Physics-informed interpretation and classification of acoustic emission signals using explainable deep convolutional neural networks

Model interpretability remains a critical challenge for deep learning applications in Acoustic Emission (AE) signal characterization, limiting their trustworthiness in structural health monitoring. The proposed approach integrates explainable Convolutional Neural Network with physics-informed segmentation to enhance both classification accuracy and interpretability. By segmenting signals based on the theoretical arrival times of fundamental Lamb wave modes ($S_0$ and $A_0$), and employing Class Activation Mapping (CAM), Gradient-weighted CAM (Grad-CAM), and Dimension-wise CAM (DCAM), we provide quantitative insights into the model's decision-making process. Using 200 AE signals from pencil break tests, our model identifies distinct features for different events. Visualizations show the model focuses on $S_0$-$A_0$ transition and post-A₀ regions, with DCAM highlighting significant importance in the $S_0$-$A_0$ transition region for the closest test point. In this paper, we address the 'black box' nature of deep learning by offering quantitative, physics-based intuition, correlating model outputs to specific AE signal segments and underlying physical processes such as Lamb wave mode interactions and dispersion. By bridging the gap between deep learning performance and human-interpretable insights, our method enhances the reliability of AE-based structural health monitoring.