Personalized local initialization based federated learning for bearing remaining useful life prediction

Existing bearing remaining useful life (RUL) prediction methods often rely on specific datasets to train models optimized for particular operating conditions, disregarding samples from different conditions. Given the similarities in deterioration mode and vibration characteristics among bearings, common degradation features can be beneficial across machines. To enhance prediction models by leveraging information from diverse datasets, we propose a personalized local initialization based federated learning (PLI-FL) method. This method aims to develop individualized RUL prediction models for bearings operating under different conditions. The framework involves a central server and multiple clients, with each client possessing monitoring data generated under a specific working condition. Initially, clients train their local models, which are then aggregated by the server using federated averaging to create a global model. Clients subsequently combine the global model with their local models in an adaptive manner to establish personalized initialized models for the next round of local training. Through several iterations of local combination, training, and central aggregation, each client retains desirable degradation features from the global model and individualized features from the local model. Thus, personalized models are constructed for clients in a global model-assisted manner. Experimental results using bearing data demonstrate the effectiveness of the PLI-FL method in improving RUL prediction.