Bayesian neural networks with physics-based priors for robust battery health assessment
Jokin Alcibar、Aitor Aguirre-Ortuzar、Ekhi Zugasti、Chetan S. Kulkarni、Javier Fernández、Jose I. Aizpurua
1. Bayesian neural networks with physics-based priors for robust battery health assessment
- 标题:Bayesian neural networks with physics-based priors for robust battery health assessment
- DOI/URL:https://doi.org/10.1038/s41598-026-63104-y
- 期刊/会议/预印本平台:Scientific Reports
- 发表日期:2026-07-23
- 作者和机构:Jokin Alcibar、Aitor Aguirre-Ortuzar(Mondragon University);Ekhi Zugasti、Jose I. Aizpurua(University of the Basque Country/UPV/EHU,Aizpurua 亦属 Ikerbasque);Chetan S. Kulkarni(KBR, NASA Ames Research Center);Juan Fernandez(Quantia Ingeniería y Consultoría S.L.)。
- 通讯作者:Jokin Alcibar(Nature 页面明确 correspondence to Jokin Alcibar)
- 脚本评分:83
纳入理由和命中证据:题目和摘要均明确指向 lithium-ion batteries 的 health assessment / prognostics;摘要称方法用于可靠、uncertainty-aware 的电池健康预测,并在 NASA 与 Wenzhou pack degradation 数据上评估。
创新点:把电化学仿真得到的物理先验嵌入 Bayesian Neural Network 的变分推断目标,通过 KL divergence 项实现物理一致性正则;重点不是单纯点预测,而是 posterior predictive distribution 与不确定性量化。
主要内容:面向数据驱动模型缺少解释性/可靠不确定性、物理模型依赖精确参数且泛化有限的问题,提出物理先验 BNN 框架,在 cell 与 pack 两类尺度数据上验证健康预测能力。
模型架构/方法流程:可访问页面确认是 Bayesian Neural Network,物理先验来自 electrochemical battery simulations,并通过 variational objective 中的 KL divergence 项注入。具体网络层数、输入窗口、训练细节未能从已抓取页面确认。
输入特征与输出目标:未能从可访问内容确认具体输入特征;输出为 battery health prognostics/health assessment 的预测分布。
数据来源/数据结构/电池类型/循环协议:NASA battery dataset;Wenzhou Pack Degradation dataset。摘要称覆盖 diverse discharge profiles、operating conditions、battery scales;具体电池化学体系和循环协议未能从可访问内容完全确认。
结果指标:摘要称在预测准确性、概率校准、鲁棒性上相对 standard BNN conventional priors 和 state-of-the-art physics-guided Bayesian approaches 有持续改进;具体 MAE/RMSE/校准指标数值未在已抓取内容中确认。
Baseline/对比方法:standard BNNs with conventional priors;state-of-the-art physics-guided Bayesian approaches。
泄漏风险/局限性/复现性:页面可确认开放获取与补充材料链接,但抓取到的主体主要为摘要/作者信息,缺少训练/切分细节;需要下载 PDF/补充材料后核验是否存在同电池跨周期随机切分、是否有跨 pack 外推测试,以及物理先验是否由目标测试条件间接生成。
对哥哥研究/落地的启发和后续动作:适合借鉴“物理先验作为 Bayesian regularizer”思路,用在哥哥现有 SOH/RUL 模型的不确定性校准;建议后续重点复核指标表、先验生成流程和代码/补充材料。
查看关键词分类依据
- PINN / physics-informed · 摘要包含「physics-informed」
nsistency. The proposed framework produces a physics-informed posterior predictive distribution and enables end-to-end uncertainty quantifica - Generic neural network / deep learning · 标题包含「neural network」
Bayesian neural networks with physics-based priors for robust battery health assessment - Bayesian · 标题包含「bayesian」
Bayesian neural networks with physics-based priors for robust battery health assessment - ECM / electrochemical model · 摘要包含「electrochemical」
orporating physics-based priors derived from electrochemical battery simulations. These priors embed electrochemical domain knowledge direct - NASA · 摘要包含「nasa」
mance is evaluated on the publicly available NASA battery dataset and the Wenzhou Pack Degradation dataset, covering diverse disc - Li-ion / LIB · 摘要包含「lithium-ion」
e and uncertainty-aware health assessment of lithium-ion batteries is essential for ensuring the reliability, safety, and efficiency of