Review

Structural Health Monitoring

2021, Vol. 0(0) 1–50
Machine learning and structural health © The Author(s) 2021

monitoring overview with emerging Article reuse guidelines:

sagepub.com/journals-permissions
technology and high-dimensional data DOI: 10.1177/14759217211036880
journals.sagepub.com/home/shm
source highlights

Arman Malekloo1, Ekin Ozer2 , Mohammad AlHamaydeh3  and Mark Girolami4,5

Abstract
Conventional damage detection techniques are gradually being replaced by state-of-the-art smart monitoring and decision-
making solutions. Near real-time and online damage assessment in structural health monitoring (SHM) systems is a
promising transition toward bridging the gaps between the past’s applicative inefficiencies and the emerging technologies of
the future. In the age of the smart city, Internet of Things (IoT), and big data analytics, the complex nature of data-driven civil
infrastructures monitoring frameworks has not been fully matured. Machine learning (ML) algorithms are thus providing the
necessary tools to augment the capabilities of SHM systems and provide intelligent solutions for the challenges of the past.
This article aims to clarify and review the ML frontiers involved in modern SHM systems. A detailed analysis of the ML
pipelines is provided, and the in-demand methods and algorithms are summarized in augmentative tables and figures.
Connecting the ubiquitous sensing and big data processing of critical information in infrastructures through the IoT
paradigm is the future of SHM systems. In line with these digital advancements, considering the next-generation SHM and
ML combinations, recent breakthroughs in (1) mobile device-assisted, (2) unmanned aerial vehicles, (3) virtual/augmented
reality, and (4) digital twins are discussed at length. Finally, the current and future challenges and open research issues in
SHM-ML conjunction are examined. The roadmap of utilizing emerging technologies within ML-engaged SHM is still in its
infancy; thus, the article offers an outlook on the future of monitoring systems in assessing civil infrastructure integrity.

Keywords
Structural health monitoring, machine learning, internet of things, big data, emerging technologies

Introduction inspection either carried out using non-destructive testing

/non-destructive evaluation (NDE) or by visual observation.
Machinery equipment and structures, particularly lifelines, The latter method, although it performs well for straight-
fabricate the most critical components in this modern age, forward applications, is susceptible to subjectivity, human
and they have become an indispensable part of the present errors, prolonged duration, and occupant’s safety for more
day. In the case of utility lifelines, such as roadways, complex systems. Prior knowledge of the damaged area is
bridges, and powerlines, any threats that could cause a
failure in any part of the system, no matter the extent, can
eventually lead to the disruption of a whole city or a country. 1
Department of Civil Engineering, Middle East Technical University,
This means that if it was possible to predict future failures Ankara, Turkey
2
and detect the existing ones, this could potentially lead to a Department of Civil & Environmental Engineering, University of
Strathclyde, Glasgow, UK
reduction in direct and indirect economic costs and human 3
Department of Civil Engineering, College of Engineering, American
life fatalities. The key to doing so lies in identifying damage University of Sharjah, Sharjah, UAE
in structures. Damage is typically defined, in simple terms, 4
Department of Engineering, University of Cambridge, Cambridge, UK
5
as any change to the material or geometry, such as the The Alan Turing Institute, London, UK
boundary condition that can alter the dynamic properties or
Corresponding author:
the response of the structure,1 thus adversely affecting the Ekin Ozer, Department of Civil & Environmental Engineering, University of
current or future performance of the system.2 In the past, Strathclyde, 75 Montrose Street, Glasgow G1 1XJ, UK.
identifying damage was only based on a periodical Email: [email protected]
2 Structural Health Monitoring 0(0)

necessary for such techniques which would be impossible Model-driven SHM versus data-driven SHM
for small and unreachable regions without completely
dismantling part of that area first. Such damage detection is As stated earlier, to identify the damage, the undamaged
localized, meaning it cannot represent the global behavior or state of the structure must either be assumed or developed.
the system’s response. Similarly, the extent of the damage is nearly impossible to
The impracticality of visual inspection for large and quantify or assess if the previous “undamaged state” is
complex civil infrastructures and long biennial inspection unknown. Therefore, the ability to identify a damaged
intervals has opened up the possibility of incorporating structure from the given measurements ultimately lies in
condition-based assessment techniques. As such, struc- realizing the previously recorded information and the pat-
tural health monitoring (SHM) has emerged to provide tern of changes it follows throughout the measuring period.
the transition from offline damage identification to near In certain SHM applications, a prior model, typically the
real-time and online damage assessment. In layman’s finite element model (FEM) of the structure, is useful as a
terms, SHM is a damage detection strategy that can baseline. Model updating is then performed, replacing the
observe a structure over a long period using a series initial assumptions with the measured values. This is then
of continuous measuring devices. Sensitive features considered as the original state of the structure. Further
extracted from these continuous measurements and the updating of the model can, therefore, identify the damage by
statistical analysis of such measures can provide the considering the structural changes. This process of SHM
ability to assess the current performance of structures. implementation is a model-driven method. Therefore, an
Figure 1 represents the typical components of an SHM accurate analytical model of the structure is required.3
system. It starts with a selection of sensors and the There are numerous works related to model-driven
placement of them in strategic locations on the structure. SHM. To name a few, Cao et al.4 developed a piezoelec-
The collected data through the data acquisition system are tric impedance measurement for structural damage identi-
transmitted to the processing unit and stored and managed fication through an inverse analysis. Similarly, Moore et al.5
in a database system. The evaluation of the collected data identified cracks in a thin plate by model updating. Gen-
and the health state of the system is determined through erally, coming up with an accurate model is burdensome.
several techniques and algorithms. In the end, based on Model discrepancies, especially for complex structures, are
the location and severity of the identified damage and inevitable with little to no information about joints and
how it can propagate in the future, inspection and bonds. Such an inverse problem is not well-posed6 and
maintenance during the decision-making process will be requires regularization and simplification.7 An alternative to
decided and carried out. a model-driven SHM system is a data-driven model. Other
than relying on the physical model of the structure, the
model construction is dependent on statistical pattern rec-
ognition (PR), which is usually applied by machine learning
(ML) algorithms.
In contrast to having an FEM and updating the model
later, the sensing devices’ data from the structures are used
more conveniently in the undamaged state and under few
circumstances in the damaged state. In cases where insuf-
ficient labeled data exists, the data-driven approach can take
an unsupervised form, or a hybrid model can be utilized for
generating additional data. Augmentation of data-driven
SHM systems with FEM can generate labeled datasets
for training validation and testing phases. However, it is
crucial to highlight that physical models are computation-
ally intensive and need validation with experimental re-
sults.8 On the other hand, not every ML algorithm is capable
of damage prognosis, meaning data-driven approaches are
not always predictive models. Therefore, the decision be-
tween employing model-driven or data-driven SHM sys-
tems or both ultimately boils down to realizing (1) the
proposed system’s requirements, (2) the complexity of
the application where the system is deployed, and (3) if the
existing data and models can support and provide valuable
Figure 1. Typical components of SHM. inferences about the health state of the structure. For
Malekloo et al. 3

example, suppose one prefers a hybrid combination of the cases, and the current information for damage state is
two methods. In that case, the system’s predictive accuracy limited, if not unavailable. For such situations, there exists a
depends on the performance of the physics-based model and method called unsupervised learning. Instead of learning the
if the measured data from the data-driven approach is rel- models and train based on the data, a relatively simple
evant and usable for training and validation. approach, novelty, or outlier detection is applied.11 An
initial baseline of the model is therefore created assuming
normal operative conditions. Later, upon receiving new
Damage definition and identification data from the sensing systems during the operation mode,
A vertical hierarchy is typically considered in order to the algorithm detects any outlier given the set threshold
identify damage. A pioneered damage typology scheme was defined by the system.12 One example of an unsupervised
offered by Rytter.9 Damage state was categorized into four algorithm was tested on an aircraft fuselage and multi-
levels, namely: layered carbon fiber–reinforced plastic (CFRP) plate for
damage detection.13
1. Existence of damage—Detection Compared to supervised learning, the unsupervised
2. Position of damage—Location method provides a clear advantage as it no longer requires
3. Severity of damage—Extent having prior information about the damaged state of the
4. Prognosis of damage—Prediction structure. However, this learning model can only be used to
detect and sometimes, but not always, locate damage.14 In
In such a hierarchy, knowledge of the previous level is addition, many of the implemented ML approaches for
generally essential for complete damage identification. damage detection do not consider environmental and op-
Thus, the success at each level is likely to depend on the erational factors (EOFs) and only rely on the severe
performance of the lower levels. With the advent of ML and damages that occur on the structures. Temperature effect
PR algorithms, a new level can be added to the above. and traffic loading are a few of the neglected variabilities
Determination of the type or classification of damage is the that, in reality, have a significant influence on the in-service
level that is possible through the use of ML algorithms.10 structures’ response.15 Some works16–19 have extensively
This new step lies between steps 2 and 3 introduced above. studied the effect of such variabilities for an extended period
Figure 2 depicts the 5-step hierarchical damage identifi- ranging from 1 to 2 years. Thus, an unsupervised approach
cation from detection to prediction transactions. cannot effectively be used on its own when external factors’
Given that both damaged and undamaged information is dependency requires consideration while identifying dam-
available, a supervised learning algorithm can effectively go age.20 Rather, a coupled approach of model-driven and data-
through all five damage detection levels. This, as explained driven algorithms can work together to achieve a reasonable
before, requires extensive data to be readily available from damage identification level.19
the sensing systems, the physical-based models, or the
experiments. Nevertheless, this is not possible in many
Objectives of this study
The application of PR is not a new topic and dates back to
the early 70s and 80s. In simple terms, PR is a tool to
represent and recognize regularities in data. Sometimes,
simple mathematical models based on a shared domain
about a specific application can be used to infer patterns
from a set of data and classify accordingly. During the
1990s, however, instead of relying on models derived by an
expert (usually researchers) to classify data, machines were
used to learn from the data, generate the most probable
outcome, and validate the model based on unseen set data.
The most likely outcome is a result of statistical PR al-
gorithms, which are generally referred to as ML techniques.
This review aims to generalize these applications har-
moniously using ML and SHM frameworks. Many methods
with different results exist in the rich body of literature.
Several approaches and techniques for feature extraction,
data normalizations, and dimensionality reductions are
employed for various civil infrastructures. This review
Figure 2. Five-step hierarchical damage identification scheme. brings a systematic collection of different SHM applications
4 Structural Health Monitoring 0(0)

compatible with the statistical PR perspective. The readers, 1. The pipeline of ML in each component that makes up
therefore, are introduced to the concept of ML and its SHM systems.
utilization in the SHM paradigm. Moreover, model-driven 2. The different tools and algorithms used in ML and
and data-driven approaches in SHM will be discussed, but DL processes for each level of SHM damage
an emphasis will be placed on data-driven SHM ap- identification.
proaches. In addition, tables and figures refine the ML 3. The different learning algorithms proposed for
taxonomy behind the vast SHM literature complementing context-dependent applications.
the article. Next-generation SHM potentials such as un- 4. Extension into IoT age-related and next-generation
manned aerial vehicle (UAV)-assisted SHM, mobile-SHM, emerging technologies and data science prospects for
and virtual/augmented reality–supported SHM are also SHM.
addressed in this study together with the digital twin, smart
city, and big data era.
Comparison of past relevant reviews
For better readability, the abbreviations used in this
survey, along with their definitions, are provided in Table 1. Many works in the past years or so have reviewed ML
In summary, the review aims to consider: aspects of SHM. Nevertheless, there is a growing need for

Table 1. List of used abbreviation.

Abbreviation Definition Abbreviation Definition

AANN Auto-associative neural network MLR Multiple linear regression

AI Artificial intelligence mRMR Minimum redundancy feature selection
ANN Artificial neural networks MSD Mahalanobis squared distance
ANOVA Analysis of variance NB Naı̈ve Bayes
AR Augmented reality NDT Non-destructive testing
ARMA Autoregressive moving-average model NLPCA Nonlinear principal component analysis
BD Big data NN Neural networks
CART Classification and regression tree PCA Principal component analysis
CFRP Carbon fiber–reinforced plastic PCR Principal component regression
CNN Convolutional neural network PPCA Probabilistic principal component analysis
CCF Cross-correlation function PR Pattern recognition
CSD Cross spectral density PSD Power spectral density
DA Deep autoencoders QDA Quadratic discriminant analysis
DL Deep learning QP Quadratic programming
DT Decision tree RDA Regularized discriminant analysis
EOF Environmental and operational factor RF Random forest
FA Factor analysis ROC Receiver operating characteristic
FDD Fervency domain decomposition SA Stacked autoencoders
FEM Finite element model SFA Slow feature analysis
FRF Frequency response function SHM Structural health monitoring
GA Genetic algorithms SNR Signal to noise ratio
GKPCA Greedy kernel principal component analysis SSI Stochastic subspace identification
GMM Gaussian mixture modeling SVD Singular value decomposition
GP Gaussian process SVDD Support vector data description
ICA Independent component analysis SVM Support vector machine
IoT Internet of things SVM Support vector machine
IRF Impulse response function SVR Support vector regression
kNN K-nearest neighbor UAV Unmanned aerial vehicle
KPCA Kernel principal component analysis UmRmR Unsupervised redundancy–maximum
relevance
LDA Linear discriminant analysis VR Virtual reality
LLE Locally linear embedding WIM Weigh-in-motion
LQD-RKHS Log quantile-density of reproducing kernel Hilbert space WSN Wireless sensor network
ML Machine learning — —
MLP Multilayer perceptrons — —
Malekloo et al. 5

an up-to-date survey connecting the algorithmic advance- very similar to the previous paper and can be considered a
ments with the forthcoming SHM trends utilizing emerging comparable survey to ours. The authors acknowledged the
technologies in a collective form. Noel et al.21 and Arcadius vital role of AI and big data in SHM applications. Analysis
Tokognon et al.22 provided a general review of SHM in the of ML and PR procedures is also depicted in a general way.
context of wireless sensor networks (WSNs) and IoT, re- However, alternative applications in this era were not re-
spectively. Data acquisitions, processing, and network viewed, and the connection to the ultimate goal of SHM in
connectivity were among the topics that were discussed in the context of smart cities and emerging technologies was
detail. However, there was no explicit link between SHM not recognized. Data-driven SHM damage identification
and ML in these two articles, although some basic aspects with DL was reviewed in a recent survey by Azimi.38 The
were defined. Gomes et al.23 and Fan and Qiao24 reviewed authors in their paper discussed in great length the usage of
vibration-based damage detection techniques in the model- DL and machine vision and the new methods of monitoring
driven approach through optimization algorithms and ar- damages, that is, mobile sensors and UAVs in SHM ap-
tificial neural networks (ANN). In other works, Feng and plication. Avci et al.39 reviewed the vibration-based damage
Feng25 and Ye et al.26 summarized data-driven SHM detection in the literature while considering ML and DL
methods using machine vision in the literature. In a most algorithms. In the paper by Tibaduiza Burgos,40 a brief
recent and comprehensive review of computer vision for overview of data-driven SHM applications and a summary
SHM at both local and global levels, Dong and Catbas27 of ML procedures were presented. In contrast to the papers
presented a detailed breakdown of DL and the challenge and above, the authors discussed some of the implementation
opportunities in the field. The work by Ghiasi et al.28 re- steps of data-driven SHM. However, some important as-
viewed artificial intelligence (AI) techniques for damage pects of ML processes, such as feature selection and ex-
detection and applied common prognostics algorithms for traction, were not comprehensively analyzed. Hou and
several case studies. The authors did not explicitly link their Xia41 reviewed vibration-based damage identification for
findings to the benefits AI could offer in global damage civil engineering structures in the last decade. A thorough
detection for large-scale SHM despite their several tests. A learning algorithms analysis in different steps of ML-
short review of prognostics methods based on both model- enhanced SHM system was not present. Sony et al.42 re-
driven and data-driven approaches was discussed by An viewed the next-generation smart sensing technology in
et al.29 The authors described some ML benefits and SHM. In their paper, the authors included emerging tech-
standard algorithms in SHM; however, a more compre- nologies for collecting data from structures. Alavi and
hensive analysis was not conducted. Moughty and Casas30 Buttlar43 overviewed smartphones’ deployment in major
overviewed the model-driven damage detection and alter- civil engineering areas. An emphasis was placed on the
native solutions. Kerle et al.31 reviewed how UAV-based sensing capabilities of smartphones and their crowdsourc-
damaged mapping could provide a flexible solution for ing power in SHM applications.
damage identification with advanced deep learning (DL) As summarized in Table 2, the majority of the past
approaches. Khan and Yairi32 used a different approach to surveys did not assess some of the important aspects of ML
review DL research on system health management. Al- in SHM systems. Necessary details, systemic explanation of
though the authors centered their systematic review on implementation steps, and available methods for an ML-
aerospace applications, yet it still provides plausible benefits engaged SHM system are the notable items that were
for fault diagnosis and prognostics in different disciplines. missing or partially provided by the previous works.
Similarly, Ye et al.33 reviewed DL application in SHM of Moreover, the shift to the new era of the Internet of Things
civil infrastructures, including the history and how it de- (IoT) and smart city necessitates a connection to be drawn
veloped throughout the years. Both reviews lack extension that links data-driven SHM systems to the future paradigm
to mobile data-driven and UAV-assisted SHM systems. Hou and emerging technologies. ML, DL, and machine vision
et al.34 examined the usefulness of DL applications to can provide the necessary algorithms for condition moni-
address structure inspection and on-site safety monitoring toring of many structures for SHM purposes, with big data
challenges in the current era. In contrast to the previous DL and digital twins as the cornerstone of the future.44
papers, the recent review by Sony et al.35 emphasizes on the The authors imagine that the roadmap of utilizing new
convolutional neural network (CNN) application in SHM. technologies such as ML is not limited to cases where
Owning to its specificity, major key points that were damage is to be detected. The method used for data col-
overlooked by the previous DL reviews were comprehen- lection, either through WSNs or UAVs, the novelty of the
sively covered. processing type, and lastly, the expected utilization of the
Salehi and Burgueño36 discussed, in great detail, the result obtained are the factors that this review article hopes
power of AI in ML, PR, and DL. However, a specific to achieve.
breakdown of ML and PR, especially in SHM, was not Connecting the paradigm of sensing and processing
explicitly examined. The recent review by Sun et al.37 is critical information in infrastructures in a new domain with
6 Structural Health Monitoring 0(0)

Table 2. Summary of related surveys.

Novel
Detailed applications Smart
Model- Data- Application of ML/DL Mobile Machine (UAV, VR, AR, city and Open
Reference driven driven ML and DL breakdown applications vision etc.) IoT issues

Noel et al.21 3* 3** — — — — — — 3

Tokognon 3* 3** — — — — — 3 —
et al.22
Gomes et al.,23 3 — 3* — — — — — —
Fan and
Qiao24
Feng and — — 3 — — 3 — — —
Feng25
Ye et al.26 — — 3* — — 3* — — —
Dong and — 3 3 3 — 3 3 — 3
Catbas27
Sony et al.35 — 3 3 3 3** 3 3 — 3
Ghiasi et al.28 — 3 3 — — — — — —
An et al.29 3* 3* 3* — — — — — —
Moughty and 3* 3* 3* — — — — — —
Casas30
Kerle et al.31 — — 3 — — 3* UAV only — —
Khan and — 3 3 3 — 3* — — 3
Yairi32
Ye et al.33 — 3 3 3 3** 3 — — —
Hou et al.34 — 3* 3 — — — — — 3*
Salehi and — 3 3* 3** — 3** — 3** 3
Burgueño36
Sun et al.37 — 3 3 3* — — — — 3
Azimi38 — 3 3 — 3** 3 3** — 3
Avci et al.39 — 3 3 — — — — — 3*
Tibaduiza 3* 3 — 3* — — — — —
Burgos40
Hou and Xia41 3 3 3 — — — — — 3
Sony et al.42 — — — — 3* — UAV only 3** 3*
Alavi and — — — — 3 — — 3* 3
Buttlar43
Our survey 3* 3 3 3 3 3* 3 3 3
— indicates not considered. 3 indicates fully discussed. 3* indicates partially discussed. 3** indicates an insufficient discussion or was discussed briefly in
the open research issues.

the features of the current IoT era, such as cloud (edge) ML/DL enhancements” section, three next-generation SHM
computing or Industry 4.0, such as digital twin modeling applications are briefly summarized. Finally, an open re-
and blockchain, is the next stage of SHM. In light of these search discussion for the future of SHM and ML is provided
achievements, this article is organized such that: the “SHM in the “Open research issues” section, and the “Conclusion”
and machine learning, a detailed overview” section exhibits section concludes the review.
a detailed overview of the ML and DL pipeline in each
component-level of SHM; ML learning algorithms are SHM and machine learning,
discussed in length in the “ML-supported pattern recog-
a detailed overview
nition techniques” section and connection to each level of
damage identification in SHM is also provided; in the “IoT- Ubiquitous data is everywhere. Given the amount of data
related applications” section, related IoT, big data, and gathered from numerous possible sources, it is essential to
hybrid approaches applications in SHM are reviewed. understand the pattern that underlines it. Day by day, with
Similarly, in the “Next-generation SHM applications with the increasing complexity of structures and the sheer
Malekloo et al. 7

amount of data collected, without automatic (or sometimes to take. Figure 3 illustrates a statistical PR classification
semiautomatic) processes to discover patterns using com- model for a typical damage assessment scenario.
puters, such tasks would be infeasible and impractical. ML The subsections below describe the necessary proce-
is considered a tool to recognize/classify information based dures for any ML application, emphasizing SHM statistical
on a learned pattern through the use of different algorithms. PR. It should be noted that there are several methods
In general, ML algorithms are based on either (1) statistical, available in any of the following procedures; however, only
(2) neural, or (3) synthetic approaches. The first two the popular methods are expressed here. Figure 4 sum-
methods are generally considered as the main pattern marizes the necessary steps that an ML model for SHM has
classifiers for SHM.8 In detecting damages using ML, to go through, along with a few examples of the techniques
initially, a pattern class or category is defined. For SHM, one used in the literature. As shown in the figure, the first seven
establishes training data through which all the attributes steps are introduced in this section, whereas step 8 is
defining the structure are gathered (sensing). At this stage, presented in a more detailed manner in the “ML-supported
depending on the collected data, class labels may or may not pattern recognition techniques” section. The material pre-
be assigned to the data. These data are then pre-processed to sented here offers the readers the chance to discover some of
remove any noise or outliers and to reduce the dimensions the common ML methods and techniques utilized in the vast
of the damage vectors (pre-processing). The next stage is body of SHM literature. Methods and techniques are in-
feature extraction. At this step, damage-sensitive features troduced and summarized such that, in essence, the readers
are selected either based on engineering judgment or can prioritize and quickly grasp how each part of the ML
mathematical and transformation procedures, or a combi- process is implemented in SHM systems specific to their
nation of both. Postprocessing may also be applied after needs.
feature extraction to further compress, normalize, or fuse
data as needed. After these stages, an algorithm is used to
Excitation methods
identify the damage state using one or more of the following
techniques: The very first step in detecting the existence of damage is to
excite the structure in place. Bridge condition monitoring,
1. Classification: Discrete class label (damaged/ seismic performance assessment of bridges, and verification
undamaged) of the numerical models with the measured data are some of
2. Regression (location of the damage, size of the fa- the factors, especially in vibrational-based applications, that
tigue crack, etc.) require careful consideration of the method of excitation. In
3. Novelty/outlier detection general, there are two ways to achieve dynamic excitation
of civil infrastructures,45 (1) ambient excitation and
Finally, from the processed output data, one can deter- (2) measured-input test. Ambient excitation is suitable for
mine, if necessary, whether a decision has to be made to real structures, whereas measured-input test is more often
rectify the identified damages and which subsequent actions restrained to a laboratory experiment. Damage is considered

Figure 3. A general statistical PR classification model.

