Received: 18 June 2024 Accepted: 4 September 2024

DOI: 10.1002/pc.29055

RESEARCH ARTICLE

An explainable artificial intelligence-based approach

for reliable damage detection in polymer composite
structures using deep learning

Muhammad Muzammil Azad | Heung Soo Kim

Department of Mechanical, Robotics and

Energy Engineering, Dongguk University- Abstract
Seoul, Seoul, Republic of Korea Artificial intelligence (AI) techniques are increasingly used for structural
health monitoring (SHM) of polymer composite structures. However, to be
Correspondence
Heung Soo Kim, Department of confident in the trustworthiness of AI models, the models must be reliable,
Mechanical, Robotics and Energy interpretable, and explainable. The use of explainable artificial intelligence
Engineering, Dongguk University-Seoul,
(XAI) is critical to ensure that the AI model is transparent in the decision-
30 Pildong-ro 1-gil, Jung-gu, Seoul 04620,
Republic of Korea. making process and that the predictions it provides can be trusted and under-
Email: [email protected] stood by users. However, existing SHM methods for polymer composite
Funding information
structures lack explainability and transparency, and therefore reliable dam-
National Research Foundation of Korea, age detection. Therefore, an interpretable deep learning model based on an
Grant/Award Number: explainable vision transformer (X-ViT) is proposed for the SHM of compos-
2020R1A2C1006613
ites, leading to improved repair planning, maintenance, and performance.
The proposed approach has been validated on carbon fiber reinforced
polymers (CFRP) composites with multiple health states. The X-ViT model
exhibited better damage detection performance compared to existing popular
methods. Moreover, the X-ViT approach effectively highlighted the area of
interest related to the prediction of each health condition in composites
through the patch attention aggregation process, emphasizing their influence
on the decision-making process. Consequently, integrating the ViT-based
explainable deep-learning model into the SHM of polymer composites
provided improved diagnostics along with increased transparency and
reliability.
Highlights
• Autonomous damage detection of polymer composites using vision trans-
former based deep learning model.
• Explainable artificial intelligence by highlighting region of interest using
patch attention.
• Comparison with the existing state of the art structural health monitoring
methods.

This is an open access article under the terms of the Creative Commons Attribution License, which permits use, distribution and reproduction in any medium, provided
the original work is properly cited.
© 2024 The Author(s). Polymer Composites published by Wiley Periodicals LLC on behalf of Society of Plastics Engineers.

1536 wileyonlinelibrary.com/journal/pc Polymer Composites. 2025;46:1536–1551.

 15480569, 2025, 2, Downloaded from https://4spepublications.onlinelibrary.wiley.com/doi/10.1002/pc.29055 by The Director National Institut, Wiley Online Library on [16/05/2025]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
AZAD and KIM 1537

KEYWORDS
composite structures, damage detection, explainable artificial intelligence, laminated
composites, polymer composites, structural health monitoring, vision transformer

1 | INTRODUCTION SVM. Rautela et al.18 used GW signals to identify

delamination in aerospace applications and suggested an
Laminated polymer composites are extensively used in unsupervised approach that combines SVM with principal
mechanical and aerospace equipment due to their excel- and independent component analysis.
lent specific strength and specific stiffness.1,2 However, In addition to these traditional ML approaches,
during the production and service life of these composites, deep learning (DL)-based approaches have also gained
defects like matrix cracking, debonding, delamination, significant popularity. DL approaches generally use
and inclusion are inevitable and pose a threat to the neural networks to independently extract deep features
structure.3–5 Thus, damage detection in polymer compos- from composite datasets.19 Numerous techniques, mainly
ites is an important research topic in both academia and experimental, have been established to automate the
industry. An effective way to prevent catastrophic struc- process of detecting damage in composite structures.
tural collapses is to execute timely maintenance based on Asif et al.8 predicted delamination in smart laminated
inspection results. However, the orthotropic nature of composites by exploiting structural vibrations using a con-
composite restricts inspection-based maintenance as dam- volutional neural network (CNN). An electromechanically
age can occur invisibly in any layer. Therefore, different coupled mathematical model was then employed to obtain
non-destructive testing (NDT) methods are widely used to sufficient training data for both healthy and delaminated
perform their structural health monitoring (SHM).6,7 composites. Wang et al.20 achieved an accuracy of 87.8% in
Vibrational signals serve as a commonly employed NDT detecting four types of damages in composite wind turbine
technique to monitor the structural health of composites. blades utilizing vibration data as input to a CNN model.
They offer insight into the dynamic behavior of structures Rautela and Gopalakrishnan21 extracted GWs from com-
under different operating conditions and environmental posite laminates, converted them into time–frequency
circumstances.8–10 Furthermore, vibration-based damage images using continuous wavelet transform (CWT), and
diagnosis exhibits additional features of convenience, used a 2D-CNN classifier for damage identification along
feasibility, and cost-effectiveness of SHM in engineering with a long short–term memory (LSTM) regressor for
applications.11,12 damage localization. They concluded that DL models,
The field of composite damage detection has largely even when exposed to varying degrees of noise, provide
adopted data-driven machine learning (ML) algorithms, better performance and generalization skills. Using soli-
which have the potential to discover latent correlations tary waves, Yoon et al.22 carried out binary classification
in a given dataset.13 These techniques allow for continu- using CNN and achieved an accuracy of over 90% in find-
ous and real-time monitoring of composite structures, ing defects in composites. Similarly, CNN has also shown
helping to detect damage early and prevent potential better performance compared to other ML models in the
catastrophic failures. Therefore, numerous articles in structural diagnosis of composite panels using both real
the literature discuss the significance of using ML-based and synthetic data.23 Lee et al.24 developed an autonomous
data-driven techniques for composite damage detection. damage detection method utilizing Lamb waves and auto-
Jakkamputi et al.14 evaluated multiple ML models, includ- encoder, implementing singular value decomposition in
ing decision tree (DT), random forest (RF), naive Bayes the feature space of the autoencoder model to condense
(NB), and Bayesian network, to diagnose damage in com- feature dimensionality, resulting in accurate damage
posite structures from vibration data. Das et al.15 classified detection and classification. Park et al.25 autonomously
multiple damage conditions using time-embedding and extracted delamination location in composite laminates
time–frequency methods, revealing superior performance by training a convolutional autoencoder (CAE) on light
of the support vector machine (SVM) with time–embedded emission images obtained from the mechanoluminescent
features over time–frequency features. Another SVM- sensors.
based damage detection model aimed to quantitatively Although the ML and DL models provide good damage
identify, classify, and localize damage in multiple areas of detection performance for laminated composites, all of
composite using guided waves (GWs) with sparse sensor these models are purely black–box models and their predic-
arrays.16 Qiao et al.17 used acoustic emission (AE) for tions may be unreliable, and incorrect predictions may
damage model identification in polymer composites using result in catastrophic failure. Therefore, these models suffer
 15480569, 2025, 2, Downloaded from https://4spepublications.onlinelibrary.wiley.com/doi/10.1002/pc.29055 by The Director National Institut, Wiley Online Library on [16/05/2025]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
1538 AZAD and KIM

