VProtoNet: Intelligent damage perception in composite structures with joint time-frequency encoding and multi-objective decoupling

Accurate damage characterization in the structural health monitoring of composite structures is essential for informed decision-making, as well as for improving safety and durability while minimizing maintenance costs. Acoustic emission techniques, which offer millisecond-level localization and quantification of micro-damage events, are well-suited for long-term monitoring. Nevertheless, structural health monitoring based on acoustic emission is challenged by pronounced data imbalance arising from sparse damage events and variable signal-to-noise ratios, as well as by the need to jointly yet distinctly resolve correlated impact location and severity estimation tasks. To address these challenges, an adaptive VProtoNet model is proposed. A variational autoencoder is first employed as an anomaly detector, trained in an unsupervised manner to learn the distribution of normal signals and reject samples that deviate from it, thereby isolating potential damage signatures. The filtered signals are subsequently analyzed using a prototype network, which enables unified feature learning and coordinated multi-task prediction via heterogeneous output heads. A two-stage optimization scheme with conservative data augmentation is introduced to enhance model robustness. Experimental results on carbon fiber-reinforced polymer plates demonstrate that the proposed approach substantially improves impact localization accuracy and severity prediction, demonstrating its effectiveness for long-term monitoring applications.