Reproducibility of battery state of health estimation using short discharge tests across independent experimental platforms

Accurate estimation of battery state of health (SOH) is essential for practical battery diagnostics and for screening retired lithium-ion batteries for second-life applications. Although short-duration discharge-based data-driven approaches show promising performance, their reproducibility across independently acquired datasets remains insufficiently explored. This study evaluates a previously developed short-duration discharge-based cell-level SOH estimation framework using two independently acquired datasets generated under methodologically matched aging and diagnostic procedures. The same commercial cell model, aging procedure, diagnostic protocol, preprocessing method, and LSTM-based model architecture are used for both datasets. Within each dataset, short partial-discharge voltage features maintain a stable relationship with SOH, enabling reliable estimation under same-dataset learning conditions. In contrast, direct cross-dataset application of the learned relationship leads to less consistent estimation behavior. Further analysis reveals noticeable differences in voltage characteristics between the two datasets despite the use of the same cell model and diagnostic protocol. To account for this variability, an integrated learning condition is also examined by combining the two datasets into a unified training set. The present findings were obtained using the cells and experimental conditions examined in this study, and further validation under broader cell populations and operating conditions is needed to establish wider applicability.