Original Article

Journal of Reinforced Plastics and

Composites
Sensory methods and machine learning 2022, Vol. 0(0) 1–28
© The Author(s) 2022
based damage identification of Article reuse guidelines:

fibre-reinforced composite structures: An

sagepub.com/journals-permissions
DOI: 10.1177/07316844221145972
journals.sagepub.com/home/jrp

introductory review

Senaka Bandara1 , Madhubhashitha Herath2,3  and Jayantha

Epaarachchi3,4

Abstract
Fibre-reinforced composite materials are extensively used for manufacturing critical engineering components in diverse
applications, which demands intelligent and reliable structural health monitoring (SHM) schemes to prevent catastrophic
failures associated with composite structures. Composite materials have complex failure mechanisms, and it is essential to
employ reliable SHM methods with high accuracy to detect damages at the incipient stage. Although there are several SHM
technologies available, no single strategy is impeccable for tackling all damage types due to the incredibly complex failure
mechanisms of the composite materials. Machine learning (ML) methods are frequently integrated to compensate for the
limitations of the traditional SHM methods. This paper presents the state-of-the-art sensory methods and deep learning
(DL) techniques while emphasizing the future directions for the engineering and scientific community interested in
developing novel SHM systems for fibre-reinforced polymer composite structures intended for civil, aerospace, auto-
motive, marine, oil and gas exploration industries.

Keywords
Structural health monitoring, composite structures, failure mechanisms, machine learning, damage identification

Introduction improved corrosion resistance etc., are some notable ad-

vantages offered by composites over conventional materi-
Structural health monitoring (SHM) is a potent diagnostic als.8 As a result, composites are preferred for safety-critical
tool used to detect damages and avoid catastrophic failures applications even in hostile environments. Composite
of structures. According to the SAE Standard ARP6461,
SHM is defined as ‘the process of acquiring and analysing
data from onboard sensors to evaluate the health of 1
Division of Mechanical Engineering Technology, Institute of Technology,
a structure’.1 SHM systems comprise a network of sensors University of Moratuwa, Homagama, Sri Lanka
2
attached to or embedded in the structure for continuous Department of Engineering Technology, Faculty of Technological Studies,
monitoring. Moreover, SHM systems consist of an onboard Uva Wellassa University, Badulla, Sri Lanka
3
Centre for Future Materials, Institute for Advanced Engineering and Space
network coordinator and algorithms to process large Sciences, University of Southern Queensland, Toowoomba, QLD,
amounts of data in real-time to diagnose damage after Australia
4
correcting for environmental and operational noise. SHM School of Engineering, Faculty of Health Engineering and Sciences,
consists of five levels that include damage detection, lo- University of Southern Queensland, Toowoomba, QLD, Australia
calization, classification, quantification and prognosis.
Corresponding authors:
Figure 1 encapsulates the aforementioned five levels of Madhubhashitha Herath, Department of Engineering Technology,
SHM and the recent advancements (smart sensing tech- Faculty of Technological Studies, Uva Wellassa University, Passara Road,
nologies,2 cloud-based computing and rapid growth of Badulla 90000, Sri Lanka.
machine learning techniques3–5) that have revolutionized Email: madhubhashitha@uwu.ac.lk
intelligent SHM systems.
Jayantha Epaarachchi, School of Engineering, Faculty of Health Engineering
Composite materials have become increasingly popular and Sciences, University of Southern Queensland, West Street,
across numerous industries due to their superior material Toowoomba, QLD 4350, Australia.
properties.6,7 Lightweight, high strength and stiffness, Email: jayantha.epaarachchi@usq.edu.au
2 Journal of Reinforced Plastics and Composites 0(0)

material is a synthesis of a base material known as the range.30–33 If developed, multi-physics-based fibre optics
matrix which serves as the bonding medium and filler SHM methods will be able to detect damages with high
material which reinforces the composite by providing levels of sensitivity and accuracy. Piezoelectric wafer active
strength and stiffness via fragments, particles, fibers or sensing (PWAS)34 technologies have been extensively
whiskers of a different material.9,10 Fibre-reinforced poly- applied in SHM both in active and passive modes. Guided
mer composite materials are susceptible to low-velocity wave-based35 and electromechanical impedance (EMI)36,37
impact damages (LVID)11,12 during their service life based SHM techniques are regarded as active methods as
which can engender barely visible impact damages the structure under investigation is continuously excited for
(BVID).13 Matrix cracking, debonding, delamination and active interrogation. Sensor placement is critical in guided
fibre breakage are key failure mechanisms (Figure 1) of wave-based techniques to identify structural damages ac-
fibre-reinforced composites due to excessive loading and curately. Guided wave-based techniques are not effective
impact events.14 The anisotropic nature of composite ma- for the damage detection of complex composite structures
terials makes it challenging to analytically predict damage having complex geometries and structural elements like
propagation and durability of composite structures. Sudden stiffeners and stringers due to boundary reflections. In
catastrophic failures can be avoided by detecting incipient contrast, acoustic emission (AE)38–40 is a passive PWAS
damage via employing reliable SHM techniques. technique used to identify growing cracks and impacts on
Vibration-based SHM methods have the ability to locate the composite structure. While the accuracy of the acoustic
both global and local damages in structures. Traditional emission method is affected by environmental conditions
modal analysis techniques which include modal displace- and noise, sensitivity is dependent on the number of sensors
ments, modal curvatures and modal strain energy methods in the network and their placement.
can easily monitor the entire structure.15 Signal processing Automated SHM techniques are categorized as physics-
techniques such as the fast Fourier transform (FFT), en- based methods that require expertise in system physics and
velope, spectrum and wavelet analysis can be used to an- data-driven4 approaches ideal for complex structures.
alyze the vibration response. This processed data is being Machine learning (ML) techniques 3 have been widely used
used as features for classifiers like artificial neural networks in structural damage detection during the last decade. ML
(ANNs), support vector machines (SVMs), principal models map patterns from the input sensor data to output
component analysis (PCA) and k-means algorithms prior to targets when identifying damage scenarios. Conventional
the introduction of deep learning.16 A major concern with ML techniques require handcrafted damage sensitive fea-
most of the vibration-based SHM techniques is that the tures and cannot process large amounts of raw data.
accuracy of damage detection gets compromised by envi- However, deep learning (DL)41–44 based SHM approaches
ronmental noise and varying environmental conditions. are extensively adopted due to their ability to process large
Correct sensor placement is crucial for recording meas- amounts of raw data while learning complex features au-
urements, and this may not be possible for large complex tomatically. Stacked autoencoder,43 convolutional neural
structures having inaccessible areas. Strain-based SHM networks (CNNs)45–47 and recurrent neural networks
techniques are popular for damage detection in civil and (RNNs) are a few DL algorithms popular among researchers
aerospace composite structures. Strain gauges used for in composite SHM. Environmental noise is a major concern
strain-based monitoring are being gradually supplanted by that needs to be addressed for these methods as well. Most
fibre optic sensors (FOS) due to several practical limi- of these research works have been carried out in laboratory
tations, including difficulties with deployment across large settings, and it must be noted that these studies must be
structures, cumbersome wiring, operational time issues and extended to complex real structures to ensure their feasi-
inability to withstand harsh environments.17 Multiplexing bility in real-time damage detection.
capability, higher sensitivity, electro-magneto immunity Over the years, catastrophic accidents have been reported
and durability have made fibre Bragg grating (FBG) from various fields due to composite structural failures. In
sensors18–22 an attractive candidate for the real-time 2001, American Airlines Flight 587 crashed as the vertical
monitoring of aircraft structures.23–25 Distributed fibre tail (Figure 2(a)) separated from the fuselage due to ex-
optic sensing (DFOS)26 technology capable of obtaining cessive rudder pedal inputs.48 Space Shuttle Columbia (OV-
strain measurements along the entire length of the fibre has 102) was destroyed upon re-entry into Earth’s atmo-
been developed significantly over the last decade and is sphere.49 This was due to impact damage taken place at the
being used in large structures like bridges, high-rise leading edge of the left-wing of the orbiter during the launch
buildings and tunnels.1 Strain, temperature and vibration (Figure 2(b)). Crack propagation on the reinforced carbon–
are the physical parameters monitored for damage detection carbon (RCC) panels had been accelerated by the extreme
in fibre optic sensing techniques. Developments in the stress and temperature exerted during the re-entry. Thermal
optical time-domain reflectometer (OTDR)27,28 and optical protection systems (TPSs) are at risk of high velocity impact
frequency-domain reflectometer (OFDR)29 techniques have damages by foreign objects. The health condition moni-
enhanced the measurement resolution and the sensing toring of TPSs became more vital following the Columbia
Bandara et al. 3

Figure 1. Five levels of SHM, advancements of intelligent SHM and key failure mechanisms of fibre-reinforced composites.

shuttle incident. The UK registered sailing yacht, Cheeki project (Figure 3(c)) was developed to reduce risks related
Rafiki, had reportedly capsized in the north Atlantic killing to wind turbine blade failure during operation.56 A fibre
all four crew members, following a detachment of its keel optic sensing based SHM system having multi-sensing
(Figure 2(c)).50 Investigations revealed that the keel and the capabilities including strains, vibrations and acoustic
glass-reinforced plastic (GRP) matrix had endured damage emission was used for this project. These are only a few
during previous groundings. In addition to these fatal ac- recent composite SHM applications utilized across different
cidents, wind turbines are prone to composite structural industries. These monitoring systems will gain more pop-
failure. Since 2000, more than 10% of wind turbine ularity with the adoption and growth of concepts such as IoT
breakdowns had occurred due to blade failure.51 Further, and smart cities.
lightning strikes, bird attacks and acid rains can deteriorate This paper presents an overview of the recent ad-
the structural integrity of the composite wind turbine blades vancements in fibre-reinforced composite SHM techniques.
which leads to failure (Figure 2(d)). It is evident from the The main goal of this review is to present the recent research
above incidents that these types of catastrophic failures can findings in this field with an emphasis on the last 5 years.
be mitigated by integrating a reliable SHM system in the The section health monitoring of composite structures re-
composite structure. views recent state-of-the-art developments in vibration-
It is worth mentioning a few recent composite SHM based, strain-based and piezoelectric wafer active sensor
projects. CFRP is extensively used in the structure of the based SHM techniques. The section Machine learning for
Airbus A350 XWB (Figure 3(a)).54 High priority was given composite SHM systems encapsulates the impact of ML and
to develop an SHM system to quickly detect and estimate deep-learning techniques towards enhancing the perfor-
random ground handling impact damage to CFRP in the mance and reliability of intelligent damage detection of
vicinity of doors as the door surrounding areas show a high complex composite structures. In addition, possible future
concentration of impact scenarios during ground handling. directions in this area are also presented. A summary and
Guided wave-based piezoelectric active sensing method conclusions are finally presented in the final section.
was proposed due to its ability to detect delamination and
debonding. Semmering Base Tunnel in Austria (Figure
Health monitoring of composite structures
3(b)) is one of the largest tunnel construction projects in
Europe.55 A distributed fibre-optic sensing system is used in Different SHM techniques are used for monitoring com-
this project to monitor tunnels, shafts, reinforced earth posite structures depending on the nature of the application.
structures and pipelines. European union funded BladeSave These techniques have their advantages and shortcomings,
4 Journal of Reinforced Plastics and Composites 0(0)

