State of Health Estimation of Lithium Iron Phosphate Batteries Based on Degradation Knowledge Transfer Learning

Abstract

Accurate state of health (SOH) estimation constitutes a critical task for systems employing lithium-ion (Li-ion) batteries. However, many current studies that focus on data-driven SOH estimation methods ignore the battery degradation modes (DMs). This paper proposes a two-stage framework to develop a SOH estimation model for Li-ion batteries considering the transferred DM knowledge. Firstly, a battery DM regression model is designed to diagnose the contributions of three DMs by transferring the DM knowledge. Since the real and synthetic datasets are independent and identically distributed, DM regression model trained with the synthetic dataset cannot be directly applied to the real dataset. To bridge the gap, this paper proposes a deepCoral-based domain adaptation method to minimize the regression loss and domain adaptation loss between the source domain (synthetic) and the target domain (real) such that the degradation knowledge learned from the synthetic batteries can be transferred to the real batteries. The model’s structure and parameters are optimized through simulation tests to improve the diagnosis accuracy. Secondly, we propose a new deep learning model, conditional time series generative model (CTSGAN), which can effectively preserve temporal dynamics during battery degradation. With the DM and other related conditions, a CTSGAN-based SOH estimation model is constructed, which shows good estimation performance. Finally, case studies verify the effectiveness and superiority of degradation knowledge transfer learning and the SOH estimation for synthetic and real battery datasets.