A physics-guided transformer for robust state of health estimation in real-world electric vehicles

Accurate prediction of the state of health (SOH) for lithium-ion batteries is crucial for ensuring the safe operation and maintenance decisions of electric vehicles (EVs). However, real-world operation data are often susceptible to significant interference from dynamic load noise, which diminishes the observability of aging characteristics. Furthermore, existing physics-guided approaches often face conflicting optimization objectives between data fitting accuracy and physical consistency during joint training. To address these challenges, a Physics-Guided Multi-Task Constrained Deep Learning Model (P-MCDM) is proposed. First, a feature refinement module integrating Inverse Gaussian distribution modeling and a Strong Tracking Extended Kalman Filter (STEKF) is constructed to decouple intrinsic electrochemical features from noisy cell-voltage and capacity sequences. Second, a Physics-Semantic Decoupling PatchTST (PSD-PatchTST) encoder incorporating Mechanism-Guided Cross-Attention (MG-CA) is developed, which performs channel-independent patching over event-level feature sequences to facilitate semantic alignment between physical priors and data representations. Finally, PCGrad is introduced to mitigate gradient conflicts between the data-fitting and physical regularization objectives, supporting learning that is consistent with the macroscopic degradation trend of battery aging. Experimental validation on a large-scale dataset of 300 real-world EVs shows that P-MCDM achieves an RMSE of 2.16% while maintaining physical consistency under complex operating conditions.