Recently, different industry sectors have focused on producing small-scale products that are multifunctional and efficient. To achieve those miniature products, composite materials have gained popularity among others as they offer high stiffness, high resistance to corrosion, and toughness while being lightweight [1,2]. As the name suggests, composite materials are made of two or more distinct materials on the macroscopic level. One of the materials acts as reinforcement, which is stiffer and stronger, whereas the other plays the role of a matrix, which is less stiff and weaker [3,4]. The demand for composite materials is high in aerospace [5,6], automotive industries [7,8], biomedical applications [9,10], construction materials [11], and wind turbines [12], to name a few. Composite materials offer more advantages than conventional materials, such as less weight, higher strength and stiffness, and resistance to corrosion and other chemical substances. However, the particularity of the manufacturing process and the very expensive materials required to make these materials are their main weaknesses. The complex manufacturing process of composite materials usually comes with internal flaws that are not easily detected by visual inspection. Some defects are due to human factors such as poor material handling, and environmental factors (impurities, temperatures, etc.) [13,14].

Laminated composite materials are made up of more than two layers stacked together, and each layer is called a lamina. The occurrence of damage mechanisms in composite laminates is categorized into interlaminar and intralaminar damages [15]. Delamination is the most dominant form of failure in composite materials, and it mostly happens due to high interlaminar stresses between two lamina [16,17]. Other common types of failures are within a lamina itself, such as matrix cracking, fibre pull-out, fibre breakage, and fibre/matrix debonding [18–20]. Since composite materials are prone to failures that are hardly detected, a routine inspection should be performed to examine the possible damages during their initiation stage using appropriate techniques such as nondestructive testing and evaluation (NDT&E). Among NDT&E and structural health monitoring (SHM) techniques, the common methods used are ultrasonic waves, acoustic emission, strain measurement, vibration, thermography, and machine vision, to name a few [21–26]. The list of representative articles that adopt the above methods for damage detection is listed in Table 1. For brevity and accuracy, the remainder of this paper will focus on Lamb waves, particularly guided waves, for damage detection.

At least two decades ago, enormous research about damage localization and quantification based on Lamb wave was conducted. Sun et al. [27] proposed a data-driven method based on the Lamb wave for damage quantification using a genetic algorithm and least squares support vector machine (LS-SVM). The damage was estimated based on three features: normalized amplitude, phase changes, and correlation coefficient of Lamb wave changes. Their proposed method was proven to predict the crack size at the lamp joint. Eremin et al. [28] proposed the idea of using scattering resonance frequencies that are sensitive to the size of the crack and depth variation to express a non-destructive estimation of the crack severity. This method was based on matching the resonance frequencies obtained both experimentally and numerically. Nicassio et al. [29] presented a novel method based on nonlinear Lamb wave characteristics and local defect resonance for locating and evaluating single-lap joint disbond. In this work, the damage location was estimated based on the time of flight (ToF) of each damage signal and the total probability function. Qiang et al. proposed a new method for crack localization on beam structures using piezoelectric sensors. In this work, the crack damage and localization were determined by the presence of both antisymmetric and symmetric modes (A_0/S_0). Most of the work done for damage detection has reported that the Lamb wave generates multiple reflections when interacting with the damage. To overcome those complications, parametric and non-parametric methods were proposed for damage detection. However, the parametric method requires a full understanding of physics and mathematical modeling, which is not easily obtained for complex models. On the other side, the non-parametric method known as the data-driven approach is widely used since it is easily applicable to complex structures [30].

