Scholarlink

<p>State of Health (SOH) estimation, as one of the core functionalities in battery management, is essential to ensuring the safety and stability of energy storage systems. Significant progress has been made in SOH estimation using convolutional neural networks (CNNs) and transformer networks. However, battery characteristic data exhibit high spatial coupling, and relying solely on multi-layer convolutions for feature extraction limits the capture of comprehensive global features. To address this challenge, we propose a spatial decoupling and multi-scale-assisted cross-domain learning transformer (MCDLT) method based on color names (CN) feature mapping, that effectively integrates original convolutional and diffusion features, thereby reducing spatial coupling in aging features. Firstly, the CN feature approach projects low-dimensional battery aging features into a higher-dimensional space, enabling greater spatial dispersion among distinct aging characteristics. This is followed by the extraction of joint attention-diffusion features through principal component analysis (PCA) and an attention-fusion module. Secondly, a multi-scale CNN module comprehensively extracts multi-scale original convolutional features using various kernel sizes, fusing with the diffusion features. Finally, this paper designs an SOH estimation mechanism based on the transformer network for feature encoding and prediction. The proposed algorithm is validated using NASA and Oxford datasets, achieving a low mean absolute error (MAE) of 0.022% and root mean square error (RMSE) of 0.233%, respectively.</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 lithium-ion batteries based on color names spatial decoupling and multi-scale-assisted cross-domain learning transformer method

  • 期刊论文

  • 发表日期：

    2025年7月17日

查看全文
  • Yunji Zhao,
  • Yuchen Liu,
  • Kai Liu,
  • Yong Wang,
  • Hui Guo

摘要

State of Health (SOH) estimation, as one of the core functionalities in battery management, is essential to ensuring the safety and stability of energy storage systems. Significant progress has been made in SOH estimation using convolutional neural networks (CNNs) and transformer networks. However, battery characteristic data exhibit high spatial coupling, and relying solely on multi-layer convolutions for feature extraction limits the capture of comprehensive global features. To address this challenge, we propose a spatial decoupling and multi-scale-assisted cross-domain learning transformer (MCDLT) method based on color names (CN) feature mapping, that effectively integrates original convolutional and diffusion features, thereby reducing spatial coupling in aging features. Firstly, the CN feature approach projects low-dimensional battery aging features into a higher-dimensional space, enabling greater spatial dispersion among distinct aging characteristics. This is followed by the extraction of joint attention-diffusion features through principal component analysis (PCA) and an attention-fusion module. Secondly, a multi-scale CNN module comprehensively extracts multi-scale original convolutional features using various kernel sizes, fusing with the diffusion features. Finally, this paper designs an SOH estimation mechanism based on the transformer network for feature encoding and prediction. The proposed algorithm is validated using NASA and Oxford datasets, achieving a low mean absolute error (MAE) of 0.022% and root mean square error (RMSE) of 0.233%, respectively.