Bearing Remaining Useful Life Prediction Using Personalized Soft Aggregation in Federated Learning

Abstract Data-driven bearing prognostic methods have achieved significant success, but existing methods often rely on a single data resource, which overlooks the practical scenario in industrial settings where condition monitoring data is usually distributed across multiple locations and generated under diverse operating conditions. To address this issue, this paper proposes a bearing remaining useful life (RUL) prediction method using personalized soft aggregation in federated learning, called PSA-FL. The proposed method involves a central server and multiple clients, where each client possesses monitoring data collected under a specific working condition. Specifically, clients conduct local training of their models and upload them to the central server. The server then uses the personalized soft aggregation algorithm to create an ensemble model for each client, taking advantage of the aggregated degradation features contained in heterogeneous data scenarios. Subsequently, the ensemble model is returned to each client for further rounds of local training. By iteratively repeating the process of local training and personalized soft aggregation, each client obtains its personalized prognostic model. Experiments with bearing data show that the PSA-FL method is effective in RUL prediction tasks. Additionally, the PSA-FL method demonstrates inherent robustness by successfully limiting client drift in real-world settings.