ConvMPAnet: A Novel End-to-End Lightweight Damage Localization Framework Under Heavy Noise in Composite Structures

Ultrasonic guided wave (UGW)-based damage detection is considered as one of the most promising structural health monitoring (SHM) technologies for assessing the safety and integrity of composite structures. However, existing UGW-based monitoring methods using machine learning face great challenges due to noisy interference and limited computing resources. To address these issues, a novel end-to-end lightweight multiscale point-level attention (MPA)-based convolutional neural network is proposed for real-time damage localization in composite structures under heavy noise conditions. In the proposed method, first, an efficient improved differential-driven piecewise aggregate approximation (IDPAA) is developed to compress multipath-guided wave signals for improving calculation efficiency. Next, considering the impact of the damage location on different paths and the coupling relationships between these paths, a feasible method is proposed for enhancing the damage signal and fusing multipath data. Finally, by incorporating specially the designed MPA mechanism, a lightweight convolution network with multiscale point-level attention (ConvMPAnet) is developed for real-time damage localization under a heavy noise environment. The performance of the proposed method is verified on real-world guided wave experiments. Experimental results demonstrate that the proposed approach has exceptional anti-noise capability performance, surpassing the state-of-the-art damage detection methods in both accuracy and lightweight performance.