8 Structural Health Monitoring 0(0)

Figure 4. A summary of the necessary and optional steps and techniques involved in a data-driven SHM system with ML augmentation.

a local phenomenon; however, local excitations tend to have advantages and disadvantages of these excitation methods
little to no effect on low-frequency global response. Par- can be found in Farrar et al.45 and Maas et al.47
ticularly in large-scale structures, ambient excitation is the
only source that is capable of providing the necessary inputs
Data acquisition
that lead to identifying damage in terms of global behavior.
However, the problem with such excitations is that, unlike It is well understood that without sufficient and accurate
local excitation, they are nonstationary, and the variabilities data, a clear understanding of the damage-sensitive features,
of their inputs need to be taken into account. Thus, for excitation methods, types of sensors, and, lastly, sensor
small-scale structures, measuring devices such as piezo- configuration, SHM may not reveal the optimal informa-
electric materials that can act as both force transducers and tion. For example, the method of global damage detection
actuators can simplify the first two steps of implementing an through standard accelerometers, which often are described
SHM system. as continuous measuring devices, differs from other sources
Dynamic excitation of the civil structures, such as of data collection such as strain sensors, a local measuring
bridges, can be achieved with (1) the ongoing vehicle or instrument, where an average data over a short period is
pedestrian traffic on the structure, (2) ambient wind and gathered. In other words, monitoring of dynamic parameters
waves excitation, and lastly, (3) seismic ground motion, that may imply different needs compared with the static ones.
is, earthquake or micrometer excitation. These types of Thus, it is essential to identify how the data can be collected
excitations are output-only modal-based analyses that can and utilized. Therefore, the performance indicators of any
be used to estimate modal parameters such as mode shape or sensors, as listed below,48 need to be scrutinized.
resonance frequencies. In terms of local excitation, popular
methods are (1) shakers with a variety of input patterns and 1. Sensitivity
frequencies, (2) direct impact to the structure at the point of 2. Dynamic range
interest, and nowadays, (3) input–output sensors capable of 3. Accuracy
producing known input forces with high signal to noise ratio 4. Linearity
(SNR). Often, in these input–output methods, given the 5. Stability
nature of the structure, disruption of traffic or on-structure 6. Response time
activities are followed. Therefore, it is not practical to
implement such excitation methods on a large structure. In addition to the items listed, two essential elements of
However, in exceptional cases and depending on the extent data acquisition: (1) the number of sensors and (2) their
of the damage, it may be necessary to perform both exci- locations must also be optimized. Redundant and unnec-
tation techniques in a combined approach.46 Further essary data would burden the data acquisition system and
Malekloo et al. 9

hinder the subsequent processes of the SHM system. EOFs. Examples of standard active sensors are piezoelectric
Moreover, these sensors are typically permanently installed, ultrasonic sensors which can be utilized either as an
and consequently, any unsystematic approach to designing impedance-based method54,55 or Lamb wave-propagation
a data acquisition system would introduce additional method.56–59
challenges. ML adaptations in data acquisition systems are generally
Traditionally, wired-based sensing equipment was observed in sensor layout optimization. The goal is to use a
widespread over other means. Especially for important minimal number of sensors while ensuring maximum
structures and long-term monitoring, expensive wired- damage sensing capabilities or, as referred by Farrar et al.,10
based systems are still preferred over wireless ones.21 “maximizing damage observability.” A range of input
However, with the advancement of technology in many sensors at different locations is trained, and based on the
aspects of WSNs (size, energy storage/generation, etc.), most relevant feature an optimal sensor position is located.
they are becoming increasingly attractive to researchers Recent studies tackled this daunting task. In the paper by
as they provide portable, practical, and efficient Bigoni et al.,60 the authors utilized a sparse Gaussian
alternatives.49,50 Generally, sensors are categorized into process and one-class support vector machine (SVM) for a
passive or active sensors. Usually, a combination of the reduced model. The most widely used ML algorithms in
sensing technologies is used for SHM solutions. A detailed sensor optimizations are genetic algorithms (GA) in coor-
review of the currently used SHM sensors and next- dination with neural networks (NN),61 as shown in Table 3.
generation smart sensing technologies is given in Different computational methodologies in optimal sensor
Moreno-Gomez et al.51 and Sony et al.42 Additionally, placement along with the advantages and disadvantages of
Abdulkarem et al.52 discussed the state-of-the-art WSNs in each algorithm are reviewed in Bigoni et al.62 and Soman
SHM from a different perspective such as academic and et al.63 In practice, ML models are typically adapted to
commercial wireless platform technologies. reposition the existing sensor locations to increase overall
Passive operation of sensors is described as a mea- system performability in terms of damage detection. It
surement that inflicts no input energy to the structure. offers simple, adaptable, and low-cost solutions over tra-
Accelerometers, strain gauges, and acoustic emission are ditional approaches.
examples of this type of sensor. They only detect damages
with no interaction with the actual structure. Usually, these
Data normalization
nonstationary sensors cannot precisely determine the dy-
namic response of the structure53 that could otherwise be Some of the collected data for learning is generated by
due to EOFs. Additionally, in the early stages of SHM numerical simulations, which often disregard EOFs (e.g.,
applications, passive sensors had difficulties directly temperature and traffic load, respectively). Many re-
identifying the damage. It relies on variable ambient ex- searchers investigated these nonstationary sources of var-
citation, which may or may not output the desired data to be iations. Several works15,19,72–74 have shown that the
evaluated. dynamic performance of bridges varies significantly de-
On the other hand, very similar to the NDE approach, pending on the condition that the bridge is subjected to
active sensors localize the excitation tailoring the overall daily. The effect of these trends on damage-sensitive fea-
damage detection process. The main advantage over passive tures could be removed by utilizing different linear and
systems is that, with known excitation force and location, it nonlinear correction models. Data normalization at the first
is much easier to detect damages and minimize the effect of stage tries to bring every data at a common scale since the

Table 3. A few examples of sensor layout optimization algorithms for SHM systems.

ML algorithms References Advantages Disadvantages

61 63
GA Banik and Das, Soman et al., Gomes Blind search method Premature convergence with larger search
et al.,64 Yang et al.65 Highly parallel space
Global optimum and
discreteness
NN Banik and Das,61 Blanloeuil et al.,66 Mallardo Nonlinearity Unexplained behavior
and Aliabadi67 Robustness Over-training
Fuzzy information
Bayesian Flynn and Todd,68,69 Capellari et al.70 Easy to implement Computationally expensive
models Easy to interpret Assumptions about prior
Others Bigoni et al.,60 Semaan71 - -
10 Structural Health Monitoring 0(0)

data from various sensors are inconsistent depending on the Data cleaning
size and location of the damage. The simplest way is the Z-
score normalization as shown in equation (1), where x One should not expect that collected data from sensors are
represents the original feature vector, x represents the mean always satisfactory and up to par. Loosely mounted sensors
of the feature vector, and σ represents the standard devia- or external effects can reduce the quality of the data. Data
tion. The numerator is also known as DC offset filtering cleaning, in simple terms, refers to implying hard limits in
which the data is not usable and must be discarded (noise/
xx
x0 ¼ (1) outlier treatment) or missing data to be imputed.37,79 In the
σ context of SHM and big data, five distinct quality standard
The next stage is to take into account the EOFs. De- indicators,80 namely: (1) availability, (2) usability, (3)
pending on the presence of the variabilities, either the reliability, (4) relevance, and (5) presentation-quality, must
damage-sensitive features are parameterized as a function of be taken into consideration before any SHM im-
the EOFs (measured and known variables) and later com- plementation. Unification of data ensures the efficiency
pared with a new set of extracted data, or they are developed and accuracy of an ML algorithm. In SHM, typical data
indirectly with the help of ML algorithms. Table 4 shows cleaning is performed via either software or hardware
some examples in dealing with EOFs in linear and nonlinear filtering methods.81 Embedded in data acquisition devices,
behavior for data normalization. To achieve full data-driven noise rejections with low/high bandpass filters, resam-
analysis, EOFs have to be analyzed in a data-driven manner pling, or other techniques can be employed. Unusable data
as well. However, certain conditions, such as long-term can be decimated, and missing data can be statistically
behavior of EOFs, lack of data, and insufficient data nor- imputed, if necessary. Due to the requirement of com-
malization, necessitate the usage of model-driven analysis. pression for big data, in many applications, data recon-
In hybrid use-cases, the FEM of a structure is used to struction using ML algorithms is utilized to make
generate damage scenarios resulting from EOFs. These incomplete data (irregular) to corresponding complete data
scenarios are then fed to an ML algorithm to classify the (regular). DL applications82,83 in data imputation work
existence of the damage better. Moreover, to validate the great for categorical and non-numerical features such as
results, damage data from FEM are used as test data. the case in SHM. However, they are rarely used due to their
Minimizing the misclassification of both Type I and Type II slow nature when dealing with extensive datasets. Other
errors heavily depend on the model that is used to normalize methods such as k-nearest neighbors (kNN), stochastic
the data from EOFs75—Type I refers to a false-positive regression, extrapolation and interpolation, and many
indication of damage, whereas Type II is a false-negative others can be employed to correct or remove irrelevant
indication of damage. In their article, the authors’ nonlinear data. Recently, Tan et al.84 investigated the effectiveness of
ML algorithms performed better when compared with their multiple supervised learning methods (Ridge, RF, SVR,
linear counterparts (less than 3% of total error for combined MLP, and XGBoost) for data augmentation under different
Type 1 and Type 2), meaning it better illustrated the con- missing rates of the inputted database. All models were
formity to the nonlinearity of EOFs. It appears that the able to capture the missing trend when the missing data are
consideration of EOFs and their patterns appear the be uniformly distributed, and SVR and MLP performed best
nonlinear in nature. For example, freezing temperature can on average with root mean square error (RMSE) of less
heavily influence the natural frequency, and therefore, a than 2. Some examples of data cleaning and data recon-
linear correlation between the EOFs and damage cannot be struction techniques are presented in Table 5.
assumed. The best performance models, according to the Perhaps, one of the critical challenges that data cleaning
works, are those that utilize nonlinearly separable clustering faces is when the process is scaled to large and complex
techniques such as KPCA or GKPCA. structures.95 In the context of big data analysis, traditional

Table 4. Linear and nonlinear algorithms for EOFs.

EOF behavior Reference Models Performance indicator Best performance

Linear Figueiredo et al.20 FA, MSD, SVD ROC MSD

Figueiredo and Cross76 PCA, MSD, GMM ROC GMM
Nonlinear Flexa et al.77 AANN, KPCA GKPCA, ROC KPCA GKPCA
Santos et al.75 Figueiredo et al.20 + SVM, SVDD, KPCA, GKPCA ROC KPCA, GKPCA
Figueiredo et al.20 AANN ROC —
Figueiredo and Cross76 NLPCA ROC —
Oh et al.78 KPCA N/A —
Malekloo et al. 11

Table 5. ML-based data decimation and imputation techniques.

Reference Decimate/Impute Method(s) Application

85
Ren et al. Impute Bayesian tensor learning Strain and temperature records of a concrete bridge
Chen et al.86 Impute Kernel regression Strain responses between sensors in SHM
Chen et al.87 Impute LQD-RKHS regression Probability distributions of missing SHM data
Martinez-Luengo et al.88 Impute ANN Offshore wind turbine
Oh et al.89 Impute CNN Strain structural response
Fan et al.90 Impute CNN Recovering lost vibration data
Li et al.91 Impute LSTM Stacked DL-based imputation framework for dams
Fan et al.92 Decimate Residual CNN Denoising SHM vibration data
Yang et al.93 Decimate/Impute 1) Chebyshev inequality Wind turbine
2) SVR
Batista et al.94 Decimate/Impute Ten different methods Balancing training models

data cleaning working sequentially cannot easily be applied The most significant limitation in ML algorithms is when
to ever-growing complicated structures.96 Thus, the parallel they are used to learn from high-dimensional data vectors
execution of any method should be in line with the five big with limited exogenous variables.98 Data compression
data quality standards discussed previously. The data should not come at the cost of losing the ability to learn a
cleaning process is always performed before initiating the pattern. Without enough features extracted following
subsequent ML processes. However, it may also appear compression, it is not possible to deduce whether an al-
after extracting features as well. In an ideal situation, one gorithm can serve as a damage identifier. An exponential
data cleaning pipeline is enough assuming the selected increase in the dimension of the data would blow up
features, in the beginning, would lead to identifying enormously in the amount of training data needed to achieve
damages more efficiently. However, in reality, given the a reasonable and small error on the estimation; an issue
judgment of the engineer and the required outcome from the commonly referred to as Bellman’s curse of dimensional-
algorithm, it may be deliberately assumed that certain ity.99 Therefore, with inadequate training data, one cannot
features, although having passed the initial filtering process, achieve an ML implementation with a high degree of ac-
would not be helpful in determining damage. Therefore, the curacy. This effect, however, can be mitigated by im-
second round of cleaning, referred to as postprocessing, is plementing linear or nonlinear projection transformation of
carried out to ensure that individual decisions about the p-dimensional feature vectors onto a q-dimensional plane.
nature of the ML implementation conform to the model’s The most classical method used is the linear principal
output. An example of how postprocessing led to better component analysis (PCA). One might argue that, with the
damage identification was presented in the paper by Li addition of EOFs into the data vector, due to the nonlinear
et al.97 behavior of temperature and external loading, linear PCA
would not be a feasible solution to reduce the dimensions.
This was, however, proven to be false in real-life scenarios
Data compression as comprehensively analyzed in the study by Van Der
Structures equipped with SHM consist of tens or hundreds Maaten and Postma.100 In their paper, the authors found out
of different sensors. Each produces single or multi-feature the linear PCA works comparatively better than their
data continuously with various sampling rates ranging from nonlinear counterparts. With enough analysis and perfor-
around 10 Hz to 10–50 kHz. Over an extended period of mance measures of different variations of nonlinear data
monitoring, a multitude of data is generated, although not compression, future research can create better techniques to
every generated feature is usable in the analysis. Adding to identify nonlinear behaviors of structures as well as EOFs
this, EOFs would also play an essential role in increasing the that can lead to better data compression. Some recent ex-
features’ dimensions.8 In this regard, data compression, or amples of such techniques utilizing ML algorithms are
simply put, dimensionality reduction of the features, allows introduced.77,101,102 Figure 5 depicts how dimensionality
only the most statistically significant and damage-sensitive reduction models are categorized. In particular, to reflect the
features to be extracted. One way to tackle this is fusing nonlinearity of data, there are two approaches where
sensor arrays extracting similar features such as different the high-dimensional data is transformed onto a low-
mode shapes that are collected at each sensor node that is dimensional space work.103 In a local perspective, the lo-
later compressed to produce a low-dimensional feature cal geometry of data is preserved, and the model attempts to
vector containing only the first few mode shapes. map nearby points on a set of closely related points in the
12 Structural Health Monitoring 0(0)

Although, it needs to be stated that these free parameters are

actually advantageous in various cases as they promise
flexibility in the reduction process. The other issue of non-
parametric techniques is in the computational process since
they require more data for better performance. One of the
important requirements in dimensionality reduction models
is the out-of-sample extension abilities. Simply, the out-of-
sample extension performs the reduction process on the
training set and applies the mapping directly on the test set
to lower the dimensions. Such a capability is very crucial to
SHM systems, as different signals can be embedded during
the monitoring process. This eliminates the need for re-
training the whole dataset to learn new mapping functions,
which is less computationally expensive. However, not
every dimensionality reduction technique contains out-of-
sample extensions (Isomap, LLE, etc.). The non-parametric
out-of-sample extension is therefore required for all non-
Figure 5. Categorization of dimensionality reduction models. linear models. The approximation in the out-of-sample
extensions leads to an estimation error,104 so great care
vicinity. In the global approach, in addition to the mapping must be taken in these cases. The out-of-sample extension is
of nearby points, faraway points are also mapped to far- not one of the deciding factors in selecting dimensionality
away points, essentially keeping the geometry intact in all reduction models in the SHM application. In fact, only two
scales. studies, Langone et al.,105 and Liu et al.,106 have directly
Typically, there exist three methods to achieve data considered this capability in their proposed systems.
compression: (1) linear transformation such as PCA, (2) The majority of the examples of dimensionality reduc-
nonlinear transformation such as NPCA, and (3) au- tion techniques used in the SHM domain do not provide a
toencoders, where it is a nonlinear transformation based definitive reason as to why one technique was preferred over
mostly on ANN. Table 6 exhibits the most recent utilization the other, except those that research and compare different
of widely used dimensionality reduction models in litera- methods for reducing dimensions and filtering unwanted
ture, emphasizing SHM applications. From reviewing the data. Therefore, it becomes increasingly difficult to rec-
literature, it was found that most SHM applications are ommend one technique. Based on the authors’ observation,
utilizing PCA, LDA, QDA, and ICA in the linear approach linear and nonlinear variations of PCA can yield, in most
and NPCA, LLE, and AANN in the nonlinear approach for cases, an acceptable level of data compression. Nonlinear
dimensionality reduction. and clustering approaches are mostly preferred in situations
There is no clear cut between which dimensionality where the data is highly irregular in nature, such as EOFs.
reduction model is suitable. The variability in the nature of Having said that, other approaches can produce better re-
the collected data and the size and resolution of the input sults as this is highly dependent on the dataset and the
signals greatly influence the system’s overall performance. available computational resources. Interested readers are
For example, very few techniques (PCA and its variants, referred to the recent comparative study of dimensionality
autoencoders, etc.) are parametric; that is, there is a direct reduction techniques by Ayesha et al.107 for more detail.
mapping from the high-dimensional to low-dimensional Finally, the considerations below and the remarks in Table 6
space. Therefore, it enables to verify, to some extent, can facilitate selecting an appropriate dimensionality re-
how much of the high-dimensional space was preserved duction technique:
during the space reduction process. Being that majority of
the other techniques fall in the non-parametric domain, it 1. Understanding that some techniques are supervised
indicates a disadvantage along with other problems with or unsupervised entailing their own limitations and
non-parametric modeling, such as the curse of dimen- considerations.
sionality. Moreover, the presence of free parameters in non- 2. Deciding which dimensions to retain and the reali-
parametric techniques (learning rate and the number of zation and comprehension of the reduced model and
iterations), which can impact the cost function in the new dimensions.
nonlinear convex optimization, introduces another burden. 3. Recognizing that these dimensionality reduction
The performance of the dimensionality reduction technique techniques can sometimes negatively impact the
depends on the optimization of the free parameters. performance of the classifier.
Malekloo et al. 13

Table 6. Widely used dimensionality reduction models in the literature.

Technique References Model Remarks

76 106
Linear Figueiredo and Cross, Liu et al., Güemes PCA Highly interpretable
et al.,108 Roberts et al.,109 Mei et al.,110 Li Requires data standardization
et al.,111 Akintunde et al.,112 Deraemaeker and Assumption of orthogonality
Worden,113 Kullaa,114 Garcı́a-Macı́as and
Ubertini115
Huang et al.,116 Li et al.,117 Zhu et al.,118 Yao ICA Assumption of statistically independent and
et al.119 normally distributed variables order of the
independent component is difficult to be
determined
Zhang et al.,120 Avendaño-Valencia et al.,121 Hu PCR Perform well for highly correlated and colinear
et al.122 data
Imposes constraints on the coefficients of
nonrelated explanatory variables
Jiménez et al.,102 Yanez-Borjas et al.,123 Mboo and LDA Interpretable
Hameyer,124 Mangalathu et al.,125 Zheng and Small sample size
Qian,126 Mishra et al.127 Requires normal distribution assumption
Mangalathu et al.,125 Mishra et al.,127 Jiménez QDA Worse performance than LDA for large datasets
et al.102,128
Mishra and colleagues127,129 RDA A comprise of LDA and RDA model
Noel et al.,21 Figueiredo et al.,20 Deraemaeker and FA Unable to produce a meaningful pattern for
Worden,113 Garcı́a-Macı́as and Ubertini115 unrelated explanatory variables
Guo et al.,130 Zhao and Huang131 SFA Guaranteed optimal solution
Mapping is provided as functions directly
Nonlinear Figueiredo and Cross,76 Dervilis et al.,101 Li NLPCA Does not require the a priori specification of a
et al.,132 Hsu and Loh,133 Ye et al.,134 Tibaduiza time series
et al.,135 Silva et al.136 Performs poor in very large datasets
It incorporates nominal and ordinal variables
Borate et al.,137 Yang et al.138 PPCA Extends the scope of conventional PCA
It enables uncertainty assessment of the model
Fuentes et al.,139 Jeong et al.140 Isomap Preserves “true” relationship between data
points
Preserves the global data structure
Computationally expensive sensitive to “noise”
examples
Liu et al.106 Laplacian No local optima
Eigenmaps Less geometrically intuitive
Zhang et al.,120 Sun et al.,141 Chaabane et al.142 Partial least Can handle multicollinearity
squares Lack of model test statistics
regression
Flexa et al.,77 Oh et al.,78 Ghoulem et al.143 KPCA and its Mapping function does not need to be known a
variations priory
Choice of the kernel and multiple refitting are
required
Dervilis et al.,144 Xiao et al.145 Nonlinear ICA Back projection/reconstruction can be
and its implemented
variations More complex than ICA
Dervilis and colleagues,146,147 Sun et al.148 LLE Accurate in preserving local structure
Less accurate in preserving global structure
difficulty on non-convex manifolds
Flexa et al.,77 Garcı́a-Macı́as and Ubertini,115 AANN Mapping function does not need to be known a
Nguyen et al.,149 Zhang et al.150 priory
High computational complexity
Autoencoders Ma et al.,151 Wang et al.152 DA Can find different levels of features
Liu et al.,106 Mboo and Hameyer124 SA Can be inefficient for massive data
14 Structural Health Monitoring 0(0)

Feature extraction/selection illustrates the three different approaches in feature selection/

extraction.
Identifying damage-sensitive features from the collected Other than data-driven and model-driven techniques,
data is not a trivial task. Not every quantity is significant in feature extraction can also be done using wave-propagation
indicating the presence of damage, nor do they correlate in and impedance-based methods. They are a subset of data-
any way that leads to detecting damage even with the most driven techniques, but due to the unique devices and their
advanced ML algorithms. Feature extraction is a process specific extraction methods, they can be quite different
that enables the transformation of the collected data to a relative to data-driven and model-driven methods. How-
form that is more identifiable and quickly picked up by any ever, there have been many studies in the past that enhanced
simple ML algorithm. The most critical aspect of this step of the capabilities of these systems with ML and DL. A review
any ML-based SHM implementation is finding ways to of the guided wave-based SHM is provided by Mitra and
extract and select sensitive features that positively correlate Gopalakrishnan.153 Extraction of damage-sensitive features
to damage. The challenge in this regard is that the extracted in data-driven approaches can be carried out by using (1)
features may also be vulnerable to changes in the system’s time-domain, (2) frequency-domain, (3) time-frequency
response that do not necessarily relate to damage. Figure 6 domain, or (4) ML algorithms. In classical time-series

Figure 6. Three different approaches at the feature extraction step.

Malekloo et al. 15

analysis, the variety of changes in the structures can be fitted updating process can be developed to provide a better
to a model that identifies damage. Low-order (dimension) representation of the system from both a global and local
time-series modeling, techniques such as autoregressive perspective.169 One of the benefits of non-modal-based
moving-average model (ARMA),154 cross-correlation FEM updating is that inherent to the classical modal
function (CCF),155 and stochastic subspace identifica- FEM updating, the strict assumption that structures must
tion (SSI)156,157 can be proven useful in extracting highly exhibit linearity, reciprocity, and time–invariant properties
damage-sensitive feature vectors. In scenarios where the can therefore be lifted and eased for the systems that show
dimension of the vectors is high, using high-order time- nonlinear and nonstationary response.
series modeling to capture the variations may result in After extracting feature vectors that correlate to damage,
fitting the noise in the collected data. one must select the most appropriate feature(s) to be trans-
In these cases, extracting features in the frequency do- ferred to the damage detection process. Though previously an
main can be used. Methods such as power spectral density implicit explanation of selection procedures was given, de-
(PSD),158 cross spectral density (CSD),159 impulse response termining a subset of collected data for detection purposes
function (IRF),160 frequency response function (FRF),161 can be established from mathematical models or based on
and frequency domain decomposition (FDD),162 are few of intuitive and engineering judgment. An overview of feature
the methods that are frequently used. New techniques other selection techniques is given in the studies by Khalid et al.168
than the two described above are being developed and tested and Chandrashekar and Sahin.170 It should be noted that
to identify damage-sensitive features better. These methods while both feature extraction and feature selection are the
are based on time-frequency domain, such as wavelet or techniques used to reduce features and eliminate redundant
phase shift in a linear and nonlinear fashion.163–166 A de- and irrelevant data, the contrasting point between the two is
tailed walkthrough of the wavelet technique is provided in that the former creates a brand-new set of data. In contrast, the
the paper by Taha et al.167 The constraints in time-domain, latter creates a subset of the original data. This way, there is
and frequency-domain waveform analysis rises in signifi- no clear boundary between them, and it ultimately boils down
cance as the dimension of the data of these output-only to the application domain and the system’s requirement.
models increases. Moreover, they do not precisely indicate Moreover, feature selection and extraction can work co-
the location of the damage and require a high quantity of herently and synergistically to inform one another in the form
data for sensitivity analysis as the reproducibility of the of change detectors.171 Few ML algorithms, such as Lasso or
models in different time frames are inconsistent when random forest (RF), have built-in selection algorithms that
factoring the EOFs. These shortcomings can be overcome fall under the category of the embedded selection method.
by incorporating ML algorithms with their inherent features Filter and wrapper methods are the other two selection
in extracting data (dimensionality reduction) or feature procedures.172 Filters are employed in producing a most
selection (filter, wrappers, and embedded).168 promising subset before passing on to the damage detection
Model-driven techniques, on the other side of the process as part of the pre-processing step. Whereas in the case
spectrum, solely depend on accurate physical models to of the wrapper, as the name suggests, the selection procedure
identify damage. They usually fall into the model updating is “wrapped” into an algorithm that is trained based on a
paradigm (linear and nonlinear), where mathematical model. It either begins with no features and with each iter-
models with input–output or output-only measurements are ation, features are added that translate to the best performing
used to identify modal parameters (e.g., mode shape, mode model (forward selection). Alternatively, the model starts
shape curvature, and resonance frequencies) given an un- with all features intact, and at each iteration, the least sig-
known system and then calibrate the physical parameters. nificant feature is eliminated, thus improving the overall
When a high fidelity physical model of the system is es- model performance (backward elimination). Lastly, a form of
tablished, numerical analysis such as FEM can be used to the greedy optimization algorithm can be used to rank the
update the initial model on the grounds of the identified feature subset at each iteration (recursive feature elimination).
parameters from the system. Model-driven techniques, al- To put it simply, these wrapper methods are used as a search
though proven useful in scenarios where sufficient data is strategy, and the performance of each is dependent on the
unavailable, would become challenging when EOFs are quality of the given algorithm.
incorporated into the FEMs as described earlier. In an at- Additionally, these approaches can be combined to
tempt to develop accurate feature extraction methods, a form an ensemble learning method built on the output of
hybrid model of both data-driven and model-driven tech- different algorithms or learners and is believed to produce
niques is used that could overcome the deficiency of each a better selection process.173 Though not many cases of
approach. The two systems can be used jointly to validate ensemble feature selection are present in SHM, the pro-
the presence of damage, or the model-driven system can be liferation and advancement in this recent field can enable a
exploited to generate training/testing data for the data- selection of features from large and complex structures that
driven system. Through this, a non-modal-based FEM comprise different sensors and measurement devices with
16 Structural Health Monitoring 0(0)

high-dimensional datasets. Moreover, many feature selec- sensors are subject to spatial and typological variation.
tion techniques are based on supervised methods of Multisensory systems are receiving increasing attention
searching for the best subset of features. With no prior since they provide a spectrum of advantageous features.
knowledge about the collected features, the risk of over- Higher SNR, higher data resolution, data redundancy, and
fitting the selected features can be reduced if unsupervised complementarity and timeliness are some of the potentials
or semi-supervised methods can be used instead.174 Similar of such systems.8,195 Additionally, the spatial distribution of
to the supervised models described earlier, unsupervised these sensors can enable engineers to increase the SHM
feature selection techniques essentially follow the same system’s observability. Additional information collected
routine. A recent review of unsupervised feature selection from different sensors located on the structure can lead to
methods is provided by Solorio-Fernández et al.174 In the enhancement in the identifiability of the SHM system. Data
context of SHM and ML, Bull et al.175 utilized an unsu- fusion is a technique of combing information in such a way
pervised feature selection procedure in ensemble analysis, that the aforementioned benchmarks, that is, observability
specifically bagging and feature bagging. An unsupervised and identifiability, would considerably improve the system
filter method, namely unsupervised minimum redundancy– performance.196–198 As it can be derived from the norm of
maximum relevance (UmRmR), was used in the study by the data fusion process, the steps taken to ensure high
Zugasti et al.176 to minimize redundant and irrelevant data system performance are correlated to data normalization as
from offshore wind turbines. Based on our analysis, un- well as data cleaning techniques that were discussed in the
supervised feature selection has not received enough at- prior sections. Similar ML techniques can be used in the
tention. As discussed in this section, classification based on same way, and interested readers are recommended to read
unsupervised learning is one of the hot topics in damage the extensive review of the data fusion process by Wu and
identification in civil infrastructures through ML and SHM. Jahanshahi199 for ML-SHM conjunct review.
Table 7 shows some of the supervised feature selection There are, in general, two ways to amalgamate data.
techniques used in the literature, emphasizing SHM. Multiples of a single chain of data processing scheme, for
example, steps 1 to 6 in Figure 4, are fused to have an
arbitrary number of chains with each processing their data
Data fusion individually and feed-forwarding their outputs to a PR unit.
As discussed previously, structures comprise a plethora of In the other type, a centralized batch processing system
sensors. In order to perform global damage identification, accepts the chains of the sensor up to the data compression

Table 7. Supervised feature selection techniques in the literature.

