Bearing fault diagnosis for variable operating conditions based on KAN convolution and dual branch fusion attention

This paper proposes a bearing fault diagnosis method based on Kolmogorov-Arnold Convolutional Network: Adaptive Context-aware Graph Channel Attention with Squeeze-and-Excitation Networks (KANConv-ACGCA-SENet). Firstly, a new structure of KANs is applied to Convolutional Neural Networks (CNN) for replacing traditional linear convolutional kernels. Secondly, a dual-branch fusion attention module, comprising the ACGCA modules, is proposed for use in learning fault features. This is achieved by capturing feature differences and utilising non-local(NL) operations, thereby enhancing the feature representation ability under different working conditions. Subsequently, context-aware features and non-local aggregation features are combined with the objective of obtaining global features. Finally, the SENet module is introduced with the aim of further enhancing the key information in the global features and improving the robustness of the model. The experimental results demonstrate that the method proposed in this paper achieves an average accuracy of 99.63% in a single load scenario and 99.05% in a variable working condition scenario. It exhibits high diagnostic accuracy and a superior capacity for generalization, proves that the KANConv represents a formidable alternative to the existing CNN-based variants for bearing fault diagnosis.