Railway Engineering Science

https://doi.org/10.1007/s40534-025-00393-5

Drive‑by damage detection and localization exploiting continuous

wavelet transform and multiple sparse autoencoders
Lorenzo Bernardini1 · Francesco Morgan Bono1 · Andrea Collina1

Received: 2 December 2024 / Revised: 9 April 2025 / Accepted: 10 April 2025

© The Author(s) 2025

Abstract
Drive-by techniques for bridge health monitoring have drawn increasing attention from researchers and practitioners, in
the attempt to make bridge condition-based monitoring more cost-efficient. In this work, the authors propose a drive-by
approach that takes advantage from bogie vertical accelerations to assess bridge health status. To do so, continuous wavelet
transform is combined with multiple sparse autoencoders that allow for damage detection and localization across bridge
span. According to authors’ best knowledge, this is the first case in which an unsupervised technique, which relies on the
use of sparse autoencoders, is used to localize damages. The bridge considered in this work is a Warren steel truss bridge,
whose finite element model is referred to an actual structure, belonging to the Italian railway line. To investigate damage
detection and localization performances, different operational variables are accounted for: train weight, forward speed and
track irregularity evolution in time. Two configurations for the virtual measuring channels were investigated: as a result,
better performances were obtained by exploiting the vertical accelerations of both the bogies of the leading coach instead
of using only one single acceleration signal.

Keywords Drive-by · Sparse autoencoder · Steel truss railway bridge · Continuous wavelet transform · Damage detection ·
Damage localization

1 Introduction accessibility challenges, and a strong dependence on inspec-

tors’ skills and experience [4, 5].
Bridges and viaducts play a vital role in land transporta- In response to these challenges, direct structural health
tion, serving as essential components for fostering social monitoring (SHM) systems have emerged as crucial tools
activities and driving economic growth. Ensuring their safe to complement visual inspections, gaining significant atten-
operation consistently remains a critical priority for infra- tion in both academic and industrial contexts. These sys-
structure managers, particularly as an increasing number of tems require the installation of sensors directly on the bridge
these structures is approaching or exceeding their original structure being monitored [6]. However, direct SHM systems
design lifespans [1, 2]. Traditionally, the monitoring of high- are inherently limited by their low flexibility, lack of mobil-
way and railway bridges has predominantly relied on visual ity, high installation costs, and safety risks for operators
inspections [3]. While these inspections are standardized and during installation and maintenance [7]. To address some
widely adopted, they are associated with notable limitations, of these limitations, indirect SHM systems have recently
including qualitative outputs, high costs, safety concerns, gained popularity as a promising alternative [8], following
the pioneering work by Yang et al. [9], where the authors
* Lorenzo Bernardini showed the possibility to extract bridge fundamental fre-
[email protected] quencies from a passing vehicle. Indirect SHM systems
Francesco Morgan Bono utilize instrumented vehicles that act as both exciters and
[email protected] response receivers for the bridge’s vibrations. Structural
Andrea Collina changes in the bridge deck due to damage manifest as mod-
[email protected] ifications in the dynamic interaction between the vehicle
and the bridge [10], which in turn influence the vehicle’s
1
Department of Mechanical Engineering, Politecnico di dynamic response. Often referred to as drive-by approaches,
Milano, Via G. La Masa 1, 20156 Milan, Italy

Vol.:(0123456789)
 L. Bernardini et al.

these systems offer potential cost advantages [11]: a single limitations: the approach was tested numerically without
vehicle can monitor multiple bridges along its daily route, considering variations in vehicle properties or travel speed.
requiring fewer sensors, minimizing or eliminating traffic As in [20], the vehicle in [21] is assumed to move at a con-
disruptions, and providing greater operational flexibility. stant speed of approximately 2 m/s, which could necessitate
According to Malekjafarian et al. [7], drive-by monitoring traffic disruptions or restrictions, undermining some of the
approaches can be broadly categorized into two main types: advantages potentially deriving from drive-by methods. A
modal identification and bridge condition monitoring. The laboratory-scale vehicle–bridge interaction model is used
latter can be further divided into modal-based and non- by Li et al. [22] who present a damage detection method-
modal-based approaches: this study lies in the second sub- ology using a deep autoencoder trained to reconstruct the
family. Nonmodal-based methodology can exploit different short-time Fourier transform (STFT) of the moving vehicle’s
algorithms and signal processing tools, among which con- response.
tinuous wavelet transform (CWT) has been widely adopted Moving from highway/roadway to railway applications,
in the drive-by literature, thanks to its features and promising the authors in [16] introduce a deep autoencoder frame-
results. Fitzgerald et al. [12] adopted a CWT-based damage work designed to detect scour defect, processing raw ver-
detection and localization approach to identify scour defect tical acceleration responses from a 2D train–track–bridge
in multi-span bridges with groups of train transits, exploiting interaction model under varying operational conditions.
bogie vertical accelerations. Wavelet coefficients moduli are A drive-by damage detection methodology, tailored for
averaged considering a predetermined frequency range, from high-speed railways, is proposed in [17], utilizing sparse
0.5 to 15 Hz. As well as the work by Bernardini et al. [13], autoencoders (SAEs) and Mel-frequency cepstral coef-
which is focused on single train passages on a Warren truss ficients (MFCC) to detect two distinct damage types: one
bridge, also [12] requires a reference baseline. The latter is involving end supports degradation and another, more
not required by the vehicle-scanning method described by commonly addressed in the literature, involving flexural
Demirlioglu and Erduran [14], who proposed an indirect stiffness reduction of the bridge’s beam elements. The
approach for damage detection and localization based on the method demonstrated high robustness against environ-
use of wavelet coefficients chosen to be in a frequency band mental and operational variations (EOVs), as validated
sufficiently far from bridge main bending natural frequency. through 3D train–track–bridge interaction simulations.
In the last decade, a marked trend toward the massive use It is important to stress that both [16, 17] refer to con-
of data-driven approaches has been observed in the field of crete bridge structures. Steel truss structures, despite their
SHM, applied to civil structures [15]. In this case, unsu- widespread use, global significance, and average age [2,
pervised learning algorithms are advantageous with respect 23], have been less extensively studied in the related liter-
to supervised ones, since for large civil structures such as ature. Addressing this gap, Bragança et al. [24] evaluated
bridges, dams or viaducts, comprehensive data set are often the use of freight train accelerations to identify damage
unavailable [16]. This aspect led to the fact that a number of in the side diagonals and lower chords of a Warren truss
authors has started focusing on the use of autoencoders for bridge. Acceleration data were collected at eight loca-
drive-by damage detection, whose majority of applications tions on the freight wagon, from which wavelet scattering
are still related to road vehicles [17]. In [18], the authors coefficients were derived. These coefficients were then
introduce a deep autoencoder architecture that exploits reconstructed using eight separate autoencoders. The nor-
multiple road vehicle passages to monitor the health condi- malized reconstruction errors were subsequently fused to
tion of a highway bridge. The selected damage index (DI) create a highly sensitive damage indicator, demonstrating
is based on the Kullback–Leibler (KL) divergence, which the numerical effectiveness of the approach.
evaluates changes in the distribution of mean reconstruc- It is clear that in drive-by literature, SAEs have been typi-
tion error values (MAE). Similarly, Kaur et al. [19] utilize cally adopted for damage detection, but not used for damage
an adversarial autoencoder (AAE) to detect and localize localization, which is crucial in the context of direct and
damage using continuous wavelet transform (CWT) maps indirect SHM. In this work, we aim to present a drive-by
derived from road vehicle responses. In alignment with [20], damage detection and localization methodology based on
the analyzed frequency band in the CWT map is chosen to wavelet coefficients, computed from vertical bogie accel-
include the first bridge bending frequency component. Hur- erations and multiple sparse autoencoders. According to
tado et al. [21] proposed an adversarial autoencoder (AAE) authors’ best knowledge of the presented literature, main
approach to indirectly detect damage in highway bridges. contributions and differences with respect to past works can
Their method utilizes the residual acceleration spectrum, be listed as follows:
processed through filtering and averaging, as input to the
AAE. While the AAE demonstrated superior performance • As mentioned before, the majority of previous works deal
compared to stacked autoencoders, the study has certain with drive-by damage detection methodologies, while in

Railway Engineering Science

Drive‑by damage detection and localization exploiting continuous wavelet transform and…

this case, damage localization is also performed. Com- 2 Models

pared to [19], several differences can be observed, in
terms of methodology, signal processing and application. The adopted simulation tool, named ADTreS, is a non-
An important distinction relies in the simulated forward commercial software developed, over the years, within the
speed, which in this work is at least 40 times larger than Department of Mechanical Engineering at Politecnico di
the one adopted in [19]; Milano [26]. Concerning rail vehicle dynamics, compari-
• The wavelet coefficients band of interest, differently from sons between numerical results, obtained through ADTreS,
[12, 14], are directly suggested by the modular structure and experimental data are reported in [27, 28]. Instead, a
of the truss bridge; comparison between numerical and experimental data, in
• In this work, Hotelling’s distance is used for damage terms of bridge vertical displacement during train transit, is
detection and localization. While T 2 statistic, computed reported by Bruni et al. [29]. This software provides capa-
at latent space level, has already been used in direct SHM bilities for:
(e.g., [25]), but its adoption for railway bridge health
assessment cannot be found yet in published drive-by 1. Multi-body modeling of each vehicle within the train;
literature. 2. Finite element modeling of bridges and track structures;
3. Time integration of the equations of motion, accounting
The simulations were conducted using 3D TTB dynamic for transient phenomena and system nonlinearities.
interaction software, incorporating a 3D model of a
Warren truss bridge that has been dynamically validated The main components of the adopted software and their
through experimental measurements, along with a set of interaction are schematically shown in Fig. 1, and they are
train models designed to reflect the characteristics of explained in detail in the following subsections.
actual trains operating on the actual structure. The paper
is structured as follows: Section 2 is dedicated to the 2.1 Train and track models
illustration of the models adopted. Section 3 describes
in detail the methods for damage detection and localiza- The rail vehicle (see Fig. 2) is modeled using a multi-body
tion. Outcomes are shown in Sect. 4 and 5, while main approach, based on a standard configuration in which each
achievements and limitations are critically discussed in carbody is supported by two bogies and four wheelsets.
Sect. 6. In this analysis, the carbody, bogies, and wheelsets are