Feature selection method Model Application and reference

Filters Pearson’s correlation (i) Sousa Tomé et al.177

(ii) Monaco et al.178
Mutual information (iv) Trendafilova et al.179
(v) Fang Qian and Gang Niu180
(vi) Zhao et al.181
ANOVA (v) Sbarufatti,182
(ii) Kessler and Agrawal183
mRMR (i) Babajanian Bisheh et al.184
ReliefF (i) Babajanian Bisheh et al.184
(vi) Yan and Jia185
Chi-squared (iii) Hoell and Omenzetter186
Wrappers Forward selection (i) Babajanian Bisheh et al.184
(i) Yun et al.187
(v) Park et al.188
Backward elimination (i) Babajanian Bisheh et al.184
Recursive feature elimination (v) Zhou et al.189
Embedded SVM-RFF (i) Babajanian Bisheh et al.184
(v) Mustapha et al.190
Lasso (i) Ni et al.191
(iv) Mustapha et al.192
(iv) Michau et al.193
Ridge (vi) Ma et al.194

Note: (i) Bridge; (ii) Composite plate; (iii) Wind turbine; (iv) Simulation; (v) Experimental testbed; (vi) Mechanical machinery.
Malekloo et al. 17

unit, for example, step 5 in Figure 4, then feeds all the damage-sensitive features fed into the PR unit have to
collected data to a feature extraction unit and finally passes correlate to the choice of the algorithm used to learn from
them to a PR step. It is also possible for data fusion to be the features. In general, there are four different types of
applied at each level individually or collectively if desired.8 learning in the domain of ML, as depicted in Figure 7.
Raw sensor data can be collected from multiple units and In the framework of SHM and statistical PR, the most
combined to produce a more uniform data set (initial fu- common learning algorithms are supervised, unsupervised,
sion). It may also be applied in the same way as the data and semi-supervised. In rare scenarios where both damaged
normalization or data compression level, meaning the and undamaged data of a structure is available for engi-
feature vectors can be merged to produce a single vector neering structures, supervised learning is the preferred
(feature-level fusion). Furthermore, many feature vectors learning method. In this case, group classification and re-
can be passed to a single PR algorithm (pattern-level fu- gression analysis are the primary methods of supervised
sion). Lastly, damage classifiers resulting in the health state learning. However, for larger and more complex civil en-
of the structure can be fused to provide a high degree of gineering structures, unsupervised learning is required due
damage identification confidence level (decision-level to the lack of damaged data.20 This is regarded as the go-to
fusion). method for most civil infrastructures, such as bridges, where
there is limited availability of damaged data from the
ML-supported pattern structure, or it is found not feasible to collect data on a
global scale in the first place. In this context, the unsu-
recognition techniques pervised method is commonly referred to as novelty de-
The prior section described the seven initial steps that are tection or outlier detection.14,200–202 A more recent learning
necessary for implementing a data-driven SHM system. As method, semi-supervised learning, has been introduced
it was previously shown in Figure 4, the last step of an ML- where data label from damaged and undamaged states of the
augmented SHM system is the PR, that is, the identification system is partially available—a very common occurrence
of the health state of the structure. The 5-step damage with engineering structures.203
identification hierarchy was introduced at the beginning of It is essential to understand the criteria for choosing an
this article. More specifically, in the context of statistical PR, algorithm. Specific to the problem and the type of data
to accurately classify damage, one needs to ensure data available, different algorithms, or even a combination of
availability from either damaged, undamaged, or even both algorithms, are required, for example, the effect of EOFs on
conditions. Additionally, in order to be able to assess and identifying damages. The number of training data available,
predict the damage caused to the system, the selected the expected reasonable training time, and the amount of

Figure 7. Four different ML learning algorithms.

18 Structural Health Monitoring 0(0)

accuracy required from a learner are some of the consid- Decision tree (supervised)
erations that must be meticulously thought about before
selecting an algorithm. Moreover, in a simple SHM for Decision tree (DT) is a well-established learning method
small-scale systems, the number of parameters and fea- capable of partitioning datasets from a non-parametric point
tures is typically low. However, for many of the im- of view. They are capable of targeting categorical variables
plemented SHM systems, these two criteria can ultimately (classification: damaged/undamaged) and continuous vari-
make one algorithm superior to the other. Various papers ables (regression: source signal comparison with the healthy
have shown the different cases of these ML algorithms in state of the system). A tree starts with the root node rep-
their reviews.36,37 However, a clear cut between the uti- resenting the input feature(s), such as acceleration data.
lization of different learners is somewhat hidden in the Given a threshold set by the algorithm, the root node is
core of their review strategy. We, therefore, aim to expand partitioned into many child nodes (internal nodes). The
on why such leaners are chosen as the method of im- segmentation of each node is based on the node that results
plementing a data-driven SHM system. Also, several new in the most significant information gain, called purity. The
methods specific to the deep NN are introduced. When- process is repeated until the last node (leaf node) is reached
ever available, the authors will refer the readers to a such that the node becomes impure. In the context of SHM,
particular review paper about each type of ML algorithms. the input feature at the node could be wind direction and
In the subsequent sections, some of the widely used ML wind speed, and the testing attribute for splitting each node
algorithms are explained. The implementations of each is the observed PSD of the measured vertical acceleration.
method provided are primarily from mid-2019 onwards. This example was demonstrated by Li et al.204 (Level 1) as a
In order to realize how each implementation relates to means of classifying vortex-induced vibrations that may
different SHM damage identification levels, a snippet result in long-term fatigue damage of long-span bridges
next to each cited reference is provided, indicating the subjected to crosswinds. In their methodology, the root node
SHM level (e.g., Levels 1–3). Furthermore, a separate was assumed to be wind speed.
section is dedicated to SHM level 5 (damage prognosis), However, in cases where N multiple damage-sensitive
where the recent ML-supported applications are sum- features are present, selecting the root node and the internal
marized. Finally, based on the reviewed papers, a rec- nodes is not trivial. A random selection can undoubtedly
ommended SHM system with the best examples of ML lead to poor results. For such cases, there exist many sta-
and DL for each SHM component up to damage prognosis tistical attribute selection measures that can be used to solve
stage is discussed at the end of the section. A summary of this issue. Entropy, Gini index, and Chi-square are some of
the most common learning algorithms is depicted in the ways to achieve selecting the most feasible starting root
Figure 8. node and internal nodes. Typically, in the SHM system, DT

Figure 8. Three common learning algorithms in SHM.

Malekloo et al. 19

learning is of the classification type due to the large high- and DT. However, the input–output model of a slender
dimensional feature vector and stochastic behavior of coastal bridge presented by Lu et al.209 (Level 1-2,4) for
structures given ambient excitation. The study by Gordan fatigue damage assessment due to EOFs showed the ran-
et al.205 (Levels 1-2,4) on a slab-on-girder bridge showed domness behavior of RF learner with a high degree of
that the classification and regression tree (CART) learning variability (high standard deviations) when compared
method is incapable of competing with AI algorithms such strictly against the regressor function of SVM and Gaussian
as ANN due to lack of capacity, flexibility, and complexity. process (GP). Relatively speaking, a higher amount of
Another area that decision trees can be applied to is in training data and a better choice of the kernel would lead to
combining other ML models such that a higher accuracy can better results for SVR as it is susceptible to the initial pa-
be achieved sequentially by traversing down the tree by rameters. This could explain the reason why these two
assigning a model as a function of the input variables of the papers appear to contradict. Chencho et al.210 (Level 1-2,4)
preceding node.200 Zhang et al.206 (Level 1) demonstrated developed a structural damage quantification based on RF
that for real-time visible fatigue crack growth detection with and PCA for dimensionality reduction. The authors
computer vision, DT performs the best when compared with achieved an R-score of 89.2 and 95.3% for single-element
RF, kNN, NB, RF, NN, and an ensemble model. They and two-element damage cases, respectively. Unlike DT, RF
concluded that other than DT, the rest of the ML algorithms is not at all interpretable, and for large datasets, they can take
used tend to overfit. Mariniello et al.207 (Levels 1-2) pro- a long time to train.
posed DT ensemble damage detection and localization
down to the single structural elements. The authors achieved
Support vector machine (supervised)
an accuracy score of 90% or more while recording limited
localization errors. As it can be observed, with the simplicity Perhaps, one of the most widely used types of ML algo-
of DT and flexibility it provides, damage detection and rithms is SVM, both in its classifier and regressor form. It
partial localization can be achieved. However, there is still can map both linear and nonlinear data to an n-dimensional
no research on the utilization of DT when EOFs are feature vector where the hyperplane separates features into
included. separate classes while maximizing the margin distance
The main problem with DTs in a high-dimensional space, between them. Kernel functions achieve the transformation
such as the cases in engineering structures, is the overfitting of the data into a higher space. As discussed before, the
issue. In this situation, the model memorizes the training accuracy of SVM highly depends on the choice of kernel.
data and returns an unrealistic representation of the domain Generally, in SHM applications where the features are both
leading to poor predictions when tested with new data. abundant in amount and exist in the high-dimensional
Although there are methods to overcome overfittings such domain, SVM usually outperforms supervised ML algo-
as cross-validation of a parameter204 or hyperparameter rithms when provided with a suitable choice of kernel
optimization, a common way to tackle this is to use indi- functions—which is not always trivial, as demonstrated in
vidual trees as an ensemble, referred to as random forest. the study by Lu et al.209 A polynomial kernel function was
used in Gordan et al.205 (Level 1-2,4) for damage identi-
fication of slab-on-girder bridge. The authors concluded that
Random forest (supervised)
SVM proved to be superior due to its capacity to perform
As explained in the previous subsection, RF is an ensemble high-quality predictions compared to a CART method. One
of DTs capable of solving regression and classification of the downsides of SVM is that although the increased size
problems. The issue with DT with extensive features can be of training data leads to better predictions, the training time
overcome by using RF, which works quite well with high- increases exponentially as well. In order to overcome this
dimensional sparse data. The main advantage of RF com- limitation, least-squares SVM (LS-SVM) was proposed.
pared to DT is that each tree is constructed from a random This method finds the solution by optimizing a set of linear
set of training data and the splitting of nodes also happens equations rather than the quadratic programming (QP)
for a random subset of attributes. In the end, a simple method used in SVM. An implementation of LS-SVM as a
average of all the predictions can be used to find the most hybrid model-driven and data-driven SHM system was
probable outcomes of the model. RF works best if the trees presented by Deng et al.211 (Level 1-2,4-5). In their paper,
are not strongly dependent on each other and show a weak the authors showed the daily fatigue damage due to traffic
correlation between the attributes selected at the splitting using the weigh-in-motion (WIM) and regressor function of
node. An example of RF was demonstrated by Laory SVM for highway suspension bridge hangers. SVM can
et al.208 (Level 1) They showed the effect of EOFs on the also enhance DL damage detection. In identifying damage
natural frequency of a suspension bridge, where they found using DL, one of the drawbacks of such a method is the
RF and support vector regression (SVR) to be more suitable misclassification of the input data of an unlearned pattern as
when compared to multiple linear regression (MLR), ANN, that of a learned pattern. A validation of this method using a
20 Structural Health Monitoring 0(0)

shaking table and simulated training data of a steel frame of measuring conditional probabilities of certain statements
structure using deep NN was established by Kohiyama given other statements, as shown in equation (2)
et al.212(Level 1). SVM was used to detect an unlearned
Prior × Likelihood
damage pattern based on the feature data of a DNN. An Posterior ¼ (2)
efficient and precise nonlinear multiclass SVM (NMSVM) Evidence
of nonlinear time-varying structures was proposed by Bayesian approach is widely used in SHM applications.
Chong et al.213 (Level 1). Their algorithms were trained They are either used alone or are integrated with different
using many wavelet-based autoregressive coefficients that ML algorithms such as Bayesian clustering. The Bayesian
were found from applying wavelet transform to signals method can bring about numerous benefits, including
generated from healthy and damaged structures under probabilistic inferences. For example, a pro-active SHM
random excitation. Other than the above, SVM can also be solution using FEM and Bayesian network was proposed by
used in other parts of ML procedures such as sensor Sousa et al.218 (Level 1-4). The authors developed a
placement optimization, data normalization, and feature monitoring solution based on numerical analysis of a real
extraction/selection, as discussed in the previous sections. bridge and achieved acceptable performance when com-
pared with actual damage data. Their pro-active tool
k-Nearest neighbor (supervised) managed to demonstrate the first four levels of damage
identification, which can provide useful information for
k-Nearest neighbor (kNN) is one of the earliest and simplest online bridge management and monitoring.
supervised learner methods. Similar to the previous algo- Naı̈ve Bayes (NB) classification is one of the methods of
rithms, it works well for regression or classification tasks. It Bayes’ theorem. In this method, it is assumed there is no
classifies the input training features based on their distance dependency between the features. In the study by Man-
from the testing set. kNN is based on the idea of similarities galathu et al.219 (Level 1), eight different ML algorithms,
in properties of features discriminated from the feature including NB, kNN, DT, RF, and others, were used to
space. The selection of how many neighbors to consider is a establish a classification model of seismic failure of RC
function of noise in the data. Low-dimensional feature space shear walls. It was found out RF had better accuracy while
requires less training data. In SHM cases where the di- NB classifier fell short and was ranked sixth. They stated
mension of the features is high, higher training data is re- that the low accuracy of their parametric methods, that is,
quired, which results in a computationally expensive NB, was because of the existence of a nonlinear decision
process. kNN algorithm with Euclidean distance mea- boundary between the failure methods. In a similar study by
surement with six neighbors for damage detection of a Mangalathu et al.125 (Level 1), a rapid seismic assessment of
scaled-down cable-stayed bridge was demonstrated in Li a two-span box girder bridge was analyzed based on
et al.214 (Level 1). The authors showed the performance of simulations from Open System for Earthquake Engineering
traditional ML algorithms, mainly DT, RF, SVM, and kNN, Simulation (OpenSees) platform. In their paper, the authors
against their proposed CNN model. It was observed that the evaluated their models using NB, kNN, QDA, and RF.
accuracy imbalance of kNN was the most severe compared Analogous to their previous paper, RF performed the best,
to others due to the lack of complexity of the learner. The while NB performed better while classifying the bridge as
study by Dogan et al.215 (Level 1-2,4) developed a model for unsafe. Although Bayesian analysis is found superior to
determining post-earthquake damage to RC columns. In the highly complicated learners, the main disadvantage is that
form of cracks, the damage was identified employing a since it assumes no interdependency between the features,
camera and evaluated against the allowable ranges from the the estimation of probabilities may not be accurate if the
building code. kNN, DT, SVM, and LDA were checked assumption does not hold.
against the ensemble of these algorithms. The success rate An extension of NB, Gaussian NB (GNB) was used in
of each one resulted such that the ensemble method was the study by Nazarian et al.220 (Level 1-2,4) for a turn-of-
found to be the best one, followed by kNN with a small the-century building structure that was damaged due to
margin of error. One area in that kNN works surprisingly settlement of its foundation. The authors employed FEM to
well is data imputation in SHM systems. Inherent in its generate stiffness and strain dataset and later train GNB in
algorithm, it looks for the closest data to infer the missing addition to SVM and NN algorithms to effectively find the
value as demonstrated in different studies.216,217 location and the severity of the damage in each structural
component. Out of the three algorithms, NN yielded better
results; however, GNB was not far off. Soyoz et al.221
Bayesian (supervised)
incorporated Bayesian updating into reliability estimation
Bayes analysis is a probabilistic parametric learning method of bridges through vibration-based SHM readings before
and is considered a statistical learning approach on the basis and after damage. While conditional independency in NB is
Malekloo et al. 21

imposed only on the selected features in the Bays nets, variables at each input neuron, the processing and the
Bayesian network, on the other hand, assumes an inde- calculation depends on variables at hidden and output
pendent relationship in every class. Although NB makes a layers(s), respectively. The model is trained with an error
simplified assumption, both can perform equally in many propagation algorithm. This method of training is consid-
scenarios, provided that both are used for inference pur- ered to be a supervised learning approach. In a recent study,
poses. Lee and Song222 (Level 1) demonstrated the Bayesian Hekmati Athar et al.228 (Level 1) experimentally collected
network approach for system identification in a numerical sensor data from both contact-based and contactless-based
example. Compared with FEM updating and maximum sensors to identify damage on a lab-scaled bridge through an
likelihood estimation, the authors’ approach provided a ANN model. In another work, Malekjafarian et al.229 (Level
more robust and stable system identification scheme. A 1-2,4) proposed a two-stage bridge damage detection based
Bayesian network for near real-time seismic damage as- on the response of a moving vehicle. In the first stage, ANN
sessment was proposed by Tubaldi et al.223 (Level 1-2,4). It is trained with backpropagation to predict the response of
incorporates multiple heterogeneous data sources, namely, the passage of a vehicle on the bridge. In the second stage,
ShakeMap, GPS, and accelerometers placed on a structure however, with the help of a Gaussian process, the change in
for the response, damage, and loss estimation by comparing the prediction errors’ distribution is detected; hence, the
prior and posterior statistical distributions. Each of the three damage is indicated. Although MLP and backpropagation
sources of information reduces different engineering de- learning methods are typically employed in damage de-
mand parameters defined in the article when used indi- tection of civil infrastructure, unsupervised NNs have also
vidually. However, the proposed Bayesian network data been considered. Self-organizing map (SOM) approach is
fusion techniques resulted in uncertainty reduction. Simi- an example of unsupervised NN. SOMs are observed as
larly, Bayesian techniques can quantify uncertainties of grids of neurons where they attempt to show high-
damage-sensitive features, for example, modal character- dimensional data in a 2D or 3D map while preserving
istics224 and can merge multiple techniques in an ensemble the original feature(s) properties.172 The advantage of the
learning form to reduce algorithm-induced false positives SOM learning method in comparison to the more traditional
and negatives.225 For more detail in Bayesian methods, ANN approach is that SOM training depends only on the
especially considering natural hazards engineering, inter- internal structure of the inputs rather than the input–output
ested readers are, therefore, referred to the very recent re- samples with error propagation as there is no target defined.
view by Zheng et al.226 Tibaduiza et al.230 (Level 1-4) used SOM by applying PCA
As opposed to the “black-box” nature of other ML al- data reduction to classify damage on an aluminum plate in a
gorithms, in contrast, since Bayesian methods assume the two-stage validation and diagnosis mode. In another work
prior knowledge or assumption of a hypothesis, this enables by Avci et al.231 (Level 1-2), SOM was applied on a grid
for a more transparent statistical inference. Being that it structure based on the stiffness reduction and boundary
represents a probabilistic distribution for both data and the condition changes to identify and quantify the damage.
model, various data types and parameters can be easily Even with optimized learning methods, a large amount of
integrated for a robust and flexible classifier. Although training time and being computationally heavy can deter the
different Bayesian approaches, as reported earlier, can implementation of ANN in some SHM applications.
benefit many SHM systems, it still does not offer out-of-the- Other than ANN and MLP, several other shallow and
box solutions for issues such as subjective selection of prior deep NN algorithms exist in the literature, although very
probability distributions and the computationally expensive few are popular in the SHM community. Perhaps the most
procedures needed for integrations over uncertain param- common DL approach is CNN. Recently, more research is
eters in the distribution. being carried out in CNN for SHM systems. This is ex-
haustively covered in numerous review papers,25,26,32–35,41
as indicated before. Therefore, other deep NN algorithms
Neural network (supervised/unsupervised)
are introduced here. It should be noted that many appli-
ANN approaches in damage detection take after the cations can utilize both CNN and different deep NN al-
working components in a human brain. In general, ANN gorithms together. The benefits that such combined methods
consists of at least three layers, namely, (1) input layer, (2) provide are sometimes far more superior when CNN is used
hidden layer(s), and (3) output layer. ANNs with one or alone. Therefore, the authors aim to provide examples of
more hidden layers are called multilayer perceptrons those papers that use deep NN as their core damage de-
(MLP). ANN can be viewed as an optimization process that tection technique.
identifies a set of network weights that minimize the cost Sequential data or time series is a major part of many
function.227 Such an approach has been widely used in past SHM systems. In contrast to the classical feed-forwarding
works as it allows for various inputs and outputs to be ANN, recurrent neural networks (RNN) can use their in-
included. In a feed-forward ANN, with independent ternal memory to loop the output back into the prediction
22 Structural Health Monitoring 0(0)

continuously. Although the looping can be computationally datasets. The superiority of GRU for smaller datasets was
expensive, it ensures that the sequence of data dependent on shown in a recent paper by Choe et al.237 (Level 1-2).
each other, unlike ANN, which can prove to be a useful Compared to LSTM and stacked LSTM, GRU managed to
feature in identifying damage in situations where the cor- achieve 10–30% in accuracy for structural damage detection
relation between multiple sources of signals and direct of floating offshore wind turbine blades. One limitation of
relation with external factors, for example, EOF exist. RNNs and its variants is unidirectionality of network; that
Mousavi and Gandomi232 (Level 1) used RNN to capture is, the output at a particular time step depends only on the
and predict temperature variations. The numerical anal- past information in the input sequence. To mitigate this,
ysis of their nonlinear system showed that damage could bidirectional RNN was proposed. To fix the problems of
be identified when prediction errors of the temperature variation in structural response due to initial residual stress,
signal deviate significantly from the expected value of the coupling effects of structure damage, and external loads,
error. In a similar study by Mousavi233 (Level 1), damage Tian et al.238 (Level 1) proposed a global and partial bi-
was identified under the conditions of EOFs when Jo- directional LSTM model to relate the girder vertical de-
hansen cointegration of the frequency signals that were flection to cable tension. It was found out that the partial
used to train an RNN model failed to identify a relationship model performed better with relative root mean square error
among the frequency signals. Their implementation per- (RRMSE) of 3.24% in addition to performing consistently
formed well under noisy conditions when tested against with noise levels and traffic volumes under normal oper-
two experimental samples. However, as confirmed by ational conditions. Although very popular in natural lan-
Zhang et al.,234 RNN suffers from exploding and guage processing (NLP), it appears that bidirectional deep
gradient-vanishing; that is, when the weights are assigned NN algorithms, even though they provide better flexibility
at the node, they are either really small, which effectively in some aspects of SHM systems, are not widely adapted by
stops the training process (vanishing), or the weights the community. This is perhaps due to the issue that the
become too large which may lead to an unstable network entire sequence must be available before making predictions
(exploding). and high computing cost of running such complicated
In order to remedy the gradient-vanishing issue, long models.
short-term memory (LSTM) network was introduced. In The final deep NN to discuss is the generative adversarial
short, LSTM has a special architecture allowing to re- networks (GAN). The idea of GAN is that two sub-models,
member information for long periods of time, which allows namely, generator and discriminator, produce and distin-
for learning long-term temporal dependencies. The work guish fake images given a latent vector and the original
by Zhang et al.235 (Level 1-2,4) is of the first im- dataset. With training, the generator improves and pro-
plementations of LSTM network with limited data for duces images that are more real. The idea of using GAN
seismic response modeling of highly nonlinear complex for SHM systems was studied by Tsialiamanis et al.239 It
dynamic systems. Due to limited data, K-means clustering was demonstrated that with prior knowledge, GAN can
was used to partition the dataset for generating training and reflect damage characteristics via categorical and con-
testing data. Their stacked LSTM network scheme per- tinuous variables despite the presence of EOFs. There-
formed well with the prediction error of ±10% with fore, GAN can be promising in training large datasets.
confidence intervals of 91%, 86%, and 84%, respectively. The current SHM system may benefit a lot by considering
Their approach, however, is computationally costly, with GAN in the pipeline. For example, the very recent paper
50,000 epochs for training. A novel deep RRN encoder– by Fan et al.240 (Level 1) demonstrated the applicability
decoder approach with LSTM in sequence-to-sequence of GAN for structural dynamic response reconstruction
(seq2seq) modeling was presented by Li et al.236 (Level under ambient excitations or seismic loadings. Although
1-2). Their online SHM monitoring under seismic exci- damage detection was a preliminary investigation and not
tation performed reliably to predict dynamic responses the goal of the article, the authors confirmed that such a
subjected to future earthquakes when compared with model can be used for identifying damage in SHM. Their
seven state-of-the-art methods for sequence learning and dynamic response reconstruction error was 15.7%
prediction—reduction in prediction error and the standard compared to the traditional CNN model with 69%.
deviation by at least 13% and 15%. The major drawback of
LSTM is the need for huge resources and long training
time to perform well.
K-means (unsupervised)
Another variant of LSTM architecture, named gated Clustering is a technique in which subgroups are assembled
recurring units (GRU) also exist. Having one fewer gate based on either features or samples. It performs a partition of
than LSTM makes GRU’s internal structure much simpler; data into K non-overlapping clusters. In an iterative process,
therefore, it becomes easier to train with fewer computa- each element is assigned to a partition considering the
tions. Generally, GRU performs better than LSTM for small minimum distance between the element and the centroid of
Malekloo et al. 23