Figure 2. Disasters of composite structures: (a) vertical stabilizer as recovered from Flight 587;48 (b) Columbia shuttle impact event;52
(c) keel detached area of Cheeki Rafiki;50 (d) wind turbine blade damage.53

and the composite SHM techniques have evolved re- Vibration-based SHM techniques consist of physically
markably during the past decade. In this review, priority will based and non-physically based methods57 as shown in
be given to the most notable advancements made during the Figure 4. Modal-based method and local diagnostic method
last 5 years.1 Recent developments of vibration-based, are physically based damage detection techniques. Fre-
strain-based and piezoelectric active wafer sensor-based quency response method,58 mode shape analysis59,60 and
composite SHM techniques are briefly discussed in the strain energy method61 are investigated under the modal-
following sections. based method. Although these modal-based techniques
successfully detected the damage, determining the number
of sensors, uniformly placing them on the structure58 and
Vibration-based methods the effect of environmental noise on the outcome are
Most of the SHM techniques used for composite structures identified as significant drawbacks. Damage detection based
are local detection methods in nature. The damaged zone on the frequency response function (FRF) method is based
should be approximately known beforehand, and the in- on the change in frequency of the structure. For this method
vestigated area should be easily accessible for these local to give accurate results, environmental parameters must be
detection techniques to be applied. The underlying principle constant and environmental noise should be less than 10%.
of vibration-based damage detection,15 which is a global These environmental interferences can be mitigated by
detection technique, is that the damage will modify the incorporating a fuzzy clustering method, which uses nor-
stiffness, mass or energy dissipation characteristics. These malized data. Sensor placement is important in this method
modifications correspond to a change in the dynamic re- for measuring data, and this method is not suited for large
sponse of the system, and the damage can be assessed based complex structures where certain areas are inaccessible. In
on the changes in these dynamic characteristics. Although the mode shape analysis method, higher order derivatives
the concept of vibration-based SHM seems relatively are preferred as they are more sensitive to structural
simple, the application can be tricky due to technical anomalies. The strain energy method uses stiffness and
challenges. mode shape matrix for damage detection, and
Bandara et al. 5

Figure 3. Recent composite SHM projects: (a) CFRP impact damage detection system employed in Airbus A350 XWB;54 (b)DFOS
system in the Semmering Base Tunnel;55 (c) BladeSave project.56

environmental conditions do not affect the determination of vibration-based health monitoring of large structures.66 An
the damage location. impulse hammer was used to check the dynamic sensing
The local diagnostic method utilizing piezoelectric capabilities, and it was found that the sensor can capture
sensors62 was used to counter the noise interference natural frequencies up to the third mode. The results of the
problem. The notable disadvantage of this method is careful developed sensor were compliant with that of ready-made
placement of the sensors is required to detect damages accelerometers. Cement-based nanocomposite (CNC)
accurately and failure to do so would not yield the intended sensor is an effective sensor for vibration-based health
results. The non-parametric method and time series method monitoring and condition assessment of structures as it only
belong to the non-physically based category. Under the non- feeds off ambient vibration information.
parametric method, probabilistic approaches such as Sepehry et al.67 studied the application of the scaled
Bayesian probabilistic method, autoregressive moving av- boundary finite element method (SBFEM) for vibration-
erage (ARMA)63 model and vector autoregression moving based SHM of breathing cracks. High-frequency actuation
average (VARMA)64 method were studied. These non- is required to detect a minute breathing crack, and simu-
parametric methods were unable to detect non-linear lation is also vital to understand the effect of microcracks on
damage in the structure. Accurate damage detection is the dynamic response. SBFEM was applied to reduce the
achieved by combining the time series model65 with non- resulting large mass and stiffness matrices from the tradi-
parametric-based methods such as the ARMA. A com- tional finite element method (FEM), and the node-to-node
parative analysis of these methods can be found in a detailed contact strategy was used to capture the contact of the
review of vibration-based damage detection techniques breathing crack during vibration. The healthy and non-
published by Das et al.57 contact model only includes the first harmonic, whereas
Rao and Sasmal investigated the applicability of smart in the contact model, a higher harmonic is also observed
nano-engineered cementitious composite sensors for (Figure 5(a)). According to the frequency response of the
6 Journal of Reinforced Plastics and Composites 0(0)

cracked structure, the frequency response of the displace- normalize the result and quantify the damage. A steel beam
ment is symmetric around the natural frequency under low was used in the experiment along with three accelerometers
forces (Figure 5(b)). Nevertheless, this symmetricity has placed at node 2 in the first mode and node 1 and node 3 of
disappeared with increasing force due to the nonlinearity of the second mode (Figure 6(a)). The amplitude spectrum of
the contact phenomenon. these nodes was investigated in their respective mode
The results suggest that the scaled boundary method is frequencies under three different temperatures (0°C, 30°C
more efficient than the conventional FEM and can capture and 60°C) for the healthy and damaged structure where the
the higher harmonic generation, frequency shift and vi- saw cut depth was increased from 2 mm to 16 mm.
broacoustic modulation resulting from the non-linear effects As shown in Figure 6(b), the influence of both tem-
of breathing crack. The suggested scaled boundary finite perature and actual structural damage on the natural fre-
element technique has the potential to improve early quency makes it challenging to discern the two using
damage detection. frequency-based methods. Figure 6(c) shows that DI in-
Different SHM techniques are widely used for the creases with damage size and the value at node 3 is higher
condition assessment and damage detection of wind tur- than that of node 1 as the damage is located closer to node 3.
bines. Ou et al.68 investigated a vibration-based damage In addition, it can be seen that node 1 which is extracted
detection method for a small-scale wind turbine blade. The from a higher frequency mode is more sensitive to damage
blade was tested under artificially induced damages, and the than node 2 which confirms that higher frequency nodes are
response was monitored using accelerometers. The captured more sensitive to small damage. This method can be applied
signals were used in two damage detection frameworks, to more complex structures using a laser vibrometer to
namely, statistical and modal damage detection methods. locate the nodes.
The statistical method corresponds to the quantitative es-
timation of relevant statistical performance metrics ex-
tracted from dynamic response data, and the modal
Strain-based methods
detection technique tracks the variations of modal param- Early-stage structural damage can usually be identified by
eters such as natural frequencies, damping ratios, mode local strain-field anomalies that are being monitored to
shapes and mode shape curvatures. The tracked quantities achieve reliable and accurate damage detection, locali-
were assessed using a baseline-dependent method. These zation and evaluation of damage extent.71 However, it is
quantities were determined using non-parametric and worth noting that the global strain field changes only
parametric identification techniques. The power spectral slightly in an event of a local crack and is only pronounced
density (PSD) and the FRF were used in the parametric at the crack tip.72 Traditionally, electrical strain gauges
approach, and vector autoregressive (VAR) and vector were used for strain monitoring of structures. Due to
autoregressive with exogenous inputs (VARX) were utilized various limitations of strain gauges, including the complex
in the non-parametric approach. Statistical damage de- wiring involved, researchers shifted their interest towards
tection methods are found to be more suited for the fibre optical sensors (FOS). FOS-based SHM techniques
maintenance processes of structures. Dynamic loading, became increasingly popular due to their notable advan-
complex geometry, material characteristics of blades and the tages including higher sensitivity, immunity to electro-
effect of varying environmental conditions are challenges to magnetic interruption, durability and multiplexing
the application of SHM to wind turbine blades. capability.73 Fibre sensors can be integrated into com-
Structural health monitoring techniques that rely on posite materials during the layup process. This makes it
baseline data require a large data storage, and the fact that possible to monitor composite structures continuously
current data needs to be compared with the stored reference during their life cycle, enhancing reliability and safety
data makes it harder to be applied for online real-time health while extending the operational life. Interested readers are
monitoring. To overcome these problems, researchers are referred to a comprehensive review on FOS technologies
working on baseline-independent SHM techniques which which was recently published by Di Sante.17
do not require past measured data from the intact structure. Fibre optic sensors can be categorized into three to-
Some researchers noticed that structural damage could pologies based on their working principle. Point sensors can
displace nodes in a mode shape at its natural frequency. An only detect variations of the measured quantity in the
effective structural damage indicator (SDI) can be derived proximity of a single sensor. In multipoint or quasi-
from the vibration amplitude at these nodes where the initial distributed sensors, multiple localized sensors are placed
measurement reveals information about the health of the at regular intervals along the fiber length. The optical fibre
structure. A baseline-free damage detection approach based works simultaneously as a transmitter and a sensor in
on the nodal displacement in mode shape is proposed by distributed sensing and can capture variations in local ex-
Huang et al.69,70 This proposed new approach was named as ternal variables like strain and temperature. Single-point
node displacement (NODIS) method, and a transfer FOS are also known as interferometric sensors, and quasi-
function-based damage indicator (TFDI) was utilized to distributed sensors are grating-based.
Bandara et al. 7

Figure 4. Vibration-based SHM techniques.

Interferometric sensors. An interferometric sensor is pro- the results revealed that it is more sensitive to applied strain
duced by an intrinsic or extrinsic interferometric cavity than temperature variations (Figure 7(a) and (b)). A wave-
along the optical path. Structural anomalies are identified by length shift of 0.071 nm was observed over a temperature
changes in the optical phase difference between two light range of 100°C, while an applied strain of 1000 µε resulted in
waves having the same frequency. A notable advantage of a wavelength shift of 0.58 nm.
interferometric sensors is their high resolution in strain Xu and Khodaei75 developed a novel FBG- based FP
measurement (με). Fabry–Perot (FP) and low coherent sensor to monitor composite structures to detect damage
interferometric sensors (SOFO) are the most popular in- with high reliability under varying operational conditions,
terferometric sensor types used in strain-related monitoring such as fluctuating noise levels and loading. First, the
applications. spectrum was simulated and subsequently experimentally
Fabry–Perot sensors. Kaur et al.74 developed an extrinsic validated before being tested on a composite panel. The
FP interferometric sensor to measure strain at high ambient newly developed sensor was able to record signals of the
temperatures. The sensor was fabricated using a femtosecond pristine structure under varying environmental conditions
(fs) laser and found to be thermally stable at temperatures as with high repeatability. When artificial damage was induced
high as 800°C. This sensor was tested by embedding in in the composite structure, the FBG-based FP sensor
carbon fibre/bismaleimide (BMI) composite laminates, and showed high amplitudes compared to conventional FBG
8 Journal of Reinforced Plastics and Composites 0(0)

Figure 5. SHM of breathing cracks using the contact method: (a) displacement versus harmonic number; (b) frequency response for
range 13.9–14.6 kHz for crack without contact and crack in contact with different forces.67

sensors. The effect of damage was more significant in the wavelength (λB), and permits the rest of the incoming light
FBG-based FP sensor, largely due to higher SNR and to pass through without modifying its property. The Bragg
amplitude (Figure 7(c) and (d)). The authors expect to test wavelength of the reflected component is a function of
the newly developed sensor for BVID on composite temperature and/or strain.
structures and to test its multiplexing capability. Fibre Briggs grating sensors are being used in large-scale
SOFO sensors. Blazejewski et al.76 investigated the ap- applications across a wide range of industries such as
plicability of FOS for strain monitoring of composite high- aerospace, civil and oil and gas. Variations of load on an
pressure vessels. Both FBG sensors and SOFO sensors were aircraft wing cause it to deform, which induce deformations
used for this study, and the sensors were installed between in the aircraft baseline. Therefore, it is crucial to determine
GFRC layers during the braiding process of the composite the wing shape deformations accurately for the safety of the
pressure vessels. Although the main objective of this study aircraft.24 A study was conducted by Nicolas et al. to de-
was focused on monitoring the health of the composite termine the deflected wing shape and the out-of-plane loads
structure, these sensors could also be used to monitor strain of a 5.5 m full-scale composite wing of an ultralight aircraft
during the vessel manufacturing process. subjected to concentrated and distributed loads.79 Strain
SOFO interferometric sensors are long-gauge sensors data from 780 FBG sensors were used in this experiment,
used in various structures like bridges, buildings, oil pipes and the calculated results were within 4.2% of the recorded
and tunnels. These sensors have a wide measurement range data. Both deflection and load algorithms used in this in-
extending up to 100 m with a resolution in the micrometre vestigation were based on the classical beam theory and
level. SOFO sensors are insensitive to temperature and have only the case of pure bending was considered. The mea-
high precision and stability. Since the dynamic measure- sured strains (Figure 8(a)) from the FBG sensors at the
ment range is considerably low (0.1 Hz–1 Hz), these sensors maximum resultant load of 2269 N were used to determine
are unsuitable for monitoring impact damages in aircraft the elastic deflection curve of the wing (Figure 8(b)). This
structures.77 computed deflection curve is superimposed on the actual
deflected wing to show high accuracy as shown in Figure
Grating-based sensors. Fibre Briggs grating sensors are the 8(c). In addition, the in-plain strain data acquired from FBG
most popular grating-based fibre optics SHM technique sensors were used to calculate the out-of-plane loads (Figure
used in composite materials.78 FBGs are made by exposing 8(d)). The computation of both deflected wing shape and
the core of single-mode fibre to a periodic pattern of intense out-of-plane loads is possible due to the high spatial res-
laser or UV light laterally. The refractive index of the fibre’s olution of the FBG strain measurements.
core is permanently changed, depending on the pattern of Delaminations of composite structures used in the
exposure, and this fixed index modulation is called a grat- aeronautical and wind energy sector are mostly caused by
ing. An FBG sensor reflects a portion of the incoming light impact damages. Lamberti et al.25 researched to dem-
of a particular wavelength, which is called Bragg onstrate the feasibility of using FBG sensors embedded in
Bandara et al. 9

