Scholarlink

<p>Accurate state of health (SOH) prediction of electric vehicle batteries is critical for safety assurance, lifespan extension, and energy management optimization. In the context of federated learning (FL), this paper addresses the data privacy-protection issues of existing SOH prediction methods. We propose a federated learning–based capacity prediction model (FL-CPM). A global and local feature extraction network is built via convolutional neural network-long short-term memory (CNN-LSTM). Funk-singular value decomposition (Funk-SVD) gradient decomposition and hierarchical noise injection are introduced to block privacy leakage risks. Also, the server-side weight aggregation strategy is improved by incorporating a client detection algorithm, enhancing model prediction accuracy and robustness. Experimental results show that the proposed FL-CPM yields a mean absolute error (MAE) of 0.015 and a coefficient of determination (<i>R</i><sup>2</sup>) of 0.8035, demonstrating superior fit compared to baseline models. While ensuring forecasting accuracy, FL-CPM preserves training data privacy, thus highlighting its remarkable capability in tackling data silos and privacy issues.</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

A study on data privacy protection in power battery SOH prediction based on federated learning

  • 期刊论文

  • 发表日期：

    2025年8月22日

查看全文
  • Weidong Fang,
  • Ji Zhang,
  • Xibin Lin,
  • Jiacheng Hu,
  • Linrun Huang,
  • Guangqian Yuan

摘要

Accurate state of health (SOH) prediction of electric vehicle batteries is critical for safety assurance, lifespan extension, and energy management optimization. In the context of federated learning (FL), this paper addresses the data privacy-protection issues of existing SOH prediction methods. We propose a federated learning–based capacity prediction model (FL-CPM). A global and local feature extraction network is built via convolutional neural network-long short-term memory (CNN-LSTM). Funk-singular value decomposition (Funk-SVD) gradient decomposition and hierarchical noise injection are introduced to block privacy leakage risks. Also, the server-side weight aggregation strategy is improved by incorporating a client detection algorithm, enhancing model prediction accuracy and robustness. Experimental results show that the proposed FL-CPM yields a mean absolute error (MAE) of 0.015 and a coefficient of determination (R2) of 0.8035, demonstrating superior fit compared to baseline models. While ensuring forecasting accuracy, FL-CPM preserves training data privacy, thus highlighting its remarkable capability in tackling data silos and privacy issues.