Rail vehicle model Wheel–rail contact forces

Track superstructure Bridge FE model

Fig. 1  Main components of ADTreS software [30]. From top-left, in clockwise direction, coach multi-body model, contact forces, bridge FE
model and track superstructures are depicted

Railway Engineering Science

 L. Bernardini et al.

ide view ear view

βc
βc c
c
c c c

β2 β1 2
2 2 1 β2 2
1 1
2 1 2
a
βa,4 βa,3 βa,1
βa,2 a,4 a,4
a,4

op view

Fig. 2  Rail coach 3D scheme with its degrees of freedom. Side, top and rear views. Vertical degrees of freedom are highlighted by the letter z,
lateral with y, while 𝜌, 𝛽 and 𝜎 stand for roll, pitch and yaw rotations. Please notice that 𝛽a,i , which represents the rotation about the ith wheelset
axis, accounts for contributions to creepage associated with curve negotiations

considered rigid bodies. Each coach in the train is repre- neglected, and vertical stiffness is set to 120 MN/m. Outside
sented with a three-dimensional model, where primary and the bridge, concrete sleepers are used, modeled as concen-
secondary suspensions are simulated using linear springs trated masses of 315 kg, with rail pad properties assumed
and dampers. The coach’s motion is characterized by 35 to match those on the bridge. The sleeper spacing is set to
degrees of freedom, with each component (i.e., wheelsets, 0.6 m both on and off the bridge, with rail mass per unit
bogies, and carbody) contributing for 5 degrees of freedom: length set equal to 60 kg/m.
two translational (vertical and lateral) and three rotational The initial ballasted section is specifically designed
(yaw, pitch and roll rotations). The longitudinal degree of to prevent artificial transients in the vehicle’s vertical
freedom is omitted, since the train is assumed to travel at a acceleration.
constant forward velocity.
The train configuration used in this study represents a 2.2 Bridge model
rail vehicle with 32 axles crossing the analyzed structure.
It consists of eight coaches, each 24.9 m in length, with a The finite element (FE) model examined in this study rep-
wheelbase of 2.85 m and a spacing of 17.4 m between the resents an existing bridge structure on a regional railway
bogie centers of mass. This model is a sort of digital replica line in northern Italy (see Fig. 3). This structure, which is
of an actual passenger train that regularly operated on the still in operation, is composed of four Warren truss spans,
bridge. This specific commuter train type was also identified and it was constructed in 1946. More in detail, the bridge
by the weigh-in-motion (WIM) system installed near the comprises two parallel structures, one for each travel direc-
entrance of the bridge [31]. tion, supported by a shared pier. These structures include
The bridge features a ballastless track system with two spans with different lengths: the longer span crosses the
wooden sleepers, which are treated as infinitely rigid in riverbed, while the shorter span extends over the floodplain.
the model. Rail pads are represented by 3D springs, with a To reduce computational demands, only the longer span
vertical stiffness of 350 MN/m and a lateral stiffness of 67 was modeled in this work, neglecting both the shorter span
MN/m, connecting the rail nodes to the beam elements form- and the central pier, as their contributions were considered
ing the bridge stringers. Beyond the bridge, however, track negligible for the purpose of this study. Figure 4 shows the
sections with ballast are modeled, where the ballast mass is finite element (FE) model of the bridge used in the present

Railway Engineering Science

Drive‑by damage detection and localization exploiting continuous wavelet transform and…

Bridge side of interest

Fig. 3  Actual truss railway bridge. The longer spans are enboxed in red: one of them, precisely the one on the upstream side, is the one modeled
in this work

L1A = 2.35 Hz V1= 4.28 Hz L1B=4.31 Hz

z
L2A=5.03 Hz L2B=8.07 Hz
y x

Fig. 4  Bridge FE model of the longer span (see Fig. 3) and first five main numerical mode shapes

paper, together with the first five numerical mode shapes. given the year of construction of the bridge, these proper-
The latter are defined with the following nomenclature: the ties are considered reasonable. To account for components
first letter, L or V, refers respectively to lateral or vertical not included in the net cross-sectional area, such as bolts
mode shapes. Precisely, with L, we identify modes that are and connecting plates, the density was artificially increased
mainly lateral and the same goes for vertical ones with V. by the 15%. Additional concentrated masses were applied
This first letter is then followed by a number which iden- to the bridge deck to represent non-structural elements like
tifies the order of the considered mode. Finally, a letter, timber sleepers, track plates, handrails, and footbridges
A or B, is used in case different modes of the same order, along the bridge’s sides. Proportional damping, using the
with same predominant motion (lateral or vertical), appear. Rayleigh model [33], was applied, with coefficients set to
The structure is modeled with ideal hinge and roller achieve damping ratios consistent with the recommendations
boundary conditions and features a ballastless track, as men- presented in [34].
tioned before. The model is mainly made of Euler-Bernoulli The model was calibrated and validated primarily through
beam elements, each with six degrees of freedom per node dynamic measurements [35] obtained from a permanent
[32]. Geometric properties were derived from previous field monitoring system installed on the bridge [36]. As part of
surveys and technical drawings, referred to the bridge’s long a project with the infrastructure manager [37], a perma-
span. nent monitoring system was set up on the upstream bridge
The steel material is assigned a Young’s modulus of 200 (both on short and long spans), in April 2022. As shown in
GPa, a Poisson’s ratio of 0.3, and a density of 7850 kg/m3: Table 1, there is a strong agreement between the numerical

Railway Engineering Science

 L. Bernardini et al.

model their dynamic interaction. The equations of motion

0.03
Num 2.35 Hz 0.99 0.01 0.37 0 0.01 0.03
for the ith vehicle are given by

Num 4.28 Hz 0 0.99 0.01 0 0 0 0 Mvi ẍ vi + Cvi ẋ vi + K vi xvi

( ) (1)
= Qcvi xs , ẋ s , ẋ v1 , ẋ v1 , xv2 , ẋ v2 , … , xvi , ẋ vi , … , xv8 , ẋ v8 , t .
Num 4.31 Hz 0.38 0.01 0.99 0.01 0 0.06 0
where Mvi , Cvi , and K vi represent the mass, damping, and
Num 5.03 Hz 0 0 0 0.98 0.56 0 0 stiffness matrices of the ith vehicle, and xvi denotes the vec-
tor collecting vehicle’s degrees of freedom. The vector Qcvi
Num 8.07 Hz 0 0 0 0.6 1 0 0
contains the generalized contact forces for the vehicle.
For the bridge–track structure, the equations of motion
Num 9.07 Hz 0.08 0 0.09 0.01 0 0.99 0.74
can be expressed as follows:
Num 11.80 Hz 0 0 0.01 0 0 0.85 0.99 Ms ẍ s + Cs ẋ s + K s xs
( )

Exp 11.32 Hz
Exp 2.24 Hz

Exp 4.20 Hz

Exp 4.42 Hz

Exp 5.00 Hz

Exp 8.10 Hz

Exp 8.71 Hz
= Qcsv xs , ẋ s , ẋ v1 , ẋ v1 , xv2 , ẋ v2 , … , xvi , ẋ vi , … , xv8 , ẋ v8 , t .
(2)
where Ms , Cs , and K s are the structure’s mass, damping,
and stiffness matrices, and xs represents the vector of the
structure’s free coordinates. The vector Qcsv contains the
Fig. 5  MAC matrix comparing numerical and experimental mode
shapes for the first seven frequencies generalized contact forces defined by the relative motion of
the structure and the vehicle.
The coupling between the train and the track–bridge
structure is achieved through contact forces exchanged at
Table 1  Experimental and numerical frequencies for the first seven
mode shapes each wheel–rail interface. The resolution of wheel–rail
contact is split into normal and tangential problems. In the
Modes Exp. Freq. (Hz) Num. Freq. (Hz) Difference (%)
normal problem, a multi-Hertzian [39] wheel–rail contact
L1A 2.24 2.35 4.9 model is applied. Before simulation, either measured or
V1 4.20 4.28 1.9 theoretical wheel and rail profiles are used to calculate a
L1B 4.42 4.31 −2.5 contact table [40] containing parameters needed for cal-
L2A 5.00 5.03 0.6 culating contact forces. This evaluation is based on the
L2B 8.10 8.07 −0.4 transverse profiles of both the wheel and rail, generating a
L3A 8.71 9.07 4.1 table of geometric contact parameters as a function of the
L3B 11.32 11.80 4.2 relative lateral displacement between the wheel and rail.
Key parameters include the contact angle, wheel rolling
radius, and the local curvature radii of the wheel and rail
profiles [40].
and experimental natural frequencies with maximum dis-
Once the normal force is calculated, tangential forces
crepancies under 5%, for the structure’s first seven modes.
are derived using the Shen-Hedrick-Elkins creepage-based
Concerning the latter, Fig. 5 shows the modal assurance cri-
formulation [41]. ADTreS supports multiple simultaneous
terion (MAC) [38] matrix obtained by comparing numerical
wheel–rail contact points, with options to incorporate both
and experimental modes: high similarity can be observed,
new and worn wheel–rail profiles and to account for track
with maximum differences of 2%.
geometry irregularities. Time integration is performed
with a modified Newmark approach [32], which is now
2.3 Equations of motion and time integration
briefly recalled.
Let us consider zk as the vector containing structure or
A s p rev i o u s ly m e n t i o n e d , t h e s i m u l a t i o n o f
the ith vehicle degrees of freedom at the kth time step. The
train–track–bridge (TTB) dynamic interaction is performed
method begins computing the solution for accelerations
using ADTreS [26]. This software enables finite element
(vector z̈0 ) corresponding to the first time step, named
(FE) modeling of track and bridge structures, allowing flex-
t0 . Under this regard, it is necessary to set the starting
ibility in selecting different FE element types, such as Euler-
positions and speeds in the vectors z0 and ż 0 respectively.
Bernoulli beam, plate, or bar elements. The rail vehicle, on
Subsequently, through an iterative procedure starting from
the other hand, is modeled using a multi-body approach.
the solution computed at the previous time instant, the
The structure and train are represented as two distinct
solution is calculated at the following one. The adopted
sub-systems, coupled through contact forces, to properly

