Applied Composite Materials

https://doi.org/10.1007/s10443-023-10161-5

On the Use of Machine Learning for Damage Assessment

in Composite Structures: A Review

Ronny Francis Ribeiro Junior1 · Guilherme Ferreira Gomes1

Received: 28 June 2023 / Accepted: 21 August 2023

© The Author(s), under exclusive licence to Springer Nature B.V. 2023

Abstract
Composite materials are those formed by combining two or more different materials to
take advantage of the best characteristics of each one. However, due to this heterogene-
ity, composite materials suffer from non-linear failure modes. Because of this complexity,
damage to composite structures cannot be identified by visual inspection or conventional
techniques. Therefore, several complex techniques are employed in this type of material,
with Machine Learning being the main way of dealing with the excessive data extracted
from these techniques. Given the rapidly increasing use of composite materials in real-
world applications, the demand for damage assessment (detection, quantification, and
localization) methods is increasingly high. This article reviews the main and most recent
works on ML methods for the damage assessment of composite structures. The selected
studies are then covered in detail to provide researchers with an in-depth comprehension
of what is new in ML algorithms for the damage assessment of composite structures. From
2019 to now, there has been a large increase in the number of publications related to dam-
age assessments of composite materials, with a strong predominance of Artificial Neural
Networks, Convolutional Neural Networks, and Principal Component Analysis techniques.
However, there is still a lack of studies with real cases in real environments. Finally, future
research directions and a summary of all selected works are suggested, presenting possible
improvements of the state of the art.

Keywords Machine learning · Composite structures · Smart composite · Damage assessment ·

Structural health monitoring

Acronyms
AE Acoustic Emission
ANN Artificial Neural Network
AR Models Autoregressive Models
ARIMA Autoregressive Integrated Moving Average
CNN Convolutional Neural Network
CI Cornwell Indication
CT Computed Tomography

* Ronny Francis Ribeiro Junior

[email protected]
1
Mechanical Engineering Institute, Federal University of Itajubá - UNIFEI, Itajubá, Brazil

13
Vol.:(0123456789)
 Applied Composite Materials

CLC Conventional Load-Carrying

DIC Digital Image Correlation
EI Electromechanical Impedance
FBG Fiber Bragg Grating
FGM Functionally Graded Material
FEM Finitite Element Analysis
GRFP Glass Fiber Reinforced Polymer
IGA Isogeometric Analysis
IDT Impact Damage Tolerance
IC Incremental Clustering
KNN K-nearest Neighbors
LFRR Local Frequency Response Ratio
ML Machine Learning
MLP Multilayer Perceptron
NFC Natural Fiber-reinforced Composites
NDT Non-Destructive Testing
PCA Principal Component Analysis
PZT Piezoelectric Sensor
RAPID Reconstruction Algorithm for Probabilistic Inspection Damage
RLCP Rectangular Laminated Composite Plates
PSO Particle Swarm Optimization
SVM Support Vector Machine
SEM Scanning Electron Microscopy
TDCNN Temporal Derivative Cellular Neural Network
UHMWPP Ultra-high Molecular Weight Polypropylene
WGFRP Woven Glass Fiber Reinforced Polymer

1 Introduction

Composite materials are produced by blending two or more materials that possess distinct
physical or chemical properties, without any chemical reaction or solubility, to create new
materials that exhibit superior qualities of their individual components. The idea of com-
posite materials was inspired by natural examples like wood, composed of cellulose and
lignin fibers. Currently, composite materials are being increasingly adopted in various sec-
tors like aerospace, defense, marine, civil engineering, and automotive, gradually replacing
traditional metallic materials [1–5].
Composite structures are extensively used due to their superior properties such as
increased stiffness, strength, corrosion resistance, fatigue life, and wear resistance, as well
as improved thermal properties and reduced weight [6, 7]. However, composite structures
are susceptible to different forms of damage, such as delamination, fiber breakage, matrix
cracking, core detachment, porosity, and fiber misalignment [8–11], which can occur both
in the manufacturing process of the composite or in service [12]. In addition, the damage
is difficult to detect with a visual inspection, and most of the time it occurs abruptly. The
timely diagnosis of damage in composite structures is critical to avoiding unexpected dam-
age and structural failure.

13
Applied Composite Materials

The ability to identify, measure and pinpoint damage is crucial for the safe and depend-
able use of composite structures in practical situations. As a result, several techniques that use
ultrasound [13], x-rays [14], vibration [15], acoustic emission [16], modal analysis [17], opti-
cal methods [18], guided waves [19], and others [20–23] have been used for this purpose over
time, and it is quite common to apply more than one method to monitor composite structures.
In this particular situation, machine learning methods are attracting the attention of research-
ers due to their ability to quickly and effectively analyze diagnostic data based on certain param-
eters. Machine learning is a scientific discipline that focuses on creating mathematical algorithms
that can establish useful connections between variables using collected data [24].
This manuscript focuses on the use of machine learning techniques applied to damage
detection in composite structures. This paper discusses several recently published studies,
selected following a criterion of quality and relevance to the scientific community. There
are various machine learning approaches that may be applied, and one learning strategy
may correctly characterize one type of damage while failing to characterize another [25].
This manuscript will concentrate on the main ones that are: artificial neural networks,
autoregressive models, Bayesian Classifiers, convolution neural networks, K-means and
K-nearest neighbors, Principal Component Analysis, and Support Vector Machine. As the
article will focus on the mentioned techniques, new techniques that do not fit into one of
the above will be ignored.
The study also plays an important role in identifying future research directions and indi-
cating possible improvements and enhancements to existing techniques. As technology
and data science advance, the use of machine learning in damage detection in composite
structures may become even more accurate and reliable. Every year, new techniques and
methods may emerge to solve several past problems. By highlighting the key aspects of
the reviewed research, the study provides valuable insights for scholars and professionals
interested in the field of composite structures and machine learning, fostering the ongoing
advancement of knowledge in this area so that the reader can select the appropriate combi-
nation of features and ML techniques for the problem of interest.
The paper is organized as follows: Section 2 presents the basic concepts of the main
machine learning methods. In Section 3 the methodological procedure used in the review is
presented. The results of this review are systematically classified and discussed in Section 4.
Section 5 provides research directions for the future. Finally, Section 6 concludes the study
by highlighting the main aspects of the reviewed studies.

2 Fundamental Concepts

This section discusses the fundamental ideas of machine learning and provides an over-
view of the main learning algorithms used in composite structure damage detection.

2.1 Machine Learning

Machine learning refers to an algorithm that creates a model by using a set of sample
data, known as training data, to make predictions or decisions without requiring explicit
programming [26]. There are two main types of machine learning methods: super-
vised and unsupervised, which differ based on the use of labeled datasets. Essentially,

13
 Applied Composite Materials

supervised learning algorithms use labeled input and output data, while unsupervised
learning algorithms do not rely on labeled data. Within the category of unsupervised
algorithms, there is a specific subgroup known as clustering, which is utilized in multi-
variate data sets to identify related groups of elements. Meanwhile, supervised methods
are broadly categorized into two types: classification and regression approaches. Clas-
sification methods are used to predict class labels, while regression approaches are used
to determine the most appropriate model for the given data. Figure 1 provides a visual
representation of the divisions and subdivisions of machine learning algorithms.
A new type of learning, known as reinforcement learning, has been developed for par-
ticular tasks. This approach involves the algorithm interacting with a changing environment
to accomplish a specific objective, such as playing a game [27]. The objective of this form of
learning is to increase the number of positive outcomes as the algorithm “plays the game".
However, it is a very recent type of learning that is still in development, so, there are still not
enough cases applied to composite structures to make a review.
Two methods are being used to enable intelligent structural damage diagnosis. The first
method involves utilizing modern sensor technology in combination with numerical simu-
lation approaches to collect monitoring data that can help determine the state of structural
deterioration. The second method involves using machine-learning algorithms to extract con-
cealed features from the monitoring data for intelligent structural damage diagnosis [25, 28].
In the upcoming sections, we will present a general introduction to the key machine-
learning techniques utilized to assess damage to composite structures. As the focus of
this paper is to provide a practical discussion of the application of the method, we will
not delve into the mathematical details of the algorithms.

Fig. 1  Divisions and subdivisions of machine learning algorithms

13
Applied Composite Materials

2.2 Artificial Neural Network

Artificial neural networks are computer models that mimic the activity of the human brain.
They use basic processing units called neurons, which work like biological neurons, to per-
form calculations. These neurons are connected by weights, and during the training process,
these weights are adjusted based on the data supplied to the network [29]. Figure 2 depicts a
simple neuron model.
The way an Artificial Neural Network behaves is influenced by both the arrangement of
connections (topology) between its neurons and the strength of those connections (synaptic
values). The parameters that determine the entire functioning of the ANN are [30]:

• Number of neuron layers

• Number of neurons per layer
• Types of connections between neurons
• Degree of connectivity between neurons

2.3 Autoregressive Models

Autoregressive models are statistical models that aim to predict the value of a variable of inter-
est. These models are commonly used when there is a correlation between past and future data
in a series [31]. AR models are essentially linear regressions of data, where the future value is
predicted based on one or more previous data points within the same series. However, due to
the stochastic nature of AR models, there is a level of uncertainty and unpredictability associ-
ated with them, meaning that every forecast should be considered a trend with some level of
confidence attached to it. Therefore, it is impossible to make predictions with 100% certainty.
AR model is typically denoted by AR(p), where p is the order of the model. The order of
an AR model refers to the number of past values in the series that are used to predict its future
values. For instance, an AR(1) model is a first-order model that makes use of only one previ-
ous measurement to forecast future values. The AR(p) model is defined by the Eq. 1 [32]:
p
∑
Xt = 𝜑i Xt−1 + 𝜀t (1)
i=1

where: 𝜑i are the AR Coefficients and 𝜀t is a Gaussian white-noise.

Fig. 2  Artificial neuron representation (adapted from [30])

13
 Applied Composite Materials

There are currently various autoregressive models that integrate different approaches to
improve prediction performance. One such model is ARIMA, which is particularly note-
worthy for its effectiveness [33].

2.4 Bayesian Classifier

A Bayesian classifier is a type of model that uses probability to predict classes within
a given set of data. The model also enables the calculation of probabilities for both the
classes and their associated variables. When there is no specific knowledge about the data,
an approach commonly used is to assume “uninformed priors", which assigns equal prob-
abilities to each class and attribute value [34].
The Bayesian classifier is a fundamental algorithm for classifying an instance by deter-
mining the probability of it being assigned to a specific class, such as Q1. This involves the
computation of probabilities according to Eq. 2 [35]:
P(Ci |A1 = V1j & ⋅ ⋅ ⋅ &AN = VNj ) (2)

where an example is represented by attribute-value pairs of the form Ai = Vi .

If there are N independent attributes, then the probability is equal to Eq. 3 [35]:
∏
P(Ci ) P(Ak = Vkj |Ci ) (3)
k

When this independence assumption is made, the classifier is called Naive Bayes.
Naive Bayesian classifiers have a significant advantage over probabilistic classification
algorithms due to their straightforward and commonly used approach for performing clas-
sification tasks. This advantage is derived from their ability to calculate class and condi-
tional probabilities, making them robust to noise, and the statistical foundation that enables
them to scale effectively across domains with many irrelevant features [36].
The naive Bayesian classifier relies on two important assumptions. Firstly, it assumes
that each class can be represented by a single probability distribution, and secondly, it
assumes that the probability distributions of the features within each class are independent
of each other. However, these assumptions may limit the model and result in reduced clas-
sifier accuracy in certain scenarios [36, 37].

