Federated domain generalization with global robust model aggregation strategy for bearing fault diagnosis

Abstract

Federated learning ensures the privacy of fault diagnosis by maintaining a decentralized and local training data approach, eliminating the need to share confidential information with a central server. However, the performance of trained models tends to significantly deteriorate when applied to completely unseen domains. To address this challenge, this paper introduces a novel method called federated domain generalization with a global robust model aggregation strategy. Our proposed method can collaboratively train a model with outstanding generalization ability and robustness to unseen target domains in a data-protecting way. Specifically, a maximum mean discrepancy is introduced in the central server to reduce the discrepancy of features from different source clients. Meanwhile, classification loss across source domains is designed as the weights for local model aggregation in the central server. Finally, experimental results under two different bearing fault datasets show that our method obtains higher classification accuracy than other compared methods, which demonstrates that the proposed method has better generalization ability and is promising in real industrial applications.