Improving Battery Life Prediction With Unlabeled Data: Confidence-Weighted Semi-Supervised Learning With Label Propagation
Song Zhang, Yannan Li, Jinpeng Tian, Zhihong Man, C. Y. Chung, Weixiang Shen
Abstract
Accurately predicting the remaining useful life (RUL) of lithium-ion batteries is crucial for the safety and reliability of electric vehicles (EVs). Although data-driven approaches have been extensively used with high accuracy, they need to be trained on massive data with RUL labels, leading to prohibitive data collection costs. In this article, we propose a semi-supervised learning method that can integrate battery operating data without RUL labels into model training to enhance the RUL prediction performance while relaxing the data demand. First, a label propagation (LP) strategy is developed to generate pseudo-RUL labels for unlabeled samples, enabling the incorporation of unlabeled samples into the existing supervised training framework. Afterward, confidence-weighted training is proposed to assign different levels of confidence to the generated pseudo-labeled samples, reducing the negative impact of inaccurate pseudo labels on model training. The proposed method’s effectiveness is validated on various battery aging datasets, covering different battery types, charging/discharging policies, temperatures, and model structures. Compared to conventional supervised learning strategies, the proposed method reduces the average root mean squared errors (RMSEs) up to 80% with limited labeled data.