Figure 6. NODIS method. (a) FRF and mode shape; (b) amplitude spectrum under different temperatures and saw cut depths; (c) DI
under different temperatures and saw cut depths.70

composite structures to detect impacts. A modal hammer Zeng et al.81 investigated the health monitoring of
was used to make 164 impacts on 41 locations of a carbon sandwich L-joints using FBG sensors. The safety of the
fibre plate embedded with FBG sensors. Captured data connections is crucial during the design stage as well as
was then processed using an improved fast phase cor- throughout the service life. In this study, FBG sensors were
relation (FPC) algorithm in tandem with a variable se- used to monitor the skin–core interfacial damage, and local
lective least square (VS-LS) inverse solver method. The strains were measured using resistance strain gauges.
proposed technique performed better in 70.7% of the Strain–displacement curves obtained in the experiment
cases than the P-Inv-based technique. Zhu et al.80 in- suggest that FBG sensors are more sensitive to internal
vestigated the damage detection capability of FBG sen- damage. Although FBG sensors are quick to reveal in-
sors in composite materials under low-frequency cycling. formation about the inception and growth of the damage, the
FBG-embedded composite cantilever beam was used in accuracy and reliability can still be improved. Although
this study for both undamaged and damaged scenarios. FBG sensors may slightly affect the mechanical perfor-
By analyzing the experimental results with the simulation mance of sandwich structures, more research needs to be
results obtained from ANSYS simulation, it was con- conducted on the application of FBG sensors for curved
cluded that FBG sensors are suitable for damage de- interfaces. The centre wavelength is used as the main in-
tection in composite structures under low-frequency dicator, and when the complexity of the reflected spectrum
cycling. increases, the centre wavelength is undetectable. For such
10 Journal of Reinforced Plastics and Composites 0(0)

Figure 7. Application of Fabry–Perot sensors in composite SHM: (a) response of the embedded sensor to ambient temperature
change; (b) response of the embedded sensor to applied strain;74 (c) damage signals recorded by FBG-based FP sensors; (d) damage signals
recorded by FBG sensors.75

cases, strain-measuring methods based on full spectral the interrogator unit. This makes it more cost-effective and
signals are more suited. desirable for a wide range of applications that require
Oil and gas pipeline structures in offshore applications continuous monitoring. Information about the fibre prop-
succumb to various failures due to harsh operating con- erties and physical parameters such as strain and temper-
ditions. Early detection of leakage is crucial for preventing ature can be estimated by analyzing the back scattered light.
uncalled disasters. Although different monitoring techni- The optical fibre is embedded in the structure being mon-
ques have been used, the high sensitivity and accuracy make itored, and one end of the fibre is connected to the in-
FBG sensors more reliable for marine applications. Bedi terrogator, which is used to generate and process optical
et al. proposed a new method to detect damages in pipeline signals.
structures using FBG sensors.82 When a crack is initiated Most of the previously discussed strain-based composite
and propagates through the structure, ultrasonic waves are SHM techniques have been tested on simple composite
generated. Mechanical strains are induced on the optical plates under laboratory conditions. However, composite
fibre upon contact of these waves with the FBG sensors structures typically employed in external structures of an
which causes a shift in the wavelength. An optical spectrum aircraft are curved and complex. Complex manufacturing
analyzer (OSA) is used in the experimental setup to detect processes and rigorous in-service mechanisms may cause
anomalies in the signal. The wavelength shift varies linearly structural failure. Shan et al.84 studied the application of
with the applied stress, and this linear dependency aids to DFOS for damage diagnosis of complex curved composite
locate damages on pipeline structures. structures. A DFOS network was applied to a composite
winglet of a real aircraft and observed that it is capable of
Distributed fibre optic sensors. Distributed fibre optic sensing identifying and localizing the damage precisely in the vi-
(DFOS) enables continuous, real-time measurements along cinity of the fibre installation path (Figure 9).
the entire length of a fibre optic cable. Precise quantification Both DFOS and strain gauges were used in this study.
of the structure’s health is attained using a dense layout of Graphs of strain measurements obtained from both classic
these sensors.83 A key advantage of DFOS is that it only strain gauges and distributed FOS complement well with
requires a single cable to transfer the acquired data back to each other. Strain distributions recorded by DFOS under
Bandara et al. 11

Figure 8. Determination of in-flight loads and shape of composite wing. (a) Strain distribution under 2269 N; (b) elastic deflection of
the wing; (c) displacement results superimposed on the deflected wing; (d) out-of-plane load per FBG sensor for load of 2269 N.79

varying loads show the linear variation of strain with in- a large effective mode area and low intermodal dispersion
creasing loads. DI calculated using strain data 40% of the for Raman-distributed temperature sensors (RDTS) to gain
ultimate load exhibited the existence of damage on the high spatial and temperature resolution over a long range.30
aircraft wing. A temperature resolution of 1°C along with a spatial res-
In DFOS, scattering is the interaction between light and olution of 1.13 m at a range of 25 km was achieved.
an optical medium. Raman,27,30 Brillouin17,26,33,85 and However, the temperature resolution increased to 0.7°C
Rayleigh17,29 scattering are the three modes of scattering while the spatial resolution degraded to 2.58 m when
that occur in DFOS. Raman scattering is influenced by the standard multimode fibers (MMFs) were used over the same
temperature of the fiber, and Brillouin scattering depends on length.
the fiber density, which is directly related to the temperature Brillouin optical time-domain reflectometers (BOTDRs)
and strain change of the optical fibre. Rayleigh scattering is determine the strain or temperature based on Brillouin
independent of external physical parameters and is used to scattering of a single pulse. BOTDR is suited for distributed
measure propagation effects. Techniques used when ap- sensing of large structural applications due to its high
plying DFOS to different SHM applications are discussed in sensitivity. Sheng et al.31 proposed a novel technique based
the following sections. on BOTDR to determine the distributed measurement of
temperature and strain simultaneously. The strain was kept
Optical time-domain reflectometry. In the optical time- unchanged when a segment of the optical fibre was heated to
domain reflectometer technique, a short optical pulse is 40°C in determining the temperature sensitivities. Then the
sent through the fiber and a photodetector processes the strain was gradually increased in a predetermined fibre
backscattered light. In this technique, the loss occurs due to section at room temperature to find the strain sensitivities.
the Rayleigh scattering, which is resulted from variations in Brillouin optical time-domain analysis (BOTDA) employs
the refractive index of the fibre core. The spatial resolution a more complex method known as stimulated Brillouin
of an OTDR instrument is the smallest distance between two scattering (SBS). Esmailzadeh Noghani et al.26 proposed
scatters that can be determined. Spatial resolution obtained a novel method called the multiple short pulse BOTDA
from commercially available OTDR instruments is rela- (MSP-BOTDA) which yielded sub-centimetre spatial res-
tively low. Reducing the pulse width can increase the spatial olution and kilometre dynamic range simultaneously. This
resolution, which in turn decreases the pulse energy. The technique can detect a wide range of temperature/strain
signal-to-noise ratio is weakened, as the pulse width is changes as the effective Brillouin gain bandwidth of the
reduced, which is a problem for long-range sensing ap- pump pulse train is greater. Different methods including the
plications. Therefore, the pulse width and pulse energy Brillouin optical correlation domain analysis (BOCDA)32
should be optimized based on the application.86 Liu et al. and optimized algorithms39,40 have been studied to enhance
used a graded-index few-mode fiber (GI-FMF) having the spatial resolution.
12 Journal of Reinforced Plastics and Composites 0(0)

Figure 9. Experimental setup including (a) the sensor network, (b) the measurement path arrangement and (c) the hydraulic
system.84

Optical frequency-domain reflectometry. The limitations of form of an elastic wave and is received by the other PZTs.
OTDR-based sensors paved the way for the development of The signals received by the PZTs for the pristine structure
Rayleigh scattering based OFDR, also called optical are stored and compared to the signals received throughout
backscattered reflectometer (OBR). The OBR-based dis- the service life of the structure. Any structural damage in the
tributed sensors are ideal for short sensing lengths, pref- path between the actuator and receiver PZTs will cause
erably less than 100 m. Nevertheless, long sensing lengths distortions of the new received signal. Most of the recent
can be achieved, at the expense of spatial resolution and research in this area has been focused on developing new
temperature/strain resolution. A comparison of different damage imaging algorithms for damage localization and
distributed FOS is presented in Table 1. characterization.34
Although guided wave-based SHM techniques are easily
applied to flat plates and pipes,88 it becomes challenging
Piezoelectric wafer active sensors
when applied to structures having boundaries, stiffeners and
PWAS-based SHM techniques are broadly classified as thickness variations. False-positive signals are generated
active and passive sensing. PWAS active sensing involves due to multiple reflections and scattering from the afore-
continuous active interrogation with the structures as re- mentioned interfaces.89 Spada et al.90 proposed a combined
sponses are recorded upon specified excitations. In contrast, global-local (GL) method that is effective for composite
PWAS passive sensing techniques only receive signals upon structures having geometrical and/or material dis-
the occurrence of damage or the incident.34 PWAS are used continuities. This new method was initially tested for
for (i) active sensing of mid-range damage detection using a composite skin-to-stringer assembly employed in the
guided waves, (ii) active sensing of low-range damage construction of modern aircraft. A skin-to-stringer assembly
detection using EMI technique and (iii) passive sensing of was studied for four different defects: delamination between
acoustic emissions of growing cracks and low-velocity skin and stringer (defect 1), horizontal fracture in the
impacts.1 stringer (defect 2), vertical fracture in the stringer (defect 3)
and delamination between second and third laminae in the
Active sensing using guided waves. Guided wave-based SHM skin (defect 4). Scattered spectra were acquired for both
techniques87 are popular in the aerospace industry due to flexural and axial wave modes. The recorded scattered
their ability to detect very small damages.1 Multiple PZTs spectra are useful for determining suitable wave mode-
that are embedded or bonded to the structure are used in the frequency combinations and identifying specific defects
guided wave method. A PZT actuator generates a short when guided wave-based SHM is applied to similar
ultrasonic pulse that propagates through the structure in the components.
Bandara et al. 13

