Scholarlink

<p>To address the substantial differences between laboratory battery aging data and onboard battery state data in operating-condition complexity, sampling completeness, and statistical distributions, this paper proposes a cross-domain transfer learning method for state-of-health (SOH) estimation of lithium-ion batteries under real-world vehicle conditions. Using laboratory battery aging data from the University of Maryland as the source domain, an LSTM–Transformer-based temporal prediction model is constructed. A dual-layer adaptive structure, consisting of an input-level Adapter and a vehicle-aware feature-wise linear modulation module (V-FiLM), is further introduced to achieve feature alignment across domains and to capture vehicle-level heterogeneity. During transfer fine-tuning, a staged freeze–unfreeze strategy is adopted to improve training stability, and a lightweight statistical alignment constraint is incorporated to further mitigate distribution shift. Validation on a target-domain dataset containing operational data from 20 real-world vehicles shows that the proposed method achieves low-error battery-pack SOH prediction without relying on complete charge–discharge curves and using only proxy labels for target-domain modeling. The overall MAE, RMSE, and MAPE are 0.0135, 0.0175, and 1.7%, respectively. Paired bootstrap analysis based on vehicle-level error metrics further indicates that the performance gains brought by the proposed dual-layer adaptive structure are statistically stable. Under the current data conditions, the proposed method improves SOH prediction when transferring laboratory battery aging knowledge to real-world vehicle scenarios and demonstrates promising potential for cross-domain application. Nevertheless, given the limitations of source-domain sample size, proxy-label construction, and deployment conditions, its cross-scenario generalization ability and real-time onboard applicability still require further validation.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

State of health estimation for onboard lithium-ion batteries based on cross-domain transfer learning

  • 期刊论文

  • 发表日期：

    2026年5月4日

查看全文
  • Jiachen Huang,
  • Jun Chen,
  • Yanping Zheng

摘要

To address the substantial differences between laboratory battery aging data and onboard battery state data in operating-condition complexity, sampling completeness, and statistical distributions, this paper proposes a cross-domain transfer learning method for state-of-health (SOH) estimation of lithium-ion batteries under real-world vehicle conditions. Using laboratory battery aging data from the University of Maryland as the source domain, an LSTM–Transformer-based temporal prediction model is constructed. A dual-layer adaptive structure, consisting of an input-level Adapter and a vehicle-aware feature-wise linear modulation module (V-FiLM), is further introduced to achieve feature alignment across domains and to capture vehicle-level heterogeneity. During transfer fine-tuning, a staged freeze–unfreeze strategy is adopted to improve training stability, and a lightweight statistical alignment constraint is incorporated to further mitigate distribution shift. Validation on a target-domain dataset containing operational data from 20 real-world vehicles shows that the proposed method achieves low-error battery-pack SOH prediction without relying on complete charge–discharge curves and using only proxy labels for target-domain modeling. The overall MAE, RMSE, and MAPE are 0.0135, 0.0175, and 1.7%, respectively. Paired bootstrap analysis based on vehicle-level error metrics further indicates that the performance gains brought by the proposed dual-layer adaptive structure are statistically stable. Under the current data conditions, the proposed method improves SOH prediction when transferring laboratory battery aging knowledge to real-world vehicle scenarios and demonstrates promising potential for cross-domain application. Nevertheless, given the limitations of source-domain sample size, proxy-label construction, and deployment conditions, its cross-scenario generalization ability and real-time onboard applicability still require further validation.