Interpretable real-time monitoring of pipeline weld crack leakage based on wavelet multi-kernel network

The deep learning technology used for pipeline weld crack leakage monitoring lacks physical interpretability, which makes it difficult to provide theoretical support to decision-makers, reducing the practicality and applicability of the technology. To solve this problem, a deep learning framework that combines interpretability and feature fusion is proposed to solve the real-time monitoring problem of pipeline leaks. First, three different wavelet basis functions are used to analyze the leakage acoustic emission signal, and the resulting time-frequency features exhibit correlation and complementarity. Based on this, an interpretable neural network with different wavelet convolution kernels is designed to extract abstract feature details of wideband acoustic emission signals through multi-level dynamic receptive fields. Then, a feature fusion module based on channel importance weighting was designed to optimize the learning process of the network by highlighting the contribution of different channels. Finally, the effectiveness and superiority of the proposed method were verified through pipeline leakage acoustic emission monitoring experiments. The results show that the proposed method can effectively extract distinguishing features of leak acoustic emission signals, and the recognition accuracy of different leakage states can reach 98.32%, which is significantly higher than that of typical deep learning methods. In addition, feature map visualization proves that the proposed method has physical interpretability in abstract feature extraction, which will provide a reference basis for correct decision-making by industrial technicians.