1 1 1 1 1 Review On The New Development of Vibration-Based Damage2010 2019

Journal of Sound and Vibration 491 (2021) 115741
Contents lists available at ScienceDirect
Journal of Sound and Vibration

journal homepage: www.elsevier.com/locate/jsv
Review on the new development of vibration-based damage

identification for civil engineering structures: 2010–2019
Rongrong Hou, Yong Xia∗
Department of Civil and Environmental Engineering, The Hong Kong Polytechnic University, Kowloon, Hong Kong, China
a r t i c l e i n f o a b s t r a c t
Article history: Structural damage identification has received considerable attention during the past
Received 7 January 2020 decades. Although several reviews have been presented, some new developments have
Revised 26 July 2020
emerged in this area, particularly machine learning and artificial intelligence techniques.
Accepted 22 September 2020
This article reviews the progress in the area of vibration-based damage identification
Available online 23 September 2020
methods over the past 10 years. These methods are classified in terms of different dam-
Keywords: age indices and analytical/numerical techniques used with discussions of their advantages
Damage identification and disadvantages. The challenges and future research for vibration-based damage iden-
Vibration method tification are summarised. This review aims to help researchers and practitioners in im-
Model updating plementing existing damage detection algorithms effectively and developing more reliable
Review and practical methods for civil engineering structures in the future.
© 2020 Elsevier Ltd. All rights reserved.
1. Introduction
Civil engineering structures are exposed to natural and manmade hazards, which may cause structural damage or even
collapse. An unpredicted structural failure can be catastrophic not only in terms of life and economic losses but also in terms
of the subsequent societal impacts. Therefore, structural damage detection is important, especially in the early damage state,
to avoid sudden failures and improve the safety and reliability of structures.
Structural damage identification has been widely explored over the past decades. Rytter [1] classified damage identi-
fication into four levels: determination of damage existence, determination of damage location, quantification of damage
severity, and finally prediction of the remaining service life of structures, which are referred to as Levels 1 to 4, respectively.
Most current studies focus on the first three levels [2].
Structural damage detection methods can be divided into two categories, namely, non-destructive testing (NDT) and
vibration-based methods. The former are local methods that cannot easily detect damage located inside structures (e.g.
cracks in concrete and/or corrosion of steel bars) or damage enclosed by non-structural components (e.g. decorations of
buildings). By contrast, vibration-based damage identification methods examine changes in structural global vibration char-
acteristics and are thus regarded as global methods and have attracted considerable attention during the past decades. In
this regard, only vibration-based damage detection methods are reviewed in this paper.
Early vibration-based damage detection methods were reviewed and summarised in several literatures. For example,
Doebling et al. [2] and Sohn et al. [3] comprehensively reviewed vibration-based damage detection methods and their ap-
plications to various types of structures before 1996 and between 1996 and 2001, respectively. Salawu [4] reviewed damage
∗
Corresponding author.
E-mail address: [email protected] (Y. Xia).
https://doi.org/10.1016/j.jsv.2020.115741
0022-460X/© 2020 Elsevier Ltd. All rights reserved.
R. Hou and Y. Xia Journal of Sound and Vibration 491 (2021) 115741
List of abbreviations
AANN Auto-associative neural network

ABC Artificial bee colony
AI Artificial intelligence
ANN Artificial neural network
AR Auto-regressive
ARD Automatic relevance determination
ARMA Auto-regressive moving average
ARMAX Auto-regressive moving average with exogenous input
ARX Autoregressive with exogenous input
BPNN Back propagation neural network
CS Compressive sensing
CMS Component mode synthesis
DOF Degree of freedom
EKF Extended Kalman filter
EM Expectation–maximization
EMD Empirical mode decomposition
FD Fractal dimension
FE Finite element
FRF Frequency response function
GA Genetic algorithm
GS Gibbs sampling
HHT Hilbert–Huang transform
IMF Intrinsic mode function
KPCA Kernel principal component analysis
Lasso Least absolute shrinkage and selection operator
MAC Modal assurance criterion
ML Machine learning
MSC Mode shape curvature
MSD Mahalanobis squared distance
MSE Modal strain energy
NDT Non-destructive testing
ODS Operational deflection shape
OSP Optimal sensor placement
PCA Principal component analysis
PDF Probability density function
PSD Power spectral density
PSO Particle swarm optimization
RBF Radical basis function
RD Random decrement
RF Random forest
RSM Response surface methodology
SBL Sparse Bayesian learning
SHM Structural health monitoring
SVD Singular value decomposition
SVM Support vector machine
TSD Traffic speed deflectometer
WT Wavelet transformation
identification methods using natural frequency changes. Carden and Fanning [5] focused on publications between 1996 and
2003. Fan and Qiao [6] compared several damage detection methods. Hakim and Razak [7] and Chen et al. [8] reviewed the
applications of artificial neural networks (ANNs) and Hilbert–Huang transform (HHT) for damage identification over the last
20 years, respectively.
A recent literature survey was conducted by Kong et al. [9], who discussed the prediction of the remaining life of struc-
tures and related decision making. Cao et al. [10] provided an overall survey of damping-based damage detection methods.
Huang et al. [11] presented a complete review on the recent development of Bayesian inference for structural damage de-
tection and assessment. Weng et al. [12] reviewed dynamic substructuring methods for damage identification of large-scale
structures. In terms of the massive collected data, Gordan et al. [13] intensively reviewed the applications of data min-
2
ing techniques in damage identification and structural health monitoring (SHM) that have been conducted since 20 0 0. Bao
et al. [14] reviewed advancements in data science and engineering in SHM. An et al. [15] summarised damage identification
methods for bridge structures between 2011 and 2017.
The fast development in information technology, particularly sensing technology, signal processing techniques, machine
learning (ML) and artificial intelligence (AI) technologies, advances vibration-based damage identification methods over the
past decade. In addition, damage identification under operational and environmental conditions, output-only identification,
statistical damage detection, real-time identification for on-line SHM, and optimal sensor placement (OSP), etc., are also
important issues. A number of techniques have been developed to deal with these problems.
This work aims to review the latest vibration-based damage detection methods between 2010 and 2019. Although hun-
dreds of papers are published every year, only representative papers are reviewed in Section 2 because of the length limit.
The challenges and future trends of the vibration-based damage identification methods are discussed in Section 3, followed
by Conclusion.
In Section 2, damage identification methods are classified according to the damage index used and analytical techniques.
First, methods using various modal parameters are reviewed in Subsection 2.1. Different signal processing based damage
identification methods are reviewed in Subsection 2.2. Subsection 2.3 discussed finite element (FE) model updating methods,
including conventional model updating, substrcuturing methods, regularisation techniques, and sparse recovery techniques.
The optimisation algorithms such as genetic algorithm (GA) and ANN are reviewed in Subsection 2.4. Subsection 2.5 intro-
duces the statistical time series methods based on auto-regressive (AR) models and their variants. Subsequently, the cutting-
edge ML methods are comprehensively reviewed in Subsection 2.6, which includes supervised, unsupervised and semi-
supervised learning. Subsection 2.7 introduces the probabilistic damage identification approaches, particularly the Bayesian
methods. Subsection 2.8 addresses damage identification under varying environmental conditions. The methods are divided
into two categories, depending on whether the environmental variables are measured or not. Subsection 2.9 reviews exist-
ing algorithms with consideration of nonlinear structural behaviours. In Subsection 2.10, some techniques that are excluded
from the abovementioned categories are introduced, such as the newly developed drive-by damage identification methods
and OSP. Finally, Subsection 2.11 focuses on the papers that compared different damage identification methods.
2. Vibration-based damage identification methods
Vibration-based damage detection methods have been first developed and applied in aerospace and mechanical engi-
neering; the civil engineering community has studied the vibration-based damage detection of bridge structures since the
early 1980s [16]. The basic idea of vibration-based damage detection methods is that structural damage may induce changes
in vibration characteristics, such as frequencies and mode shapes [2].
Vibration-based damage detection methods can be categorised into three domains based on vibration parameters: time
domain, frequency domain and time–frequency domain methods. In time domain methods, time–history responses are
used. In frequency domain methods, modal parameters are utilised. Time–frequency domain methods are based on time–
frequency analytical tools. In terms of algorithms used, damage detection methods can be classified into nonmodel-based or
data-driven and model-based methods.
The performance of damage detection largely relies on the choice of damage sensitive features. Therefore, vibration-based
damage detection methods involving different damage indices are first introduced.
2.1. Modal parameter-based methods
With the development of modal analysis technology, the majority of vibration-based methods fall into frequency do-
mains. Modal parameters, such as natural frequencies, mode shapes and their variants, have been commonly used. However,
the use of natural frequencies only as damage indices is no longer popular in recent years because they are insensitive to
local damage and the number of available frequencies is limited, generally less than 10.
2.1.1. Mode shapes

Yoon et al. [17] applied their previously proposed global fitting method [18] to identify damage in 2D plate-like struc-
tures by using the mode shape data only from a damaged structure. In comparison with the gapped smoothing method,
which locally fits mode shape curvature (MSC), global fitting involves the use of a generic mode shape form to globally fit
mode shapes, thus eliminating the smearing effects and reducing false detection. Zhang et al. [19] approximately extracted
mode shapes from the acceleration responses of a passing vehicle with sinusoidal tapping force. Damage location was then
determined on the basis of the difference between damaged and intact mode shape squares. Although baseline information
is necessary, the proposed approach does not require many preinstalled sensors and solving eigenvector or singular value
problems. Feng and Feng [20] extracted a first-order mode shape from the vehicle-induced displacement response, which
was utilised as a damage index to determine damage location and quantitatively monitor the damage progression of bridges.
2.1.2. Natural frequencies and mode shapes

Some researchers combined natural frequencies and mode shapes for damage detection. Sun et al. [21] used a normalised
uniform load surface curvature, which was estimated from modal flexibility, to locate damage for beam-like structures.
3
The proposed method performed better in identifying single and multiple damage locations than the uniform load surface
curvature and MSC methods. However, this method was only applicable to beam-like structures following Bernoulli–Euler
beam theory. Zhao and Zhang [22] utilised the changes in natural frequencies and mode shapes for damage localisation and
quantification. The modal assurance criterion (MAC) was used to analyse the sensitivity of mode shapes between different
orders, and mode shapes with high sensitivities to damage were employed to calculate the damage index.
Radzieński et al. [23] compared six widely used damage detection methods based on modal parameters, including MSC,
coordinate MAC, strain energy damage index, gapped smoothing method, fractal dimension (FD) and wavelet transformation
(WT). However, only the generalised FD and WT damage indicators were able to locate damage position accurately in the
presence of measurement noise. In this regard, the authors proposed a new damage indicator based on the change in natural
frequencies and any one mode shape (measured or modelled). Capecchi et al. [24] combined natural frequencies, mode
shapes and MSCs for damage identification in a parabolic arch.
Single crack identification using modal parameters has been intensively studied. However, relatively few studies have
addressed multicrack identification problems. Caddemi and Caliò [25] derived a closed form expression of the exact dynamic
stiffness matrix of a multi-cracked Euler–Bernoulli beam based on their previous work [26] and then extended it to frame
structures. The natural frequencies and mode shapes of undamaged and damaged frames were calculated on the basis of
the Wittrick–Williams algorithm and further used for damage identification. Later, Khiem and Tran [27] derived a simplified
closed-form expression of the vibration modes of multiple cracked beams. The shifts of natural frequencies and mode shapes
were explicitly expressed in terms of crack locations and magnitudes. An iterative procedure was developed to determine
not only the position and severity of cracks but also their quantity. Khiem and Toan [28] proposed an explicit expression of
natural frequencies in terms of crack positions and sizes for multiple cracked beams; their proposed method differed from
the earlier one by including nonlinear terms with respect to the crack magnitude. The non-uniqueness problem in damage
detection under a symmetrical boundary condition was overcame by incorporating nonlinear terms.
2.1.3. Damping
In comparison with natural frequencies and mode shapes, damping has been less commonly used for damage identi-
fication because of the complexity of its measurement and mechanism. Frizzarin et al. [29] used the nonlinear damping
identified from ambient vibration responses to locate damage in concrete structures without any reference to an undam-
aged baseline. Mustafa et al. [30] proposed an energy-based damping evaluation method for damage localisation. Ay et al.
[31] estimated the damage-induced changes in the overall damping behaviour of a free-vibration dynamic system within a
statistical framework.
The damping model used for damping estimation is critical for damping-based methods [10]. Most studies have adopted
a Rayleigh damping model because of its mathematical simplicity. However, classical Rayleigh damping may be an inappro-
priate assumption for most civil structures [32]. In this regard, Liu et al. [33] proposed a technique to identify damage to
non-classically damped shear buildings. A novel modal identification technique was developed to identify complex modal
parameters from vibration measurements under harmonic excitations. The locations and magnitudes of damage with re-
spect to stiffness reduction and damping defect were then simultaneously identified through the sensitivity-based model
updating.
Damage identification methods based on modal parameters possess the merit of direct physical interpretation. However,
modal identification is susceptible to measurement noise, especially for damage-sensitive high modes. This process may
produce some unavoidable errors and render the damage identification results unreliable. In this regard, some researchers
directly utilised the measured data, such as frequency response function (FRF), for damage identification.
2.1.4. FRFs and their variants

In contrast to modal data, FRFs are calculated over the entire frequency range and can provide more information about
damage. Limongelli [34] proposed a damage detection method by using the difference between the FRF and its spline in-
terpolation. The method was later successfully applied to frame structures under seismic excitation [35] and a single-span
reinforced concrete bridge [36].
An operational deflection shape (ODS) is usually defined as the deflection of a structure at a particular frequency [37].
On the basis of their previous work, Zhang et al. [38] used an ODS curvature to locate damage. The ODS curvature of an
undamaged structure was approximated by a smooth line under the assumption that an intact structure was homogeneous
and uniform to avoid the requirement of a baseline. Damage location was identified by comparing ODS curvatures before
and after damage.
The major problem of FRF-based damage detection methods lies in the choice of the optimum frequency range for anal-
ysis. Moreover, the FRF requires the measurement of excitation forces and structural responses simultaneously.
As a substitute of the FRF, transmissibility is defined as the relationship between two sets of responses and independent
of input excitations. The transmissibility is a local quantity and is of high sensitivity to damage [39].
Maia et al. [39] utilised the correlations of acceleration response transmissibilities to detect and quantify structural dam-
age. The transmissibility was found more sensitive than FRFs in damage detection and quantification. Li et al. [40] recon-
structed the auto-spectral density functions by using power spectral density (PSD) transmissibility. Damage identification
was then conducted by minimising the difference between the measured and reconstructed PSD functions. Kong et al.
[41] used the transmissibility of vehicle responses in a vehicle-bridge coupled system to detect bridge damage.
4
Table 1
Modal parameter-based damage identification methods.
Features Authors Damage identification level Applications Remarks
Mode shapes Yoon et al. [17] 1-3 Experimental 1D beam and 2D M

plate
Zhang et al. [19] 1 and 2 Experimental plywood/plate B
Feng and Feng [20] 1 and 2∗ Numerical simply supported Output
bridge only
Frequencies and Sun et al. [21] 1 and 2 Numerical beam-like structures B
mode shapes
Zhao and Zhang [22] 1-3 Numerical planar truss beam B, M
Radzieński et al. [23] 1 and 2 Experimental beam-like structures B
Capecchi et al. [24] 1-3 Experimental parabolic arch M
Caddemi and Caliò [25] 1-3 Numericalframe B, M
Khiem and Tran [27] 1-3 Numerical beam-like structures M
Damping Frizzarin et al. [29] 1 and 2 Experimental RC bridge Output
only
Mustafa et al. [30] 1 and 2 Steel truss bridge (viscous M
damping)
Ay et al. [31] 1-3 Experimental steel bridge (viscous B
damping)
Liu et al. [33] 1-3 Numerical shear M
building(non-classically
damping)
FRFs and their Limongelli [34] 1 and 2 I40 bridge B
variants
Limongelli [35] 1 and 2 Numerical frameunder seismic B
excitation
Dilena et al. [36] 1 and 2∗ RC bridge B
Zhang et al. [38] 1 and 2 Experimental plywood beam and Output
plate only
Maia et al. [39] 1-3 Experimental steel beam B
Li et al. [40] 1-3 Experimental steel plane frame M
Kong et al. [41] 1 and 2 Numerical simply supported beam M
Note: ∗ – Quantitatively indicate the relative severity of damage;

B – Baseline information is required;
M – Analytical model is required.
The aforementioned damage detection methods can be classified as damage index methods and are summarized in
Table 1. In the subsequent subsection, different signal processing techniques for damage detection are reviewed.
2.2. Signal processing-based methods
In order to improve damage sensitivity, dynamic responses should be further processed to extract hidden information.
In this regard, a number of signal processing techniques, such as WT, HHT and FD, have been developed and applied for
structural damage detection.
Yang and Nagarajaiah [42] combined independent component analysis with WT for output-only damage identification.
Structural vibration responses were transformed into a wavelet domain and then fed as mixtures into a blind source separa-
tion model, which was examined through independent component analysis. Consequently, the damage information hidden
in wavelet-domain signals was clearly revealed by a sharp spike.
B-spline wavelets have been widely applied to signal processing since its introduction by Chui and Wang [43]. Katunin
[44] derived the analytical formulation of high-order (e.g. fifth-, sixth- and seventh-order) B-spline wavelets, in addition to
the first four B-spline wavelets. The discrete WT with the sixth-order B-spline wavelets was then applied to identify damage
in composite beams by evaluating the singularities of detailed coefficients. In another study, Katunin [45] applied the sixth-
order B-spline wavelets to detect damage to composite plates. Numerical and experimental results showed that high-order
B-spline wavelets could improve the sensitivity and accuracy of damage detection and localisation.
Identification of multiple damage is more challenging than the identification of single damage. Cao et al. [46,47] applied
WT to a curvature mode shape to alleviate noise effect. A Teager energy operator was then implemented to intensify the
local singularities of the signal. With this proposed technique, slight and multiple cracks in beams could be detected even
under high-noise conditions. Recently, Shahsavari et al. [48] presented a statistical procedure to detect low levels of multiple
damage in beams. Continuous WT was first applied to the first mode shape. A damage indicator was then extracted from
the wavelet coefficients through principal component analysis (PCA). Once damage was detected, a likelihood ratio test was
further conducted to determine the likely location.
HHT [49] is a time–frequency analytical technique that is able to process nonlinear and nonstationary signals. Dong et al.
[50] used a vector AR moving average (ARMA) model, unlike conventional HHT, to represent intrinsic mode functions (IMFs)
5
obtained from the empirical mode decomposition (EMD) of vibration signals. A damage index was defined on the basis
of vector ARMA coefficients, which indicated the occurrence and relative severity of damage. Bao et al. [51] developed a
multistage output-only identification scheme based on an improved HHT, which was able to provide Levels 1–3 damage
identification. The improved HHT could decompose structural response data under ambient excitations and possess better
robustness to noise compared with the traditional HHT. Han et al. [52] used HHT for modal identification and damage
detection. Modal parameters were identified through HHT in combination with other techniques, such as random decrement
(RD), natural excitation and stochastic subspace identification.
Aied et al. [53] applied ensemble EMD to the acceleration responses of a bridge to a moving load to detect a sudden
stiffness change. Experimental results revealed that changes in the stiffness were successfully identified even in the presence
of rough profiles, high vehicle speeds and noisy signals.
The FD introduced by Mandelbrot [54] is an effective indicator to characterise irregularities in nonlinear systems. In
comparison with WT that needs high spatial resolution in measurements, FD-based methods only require a small number
of measured points; as such, they are convenient and effective for online data processing and structural damage detection
over the past 10 years.
Li et al. [55] expressed the difference in the angles of sliding windows between two successive points of a displacement
mode shape in FD, and then utilised the change in angle for damage localisation. A damage quantification index was devel-
oped on the basis of the relationship between the angle and modal strain energy (MSE). Bai et al. [56] extended the previous
FD-based methods to a higher mode shape, which was transformed into a new mode shape through affine transformation.
The newly generated mode shape preserved the original damage information and eliminated the local extrema that may
cause false damage identification. The developed method was then successfully applied to detect damage in beam and plate
structures [57].
An intrinsic deficiency of the FD analysis is its susceptibility to measurement noise. To address this problem, Bai et al.
[58,59] applied WT to decompose a mode shape into scale mode shapes before the FD analysis. Damage information and
noise were thus separated because they had different scale characteristics.
The aforementioned studies applied the FD analysis to mode shapes. An and Ou [60] directly analysed acceleration data
through FD to locate damage in order to avoid modal identification errors. FDs were computed at each measured node.
The curvature difference in FDs before and after damage indicated damage location. The proposed method exhibited a high
robustness to noise and was still feasible even if the noise level was up to 15%. Li et al. [61] combined the time–frequency
analysis and FD to identify seismic damage for shear-type building structures by using the acceleration data only. WT was
applied to determine the time–frequency feature, whose FD was then calculated using a box-counting method. Damage-
induced nonlinearity was localised by comparing FDs in different stories.
A bridge and moving loads on it is a nonstationary dynamic system. The interaction between the loads and structure is
closely related to the safety condition of the bridge and thus has been used for damage identification. Hester and González
[62] applied WT to the acceleration response of a bridge to a moving vehicle. The wavelet energy content was employed
as a damage indicator, which was calculated on the basis of a range of scales rather than a given scale, to improve the
sensitivity of the wavelet coefficient to damage. Roveri and Carcaterra [63] used the HHT to identify damage in bridge
structures under a travelling load. A single point response was measured and processed through HHT, and the peak of the
first instantaneous frequency indicated the damage location. Kunwar et al. [64] adopted HHT to locate damage in a bridge
model under transient vibration loads. The relative amplitude of a marginal Hilbert spectrum was used to identify damage
location, and the joint time–frequency distribution referred to damage evolution.
Signal processing-based methods are typically nonparametric and only require experimental data from a damaged struc-
ture. Although these methods are efficient for practical applications without resorting to a baseline model, they are mostly
limited to Level 2 damage identification, that is, damage location. This is because the quantitative relation between the sig-
nal and damage severity cannot be established. Table 2 compares and summarizes the signal processing-based methods that
have been reviewed in this subsection.
2.3. Finite element model updating methods
Model updating methods modify model property matrices, such as mass, stiffness and damping matrices, to ensure that
the analytical predictions of the updated model resemble experimental data as closely as possible [65]. When undamaged
and damaged measurement data are available, changes in structural parameters can be utilised to detect the presence of
damage, identify damage location and quantify damage extent (Levels 1–3 damage identification).
2.3.1. Conventional model updating