The data driven SHM approach requires baseline data from the pristine structure and data from the damaged structure. Data-driven modeling based on statistical models combined with classical machine learning has proven that they can’t cope with complexity of the generated Lamb wave signals to detect and localize the induced damages [31]. Moreover, the conventional method requires extra skills for the professionals to extract the important features for pattern recognition, and most of conventional methods will fail to detect the damage in case a new pattern is introduced [32]. The current advancement of artificial intelligence (AI) comes with a need for a solution in the field of SHM, particularly deep learning, which needs a large volume of data and a complex algorithm to deal with those abovementioned complexities. As aforementioned, the classical machine learning algorithms were not powerful enough to analyze the complexities relying on the large dataset, but with deep learning, the complex relationship between healthy and damaged signal features can be easily mapped and classified [33,34].

Much research has been conducted on applying deep learning based on vibration signals for SHM of structures, and many related review papers have been reported. However, review papers related to research applying Lamb wave signals to deep learning have not been reported. This review paper is going to focus on the application of Lamb waves in SHM, with an emphasis on delamination and matrix crack identification of composite materials. Moreover, this paper is going to explore the current findings of deep learning algorithms applied to structure health monitoring to assist in detecting and localizing the damages.

The cited references in this work were gathered from Scopus and Web of Science databases, covering the period from 2018 to 2024, using the following search criteria: ((“Structural health monitoring”) AND (“Damage detection” OR “Damage localization” OR “Damage quantification”) AND (“Delamination” OR “Matrix crack”) AND (“Composite materials” OR “Composite plate” OR “CFRP”) AND (“Machine learning” OR “Deep learning”) AND (“Lamb waves” OR “Guided waves”)). Fig. 1 shows a steady increase in the number of published articles over the past six years, indicating the growing appeal of SHM to researchers. To analyze the co-occurrence of keywords and identify patterns among frequently used terms, VOSviewer was utilized. This analysis highlights areas of intensive research activity. A bibliometric map was generated with a minimum threshold of 15 occurrences per keyword. Out of 3,906 keywords, only 43 met this threshold. The keyword co-occurrence map, shown in Fig. 2, comprises 6 clusters and 1,384 links between keywords. The dominant keywords for each cluster are ‘structural health monitoring’, ‘machine learning’, ‘damage identification’, ‘acoustic emission’, ‘deep learning’, and ‘composite materials’, respectively.

Lamb waves are a special type of ultrasonic guided wave, and they are often generated in a thin plate structure with traction-free boundaries. Based on the vibration displacement of ultrasonic particles, Lamb waves have two distinct propagation modes, namely symmetric (Sn) and antisymmetric (An) as shown in Figs. 3(a) and 3(b) respectively [35]. The propagation of Lamb waves is highly dispersive, and their phase velocity changes with frequency and material thickness [36]. Apart from the dispersive curve, another important characteristic of Lamb wave is the group velocity curve. The group velocity comes in handy when examining the wave-packets travel [37,38]. Fig. 4 clearly illustrates that as the value of (f.h) product increase, both the phase and group velocities asymptotically approach the wave propagation velocity of Lamb wave. In addition, the first propagation wave modes (S0 and A0) can be obtained at low frequencies, and the symmetric mode (S0) has a higher group velocity than the antisymmetric mode (A0). To understand clearly the existing modes, Sn and An, the analytical dispersion curves and group velocities are presented as follow [39]:

with

p and q are placeholder variables, ω, h, C_L, C_T, λ, k are the wave angular frequency, plate half thickness, longitudinal velocity, transverse velocity, wavelength, and wave number, respectively.

The longitudinal and transversal velocities usually depend on the properties of material and are given by:

Where, E and v represent the young’s modulus and Poisson’s ratio, and ρ is the density of the material.

The conceptual diagram of Lamb wave based SHM is shown in Fig. 5. Lamb wave-based SHM was proven to be effective in terms of damage detection due to its actuation simplicity, ability to travel larger areas with small attenuation, and sensitivity to minor damages [24,40,41]. The Lamb wave signals obtained via the stated methods in Table 1, can be further processed under feature extraction to identify, localize, and quantify the material’s damage. Different classical damage detection algorithms for SHM are provided in Table 2 [42–52].