from transparency during the decision-making process. 2 | METHODOLOG Y

Understanding why the AI model accurately predicts spe-
cific defects is crucial, as it provides valuable insight into This paper aims to develop an explainable deep learning
the structure being analyzed. Thus, an additional approach model for more robust and reliable damage detection in
or methodology is required to effectively open up the opera- composite structures. To achieve this goal, the intelligent
tion of the black–box DL models and understand the rea- damage detection method is unified with explainable
soning behind the model's predictions. Consequently, the damage detection. The complete workflow of the pro-
development of an explainable artificial intelligence (XAI) posed framework is shown in Figure 1. The entire process
model is imperative to elucidate the rationale behind these comprises four steps: (1) obtaining vibration-based exper-
predictions.26 Moreover, enabling transparency between imental data from multiple health states of the laminated
engineers and AI models also requires the creation of composite, (2) implementing continuous wavelet trans-
explainable models27,28; because end users are likely to formation to convert the 1D raw signal into images,
adopt and prefer interpretable solutions. Currently, several (3) performing data augmentation by adding zero mean
frameworks are used to implement XAI in various domains, Gaussian noise, and (4) development of the X-ViT mod-
such as Local Interpretable Model-Agnostic Explanations ule by integrating damage detection and explanation for
(LIME) and SHapley Additive exPlanation (SHAP).29–32 more reliable intelligent diagnosis. The methodological
These frameworks interpret the results of data-driven contribution of the proposed framework comes from the
models by explaining their predictions. However, the appli- two aspects. On the one hand, the ViT model aims to
cations of these models in damage detection of composite independently extract deep features from the input data.
structures have not yet been explored. Therefore, it is cru- The ViT model works by taking in the scalogram images
cial to decode the black–box AI models and demonstrate and the augmented data. The model autonomously learns
their robustness, explainability, and reliability for laminated the time-frequency domain features present in the images
composites. Since AI may augment human intellect, and adjusts its weights and biases to perform classification.
explainability, and robustness can increase confidence and During classification, the model outputs the health status
dependability while preserving human control.33,34 of composites and predicts either the healthy status or the
In this paper, a vision transformer (ViT) based type of damage occurring in the composites. On the other
explainable damage detection scheme is developed for hand, the developed ViT model is made explainable
polymer composites. The ViT-based DL model is pro- (X-ViT) by calculating the patch weights used to classify
posed because of its excellent performance in several the specific health states. The X-ViT model integrates
computer vision applications.35–37 Moreover, the pro- patch attention across the entire dataset and highlights the
posed ViT model is made interpretable by implementing region of interest to predict each health state. The detailed
the concept of prediction explainability using a patch description of the ViT model and the incorporation of XAI
attention mechanism. The explainable vision transformer to improve the interpretability and reliability of damage
(X-ViT) model aggregates patch attention from the ViT detection in composite structures are presented in the next
model for each patch, identifies the patch receiving the sections.
most attention, and counts its occurrences for each pre-
dicted health state of composites, offering insight into the
model's decision-making process. Therefore, the contri- 2.1 | Vision transformer (ViT)
bution of this study is twofold: (1) developing ViT for the
first time to detect damage in polymer composite lami- Generally, CNN is extensively used for damage detection
nates, and (2) interpreting the predictions using the in engineering applications. However, recently, trans-
X-ViT model. Experimental data acquired from polymer former networks have shown advantages over CNN
composites are used to validate the proposed approach. models due to the use of a self-attention mechanism
The dataset covers multiple health states, including cases to obtain long-range dependencies in the dataset. The
of delamination damage, which is the most severe dam- transformer networks comprise two key components:
age in laminated composites. Moreover, to enhance the (1) attention mechanism, and (2) position encoding. The
performance of the X-ViT model, the raw vibration sig- transformer network was first introduced in 2017 for use
nals are converted to images using CWT and augmented in natural language processing (NLP), based entirely on
using Gaussian noise. Various hyperparameters of the attention to drawing global dependencies between input
ViT model were optimized to further improve the model and output.38 With further progress in the field, a trans-
performance. The performance of the model is assessed former network for computer vision tasks called a vision
using multiple evaluation metrics and compared with transformer, was first realized in 2021.39 The vision trans-
existing popular methods. former network utilizes attention processes that go beyond
 15480569, 2025, 2, Downloaded from https://4spepublications.onlinelibrary.wiley.com/doi/10.1002/pc.29055 by The Director National Institut, Wiley Online Library on [16/05/2025]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
AZAD and KIM 1539

FIGURE 1 Proposed workflow for developing an explainable ViT-based deep learning model for damage detection in laminated
composites.

typical CNN with local receptive fields. This is achieved by Each input sequence vector is transformed into three
directing attention towards distinct regions of images and separate vectors q, k, and v, called query vector, key vec-
combining feature information from the whole image. tor, and value vector, respectively. Therefore, the atten-
Due to these advantages, the attention-based ViT models tion function (AF) with vector dimension d can be
have been utilized in numerous other engineering applica- expressed as:
tions36,40,41 However, this study is the first implementation
 T
of the ViT model for damage detection in polymer com- qk
AF ðq, k,vÞ ¼ SoftMax pffiffiffiffiffi v ð1Þ
posite structures. dk
To recognize the principle and structure of the ViT
model, it is important to first understand the single
attention mechanism. For a given input sequence where non-linearity is introduced in the AF using the Soft-
X ¼ ðx 1 , x 2 , x 3 ,…, x n Þ, the self-attention mechanism Max activation to capture intricate patterns and representa-
acquires knowledge about the entities of a given sequence tions.42 This study utilized the advantages of a ViT-based
and provides the output sequence Z ¼ ðz1 , z2 , z3 , …, zn Þ. deep learning model using a 2D image-based dataset. The
 15480569, 2025, 2, Downloaded from https://4spepublications.onlinelibrary.wiley.com/doi/10.1002/pc.29055 by The Director National Institut, Wiley Online Library on [16/05/2025]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
1540 AZAD and KIM

FIGURE 2 Conceptual framework of the ViT model for damage detection.

conceptual framework of the image-based ViT model for contain both information and the location of the patch.
this study is shown in Figure 2. In this case, the input Moreover, an extra class embedding vector is also
image x  RHWK is divided into a series of smaller introduced to track the image class during learning.
patches x P  RNðq K Þ (such as 16 patches in a 4  4 grid).
2
Ultimately, the data is fed into a conventional trans-
Herein, H and W represents the height and width of the former architecture to capture global correlations and
raw image in pixels, K is the number of image channels, features for each class which also consists of a multi-
and ðq  qÞ represents the patch sizes in pixels. layer perceptron (MLP) layer that helps introduce
Thus, the length of the input sequence for the ViT net- non-linearity to extract complex features. These features
work equals the total number of small patches are then input to the classifier, which may contain several
given as: layers to detect and classify damage.

