Engineering Failure Analysis 118 (2020) 104845

Contents lists available at ScienceDirect

Engineering Failure Analysis

journal homepage: www.elsevier.com/locate/engfailanal

Model-based damage identification of railway bridges using

T
genetic algorithms
Vinicius N. Alvesa, Matheus M. de Oliveiraa, Diogo Ribeirob, , Rui Calçadac,
⁎

Alexandre Curyd
a
Department of Civil Engineering, School of Mines, Federal University of Ouro Preto, Ouro Preto, Brazil
b
CONSTRUCT-LESE, School of Engineering, Polytechnic of Porto, Porto, Portugal
c
CONSTRUCT-LESE, Faculty of Engineering, University of Porto, Porto, Portugal
d
Post-Graduate Program in Civil Engineering, Federal University of Juiz de Fora, Juiz de Fora, Brazil

ARTICLE INFO ABSTRACT

Keywords: The assessment of structural integrity via numerical model updating has been drawing attention
Railway bridges in several areas of engineering over the last years. Basically, it consists in an optimization process
Damage identification based on the minimization of the residuals between measured and estimated numerical re­
Modal parameters sponses. In such methodologies, several factors influence the success of both localization and
Model updating
quantification of structural damage, such as: the damage features used in the formulation of the
Genetic algorithm
objective function, the optimization algorithm and the adopted updating parameters. Many ex­
Features selection
isting studies using these methods are applied to simple structural systems, e.g., beams, frames
and trusses. However, few studies applied to large and complex structures are found in the lit­
erature. In this context, this work proposes to assess the performance of a genetic algorithm-
based approach applied to two case studies. The first case refers to a two-dimensional model of a
hypothetical railway bridge, where the efficiency and robustness of five different indicators are
assessed considering three damage scenarios. In the second case, a real railway bridge is con­
sidered. The results obtained show that the proposed approach is able to detect, locate and
quantify multiple damage with several updating parameters and few target responses.

1. Introduction

Currently, many civil engineering structures are reaching their estimated life cycle, showing visible signs of deterioration over
time. In many situations, such degradation is aggravated by the occurrence of unforeseen extreme events, e.g., earthquakes, ex­
plosions, among others [1]. Since the replacement of these structures is economically unfeasible, many damage detection procedures
have been designed so that structural life cycle is safely extended. Such techniques aim to identify damage at early ages, foreseeing
eventual maintenance and reinforcement operations to guarantee both structural safety and operationality, thus avoiding serious
social, environmental and economic consequences that could result from a collapse.
Most structural damage identification techniques are based on vibration data. In general, damage can be identified through
modifications in the structure’s dynamic responses. This is due to the fact that damage oftentimes cause changes in the system’s mass,
stiffness and energy dissipation characteristics [2].
According to Rytter [3], the problem of structural damage identification can be divided into five levels: level 1 - damage de­
tection, level 2 - damage location, level 3 - damage classification, level 4 - damage quantification, and level 5 - residual life prediction.

⁎
Corresponding author at: CONSTRUCT-LESE, School of Engineering, Polytechnic of Porto, Porto, Portugal.

https://doi.org/10.1016/j.engfailanal.2020.104845
Received 18 April 2020; Received in revised form 7 August 2020; Accepted 10 August 2020
Available online 25 August 2020
1350-6307/ © 2020 Elsevier Ltd. All rights reserved.
V.N. Alves, et al. Engineering Failure Analysis 118 (2020) 104845

In fact, two different approaches can be performed for structural damage evaluation. The first is the data-based approach, which
essentially relies on experimental data; the second is the model-based approach, based on analytical or numerical models (inverse
problem). The approach chosen for damage assessment influences on the accuracy of the structural diagnosis.
Usually, data-based approaches resort to statistical analysis and pattern recognition techniques to evaluate the existence of
structural damage [4]. For methods using such a methodology, the knowledge of the current structural condition is achieved through
comparisons with previous structural conditions.
Several techniques that rely on this premise have proven to be effective in detecting structural damage, e.g., Artificial Neural
Networks [5], data clustering procedures [6–9], Principal Component Analysis (PCA) [10], Support Vector Machines (SVM) [11],
among others. Despite their proven efficiency, Barthorpe [12] cites some potential limitations of data-based techniques. For instance,
if only the detection of damage is required, these methods are very effective as only the structure’s experimental data are needed.
However, when it comes to the assessment of damage location, classification or quantification, it is unlikely that information from all
undamaged and damaged structures for the training of the classification techniques will be available, especially for large scale
structures. Even if it is possible to acquire data from the damaged structures for training, it will be difficult to acquire data in the
several damaged locations, mainly because the amount of data needed would be very large, in addition to the difficulty in predicting
the exact damage location.
The model-based approach, which has gained more popularity in recent years, consists in a finite element (FE) model updating in
which parameters of a preliminary model of an undamaged structure are modified to reach an acceptable correspondence to the
actual structure, possibly damaged. Once this correspondence is reached, it is possible to infer about damage location and quanti­
fication. Typically, seeking this agreement consists in minimizing the differences between the dynamic responses measured at the
actual structure and those obtained through its FE model counterpart, which is generally described as an optimization problem [2].
Within methodologies that adopt this type of approach, three factors significantly influence the accuracy of location and quantifi­
cation of structural damage: the objective function, the optimization algorithm, and the selected updating parameters [13].
Studies evaluating the efficiency of methodologies based on FE model updating have been carried out by different authors in
various applications, such as in beams [6,7,14,15], portal frames [16,17], trusses [13,18], and shells [19,20].
Most of the existing research using numerical models is applied to relatively simple structures, while its application to large scale
structures is relatively limited [21]. Structural damage identification in complex structures presents an enormous challenge asso­
ciated with the large amounts of degrees of freedom, which intrinsically leads to numerous nodes and elements. Another difficulty
lies on the fact that typically small damage in some specific structural elements can have negligible influence on the overall struc­
ture’s dynamic behavior.
Among the several existing optimization algorithms, evolutionary algorithms, such as genetic algorithms [22], particle swarm
optimization (PSO) [23] and ant colony optimization (ACO) [24], are the most used techniques to deal with complex and large
structures, especially when many design variables (mostly non-linear) are considered. Additionally, the evolutionary algorithms
intrinsic characteristics increase their ability to detect damage, particularly when damage locations are distributed throughout the
structure.
In this context, this paper proposes some important contributions concerning features that are often insufficiently addressed in the
literature, such as:

- The development of a damage identification strategy based on genetic algorithms that is not only able to detect damage, but also
has the potential to locate and quantify their extension.
- The development of an efficient methodology to detect multiple damage scenarios with a considerable number of unknown
numerical parameters and in large-scale FE models of real civil engineering structures.

Thus, this study focuses on assessing the efficiency and robustness of a methodology based on FE model updating using genetic
algorithms to identify multiple damage scenarios in large-scale structures. To this end, two applications are considered. In the first
application, the efficiency of different damage indicators is evaluated considering a hypothetical railway bridge subjected to multiple
damage. After identifying the most efficient indicators, a second case study is carried out considering a real railway bridge.

2. Damage identification methodology

The proposed damage identification strategy relies on comparing the reference modal/dynamic responses of the bridge’s nu­
merical model to the simulated damage model’s counterpart. This strategy is performed by minimizing an objective function defined
by one or several damage indicators based on the modal/dynamic responses sensitive to structural change. A flowchart illustrating
the computational implementation of the damage identification methodology based on genetic algorithms, which was developed
using the commercial software MATLAB and ANSYS, is presented in Fig. 1. Genetic algorithms have recognized advantages, such as:
the non-dependence of the initial starting point, capability to manage a large number of parameters and constraints, possibility to
handle discrete and binary variables, ability to find the global minimum in functions with several local minima and the possibility to
accept failed designs [25].
In this study, the damage scenario is artificially induced. For this, reference parameters are used to simulate the damage scenario
before the minimization process starts. Therefore, the FE model for the damage scenario is defined in ANSYS software, where a
modal/dynamic analysis is carried out and the reference damage indicators (DIref) are obtained. After this stage, the optimization
process is performed. In the iterative process, the first generation (k = 1) is created based on random generation of individuals that

2
V.N. Alves, et al. Engineering Failure Analysis 118 (2020) 104845

Fig. 1. Flowchart of the damage identification methodology.

are candidates to form the initial population using the MATLAB software. Each generation is formed by n individuals, in which each
individual is a possible solution to the optimization problem. An individual is formed by j updating parameters values, whose
modifications are responsible for matching between the simulated and reference models. The individuals of generation k + 1 are
created from individuals of generation k, based on genetic operators such as selection, crossover, mutation and substitution [26].
From the numerical parameters defined for each individual, the simulated numerical model is developed, and a modal/dynamic
analysis is performed to determine the respective damage indicators (DIsim). The damage identification ends when a stopping con­
dition is satisfied. Typically, this is achieved by limiting the number of generations or by specifying a convergence criterion, i.e.,
when the value of the objective function is less than a predefined tolerance. Thus, the updating parameters of the individual with the
lowest value of the objective function are representative of the damage scenario. In this way, the set of parameters of this individual
with the least objective function will be responsible for indicating the position and magnitude of the damages along the structure. In
real structures, the modal/dynamic data derived from an artificially induced damage scenario can be replaced by the inclusion of
measured experimental data.

3. Damage indicators

One of the most relevant aspects within the damage identification procedure is to select appropriate dynamic responses that can
serve as indicators of structural damage. When updating numerical models, one or more residuals from the dynamic responses can be
used to define the objective function, which in this case represents the differences between the responses of the simulated damaged
model and the reference damaged model. Alkayem et al. [27] reviewed a series of dynamic responses that are generally used to
identify structural damage. The damage indicators used in this paper for the formulation of the objective function are presented in the
following sections.

3.1. Natural frequencies (DI1)

Natural frequencies can be use as sensitive parameters to identify structural modifications since they may vary due to the
structure’s mass and stiffness characteristics, as well as due to its boundary conditions. However, their use may impose practical
limitations in various applications, especially for large-scale structures, for which detectable changes in frequency may require high
levels of damage. Additionally, as natural frequencies are highly sensitive to environmental effects (e.g., temperature, wind, hu­
midity, etc.), they could indicate the occurrence of false-positive damage scenarios.

