Hybrid Physics-Informed UKF–Transformer Framework for Lithium-Ion Battery SOH Estimation

Abstract

The accurate estimation of the State of Health (SOH) of lithium-ion batteries is essential for ensuring reliable operation and enabling prognosis in energy storage systems. Model-based approaches such as the Unscented Kalman Filter (UKF) provide physically interpretable estimates and stability properties that can be analyzed under standard modeling and noise assumptions. However, their performance is constrained by the need for an explicit and accurate battery model, as well as careful tuning of process and measurement noise covariances. As a result, standalone UKF implementations may strugglw in the presence of nonlinear aging effects, parameter drift, and real-world operating variability. On the other hand, data-driven approaches—particularly transformer-based architectures—excel at modeling nonlinear systems and capturing long-range temporal dependencies. However, they typically require large and diverse datasets, are sensitive to the scenario distribution used during training, and may lack stability or physical interpretability. Despite these challenges, transformers can reduce the overall data requirement by efficiently learning both temporal relationships and cross-feature interactions within the input signals. In this context, we propose a hybrid physics-informed framework that combines a UKF with a transformer model. The UKF provides a physically grounded intermediate SOH estimate, while the transformer compensates for unmodeled dynamics and nonlinear degradation patterns. The transformer further improves the estimator’s efficiency by capturing temporal dependencies and cross-feature relationships, which reduces the amount of training data required while maintaining robustness and generalization. Using the strengths of both techniques, the proposed hybrid approach improves estimation accuracy, robustness to noise, and generalization capability compared to either method used alone.