PSI - Issue 77

Muhammad Jahanzeb Zia et al. / Procedia Structural Integrity 77 (2026) 111–118 Muhammad Jahanzeb Zia et al. / Structural Integrity Procedia 00 (2026) 000–000

113

3

AE-based approaches are well established for detecting and classifying composite damage under torsional and compressive loads. Distinct frequency bands are associated with specific damage types: matrix cracking (125–187.5 kHz), delamination (187.5–250 kHz), fibre/matrix debonding (250–312.5 kHz), and fibre breakage (312.5–375 kHz). Gholizadeh et al. (2015) noted that AE signals in epoxy–glass composites arise from micro-mechanical events with unique spectral signatures, enabling damage identification. Paipetis et al. (2011) further distinguished between low amplitude, high-frequency signals from matrix cracking and high-amplitude events from fibre breakage. Zhou et al. (2018) linked AE parameters such as amplitude, frequency, and cumulative counts with compression-induced damage, while Zhang et al. (2024)identified matrix cracking, debonding, and delamination as dominant torsional mechanisms. Pei and Xiang (2023) applied AE clustering to torsion-loaded carbon composites, successfully distinguishing between fibre breakage, matrix cracking, and debonding. Their work also showed that pre-delaminated specimens were more susceptible to delamination. Roundi et al. (2018) validated the role of low-frequency AE in detecting matrix cracking and fibre breakage in glass-fibre reinforced plastics, highlighting the potential of AE for real-time SHM. Static tensile testing continues to be a cornerstone for evaluating composite failure. In this context, AE techniques provide sensitive, real-time monitoring of microstructural events. However, AE datasets are high-dimensional and noisy, requiring machine learning (ML) for automated classification. Among ML methods, Support Vector Machines (SVMs) have been widely adopted for AE analysis due to their robustness with limited training sets and ability to model nonlinear feature spaces using kernel functions (Liu et al., 2018; Li et al., 2011). Feature extraction typically employs spectral descriptors, statistical measures, and time–frequency transformations and achieved accurate crack classification in rotating shafts using PWVD-based features combined with SVMs. Similarly, Barbero et al. (2013) and Chiu et al. (2016) highlighted the value of FE-derived AE datasets for enhancing model training. To address the scarcity of experimental AE data, transfer learning and synthetic augmentation have emerged as promising strategies. Dong et al. (2022) and Xiao et al. (2019) demonstrated that transfer learning, from weight freezing to fine-tuning, substantially improves classifier performance when leveraging simulated AE signals. Recent work confirms SVM effectiveness in composite SHM: Guo et al. (2021) and Zhang et al. (2015) reported high accuracies for damage classification, while Chen et al. (2022) proposed a multi-class SVM framework capable of distinguishing multiple mechanisms simultaneously. Collectively, these approaches achieve classification accuracy exceeding 90%, particularly when combined with advanced feature extraction such as Recurrence Quantification Analysis (RQA) and Kernel Principal Component Analysis (KPCA). More recently, deep learning architectures, including Artificial Neural Networks, Convolutional Neural Networks, and Autoencoders have been applied to AE data, offering automated feature extraction directly from raw signals (Azad et al., 2024). Similar to these experimentations, the raw signal was generated by FEA for the training of an intelligent model. In the present study, finite element models were used to simulate AE signals associated with matrix cracking and fibre breakage under varying input conditions. Damage initiation was modelled using the Hashin criteria, with the Benzeggagh-Kenane damage evolution law applied to capture energy release post-initiation. The simulated AE signals formed the training dataset for an SVM classifier, which was subsequently validated on previously unseen AE events. Importantly, raw AE signals rather than spectrograms or FFT representations were employed to reduce computational overhead, thereby supporting the feasibility of real-time SHM implementation.

2. Numerical methodology 2.1. Geometry and material

An epoxy-glass composite material was employed to investigate the acoustic emission signals generated during matrix and fibre cracking under tensile loading. The ABAQUS/Explicit 2024 solver was used to perform failure analysis on a 20×10 mm strip consisting of 24 unidirectional plies, each with a thickness of 0.14 mm. A notch was introduced at the centre of the strip to act as a stress concentration region, as illustrated in Fig. 1. The material properties used in the analysis are summarized in Table 1 (Singh et al., 2015). Table 1. Mechanical properties of E-Glass composite (Singh et al., 2015). 1 (GPa) 2 (GPa) 3 (GPa) 1 2 3 12 (GPa) 13 (GPa) 23 (GPa) 40 10 10 0.3 0.3 0.21 3.15 3.15 4.32