Federated contrastive prototype learning: An efficient collaborative fault diagnosis method with data privacy

Data-driven fault diagnosis approaches have attracted considerable attention in the past few years, and promising diagnostic performance has been achieved with sufficient monitoring data. However, in real industrial scenarios, individual users often struggle to collect enough labeled data. Meanwhile, direct data aggregation from multiple users is not always feasible due to data privacy concerns and conflicts of interest. To solve this issue, a novel federated contrastive prototype learning scheme is proposed for collaborative fault diagnosis of rotating machinery. The collaborative modeling between the central server and multiple clients is implemented to establish a global fault diagnostic model with data privacy. A contrastive prototype learning module is designed to align the prototypes of the same classes across different clients while separating them away from other class prototypes, thus effectively eliminating distribution discrepancies across clients and learning domain invariant discriminative features. To remove the bias of the global model during federated communication, an unbiased prototype learning module is constructed, which aligns the class prototypes of different clients to the global prototype center and enhances the generalization ability of the proposed approach under unseen conditions. Experimental results on two self-built testbeds and a laboratory dataset demonstrate that the proposed approach is a potential solution for real-world fault diagnosis applications.