Suggested Citation:
Lin, Guangyang, Hui Pang, Jiarong Du, Xudong Qu, Yushan Leng, Jingyuan Zhao (2026)

Observer-Guided Deep Learning for Joint State of Charge and State of Health Estimation of Lithium-Ion Batteries

Accurate and robust estimation of the state of charge (SOC) and state of health (SOH) is fundamental for the safety, efficiency, and lifetime management of lithium-ion batteries, yet remains challenging because of their strongly coupled and time-dependent dynamics. Here we present a hybrid data–physics framework that integrates a dual extended Kalman filter (DEKF) with a bidirectional gated recurrent unit enhanced by self-attention (STBiGRU). The DEKF captures dynamic state variations and generates physically consistent features, which are then fused with deep temporal representations to jointly estimate SOC and SOH. A cycle-feature method further quantifies SOH by linking the SOC drop rate to capacity fading. Across dynamic drive cycles (4C, US06, DST, FUDS), the framework reduces SOC estimation error by up to 60% in RMSE relative to GRU baselines, while under cycle-aging conditions it achieves <2% RMSE in SOC and < 4% error in SOH. These results demonstrate that coupling explainable circuit models with attention-based recurrent networks can overcome error accumulation and feature extraction limits of existing methods. Beyond methodological advances, this study provides a scalable pathway toward physics-informed battery management systems capable of safe, accurate, and adaptive monitoring in electric vehicles and grid applications.