Temperature-resilient state-of-health estimation for electric vehicles using domain-corrected fast-charging features

초록

Reliable battery management systems (BMSs) are essential for electric vehicle safety, but their performance often degrades under variable operating conditions. Accurate state-of-health (SOH) estimation remains challenging because of limited labeled data and distribution shifts induced by diverse fast-charging and temperature profiles. This study proposes an AI framework for improving generalization through structured feature learning, temperature-calibrated domain correction, and physics-informed neural networks (PINNs). Based on reference performance tests (RPTs) and reproduced charging logs, 31 health indicators (HIs) are extracted. Feature selection identifies an 11-HI set (F11), which is further reduced to a temperature-calibrated 7-HI set (F7) through regression to a 25 °C reference. The PINN achieves a low absolute percentage error (APE) of approximately 0.2% in an unseen-vehicle validation case, yielding a smaller error than the commercial diagnostic scanner in that case. Under representative real-world monitoring with rapid temperature variations, temperature calibration reduces the APE from 0.9% to 0.3% in the tested field case. These results suggest that domain-calibrated PINNs can support practical battery diagnostics using BMS monitoring data under temperature-varying fast-charging conditions, while broader multi-vehicle validation is required for generalization.

키워드