Transfer learning-based SOH estimation from partial charging data with polarization-aware modeling

Yuxia Li、Yuanyuan Gao、Xianbao Wang、C Wang、Zi Wang、Chuxiong Wu、Yi-Xiang Wang

3. Transfer learning-based SOH estimation from partial charging data with polarization-aware modeling

  • 标题:Transfer learning-based SOH estimation from partial charging data with polarization-aware modeling
  • DOI/URL:https://doi.org/10.1038/s41598-026-48906-4
  • 期刊/会议/预印本平台:Scientific Reports
  • 发表日期:2026-07-23
  • 作者和机构:Yongxu Li(College of Arts and Information Engineering, Dalian Polytechnic University);Yuanyuan Gao、Xianbao Wang、Cui Wang、Zhenyi Wang、Chuxiong Wu、Jiali Wang(School of Mechanical and Electrical Engineering, Qingdao City University)。注意脚本元数据作者名与 Nature 页面略有出入,报告以 Nature 页面为准。
  • 通讯作者:Yongxu Li;Jiali Wang(Nature 页面明确 Correspondence to Yongxu Li or Jiali Wang)
  • 脚本评分:60

纳入理由和命中证据:全文明确 SOH estimation from incomplete/partial charging data;SOH 定义为当前可充容量与额定容量比值;电池为 lithium-ion,任务直接对应健康状态估计与寿命管理。

创新点:不是直接用 LSTM 从片段回归 SOH,而是先用电压预测任务进行密集监督预训练,迫使 LSTM hidden/cell state 学到极化动态与近期历史,再替换输出头微调用于 SOH 回归。

主要内容:针对实际 EV/储能中充电片段起止 SOC 不固定、历史极化未知导致 SOH 估计不稳的问题,提出 polarization-aware transfer learning。作者强调 source task 不是泛泛迁移,而是与短时电压演化/内部极化状态物理相关。

模型架构/方法流程:LSTM 网络;预训练阶段以 current、voltage、capacity-step 序列预测下一步电压,训练 50 epochs;迁移阶段保留 recurrent parameters,更换输出头进行 SOH regression,再训练 50 epochs。在线测试阶段用 fine-tuned model 从部分充电片段估计 SOH。

输入特征与输出目标:输入为观测的电流、电压、容量步长/容量区间片段;预训练输出为未来 7% nominal capacity 窗口内电压/下一步电压;下游输出为 cycle-level SOH。

数据来源/数据结构/电池类型/循环协议:Severson et al. fast-charging 数据集,124 个商用 LFP/graphite cells,额定容量 1.1 Ah;EOL 定义 80% nominal capacity;充电通常两段 CC 到 3.6 V cutoff 后 CC/CV 到 80% capacity,放电 4C 到 2.0 V;30 °C;cycle life 约 150–2300 cycles。排除一个异常短寿命 outlier,99 个训练 cell、24 个测试 cell。每 cycle 生成 20% nominal capacity 观测窗口,起点从 0% 到 50% 以 1% 滑动。

结果指标:电压预测预训练 MAE 4.3 mV、RMSE 10.2 mV;SOH 估计 MAE 0.91%、RMSE 1.30%。

Baseline/对比方法:直接训练 LSTM:MAE 1.75%、RMSE 2.35%;1D CNN:MAE 0.99%、RMSE 1.41%;Random Forest:MAE 1.80%、RMSE 2.50%。作者也提醒跨论文比较因数据切分、目标、窗口定义不同,不宜直接 apples-to-apples。

泄漏风险/局限性/复现性:单一公开数据集;温度影响未显式建模;固定 train/test split,缺少重复交叉验证;大量同一 cycle/cell 片段样本可能提高近邻相关性,需要确认训练/测试严格按 cell 分离(页面写 99/24 cell split,是积极信号)。未来需要跨化学体系、温度、真实车辆片段验证。

对哥哥研究/落地的启发和后续动作:这篇很适合复现:把“辅助任务密集监督预训练”嵌入现有 SOH/RUL 管线;可以尝试用电压预测、dV/dQ 重建或 relaxation prediction 作为 pretext task,再比较 downstream SOH/RUL 泛化。

查看关键词分类依据
  • SOH estimation · 标题包含「soh」
    Transfer learning-based SOH estimation from partial charging data with polarization-aware modeling
  • Transfer learning / domain adaptation · 标题包含「transfer learning」
    Transfer learning-based SOH estimation from partial charging data with polarization-aware modelin
  • Transformer / attention · 摘要包含「transformer」
    against recent partial-charging, CNN-based, Transformer-based, and transfer-learning studies. The results support the value of mechanis
  • LSTM · 摘要包含「lstm」
    framework in which a long short-term memory (LSTM) network is first pretrained on a voltage-prediction task and is then fine-tune
  • CNN · 摘要包含「cnn」
    the method against recent partial-charging, CNN-based, Transformer-based, and transfer-learning studies. The results support th
  • Generic neural network / deep learning · 摘要包含「data-driven」
    larization history is unknown. Many existing data-driven methods either rely on handcrafted health indicators or assume relatively fixed
  • MIT / Severson · 摘要包含「severson」
    to future battery response. Using the public Severson fast-charging dataset of 124 commercial lithium iron phosphate/graphite cells,
  • Self-built / experimental dataset · 摘要包含「in-house」
    arying starting positions. Under the matched in-house protocol used in this study, transfer learning reduces the SOH estimation mean
  • LFP / LiFePO4 · 摘要包含「lithium iron phosphate」
    rson fast-charging dataset of 124 commercial lithium iron phosphate/graphite cells, we evaluate the proposed framework on 20%-capacity charging seg