In guided wave-based damage detection techniques, approach for sensors as this method requires a large
wave propagation is affected by environmental temperature number of sensors.
variations.91 Gao et al.92 investigated the effect of tem-
perature on PZT-based guided wave propagation and dis- Active sensing using electromechanical impedance. In recent
covered that the wave speed of the guided wave signal is times, the application of EMI-based SHM techniques to
more sensitive to temperature than signal fluctuation composite structures has gained significant attention from
characteristics. A damage index based on Spearman rank the scientific community. Research has been conducted
correlation coefficient that is suitable for damage detection using data-based and model-based methods. Although the
and localization under varying temperatures was proposed. use of ML and artificial neural networks (ANNs)97 has
The covariance DI is more sensitive to temperature varia- enhanced the data-based techniques, the requirement of
tions than the Spearman DI as shown in Figure 10(a). a large amount of training data was seen as a challenge.
Further, the results obtained show that the covariance DI The most popular damage indexes used in EMI-based
based method has a larger location error than the Spearman SHM are root mean square deviation (RMSD), mean ab-
DI based method where the baseline signal at 40°C covers solute percentage deviation (MAPD), covariance (Cov) and
a wider temperature range (Figure 10(b)). correlation coefficient deviation (CCD).37 The RMSD and
The applicability of the guided wave-based SHM MAPD metrics are more sensitive to magnitude fluctuations
techniques to detect debonding in CFRP-reinforced steel of the spectra, and the Cov and CCD metrics are more
structures was studied using CFRP-reinforced steel plates sensitive to changes affecting the signature shape, like
having varying debonding sizes.93 Anti-symmetrical Lamb frequency shifts.98 Therefore, RMSD and MAPD are
wave mode was used in this study due to its higher sen- suitable for the localization and characterization of incipient
sitivity to interfacial damage. A theoretical relationship damage, whereas Cov and CCD are preferred for de-
between the time of flight (ToF) of the first wave packet of termining the extent of damage at a fixed location.37
the received signal and the size of the debonding was de- Furthermore, when applying the model-based EMI
rived. According to the theoretical model, the ToF decreases technique to complex structures, it is difficult to establish
linearly with the extent of the debonding which was later constitutive correlations on which the analytical models
verified by conducting numerical and experimental studies. are based. As a solution, Zhu et al.99 developed a com-
De Luca et al.94 developed two numerical models based prehensive active monitoring scheme (CAMS), com-
on FEM for the simulation of Lamb wave propagation in bining DCMI (direct coupling mechanical impedance)
a damaged winglet (Figure 11(a)–(d)) made of glass fibre signatures and RMSD (root mean square deviation)
reinforced polymer composite. Damages were numerically damage index to locate and quantify interfacial dis-
modelled inside the laminate by reducing the elastic bonding of honeycomb sandwich composites. Although
properties of the material by 70% in one model and by the EMI method is an effective composite SHM tech-
adding a notch to the other. Pitch-catch approach was used nique, the signal gets corrupted due to environmental
for detecting the damage as it allowed to compare the Lamb noise and vibrations. Castro et al.36 carried out an ex-
wave propagation in both undamaged and damaged winglet perimental study on the effect of white noise on the
configurations.95 A total of 12 sensors were used in these impedance monitoring of a CFRP composite plate. For
experiments, and low-velocity impact damages were in- this study, correlation coefficient deviation metric
duced in the area encompassed by sensors 4, 5, 7 and 8 by (CCDM) and cross-correlation square deviation (CCSD)
dropping a mass of 0.8 kg. Two central frequencies damage indexes were applied to the raw response signal
(100 kHz and 150 kHz) were used in this study, and the root using the pitch-catch method. According to the results,
mean square deviation (RMSD) damage index was used to the CCSD index proved to be more effective than the
quantify the damage. Results obtained from the numerical CCDM index in a noisy environment. Two damage sizes
modelling showed a good correlation with the experimental were inserted into the composite plate, and the sensitivity
results. of the two indices was tested for healthy and damaged
Memmolo et al.96 investigated a multi-path and multi- (damage A: 1 mm and damage B: 10 mm) structure under
parameter damage detection approach competent in ob- noisy and no-noise conditions. Under the no-noise
serving multiple wave characteristics. This proposed condition, both indices produced the expected results
technique includes a multi-path and multi-parameter ap- by showing higher index values for larger damage sizes.
proach effective in damage identification and localization of However, under the noisy condition, the healthy structure
complex composite structures. This method was applied for showed CCDM values greater than the structure with
the detection and localization of composite debonding. damage A. CCSD variations were well clustered under
Good results were obtained even though holes and stiffeners noisy conditions, and values of the healthy structure were
were present in the monitored area, and a high probability of less than both damaged cases. Tang et al.100 proposed the
detection (POD) was observed for all investigated damage very first EMI-based FE and analytical models to monitor
conditions. It is advised to use a proper self-diagnosis the curing process of structural adhesives used with
14 Journal of Reinforced Plastics and Composites 0(0)

Table 1. Comparison of different distributed fibre optic sensing techniques.

Sensing technology Key measurand Measurement resolution Sensing range Spatial resolution Reference

Raman OTDR Temperature 1°C 25 km 1.13 m 30

0.7°C 25 km 2.58 m
BOTDR Temperature and strain 0.93°C/19.48 µε 10 km 3m 31
BOTDA Temperature and strain 2°C 2 km 2 cm 85
1.5°C/30 µε 150 km 2m 33
BOCDA Temperature and strain — 3 km 2 mm 32
Rayleigh OFDR/OBR Temperature and strain 0.8°C/7 µε 170 m 6.5 mm 29

Figure 10. Guided wave-based composite damage detection method under varying temperatures. (a) DIs under varying
temperatures. (b) Location error of damage diagnosis methods.92

fibre-reinforced polymer composites in concrete structural picker to locate impact damages on composite structures
elements. These developed models can determine the elastic with high accuracy. A composite panel was tested upon
modulus and strength of adhesives across different curing impact damage to compare the accuracy of the ToA,
durations. In addition, the acquired conductance signatures Delta-t and AIC Delta-t methods in localizing the de-
were used to develop empirical relationships to find the laminated area. Figure 12(a) delineates the high accuracy
ultimate tensile strength of the adhesive layer. and precision of the proposed AIC Delta-t method in
Some damage indexes used in EMI-based SHM are locating the impact damage, and the low Euclidean
tabulated below (Table 2). distance error for each impact event shown in Figure
12(b) reassures the reliability of this technique.
Acoustic emission-based methods. In SHM, AE measure- Zhou et al.104 proposed a tri-variate kernel density
ments are used for the detection, localization and clas- estimator-based method to further enhance the AE source
sification of damages. Previous studies confirm that when localization. Researchers focus on developing velocity-free
applied in noiseless surroundings, AE can effectively AE source localization techniques based on artificial neural
detect early stages of delamination and debonding in networks, time reversal and probabilistic approaches. Godin
composites.38,39 As AE-based methods can observe et al.105 proposed a novel method to identify a critical time
elastic waves having frequencies up to 1 MHz, these are with acoustic emission monitoring of ceramic matrix
capable of detecting fibre breakage, matrix cracking and composites during static fatigue tests. Beyond the point
fibre pull-out of composite structures.40 The time of ar- characterized by this critical time, acoustic energy release
rival (ToA) method is not suitable to locate the source in can be extrapolated using the Benioff law to evaluate the
complex composite structures as the anisotropic nature time of failure. The evolution of both the emission co-
makes it difficult to determine the wave velocities. efficient RAE and the attenuation coefficient B was con-
Pearson et al.103 proposed a novel method combing sidered as indicators for composite lifetime prediction in
Delta-t mapping with Akaike information criteria (AIC) this study.
Bandara et al. 15

Figure 11. FEM for the simulation of Lamb wave propagation.94 (a) FEM damage configurations; (b) investigated winglet with
the damaged area in red; (c) baseline signal versus damaged configuration signal received by sensor 7 at 100 kHz central frequency;
(d) numerical versus experimental DIs at 100 kHz.94

Machine learning for composite versatile learning procedures. DL-based SHM attracted
SHM systems significant attention recently due to advancements in big
data and cloud-based computing, computer hardware and
Automated damage detection in SHM can be classified as software, data science and transfer learning.4
physics-based and data-driven techniques. Although
physics-based methods are suitable for monitoring simple
structures in well-controlled environments, the dependency
on system physics makes these methods ill-suited for
Application of deep learning in composite SHM
complex real-world structures.3 As the complexity of the Traditional ML methods require domain-based knowledge
structure increases, an SHM technique that is primarily data- and expertise for the design and selection of appropriate
driven is preferred, as shown in Figure 13. Data-driven features which can be problematic when applied to the
methods are preferred over physics-based techniques as the monitoring of complex structures. DL methods intend to
rapid growth in ML techniques3,5 allowed for circum- develop fully automated damage detection by providing an
venting challenges inherent in the latter approach. Artificial end-to-end system that does not require any human in-
neural networks have been widely used for automated volvement. Further, the feature extraction module and
damage detection in recent years. Different system archi- pattern classification/regression module can be trained
tectures like the standard feed-forward multi-layer per- concurrently.42
ceptron (MLP) and the radial basis function (RBF) are used Stacked autoencoder, restricted Boltzmann machine
with ANNs.106 Various hand-engineered damage-sensitive (RBM), CNN and recurrent neural network (RNN) are some
features used for artificial damage detection with ANNs of the frequently used DL techniques. Autoencoders are
may be application/structure dependant. DL, considered used for unsupervised learning of a dataset, and di-
a sub-branch of ML, provides a solution to this problem mensionality reduction of input data is achieved despite
owing to the ability to extract features from raw data using having noise. Deep belief networks (DBNs) and deep
16 Journal of Reinforced Plastics and Composites 0(0)

Table 2. Damage indexes used for EMI-based SHM.

Damage index Remark Reference

Root mean square deviation It is suitable for localization and characterization of incipient damage. It can only 37,101
(RMSD) characterize the correlation between two variables in a single frequency range at a time
Mean absolute percentage It is suitable for localization and characterization of incipient damage. Sensible to 35,37
deviation (MAPD) environmental noise
Covariance (Cov) It is suitable for determining the size of growing damage 37
Correlation coefficient deviation It is suitable for determining the size of growing damage 36,37
(CCD)
Cross-correlation square It is more effective than the CCD index in a noisy environment 36
deviation (CCSD)
Chessboard distance (CB) It is sensitive to growing damage at varying temperatures 37,102
Ellipse Damage Index (EDI) It is effective and reliable to identify the severity of incipient damage. Better at damage 101
quantification than conventional RMSD