Early model updating methods are one-step approaches that directly reconstruct the stiffness and mass matrices of an
analytical model to reproduce the measured modal data [66,67]. The main drawback of these methods is that the updated
mass and stiffness matrices have minor physical meaning, that is, they cannot be related to the changes in the parameters
of the original model. Nowadays, model updating methods are iterative approaches that repeatedly modify the physical
parameters of a FE model. This approach directly changes matrices and adjusts the physical parameters at an elemental or
substructural level. System stiffness and mass matrices are assembled from all elements in a discrete FE model. Therefore,
6
Table 2
Signal processing-based methods.
Authors Methods/Features Damage identification level Applications Remarks
Yang and WT+BSS/approximated and 1, 2, and damage instant Real seismic-excited Output only
Nagarajaiah [42] detailed components building
Katunin [44] WT using fifth-order B-spline 1 and 2 Experimental beams Output only, on-line
wavelet/detail coefficients
Katunin [45] WT using sixth-order B-spline 1, 2 and shape Experimental plates Output only, on-line
wavelet/detail coefficients
Cao et al. [46,47] WT+Teager energy 1 and 2∗ Experimental beams Output only
operator/curvature mode
shape
Shahsavari et al. W T+PCA/W T coefficients 1 and 2 Experimental beams Output only
[48]
Dong et al. [50] EMD+ARMA/ARMA 1∗ Benchmark structures Output only
coefficients
Bao et al. [51] HHT/instantaneous phase and 1, 2∗ , and damage instant Experimental three-storey Output only, on-line
frequency, Hilbert marginal frame
spectrum
Han et al. [52] HHT/instantaneous frequency 1 and 2∗ Experimental 12-story B
and energy, Hilbert marginal reinforced concrete
spectrum frame model
Aied et al. [53] EMD/IMF1 Damage instant and Numerical 3D VBI model Output only
duration
Li et al. [55] FD/FD-based indices 1-3 Experimental beam B
Bai et al. [56] FD+affine transformation/FD 1 and 2∗ Experimentalbeam Output only
trajectory
Bai et al. [57] FD+affine transformation/FD 1 and 2 Experimental plate Output only
trajectory
Bai et al. [58] FD+WT/scale FD trajectory 1 and 2∗ Experimentalbeam Output only
Bai et al. [59] FD+WT/scale FD trajectory Location of delamination Experimental composite Output only
plate
An and Ou [60] FD/waveform FD 1 and 2 Experimental beam B
Li et al. [61] FD+WT/FDs 1 and 2 Numerical building Output only
Hester and WT/wavelet energy 1 and 2∗ Numerical bridge beam Output only
González [62]
Roveri and HHT/first instantaneous 1 and 2 Numerical bridgebeam Output only
Carcaterra [63] frequency
Kunwar et al. [64] HHT/marginal Hilbert 1 and 2∗ Experimental bridge B
spectrum
Note: ∗ – Quantitatively indicate the relative severity of damage;
VBI – vehicle–bridge interaction;
B – Baseline information is required;
M – Analytical model is required.
(1) the matrix properties of symmetry, sparseness and positive definiteness are guaranteed; (2) structural connectivity is
preserved; and (3) changes in the updated global matrices are represented by changes in the updated parameters [68].
Model updating methods in mathematics are regarded as an optimisation problem that minimises the difference between
the measured and predicted responses or referred to as an error function. Similar to data-based methods, measurement data
can be FRFs [69], natural frequencies and mode shapes [70], time histories [71,72], dynamic strain responses [73-77] or a
combination of static and modal test data [78].
2.3.2. Substructuring techniques

Substructuring techniques have been developed and employed efficiently in the structural analysis of large-scale struc-
tures since the 1960s [79]. In these techniques, a global structure is divided into small manageable substructures, each of
which is analysed independently to obtain its designated solution. These solutions are then assembled to recover the so-
lutions of the global structure by imposing constraints at the interfaces. The component mode synthesis (CMS) method is
a popular substructuring technique. It can be classified as the free interface, fixed interface and hybrid methods, accord-
ing to the interface condition of the substructures. Yu et al. [80] proposed a free interface CMS method and applied it to
the element-by-element model updating of large-scale structures. Later, Wang et al. [81] improved the free interface CMS
method for model updating, in which the residual flexibility attachment matrix was constructed without inverting the stiff-
ness matrix. Liu et al. [82] used the CMS method developed in [83] to update the FE model of a scaled arch bridge model.
Papadimitriou and Papadioti [84] proposed a fixed interface CMS method and applied it to damage detection of a highway
bridge, in which each substructure had one unknown parameter for updating.
Weng et al. [85,86] proposed an iterative substructuring method for FE model updating on the basis of Kron’s substruc-
turing concept [87]. During model updating, each substructure was independently handled. When a structure was damaged
in a local area, some specific substructures were re-analysed and assembled with other unchanged substructures to recover
7
the solutions of the global structure without repeatedly analysing the global structural properties. Later, Weng et al. [88] de-
veloped an inverse substructuring method. In their study, the substructural flexibility was extracted from the experimental
modal data of the global structure and then used as a reference for updating the substructural FE model. Eigenvalues and
eigenvectors decomposed from substructural flexibility were employed as damage indicators [89]. Substructural properties
were more sensitive to damage than the global ones because damage typically occurred in a local area.
Yuen and Huang [90] developed an improved Bayesian substructure identification approach based on their previous work
[91]. The improvement was made by modelling the boundary force as filtered white noise, which imposed extra constraints
and thus enhanced the identifiability of the inverse problem.
2.3.3. Regularisation techniques

FE model updating is an inverse problem and typically ill conditioned. Moreover, the number of available measurements
is usually less than that of unknown parameters, resulting in an underdetermined problem. To solve the problems, the
regularisation technique has been developed in the model updating by including an additional item in the error function,
usually a 2-norm item, which leads to a convex error function. This technique is also referred to as Tikhonov regularisation
or l2 regularisation.
Wang and Yang [92] presented a modified Tikhonov regularization in model updating. Since the structural system con-
tains modelling errors and measurement noise, the identified results may diverge after several iterations. In this regard,
they imposed limits on the identified parameters based on their physical meanings to circumvent the divergence problem.
Li and Law [93] proposed an adaptive Tikhonov regularisation, which forced the stiffness reduction factors of intact elements
close to zeros in each iteration. Comparative studies showed that the proposed approach had an obvious advantage over the
traditional Tikhonov regularisation, especially when measurements contained considerable noise. Zhu et al. [94] developed
a sensitivity-based model updating approach with the Tikhonov regularisation in a state-space domain without the need
for input measurements. Damage was identified by minimising the difference between the measured and reconstructed
responses based on transmissibility.
The regularisation parameter plays a critical role in regularisation problems, which controls the trade-off between the
data fidelity and solution sparsity. In general, it is an unknown a priori and problem dependent. A number of methods,
including discrepancy principle [95,96], ordinary and generalised cross validations [97], min–max rules [98] and L-curve
criterion [99], have been developed to determine the optimal regularisation parameter for inverse problems in mathematics.
In structural damage identification, the optimal regularisation parameter of l2 regularisation is typically determined using
the L-curve criterion [92-94, 100]. This criterion utilised a parametric plot of the solution norm versus the residual norm,
and the corner of the curve is regarded as a good choice of the regularisation parameter, which simultaneously satisfies a
small solution norm and a small residual norm [99]. Another widely used technique is generalised cross-validation [101]. In
this technique, the regularisation parameter is calculated by minimising the overall prediction error based on the leave-one-
out rule without any knowledge of the noise variance [97,102].
The Tikhonov regularisation is convenient for implementation, and it has received wide applications in structural damage
detection. However, it tends to produce over smooth solutions, that is, the damage identification results are distributed to
many structural elements, which are inconsistent with practical situations.
2.3.4. Exploiting structural damage sparsity

Structural damage often occurs at several locations only, especially at the early stage, which is sparser than the large total
number of elements in the entire structure. Structural damage sparsity is regarded as essential prior information that can
be utilised for more accurate damage identification. The preliminary use of the sparsity for damage detection can date back
to earlier years in the 1990s, where sparsity prior information is exploited by the minimum rank model updating method
[103-105].
In this connection, structural damage identification can be treated as a sparse recovery problem. Sparse recovery theory,
particularly compressive sensing (CS), has recently attracted considerable interest in a wide range of applications [106-108].
The basic idea manifests that an unknown sparse vector can still be accurately recovered when the number of measurements
is smaller than the size of the entire vector, provided that the sensing matrix satisfies certain incoherence properties [106].
However, this theory has been introduced to structural damage detection since the 2010s. One possible reason is most CS
dealing with linear problems where damage detection is generally a nonlinear one.
Sparse recovery theory adopts the lp (0 ≤ p ≤ 1) regularisation instead of the l2 regularisation [109], that is, a p-norm
item is added to the error function. For a small item of a vector, the p-norm retains a relatively larger weight than the
2-norm. Therefore, a small item significantly contributes to the error function and tends to be penalised by being pushed to
zero, resulting in a sparse solution [110]. Chen et al. [111] and Chartrand [112] demonstrated that lp (0 ≤ p< 1) regularisation
can provide a sparser solution by using fewer measurement data than l1 regularisation. In addition, the lp regularisation is
robust to noise. However, for 0 ≤ p < 1, the corresponding nonconvex optimisation problem is NP-hard. Solving this NP-hard
problem requires a combinatory search and is thus computationally infeasible for large-scale problems. Moreover, a globally
optimal solution cannot be obtained for a nonconvex optimisation problem [113]. For these reasons, the l1 regularisation is
more widely used than thelp regularisation (0 ≤ p < 1) counterpart.
Bao et al. [114] first introduced the CS technology to structural damage detection. Hernandez [115] expanded the
sensitivity-based model updating by using l1 norm minimisation. Following a similar idea, Zhou et al. [116] developed an l1
8
regularisation approach by using the first few frequency data. The proposed technique outperformed the conventional l2 reg-
ularisation through numerical and experimental studies. The effects of the measurement number, damage severity, number
of damage and noise level on the damage detection results were numerically investigated. Zhang and Xu [117] compared
the Tikhonov regularisation with sparse regularisation by using time history data. The reweighted l1 -norm regularisation
technique was adopted to enhance sparsity in the solution. This comparison showed that the proposed sparse regularisa-
tion exhibited certain superiority to the Tikhonov regularisation in terms of the identification accuracy and computational
efficiency. Hou et al. [118] further extended this technique by using frequencies and mode shapes. Wu and Zhou [119] devel-
oped a l1 -regularised one-step model updating approach in which the measured modal data before and after damage was
compared directly. In this manner, model updating at an undamaged stage was unnecessary.
Zhang et al. [120] combined the extended Kalman filter (EKF) and l1 regularisation for damage identification by using free
vibration responses. The original unconstrained optimisation problem was transformed into an optimisation problem with
the l1 -norm constraint, and a pseudo-measurement technique was utilised to enforce the constraint into each recursive
step of EKF. Recently, Huang et al. [121] proposed an improved EKF method based on the lp regularisation. A novel L-
surface approach was used to determine an appropriate p. Numerical and experimental examples showed that the proposed
improved EKF method was superior to EKF with Tikhonov and l1 regularisation methods in terms of identification accuracy
and required measurement quantities.
Zhou et al. [122] presented an iteratively reweighted l1 regularisation algorithm, which closely resembled the l0 regular-
isation. The nonconvex l0 regularisation problem was solved by transforming it into a series of weighted l1 regularisation
problems. An experimental example demonstrated that the proposed algorithm was able to provide sparser damage identi-
fication results with higher accuracy than the l1 regularisation.
Chen et al. [123] employed a weighted strategy and trace the least absolute shrinkage and selection operator (Lasso)
for FE model updating by using modal parameters. Different weight coefficients were used to balance information from
frequencies and mode shapes. The trace Lasso improved the accuracy and stability of the l1 regularisation, especially when
unknown variables were highly correlated. The ant lion technique was employed to solve the optimisation problem. Later,
Chen and Yu [124] combined the optimiser with an improved Nelder–Mead algorithm to improve the local searching ability.
A comparison study indicated that the proposed algorithm was more robust and accurate than the ant lion optimiser and
required more computational time.
Wang and Lu [125] proposed a new error function that was decoupled from the damage parameters. The new error
function, including an l1 regularisation term, was solved using an alternative minimisation approach without a sensitivity
analysis. Ding et al. [126] proposed a novel error function based on sparse regularisation and Bayesian inference by consid-
ering the uncertainty effect and a limited number of measurements. A new heuristic algorithm, namely, the Jaya algorithm,
was employed for the optimisation. Comparison results showed that the proposed error function yielded more reliable and
accurate identification results than those with either sparse regularisation or Bayesian inference alone.
Numerous methods have been developed to select the regularisation parameter for l2 regularisation, whereas few meth-
ods have been devised for lp regularisation (0 ≤ p ≤ 1) [109]. This gap is due to the fact that the former has a closed-form
solution, whereas the latter does not. In SHM and damage identification, an appropriate regularisation parameter of an
l1 -regularised problem is typically selected on the basis of experience. Mascarenas et al. [127] heuristically selected the
regularisation parameter as a unit. Yang and Nagarajaiah [128] set the regularisation parameter as√ 0.01 in CS-based modal
identification. Yang and Nagarajaiah [129] calculated the regularisation parameter by using β = 1/ N, where N is the num-
ber of the time history sampling points corresponding to the dimension of an unknown vector. Zhang and Xu [117] used
the reweighted l1 regularisation technique to determine the regularisation parameter. Yao et al. [130] showed that the plot
of the residual term versus the regularisation term on a linear scale resembled an “L” shape. Afterwards, they selected the
regularisation parameter corresponding to the corner of the L curve.
More recently, Hou et al. [131] developed two strategies to select the regularisation parameter of the l1 regularisation
problem. The first method utilized the residual and solution norms of the optimisation problem and ensured that they are
both small. The other selection criterion was based on the discrepancy principle, which required that the variance of the
discrepancy between the calculated and measured responses was close to the variance of the measurement noise. A range
of the regularisation parameter, rather than one single value, could be determined using these two strategies. Wang and Lu
[125] selected the optimal regularisation parameter based on the threshold setting method, which was closely related to
two threshold parameters determined through numerical studies.
In l1 -regularised damage detection, the sensitivity matrix serves as the sensing matrix and is directly related to sensor
locations. Sensor placement is a typical combinatorial problem, and the global optimum is difficult to obtain using conven-
tional techniques. In this regard, Hou et al. [132] developed a GA-based OSP technique such that the columns of the resulting
sensitivity matrix have maximum independence. Although the optimal sensor location is generally damage-dependent, the
proposed technique worked on the sensitivity matrix in the undamaged state and did not need the prior knowledge of
damage location and severity.
Although the use of FE model updating for damage detection has been significantly developed, these methods have some
limitations. For example, the performance of these methods largely depends on the accuracy of the analytical FE model. In
terms of optimising a large-scale and complex structure, constructing a well-conditioned sensitivity matrix is difficult, and
the computational load is heavy. The FE model updating methods reviewed in this subsection have been summarised in
Table 3.
9
Table 3
Finite element model updating methods.
Method Feature Application Authors
Conventional model updating