Railway Engineering Science

Drive‑by damage detection and localization exploiting continuous wavelet transform and…

version of the integration resolves first in acceleration, and ż k+1 = ż k + (1 − 𝛽)Δẗzk + 𝛽Δẗzk+1 . (6)
then it calculates velocity and displacement. Therefore,
after choosing suitable values for Δt , 𝛼 and 𝛽 parameters
[42], main steps to solve the equations of motion are writ- zk+1 = zk + Δtż k + (0.5 − 𝛼)Δt2 z̈k + 𝛼Δt2 z̈k+1 . (7)
ten as follows:
Please notice that M , C and K are respectively mass,
damping and stiffness matrices of the ith vehicle or of the
1. Starting solution is obtained by solving the system:
structure. To properly account for nonlinearities related
M z̈0 = (Q0 − C ż 0 − K z0 ), (3) to contact forces and suspensions/track systems, a modi-
fied time step approach must be considered [32], as shown
where the vector Q0 contains the initial values of the schematically in Fig. 6. More in detail, with respect to a
generelized lagrangian components written on the right traditional Newmark integration procedure, adopted for
hand side of the Lagrange equation. linear systems [32], in this case an internal loop is added,
2. Calculation of acceleration vector z̈k+1 for the subsequent enboxed in green in Fig. 6. This part of the numerical
time step solving the system: process is executed considering frozen time integration
( ) which is necessary being contact forces nonlinear. There-
D z̈k+1 =Qk+1 − K zk + Δtż k + (0.5 − 𝛼)Δt2 z̈k fore, at the beginning of the generic jth time instant, the
( ) (4) first approximation (q = 1) of the contact point motion
− C ż k + (1 − 𝛽)Δẗzk ,
is calculated, considering the wheels in the new position.
After generalized contact forces computation, Equations
where D matrix is defined as (1)–(2) can be integrated. Subsequently, a new approxima-
D = M + 𝛽ΔtC + 𝛼Δt2 K. (5) tion (at q = q + 1) of the contact point motion is obtained
and the contact forces are calculated again, considering
3. Calculation of velocity and displacement at k+1 step: frozen time. Therefore, an iterative procedure is performed
within the j-th time step. When convergence is reached for

Generalised forces on each Generalised forces on the

Contact forces calculation: vehicle (Eq.(1)) structure (Eq.(2))
, tk , l

Integration of each rail Integration of structure

vehicle equation (Eq.(1)) equation (Eq.(2))
Contact point motion

Contact point
vi ( +1 ) s ( +1 )
s( + Δ ) ≈ s( ) motion
v( +Δ )≈ v ( )
ξm , +1 = ξm , + V ∙ Δ
Normal contact force
calculation

+1 = +Δ
Longitudinal and transverse contact forces (in the local contact
plane) calculation: l , t

No Contact forces
= +1
convergence
Yes

Fig. 6  Iteration loop procedure to handle contact nonlinearities. 𝜉m represents the position of the mth wheel along the rail, while V is the train
forward speed. Please notice that xs and xvi represent the vectors containing structure and the ith vehicle free coordinates, respectively. Moreover,
Nk , Flk and Ftk identify normal, longitudinal and transverse contact force components at the mth wheel

Railway Engineering Science

 L. Bernardini et al.

the contact forces, it is possible to exit the internal loop, 3 Methods

meaning that the time step can be now incremented and the
integration can proceed further. This section is devoted to the detailed description of the
developed damage detection and localization algorithm. The
2.4 Damage model procedure, schematized in Fig. 8, is mainly composed of the
following steps:
This study focuses on simulated damage scenarios associated
with corrosion, one of the primary causes of material degrada- 1. Vertical bogie accelerations undergo continuous wavelet
tion in steel structures [43]. Corrosion is an electrochemical transform [45], considering a complex Morlet wavelet.
process in which metals react with the environment, leading to 2. Average wavelet coefficients curve is constructed consid-
a reduction in the cross-sectional area of affected elements. In ering a frequency range defined by the module passing
steel structures, this degradation compromises both structural frequency [46]. The latter is computed as train forward
integrity and performance [44]. speed divided by bridge module length.
In this work, corrosion is modeled as a simultaneous 3. Resampling in space, according to a constant spatial
reduction in both density and Young’s modulus to account step.
for material loss as well as reductions in flexural and 4. Adoption of a number of sparse autoencoders (SAEs)
axial stiffness. A set of damage scenarios, outlined in [47] equal to the number of bridge modules, trained to
Fig. 7, is investigated. Four damage levels were consid- reconstruct different parts of the curve constructed at
ered, 10%, 15%, 25%, and 50%, with the damage applied point 2.
to two distinct structural components: cross-girders and 5. Calculation of Hotelling’s distances [48], at autoencoder
stringers, each analyzed individually. bottleneck, to extract information regarding damage
For damage type SD (applied to stringers), both string- presence and its location.
ers were degraded by the same percentage, with two dif-
ferent locations investigated (denoted by subscripts 2 and In other words, the idea is to divide the average wavelet coef-
3). For damage type CG (applied to cross-girders), three ficients curve, computed over the frequency band of interest,
positions were considered (denoted by subscripts 1, 2, according to the number of modules composing the bridge.
and 3) to evaluate the influence of damage location on By training N different sparse autoencoders (SAE) networks
the detection algorithm.

CG3,10%
Cross girder CG3,15%
CG1,50% CG2,50%
CG3,25%
CG3,50%

SD3,10%
SD3,15%
SD2,50% SD3,25%
SD3,50%

Fig. 7  Corrosion damage scenarios: types, positions, and intensities. Top view of the bridge deck FE model

Railway Engineering Science

Drive‑by damage detection and localization exploiting continuous wavelet transform and…

(T 2 )

(T 2 )

(T 2 )

Fig. 8  Damage detection and localization algorithm flowchart, with its main steps. The procedure starts from bogie vertical acceleration, con-
verted into a scalogram through CWT operation. Subsequently a frequency range of interest is selected, based on the value of fp which repre-
sents the module passing frequency, defined by the ratio between train speed (V) and the bridge module length (d)

on N different segments of the input curve, where N is the signal (frequency components with their magnitudes and
number of modules composing the truss structure, it is pos- phases) without any time-localization capability, making
sible to derive region located information that can be used it generally less effective for identifying temporally (and
to localize damage. It is important to notice that the average thus spatially) damage locations. The difference between
wavelet coefficients curve is actually considered on a spatial a fast Fourier and wavelet transforms mainly lies in the
window that is 5% larger than the bridge span. Therefore, the orthogonal basis functions, which in the former case are
seven segments in which it is divided does not exactly coin- represented by sines and cosines, while for the latter are
cide with the seven modules composing the Warren truss small oscillating waves [45]. Wavelet transform can be
bridge, but the difference is considered absolutely negligible. written as follows:
+∞

�s� ∫−∞
3.1 Continuous wavelet transform 1 t−𝜏
� �
W(𝜏, s) =< f (t), 𝜓 ∗ >= √ ⋅ f (t) ⋅ 𝜓 ∗ dt,
s
As shown in Fig. 8, for each simulated scenario, the ver- (8)
tical bogie acceleration data is processed using the 1D where f(t) is the time-series signal we aim to transform (i.e.,
continuous wavelet transform (CWT) [45], a powerful and bogie vertical accelerations, see Fig. 8), while 𝜓 ∗ represents
well-known time-frequency analysis tool. This choice is the basis function that can be changed in width (enlarged or
mainly due to the capability of the CWT to be intrinsically shrunk), by scaling with s, and translated in time through 𝜏 .
better suited for the processing of non-stationary signals. More precisely, higher scale values lead to expanded wave-
In fact, CWT offers several advantages over the fast Fou- lets, while small values gives shrunken wavelets. As a result
rier transform (FFT) when analyzing time-series data for W(𝜏, s) represents the wavelet coefficient associated to the
extracting damage-sensitive features. Specifically, CWT scale
( s )and to the time translation 𝜏 . It is now clear that
provides a time-frequency representation of the signal,
𝜓∗ t−𝜏
is nothing but the scaled and shifted version of the
enabling the analysis of its frequency content alongside s

its temporal evolution, in a multi-resolution fashion [49]. mother wavelet 𝜓(t). In the present paper, a complex Morlet
In contrast, FFT only delivers frequency information of a wavelet was adopted, described by [12]

