A Meta-Learning Method for Few-Shot Multidomain State-of-Health Estimation of Lithium-Ion Batteries
Xiaoyu Zhao, Zuolu Wang, Te Han, Wenxian Yang, Fengshou Gu, Andrew D. Ball
Abstract
Diverse electrochemical characteristics and complex operational conditions of the lithium-ion battery cause multidomain discrepancies in practical applications, which poses huge challenges to the robust state-of-health (SOH) estimation based on small samples. This article proposes a novel meta-learning method for few-shot multidomain battery SOH estimation using relaxation voltages (RVs). First, a convolutional neural network (CNN)-Attention-based parallel network is developed to enhance the extraction of transferable health features across multiple domains. Second, the loss interaction difference of multiple target domain tasks is proposed to improve the meta-learning method for comprehensive task judgment. Finally, the cross-domain validation is conducted on two types of batteries operating under three working temperatures. The results reveal that the proposed method can provide higher estimation accuracy compared to state-of-the-art network architectures. By only using six cycles from one target battery, it achieves lower average root-mean-square error (RMSE) and mean absolute error (MAE) of 2.28% and 1.79% for NCA batteries and 1.38% and 1.14% for NCM batteries, outperforming traditional methods without pretraining and transfer learning (TL).