Federated Learning with a Temporal-Degradation-Aware Transformer for Robust Joint SOC and SOH Estimation in Electric Vehicle Batteries

Accurate estimation of State-of-Charge (SOC) and State-of-Health (SOH) is critical for the safe and reliable operation of Electric Vehicle (EV) batteries. However, centralized approaches face challenges such as data privacy, communication overhead, and poor scalability under non-independent and identically distributed (non-IID) data. Although some methods attempt joint estimation, most rely on centralized learning, lack computational efficiency for embedded systems, and overlook uncertainty quantification, limiting real-world applicability. To address these limitations, this paper proposes a federated framework for joint SOC and SOH estimation that is privacy-preserving, lightweight, and uncertainty-aware. The core model is a Temporal-Degradation-Aware Informer (TDA-Informer), which incorporates an Informer-Lite encoder and a multi-task prediction structure. A Battery-Focused Multi-Feature Attention (BFMFA) module emphasizes critical inputs such as voltage, current, and temperature. The Stochastic Controlled Averaging for Federated Learning (SCAFFOLD) algorithm is employed to mitigate client drift under non-IID conditions, enhancing convergence across distributed EV clients. Uncertainty Quantification is achieved via Monte Carlo Dropout to provide confidence intervals for SOC and SOH predictions. The framework is trained and validated on the NASA battery dataset and tested for generalization on the unseen Oxford dataset, with realistic sensor noise added to simulate practical conditions. Extensive experiments, including ablation studies and comparisons with baseline models, confirm its superior accuracy, robustness, and scalability. While physical deployment remains future work, the results confirm the framework’s potential for intelligent, scalable Battery Management Systems (BMS).