Machine learning based window agnostic state-of-health estimation of Li-ion batteries

Accurate estimation of battery state of health (SoH) from partial charging data is essential for practical battery management systems, where complete charge cycles are rarely available. Existing partial-charging methods generally rely on fixed observation windows or allow the window boundary to be implicitly absorbed by the model, making changes in window width difficult to accommodate without retraining. This paper introduces a window-agnostic feature extraction mechanism based on two coupled design choices. A fixed 3.6–3.7 V anchor sub-window, contained within all evaluated partial windows, and the explicit inclusion of the window exit voltage v exit as a model input. Because the anchor sub-window is always contained within the observed window, the features it provides remain invariant to extension of the window, while v exit explicitly conveys the available window extent to the regressor. Together with two capacity-related features normalized by voltage span and two thermal indicators, these quantities form a seven-dimensional physics-guided feature vector that enables a single trained model to estimate SoH across partial windows with exit voltages of 3.9–4.2 V without retraining, latent-space encoding, or curve reconstruction. Eighteen machine learning and deep learning models are evaluated using leave-one-cell-out cross-validation on the Oxford Battery Degradation Dataset, comprising eight NMC cells and 518 characterization cycles spanning 60%–100% SoH. A Gaussian process regressor with a Matérn-5/2 kernel achieves RMSE of 0.69 ± 0.44%, MAE of 0.57%, and R 2 = 0.985 , corresponding to a 47.9% improvement over the reference baseline. The proposed framework requires no differentiation, peak detection, or iterative curve fitting and has negligible estimated inference cost, supporting its potential implementation on resource-constrained battery management hardware.