An interpretable fault diagnosis method for rotating machinery based on adaptive fusion wavelet networks

In the field of fault diagnosis, traditional deep learning models suffer from the problem of non-interpretable processes. Moreover, traditional convolutional neural networks (CNNs) and fixed wavelet bases have difficulty in adaptively extracting high-frequency features from complex non-stationary signals, which greatly limits their practical application effectiveness. To solve this problem, a fault-explainable method of rotating machinery based on adaptive fusion wavelet network (AFWNet) is proposed. AFWNet is mainly composed of adaptive fusion wavelet convolution layer (AFWConv), residual module, and one-dimensional signal attention mechanism. The AFWConv structure is designed to extract multi-scale features by leveraging the inner product mechanism inherent in wavelet transforms, enabling the automatic selection of the optimal wavelet basis. The introduction of a residual module improves the convergence of the model, while the optimized one-dimensional signal attention mechanism automatically adjusts training parameters, so that the model pays more attention to important features. Public bearing data sets were used to validate the AFWNet method, and the visual feature extraction results of vibration signals under different convolution layers were compared with the fault classification results to verify the reliability and feasibility of the AFWNet method. In addition, in terms of diagnostic performance, AFWNet’s fault classification accuracy can reach 99.98%, the effective training cycle is shorter than that of CNN and other models, and it also has certain anti-noise ability.