Damage detection of aircraft skin in complex regions remains a persistent challenge in the aerospace industry due to the compact structures and limited accessibility of areas such as wing-fuselage junctions, engine nacelles, and flaps. This article presents a miniaturized damage detection device tailored for such environments and investigates the physical characteristics of the acquired signals, including impact-response behavior, energy-induced temporal causality, and spatial spike distributions. These characteristics are leveraged to design a spatio-temporal neural network architecture. The detection process is modeled as an impact-response dynamic system to guide the construction of wavelet-based convolutional kernels with learnable scale, translation, damping, and phase parameters. A causal convolutional network captures temporal dependencies of spike features, while a graph attention network extracts spatial features based on transformed spike intervals and an adjacency matrix. A point-wise attention mechanism is then applied to fuse multi-dimensional features for final classification. Experimental results demonstrate that the proposed method outperforms existing approaches and provides strong interpretability.
Zhou et al. (Mon,) studied this question.