A Bearing Fault Diagnosis Framework Based on Few-Shot Learning with Distribution Consistency and Structural Reparameterization

Nowadays, end-to-end deep learning networks inherit the development of convolutional neural networks to have good performance, but they can only do well when they are learned on a large number of samples. However, building a large dataset is difficult for specific data, such as bearing faults. In this paper, a novel end-to-end few-shot model is proposed for bearing fault diagnosis even in the case of limited data with artificial and real faults. The architecture is designed to optimally exploit the feature extraction process as well as the covariance metric based on the distribution consistency for the few-shot classification tasks. In particular, we propose a module for the feature extraction stage with structural re-parameterization combined with a self-attention mechanism. Specifically, the input data goes through the STEM block and is then put into three stages, which consist of two stages called RepMixer block using re-parameterizing skip connections and the other stage called Attention block. We also use a covariance metric to help the feature extractor learn more precise features and enhance its discriminative performance during training. Experimental results based on the public imbalanced dataset CWRU, with different degrees of training samples, show that our model can significantly improve the accuracy of identifying faults-bearing to guarantee safety in industrial systems. Our code is released at https://github.com/giabao804/few-shot-structural-rep .