Pulse-based state-of-charge and state-of-health estimation of randomly retired lithium-ion batteries for second-life applications

Retired lithium-ion batteries are an important resource for second-life energy storage, but their practical reuse requires rapid and efficient state assessment under unknown usage histories. To address this problem, we develop a pulse-based framework for joint state-of-health (SOH) and state-of-charge (SOC) estimation of randomly retired lithium-ion battery cells. A set of voltage-response features is extracted from short bipolar pulse tests and used as diagnostic descriptors. Based on these features, two machine learning models, extremely randomized trees (ExtraTrees) and a tabular prior-data fitted network (TabPFN), are employed for battery state estimation and evaluated using data from 270 retired cells that cover three chemistry types and four capacity classes. The results show that TabPFN clearly outperforms ExtraTrees and achieves mean absolute errors ranging from 0.0076 to 0.0306 for SOH estimation and below 0.01 for SOC estimation across all cell groups. To reduce the testing burden, a subset of SOC-dependent pulse tests is removed, and the missing feature points are reconstructed using interpolation methods. The reduced-test results show that near-baseline estimation accuracy can be retained even after removing 40%–70% of SOC-dependent pulse tests, depending on the target state and accuracy requirements. These results demonstrate a practical route toward faster, lower-cost, and scalable state assessment for second-life battery applications.