A frequency-adaptive and interpretable neural network for intelligent fault diagnosis of high-speed motor bearings

Abstract Convolutional neural networks (CNNs) have significant application value in industrial equipment condition monitoring and fault diagnosis, but they face notable challenges in practical engineering scenarios: feature distribution drift caused by varying operating conditions, on-site noise interfering with feature extraction, and a lack of interpretability in model decisions, which directly affect the reliability and maintainability of on-site industrial diagnostics. To address these engineering pain points, this study integrates the advantages of continuous wavelet convolution and discrete wavelet transform to construct an interpretable wavelet convolution intelligent diagnosis network (WCIDN). First, an adaptive condition convolutional layer (ACC layer) is proposed, which achieves dynamic weight configuration under varying operating conditions by setting frequency kernel functions, channel kernel functions, and filter kernel functions. Subsequently, a discrete wavelet transform layer is employed to convert time-domain signals into the wavelet domain, and an adaptive wavelet attention mechanism (AWAM) is designed. This mechanism can adaptively match fault frequencies, effectively filter noise, and retain fault signals. Finally, the visualization of kernel functions under varying operating conditions and the spectral analysis of the attention mechanism contribute to explaining the decision-making mechanism of the network. The potential engineering practical value of this method was demonstrated using spindle bearing fault data from Jilin University and motor bearing data from the Polytechnic University of Turin.