Scholarlink

<p>Accurate estimation of batteries’ state of health (SOH) is essential for ensuring reliable driving and evaluating potential maximal capacity. Voltage relaxation has emerged as a pivotal SOH indicator due to its unique capabilities. It directly reveals electrochemical equilibrium states and ion diffusion dynamics, while requiring only single-cycle measurements for data collection. Traditional SOH models adopt statistical or machine learning approaches to extract sophisticated features from voltage relaxation. However, these methods potentially reduce the data resolution and ignore the tendency of voltage decay curve, leading to low accuracy. To solve these challenges, we introduce a transformer-based architecture—DIPformer that leverages an inverted patch embedding approach combined with a multivariate attention mechanism. This model can capture the sequential and multivariate dependencies intrinsic to spatial and temporal features of voltage relaxation. To enhance the robustness and accuracy, we integrate local features identified by attention with the global degradation trends observed throughout the battery’s life cycle in LayerNorm and feed-forward network. Experimental results show that DIPformer achieves MAE of 0.50% and RMSE of 0.81%, representing a significant improvement of 32.4% and 19.8% in accuracy compared to previous state-of-the-art models. The code and data are available in <a href="https://github.com/Danker0/DIPformer">https://github.com/Danker0/DIPformer</a>.</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

DIPformer: a deep inverted patching transformer for battery state of health estimation on voltage relaxation

  • 期刊论文

  • 发表日期：

    2025年6月25日

查看全文
  • Yaoyidi Wang,
  • Dingyu Yang,
  • Niansheng Chen,
  • Lei Rao,
  • Yingzhou Yu

摘要

Accurate estimation of batteries’ state of health (SOH) is essential for ensuring reliable driving and evaluating potential maximal capacity. Voltage relaxation has emerged as a pivotal SOH indicator due to its unique capabilities. It directly reveals electrochemical equilibrium states and ion diffusion dynamics, while requiring only single-cycle measurements for data collection. Traditional SOH models adopt statistical or machine learning approaches to extract sophisticated features from voltage relaxation. However, these methods potentially reduce the data resolution and ignore the tendency of voltage decay curve, leading to low accuracy. To solve these challenges, we introduce a transformer-based architecture—DIPformer that leverages an inverted patch embedding approach combined with a multivariate attention mechanism. This model can capture the sequential and multivariate dependencies intrinsic to spatial and temporal features of voltage relaxation. To enhance the robustness and accuracy, we integrate local features identified by attention with the global degradation trends observed throughout the battery’s life cycle in LayerNorm and feed-forward network. Experimental results show that DIPformer achieves MAE of 0.50% and RMSE of 0.81%, representing a significant improvement of 32.4% and 19.8% in accuracy compared to previous state-of-the-art models. The code and data are available in https://github.com/Danker0/DIPformer.