.05.003. on Information Engineering for Mechanics and Materials. 2011;80–81:490–4. htt
[20] Sharafi H, Ebtehaj I, Bonakdari H, Zaji AH. Design of a support vector machine ps://doi.org/10.4028/www.scientific.net/AMM.80–81.490.
with different kernel functions to predict scour depth around bridge piers. Nat
Hazards 2016;84(3):2145–62. https://doi.org/10.1007/s11069-016-2540-5.

16

View publication stats