面向少样本删失数据的三参数Weibull蓄电池寿命评估方法——基于NASA公开数据的阈值敏感性验证
A Three-Parameter Weibull Method for Evaluating Battery Life with Few Samples and Missing Data—Threshold Sensitivity Validation Based on Publicly Available NASA Data

在储能与电力系统运行中，蓄电池寿命评估对预防性维护与更换决策具有重要意义，但工程上可获得的寿命样本往往规模有限且包含右删失数据，导致传统两参数Weibull模型在早期可靠度评估上可能出现偏差。本文提出一种面向少样本删失数据的三参数Weibull蓄电池寿命评估方法：引入位置参数γ表征非老化期，并采用“γ外循环搜索 + Levenberg-Marquardt (LM)内回归”的稳定估计流程。为提高可复现性，本文基于NASA Ames PCoE Li-ion Battery Aging Dataset的多电芯循环退化数据构造寿命样本，以SOH阈值定义EOL，并在SOH = 0.85/0.80/0.75下进行阈值敏感性分析。结果表明：在较早阈值(SOH ≤ 0.85)下，三参数模型在赤池信息准则、贝叶斯信息准则以及相对KM经验曲线的早期加权误差方面更优，适用于早期预警型寿命评估；在SOH ≤ 0.80/0.75条件下两参数模型已能取得更低的信息准则，说明该阈值下数据对非老化期的统计识别较弱。本文进一步给出基于AIC/BIC与早期加权误差的模型选取建议，可为蓄电池寿命建模与运维决策提供可复现的工程方法。
In energy storage and power system operations, battery life assessment is crucial for preventive maintenance and replacement decisions. However, the life samples available in engineering practice are often limited in size and contain right-censored data, which may lead to biases in early reliability assessments using the traditional two-parameter Weibull model. This paper proposes a three-parameter Weibull battery life assessment method tailored for small-sample censored data: it introduces the location parameter γ to characterize the non-aging period and employs a stable estimation workflow consisting of “γ outer loop search + Levenberg-Marquardt (LM) inner regression”. To enhance reproducibility, this study constructs a lifespan sample based on multi-cell cycle degradation data from the NASA Ames PCoE Li-ion Battery Aging Dataset, defines End-of-Life (EOL) using State of Health (SOH) thresholds, and performs threshold sensitivity analysis at SOH = 0.85/ 0.80/0.75. The results indicate that at earlier thresholds (SOH ≤ 0.85), the three-parameter model performs better in terms of AIC/BIC and early-weighted error relative to the KM empirical curve, making it suitable for early-warning-type lifespan assessment; under conditions of SOH ≤ 0.80/0.75, the two-parameter model already achieves lower information criteria, suggesting that data at these thresholds provide weaker statistical identification of the non-aging period. This paper further provides model selection recommendations based on AIC/BIC and early weighted error, offering a reproducible engineering method for battery life modeling and O&M decision-making.