Railway Engineering Science

 L. Bernardini et al.

2
1 − Ft
𝜓(t) = √ ⋅e B ⋅ ei2πFc t , (9)
π ⋅ FB

where two input parameters can be clearly distinguished,

namely the bandwidth FB and the center frequency Fc . As
reported in [50], FB determines how localized is the wavelet
in frequency domain, while Fc represents the frequency that
characterizes wavelet global behavior. Despite there are no
predetermined rules for choosing one type of wavelet with
respect to another, it is true that there exists a number of
studies in drive-by literature where the authors utilized a Fig. 9  Autoencoder schematic representation. From left to right,
encoder, latent space (bottleneck), and decoder components are
Morlet wavelet (e.g., [12, 13]). Further theoretical insights shown
on wavelet transform can be found directly in [45].
Going into detail with the present work, we assumed
the same center frequency and bandwidth adopted by the encoder is in charge of compressing input data (i.e., Wavg, see
authors in [12], meaning that a bandwidth FB of 1 and a Eq. (10)) into a lower-dimensional representation that captures
center frequency Fc equal to 1.5 were set. The CWT was the essential features related to inputs. Conversely, the decoder
performed exploiting the Python function cmor, based on reconstructs the input signal starting from the lower-dimen-
the work by Torrence and Compo [51]. Compared to [12], sional input representation at the bottleneck. The autoencoder
in this work, the focus is on a frequency range which is not is trained to minimize discrepancies between original signal
predetermined (0.5–15 Hz in [12]), but rather depends on and its reconstructed version, considering the mean absolute
two factors, which are the bridge modular geometry (the error (MAE).
extension of each module of the truss structure) and train Sparse autoencoders (SAEs) have a strong capability in
forward speed. Therefore, the passing frequency ( fp = V∕d ), learning sparse representation of data, outlining most relevant
whose concept was introduced in the drive-by context by features during encoding process [54, 55]. This is obtained by
Bernardini et al. [46], is the main driver of this stage of the introducing a sparsity constraint during its training, with the
methodology (shown in all its steps in Fig. 8). Precisely, purpose to enable the model to outline fundamental details and
wavelet coefficients (W in Eq. (8)) modules obtained in the aspects of the input signal, distinguishing them from noise and
frequency band ranging from 0.5 ⋅ fp to 2 ⋅ fp are extracted disturbances [54]. Assuming that the autoencoders learning is
and, subsequently, averaged among frequency direction, as accomplished by minimizing a function P(z, g(f (z))), a sparse
explicated by Eq. (10): penalty 𝛺(x) function is included in the function to be mini-
mized [25, 54], called L, and defined as follows:
Nf
1 ∑|
Wavg (𝜉) = W(𝜉, fk )||, (10) L = P(z, g(f (z))) + 𝛺(x), (11)
Nf k=1 |
where z is a vector representing the network inputs, x = f (z)
where Nf is the number of frequencies contained in the is the output of the encoder (i.e., the latent space), being f (z)
frequency interval of interest, while 𝜉 defines the spatial and g(x) the functions that represent encoder and decoder
abscissa, once resampled. Each Wavg curve (see Eq. 10) is operations, respectively. The loss function P is based on
composed of 1400 points that ensures a spatial discretization mean absolute error (MAE) between SAE inputs (actual)
lower than 0.05 m. and outputs (predicted), while 𝛺(x) introduces a constraint
on the latent layer, limiting the amount of active neurons in
3.2 Multiple sparse autoencoders the bottleneck x [25, 56]. The layers forming both encoder
and decoder are shown in Fig. 10. This NN structure was
An autoencoder (AE) [47] is a type of unsupervised artificial tuned through Bayesian optimization process. It is important
neural network (ANN), since automatically learns features to stress that the encoder input, i.e., Wavg is processed with
from unlabeled data, which, as mentioned before, is a great z-score transformation, expressed as follows:
advantage when dealing with damage detection and locali-
zation applied to large civil structures [52, 53]. Autoencoder Wavg − 𝜇
zw = , (12)
models try to describe large amount of data in an efficient 𝜎
and compact manner, i.e., with as few meaningful features as
where 𝜇 and 𝜎 are respectively the mean value and the stand-
possible. The basic architecture, shown in Fig. 9, is composed
ard deviation of the Wavg population in the training set. In
of three parts: encoder (purple part), latent space (even called
this way, the inputs distribution in the training have null
bottleneck, colored in blue) and decoder (drawn in green). The

Railway Engineering Science

Drive‑by damage detection and localization exploiting continuous wavelet transform and…

tion to ensure compatibility with the continuous nature

of the output data.
• Regularization term coefficient: A value of 𝜆 = 0.001 was
incorporated to prevent overfitting and ensure a smooth
Conv1D
optimization process.
MaxPooling1D • Sparsity regularization coefficient: A value of 𝛽 = 0.05
was applied to encourage the sparse activation of hidden
units, aligning with the model’s sparse nature. Please
Conv1D

MaxPooling1D notice that this 𝛽 parameter is not related to the one con-
cerning Newmark integration method explained before
Dropout
(see Sect. 2.3).
Flatten • Percentage of activation of hidden units: The sparsity
constraint aimed to activate 10% of the hidden units,
ensuring a compact representation while maintaining
Dense sufficient feature extraction capability.
• The model was trained for a maximum of 200 epochs,
with early stopping applied based on validation loss to
Dense
avoid overfitting with a patience of 20 epochs.
Flatten

Reshape 3.3 Hotelling’s T2 distance

Upsampling1D
Multivariate control charts are valuable tool in the context of
Conv1D
structural health monitoring (SHM), given their capability to
detect anomalies or deviations from baseline behavior, when
MaxPooling1D multiple features and variables are taken into account. An
Conv1D
interesting multivariate statistical process control (MSPC)
tool is represented by Hotelling’s T 2 control chart [48, 57,
58]. This statistical parameter has been used for direct SHM
approaches, such as [57, 59, 60], to monitor non-modal or
modal parameters. Therefore, in the context of civil struc-
Fig. 10  Architecture of the adopted sparse autoencoder, whose layers tures, this tool can be used to monitor modal or non-modal
are reported
(present case) features. The versatility and broad applicabil-
ity of T 2 statistic distance establish it as a powerful method
for detecting and quantifying changes in analyzed signals.
In the present work, this distance is computed at the latent
mean and unitary standard deviation. The architecture shown space level, in which a lower-dimensional representation of
in Fig. 10 takes in input the normalized version of Wavg the autoencoders input is found. Therefore, the idea is to
(i.e., zw ), where the latter is composed of 1400 points, as detect anomalies by looking to changes in the outputs pro-
explained in Sect. 3.1. Then, this number normalized curve vided by the encoder, quantified by the T 2 distance. More in
is divided into seven segments, composed of 200 points detail, considering a matrix XNt XM containing Nt observa-
each, which are then reduced, for each SAE, to a total num- tions (corresponding to Nt train passages) of the M features
ber of features, at the bottleneck layer, equal to multiples of contained in the latent space of the SAE, the kth T 2 statistic
16 (up to 64). These features are then used by the decoder is defined as follows [25]:
to reconstruct the input signal.
To ensure clarity and reproducibility, the key character- Tk2 = r(̄x − x̄ )T 𝜮 −1 (̄x − x̄ ) k = 1, 2, … , Nt ∕r, (13)
istics of the sparse autoencoder (SAE) model, used in this
study, are detailed below: where r is an integer parameter referred to as subgroup size
that determines the Nt ∕r number of Tk2 that can be computed
• Activation function: The rectified linear unit (ReLU) from the matrix XNt XM , and x̄ is the vector containing the
activation function was employed for the encoding and mean values of the latent space for the kth subgroup contain-
intermediate decoding layers to improve computational ing r observations. Moreover, x̄ and 𝜮 are respectively the
efficiency and mitigate vanishing gradient issues. The mean values vector and the covariance matrix of the latent
last layer of the decoder uses a linear activation func- space statistically estimated in the training period. To detect

Railway Engineering Science

 L. Bernardini et al.

anomalies, first step is the definition of an in-control region: under varying operational conditions. To achieve this, the
in fact, any deviation from normal processes that can be due authors considered the following key variables:
to damage presence is detected if data points fall outside
this region. The latter is defined by the upper control limit • track irregularity,
(UCL), which simply represents the statistical distance cor- • train mass, and
responding to a certain significance level 𝛼 , defined from • train speed.
the training data set. In other words, the in-control region
is the interval [0, UCL], given that computed distances are More in detail, a first track irregularity profile was obtained
inherently positive quantities. In the present paper, the UCL as a random spatial realization of the PSD curve referring
is set as the T 2 95th percentile calculated considering the to longitudinal profile provided by [61], accounting for low-
training set, meaning that, in the training set, values greater level defects. Precisely, the adopted PSD curve is expressed
than the UCL can be observed only the 5% of the time by as a function of the spatial circular frequency Ω:
chance [25]. This assumption was made to have an appropri-
ate sensitivity to the onset of a damage as well as robustness AV ⋅ 𝛺c2
SV (𝛺) = ( ), (14)
to false positives. This choice is in line with what proposed
) (
𝛺2 + 𝛺r2 ⋅ 𝛺2 + 𝛺c2
by other papers, such as [25, 60].
The drive-by methodology is tested considering a single where 𝛺c is equal to 0.8246 rad/m, 𝛺r to 0.0206 rad/m and
(first bogie vertical acceleration) and two virtual measure- AV to 4.032 × 10−7 m ⋅rad [61]. The last parameter, AV ,
ment channels (first coach bogies vertical accelerations), denotes that as mentioned before, low-level vertical irregu-
respectively. In the first case, the encoder receives the Wavg larity was considered in this work. This assumption is rea-
curve as input and is trained to extract a series of features sonable since on the truss bridge the track is ballastless.
from it, which the decoder then uses for reconstruction. In Starting from this first reference profile, other four profiles
the second case, the operating principle is exactly the same, are realized, considering an increase in amplitude and phase,
but, here, the encoder receives two curves as input (instead with the aim to model track evolution in time. The increase
of one) and extracts a series of features (at the bottleneck in terms of amplitude and phase depends on the spatial
layer) which the decoder uses to reconstruct both of them. wavelength of the defect taken into consideration. In par-
ticular, reference amplitudes and phases are increased of a
quantity whose maximum value is 5% for defects larger than
3.4 Plan of simulations 3 m, while 10% was set for defects lower than this value. The
idea is to model a differential evolution in time for smaller
Before presenting and discussing the outcomes of this work, and larger defects, where smaller defects are assumed to
the plan of simulations is recalled. The methodology out- change faster than larger.
lined in this work was validated by analyzing its performance

180

170

160
Axle load (kN)

150

140

130 Train 1
Train 2
120 Train 3
Train 4
Train 5
110
1 4 5 8 9 12 13 16 17 20 21 24 25 28 29 32
Axle No.

Fig. 11  Axle loads for the five passenger trains included in the simulation set (train 1 to 5) are presented. The solid blue line represents the mean
axle load for the convoy, calculated from a dataset of 268 corresponding real train passages, while the magenta dashed lines indicate the ±3𝜎
confidence intervals

Railway Engineering Science

Drive‑by damage detection and localization exploiting continuous wavelet transform and…

Concerning train mass, five different convoys are consid- is plotted for each segment, and each damage scenarios of
ered in this work, whose axle loads, shown in Fig. 11, are interest, as a function of train groups. In particular, in this
plotted against the experimental values, estimated through work, batches of 40 train transits are considered to make
a WIM system that was installed before the entrance of the the methodology more robust against disturbing operational
actual bridge [31, 37]. In addition, a set of train speeds is variables. With the term normalized T 2 distance (Tnd
2
), in this
considered in the range 80 and 130 km/h. work, we referred to the following quantity:

