Temperature-compensated autoencoders with sequences of raw time-domain signals for damage detection and localization in active and passive SHM

This study investigates an autoencoder trained with short-term sequences of raw time-domain signals for unsupervised damage detection and localization under varying temperatures. The approach is designed to overcome the lack of transferability in feature-based methods and is therefore tested on both active (ultrasonic guided waves) and passive (vibrational responses) structural health monitoring systems. Both systems are highly sensitive to temperature variations, which alter structural responses and wave propagation properties without inducing permanent changes, thereby necessitating robust normalization strategies. For a cantilever beam in a climate chamber and an active piezoelectric system placed on a composite plate, the data-driven strategy successfully compensated for temperature effects, enabling sensitive damage analysis. In vibration-based monitoring, the model performed best when trained on temperature ranges rather than discrete states. For guided waves, damage was localized with consistently low error by integrating the autoencoder’s residual covariances with the Reconstruction Algorithm for Probabilistic Inspection of Damage (RAPID) algorithm. Critically, this was achieved without requiring a comprehensive intact-state data set across all temperatures. These findings demonstrate that the autoencoder framework is robustly applicable across both active and passive SHM domains, and the developed enhancements are fully transferable.