AI is mainly concerned with building systems that can imitate human behavior. Under artificial intelligence, machine learning as a subfield is also mostly known for its capacity to make the computer learn from the data and decide without being explicitly programmed. From Fig. 6, at the core of AI, there is deep learning, which is also subset of machine learning and deep learning is currently the most powerful and practical model among others for its self-learning capabilities.

Machine learning technique is aimed at developing algorithms that can learn from experimentally recorded or simulated data to predict the system’s future performance [53,54]. Machine learning is mostly used in different applications such as text mining, image classification, matching news items, and object identification, to name a few. Machine learning is usually divided into three categories namely supervised learning, unsupervised learning and reinforcement learning. Due to the lack of ability to analyze raw natural and complex data of classical machine learning techniques, a deep learning method has been introduced to solve the reliance on manual feature extraction in classical machine learning models.

Deep learning is a subfield of machine learning that can extract important features immediately from raw data without human intervention. Moreover, deep learning can be interpreted as a representation learning method with multiple levels of representation by which the data is fed through stacked layers of a deep neural network (DNN), and each layer is capable of extracting features gradually and inputting them to the next layer without human intervention [55–58]. There are three major classifications of deep learning techniques: supervised, unsupervised, and reinforcement learning. The supervised deep learning-based SHM uses the labeled data to train the model. This technique requires a collection of inputs with their corresponding outputs (x_i, y_i) [59]. Contrary to the supervised deep learning, the unsupervised deep learning algorithm uses unlabeled data and learns the recurring patterns. The training dataset of unsupervised model contains only x_i variables and its main objective is to find the data with common features and classify them under groups or clusters. Also, it has the ability of labelling the data based on the fundamental algorithm structure [57,58,60]. Finally, reinforcement learning learns the optimal behavior in an environment to get the utmost reward. The reward is maximized by a trial-and-error search. As we mentioned before, traditional machine learning has computational limitations when dealing with a large dataset and that issue can be overcome by deep learning. Deep learning-based SHM data are collected and analyzed either as 1D time-series datasets or 2D data such as images [61–63]. The only key challenge for the researchers in deep learning is the substantial computational resources and larger data required to train the model effectively. A detailed configuration of the deep learning structure for both types of datasets is illustrated in Fig. 7. Convolutional Neural Network (CNN) relies on the concept of mathematical principles for convolution, rectified linear step, the pooling step, flattening step and the full connected layer.

As mentioned earlier, interlaminar delamination is a common type of damage in laminated composite materials. The basic concepts of interlaminar damage and intralaminar damage in laminated composites are presented in Fig. 10. Its presence has a direct impact on the stiffness and strength of the material, which may lead to catastrophic failure in the absence of earlier detection. The initiation and evolution of interlaminar and intralaminar damages was studied by [70–72]. Most of the reviewed papers in this section have used a data-driven approach rather than physics-based approach for delamination assessment due to the availability of high-performance computational equipment and powerful machine learning algorithms such as deep learning to process the big data. Asif Khan et al. [73] proposed a synthetic data augmentation strategy to cope with the issue of scarce data for deep learning assessment of delamination in stuck laminated composites. Gao et al. [74] suggested a damage localization method based on Lamb wave and modular artificial neural network (M-ANN) and the proposed model exhibited a lower MSE, MAPE, MAE and high R². The pre-augmented and augmented datasets were trained via transfer learning, and the customized deep learning training accuracy for augmented data outperformed the k-nearest neighbor (KNN), linear discriminant analysis (LDA), and SVM classifiers trained on the original dataset. In SHM procedures, collecting data for all damaged scenarios is very cumbersome and time-consuming. Rautela et al. [75] proposed two different approaches for unsupervised learning where the data were trained based on the baseline situations only. On the one hand, they proposed a combination of dimension reduction techniques, principal component analysis, and one-class support vector machines (PCA and ocSVM) and on the other hand, they utilized a deep learning-based deep convolutional autoencoder (CAE) to train the unsupervised data for delamination detection. The deep learning-based approach has demonstrated higher accuracy than the machine learning-based approach. Despite its high accuracy in detecting delamination, the machine learning-based method requires less computational time and a smaller amount of data than the deep learning method. Wu et al. [76] suggested a novel method for detecting and localizing the internal delamination of carbon fiber reinforced plastic (CFRP) through a deep convolutional neural network and continuous wavelet transform. The trained data was collected via lead zirconate titanate (PZT) sensor arrays, and the test data was obtained using X-ray to validate the data. The proposed method achieved high accuracy under a low actuation frequency of 150 kHz, and it was able to diagnose the damage through automatic feature extraction. However, as the number of cycles increased, the proposed method was not able to locate all the sensing paths.