FRFs Numerical six-bay truss Sipple and Sanaye [69]
Frequencies and mode shapes Experimental full-scale RC Moaveni et al. [70]
building slice
Accelerations Experimental RC bridge Jafarkhani and Masri [71]
Accelerations Experimental steel frame Li et al. [72]
Dynamic strain Experimental steel frame Li et al. [73-77]
Static strains, displacements, Benchmark bridge model Sanayei et al. [78]
slopes, frequencies and
mode shapes
Substructuring techniques
Free interface CMS Frequencies and mode shapes Numerical frame Yu et al. [80]
Free interface CMS Frequencies and mode shapes Numerical wing structure and Wang et al. [81]
bolted plate
Free interface CMS Frequencies and mode shapes Experimental arch bridge Liu et al. [82]
Fixed interface CMS Frequencies and mode shapes Numerical highway bridge Papadimitriou and Papadioti
[84]
Kron’s substructuring method Frequencies and mode shapes Numerical frame Weng et al. [86]
Inverse substructure method Frequencies and mode shapes Experimental steel frame and Weng et al. [89]
numerical tower
Bayesian substructure method Accelerations Numerical shear building Yuen and Huang [90]
Regularisation techniques
Tikhonov regularisation Accelerations Numerical 3D frame Wang and Yang [92]
Adaptive Tikhonov Accelerations Numerical plane truss Li and Law [93]
regularisation
Tikhonov regularisation Accelerations Experimental steel beam Zhu et al. [94]
L1 regularisation Frequencies and mode shapes Numerical truss structure Bao et al. [114]
L1 regularisation Frequencies Numerical beam and plate Hernandez [115]
L1 regularisation Frequencies Experimental steel beam Zhou et al. [116]
Reweighted l1 regularisation Accelerations Experimental steel beam Zhang and Xu [117]
L1 regularisation Frequencies and mode shapes Experimental steel frame Hou et al. [118]
L1 regularisation Frequencies Experimental steel beam Wu and Zhou [119]
L1 regularisation Frequencies Experimental beam and frame Zhang et al. [120]
Lp regularisation Accelerations Experimental steel shear Huang et al. [121]
building
Iteratively reweighted l1 Frequencies and mode shapes Experimental 3D steel frame Zhou et al. [122]
regularisation
Trace Lasso Frequencies and mode shapes Experimental steel beam Chen et al. [123]
Weighted trace Lasso Frequencies and mode shapes Experimental steel beam Chen and Yu [124]
L1 regularisation Frequencies and mode shapes Experimental steel beam Wang and Lu [125]
L1 regularisation Frequencies and mode shapes Experimental RC bridge Ding et al. [126]
Note: The methods listed in the table all achieve Levels 1–3 damage identification.
2.4. Optimisation algorithms
Optimisation algorithms have been employed by many researchers for damage detection and can be regarded as an effec-
tive alternative to the sensitivity-based FE model updating technique in solving inverse problems. Traditional optimisation
methods are usually gradient-based and require a good initial value, thereby limiting their potential applications. With the
development of computational intelligence, a number of optimisation algorithms, such as GA, ANN, particle swarm opti-
mization (PSO) and artificial bee colony (ABC), have been proposed. These algorithms do not rely on specific formulas for
optimisation and thus avoid the aforementioned shortcomings. Moreover, these algorithms are effective in dealing with un-
certainties and insufficient information, which are typical problems in structural damage detection. ANN is an important ML
technique and will be detailed in Subsection 2.6.
GA, developed in the 1970s [133], is based on the concept of natural selection and has been utilised for damage detection
since the 1990s. The main critical problem of GAs is the heavy computational load due to the high dimension of the search
space.
Meruane and Heylen [134] implemented a hybrid real-coded GA to locate and quantify structural damage by using five
different parameters, namely, frequency, modal displacement, MAC, MSE and modal flexibility. In comparison with conven-
tional optimisation methods, the proposed approach could reach a more precise solution. Ghodrati Amiri et al. [135] com-
pared a pattern search and GA for damage identification in plates. The numerical study indicated that the GA provided
better results than the pattern search in some cases.
Guo and Li [136] combined the evidence theory and PSO for multiple damage identification. First, an information fusion
method was applied to detect damage sites by integrating the damage localisation information from MSE and natural fre-
10
Table 4
Optimisation methods.
Method Strengths Limitations Features Damage identification level Applications Authors
GA A global Heavy
optimizer computation
Real-coded GA Frequencies, mode 1-3 Experimental 3D Meruane and
shapes and truss Heylen [134]
variants
Binary-coded GA Frequencies and 1-3 Numerical plate Ghodrati Amiri
mode shapes et al. [135]
PSO Efficient, a few Difficult to
function control the
evaluations, a balance
few between
parameters to exploration
adjust and
exploitation
Mutation PSO Frequencies and MSE 1-3 Numerical truss Guo and Li [136]
Hybird PSO Frequencies and 1-3 Phase I Chen and Yu
mode shapes IASC-ASCE [137]
benchmark
problem
ABC Simple structure, Improper
high flexibility, exploitation
good for compli-
robustness cated prob-
lems, slow
convergence
rate
Modified ABC Frequencies and 1-3 Numerical beam Ding et al. [138]
mode shapes and plate
Improved ABC Frequencies 1-3 Experimental Ding et al. [139]
steel beam
quencies. An improved PSO was then used to determine the damage extent. Chen and Yu [137] combined the PSO algorithm
and an improved Nelder–Mead method to maximise the likelihood function in Bayesian inference constructed using natural
frequencies and mode shapes. The identification results obtained by searching the local area around the optimum solution
found by PSO were more stable and accurate than those obtained by the PSO-based algorithm.
Ding et al. [138] presented a modified ABC algorithm to optimise the objective function by using modal parameters. Two
modifications were introduced to improve the convergency rate and local search ability of the ABC algorithm. The numerical
study revealed that the damage identification results of the proposed algorithm were more accurate than those of other
evolutionary algorithms, such as GA and PSO. Later, they applied the modified ABC algorithm to identify cracks in beams by
using natural frequencies only [139].
The aforementioned optimisation methods are compared and summarized in Table 4.
2.5. Statistical time series methods
Time series-based methods typically fit time series models to the measured time history data and then extract fea-
tures sensitive to variations caused by damage and insensitive to operational and environmental variations. AR models
and their variants, such as autoregressive with exogenous input (ARX) and ARMA models, are commonly used to extract
damage-sensitive features based on residual errors or AR model parameters [140]. These methods inherently account for
uncertainties and do not depend on the physical models. As such, they are more suitable for automated SHM systems than
model updating methods. Statistical time series methods mainly consist of three components: (i) random excitation and/or
response signals, (ii) statistical time series model building and (iii) statistical decision making for damage diagnosis.
Mahalanobis squared distance (MSD) is a statistical measure for outlier detection and has received wide applications be-
cause of its simplicity and computational efficiency. Mosavi et al. [141] located damage to continuous structures under am-
bient vibrations by using vector AR models and MSDs. Statistical evaluations were performed on extracted damage features
for each individual sensor location. A sensor with a significant variation was identified as the one closest to the damage
location. However, a dense array of accelerometers was required to identify the accurate damage location. Chang and Kim
[142] performed multivariate AR analysis and extracted frequencies, mode shapes and damping ratios as damage-sensitive
features. The outlier analysis was then conducted on the basis of MSD. A field experimental study on a simply supported
steel truss bridge showed that the inclusion of additional parameters in the outlier analysis might lead to more sensitive
features.
Although the MSD-based method possesses several distinguishing advantages, it requires multivariate training data from
the undamaged structure. Previous studies achieved an approximation of multinormal data by increasing the observation-
11
to-variable ratio [143,144]. However, the number of measurements is limited in many practical applications. To this end,
Nguyen et al. [145] proposed a data generation scheme based on Monte Carlo simulation. Simulation data was added as in-
put parameters to compensate for the data shortage. Thus, the computational stability and reliability of MSD-based damage
identification were improved.
Gul and Catbas [146] constructed ARX models by using the free acceleration responses of a structure and developed two
different approaches to extract damage-sensitive features. In the first approach, the B-term coefficients of the ARX models
were directly used, and numerical results showed that this approach was effective for simple and noise-free models. In the
second approach, the ARX model fit ratios were selected as the damage feature to consider the noise effect and model com-
plexity. The difference between fit ratios indicated the relative change in damage severity, although direct quantification was
not achieved. Numerical and experimental results showed that the second approach performed successfully under different
damage scenarios for complex models and test specimens. Later, they extended this method to the ambient vibration case,
in which RD was applied to obtain pseudo-free vibration data from ambient vibration time histories [147].
Yao and Pakzad [148] proposed two new damage features: one was based on the Ljung–Box statistic of the AR model
residual and the other on the Cosh spectral distance of the AR model spectrum. The results of Ljung–Box statistic were
more accurate than those of the existing algorithms based on the AR model residual variance and coefficient distance.
The Cosh spectral distance was less sensitive to changes in excitation sources. Lakshmi et al. [140] conducted a singular
spectral analysis to enhance the sensitivity of the damage features derived from the auto-regressive moving average with
an exogenous input (ARMAX) model. Shahidi et al. [149] investigated the performance of four damage features derived
from single- and multivariate regression models in detecting the timing and location of the structural damage. Sequential
normalised likelihood ratio test and two-sample control statistics were adopted to detect the change in two families of
damage features.
A critical problem for efficiently employing the AR models for feature extraction is the determination of an appropriate
model order. Traditionally, the optimal model order was selected in an ad hoc manner. In this regard, Figueiredo et al.
[150] presented four techniques based on the Akaike information criterion, partial autocorrelation function, root mean square
error and singular value decomposition (SVD) to determine the appropriate model order. An appropriate range of the model
order, rather than a unique value, could be determined using the four proposed techniques. A comparative study was carried
out to investigate the influence of the model order on damage detection results.
Although the aforementioned methods are successful in identifying the presence and location of damage, they provide
limited information about damage severity. Therefore, they belong to Levels 1−2 damage identification.
Kalman filtering is a useful technique for time series analysis and state estimation and has received wide applications in
signal processing and econometrics. Lei et al. [151] used an extended Kalman estimator to identify structural parameters and
unknown excitation sequentially. The proposed algorithm was extended to damage identification of large-scale structures
based on the substructural approach. The inter-connection effect between adjacent substructures was estimated without the
measurements of the substructure interface degrees of freedom (DOFs). This feature was a major advantage over previous
substructural identification approaches, which require all the responses at the substructure interfaces DOFs.
In a traditional EKF approach, unknown structural parameters were incorporated into an extended state vector. When
a large number of unknown parameters and extended state vectors are identified simultaneously, a divergence problem
may arise. In this regard, Lei et al. [152] proposed a two-step Kalman filter approach. The structural state vector was first
recursively estimated using the traditional Kalman filter technique with the assumed structural parameters, which were
then estimated given the state vector. The proposed algorithm reduced the number of estimated parameters in each step,
and thus improved estimation convergence. They [153] adopted the approach to detect damage of frame structures in which
a beam-column joint was modelled as in Weng et al. [154].
The aforementioned statistical time series methods are compared and summarized in Table 5.
2.6. ML methods
Structural damage identification can be treated as a pattern recognition problem, which is divided into three parts: (1)
data acquisition, (2) feature extraction and (3) feature classification [16]. Feature extraction aims to fit either a data-driven
or a physics-based model to the measured structural response data by using statistical or signal processing techniques. The
parameters of these models or model residuals are then selected as damage sensitive features. Finally, with the selected
features, the classification algorithm is utilised to determine the presence, location and severity of damage.
In recent years, a number of ML classifiers have been utilised and developed for structural damage identification [155].
These ML algorithmscan be broadly divided into supervised, unsupervised and semi-supervised learning modes.
2.6.1. Supervised learning

Most ML algorithms are based on a supervised learning manner, which requires features of both undamaged and dam-
aged states of the structure with their labels to establish a statistical model during the training process [156]. Three com-
monly used classification methods, i.e., ANN, support vector machine (SVM) and random forest (RF), will be reviewed in this
subsection, and their strengths and limitations are provided in Table 6.
The ANN technique is a widely used ML algorithm, which has been introduced to civil engineering since the 1980s [157].
ANNs have drawn considerable attention in SHM and damage identification due to their ability of pattern recognition and
12
Table 5
Statistical time series methods.
Authors Methods Features Damage Applications Remarks

identification
level
Output only Mosavi et al. [141] MSD AR coefficients 1 and 2 Experimental steel Require
beam multivariate
training data
Chang and Kim MSD Frequencies and 1 and 2 Steel truss bridge Require
[142] mode shapes multivariate
training data
Nguyen et al. [145] MSD, Monte Carlo AR vectors 1 and 2 Benchmark Compensate
simulation building for data
shortage and
multivariate
data
condition
Gul and Catbas ARX ARX coefficients, 1 and 2∗ Experimental steel Free
[146] ARX model fit grid structure acceleration
ratios responses
Gul and Catbas RD ARX model fit 1 and 2∗ Experimental steel Ambient
[147] ratios grid structure acceleration
responses
Yao and Pakzad Ljung–Box AR model residual, 1 and 2 Experimental truss Insensitive to
[148] statistic/Cosh AR model and bridge slab changes in
spectral distance spectrum excitation
sources
Lakshmi et al. Singular spectrum ARMAX models 1 and 2 Benchmark Forced
[140] analysis bookshelf acceleration
structure and responses
experimental RC
beam
Shahidi et al. [149] Control chart Coefficients of 1, 2, and damage Experimental steel Require
single-variate instant frame multivariate
regression, training data
collinear
regression, ARX
and AR
Figueiredo et al. Akaike information AR residual errors 1 Experimental Determine
[150] criterion and and model aluminum frame optimal AR
SVD parameters model order
Input-output Lei et al. [151] EKF Accelerations 1-3 Numerical study Use limited
on Phase I input mea-
IASC-ASCE surements
benchmark
problem, beam,
and truss
structure
Lei et al. [152] Two-step Kalman Accelerations 1-3 Numerical study Reduce the
filter on Phase I number of
IASC-ASCE estimated
benchmark parameters
problem and a
shear building;
Experimental
frame
Lei et al. [153] Two-step Kalman Accelerations 1-3 Experimental steel Reduce the
filter frame with joint number of
damage estimated
parameters
∗
Note: – Quantitatively indicate the relative severity of damage
error tolerance in establishing a nonlinear relationship between the inputs and outputs. For structural damage identification,
the ANN is used to establish a model representing the relationship between features extracted from structural vibration data
and structural model parameters through a training process. This trained ANN model is then capable of identifying damage
from measurement data [158]. The learning algorithm of ANNs can either be supervised or unsupervised, whilst most of
them are supervised, especially in damage detection applications.
Jiang et al. [159] proposed a two-stage approach by combining fuzzy NNs and data fusion techniques. In this approach,
structural modal parameters were fed into fuzzy NNs as inputs. The data fusion technique was applied to the outputs from
different fuzzy NNs, and a consistent and reliable damage assessment result was obtained. Dackermann et al. [160] identified
13
Table 6
Supervised learning methods.
Method Strengths Limitations
ANN Self-learning, flexible; Suitable for complex and nonlinear Computational expensive; Prone to overfitting
problem
SVM Effective for high dimension data; Fault-tolerance; Prone Computational expensive in large dataset; Poor performance
to global optimal solution for noisy datasets with overlapping classes; Binary
classification algorithm; Tricky selection of kernel function
RF Accurate and stable; Computational efficient; Can extract Sensitive to noise and outliers; Not easy to interpret; Need to
variable importance; Able to reduce variance tune hyper-parameters
the member connectivity and mass changes in a frame structure using ANNs. In their study, individual networks were first
trained with FRF data at different measurement locations. The outcomes of each network were then fused through a network
ensemble to generate final damage conditions. The proposed network ensemble technique was superior to the approach that
simply added the FRF data to train the ANN. Xu et al. [161] constructed NNs to locate and quantify joint damage by directly
using dynamic displacement responses and excitation information.
Hakim and Razak [162] trained an ANN with natural frequencies and utilised it to quantify damage severity in a steel
girder bridge model. Moreover, Hakim and Razak [163] compared an ANN with an adaptive neuro-fuzzy inference system by
using the same experimental model. The latter incorporated ANNs and fuzzy logic systems in a single framework and had
benefits of both techniques. Experimental results showed that the damage identification results of the proposed framework
were more accurate than those of the ANN.
However, a considerable amount of computational effort is required in the ANN techniques, especially when large DOFs
are involved. Therefore, the ANN-based damage identification is generally applicable to small structures with a limited num-
ber of DOFs. To this end, Bakhary et al. [158] used a multi-stage ANN method with a substructure technique to detect the
damage location and extent. The full structure was divided into substructures, and each substructure was independently
analysed to identify damage progressively. In this manner, the size of the ANN models was reduced, and the computational
effort was saved.
Principal component analysis (PCA) is a statistical technique for dimension reduction and feature extraction. This method
reduces a large set of correlated variables to a low dimension through orthogonal transformation whilst retaining the most
relevant information. PCA has been extensively applied to measured structural vibration responses for reduced-order mod-
elling, modal analysis and parameter identification. Worden et al. [164] first applied the PCA in structural damage detection.
Some researchers [160,165-167] applied PCA to reduce the dimension of the FRF data and then utilised the ANN to train the
FRF data for damage detection.
To improve damage identification accuracy, Bandara et al. [167] investigated the number of hidden layers and the number
of neurons per hidden layer to formulate an optimal architecture of ANN with minimum training and testing errors.
Recently, with the improvement of computing capacity and network architecture, deep learning algorithms evolved from
the ANN, e.g., convolutional NNs (CNNs), have been developed rapidly [168]. Abdeljaber et al. [169] proposed adaptive 1D
CNNs, which fused feature extraction and classification blocks into a single and compact learning body. Consequently, these
NNs could directly learn from the acceleration data measured under known random excitations. Since modal identification
was not required, it could be implemented near-real-time and suitable for online SHM. Duan et al. [170] used a CNN tech-
nique to detect damage in bridge hangers, in which the Fourier amplitude spectra of the acceleration responses were used
as the input. Bao et al. [171] transformed the time series signals into image data. The randomly selected and manually la-
belled image data were then used to train the deep NNs via the greedy layer-wise training technique and automatically
detect anomalies of a cable-stayed bridge.
SVM is a supervised learning model that aims to separate two classes of data. It is trained to estimate the boundary
between two classes by maximizing the margin and minimizing the misclassification [172]. SVM is becoming increasingly
popular these years for damage identification because of its superior ability to solve nonlinear, high-dimensional and small
sample problems [173]. In comparison with conventional NNs, SVM overcomes the problems of local minimisation and
inadequate statistical capabilities [174].
Kourehli [175] used the first two incomplete mode shapes and natural frequencies as input data to train the SVM. In
this approach, a radical basis function (RBF) was chosen as a kernel function. The parameters of the kernel function were
determined on the basis of the coupled simulated annealing and standard simplex method. Liu and Jiao [176] used GA to
optimise the SVM parameters with similar input data and kernel function to detect damage of bridges. Numerical studies
on a simply supported bridge have demonstrated the feasibility and superiority of the GA-SVM algorithm to RBF networks
and back propagation NNs (BPNNs) optimised by GA.
Ghiasi et al. [174] introduced the thin plate spline Littlewood–Paley wavelet kernel function to improve the learning
ability of SVM. In this approach, feature vectors, as the input of the SVM, were extracted from the acceleration responses
through the wavelet packet decomposition. A social harmony search algorithm was used to determine the parameters of the
SVM. In comparison with the SVM based on other combinational and conventional kernels, the proposed kernel achieved an
enhanced performance for multiple damage identification.
14
Gui et al. [177] compared SVM on the basis of three optimisation algorithms, namely, grid search, PSO and GA, to optimise
the penalty parameters and parameters of the RBF kernel parameters. Two types of features, namely, the parameters of the
AR model and the residual errors of the statistical parameters, were extracted from the time series data. The optimisation-
based methods significantly improved the sensitivity, accuracy and effectiveness of the conventional SVM. Using the residual
errors achieved a significantly higher accuracy than using the AR type.
RF is an ensemble classifier that consists of a large number of decision trees [178]. The model prediction is obtained
through combining the predictors of each individual tree by majority voting. Zhou et al. [179] proposed a damage detection
method by RFs and data fusion. In this method, the wavelet packet decomposition was applied to decompose acceleration
signals into energy features, which were fused into new energy features through data fusion. The obtained features were
then inputted into RFs to classify structural damage. Experimental results showed that the accuracy and stability of the
proposed method were higher than those of RF alone, SVM alone and SVM and data fusion.
The supervised ML techniques require data from undamaged and damaged structures for training purposes. However,
data associated with various damage scenarios may be unavailable in practical structures. Most studies have generated
training samples from laboratory testing or through numerical FE simulations. Therefore, the efficiency of the supervised
learning approaches depends on the model accuracy. For this reason, the development of unsupervised algorithms is of
particular interest.
2.6.2. Unsupervised learning

An unsupervised learning algorithm only requires data from the intact state of a structure for training, which belongs
to the outlier or novelty detection category. A model is trained by machine learning algorithms based on the data in the
undamaged state. The trained model is then used to evaluate the structural condition when new measurement data are
available. If the difference between the measured data and those predicted from the model exceeds a threshold, the struc-
ture is regarded as a deviation from its normal condition and is probably damaged.
Santos et al. [180] combined two statistical learning methods for online early-damage detection. Multi-layer perceptron
NNs were used for the statistical modelling of the structural responses. The unsupervised K-means clustering algorithm was
employed to classify the NNs’ estimation errors. These methods were sequentially applied in successive time windows to
realize continuous on-line damage identification.
Neves et al. [181] developed a model-free ANN-based approach for damage detection of bridge structures. ANNs were
trained with an unsupervised learning approach using accelerations collected on the healthy bridge. The prediction errors
of each network were then statistically characterized by a Gaussian process to determine a damage detection threshold.
Consequently, the structural condition, namely healthy or damaged, was determined by comparing damage indices with the
selected threshold.
Rafiei and Adeli [182] used an unsupervised restricted Boltzmann machine to extract features from the frequency domain
of the ambient vibration signals. A structural health index was established for each substructure in terms of a probability
density function (PDF), which measured the similarity between the ambient vibrations of the current state of the structure
and those of the healthy one. The larger the difference, the higher the likelihood of damage.
Cha and Wang [183] revised the original density peaks-based fast clustering algorithm to an unsupervised machine learn-
ing method to detect and locate structural damage. An intact statistical model was built by using the training points from
each sensor in the intact state of the structure. The sensor location corresponding to the novelty point was identified as the
location where damage occurred.
Avci and Abdeljaber [184] proposed an unsupervised damage detection algorithm based on self-organizing maps, which
is a class of ANNs. In their algorithm, self-organizing maps were used to extract the damage indices from the random
acceleration responses of the monitored structure. The summation of the indices indicated the overall condition of the
structure, the value of which could be used to evaluate the damage severity.
Although unsupervised learning approaches are preferred for practical damage detection, most of them are limited to
novelty detection, i.e., Level 1 damage identification, and fail to give additional information in terms of the location and
severity of damage.
2.6.3. Semi-supervised learning