2.5 Convolutional Neural Network

A method of learning called convolutional neural network, which was first proposed by
LeCun et al. [38], uses a feed-forward neural network to extract data by employing a con-
volution structure [39]. Its primary function is to classify input images into various catego-
ries. A CNN is composed of two main components: one for selecting features and another
for classification. The feature selection part consists of an input layer, convolutional layer,
and pooling layer, while the classification part comprises a fully connected layer and an
output layer. Figure 3 shows an example of a CNN architecture.
Recently, a specialized variant of CNN called the 1D CNN has emerged, which utilizes
a signal or 1D data as input instead of an image [41]. Despite this difference, the architec-
ture of 1D and 2D CNNs are highly similar, except for the filter sliding mechanism. In the
case of 1D CNN, the filter moves vertically (along the height dimension) to extract fea-
tures, and the height determines the number of sample points involved in the convolutional

13
Applied Composite Materials

Fig. 3  A basic architecture of Convolutional Neural Network (adapted from [40])

process. In contrast, the 2D CNN filter slides both horizontally and vertically (height and
width) across the entire matrix.

2.6 K‑Means and K‑Nearest Neighbors

K-means, also known as K-means clustering, is an unsupervised learning technique that

partitions data into predetermined classes through iterative convergence [42]. Initially, the
algorithm randomly selects K cluster centers or centroids, where K is predetermined. Then,
the data are allocated to the closest centroid based on the Euclidean distance [43]. The
algorithm iteratively determines new centroids until convergence is achieved while mini-
mizing the sum of the squared errors [44].
K-nearest neighbor is a type of supervised learning method that is utilized for solving
both classification and regression tasks. The fundamental principle behind KNN involves
analyzing the distance between each data point in a given dataset and its nearest neighbors,
typically by utilizing Euclidean distance calculations [43]. If the majority of the nearest
neighbors for a specific sample are of a particular category, then that sample is classified
into that category. Although KNN algorithms are simple to apply, they have certain restric-
tions when working with multi-dimensional data that has irrelevant attributes. Addition-
ally, there is no fixed rule to determine the ideal value for K [44].
K-means and KNN are two distinct algorithms that are often mistakenly interchanged
because of their similar names. To provide clarification, the “K" in K-means specifies the
number of classes, whereas the “K" in KNN refers to the number of nearest neighbors.
Nevertheless, it is common in research literature to observe the utilization of these two
algorithms in conjunction.

2.7 Principal Component Analysis

Principle component analysis is a statistical technique that utilizes a vector space transfor-
mation to decrease the number of potentially correlated variables, also known as principal

13
 Applied Composite Materials

components. PCA was initially developed for analyzing multivariate data, but it has now
found applications in a variety of other fields [45]. As a result of its many advantages,
PCA has become the most commonly used multivariate statistical approach in the scien-
tific community [46].
If we have a dataset X with standardized zero-mean values, and a matrix of size m × n,
where m denotes the number of unique parameters and n represents the number of meas-
urements for each parameter, then we can obtain the covariance matrix CX using Eq. 4 [47]:
1 T
CX = XX (4)
n
PCA involves finding an orthonormal matrix P that linearly transforms X into Y = PX ,
leading to a diagonal covariance matrix CY . The off-diagonal terms in CY should be zero,
indicating uncorrelated variables. Furthermore, the subsequent dimensions in Y must be
ordered by variance, with the primary components being the arranged set of pi elements.
Equation 5 represents the rewritten form of CY [47]:
1
CY = (PX)(PX)T = PCX PT (5)
n
Since the covariance matrix CX is symmetric, we can diagonalize it by using the ortho-
normal eigenvector matrix E, such that CX = EDET . By selecting P = ET and recognizing
that the inverse of an orthogonal matrix is its transpose, we can obtain Eq. 6 [47].

CY = P(EDET )PT = D (6)

where CY is a diagonal matrix. Thus, the principal components of X is the eigenvectors of CX.
In short, PCA aims to find a new set of meaningful and independent dimensions by
combining the original variables in a linear way. This results in a transformed data set with
reduced noise and potentially reveals hidden patterns or information.

2.8 Support Vector Machine

Supervised learning algorithms called Support Vector Machines are utilized for classifica-
tion and regression to detect patterns in data. The conventional SVM takes a dataset as input
and predicts whether the input belongs to one of two classes, without relying on probabilities,
thereby functioning as a binary linear classifier. The primary objective of SVMs is to construct
a decision boundary that maximizes the margin between the two classes [48].
In simple terms, the goal of SVM is to locate a line, or hyperplane, that can divide data
into two distinct classes. This hyperplane, as shown in Fig. 4, aims to optimize the dis-
tance between the nearest points of each class [49]. The ideal hyperplane is the one with
the greatest margin of separation between the classes. Although margin maximization is a
measure of SVM’s effectiveness, the technique prioritizes class division over margin maxi-
mization, indicating that categorizing the classes accurately is more critical than maximiz-
ing the margin.
In addition to linear classification, SVMs may do non-linear classification effectively by
employing the kernel technique, which involves implicitly mapping their inputs into high-
dimensional feature spaces [50].

13
Applied Composite Materials

Fig. 4  A graphical representation

of an SVM (adapted from [51])

2.9 Comparison of Machine Learning Algorithms

Due to the distinct characteristics and objectives of various machine learning methods,
along with their heavy reliance on data and data formats, making direct comparisons
between them to highlight advantages and disadvantages can be challenging. However,
by analyzing each method carefully, that is, evaluating where each algorithm is gener-
ally applied and how it is, it is possible to extract certain general characteristics that
indicate the advantages and disadvantages of each method. It’s worth noting that there
are cases where unconventional utilization of certain algorithms can yield favorable out-
comes. Table 1 presents a general guide to the theoretical advantages and disadvantages
of each machine learning method.

3 Review Methodology

This section presents the methodology used in the systematic review. In these steps, we
followed some guidelines to select the best and most influential articles related to the
topic of machine learning on damage assessment of composite structures.
The Scopus database was used as a screening tool to select the most suitable papers
for the study. In it, keywords were used in order to find articles that matched the desired
topic. The keywords were combined using Boolean operators, that is, AND and OR,
resulting in the following search:
(damage OR assessment OR detection OR quantification OR localization) AND
(composite OR fiber OR polymer OR laminate) AND

13
13
Table 1  Theoretical advantages and disadvantages of each machine learning algorithm. adopted from [52]
Algorithm Advantages Disadvantages Training time

Artificial Neural Network Capability to learn and understand nonlinear and complex relationships Difficult to comprehend and interpret Long
Autoregressive Models Simple Implementation When the autocorrelation coefficient is less than Average
0.5, the prediction is inaccurate
Bayesian Classifier A simple technique, with only a small amount of training data necessary There is no interdependence between the features Average
Convolutional Neural Network Obtaining major conclusions from data collection is quick and simple Difficult to comprehend and interpret Long
K-Means Simple Implementation high dependence on initial parameters Average
K-Nearest Neighbor Simple Implementation In a high-dimensional space, it is easy to overfit Quick
Principal Component Analysis Excellent downscaling ability Poor naming clarity Quick
Support Vector Machine Effective in high-dimensional spaces Kernel function sensitive Long
Bayesian Classifier A simple technique, with only a small amount of training data necessary There is no interdependence between the features Average
Applied Composite Materials
Applied Composite Materials

(machine learning OR support vector machine OR principal component analysis OR

k-means OR K-nearest neighbors OR convolutional neural network OR bayesian
classifier OR auto-regressive model OR artificial neural network).
Other filters were also used to further narrow down the results. The surveys only
included articles; conference papers, Ph.D. thesis, books, editorial comments, and techni-
cal reports were not included. This was done in order to choose articles with the highest
level of research outputs and that have been extensively peer-reviewed. Furthermore, only
publications published between January 2012 and December 2022 were evaluated in order
to avoid picking research that is no longer pertinent to the topic. After the first screening,
the abstracts of the selected articles were reviewed. Those who had an abstract that did not
discuss something related to damage assessment in composite structures using machine
learning were also excluded.

4 Results and Discussion

In this section, we present the results obtained with the systematic review in addition to the
discussions. It is comprised of Research trends, Publication venues, and the Selected stud-
ies as mentioned in Section 3.

4.1 Research Trend

So, sorted by year of publication, we can see that until 2018, there were few publications
on damage assessment using machine learning techniques. It was only after 2019 that we
began to see an increase in the number of publications on the topic. This indicates its cur-
rent relevance and ongoing development. Figure 5 shows the trend.

4.2 Publication Venue

In this subsection, we present the distribution of the selected studies in this subsection by
taking into account the published channels. The studies appeared in a total of 37, where

Fig. 5  Number of publications per year

13
 Applied Composite Materials

Composites structures journal (14 articles) takes the predominant publication on the topic.
It is followed by Structural Health Monitoring and Sensors, both with 6 articles. With 3
articles published, we have Composites Part B: Engineering, IEEE Access, Journal of
Intelligent Material Systems and Structures, and Journal of Nondestructive Evaluation. The
other journals that share the same number of publications (2 articles) are Applied Acous-
tics, Applied Composite Materials, Composites Science and Technology, Cement and Con-
crete Composites, Materials, Proceedings of the Institution of Mechanical Engineers, Part
C, and Smart Materials and Structures. Lastly, the rest journals published only one article.
Table 2 shows all publication journals, number of articles, and publishers.

4.3 Selected Studies

As previously mentioned, there has been a rise in the amount of work utilizing machine
learning approaches for damage assessment of composite structures in recent years. This
is mainly due to the greater use of composites in the industry, which consequently requires
greater monitoring of the condition of the material in order to guarantee reliability as a
whole. This subsection is dedicated to state-of-the-art approaches for damage assessment
of composite structures based on machine learning techniques. Figure 6 shows the contri-
bution of each machine-learning technique to the overall studies chosen. The majority of
research uses ANN (24.7%), followed by CNN and PCA accounting for 22.4% and 14.1%,
respectively. K-means and KNN account for 12.9% and 3.5% of the total studies, respec-
tively. SVM accounts for 11.8% of the total analyzed studies. Finally, the Bayesian classi-
fier and AR-model account for 5.9% and 4.7% of all studies, respectively.

4.3.1 Artificial Neural Network Studies

Ramasamy and Sampathkumar [53] propose using Acoustic Emission and an Artificial
Neural Network to assess drop impact damage on woven glass fiber reinforced polymer
(WGFRP) composite laminates. The ANN trained with AE parameters accurately predicts
impact damage tolerance. The average error tolerance of the estimated impact damage tol-
erance was 3.35%, indicating good agreement.
In another study by Ramasamy and Sampathkumar [54] extend their previous method
by combining AE and ultrasonic techniques for offline and online monitoring of drop
impact damage on WGFRP composite laminates. The ANN-based approach provides a
real-time prediction of Impact Damage Tolerance.
Khatir et al. [55] propose a novel approach combining Cornwell Indication, Artificial Neural
Network, and Particle Swarm Optimization to estimate damage in laminated composite plates.
The method achieves high accuracy in damage quantification and significant time savings.
Qian et al. [56] introduces a non-destructive damage detection technique for fiber-rein-
forced composite structures using Lamb waves and an ANN model. The study involved
simulating damages of varying sizes on a unidirectional carbon fiber-reinforced polymer
composite plate using rectangular Teflon tapes. The ANN accurately quantifies damage
size in these structures.
Califano et al. [57] describe a passive structural health monitoring method using strains
and ANN to detect damage in carbon fiber/epoxy composite plates. The approach success-
fully detects abnormal conditions, even if they were not structurally significant.

