Federated Transfer Learning for Bearing Fault Diagnosis Based on Averaging Shared Layers

Existing data-driven machinery fault diagnosis methods can obtain high diagnosis accuracy under the condition of abundant labeled data. However, in the actual industrial environment, complete and high-quality training data may often be distributed on multiple mechanical equipment of different regions or institutions, so-called an isolated data island problem. It is often difficult to integrate and utilize these datasets due to limitation of legal regulations or interest conflict, such as privacy protection, security risk and industry competition. Therefore, how to effectively use the separated data of multiple participants to jointly train a reliable intelligent fault diagnosis model is an urgent challenge. To address this problem, a federated transfer learning method based on averaging shared layers for bearing fault diagnosis is proposed in this study. A server-clients architecture with multiple deep transfer networks is constructed to jointly learn the global features from isolated datasets. Then, a modified federated averaging method based on shared layers is adopted to implement federated averaging of distributed feature layers from different diagnosis models, and personalized layers are updated locally. Three different bearing datasets collected by different devices are used for experimental verification. Compared with the current popular federated learning schemes, the experiment results demonstrate the effectiveness and superiority of the proposed method.