Td2
2
Tnd
ijk
, (15)
4 Results and discussion i,j,k
=
UCLi

This section of the paper is dedicated to the presentation where i refers to the ith segment, j to the jth damage sce-
and critical discussion of the results. First of all, outcomes nario, and k to the kth 40 trains group, while UCLi to the
obtained considering one single sensing point on the first upper control limit defined in the training set (see Sec. 3.3)
bogie of the leading coach are presented. Subsequently, for the i-th segment. This quantity enhances the understand-
results regarding the case with two sensing points, corre- ing of the collocation of the T 2 distance with respect to the
sponding to the two bogies of the leading vehicle, are shown. UCL , where the latter is different for each segment con-
sidered. Therefore, Eq. 15 defines the damage index (DI)
4.1 Results considering one sensing point adopted in this work. In Fig. 12, the legend points out the
type of damage and the modules that are adjacent to that
This section shows the outcomes of the presented method- degraded element, or in other words, the modules that share
ology accounting for only one sensing point (i.e., virtual this truss member. Specifically, to make an example, the
measuring point), which corresponds to the first leading CG1,50% regards a cross-girder shared by the second and the
bogie, whose vertical acceleration is used. To start, three third module. At this point, it is important to recall that the
different damage scenarios are considered, namely CG1−3,50% signal window under consideration is not exactly just limited
which regards the case of a 50% corrosion intensity affect- to bridge extension, but it accounts for a slightly larger dis-
ing three distinct cross-girders (one at a time), as shown in tance. This because the vertical acceleration are originally
Fig. 7. More in detail, in Fig. 12, the normalized T 2 distance cut considering the time instant at which the first axle of

Segment 1 Segment 2 Segment 3 Segment 4

50 50 50 50

40 40 40 40
T2 distance/UCL

T2 distance/UCL
T2 distance/UCL

T2 distance/UCL

30 30 30 30

20 20 20 20

10 10 10 10

0 0 0 0
0 5 10 0 5 10 0 5 10 0 5 10
Batch number Batch number Batch number Batch number
Segment 5 Segment 6 Segment 7
50 50 50

40 40 40
T2 distance/UCL
T2 distance/UCL

T2 distance/UCL

CG1,50% (S2–3)
30 30 30 CG2,50% (S4–5)
CG3,50% (S6–7)
20 20 20
UCLn
10 10 10

0 0 0
0 5 10 0 5 10 0 5 10
Batch number Batch number Batch number

Fig. 12  Damage scenario: corrosion affecting cross-girders, highest intensity and three different positions (see Fig. 7). Normalized T 2 distance,
plotted for the seven segments composing the bridge window of interest, as a function of the train groups. Batches of 40 transits are considered.
One vertical acceleration is used

Railway Engineering Science

 L. Bernardini et al.

Segment 1 Segment 2 Segment 3 Segment 4

10 10 10 500

8 8 8 400

T2 distance/UCL
T 2 distance/UCL

T 2 distance/UCL

T 2 distance/UCL
6 6 6 300

4 4 4 200

2 2 2 100

0 0 0 0
0 2 4 6 0 2 4 6 0 2 4 6 0 2 4 6
Batch number Batch number Batch number Batch number

Segment 5 Segment 6 Segment 7

450 450 450
400 400 400
350 350 350

T 2 distance/UCL
T2 distance/UCL

T 2 distance/UCL

300 300 300

250 250 250 SD2,50 % (S4)
200 200 200 SD3,50% (S7)
150 150 150 UCLn
100 100 100
50 50 50
0 0 0
0 2 4 6 0 2 4 6 0 2 4 6
Batch number Batch number Batch number

Fig. 13  Damage scenario: corrosion affecting stringers, highest intensity and two different positions (see Fig. 7). Normalized T 2 distance, plot-
ted for the seven segments composing the bridge window of interest, as a function of the train groups. Batches of 40 transits are considered. One
vertical acceleration is used

the bogie of interest enters the bridge, and the last axle of following ones, the number of normalized distances cal-
the same bogie exits it. Therefore, once moved to space (V culated for each damage scenario are not equal in number,
is constant), the acceleration signal extends for a quantity being the amount of simulations different.
which is equal to the bridge length plus the wheelbase dis- From this outcome, it seems that detecting damage which
tance. However, the latter represents less than the 5% of the affects a portion of the bridge that is closer to its entrance is
bridge span length, and it is not trimmed from the signal a more challenging task to be accomplished.
then processed through the CWT operation. To resume, this Figure 13 shows the results regarding the case of a dam-
means that the seven parts (i.e., seven segments) in which age affecting the stringers of two modules, precisely the
the signal is divided do not correspond exactly to the exten- fourth and the seventh one (see Fig. 7). Again, by looking
sion of the seven modules composing the bridge, but the at the legend of the figure, for each degraded scenario, the
difference is considered negligible for the purpose of the affected segment of the bridge is reported (e.g., S4 which
present work. Drawn with a black horizontal solid line, the stands for segment 4). In this case, it is possible to observe
normalized upper control limit UCLn is shown, equal to 1 a clear increase in the normalized T 2 distance at the segment
for each segment. which encompasses damaged trusses. Segment 4 presents
Figure 12 clearly shows the following aspects: for dam- highest values of the normalized distance for damage SD2,50%
ages CG2,50% and CG3,50% (green and red), it is possible to (blue squares), while segment 7 for damage scenario SD3,50%.
observe that there is a clear increase in the normalized T 2 By comparing the outcomes shown Fig. 12 with the ones
distance moving from further segments to the ones closer in Fig. 13, it is possible to notice that damages affecting
to the damage position. Precisely, for CG3,50% scenario stringers produce much larger values of T 2 ∕UCL parameter.
larger normalized distances are found for segments six This can be due to the more direct relation between the rail
and seven, which both share the degraded element. Analo- supporting the transit train (than for the cross-girders), as
gous conclusions can be drawn for damage CG2,50% , while well as the extension of the damage and the fact that when
cannot for scenario CG1,50% . In Fig. 12 as well as in the

Railway Engineering Science

Drive‑by damage detection and localization exploiting continuous wavelet transform and…

considering stringers, both the two elements under the two Segment 7. As observed before, larger values for the DI are
rails are considered and damaged of the same quantity. obtained.
The next step consists in the computation of the mean To conclude, with one sensing point, the methodology
values obtained from the normalized T 2 distances, segment shows promising results in terms of both damage detec-
by segment. The result is shown in Fig. 14, for damage sce- tion and localization capabilities when corrosion is strong.
narios featured by 50% corrosion level. Average values of However, some problems are observed: when high-intensity
normalized Hotelling’s distances are drawn as functions corrosion affect a cross-girder closer to the bridge entrance,
of the segment number. This figure further points out that, localization is not possible. Furthermore, scenarios, that are
except for CG1,50%, the normalized T 2 distance, on average, is closer to early damage conditions, were found to be critical
maximum close or in correspondence of the damaged truss. when dealing with cross-girders, even though the considered
Another aspect that deserves to be stressed is that in both train batches are formed by 40 train transits. The presented
the two cases, peaks of the damage indices appear to not be damage index does not markedly change its value when
significantly influenced by the position of the damage itself. the damage is moved, but is strongly affected by damage
Now, let us consider milder damages, with the purpose to typology.
investigate the following aspects: whether the DI is propor-
tional to corrosion intensity and how effective is the meth- 4.2 Results considering two sensing points
odology for lighter degraded scenarios. Under this regard,
corrosion of 15 and 25% are now taken into account, refer- To overcome some of the criticalities presented in the pre-
ring to one single position across bridge span, identified by vious section, the idea is now to investigate the use of two
the number 3 (being toward bridge exit, as shown in Fig. 7). virtual channels, instead of one. Starting from the consid-
Figure 15 shows the outcomes in terms of normalized T 2 eration of high-intensity damages affecting cross-girders
distance when dealing with milder corrosion intensities located in different positions across the bridge span (see
affecting the cross-girder shared by the sixth and seventh Fig. 7), the results are collected in Fig. 17. It is important to
modules of the bridge, again plotted for batches of 40 trains. recall that the plotted damage index (DI) is represented by a
The outcomes point out the capability of the methodology normalized T 2 distance, expressed by Eq. 15. Figure 17 wit-
to detect the presence of a damaged condition and localize nesses good performances in terms of both damage detection
the region of that damage. However, it is clear that the nor- and localization. The exploitation of two sensing positions
malized distance is very close to the UCLn threshold, which (instead of one) allows to overcome the problem related to
can be critical. the localization for damage CG1,50%. In fact, now higher DI
Moving to the case of damage affecting stringers, Fig. 16 values (on average) are obtained for segment 3 instead of
presents the results for milder corrosion intensities. Damage segment 1, which is different from what can be observed
position is clearly identified, as shown in the subplot entitled in Fig. 12. Regarding CG2−3,50% similar performances are
actually obtained, if compared to the case shown in Fig. 12.

Cross-girder damages Stringer damages

50 500
CG1,50% (S2–3) SD2,50% (S4)
45 CG2,50% (S4–5) 450 SD3,50% (S7)
CG3,50% (S6 –7)
40 400
35 350
(T2/UCLi)avg

(T2/UCLi)avg

30 300
25 250
20 200
15 150
10 100
5 50
0 0
1 2 3 4 5 6 7 1 2 3 4 5 6 7 8
Segment No. Segment No.