13
Applied Composite Materials

Table 2  Dissemination of the selected articles (in number of articles)

No Jounal Publisher #

1 Composite Structures Elsevier 14

2 Sensors MDPI 6
3 Structural Health Monitoring SAGE 6
4 Composites Part B: Engineering Elsevier 3
5 IEEE Access IEE 3
6 Journal of Intelligent Material Systems and Structures SAGE 3
7 Journal of Nondestructive Evaluation Springer 3
8 Applied Acoustics Elsevier 2
9 Applied Composite Materials Springer 2
10 Composites Science and Technology Elsevier 2
11 Cement and Concrete Composites Elsevier 2
12 Materials MDPI 2
13 Proceedings of the Institution of Mechanical Engineers, Part C SAGE 2
14 Smart Materials and Structures IOPscience 2
15 Acta Mechanica Sinica Springer 1
16 Advances in Engineering Software Springer 1
17 Advances in Mechanical Engineering SAGE 1
18 Applied Sciences MDPI 1
19 Engineering Structures Elsevier 1
20 Heliyon Elsevier 1
21 International Journal of Damage Mechanics SAGE 1
22 International Journal of Integrated Engineering UTHM 1
23 International Journal of Mechanical Sciences Elsevier 1
24 Iranian Journal of Science and Technology ACM 1
25 Journal of Intelligent Manufacturing Springer 1
26 Journal of Process Control Elsevier 1
27 Journal of Reinforced Plastics and Composites SAGE 1
28 Journal of The Institution of Engineers (India): Series C Springer 1
29 Journal of Vibration and Control SAGE 1
30 Journal of Vibration Engineering & Technologies Springer 1
31 Machine Vision and Applications Springer 1
32 Nondestructive Testing and Evaluation Taylor & Francis 1
33 Optik Elsevier 1
34 Structural Monitoring and Maintenance Korea Science 1
35 Structure and Infrastructure Engineering Taylor & Francis 1
36 JOM Springer 1
37 Ultrasonics Elsevier 1

13
 Applied Composite Materials

Fig. 6  Percentage of articles chosen within machine learning topics

Zenzen et al. [58] propose a new damage detection method combining the transmis-
sibility approach with ANN for improved local frequency response ratio predictions in
composite structures. The method reduces data requirements and offers accurate damage
prediction compared to conventional ANN. The training results of each damage scenario
are presented in Fig. 7.
Mardanshahi et al. [59] investigate a nondestructive assessment system for detecting core
cracking and skin/core debonding in foam core sandwich structures. A large dataset of the dam-
aged sandwich structure’s first five harmonic frequencies with varying damage amounts and
locations was generated using the updated finite element model. The combination of experimen-
tal data, ANN, and finite element models shows promise for nondestructive evaluation.
Tan et al. [60] use vibration characteristics and ANN to detect damage in composite
bridge construction. They combined the ANN with a damage index based on modal strain
energy to identify and measure damage in the steel beams. The method successfully identi-
fies and measures deterioration in composite concrete slabs on steel girder bridges.
Jang et al. [61] present a technique to identify debonding damage using natural fre-
quency and an ANN. The authors simulated over a thousand debonding damages with dif-
ferent configurations, positions, and lengths using modal analysis from a Finite element
model. The method achieves a high success rate in detecting damage.
Khatir et al. [62] propose a two-stage method for damage detection, localization, and
measurement in functionally graded material plate structures. The authors employed isoge-
ometric analysis to model the FGM plates, which is a more efficient approach than tra-
ditional finite element analysis. An improved ANN-based approach accurately identifies
damaged components.
Tran-Ngoc et al. [63] develop an efficient approach for detecting damage in laminated
composite structures using an ANN and a hybrid metaheuristic optimization algorithm.
The method surpasses standard ANNs in terms of accuracy and reduces computing time.

13
Applied Composite Materials

Fig. 7  Damage quantification in cross-ply composite plate: a scenario 1, b scenario 2, c scenario 3, and d
scenario 4.) (adapted from [58])

Nasser et al. [64] explore using the piezoresistive properties of LIG interlayered fiber-
glass composites for damage assessment using ANNs. A two-layer Bayesian regularized
ANN with 40 neurons in each layer achieved a higher level of accuracy compared to con-
ventional degradation models.
Al-Athel et al. [65] combine hybrid thermography, computational modeling, and ANN
techniques to analyze defects in composites. The hybrid technology accurately predicts
both the form and magnitude of damage.
Feng et al. [66] propose a method to locate low-velocity impacts on a composite plate
using multi-frequency properties of impact-induced guided waves. Fusion techniques
involving ANN achieve accurate localization of damage. Then the images were added to
get the localization results with multi-frequency fusion Fig. 8.
Saadatmorad et al. [67] present a method for damage quantification in rectangular lami-
nated composite plates using wavelet transforms and ANN. The findings show that the pro-
posed method can accurately quantify damage in rectangular laminated composite plates
(with R = 0.992) and that the presence of “edge noise" can be a concerning factor in dam-
age diagnosis algorithms that rely on wavelet transformations.

13
 Applied Composite Materials

Fig. 8  Imaging results of the 32 impacts (horizontal and vertical axes of each image are in the unit of mm,
color bar shows the value of probability.) (adapted from [66])

Reis et al. [68] develop a method to detect and classify damage in glass fiber-reinforced
plastic composite beams using ANN and vibration data. The technique shows promise in
identifying damage based on dislocated time series parameters.
Overall, these articles collectively contribute to the advancement of damage detection
and assessment techniques in composite structures using the most diverse sensing tech-
nologies; such as acoustic emission, strain signals, impact signals, lamb waves, and oth-
ers; which have been combined with good scalability and strong nonlinear generalization
of ANNs. In general, the ANNs used in the damage assessment of composite materials
were of the feed-forward type, with one to two hidden layers and an average of 20 neurons
per layer. MATLAB software was the most commonly used in their development. These
research efforts aim to enhance the safety, reliability, and longevity of composite structures
in various industries.

4.3.2 Autoregressive Models Studies

Mouzakis et al. [69] conducted a study on the impact of aging on carbon fiber-reinforced
composites using AR-models. They subjected composite plates to two aging scenarios
and performed impact testing on aged and non-aged composites. The results showed
that impact test data could distinguish between aged and non-aged specimens and assess
the level of damage. The study also correlated impact test behavior with changes in the
mechanical properties of aged specimens.
Nardi et al. [70] explored the use of AR models for detecting delaminations in carbon-
fiber-reinforced-plastic laminate plates caused by low-velocity impacts. They employed
piezoelectric patches for actuation and sensing and utilized the captured signal to generate
an AR Model. Linear Discriminant Analysis was applied to enhance delamination detec-
tion. The technique showed high success rates in identifying delamination-induced damage
with a small number of sensors.

13
Applied Composite Materials

da Silva et al. [71] proposed a data-driven approach using multiple AR models to pre-
dict the time-series outputs of a PZT sensor detecting Lamb waves in composite speci-
mens. They demonstrated a good correlation between the extrapolated model’s estimations
and observed damage progression. The method offers the advantage of achieving this task
without complex mathematical-physical models.
Paixão et al. [72] investigated the use of stochastically interpolated global damage indi-
ces based on Gaussian process regression for measuring delamination area. They utilized
AR models to extract damage-sensitive features from Lamb wave signals and calculated
damage indices using the Mahalanobis squared distance. The methodology was tested on
carbon epoxy laminate with simulated damage and carbon fiber-reinforced polymer cou-
pons with real delamination. It effectively predicted the affected region for both simulated
damage and actual delamination. Figure 9 shows, for the application case I, the damage
indices computed sequentially for all tested conditions.
In summary, the application of AR models has shown promising results in composite
material damage assessment. These methods have demonstrated their capability to distin-
guish between aged and non-aged specimens, detect delaminations, predict damage pro-
gression, and measure the delamination area. They offer advantages such as capturing the
effects of aging, requiring a small number of sensors (Usually Piezoelectric sensors), and
providing cost-effective solutions without the need for complex mathematical-physical
models. These findings highlight the potential of machine learning in advancing the field
of composite material damage assessment. In most of the papers studied, the AR models
used are of a high order, that is, p is greater than 20. Finally, MATLAB software was the
most used in the development of AR models.

Fig. 9  DI computed for all signals in application case I: carbon-epoxy plate (adapted from [72])

13
 Applied Composite Materials

4.3.3 Bayesian Classifier Studies

Peng et al. [73] proposed a probabilistic method for detecting delamination in carbon-
carbon composites using Lamb wave-based characteristics. The method utilized Bayesian
imaging to create a probability image of the delaminated region, and its accuracy was vali-
dated by comparing results with X-ray images.
Fendzi et al. [74] damage localization in anisotropic composite materials using a Bayesian
framework. They developed a unique formulation to represent the direction dependence of
group velocity and demonstrated its accuracy through computational and empirical testing.
The approach accurately predicted the location and size of the damage, considering different
degrees of anisotropy and estimating associated confidence intervals.
Cantero-Chinchilla et al. [75] presented an approach for identifying and locating dam-
age in composite beam structures using ultrasonic guided waves. The technique employed a
transient wave propagation model within a multilevel Bayesian framework, enabling dam-
age detection and localization without baseline comparison or additional transformations.
The method effectively reconstructed and diagnosed various types of damage in composite
beams while maintaining a reasonable computational cost.
Huo et al. [76] introduced a novel method for composite deterioration detection
using Lamb waves. Their approach combined the elliptical loci method and the RAPID
algorithm in a Bayesian framework, incorporating various damage-sensitive features
to enhance reliability and robustness. Numerical simulations and experimental inves-
tigations demonstrated the method’s superior accuracy and dependability compared to
existing approaches. Figure 10 shows that the proposed method can yield more accurate
damage localization results compared with the existing methods.
In this subsection, the Bayesian classifiers applied in the damage assessment of composite
materials were reviewed. Bayesian classifiers are a set of probabilistic techniques in which
these algorithms are based on a single principle: all classifiers assume that the value of one
feature is independent of the value of any other feature, given the class variable. These articles
collectively contribute to the advancement of damage detection techniques in composite mate-
rials, highlighting the use of Lamb waves and ultrasonic-guided waves. The proposed methods
offer improved accuracy, reliability, and robustness in detecting and localizing damage, con-
sidering various factors such as anisotropy and uncertainty estimation.

4.3.4 Convolutional Neural Network Studies

Guo et al. [77] proposed a detection model combining a full convolution network and a
gated recurrent unit for classifying ultrasonic signals from flawed 3D braided composite
specimens. The study’s findings show that the proposed model accurately identifies ultra-
sonic fault signals and outperforms six other models on the same dataset.
Nasiri et al. [78] developed a deep learning-based end-to-end convolutional neural net-
work model for online monitoring of the damage advancement process of SiCf -SiCm com-
posite tubes using AE data. The CNN model with single events achieves an average pre-
diction accuracy of 84.4%, while the RF models achieve 74%. Combining several audio
samples improves the accuracy of both models, with the RF accuracy reaching 82.8% and
the CNN accuracy reaching 86.6%.
Tabian et al. [79] developed a CNN-based metamodel for impact detection and localiza-
tion in complex composite structures. Piezoelectric sensors collect ultrasonic waves pro-
duced by external impact events, which are then converted into 2D images for analysis.

13
Applied Composite Materials

Fig. 10  Comparisons of the damage localization results obtained using different methods: a the elliptical
loci method, b the RAPID method, and c the proposed method (adapted from [76])

The detection accuracy exceeded 94%, and the approach showed scalability and applicabil-
ity to real-world applications. Figure 11 represents the confusion matrix, in which it can be
observed on the main diagonal that all impacts, except one, were correctly predicted.
Salehzadeh Nobari and Aliabadi [80] developed a machine-learning technique that uti-
lizes voltage signals obtained from piezoelectric sensors placed on a composite panel to
differentiate between hard and soft impacts.
Das et al. [81] proposed a framework for identifying and computing fracture parameters
in thin strain-hardening cement composite cracks using a customized deep convolutional
neural network, the TDCNN. To identify tiny fractures, crack parameters were derived
using an image processing approach. The TDCNN demonstrated high inference capacity
and resistance to uncertainty.
Saadatmorad et al. [82] introduced a novel technique, known as the Wavelet Transform-
based Convolutional Neural Network, for detecting damage in rectangular laminated com-
posite plates. In this method, a finite element model of the damaged RLCP is created to
generate two-dimensional signals that are utilized in the wavelet transform. The proposed
approach accurately predicted and identified the location of damage without requiring trial
and error simulations.