Naturally, CNN was built to deal with complex dataset such 2D and 3D images without a lot of transformation, which can cause the loss of temporal information, and most literatures has proven that 2D-CNN has a significant advantage in terms of generalization and robustness over 1D-CNN. Sikdar et al [77] designed a deep learning architecture that uses CNN to automatically extract the discrete image features. The raw guided wave signals were converted into images of time-frequency scalograms using the continuous wavelet transform. For achieving efficient data compression while conserving important information, Liao et al. [78] proposed a novel diagnosis method, integrating guided wave-based SHM with gramian angular field (GAF) image, and convolutional neural network. The proposed method performed well on damage localization than other compared methods. However, the authors are not sure whether the proposed model will perform well if environmental factors are considered. Moreover, the model’s performance was based on single damage rather than multiple damages.

Apart from using sensors embedded in structures for detecting damages and localizing a delamination, a pixel-wise segmentation model based on a fully convolutional neural network (FCN) was proposed [32]. The proposed method was capable of being trained end-to-end and pixel-to-pixel without a feature extraction process. In addition, pixel-wise segmentation was introduced to allow the shape precision and size estimation of the defect. The performance of the FCN method was compared to the conventional approach, which is based on wavenumber filtering using intersection-over-union, as shown in Fig. 11. Although the full wavefield using FCN can successively detect delamination and its size estimation, the time required to generate the data is unbearable, and since the full wavefield data is generated by Scanning Laser Doppler Vibrometer (SLDV), the proposed method is more suitable in the NDT field than SHM. Also, Ullah et al. [79] proposed an end-to-end deep learning model of many-to-one sequence prediction to perform pixel-wise image segmentation. The proposed model was able to predict a single delamination but poorly performed on multi-delaminated CFRP. In addition, the SLDV method showed its incapability of sensing damage that is situated far away from the bottom side of the specimen. Mechanoluminescent (ML) sensors respond to mechanical deformation by emitting light and providing the entire visual information of the entire structure, as proposed by Park et al. [80]. To detect the delamination from the images emitted via ML sensor, a convolutional autoencoder was used, and its performance was compared to the edge detection method. The authors highlighted the limitations of ML sensors such as environmental factors which might interfere with the emitted light and hinder its performance. Schackmann et al. [81] proposed a unified CNN method to detect damage and identify damage severity for broadband ultrasonic guided waves signals. Zhuojun et al. [82] evaluated the performance of transfer learning method for localization of damage areas in CFRP plate. Ijjeh et al. utilized a scanning laser doppler vibrometer to collect the datasets [83]. The collected data was used in a comparative study to detect and localize delamination in composite materials via deep learning models such as UNet, VGG16, encoder-decoder, FCNdenseNet, PSPNet, and Global Convolutional Network (GCN). The proposed model was claimed to perform well on detecting delamination. However, the authors are not sure whether their method will perform well on experimental dataset. Moreover, the dataset was collected under a single low excitation frequency of 50 kHz.