each cluster that is either defined or estimated initially.241 Association analysis (unsupervised)
Once the assignment is over, the centroid is recalculated
based on the average of all elements in the cluster. Alamdari After data clustering, the association analysis method can be
et al.242 (Level 1) implemented a spectral-based clustering used to find relationships and dependencies, that is, asso-
SHM of the Sydney Harbour Bridge. In their approach, ciation rules.172 Especially in large dataset points, it be-
offline datasets were adapted instead of live data streams comes essential to leverage the scalability of association
from the SHM devices due to the challenges in data rules to extract useful information for decision making. One
communication overhead, delay, and the overall system’s of the most widely used applications of association analysis
resiliency. At the same time, multiple nodes indicate the is the market basket data analysis. It is used to find asso-
presence of damage. The effect of traffic loading on bridges ciations between customer’s purchasing patterns that can
has been widely studied in the literature. However, a limited provide interesting information to market owners to max-
number of them have incorporated the collected data from imize profit and help design enticing advertisements. Jin
the traffic to enhance SHM applications or fabricate separate et al.246 (Level 1-2,4) employed the Pearson correlation-
sensors dedicated to detecting vehicles. The study by based association analysis method for performance as-
Burrello et al.243 (Level 1) leveraged the SHM data with an sessment of an in-service bridge by mapping the structural
anomaly detection technique to identify traffic load from the dependencies such as stress and displacement. Pearson
acceleration peaks and utilize the K-means algorithm to correlation is a simple process that is typically used as a
distinguish amplitude and damping duration associated with measure of finding linear correlations between a large
heavy traffic and cars, respectively. K-means clustering is amount of data in association analysis methods. Pearson
heavily dependent on the choice of the K value, the initial correlation only measures the strength of the association
value of the centroid, and the distance metrics used to rather than the significance. Spearman, Kendall, and Chi-
determine the distance between each element and the square tests are other measures of correlation where in the
centroid of clusters. Moreover, implementations with K- final test, one can expect to find the association’s signifi-
means clustering with high-dimensional data, including cance. Guéguen and Tiganescu247 (Level 1) considered
EOFs, can hinder the method’s effectiveness. Dimension- dynamic changes in a building concerning the temperature
ality reduction techniques such as PCA or spectral clustering effect. The authors analyzed the correlation of resonance
methods244 are recommended before applying K-means frequency using the association rule learning method to
clustering. detect damage. Finding appropriate parameters, discovering
too many rules, discovering insignificant rules, and com-
putationally inefficient numerical data are some of the
Gaussian mixture (unsupervised) drawbacks of association analysis.
Gaussian mixture model (GMM) is a parametric proba-
bility density function similar to kernel density estimates
Semi-supervised
but with a small number of components. In bridge mon-
itoring applications, the model tries to capture the primary As mentioned in previous sections, it is rare and infeasible
component that corresponds to the healthy state of the to have fully labeled datasets in SHM applications ;hence,
bridge’s condition, even under varying EOFs. In an ex- supervised learning is illogical. On the other hand, unsu-
perimental study on the Z-24 bridge, Figueiredo and pervised approaches could take a great deal of effort to
Cross76 (Level 1) applied a mixture of supervised and implement. The possibility of having a small subset of
unsupervised learning approaches for nonlinear long-term labeled data is not far-fetched203; therefore, to leverage both
monitoring of bridges. The study utilized the supervised aspects of the data, semi-supervised learning methods can
NLPCA for characterizing the interdependency of iden- be used to take advantage of labeled and unlabeled data.
tified features, as well as the consideration of EOFs, and Bull et al.248 (Level 1) employed semi-supervised para-
the unsupervised GMM for damage identification based on metric GMM for improving the performance of damage
outlier detection. Similarly, Figueiredo et al.19 (Level 1) classification in risk-based applications. Rogers et al.249
explored the integration of model-driven and data-driven (Level 1) proposed a Bayesian non-parametric clustering
systems in a hybrid approach for damage detection of the technique to apply labels online to the clusters in a semi-
Z-24 bridge. The data from FEM was fed to the GMM to supervised manner with little to no data training. Chen
improve the damage classification and the validation of the et al.250 (Level 1) presented an adaptive graph filtering for
model. GMM of acoustic emission for monitoring cracks semi-supervised damage classification in an indirect bridge
in an RC shear wall was tested in the study by Farhidzadeh SHM application with improving classification accuracy. In
et al.245 (Level 1-3). The GMM was successful in clus- another indirect bridge monitoring study, Liu et al.106 (Level
tering two hidden classes of crack mode, that is, shear and 1-2,4) analyzed stacked autoencoders as a nonlinear di-
tensile. mensionality reduction technique and a semi-supervised
24 Structural Health Monitoring 0(0)

damage severity estimation model on a laboratory bridge last stage of damage identification in many of the SHM
model. The authors concluded that their approach is feasible systems is simply because of the fact that the development
and applicable to in-service bridge structures. of SHM in different papers is not matured. Essentially,
understating damage propagation for determining the re-
maining useful life (RUL) of the structure is not feasible.
Blind source separation in SHM
Moreover, in RUL estimation, the prediction is probabilistic
Blind source separation (BSS) aims to separate individual in nature and comes with a certain degree of uncertainty.
sources of signals from a set of mixed signals with little to However, in other applications such as rotating machineries,
no information. It is assumed that the mixture of signals is one can determine the exact future operational and loading
independent, and there exists no correlation between the factors. This is proved to be very difficult in civil infra-
components. There are cases in which damage is detected, structures where slight deviations or external factors can
and it usually is accompanied by other damages.251 influence the properties of the whole system. Furthermore,
Therefore, the acquired signals from the structures, along damage prognosis depends on the accurate global and local
with the noise, make it difficult to pinpoint and identify the model representation of the structure. Minimal research
faulty component. BSS algorithm, in this sense, can be used studies exist which consider these phenomena. RUL pre-
for output-only system identification and damage detec- diction is considered still an emerging technology. In this
tion.252 Generally speaking, BSS is treated as an unsu- part, we will introduce a few examples where successful
pervised ML problem dealing with the time or frequency implementation of RUL based on ML algorithms in the
domain.253 The difficulty in the BSS approach is the un- context of SHM is provided.
certainty in the number of sensors relative to the number of Atamturktur et al.257 used SVR on historic masonry fort
sources, especially when there are fewer sensors than there considering different support settlements. The prognostic
are sources. ML algorithms are developed explicitly for evaluation of the structural condition is based on an adaptive
BSS applications to tackle different BSS issues such as weighting of the regressor classifier for settlement-induced
model complexity and heterogeneous environment and strains up to 100 mm. The study achieved as much as 50%
address real-world applications such as SHM for civil in- reduction in the prediction error compared to the vanilla
frastructures. To overcome the high amounts of transmitted SVR method. NNs are also used for RUL estimation. In a
data in SHM systems, Sadhu et al.254 (Level 1) presented a typical static NN, such as MLP, prediction is performed at
decentralized, high compression sensing tool for data re- each time step independently. Although it would be wise to
duction within BSS framework. Musafere et al.252 (Level 1) use phase-space representation such as time-windowing for
developed time-varying autoregressive modeling to obtain generating a fixed sequence of instances, however, such an
the mono-harmonic responses from the vibration data. They approach gives rise to increased dimension and, subse-
validated numerical and experimental studies, as well as on quently, the problem of the curse of dimensionality. In the
a full-scale earthquake-excited building. Liu et al.255 (Level context of prediction, it is rational that having a history of
1) developed a BSS ML-based approach for modal iden- the past data stored can greatly improve the prediction.
tification and validated their results from the SHM data of a Therefore, deep NN can help to enhance the prediction
cable-stayed bridge under ambient vibration. Ye et al.256 accuracy. This was precisely demonstrated in the study by
(Level 1) proposed an integrated ML-based single-channel Wu et al.,258 where the LSTM approach for RUL of en-
BSS algorithm for separating deflection components from gineered systems was proposed. The authors used aircraft
live load effects, temperature effects, and structural de- turbofan engines datasets from NASA with four damage
flection for prestressed concrete bridges. Their approach conditions. The performance of their approach was tested
takes advantage of ensemble empirical mode decomposition against two other deep NN methods, namely: RNN and
(EEMD), PCA, and ICA algorithms. Applications of AI and GNU. The methodology described in this article can be
DL for BSS approaches have also been studied in the past; extended to civil structures as well.
however, there seems to be very little work integrating them Other than deep NN, there are also opportunities to le-
with SHM systems. BSS can be seen as the middleware verage physics-informed data-driven SHM systems for
framework between signal processing and ML. The im- damage prognosis. For example, such a hybrid approach
portance of BBS has led to novel developments in the field. was realized in the study by Das et al.259 The authors used
dynamic mode decomposition as well as computer vision
for the prediction of cracking in a mortar cube specimen. It
An insight into damage prognosis with ML was observed that with more training frames, the L2 norm is
Many of the reviewed works in ML-augmented SHM re- substantially reduced, indicating a more accurate validation
alize the first two levels of damage identification. Some also of prognosis. Crack prediction is one of the areas that
consider the extent of the damage to some length, and very perhaps, RUL concept is more apparent. In such scenarios,
few classify the damage. The reason for not including the the spatiotemporal phenomena that exist among degrading
Malekloo et al. 25

infrastructures, especially bridges, RUL estimation could are used (some examples are provided in the “Digital twin
provide valuable information about resource allocation, and physics-guided ML-based SHM” subsection). There-
planning, and retrofitting when a target risk level is reached. fore, only the most common models were chosen and
The uncertainties involved with simplification and as- compared together. In the future, a more in-depth assess-
sumption made during the RUL estimation process, can be ment of damage classification will be studied. A summary of
rectified by integrating probabilistic measures such as the advantages and the disadvantages of the mentioned
Bayesian statistics. For instance, a two-phase gammas algorithms in this article is provided in Table 8. It is quite
process with Bayesian approach was used to predict the challenging to justify training time, as indicated in the table,
remaining useful life of corroded reinforced concrete for different algorithms stated in this section as there may be
beams.260 For more information about RUL estimation, underlying optimization for different algorithms in the lit-
interested readers are recommended to read the bottom-top erature. The size and the type of the collected data can vary
review paper by Lei et al.,261 and the deep NN methods for by a significant margin from work to work. Therefore, the
RUL estimation by Zhang et al.262 training time was rated based on the general observation of
However, the biggest challenge in achieving complete 5- the authors while reviewing different papers, as well as the
stage damage identification is creating a pipeline that can authors’ experience in the past. The results of the reviewed
run each stage with an acceptable level of performance both ML algorithms for different stages of SHM are depicted in
quantitatively and also qualitatively. Even with advances in Figure 9, noting that it does not necessarily mean that
technology, the realization of damage prognosis into civil certain algorithms are incapable of addressing other levels.
structure monitoring systems currently does not necessitates As it can be observed, Level 1 can be achieved by virtually
spending efforts. However, with challenges being intro- every ML technique. Levels 2 and 3 are also considered in
duced to the already complex structural systems, damage most of the applications. For the case of Levels 3 and 5,
prognosis cannot be left behind. however, many works did not either take into account these
levels or believed that the algorithms were not suitable or
capable for damage classification and prognosis.
Remarks on SHM system with ML pipeline Based on 11 ML algorithms reviewed in this section,
In this section, several of the widely used algorithms in only neural network (NN) algorithms can accomplish
damage assessment in civil infrastructures are demon- complete 5-stage damage identification with DL approaches
strated. There also exist different models in the literature, considered as the first choice for several studies. Therefore,
where hybrid implementation or entirely different algorithms based on the reviewed works, the authors recommend an

Table 8. Advantages and disadvantages of several ML algorithms.

Algorithm Advantages Disadvantages Training time

Decision tree Quicker pre-processing step Overfitting in high-dimensional space ***

Random forest Quick prediction/training speed Model interpretability *
Robust to outliers The high memory requirement for
extensive datasets
Support vector Effective in high-dimensional spaces. Choice of the good kernel **(exponentially more for
machine Memory efficient Difficult to understand and interpret larger datasets)
K-nearest Easy implementation Overfitting in high-dimensional space -
neighbor Sensitiveness to very unbalanced datasets
Bayesian Easy implementation No interdependency between the **
Requires a small amount of training data features
Neural network Ability to learn and model nonlinear and Hardware dependence ***
complex relationships Nonlinear functions that are not
straightforward to interpret
K-means Easy implementation Choosing the initial value manually **
Scales to large data sets Overfitting in high-dimensional space
Gaussian mixture Easy implementation Overfitting in high-dimensional space *
Requires a small number of parameters
Association Quick and easy way to get meaningful A considerable number of discovered **
analysis insights from the dataset rules
Obtaining noninteresting rules

* represents a rating system for training time, * means quick, ** means average, and *** means long.
26 Structural Health Monitoring 0(0)

Figure 9. Relation between SHM damage identification levels and ML algorithms based on reviewed papers.

SHM system with ML-/DL-pipeline, as shown in Figure 10. open-standard protocols, and many others are today’s
It is suggested to utilize multisensory systems with non- challenges to overcome. For example, for seamless inte-
linear data normalization techniques to consider EOFs. Data gration of different services and technologies, the ever-
augmentations can also be employed with DL methods to present IoT needs to be open and exchange data with
impute and increase the quality of the data that could have other platforms.263 In today’s IoT ecosystem, efforts have to
been affected in earlier stages due to loss of signals or delays be spent embracing and managing the fast-paced IoT ad-
caused by communication overhead or dense data com- vancement and integrating it with ML to expand the
pression before transmission. Data compression by means boundaries.
of dimensionality reduction is one of the important stages. It
is noticed that with the consideration of EOF, nonlinear
dimensionality reduction methods perform better compared
ML for IoT in SHM systems
to other approaches. In the phase of feature extraction/ It can be said that WSNs with any intelligent software for
selection, several methods exist. Although the selected data collection, analysis, enhanced connectivity, and ac-
path is data-driven techniques, however, as will be dis- cessibility are considered to be an IoT application by
cussed later in the article, physics-guided data-driven themselves. However, the authors believe that such IoT
models can also be adapted. However, it has been only a implementations in SHM only scratch the surface of the
few years that this technique has gained popularity, and it is capabilities that IoT, in large, can provide. It goes beyond a
still mostly under-development for large-scale SHM sys- simple data acquisition system and software solutions for
tems. Next, with data fusion techniques, a multitude of feature extraction and damage prediction. Instead, it should
collected data can be combined to increase the possibility aim for a symbiotic connection between multisensory and
of detecting patterns and identifying outliers in the last stage city-scale multi-infrastructural network monitoring sys-
of the SHM system. Finally, based on the analysis of 11 ML tems. Such a paradigm cannot simply be achieved with
and DL algorithms, the authors consider that for Level 1-4 traditional SHM as it fails to provide any ubiquitous ser-
SHM, Bayesian and NN techniques can be considered for vices and powerful processing of sensing data stream. This
PR. Whereas, for complete 5-stage SHM damage identifi- is a prime example of how ML and cloud-processing in-
cation, it is imagined that with the current sensor technology frastructure can replace the old systems and achieve a higher
and algorithms and their limitations, only deep NN would level of efficiency. The power of ML for IoT can be re-
be considered as a suitable solution. alized in many implementations. The consideration of
EOFs was discussed at length, and it was proved to be a
major challenge and limitation in many SHM systems.
IoT-related applications
Without ML, it would be challenging to isolate external
Together with IoT and smart city monitoring systems, our variations that can influence the data received from the
living conditions in terms of comfort and safety are con- sensors. Also, faulty sensors and missing data are other
tinually improving. With increasing numbers of ubiquitous aspects where ML can be beneficial. IoT is about inter-
connected devices, it is expected to face research challenges connectivity. Therefore, there may be cases where multiple
simultaneously as innovative solutions are being developed. sensors are being used to collect different data in different
Computational resources, energy management, optimal sizes and resolutions. Interoperating and correlating such
sensor placement, interoperability, security and privacy, as the massive amount of data can be computationally
Malekloo et al. 27

Figure 10. A recommended SHM system with ML and DL enhancements at each component level.

expensive. Here, ML can be used to make inferences and Big data and SHM, a symbiotic strategy
make relations between multiple sources of signals for
identifying damage. There is no doubt that with the sheer amount of data from
When discussing ML for IoT, the real challenge is SHM, many link the big data paradigm to such systems. The
finding a way to seamlessly connect the two together in a misconception with the term “big data” is that only
unified platform without sacrificing information and data. the volume of data plays the most prominent role. With the
Nonetheless, throughout this article, it has been realized significant improvement in the current SHM systems, the
that many of the proposed ML-conjunct SHM methods do volume of data is no longer the critical factor. Instead,
not always meet the first and foremost important aspect of the following additional 4 Vs combined with the volume of
IoT, and that is continuous monitoring. Many examples the data make up the big data model.265,266
just introduce new damage identification techniques but
fail to provide any objective justification if their system 1. Variety: type and nature of the data, for example,
can sustain continuous monitoring. With these obstacles, EOFs
civil structure asset management becomes really impor- 2. Velocity: how fast the data are generated and ana-
tant, and this warns for further research and understanding lyzed, for example, excitation/extraction and sensor
as systems become more intricate. According to the optimization
timeline envisioned by Xu et al.264 and Figure 11, IoT is 3. Variability: discrepancy of the data, for example,
currently placed in the middle stage of different SHM outliers
platforms. Damage identification started with fully cen- 4. Veracity: the useability of the data, for example, data
tralized traditional methods, and now it has been rapidly cleaning and feature extraction methods
evolving to fully decentralized blockchain SHM systems.
ML utilization started in IoT-based SHM systems and will Big data and SHM share common grounds. Both are
be even more prominent in later stages. Insight on how the considered to convey data-driven findings despite the un-
stage evolvement occurs can be widely beneficial for precedented computational expense and features non-trivial
design solutions that can meet the future of big data to capture. They are interchangeable to some degree, and
analysis. big data solutions can come in handy for various SHM
28 Structural Health Monitoring 0(0)

systems, such as decentralized learning with GPU parallel impact damage identification pipelines. In terms of vari-
computing. The pipeline for big data and that of SHM, as ability and veracity, complexity and evolving relationships
was previously shown in Figure 4 and discussed in detail in between the collected data introduce uncertainty and out-
previous sections, are very similar. Table 9 shows how the liers, especially for long monitoring SHM systems. Low
steps involved in big data relate to the different stages of the data quality, in terms of missing and noisy data, leads to
ML-based SHM system.79 The challenges in any SHM reduced structural integrity decisions inferred from the
system are inherently some of the significant data issues in features. Generally, the challenges in big data and SHM
the processing domain. Thus, with big data improvement, a relate to each other, one way or another, as indicated in
similar strategy can be developed for SHM, reinforcing Table 9. When appropriate data-processing techniques are
future applications. developed, it is expected that they would bridge the gap
In terms of variety, it is rarely the case where incoming between these two different fields, essentially providing
data are structured in any meaningful way. They can be value to both systems.
semi-structured, like the temperature, wind, and traffic Liang et al.271 proposed a big-data SHM platform for
loading (EOFs), or unstructured, like data streams from serviceability assessment of a bridge. In their application,
computer visions from changes in the pixels. Therefore, the authors based their big-data system on sensor tech-
feature extraction becomes a complicated task. In terms of nology for data mining. Wang et al.272 approach for big data
volume, assume a structure with multiple sensors at different in SHM was mainly focused on the data fusion (data ag-
locations, in addition to external monitoring of weather and gregation in big data) and learning stages (modeling in big
traffic with sensor and computer vision. Together they can data). Their experimental testbed of a 12-story test structure
include more than hundreds of sensors, and with 24 h showed promising results in data reduction, energy effi-
continuous monitoring of structure populations,267–270 they ciency, cost, and quality. Ni et al.191 addressed the vari-
can generate an enormous amount of data in terms of pe- ability challenge in big data, for example, outlier and data
tabytes per week. Processing this many data, let alone the fusion, based on a DL method. The data recontraction of
vast storage required, is a major challenge for continuous anomalies after data compression is a significant task as it
SHM monitoring solutions. In terms of velocity, modern may pose severe challenges for the high accuracy of the
SHM data acquisition systems with high-resolution data model after reconstruction. Although the root of abnormal
introduce a data transfer bottleneck and may cause data loss data is a complex process, identifying them before data
during transmission. Delayed or missing data can heavily compression is a vital task. The veracity of big data is

Figure 11. Stage-development of SHM systems with different performance measures.

Table 9. Similarities between SHM and big data pipeline and the main challenges, after Cremona and Santos.79.

SHM Big data Main challenges

Data acquisition Recording Volume: High data flow

Velocity: High-frequency sampling
Data normalization/cleaning Cleansing Variety: EOFs
Variability: Outlier detection
Veracity: Damage sensitive feature
Data fusion/compression Data aggregation Veracity: No dependency
Feature extraction/model learning Modeling Resource allocation
Long and heavy computations
Classification Interpretation Type I and type II errors
Malekloo et al. 29

usually linked to the feature extraction stage of SHM. High- necessary to apply smart asset management such that the
dimensional features, as present in SHM due to EOFs, can critical structures be prioritized in terms of hardware and
make any system time-consuming and complicated. En- software allocation. Moreover, in the backbone of a perfect
tezami et al.273 proposed an ARMA model to tackle this smart city paradigm, it is expected that the services are
specific issue related to SHM and big data. For a general interoperable with open data standards, making the data
overview, interested readers are recommended to read the interpretation, analysis and sharing seamless. The smart
recent review by Sun et al.,37 which investigated some monitoring solutions proposed by different researchers as
aspects of big data and AI in bridge SHM. shown throughout the article, demonstrate, in broad picture,
the capabilities of such systems for pursuing the goals of
smart infrastructure. The utilization of ML and DL has
Smart city and SHM, the bigger picture granted the opportunity to take one step closer to having an
One of the reasons for marching toward a smart city eco- autonomous, self-learning, and self-sustaining smart city.
system is to use the potential of existing technologies and
infrastructures in providing the best utility to users and
Intelligent transportation system and SHM,
improving their future. Some aspects of the involvement of
SHM in the current smart city era are reviewed by Du a complementary addition
et al.274 In their work, the authors tackled WSNs issues in With the help of IoT applications, mobility and trans-
the monitoring system and posed many open-ended ques- portation are considered to be the key influencing factors in
tions regarding the future of smart city monitoring. The sustaining our surrounding environments, especially those
components of SHM encourage applications of data-driven that utilize intelligent transportation systems (ITS).279 There
smart solutions in the context of the smart city. Together, a are two possible ways to merge these two systems into one.
smart city is expected to provide a seamless connection The data collected through the ITS can be fed to the SHM
between services and citizens, and monitoring is an essential system and, in turn, improve the system’s reliability: this is
component of this connection. Combined with the power of commonly known as the ITS-informed-SHM system. An
ML and DL, the adaptation and integrations of smart example of this approach was discussed by Lan et al.,280
monitoring applications are of increasing interest in civil which showed the impact of traffic load for fatigue damage
engineering.275 The novel approaches in SHM applications evaluation on bridges. On the other hand, when providing
involving ML and AI are becoming the pinnacle of research the data collected from the SHM system to the ITS, in-
today. As explored in different sections of this review, what formation for real-time traffic management can be utilized,
is being seen today is the result of decades of research from especially under critical events such as an earthquake. This
which the recent works have at least some elements of the form of integration is referred to as SHM-informed-ITS. In
common ground ML as their core. this approach, further enhancement can be made when the
In line with the discussion above, the current utilization system is integrated as part of smart cities,281 where the
of SHM requires a change in definition and architecture. information could be used for other services provided in this
With Industry 4.0 and cloud-based monitoring solution context. This, in turn, enables interoperability, leading to an
already implemented or on the horizon, a transition toward enhancement in the quality of service (QoS). A smart
cyber-physical system (CPS)-based SHM design is envis- pavement monitoring system based on a supervised ML
aged. A decentralized, self-sustaining CPS is said to be the algorithm was demonstrated by Praticò et al.282 Approaches
next stage of smart monitoring systems. However, some and methodologies taken in this work were based on in-
complications in the design and deployment stage have to tegration with current or future smart cities with ITS as a
be considered and studied beforehand.276 A cloud-based backbone for data collection. In the study by Huang,283
concept of bridge monitoring was presented by Furinghetti different data-driven methods to assess the transportation
et al.277 The cloud-computing interface developed for their system’s health, efficiency, and safety were used. Using big
proof of concept with analysis of the software and hardware data and ITS, the author provided decision support for
requirement was shown to be a practical and appliable practitioners. Interested readers are referred to the review
approach for the future of smart monitoring. Ozer and paper by Khan et al.284 for bridge conditions assessment
Feng278 proposed a mobile CPS-based SHM system for integrating SHM and ITS.
structural reliability estimation of bridges. As the authors
stated, the ultimate goal is to integrate such a design with Next-generation SHM applications with
cloud-computing power to increase efficiency and easy
integration with the smart city.
ML/DL enhancements
The dense sensor node structure of a smart city brings SHM discipline has come a long way in the past century
about some challenges. With the deployment of different from conventional to mobile and smart systems. With the
kinds of sensors on structures all over the city, it becomes continuous improvements in sensor technology fields,
30 Structural Health Monitoring 0(0)

unprecedented new techniques have been introduced. and digital twin.292 Figure 12 depicts a summary of the
Starting with the first-ever SHM inspiration in the late 19th timeline of advancement in SHM and computing algo-
century8,285 for detecting cracks in railroad wheels to the rithms. In the next subsections, state-of-the-art and
implementation of the first NN machine by Marvin Min- emerging technologies and services that have the potential
sky,286 SHM relied on two main tracks of advances, in- to augment the traditional existing SHM systems with next-
cluding sensing devices and methods/algorithms. The generation sensing and computation advents are introduced.
beginning of SHM systems started with the condition They are not limited to but receiving utmost attention from
monitoring of rotating machinery with the simple shock the SHM community, according to the authors.
pulse method in the 1960s and later vibration-based mea-
surements, with further extension into offshore oil plat-
forms. Beginning with aerospace and civil engineering
UAV-assisted SHM
SHM applications in 1980, major innovations in ML, AI, Drone technology, also known as UAV, has seen a vast
DL, and computer vision were initiated. The earliest use of increase in usage in recent years due to the advantages it can
PR and ML in civil engineering structures was developed by offer, especially its deployment flexibility.293 Given their
Adeli and Yeh.287 Similarly, the first-ever computer vision versatility, low cost, and ease of deployment elements of a
model for civil structures was developed by Stephen et al.288 flying piece of technology, they are becoming increasingly
and Olaszek289 in the 1990s for bridges. accretive.294,295 UAVs can easily be integrated into the
With the introduction of cloud computing in 2002 and design phase workflow of civil structures for simple im-
Industry 4.0 in 2010, a giant leap toward a new generation of aging or scene reconnaissance, monitor the work-in-
computing and monitoring solutions was taken. The tran- progress and document phases of the construction, and
sition to CPS-based SHM design, machine-to-machine lastly, it can be used for monitoring and inspection. UAVs
communication, cognitive computing, virtual, and aug- enable investors to visit hard-to-access areas of many
mented reality has enabled a paradigm shift in the last 10 structures, such as tall buildings or bridges with a river
years. The authors believe the future of sustainable SHM flowing underneath. The use-cases of UAVs exceed the
systems with damage prognosis capabilities co-exists and imaging and video capabilities. They can also be equipped
co-integrates with smart cities, big data, and services with other sensors for vibration-based approaches. One of
and technologies such as interoperability,290 blockchain,291 the best applications of UAV is in disaster damage and loss

