Intelligent fault diagnosis via ring-based decentralized federated transfer learning

Abstract

Federated transfer learning (FTL) can effectively address the data silos and domain shift that exist in data-driven rotating machinery fault diagnosis (RMFD). However, in FTL used for RMFD, the huge communication overhead, idle waiting between the source clients and the target client, and negative transfer caused by model aggregation are all pressing challenges. Therefore, a ring-based decentralized federated transfer learning (RDFTL) method for intelligent fault diagnosis is proposed. Firstly, a ring-based decentralized federated transfer learning framework is designed, which can be fully integrated with the bandwidth-optimal Ring-AllReduce algorithm, thereby greatly reducing the communication overhead. Secondly, an asynchronous domain adaptation strategy is proposed, which can effectively avoid idle waiting between the source clients and the target client in the collaborative model training, thereby improving the overall training efficiency of FTL. Thirdly, a multi-perspective distribution discrepancy aggregation (MPDDA) strategy is proposed to alleviate the negative transfer caused by model aggregation. The diagnosis performance of the local model of a source client on the target domain is evaluated from the three perspectives of statistical distance, domain adversarial, and stability, and these three evaluation metrics are jointly used to determine the aggregation weights, which can effectively improve the diagnosis performance of the global model. Finally, a series of experiments are carried out to verify the effectiveness of the proposed method. The results demonstrate that the proposed method can obtain a cross-domain fault diagnosis model with excellent performance in RMFD with data privacy at a fast training speed.