Federated transfer learning for remaining useful life prediction in prognostics with data privacy

Abstract Collaborative model training with multiple clients is becoming an effective solution for prognostic problems, due to the scarcity of the machine run-to-failure data in the real industries. However, direct data sharing and centralized learning are usually not feasible in practice, since the private local data basically cannot be exposed to the other commercial clients. Furthermore, the machines at different clients mostly have different degradation patterns and failure modes, resulting in different data distributions. That poses great challenges for data-driven knowledge transfer across clients with data privacy. To address these issues, this paper proposes a federated transfer learning method for remaining useful life predictions. The proposed prior alignment and feature adaptation schemes can achieve extraction of shared features across domains without simultaneous processing of the source and target data. The availability of the target-domain data in the whole life cycle is not required by the proposed method, which enhances the model applicability. Experiments on prognostic datasets are carried out for validations, and the results suggest the proposed method is promising for the federated transfer learning problems in the real industries.