Physics-Informed Deep Gated Recurrent Unit Network for Lithium-Ion Battery State of Health Estimation Under Field Conditions

Accurate estimation of lithium-ion battery state of health (SOH) in electric vehicles (EVs) under real-world conditions is much more challenging than using well-designed laboratory cycling data due to unreliable SOH labeling and segmented charging behavior, and the results often lack interpretability. To address these issues, this paper proposes a physics-informed deep gated recurrent unit (PIDGRU) architecture for robust and interpretable SOH estimation without requiring explicit physical modeling. First, a modified inverse ampere-hour integral method is combined with the Bayesian estimator of abrupt, seasonality, and trend (BEAST) algorithm to estimate battery capacity and characterize SOH uncertainty. A universal feature extraction and selection framework is then developed to handle segmented EV charging data, utilizing a hybrid linear-nonlinear redundancy analysis to ensure an optimal input feature set. The PIDGRU integrates empirical degradation modeling and nonlinear dynamic degradation learning through a deep gated recurrent unit network, utilizing a deep hidden physics model (DeepHPM). A Bayesian inference uncertainty-constrained (BIUC) strategy is introduced to enhance training reliability and uncertainty quantification. Extensive evaluations on both in-vehicle and cross-vehicle datasets demonstrate that the proposed method achieves high accuracy, robustness, and generalizability, with the mean absolute error and root mean squared error consistently below 1.10% and 1.31%, respectively.