부산대학교 도서관

Lithium-ion battery remaining useful life (RUL) is a key parameter on battery management system. Many machine learning methods are applied to RUL predictions, but they generally suffer from two limitations: (i) the extracted features fail to reflect the information hidden in the historical degradation status, and (ii) the accuracy cannot be guaranteed in the evaluation of battery degradation due to the non-linearity. In this paper, a new prediction method is proposed combining the time window (TW) and Gradient Boosting Decision Trees (GBDT). First, the energy (VCE) and the fluctuation index (VFI) of voltage signal are verified and selected as features. Then, a TW based feature extraction method is designed to extract features from the historical discharge process. After that, GBDT is adopted to model the relation of features and RUL. The proposed method is implemented on a recognized battery degradation dataset, and the advantages in accuracy are proven.