Deep learning enhanced lithium-ion battery nonlinear fading prognosis

With the assistance of artificial intelligence, advanced health prognosis technique plays a critical role in the lithium-ion (Li-ion) batteries management system. However, conventional data-driven early aging prediction exhibited dramatic drawbacks, i.e., volatile capacity nonlinear fading trajectories create obstacles to the accurate multistep ahead prediction due to the complex working conditions of batteries. Herein, a novel deep learning model is proposed to achieve a universal and accurate Li-ion battery aging prognosis. Two battery datasets with various electrode types and cycling conditions are developed to validate the proposed approaches. Knee-point probability (KPP), extracted from the capacity loss curve, is first proposed to detect knee points and improve state-of-health (SOH) predictive accuracy, especially during periods of rapid capacity decline. Using one-cycle data of partial raw voltage as the model input, the SOH and KPP can be simultaneously predicted at multistep ahead, whereas the conventional method showed worse accuracy. Furthermore, to explore the underlying characteristics among various degradation tendencies, an online model update strategy is developed by leveraging the adversarial adaptation-induced transfer learning technique. This work gains new sights into the comprehensive Li-ion battery management and prognosis framework through decomposing capacity degradation trajectories and adversarial learning on the unlabeled samples.