Figure 12. A timeline of advancement in technologies directly or indirectly affecting today’s SHM.
Malekloo et al. 31

estimation.31 Disaster mapping becomes essential in an area from 0.45 to 1.5. Hoskere et al.301 proposed a novel vision-
where no monitoring of critical infrastructure was sought based data extraction pipeline for measuring modal prop-
out. UAVs can combine complicated components of a erties of structures from a UAV. Compared to the fixed
stationary monitoring system and essentially create a mobile accelerometers, the UAVs were able to show good results
and portable mini-SHM. The following list shows different with 1.6% error in natural frequency and modal assurance
applications of UAVs typically used for damage detection/ criterion (MAC) values of above 0.925.
localization in civil structures: With the addition of a depth sensor such as an infrared
(IR) camera or LiDAR, the out-of-plane direction (the
• Concrete crack detection distance from the object to the camera) can also be mea-
• Pavement crack detection sured. This provided superior dynamic displacement
• Rust detection measurement on UAVs as demonstrated by Perry and
• 3D reconstruction via laser or light detection and Guo.302 Having a 3D structure model can easily enable local
ranging (LiDAR) changes to be detected, and damage can be identified.
• Displacement measurement using camera-lens Furthermore, the 3D reconstruction model can produce 3D
configurations FEM, which can effectively be integrated with different
• Displacement monitoring via lasers damage identification methods for a hybrid approach.
• GPS for Level 2 SHM, geo-tagging Conventional 3D reconstruction models require high-
• Ultrasonic beacon for Level 2 SHM, geo-tagging quality point clouds that are difficult to obtain. Therefore,
they may include defects and hence reducing the structural
Some of the above applications can take advantage of information required to obtain satisfactory results for
ML and DL. For example, crack and rust detection via damage detection. ML and DL can also be applied in these
computer vision can be carried out in two different ways, scenarios to overcome the deficiencies above, as demon-
pixel processing or patch processing. A full image is pro- strated by Hu et al.303 for a structure-aware semantic 3D
cessed in the former and based on edge detection or pixel model of a cable-stayed bridge using CNN.
separation on a single threshold is applied. In the latter, To utilize UAVs and the power of ML together, Perry
using ML or DL, the original image is segmented into et al.304 demonstrated a new approach to bridge inspection.
patches, and crack patterns are identified. Vibration-based This way, by collecting pictures from a bridge and mapping a
approaches can also be achieved using displacement 3D point-cloud and photo-realistic model, with the help of
monitoring via digital image correlation (DIC).296 computer vision and ML algorithms, it becomes more ver-
Noncontact reference-free displacement estimation is satile and efficient to detect faults with little to no human
particularly important for the railroad industry.297,298 As a interaction. A new crack detection technique based on the
low-cost solution, many researchers have directed their images taken from the UAV was proposed by Lei et al.305 In
efforts in developing new ways and utilizing emerging their approach, environmental noise is considerably reduced
technologies to capture vital information from railroad compared with the traditional edge detection methods and the
bridges. Height of railroad bridges, being located in remote, error rate of 5.43% was achieved. Augmentation of the new
irregular, and sometimes inaccessible areas, are the issues approach with ML and real-time bridge inspection is the next
that the usage of many traditional wired data collection step of the author’s work. Other than point-cloud-based
methods such as linear variable differential transducers methods of detecting faults, thermal-based imaging is also
(LVDTs) are becoming obsolete and replaced by next- used to capture information from bridges. Due to the
generation sensing devices such as laser Doppler vibr- mechanism of horizontal cracking around the rebar level, the
ometers (LDVs). These devices can also be mounted on change in the thermal properties of a bridge deck can indicate a
UAVs. Garg et al.299 installed LDVs on a UAV to collect pattern of delamination. This idea was used by Cheng et al.,306
displacement measurements of a railway bridge in the event using UAV with a supervised deep learning approach to
of a train-crossing event. Displacement measurements can capture the changes in the bridge deck. Due to the lack of
be followed by modal analysis and identification of fre- thermal images, experimental data augmentation was used in
quencies and mode shapes. With LDVs, amplitude and this study to enrich the training dataset. Some of the
frequency are extracted from the Doppler shift and en- applications above may provide a methodology for damage
hanced via ML and DL for modal analysis. For example, localization; however, dedicated units can be installed for geo-
CNN-LSTM was used to extract natural frequencies from a tagging identified damages. GPS and real-time kinematic
variety of beam samples using a shaker, and LDVs.300 global positioning system (RTK GPS) can be used to locate an
Displacement-based measurements using computer vision approximated position of the damaged location; however, they
only consider the plane perpendicular to the camera. The are limited only to the outdoor environment. In GPS-denied
superiority of the proposed technique was shown based on areas such as beneath a bridge, other methods such as ul-
the mean value for mean absolute error (MAE) that ranged trasonic beacon systems can be used for locating damage.307
32 Structural Health Monitoring 0(0)

From the studies above, it becomes clear that UAV-based generation devices. The latter tests only considered fre-
damage detection techniques can be proven promising to quency evaluation and provided error percentages around
implement on many civil infrastructures, especially bridges. 1%. Later on, Ozer et al.311 expanded the smartphone
The incorporation of UAVs into civil engineering demon- scheme into crowdsourcing and provided modal frequency
strates exceptional and practical feasibilities in terms of estimation of less than 1.3% error, whereas, old generation
scalability and automation. Being able to carry different smartphones were incapable capturing ambient vibrations.
kinds of sensors and devices makes UAVs an attractive low- For SHM applications, accuracy validation was extended to
cost option for rapid monitoring of structures. It is envi- mode shape identification and modal assurance criteria
sioned that with improvements in UAVs in the coming years values near 0.90 with smartphone data312 and higher values
and their integration into a smart city and becoming au- with multisensory data as well.313 The following lists the
tonomous, self-sustaining monitoring solutions, they will be potential applications of mobile-assisted SHM:
able to provide streamlined as-built critical SHM systems. A
summary of the UAV-related sensor technology and its • Crack detection with computer vision
application in SHM is shown in Table 10. • Vibration measurement with embedded accelerometers
• Displacement measurement with computer vision
• Drive-by sensing for indirect identification
Mobile-assisted SHM • Crowdsourcing for citizen-engaged operation
With the advancement in the IoT era, many citizens own a • Load detection from pedestrians using human activity
smart device that can easily be integrated into the moni- recognition (HAR)
toring solution. Smartphones with built-in cameras and • Supplementary information through multisensory
measuring components, such as accelerometers, show heterogeneous data feed
great potential in SHM applications. Smartphones can
capture video and images detecting faults and deformation Due to their size and portability, smartphones can be
on bridges308,309; the embedded accelerometers can attached to moving vehicles and can provide two use-cases.
identify dynamic characteristics of even very low- (1) Using their camera and computer vision algorithms,
frequency structures.310 Just like UAVs, SHM can benefit enhanced with ML and DL, they can identify different types
smartphone capabilities in various ways. They contain of road damages and measure road roughness.314 (2) They
storage, advanced microprocessors, GPS, and wide ranges can also be used for indirect monitoring through vibration
of the communication network from the cellular network to measurements and can be integrated with the first use case to
Wi-Fi and Bluetooth, in a small form factor. Smartphones include the effect of road roughness on the collected vi-
are also capable of higher spatial coverage compared to bration data.315 In addition to aforementioned mobile
stationary sensors. With data fusion, it would be possible to sensing paradigms, monitoring of pedestrians can be as-
design portable and massive network of monitoring solu- sociated with the dynamic bridge loading and cause dif-
tions being that the major advantage of smartphones is ferent levels of excitation.316 Therefore, smartphones can
scalability. Feng et al.158 performed small-scale, large-scale, greatly help to extract damage-sensitive features by ana-
and filed test to evaluate smartphone acceleration fidelity. lyzing the pedestrians’ body movement and transfer loading
Regarding small-scale shaking test, they reported accuracy mechanism to the bridge with the help of built-in accel-
errors in terms of identifying frequency and signal ampli- erometers for vibrations and gyroscope and magnetometer
tudes for different device generation, where old smartphone for direction correction. Applications of ML and DL in
generation showed up to 44% amplitude error and 5% mobile SHM systems can alleviate some of the short-
frequency error. In contrast, these error ranges reduced to comings with smartphones. In particular, for stationary
17% for amplitude and only 1% for frequency in new vibration monitoring with smartphones, sliding motion of

Table 10. Summary of UAV and smartphone sensors technology and use-cases in SHM.

SHM sensing capability Auxiliary Application type

sensing SHM
Platform Vision Accelerometer Gyroscope Magnetometer GPS capability Noncontact Contact level Application

UAV 3 — 3† 3† 3 3 **** * 1-2,4 CD, DM

Smartphone 3 3 3† 3† 3 — ** **** 1-2,4* CD, VB, DM,
LD

CD: crack detection; DM: displacement measurement; VB: vibration measurement; LD: load detection. † indicates for error correction only. * indicates
degree of applicability. In case of “SHM Level”, it indicates for some applications
Malekloo et al. 33

accelerometers due to smartphone not being fixed to the On the other hand, engaging consumer-grade devices
ground is an issue that was tackled by Na et al.317 The brings additional uncertainty due to the uncontrolled device
authors used SVM, NN, and RNN to detect the sliding operator.312,326,327 The challenges in big data are compa-
motion on a shaking table. With 93% accuracy, RNN was rable to the mobile sensing paradigms.328 Compiling and
able to classify the sliding motion correctly. Transfer analyzing tens of thousands of generated smartphone data,
learning is a popular method in DL where the pre-trained specifically when used in the context of crowdsourcing, can
model developed for a specific problem is reutilized in other sometimes challenge computational strategies in the field.
related problems. Therefore, the initial model is trained with Despite these obstacles, mobile sensing presents many
a considerable amount of data which can take a long time opportunities. Having a variety of data can increase the
and later transferred to smaller devices, such as smartphones observation sources and yield more accurate results, sup-
for damage detection, for example, cracks or concrete porting the decision-making tasks with abundant infor-
spalling in a fraction of a second, as demonstrated by Perez mation. New innovative ways of generating data from
and Tah.318 structures such as drive-by sensing329 are compatible with
Other than the indirect and mobile approaches, smart- the everyday smart devices placed in vehicles.330 If suc-
phones with computer vision can be used for measuring cessful crowdsourcing mechanisms are embedded into
displacement on different structural elements. Validation vehicular SHM, smartphones are likely to occupy more
studies for evaluating mobile-SHM were carried out by Yu space in the next decade of SHM research.311,315,331
et al.319,320 The results show the suitability of smartphones Smartphone-based SHM can digitally incorporate up-to-
for mini-SHM systems. A feasibility study of utilizing date advances in system identification, ML, and data mining
smartphone cameras for seismic structural damage detection that encompasses a fully connected smart city platform.332
was presented by Alzughaibi et al.321 With their experi- A summary of the smartphone-related sensor technology
mental vision-based solution for in-building damage de- and its application in SHM is shown in Table 10.
tection, the authors showed sub-millimeter accuracy
demonstrating the feasibility of smartphones for SHM. A Digital twin and physics-guided
vision-based approach with smartphones for obtaining
dynamic characteristics of a cable-supported structure ac-
ML-based SHM
cording to its dynamic displacement responses in the fre- The connection of the physical system (such as bridge el-
quency domain was investigated by Zhao et al.322 A 3D ements, sensors, etc.) and the cyber aspect (such as data
displacement monitoring system using the DIC technique management, processing, and communication) is tightly
was proposed by Wang et al.323 A real-time damage de- combined in modern SHM formulations. It was shown that
tection solution for masonry buildings using mobile DL was the isolation of cyber and physical aspects of a WSN-based
demonstrated by Wang et al.324 Leveraging the state-of-the- SHM solution is suboptimal.276 With the upcoming In-
art DL technique on historic buildings, the high-precision dustry 4.0, the Industrial Internet of Things (IIoT) is going
trained model was ported onto a smartphone and was to be the next major step for real-time performance mon-
successful in detecting damage. This shows that with ev- itoring and better predictive maintenance with new solu-
eryday improvement in object detection models and size tions in ML algorithms. The concept of the digital twin, as
reduction of the trained models, mobile device-based part of IIoT, has granted the ability to achieve a greater level
damage detection with DL can become a new attractive of automation and transparency for infrastructure asset
approach in SHM. Some studies also use multisensory management. A digital copy of the structure in the digital
capabilities in smartphone devices for monitoring solutions. twin domain is created to aggregate, process, and analyze
Ozer et al.313 proposed a hybrid vibration response mea- the information and generate new data.333 In some texts, a
surement and modal analysis system combining embedded digital twin is defined as a digital representation of a real-
accelerometers and cameras. The features and computa- world object in a CPS context.334 It is expected that the
tional power in smartphones can promote long-term current WSN-based SHM application is going to be inte-
monitoring of bridges using smartphones. With extremely grated with high fidelity ubiquitous digital twin in the future
low initial and running costs, and the ability to develop to eliminate the hurdles in current designs.
custom analysis software, smartphones can essentially Significant applications of digital twin are in manufactur-
provide a complete monitoring solution. Shrestha et al.325 ing, smart city, and healthcare applications.335 There is only
investigated the feasibility of a long-term bridge health a handful of digital twin modeling of civil structures. The
monitoring of Japan’s Takamatsu bridge. In more than necessary capabilities of designing the digital twin for
1 year operation period, seismic and traffic-induced vi- bridges for SHM purposes were discussed by Ye et al.336
brations from the smartphones were captured and validated The concept of a digital twin for cable-supported bridges
against the reference seismometers to verify their viability along with a pilot study was introduced by Shim et al.337
and accuracy. Using a 3D model and UAV with image processing, Shim
34 Structural Health Monitoring 0(0)

et al.338 developed a digital twin model of a long-span limited training data, K-means clustering was used to
bridge utilizing reversed 3D-surface modeling to identify partition the available data into training, testing, and pre-
the damage on the bridge. Data-driven analysis is of most diction categories. The computational efficiency and the
significant importance in the digital twin domain. In a high prediction scores can enable developing fragility
similar approach, the concept of the deep digital twin was function for building serviceability assessment. Another use
proposed by Booyse et al.292 to circumvent the practical of DL algorithm in physics-guided SHM system can be
limitation of a model-driven digital twin. The authors’ found in a previous study by the same authors. Zhang
generative adversarial network (GAN) as their deep et al.,342 proposed a deep LSTM network with the similar
learning framework was used for detection, diagnostics, and physics model approach as before. Analogous to their
prognostic of damage in a gearbox. follow-up study, same testing procedures were applied, and
There are cases where sole data-driven approaches might satisfactory performance was achieved. In modeling time
be insufficient to meet today’s SHM expectations’ re- series of complex nonlinear dynamical systems, shallow
quirements. One of the major obstacles in achieving pure NN such as ANN have distinct limitations. Physics-guided
data-driven monitoring solutions is the lack of training data. DL model such as CNN, RNN, and GAN may provide a
While in theory, one can develop high fidelity and inter- better approach, especially in the cases of constraint data.
pretable physics-based SHM systems and that the lack of In view of the authors, there is much potential in physics-
data is not a primary issue to begin with (in contrast to data- guided SHM with the integration of state-of-the-art ML and
driven methods), such a model-driven approach comprises DL algorithms. Based on the observation and the provided
uncertainty and modeling error for simplification and examples, the progressive improvements of data-driven
omission due to high computational cost. Therefore, in approaches and digital twin over the years can persuade
recent years, there have been strides toward synergistically researchers to design symbiotic systems such that actual
integrating these two approaches such that they preserve characteristics based on physics-informed mathematical
their merits while at the same time lessening the inadequacy models can enhance the digital copy of structures.
in a reasonable manner. Consequently, physics-guided ML-
based SHM systems have started to combine both damage Virtual reality and augmented reality
detection techniques. For example, in a recent study by
Zhang and Sun,339 FEM updating was used in an interactive
for SHM
manner with NN such that the physics-based loss function Virtual reality (VR) refers to computer technologies and
determines the difference between the output of the NN and interfaces to simulate a 3D and interactive environment;
the results of FEM updating. Through this interaction, the whereas, augmented reality (AR) implies layering virtual
NN can learn well and detect damage when tested against information on real-world objects.343 Initially inspired and
unseen data. The authors showed that their implementation developed for the gaming community, VR and AR are now
could both successfully improve the generality of NN (17% finding their way into other fields, such as nuclear facili-
increase) and also enhance the performance of FEM up- ties344 or medical fields.345 In SHM systems, such novel
dating by uncertainty reduction. Connecting this new wave approaches are becoming a trend. For example, the com-
to the paradigm of digital twin, in the framework proposed bined VR and information model (IM) was used to visualize
by Ritto and Rochinha,340 measurements are taken from a and access SHM data and metadata in 3D.346 Such a system
physical twin (bar structure) to calibrate a stochastic allows users to intuitively view where SHM data is gen-
computational model to simulate the system’s response erated and how it is used to assess the damage. A conceptual
considering different damage intensities and locations. The seismic impact simulator utilizing SHM, ITS, GIS, and VR
virtual domain was assumed to be an ML classifier that can was proposed by Büyüköztürk and Yu.347 In a recent study
detect damage with different classifiers such as SVM and by Bacco et al.,348 the authors proposed architecture for IoT-
DT. based remote monitoring with UAV and VR for locating
These two studies have shown that it is paramount for a various sensors attached to a structure and displaying in-
physics-based physical twin to exist together with an ML stantaneous and historical records.
classifier in a virtual domain for interpretability, flexibility, Despite being a mainstream condition assessment
and reduced complexity. For further information, readers technique, visual inspections have shortcomings: labor-
are recommended to read the recent comprehensive survey intensive, error-prone, tedious, etc. are to name a few.
about integrating physics-based modeling with ML as AR has brought the opportunity to deliver new ways of
presented by Willard et al.341 In the study by Zhang et al.,234 portraying information that was deemed far-fetched a few
the authors proposed a physics-guided DL surrogate model years ago. This new content delivery paradigm has enabled
for seismic structural response prediction. The authors engineers to simultaneously immerse themselves in a fully
employed the law of dynamics as their physics models for connected physical and digital world. An AR-enabled
training CNN on a reduced dataset. In order to alleviate infrastructure inspection interface was developed by
Malekloo et al. 35

Maharjan et al.349 The system was coupled with low-cost are apparent to a limited extent, and their purposes
smart sensors and QR code scanners to assist the workflow other than sole visualization need to be investigated.
of the inspectors. It is intuitive and much easier to apply ML • Given that each civil infrastructure has a unique
and DL algorithms and display AR devices’ results. A proof presence and complex nature, the uncertainties in
of concept of application of AR in bridge monitoring structural and material behavior need systematic
presented by Yuan et al.350 An AR framework was pre- quantification and reduction techniques. From a
sented by Athanasiou et al.351 It was shown that with the present SHM perspective, modern and innovative
holographic reinforcement visualization, overall inspection solutions have to tackle scalability and adaptiveness
time could be reduced, improving the efficiency in col- concerns.
lecting and managing data. The enhancement to the visual • Apart from the classical uncertainty problems associ-
inspection brought by AR and its integration with ML and ated with different SHM levels, novel SHM paradigm
DL is a promising technology for infrastructure inspection proposing citizen engagement is still insufficiently
based on the review by Mascareñas et al.352 addressed according to the authors. The majority of
The enhancements brought by VR and AR can offer crowdsourcing research in SHM arena is still at the
effective additions to other next-generation SHM applica- conceptual level and there is growing need to establish
tions, in particular, UAVs and smartphone-based systems, real-life applications with real citizens.
for both visualization and deployment measures. Human • Despite the fact that the comprehensiveness of the state-
computer interaction can benefit from multisensory and of-the-art system identification and damage detection
interconnected media which can deploy SHM data for the techniques existed as of today, one can observe that
operators, as well as users. With the power of ML and DL, fully automated approaches are still uncommon and
this interaction becomes real-time and more intuitive and mere. Even the data-centric methodologies rely on
brings VR/AR opportunities provided in the context of human decisions and interactions in numerous phases
smart infrastructure and smart monitoring in the future of implementations.
smart city. • Looking at the energy harvesting developments sup-
porting self-powered WSN, compromises can be seen
in terms of communication of data and remote pro-
Open research issues cessing. More advancement in energy-aware algo-
With the increased variety of SHM instrumentation and rithms or routing protocols could effectively reduce
analytics approaches, traditional monitoring solutions are nodal and global energy consumption. In addition,
becoming obsolete. The heterogeneous nature of the col- power-generating elements can attract more attention,
lected data notifies the need for multivariate and asyn- such as solar cells, piezoelectric, or thermoelectric
chronous processing strategies. One major problem related elements utilizing different energy sources such as
to data-driven SHM approaches is that it is highly im- light, vibration, and heat, respectively. Prediction and
practical to collect sufficient data to train ML algorithms in optimization of the performance of the energy har-
real life. To compensate for this, data fusion or sensor fusion vesting system is the trend for WSNs, especially in
has drawn the attention of the researchers. However, SHM applications.354
multisensory applications still need maturity for widespread • Despite tremendous efforts in literature, combined
use.353 characteristics of damage-sensitive features and EOFs
are still partially uncovered. ML approaches combined
• Concerning the data volume, big data research needs with long-term monitoring are believed to serve this
further integration into SHM to meet the smart city line of research. As expressed in the emerging tech-
demands and to meet the aggregated knowledge de- nologies discussion, the ubiquitous and remote sensing
mands from smart cities. Integration data-driven and alternatives diminish practical and financial problems
model-driven approaches still need to find an optimal related to the maintenance of permanently installed
level of contribution from each other. In that sense, the systems.
extent of physics-informed foundations is still yet to be • The major shortcoming with data-based SHM systems,
determined. apart from the influence of EOF, is the data scarcity for
• In addition to the above aspects of the modern SHM supervised PR algorithms. Potential remedies may
paradigm, real-time or online learning, identification, vary, but often include inefficient or infeasible work-
and monitoring, in general, are partially achieved so far arounds. The idea of transferring knowledge between
and expect further advancements. In parallel, the IoT similar and dissimilar structures in the context of
framework and cloud computing are believed to play a population-based SHM has provided solution as to no
vital role in minimal delays in digital twin’s perfor- longer be concerned with lack of data.267–270 With
mance. Other than these, the use-cases of VR and AR being able to generate a relatively complete damage-
36 Structural Health Monitoring 0(0)

labeled data from a set of structures in a population and thanks to the mobility and abundance of modern
developing an abstract framework of metric space of measurement devices. Finally, developments of mul-
structures for mapping, knowledge transfer is possible, tifaceted technologies and services can open the paths
facilitating the creation of a general ML/DL method for to interoperability and an open standard for data while
the entire population. This idea is still at its infancy and ensuring total security and privacy measures.
real-life application should prove its suitability as in- • Other than these primary and general directions re-
teroperable SHM solution in the smart city. quiring future attention, partner disciplines can embed
• With the increase in the number of connected devices, recent advances in SHM better. The extension of ML
the IoT ecosystem must communicate and exchange into earthquake engineering applications brings the
information with one another. SHM studies so far are promise of incorporating physical knowledge into
tailored as structure-specific; however, dense networks data-driven models in seismic studies. The next-
can enable deducing identification findings in region- generation SHM can be coupled with ML and revo-
scale frameworks. What is more, systemic features of lutionize earthquake engineering to solve some of the
the civil infrastructure population can be grasped significant challenges in the field.355

