Towards Rapid State-of-Health Estimation for Remanufactured Batteries via Stability-Aware Feature Engineering

The transition to electrified mobility has accelerated the need for sustainable End-of-Life (EoL) strategies for battery packs. Among these, remanufacturing and repairing are key approaches to extending battery lifespan and reducing waste. A critical step in this process is the estimation of the State-of-Health (SOH) after remanufacturing, which quantifies the restored capacity and performance of battery packs. While machine learning (ML) has been widely applied to SOH estimation, most models rely on data from lab-aged cells and are optimized for in-use deployment and monitoring. In contrast, real-world EoL batteries present distinct challenges, including unknown usage histories, high variability, and narrow SOH ranges—conditions under which lab-trained models often underperform. Moreover, SOH estimation following remanufacturing has received limited attention, particularly using real-world battery data. This study investigates the role of feature engineering in post-remanufacturing SOH estimation by proposing a scalable framework based on the voltage response to high-current impulse discharges, a fast and non-invasive diagnostic signal. Using a real-world dataset of 425 remanufactured EV battery packs, we combine domain-informed features with thousands of statistical descriptors extracted via TSFresh, followed by a multi-stage selection pipeline that emphasizes both predictive relevance and stability across resampling. Results show that a compact subset of interpretable features, especially those reflecting voltage recovery dynamics, charging variation, and internal resistance proxies, consistently outperforms raw signals. SHAP analysis confirms the model’s reliance on physically meaningful indicators, supporting both predictive power and explainability. Notably, the framework requires no full charge-discharge cycles or impedance measurements, making it highly practical for industrial deployment. These findings demonstrate that robust and interpretable SOH estimation is feasible using impulse diagnostics, offering a viable path for integration into high-throughput remanufacturing workflows.