State-of-Health Estimation With Anomalous Aging Indicator Detection of Lithium-Ion Batteries Using Regression Generative Adversarial Network

Accurate state-of-health (SOH) estimation for a data-driven method is still a great challenge, as real SOH is difficult to measure during the actual application of a lithium-ion battery, and the noise or sensor failure may be also involved. To face these challenges, we propose a novel regression generative adversarial network to obtain a general model for batteries with the same specifications. First, we develop the generator to automatically generate auxiliary training samples with similar but different distributions with real samples, which acts as data augmentation. Meanwhile, the discriminator is designed to detect anomalous aging indicators by learning the distribution of real samples, which is without the requirement of collecting anomalous samples. To capture shallow general aging knowledge, a shallow layer sharing mechanism between the discriminator and regressor is developed for regularization benefit. Finally, we propose a general model building rule based on the optimal correlation between SOH and features. The experimental results show our general model rule is effective for collected datasets of both LiNCM and LiFePO₄ batteries. For datasets with small correlation differences, the effectiveness of the general model is no longer limited by the selection of datasets. Besides, compared to other advanced models, our method could achieve superior prediction performance.