Figure 13. The system architecture of a ML-enhanced cloud-based SHM-GIS decision-making system for bridge monitoring
applications, modified after Malekloo et al.294.
Malekloo et al. 37

• Indirect-bridge monitoring and drive-by sensing have critical due to lack of centralized data acquisition.
gained popularity among researchers. Previous at- Transmission delays due to communication overhead
tempts only considered Level 1 SHM. However, it is and the differences between the target and achieved
envisioned that, with the utilization of ML and DL, sample rate introduced in the SHM system are exac-
higher levels of SHM can be achieved. Concerning erbated with multisensory systems.
drive-by SHM research, one can note that the utmost
effort is spent on individual vehicular data which does
not fully reflect the smart city theme. Therefore, more
Conclusion
research on vehicle-bridge-interaction (VBI) encom- This article provided an extensive overview of the ML-
passing vehicle fleets is suggested.356 engaged SHM systems with connections to the new tech-
• For the transition toward the future of sustainable SHM nologies rapidly growing in the latest decade. A detailed
systems, integration of SHM with digital twin plus breakdown of techniques, methods, and algorithms from the
blockchain is of utmost obscurity. Considering the literature is presented and examined, emphasizing ML and
many aspects of the ecosystem, the coherent and the data-centric advancements occupying the current re-
synergetic connection of multiple emerging technolo- search trends. The survey included a systematic discussion
gies is detrimental, providing better QoS to the user and of the steps taken to implement an ML model for SHM with
increase overall system efficiency and integrity. As with pathways, taxonomies, and breakdowns. Moreover, the
other IoT devices and services, creating a middleware most common algorithms proposed for context-dependent
system that enables integration and interoperability of applications were overviewed. The survey revealed that the
SHM with other parts of the smart city ecosystem is not extension of ML in SHM dramatically increased the sys-
far-fetched. tem’s capabilities, providing innovative solutions for dif-
• The fifth and coming sixth generation (5Gand 6G) ferent research challenges.
mobile network is expected to be the center of the The ML pipeline and corresponding algorithms have the
emerging IoT devices in the near future. With the ever- potential to uncover the influence of EOFs due to their
increasing applications in cloud computing and smart multivariate encapsulation capabilities. EOFs, a long-
devices, 5G promises to address the current issues of lasting problem in the SHM community, is one step
telecommunications. 5G integration with SHM is closer to a solution with ubiquitous data and their digital
becoming widespread, although security aspects such extensions. Moreover, ML solutions also draw a pathway to
as integration need to be further evaluated.357 Mobile addressing nonstationary and nonlinear sources of varia-
edge computing and fog networks as part of the ca- tions, and compression/dimensionality reduction brings
pabilities brought by 5G and 6G enable on-site device gigantic inverse problems into solvable stages.
deployment of models and algorithms used for rapid Forthcoming mobile and noncontact technologies are
assessment of civil infrastructures.358 arriving with their digital counterparts. They do not only
• Versatile SHM systems require a robust data man- offer new sources of physical parameters being observed,
agement scheme. It is one of the topics that receive but also have their own embedded intelligence from
very little attention from the community. The con- consumer-grade smart devices to UAVs. Likewise, IoT is no
tinuous increase in the volume and types of data re- longer a futuristic theme; it became a reality with the rapid
ceived from many multisensory SHM applications distribution of low-cost headless computers all over the
overwhelms the current capabilities of data acquisi- world. However, the community still has an unclear un-
tions’ storage size and computational power. Cloud derstanding of how these breakthroughs can serve the smart
computing with a NoSQL database has granted the city agenda as well as sustainability on the monitoring side.
ability to manage massive structured and structured The next decade is expected to provide alternative aspects,
SHM data,359 and provide the necessary graphical which attracted rare attention, such as visualization and
processing unit (GPU) power with parallel and multi- interfaces.
threaded computation. The system architecture of a Despite the intrinsic progress in ML, DL, and AI, there is
cloud-based bridge monitoring system is depicted in an apparent gap in unsupervised SHM frameworks. Unseen
Figure 13. conditions of real damage obstruct training possibilities,
• Multisensory SHM systems are closely related to the which can be barely fulfilled by synthetic datasets or
5 Vs of big data. In this case, time synchronization of physical-based realizations. Nevertheless, further ad-
different sensors despite clock imperfections, becomes vancements with label-free approaches such as population-
essential. Although this topic has been studied for based SHM, can find remedies to the ongoing learning
stationary WSNs in SHM systems, with the intro- problem in SHM systems. It is obsolete that a fully auto-
duction of next-generation SHM modules such as mated SHM relies on this direction yet has a long way to
smartphones and UAVs, time synchronization becomes propose its globally accepted frameworks.
38 Structural Health Monitoring 0(0)

In conclusion, to understand where the next-stage SHM 6. Barthorpe RJ. On Model- and Data-based Approaches to
is placed, this survey looks at the parallel developments in the Structural Health Monitoring. PhD, University of Sheffield,
multidisciplinary world of SHM from the microelectronics 2010, http://etheses.whiterose.ac.uk/1175/.
advancements to communication and from citizen science to 7. Friswell MI. Damage identification using inverse methods.
cloud and edge computing. Needless to say, uncertainty Philos Trans R Soc Math Phys Eng Sci 2007; 365: 393–410.
reduction is boosted by revolutionary advances in regressors, 8. Farrar CR and Worden K. Structural health monitoring: a
classifiers, and detectors. It is the authors’ opinion that the machine learning perspective. Chichester, West Sussex, UK;
new norms in SHM unite all aspects of the digital revolution Hoboken, NJ: Wiley, 2013.
and Industry 4.0 together with the traditional lines of system 9. Rytter A. Vibrational based inspection of civil engineering
identification, advanced modeling, and damage assessment. structures. PhD Thesis, Dept. of Building Technology and
Structural Engineering. Aalborg University, 1993.
Author Contributions 10. Farrar C, Park G and Worden K. Complexity: A new axiom for
structural health monitoring? Los Alamos National Lab Nm,
Conceptualization: A.M. (equal) and E.O. (equal); visualization:
2010.
A.M. (equal); writing original draft: A.M. (equal) and E.O. (equal);
11. Rizzo P, Cammarata M, Dutta D, et al. An unsupervised
writing–review and editing: A.M. (equal), M.A. (equal), M.G.
learning algorithm for fatigue crack detection in waveguides.
(equal), and E.O. (equal); supervision: E.O. (equal); all authors
Smart Mater Struct 2009; 18: 025016.
approved the final submitted draft.
12. Tran TTX and Ozer E. Automated and model-free bridge
damage indicators with simultaneous multiparameter modal
Funding
anomaly detection. Sensors 2020; 20: 4752.
The author(s) disclosed receipt of the following financial support 13. Tibaduiza D-A, Torres-Arredondo M-A, Mujica LE, et al. A
for the research, authorship, and/or publication of this article: Dr study of two unsupervised data driven statistical methodol-
Ekin Ozer would like to acknowledge the funding support from the ogies for detecting and classifying damages in structural
Horizon 2020 Project TURNkey, Grant No. 821046. Moreover, Dr health monitoring. Mech Syst Signal Process 2013; 41:
Mohammad AlHamaydeh would like to acknowledge the 467–484.
American University of Sharjah (AUS) funding support through 14. Farrar CR and Worden K. An introduction to structural health
the Faculty Research Grant program (FRG19-M-E65). This article monitoring. Philos Trans R Soc Math Phys Eng Sci 2007;
represents the opinions of the authors and does not mean to 365: 303–315.
represent the position or opinions of the funding entities. 15. Sohn H. Effects of environmental and operational variability
on structural health monitoring. Philos Trans R Soc Math
ORCID iDs
Phys Eng Sci 2007; 365: 539–560.
Arman Malekloo  https://orcid.org/0000-0003-0188-3098 16. Farrar CR, Cornwell PJ, Doebling SW, et al. Structural
Ekin Ozer  https://orcid.org/0000-0002-7177-0753 Health Monitoring Studies of the Alamosa Canyon and I-40
Mohammad AlHamaydeh  https://orcid.org/0000-0002-5004-0778 Bridges. LA-13635-MS, Los Alamos, NM: Los Alamos
Mark Girolami  https://orcid.org/0000-0003-3008-253X National Laboratory, 2000.
17. Ding Y, Li A and Liu T. Environmental variability study on
References the measured responses of Runyang Cablestayed Bridge
1. Doebling SW, Farrar CR, Prime MB, et al. Damage identi- using wavelet packet analysis. Sci China Ser E Technol Sci
fication and health monitoring of structural and mechanical 2008; 51: 517–528.
systems from changes in their vibration characteristics: A 18. Magalhães F, Cunha A and Caetano E. Vibration based
literature review. LA-13070-MS, Los Alamos, NM: Los structural health monitoring of an arch bridge: From auto-
Alamos National Laboratory, 1996. mated OMA to damage detection. Mech Syst Signal Process
2. Farrar CR, Doebling SW and Nix DA. Vibration–based 2012; 28: 212–228.
structural damage identification. Philos Trans R Soc Lond Ser 19. Figueiredo E, Moldovan I, Santos A, et al. Finite element–based
Math Phys Eng Sci 2001; 359: 131–149. machine-learning approach to detect damage in bridges under
3. Sohn H, Farrar CR, Hemez FM, et al. A review of structural operational and environmental variations. J Bridge Eng 2019;
health monitoring literature: 1996–2001. LA-13976-MS, 24: 04019061.
Los Alamos, NM: Los Alamos National Laboratory, 2003. 20. Figueiredo E, Park G, Farrar CR, et al. Machine learning
4. Cao P, Qi S and Tang J. Structural damage identification using algorithms for damage detection under operational and en-
piezoelectric impedance measurement with sparse inverse vironmental variability. Struct Health Monit Int J 2011; 10:
analysis. Smart Mater Struct 2018; 27: 035020. 559–572.
5. Moore EZ, Nichols JM and Murphy KD. Model-based SHM: 21. Noel AB, Abdaoui A, Elfouly T, et al. Structural health
Demonstration of identification of a crack in a thin plate using monitoring using wireless sensor networks: a comprehensive
free vibration data. Mech Syst Signal Process 2012; 29: 284–295. survey. IEEE Commun Surv Tutor 2017; 19: 1403–1423.
Malekloo et al. 39

22. Arcadius Tokognon C, Gao B, Tian GY, et al. Structural 38. Azimi M, Eslamlou AD and Pekcan G. Data-driven structural
health monitoring framework based on internet of things: a health monitoring and damage detection through deep
survey. IEEE Internet Things J 2017; 4: 619–635. learning: state-of-the-art review. Sensors 2020; 20: 2778.
23. Gomes GF, Mendez YAD, Alexandrino PDSL, et al. A review 39. Avci O, Abdeljaber O, Kiranyaz S, et al. A review of
of vibration based inverse methods for damage detection and vibration-based damage detection in civil structures: from
identification in mechanical structures using optimization traditional methods to machine learning and deep learning
algorithms and ANN. Arch Comput Methods Eng 2019; 26: applications. Mech Syst Signal Process 2021; 147: 107077.
883–897. 40. Tibaduiza Burgos DA, Gomez Vargas RC, Pedraza C, et al.
24. Fan W and Qiao P. Vibration-based damage identification Damage identification in structural health monitoring: a brief
methods: a review and comparative study. Struct Health review from its implementation to the use of data-driven
Monit 2010; 10: 83–111. applications. Sensors 2020; 20: 733.
25. Feng D and Feng MQ. Computer vision for SHM of civil 41. Hou R and Xia Y. Review on the new development of
infrastructure: from dynamic response measurement to vibration-based damage identification for civil engineering
damage detection – a review. Eng Struct 2018; 156: 105–117. structures: 2010–2019. J Sound Vib 2020; 491: 115741.
26. Ye XW, Dong CZ and Liu T. A review of machine vision- 42. Sony S, Laventure S and Sadhu A. A literature review of next-
based structural health monitoring: methodologies and ap- generation smart sensing technology in structural health
plications. J Sens 2016; 2016: 1–10. monitoring. Struct Control Health Monit 2019; 26: e2321.
27. Dong C-Z and Catbas FN. A review of computer vision– 43. Alavi AH and Buttlar WG. An overview of smartphone
based structural health monitoring at local and global technology for citizen-centered, real-time and scalable civil
levels. Struct Health Monit 2021; 20: 692–743. infrastructure monitoring. Future Gener Comput Syst 2019;
28. Ghiasi R, Ghasemi MR and Noori M. Comparative studies of 93: 651–672.
metamodeling and AI-based techniques in damage detection 44. Bao Y, Chen Z, Wei S, et al. The state of the art of data science
of structures. Adv Eng Softw 2018; 125: 101–112. and engineering in structural health monitoring. Engineering
29. An D, Choi JH and Kim NH. . Boston, Massachusetts: 2019; 5: 234–242.
American Institute of Aeronautics and Astronautics. DOI: 10. 45. Farrar CR, Duffy TA, Cornwell PJ, et al. Excitation Methods
2514/6.2013-1940. Options for Prognostics Methods: A re- for Bridge Structures. LA-UR-98-4579, Los Alamos, NM:
view of data-driven and physics-based prognostics54th Los Alamos National Laboratory, 1999.
AIAA/ASME/ASCE/AHS/ASC Structures, Structural Dy- 46. Agdas D, Rice JA, Martinez JR, et al. Comparison of visual
namics, and Materials ConferenceEpub ahead of print 8 April inspection and structural-health monitoring as bridge con-
2013 dition assessment methods. J Perform Constr Facil 2016; 30:
30. Moughty JJ and Casas JR. A state of the art review of modal- 04015049.
based damage detection in bridges: development, challenges, 47. Maas S, Nguyen VH, Kebig T, et al. Comparison of different
and solutions. Appl Sci 2017; 7: 510. excitation- and data sampling-methods in structural health
31. Kerle N, Nex F, Gerke M, et al. UAV-based structural damage monitoring: Kurita-Albrecht Best Scientific Paper Award
mapping: a review. ISPRS Int J Geo-Inf 2019; 9: 14. 2018 – Concrete. Civ Eng Des 2019; 1: 10–16.
32. Khan S and Yairi T. A review on the application of deep 48. Sensors for Smart Systems. In: Smart Material Systems and
learning in system health management. Mech Syst Signal MEMS. Chichester, UK: John Wiley & Sons, pp. 43–62.
Process 2018; 107: 241–265. 49. Pentaris FP, Stonham J and Makris JP. A review of the state-
33. Ye XW, Jin T and Yun CB. A review on deep learning-based of-the-art of wireless SHM systems and an experimental set-
structural health monitoring of civil infrastructures. Smart up towards an improved design. Eurocon 2013. Zagreb,
Struct Syst 2019; 24: 567–585. Croatia: IEEE, pp. 275–282.
34. Hou L, Chen H, Zhang (Kevin) G, et al. Deep learning-based 50. Choi H, Choi S and Cha H. Structural health monitoring
applications for safety management in the AEC industry: a system based on strain gauge enabled wireless sensor nodes.
review. Appl Sci 2021; 11: 821. In: 2008 5th International Conference on Networked Sensing
35. Sony S, Dunphy K, Sadhu A, et al. A systematic review of Systems. 2008, pp. 211–214.
convolutional neural network-based structural condition as- 51. Moreno-Gomez A, Perez-Ramirez CA, Dominguez-Gonza-
sessment techniques. Eng Struct 2021; 226: 111347. lez A, et al. Sensors used in structural health monitoring. Arch
36. Salehi H and Burgueño R. Emerging artificial intelligence Comput Methods Eng 2018; 25: 901–918.
methods in structural engineering. Eng Struct 2018; 171: 52. Abdulkarem M, Samsudin K, Rokhani FZ, et al. Wireless
170–189. sensor network for structural health monitoring: a contem-
37. Sun L, Shang Z, Xia Y, et al. Review of bridge structural porary review of technologies, challenges, and future di-
health monitoring aided by big data and artificial intelligence: rection. Struct Health Monit 2020; 19: 693–735.
from condition assessment to damage detection. J Struct Eng 53. Ubertini F, Laflamme S and D’Alessandro A. Smart cement
2020; 146: 04020073. paste with carbon nanotubes. Innovative Developments of
40 Structural Health Monitoring 0(0)

Advanced Multifunctional Nanocomposites in Civil and 69. Flynn EB and Todd MD. Bayesian probabilistic structural
Structural Engineering. Elsevier, pp. 97–120. modeling for optimal sensor placement in ultrasonic guided
54. Park G, Cudney HH and Inman DJ. Impedance-based health wave-based structural health monitoring. In: Peters KJ, Ecke
monitoring of civil structural components. J Infrastruct Syst W and Matikas TE (eds). San Diego. USA: California,
2000; 6: 153–160. p. 76480Z.
55. Park G, Sohn H, Farrar CR, et al. Overview of piezoelectric 70. Capellari G, Chatzi E and Mariani S. Optimal sensor
impedance-based health monitoring and path forward. Shock placement through bayesian experimental design: effect of
Vib Dig 2003; 35: 451–464. measurement noise and number of sensors. Proceedings
56. Giurgiutiu V. Tuned lamb wave excitation and detection with 2016; 1: 41.
piezoelectric wafer active sensors for structural health 71. Semaan R. Optimal sensor placement using machine learning.
monitoring. J Intell Mater Syst Struct 2005; 16: 291–305. Comput Fluids 2017; 159: 167–176.
57. Kehlenbach M and Das S. Identifying damage in plates by 72. Mao J-X, Wang H, Feng D-M, et al. Investigation of dynamic
analyzing Lamb wave propagation characteristics. In: Kundu properties of long-span cable-stayed bridges based on one-
T (ed). San Diego, CA, pp. 364–375. year monitoring data under normal operating condition.
58. Lee BC and Staszewski WJ. Lamb wave propagation mod- Struct Control Health Monit 2018; 25: e2146.
elling for damage detection: II. Damage monitoring strategy. 73. Huang J, Li D, Li H, et al. Damage identification of a large
Smart Mater Struct 2007; 16: 260–274. cable-stayed bridge with novel cointegrated Kalman filter
59. Staszewski WJ, Lee BC, Mallet L, et al. Structural health method under changing environments. Struct Control Health
monitoring using scanning laser vibrometry: I. Lamb wave Monit 2018; 25: e2152.
sensing. Smart Mater Struct 2004; 13: 251–260. 74. Li H, Li S, Ou J, et al. Modal identification of bridges under
60. Bigoni C, Zhang Z and Hesthaven JS. Systematic sensor varying environmental conditions: temperature and wind
placement for structural anomaly detection in the absence of effects. Struct Control Health Monit 2009: 495–512.
damaged states, 2020, http://infoscience.epfl.ch/record/277416. 75. Santos A, Figueiredo E, Silva MFM, et al. Machine learning
61. Banik MR and Das T. Proceedings of the International algorithms for damage detection: Kernel-based approaches.
Conference on Computing Advancements. Dhaka Bangla- J Sound Vib 2016; 363: 584–599.
deshACM, pp. 1–7.Application of Neuro-GA Hybrids in 76. Figueiredo E and Cross E. Linear approaches to modeling
Sensor Optimization for Structural Health Monitoring nonlinearities in long-term monitoring of bridges. J Civ Struct
62. Tan Y and Zhang L. Computational methodologies for op- Health Monit 2013; 3: 187–194.
timal sensor placement in structural health monitoring: A 77. Flexa C, Gomes W and Sales Junior C. Data Normalization in
review. Struct Health Monit 2019: 147592171987757. Structural Health Monitoring by Means of Nonlinear Fil-
63. Soman RN, Wee J, Peters K, et al. Optimization of sensor tering. 8th Brazilian Conference on Intelligent Systems
placement for guided waves based SHM using fiber Bragg (BRACIS). Salvador, Brazil: IEEE, 2019, pp. 204–209.
grating sensors. In: Zonta D and Huang H (eds). Sensors and 78. Oh CK, Sohn H and Bae I-H. Statistical novelty detection
Smart Structures Technologies for Civil, Mechanical, and within the Yeongjong suspension bridge under environmental
Aerospace Systems 2020. Online Only, United States: SPIE, and operational variations. Smart Mater Struct 2009; 18:
p. 51. 125022.
64. Gomes GF, de Almeida FA, da Silva Lopes Alexandrino P, 79. Cremona C and Santos J. Structural health monitoring as a
et al. A multiobjective sensor placement optimization for big-data problem. Struct Eng Int 2018; 28: 243–254.
SHM systems considering Fisher information matrix and 80. Cai L and Zhu Y. The challenges of data quality and data
mode shape interpolation. Eng Comput 2019; 35: 519–535. quality assessment in the big data era. Data Sci J 2015; 14: 2.
65. Yang C, Zhang X, Huang X, et al. Optimal sensor placement 81. Worden K and Manson G. The application of machine
for deployable antenna module health monitoring in SSPS learning to structural health monitoring. Philos Trans R Soc
using genetic algorithm. Acta Astronaut 2017; 140: 213–224. Math Phys Eng Sci 2007; 365: 515–537.
66. Blanloeuil P, Nurhazli NAE and Veidt M. Particle swarm 82. Chai X, Gu H, Li F, et al. Deep learning for irregularly and
optimization for optimal sensor placement in ultrasonic SHM regularly missing data reconstruction. Sci Rep 2020; 10:
systems. In: Yu T, Gyekenyesi AL, Shull PJ, et al. (eds). Las 3302.
Vegas. NV, United States, p. 98040E. 83. Siahkoohi A, Kumar R and Herrmann F. Seismic Data Recon-
67. Mallardo V and Aliabadi MH. Optimal sensor placement for struction with Generative Adversarial Networks. Copenhagen,
structural, damage and impact identification: A review. Struct Denmark. DOI: 10.3997/2214-4609.201801393. Epub ahead of
Durab Health Monit 2013; 9: 287–323. print 11 June 2018
68. Flynn EB and Todd MD. A Bayesian approach to optimal 84. Tan X, Sun X, Chen W, et al. Investigation on the data
sensor placement for structural health monitoring with ap- augmentation using machine learning algorithms in structural
plication to active sensing. Mech Syst Signal Process 2010; health monitoring information. Struct Health Monit 2021; 20:
24: 891–903. 147592172199623.
Malekloo et al. 41

85. Ren P, Chen X, Sun L, et al. Incremental Bayesian matrix/ 100. Van Der Maaten L, Postma E and Van den Herik J. Di-
tensor learning for structural monitoring data imputation and mensionality reduction: a comparative review. J Mach Learn
response forecasting. Mech Syst Signal Process 2021; 158: Res 2009; 10: 66–71.
107734. 101. Dervilis N, Shi H, Worden K, et al. Exploring environmental
86. Chen Z, Li H and Bao Y. Analyzing and modeling inter- and operational variations in SHM data using heteroscedastic
sensor relationships for strain monitoring data and missing Gaussian processes. In: Pakzad S and Juan C (eds). Dynamics
data imputation: a copula and functional data-analytic ap- of Civil Structures, Volume 2. Cham: Springer International
proach. Struct Health Monit 2019; 18: 1168–1188. Publishing; 2016, pp. 145–153.
87. Chen Z, Bao Y, Li H, et al. LQD-RKHS-based distribution-to- 102. Jiménez AA, Garcı́a Márquez FP, Moraleda VB, et al. Linear
distribution regression methodology for restoring the prob- and nonlinear features and machine learning for wind turbine
ability distributions of missing SHM data. Mech Syst Signal blade ice detection and diagnosis. Renew Energ 2019; 132:
Process 2019; 121: 655–674. 1034–1048.
88. Martinez-Luengo M, Shafiee M and Kolios A. Data man- 103. de Silva Vand Tenenbaum JB. Global versus local methods in
agement for structural integrity assessment of offshore wind nonlinear dimensionality reduction in Proceedings of the 15th
turbine support structures: data cleansing and missing data International Conference on Neural Information Processing
imputation. Ocean Eng 2019; 173: 867–883. Systems. Vancouver, Canada: MIT Press, 9–14 December
89. Oh BK, Glisic B, Kim Y, et al. Convolutional neural network– 2002, pp. 721–728.
based data recovery method for structural health monitoring. 104. van der Maaten L and Postma E. Dimensionality Reduction: A
Struct Health Monit 2020: 147592171989757. Comparative Review. Tilburg University. Tilburg, 2009
90. Fan G, Li J and Hao H. Lost data recovery for structural health 105. Langone R, Reynders E, Mehrkanoon S, et al. Automated
monitoring based on convolutional neural networks. Struct structural health monitoring based on adaptive kernel spectral
Control Health Monit; 26. DOI: 10.1002/stc.2433. Epub clustering. Mech Syst Signal Process 2017; 90: 64–78.
ahead of print October 2019 106. Liu J, Chen S, Bergés M, et al. Diagnosis algorithms for
91. Li Y, Bao T, Chen H, et al. A large-scale sensor missing data indirect structural health monitoring of a bridge model via
imputation framework for dams using deep learning and dimensionality reduction. Mech Syst Signal Process 2020;
transfer learning strategy. Measurement 2021; 178: 109377. 136: 106454.
92. Fan G, Li J and Hao H. Vibration signal denoising for 107. Ayesha S, Hanif MK and Talib R. Overview and comparative
structural health monitoring by residual convolutional neural study of dimensionality reduction techniques for high di-
networks. Measurement 2020; 157: 107651. mensional data. Inf Fusion 2020; 59: 44–58.
93. Yang C, Liu J, Zeng Y, et al. Real-time condition monitoring 108. Güemes A, Fernández-López A, Dı́az-Maroto P, et al.
and fault detection of components based on machine-learning Structural health monitoring in composite structures by fiber-
reconstruction model. Renew Energ 2019; 133: 433–441. optic sensors. Sensors 2018; 18: 1094.
94. Batista GEAPA, Prati RC and Monard MC. A study of the 109. Roberts C, Garcia D and Tcherniak D. A comparative study
behavior of several methods for balancing machine on data manipulation in pca-based structural health moni-
learning training data. ACM SIGKDD Explor Newsl 2004; toring systems for removing environmental and operational
6: 20–29. variations. In: Wahab MA (ed). Proceedings of the 13th
95. Chu X, Ilyas IF, Krishnan S, et al. Proceedings of the 2016 International Conference on Damage Assessment of Struc-
International Conference on Management of Data - SIG- tures. Porto, Portugal: 9-10 July 2019, Springer Singapore,
MOD ’16. San Francisco, California, USA: ACM Press, pp. 182–198.
pp. 2201–2206.Data Cleaning: Overview and Emerging 110. Mei Q, Gül M and Boay M. Indirect health monitoring of
Challenges bridges using Mel-frequency cepstral coefficients and prin-
96. Ridzuan F and Wan MNWZ. A review on data cleansing cipal component analysis. Mech Syst Signal Process 2019;
methods for big data. Proced Comput Sci 2019; 161: 119: 523–546.
731–738. 111. Li L, Zhang L, Liu G, et al. Sensor fault localization with
97. Li H, Ou J, Zhao X, et al. Structural health monitoring system accumulated residual contribution rate for bridge SHM. Clust
for the Shandong Binzhou Yellow River highway bridge. Comput 2019; 22: 7821–7829.
Comput-Aided Civ Infrastruct Eng 2006; 21: 306–317. 112. Akintunde E, Azam SE, Rageh A, et al. Full scale bridge
98. Verleysen M and François D. The Curse of Dimensionality in damage detection using sparse sensor networks, principal
Data Mining and Time Series Prediction. In: Cabestany J, component analysis, and novelty detection. Proceedings
Prieto A and Sandoval F (eds). Computational Intelligence 2019; 42: 34.
and Bioinspired Systems. Berlin, Heidelberg: Springer Berlin 113. Deraemaeker A and Worden K. A comparison of linear
Heidelberg, pp. 758–770. approaches to filter out environmental effects in structural
99. Bellman RE. Adaptive control processes: a guided tour. NJ, health monitoring. Mech Syst Signal Process 2018; 105:
USA: Princeton University Press, 1961. 1–15.
42 Structural Health Monitoring 0(0)