HW
N¼ ð2Þ
q2 2.2 | Explainable artificial
intelligence (XAI)
Next, the aforementioned single attention function is
implemented H times through the multi-head attention In recent years, XAI has gained significant attention as a
mechanism. The resulting multi-head attention (MH attention ) research topic in the area of artificial intelligence to com-
function is expressed as: prehend the reasons behind the decision of AI models.
Generally, the explanation of AI models is divided into
MH attention ðq,k, vÞ ¼ Concatenateðh1 , h2 , h3 ,…, hH ÞW O ð3Þ four axes using a hierarchical classification scheme:
(1) data explanation, (2) model explanation, (3) post-hoc
 
explanation, and (4) assessment of explanations.43 Data
hi ¼ Attention qW jq ,kW jk , vW jv , i ¼ 1, 2, 3, …, H ð4Þ
explainability encompasses a collection of methodologies
that aim to better understand the data used to develop
where W jq , W jk , W jv , and W O are the learnable hyperpara- AI models. The objective of model explainability is to
meters of the ViT module, and H represents the number develop models that possess inherent clarity or are
of attention heads. considered to be intrinsically explainable. Post-hoc
Meanwhile, a positional encoding layer is used to explainability refers to the retrospective process of
prevent the loss of the position information from the interpreting and providing human-understandable
input image. The position encoding process involves explanations for a trained model's decisions, offering
mapping an integer to a vector so that each vector can insight into factors influencing specific predictions.
 15480569, 2025, 2, Downloaded from https://4spepublications.onlinelibrary.wiley.com/doi/10.1002/pc.29055 by The Director National Institut, Wiley Online Library on [16/05/2025]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
AZAD and KIM 1541

FIGURE 3 Converting the ViT model to an explainable ViT model to develop a reliable and interpretable model for the end user.

Assessment of explanations refers to the level of explain- 3 | DATA ACQUISITION A ND

ability of AI systems and the regulatory actions required to PRE- PR OCES SING
carefully assess the reliability of automated explanations.
In this study, the explainability of the ViT model is The proposed ViT-based explainable damage diagnosis
achieved by adjusting the model architecture as shown method was developed using the vibrational signals
in Figure 3. This is accomplished by the collection of acquired from laminated composites on an experimental
weights given to the input patches of the ViT model, test bed. This section presents the fabrication of the poly-
referred to as patch attention weights. The weights are mer composites and details of the test-bed used to acquire
collected from the trainable layers of the ViT model, the data for various health states of composites. Moreover,
which include the transformer encoder and the classi- this section also discusses the various data pre-processing
fier. These attention weights are then aggregated across steps adopted in this study.
all layers and analyzed to understand which parts of
the input majorly contribute to the prediction of the
model. Therefore, the X-ViT model provides intuition 3.1 | Vibration dataset acquisition
about how the model takes into diverse parts of the
input data (patches), explaining the internal workings Composite damage detection was performed using
of the model itself. The attention weights collected dur- vibration data obtained for multiple composite health
ing training are utilized in the analysis of patch signifi- conditions. For this reason, carbon fiber reinforced poly-
cance, which contributes to the understanding of mer (CFRP) composites with a [0/90/0/90]s ply configu-
patches containing features that the model predicts are ration have been developed. The raw material used was
important for specific classes. It is worth noting that in prepreg T700SC-12k-60E (epoxy-based carbon fiber)
this modified model, the focus is on making the ViT with density = 1.8 g/cm3, tensile strength = 4900 MPa,
model reliable and trustworthy for end users. This is and tensile modulus = 230 GPa. The 350  300 mm2
because the final decision-makers are generally human composite sheets were fabricated using hot press
users. Therefore, the end users such as engineers in compression molding for the three health states as
responsible roles can gain insight into the decision- depicted in Table 1. The composite sheets were cut into
making process of the ViT model, encouraging trans- 300  50 mm2 pieces using a water jet cutting machine
parency and facilitating informed decision-making in to obtain five samples for each health condition. Each
damage detection of polymer composites. sample was tested under 10 random vibration signals to
 15480569, 2025, 2, Downloaded from https://4spepublications.onlinelibrary.wiley.com/doi/10.1002/pc.29055 by The Director National Institut, Wiley Online Library on [16/05/2025]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
1542 AZAD and KIM

T A B L E 1 Nomenclature and
State of the composites Nomenclature No. of samples No. of experiments
description of health conditions of
Healthy H 5 10 composites.
Delamination-1 D1
Delamination-2 D2

F I G U R E 4 Fabrication of CFRP laminates: (A) orientation of the plies, (B) hot press compression molding machine used to prepare
samples, and (C) health states of the composite where delaminations are seeded in the midplane of the composite at different locations.

FIGURE 5 Experimental setup comprising various units for collecting vibrational data from laminated composites.

obtain an excessive dataset. The ply orientation, fabrica- sample on the top surface to record vibration signals. The
tion setup, and damage configuration are shown in excitation system was placed in front of the shaker, while
Figure 4. During fabrication, the healthy composite was the data acquisition system was placed after the acceler-
developed by pressing eight layers in the compression ometer. The excitation system consisted of an excitation
molding machine. In the case of delaminated composites, data acquisition system (E-DAS) that acquired data from a
delaminations were seeded in the midplane by inserting a Lab-view PC using MATLAB Simulink followed by an
PTFE Teflon film with a thickness of 0.3 mm. For the amplifier and shaker. The acquisition system consisted of
vibration testing, a cantilever beam configuration with one the accelerometer, followed by the amplifier, and another
fixed and one free end was adopted. During the tests, the data acquisition system to collect response data. The experi-
fixed end was clamped with the shaker, simulating two mental configuration and its components are shown in
delamination cases: (1) delamination near the clamped Figure 5. The proposed experimental setup was built to illus-
end, and (2) delamination near the free end, as shown in trate the variability of vibrational data when delamination
Figure 4C. An accelerometer was affixed to the composite damages of the same size are positioned at distinct locations.
 15480569, 2025, 2, Downloaded from https://4spepublications.onlinelibrary.wiley.com/doi/10.1002/pc.29055 by The Director National Institut, Wiley Online Library on [16/05/2025]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
AZAD and KIM 1543

