文献类型：期刊文献

英文题名：A physics-guided transfer learning framework for real-world battery health diagnosis under data scarcity

作者：Chen, Hongxu[1];Chen, Ying[1,2];Xu, Chenmin[1];Luan, Weiling[1];Zhang, Wenjun[3];Shen, Ping[3];Huang, Lvwei[3];Chen, Haofeng[1,2]

机构：[1]East China Univ Sci & Technol, Sch Mech & Power Engn, Key Lab Adv Battery Syst & Safety CPCIF, Shanghai 200237, Peoples R China;[2]Inst Aircraft Mech & Control, Shanghai 200237, Peoples R China;[3]SAIC Motor R&D Innovat Headquarters, Shanghai 201804, Peoples R China

年份：2026

卷号：690

外文期刊名：JOURNAL OF POWER SOURCES

收录：;EI(收录号：20262621019545);WOS:【SCI-EXPANDED(收录号:WOS:001814892200004)】；

基金：This work was supported by the National Natural Science Foundation of China (52375144 and 52205153) , the Shanghai Pujiang Programme (23PJD019) , and the Shanghai Gaofeng Project for University Academic Program Development.

语种：英文

外文关键词：Lithium-ion battery; Electric vehicle (EV); State of health (SOH); Transfer learning; Physics-guided learning; Real-world validation

摘要：Data-driven approaches have shown great promise for battery state of health (SOH) diagnosis in electric vehicles (EVs). Despite the extensive data accumulated by historical fleets, reusing it to reduce reliance on costly aging tests remains challenging due to domain shifts across battery chemistries and varying conditions. This paper proposes a physics-guided transfer learning framework to facilitate cross-fleet SOH diagnosis. Drawing from approximately 300 million field data points from 30 vehicles, a feature set capturing electrochemical evolution and cumulative history is extracted to transfer aging knowledge from nickel-cobalt-manganese (NCM) commercial to LFP passenger fleets. Within a stacked autoencoder framework, gradient-based degradation laws integrating both electrochemistry and historical usage are embedded to enable noise-robust and cross-chemistry transfer across diverse real-world EV fleets. Our method achieves strong transferability, yielding a root mean square error (RMSE) of 0.98% on the target fleet. It also demonstrates high data efficiency, achieving an acceptable 1.80% RMSE using data from one target vehicle. Furthermore, physical constraints enable generalizability to unseen aging stages: finetuning with only the initial 50% or 30% of target data achieves RMSEs of 1.40% and 1.63%, respectively. This work underscores the feasibility of leveraging historical data to overcome data scarcity in practical battery health management.