114. Kullaa J. Structural health monitoring under nonlinear en- Mechanical, and Aerospace Systems Lynch JP (ed). Las
vironmental or operational influences. Shock Vib 2014; 2014: Vegas. NV, US, 20 April 2016, p. 98032H.
1–9. 128. Jiménez AA, Zhang L, Gómez Muñoz CQ, et al. Maintenance
115. Garcı́a-Macı́as E and Ubertini F. MOVA/MOSS: two inte- management based on Machine Learning and nonlinear
grated software solutions for comprehensive structural health features in wind turbines. Renew Energ 2020; 146: 316–328.
monitoring of structures. Mech Syst Signal Process 2020; 129. Mishra S. Structural health monitoring with lamb-wave
143: 106830. sensors: problems in damage monitoring, prognostics and
116. Huang H-B, Yi T-H and Li H-N. Anomaly identification of multisensory decision fusion, 2016, http://purl.flvc.org/fsu/
structural health monitoring data using dynamic independent fd/FSU_2016SU_Mishra_fsu_0071E_13346.
component analysis. J Comput Civ Eng 2020; 34: 04020025. 130. Guo F, Shang C, Huang B, et al. Monitoring of operating
117. Li L, Liu G, Zhang L, et al. Sensor fault detection with point and process dynamics via probabilistic slow feature
generalized likelihood ratio and correlation coefficient for analysis. Chemom Intell Lab Syst 2016; 151: 115–125.
bridge SHM. J Sound Vib 2019; 442: 445–458. 131. Zhao C and Huang B. A full-condition monitoring method for
118. Zhu Y, Ni Y-Q, Jin H, et al. A temperature-driven MPCA nonstationary dynamic chemical processes with cointegration
method for structural anomaly detection. Eng Struct 2019; and slow feature analysis. AIChE J 2018; 64: 1662–1681.
190: 447–458. 132. Li S, He S, Li H, et al. Scour depth determination of bridge
119. Yao X-J, Yi T-H, Qu C, et al. Blind modal identification in piers based on time-varying modal parameters: application to
frequency domain using independent component analysis for Hangzhou bay bridge. J Bridge Eng 2017; 22: 04017107.
high damping structures with classical damping: blind modal 133. Hsu T-Y and Loh C-H. Damage detection accommodating
identification in frequency domain. Comput-Aided Civ In- nonlinear environmental effects by nonlinear principal
frastruct Eng 2018; 33: 35–50. component analysis. Struct Control Health Monit 2009: 17,
120. Zhang W, Sun LM and Sun SW. Bridge-deflection estimation 338–354
through inclinometer data considering structural damages. 134. Ye X, Wu Y, Zhang L, et al. Ambient effect filtering using
J Bridge Eng 2017; 22: 04016117. NLPCA-SVR in high-rise buildings. Sensors 2020; 20: 1143.
121. Avendaño-Valencia LD, Chatzi EN, Koo KY, et al. Gaussian 135. Tibaduiza D, Torres-Arredondo MÁ, Vitola J, et al. A damage
process time-series models for structures under operational classification approach for structural health monitoring using
variability. Front Built Environ 2017; 3: 69. machine learning. Complexity 2018: 1–14.2018
122. Hu W-H, Cunha Á, Caetano E, et al. Comparison of different 136. Silva M, Santos A, Santos R, et al. Deep principal component
statistical approaches for removing environmental/operational analysis: an enhanced approach for structural damage iden-
effects for massive data continuously collected from foot- tification. Struct Health Monit 2019; 18: 1444–1463.
bridges: statistical approaches for removing environmental/ 137. Borate P, Wang G and Wang Y. Data-driven structural health
operational effects. Struct Control Health Monit 2017; 24: monitoring approach using guided lamb wave responses.
e1955. J Aerosp Eng 2020; 33: 04020033.
123. Yanez-Borjas JJ, Camarena-Martinez D, Valtierra-Rodriguez 138. Yang S, Jiang X, Xu S, et al. Information theory based
M, et al. Methodology based on statistical features and linear probabilistic approach to blade damage detection of turbo-
discriminant analysis for damage detection in a truss-type machine using sensor data. IEEE Trans Ind Electron 2019:
bridge IEEE International Autumn Meeting on Power, 67, 10887–10896.
Electronics and Computing (ROPEC). Ixtapa, Mexico, 13–15 139. Fuentes R, Howard T, Marshall M, et al. Detecting damage on
November 2019, pp. 1–6.IEEE wind turbine bearings using acoustic emissions and gaussian
124. Mboo CP and Hameyer K. Fault diagnosis of bearing damage process latent variable models, Structural health monitoring.
by means of the linear discriminant analysis of stator current Destech Publications, Lancaster, 2015.Epub ahead of print
features from the frequency selection. IEEE Trans Ind Appl 2015
2016; 52: 3861–3868. 140. Jeong M, Choi J-H and Koh B-H. Isomap-based damage
125. Mangalathu S, Hwang S-H, Choi E, et al. Rapid seismic classification of cantilevered beam using modal frequency
damage evaluation of bridge portfolios using machine changes. Struct Control Health Monit 2014; 21: 590–602.
learning techniques. Eng Struct 2019; 201: 109785. 141. Sun LM, Zhang W and Nagarajaiah S. Bridge real-time
126. Zheng W and Qian F. Promptly assessing probability of damage identification method using inclination and strain
barge–bridge collision damage of piers through probabilistic- measurements in the presence of temperature variation.
based classification of machine learning. J Civ Struct Health J Bridge Eng 2019; 24: 04018111.
Monit 2017; 7: 57–78. 142. Chaabane M, Mansouri M, Nounou H, et al. Damage de-
127. Mishra S, Vanli OA, Huffer FW, et al. Regularized dis- tection in structural health monitoring using kernel PLS based
criminant analysis for multi-sensor decision fusion and GLR, International Conference on Advanced Technologies
damage detection with Lamb waves. In: Proc. SPIE 9803, for Signal and Image Processing (ATSIP). IEEE, Fez, Mo-
Sensors and Smart Structures Technologies for Civil, rocco, 22–24 May 2017, pp. 1–5.
Malekloo et al. 43

143. Ghoulem K, Kormi T and Bel Hadj Ali N. Damage detection 158. Feng M, Fukuda Y, Mizuta M, et al. Citizen sensors for SHM:
in nonlinear civil structures using kernel principal component use of accelerometer data from smartphones. Sensors 2015;
analysis. Adv Struct Eng 2020: 23, 136943322091320. 15: 2980–2998.
144. Dervilis N, Simpson TE, Wagg DJ, et al. Nonlinear modal 159. Djili S, Moulin E, Assaad J, et al. Structural health monitoring
analysis via non-parametric machine learning tools. Strain of hollow cylinder using cross-correlation of ambient noise
2019; 55: e12297. field. NDT E Int 2019; 101: 87–93.
145. Xiao H, Lou S and Ogai H. A novel bridge structure damage 160. Lin J-F, Wang J, Wang L-X, et al. Structural damage
diagnosis algorithm based on post-nonlinear ICA and sta- diagnosis-oriented impulse response function estimation
tistical pattern recognition. IEEJ Trans Electr Electron Eng under seismic excitations. Sensors 2019; 19: 5413.
2015; 10: 287–300. 161. Kessler SS, Spearing SM, Atalla MJ, et al. Damage detection
146. Dervilis N, Wagg DJ, Green PL, et al. .Nonlinear modal in composite materials using frequency response methods.
analysis using pattern recognition Proceedings of ISMA Compos B Eng 2002; 33: 87–95.
2014. Le Mans, France, 7–12 July 2014, pp. 3017–3027 162. Brincker R, Zhang L and Andersen P. Modal identification of
147. Dervilis N, Antoniadou I, Cross EJ, et al. A non-linear manifold output-only systems using frequency domain decomposition.
strategy for SHM approaches. Strain 2015; 51: 324–331. Smart Mater Struct 2001; 10: 441–445.
148. Sun C, Wang P, Yan R, et al. Machine health monitoring 163. Staszewski WJ and Robertson AN. Time-frequency and time-
based on locally linear embedding with kernel sparse rep- scale analyses for structural health monitoring. Philos Trans R
resentation for neighborhood optimization. Mech Syst Signal Soc Math Phys Eng Sci 2007; 365: 449–477.
Process 2019; 114: 25–34. 164. Nagarajaiah S and Basu B. Output only modal identification
149. Nguyen HD, Bui TT, De Roeck G, et al. Damage detection in and structural damage detection using time frequency &
simply supported beam using transmissibility and auto- wavelet techniques. Earthq Eng Eng Vib 2009; 8: 583–605.
associative neural network. In: Abdel Wahab M (ed). Pro- 165. Pyayt A, Kozionov A, Mokhov I, et al. Time-frequency
ceedings of the 1st International Conference on Numerical methods for structural health monitoring. Sensors 2014;
Modelling in Engineering. Ghent, Belgium: Springer; 28-29 14: 5147–5173.
August 2019, pp. 177–186. 166. Pan H, Azimi M, Yan F, et al. Time-frequency-based data-
150. Zhang H, Gül M and Kostić B. Eliminating temperature driven structural diagnosis and damage detection for cable-
effects in damage detection for civil infrastructure using time stayed bridges. J Bridge Eng 2018; 23: 04018033.
series analysis and autoassociative neural networks. J Aerosp 167. Taha MMR, Noureldin A, Lucero JL, et al. Wavelet transform
Eng 2019; 32: 04019001. for structural health monitoring: a compendium of uses and
151. Ma M, Qin Y, Haile M, et al. Direct waveform extraction via a features. Struct Health Monit Int J 2006; 5: 267–295.
deep recurrent denoising autoencoder. In: Gyekenyesi AL 168. Khalid S, Khalil T and Nasreen S. A survey of feature se-
(ed). Nondestructive characterization and monitoring of lection and feature extraction techniques in machine learning,
advanced materials, aerospace, civil infrastructure, and Science and Information Conference. London, UK: IEEE,
transportation XIII. Denver: SPIE, p. 61. 2019 27–29 August 2014, pp. 372–378.
152. Wang L, Zhang Z, Xu J, et al. Wind turbine blade breakage 169. Marwala T. Finite-element-model updating using compu-
monitoring with deep autoencoders. IEEE Trans Smart Grid tional intelligence techniques. London: Springer London.
2018; 9: 2824–2833. 2010 Epub ahead of print 2010
153. Mitra M and Gopalakrishnan S. Guided wave based structural 170. Chandrashekar G and Sahin F. A survey on feature selection
health monitoring: a review. Smart Mater Struct 2016; 25: methods. Comput Electr Eng 2014; 40: 16–28.
053001. 171. Ying Y, Garrett JH, Oppenheim IJ, et al. Toward data-driven
154. Peter Carden E and Brownjohn JMW. ARMA modelled time- structural health monitoring: application of machine learning
series classification for structural health monitoring of civil and signal processing to damage detection. J Comput Civ Eng
infrastructure. Mech Syst Signal Process 2008; 22: 295–314. 2013; 27: 667–680.
155. Yang Z, Yu Z and Sun H. On the cross correlation function 172. Witten IH, Frank E, Hall MA, et al. Data mining: practical
amplitude vector and its application to structural damage machine learning tools and techniques. Burlington, MA,
detection. Mech Syst Signal Process 2007; 21: 2918–2932. Morgan Kaufmann, 2016.
156. Loh C-H, Chen M-C and Chao S-H. Stochastic subspace 173. Bolón-Canedo V and Alonso-Betanzos A. Ensembles for
identification for operational modal analysis of an arch bridge. feature selection: a review and future trends. Inf Fusion 2019;
In: Tomizuka M, Yun C-B and Lynch JP (eds). SPIE Smart 52: 1–12.
Structures and Materials and Nondestructive Evaluation and 174. Solorio-Fernández S, Carrasco-Ochoa JA and Martı́nez-
Health Monitoring, San Diego, CA, US, April 2012, p. 834504. Trinidad JF. A review of unsupervised feature selection
157. Loh CH and Chen MC. Modeling of environmental effects for methods. Artif Intell Rev 2020; 53: 907–948.
vibration-based SHM using recursive stochastic subspace 175. Bull LA, Worden K, Fuentes R, et al. Outlier ensembles: a
identification analysis. Key Eng Mater 2013; 558: 52–64. robust method for damage detection and unsupervised feature
44 Structural Health Monitoring 0(0)

extraction from high-dimensional data. J Sound Vib 2019; 188. Park C, Tang J and Ding Y. Aggressive data reduction for
453: 126–150. damage detection in structural health monitoring. Struct
176. Zugasti E, Mujica LE, Anduaga J, et al. Feature selection - Health Monit Int J 2010; 9: 59–74.
extraction methods based on PCA and mutual information to 189. Zhou Q, Zhou H, Zhou Q, et al. Structure damage detection
improve damage detection problem in offshore wind turbines. based on random forest recursive feature elimination. Mech
Key Eng Mater 2013; 569-570: 620–627. Syst Signal Process 2014; 46: 82–90.
177. Sousa Tomé E, Pimentel M and Figueiras J. Online early 190. Mustapha S, Braytee A and Ye L. Detection of surface
damage detection and localisation using multivariate data cracking in steel pipes based on vibration data using a multi-
analysis: application to a cable-stayed bridge. Struct Control class support vector machine classifier. In: Lynch JP (ed).
Health Monit; 26. e2434, 2019.Epub ahead of print No- SPIE Smart Structures and Materials and Nondestructive
vember 2019 Evaluation and Health Monitoring,, Portland, OR, US; April
178. Monaco E, Memmolo V, Boffa ND, et al. Guided waves 2017.
based SHM systems: parameters selection for better identi- 191. Ni F, Zhang J and Noori MN. Deep learning for data anomaly
fication and localisation of damages in composites stiffened detection and data compression of a long-span suspension
plates. In: Kundu T (ed). SPIE Smart Structures and Materials bridge. Comput-Aided Civ Infrastruct Eng 2020; 35: 685–700.
and Nondestructive Evaluation and Health Monitoring, 192. Smith CB and Hernandez EM. Detection of spatially sparse
Portland, OR, US; April 2017. damage using impulse response sensitivity and LASSO
179. Trendafilova I, Heylen W and Brussel HV. Measurement regularization. Inverse Probl Sci Eng 2019; 27: 1–16.
point selection in damage detection using the mutual infor- 193. Michau G, Chao MA and Fink O. Feature selecting hierar-
mation concept. Smart Mater Struct 2001; 10: 528–533. chical neural network for industrial system health monitoring:
180. Qian F and Niu G. Remaining useful life prediction using catching informative features with LASSO, Proceedings of
ranking mutual information based monotonic health indica- the Annual Conference of the PHM Society. PHM Society,
tor, Prognostics and System Health Management Conference Philadephia, PA, US, 24–27 September 2018, p. 494.
(PHM). Beijing, China: IEEE, 21–23 October 2015, pp. 1–5. 194. Ma S, Chu F and Han Q. Deep residual learning with de-
181. Zhao X, Zuo MJ and Patel T. EMD, ranking mutual information modulated time-frequency features for fault diagnosis of
and PCA based condition monitoring. Volume 5: 22nd Inter- planetary gearbox under nonstationary running conditions.
national Conference on Design Theory and Methodology; Mech Syst Signal Process 2019; 127: 190–201.
Special Conference on Mechanical Vibration and Noise. 195. Luo RC, Yih C-C and Su KL. Multisensor fusion and inte-
Montreal, Canada: 15–18 August 2010, ASMEDC, pp. 777–782 gration: approaches, applications, and future research di-
182. Sbarufatti C. Optimization of an artificial neural network for rections. IEEE Sens J 2002; 2: 107–119.
fatigue damage identification using analysis of variance: 196. Chatzis MN, Chatzi EN and Smyth AW. On the observability
ANN structure optimisation using analysis of error variance. and identifiability of nonlinear structural and mechanical
Struct Control Health Monit 2017; 24: e1964. systems. Struct Control Health Monit 2015; 22: 574–593.
183. Kessler SS and Agrawal P. Application of pattern recognition 197. Chatzis MN, Chatzi EN and Smyth AW. An experimental
for damage classification in composite laminates, Proceed- validation of time domain system identification methods with
ings of the 6th International Workshop on Structural Health fusion of heterogeneous data. Earthq Eng Struct Dyn 2015;
Monitoring. Citeseer: Stanford University, Stanford, CA, US, 44: 523–547.
11–13 September 2007. 198. Maes K, Chatzis MN and Lombaert G. Observability of
184. Babajanian Bisheh H, Ghodrati Amiri G, Nekooei M, et al. nonlinear systems with unmeasured inputs. Mech Syst Signal
Damage detection of a cable-stayed bridge using feature Process 2019; 130: 378–394.
extraction and selection methods. Struct Infrastruct Eng 199. Wu R-T and Jahanshahi MR. Data fusion approaches for
2019; 15: 1165–1177. structural health monitoring and system identification: past,
185. Yan X and Jia M. Intelligent fault diagnosis of rotating present, and future. Struct Health Monit 2020; 19: 552–586.
machinery using improved multiscale dispersion entropy and 200. Bishop C. Pattern Recognition and Machine Learning. New
mRMR feature selection. Knowl-Based Syst 2019; 163: York, NY: Springer-Verlag, 2006, https://www.springer.com/
450–471. gp/book/9780387310732.
186. Hoell S and Omenzetter P. Improved damage detectability in a 201. Worden K and Dulieu-Barton JM. An overview of intelligent
wind turbine blade by optimal selection of vibration signal fault detection in systems and structures. Struct Health Monit
correlation coefficients. Struct Health Monit Int J 2016; 15: Int J 2004; 3: 85–98.
685–705. 202. Worden K. Structural fault detection using a novelty measure.
187. Yun GJ, Ogorzalek KA, Dyke SJ, et al. A parameter subset J Sound Vib 1997; 201: 85–101.
selection method using residual force vector for detecting 203. Bull L, Worden K, Manson G, et al. Active learning for semi-
multiple damage locations. Struct Control Health Monit supervised structural health monitoring. J Sound Vib 2018;
2010; 17: 48–67. 437: 373–388.
Malekloo et al. 45

204. Li S, Laima S and Li H. Data-driven modeling of vortex- 218. Sousa H, Rozsas A, Slobbe A, et al. A novel pro-active
induced vibration of a long-span suspension bridge using approach towards SHM-based bridge management supported
decision tree learning and support vector regression. J Wind by FE analysis and Bayesian methods. Struct Infrastruct Eng
Eng Ind Aerodyn 2018; 172: 196–211. 2020; 16: 233–246.
205. Gordan M, Razak HA, Ismail Z, et al. A hybrid ANN-based 219. Mangalathu S, Jang H, Hwang S-H, et al. Data-driven
imperial competitive algorithm methodology for structural machine-learning-based seismic failure mode identification
damage identification of slab-on-girder bridge using data of reinforced concrete shear walls. Eng Struct 2020; 208:
mining. Appl Soft Comput 2020; 88: 106013. 110331.
206. Zhang L, Wang Z, Wang L, et al. Machine learning based real- 220. Nazarian E, Taylor T, Weifeng T, et al. Machine-learning-
time visible fatigue crack growth detection. Digit Commun based approach for post event assessment of damage in a turn-
Netw 2021.S2352864821000146 of-the-century building structure. J Civ Struct Health Monit
207. Mariniello G, Pastore T, Menna C, et al. Structural damage 2018; 8: 237–251.
detection and localization using decision tree ensemble and 221. Soyoz S, Feng MQ and Shinozuka M. Structural reliability
vibration data. Comput-Aided Civ Infrastruct Eng, 2020, estimation with vibration-based identified parameters. J Eng
p. 12633. Mech 2010; 136: 100–106.
208. Laory I, Trinh TN, Smith IFC, et al. Methodologies for 222. Lee S-H and Song J. Bayesian-network-based system iden-
predicting natural frequency variation of a suspension bridge. tification of spatial distribution of structural parameters. Eng
Eng Struct 2014; 80: 211–221. Struct 2016; 127: 260–277.
209. Lu Q, Zhu J and Zhang W. Quantification of fatigue damage 223. Tubaldi E, Ozer E, Douglas J, et al. Examining the contri-
for structural details in slender coastal bridges using machine bution of near real-time data for rapid seismic loss assessment
learning-based methods. J Bridge Eng 2020; 25: 04020033. of structures. Struct Health Monit 2021: 147592172199621.
210. Chencho, Li J, Hao H, et al. Development and application of 224. Au S-K, Zhang F-L and Ni Y-C. Bayesian operational modal
random forest technique for element level structural damage analysis: theory, computation, practice. Comput Struct 2013;
quantification. Struct Control Health Monit 28. e2678. 2021 126: 3–14.
Epub ahead of print March 2021 225. Tran TTX and Ozer E. Synergistic bridge modal analysis
211. Deng Y, Zhang M, Feng D-M, et al. Predicting fatigue using frequency domain decomposition, observer Kalman
damage of highway suspension bridge hangers using weigh- filter identification, stochastic subspace identification, system
in-motion data and machine learning. Struct Infrastruct Eng, realization using information matrix, and autoregressive
17, 2020: 233–248. exogenous model. Mech Syst Signal Process 2021; 160:
212. Kohiyama M, Oka K and Yamashita T. Detection method of 107818.
unlearned pattern using support vector machine in damage 226. Zheng Y, Xie Y and Long X. A comprehensive review of
classification based on deep neural network. Struct Control Bayesian statistics in natural hazards engineering. Nat
Health Monit 2020: 27, e2552. Hazards. 108, 63–91, 2021.Epub ahead of print 12 April
213. Chong JW, Kim Y and Chon KH. Nonlinear multiclass 2021
support vector machine-based health monitoring system for 227. Fang X, Luo H and Tang J. Structural damage detection using
buildings employing magnetorheological dampers. J Intell neural network with learning rate improvement. Comput
Mater Syst Struct 2014; 25: 1456–1468. Struct 2005; 83: 2150–2161.
214. Li S, Zuo X, Li Z, et al. Applying deep learning to continuous 228. Hekmati Athar SP, Taheri M, Secrist J, et al. Neural network
bridge deflection detected by fiber optic gyroscope for for structural health monitoring with combined direct and
damage detection. Sensors 2020; 20: 911. indirect methods. J Appl Remote Sens 2020; 14: 1.
215. Dogan G, Arslan MH and Baykan OK. Determination of 229. Malekjafarian A, Golpayegani F, Moloney C, et al. A machine
damage levels of RC columns with a smart system oriented learning approach to bridge-damage detection using responses
method. Bull Earthq Eng 2020; 18: 3223–3245. measured on a passing vehicle. Sensors 2019; 19: 4035.
216. Salehi H, Biswas S and Burgueño R. Data interpretation 230. Tibaduiza DA, Mujica LE and Rodellar J. Damage classifi-
framework integrating machine learning and pattern recog- cation in structural health monitoring using principal com-
nition for self-powered data-driven damage identification ponent analysis and self-organizing maps. Struct Control
with harvested energy variations. Eng Appl Artif Intell 2019; Health Monit 2013; 20: 1303–1316.
86: 136–153. 231. Avci O, Abdeljaber O, Kiranyaz S, et al. Structural health
217. Salehi H, Das S, Chakrabartty S, et al. A methodology for monitoring with self-organizing maps and artificial neural
structural health diagnosis and assessment using machine networks. In: Mains ML and Dilworth BJ (eds). Topics in
learning with noisy and incomplete data from self-powered modal analysis & testing, Volume 8. 2020, Cham: Springer
wireless sensors. In: Sohn H (ed). Sensors and Smart International Publishing, pp. 237–246.
Structures Technologies for Civil, Mechanical, and Aero- 232. Mousavi M and Gandomi AH. Deep learning for structural
space Systems 2018. Denver, CO: SPIE, p. 31. health monitoring under environmental and operational
46 Structural Health Monitoring 0(0)

