Autoencoder–GMM clustering of time–frequency AE data for damage mechanism identification in SiC/SiC composites

• Continuous 10 MHz AE recording enabled dead-time-free monitoring. • Three clusters separated matrix cracking, fiber pullout, and fiber breakage. • High-frequency AE above 600 kHz was associated with fiber breakage. In this study, we propose a method that combines an autoencoder with Gaussian mixture model (GMM) clustering to analyze the acoustic emission (AE) signals generated during the tensile testing of SiC/SiC composites. The AE signals were first transformed to the frequency domain and had noise reduced by applying a short-time Fourier transform (STFT). Subsequently, these signals were compressed into a three-dimensional latent space using an autoencoder. Clustering within this latent space allowed for the identification of distinct damage mechanisms. Quantitative analysis revealed that using three clusters enabled the classification into matrix cracking, fiber pullout, and fiber breakage. Each cluster exhibited distinct frequency characteristics that enabled damage identification. The cluster dominated by components above 600 kHz was associated with fiber breakage. The cluster showing strong low-frequency components below 400 kHz was attributed to fiber pullout, supported by SEM observations of pullout traces. Matrix cracking corresponded to signals with mid-frequency components and a stable Felicity ratio. The novelty of this study lies in the combination of an autoencoder with GMM clustering. To the best of our knowledge, this is the first application of this approach to AE data from SiC/SiC composites. Unlike conventional threshold-based AE methods, this framework requires no manual selection of event-extraction conditions or AE parameters, enabling uninterrupted monitoring throughout both loading and load-holding phases without dead time, and thereby allowing continuous tracking of concurrent damage mechanisms including matrix cracking, fiber pullout, and fiber breakage.