Fault diagnosis based on federated learning driven by dynamic expansion for model layers of imbalanced client

Federated Learning is a promising tool for fault diagnosis of critical components for electrical driving systems. However, the performance of existing method is limited by negative transferring owing to inter-client inhomogeneity related to the sample size, sample quality, and fault type. How to make balanced client benefit from imbalanced client who involves unique fault information in its poor data is crucial to encourage the balanced clients to take an active part in federation. This study aims to design an imbalanced federated learning framework based on dynamic expansion for model layers of an imbalanced client to make it powerful enough to make it powerful in extracting the unique fault feature involved. The proposed method can ensure that balanced client be benefited from imbalanced client by designing a collaboration mechanism of inter-client layer-by-layer federation and intra-client cross-scale recursive fusion. The effectiveness of the proposed method was validated using both the benchmark dataset and the data set of an actual project of Nio Inc. When there is significant inhomogeneity between clients, the proposed method can achieve an accuracy increase of 13.97% compared to existing methods.