Most of the literature articles were focused on top-down approaches, where the collected data from the experiment is used for damage detection and localization. Rautela et al. [84] adopted a bottom-up approach where they used a parallel implementation of a reduced-order spectral finite element model (SFEM) to solve the forward problem. To mimic the data as closely as the real scenario, the simulated datasets were corrupted with Gaussian random noise. In this work, damage prediction based on deep learning has surpassed the predictions of conventional machine learning algorithms. From the most peer-reviewed journal about SHM, the excitation frequency used was in the range of Hz to kHz. However, the earlier minor delamination that is hidden inside the composite material is hardly detected through ultrasonic-guided waves. Xu et al. [85] have adopted a novel technique called Terahertz (THz) to detect hidden delamination features with high resolution. The authors proposed a powerful approach called unsupervised CNN based on a deep adversarial adaptation strategy (CNN-DADA) to localize THz delamination. The overall architecture of the proposed framework is shown in Fig. 12 and the CNN-DADA method outperforms the traditional DL in reconstructing the input images. Lamb waves are usually processed either as time-series or imaging data. However, the performance of data-based imaging processing techniques is prone to deterioration, with several artefacts and unsatisfactory accuracy. Yu et al. [86] proposed a weighted delay-and-sum (DAS) imaging method based on denoising autoencoder (DAE) learning for complex composite materials to obtain better damage imaging of Lamb waves in both noisy and noise-free conditions. The proposed technique, DAE-DAS achieved high damage localization accuracy, but the timeconsuming nature of this method was the big issue addressed by the authors. Deep learning has been proven to be more efficient in damage detection and localization than conventional machine learning. However, the huge amount of data required to train a deep learning model was a challenge addressed by a lot of researchers. Different methods have been proposed to train a deep learning model based on small amounts of data. Tong et al. [87] proposed a deep learning inversion with supervision intended to reduce the scale of trained data. In addition, transfer learning techniques were adopted in different works to train a small-scale dataset [35,88]. Jin et al. [89] proposed a damage spatial imaging method to reconstruct the Lamb wave response signal under strong noise and realize the spatial localization of damage.

Usually, the model trained on experimental dataset, experienced low accuracy when applied to the actual dataset where environmental factors are presented. Zhang et al. [90] used a semi-supervised model integrating variational auto-encoder and Bayesian neural network (BNN) to predict damage and quantify uncertainties in composite plate. Wang et al. [91] Proposed a deep learning-assisted delamination evaluation technique by using guided wave time-space images. Wang et al. [92] developed an adaptive damage monitoring technique based on transfer features and 95, 93% accuracy were achieved for both damage detection and damage localization respectively. Azad et al. [93] compared different transfer learning models to classify the damages in composite laminates. The proposed EfficientNet-based lightweight TL model achieved higher accuracy compared to other TL models. Azad et al. [94] proposed an interpretable deep learning model based on explainable vision transformer (X-ViT). The proposed ViT achieved an accuracy of 98.8% in classifying damages. The articles mentioned about interlaminar damage detection are summarized in Table 4.