Therefore, to increase the reliability and accuracy of the respectively. In this case, the scale and shift of the corre-
SHM systems, it is essential to comprehend how the sponding wavelet function are controlled in the spectral
location of delamination impacts vibrational signals to and temporal domains, respectively, by the scale and
prevent damage to the composite structure. Additionally, shift factors. For an arbitrary signal of finite length
to enhance the practicality and dependability of the x ðt Þ  L2 ðRÞ, the CWT can be obtained from the formula:
suggested methodology, 10 responses were obtained from
Z  
each specimen in each health state.   1 þ∞
tb
CWT x ða, bÞ ¼ x ðt Þ, φa,b ðt Þ ¼ pffiffiffi x ðtÞφ dt
a ∞ a
ð8Þ
3.2 | Data pre-processing

The data acquired from the experiments contains 10 ran- where,

 CWT  x ða, bÞ represents the inner producttbof
dom responses for each sample. This data is processed at x ðtÞ, φa,b ðt Þ , and φ tba denotes the conjugate φ a .
the pre-processing stage to improve the reliability of the Thus, Equation (8) converts temporal vibrational signals
damage detection model. First, the raw 1D signal is trans- into 2D time–frequency images. Choosing the right
formed into 2D images to improve the representation of mother wavelet is important for CWT analysis. In this
data features. Second, the size of the dataset is increased study, we used the Morlet function because it works well
by augmenting the data by adding zero mean Gaussian in dynamic systems and helps find sensitive discriminative
noise to enhance the generalization ability of the model. features in the data due to high-frequency resolution.46
A detailed explanation of the pre-processing steps is Morlet mother function, normalized by the constant C, is
described in the following sections. written as:

t2
φðt Þ ¼ Ce 2 cos 5x ð9Þ
3.2.1 | Signal-to-image conversion

Continuous wavelet transform (CWT) is an effective Figure 6 illustrates the process of converting temporal
method for representing time series signals into time– data into a wavelet-enhanced representation. Initially, a
frequency scalograms.44 Feature extraction is commonly sliding window of 1875 data points with a shift of 625 data
conducted in any temporal or spectral domain; however, points is selected from the signals of all health states.
CWT gives considerable discriminative features in both This process produces a dataset of 2980 signal windows
domains. Additionally, the use of CWT-based images for each health state. These signal windows are then pro-
improves diagnostic accuracy through automatic feature cessed by CWT to obtain scalogram images. The signal
extraction, eliminating the need for human feature engi- windows for all health states are mapped to a scalogram
neering and extraction.18,45 The mother wavelet function (time-frequency) image, resulting in 2980 images for each
φðtÞ belonging to the L2 ðRÞ space subjected to Fourier health state. The scalograms are then resized to 112  112
transformations meets the following criteria: dimensions, which are suitable for feature extraction and
Z classification models.
þ∞
b ðωÞj2
jφ
dω < ∞ ð5Þ
∞ ω
3.2.2 | Data augmentation
b ðωÞ denotes the Fourier transform of φðt Þ, and ω
where, φ
represents the frequency. Moreover, the set of wavelet A frequently observed issue in deep learning-based SHM
functions can be derived from the following fundamental methods is insufficient data size. As the deep learning
wavelet function: approaches increase in scale and complexity, a significant
dataset is required to obtain optimal performance from
 
1 tb these models. Several data augmentation techniques are
φa,b ðt Þ ¼ pffiffiffi φ a, b  R, a > 0 ð6Þ
a a available to increase the size of training data by produc-
ing artificial samples. In this study, data augmentation
t ¼ at 0 þ b ð7Þ was performed by adding zero-mean Gaussian noise.
Adding noise to the signals acts as regularization that
prevents overfitting, resulting in improved generalization
where, φa,b ðt Þ denotes the mathematical wavelet, while a capability and robustness of the deep learning model.47
and b represent the scale factor and shift factor, The Gaussian noise contains a probability density
 15480569, 2025, 2, Downloaded from https://4spepublications.onlinelibrary.wiley.com/doi/10.1002/pc.29055 by The Director National Institut, Wiley Online Library on [16/05/2025]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
1544 AZAD and KIM

F I G U R E 6 Illustration of extracting a window from the raw signal of multiple health states, converting it into scalogram images, and
augmenting the data by adding zero-mean Gaussian noise.

function (PDF) similar to the normal distribution, and to taken into account for training, two random responses
augment the dataset, a random Gaussian function GðtÞ that is, 20% of the scalograms (3588) are used for testing,
can be incorporated into the original data to introduce and the remaining single random response that is, 10% of
noise.48 For the vibration signal x ðt Þ, the PDF can be for- the scalograms (1776) is used to validate the ViT network.
mulated as: This partitioned dataset is used in the next sections to
develop the damage detection model for composites. The
1 ðxμÞ2
model was developed on a system equipped with an
Pðx Þ ¼ pffiffiffiffiffiffiffiffi e 2σ2 ð10Þ
2πσ AMD Ryzen 53,500X processor, Windows 10 operating
system, DDR4 RAM (32 GB), and NVIDIA GeForce GTX
where, σ is the standard deviation of the augmentation 1650 (4 GB) graphics card.
filter and μ is the zero mean. The augmented data is then
obtained by adding the noise data to the original data as:
4.1 | ViT model architecture and
~x ðt Þ ¼ x ðt Þ þ Gðt Þ ð11Þ training