Boltzmann machines (DBMs) based on RBM can learn hidden layer and non-linear activation units. Mean-squared
complex data representations using limited labelled data. error (MSE) and regression value (R-value) are utilized to
CNNs are much preferred over the other DL algorithms in evaluate the reconstruction accuracy of input features and
the SHM research community as results can be obtained prediction quality of stiffness parameters during both re-
having less mathematical expertise. RNNs are compara- duction and learning phases, respectively. Layer-wise pre-
tively deeper than other neural networks and generally used training and fine-tuning have been adopted to enhance the
for applications related to dynamic systems.107 The most overall performance of the framework to handle large
recent applications of DL in SHM using these architectures amounts of training data. The proposed framework out-
are presented in the section Autoencoder to recurrent neural performed traditional ANN approaches in terms of accuracy
network. and efficiency. This method only employed natural fre-
quency and mode shapes as inputs. However, the sensitivity
Autoencoder. Autoencoder is an unsupervised neural net- of this network must be improved in order to detect minor
work that learns features via a backpropagation algorithm. damages in harsh operation conditions with random noise.
The autoencoder network structure employs encoder and Therefore, other vibrational characteristics such as FRF and
decoder functions to reconstruct the data. The input data are modal flexibility can be incorporated as input parameters to
mapped to a low-dimensional space during the encoding enhance the performance.
stage in order to acquire the apt features, which are then Pathriage et al.44 introduced a deep sparse autoencoder
mapped on to the output space during the decoding stage to framework (Figure 14(b)) which is a step-up from the
reconstruct the input, keeping the reconstruction error to previously explained AutoDNet technique. Multiple input
a minimum. Autoencoders are useful in performing nu- features may contribute extraneous information to the
merous functions, including but not limited to effective learning process when applied in a high-dimensional space.
feature learning, anomaly detection, dimensionality re- The measurement noise and data redundancy could be the
duction and data de-noising. reasons for this. As a result, the outcomes are less accurate
Pathirage et al.43 proposed an autoencoder-based than when employing compressed dimensional features. To
framework (AutoDNet) for structural damage identifica- overcome this issue, the proposed sparse framework con-
tion. Structural vibration characteristics, including the sists of three components: data pre-processing, sparse
natural frequency and mode shapes, were used as input, and nonlinear dimensionality reduction and relationship learn-
changes in structural material properties (e.g. stiffness re- ing, which are connected sequentially. The input data is
duction parameters) were considered as the output. The subjected to a data whitening process in the pre-processing
proposed framework has two components as shown in component, where data whitening is used to make the input
Figure 14(a), where the reduction component reduces the data less redundant and uncorrelated with each other. The
dimensionality while conserving the essential information, primary objective of the sparse dimensionality reduction
and the relationship learning component executes the re- component is to compress the dimensionality of the features
lationship learning between compressed features and the while keeping essential data sensitive to local structural
stiffness reduction parameters of the structure. The re- damage, system uncertainties and measurement noise. In
duction component comprises an autoencoder with a deep this component, a deep neural network based on a sparse
architecture and non-linear activation units to perform non- autoencoder model is trained, with the first hidden layer
linear dimensionality reduction, while the relationship performing the feature fusion for both structural frequencies
learning is executed using a basic autoencoder with a single and mode shapes. The subsequent hidden layers further
Bandara et al. 17

Figure 12. Improved acoustic emission source location during impact events: (a) identified area of interest and a close-up of the
damage region; (b) Euclidean distance error for each impact event.103

compress these features. Although not exactly the standard which can hamper real-time performance. Hence, the au-
deep autoencoder (DAE) model with the decoding structure, toencoder is used to effectively encode the original input
this model can be viewed as the encoding architecture of while preserving the essential data at reduced dimensions.
a typical DAE. A deep neural network is formed by Real-time localization and load estimation are achieved by
combining the sparse dimensionality reduction and re- feeding the encoded depth images into two distinct su-
lationship learning components, and the proposed sparse pervised ANNs having two hidden layers each. The pro-
autoencoder model is trained using a layer-wise pre-training posed method achieved accuracies of 94.3% and 92.7% for
method. The suggested framework significantly enhances localization and estimation, respectively. This study was
the efficacy and robustness of structural damage detection carried out in laboratory conditions on a composite UAV
compared to the AutoDNet method. The accuracy may be wing. This technique will be developed for the localization
further improved by utilizing other vibration characteristics and estimation of high dynamic loading of larger aircraft
that are more sensitive towards structural damage as input. wings.
Bilal et al.108 proposed an autoencoder-based artificial Rautela et al.109 proposed a convolutional autoencoder
neural network SHM framework to localize and estimate the (CAE) based anomaly detection algorithm to detect de-
loads causing deflections to airframe structures in real-time. laminations of complex aircraft composite panels (Figure
Depth sensors are used to obtain full field depth meas- 14(c)). Similar to a typical autoencoder, a CAE is trained to
urements of aircraft wings. An autoencoder coupled with recreate its input while compressing it to a lower di-
two distinct supervised ANN architectures is used for the mensional representation. Multiple convolutional layers are
real-time localization and estimation of loads. The depth employed by the CAE for encoding and decoding processes.
sensors provide high-quality raw data having a large size While maintaining local spatiality, convolutional layers
18 Journal of Reinforced Plastics and Composites 0(0)

Figure 13. Difference between data-driven and physics-based SHM techniques.

incorporate the principles of sparse connectivity, parameter possible damage scenarios. Random noise is an issue when
sharing and equivariant representation. The reconstruction implementing this technique with real complex structures.
error is recorded for each training example during the The learning ability can be enhanced by adding extra sparse
training phase, and a threshold is established to distinguish frameworks.
abnormal signals from regular signals based on the distri-
bution of reconstruction error. The 99th percentile of this Convolutional neural network. A CNN is a DL network ar-
distribution is utilized as the threshold in this work. If the chitecture that is inspired by the mammalian visual cortex
reconstruction error of a test set exceeds the threshold, the structure. CNNs are best known for image classification and
signal is categorized as an anomaly. This proposed un- identification, which were extensively used in object rec-
supervised methodology successfully identified de- ognition and computer vision applications. CNNs are ef-
lamination with high accuracy and is suitable for complex fective in classifying data that is not in the form of images,
cases where it is challenging to collect training data for all such as audio, time series and signal data. In contrast to
Bandara et al. 19

Figure 14. Autoencoder-based detection techniques: (a) AutoDNet framework;43 (b) sparse autoencoder framework;44 (c)
convolutional autoencoder-based approach.109

traditional models, CNNs gather data, train the model and Optimizing the CNN configuration and variables is antic-
then automatically classify the features for better classifi- ipated to improve the classification performance of the
cation. A typical CNN is composed of three components: CNNs. This CNN model may be trained to detect a wide
convolution, max pooling and activation function. range of damages with high accuracy by incorporating
In SHM, CNNs are mostly used in vision-based and simulated damage data in the training process of the CNN.
vibration-based damage detection methods. Abdeljaber Tabian et al.45 proposed a CNN-based passive sensing
et al.46 proposed a real-time vibration-based SHM system metamodel for impact detection, localization and charac-
using a 1-D CNN. An adaptive 1-D CNN architecture is terization of complex composite structures. The impacts are
employed in this study to integrate the feature extraction and monitored and recorded in real time using a distributed
learning stages of raw accelerometer data. Any discrepancy wireless sensory network, and the proposed CNN frame-
in the input layer dimensions is supported by the adaptive work is then used to process and analyze the recorded events
CNN architecture which consists of an input layer, hidden Figure 15(a)–(c). For each sensor, distinct signals captured
CNN layers and fully connected layers that terminate with by the sensor network are visualized as a surface map of the
an output layer. For each of the 30 joints equipped with voltage against time. Different surface maps are generated
accelerometers in this study, the proposed algorithm calls for each impact location and then converted into a 2D
for developing and training a unique 1-D CNN for damage image. Salient information such as the ToA and amplitude
detection and localization. Back propagation is used for of the signals is preserved while the signal is transformed to
training the CNN. Both feature extraction and classification a colour scale image. This is vital as the ToA and signal
were fused into a single learning body to perform intended amplitude strongly correlate with impact location and im-
tasks simultaneously. This proposed method is capable of pact energy level, respectively. Two separate CNNs are
extracting optimal damage sensitive features automatically employed in this study, one for impact localization and the
from raw acceleration signals. The trial and error method other for impact categorization based on the energy level.
was used to select CNN parameters including the number of The structure is segmented into sub-regions for impact
layers, neurons, kernel size and sub-sampling factor. localization, each of which indicates a localization category.
20 Journal of Reinforced Plastics and Composites 0(0)

Figure 15. CNN-based passive sensing metamodel. (a) Passive sensing with embedded wireless sensor networks and CNN.(b) 2D
image used as input to CNN for impact localization. (c) Methodology used in the proposed framework.45

Three classes are defined as safe, alert and damaged for the Seventekidis proposed an SHM framework with hier-
impact energy level prediction, based on the threshold of archical DL CNN classifiers trained with simulated FE
damage initiation in each sub-region of the structure. The data.110 The simulated learning dataset is fed to a CNN,
accuracy of impact detection was well over 94% when where the classifier is trained for each stage of the four-stage
applied to a composite stiffened panel of an aircraft fuse- hierarchical health monitoring problem. A multi-headed
lage. The scalability of the metamodel was investigated, and CNN composed of four separate convolutional filters is
the accuracy of impact detection was over 95%. This re- employed in this study. In the first stage, the SHM classifier
search confirmed that CNNs can be trained with fewer data performs binary damage detection to distinguish the healthy
by using data from only 280 impacts while achieving high state from the damaged state. Eight different damage levels
accuracy. More studies need to be conducted to assess the (D1–D8) have been classified for this study. If there is
effect of varying operational and environmental conditions damage, a subsequent SHM classifier is used in the second
on the prediction accuracy of the model. stage to execute damage classification to determine whether
Rezende et al.47 proposed a new method to apply CNNs to a single or a multi-damage scenario has been detected. The
EMI-based SHM by using a 5-layer 1-D CNN model. Its third and fourth stage classifiers then operate independently
architecture consists of one convolutional layer, one pooling for these single or multi-damage instances, screening even
layer (max pooling) and two fully connected layers. The more via binary separation until a single prediction from D1
adopted CNN architecture models the EMI signatures for two to D8 is reached. A CFRP pin-joined truss structure was
states: baseline and damaged conditions. Consequently, the used for this study, and the experiments confirmed that the
output layer of CNN model has only two neurons that rep- FE trained CNNs generalize damages with larger magnitude
resent the likelihood of that state. In this study, three aluminium more accurately with higher confidence. The proposed
beams were studied under three different temperature levels hierarchical CNN classifiers were compared with a non-
and discovered that the proposed method is capable of hierarchical direct multiclass classifier. It was concluded
identifying damage under environmental temperature varia- that the proposed SHM framework is appropriate for
tions. Although electromagnetic impedance signatures are complex structures having multiple damage scenarios.
two-dimensional arrays, this study implemented a one-
dimensional approach by only considering the impedance Recurrent neural network. An elastic wave containing vital
amplitude. In addition, each frequency point is used as input to information was generated and propagated through the
the CNN instead of damage matrices. These changes have structure upon a low-velocity impact event. This propa-
significantly reduced the computational time. gating wave, also known as the wavefield, was visualized
 Bandara et al.

Table 3. Deep learning algorithms used in SHM.

