航空学报 > 2013, Vol. 34 Issue (9): 2219-2229   doi: 10.7527/S1000-6893.2013.0223

基于贝叶斯LS-SVR的锂电池剩余寿命概率性预测

陈雄姿, 于劲松, 唐荻音, 王英勋   

  1. 北京航空航天大学 自动化科学与电气工程学院, 北京 100191
  • 收稿日期:2012-11-28 修回日期:2013-04-22 出版日期:2013-09-25 发布日期:2013-04-23
  • 通讯作者: 于劲松,Tel.:010-82338693 E-mail:yujs@buaa.edu.cn E-mail:yujs@buaa.edu.cn
  • 作者简介:陈雄姿 男, 博士研究生。主要研究方向: 预测与健康管理技术, 设备失效预测与剩余使用寿命估计。Tel: 010-82338693 E-mail: cxzbuaa@163.com;于劲松 男, 博士, 副教授。主要研究方向: 预测与健康管理技术, 自动测试技术。Tel: 010-82338693 E-mail: yujs@buaa.edu.cn;唐荻音 女, 博士研究生。主要研究方向: 预测与健康管理技术, 设备退化建模与视情维修策略。Tel: 010-82338693 E-mail: amytdy@asee.buaa.edu.cn;王英勋 男, 研究员, 博士生导师。主要研究方向: 无人机自主控制。Tel: 010-58356617 E-mail: wangyx@buaa.edu.cn
  • 基金资助:

    航空科学基金(20100751010,2010ZD11007)

Probabilistic Residual Life Prediction for Lithium-ion Batteries Based on Bayesian LS-SVR

CHEN Xiongzi, YU Jinsong, TANG Diyin, WANG Yingxun   

  1. School of Automation Science and Electrical Engineering, Beihang University, Beijing 100191, China
  • Received:2012-11-28 Revised:2013-04-22 Online:2013-09-25 Published:2013-04-23
  • Supported by:

    Aeronautical Science Foundation of China (20100751010, 2010ZD11007)

摘要:

提出了一种基于贝叶斯最小二乘支持向量回归(LS-SVR)的锂电池剩余寿命在线概率性预测方法。首先,通过滚动窗方法选取锂电池历史健康退化数据,并根据相空间重构原理建立训练样本,其中最小嵌入维数使用Cao氏方法计算获得。然后,运用贝叶斯3层推理训练LS-SVR预测模型,在迭代预测阶段,采用蒙特卡罗方法来表示和管理多步预测中的不确定性及其传递,即用一群离散粒子来近似连续分布,结合"退化轨迹不相交"原则和高斯过程假设,预测出锂电池健康状态未来时刻的发展趋势。最后结合给定的失效阈值,通过统计穿越阈值的粒子数目得到剩余寿命的概率分布。使用美国国家航空航天局阿姆斯研究中心公开的电池数据集与高斯过程回归(GPR)方法进行对比实验,多项预测性能指标结果表明贝叶斯LS-SVR方法具有更高的预测准确度和置信度。

关键词: 锂电池, 剩余寿命, 概率性预测, 最小二乘支持向量回归(LS-SVR), 不确定性管理

Abstract:

An online probabilistic prediction approach for the residual life of a lithium-ion battery is proposed by using the Bayesian least squares support vector regression (LS-SVR). First, historical degradation data of the lithium-ion battery are selected through a sliding window. Then the selected data are formed into training samples by the phase space reconstruction theory, with the minimum embedding dimension calculated by Cao's method. Secondly, a predicting model based on least squares support vector regression is trained by a three level Bayesian inference framework. Then in the iterative prediction stage, Monte Carlo method is applied to manage the uncertainty and its propagation in the multi-step prediction, which is achieved by approximating the continuous distribution with a group of discrete particles and predicting the future health status of the battery based on the principle of "non intersecting degradation trajectories" and the Gaussian process assumption. Finally, by counting the number of particles which pass through the predetermined failure threshold, the probability distribution of the battery residual life can therefore be estimated. Comparative experiments are conducted between Bayesian LS-SVR and Gaussian process regression (GPR) using the public battery data sets provided by National Aeronautics and Space Administration Ames Research Center. The results demonstrate that the Bayesian LS-SVR method has higher prediction accuracy and confidence.

Key words: lithium-ion battery, residual life, probabilistic prediction, least squares support vector regression (LS-SVR), uncertainty management

中图分类号: