LU Yixiang1, 2, QIAN Dongsheng1, 2, ZHU De1, 2, SUN Dong1, 2, ZHAO Dawei1, 2, GAO Qingwei1, 2
Journal of Vibration and Shock.
2024, 43(17):
203-213.
In practical engineering, fault diagnosis of rotating machinery often faces various complex situations such as noise interference, limited fault samples and variable working conditions, which pose new challenges to the application of data-driven deep learning methods that lack prior knowledge. Traditional fault diagnosis methods based on wavelet analysis can extract rich prior knowledge of faults, but a fixed (structured) or single wavelet basis is difficult to directly adapt to complex fault scenarios. To address these issues, a multiscale wavelet packet-inspired convolutional network (MWPICNet) was proposed for fault diagnosis of rotating machinery in this paper, inspired by traditional multiscale wavelet packet analysis. The proposed MWPICNet internally coupled the time-frequency domain conversion with filtering denoising, feature extraction and classification. First, the multiscale wavelet packet-inspired convolutional (MWPIC) layer and soft-thresholding activation (ST) layer were alternately used for signal decomposition and nonlinear transformation, extracting multiscale time-frequency fault features and filtering out the noise layer by layer. Each MWPIC layer could be approximately seen as a single-layer wavelet packet transform of the signal under multiple learnable wavelet bases, and learnable thresholds in the ST layer were used to sparse the wavelet coefficients. Then, the frequency band weighting (FBW) layer was designed to dynamically adjust the weights of each frequency band channel. Finally, a global power pooling layer (GPP) was introduced to extract discriminative frequency band energy features that were helpful for fault identification. The efficacy of the proposed MWPICNet is verified through case studies designed for different complex scenarios on three fault diagnosis datasets.