Deep learning SHM

technique Framework Algorithm remark technique Reference

Autoencoder AutoDNet • Dimensionality reduction is achieved while preserving essential information Vibration- 43
• Relationship learning between compressed features and the stiffness reduction parameters of based
the structure is performed
• Layer-wise pre-training and fine-tuning improve the training efficiency and accuracy of the
framework
• Outperformed ANNs using stochastic gradient descent (SGD) and scaled conjugate gradient
(SCG) algorithms
Sparse autoencoder • Employs more hidden layers than the AutoDNet framework Vibration- 44
• Data whitening method is used as a pre-processing mechanism based
• Sparse regularization term used in training phase to improve damage identification
Convolutional autoencoder • Useful in complex situations where supervised learning cannot be implemented because it is PZT guided 109
difficult to collect possible signatures for different damage scenarios waves
• Capable of identifying delamination in aerospace composite panels with high accuracy
• The performance of the algorithm is verified by low MSE, low MAE and high coefficient of
determination
(continued)
21
Table 3. (continued)
22

Deep learning SHM

technique Framework Algorithm remark technique Reference

CNN 1-D CNN • Both feature extraction and classification are fused into a single learning body to perform tasks Vibration- 46
simultaneously based
• Capable of extracting optimal damage sensitive features automatically from raw acceleration
signals
• The proposed algorithm is decentralized where a unique CNN is assigned to each location and
processes only the locally available data to assess the structural condition at its location
• Damage detection and localization of real skin-to-stringer composite panel used in aircrafts PZT guided 111
• Classification algorithm uses the ADAM optimization to improve performance in both training waves
and testing phases
• Generalization abilities of the CNN algorithm were tested using different damage scenarios
• Further studies should be conducted to determine the effect of varying operational and
environmental conditions
CNN-based passive sensing • Uses a novel method to transform the raw sensory data to 2D images before feeding as input to PZT guided 45
metamodel for impact detection a CNN waves
• Impact detection, localization and characterization of complex composite structures were
achieved with high accuracy
• Able to extract optimal features automatically from unprocessed data
• Can be used in complex structures
• More studies need to be performed on the automatic optimization of the metamodel and to
determine the effect of operational and environmental variables on the accuracy of the
prediction
FE-trained hierarchical CNN • FE generated data was used for both healthy and damage scenarios to train CNN classifiers Vibration- 110
classifiers • Used for damage detection and localization based
• Experiments confirmed that FE-trained CNNs able to generalize damages with larger
magnitudes
• Hierarchical CNN classifiers are suitable for complex structures with multiple damage
scenarios
• Direct multiclass (non-hierarchical) classifiers are preferred with simple SHM problem
formulations
RNN CNN–RNN network architecture • CNN component can determine the impact location by analyzing each frame of the input PZT guided 41
for impact diagnosis wavefield waves
• RNN reconstructs the impact force-time history treating the problem as a sequence problem
• Spatio-temporal information extraction allows both impact localization and reconstruction of
impact force-time history to be performed simultaneously
Journal of Reinforced Plastics and Composites 0(0)
Bandara et al. 23

using a series of images in the time-space domain. Zargar be detected only after it grows to a detectable size. SHM
and Yuan proposed a combined CNN–RNN network ar- techniques used for composite structures have their inherent
chitecture for impact diagnosis (both impact localization advantages, and limitations sensitive to specific or multiple
and reconstruction of impact force-time history) via au- damage types. Prognosis can be seen as an imperative step
tonomous wavefield analysis.41 The spatial encoding ca- in SHM as it reveals information about the remaining
pability of the CNN and the temporal learning capability of lifetime of the structure. However, the health monitoring of
the encoder–decoder RNN are combined into a single large, complex real-world composite structures using
network in a robust way to achieve the final output. CNN physics-based methods seems unreliable as varying envi-
component can determine the impact location by analyzing ronmental and operational loads can often lead to false-
each frame of the input wavefield separately, while the RNN positive alarms. Vast technological improvements in
reconstructs the impact force–time history treating the wireless sensor networks (WSNs), cloud-based computing
problem as a sequence problem. Both components of impact and ML techniques have paved the way for data-driven
diagnosis were performed simultaneously, courtesy of the SHM techniques. The integration of ML and DL techniques
spatio-temporal information extraction of the combined with the conventional SHM methods has greatly improved
CNN–RNN network architecture. Key characteristics of the the overall performance and reliability of health monitoring
DL networks reviewed in this paper are tabulated in Table 3. techniques.
Herein, an overview of the recent applications of con-
ventional composite SHM techniques across various industries
Future direction of ML-based SHM
and recent advances with an emphasis on physics-informed
A typical SHM system comprises several key subsystems data-driven techniques are presented. The application of
which include a sensory system, data acquisition and conventional ML approaches for complex composite struc-
transmission system, a data processing and analysis system, tures requires domain knowledge for the design and selection
a data storage and management system and a structural of appropriate features. Insufficient training data regarding all
health evaluation and decision-making system. Future re- the possible damage scenarios is a major problem in automated
search will focus on developing all these subsystems in SHM. Fully automated damage detection systems that do not
order to enhance the overall performance of the SHM require any human interaction can be achieved by in-
systems significantly. Energy consumption can be an issue corporating DL algorithms. Autoencoders were initially used
where large sensory networks are used in certain SHM for feature learning, dimensionality reduction and non-linear
methods. The development of self-powered sensors will regression. Damage detection capability was improved by
resolve the aforementioned issue. More research needs to be employing more hidden layers, a data whitening process as
carried out on developing and applying high-speed wireless a pre-processing mechanism and a sparse regularization
data transmission techniques for SHM systems. Environ- term during the training phase. Convolutional au-
mental and operational noise is a major problem for DL- toencoders can be applied to supervised learning models,
based SHM methods, and it affects the accuracy. Advanced where it is difficult to acquire signals for every possible
data whitening algorithms are required to counter the effects damage scenario. CNN-based SHM techniques can extract
of environmental and operational noise. With the rapid optimal damage sensitive features automatically from raw
growth of the Internet of things (IoT), it will be imperative to sensory data. The ability to perform both feature extraction
maintain open-source cloud-based databases to store large and classification simultaneously is a key advantage of
datasets related to composite structures across different CNN models. According to recent research findings, CNN-
industries. As discussed in the section Application of deep based models can be trained with less data and can suc-
learning in composite SHM, DL-based SHM techniques cessfully detect damages under varying environmental
have enhanced the performance and accuracy of SHM conditions. CNN models have been successfully tested for
systems significantly. More developments of advanced DL composite impact damage detection, and superior results
algorithms can be expected in the near future to enable fully were achieved by using a combined CNN–RNN model. It
automated in-situ real-time damage diagnosis and prognosis is of paramount importance to address the challenges and
technologies. limitations of the existing SHM techniques. Although
composite SHM has evolved markedly during the past
decade, none of these techniques can be categorized as
Conclusions
fully automated. Most research studies were conducted in
A state-of-the-art review of the development of cutting-edge idealized laboratory conditions. Further research is re-
SHM techniques being used for fibre-reinforced composite quired to apply in-situ DL-based methods for real-world
materials and intelligent failure predictions is presented. applications in the presence of varying environmental and
Damage detection and localization must be performed at operational noises. It will be vital to maintain cloud-based,
a very early stage to prevent complete structural failure. A open-source databases related to composite structures
notable hurdle of damage detection is that the damage will across numerous industries. In the future in-situ, fully
24 Journal of Reinforced Plastics and Composites 0(0)

automated, real-time health monitoring and transfer 11. Zangana S, Epaarachchi J, Ferdous W, et al. A novel hy-
learning with DL algorithms will ensure the enhanced bridised composite sandwich core with glass, kevlar and
safety of the composite structures. zylon fibres – investigation under low-velocity impact. Int J
Impact Eng 2019; 137: 103430. DOI: 10.1016/j.ijimpeng.
2019.103430.
Declaration of conflicting interests
12. Zangana S, Epaarachchi J, Ferdous W, et al. Behaviour of
The author(s) declared no potential conflicts of interest with continuous fibre composite sandwich core under low-
respect to the research, authorship and/or publication of this velocity impact. Thin Walled Struct 2020; 158: 107157.
article. DOI: 10.1016/j.tws.2020.107157.
13. Montalvão D, Ribeiro AMR and Duarte-Silva JAB. Ex-
Funding perimental assessment of a modal-based multi-parameter
method for locating damage in composite laminates. Exp
The author(s) received no financial support for the research, au- Mech 2011; 51(9): 1473–1488. DOI: 10.1007/s11340-011-
thorship and/or publication of this article. 9472-5.
14. Duchene P, Chaki S, Ayadi A, et al. A review of non-
ORCID iDs destructive techniques used for mechanical damage assess-
ment in polymer composites. J Mater Sci 2018; 53(11):
Senaka Bandara  https://orcid.org/0000-0003-2375-6778
7915–7938. DOI: 10.1007/s10853-018-2045-6.
Madhubhashitha Herath  https://orcid.org/0000-0002-6796-
15. Manoach E, Warminski J, Kloda L, et al. Numerical and
0802
experimental studies on vibration based methods for de-
tection of damage in composite beams. Compos Struct 2017;
References 170: 26–39. DOI: 10.1016/j.compstruct.2017.03.005.
1. Güemes A, Fernandez-Lopez A, Pozo AR, et al. Structural 16. Toh G and Park J. Review of vibration-based structural
health monitoring for advanced composite structures: a re- health monitoring using deep learning. Appl Sci 2020; 10(5):
view. J Compos Sci 2020; 4(1): 13. DOI: 10.3390/ 1680. DOI: 10.3390/app10051680.
jcs4010013. 17. Di Sante R. Fibre optic sensors for structural health moni-
2. Sony S, Laventure S and Sadhu A. A literature review of toring of aircraft composite structures: recent advances and
next-generation smart sensing technology in structural health applications. Sensors (Switzerland) 2015; 15(8):
monitoring. Struct Control Heal Monit 2019; 26(3): 1–22. 18666–18713. DOI: 10.3390/s150818666.
DOI: 10.1002/stc.2321. 18. Kahandawa GC, Epaarachchi JA, Wang H, et al. Use of fixed
3. Yuan F-G, Zargar SA, Chen Q, et al. Machine learning for wavelength fibre-bragg grating (FBG) filters to capture time
structural health monitoring: challenges and opportunities. domain data from the distorted spectrum of an embedded
Bellingham, DC: SPIE, 2020, p. 2. DOI: 10.1117/12. FBG sensor to estimate strain with an artificial neural net-
2561610. work. Sensors Actuat A Phys 2013; 194: 1–7. DOI: 10.1016/
4. Azimi M, Eslamlou AD and Pekcan G. Data-driven struc- j.sna.2012.12.028.
tural health monitoring damage detection through deep 19. Kahandawa GC, Epaarachchi J, Wang H, et al. Extraction
learning: state-of-the- art review. Sensor 2020; 20(10): 2778. and processing of real time strain of embedded FBG sensors
5. Flah M, Nunez I, Ben Chaabene W, et al. Machine learning using a fixed filter FBG circuit and an artificial neural net-
algorithms in civil structural health monitoring: a systematic work. Meas J Int Meas Confed 2013; 46(10): 4045–4051.
review. Arch Comput Methods Eng 2020; 28: 0123456789. DOI: 10.1016/j.measurement.2013.07.029.
DOI: 10.1007/s11831-020-09471-9. 20. Hafizi ZM, Epaarachchi J and Lau KT. Impact location
6. Xu D, Liu PF, Chen ZP, et al. Delamination analysis of determination on thin laminated composite plates using an
carbon fiber/epoxy composite laminates under different NIR-FBG sensor system. Meas J Int Meas Confed 2015; 61:
loading rates using acoustic emission. J Fail Anal Prev 2019; 51–57. DOI: 10.1016/j.measurement.2014.08.040.
19(4): 1034–1042. DOI: 10.1007/s11668-019-00691-1. 21. Kakei AA, Islam M, Leng J, et al. Use of an elasto-plastic
7. Mitsheal AD, Diogo M, Opukuro D, et al. A review of model and strain measurements of embedded fibre bragg
structural health monitoring techniques as applied to com- grating sensors to detect mode I delamination crack propa-
posite structures. Struct Durab Heal Monit 2017; 11(2): gation in woven cloth (0/90) composite materials. Struct
91–147. DOI: 10.3970/sdhm.2017.011.091. Heal Monit 2018; 17(2): 363–378. DOI: 10.1177/
8. Balaji R and Sasikumar M. Structural health monitoring 1475921717694812.
(SHM) system for polymer composites: a review. Indian J Sci 22. Kakei A, Epaarachchi JA, Islam M, et al. Evaluation of
Technol 2016; 9(41): 1–12, doi:10.17485/ijst/2016/v9i41/ delamination crack tip in woven fibre glass reinforced
85832. polymer composite using FBG sensor spectra and
9. Rajak DK, Pagar DD, Kumar R, et al. Recent progress of thermo-elastic response. Meas J Int Meas Confed 2018;
reinforcement materials: a comprehensive overview of 122: 178–185. DOI: 10.1016/j.measurement.2018.03.
composite materials. J Mater Res Technol 2019; 8(6): 023.
6354–6374. DOI: 10.1016/j.jmrt.2019.09.068. 23. Kahandawa GC, Epaarachchi J, Wang H, et al. Use of FBG
10. Aleksendrić D and Carlone P. Introduction to composite sensors for SHM in aerospace structures. Photo Sensors
materials, 2015. 2012; 2(3): 203–214. DOI: 10.1007/s13320-012-0065-4.
Bandara et al. 25

