A battery SOH estimation method based on entropy domain features and semi-supervised learning under limited sample conditions

Abstract

Accurately estimating the battery's state of health (SOH) is critical for battery efficiency and stability. Despite significant progress in data-driven methods, the accuracy of these models is limited by feature extraction strategies and the scarcity of dataset samples. To address this issue, this study develops a battery SOH estimation model tailored to the limited sample conditions. A refined composite multiscale discrete sine entropy (RCMDSE) algorithm is proposed, which combines composite multiscale approaches, Shannon entropy theory, and the discrete sine transform. This algorithm is designed to extract high-quality battery entropy domain health features (HFs) from current and voltage signals at various scales and levels. Subsequently, we introduce semi-supervised learning concepts to enhance the estimation performance of the nu-support vector regression (NuSVR) algorithm in limited sample conditions. The golden jackal optimization algorithm (GJO) is used to improve the estimation accuracy of the NuSVR algorithm in a semi-supervised framework. Comparative and ablation experiments on four datasets validate that the battery SOH estimation model maintains RMSE and MAPE values of <1 %, even when trained with only 10 % of the data. Furthermore, the proposed RCMDSE algorithm outperforms and is more robust in HF extraction than the widely used incremental capacity (IC) curve feature extraction method.