A Contribution-Aware Federated Framework for Electric Vehicle Batteries Health Estimation

Abstract

Accurate estimation of battery health can significantly enhance the safety of electric vehicle (EV) systems. Most machine learning methods employ centralized learning for battery state of health (SOH) estimation. These approaches lead to weak model generalization due to limited samples caused by data privacy, hindering adaptation to varying working conditions. Batteries of the same batch may show manufacturing variances, further constraining model generalization. Additionally, most methods only focus on the temporal nature of battery degradation, they overlook the significant impact of periodic fluctuations in capacity. Existing data-driven methods struggle with low generalization capabilities, manufacturing variances among batteries, and neglect of periodic capacity fluctuations. Therefore, we propose an advanced federated learning framework that combines a contribution-aware federated strategy (CAFS) with battery health predictor (BHP) models to more accurately estimate a battery’s SOH. Specifically, we design the BHP, consisting of a feature-enhanced autoencoder and a temporal period coupled attention mechanism, to effectively learn battery ageing information. Furthermore, we present the CAFS to significantly enhance the model’s generalizability and adaptability to the target battery. The experimental results show that our method achieves an average improvement of 20.18% in SOH estimation accuracy across diverse working conditions. Results from eight federated scenarios demonstrate that our method outperforms conventional federated learning in terms of accuracy, generalization, and training speed.