Dynamic Loss Balancing for Joint SOH and RUL Prediction of Lithium-Ion Batteries via a Rotary SOH-Injected Prior Battery Transformer
Shuhao Chen、Tianyu Shi、Yiwen Huang、Chengyi Tu
论文卡片 2
基本信息
- 标题:Dynamic Loss Balancing for Joint SOH and RUL Prediction of Lithium-Ion Batteries via a Rotary SOH-Injected Prior Battery Transformer
- DOI/URL:http://arxiv.org/abs/2607.18329v1;PDF:https://arxiv.org/pdf/2607.18329v1
- 期刊/会议/预印本平台:arXiv
- 发表日期:2026-07-19
- 作者和机构:Shuhao Chen、Tianyu Shi、Yiwen Huang、Chengyi Tu;School of Economics and Management, Zhejiang Sci-Tech University, Hangzhou, China
- 通讯作者:Chengyi Tu(PDF 首页标注 Corresponding author,邮箱 chengyitu1986@gmail.com)
- 脚本评分:73
纳入理由和命中证据
- 纳入理由:题名和摘要均直接指向 lithium-ion battery SOH/RUL joint prediction,完全符合电池寿命预测主题。
- 命中证据:PDF 摘要写明“joint prognosis of State of Health (SOH) and Remaining Useful Life (RUL)”以及“intermediate SOH estimate is directly injected into the RUL regression head”。
创新点
- 将 SOH/RUL 联合预测形式化为 Bayesian multi-task objective,用 homoscedastic uncertainty weighting 动态平衡两个任务损失。
- Transformer backbone 引入 RoPE,强调不依赖绝对 cycle step 的相对时间/退化轨迹建模。
- 使用 dual CLS tokens、per-dimension gated fusion 和 gradient detachment,避免高方差 RUL 梯度污染稳定 SOH 表征。
- 将中间 SOH 注入 RUL regression head,作为物理退化先验。
主要内容
- 模型架构/方法流程:多通道充电时序先经 Conv1D frontend;加入 SOH/RUL 双 CLS token;共享 Pre-LN Transformer + RoPE 编码;gated fusion 融合任务表征;Sigmoid head 约束 SOH,Softplus head 约束 RUL;用贝叶斯同方差不确定性损失联合优化。
- 输入特征与输出目标:PDF 图文说明输入为 Voltage、Current、Temperature 多通道 charging time-series;输出为 SOH 和 RUL。
- 数据来源/数据结构/电池类型/循环协议:NASA、MIT-Stanford、HUST;具体电池数量、化学体系、温度/倍率、EOL 阈值需进一步从全文表格确认。
- 结果指标:PDF 摘要报告 NASA 上 SOH MAE = 1.994%,Stanford 上 RUL prediction error = 62.85 cycles。
- Baseline/对比方法:PDF 引言提及 GPR、CNN、LSTM、domain-adaptive Transformer 等相关方法;完整 baseline 表需进一步核查。
泄漏风险/局限性/复现性
- 泄漏风险:中等/信息不足。跨 NASA、MIT-Stanford、HUST 的验证思路较好,但仍需确认 split 是否按 cell 留出、跨数据集 zero-shot 是否严格、归一化是否仅使用训练集。
- 局限性:arXiv 预印本;指标集中在公开数据集,真实 BMS 部署结论需看是否有 latency/memory/edge 测试。
- 复现性:信息不足。PDF 可访问,但本轮未确认代码、随机种子、完整训练配置。
对哥哥研究/落地的启发和后续动作
- 启发:SOH 作为 RUL head 的显式先验很适合哥哥把容量型 SOH 与寿命 cycle error 联合建模;不确定性 loss 可替代手调多任务权重。
- 后续动作:核查 RoPE 是否真的提升跨协议泛化;用哥哥数据做 ablation:无 SOH injection / 无 detach / 固定 loss weight。
- 跟进结论:强烈跟进。
查看关键词分类依据
- RUL prediction · 标题包含「rul」
Dynamic Loss Balancing for Joint SOH and RUL Prediction of Lithium-Ion Batteries via a Rotary SOH-Injected Prior Battery Tra - SOH estimation · 标题包含「soh」
Dynamic Loss Balancing for Joint SOH and RUL Prediction of Lithium-Ion Batteries via a Rotary SOH-Injected Prior Bat - Transformer / attention · 标题包含「transformer」
ries via a Rotary SOH-Injected Prior Battery Transformer - Bayesian · 摘要包含「bayesian」
flicts. By formulating joint prediction as a Bayesian multi-task objective, RoSIP-Batt introduces a homoscedastic uncertainty weighti - Hybrid model · 摘要包含「fusion」
ssification tokens and a per-dimension gated fusion mechanism, secured by a gradient-detachment operator to prevent high-variance R - ECM / electrochemical model · 摘要包含「electrochemical」
stable SOH representation space. To capture electrochemical degradation patterns without relying on absolute cycle steps, Rotary Position E - NASA · 摘要包含「nasa」
al degradation prior. Evaluations across the NASA, MIT-Stanford, and HUST datasets show that RoSIP-Batt significantly outperforms - MIT / Severson · 摘要包含「mit」
radation prior. Evaluations across the NASA, MIT-Stanford, and HUST datasets show that RoSIP-Batt significantly outperforms stat - HUST · 摘要包含「hust」
aluations across the NASA, MIT-Stanford, and HUST datasets show that RoSIP-Batt significantly outperforms state-of-the-art baseli - Li-ion / LIB · 标题包含「lithium-ion」
alancing for Joint SOH and RUL Prediction of Lithium-Ion Batteries via a Rotary SOH-Injected Prior Battery Transformer