Physics-guided graph convolutional network for damage severity and zone identification in industrial composites

Lamb wave (LW)-based technology, which offers long-range propagation and sensitivity to various damage types, has emerged as a promising approach to diagnose damage in composite structures. However, traditional structural health monitoring (SHM) systems face significant challenges that include reliance on dense sensor arrays, complex imaging-based processing, evaluation of damage index with respect to baseline signals, and high computational costs. To address these limitations, this study presents a physics-guided graph convolutional network (GCN) framework to integrate damage severity assessment and zone localization in carbon fiber-reinforced polymer (CFRP) laminates which possess a wide range of industrial applications. The framework transforms LW signals into graphical representations, where nodes correspond to sensing paths, while edges reflect physical relationships based on experimental configurations. Three adjacency matrix variants: the fully connected (GCN−FC), clustered by actuators (GCN−CA), and shared wave propagation paths (GCN−CP), were designed to explore their influence on GCN performance. Experimental results demonstrate that the physics-guided GCN models (GCN−CA and GCN−CP) significantly outperform the conventional GCN−FC model, using only four piezoelectric sensors to achieve (96.09 and 99.09) % accuracy for severity assessment and damage zone localization, respectively. The results demonstrate the potential of physics-guided graph structures to enhance LW-based SHM frameworks.