In practice, the acquisition of fully labelled data for training is infeasible, whilst a small number of labelled data may
be available. In such situations, semi-supervised learning can be of great practical value, which falls between unsupervised
and supervised learning, using both labelled and unlabelled data for training classifiers. Many researchers have found that
using unlabelled datain conjunction with a small amount of labelled data may considerably improve the accuracy of ML
algorithms [155]. Rather than pure novelty detection, semi-supervised learning approaches are able to locate and quantify
structural damage. However, the applications of semi-supervised ML algorithms for damage identification are very limited
in the literature.
Chen et al. [185] combined multi-resolution classification with semi-supervised learning for damage detection of bridge
structures. The features were extracted from localised time-frequency sub-bands. The adaptive graph filter classifier was
used to classify unlabelled data given previously labelled signals. A weighting algorithm was developed to combine infor-
mation from both labelled and unlabelled signals to make a global decision. Furthermore, in addition to unlabelled data, the
adaptive graph filtering was able to handle mislabelled as well as unseen signals.
15
Table 7
ML methods.
Authors Methods Inputs Damage Applications Remarks

identification
level
Supervised learning
Jiang et al. [159] Fuzzy NN+data Frequencies and 1-3 Numerical shear-type
fusion mode shapes building
Dackermann et al. ANN+PCA+NN FRF 1, 2 Experimental frame Joint damage and
[160] ensembles mass changes
Xu et al. [161] NNs Dynamic 1-3 Experimental steel
displacement frame
Hakim and Razak ANNs Frequencies 1 and 3 Numerical steel girder
[162] bridge
Hakim and Razak Fuzzy NN Frequencies 1 and 3 Experimental steel
[163] girder bridge
Bakhary et al. ANN+substructrue Frequencies and 1-3 Numerical RC slab and
[158] technique mode shapes frame
Li et al. [165] ANN+PCA+NN FRF 1-3 Experimental steel
ensembles beam
Samali et al. [166] ANN+PCA FRF 1-3 Experimental steel Notch-Type
frame damage
Bandara et al. ANN +PCA FRF 1-3 Numerical framed Optimal
[167] structure architecture of
ANN
Abdeljaber et al. 1D CNNs Acceleration under 1 and 2 Experimental steel On-line, joint
[169] random frame of a damage
excitations grandstand simulator
Duan et al. [170] CNNs Fourier amplitude 1-3 Numerical tied arch Automatic
spectra of bridge
wind-induced
acceleration
Bao et al. [171] DNNs Image vectors 1 Long-span cable-stayed Automatic,
converted from bridge real-time
acceleration
Kourehli [175] SVM Frequencies and 1-3 Numerical beam, plane RBF kernel
mode shapes frame, and function
spring-mass system
Liu and Jiao [176] SVM Mode shape ratio 1-3 Numerical simply RBF kernel
and frequency supported bridge function
rate
Ghiasi et al. [174] SVM Wavelet energy 1 and 2 Numerical Phase I Thin plate spline
spectrum IASC-ASCE Littlewood–Paley
benchmark problem wavelet kernel
and a 120-bar dome function
truss
Gui et al. [177] SVM AR coefficients and 1 and 2 Experimental RBF kernel
residual errors of benchmark frame function
the statistical
parameters
Zhou et al. [179] RF+data fusion Energy features Damage type Numerical steel
from and location benchmark frame
acceleration and experimental
steel shear frame
Unsupervised learning
Santos et al. [180] NNs+K-means Time-series 1 Experimental On-line
clustering displacements cable-stayed bridge
and rotations
Neves et al. [181] ANNs Accelerations from 1 Numerical railway
passing vehicle bridge
Rafiei and Adeli Deep Boltzmann Frequency domain 1 ExperimentalRC
[182] machine of ambient building
vibration
Cha and Wang Density Crest factor and 1 Experimental steel
[183] peaks-based fast T-continues WT structure
clustering extracted
(continued on next page)
16
Table 7 (continued)
Authors Methods Inputs Damage Applications Remarks

identification
level
Avci and Self-organizing Random 1 Phase II IASC-ASCE

Abdeljaber [184] maps acceleration benchmark problem
responses
Semi-supervised learning
Chen et al. [185] Multi-resolution Localised 1 Experimental
classification+label time-frequency bridge-vehicle
propagation+ sub-bands dynamic system
Adaptive Graph
Filter
Lai and Sparse Velocity, 1-3 Experimental steel Linear/nonlinear
Nagarajaiah identification+pseudo- acceleration and frame, benchmark damage
[186] force displacement frame, a
base-isolated
building
Rogers et al. [187] Dirichlet process FRF and 1 Benchmark building,
clustering frequencies Z24 bridge
Lai and Nagarajaiah [186] developed a semi-supervised algorithm to detect and characterize linear/nonlinear structural
damage. The baseline (undamaged) model was established using a sparse identification method based on supervised learn-
ing with input-output time history data. Damage was considered as a variation of the restoring force, and thus the damaged
system was transformed into an equivalent linear system subjected to external disturbance forces and pseudo-forces. Con-
sequently, the nonlinearity (including presence, type, and extent of damage) was represented by the pseudo-forces and
discovered in an unsupervised way without the creation of various damage scenarios.
Rogers et al. [187] used Dirichlet process clustering models for online damage detection based on features in the high
dimensional frequency domain. The algorithm learned clusters of data online without a training phase and then assigned
labels to new clusters in a semi-supervised manner. The model with already known states of the structure was continually
updated as more data were added. As time progressed, the method learned more states and the robustness of the method
increased.
In recent years, ML algorithms are growing rapidly and have received considerable attention in damage identification.
However, there are still some challenges and difficulties requiring further research. The training dataset is extremely impor-
tant for the performance of ML algorithms. Consequently, data selection, data cleaning, data compression, data fusion, data
normalization, and data labelling are inevitable to establish the appropriate datasets. The processes are time consuming and
labour intensive. Moreover, for structural damage detection, the lack of enough training samples may lead to over-fitting
problems, e.g. over-extraction of irrelevant features such as measurement noise [188]. The generalisation is another critical
problem for ML algorithms. A well trained and validated model may only perform well for a specified type of structures
and a particular pattern of damage. Table 7 compares and summarizes the ML methods that have been reviewed in this
subsection.
2.7. Bayesian methods
Civil structures are generally subjected to significant measurement noise and modelling errors, which may lead to incor-
rect damage identification [189,190]. For example, the existence of measurement noise may mask subtle structural changes
caused by damage. Consequently, deterministic methods may fail once they are applied to practical civil structures. In this
regard, many researchers proposed probabilistic damage identification approaches [191]. Amongst these methods, Bayesian
inference has attracted considerable attention since the 1990s, which explicitly quantifies the posterior probability of uncer-
tainties based on observations and prior information [192,193]. Apart from addressing uncertainties, Bayesian methods also
provide an efficient way to deal with the ill-posed inverse problem by specifying probability distributions over uncertain
parameters; this approach is equivalent to introducing a regularisation term to the optimisation problem [194].
Figueiredo et al. [195] developed a Bayesian pattern recognition approach based on a Markov-chain Monte Carlo method.
The Bayesian approach was employed to cluster structural responses into a reduced number of global state conditions by
using a finite mixture of Gaussian distributions. Outlier detection was then conducted on the basis of MSD. The applicability
of the proposed approach was demonstrated using the data sets from the Z-24 Bridge.
Arangio and Beck [196] used the Bayesian NN for bridge integrity assessment under ambient vibrations. In this method,
an optimal network architecture was determined on the basis of Bayesian model class selection. An automatic relevance de-
termination (ARD) method was applied to measure the relative importance of different inputs in NNs and separate relevant
variables from redundant ones. Comparison studies indicated that the accuracy of the optimal network model in damage
localisation and quantification was better than that of a heuristic-based model. Later, Arangio and Bontempi [197] applied
a Bayesian NN to the Tianjin Yonghe Cable-Stayed Bridge and detected cracks at the external portions of both spans and
damage at two piers. In this rare case, the accelerations of the deck before and after damage were employed.
17
Lam et al. [198] detected damage to a railway ballast by using modal parameters within a Bayesian framework. In their
study, the ballast under a concrete sleeper was uniformly divided into a number of regions with similar stiffness. The num-
ber of divided regions was determined on the basis of the Bayesian model class selection method. The posterior PDF of the
ballast stiffness in different regions was approximated by a multivariable Gaussian distribution.
Behmanesh and Moaveni [199] implemented a Bayesian FE model updating to identify damage to a full-scale structure.
Damage was simulated by adding concrete blocks onto a bridge deck. The adaptive Metropolis–Hastings algorithm was used
to sample the posterior distribution of the updating parameters. Behmanesh et al. [200] investigated the effects of the subset
of modes used on the performance of Bayesian FE model updating. The optimal subset of modes in the model updating
process was determined using Bayesian model class selection. Damage identification was then conducted for different weight
factors, and the final estimation was obtained by averaging all the results via the Bayesian model averaging technique.
Yin et al. [201] developed a probabilistic damage identification method for bolt connections by using incomplete modal
parameters. They combined a system mode-based method [202] and a dynamic model reduction method [203] to avoid us-
ing the complete mode shapes. The joint posterior PDF of the model and modal parameters was approximated via Gaussian
distribution.
In recent years, Sparse Bayesian Learning(SBL), as a supervised learning framework [204-206], has received great atten-
tion as a means of deriving sparse solutions in the context of regression and classification [207-209]. The SBL has some
similarities to sparse recovery theory that uses lp (0 ≤ p≤ 1) regularisation technique. The prior distribution in SBL can
induce sparsity in inferred parameters, which functions as the regularisation term in sparse recovery. In SBL, a parame-
terised prior, that is, the ARD prior, is adopted instead of a fixed prior to incorporate a preference for sparse parameters. An
individual hyper-parameter is assigned to each unknown parameter, resulting in sparse solutions [204,206].
SBL has several significant advantages over the deterministic sparse recovery, for example, the l1 or l0 regularisation tech-
niques and orthogonal matching pursuit. The SBL closely resembles the l0 regularisation, which typically results in a sparser
solution with higher accuracy than the l1 regularisation. In SBL, the global minimum is achieved at the maximally sparse
solution, which is a desirable property of l0 regularisation [205]. When the sensing matrix does not satisfy the incoherence
criteria, the performance of most existing CS algorithms will degrade, whilst the SBL still retains the excellent ability for
sparse recovery [210]. The SBL technique is more general and more flexible than the sparse recovery theory. The latter using
the regularisation techniques disregards the relative uncertainties between different variables and requires estimation of the
regularisation parameter. However, the hyper-parameters in SBL possess a clear physical meaning that represents the preci-
sion of the uncertainties. The hyper-parameters can be updated automatically, thereby avoiding the tricky selection of the
regularisation parameter in sparse recovery.
Although the Bayesian probabilistic approach has been introduced and applied to structural damage identification for
nearly two decades, SBL has not been utilised and explored for structural damage detection until recently. One main reason
is that the modal data are a nonlinear function of the structural damage parameters. Consequently, the integral in the
evidence of the Bayesian equation cannot be calculated directly. Analytical and numerical approaches have been developed
to tackle this difficulty. The former includes hierarchical modelling and asymptotic techniques (e.g. Laplace’s approximation),
whilst the latter includes expectation–maximization (EM) technique and sampling techniques.
Bayesian hierarchical modelling is a model written in a hierarchical form, which is particularly useful in dealing with
complex and nonlinear problems. Huang and Beck [211] developed a hierarchical SBL method by expanding the nonlinear
eigenvalue problem as multiple linear regression functions. The simulated damage was successfully detected with improved
accuracy compared with the Bayesian updating method in Yuen et al. [202]. Later, Huang et al. [212] improved the SBL
algorithm by eliminating two approximations in the theoretical formulation. The efficiency of the improved algorithm was
higher and its performance was better than that of the previous SBL method for a real structure with a significant modelling
error.
Multi-task learning is a useful tool to exploit data redundancy between different groups of measurements. To improve
the reliability of damage localisation, Huang et al. [213] used a multi-task SBL to fuse the respective strengths of two FD-
based damage indices. The linear regression models were employed to model the relationship between a damage localisation
vector and two damage indices, which were then incorporated in the likelihood function. Huang et al. [214] simultaneously
utilised multiple groups of measurements and proposed two hierarchical Bayesian models for multi-task SBL. In these mod-
els, an ARD prior was assigned across multiple tasks to characterise the shared sparseness profile. Unlike the previous multi-
task SBL algorithm [215], the prediction error precision parameters were marginalised from hierarchical models to improve
the learning robustness and characterise the posterior uncertainty.
Hou et al. [216] proposed an EM-based SBL method for damage detection. An iterative EM technique was employed to
tackle a nonlinear eigenvalue problem without performing asymptotic approximation or stochastic simulation. Wang et al.
[217] extended SBL via Laplace approximation, in which a complicated integral in the evidence was approximated as a
Gaussian PDF. Consequently, damage parameters and hyper-parameters were derived in an analytical form and iteratively
solved without sampling.
On the basis of a similar hierarchical SBL model, Huang et al. [218] proposed two Gibbs sampling (GS) algorithms to
sample the posterior PDF of uncertain parameters and provide a full treatment of the posterior uncertainty. Laplace’s ap-
proximation was used to estimate hyper-parameters. Later, they [219] developed a full GS method to characterise the pos-
terior uncertainty of hyper-parameters. The proposed partial and full GS algorithms were applied to the IASC-ASCE Phase
18
Table 8
Bayesian methods.
Authors Methods Features Bayesian inference Damage Application

detection
level
Figueiredo et al. MSD Frequencies Markov-chain 1 Z-24 Bridge

[195] Monte Carlo
Arangio and Beck NN+ARD Acceleration under Laplace’s 1-3 Numerical long-span
[196] ambient approximation suspension bridge
excitation
Arangio and NN+ARD Acceleration under Laplace’s 1 Benchmark
Bontempi ambient approximation cable-stayed bridge
[197] excitation
Lam et al. [198] Bayesian model Frequencies and Laplace’s 1-3 Experimental
class selection mode shapes approximation ballasted track
Behmanesh and Frequencies and Metropolis– 1-3 Numerical footbridge
Moaveni [199] mode shapes Hastings
algorithm
Behmanesh et al. Bayesian model Frequencies and Markov-chain 1-3 Numerical steel
[200] class selection mode shapes Monte Carlo frame
and averaging
Yin et al. [201] Frequencies and Laplace’s 1-3 Experimental bolted
mode shapes approximation frame
Huang and Beck SBL Frequencies and Hierarchical 1-3 Numerical
[211] mode shapes modelling+ shear-building,
Laplace’s Phase II IASC-ASCE
approximation benchmark
problem
Huang et al. SBL Frequencies and Hierarchical 1-3 Phase II IASC-ASCE
[212] mode shapes modelling+ benchmark
Laplace’s problem
approximation
Huang et al. Multi-task SBL FD-based damage Bayesian linear 1, 2 Experimental steel
[213] indices regression beam, a real
cable-stayed bridge
Huang et al. Multi-task SBL Frequencies and Hierarchical 1-3 Phase II IASC-ASCE
[214] mode shapes modelling+ problem
Laplace’s
approximation
Hou et al. [216] SBL Frequencies and EM 1-3 Experimental
mode shapes cantilever beam
Wang et al. [217] SBL Frequencies and Laplace’s 1-3 Experimental steel
mode shapes approximation frame
[218] mode shapes modelling+ benchmark
partial problem
GS+Laplace’s
approximation
[219] mode shapes modelling+ full benchmark
GS problem
II benchmark problem. The full GS algorithm was verified to be more reliable than the partial one for real experimental
studies.
Table 8 compares and summarizes the Bayesian methods that have been reviewed in this subsection.
2.8. Varying temperature conditions
Structural responses vary under the changing operational and environmental conditions, particularly temperature. Tem-
perature variations influence Young’s modulus of most construction materials [220] and boundary conditions [221-223].
Thus, such variations cause changes in structural dynamic properties [224,225]. Xia et al. [224] quantified the effect of tem-
perature on variations in frequencies, mode shapes and damping through a series of experiments on a continuous concrete
slab for nearly 2 years. Some studies have found that changes in structural responses due to temperature variations could
be more significant than those due to a medium degree of structural damage [4] or under normal operational loads [226].
Consequently, if the temperature effects are not fully understood, then false structural condition identification may occur.
Some techniques [227-229] have been developed to reduce the effects of temperature on damage detection or to detect
damage under different temperature conditions.
19
The approaches that consider the effects of temperature on damage identification can be divided into two categories
depending on whether or not environmental variables are measured.
2.8.1. Using temperature measurement

When temperature data are available, the correlation between temperature variables and damage features can be estab-
lished, and the effects of temperature can be removed from damage features.
Deng et al. [230] developed a six-order polynomial regression model to describe the correlations of frequency–
temperature and displacement–temperature by using long-term monitoring data. They then classified the measured changes
in the structural responses caused by damage and environmental variations by using a control chart. Bao et al. [231] in-
vestigated the relation between modal properties and temperature and applied the Dempster–Shafer data fusion technique
[232] to identify damage under varying temperature conditions. The accuracy of damage identification results was increased
by incorporating temperature variations. Magalhaes et al. [233] studied the time evolution of the modal parameters of an
arch bridge for 2 years. The regression analysis complemented with PCA was conducted to eliminate the effects of environ-
mental and operational factors on natural frequencies. The existence of damage was then successfully detected using the
control chart.
In practice, correlations between temperature and damage features are complicated and may not be well established
through regression analysis. In this regard, Zhou et al. [225,234] proposed a BPNN-based approach to eliminate the effects
of temperature. A total of 770h monitoring data of natural frequencies and temperatures obtained from the Ting Kau Bridge
in Hong Kong and their FE model was used to train and test a BPNN model. Case studies indicated that the approach could
detect the occurrence of damage when the change in damage-induced frequency was as small as 1%. Considering the high-
computational load of the BPNN, Jin et al. [235] presented NN trained with EKF to detect damage to a highway bridge under
severe temperature changes. The correlation analysis between natural frequencies and temperature was conducted on the
basis of 1-year monitoring data. The convergence of the proposed method was faster and its results were more accurate
than those of traditional BPNN. Numerical results also showed that the proposed method was superior to the multi-linear
regression approach.
Methods using temperature data have some practical issues. For example, damage features should be extracted under
a wide range of temperature condition. The shortage of temperature data may affect the accuracy of structural damage
detection.
2.8.2. Without temperature measurement