24. Ma Z and Chen X. Fiber bragg gratings sensors for aircraft 37. Na WS and Baek J. A review of the piezoelectric electro-
wing shape measurement: recent applications and technical mechanical impedance based structural health monitoring
analysis. Sensors (Switzerland) 2019; 19(1): 55. DOI: 10. technique for engineering structures. Sensors (Switzerland)
3390/s19010055. 2018; 18: 1307. DOI: 10.3390/s18051307.
25. Lamberti A, Luyckx G, Van Paepegem W, et al. Detection, 38. Michalcová L, Bělský P and Petrusová L. Composite panel
localization and quantification of impact events on a stiffened structural health monitoring and failure analysis under
composite panel with embedded fiber bragg grating sensor compression using acoustic emission. J Civ Struct Heal
networks. Sensors (Switzerland) 2017; 17(4): 1–13. DOI: 10. Monit 2018; 8(4): 607–615. DOI: 10.1007/s13349-018-
3390/s17040743. 0302-7.
26. Esmailzadeh Noghani F, Tofighi S, Pishbin N, et al. Fast and 39. Li H, Li J, Yun X, et al. Non-destructive examination of
high spatial resolution distributed optical fiber sensor. Opt interfacial debonding using acoustic emission. Dent Mater
Laser Technol 2018; 115: 277–288. DOI: 10.1016/j. 2011; 27(10): 964–971. DOI: 10.1016/j.dental.2011.06.
optlastec.2019.02.033. 002.
27. Lalam NR, Lu P, Venketeswaran A, et al. Raman-assisted 40. Crivelli D, Mario G, Mark E, et al. Localisation and
BOTDA performance improvement with the differential identification of fatigue matrix cracking and delamination
pulse-width pair technique and an artificial neural network in a carbon fibre panel by acoustic emission. Compos B
based fitting algorithm. In: Proceedings of the Autonomous Eng 2015; 74: 1–12. DOI: 10.1016/j.compositesb.2014.
Systems: Sensors, Processing, and Security for Vehicles and 12.032.
Infrastructure 2020, 27 April–8 May 2020, Online, DOI: 10. 41. Zargar SA and Yuan FG. Impact diagnosis in stiffened
1117/12.2558717. structural panels using a deep learning approach. Struct
28. Zhu H, Lei Y, Yufeng Z, et al. Optimized support vector Heal Monit 2021; 20(2): 681–691. DOI: 10.1177/
machine assisted botda for temperature extraction with ac- 1475921720925044.
curacy enhancement. IEEE Photon J 2020; 12(1): 1–14. 42. Zhao R, Yan R, Chen Z, et al. Deep learning and its ap-
DOI: 10.1109/JPHOT.2019.2957410. plications to machine health monitoring. Mech Syst Signal
29. Li W, Chen L and Bao X. Compensation of temperature and Process 2019; 115: 213–237. DOI: 10.1016/j.ymssp.2018.
strain coefficients due to local birefringence using optical 05.050.
frequency domain reflectometry. Opt Commun 2013; 311: 43. Pathirage CSN, Li J, Li L, et al. Structural damage identi-
26–32. DOI: 10.1016/j.optcom.2013.08.022. fication based on autoencoder neural networks and deep
30. Liu Y, Ma L, Yang C, et al. Raman distributed temperature learning. Eng Struct 2018; 172: 13–28. DOI: 10.1016/j.
sensor with high spatial and temperature resolution using engstruct.2018.05.109.
optimized graded-index few-mode fiber over 25 km-long 44. Pathirage CSN, Li J, Li L, et al. Development and application
distance. Opt Infobase Conf Pap 2018; Part F124(16): of a deep learning–based sparse autoencoder framework for
20562–20571. DOI: 10.1364/ofs.2018.tue80. structural damage identification. Struct Heal Monit 2019;
31. Sheng L, Ligong L, Jinpeng L, et al. Research on the si- 18(1): 103–122. DOI: 10.1177/1475921718800363.
multaneous distributed measurement of temperature and 45. Tabian I, Fu H and Khodaei ZS. A convolutional neural
strain based on Brillouin scattering effect in communication network for impact detection and characterization of com-
optical fiber. Physics 2021; 1160702: 4, DOI: 10.1117/12. plex composite structures. Sensors (Switzerland) 2019;
2582681. 19(22): 1–25. DOI: 10.3390/s19224933.
32. Ba D, Li Y, Yan J, et al. Phase-coded Brillouin optical 46. Abdeljaber O, Avci O, Kiranyaz S, et al. Real-time vibration-
correlation domain analysis with 2 mm resolution based on based structural damage detection using one-dimensional
phase-shift keying. Opt Express 2019; 27(25): 36197. DOI: convolutional neural networks. J Sound Vib 2017; 388:
10.1364/oe.27.036197. 154–170. DOI: 10.1016/j.jsv.2016.10.043.
33. Dong Y, Chen L and Bao X. Extending the sensing range of 47. de Rezende SWF, de Moura JDRV, Neto RMF, et al. Con-
Brillouin optical time-domain analysis combining. J Light volutional neural network and impedance-based SHM ap-
Technol 2012; 30(8): 1161–1167. plied to damage detection. Eng Res Express 2020; 2(3):
34. Mei H, Haider MF, Joseph R, et al. Recent advances in 35031. DOI: 10.1088/2631-8695/abb568.
piezoelectric wafer active sensors for structural health 48. Fox MR, Schultheisz CR, Reeder JR, et al. Materials ex-
monitoring applications. Sensors (Switzerland) 2019; 19(2): amination of the vertical stabilizer from American airlines
383. DOI: 10.3390/s19020383. flight 587. Mater Sci Technol 2006; 2: 171.
35. Barreto LS, Machado MR, Santos JC, et al. Damage indices 49. Bykowski M, Hudgins A, Deacon RM, et al. Failure analysis
evaluation for one-dimensional guided wave-based structural of the space shuttle Columbia RCC leading edge. J Fail Anal
health monitoring. Lat Am J Sol Struct 2021; 18(2): 1–17. Prev 2006; 6(1): 39–45. DOI: 10.1361/154770206X86509.
DOI: 10.1590/1679-78256292. 50. MAIB. MAIB report no 08/2015 - Cheeki Rafiki - very se-
36. De Castro BA, Baptista FG and Ciampa F. A comparison of rious marine casualty, 2015.
signal processing techniques for impedance-based damage 51. Reder MD, Gonzalez E and Melero JJ. Wind turbine failures
characterization in carbon fibers under noisy inspections. - tackling current problems in failure data analysis. J Phys
Mater Today Proc 2019; 34: 275–279. DOI: 10.1016/j.matpr. Conf Ser 2016; 753(7): 072027. DOI: 10.1088/1742-6596/
2020.03.470. 753/7/072027.
26 Journal of Reinforced Plastics and Composites 0(0)

52. Gabrys J, Schatz J, Kelly C, et al. The Use of LS-DYNA in structural health monitoring of breathing cracks. J Vib
the Columbia Accident Investigation and Return to Flight Control 2020; 27: 2870–2886. DOI: 10.1177/
Activities. 8th International LS-DYNA Users Conference, 1077546320968646.
Dearborn, USA, January 1, 2004, PP. 1-10. 68. Ou Y, Chatzi EN, Dertimanis VK, et al. Vibration-based
53. Shankar Verma A, Zhiyu J, Zhengru R, et al. A probabilistic experimental damage detection of a small-scale wind turbine
long-term framework for site-specific erosion analysis of blade. Struct Heal Monit 2017; 16(1): 79–96. DOI: 10.1177/
wind turbine blades: a case study of 31 dutch sites. Wind 1475921716663876.
Energy 2021; 24(11): 1315–1336, doi:10.1002/we.2634. 69. Huang T, Chaves-Vargas M, Yang J, et al. A baseline-free
54. Airbus. FAST- airbus technical magazine, 2014. Online damage detection method based on node displacement in
Available: https://aircraft.airbus.com/sites/g/files/jlcbta126/ mode shape. In: Proceedings of the 8th European Workshop
files/2022-04/FAST-054.pdf on Structural Health Monitoring, Bilbao, Spain, July, 2016,
55. Werner L, Fabian B, Ferdinand K. et al. Distributed fibre- pp. 765–774.
optic sensing applications at the semmering base tunnel, 70. Huang T, Chaves-Vargas M, Yang J, et al. A baseline-free
Austria. Smart Infrastruc Cons 2019; 172: 148. DOI: 10. structural damage indicator based on node displacement of
1680/jsmic.20.00006. structural mode shapes. J Sound Vib 2018; 433: 366–384.
56. Elena Saldaeva, Report on the demonstration activities of the DOI: 10.1016/j.jsv.2018.07.001.
BladeSave system, Risk Based Technology for Blade 71. Glisic B, Yao Y, Tung STE, et al. Strain sensing sheets for
Structural Assessment, BladeSave Project Consortium, 2020 structural health monitoring based on large-area electronics
57. Das S, Saha P and Patro SK. Vibration-based damage de- and integrated circuits. Proc IEEE 2016; 104(8): 1513–1528.
tection techniques used for health monitoring of structures: DOI: 10.1109/JPROC.2016.2573238.
a review. J Civ Struct Heal Monit 2016; 6(3): 477–507. DOI: 72. Herszberg I, Bannister MK, Li HCH, et al. Structural health
10.1007/s13349-016-0168-5. monitoring for advanced composite structures. ICCM Int
58. Su WC, Huang CS, Hung SL, et al. Locating damaged Conf Compos Mater 2007; 8: 1–15.
storeys in a shear building based on its sub-structural natural 73. Giurgiutiu V. Structural health monitoring (SHM) of aero-
frequencies. Eng Struct 2012; 39: 126–138. DOI: 10.1016/j. space composites. Amsterdam, The Netherlands: Elsevier,
engstruct.2012.02.002. 2019.
59. Wang S. Damage detection in offshore platform structures 74. Kaur A, Nagarajan S, Anandan S, et al. Embeddable fiber
from limited modal data. Appl Ocean Res 2013; 41: 48–56. optic strain sensor for structural monitoring. Bellingham,
DOI: 10.1016/j.apor.2013.02.004. DC: SPIE, 2013, p. 2021. DOI: 10.1117/12.2009462.
60. Dessi D and Camerlengo G. Damage identification techni- 75. Xu C and Khodaei ZS. A novel fabry-pérot optical sensor for
ques via modal curvature analysis: overview and compari- guided wave signal acquisition. Sensors (Basel) 2020; 20(6):
son. Mech Syst Signal Process 2015; 5253(1): 181–205. 1728. DOI: 10.3390/s20061728.
DOI: 10.1016/j.ymssp.2014.05.031. 76. Blazejewski W, Czulak A, Materials NE, et al. SMART
61. Dawari VB and Vesmawala GR. Structural damage identi- composite high pressure vessels with integrated optical fiber
fication of beam structures using two stage method based on sensors SMART composite high pressure vessels with in-
modal strain energy indicators and artificial neural networks. tegrated optical fiber sensors. In: Proceedings of SPIE - The
J Vibroeng 1877; 18(1): 119–128. International Society for Optical Engineering, San Diego,
62. Radzieński M, Krawczuk M and Palacz M. Improvement of California, United States, 31 March 2010, DOI: 10.1117/12.
damage detection methods based on experimental modal 847251.
parameters. Mech Syst Signal Process 2011; 25(6): 77. Guo H, Xiao G, Mrad N, et al. Fiber optic sensors for
2169–2190. DOI: 10.1016/j.ymssp.2011.01.007. structural health monitoring of air platforms. Sensors 2011;
63. Chen LJ and Yu L. Structural nonlinear damage identification 11(4): 3687–3705. DOI: 10.3390/s110403687.
algorithm based on time series ARMA/GARCH model. Adv 78. Ramakrishnan M, Rajan G, Semenova Y, et al. Overview of
Struct Eng 2013; 16(9): 1597–1609. DOI: 10.1260/1369- fiber optic sensor technologies for strain/temperature sensing
4332.16.9.1597. applications in composite materials. Sensors (Switzerland)
64. Mosavi AA, Dickey D, Seracino R, et al. Identifying damage 2016; 16(1): 99. DOI: 10.3390/s16010099.
locations under ambient vibrations utilizing vector autore- 79. Nicolas MJ, Sullivan RW and Richards WL. Large scale
gressive models and Mahalanobis distances. Mech Syst applications using FBG sensors: determination of in-flight
Signal Process 2012; 26(1): 254–267. DOI: 10.1016/j. loads and shape of a composite aircraft wing. Aerospace
ymssp.2011.06.009. 2016; 3(3): 18. DOI: 10.3390/aerospace3030018.
65. Yu L and Lin J-C. Cloud computing-based time series 80. Zhu Q, Xu C and Yang G. Experimental research on damage
analysis for structural damage detection. J Eng Mech 2017; detecting in composite materials with FBG sensors under low
143(1): 1–14. DOI: 10.1061/(asce)em.1943-7889.0000982. frequency cycling. Int J Fatigue 2017; 101: 61–66. DOI: 10.
66. Rao RK and Sasmal S. Smart nano-engineered cementitious 1016/j.ijfatigue.2017.03.034.
composite sensors for vibration-based health monitoring of 81. Zeng H, Yan R, Xu L, et al. Application study on fiber bragg
large structures. Sens Actuat A Phys 2020; 311: 112088. grating sensors in damage monitoring of sandwich composite
67. Sepehry N, Ehsani M, Zhu W, et al. Application of scaled joints. J Sandw Struct Mater 2020; 22(5): 1542–1563. DOI:
boundary finite element method for vibration-based 10.1177/1099636218789621.
Bandara et al. 27

