Existing data-driven methods for predicting the remaining useful life of bearings often rely on data from a specific operating condition to train the corresponding prediction model. The valuable degradation features present in data from other conditions are disregarded. To effectively capture and utilize degradation features across diverse operating conditions, this article proposes a personalized federated learning-based method for bearing remaining useful life prediction. In this method, monitoring data from bearings under different conditions are distributed among multiple clients, while a central server collaborates with these clients to develop personalized prediction models by model transfer, combination, and local updates. To integrate the global model aggregated by the central server with the local model, an adaptive local combination algorithm is introduced, which preserves useful degradation features that aid in initializing the client′ s model and enhancing prediction performance. The proposed method is evaluated by using two datasets of bearings. The results show its ability to construct high-performance prediction models for bearings operating under different operating conditions. In comparison to local training method, this method manifests a minimum decrease of 13% in root mean square error.