Due to the equal amount of positive and negative This section describes the process of developing the ViT
noise, the zero-mean value of the noise does not contrib- model architecture and defining optimal hyperparameters
ute to the net disturbance and is ultimately eliminated for damage detection of composites. The hyperparameters
from the system. Moreover, a standard deviation, or noise of the ViT model include number of patches, dimensional-
level, of 10% was chosen during data augmentation as it ity of the model (DM), number of transformer heads (NTH),
has been beneficial in detecting damage in laminated dimension of the intermediate layer of the multilayer per-
composites.49 The data augmentation process is presented ceptron (MLP) model (DMLP), and number of transformer
in Figure 6. Another 2980 images were obtained for each blocks. Moreover, the number and nature of layers in the
class through the data augmentation process. classifier also act as a hyperparameter of the model. The
optimal architecture of the proposed ViT model obtained
after several iterations is presented in Figure 7 along with
4 | R ES U L T S A N D D I S C U S S I O N its hyperparameters. In the ViT model, the input image is
first divided into smaller patches with non-overlapping
After the pre-processing stage corresponding to the D1, regions. These patches allow the ViT model to process
D2, and H classes, a total of 17,880 scalograms are images in a more scalable and computationally efficient
obtained. To develop the proposed ViT-based explainable manner, especially for large images.39 In our study, a 4  4
damage detection model, the entire dataset is split into grid size consisting of 16 square patches of pixel size
training, testing, and validation data. The first seven ran- (28  28) was found to be optimal, providing effective
dom responses that is, 70% of the scalograms (12,516) are discriminative features in the time-frequency domain.
 15480569, 2025, 2, Downloaded from https://4spepublications.onlinelibrary.wiley.com/doi/10.1002/pc.29055 by The Director National Institut, Wiley Online Library on [16/05/2025]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
AZAD and KIM 1545

F I G U R E 7 Detailed ViT model architecture with optimal hyperparameters and data dimensions for each stage of the damage
detection process.

Increasing the number of patches may provide enhanced network to capture non-linear relationships in the data
features but will increase computational costs, thus further while maintaining the computational efficiency of the
increasing the number of patches was not considered ben- model. Two such transformer blocks are added sequen-
eficial. The patches are then fed into a positional encoding tially to extract high-quality discriminative features auton-
block where the position of each patch is preserved. Addi- omously. The output of the transformer blocks contains
tional class embedding is also introduced here to learn the the features from the input data with dimensions of
class of the data. Both positional and class embedding are 16  784  8. These feature vectors are then processed by
flattened and then mapped to the DM dimension. In this the classifier to perform damage diagnosis. The classifier
study, DM = 8 was found to be optimal for the embedding block contains normalization, flattening, dropout, and sev-
space. Thus, instead of 3 RGB channels, an 8-dimensional eral dense layers. The dimensions of the respective layers
vector was found to effectively represent each patch, are shown in Figure 7. Herein, a dropout layer with a 10%
resulting in (16  784  8) as the model representation dropout was utilized to avoid overfitting the classifier.
space. The encoded feature space is used as input to the Moreover, the dimension of the final dense layer only
transformer block, where it is normalized before being fed (16  3) demonstrates that an exceptionally small final
into the multi-head attention. NTH = 2 is used in the feature vector is required. Such a small dimension of the
multi-head attention block to compute attention weights final layer indicates that the ViT model can extract high-
across multiple subspaces of the input. This is followed by quality features without the need for excessive computa-
another normalization layer and an MLP block that pro- tional resources.
cesses the output of the attention mechanism. The MLP Finally, the model was trained for 40 epochs using
block consists of a feed-forward neural network with the Adam optimizer at a learning rate of 5  105. To
DMLP = 16, controlling the complexity and capacity of the introduce non-linearity in the models, ReLU activation is
 15480569, 2025, 2, Downloaded from https://4spepublications.onlinelibrary.wiley.com/doi/10.1002/pc.29055 by The Director National Institut, Wiley Online Library on [16/05/2025]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
1546 AZAD and KIM

FIGURE 8 Accuracy and loss curves of the ViT model using training and testing datasets.

used in all dense layers except the last dense layer (classi- performs exceptionally well not only in identifying the
fication later) which uses SoftMax activation. Moreover, healthy state but also in differentiating the same type of
due to the multi-class problem, the categorical cross- D1 and D2 delaminations present at different locations. The
entropy loss function is utilized during model compila- ROC curve also shows that the precision (true positive rate)
tion. The progression of the accuracy and loss curves with which is the ratio of accurately predicted positive observa-
respect to the number of epochs of training and testing tions to all original positives is high (1) and the false
data is shown in Figure 8. Both the accuracy and loss positive rate which is the ratio of inaccurately predicted
curves showed that the ViT model manages to learn positive observations to all original negatives in the data is
excessive information in the initial epochs. After that, the low (0). This indicates that the ViT model achieves almost
model starts to converge after 25 epochs, and the curves perfect sensitivity (all positives correctly identified) with no
tend to flatten out. The trained ViT model demonstrated false positives. According to the ROC curve analysis, the
training and testing accuracy of 99.30% and 97.77%, AUC was determined to be 99.95%, 99.89%, and 99.98% for
respectively. Similarly, a loss of 2.06  102 is observed D1, D2, and H, respectively. Thus, the proposed approach
for train data, and 6.38  102 for test data. This trained showed improved performance for detecting damage in
model is saved and validated on the unseen validation the polymer composites as compared to the existing state-
dataset in the next section. of-the-art models, as shown in Table 2. Moreover,
t-distributed stochastic neighbor embedding (t  SNE) is
applied to the unseen validation data to visualize the fea-
4.2 | Validation of the ViT model ture extracted by the ViT model and compared with the
original data, as shown in Figure 10. It can be observed
During training, all model weights were assigned to the that the original data is significantly scattered and cannot
ViT model, and these weights were updated during test- be separated, especially in the case of D1 and D2. However,
ing. However, during validation, all model weights the features extracted from the ViT model form good clus-
are fixed, and the saved model is used to predict unseen ters and show clear separation not only between healthy
validation data. Moreover, instead of using only accuracy and damaged cases but also in the identification of D1 and
as the evaluation matrix, a comprehensive evaluation is D2. These results therefore confirm that the proposed ViT
performed through additional matrices such as confusion model provides better damage detection performance for
matrix, precision, recall, F1-score, receiver operating polymer composite structures.
characteristic (ROC) curve, and area under the curve
(AUC). The confusion matrix with precision (P), recall
(R), and f1-scores (F1) for each class predicted by the ViT 4.3 | Explainable vision transformer
model is shown in Figure 9. The results demonstrate that (X-ViT)
all classes are well predicted by the ViT model. The pre-
diction accuracy of D1, D2, and H is 98.82%, 98.48%, and In this study, a novel patch attention technique was
99.16%, respectively. Similarly, other evaluation metrics proposed to facilitate the explanation of the ViT model.
such as P, R, and F1 were also found to be higher than Generally, attention maps are used to interpret deep
98%. Moreover, it should be noted that the ViT model learning models and provide a non-linear explanation for
 15480569, 2025, 2, Downloaded from https://4spepublications.onlinelibrary.wiley.com/doi/10.1002/pc.29055 by The Director National Institut, Wiley Online Library on [16/05/2025]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
AZAD and KIM 1547

F I G U R E 9 Evaluation of ViT model predictions based on unseen validation data: (A) confusion matrix, precision, recall, and f1-score,
and (B) ROC curve for D1, (C) ROC curve for D2, and (D) ROC curve for H.

