Deciphering laser shock peening quality monitoring: Wavelet-driven network with interpretability

Quality monitoring of laser shock peening based on acoustic emission technology is a topical multidisciplinary issue that has received much attention in recent years. To acquire complex time-varying information in acoustic emission signals, convolutional neural networks with powerful learning capabilities have shown potential for a wide range of applications. However, the black-box property of the network imposes limitations on its further development and decision credibility. Therefore, this study proposes a wavelet-driven network with theoretical basis and physical significance. This network can cyclically utilize discrete wavelet packet transform to map input features to the wavelet domain during the learning process, thereby obtaining more robust and valuable information. This paper also constructs a novel wavelet attention mechanism that takes into account the difference between low-frequency and high-frequency information, and is able to allocate resources in both the decomposition component and the time-domain dimension. The proposed method can be seen as a multiresolution analysis technique that combines existing physical knowledge with nonlinear feature processing and feature selective enhancement. The results of the two laser shock peening cases show that the proposed method not only outperforms current state-of-the-art models in terms of monitoring performance, but also has better physical interpretability. Importantly, the proposed method has the potential to be further extended to other interpretable structural health monitoring.