Fig. 14  Average normalized T 2 values, plotted as a function of segment number. Shaded areas highlight the segments interested by the simulated
degraded scenario

Railway Engineering Science

 L. Bernardini et al.

Segment 1 Segment 2 Segment 3 Segment 4

T2 distance/UCL 15 15 15 15

T2 distance/UCL

T2 distance/UCL

T2 distance/UCL
10 10 10 10

5 5 5 5

0 0 0 0
0 2 4 6 8 0 2 4 6 8 0 2 4 6 8 0 2 4 6 8
Batch number Batch number Batch number Batch number
Segment 5 Segment 6 Segment 7
15 15 15
T2 distance/UCL

T2 distance/UCL

T2 distance/UCL
10 10 10
CG3,15% (S6– 7)
CG3,25% (S6 –7)
UCLn
5 5 5

0 0 0
0 2 4 6 8 0 2 4 6 8 0 2 4 6 8
Batch number Batch number Batch number

Fig. 15  Damage scenario: corrosion affecting cross-girders, milder intensities and one single position (see Fig. 7). Normalized T 2 distance, plot-
ted for the seven segments composing the bridge window of interest, as a function of the train groups. Batches of 40 transits are considered

However, the changes can be quantified by consider- can be observed, even though for SD3,50% a decrease in local-
ing, segment by segment, the ratio between the average ization performances are attained, but they are considered
DIs obtained with two sensing channels over the case with still acceptable, for what explained before. A quick compari-
only one virtual measuring point (see Fig. 18). As shown in son of Fig. 20 with Fig. 14 points out that in case of two vir-
Fig. 18, for each of the three considered scenarios, a sub- tual measuring channels, the position of the damage affects
stantial increase in the average DI is observed, for one of the magnitude of the normalized T 2 distance, regardless the
the two segments around the degraded truss (highlighted by affected element (i.e., stringers or cross-girders).
shaded areas). To further investigate the effectiveness of the methodology
The second damage typology, concerning stringers, is conditions closer to early damage occurrence are now ana-
now discussed. In particular, Fig. 19 depicts the outcomes lyzed, meaning that milder corrosion entities are taken into
regarding 50% corrosion damage affecting stringers (see account. As seen for Sect. 4.1, 15% and 25% corrosion sce-
Fig. 7). In this second case, damage SD2,50% can be clearly narios were modeled, for cross-girders (Fig. 21) and stringers
detected and localized properly. Regarding SD3,50%, larger (Fig. 22), focusing on a specific region of the bridge span (close
values are obtained in segment 6 than 7. Therefore, com- to the bridge exit). The normalized T 2 distances are capable to
pared to Fig. 13, a decrease in localization performances detect a milder damage and its location across bridge span, as
is observed for this second damage scenario. However, it is can be noticed in Fig. 21 and Fig. 22 respectively.
important to say that looking to the DI values regarding the By comparing the performance of using two virtual
SD3,50% case (green triangles in Fig. 19), the two segments measuring channels (i.e., two bogie vertical accelerations)
present indices markedly larger than the others, clearly high- instead of one, it is possible to notice that the former solu-
lighting an indicative location of the damage. tion outperforms the latter, as shown in Fig. 23. Precisely,
This outcome can be noticed also in Fig. 20, where, glob- while for segment 6, no significant changes are obtained,
ally, it can be observed that better performance are obtained a clear difference is found for segment 7, where the DIs
by exploiting two virtual measuring channels than using one are larger than the same obtained considering only the
(refer to Fig. 14). In fact, to recap, CG1,50% can be detected leading bogie vertical accelerations. This improvement
and localized properly in the former case, while it was not can be attributed to the increased amount of available
possible with one sensing point. An increase in DI values information, which appears to enhance the capability of

Railway Engineering Science

Drive‑by damage detection and localization exploiting continuous wavelet transform and…

Segment 1 Segment 2 Segment 3 Segment 4

50 50 50 50

40 40 40 40

T2 distance/UCL
T2 distance/UCL
T2 distance/UCL

T2 distance/UCL
30 30 30 30

20 20 20 20

10 10 10 10

0 0 0 0
0 5 10 0 5 10 0 5 10 0 5 10
Batch number Batch number Batch number Batch number

Segment 5 Segment 6 Segment 7

50 50 50

40 40 40

T2 distance/UCL
T2 distance/UCL

T2 distance/UCL

30 30 30
SD3,15% (S7)
SD3,25% (S7)
20 20 20
UCLn
10 10 10

0 0 0
0 5 10 0 5 10 0 5 10
Batch number Batch number Batch number

Fig. 16  Damage scenario: corrosion affecting stringers, milder intensities and one single position (see Fig. 7). Normalized T 2 distance, plotted
for the seven segments composing the bridge window of interest, as a function of the train groups. Batches of 40 transits are considered. One
vertical acceleration is used

the system to capture structural behavior and associated 5 Robustness to measurement noise
health condition. It is worth noticing that this observation
is similar to the findings reported in [62]. Therefore, it can Regardless of their nature, sensors are susceptible to meas-
be concluded that exploiting two bogies vertical accelera- urement noise. Therefore, in the attempt to simulate real
tions instead of one produce better overall performance. working conditions, it is important to include measure-
Assuming the case of two virtual measuring points, a ment noise in the virtual measuring chain. According to
corrosion damage of 10% was finally considered, affecting this observation, the purpose of this section is to investi-
stringers and cross-girders, taken, as before, as separated gate damage detection and localization performance when
damage scenarios. The outcomes, in terms of normalized T 2 considering measurement noise. The latter was modeled
distance (see Eq. 15), are collected and depicted in Fig. 24. according to the approach presented in [17, 18]. Precisely,
The plots, regarding segments 6 and 7 respectively, show the a random artificial noise was added to the vertical bogie
highest values of Tnd 2
for the two damage scenarios. There- accelerations (obtained from simulations), as follows:
fore, it appears that corrosion intensities of 10% are still
effectively detectable by the methodology herein proposed. Z̈ 1,n = Z̈ 1 + Ln ⋅ 𝜎(Z̈ 1 ) ⋅ N(0, 1), (16)
Furthermore, for the case of damage affecting cross-girders,
where Z̈ 1 is the vertical bogie acceleration directly obtained
it seems that a lower bound has been almost reached, while
from the simulation, while 𝜎(Z̈ 1 ) identifies its standard
there could be still some margin for damages affecting string-
deviation. Furthermore, N(0, 1) is a standard normal distri-
ers. Given the promising results, one of the main objectives
bution vector with zero mean and unit standard deviation,
of future research will be to model corrosion damage with an
while Ln is the noise level. In this paper, two noise levels
intensity below 10% in order to determine the extent to which
were adopted, set equal to 5% and 10% , respectively. To
the developed methodology can successfully be applied.
ensure clarity, the results for the specific case involving the
use of two virtual sensor channels, a 10% noise level, and

Railway Engineering Science

 L. Bernardini et al.

Segment 1 Segment 2 Segment 3 Segment 4

25 25 25 80

20 20 20

T2 distance/UCL
60
T2 distance/UCL

T2 distance/UCL

T2 distance/UCL
15 15 15
40
10 10 10
20
5 5 5

0 0 0 0
0 5 10 0 5 10 0 5 10 0 5 10
Batch number Batch number Batch number Batch number

Segment 5 Segment 6 Segment 7

80 80 80
T2 distance/UCL

60 60 60
T2 distance/UCL

T2 distance/UCL
CG1,50% (S2–3)
CG2,50% (S4–5)
40 40 40 CG3,50% (S6–7)
UCLn
20 20 20

0 0 0
0 5 10 0 5 10 0 5 10
Batch number Batch number Batch number

Fig. 17  Damage scenario: corrosion affecting cross-girders, highest intensity and three different positions (see Fig. 7). Normalized T 2 distance,
plotted for the seven segments composing the bridge window of interest, as a function of the train groups. Batches of 40 transits and two virtual
measuring channels are considered

Fig. 18  Ratios of averaged damage indices, for different damage scenarios affecting cross-girders, with a corrosion intensity equal to 50%

mildest damages (i.e., 10% of corrosion intensity) only are the mildest corrosion scenarios simulated in this work, i.e.,
now shown. CG3,10% and SD3,10%. Again, even in the presence of signifi-
Therefore, 10% corrosion scenarios are now consid- cant measurement noise, promising outcomes are achieved
ered. To this end, Fig. 25 shows the results referring to by the proposed approach. Precisely, larger normalized

Railway Engineering Science

Drive‑by damage detection and localization exploiting continuous wavelet transform and…

Segment 1 Segment 2 Segment 3 Segment 4

50 50 50 2.5

T2 distance/UCL (×104)
40 40 40 2
T2 distance/UCL

T2 distance/UCL

T2 distance/UCL
30 30 30 1.5

20 20 20 1

10 10 10 0.5

0 0 0 0
0 2 4 6 0 2 4 6 0 2 4 6 0 2 4 6
Batch number Batch number Batch number Batch number

Segment 5 Segment 6 Segment 7

500 500 500

400 400 400

T2 distance/UCL

T2 distance/UCL
T2 distance/UCL

300 300 300 SD2,50 % (S4)

SD 3,50% (S7)
200 200 200
UCLn
100 100 100

0 00 0 0
0 2 4 6 2 4 6 2 4 6
Batch number Batch number Batch number

Fig. 19  Damage scenario: corrosion affecting stringers, highest intensity and three different positions (see Fig. 7). Normalized T 2 distance, plot-
ted for the seven segments composing the bridge window of interest, as a function of the train groups. Batches of 40 transits and two virtual
measuring channels are considered

Cross-girder damages Stringer damages

70 15000
CG1,50% (S2–3) SD2,50% (S4)
CG2,50% (S4–5) SD3,50% (S7)
60 CG3,50% (S6–7)

50
10000
(T2/UCLi)avg
(T2/UCLi)avg

40

30

5000
20

10

0 0
1 2 3 4 5 6 7 1 2 3 4 5 6 7 8
No. of segment No. of segment

Fig. 20  Average normalized T 2 values, plotted as a function of segment. Shaded areas highlight the segments interested by the simulated
degraded scenario. Two virtual measuring channels are now considered

T 2 values are obtained in the seventh segment compared the presented methodology in detecting mild damages,
to the others, which agrees with the position of damaged even in the presence of measurement noise.
trusses (see Fig. 7). This result confirms the potential of

Railway Engineering Science

 L. Bernardini et al.

Segment 1 Segment 2 Segment 3 Segment 4

T 2 distance/UCL 15 15 15 15

T 2 distance/UCL

T 2 distance/UCL
T 2 distance/UCL
10 10 10 10

5 5 5 5

0 0 0 0
0 2 4 6 8 0 2 4 6 8 0 2 4 6 8 0 2 4 6 8
Batch number Batch number Batch number Batch number
Segment 5 Segment 6 Segment 7
15 15 15
T 2 distance/UCL

