Battery health estimation with degradation pattern recognition and transfer learning

• Published in: Journal of power sources Vol. 525; p. 231027
• Main Authors: Deng, Zhongwei, Lin, Xianke, Cai, Jianwei, Hu, Xiaosong
• Format: Journal Article
• Language: English
• Published: Elsevier B.V 30.03.2022
• Subjects: Degradation patterns; Lithium-ion battery; Long short-term memory; State of health; Transfer learning; Long short-term memory; Lithium-ion battery; Degradation patterns; Transfer learning; State of health
• ISSN: 0378-7753; 1873-2755
• DOI: 10.1016/j.jpowsour.2022.231027

Battery health monitoring is significant for the maintenance and safety of electric vehicles. Due to huge differences in operation conditions, batteries present diverse degradation patterns (DPs). For battery state of health (SOH) estimation based on data-driven methods, large errors may occur when the DPs of training batteries and test batteries are different. In this paper, early aging data of battery is used to achieve DP recognition and transfer learning (TL), both of which can effectively improve the SOH estimation accuracy. Four features are extracted from discharge capacity curves of battery. Two of them are verified to be highly correlated with battery lifetime and are used to distinguish the DPs of batteries. Others are proved to be closely related to battery capacity and are employed to achieve SOH estimation. Long short-term memory (LSTM) network is used to establish SOH estimation model, and its performance is compared with other machine learning algorithms. Data of 124 cells from public sets is used for verification, and the LSTM is proved to have the best estimation accuracy. Through the proposed DP recognition and TL methods, the estimation accuracy can be further improved, and the mean values of MAEs and RMSEs are only 0.94% and 1.13%. •Four features with high correlation to battery capacity and lifetime are extracted.•Battery degradation patterns are classified by an unsupervised learning method.•Early aging data of battery enable degradation recognition and transfer learning.•The proposed method is verified under public battery data set with 124 cells.