3.2. Vibration modes (DI2)

Vibration mode shapes provide spatially distributed information about the structure’s dynamic behavior and, therefore, can be

3
V.N. Alves, et al. Engineering Failure Analysis 118 (2020) 104845

more suitable to provide evidence about damage location when compared to natural frequencies [28]. However, the use of mode
shapes as a damage indicator requires the measurement of ambient vibrations at several points of the structure to ensure an adequate
spatial resolution, which is not always possible.

3.3. Natural frequencies and vibration modes (DI3)

An alternative used by several researchers is the formulation of an objective function including two or more damage indicators.
The objective function used most regularly combines natural frequencies and mode shapes [19,25,29–32].

3.4. Modal curvatures (DI4)

Damage indicators derived from vibration modes, such as modal curvatures, are widely used as an alternative to modal con­
figurations to obtain enhanced spatial information about modifications in vibration properties. The double derivation process has the
effect of amplifying any discontinuities in the modal configuration caused by a localized damage. It is observed that for beams, plates
and shells, the curvature value at a point in the structure is inversely proportional to its bending stiffness. Therefore, if the stiffness at
a region is reduced by damage, the curvature at that region will tend to increase. This can be used to efficiently detect and locate
structural damage [33].

3.5. Frequency response functions (DI5)

The Frequency Response Function H(ω) is a transfer function that relates the structure’s dynamic response to a specific action
imposed on it, both represented in the frequency domain. This function can be estimated as the ratio between the response spectrum
X(ω) and the force spectrum F(ω), according to Eq. (1).
X( )
H( ) =
F( ) (1)

4. Case study 1: Hypothetical railway bridge

A two-dimensional FE model of a hypothetical railway bridge was developed in ANSYS software (Fig. 2). The bridge consists of a
15 m-long supported deck made of reinforced concrete. The cross section has dimensions 5.0 m (width - B) and 0.85 m (height - H).
The material’s specific mass is 18700 kg/m and the concrete’s modulus of elasticity (E) was considered equal to 30 GPa.
To model the deck, 30 beam finite elements were used. The deck’s supports were modeled using spring elements with a vertical
stiffness (Kv1 and Kv2) equal to 4 × 109N/m. The track was modeled with a length of 10 m over the bridge (on both sides) to simulate
the continuity effect on adjacent embankments. The simulation of the track-bridge connection was performed using spring elements
reproducing the longitudinal stiffness of the ballast layer (Kh) equal to 30 × 106N/m. The distance between the deck’s center of
gravity and the rail (D) is 1.075 m. The parameters used were extracted or adapted from the final report of D214 Specialists’
Committee from ERRI [34] and from UIC 774–3 [35].
In general, the presence of damage in a structure results in changes in structural stiffness and mass. In this way, for this case study,
the damage scenarios are simulated considering changes in: i) the deck’s mass (m), which reproduces ballast mass variations due to
geometrical levelling operations of the track; ii) the vertical stiffness of the right-side support (KV2), which reflects the effect of the
support bearing degradation; iii) the modulus of elasticity (E1, E2, E3, E4) in four specific deck sections (x1, x2, x3, x4) in order to
simulate the presence of local damage. Table 1 presents the three scenarios with the respective values used to simulate damage. These
scenarios differ by the amount and intensity of local damage on the deck. Table 1 also shows the limits adopted for the variation of
each parameter used in the genetic algorithm (GA). It is important to highlight that in this case the damage position is associated with
the element number, which can vary from 1 to 30, since the deck is 15 m long and was modeled with 30 elements with a length of
0.5 m each.
In the study, all damage indicators presented in Section 3 were evaluated, namely: natural frequencies (DI1), vibration modes

Fig. 2. 2D FE model of a hypothetical railway bridge (damaged parameters marked in red). (For interpretation of the references to color in this
figure legend, the reader is referred to the web version of this article.)

4
V.N. Alves, et al. Engineering Failure Analysis 118 (2020) 104845

Table 1
Baseline and simulated damage scenarios.
Parameters Unit Baseline scenario Simulated damage scenarios Genetic algorithm

C1 C2 C3 Limits (lower/upper)

m kg/m 18700 17700 17700 17700 10000/25000

KV2 × 109N/m 4 4.5 4.5 4.5 1/15
x1 (element number) – – 10 10 10 1/30
E1 GPa 30 25 25 25 15/30
x2 (element number) – – 15 15 15 1/30
E2 GPa 30 20 20 20 15/30
x3 (element number) – – – 20 20 1/30
E3 GPa 30 – 28 28 15/30
x4 (element number) – – – – 5 1/30
E4 GPa 30 – – 26 15/30

shapes (DI2), natural frequencies and vibration modes shapes (DI3), modal curvatures (DI4), and frequency response functions (DI5).
For the first four indicators, the first five bending vibration modes of the deck are considered. Since frequency response functions are
transfer functions that relate inputs (loads) to outputs (structure’s responses), the last indicator is evaluated considering a known
impulsive action of 100 kN applied close to the midspan section of the deck and the acceleration responses measured in five different
positions (Fig. 3).

