Machine learning based damage identification in SiC/SiC composites from acoustic emissions using autoencoders

Developing the ability to leverage machine learning (ML) to identify damage mechanisms in heterogeneous materials from their acoustic emissions (AE) has wide-reaching ramifications for multi-modal experimentation. It would allow researchers to augment damage triangulation, lifetime prediction, and high-resolution optical studies with complementary mechanism-informed data streams. However, developing this capability hinges on the collection of ground truth acoustic libraries from damage in realistic geometries. Due to time and monetary considerations, there is a dearth of ground truth libraries which can be used to robustly characterize ML mechanism identification frameworks. Addressing this gap, we present a multi-modal acoustic emission and x-ray computed tomography study where AE is gathered, and subsequently labeled, from SiC/SiC unidirectional composites under monotonic tension. This library is used to demonstrate that acoustic signals from early fiber breaks are obscured by matrix cracking. A first-order micromechanical model is used to explain the origin of this obscuring effect, and identify fundamental limitations of unsupervised frameworks. An autoencoder-based anomaly detector approach is used for the first time to overcome these limitations, additionally demonstrating that the frequency distribution of fiber break acoustic signals is narrow. Implications of these findings for enhanced multi-modal testing and online health monitoring are discussed, and strategies for implementation of supervised damage mechanism identification frameworks are proposed.