Lithium-ion battery state of health estimation based on statistical features derived from voltage and temperature probability density functions under realistic randomized cycling conditions

Accurate battery state of health (SoH) estimation remains challenging, particularly under the variable load conditions typical of real-world operations. Many data-driven methods rely on features derived from complete constant current or constant voltage cycles, such as peaks and valleys observed in incremental capacity and differential voltage curves. However, such idealized cycles are rarely obtainable in practical applications involving electric vehicles or energy storage systems, significantly limiting the applicability of these features. To address this gap, this study proposes an innovative SoH estimation method based on features extracted from the probability density functions (PDFs) of battery voltage and temperature measurements. Specifically, the NASA random walk (RW) charging and discharging dataset was utilized to realistically represent real-world operational conditions, as the decisions to charge or discharge, as well as current magnitudes, were randomly determined for battery cells RW9–RW11. Analysis revealed that the voltage data consistently exhibited bimodal distributions, while the temperature data were predominantly normally distributed. Maximum likelihood estimation was applied to identify these distribution parameters, which were then combined with other statistical descriptors of the PDFs to construct a comprehensive feature set. A support vector regression model trained using this feature set achieved a mean absolute error (MAE) of 2.63 % and a root mean square error (RMSE) of 3.31 % for SoH prediction. Additionally, downsampling analysis showed that increasing the sampling interval to 5 s resulted in a maximum reduction of 0.58 % in feature correlation and increases of 3.91 % in MAE and 4.26 % in RMSE for SoH estimation. These results demonstrate that the proposed PDF-based approach provides accurate and robust SoH estimation, highlighting its practical value for real-world battery management. Extraction of battery health-related features from identified probability density functions of voltage and temperature data obtained from random walk tests, enabling accurate battery health diagnosis in real-world applications. • SoH estimation under realistic random walk cycling conditions. • Identification of data distributions using maximum likelihood estimation. • Feature extraction from voltage and temperature probability density functions. • Robustness evaluation of feature sensitivity through data downsampling.