13
 Applied Composite Materials

Fig. 11  a The confusion matrix corresponding to run D6 in the Table 2. Each class represents a range of
energy levels. b Energy level prediction for 6 locations, for 4 distinct energy levels. (adapted from [79])

Wu et al. [83] integrated a CNN with continuous wavelet transform for detecting inter-
nal delamination in carbon fiber-reinforced plastics. The approach integrates a convolu-
tional neural network with the continuous wavelet transform, eliminating the need for com-
plex feature extraction and allowing for the efficient utilization of large amounts of data.
The method accurately detected and located delamination damage in composite structures.
Djavadifar et al. [84] described a process for comparing four high-performing convo-
lutional neural network models for detecting geometrical flaws in composites. In the best
situation, wrinkle detection initially obtains a substantially lower IoU score of 0.40. The
model is then assessed as a binary predictor based on per-component detection success;
the model obtains a recall rate of 0.71 and an accuracy score of 0.76 (i.e., the ratio of those
identified being genuinely wrinkled). Based on the data provided, this model can outper-
form a human operator in practice.
Augustin et al. [85] proposes a 3D laser scanner-based approach for the detection of inclu-
sion on flat and curved surfaces. Using the portable laser scanner, the surfaces of each layer
are scanned, and compared the resulting point clouds using reference layer data. Thicknesses
between two surfaces are computed with Cloud to Cloud, Mesh to Cloud, and Hausdorff dis-
tance to enhance the visibility of inclusions. These enhanced features are used to train a multi-
view convolutional neural network to mark the inclusion regions in a fast and efficient way.
Miller and Ziemiański [86] investigated the feasibility of detecting material degradation in
composite structures using vibration mode shapes. A multi-layer composite cylinder was stud-
ied, with a square area of the cylinder’s lateral surface designated as the material degradation
zone. A CNN was employed to accurately identify the size and location of the degraded zone.
Azuara et al. [87] developed a technique to predict the distance-to-damage values using a
convolutional neural network, which takes the received experimental data as input and trans-
forms it into a two-dimensional picture using wavelet transform. The training process of the
models resulted in an accuracy of approximately 90%, and the results obtained using the imag-
ing method demonstrated the reliability of the whole model (CNN plus imaging algorithm) for
accurate damage localization. The results for tests considering a grid size of 100 × 100 pixels
(each pixel area is 9 mm2), and a 𝜎 value of 12.5 mm, are shown in Fig. 12.

13
Applied Composite Materials

Fig. 12  Results from the regression-based model for every test point (coordinates in m). Crosses point to
the actual location of the damage, and circles to the estimated locations. (adapted from [87])

Machado et al. [88] proposed a machine learning approach using a convolutional neu-
ral network to automatically analyze the void content in optical microscope images. This
approach is designed to be robust and requires no parameter adjustment. The authors dem-
onstrate that their approach outperforms a traditional thresholding algorithm for parsing
void content from microscopy images, even across different types of laminates. While indi-
vidual void statistics may show lower-than-expected accuracy, the study indicates that the
proposed technique is promising.
Wang et al. [89] used a CNN-based image semantic segmentation method for pixel-
level identification of Digital Image Correlation strain field images. The training dataset is
generated through finite element simulation and validated by the model. The trained CNN
model showed high accuracy for strain field segmentation.
Cui et al. [90] developed an ultrasonic-guided wave approach using the most sensitive
wave characteristics selected by the learned training data to detect and locate structural
degradation in a stiffened composite panel. The results demonstrated the technique’s ability
to detect damage in critical areas of the stiffened composite test panel, including the skin
region, stringer flange region, and stringer cap region.
Yang et al. [91] proposed a novel approach combining CNN and algebraic reconstruc-
tion techniques for ultrasound tomographic imaging of composites. The proposed method
uses blurred ART images as inputs to a CNN with an encoder-decoder architecture, which
segments the images using convolution and max-pooling to retrieve defect-modulated
image characteristics. The approach accurately detected anomalies and delamination, even
with limited sensing capabilities.
Helwing et al. [92] analyzed the damage evolution of GFRP composites using in situ
x-ray CT analysis and a CNN-based fault segmentation approach, improving defect detec-
tion quality and sensitivity. Detailed characteristic values for the damage behavior of GFRP
at different stiffness decreases are obtained by selectively identifying faults according to
the underlying damage mechanism. This results in a mechanism-differentiated quantifica-
tion of damage progression as stiffness decreases.

13
 Applied Composite Materials

Barile et al. [93] trained a deep convolutional network for image-based acoustic emis-
sion waveform classification. Spectrograms of AE Waveforms from four distinct damage
types, matrix cracking, delamination, debonding, and fiber breaking, were acquired in
Mel scale and utilized as training and test data for the CNN. The CNN’s overall prediction
accuracy is 97.9%, whereas fiber breakage and delamination events may be predicted with
100% accuracy.
CNNs are a class of artificial neural networks, most commonly applied to analyze 2D
signal images but also used in 1D signals. In this subsection, studies that use convolu-
tional neural networks for damage assessment of composite materials are reviewed. In the
reviewed studies, a lot of convolutional networks used defined architectures (like U-net,
Mask R-CNN, Deeplab V3+, IC-net, and others) as well as networks built part by part,
that is, inserting each layer in the desired sequence. CNNs have a high data-driven learning
capacity and thus perform well in most problems. In general, most of the studied papers
used the Python programming language in their development.

4.3.5 K‑Means and K‑Nearest Neighbors Studies

Barile et al. [94] investigated damage characteristics in CFRP composites using acous-
tic emission. They used Laplacian scores to identify appropriate AE data features and
employed k-means clustering to establish their relationship. They found a strong correla-
tion between AE signal amplitude and frequency centroid (C-Freq) for characterizing dam-
age modes. This technique enables a comprehensive understanding of the damage modes
in CFRP composites using AE analysis.
Pashmforoush et al. [95] developed a method for monitoring delamination in sandwich
composites using AE signals. They combined k-means clustering with the genetic algo-
rithm to categorize damage processes and identify distinct failure scenarios. The research-
ers validated their results through SEM observation and concluded that their approach
effectively classifies damage in sandwich composites.
Lui et al. [96] investigated the behavior of a 59.5-meter-long composite wind turbine
blade under accelerated fatigue stresses using an acoustic emission approach. They ana-
lyzed AE signal components through spectrum analysis and used a time-difference method
for detecting and locating defect sources. They applied to bisect k-means clustering to
identify distinct damage types. The approach was successful in detecting and character-
izing defects in composite blades.
Shrifan et al. [97] developed a microwave-based non-destructive testing approach using
the k-means algorithm for defect identification in GFRP composites. They utilized an
open-ended rectangular waveguide and mean measurements to reduce the effect of per-
mittivity changes. Unsupervised machine learning with k-means clustering successfully
detected defects as small as 1 mm.
Ech-Choudany et al. [98] proposed an incremental clustering and k-means approach for
analyzing AE data in glass fiber-reinforced composites. IC identified information-carrying
signals and determined damage mechanisms without additional knowledge. The strong
performance of the supervised method, as indicated by AUC values greater than 0.9 for
each damage type, validated the reliability of the generated learning database. Figure 13
shows the results obtained by using 4 clusters.
Zeng et al. [99] proposed a real-time health monitoring approach for 2D C/SiC compos-
ites. They employed unsupervised recognition (k-means++) for the spectrum properties of AE
signals and supervised identification for damage identification. The study identified primary

13
Applied Composite Materials

Fig. 13  Projection of UD0 composite clustering for 4 clusters obtained by the k-means method. (adapted
from [98])

damage mechanisms and examined damage evolution using AE energy accumulation curves.
The proposed approach showed improved sensitivity and accuracy compared to previous stud-
ies. Figure 14 shows that the method can effectively identify the damage modes.
Essasi et al. [100] discussed the behavior of an eco-sandwich composite with an auxetic
core. Different configurations were explored, and damage behavior was monitored using
AE. The k-means method was used for categorization, revealing core cracking, matrix
cracking, and fiber/matrix debonding as the main occurrences for each configuration.
This subsection reviewed papers that use the KNN and K-means algorithms. Despite
their names, the two algorithms have significant differences, with K-means being an unsu-
pervised ML method and KNN being a supervised method. Therefore, K-means presents
lower performance compared to KNN, but it has the advantage of being capable of anom-
aly detection. In general, most studies used both methods together with acoustic emission
signals, and some studies combined clustering algorithms with other methods like genetic
algorithms or time-difference methods for enhanced analysis.

13
 Applied Composite Materials

Fig. 14  Spectral characteristics distribution of 2D C/SiC at ambient temperature related to time. (adapted
from [99])

4.3.6 Principal Component Analysis Studies

Taghizadeh and Ahmadi [101] investigated delamination in ultra-high molecular weight

polypropylene/epoxy composites with various topologies using acoustic emission signal
analysis. They performed tensile tests and Mode I delamination on epoxy resin, UHMWPP
fiber bundles, and UHMWPP/epoxy specimens and managed to identify damage modes
including matrix cracking, fiber-matrix debonding, fiber breakage, and fiber pull-out using
principal component analysis.
Rao et al. [102] proposed a diagnostic technique based on frequency response functions
and principal components to detect delamination in laminated composites. Finite element
models of laminated composite beams and plates were used to produce vibration data for
healthy and damaged structures. Their method can identify the instant of damage, spatial
location, amount of delamination, and matrix cracking.
Sierra-Pérez et al. [103] developed a method using fiber Bragg gratings as strain sensors and
principal component analysis to detect damage in composite structures. The proposed methodol-
ogy was tested on a 1.5-meter-long composite unmanned aerial vehicle wing piece, and it was
able to identify variations between the baseline and different damage scenarios.
Yan et al. [104] introduced a stable principal component pursuit strategy combined with a
moving-window strategy for analyzing thermographic data. Thermal Their method effectively
extracts background information, reduces noise, and identifies defects in tested specimens.
Mohamad [105] proposed a damage classification method for natural fiber-reinforced
composites using signal processing parameters and principal component analysis. They
analyzed strain data from an impact event and observed that the network could learn and

13
Applied Composite Materials

Fig. 15  PCA plot for the combination of all time-domain features. (adapted from [105])

classify the size of damage in the panel. The combination of features retained about 84.5%
of the variance. Figure 15 shows the results for five impact severity classes.
Lu et al. [106] developed a damage detection method using fiber Bragg grating sen-
sors and principal component analysis in carbon fiber reinforced polymer structures. They
first excited the undamaged carbon fiber-reinforced polymer structure and measured its
dynamic response signal. The frequency response was extracted using the Fourier trans-
form method as the damage feature, which was then reduced in dimension using PCA.
Their approach accurately detects structural damage using only undamaged state samples.
Malinowski et al. [107] explored the use of electromechanical impedance and princi-
pal component analysis for damage detection in glass fiber-reinforced polymer composites.
They investigated a beam and a plate with simulated delamination damage. Their data pro-
cessing tool utilizing PCA showed promise in managing sensor network data, differentiat-
ing between damage scenarios, and detecting damage with missing data.
Lu et al. [108] proposed a method for assessing damage in carbon fiber-reinforced polymer
laminates using Synchrosqueezed Wavelet Transform and ensemble Principal Component Anal-
ysis. They developed an approach for extracting damaged features based on Synchrosqueezed
Wavelet Transform and Singular Spectrum Analysis to reduce feature dimension and success-
fully extract damaged features. Their approach effectively extracted damaged features and
assessed damage in the presence of noise interference and limited training samples.
All the studies reviewed in this subsection show that PCA is a technique that is
extremely useful for dealing with large datasets, but as a disadvantage, the feature visuali-
zation is impaired when the problem features many classes. It was also possible to notice
that the PCA was used in conjunction with several sensing techniques (impact testing,
vibration, AE signals, thermography, strain, lamb waves, and EMI). Although PCA has a
lower accuracy than the other techniques, it has the advantage of not requiring knowledge
of the defect, only the healthy state.

