Lamb wave damage localization method for carbon fiber-reinforced polymer plates based on 1DCNN-KAN and physics-informed data augmentation

Lamb wave-based structural health monitoring (SHM) technology provides an effective means for online damage monitoring of carbon fiber-reinforced polymer (CFRP) plates. However, affected by the complex multimodal and dispersive characteristics of Lamb waves, feature extraction from damage monitoring signals heavily relies on manual experience. To address this issue, this paper proposes an automatic Lamb wave damage localization method integrated with deep learning, where an end-to-end deep learning framework fusing the one-dimensional convolutional neural network and the Kolmogorov-Arnold Network (1DCNN-KAN) is developed. Taking raw damage monitoring signals as input, the 1D-CNN is adopted to extract local features of the signals, while the KAN decision layer explicitly models the mapping relationship between features and damage locations through learnable one-dimensional spline functions, thereby realizing structural damage localization without complex manual signal feature extraction. For damage location prediction, the damage monitoring area is gridded, and random damage is mapped to different grid nodes to achieve classification-based damage localization. To tackle the difficulty in acquiring monitoring signals of damage at arbitrary locations, a physics-informed data augmentation method is proposed, which generates Lamb wave monitoring signals covering random damage locations across the entire monitoring area based on the signals acquired at grid nodes. Finally, damage localization experiments are carried out on a CFRP plate. For damage of different sizes at 40 random locations (the minimum damage size is 10 mm × 5 mm × 20 mm), the 1DCNN-KAN model achieves grid-cell-level positioning for all test samples, and the predicted damage location and the actual location are both within the same grid cell (with a grid cell resolution of 20 mm × 20 mm). The classification accuracy of the damage location reaches 100%. Comparisons with different models demonstrate the superior performance of 1DCNN-KAN. This study provides an intelligent solution for damage localization of CFRP plates.