Enhanced robust capacity estimation of lithium-ion batteries with unlabeled dataset and semi-supervised machine learning

Abstract

The capacity estimation is a crucial task in battery health and safety management. The majority existing capacity estimation methods heavily rely on supervised learning with accurately labelled dataset collected at room temperature. However, the unlabeled dataset and the influence of ambient temperature on lithium-ion batteries degradation are rarely considered in the existing works. To address these issues, semi-supervised machine learning based noise-immune capacity estimation approach is proposed by utilizing both labelled and unlabeled datasets. By considering both the deployment and the accuracy, the optimal health indicator is extracted based on detailed analysis of limited labelled dataset. Then, the semi-supervised machine learning is proposed to improve the capacity estimation by using abundant unlabeled dataset. Specifically, low-complexity network with different training algorithms are employed to map the nonlinear relationship between the feature and capacity with the total loss of labelled and unlabeled datasets. By further exploiting the unlabeled dataset, the maximum errors of capacity estimation at three different temperatures are improved by 85%, 13%, and 48% with strong robustness, respectively. The proposed approach shows superior performance to the state-of-the-art supervised methods. The encouraged performance indicates the effectiveness of using large unlabeled battery dataset to improve the capacity estimation in real-world applications.