4.3.7 Support Vector Machines Studies

Fredo et al. [109] developed a method for classifying damages in composite materials using
photographs taken from the front and back sides of the materials after impingement of dif-
ferent sizes. They utilized anisotropic diffusion filtering and Fuzzy C-Means clustering for
segmentation. Geometrical characteristics and Zernike Moments were used as features, and
a Support Vector Machine classifier was employed. Zernike Moments performed better in
both global and local damage classification.

13
 Applied Composite Materials

Forero-Ramírez et al. [110]. investigated the effectiveness of Support Vector Machines

in detecting internal defects using thermal contrast information obtained from the Back-
ground Thermal Compensation by Filtering method. They examined a 20 × 20 cm CFRP
slab with 25 square Teflon insertions to simulate flaws, and IR images were captured using
Active Pulsed Thermography equipment under two different conditions. They achieved
higher sensitivity in flaw detection by combining the Background Thermal Compensation
by Filtering contrast method with SVM.
Shyamala et al. [111] used Support Vector Machines to detect damage in composites.
The algorithm locates the damage and evaluates its severity based on changes in dynamic
responses. The results demonstrate that the proposed algorithm effectively detects a wide
range of damage scenarios.
Gillespie et al. [112] studied the impact of delaminations on the transient heat conduc-
tion profile of carbon fiber-reinforced polymer laminates. They employed a supervised
Support Vector Classification method to detect delaminations using temperature data col-
lected from either the perimeter or the center of the defect during a 140-second ramping
heating session up to 80 degrees Celsius. The findings show high accuracy in classifying
delaminations or no delamination. Figure 16 shows the results for a studied sample.
Yue et al. [113] developed a constitutive model for plastic damage in steel fiber rein-
forced concrete. They conducted three-point bending tests on 27 concrete specimens with
varying fiber volume fractions and concrete matrix strengths and monitored the acoustic
emissions during fracture to analyze the released strain energy in the fracture process zone.
They used SVM to classify acoustic emission data into tension-crack mode and shear-crack
mode and established correlations between fiber volume fraction, characteristic length, and
damage variables.
Xu et al. [114] proposed a method for determining the degree of damage in carbon
fiber-reinforced polymer cables using a PSO-SVM algorithm based on nine parameters of
acoustic emission signals. Acoustic emission signals were analyzed, and the trained sup-
port vector machine accurately predicted the damage degree. These findings indicate that

Fig. 16  Display of overall percentage score (0-1) of the 140 time intervals for each RTD probes which clas-
sified as ‘No Delamination’ a Corresponding RTD probe locations, b Gaussian filter application to present
images similar to traditional IR thermography methods. (adapted from [112])

13
Applied Composite Materials

acoustic emission combined with machine learning is a feasible approach to determining

the damage degree in carbon fiber-reinforced polymer cables.
Sheng et al. [115] presented a novel approach for categorizing physical damage in car-
bon fiber-reinforced polymer. using AdaBoost and Support Vector Machine. The experi-
mental results revealed that the modified AdaBoost methods with distance-based weighted
least square support vector machine and filter factor performed better than traditional algo-
rithms. The authors also discussed the impact of the new weighted update technique and
filter factor on the algorithm.
Rajiv et al. [116] employed a multi-support vector machine for durability and damage
classification of composite materials using filtered images. Experimental validation on
various composite materials was conducted using real-time images, and the results dem-
onstrate that the proposed approach achieves higher accuracy in image classification than
other models. Additionally, the longevity of the material is assessed based on the material
removal rate, wear resistance rate, and material strength.
Alhammad et al. [117] used pulsed thermography and machine learning for the auto-
mated detection of damage in composite materials. The classification findings varied from
78.7% to 93.5%, opening the way for the development of an automated model to efficiently
evaluate the damage to composite materials using the Non-Destructive Testing approach.
Support vector machines are non-probabilistic binary linear classifiers used to perform
classifications and regressions. In this subsection, the analyzed studies generally combine
SVM with mechanical failure tests, such as the three-point flexion test, cumulative energy
test, fatigue test, and others. The SVM has a great advantage in analyzing complex data-
sets; however, if this dataset is very large, there may be a high computational cost.

4.3.8 Miscellaneous

This subsection of the article provides a review of the machine learning frameworks that
are developed from a combination of basic machine learning tools. These combinations
occur mainly to combine the positive points of each technique and/or to supply any defi-
ciency that some techniques have by incorporating a new one.
Pashmforoush et al. [118] explored acoustic emission monitoring to identify and catego-
rize damage modes in glass/polyester composites. They used PCA and k-means algorithms
to cluster acoustic data, successfully classifying three distinct damage processes.
Selva et al. [119] developed a system using electromechanical impedance technology
for monitoring carbon fiber-reinforced plates. To simulate damage, numerical simulations
based on finite element analysis are used. The input data for the ANNs are obtained from
a principal component analysis of the damage measurements, and the ANN is shown to
accurately predict the location of a single damaged patch in laminated composite plates.
Crivelli et al. [120] proposed a neural network-based approach for classifying acoustic
emission signals during glass-fiber tensile testing. They trained a self-organizing map and
used k-means clustering to accurately categorize specimens with different material layups
and recognize signals from notched specimens. These findings suggest the potential of the
proposed method for real-world structural health monitoring applications.
Mardanshahi et al. [121] developed a guided wave propagation-based method for detect-
ing and classifying matrix cracking in composites. Glass/epoxy cross-ply laminated com-
posites were fabricated, and matrix cracking of different densities was induced in 90°
layers. Linear discriminant analysis was used to reduce the dimensions of the data, and
support vector machines, linear vector quantization ANN, and MLP ANN were employed

13
 Applied Composite Materials

for classification. The results showed that SVM had the highest accuracy rate (91.7%),
followed by linear vector quantization ANN (88.9%) and MLP ANN (77.8%). Figure 17
shows the estimation of the parameters of SCD damage obtained from the NN regression
fitting on the validation experimental data in comparison with the actual values.
Chen et al. [122] compared the effectiveness of an artificial neural network (ANN) and
a convolutional neural network in detecting defects in fiber-reinforced composite materials.
They used micro-CT scanning and the binarized statistical image features approach, find-
ing that CNNs outperformed ANNs in defect prediction.
Hamdi et al. [123] investigated compression damage in CFRP composites with different res-
ins. They used CLC experiments and combined DIC, AE, and in-situ microscopy to track dam-
age progression. K-means, KNN, and PCA clustering of AE signals identified primary clusters
related to potential damage situations. The approach helped identify the onset of severe damage.
Figure 18 shows acoustic activity in terms of Absolute Energy during CLC testing. This activity
is represented for each acoustic class labeled as mentioned below
Dziendzikowski et al. [124] presented a method for impact damage detection in com-
posite structures using PZT sensors and a Bayesian technique for data categorization. The
method can differentiate between pathways responsive to the transmission mode of elastic
wave interaction and those sensitive to the reflection mode. The proposed method showed
better performance in detecting damage pathways and produced fewer false-positive indi-
cations compared to the KNN algorithm.

Fig. 17  Estimation of the parameters of SCD damage obtained from the NN regression fitting on the valida-
tion experimental data in comparison with the actual values. (adapted from [121])

13
Applied Composite Materials

Fig. 18  Normalized values radar chart for the R8 composites for each class and in terms of: CNTS =
counts, AMPL = Amplitude, AFRQ = average frequency, PCNT = counts to peak, RISE = Risetime,
DURA = Duration, ABEN = Absolute Energy.Class 1 (labeled as noise & matrix micro-cracking), Class 2
(labeled as delamination), and Class 3 (labeled as fibers break). (adapted from [123])

Scholz et al. [125] proposed a method for detecting and locating damage in composite
rotors based on structural vibration analysis. They created a databank of 720 simulated test
cases in various damage conditions to generate several data sets and then used it to train
a fully connected neural network and a CNN. They evaluated the trained models using
k-fold cross-validation and measured their sensitivity, specificity, and accuracy. The con-
volutional neural networks showed considerably better performance, achieving up to 99.3%
accuracy in damage localization and quantification.
Lee et al. [126] developed a system for monitoring the health of CFRP structures using
electromechanical properties. They employed a non-destructive self-sensing approach and
utilized k-means clustering and principal component analysis for damage identification. The
system successfully detected various forms of damage in composite structures using only
electromechanical activity.

5 Future Research Directions

Machine learning techniques are increasingly present in the damage assessment of com-
posite structures. However, there are still several challenges and constraints for a full appli-
cation of these techniques, and some of them will be discussed below:

• Noise signal: In industrial applications, an environment with a high level of noise is nor-
mally encountered. As most work is done on simulated data or in a controlled environment,
it is hard to guarantee that the algorithms will perform well in industrial environments.

13
 Applied Composite Materials

• Unbalanced data: In real applications, it is expected to have a greater amount of data in

the healthy condition than in the defect condition, which can affect the training process
and consequently the accuracy of ML algorithms.
• Generalization: Because most ML algorithms are purely data-oriented, they can work
well for similar data sets; however, considering generalization, it still takes a long time
for the algorithms to reach a reasonable behavior. Therefore, it is necessary to imple-
ment other training approaches that are not faithfully data-based.
• Interpretability: Although some methods already have some way of visualizing internal mech-
anisms, known as feature visualization and/or coarse localization maps, it is still not possible
to associate this with physical phenomena. Being able to associate these internal controls with
physical laws can help both in the interpretation of ML methods and in generalization.

ML approaches are still under development, requiring a larger refinement of the tech-
niques, in general, to be fully applied to most real-world issues. Some recent studies evalu-
ate the use of physics-informed neural networks (PINNs) to increase robustness and gener-
alization; however, most studies are still in the early stage of development.

6 Conclusion

The use of composite structures is becoming increasingly frequent and in a greater num-
ber of applications, which demands a greater concern with the reliability of these materi-
als. Consequently, this article reviewed the ML methods applied for damage assessment of
composite structures, classifying them according to the type of ML technique used.
Despite the fact that there are numerous machine learning-based techniques, it is chal-
lenging to identify one that has high damage sensitivity, robustness, portability, and strong
anti-interference properties due to the variety of structural forms, significant variations in
service environments, damage types, sensor networks, and sensing signals. For instance, a
regression or classification approach would work better with an AE methodology than an
object recognition method. An object recognition tool, such as a CNN, would be advan-
tageous for scanning methods like thermography. Therefore, it is extremely important to
know the physics of the problem in order to choose the best method.
Another point is that most studies use either simulated data or test specimens, with few
studies using real structures. In addition, as expected, studies that use real structures have
lower accuracy compared to other types of studies. Therefore, the next step would be to
implement ML techniques in real environments to improve their effectiveness and robust-
ness. Also, with the review of the articles, it was noticed that the main programming lan-
guages used were MATLAB (mainly for ANN) and Python (mainly for CNN).
In addition, supervised learning methods showed the best results for damage classifica-
tion tasks and consequently are the most used learning paradigm. Furthermore, important
present challenges and constraints are discussed, providing research insights into the most
recent ML approaches. The following is a summary of the reviewed articles:

• This study observed that publications became more frequent from 2019 forward.
• ANN, CNN, and PCA were the most commonly used techniques, accounting for nearly
62% of all studies. Selecting the most recent studies (from 2019 onwards), this domi-
nance is even greater.

13
Applied Composite Materials

• To perform the training, the favorite inputs were acoustic emission, vibration, guided waves
(lamb and ultrasonic), and signal images (thermal, strain, and time-frequency maps).
• In general, ML techniques do not require a large data set to reach a reasonable level of
accuracy, making most of the algorithms have a low computational cost.
• ML models are still “black boxes". Thus, future studies that allow the incorporation of
physical laws in the models, thus increasing the interpretability and generalization of
the methods, are very promising.

Funding The authors would like to acknowledge the financial support from the Brazilian agency CNPq
(Conselho Nacional de Desenvolvimento Científico e Tecnológico - 405598/2022-0), CAPES (Coordenação
de Aperfeicoamento de Pessoal de Nível Superior) and FAPEMIG (Fundação de Amparo à Pesquisa do
Estado de Minas Gerais - APQ-00385-18).

Data Availability Statement No new data were created or analyzed during this study. Data sharing is not
applicable to this article.

