Federated domain generalization for intelligent fault diagnosis based on pseudo-siamese network and robust global model aggregation

Federated learning (FL) based intelligent fault diagnosis has developed rapidly in recent years owing to the need for data privacy. However, models trained using FL may suffer from performance degradation when applied to unseen domains. In this regard, we propose a federated domain generalization approach using a pseudo-Siamese network (PSN) and robust model aggregation for intelligent fault diagnosis. Firstly, the proposed method employs PSN to calculate the discrepancy between client and global models at the local clients. This enhances the feature space boundary of fault diagnosis models. Then the proposed method computes cross-classification losses of locally trained global models on the central server for robust model aggregation. Finally, we evaluate our approach through experiments where local clients contain data from varying datasets. Experimental results on the proposed method and other transfer learning and federated learning methods prove the outperformance of the proposed method.