variations. In: Yu T-Y and Gyekenyesi AL (eds). Nonde- 246. Jin Q, Liu Z, Bin J, et al. Predictive analytics of in-service
structive Characterization and Monitoring of Advanced bridge structural performance from SHM data mining per-
Materials, Aerospace, Civil Infrastructure, and Trans- spective: a case study. Shock Vib 2019; 2019: 1–11.
portation XV. 2021. Bellingham, WA: SPIE, p. 16. 247. Guéguen P and Tiganescu A. Consideration of the effects of air
233. Mousavi M and Gandomi AH. Prediction error of Johansen temperature on structural health monitoring through traffic light-
cointegration residuals for structural health monitoring. Mech based decision-making tools. Shock Vib 2018; 2018: 1–12.
Syst Signal Process 2021; 160: 107847. 248. Bull LA, Worden K and Dervilis N. Towards semi-supervised
234. Zhang R, Liu Y and Sun H. Physics-guided convolutional and probabilistic classification in structural health monitor-
neural network (PhyCNN) for data-driven seismic response ing. Mech Syst Signal Process 2020; 140: 106653.
modeling. Eng Struct 2020; 215: 110704. 249. Rogers TJ, Worden K, Fuentes R, et al. A Bayesian non-
235. Zhang R, Chen Z, Chen S, et al. Deep long short-term parametric clustering approach for semi-supervised structural
memory networks for nonlinear structural seismic response health monitoring. Mech Syst Signal Process 2019; 119:
prediction. Comput Struct 2019; 220: 55–68. 100–119.
236. Li T, Pan Y, Tong K, et al. Attention-based sequence-to- 250. Chen S, Cerda F, Rizzo P, et al. Semi-supervised multi-
sequence learning for online structural response forecasting resolution classification using adaptive graph filtering with
under seismic excitation. IEEE Trans Syst Man Cybern Syst application to indirect bridge structural health monitoring.
2021: 1–17. IEEE Trans Signal Process 2014; 62: 2879–2893.
237. Choe D-E, Kim H-C and Kim M-H. Sequence-based mod- 251. Yang D, Yi C, Xu Z, et al. Improved tensor-based singular
eling of deep learning with LSTM and GRU networks for spectrum analysis based on single channel blind source
structural damage detection of floating offshore wind turbine separation algorithm and its application to fault diagnosis.
blades. Renew Energ 2021; 174: 218–235. Appl Sci 2017; 7: 418.
238. Tian Y, Xu Y, Zhang D, et al. Relationship modeling between 252. Musafere F, Sadhu A and Liu K. Towards damage detection
vehicle-induced girder vertical deflection and cable tension using blind source separation integrated with time-varying
by BiLSTM using field monitoring data of a cable-stayed auto-regressive modeling. Smart Mater Struct 2016; 25:
bridge. Struct Control Health Monit 28. e2667, 2021.Epub 015013.
ahead of print February 2021 253. Chien J.. Source separation and machine learning. Academic
239. Tsialiamanis G, Chatzi E, Dervilis N, et al. An application of Press, Cambridge, MA, 2018.
generative adversarial networks in structural health moni- 254. Sadhu A, Hu B and Narasimhan S. Blind source separation
toring. XI International Conference on Structural Dynamics, towards decentralized modal identification using compressive
Athens, Greece, pp. 3816–3831. November 2020 sampling, 11th International Conference on Information
240. Fan G, Li J, Hao H, et al. Data driven structural dynamic Science, Signal Processing and their Applications (ISSPA).
response reconstruction using segment based generative Montreal, Canada: IEEE, 2–5 July 2012, pp. 1147–1152.
adversarial networks. Eng Struct 2021; 234: 111970. 255. Liu D, Tang Z, Bao Y, et al. Machine-learning-based methods
241. Diez A, Khoa NLD, Makki Alamdari M, et al. A clustering for output only structural modal identification. Ar-
approach for structural health monitoring on bridges. J Civ Xiv200407644 Cs Eess Stat, 2020, http://arxiv.org/abs/2004.
Struct Health Monit 2016; 6: 429–445. 07644 accessed 12 October 2020).
242. Alamdari MM, Rakotoarivelo T and Khoa NLD. A spectral- 256. Ye X, Chen X, Lei Y, et al. An integrated machine learning
based clustering for structural health monitoring of the algorithm for separating the long-term deflection data of
Sydney harbour bridge. Mech Syst Signal Process 2017; 87: prestressed concrete bridges. Sensors 2018; 18: 4070.
384–400. 257. Atamturktur S, Farajpour I, Prabhu S, et al. Adaptively
243. Burrello A, Brunelli D, Malavisi M, et al. Enhancing weighted support vector regression: prognostic application to
structural health monitoring with vehicle identification and a historic masonry fort. J Perform Constr Facil 2015; 29:
tracking, IEEE International Instrumentation and Measure- 04014057.
ment Technology Conference (I2MTC). Dubrovnik, Croatia: 258. Wu Y, Yuan M, Dong S, et al. Remaining useful life esti-
IEEE, 25–28 May 2020, pp. 1–6. mation of engineered systems using vanilla LSTM neural
244. Jia H, Ding S, Xu X, et al. The latest research progress on networks. Neurocomputing 2018; 275: 167–179.
spectral clustering. Neural Comput Appl 2014; 24: 259. Das S, Dutta S, Putcha C, et al. A data-driven physics-
1477–1486. informed method for prognosis of infrastructure systems:
245. Farhidzadeh A, Dehghan-Niri E and Salamone S. Gaussian theory and application to crack prediction. ASCE-ASME J
mixture modeling of acoustic emissions for structural health Risk Uncertain Eng Syst Part Civ Eng 2020; 6: 04020013.
monitoring of reinforced concrete structures. In: Lynch JP, 260. Prakash G, Narasimhan S and Al-Hammoud R. A two-phase
Yun C-B and Wang K-W (eds). SPIE Smart Structures and model to predict the remaining useful life of corroded re-
Materials + Nondestructive Evaluation and Health Moni- inforced concrete beams. J Civ Struct Health Monit 2019; 9:
toring, San Diego, CA, US: April 2013, p. 86920B. 183–199.
Malekloo et al. 47

261. Lei Y, Li N, Guo L, et al. Machinery health prognostics: a 278. Ozer E and Feng MQ. Structural reliability estimation with
systematic review from data acquisition to RUL prediction. participatory sensing and mobile cyber-physical structural
Mech Syst Signal Process 2018; 104: 799–834. health monitoring systems. Appl Sci 2019; 9: 2840.
262. Zhang Y, Hutchinson P, Lieven NAJ, et al. Remaining useful 279. Bibri SE. The IoT for smart sustainable cities of the future: an
life estimation using long short-term memory neural networks analytical framework for sensor-based big data applications
and deep fusion. IEEE Access 2020; 8: 19033–19045. for environmental sustainability. Sustain Cities Soc 2018; 38:
263. Al-Turjman F and Malekloo A. Smart parking in IoT-enabled 230–253.
cities: a survey. Sustain Cities Soc 2019; 49: 101608. 280. Lan C, Li H and Ou J. Traffic load modelling based on
264. Xu J, Liu H and Han Q. Blockchain technology and smart structural health monitoring data. Struct Infrastruct Eng 2011;
contract for civil structural health monitoring system. 7: 379–386.
Comput-Aided Civ Infrastruct Eng 2021: 12666. 281. Jin J, Gubbi J, Marusic S, et al. An information framework for
265. Jagadish HV. Big data and science: myths and reality. Big creating a smart city through internet of things. IEEE Internet
Data Res 2015; 2: 49–52. Things J 2014; 1: 112–121.
266. Hilbert M. Big data for development: a review of promises 282. Praticò FG, Fedele R, Naumov V, et al. Detection and
and challenges. Dev Pol Rev 2016; 34: 135–174. monitoring of bottom-up cracks in road pavement using a
267. Bull LA, Gardner PA, Gosliga J, et al. Foundations of machine-learning approach. Algorithms 2020; 13: 81.
population-based SHM, part I: homogeneous populations and 283. Huang T.. Big data driven diagnostics for intelligent trans-
forms. Mech Syst Signal Process 2021; 148: 107141. portation systems. Iowa State University, Ames, IA, 2018,
268. Gosliga J, Gardner PA, Bull LA, et al. Foundations of https://lib.dr.iastate.edu/etd/17211.
population-based SHM, part II: heterogeneous populations— 284. Khan SM, Atamturktur S, Chowdhury M, et al. Integration of
graphs, networks, and communities. Mech Syst Signal Pro- structural health monitoring and intelligent transportation
cess 2021; 148: 107144. systems for bridge condition assessment: current status and
269. Gardner P, Bull LA, Gosliga J, et al. Foundations of population- future direction. IEEE Trans Intell Transp Syst 2016; 17:
based SHM, part III: heterogeneous populations—mapping 2107–2122.
and transfer. Mech Syst Signal Process 2021; 149: 107142. 285. Higgins S. Inspection of steel-tired wheels. Proc N Y Railr
270. Tsialiamanis G, Mylonas C, Chatzi E, et al. Foundations of Club 1895; 5: 988–989.
population-based SHM, part IV: the geometry of spaces of 286. Crevier D. AI: the tumultuous history of the search for ar-
structures and their feature spaces. Mech Syst Signal Process tificial intelligence. New York, NY: Basic Books, 1993.
2021; 157: 107692. 287. Adeli H and Yeh C. Perceptron learning in engineering de-
271. Liang Y, Wu D, Liu G, et al. Big data-enabled multiscale sign. Comput-Aided Civ Infrastruct Eng 1989; 4: 247–256.
serviceability analysis for aging bridges. Digit Commun Netw 288. Stephen GA, Brownjohn JMW and Taylor CA. Measurements
2016; 2: 97–107. of static and dynamic displacement from visual monitoring of
272. Wang T, Bhuiyan MZA, Wang G, et al. Big data reduction for the Humber bridge. Eng Struct 1993; 15: 197–208.
a smart city’s critical infrastructural health monitoring. IEEE 289. Olaszek P. Investigation of the dynamic characteristic of
Commun Mag 2018; 56: 128–133. bridge structures using a computer vision method. Mea-
273. Entezami A, Sarmadi H, Behkamal B, et al. Big data analytics surement 1999; 25: 227–236.
and structural health monitoring: a statistical pattern 290. Theiler M and Smarsly K. IFC monitor—an IFC schema
recognition-based approach. Sensors 2020; 20: 2328. extension for modeling structural health monitoring systems.
274. Du R, Santi P, Xiao M, et al. The sensable city: a survey on the Adv Eng Inform 2018; 37: 54–65.
deployment and management for smart city monitoring. IEEE 291. Jo B, Khan R and Lee Y-S. Hybrid blockchain and internet-
Commun Surv Tutor 2019; 21: 1533–1560. of-things network for underground structure health moni-
275. Luckey D, Fritz H, Legatiuk D, et al. Artificial intelligence toring. Sensors 2018; 18: 4268.
techniques for smart city applications. In: Toledo Santos E 292. Booyse W, Wilke DN and Heyns S. Deep digital twins for
and Scheer S (eds). Proceedings of the 18th International detection, diagnostics and prognostics. Mech Syst Signal
Conference on Computing in Civil and Building Engineering. Process 2020; 140: 106612.
São Paulo, Brazil: Springer International Publishing; 18–20 293. Al-Turjman F, Abujubbeh M and Malekloo A. Deployment
August 2021, pp. 3–15. strategies for drones in the IoT Era: a survey. In: Drones in
276. Bhuiyan MZA, Wu J, Wang G, et al. Towards cyber-physical IoT-enabled spaces. Boca Raton, FL: CRC Press/Taylor &
systems design for structural health monitoring: hurdles and Francis Group, 2019. pp. 7–42.
opportunities. ACM Trans Cyber-Phys Syst 2017; 1: 1–26. 294. Malekloo A, Ozer E and Al-Turjman F. Combination of GIS
277. Furinghetti M, Pavese A, Lunghi F, et al. Strategies of and SHM in prognosis and diagnosis of bridges in
structural health monitoring for bridges based on cloud earthquake-prone locations, Smart Grid in IoT-Enabled
computing. J Civ Struct Health Monit 2019; 9: 607–616. Spaces. CRC Press, Boca Raton, FL, pp. 139–158.
48 Structural Health Monitoring 0(0)

295. Al-Turjman F, Abujubbeh M, Malekloo A, et al. UAVs as- 311. Ozer E, Feng M and Feng D. Citizen sensors for SHM: towards
sessment in software-defined IoT networks: an overview. a crowdsourcing platform. Sensors 2015; 15: 14591–14614.
Comput Commun 2020; 150: 519–536. 312. Ozer E and Feng MQ. Synthesizing spatiotemporally sparse
296. Chen G, Liang Q, Zhong W, et al. Homography-based smartphone sensor data for bridge modal identification. Smart
measurement of bridge vibration using UAV and DIC Mater Struct; 25. 085007, 2016.Epub ahead of print 1 August
method. Measurement 2021; 170: 108683. 2016
297. Gomez JA, Ozdagli AI and Moreu F. Reference-free dynamic 313. Ozer E, Feng D and Feng MQ. Hybrid motion sensing and
displacements of railroad bridges using low-cost sensors. experimental modal analysis using collocated smartphone
J Intell Mater Syst Struct 2019; 30: 1291–1305. camera and accelerometers. Meas Sci Technol 2017; 28:
298. Liu B, Ozdagli AI, Moreu F, et al. Hybrid reference-free total 105903.
displacement for railroad bridge campaign monitoring. Meas 314. Luo L, Feng MQ, Wu J, et al. Autonomous pothole detection
Sci Technol 2019; 30: 095901. using deep region-based convolutional neural network with
299. Garg P, Moreu F, Ozdagli A, et al. Noncontact dynamic cloud computing. Smart Struct Syst 2019; 24: 745–757.
displacement measurement of structures using a moving laser 315. Mei Q and Gül M. A crowdsourcing-based methodology
Doppler vibrometer. J Bridge Eng 2019; 24: 04019089. using smartphones for bridge health monitoring. Struct
300. Yang R, Singh SK, Tavakkoli M, et al. CNN-LSTM deep Health Monit 2019; 18: 1602–1619.
learning architecture for computer vision-based modal fre- 316. Mustapha S, Kassir A, Hassoun K, et al. Estimation of crowd
quency detection. Mech Syst Signal Process 2020; 144: 106885. flow and load on pedestrian bridges using machine learning
301. Hoskere V, Park J-W, Yoon H, et al. Vision-based modal with sensor fusion. Autom Constr 2020; 112: 103092.
survey of civil infrastructure using unmanned aerial vehicles. 317. Na Y, El-Tawil S, Ibrahim A, et al. Stick-slip classification
J Struct Eng 2019; 145: 04019062. based on machine learning techniques for building damage
302. Perry BJ and Guo Y. A portable three-component displace- assessment. J Earthq Eng 2021: 1–18.
ment measurement technique using an unmanned aerial ve- 318. Perez H and Tah JHM. Deep learning smartphone application
hicle (UAV) and computer vision: a proof of concept. for real-time detection of defects in buildings. Struct Control
Measurement 2021; 176: 109222. Health Monit. 28, e2751, 2021.Epub ahead of print 4 May
303. Hu F, Zhao J, Huang Y, et al. Structure-aware 3D recon- 2021
struction for cable-stayed bridges: a learning-based method. 319. Yu Y, Han R, Zhao X, et al. Initial validation of mobile-
Comput-Aided Civ Infrastruct Eng 2021; 36: 89–108. structural health monitoring method using smartphones. Int J
304. Perry BJ, Guo Y, Atadero R, et al. Streamlined bridge in- Distrib Sens Netw 2015; 11: 274391.
spection system utilizing unmanned aerial vehicles (UAVs) 320. Li J, Xie B and Zhao X. Measuring the interstory drift of
and machine learning. Measurement 2020; 164: 108048. buildings by a smartphone using a feature point matching
305. Lei B, Wang N, Xu P, et al. New crack detection method for algorithm. Struct Control Health Monit; 27. e2492, 2020. Epub
bridge inspection using UAV incorporating image processing. ahead of print April 2020
J Aerosp Eng 2018; 31: 04018058. 321. Alzughaibi AA, Ibrahim AM, Na Y, et al. Feasibility of
306. Cheng C, Shang Z and Shen Z. Automatic delamination utilizing smart-phone cameras for seismic structural damage
segmentation for bridge deck based on encoder-decoder deep detection, IEEE International Instrumentation and Measure-
learning through UAV-based thermography. NDT E Int 2020; ment Technology Conference (I2MTC). Dubrovnik, Croatia:
116: 102341. IEEE, 25–28 May 2020, pp. 1–5.
307. Kang D and Cha Y-J. Autonomous UAVs for structural health 322. Zhao X, Ri K and Wang N. Experimental verification for
monitoring using deep learning and an ultrasonic beacon cable force estimation using handheld shooting of smart-
system with geo-tagging: autonomous UAVs for SHM. phones. J Sens 2017; 2017: 1–13.
Comput-Aided Civ Infrastruct Eng 2018; 33: 885–902. 323. Wang N, Ri K, Liu H, et al. Structural displacement moni-
308. Ji X, Miao Z and Kromanis R. Vision-based measurements of toring using smartphone camera and digital image correlation.
deformations and cracks for RC structure tests. Eng Struct IEEE Sens J 2018; 18: 4664–4672.
2020; 212: 110508. 324. Wang N, Zhao X, Zhao P, et al. Automatic damage detection
309. Kromanis R, Xu Y, Lydon D, et al. Measuring structural of historic masonry buildings based on mobile deep learning.
deformations in the laboratory environment using smart- Autom Constr 2019; 103: 53–66.
phones. Front Built Environ; 5. 2019.Epub ahead of print 4 325. Shrestha A, Dang J, Wang X, et al. Smartphone-based bridge
April 2019 seismic monitoring system and long-term field application
310. Ozer E, Purasinghe R and Feng MQ. Multi-output modal tests. J Struct Eng 2020; 146: 04019208.
identification of landmark suspension bridges with distributed 326. Ozer E and Feng MQ. Biomechanically influenced mobile
smartphone data: golden gate bridge. Struct Control Health and participatory pedestrian data for bridge monitoring. Int J
Monit; 27. e2576, 2020.Epub ahead of print October 2020 Distrib Sens Netw; 13. 2017.Epub ahead of print April 2017
Malekloo et al. 49

327. Ozer E and Feng MQ. Direction-sensitive smart monitoring Stat, 2020, http://arxiv.org/abs/2003.04919 accessed 16 May
of structures using heterogeneous smartphone sensor data and 2021).
coordinate system transformation. Smart Mater Struct; 26. 342. Zhang R, Liu Y and Sun H. Physics-informed multi-LSTM
045026 2017.Epub ahead of print 1 April 2017 networks for metamodeling of nonlinear structures. Comput
328. Gulgec NS, Shahidi GS, Matarazzo TJ, et al. Current chal- Methods Appl Mech Eng 2020; 369: 113226.
lenges with bigdata analytics in structural health monitoring. 343. Ong SK and Nee AYC (eds). Virtual and augmented reality
In: Niezrecki C (ed). Structural health monitoring & damage applications in manufacturing. London: Springer London;
detection, Volume 7. Cham: Springer International Publish- 2004.Epub ahead of print
ing; 2017, pp. 79–84. 344. Mascareñas D, Harden T, Morales J, et al. Augmented reality
329. OBrien EJ and Keenahan J. Drive-by damage detection in for enabling smart nuclear infrastructure. Front Built Environ;
bridges using the apparent profile. Struct Control Health 5. 2019.Epub ahead of print 25 June 2019
Monit 2015; 22: 813–825. 345. Navab N, Blum T, Wang L, et al. First deployments of
330. Shirzad-Ghaleroudkhani N, Mei Q and Gül M. Frequency augmented reality in operating rooms. Computer 2012; 45:
identification of bridges using smartphones on vehicles with 48–55.
variable features. J Bridge Eng 2020; 25: 04020041. 346. Napolitano R, Anna B and Glisic B. Virtual environments for
331. Matarazzo TJ, Santi P, Pakzad SN, et al. Crowdsensing visualizing structural health monitoring sensor networks,
framework for monitoring bridge vibrations using moving data, and metadata. Sensors 2018; 18: 243.
smartphones. Proc IEEE 2018; 106: 577–593. 347. Büyüköztürk O and Yu T-Y. Structural health monitoring and
332. Ozer E and Feng MQ. Structural health monitoring, Start-up seismic impact assessment, Proceedings of the 5th National
creation. Elsevier, Amsterdam, 2020, pp. 345–367. Conference on Earthquake Engineering. Istanbul, Turkey,
333. Tao F, Qi Q, Wang L, et al. Digital twins and cyber-physical 26–30 May 2003.
systems toward smart manufacturing and industry 4.0: cor- 348. Bacco M, Barsocchi P, Cassara P, et al. Monitoring ancient
relation and comparison. Engineering 2019; 5: 653–661. buildings: real deployment of an IoT system enhanced by
334. Gabor T, Belzner L, Kiermeier M, et al. A simulation-based UAVs and virtual reality. IEEE Access 2020; 8: 50131–50148.
architecture for smart cyber-physical systems IEEE Inter- 349. Maharjan D, Agüero M, Mascarenas D, et al. Enabling human-
national Conference on Autonomic Computing (ICAC). infrastructure interfaces for inspection using augmented reality.
Wuerzburg, Germany: IEEE, 17–22 July 2016, pp. 374–379. Struct Health Monit, 20, 2020: 147592172097701.
335. Fuller A, Fan Z and Day C. Digital twin: enabling technology, 350. Yuan F-G, Zargar SA, Chen Q, et al. Machine learning for
challenges and open research, 2020, ArXiv191101276 Cs2019, structural health monitoring: challenges and opportunities. In:
http://arxiv.org/abs/1911.01276 accessed 1 January 2020) Zonta D, Sohn H and Huang H (eds). Sensors and Smart
336. Ye C, Butler L, Bartek C, et al. A digital twin of bridges for Structures Technologies for Civil, Mechanical, and Aero-
structural health monitoring, 12th International Workshop space Systems 2020. Bellingham, WA: SPIE, p. 2.
on Structural Health Monitoring. Stanford, CA, US: Stan- 351. Athanasiou A and Salamone S. Acquisition and management
ford University (CA), 10–12 September 2019, https:// of field inspection data using augmented reality. In: Dennison
web.stanford.edu/group/sacl/workshop/IWSHM2019/ MS (ed). Virtual, Augmented, and Mixed Reality (XR)
specialsession.html 1 January 2020). Technology for Multi-Domain Operations. 2020, Bellingham,
337. Shim C, Kang H and Dang N. Digital twin models for WA: SPIE, p. 7.
maintenance of cable-supported bridges, International Con- 352. Mascareñas DD, Ballor JP, McClain OL, et al. Augmented
ference on Smart Infrastructure and Construction 2019 (IC- reality for next generation infrastructure inspections. Struct
SIC). Cambridge, UK: 8–10 July 2019, ICE Publishing, Health Monit, 20, 2020: 147592172095384.
pp. 737–742. 353. Ozer E.. . Multisensory smartphone applications in vibration-
338. Shim C-S, Dang N-S, Lon S, et al. Development of a bridge based structural health monitoring. Columbia University;
maintenance system for prestressed concrete bridges using 3D New York, NY, 2016.PhD Thesis
digital twin model. Struct Infrastruct Eng 2019; 15: 1319–1332. 354. Zhang Y, Cai SC and Deng L. Piezoelectric-based energy
339. Zhang Z and Sun C. Structural damage identification via harvesting in bridge systems. J Intell Mater Syst Struct 2014;
physics-guided machine learning: a methodology integrating 25: 1414–1428.
pattern recognition with finite element model updating. Struct 355. Xie Y, Ebad Sichani M, Padgett JE, et al. The promise of
Health Monit, 20, 2020: 147592172092748. implementing machine learning in earthquake engineering:
340. Ritto TG and Rochinha FA. Digital twin, physics-based a state-of-the-art review. Earthq Spectra, 36, 2020:
model, and machine learning applied to damage detection 875529302091941.
in structures. Mech Syst Signal Process 2021; 155: 107614. 356. Keenahan J, Ren Y and OBrien EJ. Determination of road
341. Willard J, Jia X, Xu S, et al. Integrating physics-based mod- profile using multiple passing vehicle measurements. Struct
eling with machine learning: a survey. ArXiv200304919 Phys Infrastruct Eng 2020; 16: 1262–1275.
50 Structural Health Monitoring 0(0)

357. Champneys MD, Green A, Morales J, et al. On the vulnerability condition assessment of infrastructures. Comput-Aided Civ
of data-driven structural health monitoring models to adver- Infrastruct Eng 2019; 34: 774–789.
sarial attack. Struct Health Monit, 20, 2020: 147592172092023. 359. Jeong S, Hou R, Lynch JP, et al. A scalable cloud-based
358. Wu R, Singla A, Jahanshahi MR, et al. Pruning deep con- cyberinfrastructure platform for bridge monitoring. Struct
volutional neural networks for efficient edge computing in Infrastruct Eng 2019; 15: 82–102.