Few-shot fault diagnosis of rolling bearing via mutual centralized learning combining simple and parameter-free attention

The development of deep learning has led to great success in the bearing fault diagnosis. However, the issue of limited fault samples impedes the extensive application of most fault diagnosis approaches based on deep learning. To address this challenge, a new few-shot fault diagnosis method based on mutual centralized learning (MCL) and simple and parameter-free attention (SimAM) is put forward in this paper. First, MCL is adopted to diagnose the bearing fault with small samples, which employs a bidirectional approach rather than the traditional unidirectional method to better learn mutual affiliations between the fault features, having better few-shot classification ability. Furthermore, a new feature extractor module is constructed through the SimAM to improve the feature extraction capability of the MCL model by providing better feature maps for classification. The effectiveness of the proposed method is tested on CWRU bearing dataset and our own bearing dataset. The experimental results show that the proposed MCL-SimAM model can effectively recognize the bearing fault with few samples. Additionally, the comparison experiments demonstrate that the proposed model is superior to the comparable models [relation network (RN), prototypical network (PN), and matching network (MN), deep subspace networks (DSN), and ridge regression differentiable discriminator (R2D2)], which has a better recognition accuracy in few-shot scenarios.