Federated transfer learning with consensus knowledge distillation for intelligent fault diagnosis under data privacy preserving

Fault diagnosis with deep learning has garnered substantial research. However, the establishment of a model is contingent upon a volume of data. Moreover, centralizing the data from each device faces the problem of privacy leakage. Federated learning can cooperate with each device to form a global model without violating data privacy. Due to the data distribution discrepancy for each device, a global model trained only by the source client with labeled data fails to match the target client without labeled data. To overcome this issue, this research suggests a federated transfer learning method. A consensus knowledge distillation is adopted to train the extended target domain model. A mutual information regularization is introduced to further learn the structure information of the target client data. The source client and the extended target models are aggregated to improve model performance. The experimental results demonstrate that our method has broad application prospects.