T A B L E 2 The performance
Model Accuracy Precision Recall F1-score Explainability
comparison of the proposed approach
42
with existing state-of-the-art models for SVM 85.00 84.87 84.46 84.50
42
damage detection of polymer XGBoost 84.67 84.50 84.42 84.34
composites. 42
CNN 95.44 95.57 95.49 95.40
50
VGG-16 93.66 93.67 93.67 94.00
50
VGG-19 91.33 91.33 91.33 90.33
50
Xception 84.66 85.00 84.67 84.33
50
ResNet 96.67 96.67 96.00 96.33
Proposed X-ViT 98.82 98.82 98.82 98.82

F I G U R E 1 0 Visual comparison of model features using t-SNE for (A) the original validation dataset, and (B) features of validation data
from the last layer of the ViT model.
 15480569, 2025, 2, Downloaded from https://4spepublications.onlinelibrary.wiley.com/doi/10.1002/pc.29055 by The Director National Institut, Wiley Online Library on [16/05/2025]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
1548 AZAD and KIM

F I G U R E 1 1 End-user interpretation of the X-ViT model (A) normalized patch attentions, and (B) patch attentions overlap each health
state of the composites.

predictions.31,51 This implies that attention maps provide states, which means that most of the composite possesses
localized attention that may not effectively capture global healthy status. None of the other patches contributed suffi-
dependencies, focusing more on local regions in a non- ciently to the healthy state of H. However, in the case of D1
linear manner. Moreover, while attention maps provide and D2, several other patches are receiving high attention.
insight into the model's decision-making process, inter- In the case of D1 patches 4,5, and 7 also contribute 0.07,
preting non-linear ratings can be challenging for the end 0.06, and 0.06, respectively, to the D1 delamination classifi-
user, especially for complex models with multiple layers. cation. While patches 5–8 and 12 contribute to the D2 clas-
Therefore, this study aims to provide a linear and easy- sification. Moreover, the combined eight patches in the
to-understand interpretation based on patch attention middle contribute significantly to the damage cases (0.42
information. Since patching is an essential part of the for D1 and 0.52 for D2), while only 0.01 for H, indicating
ViT model, patch attention identifies patches or regions that the middle region of the scalogram images is impor-
that contain significant characteristics of a specific health tant for delamination detection. Similarly, the bottom fig-
state. In our study, patch attention is calculated using the ures highlight the importance of these patches in a way
weights of each layer of the ViT model. Generally, that the end user can understand while making decisions.
the attention mappings in each layer of a multi-layer ViT The increased transparency shows the area of interest for
model are learned individually, adding a significant num- each class. High transparency can be observed in the first
ber of parameters and limiting the generalization of the 4 patches of the healthy state, indicating that only the top
model. However, in the proposed X-ViT model, patch region is sufficient to predict H. However, the transparency
attentions from various layers are connected, and the sub- of the 8 patches in the middle increases in the scalograms
sequent layer directly uses the information from earlier with delamination indicating that the middle patches con-
layers to create a more beneficial dependency structure. tain the knowledge of delaminations. The bottom 4 patches
Next, this knowledge is aggregated across the entire dataset are highly opaque for all classes, which indicates that they
and significant patches for each health state are identified do not contribute to the prediction of any of the health
establishing the X-ViT model. The normalized patch atten- states D1, D2, or H. Thus, these visual interpretations are
tion and its overlap with each health state of the compos- beneficial in increasing the transparency and reliability of
ites are shown in Figure 11. This indicates that the patch the deep learning models, making it easier for the end user
attention is high for the first four patches in all health to make a decision.
 15480569, 2025, 2, Downloaded from https://4spepublications.onlinelibrary.wiley.com/doi/10.1002/pc.29055 by The Director National Institut, Wiley Online Library on [16/05/2025]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
AZAD and KIM 1549