T 2 distance/UCL
T2 distance/UCL

10 10 10
CG3,15% (S6–7)
CG3,25% (S6–7)
5 5 5 UCLn

0 0 0
0 2 4 6 8 0 2 4 6 8 0 2 4 6 8
Batch number Batch number Batch number

Fig. 21  Damage scenario: corrosion affecting cross-girders, milder intensities and one single position (see Fig. 7). Normalized T 2 distance, plot-
ted for the seven segments composing the bridge window of interest, as a function of the train groups. Batches of 40 transits are considered and
two virtual measuring channels

Segment 1 Segment 2 Segment 3 Segment 4

50 50 50 50

40 40 40 40
T2 distance/UCL
T2 distance/UCL

T2 distance/UCL

T2 distance/UCL

30 30 30 30

20 20 20 20

10 10 10 10

0 0 0 0
0 5 10 0 5 10 0 5 10 0 5 10
Batch number Batch number Batch number Batch number
Segment 5 Segment 6 Segment 7
50 50 50

40 40 40
T2 distance/UCL
T2 distance/UCL
T2 distance/UCL

30 30 30 SD3,15% (S7)
SD3,25% (S7)
20 20 20
UCLn
10 10 10

0 0 0
0 5 10 0 5 10 0 5 10
Batch number Batch number Batch number

Fig. 22  Damage scenario: corrosion affecting stringers, milder intensities and one single position (see Fig. 7). Normalized T 2 distance, plotted
for the seven segments composing the bridge window of interest, as a function of the train groups. Batches of 40 transits are considered and two
virtual measuring channels

Railway Engineering Science

Drive‑by damage detection and localization exploiting continuous wavelet transform and…

Segment 6 Segment 7
15 15

CG3,15% (S6–7) - 2 sensors

CG3,25% (S6–7) - 2 sensors
CG3,15% (S6– 7) - 1 sensor
CG3,25% (S6–7)- 1 sensor
10 UCLn
10
T2 distance/UCL

T2 distance/UCL
5
5

0 0
0 2 3 4 5 6 7 8 0 2 3 4 5 6 7 8
Batch number Batch number

Fig. 23  Comparison of the normalized T 2 distance in case of milder damages affecting cross-girders, using one or two virtual measuring chan-
nels. The focus is put on segments 6 and 7, the one around the position of the considered degraded truss

Segment 1 Segment 2 Segment 3 Segment 4

15 15 15 15
T2 distance/UCL

T2 distance/UCL
T2 distance/UCL

T2 distance/UCL

10 10 10 10

5 5 5 5

0 0 0 0
0 2 4 6 8 10 0 2 4 6 8 10 0 2 4 6 8 10 0 2 4 6 8 10
Batch number Batch number Batch number Batch number

Segment 5 Segment 6 Segment 7

15 15 15
T2 distance/UCL

T2 distance/UCL

T2 distance/UCL

10 10 10

5 5 5
CG3,10% (S6–7)
SD3,10% (S7)
UCLn
0 0 0
0 2 4 6 8 10 0 2 4 6 8 10 0 2 4 6 8 10
Batch number Batch number Batch number

Fig. 24  Damage scenario: corrosion affecting stringers and cross-girder, 10% corrosion intensity and one single position (see Fig. 7). Normal-
ized T 2 distance, plotted for the seven segments composing the bridge window of interest, as a function of the train groups. Batches of 40 transits
are considered and two virtual measuring channels

Finally, Fig. 26 shows normalized Hotelling’s distances, more pronounced negative effect on damage identification
obtained for segment 7, considering CG3,10% and SD3,10% and localization performances when the defect involves
scenarios (see Fig. 7), with and without the consideration the cross-girder, making its identification less clear. How-
of measurement noise ( Ln equal to 10%). The latter has a ever, six DI values over nine are still above the UCLn ,

Railway Engineering Science

 L. Bernardini et al.

Segment 1 Segment 2 Segment 3 Segment 4

15 15 15 15
T2 disTance/UCL

T2 distance/UCL
T2 distance/UCL

T2 distance/UCL
10 10 10 10

5 5 5 5

0 0 0 0
0 5 10 0 5 10 0 5 10 0 5 10
Batch number Batch number Batch number Batch number

Segment 5 Segment 6 Segment 7

15 15 15
T2 distance/UCL

T2 distance/UCL

T2 distance/UCL
10 10 10
CG3,10% (S6–7)
SD3,10% (S7)
5 5 5 UCLn

0 0 0
0 5 10 0 5 10 0 5 10
Batch number Batch number Batch number

Fig. 25  Mildest corrosion scenarios, considering measurement noise (level of 10%). Normalized T 2 distance, plotted, for the seven segments
composing the bridge window of interest, as a function of the train groups. Batches of 40 transits are considered and two virtual measuring
channels

where nine is the total number of batches encompassing

40 trains each.

6 Conclusions

In this work, the authors introduced a drive-by damage

detection and localization algorithm based on wavelet coef-
ficients and multiple sparse autoencoders (SAEs). In particu-
lar, starting from the scalogram obtained from vertical bogie
accelerations, using a complex Morlet wavelet, a frequency
band of interest is identified, related to the module passing
frequency. In fact, the methodology herein discussed takes
advantage of the typical modular configuration of Warren
truss bridges to extract relevant information about their con-
dition. Once wavelet coefficients are averaged in frequency,
in an interval ranging from 0.5 to 2 times the module pass-
ing frequency fp (V/d), and resampled in space, the obtained
Fig. 26  Mildest corrosion scenarios, with and without considering
measurement noise (level of 10%). Normalized T 2 distance, plotted curve (i.e., Wavg ) is input to a set of sparse autoencoders
for the seventh segment, as a function of the train batch (considering (SAEs). In particular, the Wavg corresponding to the train
groups of 40 trains) transit over the steel truss bridge is isolated and divided

Railway Engineering Science

Drive‑by damage detection and localization exploiting continuous wavelet transform and…

in seven sections, indicatively corresponding to the seven bridge position is required for the technique to be effective.
modules composing the bridge. An equal number of SAEs However, given the promising outcomes of this research,
are then trained considering almost 760 simulations over investigations on the effects played by speed and position
the healthy structure. The damage index is then constructed estimation inaccuracies will be conducted in the next future.
considering the normalized Hotelling’s T 2 distance, com- In parallel, corrosion damages below 10% will be modeled
puted at the bottleneck layer of each sparse autoencoder. to investigate how far the methodology can be further suc-
The rationale behind this methodology is that autoencoders cessfully applied. Moreover, an analysis on the impact of
related to segments in the region of damaged elements will damage position on its detection and localization, when
lead to larger normalized T 2 distances, outlining the pres- dealing with mild corrosion intensities, will be carried out.
ence of a degraded condition, and its indicative region. The
methodology was numerically tested considering a set of
Open Access This article is licensed under a Creative Commons Attri-
operational variables, such as train mass, forward speed and bution 4.0 International License, which permits use, sharing, adapta-
track irregularity evolution in time. Two configurations are tion, distribution and reproduction in any medium or format, as long
investigated: considering one acceleration virtual measuring as you give appropriate credit to the original author(s) and the source,
channel or two, always referring to the bogies of the leading provide a link to the Creative Commons licence, and indicate if changes
were made. The images or other third party material in this article are
coach of the rail convoy. The work provided the following included in the article’s Creative Commons licence, unless indicated
main findings: otherwise in a credit line to the material. If material is not included in
the article’s Creative Commons licence and your intended use is not
• Multiple sparse autoencoders can be effectively used permitted by statutory regulation or exceeds the permitted use, you will
need to obtain permission directly from the copyright holder. To view a
to detect damage occurrence and its indicative position copy of this licence, visit http://​creat​iveco​mmons.​org/​licen​ses/​by/4.​0/.
across bridge span. The exact longitudinal location of
the degraded truss cannot be provided, while indications
of the region to be further inspected are given, which
can represent a significant aid for bridge infrastructure References
managers;
• The normalized T 2 distance, computed at the bottleneck 1. Aflatooni M, Chan T, Thambiratnam D et al (2013) Synthetic
of the trained sparse autoencoders, is globally effective rating system for railway bridge management. J Civ Struct Health
Monit 3(2):81–91
for damage detection and localization across the bridge
2. Lin W (2018) Rehabilitation and strengthening of aged steel rail-
span. Hotelling’s distance is a well-known concept in way bridges in Japan. J Civil Environ Eng 8(2):1000305
direct SHM, but, according to best authors’ knowledge, 3. Agdas D, Rice JA, Martinez JR et al (2016) Comparison of visual
there are no applications in the field of indirect bridge inspection and structural-health monitoring as bridge condition
assessment methods. J Perform Constr Facil 30(3):04015049
health monitoring;
4. Lynch JP, Farrar CR, Michaels JE (2016) Structural health moni-
• Using the vertical accelerations of the two bogies of the toring: technological advances to practical implementations. Proc
first coach outperforms the case in which only one accel- IEEE 104(8):1508–1512
eration is adopted. Precisely, corrosion damages of 15% 5. Iacovino C, Turksezer ZI, Giordano PF et al (2022) Comparison
of bridge inspection policies in terms of data quality. J Bridge Eng
can be detected and localized with higher robustness;
27(3):04021115
• Promising detection and localization outcomes were 6. He Z, Li W, Salehi H et al (2022) Integrated structural health
achieved, even in the presence of measurement noise, monitoring in bridge engineering. Autom Constr 136:104168
when considering a corrosion level of the 10%; 7. Malekjafarian A, Corbally R, Gong W (2022) A review of mobile
sensing of bridges using moving vehicles: progress to date, chal-
• The requirement for a maximum number of two acceler-
lenges and future trends. Structures 44:1466–1489
ometers make the methodology cost effective for bridge 8. Malekjafarian A, McGetrick PJ, OBrien EJ (2015) A review of
infrastructure monitoring. indirect bridge monitoring using passing vehicles. Shock Vib
1:286139
9. Yang Y-B, Lin C, Yau J (2004) Extracting bridge frequencies
On the other hand, some limitations regarding this work
from the dynamic response of a passing vehicle. J Sound Vib
must be pointed out: first of all, the methodology was not 272(3–5):471–493
tested considering inaccuracies in the identification of vehi- 10. Souza EF, Bragança C, Meixedo A et al (2023) Drive-by method-
cle speed, which is needed to define the frequency range ologies applied to railway infrastructure subsystems: a literature
review: Part I: bridges and viaducts. Appl Sci 13(12):6940
of interest to calculate Wavg curve. Another aspect is repre-
11. Corbally R, Malekjafarian A (2023) Detecting changes in the
sented by the fact that “low level” track irregularity defects structural behaviour of a laboratory bridge model using the con-
were modeled, not accounting for large defects due, for tact-point response of a passing vehicle. J Struct Integr Maint
instance, to transition zones at the bridge ends. Moreover, 8(4):226–238
as for the majority of drive-by literature, constant traveling
speed is assumed. In addition, a precise identification of

