MelNet: an end-to-end adaptive network with adjustable frequency for preprocessing-free broadband acoustic emission signals

Deep learning can obtain discriminative abstract information from the artificial features of acoustic emission signals and complete the target task in the form of high-precision recognition. However, mainstream deep learning is only a mechanical implementation work, but cannot provide valuable feedback on signal analysis. To overcome this problem, this paper innovatively proposes an end-to-end interpretable model that highly integrates two-dimensional time-frequency representation with neural network feature extraction. The proposed model can be considered as an initial tool for signal analysis. Specifically, the core frequency used to control the resolution of time-frequency features will be fully involved in model training and gradient propagation. The adaptive learning algorithm not only accurately captures the key frequency characteristics of the signal, but also identifies and skips sub-optimal frequency points that do not bring significant performance improvements. More importantly, this visualization of the adaptive frequency selection process ensures that the extracted features are highly relevant to the task, thus improving the interpretability of the feature extraction stage. The feasibility of the proposed method was verified in two different cases: key equipment service monitoring and manufacturing process monitoring. The results show that the proposed method can optimize the optimal frequency component and obtain ideal monitoring accuracy without relying on any expert experience. While the deep learning model itself may not be inherently interpretable, its role in guiding the feature extraction process via gradient optimization introduces a level of interpretability absent in conventional methods.