Semi-supervised learning for explainable few-shot battery lifetime prediction

Abstract

Accurate prediction of battery lifetime is critical for ensuring timely maintenance and safety of batteries. Although data-driven methods have made significant progress, their model accuracy is often hampered by a scarcity of labeled data. To address this challenge, we developed a semi-supervised learning technique named partial Bayesian co-training (PBCT), enhancing the modeling of battery lifetime prediction. Leveraging the low-cost unlabeled data, our model extracts hidden information to improve the understanding of the underlying data patterns and achieve higher lifetime prediction accuracy. PBCT outperforms existing approaches by up to 21.9% on lifetime prediction accuracy, with negligible overhead for data acquisition. Moreover, our research suggests that incorporating unlabeled data into the training process can help to uncover critical factors that impact battery lifetime, which may be overlooked with a limited number of labeled data alone. The proposed semi-supervised approach sheds light on the future direction for efficient and explainable data-driven battery status estimation.