4.1. Modal parameters

Fig. 4 depicts the configurations of the first five bending mode shapes of the bridge with their respective natural frequencies for
the baseline scenario (without damage). The values of the natural frequencies and the Modal Assurance Criterion (MAC) parameters
for the models with and without damage are presented in Table 2. It is possible to verify that the frequencies were clearly altered in
the presence of damage, while the MAC values were less affected, maintaining correlation values close to 1.0 for all vibration modes.

4.2. Performance of damage indicators

To attest the efficiency of the five damage indicators, the following considerations were made:

- For indicator DI1, the natural frequencies related to the first five mode shapes were considered,
- For indicator DI2, the mode shapes were obtained based on the amplitudes of all nodes of the FE mesh (31 nodes) corresponding to
the bridge deck,
- For indicator DI4, the curvatures of the mode shapes were calculated using the finite central difference method,
- For indicator DI5, the FRFs were evaluated considering the excitation force applied near the deck’s midspan and the acceleration
responses were measured in five different locations (see Fig. 3).

Regarding the genetic algorithm, an initial population comprised 30 individuals randomly generated through the Monte Carlo
method, considering 300 generations, in a total of 9000 individuals. The number of elites was considered equal to 1 and the number
of substitute individuals was also defined equal to 1. The crossing and mutation rates were considered equal to 50% and 15%,
respectively. In case of mutation rate, a standard deviation variable along the optimization, between 0.10 and 0.01, was defined.
Fig. 5 presents the evolution of the updating parameters as a function of the number of generations for the five damage indicators
(C2 damage scenario). In the ordinate axes, it is possible to observe lower and upper limit values of each parameter and the target
reference value (corresponding to the damage scenario) highlighted with an asterisk. In the abscissa axes, the limit of 150 generations
was adopted, as no variations in the results of the different parameters occurred between generations 150 to 300.
By evaluating the mass component (m), one notices that the different damage indicators can find the reference value before

Fig. 3. Position of the control points and location of the excitation force.

5
V.N. Alves, et al. Engineering Failure Analysis 118 (2020) 104845

Fig. 4. Mode shapes and respective natural frequencies for the bridge’s first five vibration modes (baseline scenario).

Table 2
Natural frequencies and MAC values for the baseline and damage scenarios.
Vibration modes Baseline scenario Simulated damage scenarios

C1 C2 C3

Freq. (Hz) Freq. (Hz) MAC Freq. (Hz) MAC Freq. (Hz) MAC

1 4.69 4.74 1.000 4.73 1.000 4.73 1.000

2 17.24 17.67 1.000 17.65 1.000 17.60 1.000
3 36.34 37.02 0.999 37.01 0.999 36.87 0.999
4 59.33 61.20 0.998 61.09 0.998 60.98 0.998
5 83.91 85.97 0.996 85.92 0.996 85.92 0.996

Fig. 5. Evolution of the parameters as a function of the number of generations (scenario C2).

6
V.N. Alves, et al. Engineering Failure Analysis 118 (2020) 104845

Fig. 6. Errors of damage indicators for damage estimation for each scenario: a) scenario C1, b) scenario C2, c) scenario C3.

reaching 100 generations. For the KV2 parameter, for instance, all damage indicators found the reference value within the first
generations. In the case of the parameters related to local damage along the deck, it is possible to observe that the algorithm tends to
efficiently find the most severe damage first, that is, it initially identifies the damage of the element in position x2, which refers to a
reduction of 33% of the E-modulus; then, the damage of the element in position x1, which has a reduction of 17% of the E-modulus;
finally, the damage of the element in position x3 with a reduction of 7% of the E-modulus. It is also noted that the DI1 indicator is not
effective in estimating all local damage parameters. Moreover, the DI5 indicator, despite its good performance in identifying all
damage associated to scenario C2, it presents a lot of instability in the estimates over the generations.
Fig. 6 shows the relative errors of the parameters’ estimation for the three damage scenarios and considering distinct damage
indicators. The relative error (ε) of the parameters’ estimation is calculated using the difference of the values obtained for the
parameters that reproduce the damage (Xsim) and the target reference damage values (Xref) in relation to the target values:

X sim X ref
(% ) = × 100
X ref (2)

Comparing the three damage scenarios, it is observed that the algorithm converges slower to the reference values as more
parameters are used. When scenarios C1 and C2 are evaluated with 6 and 8 parameters, respectively, the reference value was found
for most damage indicators. However, for the scenario C3 (considering 10 parameters), errors appear in some parameters, regardless
of the damage indicator analyzed.
When it comes to the performance of the different damage indicators, the one based on natural frequencies (DI1) presents the
largest differences in relation to the reference, regardless of the damage scenario and the parameter considered. For example, it
presents errors around 66.7% for determining location x4, and 43% for the modulus of elasticity E4 in scenario C3. It is also possible to
observe that the DI5 indicator presents a relative error of 11% for the quantification of parameter E1 in the C3 damage scenario, not
shown in other scenarios. For most parameters, the reference value was obtained with a zero error or close to it.
Fig. 7 presents the local damage identified by indicator DI1, where the red color represents the locations of the artificially
simulated damage and the green color indicates the locations of the damage estimated by the proposed methodology. It is possible to

7
V.N. Alves, et al. Engineering Failure Analysis 118 (2020) 104845

Fig. 7. Location of damage obtained by the optimization algorithm for ID1: a) scenario C1, b) scenario C2, c) scenario C3.