Declarations
Conflict of Interest The authors have no conflicts of interests/competing interests to declare that are relevant
to the content of this article.

References
1. Gay, D., Hoa, S.V.: Composite materials: design and applications. CRC press (2007)
2. Elmarakbi, A.: Advanced composite materials for automotive applications: Structural integrity and
crashworthiness. John Wiley & Sons (2013)
3. Chung, D.D.: Composite materials: science and applications. Springer Science & Business Media (2010)
4. Pereira, J.L.J., Francisco, M.B., Ribeiro, R.F., Cunha, S.S., Gomes, G.F.: Deep multiobjective design
optimization of CFRP isogrid tubes using lichtenberg algorithm. Soft. Comput. 26(15), 7195–7209
(2022). https://​doi.​org/​10.​1007/​s00500-​022-​07105-9
5. Demircan, G., Ozen, M., Kisa, M., Acikgoz, A., Işıker, Y.: The effect of nano-gelcoat on freeze-
thaw resistance of glass fiber-reinforced polymer composite for marine applications. Ocean Eng. 269,
113589 (2023)
6. Hassani, S., Mousavi, M., Gandomi, A.H.: Structural health monitoring in composite structures: a
comprehensive review. Sensors 22(1), 153 (2021). https://​doi.​org/​10.​3390/​s2201​0153
7. Kar, K.K.: Composite materials: processing, applications, characterizations. Springer (2016)
8. Smith, R.: Composite defects and their detection. Mater. Sci. Eng. 3(1), 103–143 (2009)
9. Talreja, R., Singh, C.V.: Damage and failure of composite materials. Cambridge University Press (2012)
10. Ozen, M., Demircan, G., Kisa, M., Acikgoz, A., Ceyhan, G., Işıker, Y.: Thermal properties of surface-
modified nano-al2o3/kevlar fiber/epoxy composites. Mater. Chem. Phys. 278, 125689 (2022)
11. Demircan, G., Kisa, M., Ozen, M., Acikgoz, A.: Quasi-static penetration behavior of glass-fiber-
reinforced epoxy nanocomposites. Mech. Compos. Mater. 57, 503–516 (2021). https://​doi.​org/​10.​
1007/​s11029-​021-​09973-y
12. Ghatage, P.S., Kar, V.R., Sudhagar, P.E.: On the numerical modelling and analysis of multi-directional func-
tionally graded composite structures: a review. Compos. Struct. 236,111837 (2020)
13. Polimeno, U., Meo, M., Almond, D., Angioni, S.: Detecting low velocity impact damage in compos-
ite plate using nonlinear acoustic/ultrasound methods. Appl. Compos. Mater. 17(5), 481–488 (2010).
https://​doi.​org/​10.​1007/​s10443-​010-​9168-5
14. Yu, B., Blanc, R., Soutis, C., Withers, P.: Evolution of damage during the fatigue of 3D woven
glass-fibre reinforced composites subjected to tension-tension loading observed by time-lapse
x-ray tomography. Compos. A: Appl. Sci. Manuf. 82, 279–290 (2016). https://​doi.​org/​10.​1016/j.​
compo​sitesa.​2015.​09.​001

13
 Applied Composite Materials

15. Lakhdar, M., Mohammed, D., Boudjemâa, L., Rabiâ, A., Bachir, M.: Damages detection in a com-
posite structure by vibration analysis. Energy Procedia 36, 888–897 (2013). https://​doi.​org/​10.​1016/j.​
egypro.​2013.​07.​102
16. Saeedifar, M., Zarouchas, D.: Damage characterization of laminated composites using acoustic emis-
sion: a review. Compos. Part B Eng. 195, 108039 (2020)
17. Hu, H., Wang, B.-T., Lee, C.-H., Su, J.-S.: Damage detection of surface cracks in composite laminates
using modal analysis and strain energy method. Compos. Struct. 74(4), 399–405 (2006). https://​doi.​
org/​10.​1016/j.​comps​truct.​2005.​04.​020
18. Leng, J., Asundi, A.: Structural health monitoring of smart composite materials by using EFPI and
FBG sensors. Sens. Actuators A Phys. 103(3), 330–340 (2003). https://​doi.​org/​10.​1016/​S0924-​
4247(02)​00429-6
19. De Luca, A., Perfetto, D., De Fenza, A., Petrone, G., Caputo, F.: Guided wave SHM system for dam-
age detection in complex composite structure. Theor. Appl. Fract. Mech. 105, 102408 (2020)
20. Bandara, S., Herath, M., Epaarachchi, J.: Sensory methods and machine learning based damage
identification of fibre-reinforced composite structures: an introductory review. J. Reinf. Plast. Com-
pos. 07316844221145972 (2022). https://​doi.​org/​10.​1177/​07316​84422​11459​72
21. Nasiri, S., Khosravani, M.R.: Applications of data-driven approaches in prediction of fatigue and
fracture. Mater. Today Commun. 33, 104437 (2022)
22. Luo, X.-L., Ye, J.-Y., Ma, P.-S., Zhang, L.-W.: Data-driven enhanced phase field models for highly accurate
prediction of mode I and mode II fracture. Comput. Methods Appl. Mech. Eng. 400, 115535 (2022)
23. Demircan, G., Kisa, M., Ozen, M., Aktas, B.: Surface-modified alumina nanoparticles-filled aramid
fiber-reinforced epoxy nanocomposites: preparation and mechanical properties. Iran. Polym. J. 29,
253–264 (2020). https://​doi.​org/​10.​1007/​s13726-​020-​00790-z
24. Cherkassky, V., Mulier, F.M.: Learning from data: concepts, theory, and methods. John Wiley & Sons (2007)
25. Khan, A., Kim, N., Shin, J.K., Kim, H.S., Youn, B.D.: Damage assessment of smart composite struc-
tures via machine learning: a review. JMST Advances 1(1), 107–124 (2019). https://​doi.​org/​10.​1007/​
s42791-​019-​0012-2
26. Nasiri, S., Khosravani, M.R.: Machine learning in predicting mechanical behavior of additively manufac-
tured parts. J. Mater. Res. Technol. 14, 1137–1153 (2021). https://​doi.​org/​10.​1016/j.​jmrt.​2021.​07.​004
27. Sutton, R.S., Barto, A.G.: Reinforcement learning: An introduction. MIT press (2018)
28. Avci, O., Abdeljaber, O., Kiranyaz, S., Hussein, M., Gabbouj, M., Inman, D.J.: A review of vibration-
based damage detection in civil structures: From traditional methods to machine learning and deep
learning applications. Mech. Syst. Signal Process. 147, 107077 (2021)
29. Dayhoff, J.E., DeLeo, J.M.: Artificial neural networks: opening the black box. Cancer: Interdiscipli-
nary International Journal of the American Cancer Society 91(S8), 1615–1635 (2001). https://​doi.​org/​
10.​1002/​1097-​0142(20010​415)​91:8+​<​1615::​AID-​CNCR1​175>3.​0.​CO;2-L
30. Ribeiro Junior, R.F., de Almeida, F.A., Gomes, G.F.: Fault classification in three-phase motors based
on vibration signal analysis and artificial neural networks. Neural Comput. Applic. 32(18), 15171–
15189 (2020). https://​doi.​org/​10.​1007/​s00521-​020-​04868-w
31. Wang, X., Makis, V.: Autoregressive model-based gear shaft fault diagnosis using the kolmogorov-
smirnov test. J. Sound Vib. 327(3–5), 413–423 (2009). https://​doi.​org/​10.​1016/j.​jsv.​2009.​07.​004
32. Akaike, H.: Autoregressive model fitting for control. In: Selected Papers of Hirotugu Akaike,
pp. 153–170. Springer (1998)
33. Regis, M., Serra, P., van den Heuvel, E.R.: Random autoregressive models: A structured overview.
Econ. Rev. 41(2), 207–230 (2022). https://​doi.​org/​10.​1080/​07474​938.​2021.​18995​04
34. John, G.H., Langley, P.: Estimating continuous distributions in Bayesian classifiers. Preprint at http://​
arxiv.​org/​abs/​1302.​4964 (2013). https://​doi.​org/​10.​48550/​arXiv.​1302.​4964
35. Keogh, E.J., Pazzani, M.J.: Learning the structure of augmented bayesian classifiers. Int. J. Artif.
Intell. Tools 11(04), 587–601 (2002). https://​doi.​org/​10.​1142/​S0218​21300​20010​52
36. Langley, P., Sage, S.: Induction of selective bayesian classifiers. In: Uncertainty Proceedings 1994,
pp. 399–406. Elsevier (1994)
37. Bielza, C., Larranaga, P.: Discrete Bayesian network classifiers: A survey. ACM Comput. Surv.
(CSUR) 47(1), 1–43 (2014). https://​doi.​org/​10.​1145/​25768​68
38. LeCun, Y., Bengio, Y., et al.: Convolutional networks for images, speech, and time series. The
Handbook of Brain Theory and Neural Networks 3361(10), 1995 (1995)
39. Li, Z., Liu, F., Yang, W., Peng, S., Zhou, J.: A survey of convolutional neural networks: analy-
sis, applications, and prospects. IEEE Trans. Neural Netw. Learn. Syst. (2021). https://​doi.​org/​10.​
1109/​TNNLS.​2021.​30848​27
40. Ribeiro Junior, R.F., dos Santos Areias, I.A., Campos, M.M., Teixeira, C.E., da Silva, L.E.B.,
Gomes, G.F.: Fault detection and diagnosis in electric motors using convolution neural network

