Federated Clustering of Electric Vehicle’s Usage Patterns for Personalized State-of-Health Estimation

Abstract

Battery-health degradation varies significantly among electric vehicles, complicating accurate fleet-wide monitoring. This study investigates key factors influencing battery ageing and incorporates them into a dynamic clustered federated-learning framework, combining K-means initialization and Affinity Propagation (AP) clustering. Evaluated on NASA laboratory cells and telemetry from nine electric vehicles used in heterogeneous driving conditions, the approach achieves faster convergence and lower prediction errors compared to single trained models. Whereas clustering solely on ambient temperature profiles produced the best results, highlighting temperature as the clearest indicator for personalized state-of-health estimation. Adding more features led to overlapping clusters and reduced model performance. Despite room for improvement in the clustering algorithm and the limited dataset, the results demonstrate promising potential. Recommended future steps include implementing richer sequence models, automated outlier filtering, and validation in larger electric vehicles fleets to provide robust, privacy-preserving battery diagnostics in real-world scenarios.