A credible granular feature contrastive learning network for fault diagnosis with semi-supervised domain generalization

Fault diagnosis methods based on domain generalization (DG) have gained significant attention for their ability to improve model generalization and perform fault diagnosis in unknown target domains. However, in practical applications, unlabeled data is more common, making it crucial to enhance model generalization using unlabeled data. In this paper, we propose a credible granular feature contrastive learning network (CGFCLN) for semi-supervised DG (SSDG) in fault diagnosis of rotating machinery, which further enhances the practicality of DG methods. CGFCLN leverages both deep features and credible labels to fully utilize unlabeled data, thereby improving model performance. First, CGFCLN contrasts deep features from different source domains, enabling the model to extract domain-invariant features from granular features. Additionally, CGFCLN contrasts deep features and final outputs of both original and augmented data, which helps improve the robustness of the model. Second, we introduce credible pseudo-labels, which minimize the negative impact of incorrect pseudo-labels on the model and enhance classification and generalization performance. To validate the effectiveness of CGFCLN, extensive experiments were conducted on two rotating machinery datasets, achieving average accuracies of 93.85% and 92.70%, respectively. Compared to existing SSDG methods, CGFCLN demonstrates superior diagnostic and generalization performance.