WRNN and CAT-Based Few-Shot Learning Fault Diagnosis for Bearing

This paper proposes a fault diagnosis method based on Few-shot Learning(FSL) for bearings. Acquiring adequate fault samples for bearing diagnostics remains challenging in real-world industrial settings. The proposed diagnostic framework demonstrates strong capability in addressing the challenge of limited samples in actual industrial settings. The FSL fault diagnosis method combines with Continuous Wavelet Convolution( CWConv), Weight-Reduce K-Nearest Neighbor(WRNN) and Cross Attention Transformer(CAT). Firstly, the CWConv is employed to capture features from limited fault diagnosis samples (1-shot, 5-shots). The features of the few samples are learned to achieve better fault diagnosis results. Secondly, in the module dedicated to learning global features, the consumption of computing resources is decreased while the time for model training is saved by reducing the dimensions of the input data to lightweight CAT. Thirdly, in the module dedicated to learning local features, this paper proposes a new K-Nearest Neighbor(KNN) method for local feature similarity learning, which is called WRNN. This method dynamically filters out redundant samples and subsequently assigns distance-based weights to the preserved samples for precise similarity evaluation. Finally, the proposed method is applied to the bearing to verify the accuracy of fault diagnosis. The experimental results show that the accuracy of fault diagnosis can reach more than 90% under both 1-shot and 5-shots conditions.