A number of methods have been developed to alleviate the need for direct measurement of temperature variations by
using the measured response data only under varying environmental conditions. In these methods, temperature variations
are treated as embedded variables. These methods are typically based on the assumption that variations in structural vibra-
tion characteristics due to damage behave differently from those due to varying temperature conditions [227]. In this case,
ML algorithms and statistical pattern recognition techniques are typically used to derive a robust damage index for outlier
analysis.
Kullaa [236] proposed an approach to distinguish three sources of variability, namely, environmental or operational ef-
fects, sensor faults and structural damage. Structural damage was global, and sensor faults were local. Accordingly, sensor
fault and structural damage were distinguished through sensor isolation. Measurement data under different environmental
or operational conditions were included in the training data to consider environmental or operational effects. Damage or
sensor fault localisation was detected on the basis of the maximum likelihood ratio.
Shokrani et al. [237] presented a PCA-based approach for damage localisation under varying environmental conditions.
During the training stage, the MSC data matrix was formulated across a representative operational period and the statis-
tical characteristics of the operational variations on a curvature were then extracted via PCA transformation. The residual
of MSC matrices between inspected and baseline structural states was used as the damage index based on a hypothesis
test. Numerical studies indicated that the proposed method was effective in the case of linear or weakly nonlinear situation.
Kostić et al. [238] integrated a sensor-clustering-based time-series analysis method with ANNs to compensate for the effects
of temperature. A sensor-clustering-based ARX method was applied to the free vibration acceleration data to calculate the
damage features. Multilayer ANNs were then trained using the obtained damage features resulting from different temper-
ature scenarios. Differences between the damage features from the time series and ANN analyses were used for damage
detection. Numerical results demonstrated that the proposed method could successfully determine the existence, location
and relative severity of damage under varying temperature conditions. Fallahian et al. [239] combined couple sparse coding
and deep NNs to assess damage by considering uncertainties, such as noise and temperature. The simulated FRF data was
first generated from a numerical model, and PCA was applied to decrease the dimension of FRF data and extract the fea-
tures. The couple sparse coding and deep NN were then individually trained. The outputs were combined with the weight
majority voting method to make a better decision about the healthy state of the structure.
Liang et al. [240] proposed a novel frequency-based technique to eliminate the interference of varying environmental
conditions. The non-stationary frequencies sensitive to environmental variation were transformed to a stationary sequence
by using the co-integration algorithm. The co-integration residual was then employed as the damage feature, which would
display a noticeable jump once damage occurred. Erazo et al. [241] used a Kalman filter to decouple structural damage
20
from temperature variations. According to the properties of the filtering residual, the residual spectral density was approxi-
mately constant under global changes caused by environmental variations and significantly affected by local changes caused
by structural damage. Therefore, damage location and severity were defined on the basis of the spectral moments of the
residual spectral density.
Most previous studies have assumed that environmental effects are linear or piecewise linear [242]. However, the rela-
tionship between damage features and unknown environmental factors may be nonlinear in practice. Nonlinear analysis can
be conducted in several ways, including auto-associative NN (AANN), kernel PCA (KPCA) and principal curves.
KPCA is a nonlinear PCA. It nonlinearly maps the input variables into a high-dimensional linear space through incor-
porating the kernel functions, where normal PCA can be conducted [243]. Hsu and Loh [244] conducted KPCA by using
AANN instead of SVM to locate and quantify structural damage. The element stiffness identified from natural frequencies
and mode shapes was used as features to conduct nonlinear PCA. Nguyen et al. [245] successfully applied KPCA to identify
damage to a bridge by considering temperature and soil variations. A global nonlinear model that described the relationship
of damage-sensitive features with variations in environmental conditions was developed using Gaussian KPCA.
KPCA requires the specification of two parameters, namely, the kernel width and the number of extracted principle
components, which may have a profound effect on the algorithm performance. Reynders et al. [242] proposed an improved
KPCA to consider the nonlinear environmental and operational effects, in which these two parameters were automatically
determined. Experimental results showed that the proposed method successfully identified damage to a three-span pre-
stressed concrete bridge, whereas linear PCA could not.
In the algorithms of the second category (e.g. environmental variables are not measured), all sources of environmental
variability should be efficiently characterised by the training data sets. Otherwise, these algorithms may be unreliable when
they are applied to new data corresponding to the environmental conditions excluded in the training phase. In addition, if
changes in the structural dynamic characteristics due to damage are analogous to those due to varying environmental con-
ditions, then the effectiveness of these algorithms cannot be guaranteed. In this circumstance, the first-category algorithms
are suggested.
The aforementioned approaches considering the effects of temperature are summarized in Table 9.
2.9. Nonlinear methods
Most existing vibration-based damage identification methods assume that a structure behaves linearly before and after
damage. However, in practice, structures may behave nonlinearly in the undamaged state because of complex joints and
interfaces [246]. In these situations, fitting a linear model to the data measured from an intrinsically nonlinear structural
system results in the biased estimation of parameters. This circumstance may result in false damage identification. Therefore,
structural nonlinearity should be appropriately considered for accurate and reliable damage identification.
The breathing crack or delamination is a typical type of damage that may induce nonlinearity to a structure [247]. The
Volterra series may be an effective method of describing nonlinear behaviours. As a generalization of the linear convolution,
the Volterra series can separate a system response into linear and nonlinear components [248]. Chatterjee [249] analysed
the nonlinear response of a cantilever beam with a breathing crack by using the Volterra series. They also developed a
nonlinear dynamic model by utilising high-order FRFs and estimated crack severity based on the first and second harmonic
amplitudes. Andreaus and Baragatti [250] exploited the nonlinear features of harmonic forced vibration and determined the
crack location and depth by analysing sub- and super-harmonic components in the Fourier spectra and phase portrait distor-
tions. Later, the proposed method was experimentally verified using a steel cantilever beam [251]. Peng et al. [252] applied
a nonlinear ARMAX to establish a nonlinear ARX model, from which nonlinear output FRFs and an associated index were
determined. Structural damage was then identified by comparing the nonlinear output FRF indices of the inspected structure
before and after damage.
Another common type of nonlinearity in structural systems is the yielding of steel members during an earthquake [246].
Chanpheng et al. [253] proposed a nonlinear feature, that is, the degree of nonlinearity, for damage detection due to earth-
quakes. The degree of nonlinearity was calculated from the data of the ground motion and structural vibration based on the
Hilbert transform, which indicated whether damage occurred. Wang et al. [254] proposed an analytical mode decomposition
method in combination with the Hilbert transform for structural nonlinearity quantification and damage detection during
earthquakes. The measured structural responses were decomposed into IMFs by using the proposed decomposition method,
and the instantaneous frequencies were extracted using the Hilbert transform. The instantaneous frequency was integrated
over time duration to eliminate the effects due to nonlinearity. On this basis, the degree of nonlinearity index was defined
to represent damage severity.
Since many structures behave nonlinearly even in an undamaged state, previous linear methods may not be suitable to
handle these initially nonlinear systems. Bornn et al. [255] applied an AR-SVM approach to time-series data for detecting
damage to an initially nonlinear system, which was experimentally simulated with a column being suspended from the
top floor and a bumper being placed on the second floor. Shiki et al. [256] used a discrete Volterra model to represent
the behaviour of a magneto-elastic beam with nonlinearity even in a reference state. Input and output data were utilised
to estimate the Volterra kernels. The prediction error of the model, together with hypothesis testing, was used to detect
damage to the system during the linear and nonlinear regime of motion. Experimental studies showed that the nonlinear
index could detect structural changes in both regimes of motion, while the linear one failed during the nonlinear regime.
21
Table 9
Damage identification methods considering the effect of temperature.
Authors Methods Features Relation Damage Applications

detection
level
Using the measurement of temperature

Deng et al. [230] Control chart, Frequencies and Six-order 1 Numerical
online displacement polynomial suspension
regression model bridge
Bao et al. [232] Dempster–Shafer Frequencies and mode Linear regression 1 and 2 Experimental steel
data fusion shapes model frame
Magalhaes et al. Control chart, Frequencies Linear regression 1 Numerical concrete
[234] online model arch bridge
Zhou et al. [225] AANN Frequencies BPNN correlation 1 Numerical
model cable-stayed
bridge
Jin et al. [236] NN+EKF Frequencies - 1 Numerical
composite bridge
Without direct measurement of temperature
Kullaa [237] Likelihood ratio Accelerations 1 Experimental
test+ control bridge
chart
Shokrani et al. PCA MSC 1 and 2 Numerical
[238] spring–mass
chain and bridge
Kostic et al. ARX+ANN Accelerations 1 and 2∗ Numerical
[239] footbridge
Fallahian et al. PCA+ deep NN+ FRF 1-3 I-40 bridge
[240] couple sparse
coding
Liang et al. [241] Co-integration Frequencies 1 Real steel truss
algorithm bridge
Erazo et al. [242] Kalman filter+ PSD residual 1-3 Numerical bridge
Bayesian
whiteness test
Hsu and Loh KPCA+AANN Frequencies, mode shapes 1-3 Numerical bridge
[245]
Nguyen et al. KPCA Frequencies 1 Real bridge
[246]
Reynders et al. Improved KPCA Frequencies 1 Z-24 bridge
[245]
Note: ∗ :Quantitatively indicate the relative severity of damage
Villani et al. [257] presented a stochastic version of the Volterra series to describe the nonlinear behaviour to consider
uncertainties. The presence of damage was detected on the basis of the MSD. Numerical results showed that the perfor-
mance of the present method was better than that of linear analysis and could detect small cracks even in the presence of
uncertainties.
Recently, nonlinear FE model updating has been developed to identify structural nonlinear parameters. Ebrahimian et al.
[258] and Ebrahimian et al. [259] used an EKF and a batch of Bayesian approaches, respectively, to estimate time-invariant
parameters of the nonlinear FE models of frame-type structures by using input excitation and dynamic response of a struc-
ture during earthquake events.
However, input excitations to a structure are difficult to be measured accurately in practice. In this regard, Ebrahimian
et al. [260] proposed an output-only nonlinear FE model updating method by using the measured time history responses
during an earthquake event. The proposed methodology simultaneously estimated the unknown FE model parameters and
input excitations based on a sequential maximum likelihood estimation approach and a sequential maximum a posteriori
estimation approach, respectively. Astroza et al. [261] used an unscented Kalman filter to solve the nonlinear state-space
equation and circumvent the computation of FE response sensitivities with respect to model parameters and input excita-
tions.
Although a number of methods have been developed for nonlinear damage identification, a general model that can char-
acterise the nonlinear behaviour of a structure is still challenging to construct. For example, when high levels of excitation
are applied, structural responses may exhibit high-order nonlinearities that are inaccurately approximated using a small
number of terms in the Volterra expansion. Nonlinear analysis is computationally intensive and time consuming. Conse-
quently, a nonlinear FE model updating approach is difficult to be applied to practical structures that contain large DOFs.
The substructuring approach may be integrated to solve the challenge. Table 10 compares and summarises the nonlinear
methods that have been reviewed in this subsection.
22
Table 10
Nonlinear methods.
Authors Methods Features Damage type and Applications