Railway Engineering Science

 L. Bernardini et al.

12. Fitzgerald PC, Malekjafarian A, Cantero D et al (2019) Drive- 31. Radicioni L, Bono FM, Benedetti L et al (2023) Overcoming
by scour monitoring of railway bridges using a wavelet-based strain gauges limitation in the estimation of train load pass-
approach. Eng Struct 191:1–11 ing on a bridge through deep learning. In: NDE 4.0, Predictive
13. Bernardini L, Carnevale M, Collina A (2021) Damage identifi- Maintenance, Communication, and Energy Systems: The Digi-
cation in warren truss bridges by two different time-frequency tal Transformation of NDE Long Beach, pp 124890
algorithms. Appl Sci 11(22):10605 32. Cheli F, Diana G (2015) Advanced dynamics of mechanical
14. Demirlioglu K, Erduran E (2024) Drive-by bridge damage detec- systems. Springer, Berlin
tion using continuous wavelet transform. Appl Sci 14(7):2969 33. Silva CW (2006) Vibration: fundamentals and practice, 2nd edn.
15. Flah M, Nunez I, Ben Chaabene W et al (2021) Machine learn- CRC Press, Boca Raton
ing algorithms in civil structural health monitoring: a systematic 34. British Standards (2006) Eurocode 1: Actions on structures.
review. Arch Comput Meth Eng 28(4):2621–2643 Technical report, BS EN 1991
16. Fernandes T, Lopez R, Ribeiro D (2024) Drive-by scour dam- 35. Bernardini L, Matsuoka K, Collina A (2024) Indirect frequency
age detection in railway bridges using deep autoencoder and estimation by time-shifted accelerations subtraction: generaliza-
different sensor placement strategies. J Civ Struct Health Monit tion of the methodology and numerical application on a warren
14(8):1895–1916 truss bridge. J Sound Vib 590:118491
17. de Souza EF, Bragança C, Ribeiro D et al (2024) Drive-by dam- 36. Argentino A, Bono FM, Bernardini L et al (2024) Automated
age detection methodology for high-speed railway bridges using OMA through SSI-COV algorithm of a warren truss railway
sparse autoencoders. Railw Eng Sci. https://​doi.​org/​10.​1007/​ bridge exploiting free decay response. In: Proceedings of the
s40534-​024-​00347-3 10th International Operational Modal Analysis Conference
18. Sarwar MZ, Cantero D (2021) Deep autoencoder architecture for (IOMAC 2024). Naples, pp 600–608
bridge damage assessment using responses from several vehicles. 37. Benedetti L, Argentino A, Bernardini L et al (2024) A three-
Eng Struct 246:113064 year project on structural health monitoring of railway bridges:
19. Kaur K, Alamdari MM, Chang KC et al (2023) Damage detection main results and lessons learnt. In: Proceedings of the 11th
and localization for indirect bridge monitoring exploiting adver- European Workshop on Structural Health Monitoring (EWSHM
sarial autoencoder and wavelet transform. In European Workshop 2024). Potsdam, pp 1–10
on Structural Health Monitoring, EWSHM 2022. Springer, Cham 38. Pastor M, Binda M, Harcarik T (2012) Modal assurance crite-
20. Calderon Hurtado A, Makki Alamdari M, Atroshchenko E et al rion. Procedia Eng 48:543–548
(2023) An unsupervised learning method for indirect bridge struc- 39. Pascal J, Sauvage G (1992) New method for reducing the multi-
tural health monitoring. In Experimental Vibration Analysis for contact wheel/rail problem to one equivalent contact patch. Veh
Civil Engineering Structures, EVACES 2023. Springer, Cham Syst Dyn 20(sup1):475–489
21. Calderon Hurtado A, Kaur K, Alamdari MM et al (2023) 40. Di Gialleonardo E, Braghin F, Bruni S (2012) The influence of
Unsupervised learning-based framework for indirect structural track modelling options on the simulation of rail vehicle dynam-
health monitoring using adversarial autoencoder. J Sound Vib ics. J Sound Vib 331(19):4246–4258
550:117598 41. Shen ZY, Hedrick JK, Elkins JA (1983) A comparison of alter-
22. Li Z, Lin W, Zhang Y (2023) Real-time drive-by bridge damage native creep force models for rail vehicle dynamic analysis. Veh
detection using deep auto-encoder. Structures 47:1167–1181 Syst Dyn 12(1–3):79–83
23. Dinas A, Nikolaidis TN, Baniotopoulos C (2017) Sustainable 42. Newmark NM (1959) A method of computation for structural
restoration criteria for a historical steel railway bridge. Procedia dynamics. J Eng Mech Div 85(3):67–94
Environ Sci 38:578–585 43. Venturi G, Simonsson P, Collin P (2021) Strengthening old steel
24. Bragança C, de Souza EF, Ribeiro D et al (2024) Drive-by early railway bridges: a review. In IABSE Congress, Ghent 2021:
damage detection in railway bridges using wavelets and autoen- Structural Engineering for Future Societal Needs. Ghent, pp
coders. In: Proceedings of the 6th International Conference on 1718–1727
Railway Technology: Research, Development and Maintenance. 44. Simoncelli M, Aloisio A, Zucca M et al (2023) Intensity and
Prague, Paper 15.7 location of corrosion on the reliability of a steel bridge. J Constr
25. Finotti RP, Gentile C, Barbosa FS et al (2020) Vibration-based Steel Res 206:107937
anomaly detection using sparse auto-encoder and control charts. 45. Teolis A (1998) Computational signal processing with wavelets.
In Proceedings of the XI International Conference on Structural Birkhäuser, Boston
Dynamics. Athens, pp 1335–1347 46. Bernardini L, Collina A, Somaschini C et al (2023) An applica-
26. Diana G, Cheli F, Bruni S et al (1995) Dynamic interaction tion of drive-by approach on a railway warren bridge. Life-cycle
between rail vehicles and track for high speed train. Veh Syst of structures and infrastructure systems. CRC Press, Boca Raton
Dyn 24(sup1):15–30 47. Ng A (2011) Sparse autoencoder. CS294A lecture notes 72:1–19
27. Bruni S, Collina A, Diana G et al (1999) Lateral dynamics of a 48. Hotelling H (1947) Multivariate quality control, illustrated by
railway vehicle in tangent track and curve: tests and simulation. the air testing of sample bombsights. Techniques of statistical
Veh Syst Dyn 33(sup1):464–477 analysis. McGraw Hill, New York
28. Alfi S, Bruni S (2009) Mathematical modelling of train-turnout 49. Martinez-Ríos EA, Bustamante-Bello R, Navarro-Tuch S, Perez-
interaction. Veh Syst Dyn 47(5):551–574 Meana H (2022) Applications of the generalized morse wavelets:
29. Bruni S, Collina A, Corradi R, et al (2004) Numerical simula- a review. IEEE Access 11:667–688
tion of train-track-bridge dynamic interaction. In: Sixth World 50. Gomez-Luna E, Aponte Mayor G, Pleite Guerra J (2014) Applica-
Congress on Computational Mechanics (WCCM VI) in conjunc- tion of wavelet transform to obtain the frequency response of a
tion with the Second Asian-Pacific Congress on Computational transformer from transient signals: Part II: practical assessment
Mechanics (APCOM’04). Beijing, pp 237–242 and validation. IEEE Trans Power Deliv 29(5):2231–2238
30. Diana G, Cheli F, Bruni S (2000) Railway runnability and 51. Torrence C, Compo GP (1998) A practical guide to wavelet analy-
train-track interaction in long span cable supported bridges. In: sis. Bull Amer Meteor Soc 79(1):61–78
International Conference on Advances in Structural Dynamics. 52. Teng S, Liu A, Chen B et al (2024) Bridge progressive damage
Hongkong, pp. 43–54 detection using unsupervised learning and self-attention mecha-
nism. Eng Struct 301:117278

Railway Engineering Science

Drive‑by damage detection and localization exploiting continuous wavelet transform and…

53. Bono FM, Radicioni L, Cinquemani S et al (2023) A compari- 59. García-Macías E, Ubertini F (2020) MOVA/MOSS: two inte-
son of deep learning algorithms for anomaly detection in discrete grated software solutions for comprehensive Structural Health
mechanical systems. Appl Sci 13(9):5683 Monitoring of structures. Mech Syst Signal Process 143:106830
54. Spínola Neto M, Finotti R, Barbosa F et al (2024) Structural dam- 60. Borlenghi P, Saisi A, Gentile C (2024) Vibration monitoring of
age identification using autoencoders: a comparative study. Build- masonry bridges to assess damage under changing temperature.
ings 14(7):2014 Dev Built Environ 20:100555
55. Goodfellow I, Bengio Y, Courville A (2016) Deep learning. MIT 61. International Union of Railways (1989). Bogies with steered or
Press, Cambridge steering wheelsets. Report No. 1: specifications and preliminary
56. Hoefler T, Alistarh D, Ben-Nun T et al (2021) Sparsity in deep studies, specification for a bogie with improved curving charac-
learning: pruning and growth for efficient inference and training teristics. Volume 2. B176
in neural networks. J Mach Learn Res 22(241):1–124 62. Li Z, Lan Y, Lin W (2024) Bridge damage classification using
57. Wang Z, Ong K (2008) Autoregressive coefficients based hotel- multiple responses of vehicles and 1-D convolutional neural net-
ling’s t2 control chart for structural health monitoring. Comput works. CRC Press, Boca Raton
Struct 86(19–20):1918–1935
58. Montgomery DC (2019) Introduction to statistical quality control,
8th edn. John Wiley & Sons, Hoboken

Railway Engineering Science