5 | C ON C L U S I ON ORCID
Muhammad Muzammil Azad https://orcid.org/0000-
In this study, significant progress has been made in the 0002-1216-979X
SHM of laminated composites by integrating explainable
AI. Unlike existing black box AI models, this study intro- RE FER EN CES
duces an X-ViT model that addresses the critical need for 1. Sharma A, Mukhopadhyay T, Rangappa SM, Siengchin S,
reliability, interpretability, and transparency of AI models Kushvaha V. Advances in computational intelligence of poly-
deployed for SHM of composite structures. To achieve this mer composite materials: machine learning assisted modeling,
objective, first, the ViT model was developed using a trans- analysis and design. Arch Comput Methods Eng. 2022;29(5):
3341-3385. doi:10.1007/s11831-021-09700-9
former architecture utilizing attention mechanism and
2. Azad MM, Shah AU, Prabhakar MN, Kim HS. Deep learning-
positional encoding. The ViT model is then optimized by based fracture mode determination in composite laminates.
choosing hyperparameters that provide optimal damage Journal of the Computational Structural Engineering Institute of
detection results for laminated composites and validated Korea. 2024;37(4):225-232. doi:10.7734/COSEIK.2024.37.4.225
on experimental data comprising various health states, 3. Gayathri P, Umesh K, Ganguli R. Effect of matrix cracking and
including delamination damage cases. Moreover, the per- material uncertainty on composite plates. Reliab Eng Syst Saf.
formance and generalization ability of the model were 2010;95(7):716-728. doi:10.1016/J.RESS.2010.02.004
improved by transforming the raw vibration signals into 4. Khan A, Azad MM, Sohail M, Kim HS. A review of physics-based
models in prognostics and health management of laminated com-
images using CWT and adding Gaussian noise to augment
posite structures. Int J Precis Eng Manuf-Green Technol. 2023;
the data. The results indicated that the proposed ViT model 10(6):1615-1635. doi:10.1007/s40684-023-00509-4
performs well compared to existing state of the art methods 5. Khan A, Kim HS. A brief overview of delamination localization
in terms of numerous evaluation metrics including accu- in laminated composites. Multiscale Sci Eng. 2022;4(3):102-110.
racy, precision, recall, and f1-score. Overall, the proposed doi:10.1007/s42493-022-00085-w
model showed an accuracy of 98.82%, 98.48%, and 99.16%, 6. Giannakeas IN, Mazaheri F, Bacarreza O, Khodaei ZS,
for D1, D2, and H, respectively, using the unseen validation Aliabadi FMH. Probabilistic residual strength assessment of
smart composite aircraft panels using guided waves. Reliab Eng
dataset. The developed ViT-based deep learning model can
Syst Saf. 2023;237:109338. doi:10.1016/j.ress.2023.109338
then be explained by analyzing the patch attention in each
7. Yang F, Wu D, Feng H. A new probe-based electromagnetic
layer of the model. By aggregating patch attention and non-destructive testing method for carbon fiber-reinforced
identifying key patches influencing predictions, the X-ViT polymers utilizing eddy current loss measurements. Polym
model offers valuable insight into the decision-making pro- Compos. 2023;44(10):6661-6675. doi:10.1002/pc.27587
cess, enhancing trust and reliability among end users. 8. Khan A, Ko DK, Lim SC, Kim HS. Structural vibration-based
Therefore, the interpretable X-ViT model signifies a major classification and prediction of delamination in smart compos-
contribution to the field by offering a robust and transpar- ite laminates using deep learning neural network. Compos Part
B Eng. 2019;161:586-594. doi:10.1016/j.compositesb.2018.12.118
ent solution for damage detection in composite laminates.
9. Khan A, Kim H. Active vibration control of a piezo-bonded
However, the current research focuses on a single dataset laminated composite in the presence of sensor partial debond-
covering three health states, future work will focus on ing and structural Delaminations. Sensors. 2019;19(3):540. doi:
developing another dataset with complex damage scenar- 10.3390/s19030540
ios, to not only detect damage but also perform explainable 10. Sohn JW, Choi SB, Kim HS. Vibration control of smart hull
damage localization for laminated composite structures. structure with optimally placed piezoelectric composite actua-
tors. Int J Mech Sci. 2011;53(8):647-659. doi:10.1016/j.ijmecsci.
A U T H O R C ON T R I B U T I O NS 2011.05.011
11. Yang JS, Liu ZD, Schmidt R, Schröder KU, Ma L, Wu LZ. Vibra-
Muhammad Muzammil Azad: Conceptualization, Data
tion-based damage diagnosis of composite sandwich panels with
curation, Methodology, Validation, Writing—original bi-directional corrugated lattice cores. Compos Part A Appl Sci
draft. Heung Soo Kim: Supervision, Writing—review & Manuf. 2020;131:105781. doi:10.1016/j.compositesa.2020.105781
editing, Funding acquisition. 12. Essedik LM, Rachid T, Madjid E, et al. Active vibration control
of piezoelectric multilayers FG-CNTRC and FG-GPLRC plates.
A C K N O WL E D G M E N T S Polym Compos. 2024;11:9573-9587. doi:10.1002/pc.28428
This work was supported by a National Research 13. Azad MM, Kim S, Bin CY, Kim HS. Intelligent structural health
monitoring of composite structures using machine learning,
Foundation of Korea (NRF) grant, funded by the Korea
deep learning, and transfer learning: a review. Adv Compos
government (MSIT) (No. 2020R1A2C1006613).
Mater. 2024;33(2):162-188. doi:10.1080/09243046.2023.2215474
14. Jakkamputi L, Devaraj S, Marikkannan S, et al. Experimental
DATA AVAILABILITY STATEMENT and computational vibration analysis for diagnosing the
The data that support the findings of this study are avail- defects in high performance composite structures using
able from the corresponding author upon reasonable machine learning approach. Appl Sci. 2022;12(23):12100. doi:
request. 10.3390/app122312100
 15480569, 2025, 2, Downloaded from https://4spepublications.onlinelibrary.wiley.com/doi/10.1002/pc.29055 by The Director National Institut, Wiley Online Library on [16/05/2025]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
1550 AZAD and KIM