identification level
Chatterjee [249] Volterra series First and second response Presence and size of Numerical
harmonic amplitudes breathing crack of spring-mass-damper
cantilever beam system
Andreaus and Baragatti - Phase portrait distortions, Location and depth of Numerical cantilever beam
[250] sub- and super-harmonic breathing crack of
components and curved cantilever beam
shape of the modal line
Andreaus and Baragatti - Phase portrait distortions Presence and size of Experimental steel
[251] and sub- and breathing crack of cantilever beam
super-harmonic cantilever beam
components
Peng et al. [252] Nonlinear ARMAX FRFs Presence and degree of Experimental aluminium
structural nonlinearity plates
Chanpheng et al. [253] FRFs Degree of nonlinearity Simulation data
Wang et al. [254] Analytical mode Instantaneous frequency Nonlinearity quantification Experimental building
decomposition+ Hilbert
transform
Bornn et al. [255] AR-SVM Residual error of AR-SVM Presence of damage for Experimental aluminum
model initially nonlinear plates and columns
systems connected with bolted
joints
Shiki et al. [256] Discrete-time Volterra Prediction error of Presence of damage for Experimental aluminum
series Volterra model initially nonlinear beam
systems
Villani et al. [257] Stochastic Volterra series+ Volterra kernel coefficients Presence of damage for Experimental aluminum
MSD and contribution initially nonlinear beam
systems
Ebrahimian et al. [258] EKF Input-output time Levels 1-3 Numerical steel bridge
histories data under column and 2D moment
earthquake excitation resisting steel frame
Ebrahimian et al. [259] Batch Bayesian approach Input-output time Levels 1-3 Numerical steel bridge
histories data under column and 2D moment
earthquake excitation resisting steel frame
Ebrahimian et al. [260] Bayesian inference Output acceleration under Levels 1-3 Numerical 3D RC building
earthquake excitation
Astroza et al. [261] Unscented Kalman filter Output acceleration under Levels 1-3 Numerical 3D RC frame
earthquake excitation
2.10. Other methods
In addition to the methods introduced earlier, some other damage identification techniques are reviewed here.
Drive-by (or indirect) damage identification methods have been proposed for bridge structures more than a decade. These
methods extract the dynamic properties of the bridge, such as natural frequencies, from the measured responses of a passing
vehicle instrumented with sensors [262,263]. As a vehicle passing over a bridge, the moving vehicle can be considered as
both exciter and receiver due to the bridge-vehicle interaction [264]. The major advantage of the drive-by approach is that
it uses sensors installed on the vehicle only and thus minimises the effect on the normal traffic of the bridge. It is more
economical, efficient, and mobile compared with conventional methods. In addition, the vehicle passing over the bridge as
a “moving sensor” can provide higher spatial resolution than fixed sensors [264].
Siringoringo and Fujino [265] estimated the first natural frequency of a bridge using the vehicle response as the driving
velocity was below 30 km/h. However, the vehicle frequency usually appears as a dominant peak in the spectrum, making
the identification of the bridge frequency challenging. In this connection, Yang et al. [266] combined the band-pass filter and
singular spectrum analysis to filter out the vehicle frequency from the spectrum and thus enhance the visibility of bridge
frequencies. Zhang et al. [267] approximately extracted structural mode shape squares from the acceleration of a passing
vehicle with tapping devices. However, acceptable damage identification accuracy was achieved only for the vehicle speed
as low as 2 m/s. Oshima et al. [268] estimated bridge mode shapes from the dynamic responses of moving vehicles based
on SVD. The numerical study showed that damage could be detected accurately for vehicle speeds varying from 5 to 15
m/s. Nguyen and Tran [269] applied the WT to the displacement response of a moving vehicle with low speed to determine
the existence and location of cracks for beam-like structures. Obrien et al. [270] used EMD to decompose the response
measured in a passing vehicle into three components. The damage location was detected using the IMFs corresponding to
the pseudo-frequency component.
The vehicle speeds for the aforementioned methods are all lower than the highway speed range, which thus may require
temporary bridge and/or lane closures [264]. In 2004, traffic speed deflectometer (TSD) was developed for pavement deflec-
23
tion measurementsat speeds of up to 80km/h. OBrien and Keenahan [271] used a TSD vehicle containing two displacement
sensors for damage detection for short-medium span bridges. Keenahan and Obrien [272] employed a TSD vehicle containing
five displacement sensors. The time-shifted curvature derived from the displacements was selected as the damage indicator.
Numerical studies showed that the proposed method was more robust to noise than the algorithm proposed in OBrien and
Keenahan [271].
Although the drive-by damage detection methods possess some distinct advantages over the conventional direct meth-
ods, they are primarily limited to Levels 1 and 2 damage identification. Moreover, most of these methods require the vehi-
cles”speed to be slow, which may cause traffic congestion and disruption. Some researchers used the TSD devices to address
this problem, while the high cost of the equipment hinders their wide applications.
Response surface methodology (RSM), which is a combination of mathematical and statistical techniques, can provide an
approximate mathematical model mapping the input parameters of a physical system to its output responses [273]. Linear
or second-order polynomial models are typically employed as the fundamental structures of RS models to describe a studied
system. In comparison with conventional FE model updating methods, RSM is a good alternative to solve inverse problems
without using a sensitivity matrix. This method also requires low modelling and updating efforts.
Fang and Perera [274] used D-optimal designs to establish RS models for screening out non-significant updating pa-
rameters, which required few samples for the desired RS modelling. Damage was identified by minimising the discrepancy
between reference-state RS and experimental models.
Kim et al. [275] combined global vibration-based and local impedance-based approaches to distinguish two typical dam-
age types, namely, girder damage and tender damage, for pre-stressed concrete girder bridges. First, damage occurrence was
detected on the basis of changes in frequency responses. Electromechanical impedance was then used to monitor whether
tendon damage occurred. The location and extent of damage were estimated using natural frequencies or mode shapes
based on damage type.
Yang et al. [276] proposed a 1D generalised local entropy method to detect cracks in beam structures. Yang et al.
[277] further developed 2D generalised local entropy combined with statistical analysis and AI to identify damage in plate-
like structures. They then evaluated the damage severities at the identified locations by using the ABC algorithm with the
objective function defined by the combination of natural frequencies.
Considering the sparsity of structural damage, some researchers developed CS-based damage identification methods. Yang
and Nagarajaiah [129] combined blind feature extraction and sparse representation classification to locate and quantify struc-
tural damage. They extracted the modal features of structures by using the unsupervised complexity pursuit algorithm and
expressed the test modal feature as a linear combination of the bases of the overcomplete reference feature dictionary. The
resulting highly underdetermined linear system was correctly solved via the l1 minimisation. Wang and Hao [278] proposed
a CS-based pattern classification algorithm by constructing the feature matrix based on the sparse representation of nu-
merically simulated time domain data. The existence, location and extent of damage were then determined sequentially by
solving the l1 optimisation problem.
The number of available sensors is always limited in practice due to the economic and technological considerations. Given
a total number of sensors, the sensor locations should be determined such that the measured data contain useful features
of the structure as much as possible. Consequently structural damage can be detected accurately. Although a number of OSP
techniques have been developed for decades, most of these techniques are devoted to modal identification. OSP methods for
damage identification are still far from the end.
Zhou et al. [279] introduced a sensor placement index in terms of the ratio of two parameters, namely, the contribution
of the measurementpoints to the Fisher information matrix and the damage sensitivity to the measurement noise. Li et al.
[280] proposed a two-phase OSP scheme based on the Fisher information matrix. The first phase was to find out the sensor
locations that reconstructed accurate responses. In the second phase, the optimal sensor locations were determined based on
the sensitivity analysis with respect to the elemental stiffness parameter. Lin et al. [281] employed two objective functions
for multi-type sensor placement based on the simultaneous optimization of the response covariance sensitivity and the
response independence. Later, they [282] applied the multi-type OSP method for damage detection in a nine-bay 3D frame
structure.
The aforementioned sensitivity-based OSP methods calculate the sensitivity analysis with respect to the model pa-
rameters of each structural element, which are not applicable for large-scale structures. Guo et al. [283] developed an
information-entropy-based OSP method targeting damage detection of large-scale bridges subject to ship collision. The sen-
sor configuration was optimized by a multi-objective optimization algorithm, which simultaneously minimized the infor-
mation entropy index for each possible ship-bridge collision scenario. The proposed method was applicable in practice to
determine the OSP prior to field testing. Beygzadeh et al. [284] proposed an improved GA algorithm for OSP in space struc-
tures damage detection. A numerical study showed that the proposed algorithm performed better than the GA in terms of
convergence speed and damage detection accuracy.
2.11. Comparative studies
Over the past 10 years, some researchers have compared existing vibration-based damage detection methods in terms of
different aspects. Different techniques may behave differently on different types of structures. Comparative studies may give
insight on the performance, applicability and feasibility of each technique.
24
Kopsaftopoulos and Fassois [285] experimentally compared several statistical time series methods via application to
a lightweight aluminium truss structure. They assessed two nonparametric methods (i.e. PSD- and FRF-based methods)
and four parametric methods (e.g. a model parameter, a residual variance, a residual likelihood function and a residual
uncorrelatedness-based method). All the methods were effective for various damage scenarios, but the accuracy of paramet-
ric methods was higher than that of nonparametric methods.
Talebinejad et al. [286] evaluated four mode shape-based techniques to identify damage to long-span cable-stayed bridges
by using simulated acceleration data. These methods included an enhanced coordinate MAC, a damage index, MSC, and
modal flexibility methods. The performance of the damage index and MSC was better than that of the two other methods.
When measurement noise was considered, only high-intensity damage could be detected using the damage index and MSC,
and none of them could identify multiple damage to the deck.
Saeed et al. [287] compared the performance of single and multiple ANNs and multiple adaptive neuro-fuzzy inference
systems in detecting the location and length of a crack in curvilinear beams. Natural frequencies and FRFs were employed as
inputs, and PCA was conducted to reduce the size of FRF data. Multiple ANN models produced the lowest average prediction
errors. Multiple adaptive neuro-fuzzy inference systems were less sensitive to noise than other classifier models.
Dessi and Camerlengo [288] compared nine damage identification methods based on natural frequencies, MSCs and MSE.
The selected techniques were divided into two categories: one required baseline data, whilst the other did not need any
reference to the undamaged state. Their performances were evaluated on the basis of the same test of a simply supported
Euler–Bernoulli beam. They found that a damage index performed well in identifying the damage location may be not so
accurate in estimating the damage severity and vice versa. The position of damage along the beam affected the accuracy of
damage identification for all methods.
Figueiredo et al. [289] comparatively studied four ML algorithms, including AANN, factor analysis, MSD and SVD, for dam-
age detection under operational and environmental variabilities. These algorithms relied only on the measured time-series
responses and did not require a direct measurement of parameters related to operational and environmental variations. An
MSD-based method was proven to be an optimal algorithm in terms of classification performance and computational com-
plexity. Santos et al. [290] presented four kernel-based algorithms, i.e. one-class SVM, support vector data description, KPCA
and greedy KPCA. They compared the performance of these algorithms by using a benchmark three-story frame structure.
Experimental results showed that the proposed kernel-based algorithms had an optimal performance, especially when a
nonlinear temperature–stiffness relationship was present.
3. Challenges and future research
Numerous vibration-based damage detection methods have been developed over the past decades. However, their ap-
plications in practical civil engineering structures are still immature. Some challenging issues deserve further research to
achieve accurate and practical damage identification.
(1) Although some researchers comparatively studied different damage identification methods, no general consensus has
been achieved regarding which type of data is a good damage indicator and which identification method is most ef-
fective. ASCE organized a benchmark study on damage identification of a lab frame 20 years ago [291,292]. Benchmark
studies on real structures are necessary.
(2) For civil engineering structures, nonlinearity often exists and may be induced by connections or the presence of dam-
age. However, few existing algorithms consider nonlinear structural behaviour in the reference and damaged states.
Nonlinear damage identification methods are preferable. Given the heavy computation load, nonlinear analysis can be
combined with substructuring methods to reduce computational effort and improve efficiency when it is applied to
large-scale structures.
(3) Structural damage detection inevitably entails uncertainties, such as measurement noise and modelling errors, due
to inaccurate physical parameters, non-ideal boundary conditions and structural nonlinear properties. The existence
of uncertainties may cause large variations in structural vibration characteristics, which may lead to incorrect dam-
age identification. Therefore, statistical damage identification methods that appropriately consider these uncertainties
should be continually developed. Moreover, noise quantification and elimination need to be explored further.
(4) Operational and environmental variations, particularly temperature, cause changes in structural vibration properties.
Many field studies have found that varying temperatures may cause more significant changes in the structural vibra-
tion properties than those due to damage. Although researchers have proposed various techniques to consider the
effects of temperature on damage detection, no consensus and generally effective methods have been developed.
(5) Although vibration-based damage identification methods have been successfully applied to mechanical and aerospace
structures, the applications of these methods to practical civil structures are far from maturity due to the complex-
ity and uniqueness of civil structures. Civil engineering structures are typically different because of the variability in
the materials and construction processes, the uniqueness of the soil conditions, and environmental conditions. Conse-
quently, it is difficult to extend a well-established method to other structures.
(6) In practice, a large volume of data is generated from an SHM system. The present data processing and damage identi-
fication are manually conducted by practitioners, thereby causing inefficiency or ineffectiveness or false identification
25
results. Automatic data processing and condition assessment are lacking. Rapid developing big data and AI techniques
may be a direction to achieve this purpose. Many interesting studies in the topic are underway.
(7) In recent years, ML methods have received considerable attention for structural damage identification. However, most
of the studies use laboratory or numerical data for training as real data in damaged structures are very rare. Damage
identification problems have poor data conditions when they are casted into a supervised learning framework. The
introduction of semi-supervised and unsupervised algorithms is expected in the future.
(8) Structural damage detection is essentially a multidiscipline area, involving sensor technology, data processing, mathe-
matics, and structural analysis. A successful damage identification method needs close collaboration among different
disciplines.
4. Conclusion
The vibration-based damage detection methods between 2010 and 2019 have been reviewed. The applicability and ef-
fectiveness of different available techniques depend on damage type, structural configuration and available data. However,
a universal methodology that can identify all damage types of different structures has yet to be developed. In addition,
few existing algorithms can predict the remaining service life of structures, which is regarded as Level 4 structural damage
detection.
In summary, there is a pressing need to develop more accurate and reliable damage identification methods for practical
civil engineering structures by using vibration measurement data. The prognosis of damage to a monitored structure also
requires extensive investigations.
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
This research was supported by the PolyU Project of Strategic Importance (Project No. 1-ZE1F), the FCE Postdoctoral
Fellowship Scheme (Project No. 1-ZVP2), NSFC Joint Research Fund for Overseas and Hong Kong and Macao Scholars (Project
No. 51629801) and the Research Grants Council-General Research Fund (Project No. 15201920).
References
[1] A. Rytter, Vibration based inspection of civil engineering structures, PhD Thesis, Aalborg University, Denmark, 1993.
[2] S.W. Doebling, C.R. Farrar, M.B. Prime, D.W. Shevitz, Damage identification and health monitoring of structural and mechanical systems from changes
in their vibration characteristics: a literature review, Los Alamos National Laboratory Report LA-13070-MS (1996).
[3] H. Sohn, C.R. Farrar, F.M. Hemez, D.D. Shunk, D.W. Stinemates, B.R. Nadler, A review of structural health monitoring literature: 1996–2001, Los Alamos
National Laboratory Report (2003).
[4] O.S. Salawu, Detection of structural damage through changes in frequency: a review, Eng. Struct. 19 (9) (1997) 718–723.
[5] E.P. Carden, P. Fanning, Vibration based condition monitoring: a review, Struct. Health Monit. 3 (4) (2004) 355–377.
[6] W. Fan, P.Z. Qiao, Vibration-based damage identification methods: a review and comparative study, Struct. Health Monit. 10 (1) (2011) 83–111.
[7] S.J.S. Hakim, H.A. Razak, Modal parameters based structural damage detection using artificial neural networks - a review, Smart Struct. Syst. 14 (2)
(2014) 159–189.
[8] B. Chen, S.L. Zhao, P.Y. Li, Application of Hilbert-Huang transform in structural health monitoring: a state-of-the-art review, Math. Probl. Eng. (2014)
317954.
[9] X. Kong, C.S. Cai, J.X. Hu, The state-of-the-art on framework of vibration-based structural damage identification for decision making, Appl. Sci. 7 (5)
(2017) 497.
[10] M.S. Cao, G.G. Sha, Y.F. Gao, W. Ostachowicz, Structural damage identification using damping: a compendium of uses and features, Smart Mate. Struct.
26 (2017) 043001.
[11] Y. Huang, C. Shao, B. Wu, J.L. Beck, H. Li, State-of-the-art review on Bayesian inference in structural system identification and damage assessment,
Adv. Struct. Eng. 22 (6) (2019) 1329–1351.
[12] S. Weng, H.P. Zhu, Y. Xia, J.J. Li, W. Tian, A review on dynamic substructuring methods for model updating and damage detection of large-scale
structures, Adv. Struct. Eng. (2019) 1–17.
[13] M. Gordan, H.A. Razak, Z. Ismail, K. Ghaedi, Recent developments in damage identification of structures using data mining, Lat. Am. J. Solids Stru. 14
(13) (2017) 2373–2401.
[14] Y.Q. Bao, Z.C. Chen, S.Y. Wei, Y. Xu, Z.Y. Tang, H. Li, The state of the art of data science and engineering in structural health monitoring, Engineering
5 (2019) 234–242.
[15] Y.H. An, E. Chatzi, S.H. Sim, S. Laflamme, B. Blachowski, J.P. Ou, Recent progress and future trends on damage identification methods for bridge
structures, Struct. Control Health Monit. 26 (10) (2019) e2416.
[16] C.R. Farrar, S.W. Doebling, D.A. Nix, Vibration–based structural damage identification, Philos. Trans. R. Soc. A, Phys. Eng. Sci.359 (1778) (2001) 131–149.
[17] M.K. Yoon, D. Heider, J.W. Gillespie Jr., C.P. Ratcliffe, R.M. Crane, Local damage detection with the global fitting method using operating deflection
shape data, J. Nondestruct. Eval. 29 (1) (2010) 25–37.
[18] M.K. Yoon, D. Heider, J.W. Gillespie Jr., C.P. Ratcliffe, R.M. Crane, Local damage detection with the global fitting method using mode shape data in
notched beams, J. Nondestruct. Eval. 28 (2) (2009) 63–74.
[19] Y. Zhang, L.Q. Wang, Z.H. Xiang, Damage detection by mode shape squares extracted from a passing vehicle, J. Sound Vib. 331 (2) (2012) 291–307.
[20] D.M. Feng, M.Q. Feng, Output-only damage detection using vehicle-induced displacement response and mode shape curvature index, Struct. Control
Health Monit. 23 (8) (2016) 1088–1107.
[21] S.H. Sun, H.J. Jung, H.Y. Jung, Damage detection for beam-like structures using the normalized curvature of a uniform load surface, J. Sound Vib. 332
(6) (2013) 1501–1519.
26
[22] J.H. Zhao, L. Zhang, Structural damage identification based on the modal data change, Int. J. Eng. Man. 4 (2012) 59–66.
[23] M. Radzieński, M. Krawczuk, M Palacz, Improvement of damage detection methods based on experimental modal parameters, Mech. Syst. Signal
Process. 25 (6) (2011) 2169–2190.
[24] D. Capecchi, J. Ciambella, A. Pau, F. Vestroni, Damage identification in a parabolic arch by means of natural frequencies, modal shapes and curvatures,
Meccanica 51 (2016) 2847–2859.
[25] S. Caddemi, I. Caliò, The exact explicit dynamic stiffness matrix of multi-cracked Euler–Bernoulli beam and applications to damaged frame structures,
J. Sound Vib. 332 (12) (2013) 3049–3063.
[26] S. Caddemi, I. Caliò, Exact closed-form solution for the vibration modes of the Euler–Bernoulli beam with multiple open cracks, J. Sound Vib. 327 (3)
(2009) 473–489.
[27] N.T. Khiem, H.T. Tran, A procedure for multiple crack identification in beam-like structures from natural vibration mode, J. Sound Vib. 20 (9) (2014)
1417–1427.
[28] N.T. Khiem, L.K. Toan, A novel method for crack detection in beam-like structures by measurements of natural frequencies, J. Sound Vib. 333 (18)
(2014) 4048–4103.
[29] M. Frizzarin, M.Q. Feng, P. Franchetti, S. Soyoz, C. Modena, Damage detection based on damping analysis of ambient vibration data, Struct. Control
Health Monit. 17 (4) (2010) 368–385.
[30] S. Mustafa, Y. Matsumoto, H. Yamaguchi, Vibration- based health monitoring of an existing truss bridge using energy-based damping evaluation, J.
Bridge Eng. 23 (1) (2017) 04017114.
[31] A.M. Ay, S. Khoo, Y. Wang, Probability distribution of decay rate: a statistical time-domain damping parameter for structural damage identification,
Struct. Health Monit. 18 (1) (2019) 66–86.
[32] S. Adhikari, Structural Dynamic Analysis with Generalized Damping Models: Analysis, Wiley-ISTE, London, UK, 2014.
[33] J. Liu, Z.R. Lu, M.L. Yu, Damage identification of non-classically damped shear building by sensitivity analysis of complex modal parameter, J. Sound
Vib. 483 (6) (2019) 457–475.
[34] M.P. Limongelli, Frequency response function interpolation for damage detection under changing environment, Mech. Syst. Signal Process. 24 (8)
(2010) 2898–2913.
[35] M.P. Limongelli, The interpolation damage detection method for frames under seismic excitation, J. Sound Vib. 333 (22) (2011) 5474–5489.
[36] M. Dilena, M.P. Limongelli, A. Morassi, Damage localization in bridges via the FRF interpolation method, Mech. Syst. Signal Process. 52 (2015) 162–180.
[37] B.J. Schwarz, M.H. Richardson, Introduction to operating deflection shapes, CSI Reliability Week, Orlando, FL, 1999.
[38] Y. Zhang, S.T. Lie, Z. Xiang, Damage detection method based on operating deflection shape curvature extracted from dynamic response of a passing
vehicle, Mech. Syst. Signal Process. 35 (1) (2013) 238–254.
[39] N.M.M. Maia, R.A.B. Almeida, A.P.V. Urgueira, Damage detection and quantification using transmissibility, Mech. Syst. Signal Process. 25 (7) (2011)
2475–2483.
[40] J. Li, H. Hao, J.V. Lo, Structural damage identification with power spectral density transmissibility: numerical and experimental studies, Smart Struct.
Syst. 15 (1) (2015) 15–40.
[41] X. Kong, C.S. Cai, B. Kong, Damage detection based on transmissibility of a vehicle and bridge coupled system, J. Eng. Mech. 141 (1) (2015) 04014102.
[42] Y.C. Yang, S. Nagarajaiah, Blind identification of damage in time-varying systems using independent component analysis with wavelet transform,
Mech. Syst. Signal Process. 47 (1) (2014) 3–20.
[43] C.K. Chui, J. Wang, A general framework of compactly supported splines and wavelets, J. Approx. Theory 71 (3) (1992) 54–68.
[44] A. Katunin, The construction of high-order B-spline wavelets and their decomposition relations for faults detection and localization in composite
beams, Scientific Probl. Mach. Op. Maint. 3 (167) (2011) 43–59.
[45] A. Katunin, Damage identification in composite plates using two-dimensional B-spline wavelets, Mech. Syst. Signal Process. 25 (8) (2011) 3153–3167.
[46] M.S. Cao, M. Radzieński, W. Xu, W. Ostachowicz, Identification of multiple damage in beams based on robust curvature mode shapes, Mech. Syst.
Signal Process. 46 (2) (2014) 468–480.
[47] M.S. Cao, W. Xu, W. Ostachowicz, Z.Q. Su, Damage identification for beams in noisy conditions based on Teager energy operator-wavelet transform
modal curvature, J. Sound Vib. 333 (6) (2014) 1543–1553.
[48] V. Shahsavari, L. Chouinard, J. Bastien, Wavelet-based analysis of mode shapes for statistical detection and localization of damage in beams using
likelihood ratio test, Eng. Struct. 132 (2017) 494–507.
[49] N.E. Huang, The Hilbert-Huang Transform in Engineering, Taylor and Francis Group, New York, NY, USA, 2005.
[50] Y.F. Dong, Y.M. Li, L. Ming, Structural damage detection using empirical-mode decomposition and vector autoregressive moving average model, Soil
Dyn. Earthq. Eng. 30 (3) (2010) 133–145.
[51] C.X. Bao, H. Hao, Z.X. Li, Multi-stage identification scheme for detecting damage in structures under ambient excitations, Smart Mate. Struct. 22 (4)
(2013) 045006.
[52] J.P. Han, P.J. Zheng, H.T. Wang, Structural modal parameter identification and damage diagnosis based on Hilbert-Huang transform, Earthq. Eng. Eng.
Vib. 13 (2013) 101–111.
[53] H. Aied, A. Gonzáleza, D. Cantero, Identification of sudden stiffness changes in the acceleration response of a bridge to moving loads using ensemble
empirical mode decomposition, Mech. Syst. Signal Process. 66 (2016) 314–338.
[54] B.B. Mandelbrot, How long is the coast of Britain? Statistical self-similarity and fractional dimension, Science 156 (1967) 636–638.
[55] H. Li, Y. Huang, J.P. Ou, Y.Q. Bao, Fractal dimension-based damage detection method for beams with a uniform cross-section, Comput-Aided Civ. Inf.
26 (3) (2011) 190–206.
[56] R.B. Bai, M.S. Cao, Z.Q. Su, W. Ostachowicz, H. Xu, Fractal dimension analysis of higher-order mode shapes for damage identification of beam struc-
tures, Math. Probl. Eng. (2012) 454568.
[57] M.S. Cao, W. Ostachowicz, R.B. Bai, M. Radzieński, Fractal mechanism for characterizing singularity of mode shape for damage detection, Appl. Phys.
Lett. 103 (2013) 221906.
[58] R.B. Bai, O. Ostachowicz, M.S. Cao, Z. Su, Crack detection in beams in noisy conditions using scale fractal dimension analysis of mode shapes, Smart
Mate. Struct. 23 (2014) 065014.
[59] R.B. Bai, M. Radzieński, M.S. Cao, W. Ostachowicz, Z Su, Non-baseline identification of delamination in plates using wavelet-aided fractal analysis of
two-dimensional mode shapes, J. Intel. Mate. Syst. Struct. 26 (17) (2015) 2338–2350.
[60] Y.H. An, J.P. Ou, Experimental and numerical studies on damage localization of simply supported beams based on curvature difference probability
method of waveform fractal dimension, J. Intel. Mat. Syst. Str. 23 (4) (2012) 415–426.
[61] H. Li, D.W. Tao, Y. Huang, Y.Q. Bao, A data-driven approach for seismic damage detection of shear-type building structures using the fractal dimension
of time–frequency features, Struct. Control Health Monit. 20 (9) (2013) 1191–1210.
[62] D. Hester, A. González, A wavelet-based damage detection algorithm based on bridge acceleration response to a vehicle, Mech. Syst. Signal Process.
28 (2012) 145–166.
[63] N. Roveri, A. Carcaterra, Damage detection in structures under traveling loads by Hilbert–Huang transform, Mech. Syst. Signal Process. 28 (2012)
128–144.
[64] A. Kunwar, R. Jha, M. Whelan, K. Janoyan, Damage detection in an experimental bridge model using Hilbert–Huang transform of transient vibrations,
Struct. Control Health Monit. 20 (1) (2013) 1–15.
[65] M. Friswell, J.E. Mottershead, Finite element model updating in structural dynamics, Kluwer Academic Publishers, Dordrecht, Netherlands, 1995.
[66] M. Baruch, Optimization procedure to correct stiffness and flexibility matrices using vibration tests, AIAA J. 16 (11) (1978) 1208–1210.
[67] A. Berman, E.J. Nagy, Improvement of large analytical model using test data, AIAA J. 21 (8) (1983) 1168–1173.
27
[68] Y.L. Xu, Y. Xia, Structural health monitoring of long-span suspension bridges, Spon Press: London (2012).
[69] J.D. Sipple, M. Sanayei, Finite element model updating using frequency response functions and numerical sensitivities, Struct. Control Health Monit.
21 (2014) 784–802.
[70] B. Moaveni, X.F. He, J.P. Conte, J.I. Restrepo, Damage identification study of a seven-story full-scale building slice tested on the UCSD-NEES shake
table, Struct. Saf. 32 (5) (2010) 347–356.
[71] R. Jafarkhani, S.F. Masri, Finite element model updating using evolutionary strategy for damage detection, Comput-Aided Civ. Inf. 26 (2011) 207–224.
[72] J. Li, S.S. Law, Y. Ding, Substructure damage identification based on response reconstruction in frequency domain and model updating, Eng. Struct.
41 (2012) 270–284.
[73] X.H. Li, M. Kurata, M. Nakashima, Evaluating damage extent of fractured beams in steel moment-resisting frames using dynamic strain responses,
Earthq. Eng. Struct. Dyn. 44 (2015) 563–581.
[74] X.H. Li, M. Kurata, M. Nakashima, Simplified derivation of a damage curve for seismically induced beam fractures in steel moment-resisting frames,
J. Struct. Eng. 142 (6) (2016) 04016019.
[75] X.H. Li, M. Kurata, A. Suzuki, Decoupling algorithm for evaluating multiple beam damages in steel moment-resisting frames, Earthq. Eng. Struct. Dyn.
46 (2017) 1045–1064.
[76] A. Suzuki, M. Kurata, X.H. Li, S. Shimmoto, Residual structural capacity evaluation of steel moment-resisting frames with dynamic-strain-based model
updating method, Earthq. Eng. Struct. Dyn. 46 (2017) 1791–1810.
[77] T.J. Matarazzo, M. Kurata, H. Nishino, A. Suzuki, Postearthquake strength assessment of steel moment-resisting frame with multiple Beam-Column
fractures using local monitoring data, J. Struct. Eng. 144 (2) (2018) 04017217.
[78] M. Sanayei, A. AliKhaloo, M. Gul, F.N. Catbas, Automated finite element model updating of a scale bridge model using measured static and modal
test data, Eng. Struct. 102 (2015) 66–79.
[79] R.R. Craig, M. Bampton, Coupling of substructures for dynamic analyses, AAAA J. 6 (7) (1968) 1313–1319.
[80] J.X. Yu, Y. Xia, W. Lin, X.Q. Zhou, Element-by-element model updating of large-scale structures based on component mode synthesis method, J. Sound
Vib. 362 (2016) 72–84.
[81] T. Wang, H. He, W. Yan, G.P. Chen, A model-updating approach based on the component mode synthesis method and perturbation analysis, J. Sound
Vib. 433 (2018) 349–365.
[82] Y. Liu, H. Sun, D.J. Wang, Updating the finite element model of large-scaled structures using component mode synthesis technique, Intell. Autom.
Soft Comput. 19 (1) (2013) 11–21.
[83] Y. Aoyamai, G. Yagawa, Component mode synthesis for large-scale structural eigenanalysis, Compu. Struct. 79 (2001) 605–615.
[84] C. Papadimitriou, D.C. Papadioti, Component mode synthesis techniques for finite element model updating, Comput. Struct. 126 (2013) 15–28.
[85] S. Weng, Y. Xia, Y.L. Xu, H.P. Zhu, An iterative substructuring approach to the calculation of eigensolution and eigensensitivity, J. Sound Vib. 330 (14)
(2011) 3368–3380.
[86] S. Weng, Y. Xia, Y.L. Xu, H.P. Zhu, Substructure based approach to finite element model updating, Comput. Struct. 89 (9) (2011) 772–782.
[87] A. Simpson, A generalization of Kron’s eigenvalue procedure, J. Sound Vib. 26 (1973) 129–139.
[88] S. Weng, Y. Xia, X.Q. Zhou, Y.L. Xu, H.P. Zhu, Inverse substructure method for model updating of structures, J. Sound Vib. 331 (25) (2012) 5449–5468.
[89] S. Weng, H.P. Zhu, Y. Xia, L. Mao, Damage detection using the eigenparameters decomposition of substructural flexibility matrix, Mech. Syst. Signal
Process. 34 (122) (2013) 19–38.
[90] K.V. Yuen, K. Huang, Identifiability-enhanced Bayesian frequency-domain substructure identification, Comput-Aided Civ. Inf. 33 (2018) 800–812.
[91] K.V. Yuen, L.S. Katafygiotis, Substructure identification and health monitoring using noisy response measurements only, Comput-Aided Civ. Inf. 21 (4)
