Early fatigue failure detection in composites using autoencoder-based anomaly detection

Abstract

Despite the widespread adoption of composite materials across various industries, accurately evaluating their durability—particularly under fatigue loading—remains a major challenge. A key difficulty lies in the substantial scatter in fatigue life among seemingly identical specimens. This variability elevates the risk of sudden, catastrophic failures and necessitates conservative, schedule-based maintenance plans that are designed around worst-case scenarios. This study presents a novel approach to identify composite specimens with short fatigue lives at the early stage of loading by integrating piezo-resistivity-based structural health monitoring (SHM) with autoencoder-based anomaly detection techniques. Glass fiber–epoxy composites, made electrically conductive by incorporating carbon nanotubes (CNTs), were subjected to fatigue loading until failure, while their electrical resistance (ER) was continuously monitored. The ER data from the early stage of loading were extracted and used to train and optimize autoencoders to detect early fatigue failure (i.e., short-life specimens). The results demonstrated an F1 score of 95 % and an accuracy of 97 % in correctly identifying short-life specimens, underscoring the effectiveness of the proposed approach.