Dual prototypical contrastive network: a novel self-supervised method for cross-domain few-shot fault diagnosis

Data-driven methods have pushed mechanical fault diagnostics to an unprecedented height recently. However, their satisfactory performance heavily relies on the availability of abundant labeled data, which poses a challenge given the scarcity of fault data in industrial scenarios. In this paper, a novel self-supervised method named dual prototypical contrastive network (DPCN) is proposed for cross-domain few-shot fault diagnosis. The proposed method contains two different prototypical contrastive learning stages. Specifically, in the first stage, intra-domain prototypical contrast guides the model to learn class-wise discriminative features by enhancing prototype-instance compactness within the same domain. Subsequently, in the second stage of cross-domain prototypical contrast, the model learns domain-invariant features by realizing cross-domain prototype-instance matching and proximity. Besides, mutual information maximization is applied to ensure the reliability of the predicted result. We undertake few-shot fault diagnosis experiments involving cross-load and cross-speed scenarios in two case studies. The extensive experimental results validate the superiority of the proposed method compared with the comparative methods.