Fig. 8. Location of damage obtained by the optimization algorithm for ID3: a) scenario C1, b) scenario C2, c) scenario C3.

identify that for scenarios C1 and C3 the indicator estimated the damage in a position symmetrical to the real simulated position,
which confirms the limitations of the frequencies in providing accurate information about the damage location. The same is done for
indicator ID3 (Fig. 8), where it is possible to observe for this damage indicator, the methodology made it possible to correctly find the
damage locations for all damage scenarios.

5. Case study 2: São Lourenço railway bridge

This second case study refers to the São Lourenço railway bridge, a bowstring-type bridge situated on the northern line of
Portuguese railways (Fig. 9a). The bridge structure consists of two independent reinforced concrete decks with a length of 42 m. Each
deck consists of a lower slab laterally suspended by two metallic arches through hangers and diagonals. The deck loads are trans­
ferred to the abutments via support bearing, two in each abutment, positioned in the alignments of the main girders. Each half-deck
cross section, with a total width of 7.35 m, consists of a concrete slab laterally supported by two main girders, forming a U-section,
and a side footway. The track is formed by a ballast layer positioned over the deck slab, single-block concrete sleepers, rail pads and
rails.
A 3D FE model including the deck, arches, hangers, diagonals, support bearings and track, is used for the application of the
damage identification methodology. This model was previously developed by Ribeiro et al. [25] in ANSYS software (Fig. 9b). Similar
FE models are proposed by Guo et al. [36] and Liu et al. [37].
The deck slab was modelled using solid elements. The arches, hangers, diagonals and bracings were modelled with beam ele­
ments. The support bearings were modeled using spring-dashpot assemblies. To correctly reproduce the deformability length of the
hangers and diagonals, rigid elements were introduced in its extremities in the connections with the arch and main girders.
Four damage scenarios of the bridge structure were simulated, namely:

- damage located in a central section of the slab at the midspan of the bridge, by reducing the concrete's modulus of elasticity,
- damage in one of the support bearings by increasing its vertical stiffness,
- damage in a hanger/diagonal-deck connection by reducing the modulus of elasticity of the connection element between the
hanger/diagonal-deck,
- the simultaneous occurrence of the three previous damage scenarios.

8
V.N. Alves, et al. Engineering Failure Analysis 118 (2020) 104845

Fig. 9. São Lourenço railway bridge: a) general overview, b) FE numerical model.

The values of the numerical parameters referring to the baseline and damaged scenarios are shown in Table 3 and the exact
damage locations illustrated in Fig. 10. The damage on the deck slab is simulated in order to become less intense as it moves away
from the location of the affected finite element. The damage is simulated by reducing the concrete's modulus of elasticity of the deck
so that the central element, indicated by the coordinates x1 and z1, presents 80% of the material’s modulus of elasticity, whereas the
adjacent elements (distributed concentrically) present, respectively, 85% and 90%, representing a damage area of 2.25 m2 (Fig. 10b).
Damage in support is introduced in one of four supports bearing (pos1), and damage in a hanger/diagonal-deck connection is
simulated in one of 52 possible connections (pos2) of this type existing on the bridge.

5.1. Modal parameters

The bridge’s natural frequencies are associated to global and local vibration modes. Global vibration modes involve global dis­
placements of the deck and the arches. Local vibration modes are related to diagonal and hanger vibrations without any significant
movement of the deck or arches. From a modal analysis performed for the first 30 modes of vibration of the structure, five global
modes of vibration were identified, shown in Fig. 11. These modes shapes will be later used for damage identification.
The modal configurations were described using the 3 DOFs of the 60 nodes of the numerical model identified in Fig. 12. These
nodes correspond to the positions where the accelerometers were installed in an ambient vibration test of the bridge, described in
Ribeiro et al. [25]. The positioning of the sensors was conditioned to 2.80 m in height due to the safety distance relative to the
catenary.

