A semi-supervised prototypical network with dual correction for few-shot cross-machine fault diagnosis

Abstract Meta-learning has been widely applied and achieved certain results in few-shot cross-domain fault diagnosis due to its powerful generalization and robustness. However, existing meta-learning methods mainly focus on cross-domain fault diagnosis within the same machine, ignoring the fact that there are more significant domain distribution differences and sample imbalance problems between different machines, leading to poor diagnostic performance. This paper proposes a semi-supervised prototypical network with dual correction to address these issues. First, a dual-channel residual network is utilized to comprehensively extract sample features, capturing deep and shallow information. Then, correct the semi-supervised prototypical network by weighting the features and adding a shift term on support set samples and query set samples, respectively, to diminish its intra-class and extra-class bias. Meanwhile, a regularization term is introduced into the model to balance the distribution among different class prototypes, enhancing distinctiveness. Finally, few-shot cross-machine fault diagnosis experiments are conducted on three datasets to validate the method’s effectiveness. Additionally, an interpretability analysis of the model is conducted using the gradient-weighted class activation mapping (Grad-CAM) technique to discern its primary regions of focus in the classification tasks.