Improving state-of-health estimation for lithium-ion batteries via unlabeled charging data

Abstract

The state-of-health (SOH) estimation is an important and open issue in battery health management. Most existing data driven SOH estimation methods are based on supervised learning algorithms, relying on large and precious labeled data. However, unlabeled charging data are abundant and readily available, but are rarely used to estimate SOH. To solve these problems, a semi-supervised learning (SSL) based SOH estimation approach is proposed in this paper. By exploiting unlabeled data, the proposed SSL based method can effectively alleviate the labeled data scarcity. Specifically, two regressors are used to learn the mapping between health indicators (HIs) and SOH. The pseudo-labels are predicted for unlabeled data based on semi-supervised co-training to augment the training samples. The final prediction is realized by combining two regressors. Analysis and experiments show that the proposed SSL based method can significantly improve the SOH estimation performance. Using labeled data of only one cell, the average root-mean-square error (RMSE) of SOH estimation for the other seven cells is 0.55%. Compared to two benchmarks without using unlabeled data, the average prediction accuracy is improved by 53% and 26%, respectively. The proposed SSL method is encouraging to surpass a state-of-the-art supervised learning based SOH estimation method. Moreover, physical interpretations for the selected three short-time HIs are provided. This work highlights the promise of combining large-volume unlabeled industrial data with limited labeled laboratory data to estimate the battery SOH.