5.2. Damage identification

The set of natural frequencies and mode shapes was used as a damage indicator (DI3) due to its easiness of implementation in a
three-dimensional model, and mainly due to the good performance presented in the hypothetical bridge study case.
For DI3 indicator the objective function (f) comprises two terms, one related to the residuals of natural frequencies and other
related to the residuals of mode shapes through the MAC parameter [25]:

N |firef fisim | N ref sim

f=a +b |MAC ( i , i ) 1|
i=1 firef i=1
(3)
where firef and fisim are
the frequencies values regarding the reference and simulated damage scenarios for mode i, and are i
ref
i
sim

the vectors containing the reference and simulated modal information regarding mode i, N is the number of mode shapes, and a and b
are weighing factors of the terms of the objective function. In this application, N is equal to 5 and terms a and b are both equal to 1.
Like the previous case, concerning the genetic algorithm, an initial population comprised 30 individuals randomly generated
through the Monte Carlo method, considering 300 generations, in a total of 9000 individuals. The values of the genetic operators’

Table 3
Parameters for different damage scenarios of the São Lourenço bridge’s numerical model.
Parameters Unit Baseline scenario Damage scenarios Genetic algorithm

C1 C2 C3 C4 Limits (lower/upper)

Deck x1 m – 2.3 – – 2.3 0/4.6

z1 m – 20.63 – – 20.63 0/42
Ebd GPa 38.1 30.5* – – 30.5* 19.0/40.0
Support pos1 – – 2 – 2 1/4
Kv N/m 5000 – 5500 – 5500 4500/9500
Hangers pos2 – – – 45 45 1/52
Esc GPa 210 – – 189 189 100/210

* in the central finite element.

9
V.N. Alves, et al. Engineering Failure Analysis 118 (2020) 104845

Fig. 10. Location of damage in São Lourenço bridge model: a) side view, b) top view and detail of the localized deck damage.

Fig. 11. Global mode shapes and corresponding natural frequencies of São Lourenço bridge.

Fig. 12. Nodes considered to represent modal configurations.

10
V.N. Alves, et al. Engineering Failure Analysis 118 (2020) 104845

Table 4
Relative errors of the parameter’s estimation values in the damage identification of São Lourenço bridge.
Damage scenario Relative errors (%)

x1 z1 Ebd pos1 Kv pos2 Esc

C1 0.0 0.0 0.0 – – – –

C2 – – 0.0 0.0 – – –
C3 – – – – – 0.0 0.0
C4 0.0 0.0 0.0 0.0 0.0 15.7 5.4

parameters were defined the same as those of the case study 1.

Table 4 shows the relative errors of the parameter’s estimation values, between simulated and target damage scenarios, estimated
through Eq. (2) for the four damage scenarios. In the situations where a single damage is considered (scenarios C1, C2 and C3), the
proposed method located it successfully, with an error value of 0%. For scenario C4, in which the damage from the first three
scenarios is considered simultaneously, errors were found in determining the damage location of the deck-arch connection, in the
amount of 15.7%, and also in evaluating the modulus of elasticity, with a relative error of 5.4%. Finally, in Fig. 13 it is possible to
observe the location of the target damage and those obtained using the proposed strategy for scenario C4.
It is important to highlight that, from all the damage scenarios considered, the damage to the deck-arch connection is the one that
less affects the global structure’s behavior, which justifies the greater difficulty in identifying this damage. The changes in frequencies
and global vibration modes caused by damage to the deck-arch connection are smaller than those caused on the deck and support
bearing. In addition, the obtained location for this specific damage is symmetrical in relation to the reference damage, which can be
justified by the fact that damage in symmetrical positions cause the same effect in modifying the frequencies of symmetric vibration
modes. If global asymmetric vibration modes had been identified, which did not happen in the frequency range of interest, the
algorithm would have adequately identified this damage as well.

6. Conclusions and future developments

This paper presented a damage identification strategy based on the updating of numerical models using genetic algorithms able to
detect, locate and quantify damage. The proposed methodology was assessed through two applications. The first case – a FE model of
a bridge – aimed to attest the efficiency of five different damage indicators, and in the second case, the methodology was applied to a
real structure to assess its efficiency in a large-scale structure.
In the first application, a FE numerical model of a hypothetical railway bridge was simulated considering three damage scenarios
aiming to evaluate five damage indicators: natural frequencies (DI1), vibration mode shapes (DI2), natural frequencies and mode
shapes (DI3), modal curvatures (DI4), frequency response functions (DI5). All five indicators allowed to precisely detect, locate and
quantify the simulated damage. Only the indicator (DI1), however, showed minor levels of error even for the damage scenarios C1
and C2, in which few parameters simulating the damage were considered. The other indicators allowed identifying most of the
damage cases, only presenting some level of error for the C3 damage scenario.
The second study was based on the numerical model of São Lourenço railway bridge. For this case, four damage scenarios were
considered: three scenarios related to isolated damage associated to a local damage on the deck slab, a malfunctioning of a support
bearing, and an anomaly on a deck-arch connection (scenarios C1, C2 and C3), and all these damages simultaneously (scenario C4).
For this assessment, only the DI3 indicator was used, since it presented the best performance in the first case study. The results

Fig. 13. Location of damage obtained in São Lourenço bridge model for ID3: a) lateral view, b) top view.

