零寿命标签下退化设备剩余寿命预测方法  

Prognostic Method of Remaining Useful Life for Degraded Equipment Under Zero Life Label

作　　者：裴洪 司小胜 胡昌华 郑建飞 张建勋 董青 PEI Hong;SI Xiao-sheng;HU Chang-hua;ZHENG Jian-fei;ZHANG Jian-xun;DONG Qing(Department of Automation,Rocket Force University of Engineering,Xi'an,Shaanxi 710025,China)

机构地区：[1]火箭军工程大学控制工程系,陕西西安710025

出　　处：《电子学报》2023年第7期1939-1948,共10页Acta Electronica Sinica

基　　金：国家自然科学基金(No.62103433,No.62233017,No.62227814,No.62073336)。

摘　　要：考虑到安全性与经济性因素,同类历史设备的性能退化数据大多属于截尾型,采用深度学习训练时往往面临零寿命标签的挑战,量化剩余寿命(Remaining Useful Life,RUL)不确定性更是难上加难,并且现有深度学习模型进行RUL预测时忽略了首达与非首达时间意义之间的区别.为了克服以上困难,本文提出一种零寿命标签下退化设备RUL预测方法,采用数据预处理技术生成以退化信息为标签的样本,利用贝叶斯双向长短期记忆(Bayesian Bidirectional Long Short-Term Memory,B-Bi-LSTM)模型描述设备性能退化演变规律,同时借助变分推断技术实现了性能退化的不确定性度量.进一步,从可靠性角度分析了性能退化预测分布与RUL分布间的关系,分别围绕首达与非首达两类情形推导设备RUL概率分布,通过锂电池案例对所提方法进行实例验证.实验结果表明,所提方法能够提供RUL预测的点估计与概率分布式结果,有效确保了预测结果的科学性.Considering the safety and economic factors,most of the performance degradation data of historical equip⁃ment are truncated type.The challenge of zero life label is encountered when deep learning training is adopted,quantifying the uncertainty of remaining useful life(RUL)is even more difficult,and what is more,the existing deep learning models ignore the difference between the first hitting and non-first hitting time meanings when predicting the RUL.To overcome the above difficulties,this paper proposes a method over RUL prediction for degraded equipment under zero life label.Da⁃ta preprocessing technology is utilized to generate samples labeled with degradation information,and the evolution law of equipment performance degradation is described by Bayesian bidirectional long short-term memory(B-Bi-LSTM)model.At the same time,the uncertainty measurement of performance degradation is realized by means of variational inference technology.Furthermore,the relationship between the performance degradation prediction distribution and the RUL distri⁃bution is analyzed from the perspective of reliability,and the RUL probability distribution of the equipment is derived from the point of the first hitting and non-first hitting time respectively.The proposed method is verified by a case of lithium bat⁃tery.The experimental results show that the proposed method can provide the point estimation and probability distribution results of RUL prediction,which can effectively ensure the scientificity of the prediction results.

关 键 词：剩余寿命预测 零寿命标签 贝叶斯双向长短期记忆模型 退化设备 首达时间 非首达时间 

分 类 号：TP18[自动化与计算机技术—控制理论与控制工程] TB114.3[自动化与计算机技术—控制科学与工程] V239[理学—概率论与数理统计]