13
Applied Composite Materials

and short-time Fourier transform. J. Vib. Eng. Technol. 1–12 (2022). https://​doi.​org/​10.​1007/​
s42417-​022-​00501-3
41. Junior, R.F.R., dos Santos Areias, I.A., Campos, M.M., Teixeira, C.E., da Silva, L.E.B., Gomes,
G.F.: Fault detection and diagnosis in electric motors using 1d convolutional neural networks
with multi-channel vibration signals. Measurement 190, 110759 (2022). https://​doi.​org/​10.​1016/j.​
measu​rement.​2022.​110759
42. Gangwar, A.K., Mahela, O.P., Rathore, B., Khan, B., Alhelou, H.H., Siano, P.: A novel k-means
clustering and weighted k-nn-regression-based fast transmission line protection. IEEE Trans. Ind.
Inf. 17(9), 6034–6043 (2020). https://​doi.​org/​10.​1109/​TII.​2020.​30378​69
43. Kumar, D., Fet, D.: Performance analysis of various data mining algorithms: A review. Int. J.
Comput. Appl. 32(6), 9–16 (2011)
44. Karegowda, A.G., Jayaram, M., Manjunath, A.: Cascading k-means clustering and k-nearest neighbor
classifier for categorization of diabetic patients. Int. J. Eng. Adv. Technol. 1(3), 147–151 (2012)
45. Richardson, M.: Principal component analysis. Aleš Hladnik Dr., Ass. Prof., Chair of Informa-
tion and Graphic Arts Technology, Faculty of Natural Sciences and Engineering, University of
Ljubljana, Slovenia [email protected], vol. 6, p. 16, 2009. http://​people.​maths.​ox.​ac.​uk/​
richa​rdsonm/​Signa​lProc​PCA.​pdf (last access: 3.5. 2013)
46. Abdi, H., Williams, L.J.: Principal component analysis. Wiley Interdiscip. Rev. Comput. Stat.
2(4), 433–459 (2010). https://​doi.​org/​10.​1002/​wics.​101
47. Teixeira, C.E., da Silva, L.E.B., Veloso, G.F., Lambert-Torres, G., Campos, M.M., Noronha, I.,
Bonaldi, E.L., de Oliveira, L.E.L.: An ultrasound-based water-cut meter for heavy fuel oil. Meas-
urement 148, 106907 (2019). https://​doi.​org/​10.​1016/j.​measu​rement.​2019.​106907
48. Mammone, A., Turchi, M., Cristianini, N.: Support vector machines. Wiley Interdiscip. Rev. Com-
put. Stat. 1(3), 283–289 (2009). https://​doi.​org/​10.​1002/​wics.​49
49. Wang, L.: Support vector machines: theory and applications, vol. 177. Springer Science & Busi-
ness Media (2005)
50. Dioşan, L., Rogozan, A., Pecuchet, J.-P.: Improving classification performance of support vector
machine by genetically optimising kernel shape and hyper-parameters. Appl. Intell. 36(2), 280–
294 (2012). https://​doi.​org/​10.​1007/​s10489-​010-​0260-1
51. García-Gonzalo, E., Fernández-Muñiz, Z., García Nieto, P.J., Bernardo Sánchez, A., Menéndez
Fernández, M.: Hard-rock stability analysis for span design in entry-type excavations with learn-
ing classifiers. Materials 9(7), 531 (2016). https://​doi.​org/​10.​3390/​ma907​0531
52. Qing, X., Liao, Y., Wang, Y., Chen, B., Zhang, F., Wang, Y.: Machine learning based quantita-
tive damage monitoring of composite structure. Int. J. Smart Nano Mater. 13(2), 167–202 (2022).
https://​doi.​org/​10.​1080/​19475​411.​2022.​20548​78
53. Ramasamy, P., Sampathkumar, S.: Prediction of impact damage tolerance of drop impacted wgfrp
composite by artificial neural network using acoustic emission parameters. Compos. Part B Eng.
60, 457–462 (2014). https://​doi.​org/​10.​1016/j.​compo​sitesb.​2013.​12.​028
54. Ramasamy, P., Sampathkumar, S.: Offline and online investigation of drop impact damage on
GFRP composite using non-destructive data by artificial neural network. Iran. J. Sci. Technol. -
Trans. Mech. Eng. 39(M1), 29 (2015)
55. Khatir, S., Tiachacht, S., Thanh, C.-L., Bui, T.Q., Wahab, M.A.: Damage assessment in composite
laminates using ANN-PSO-IGA and Cornwell indicator. Compos. Struct. 230, 111509 (2019)
56. Qian, C., Ran, Y., He, J., Ren, Y., Sun, B., Zhang, W., Wang, R.: Application of artificial neural net-
works for quantitative damage detection in unidirectional composite structures based on lamb waves.
Adv. Mech. Eng. 12(3), 1687814020914732 (2020). https://​doi.​org/​10.​1177/​16878​14020​914732
57. Califano, A., Chandarana, N., Grassia, L., D’Amore, A., Soutis, C.: Damage detection in compos-
ites by artificial neural networks trained by using in situ distributed strains. Appl. Compos. Mater.
27(5), 657–671 (2020). https://​doi.​org/​10.​1007/​s10443-​020-​09829-z
58. Zenzen, R., Khatir, S., Belaidi, I., Le Thanh, C., Wahab, M.A.: A modified transmissibility indica-
tor and artificial neural network for damage identification and quantification in laminated compos-
ite structures. Compos. Struct. 248, 112497 (2020)
59. Mardanshahi, A., Mardanshahi, M., Izadi, A.: Damage quantification in foam core sandwich com-
posites via finite element model updating and artificial neural networks. Proc. Inst. Mech. Eng. C J.
Mech. Eng. Sci. 234(21), 4288–4304 (2020). https://​doi.​org/​10.​1177/​09544​06220​92120
60. Tan, Z.X., Thambiratnam, D.P., Chan, T.H., Gordan, M., Abdul Razak, H.: Damage detection in
steel-concrete composite bridge using vibration characteristics and artificial neural network. Struct.
Infrastruct. Eng. 16(9), 1247–1261 (2020). https://​doi.​org/​10.​1080/​15732​479.​2019.​16963​78

13
 Applied Composite Materials

61. Jang, Y.-J., Kim, H.-J., Kim, H.-G., Kang, K.-W.: Identification of debonding damage at spar cap-
shear web joints by artificial neural network using natural frequency relevant key features of compos-
ite wind turbine blades. Appl. Sci. 11(12), 5327 (2021). https://​doi.​org/​10.​3390/​app11​125327
62. Khatir, S., Tiachacht, S., Le Thanh, C., Ghandourah, E., Mirjalili, S., Wahab, M.A.: An improved
artificial neural network using arithmetic optimization algorithm for damage assessment in FGM
composite plates. Compos. Struct. 273, 114287 (2021)
63. Tran-Ngoc, H., Khatir, S., Ho-Khac, H., De Roeck, G., Bui-Tien, T., Wahab, M.A.: Efficient artifi-
cial neural networks based on a hybrid metaheuristic optimization algorithm for damage detection in
laminated composite structures. Compos. Struct. 262, 113339 (2021)
64. Nasser, J., Groo, L., Sodano, H.: Artificial neural networks and phenomenological degradation mod-
els for fatigue damage tracking and life prediction in laser induced graphene interlayered fiberglass
composites. Smart Mater. Struct. 30(8), 085010 (2021)
65. Al-Athel, K.S., Alhasan, M.M., Alomari, A.S., Arif, A.F.M.: Damage characterization of embedded
defects in composites using a hybrid thermography, computational, and artificial neural networks
approach. Heliyon 8(8), e10063 (2022). https://​doi.​org/​10.​1016/j.​heliy​on.​2022.​e10063
66. Feng, B., Ribeiro, A.L., Pasadas, D.J., Ramos, H.G.: Locating low velocity impacts on a composite
plate using multi-frequency image fusion and artificial neural network. J. Nondestruct. Eval. 41(2),
1–9 (2022). https://​doi.​org/​10.​1007/​s10921-​022-​00865-2
67. Saadatmorad, M., Jafari-Talookolaei, R.-A., Pashaei, M.-H., Khatir, S.: Damage detection in rectan-
gular laminated composite plate structures using a combination of wavelet transforms and artificial
neural networks. J. Vib. Eng. Technol. 1–18 (2022). https://​doi.​org/​10.​1007/​s42417-​022-​00471-6
68. Reis, P.A., Iwasaki, K.M., Voltz, L.R., Cardoso, E.L., Medeiros, R.D.: Damage detection of com-
posite beams using vibration response and artificial neural networks. Proceedings of the Institution
of Mechanical Engineers, Part L: Journal of Materials: Design and Applications 236(7), 1419–1430
(2022). https://​doi.​org/​10.​1177/​14644​20721​10413​26
69. Mouzakis, D.E., Dimogianopoulos, D.G., Zaoutsos, S.: Damage assessment of carbon fiber rein-
forced composites under accelerated aging and validation via stochastic model-based analysis. Int. J.
Damage Mech. 23(5), 702–726 (2014). https://​doi.​org/​10.​1177/​10567​89513​508798
70. Nardi, D., Lampani, L., Pasquali, M., Gaudenzi, P.: Detection of low-velocity impact-induced delami-
nations in composite laminates using auto-regressive models. Compos. Struct. 151, 108–113 (2016).
https://​doi.​org/​10.​1016/j.​comps​truct.​2016.​02.​005
71. da Silva, S., Paixão, J., Rébillat, M., Mechbal, N.: Extrapolation of ar models using cubic splines for
damage progression evaluation in composite structures. J. Intell. Mater. Syst. Struct. 32(3), 284–295
(2021). https://​doi.​org/​10.​1177/​10453​89X20​963171
72. Paixão, J., da Silva, S., Figueiredo, E., Radu, L., Park, G.: Delamination area quantification in com-
posite structures using Gaussian process regression and auto-regressive models. J. Vib. Control.
27(23–24), 2778–2792 (2021). https://​doi.​org/​10.​1177/​10775​46320​966183
73. Peng, T., Saxena, A., Goebel, K., Xiang, Y., Sankararaman, S., Liu, Y.: A novel bayesian imaging
method for probabilistic delamination detection of composite materials. Smart Mater. Struct. 22(12),
125019 (2013)
74. Fendzi, C., Mechbal, N., Rebillat, M., Guskov, M., Coffignal, G.: A general bayesian framework
for ellipse-based and hyperbola-based damage localization in anisotropic composite plates. J. Intell.
Mater. Syst. Struct. 27(3), 350–374 (2016). https://​doi.​org/​10.​1177/​10453​89X15​571383
75. Cantero-Chinchilla, S., Malik, M.K., Chronopoulos, D., Chiachio, J.: Bayesian damage localization
and identification based on a transient wave propagation model for composite beam structures. Com-
pos. Struct. 267, 113849 (2021)
76. Huo, H., He, J., Guan, X.: A bayesian fusion method for composite damage identification using lamb
wave. Struct. Health Monit. 1475921720945000 (2020). https://​doi.​org/​10.​1177/​14759​21720​945000
77. Guo, Y., Xiao, Z., Geng, L., Wu, J., Zhang, F., Liu, Y., Wang, W.: Fully convolutional neural network
with gru for 3d braided composite material flaw detection. IEEE Access 7, 151180–151188 (2019).
https://​doi.​org/​10.​1109/​ACCESS.​2019.​29464​47
78. Nasiri, A., Bao, J., Mccleeary, D., Louis, S.-Y.M., Huang, X., Hu, J.: Online damage monitoring of
sicf-sicm composite materials using acoustic emission and deep learning. IEEE Access 7, 140534–
140541 (2019). https://​doi.​org/​10.​1109/​ACCESS.​2019.​29432​10
79. Tabian, I., Fu, H.: and Z. Sharif Khodaei, A convolutional neural network for impact detection and
characterization of complex composite structures. Sensors 19(22), 4933 (2019). https://​doi.​org/​10.​
3390/​s1922​4933
80. Salehzadeh Nobari, A.E., Aliabadi, M.F.: A multilevel isolation forrest and convolutional neural net-
work algorithm for impact characterization on composite structures. Sensors 20(20), 5896 (2020).
https://​doi.​org/​10.​3390/​s2020​5896

