WCNN-KAN: A Novel Interpretable Multiscale Attention Feature Fusion Network for Rotating Machinery Fault Diagnosis

Due to the extremely harsh environment in which rotating machines operate, existing fault diagnosis algorithms face significant challenges in extracting key features: (1) Traditional deep learning methods make it difficult to obtain fault information from a non-stationary signal. (2) Extracting critical faults from single-scale features can be hard in complex working environments. (3) Traditional Multilayer Perceptron (MLP) structures limit the learning ability of models. To address these issues, this paper proposes a Wavelet Convolutional Neural Network with Kolmogorov-Arnold Networks (WCNN-KAN) framework for intelligent fault diagnosis. Firstly, signals are converted into wavelet time-frequency diagrams, and the multi-scale wavelet characteristic enhancement layer is constructed by using wavelet convolution. Secondly, a feature purification network based on the KAN attention mechanism, coupled with a multi-stage feature fusion module, is established to better extract discriminative features. Finally, KAN is introduced as a classifier to further improve the diagnostic performance of the model. Two bearing datasets are utilized to validate the efficiency of WCNN-KAN, and the results demonstrate that WCNN-KAN is superior to existing fault diagnosis methods.