State of Health Estimation of Lithium-Ion Batteries Using Data Augmentation and Feature Mapping

Accurate state of health (SOH) estimation is crucial for the operating safety and reliability of lithium-ion batteries (LIBs). Currently, data-driven SOH estimation methods have been widely used in both academic and industrial communities. Nevertheless, they are still confronted with challenges, including time-consuming data collection and poor adaptability to different batteries. This article proposes a deep learning framework with a generative adversarial network (GAN) for data augmentation and a feature mapping module. To make the GAN’s generator learn feature distribution of limited original data better, consistent Gaussian noise is applied to the same cycle sequence during data generation. In consideration of the prominent noises introduced by the GAN to short-cycle batteries, a feature mapping module is proposed to reduce the difference in features of different battery datasets. Experiments using estimators trained on different batteries and directly transferred to other batteries are carried out. Results indicated that the accuracy of about 87% of the total 60 samples is improved after data augmentation, and the average absolute error (AE) is below 2%. Furthermore, the embedded feature mapping module achieves a root mean square error (RMSE) of lower than 1%, and half of them are below 0.5%.