13
Applied Composite Materials

81. Das, A.K., Leung, C.K., Wan, K.T.: Application of deep convolutional neural networks for auto-
mated and rapid identification and computation of crack statistics of thin cracks in strain harden-
ing cementitious composites (SHCCS). Cem. Concr. Compos. 122, 104159 (2021)
82. Saadatmorad, M., Jafari-Talookolaei, R.-A., Pashaei, M.-H., Khatir, S.: Damage detection on rec-
tangular laminated composite plates using wavelet based convolutional neural network technique.
Compos. Struct. 278, 114656 (2021)
83. Wu, J., Xu, X., Liu, C., Deng, C., Shao, X.: Lamb wave-based damage detection of composite
structures using deep convolutional neural network and continuous wavelet transform. Compos.
Struct. 276, 114590 (2021)
84. Djavadifar, A., Graham-Knight, J.B., Korber, M., Lasserre, P., Najjaran, H.: Automated visual
detection of geometrical defects in composite manufacturing processes using deep convolutional
neural networks. J. Intell. Manuf. 1–19 (2021). https://​doi.​org/​10.​1007/​s10845-​021-​01776-1
85. Augustin, M., Ramesh, V., Prasad, R.K., Gupta, N., Kumar, M.R.: Detection of inclusion by using
3D laser scanner in composite prepreg manufacturing technique using convolutional neural net-
works. Mach. Vis. Appl. 32(6), 1–10 (2021). https://​doi.​org/​10.​1007/​s00138-​021-​01241-2
86. Miller, B., Ziemiański, L.: Detection of material degradation of a composite cylinder using mode
shapes and convolutional neural networks. Materials 14(21), 6686 (2021). https://​doi.​org/​10.​3390/​
ma142​16686
87. Azuara, G., Ruiz, M., Barrera, E.: Damage localization in composite plates using wavelet trans-
form and 2D convolutional neural networks. Sensors 21(17), 5825 (2021). https://​doi.​org/​10.​3390/​
s2117​5825
88. Machado, J.M., Tavares, J.M.R., Camanho, P.P., Correia, N.: Automatic void content assessment
of composite laminates using a machine-learning approach. Compos. Struct. 288, 115383 (2022)
89. Wang, Y., Luo, Q., Xie, H., Li, Q., Sun, G.: Digital image correlation (DIC) based damage detec-
tion for CFRP laminates by using machine learning based image semantic segmentation. Int. J.
Mech. Sci. 230, 107529 (2022)
90. Cui, R., Azuara, G., Lanza di Scalea, F., Barrera, E.: Damage imaging in skin-stringer composite air-
craft panel by ultrasonic-guided waves using deep learning with convolutional neural network. Struct.
Health Monit. 21(3), 1123–1138 (2022). https://​doi.​org/​10.​1177/​14759​21721​10239​34
91. Yang, J., Su, Y., He, Y., Zhou, P., Xu, L., Su, Z.: Machine learning-enabled resolution-lossless
tomography for composite structures with a restricted sensing capability. Ultrasonics 125, 106801
(2022). https://​doi.​org/​10.​1016/j.​ultras.​2022.​106801
92. Helwing, R., Hülsbusch, D., Walther, F.: Deep learning method for analysis and segmentation of
fatigue damage in x-ray computed tomography data for fiber-reinforced polymers. Compos. Sci.
Technol. 230, 109781 (2022)
93. Barile, C., Casavola, C., Pappalettera, G., Kannan, V.P.: Damage monitoring of carbon fibre
reinforced polymer composites using acoustic emission technique and deep learning. Compos.
Struct. 292, 115629 (2022)
94. Barile, C., Casavola, C., Pappalettera, G., Kannan, V.P.: Laplacian score and k-means data clus-
tering for damage characterization of adhesively bonded CFRP composites by means of acoustic
emission technique. Appl. Acoust. 185, 108425 (2022)
95. Pashmforoush, F., Khamedi, R., Fotouhi, M., Hajikhani, M., Ahmadi, M.: Damage classification
of sandwich composites using acoustic emission technique and k-means genetic algorithm. J. Non-
destruct. Eval. 33(4), 481–492 (2014). https://​doi.​org/​10.​1007/​s10921-​014-​0243-y
96. Liu, P., Xu, D., Li, J., Chen, Z., Wang, S., Leng, J., Zhu, R., Jiao, L., Liu, W., Li, Z.: Damage
mode identification of composite wind turbine blade under accelerated fatigue loads using acous-
tic emission and machine learning. Struct. Health Monit. 19(4), 1092–1103 (2020). https://​doi.​org/​
10.​1177/​14759​21719​878259
97. Shrifan, N.H., Jawad, G.N., Isa, N.A.M., Akbar, M.F.: Microwave nondestructive testing for
defect detection in composites based on k-means clustering algorithm. IEEE Access 9, 4820–4828
(2020). https://​doi.​org/​10.​1109/​ACCESS.​2020.​30481​47
98. Ech-Choudany, Y., Assarar, M., Scida, D., Morain-Nicolier, F., Bellach, B.: Unsupervised clustering for
building a learning database of acoustic emission signals to identify damage mechanisms in unidirec-
tional laminates. Appl. Acoust. 123, 123–132 (2017). https://​doi.​org/​10.​1016/j.​apaco​ust.​2017.​03.​008
99. Zeng, X., Shao, H., Pan, R., Wang, B., Deng, Q., Zhang, C., Suo, T.: Real-time damage analysis of
2D c/sic composite based on spectral characters of acoustic emission signals using pattern recogni-
tion. Acta Mech. Sinica 38(10), 1–18 (2022). https://​doi.​org/​10.​1007/​s10409-​022-​22177-x
100. Essassi, K., Rebiere, J.-L., Mahi, A.E., Amine Ben souf, M., Bouguecha, A., Haddar, M.: Health moni-
toring of sandwich composites with auxetic core subjected to indentation tests using acoustic emis-
sion. Struct. Health Monit. 21(5), 2264–2275 (2022). https://​doi.​org/​10.​1177/​14759​21721​10539​91

13
 Applied Composite Materials

101. Taghizadeh, J., Ahmadi, M.: Identification of damage modes in polypropylene/epoxy composites by
using principal component analysis on ae signals extracted from mode i delamination. Nondestruct.
Test. Eval. 27(2), 151–170 (2012). https://​doi.​org/​10.​1080/​10589​759.​2011.​633607
102. Rao, A.R.M., Lakshmi, K., Kumar, S.K.: Detection of delamination in laminated composites with
limited measurements combining pca and dynamic QPSO. Adv. Eng. Softw. 86, 85–106 (2015).
https://​doi.​org/​10.​1016/j.​adven​gsoft.​2015.​04.​005
103. Sierra-Pérez, J., Güemes, A., Mujica, L.E., Ruiz, M.: Damage detection in composite materials struc-
tures under variable loads conditions by using fiber bragg gratings and principal component analy-
sis, involving new unfolding and scaling methods. J. Intell. Mater. Syst. Struct. 26(11), 1346–1359
(2015). https://​doi.​org/​10.​1177/​10453​89X14​541493
104. Yan, Z., Chen, C.-Y., Luo, L., Yao, Y.: Stable principal component pursuit-based thermographic data
analysis for defect detection in polymer composites. J. Process Control 49, 36–44 (2017). https://​doi.​
org/​10.​1016/j.​jproc​ont.​2016.​11.​008
105. Mohamad, Z.: Classification for damage severity in natural fibre composites using principal compo-
nent analysis. Int. J. Integr. Eng. 10(8) (2018). https://​doi.​org/​10.​30880/​ijie.​2018.​10.​08.​023
106. Lu, S., Jiang, M., Wang, X., Yu, H.: Damage detection method of cfrp structure based on fiber bragg
grating and principal component analysis. Optik 178, 858–867 (2019). https://​doi.​org/​10.​1016/j.​ijleo.​
2018.​10.​055
107. Malinowski, P.H., Wandowski, T., Singh, S.K.: Employing principal component analysis for assess-
ment of damage in GFRP composites using electromechanical impedance. Compos. Struct. 266,
113820 (2021)
108. Lu, S., Dong, H., Yu, H.: Interlaminar damage assessment method of CFRP laminate based on syn-
chrosqueezed wavelet transform and ensemble principal component analysis. Compos. Struct. 276,
114581 (2021)
109. Fredo, A.J., Abilash, R., Femi, R., Mythili, A., Kumar, C.S.: Classification of damages in compos-
ite images using zernike moments and support vector machines. Compos. Part B Eng. 168, 77–86
(2019). https://​doi.​org/​10.​1016/j.​compo​sitesb.​2018.​12.​064
110. Forero-Ramírez, J.-C., Restrepo-Girón, A.-D., Nope-Rodríguez, S.-E.: Detection of internal
defects in carbon fiber reinforced plastic slabs using background thermal compensation by filter-
ing and support vector machines. J. Nondestruct. Eval. 38(1), 1–11 (2019). https://​doi.​org/​10.​1007/​
s10921-​019-​0569-6
111. Shyamala, P., Mondal, S., Chakraborty, S.: Numerical and experimental investigation for dam-
age detection in FRP composite plates using support vector machine algorithm. Struct. Monit.
Maint. 5(2), 243–260 (2018). https://​doi.​org/​10.​12989/​smm.​2018.5.​2.​243
112. Gillespie, D.I., Hamilton, A.W., Atkinson, R.C., Bellekens, X., Michie, C., Andonovic, I., Tachtatzis, C.:
Composite laminate delamination detection using transient thermal conduction profiles and machine learn-
ing based data analysis. Sensors 20(24), 7227 (2020). https://​doi.​org/​10.​3390/​s2024​7227
113. Yue, J., Wang, Y., Beskos, D.: Uniaxial tension damage mechanics of steel fiber reinforced concrete
using acoustic emission and machine learning crack mode classification. Cem. Concr. Compos. 123,
104205 (2021)
114. Xu, J., Liu, X., Han, Q., Wang, W.: A particle swarm optimization-support vector machine hybrid
system with acoustic emission on damage degree judgment of carbon fiber reinforced polymer cables.
Struct. Health Monit. 20(4), 1551–1562 (2021). https://​doi.​org/​10.​1177/​14759​21720​922824
115. Sheng, W., Liu, Y., Söffker, D.: A novel adaptive boosting algorithm with distance-based weighted least
square support vector machine and filter factor for carbon fiber reinforced polymer multi-damage classifica-
tion. Struct. Health Monit. 4759217221098173 (2022). https://​doi.​org/​10.​1177/​14759​21722​10981​73
116. Rajiv, B., Kalos, P., Pantawane, P., Chougule, V., Chavan, V.: Classification of damages in composite
material using multi-support vector machine. J. Inst. Eng. (India) C. 1–7 (2022). https://​doi.​org/​10.​
1007/​s40032-​022-​00811-1
117. Alhammad, M., Avdelidis, N.P., Ibarra-Castanedo, C., Torbali, M.E., Genest, M., Zhang, H., Zolotas,
A., Maldgue, X.P.: Automated impact damage detection technique for composites based on thermo-
graphic image processing and machine learning classification. Sensors 22(23), 9031 (2022). https://​
doi.​org/​10.​3390/​s2223​9031
118. Pashmforoush, F., Fotouhi, M., Ahmadi, M.: Acoustic emission-based damage classification of glass/
polyester composites using harmony search k-means algorithm. J. Reinf. Plast. Compos. 31(10), 671–
680 (2012). https://​doi.​org/​10.​1177/​07316​84412​44225
119. Selva, P., Cherrier, O., Budinger, V., Lachaud, F., Morlier, J.: Smart monitoring of aeronautical com-
posites plates based on electromechanical impedance measurements and artificial neural networks.
Eng. Struct. 56, 794–804 (2013). https://​doi.​org/​10.​1016/j.​engst​ruct.​2013.​05.​025

13
Applied Composite Materials

120. Crivelli, D., Guagliano, M., Monici, A.: Development of an artificial neural network processing tech-
nique for the analysis of damage evolution in pultruded composites with acoustic emission. Compos.
Part B Eng. 56, 948–959 (2014). https://​doi.​org/​10.​1016/j.​compo​sitesb.​2013.​09.​005
121. Mardanshahi, A., Nasir, V., Kazemirad, S., Shokrieh, M.: Detection and classification of matrix
cracking in laminated composites using guided wave propagation and artificial neural networks.
Compos. Struct. 246, 112403 (2020)
122. Chen, G.L., Yanamandra, K., Gupta, N.: Artificial neural networks framework for detection of defects
in 3D-printed fiber reinforcement composites. JOM 73(7), 2075–2084 (2021). https://​doi.​org/​10.​
1007/​s11837-​021-​04708-9
123. Hamdi, K., Moreau, G., Aboura, Z.: Digital image correlation, acoustic emission and in-situ microscopy in
order to understand composite compression damage behavior. Compos. Struct. 258, 113424 (2021)
124. Dziendzikowski, M., Heesch, M., Gorski, J., Dragan, K., Dworakowski, Z.: Application of PZT
ceramic sensors for composite structure monitoring using harmonic excitation signals and bayesian
classification approach. Materials 14(19), 5468 (2021). https://​doi.​org/​10.​3390/​ma141​95468
125. Scholz, V., Winkler, P., Hornig, A., Gude, M., Filippatos, A.: Structural damage identification of
composite rotors based on fully connected neural networks and convolutional neural networks. Sen-
sors 21(6), 2005 (2021). https://​doi.​org/​10.​3390/​s2106​2005
126. Lee, I.Y., Roh, H.D., Park, H.W., Park, Y.-B.: Advanced non-destructive evaluation of impact damage
growth in carbon-fiber-reinforced plastic by electromechanical analysis and machine learning cluster-
ing. Compos. Sci. Technol. 218, 109094 (2022)

Publisher’s Note Springer Nature remains neutral with regard to jurisdictional claims in published maps and
institutional affiliations.

Springer Nature or its licensor (e.g. a society or other partner) holds exclusive rights to this article under a
publishing agreement with the author(s) or other rightsholder(s); author self-archiving of the accepted manuscript
version of this article is solely governed by the terms of such publishing agreement and applicable law.

13