Classification of Acoustic Emission Signals from Failure Modes in Thermoplastic Composites Using Machine Learning

The evolution of the damage process in PAEK-type thermoplastic carbon fiber-reinforced polymer (CFRP) quasi-isotropic composite tensile specimens was investigated using acoustic emission (AE) monitoring under quasi-static loading to failure. Key objectives were to apply a deep learning (DL)-machine learning (ML) framework to the AE signals from tested specimens, identify the participating signal classes using unsupervised classification, and quantify their relative extent at different stages of the loading cycle. Unlike prior studies that rely on conventional AE features (i.e., amplitude, rise time, etc.), this work automated feature extraction directly from AE signals using a deep learning model. A 1D convolutional neural network (CNN) automatically extracted frequency and temporal features from AE events recorded from nominally identical specimens tested according to ASTM D3039. Principal component analysis reduced the high-dimensional feature space while retaining 85% of the variance, and spectral clustering identified four distinct classes validated by Davies–Bouldin index < 1.2 and Calinski–Harabasz index > 200 across both specimens. Fast Fourier transform analysis revealed that classes are primarily distinguished by frequency content. Low- and mid-frequency signals (<900 kHz), typically associated with delamination and matrix cracking in published literature, were relatively few. Two high-frequency classes emerged. One had mid-high frequency events (900–1600 kHz) with high amplitude, increasing steadily up to 70% of ultimate load, and the other very high-frequency events (>1600 kHz) with low amplitude, surging rapidly beyond 80% of ultimate load until failure. Cross-correlation analysis revealed minimal formation of groups of fiber breaks called the ‘multiplets’ characteristic, suggesting the ductile PAEK matrix suppresses clustered fiber damage. The proposed DL-ML framework exhibited promising results, with the prospect of eliminating dependence on subjective, test-setup-sensitive parameters and providing a reproducible methodology for damage classification in thermoplastic composites. Besides, automated feature extraction could improve the likelihood of early detection of real-life damage progression before final failure. As a potential mathematical model, the scope for further improvement includes that the performance of the current framework depends significantly on the preprocessing of the data (i.e., the right size of signal that has substantial information on damage modes of both short and long type with least noise, elimination of the amplitude parameter by choosing manual amplitude bin). Direct validation of frequency-to-damage mode assignments through synchronized in situ imaging remains necessary to confirm the damage modes.