(2006) 280–291.
[92] J. Wang, Q.S. Yang, Modified Tikhonov regularization in model updating for damage identification, Struct. Eng. Mech. 44 (5) (2012) 585–600.
[93] X.Y. Li, S.S. Law, Adaptive Tikhonov regularization for damage detection based on nonlinear model updating, Mech. Syst. Signal Process. 24 (6) (2010)
1646–1664.
[94] H.P. Zhu, L. Mao, S. Weng, A sensitivity-based structural damage identification method with unknown input excitation using transmissibility concept,
J. Sound Vib. 333 (2014) 7135–7150.
[95] D.L. Phillips, A technique for the numerical solution of certain integral equations of the first kind, J. ACM 9 (1) (1962) 84–97.
[96] V.A. Morozov, Methods for solving incorrectly posed problems, Springer Verlag, New York, 2012.
[97] G.H. Golub, M. Heath, G. Wahba, Generalized cross–validation as a method for choosing a good ridge parameter, Technometrics 21 (2) (1979) 215–223.
[98] I.M. Johnstone, On minimax estimation of a sparse normal mean vector, Ann. Stat. 22 (1) (1994) 271–289.
[99] P.C. Hansen, Analysis of discrete ill-posed problems by means of the L-curve, SIAM Rev. 34 (4) (1992) 561–580.
[100] Y.Z. Fu, Z.R. Lu, J.K. Liu, Damage identification in plates using finite element model updating in time domain, J. Sound Vib. 332 (2013) 7018–7032.
[101] B. Weber, P. Paultre, Damage Identification in a Truss Tower by Regularized Model Updating, J. Struct. Eng. 136 (3) (2010) 307–316.
[102] N. Nguyen, P. Milanfar, G. Golub, Efficient generalized cross-validation with applications to parametric image restoration and resolution enhancement,
IEEE Trans. Image Process 10 (9) (2001) 1299–1308.
[103] K. Kaouk, D.C. Zimmerman, Structural damage assessment using a generalized minimum rank perturbation theory, AIAA J. 32 (4) (1994) 836–842.
[104] D.C. Zimmerman, M. Kaouk, Structural damage detection using a minimum rank update theory, J Vib. Acoust. 116 (2) (1994) 222–231.
[105] S.W. Doebling, Minimum-rank optimal update of elemental stiffness parameters for structural damage identification, AIAA J. 34 (12) (1996)
2615–2621.
[106] E.J. Candès, J. Romberg, T. Tao, Robust uncertainty principles: exact signal reconstruction from highly incomplete frequency information, IEEE Trans.
Inf. Theory 52 (2) (2006) 489–509.
[107] D.L. Donoho, Compressed sensing, IEEE Trans. Inf. Theory 52 (4) (2006) 1289–1306.
[108] R.G. Baraniuk, More is less: signal processing and the data deluge, Science 331 (6018) (2011) 717–719.
[109] D. Malioutov, M. Çetin, A.S. Willsky, A sparse signal reconstruction perspective for source localization with sensor arrays, IEEE Trans. Signal Process.
53 (8) (2005) 3010–3022.
[110] S. Theodoridis, Y. Kopsinis, K. Slavakis, in: Sparsity-aware learning and compressed sensing: An overview, 1, Academic Press Library in Signal Pro-
cessing, 2014, pp. 1271–1377.
[111] X.J. Chen, F.M. Xu, Y.Y. Ye, Lower bound theory of nonzero entries in solutions of l2 –lp minimization, SIAM J. Sci. Comput. 32 (5) (2010) 2832–2852.
[112] R. Chartrand, Exact reconstruction of sparse signals via nonconvex minimization, IEEE Signal Process. Lett. 14 (10) (2007) 707–710.
[113] R. Chartrand, W. Yin, in: Iteratively reweighted algorithms for compressive sensing, Proc. IEEE ICASSP, Las Vegas, Nevada, USA, 2008, pp. 3869–3872.
31 March–4 April.
[114] Y.Q. Bao, H. Li, J.P. Ou, Emerging data technology in structural health monitoring: compressive sensing technology, J. Civ. Struct. Health Monit. 4 (2)
(2014) 77–90.
[115] E.M. Hernandez, Identification of isolated structural damage from incomplete spectrum information using l1 -norm minimization, Mech. Syst. Signal
Process. 46 (1) (2014) 59–69.
[116] X.Q. Zhou, Y. Xia, S. Weng, L1 regularization approach to structural damage detection using frequency data, Struct. Health Monit. 14 (6) (2015)
571–582.
[117] C.D. Zhang, Y.L. Xu, Comparative studies on damage identification with Tikhonov regularization and sparse regularization, Struct. Control Health
Monit. 23 (3) (2016) 560–579.
[118] R.R. Hou, Y. Xia, X.Q. Zhou, Structural damage detection based on l1 regularization using natural frequencies and mode shapes, Struct. Control Health
Monit. 25 (3) (2018) e2107.
28
[119] Y.H. Wu, X.Q. Zhou, L1 regularized model updating for structural damage detection, Int. J. Struct. Stab. Dyn. 18 (12) (2018) 1850157.
[120] C. Zhang, J.Z. Huang, G.Q. Song, L. Chen, Structural damage identification by extended Kalman filter with l1 -norm regularization scheme, Struct.
Control Health Monit. 24 (11) (2017) e1999.
[121] J.Z. Huang, D.S. Li, C. Zhang, H. Li, Improved Kalman filter damage detection approach based on lp regularization, Struct. Control Health Monit. 26
(10) (2019) e2424.
[122] X.Q. Zhou, R.R. Hou, Y.H. Wu, Structural damage detection based on iteratively reweighted l1 regularization algorithm, Adv. Struct. Eng. 22 (6) (2019)
1479–1487.
[123] C.B. Chen, C.D. Pan, Z.P. Chen, L. Yu, Structural damage detection via combining weighted strategy with trace Lasso, Adv. Struct. Eng. 22 (3) (2019)
597–612.
[124] C.B. Chen, L. Yu, A hybrid ant lion optimizer with improved Nelder–Mead algorithm for structural damage detection by improving weighted trace
lasso regularization, Adv. Struct. Eng. (2019) 1–17.
[125] L. Wang, Z.R. Lu, Sensitivity-free damage identification based on incomplete modal data, sparse regularization and alternating minimization approach,
Mech. Syst. Signal Process. 120 (1) (2019) 43–68.
[126] Z.H. Ding, J. Li, H. Hao, Structural damage identification using improved Jaya algorithm based on sparse regularization and Bayesian inference, Mech.
Syst. Signal Process. 132 (1) (2019) 211–231.
[127] D. Mascarenas, A. Cattaneo, J. Theiler, C. Farrar, Compressed sensing techniques for detecting damage in structures, Struct. Health Monit. 12 (2013)
325–338.
[128] Y.C. Yang, S. Nagarajaiah, Output-only modal identification by compressed sensing: non-uniform low-rate random sampling, Mech. Syst. Signal Pro-
cess. 56 (2015) 15–34.
[129] Y.C. Yang, S. Nagarajaiah, Structural damage identification via a combination of blind feature extraction and sparse representation classification, Mech.
[130] H.J. Yao, P. Gerstoft, P.M. Shearer, C. Mecklenbräuker, Compressive sensing of the Tohoku-Oki Mw 9.0 earthquake: frequency-dependent rupture
modes, Geophys. Res. Lett. 38 (20) (2011) L20310.
[131] R.R. Hou, Y. Xia, Y.Q. Bao, X.Q. Zhou, Selection of regularization parameter for l1 -regularized damage detection, J. Sound Vib. 423 (9) (2018) 141–160.
[132] R.R. Hou, Y. Xia, Q. Xia, X.Q. Zhou, Genetic algorithm based optimal sensor placement for L1 -regularized damage detection, Struct. Control Health
Monit. 26 (1) (2019) e2274.
[133] J.H. Holland, Adaptation in Natural and Artificial Systems, The University of Michigan Press, Ann Arbor, 1975.
[134] V. Meruane, W. Heylen, An hybrid real genetic algorithm to detect structural damage using modal properties, Mech. Syst. Signal Process. 25 (5) (2011)
1559–1573.
[135] G. Ghodrati Amiri, S.A. Seyed Razzaghi, A. Bagheri, Damage detection in plates based on pattern search and Genetic algorithms, Smart Struct. Syst. 7
(2) (2011) 117–132.
[136] H.Y. Guo, Z.L. Li, Structural damage identification based on evidence fusion and improved particle swarm optimization, J. Vib. Control 20 (9) (2014)
1297–1292.
[137] Z.P. Chen, L. Yu, A novel PSO-based algorithm for structural damage detection using Bayesian multi-sample objective function, Struct. Eng. Mech. 63
(6) (2017) 825–835.
[138] Z.H. Ding, R.Z. Yao, J. Li, Z.R. Lu, Structural damage identification based on modified Artificial Bee Colony algorithm using modal data, Inverse Probl.
Sci. Eng. 26 (3) (2017) 422–442.
[139] Z.H. Ding, Z.R. Lu, M. Huang, J. Liu, Improved artificial bee colony algorithm for crack identification in beam using natural frequencies only, Inverse
Probl. Sci. Eng. 25 (2) (2017) 218–238.
[140] K. Lakshmi, A.R.M. Rao, N. Gopalakrishnan, Singular spectrum analysis combined with ARMAX model for structural damage detection, Struct. Control
Health Monit. 24 (9) (2017) e1960.
[141] A.A. Mosavi, D. Dickey, R. Seracino, S. Rizkalla, Identifying damage locations under ambient vibrations utilizing vector autoregressive models and
Mahalanobis distances, Mech. Syst. Signal Process. 26 (2012) 254–267.
[142] K.C. Chang, C.W. Kim, Modal-parameter identification and vibration-based damage detection of a damaged steel truss bridge, Eng. Struct. 122 (1)
(2016) 156–173.
[143] R.A. Johnson, D.W. Wichern, Applied multivariate statistical analysis, 5th edn., Prentice Hall, Upper Saddle River, New Jersey, 2002.
[144] A.C. Rencher, Methods of multivariate analysis (2002).
[145] T. Nguyen, T.H.T. Chan, D.P. Thambiratnam, Controlled Monte Carlo data generation for statistical damage identification employing Mahalanobis
squared distance, Struct. Health Monit. 13 (4) (2014) 461–472.
[146] M. Gul, F.N. Catbas, Structural health monitoring and damage assessment using a novel time series analysis methodology with sensor clustering, J.
Sound Vib. 330 (6) (2011) 1196–1210.
[147] M. Gul, F.N. Catbas, Damage assessment with ambient vibration data using a novel time series analysis methodology, J. Struct. Eng. 137 (12) (2010)
1518–1526.
[148] R.G. Yao, S.N. Pakzad, Autoregressive statistical pattern recognition algorithms for damage detection in civil structures, Mech. Syst. Signal Process. 31
(2012) 355–368.
[149] S.G. Shahidi, M.B. Nigro, S.N. Pakzad, Y.C. Pan, Structural damage detection and localisation using multivariate regression models and two-sample
control statistics, Struct. Infrastruct. Eng. 11 (10) (2015) 1277–1293.
[150] E. Figueiredo, J. Figueiras, G. Park, C.R. Farrar, K. Worden, Influence of the autoregressive model order on damage detection, Comput-Aided Civ. Inf.
26 (3) (2011) 225–238.
[151] Y. Lei, Y.Q. Jiang, Z.Q. Xu, Structural damage detection with limited input and output measurement signals, Mech. Syst. Signal Process. 28 (2012)
229–243.
[152] Y. Lei, F. Chen, H. Zhou, An algorithm based on two-step Kalman filter for intelligent structural damage detection, Struct. Control Health Monit. 22
(4) (2015) 694–706.
[153] Y. Lei, Q. Li, F. Chen, Z. Chen, Damage identification of frame structures with joint damage under earthquake excitation, Adv. Struct. Eng. 17 (8) (2014)
1075–1088.
[154] J.H. Weng, C.H. Loh, J.N. Yang, Experimental study of damage detection by data-driven subspace identification and Finite-Element model updating, J.
Struct. Eng. 135 (12) (2009) 1533–1544.
[155] C.R. Farrar, K. Worden, Structural Health Monitoring: A Machine Learning Perspective, Wiley, West Sussex, UK, 2013.
[156] C.M. Bishop, Pattern recognition and machine learning, Springer, Berlin, 2006.
[157] H. Adeli, C. Yeh, Perceptron learning in engineering design, Comput-Aided Civ. Inf. 4 (4) (1989) 247–256.
[158] N. Bakhary, H. Hao, A.J. Deeks, Structure damage detection using neural network with multi-stage substructuring, Adv. Struct. Eng. 13 (1) (2010)
95–110.
[159] S.F. Jiang, C.M. Zhang, S. Zhang, Two-stage structural damage detection using fuzzy neural networks and data fusion techniques, Expert Syst. Appl.38
(1) (2011) 511–519.
[160] U. Dackermann, J.C. Li, B.J. Samali, Identification of member connectivity and mass changes on a two-storey framed structure using frequency re-
sponse functions and artificial neural networks, J. Sound Vib. 332 (16) (2013) 3636–3653.
[161] B. Xu, G. Song, S.F. Masri, Damage detection for a frame structure model using vibration displacement measurement, Struct. Health Monit. 11 (3)
(2011) 281–292.
29
[162] S.J.S. Hakim, H.A. Razak, Structural damage detection of steel bridge girder using artificial neural networks and finite element models, Steel Compos.
Struct. 14 (4) (2013) 367–377 a.
[163] S.J.S. Hakim, H.A. Razak, Adaptive Neuro Fuzzy Inference System (ANFIS) and Artificial neural networks (ANNs) for structural damage identification,
Struct. Eng. Mech. 45 (6) (2013) 779–802 b.
[164] K. Worden, G. Manson, N.R.J. Fieller, Damage detection using outlier analysis, J. Sound Vib. 229 (3) (20 0 0) 647–667.
[165] J.C. Li, U. Dackermann, Y.L. Xu, B. Samali, Damage identification in civil engineering structures utilizing PCA-compressed residual frequency response
functions and neural network ensembles, Struct. Control Health Monit. 18 (2) (2011) 207–226.
[166] B. Samali, U. Dackermann, J. Li, Location and severity identification of notch-type damage in a two-storey steel framed structure utilising frequency
response functions and artificial neural network, Adv. Struct. Eng. 15 (5) (2012) 743–757.
[167] R.P. Bandara, T.H.T. Chan, D.P. Thambiratnam, Frequency response function based damage identification using principal component analysis and pat-
tern recognition technique, Eng. Struct. 66 (1) (2014) 116–128.
[168] X.W. Ye, T. Jin, C.B. Yun, A review on deep learning based structural health monitoring of civil infrastructures, Smart Struct. Syst. 24 (5) (2019)
567–586.
[169] O. Abdeljaber, O. Avci, S. Kiranya, M. Gabbouj, D.J. Inman, Real-time vibration-based structural damage detection using one-dimensional convolutional
neural networks, J. Sound Vib. 388 (3) (2017) 154–170.
[170] Y.F. Duan, Q.Y. Chen, H.M. Zhang, C.B. Yun, S.K. Wu, Q. Zhu, CNN-based damage identification method of tied-arch bridge using spatial-spectral
information, Smart Mate. Struct. 23 (5) (2019) 507–520.
[171] Y.Q. Bao, Z.Y. Tang, H. Li, Y.F. Zhang, Computer vision and deep learning–based data anomaly detection method for structural health monitoring,
[172] V.N. Vapnik, Statistical Learning Theory, Wiley, New York, 1998.
[173] C. Cortes, V. Vapnik, in: Support-vector networks, 20, Mach. Learn, 1995, pp. 273–297.
[174] R. Ghiasi, P. Torkzadeh, M. Noori, A machine-learning approach for structural damage detection using least square support vector machine based on
a new combinational kernel function, Struct. Health Monit. 15 (3) (2016) 302–316.
[175] S.S. Kourehli, LS-SVM regression for structural damage diagnosis using the iterated improved reduction system, Int. J. Struct. Stab. Dyn. 16 (6) (2016)
1550018.
[176] H.B. Liu, Y.B. Jiao, Application of genetic algorithm-support vector machine (GA-SVM) for damage identification of bridge, Int. J. Comput. Intell. Appl.
10 (4) (2011) 383–397.
[177] G.Q. Gui, H. Pan, Z.B. Lin, Y.H. Li, Z.J. 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.
[178] L. Breiman, Random forests, Mach. Learn. 45 (1) (2004) 5–32.
[179] Q.F. Zhou, Y.P. Ning, Q.Q. Zhou, L.K. Luo, J.Y. Lei, Structural damage detection method based on random forests and data fusion, Struct. Health Monit.
12 (1) (2012) 48–58.
[180] J.P. Santos, C. Crémona, L. Calado, P. Silveira, A.D. Orcesi, On-line unsupervised detection of early damage, Struct. Control Health Monit. 23 (2016)
1047–1069.
[181] A.C. Neves, I. González, J. Leander, R. Karoumi, Structural health monitoring of bridges: a model-free ANN-based approach to damage detection, J.
Civil Struct. Health Monit. 7 (2017) 689–702.
[182] M.H. Rafiei, H. Adeli, A novel unsupervised deep learning model for global and local health condition assessment of structures, Eng. Struct. 156 (2018)
598–607.
[183] Y.J. Cha, Z.L. Wang, Unsupervised novelty detection–based structural damage localization using a density peaks-based fast clustering algorithm, Struct.
Health Monit. 17 (2) (2018) 313–324.
[184] O. Avci, O. Abdeljaber, Self-organizing maps for structural damage detection: a novel unsupervised vibration-based algorithm, J. Perform. Constr. Facil.
30 (3) (2016) 04015043.
[185] S. Chen, F. Cerda, P. Rizzo, J. Bielak, J.H. Garrett, J. Kovacevic, Semi-supervised multiresolution classification using adaptive graph filtering with appli-
cation to indirect bridge structural health monitoring, IEEE Trans. Signal Process. 62 (2014) 2879–2893.
[186] Z. Lai, S. Nagarajaiah, Semi-supervised structural linear/nonlinear damage detection and characterization using sparse identification, Struct. Control
Health Monit. 26 (3) (2019) e2306.
[187] T.J. Rogers, K. Worden, R. Fuentes, N. Dervilis, U.T. Tygesen, E.J. Cross, A Bayesian non-parametric clustering approach for semi-supervised Structural
Health Monitoring, Mech. Syst. Signal Process. 119 (2019) 100–119.
[188] X.W. Ye, T. Jin, C.B. Yun, A review on deep learning-based structural health monitoring of civil infrastructures, Smart Struct. Syst. 24 (5) (2019)
567–586.
[189] Y. Xia, H. Hao, J.M.W. Brownjohn, P.Q. Xia, Damage identification of structures with uncertain frequency and mode shape data, Earthq. Eng. Struct.
Dyn. 31 (5) (2002) 1053–1066.
[190] Y. Xia, H. Hao, Statistical damage identification of structures with frequency changes, J. Sound Vib. 263 (4) (2003) 853–870.
[191] E. Simoen, G. De Roeck, G. Lombaert, Dealing with uncertainty in model updating for damage assessment: a review, Mech. Syst. Signal Process. 56
(2015) 123–149.
[192] J.L. Beck, L.S. Katafygiotis, Updatingmodels and their uncertainties. I: Bayesian statistical framework, J. Eng. Mech. 124 (4) (1998) 455–461.
[193] L.S. Katafygiotis, J.L. Beck, Updating models and their uncertainties. II: Model Identifiability, J. Eng. Mech. 124 (4) (1998) 463–467.
[194] P.M. Williams, Bayesian regularization and pruning using a Laplace prior, Neural Comput. 7 (1) (1995) 117–143.
[195] E. Figueiredo, L. Radu, K. Worden, C.R. Farrar, A Bayesian approach based on a Markov-chain Monte Carlo method for damage detection under
unknown sources of variability, Eng. Struct. 80 (1) (2014) 1–10.
[196] S. Arangio, J.L. Beck, Bayesian neural networks for bridge integrity assessment, Struct. Control Health Monit. 19 (1) (2012) 3–21.
[197] S Arangio, F. Bontempi, Structural health monitoring of a cable-stayed bridge with Bayesian neural networks, Struct. Inf. Eng. 11 (4) (2015) 575–587.
[198] H.F. Lam, Q. Hu, M.T. Wong, The Bayesian methodology for the detection of railway ballast damage under a concrete sleeper, Eng. Struct. 81 (15)
(2014) 289–301.
[199] I. Behmanesh, B. Moaveni, Probabilistic identification of simulated damage on the Dowling Hall footbridge through Bayesian finite element model
updating, Struct. Control Health Monit. 22 (3) (2015) 463–483.
[200] 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.
[201] T. Yin, Q.H. Jiang, K.V. Yuen, Vibration-based damage detection for structural connections using incomplete modal data by Bayesian approach and
model reduction technique, Eng. Struct. 132 (1) (2017) 260–277.
[202] K.V. Yuen, J.L. Beck, L.S Katafygiotis, Efficient model updating and health monitoring methodology using incomplete modal data without mode match-
ing, Struct. Control Health Monit. 13 (1) (2006) 91–107.
[203] H.F. Lam, T. Yin, Dynamic reduction-based structural damage detection of transmission towers: practical issues and experimental verification, Eng.
Struct. 33 (5) (2011) 1459–1478.
[204] M.E. Tipping, Sparse Bayesian learning and the relevance vector machine, J. Mach. Learn. Res. 1 (2001) 211–244.
[205] D.P. Wipf, B.D. Rao, Sparse Bayesian learning for basis selection, IEEE Trans. Signal Process. 52 (8) (2004) 2153–2164.
[206] S. Ji, Y. Xue, L. Carin, Bayesian compressive sensing, IEEE Trans. Signal Process. 56 (6) (2008) 2346–2356.
[207] O. Williams, A. Blake, R. Cipolla, Sparse Bayesian learning for efficient visual tracking, IEEE T. Pattern Anal. 27 (8) (2005) 1292–1304.
30
[208] Z. Zhang, B.D. Rao, Sparse signal recovery with temporally correlated source vectors using sparse Bayesian learning, IEEE J. Sel. Topics Signal Process.
5 (5) (2011) 912–926.
[209] J. Lin, M. Nassar, B.L. Evans, Impulsive noise mitigation in powerline communications using sparse Bayesian learning, IEEE J. Sel. Area. Comm. 31 (7)
(2013) 1172–1183.
[210] D.P. Wipf, Sparse estimation with structured dictionaries, Proc. 24th Advances in NIPS, Granada, Spain, 12–15 December, (2011) 2016–2024.
[211] Y. Huang, J.L. Beck, Hierarchical sparse Bayesian learning for structural health monitoring with incomplete modal data, Int. J. Uncertain. Quan. 5 (2)
(2015) 139–169.
[212] Y. Huang, J.L. Beck, H. Li, Hierarchical sparse Bayesian learning for structural damage detection: Theory, computation and application, Struct. Saf. 64
(2017) 37–53.
[213] Y. Huang, H. Li, S. Wu, Y. Yang, Fractal dimension based damage identification incorporating multi-task sparse Bayesian learning, Smart Mate. Struct.
27 (7) (2018) 075020.
[214] Y. Huang, J.L. Beck, H. Li, Multi-task sparse Bayesian learning with applications in structural health monitoring, Comput-Aided Civ. Inf. 34 (9) (2019)
732–754.
[215] S. Ji, D. Dunson, L. Carin, Multi-task compressive sensing, IEEE Trans. Signal Process. 57 (1) (2009) 92–106.
[216] R.R. Hou, Y. Xia, X.Q. Zhou, Y. Huang, Sparse Bayesian learning for structural damage detection using expectation–maximization technique, Struct.
Control Health Monit. 26 (5) (2019) e2343.
[217] X.Y. Wang, R.R. Hou, Y. Xia, X.Q. Zhou, Laplace approximation in sparse Bayesian learning for structural damage detection, Mech. Syst. Signal Process.
140 (2020) 106701 under review.
[218] Y. Huang, J.L. Beck, H. Li, Bayesian system identification based on hierarchical sparse Bayesian learning and Gibbs Sampling with application to
structural damage assessment, Comput. Method. Appl. M. 318 (1) (2017) 382–411.
[219] Y. Huang, J.L. Beck, Full Gibbs Sampling procedure for Bayesian system identification incorporating sparse Bayesian learning with Automatic Relevance
Determination, Comput-Aided Civ. Inf. 33 (9) (2018) 712–730.
[220] R.D. Adams, P. Cawley, C.J. Pye, B.J. Stone, A vibration technique for non-destructively assessing the integrity of structures, J. Mech. Eng. Sci. 20 (2)
(1978) 93–100.
[221] H. Sohn, Effects of environmental and operational variability on structural health monitoring, Philos. Trans. R. Soc. A 365 (1851) (2007) 539–560.
[222] Z.D. Xu, Z.S. Wu, Simulation of the effect of temperature variation on damage detection in a long-span cable-stayed bridge, Struct. Health Monit. 6
(3) (2007) 177–189.
[223] H. Li, S.L. Li, J.P. Ou, S.L. Li, Modal identification of bridges under varying environmental conditions: Temperature and wind effects, Struct. Control
Health Monit. 17 (5) (2010) 499–512.
[224] Y. Xia, H. Hao, G. Zanardo, A. Deeks, Long term vibration monitoring of an RC slab: temperature and humidity effect, Eng. Struct. 28 (3) (2006)
441–452.
[225] H.F. Zhou, Y.Q. Ni, J.M. Ko, Eliminating temperature effect in Vibration-Based structural damage detection, J. Eng. Mech. 137 (12) (2011) 785–796.
[226] Y.L. Xu, B. Chen, C.L. Ng, K.Y. Wong, W.Y. Chan, Monitoring temperature effect on a Long suspension bridge, Struct. Control Health Monit. 17 (2010)
632–653.
[227] A.M. Yan, G. Kerschen, P. De Boe, J.C. Golinval, Structural damage diagnosis under changing environmental conditions – Part 1: linear analysis, Mech.
[228] D.F. Giraldo, S.J. Dyke, J.M. Caicedo, Damage detection accommodating varying environmental conditions, Struct. Health Monit. 5 (2) (2006) 155–172.
[229] A. Deraemaeker, E. Reynders, G. De Roeck, J. Kullaa, Vibration-based structural health monitoring using output–only measurements under changing
environment, Mech. Syst. Signal Process. 22 (1) (2008) 34–56.
[230] Y. Deng, Y.L. Ding, A.Q. Li, Structural condition assessment of long-span suspension bridges using long-term monitoring data, Earthq. Eng. Eng. Vib.
9 (1) (2010) 123–131.
[231] Y.Q. Bao, Y. Xia, H. Li, Y.L. Xu, P. Zhang, Data fusion-based structural damage detection under varying temperature conditions, Int. J. Struct. Stab. Dyn.
12 (6) (2012) 1–27.
[232] Y.Q. Bao, H. Li, Y.H. An, J.P. Ou, Dempster–Shafer evidence theory approach to structural damage detection, Struct. Health Monit. 11 (1) (2011) 13–26.
[233] F. Magalhães, A. Cunha, E. Caetano, Vibration based structural health monitoring of an arch bridge: from automated OMA to damage detection, Mech.
Syst. Signal Process. 28 (2012) 212–228.
[234] H.F. Zhou, Y.Q. Ni, J.M. Ko, Constructing input to neural networks for modeling temperature-caused modal variability: mean temperatures, effective
temperatures, and principal components of temperatures, Eng. Struct. 32 (6) (2010) 1747–1759.
[235] C.H. Jin, S. Jang, X.R. Sun, J. Li, R. Christenson, Damage detection of a highway bridge under severe temperature changes using extended Kalman filter
trained neural network, J. Civ. Struct. Health Monit. 6 (3) (2016) 545–560.
[236] J. Kullaa, Distinguishing between sensor fault, structural damage, and environmental or operational effects in structural health monitoring, Mech.
[237] Y. Shokrani, V.K. Dertimanis, E.N. Chatzi, M.N. Savoia, On the use of mode shape curvatures for damage localization under varying environmental
conditions, Struct. Control Health Monit. 25 (4) (2018) e2132.
[238] B. Kostić, M. Gül, Vibration-based damage detection of bridges under varying temperature effects using time-series analysis and artificial neural
networks, J. Bridge Eng. 22 (10) (2017) 04017065.
[239] M. Fallahian, F. Khoshnoudian, V. Meruane, Ensemble classification method for structural damage assessment under varying temperature, Struct.
Health Monit. 17 (4) (2018) 747–762.
[240] Y.B. Liang, D.S. Li, G.B. Song, Q. Feng, Frequency co-integration-based damage detection for bridges under the influence of environmental temperature
variation, Measurement 125 (2018) 163–175.
[241] K. Erazo, D. Sen, S. Nagarajaiah, L.M. Sun, Vibration-based structural health monitoring under changing environmental conditions using Kalman
filtering, Mech. Syst. Signal Process. 117 (15) (2019) 1–15.
[242] E. Reynders, G. Wursten, G. De Roeck, Output-only structural health monitoring in changing environmental conditions by means of nonlinear system
identification, Struct. Health Monit. 13 (1) (2014) 82–93.
[243] B. Schölkopf, A. Smola, K.R. Müller, Nonlinear component analysis as a kernel eigenvalue problem, Neural Comput. 10 (5) (1998) 1299–1319.
[244] T.Y. Hsu, C.H. Loh, Damage detection accommodating nonlinear environmental effects by nonlinear principal component analysis, Struct. Control
Health Monit. 17 (3) (2010) 338–354.
[245] V.H. Nguyen, J. Mahowald, J.C. Golinval, S. Maas, Damage detection in civil engineering structure considering temperature effect, Dyn. Civ. Struct. 4
(2014) 187–196.
[246] K. Worden, C.R. Farrar, J. Hayward, M.D. Todd, A review of applications of nonlinear dynamics to structural health monitoring, Struct. Control Health
Monit. 15 (2007) 540–567.
[247] G.D. Chen, X.B. Yang, X.F. Ying, A. Nanni, Damage detection of concrete beams using nonlinear features of forced vibration, Struct. Health Monit. 5
(2) (2006) 125–141.
[248] M. Schetzen, The Volterra and Wiener Theories of Nonlinear Systems, Krieger Publishing Company, Florida, USA, 2006.
[249] A. Chatterjee, Structural damage assessment in a cantilever beam with a breathing crack using higher order frequency response functions, J. Sound
Vib. 329 (16) (2010) 3325–3334.
[250] U. Andreaus, P. Baragatti, Cracked beam identification by numerically analysing the nonlinear behaviour of the harmonically forced response, J. Sound
Vib. 330 (4) (2011) 721–742.
31
[251] U. Andreaus, P. Baragatti, Experimental damage detection of cracked beams by using nonlinear characteristics of forced response, Mech. Syst. Signal
Process. 31 (2012) 382–404.
[252] Z.K. Peng, Z.Q. Lang, C. Wolters, S.A. Billings, K. Worden, Feasibility study of structural damage detection using NARMAX modelling and Nonlinear
Output Frequency Response Function based analysis, Mech. Syst. Signal Process. 25 (2011) 1045–1061.
[253] T. Chanpheng, H. Yamada, H. Katsuchi, E. Sasaki, Nonlinear features for damage detection on large civil structures due to earthquakes, Struct. Health
Monit. 11 (4) (2012) 482–488.
[254] Z.C. Wang, D. Geng, W.X. Ren, G.F. Zhang, Damage detection of nonlinear structures with analytical mode decomposition and Hilbert transform, Smart
Struct. Syst. 15 (1) (2015) 1–13.
[255] L. Bornn, C.R. Farrar, G. Park, Damage detection in initially nonlinear systems, Int. J. Eng. Sci. 48 (10) (2010) 909–920.
[256] S.B. Shiki, S. da Silva, M.D. Todd, On the application of discrete-time Volterra series for the damage detection problem in initially nonlinear systems,
[257] L.G.G. Villania, S. Da Silva, A. Cunha Jr, Damage detection in uncertain nonlinear systems based on stochastic Volterra series considering data variation
caused by the uncertainties, Mech. Syst. Signal Process. 125 (2019) 288–310.
[258] H. Ebrahimian, R. Astroza, J.P. Conte, Extended Kalman filter for material parameter estimation in nonlinear structural finite element models using
direct differentiation method, Earthq. Eng. Struct. Dyn. 44 (10) (2015) 1495–1522.
[259] H. Ebrahimian, R. Astroza, J.P. Conte, R.A. de Callafonc, Nonlinear finite element model updating for damage identification of civil structures using
batch Bayesian estimation, Mech. Syst. Signal Process. 84 (2017) 194–222.
[260] H. Ebrahimian, R. Astroza, J.P. Conte, C. Papadimitriou, Bayesian optimal estimation for output-only nonlinear system and damage identification of
civil structures, Struct. Control Health Monit. 25 (4) (2017) e2128.
[261] R. Astroza, H. Ebrahimian, Y. Li, J.P. Conte, Bayesian nonlinear structural FE model and seismic input identification for damage assessment of civil
structures, Mech. Syst. Signal Process. 93 (1) (2017) 661–687.
[262] Y.B. Yang, C.W. Lin, J.D. Yau, Extracting bridge frequencies from the dynamic response of a passing vehicle, J. Sound Vib. 272 (3-5) (2004) 471–493.
[263] Y.B. Yang, C.W. Lin, Vehicle-bridge interaction dynamics and potential applications, J. Sound Vib. 284 (1-2) (2005) 205–226.
[264] A. Malekjafarian, P.J. McGetrick, E.J. Obrien, A review of indirect bridge monitoring using passing vehicles, Shock Vib. (2015) 286139.
[265] D.M. Siringoringo, Y. Fujino, Estimating bridge fundamental frequency from vibration response of instrumented passing vehicle: analytical and exper-
imental study, Adv. in Struct. Eng. 15 (3) (2012) 417–433.
[266] Y.B. Yang, K.C. Chang, Y.C. Li, Filtering techniques for extracting bridge frequencies from a test vehicle moving over the bridge, Eng. Struct. 48 (2013)
353–362.
[267] Y. Zhang, L.Q. Wang, Z.H. Xiang, Damage detection by mode shape squares extracted from a passing vehicle, J. Sound Vib. 331 (2) (2012) 291–307.
[268] Y. Oshima, K. Yamamoto, K. Sugiura, Damage assessment of a bridge based on mode shapes estimated by responses of passing vehicles, Smart Struct.
Syst. 13 (5) (2014) 731–753.
[269] K.V. Nguyen, H.T. Tran, Multi-cracks detection of a beam-like structure based on the on-vehicle vibration signal and wavelet analysis, J. Sound Vib.
329 (21) (2010) 4455–4465.
[270] E.J. Obrien, A. Malekjafarian, A. González, Application of empirical mode decomposition to drive-by bridge damage detection, Eur. J Mech. A/Solids
61 (2017) 151–163.
[271] E.J. Obrien, J. Keenahan, Drive-by damage detection in bridges using the apparent profile, Struct. Control Health Monit. 22 (2015) 813–825.
[272] J.C. Keenahan, E.J. Obrien, Drive-by damage detection with a TSD and time-shifted curvature, J. Civ. Struct. Health Monit. 8 (3) (2018) 383–394.
[273] R.H. Myers, D.C. Montgomery, Response Surface Methodology: Process and Product Optimization Using Designed Experiments, 2nd edn., Wiley, New
York, 2002.
[274] S.E. Fang, R. Perera, Damage identification by response surface based model updating using D-optimal design, Mech. Syst. Signal Process. 25 (2)
(2011) 717–733.
[275] J.T. Kim, J.H. Park, D.S. Hong, W.S. Park, Hybrid health monitoring of prestressed concrete girder bridges by sequential vibration-impedance ap-
proaches, Eng. Struct. 32 (1) (2010) 115–128.
[276] Z.B. Yang, X.F. Chen, Y.Y. Jiang, Z. He, Generalised local entropy analysis for crack detection in beam-like structures, Nondestruct, Test. Eva. 29 (2)
(2014) 133–153.
[277] Z.B. Yang, X.F. Chen, Y. Xie, H.H. Miao, J.J. Gao, K.Z. Qi, Hybrid two-step method of damage detection for plate-like structures, Struct. Control Health
Monit. 23 (2016) 267–285.
[278] Y. Wang, H. Hao, Damage identification scheme based on compressive sensing, J. Comput. Civ. Eng. 29 (2) (2015) 1–10.
[279] X.Q. Zhou, Y. Xia, H. Hao, Sensor placement for structural damage detection considering measurement uncertainties, Adv. Struct. Eng. 16 (5) (2013).
[280] J. Li, H. Hao, Z.W. Chen, Damage identification and optimal sensor placement for structures under unknown traffic-induced vibrations, J. Aerosp. Eng.
30 (2) (2017) B4015001.
[281] J.F. Lin, Y.L. Xu, S.S. Law, Structural damage detection-oriented multi-type sensor placement with multi-objective optimization, J. Sound Vib. 422
(2018) 568–589.
[282] J.F. Lin, Y.L. Xu, S. Zhan, Experimental investigation on multi-objective multi-type sensor optimal placement for structural damage detection, Struct.
Health Monit. 18 (3) (2019) 882–901.
[283] Y.L. Guo, Y.Q. Ni, S.K. Chen, Optimal sensor placement for damage detection of bridges subject to ship collision, Struct. Control Health Monit. 24 (9)
(2017) e1963.
[284] S. Beygzadeh, E. Salajegheh, P. Torkzadeh, J. Salajegheh, S.S. Naseralavi, An improved genetic algorithm for optimal sensor placement in space struc-
tures damage detection, Int. J. Space Struct. 29 (3) (2014).
[285] F.P. Kopsaftopoulos, S.D. Fassois, Vibration based health monitoring for a lightweight truss structure: experimental assessment of several statistical
time series methods, Mech. Syst. Signal Process. 24 (7) (2010) 1977–1997.
[286] I. Talebinejad, C. Fischer, F. Ansari, Numerical evaluation of vibration-based methods for damage assessment of cable-stayed bridges, Comput-Aided
Civ. Inf. 26 (3) (2011) 239–251.
[287] R.A. Saeed, A.N. Galybin, V. Popov, Crack identification in curvilinear beams by using ANN and ANFIS based on natural frequencies and frequency
response functions, Neural Comput. Appl. 21 (7) (2012) 1629–1645.
[288] D. Dessi, G. Camerlengo, Damage identification techniques via modal curvature analysis: overview and comparison, Mech. Syst. Signal Process. 52
(2015) 181–205.
[289] E. Figueiredo, G. Park, C.R. Farrar, K. Worden, J. Figueiras, Machine learning algorithms for damage detection under operational and environmental
variability, Struct. Health Monit. 10 (6) (2010) 559–572.
[290] A. Santos, E. Figueiredo, M.F.M. Silva, C.S. Sales, J.C.W.A. Costa, Machine learning algorithms for damage detection: Kernel-based approaches, J. Sound
Vib. 363 (2016) 584–599.
[291] S.J. Dyke, D. Bernal, J.L. Beck, C. Ventura, in: An experimental benchmark problem in structural health monitoring, In: Proc. of the 3rd Int. Workshop
on Struct. Health Monit., Stanford, Calif., 2001.
[292] S.J. Dyke, D. Bernal, J.L. Beck, C. Ventura, Experimental phase of the structural health monitoring benchmark problem, Proc. 16th Eng. Mech. Conf.
ASCE Reston, Va. (2003).
32

