Federated Transfer Learning for Intelligent Fault Diagnostics Using Deep Adversarial Networks with Data Privacy

Abstract

Intelligent data-driven machinery fault diagnosis methods have been popularly developed in the past years. While fairly high diagnosis accuracies have been obtained, large amounts of labeled training data are mostly required, which are difficult to collect in practice. The promising collaborative model training solution with multiple users poses high demands on data privacy due to conflict of interests. Furthermore, in the real industries, the data from different users can be usually collected from different machine operating conditions. The domain shift phenomenon and data privacy concern make the joint model training scheme quite challenging. To address this issue, a federated transfer learning method for fault diagnosis is proposed in this study. Different models can be used by different users to enhance data privacy. A federal initialization stage is introduced to keep similar data structures in distributed feature extractions, and a federated communication stage is further implemented using deep adversarial learning. A prediction consistency scheme is also adopted to increase model robustness. Experiments on two real-world datasets suggest the proposed federated transfer learning method is promising for real industrial applications.