A Novel Transfer Learning Approach for State-of-Health Prediction of Lithium-Ion Batteries in the Absence of Run-to-Failure Data

• Published in: IEEE transactions on instrumentation and measurement 2024, Vol.73, p.1-12
• Main Authors: Song, Lei, Gui, Xuanang, Du, Junrong, Fan, Zimeng, Li, Mingwei, Guo, Lili
• Format: Article
• Language: English
• Subjects: Data models; Degradation; Distribution discrepancy; Lithium-ion batteries; Predictive models; state-of-health (SOH) prediction; Task analysis; temporal covariate shift; Training; transformer; Transformers
• ISSN: 0018-9456; 1557-9662
• DOI: 10.1109/TIM.2024.3450095

Accurate and reliable state-of-health (SOH) prediction becomes increasingly vital to ensure the safe and reliable operation of lithium-ion batteries (LIBs). The existing data-driven methods for LIBs' SOH prediction are developed with an ideal database, i.e., a huge run-to-failure data with a consistent distribution of training and testing sets. However, due to individual quality differences and complex operating conditions, data distribution shifts among different batteries may be obvious, and the entire life-cycle samples are difficult to collect in real world. Therefore, there exists a distribution discrepancy between source and target domains. Besides, temporal distribution shift may also exist with incomplete target domain, called time covariate shift (TCS). Thus, the model trained with source and incomplete target domains will cause prediction bias in unseen target data. To address such issues, a novel transfer learning (TL) approach using multiple feature alignment transformer (MFA-Transformer) model is conducted for the SOH prediction of LIBs. First, a multilayer feature alignment is performed via encoder-decoder structure of transformer framework, and multikernel maximum mean discrepancy (MK-MMD) is adopted to tackle data distribution discrepancy. Then, a new loss item based on Weibull distribution is utilized to enhance the data alignment effect. Moreover, a shift compensation strategy using shape-based distance (SBD) estimation is designed to dynamically eliminate the prediction bias resulting from TCS. Finally, experiments on two public LIBs datasets validate the effectiveness of the proposed method, which can offer a promising solution for industrial prognostic without entire life-cycle data.