A novel aging characteristics-based feature engineering for battery state of health estimation. (15th June 2023)

• Record Type: Journal Article
• Authors: Wang, Jinyu; Zhang, Caiping; Zhang, Linjing; Su, Xiaojia; Zhang, Weige; Li, Xu; Du, Jingcai
• Is Part Of: Energy. Volume 273(2023)
• Publication Date: 2023-06-15
• Subjects: Lithium-ion battery -- Feature engineering -- Aging features -- Feature selection -- State of health -- Machine learning; Power resources -- Periodicals; Power (Mechanics) -- Periodicals; Energy consumption -- Periodicals; 333.7905
• DOI: 10.1016/j.energy.2023.127169
• Languages: English
• ISSNs: 0360-5442

Abstract

Abstract: State of health (SOH) estimation is essential for lithium-ion battery systems to ensure safe and reliable operation. The existing SOH estimation considers a few available signals, such as voltage and current, to extract specified and limited capacity-related features. Once the cell or materials is changed, features require manual re-built as the construction is specific and unsystematic. This paper proposes a novel aging information-based feature engineering framework for SOH diagnosis, which combines a comprehensive feature library driven by three-step construction strategy and an automatic feature selection pipeline fused with embedded-based and filter-based methods. In the feature space, the role played by each feature type and the extent to which the combination of features affects SOH estimation are explored by accuracy and robustness. For the collected datasets, a library of 206 features is generated as inputs for feature selection which eventually output a space with 7 features to track SOH change. These features perform well under all three typical machine learning models, with the maximum absolute error within 1% and the root mean square error (RMSE) below 0.29% for all cells of transfer operations. Compared to the existing literature using the features of discharge capacity differences between two cycles [ΔQ(V) curve], the RMSE is reduced by up to 85.1%. The approach is automated to produce a highly robust feature subset for accurate SOH estimation across usage protocols and multiple battery chemistries due to the wide range of feature sets and the superiority of feature selection. Highlights: A comprehensive feature library driven by a three-step construction method is generated. An automatic feature selection pipeline for generating an algorithm-free feature subset is developed. The performance of feature space-based SOH estimation is verified by three machine learning models on cross-service cells. The role played by each feature type and the influence of feature combination on SOH are investigated.