Apart from delamination, which is the most common interlaminar damage, matrix cracking and fibre-matrix splitting are categorized as intralaminar damage and the initiation of intralaminar damage triggers other types of damage such as delamination [95]. In this section, current research about detection and localization through deep learning for matrix cracking and fibre pullout was highlighted and summarized in Table 5. Sampath et al. [96] proposed a hybrid method that combines long short-term memory and trispectrum (LSTM-TS) based on a deep learning model to accurately detect earlier fatigue cracks in structures in a noisier environment. The proposed method outperformed other methods by providing low values of mean absolute error (MAE), mean absolute percentage error (MAPE), and root mean square error (RMSE) and exhibiting a high signal-to-noise ratio (SNR). Nguyen et al. [97] identified multiple cracks in a structure using wavelet analysis based on displacement signals. They used deep learning to combine two independent algorithms in wavelet analysis of deflected static and dynamic elements signals. However, the authors did not clearly explain the benefit of dynamic load over static load in detecting cracks. A well-structured and understandable article about microscale and macroscale damages was conducted in [98]. In their study, they proposed various transfer learning algorithms for the quantitative assessment of visual detectability of both in-service aircraft and wind turbines. However, none of the models was capable of classifying the damage from low-impact energy. Usually, the evolution of matrix cracking leads to delamination under applied fatigue load. An automated and robust damage detection technique using a deep autoencoder was proposed by [99] with the aim of detecting and classifying fatigue damage in CFRP structures. In this work, only one dominant feature was classified as matrix cracking, while others were categorized as intact and delaminated features. Miorelli et al. [100] proposed a DNN architecture to automatically invert defect position and size based on guided wave images (GWI) images obtained through simulations. Zhao et al. [101] conducted simulations and experimental studies for early crack detection using nonlinear output frequency response functions (NOFRFs) and CNN-LSTM methods. Among NDT techniques, the data-driven method has proven to be effective for damage detection in composite materials. However, the serious issue with this technique has been the larger dataset collection and its quality. Dabetwar et al. [102] improved damage diagnostics of complex materials by developing a data fusion technique along with data augmentation via a deep neural network. Also, a deep neural network was combined with an existing technique, such as unsupervised dimensionality reduction, that projects very high dimensional data into two- or three-dimensional space. The improved technique by CNN was able to effectively cluster any incoming unknown samples [103]. Lee et al. [104] proposed real-time self-sensing data using electromechanical behavior data images and CNN architecture. They were aiming to solve the issue encountered in their previous research, where a huge number of embedded electrodes were used to sense the damage on a small scale.

Deep learning algorithms have proven to detect damage with higher accuracy than conventional machine learning, despite their need for larger datasets and high computational cost. Most researchers are interested in knowing how a deep learning model classifies test data as normal or unhealthy based on critical feature contributions. Zhang et al. [105] come up with a method, namely attention-based interpretable prototypical networks for deep learning models trained on a small dataset. The proposed method is illustrated in Fig. 13. They introduced an attention channel module to extract the highly important features and disregard the non-relatable features to alleviate the overfitting issues of scarce datasets. Saveeth et al. [106] proposed a modified crow deep neural network (McrowDNN) and this approach helped to choose the most appropriate weights and prejudices. McrowDNN provides higher accuracy for crack detection than ANN and RNN. However, a larger amount of image datasets with different types of defects should be used to confirm the efficiency of the proposed model. To judge whether the deep learning model did not cheat on the trained data, a physical local interpretable and explainable algorithm namely Local Interpretable Model-Agnostic Explanation (LIME) was performed. Moreover, LIME was adopted to interpret a 1D-CNN model in terms of damage feature contributions and find the interpretation corresponds to damage signature [107]. Guo et al. [108] developed ultrasonic guided HMC network and its lightweight version to detect different type of defects such as pinhole, crack and corrosion.

During crack detection in sandwich material, different vision-based technics have been used to identify superficial damages from the scanned images. Manujesh et al. [109] compared classical machine learnings with DCNN performance for damage detection and classification on sandwiched composites. The trained dataset was collected via scanning electron microscope (SEM) images and processed with different techniques such as grey scaling before feature extractions. A deep neural networks-based on seismic-wave for crack detection inside a composite plate was proposed by [110]. Moreover, Gao et al. [111] applied deep learning to process the nano computed tomography (CT) scan images to obtain complex 3D composite features with the help of image segmentation using ORS dragonfly software. With the help of CNN, the damaged features such as matrix cracks, pull-out fibres and holes were accurately detected at the macro-scale, mesoscale and micro-scale level. Most of the works cited above were mostly offline conducted. A real-time damage detection of additive manufacturing was conducted for ensuring the quality of a fabricated CFRP composite [112]. In this work, a deep learning model was used to detect, classify and evaluate the defect in real time with high accuracy. To improve the prediction accuracy of SHM in real-time, Du et al. [113] proposed an end-to-end SHM approach employing acoustic emission technique and a deep learning model with attention mechanism.