82. Kumar S, Bedi A and Kothari V. Design and analysis of FBG 97. de Oliveira MA, Araujo NVS, da Silva RN, et al. Use of
based sensor for detection of damage in oil and gas pipelines Savitzky-Golay filter for performances improvement of
for safety of marine life. Proceed SPIE 2018: 31: 104880X6. SHM systems based on neural networks and distributed PZT
DOI: 10.1117/12.2285923. sensors. Sensors (Switzerland) 2018; 18(1): 152. DOI: 10.
83. Loayssa A. (2020) Structural Health Monitoring Using 3390/s18010152.
Distributed Fibre-Optic Sensors. In Optical Fibre Sensors 98. Malinowski PH, Wandowski T and Singh SK. Employing
(eds I. Villar and I.R. Matias). https://doi.org/10.1002/ principal component analysis for assessment of damage in GFRP
9781119534730.ch5 composites using electromechanical impedance. Compos Struct
84. Shan Y, Xu H, Zhou Z, et al. State sensing of composite 2021; 266: 113820. DOI: 10.1016/j.compstruct.2021.113820.
structures with complex curved surface based on distributed 99. Zhu J, Wang Y and Qing X. Modified electromechanical
optical fiber sensor. J Intell Mater Syst Struct 2019; 30(13): impedance-based disbond monitoring for honeycomb
1951–1968. DOI: 10.1177/1045389X19849287. sandwich composite structure. Compos Struct 2019; 217:
85. Dong Y, Zhang H, Chen L, et al. 2 cm spatial-resolution and 2 175–185. DOI: 10.1016/j.compstruct.2019.03.033.
km range Brillouin optical fiber sensor using a transient 100. Tang ZS, Lim YY, Smith ST, et al. Modelling of the elec-
differential pulse pair. Appl Opt 2012; 51(9): 1229–1235. tromechanical impedance technique for prediction of elastic
DOI: 10.1364/AO.51.001229. modulus of structural adhesives. Struct Heal Monit 2020;
86. Barrias A, Casas JR and Villalba S. A review of distributed 20(5): 2245–2260.DOI: 10.1177/1475921720916924.
optical fiber sensors for civil engineering applications. 101. Wang D, Wang Q, Wang H, et al. Experimental study on
Sensors (Switzerland) 2016; 16(5): 748. DOI: 10.3390/ damage detection in timber specimens based on an elec-
s16050748. tromechanical impedance technique and RMSD-based ma-
87. Mitra M and Gopalakrishnan S. Guided wave based halanobis distance. Sensors (Switzerland) 2016; 16: 1765.
structural health monitoring: a review. Smart Mater Struct DOI: 10.3390/s16101765.
2016; 25(5): 053001. DOI: 10.1088/0964-1726/25/5/ 102. Wandowski T, Opoka S, Malinowski PH, et al. The per-
053001. formance of three Electromechanical Impedance damage
88. Cawley P. Structural health monitoring: closing the gap between indicators on structural element with bolted joints. e-Journal
research and industrial deployment. Struct Heal Monit 2018; of Nondestructive Testing 2015; 1.
17(5): 1225–1244. DOI: 10.1177/1475921717750047. 103. Pearson MR, Eaton M, Featherston C, et al. Improved acoustic
89. Memmolo V, Monaco E, Boffa ND, et al. Guided wave emission source location during fatigue and impact events in
propagation and scattering for structural health monitoring of metallic and composite structures. Struct Heal Monit 2017;
stiffened composites. Compos Struct 2017; 184: 568–580. 16(4): 382–399. DOI: 10.1177/1475921716672206.
DOI: 10.1016/j.compstruct.2017.09.067. 104. Zhou Z, Riyan L, Yichao R, et al. A new acoustic emission
90. Spada A, Capriotti M and Lanza di Scalea F. Global-local source location method using tri-variate kernel density es-
model for guided wave scattering problems with application timator. IEEE Access 2019; 7: 158379–158388. DOI: 10.
to defect characterization in built-up composite structures. Int 1109/ACCESS.2019.2950225.
J Sol Struct 2020; 182–183: 267–280. DOI: 10.1016/j. 105. Godin N, Reynaud P, R’Mili M, et al. Identification of
ijsolstr.2019.08.015. a critical time with acoustic emission monitoring during
91. Zhang J, Liu K, Gao C, et al. Multi-mode guided waves static fatigue tests on ceramic matrix composites: towards
based reference-free damage diagnostic imaging in plates. lifetime prediction. Appl Sci 2016; 6(2): 43. DOI: 10.3390/
SDHM Struct Durab Heal Monit 2019; 13(1): 41–59. DOI: app6020043.
10.32604/sdhm.2019.05142. 106. Alaimo A, Barracco A, Milazzo A, et al. Artificial neural
92. Gao D, Ma Y, Wu Z, et al. Guided wave based damage networks comparison for a SHM procedure applied to com-
detection method for aircraft composite structures under posite structure. In Proceedings of the Italian Association of
varying temperatures. Struct Durab Heal Monit 2021; 15(1): Aeronautics and Astronautics XXII Conference, Naples, Italy,
23–37. DOI: 10.32604/sdhm.2021.013737. September 2012, pp. 9–12. DOI: 10.13140/2.1.1538.4006.
93. Li J, Lu Y and Lee YF. Debonding detection in CFRP- 107. Sony S, Dunphy K, Sadhu A, et al. A systematic review of
reinforced steel structures using anti-symmetrical guided convolutional neural network-based structural condition
waves. Compos Struct 2020; 253: 112813. DOI: 10.1016/j. assessment techniques. Eng Struct 2020; 226: 111347. DOI:
compstruct.2020.112813. 10.1016/j.engstruct.2020.111347.
94. De Luca A, Perfetto D, Petrone G, et al. Guided-waves in a low 108. Bilal DK, Unel M, Yildiz M, et al. Realtime localization and
velocity impacted composite winglet. Key Eng Mater 2018; 774: estimation of loads on aircraft wings from depth images.
343–348. DOI: 10.4028/www.scientific.net/KEM.774.343. Sensors (Switzerland) 2020; 20(12): 1–24. DOI: 10.3390/
95. Yue N, Khodaei ZS and Aliabadi MH. Damage detectability s20123405.
model of pitch-catch configuration in composite plates. Key 109. Rautela M, Monaco E and Gopalakrishnan S. Delamination
Eng Mater 2017; 754: 387–390. DOI: 10.4028/www. detection in aerospace composite panels using convolutional
scientific.net/KEM.754.387. autoencoders. In Proceedings of the Health Monitoring of
96. Memmolo V, Boffa ND, Maio L, et al. Damage localization Structural and Biological Systems XV, Bellingham, DC,
in composite structures using a guided waves based multi- March 2021. DOI: 10.1117/12.2582993.
parameter approach. Aerospace 2018; 5: 4. DOI: 10.3390/ 110. Seventekidis P and Giagopoulos D. A combined finite ele-
aerospace5040111. ment and hierarchical deep learning approach for structural
28 Journal of Reinforced Plastics and Composites 0(0)

health monitoring: test on a pin-joint composite truss FFT Fast Fourier transform
structure. Mech Syst Signal Process 2021; 157: 107735. DOI: FOS Fibre optic sensing
10.1016/j.ymssp.2021.107735. FPC Fast phase correlation
111. Cui R, Azuara G, Lanza di Scalea F, et al. Damage imaging in FRF Frequency response function
skin-stringer composite aircraft panel by ultrasonic-guided
GE General electric
waves using deep learning with convolutional neural net-
GI-FMF Graded-index few-mode fiber
work. Struct Heal Monit 2021; 21(3): 1–16. DOI: 10.1177/
14759217211023934. GL Global-local
GRP Glass-reinforced plastic
IoT Internet of things
LVID Low-velocity impact damages
Appendix 1 MAE Mean absolute error
MAIB Marine accident investigation branch
Nomenclature MAPD Mean absolute percentage deviation
AE Acoustic emissions ML Machine learning
AIC Akaike information criteria MLP Multi-layer perceptron
ANN Artificial neural network MSE Mean-squared error
ARMA Autoregressive moving average MSP Multiple short pulse
BOTDA Brillouin optical time-domain analysis NAW North American Windpower
BVID Barely visible impact damage NODIS Node displacement
CAE Convolutional autoencoder OFDR Optical frequency-domain reflectometer
CAMS Comprehensive active monitoring scheme OTDR Optical time-domain reflectometer
CB Chessboard distance PCA Principal component analysis
CCDM Correlation coefficient deviation metric POD Probability of detection
CCSD Cross-correlation square deviation PSD Power spectral density
CDS Characteristic damage state PWAS Piezoelectric wafer active sensors
CFRC Carbon fibre-reinforced composites RBF Radial basis function
CFRP Carbon fibre-reinforced polymer RBM Restricted Boltzmann machine
CNC Cement-based nanocomposite RCC Reinforced carbon–carbon
CNN Convolutional neural network RMSD Root mean square deviation
Cov Covariance RNN Recurrent neural network
CWT Continuous wavelet transformation SBFEM Scaled-boundary finite element method
DAE Deep auto encoder SDI Structural damage indicator
DBM Deep Boltzmann machines SHM Structural health monitoring
DBN Deep belief networks SOFO Low coherent interferometric sensors
DCMI Direct coupling mechanical impedance SVM Support vector machines
DFOS Distributed fibre optic sensing TFDI Transfer function-based damage indicator
DI Damage index ToA Time of arrival
DL Deep learning ToF Time of flight
EDI Ellipse damage index VAR Vector autoregressive
EMI Electro mechanical impedance VARMA Vector autoregression moving average
FBG Fibre Briggs grating VARX Vector autoregressive with exogenous inputs
FEM Finite element method WSN Wireless sensor networks.