11
V.N. Alves, et al. Engineering Failure Analysis 118 (2020) 104845

obtained proved the efficiency and robustness of the methodology to successfully detect all the damage and their locations, with
exception of the anomaly on the deck-arch connection in scenario C4. However, for this last case, the damage location indicated by
the algorithm was in a symmetrical position in relation to the induced damage, possibly due to the fact that damage in symmetrical
positions of the structure cause the same effect in changing the frequencies of symmetrical vibration modes.
As future developments, the authors intend to perform the damage identification based on damage indicators related to the bridge
dynamic responses under the passage of traffic loads, such as, accelerations, displacements and strains. The dynamic responses will be
obtained from complex train-bridge interaction models including the track irregularities.
Additionally, the proposed damage identification methodology will also be applied for experimental data derived from a
Structural Health Monitoring (SHM) system. For that, the reference damage scenario, which was artificially induced in the numerical
model, should be replaced by a baseline damage scenario, derived from the experimental data, and representing the actual condition
of the real structure. The regular updating of numerical models of railway bridges, based on experimental information provided by
SHM systems, will allow upgrading its ability to reliably reproduce the real structural condition, considering the existing damages.
The baseline numerical model of the bridge can be used for a more precise assessment of the structural safety and train running
safety, and planning maintenance operations.

Declaration of Competing Interest

The authors declare that they have no known competing financial interests or personal relationships that could have appeared to
influence the work reported in this paper.

Acknowledgements

The authors are grateful for the support from FAPEMIG (grants PPM-00002-16 and PPM-00001-18), CAPES, CNPq (grant
#304329/2019-3) and UFOP. The third author would like to acknowledge the support of the Base Funding - UIDB/04708/2020 of the
CONSTRUCT - Instituto de I&D em Estruturas e Construções - funded by Portuguese national funds through the FCT/MCTES
(PIDDAC).

References

[1] U. Dackermann, J. Li, B. Samali, Dynamic-Based Damage Identification Using Neural Network Ensembles and Damage Index Method, Adv. Struct. Eng. 13 (6)
(2010) 1001–1016.
[2] N.F. Alkayem, M. Cao, M. Ragulskis, Damage localization in irregular shape structures using intelligent FE model updating approach with a new hybrid objective
function and social swarm algorithm, Appl. Soft Comput. 83 (2019) 105604, https://doi.org/10.1016/j.asoc.2019.105604.
[3] Rytter, A. (1993). Vibrational Based Inspection of Civil Engineering Structures, Dept. of Building Technology and Structural Engineering, Aalborg University,
Aalborg.
[4] Meisam Gordan, Hashim Abdul Razak, Zubaidah Ismail, Khaled Ghaedi, Recent Developments in Damage Identification of Structures Using Data Mining, Lat.
Am. J. solids struct. 14 (13) (2017) 2373–2401, https://doi.org/10.1590/1679-78254378.
[5] Finotti, R.P., Cury, A.A., and Barbosa, F. de S. (2019). An SHM approach using machine learning and statistical indicators extracted from raw dynamic mea­
surements: Latin American Journal of Solids and Structures, Vol. 16, No. 2, DOI: 10.1590/1679-78254942.
[6] V. Alves, A. Cury, N. Roitman, C. Magluta, C. Cremona, Novelty detection for SHM using raw acceleration measurements: Novelty Detection for SHM Using Raw
Acceleration Measurements, Struct. Control Health Monit. 22 (9) (2015) 1193–1207.
[7] V. Alves, A. Meixedo, D. Ribeiro, R. Calçada, A. Cury, Evaluation of the Performance of Different Damage Indicators in Railway Bridges, Procedia Eng. 114
(2015) 746–753.
[8] João P. Santos, Christian Crémona, Luís Calado, Paulo Silveira, André D. Orcesi, On-line unsupervised detection of early damage: On-line Unsupervised Detection
of Early Damage, Struct. Control Health Monit. 23 (7) (2016) 1047–1069.
[9] R. de Almeida Cardoso, A. Cury, F. Barbosa, Automated real-time damage detection strategy using raw dynamic measurements, Eng. Struct. 196 (2019) 109364,
https://doi.org/10.1016/j.engstruct.2019.109364.
[10] Rocco Langone, Edwin Reynders, Siamak Mehrkanoon, Johan A.K. Suykens, Automated structural health monitoring based on adaptive kernel spectral clus­
tering, Mech. Syst. Sig. Process. 90 (2017) 64–78.
[11] Guoqing Gui, Hong Pan, Zhibin Lin, Yonghua Li, Zhijun Yuan, Data-driven support vector machine with optimization techniques for structural health monitoring
and damage detection, KSCE J Civ Eng 21 (2) (2017) 523–534.
[12] Barthorpe, R.J. (2011). On Model and Data-based Approaches to Structural Health Monitoring, University of Sheffield, p. 273.
[13] Milad Jahangiri, M.A. Najafgholipour, S.M. Dehghan, M.A. Hadianfard, The efficiency of a novel identification method for structural damage assessment using
the first vibration mode data, J. Sound Vib. 458 (2019) 1–16.
[14] A.C. Altunışık, F.Y. Okur, V. Kahya, Modal parameter identification and vibration based damage detection of a multiple cracked cantilever beam, Eng. Fail. Anal.
79 (2017) 154–170.
[15] Roumaissa Zenzen, Idir Belaidi, Samir Khatir, Magd Abdel Wahab, A damage identification technique for beam-like and truss structures based on FRF and Bat
Algorithm, Comptes Rendus Mécanique 346 (12) (2018) 1253–1266.
[16] N.F. Alkayem, M. Cao, Damage identification in three-dimensional structures using single-objective evolutionary algorithms and finite element model updating:
evaluation and comparison, Eng. Optim. 50 (10) (2018) 1695–1714.
[17] I. Behmanesh, B. Moaveni, C. Papadimitriou, Probabilistic damage identification of a designed 9-story building using modal data in the presence of modeling
errors, Eng. Struct. 131 (2017) 542–552.
[18] Leila Katebi, Mohsen Tehranizadeh, Negar Mohammadgholibeyki, A generalized flexibility matrix-based model updating method for damage detection of plane
truss and frame structures, J Civil Struct Health Monit 8 (2) (2018) 301–314.
[19] Dae-Sung Jung, Chul-Young Kim, Finite element model updating on small-scale bridge model using the hybrid genetic algorithm, Struct. Infrastruct. Eng. 9 (5)
(2013) 481–495.
[20] Masoud Pedram, Akbar Esfandiari, Mohammad Reza Khedmati, Frequency domain damage detection of plate and shell structures by finite element model
updating, Inverse Prob. Sci. Eng. 26 (1) (2018) 100–132.
[21] Jianxiao Mao, Hao Wang, Jian Li, Bayesian Finite Element Model Updating of a Long-Span Suspension Bridge Utilizing Hybrid Monte Carlo Simulation and
Kriging Predictor, KSCE J Civ Eng 24 (2) (2020) 569–579.

