Distribution domain health descriptor fusion for robust battery state of health estimation

Abstract

Robust state of health estimation of lithium-ion batteries requires health descriptors that are insensitive to transient signal fluctuations while remaining sensitive to degradation evolution. To this end, this study proposes a distribution-domain health descriptor fusion framework for battery state of health estimation. Charge and discharge measurements are first transformed into probability density distributions through nonparametric kernel density estimation. The modal peaks of the estimated distributions are then extracted as compact health descriptors, representing the dominant operating state of the battery within each cycle and reducing the influence of local disturbances. To enhance degradation relevance, correlation analysis and sensitivity analysis are jointly used to select informative descriptors. The selected descriptors are further integrated through an adaptive fusion strategy to construct a unified composite health indicator from complementary charge and discharge information. Experiments on two public lithium-ion battery aging datasets show that the proposed method achieves consistently accurate state of health estimation and outperforms representative comparison methods, with root mean square errors ranging from 0.22% to 1.36%. The results demonstrate that distribution-domain descriptor construction provides an effective and robust feature extraction route for battery health estimation from routinely measurable operational data.