15. Das S, Chattopadhyay A, Srivastava AN. Classifying induced 31. Noh YR, Khalid S, Kim HS, Choi SK. Intelligent fault
damage in composite plates using one-class support vector diagnosis of robotic strain wave gear reducer using area-
machines. AIAA J. 2010;48(4):705-718. doi:10.2514/1.37282 metric-based sampling. Mathematics. 2023;11(19):4081. doi:
16. Tang L, Li Y, Bao Q, et al. Quantitative identification of dam- 10.3390/math11194081
age in composite structures using sparse sensor arrays and 32. Yagin FH, Cicek IB,_ Alkhateeb A, et al. Explainable artificial
multi-domain-feature fusion of guided waves. Measurement. intelligence model for identifying COVID-19 gene biomarkers.
2023;208:112482. doi:10.1016/j.measurement.2023.112482 Comput Biol Med. 2023;154:106619. doi:10.1016/j.compbiomed.
17. Qiao S, Huang M, Liang Y, Zhang S, Zhou W. Damage mode 2023.106619
identification in carbon/epoxy composite via machine learning 33. Liu H, Wang Y, Fan W, et al. Trustworthy AI: a computational
and acoustic emission. Polym Compos. 2023;44(4):2427-2440. perspective. ACM Trans Intell Syst Technol. 2023;14(1):1-59.
doi:10.1002/pc.27254 doi:10.1145/3546872
18. Rautela M, Senthilnath J, Monaco E, Gopalakrishnan S. Delami- 34. Holzinger A. The Next Frontier: AI We Can Really Trust. 2021:
nation prediction in composite panels using unsupervised-feature 427–440. doi:10.1007/978-3-030-93736-2_33
learning methods with wavelet-enhanced guided wave representa- 35. Manzari ON, Ahmadabadi H, Kashiani H, Shokouhi SB,
tions. Compos Struct. 2022;291:115579. doi:10.1016/j.compstruct. Ayatollahi A. MedViT: a robust vision transformer for general-
2022.115579 ized medical image classification. Comput Biol Med. 2023;157:
19. Ullah S, Ijjeh AA, Kudela P. Deep learning approach for delami- 106791. doi:10.1016/j.compbiomed.2023.106791
nation identification using animation of lamb waves. Eng Appl 36. Liang P, Yu Z, Wang B, Xu X, Tian J. Fault transfer diagnosis
Artif Intell. 2023;117:105520. doi:10.1016/j.engappai.2022.105520 of rolling bearings across multiple working conditions via sub-
20. Wang MH, Lu SD, Hsieh CC, Hung CC. Fault detection of domain adaptation and improved vision transformer network.
wind turbine blades using Multi-Channel CNN. Sustainability. Adv Eng Inform. 2023;57:102075. doi:10.1016/j.aei.2023.102075
2022;14(3):1781. doi:10.3390/su14031781 37. Wu H, Triebe MJ, Sutherland JW. A transformer-based approach
21. Rautela M, Gopalakrishnan S. Ultrasonic guided wave based for novel fault detection and fault classification/diagnosis in
structural damage detection and localization using model manufacturing: a rotary system application. J Manuf Syst. 2023;
assisted convolutional and recurrent neural networks. Expert 67:439-452. doi:10.1016/j.jmsy.2023.02.018
Syst Appl. 2021;167:114189. doi:10.1016/j.eswa.2020.114189 38. Vaswani A, Shazeer N, Parmar N, et al. Attention is all you
22. Yoon S, (Song-Kyoo) Kim A, Cantwell WJ, et al. Defect detec- need. Advances in Neural Information Processing Systems.
tion in composites by deep learning using solitary waves. Int J NeurIPS Proceedings; 2017.
Mech Sci. 2023;239:107882. doi:10.1016/j.ijmecsci.2022.107882 39. Dosovitskiy A, Beyer L, Kolesnikov A, et al. Jakob Uszkoreit
23. Modarres C, Astorga N, Droguett EL, Meruane V. Convolu- NH. An image is worth 16  16 words: transformers for image
tional neural networks for automated damage recognition and recognition at scale. International Conference on Learning Rep-
damage type identification. Struct Control Health Monit. 2018; resentations; 2021. 10.48550/arXiv.2010.11929
25(10):e2230. doi:10.1002/stc.2230 40. Ding Y, Jia M, Miao Q, Cao Y. A novel time–frequency trans-
24. Lee H, Lim HJ, Skinner T, Chattopadhyay A, Hall A. Auto- former based on self–attention mechanism and its application
mated fatigue damage detection and classification technique in fault diagnosis of rolling bearings. Mech Syst Signal Process.
for composite structures using lamb waves and deep autoenco- 2022;168:108616. doi:10.1016/j.ymssp.2021.108616
der. Mech Syst Signal Process. 2022;163:108148. doi:10.1016/j. 41. Alshahrani H, Sharma G, Anand V, et al. An intelligent
ymssp.2021.108148 attention-based transfer learning model for accurate differenti-
25. Park S, Song J, Kim HS, Ryu D. Non-contact detection of ation of bone marrow stains to diagnose hematological disor-
delamination in composite laminates coated with a mechano- der. Life. 2023;13(10):2091. doi:10.3390/life13102091
luminescent sensor using convolutional AutoEncoder. Mathe- 42. Azad MM, Kim HS. Hybrid deep convolutional networks for the
matics. 2022;10(22):4254. doi:10.3390/math10224254 autonomous damage diagnosis of laminated composite struc-
26. Rudin C, Radin J. Why are we using black box models in AI when tures. Compos Struct. 2024;329:117792. doi:10.1016/j.compstruct.
we Don't need to? A lesson from an explainable AI competition. 2023.117792
Harvard Data Sci Rev. 2019;1(2):1-9. doi:10.1162/99608f92.5a8a3a3d 43. Ali S, Abuhmed T, El-Sappagh S, et al. Explainable artificial
27. Naser MZ. An engineer's guide to eXplainable artificial intelli- intelligence (XAI): what we know and what is left to attain
gence and interpretable machine learning: navigating causality, trustworthy artificial intelligence. Inf Fusion. 2023;99:101805.
forced goodness, and the false perception of inference. Autom doi:10.1016/j.inffus.2023.101805
Constr. 2021;129:103821. doi:10.1016/j.autcon.2021.103821 44. Zhou J, Li Z, Chen J. Damage identification method based on
28. Bogue R. What are the prospects for robots in the construction continuous wavelet transform and mode shapes for composite
industry? Ind Robot an Int J. 2018;45(1):1-6. doi:10.1108/IR-11- laminates with cutouts. Compos Struct. 2018;191:12-23. doi:10.
2017-0194 1016/j.compstruct.2018.02.028
29. Guilleme M, Masson V, Roze L. Termier A. Agnostic Local 45. Sun S, Zhang T, Li Q, et al. Fault diagnosis of conventional cir-
Explanation for Time Series Classification. In: 2019 IEEE 31st cuit breaker contact system based on time–frequency analysis
International Conference on Tools with Artificial Intelligence and improved AlexNet. IEEE Trans Instrum Meas. 2021;70:1-
(ICTAI). IEEE. 2019:432–439. doi:10.1109/ICTAI.2019.00067 12. doi:10.1109/TIM.2020.3045798
30. Kumar P, Hati AS. Deep convolutional neural network based 46. Liu Y, Li Z, Zhang W. Crack detection of fibre reinforced compos-
on adaptive gradient optimizer for fault detection in SCIM. ISA ite beams based on continuous wavelet transform. Nondestruct
Trans. 2021;111:350-359. doi:10.1016/j.isatra.2020.10.052 Test Eval. 2010;25(1):25-44. doi:10.1080/10589750902744992
 15480569, 2025, 2, Downloaded from https://4spepublications.onlinelibrary.wiley.com/doi/10.1002/pc.29055 by The Director National Institut, Wiley Online Library on [16/05/2025]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
AZAD and KIM 1551

47. Zhu S, Yu C, Hu J. Regularizing deep neural networks for experimental data. J Mater Res Technol. 2024;29:3024-3035.
medical image analysis with augmented batch normaliza- doi:10.1016/j.jmrt.2024.02.067
tion. Appl Soft Comput. 2024;154:111337. doi:10.1016/j.asoc. 51. Innat M, Hossain MF, Mader K, Kouzani AZ. A convolutional
2024.111337 attention mapping deep neural network for classification and
48. Soltanieh S, Etemad A, Hashemi J. Analysis of Augmenta- localization of cardiomegaly on chest X-rays. Sci Rep. 2023;
tions for Contrastive ECG Representation Learning. In: 13(1):6247. doi:10.1038/s41598-023-32611-7
2022 International Joint Conference on Neural Networks
(IJCNN). IEEE. 2022:1–10. doi:10.1109/IJCNN55064.2022.
9892600
49. Sikdar S, Ostachowicz W, Kundu A. Deep learning for How to cite this article: Azad MM, Kim HS. An
automatic assessment of breathing-debonds in stiffened
explainable artificial intelligence-based approach
composite panels using non-linear guided wave signals.
Compos Struct. 2023;312:116876. doi:10.1016/j.compstruct.
for reliable damage detection in polymer composite
2023.116876 structures using deep learning. Polym Compos.
50. Azad MM, Kumar P, Kim HS. Delamination detection in 2025;46(2):1536‐1551. doi:10.1002/pc.29055
CFRP laminates using deep transfer learning with limited