12
V.N. Alves, et al. Engineering Failure Analysis 118 (2020) 104845

[22] Hanbing Liu, Xianqiang Wang, Yafeng Gong, Yubo Jiao, Damage Identification of Urban Overpass Based on Hybrid Neurogenetic Algorithm Using Static and
Dynamic Properties, Mathematical Problems in Engineering 2015 (2015) 1–10.
[23] Hui Yong Guo, Zheng Liang Li, Structural damage identification based on evidence fusion and improved particle swarm optimization, J. Vib. Control 20 (9)
(2014) 1279–1292.
[24] Aditi Majumdar, Dipak Kumar Maiti, Damodar Maity, Damage assessment of truss structures from changes in natural frequencies using ant colony optimization,
Appl. Math. Comput. 218 (19) (2012) 9759–9772.
[25] D. Ribeiro, R. Calçada, R. Delgado, M. Brehm, V. Zabel, Finite element model updating of a bowstring-arch railway bridge based on experimental modal
parameters, Eng. Struct. 40 (2012) 413–435.
[26] Kaazem Moslem, Ramin Nafaspour, Structural Damage Detection by Genetic Algorithms, AIAA Journal 40 (7) (2002) 1395–1401.
[27] N.F. Alkayem, M. Cao, Y. Zhang, M. Bayat, Z. Su, Structural damage detection using finite element model updating with evolutionary algorithms: a survey,
Neural Comput & Applic 30 (2) (2018) 389–411.
[28] Z.Y. Shi, S.S. Law, L.M. Zhang, Damage Localization by Directly Using Incomplete Mode Shapes, J. Eng. Mech. 126 (6) (2000) 656–660.
[29] Gyeong-Ho Kim, Youn-Sik Park, An improved updating parameter selection method and finite element model update using multiobjective optimisation tech­
nique, Mech. Syst. Sig. Process. 18 (1) (2004) 59–78.
[30] Seung-Seop Jin, Soojin Cho, Hyung-Jo Jung, Jong-Jae Lee, Chung-Bang Yun, A new multi-objective approach to finite element model updating, J. Sound Vib.
333 (11) (2014) 2323–2338.
[31] Joel Malveiro, Diogo Ribeiro, Carlos Sousa, Rui Calçada, Model updating of a dynamic model of a composite steel-concrete railway viaduct based on experi­
mental tests, Eng. Struct. 164 (2018) 40–52.
[32] Ladislao R. Ticona Melo, Diogo Ribeiro, Rui Calçada, Túlio N. Bittencourt, Validation of a vertical train–track–bridge dynamic interaction model based on limited
experimental data, Struct. Infrastruct. Eng. 16 (1) (2020) 181–201.
[33] A.K. Pandey, M. Biswas, M.M. Samman, Damage detection from changes in curvature mode shapes, J. Sound Vib. 145 (2) (1991) 321–332.
[34] European Rail Research Institute (ERRI) (2001). ERRI D214/RP9. Railway bridges for speeds > 200km/h. Final Report., Utrecht, Netherlands.
[35] International Union of Railways (2001). UIC 774-3. Track/bridge interaction: recommendations for calculations, Paris, France.
[36] T. Guo, Z. Liu, S. Pan, Z. Pan, Cracking of longitudinal diaphragms in long-span cable-stayed bridges, J. Bridg. Eng. 20 (2015) 04015011, https://doi.org/10.
1061/(ASCE)BE.1943-5592.0000771.
[37] Z. Liu, T. Guo, L. Huang, Z. Pan, Fatigue life evaluation on short suspenders of long-span suspension bridge with central clamps, J. Bridg. Eng. 22 (2017)
04017074, https://doi.org/10.1061/(ASCE)BE.1943-5592.0001097.

13