Lamb wave-based damage detection has proven to be efficient in damage detection with great advantages such as actuation simplicity, long travel with low attenuation, and higher sensitivity to minor damages. Despite all those potentials, SHM based on Lamb waves still faces several challenges and limitations that need attention from the researchers. The relevant challenge issues with their respective future prospect have been identified as follow:

Data-driven structural health monitoring approach uses Lamb wave signals, mostly acquired from a transmitter through the composite surface up to the receiver. The dispersive nature and mode conversion of Lamb waves and boundary reflections of material complicate the analysis of acquired Lamb wave signals, which impose hindrance and may lead to false damage prediction. Mode conversion and wave scattering result when incident waves interact with damage, and this phenomenon can complicate the detection of the presence damage [114]. Based on the complexity of the generated signal, smart and trustworthy preprocessing techniques should be developed to avoid poor performance in the damage identification process.

Lamb wave propagation is highly affected by environmental factors such as humidity, external noise, vibration, temperature, and loading conditions. Gorgin et al [115], presented a deep review about environmental and operational conditions effects on Lamb waves, and they mentioned that temperature has been the most significant factors while others were mostly ignored. However, the effects of other environmental factors on Lamb waves in real application are unavoidable. In addition, temperature can affect the propagation of the Lamb wave in two distinctive manners, which are the effect on material parameters and the influence on the thermal stress of PZT [116]. The variation of environmental factors during data collection can lead to changes in wave velocity, signal attenuation, or amplification, which can give false negatives or positives damage detection. For collecting accurate data and reliable damage detection and localization, environmental factors should be taken into consideration.

Gathering data in a real-world environment for structure health management applications is not an easy task since it requires testing the structure up to costly failures, it takes a long time and is almost impossible for some application areas. In addition, a larger number of sensors installed on the structures to capture all the necessary sensing paths is not desired where weight is an issue. Moreover, the large dataset collected by sensor arrays imposes high computational challenges and expensive hardware. There is still room for researchers to develop real-time data processing techniques that are efficient and light-weighted algorithms for real-time processing data.

In less than two decades, deep learning was introduced with feature learning ability as an advantage compared to traditional methods. Moreover, the features extracted by deep learning models can be used by traditional machine learning for classification tasks. Despite the benefit offered by deep learning in structural health monitoring, its shortcomings are also undeniable. For high accuracy expectancy, most deep learning models rely on larger and balanced datasets to train the data, which may inhibit its applicability.

To alleviate the issue of data scarcity when training a deep learning model, different methods such as transfer learning [117,118], generative models [119–121] and data augmentation [73] have been addressed. However, all the methods tried to enhance the trained data which still requires long training hours and powerful hardware. Researchers should focus more on developing lightweight and powerful deep learning models that are able to train small, balanced datasets with clear interpretability and explainability of the model’s prediction accuracy without an overfitting issue. In addition, digital twin technology can enable intelligent decision-making based on rich digital data and real-time updates from the physical world [122].

This review explored a larger number of the latest studies related to the application of Lamb waves and deep learning to detect and localize both intralaminar and interlaminar damages in composite materials. Traditional machine learning techniques have shown some unbearable weaknesses as they require experts in statistics and signal processing before training the models. However, it offers lower complexity and a faster processing time than deep learning. Deep learning application is overtaking traditional machine learning in structure health monitoring due to their self-feature learning capability, and it has proven its effectiveness in detecting damages with higher accuracy than conventional machine learning. Although deep learning offers great advantages over conventional machine learning techniques, its shortcomings, such as larger dataset dependence, complexity, and legion hyperparameters of the trained model make it difficult to optimize. Finally, scholars should aim to develop deep learning models that can also perform well on small datasets without relying on data augmentation techniques.