Federated adversarial domain generalization network: A novel machinery fault diagnosis method with data privacy

Abstract

Domain generalization (DG) methods have been successfully proposed to enhance the generalization ability of the intelligent diagnosis model. However, these methods hardly achieve data privacy-preserving generalization diagnosis tasks due to potential conflicts of interest or data privacy protocol. To this end, this study proposes a novel federated adversarial DG network for machinery fault diagnosis. The collaborative training between the central server and several clients is implemented in the proposed network, which aims to build a global fault diagnosis model for multiple clients under data privacy conditions. To eliminate the distribution discrepancies of different clients, a multiclient feature alignment module with adversarial learning is designed. In this module, the generative adversarial network with class-wise information is introduced to generate a reference distribution adaptatively, and the adversarial training strategy between reference distribution and real distribution is implemented to learn the DG features from different clients. Experimental results on three cases show that the proposed network can achieve more than 10% performance improvement for diagnosis methods with federated learning. The small performance gap between the proposed network and the no-federated learning method suggests that the proposed network is promising for solving data privacy-preserving generalization diagnosis tasks.