Explainable deep learning framework for residual stress monitoring in laser shock peening via acoustic emission

In the field of intelligent manufacturing, leveraging acoustic emission (AE) technology for quality monitoring and assurance during laser manufacturing processes is paramount. Despite this, current monitoring techniques struggle to accurately characterize the time-frequency distribution and transient dynamics of AE signals, and there exists a paucity of neural network models tailored for these specific analytical tasks. To bridge this gap, this paper presents a cutting-edge monitoring approach that integrates a Bi-Differential Convolutional Network (BDCN) with a Frequency Bands Recalibration Spectrogram (FBRS). Firstly, a novel analytical technique employing FBRS for transient AE signals is introduced, which adaptively redistributes frequency resolution to highlight informative components within a constrained pixel space. The BDCN, a groundbreaking nonlinear network model, jointly performs directional feature processing and stress state classification by incorporating two specialized functional modules designed for horizontal and vertical differencing. The model emphasizes directional texture and gradient patterns while mitigating low-frequency feature loss through complementary enhancement strategies. The efficacy of the proposed methodology has been empirically confirmed through rigorous testing on aluminum alloy 7075 and titanium alloy TC4. When juxtaposed with state-of-the-art networks, the presented monitoring strategy exhibits enhanced discriminative precision and robustness, signifying its potential in the domain of intelligent manufacturing quality assurance.