1 1 1 1 1 Review On The New Development of Vibration-Based Damage2010 2019

Uploaded by

Copyright:

Available Formats

1 1 1 1 1 Review On The New Development of Vibration-Based Damage2010 2019

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

1 1 1 1 1 Review On The New Development of Vibration-Based Damage2010 2019

Uploaded by

Copyright:

Available Formats

Journal of Sound and Vibration 491 (2021) 115741

Contents lists available at ScienceDirect

Journal of Sound and Vibration

Review on the new development of vibration-based damage

AANN Auto-associative neural network

2. Vibration-based damage identiﬁcation methods

2.1. Modal parameter-based methods

2.1.1. Mode shapes

2.1.2. Natural frequencies and mode shapes

2.1.4. FRFs and their variants

Features Authors Damage identiﬁcation level Applications Remarks

Mode shapes Yoon et al. [17] 1-3 Experimental 1D beam and 2D M

Note: ∗ – Quantitatively indicate the relative severity of damage;

2.2. Signal processing-based methods

2.3. Finite element model updating methods

2.3.1. Conventional model updating

Authors Methods/Features Damage identiﬁcation level Applications Remarks

2.3.2. Substructuring techniques

2.3.3. Regularisation techniques

2.3.4. Exploiting structural damage sparsity

Method Feature Application Authors

Conventional model updating

2.4. Optimisation algorithms

Method Strengths Limitations Features Damage identiﬁcation level Applications Authors

2.5. Statistical time series methods

2.6.1. Supervised learning

Authors Methods Features Damage Applications Remarks

Method Strengths Limitations

2.6.2. Unsupervised learning

2.6.3. Semi-supervised learning

Authors Methods Inputs Damage Applications Remarks

Authors Methods Inputs Damage Applications Remarks

Avci and Self-organizing Random 1 Phase II IASC-ASCE

2.7. Bayesian methods

Authors Methods Features Bayesian inference Damage Application

Figueiredo et al. MSD Frequencies Markov-chain 1 Z-24 Bridge

2.8. Varying temperature conditions

2.8.1. Using temperature measurement

2.8.2. Without temperature measurement

2.9. Nonlinear methods

Authors Methods Features Relation Damage Applications

Using the measurement of temperature

Authors Methods Features Damage type and Applications

2.10. Other methods

2.11. Comparative studies

3. Challenges and future research

Declaration of Competing Interest

You might also like