Tokenizing battery degradation: a physics-informed and uncertainty-aware framework for state-of-health estimation

Accurate and trustworthy estimation of State of Health (SOH) is essential for the safe and efficient operation of modern battery management systems. While data-driven methods have achieved high accuracy, they often lack physical grounding and fail to provide reliable confidence metrics for their predictions, posing risks in safety-critical applications. In this work, we introduce TBD-SOH, an SOH estimation framework that combines physics-informed feature construction with uncertainty-aware regression. Specifically, each charging cycle is represented as a fused cycle-level token constructed from two complementary sources: statistical descriptors of charging behavior and a hardware-free pseudo-impedance representation derived from routine voltage and current measurements in the frequency domain. The pseudo-impedance branch is introduced as a degradation-sensitive proxy rather than a direct electrochemical impedance spectroscopy measurement. In addition, an MC dropout-based regression head is adopted to provide confidence intervals together with point estimates. Comprehensive experiments on the public MIT and HUST battery datasets demonstrate that TBD-SOH achieves high accuracy, outperforming conventional deep learning models. More importantly, its uncertainty estimates are shown to be dynamically responsive to prediction difficulty, and provide a new reference point for developing trustworthy and reliable battery health monitoring systems.