Decoding battery aging in fast-charging electric vehicles: An advanced SOH estimation framework using real-world field data

• Vehicle battery SOH estimation framework is developed via a user charging behavior-driven feature engineering technique. • Customized data window approach mitigates the variability in EV user behaviors. • Hierarchical feature extraction at the vehicle-pack-cell levels enhances battery aging decoding. • Minimal data is required for training, with a reliable estimation using data from just one vehicle. Accurately estimating the state of health (SOH) of in-vehicle batteries is critical for advancing electric vehicle (EV) technology. However, higher charging rates and more complex driving conditions have posed major challenges, with significant variations from vehicle-to-vehicle and cycle-to-cycle. In this study, we developed a SOH estimation framework to monitor battery capacity degradation, in EVs with multi-step constant-current fast charging and voltage balancing technology. The framework employs a customized data window approach, informed by a thorough analysis of EV charging behavior, and extracts hierarchical features from vehicle-, pack- and cell-levels for tracking battery aging. We collected real-world charging data from 300 pure EVs over 1.5 years, resulting in 193,180 samples for validation. The best-performing machine learning models achieved an absolute error of less than 2 % for 93.7 % of samples, a root mean square error (RMSE) of 1.05 %, and a maximum error of only 3.73 % whilst using only 30 % data for training. Our analysis indicates that the proposed model can be effectively developed without the need to pre-select vehicles based on specific driving habits or operating conditions. Notably, reliable and accurate estimations were produced using data from just one vehicle, achieving an RMSE of 1.82 %. Our results highlight the potential of user behavior-assisted feature engineering to decode battery pack aging under dynamically changing vehicle profiles. This work underscores the promise of developing accurate SOH estimation modules for battery management systems using minimal vehicle data. We collected a comprehensive dataset of 193,180 real-world charging events from 300 pure electric vehicles (EVs), meticulously recorded over 1.5 years. Through a thorough analysis of EV charging behaviors, we tailored data windows for frequency-based battery health evaluations. Advanced multi-level feature extraction was performed at the vehicle, pack, and cell levels within the customized data window, yielding high accuracy predictions with a RMSE of 1.05 % using 30 % data, 1.13 % using 5 % data, and impressively 1.82 % using data from only ONE vehicle. Our approach underscores the potential of user behavior-assisted feature engineering to effectively decode real-world EV battery degradation.