基于DPCA-IM的动态过程监测方法  

Dynamic process monitoring based on estimation error of missing variable

作　　者：孟生军 童楚东[1] Meng Shengjun;Tong Chudong(Faculty of Electrical Engineering&Computer Science,Ningbo University,Ningbo Zhejiang 315211,China)

机构地区：[1]宁波大学信息科学与工程学院,浙江宁波315211

出　　处：《计算机应用研究》2021年第1期175-178,共4页Application Research of Computers

基　　金：国家自然科学基金资助项目(61773225,61803214)。

摘　　要：动态主成分分析(DPCA)通过增广矩阵或向量的方式来挖掘采样数据间的时序自相关性。然而,DPCA对自相关的特征成分与残差直接实施监测是不合理的,故其故障检测效果较差。为了剔除采样数据的自相关性以提高故障检测效果,提出一种基于估计误差的动态过程监测方法。首先,通过逐个假设各个过程变量的测量数据缺失,并在已建立的DPCA模型中引入迭代方法(IM)计算得到相应变量缺失数据的估计值。由于该估计值在仅缺失一个变量数据的条件下能较大程度地逼近原测量数据,两者之差(即估计误差)不再存在显著的自相关性,而且该估计误差的变化可直接反映出采样数据变化情况的异常,所以可利用估计误差监测动态过程。最后,通过两个动态过程实例,即动态数值仿真过程与田纳西—伊斯曼(TE)标准测试平台的仿真结果表明,该方法能剔除采样数据间的自相关性,并能有效地提高故障检测效果,验证了该方法不仅可行,而且具有良好的优越性。Dynamic principal component analysis(DPCA)extracts the time-serial auto-correlation inherited from sampled data through enhancement matrix or vector.However,the way of monitoring the auto-correlated features and residuals directly in the DPCA model is not appropriate,given that the negative influence caused by the auto-correlation on the monitoring statistics is ignored.Therefore,on the basis of the DPCA model that exacts time-serial auto-correlation,how to eliminate the auto-correlation inherited in the sampled data is further required.This paper proposed a dynamic process monitoring method based on estimated errors.Through sequentially assuming the measured data of each process variable was missing,it introduced and incorporated the iteration method(IM)with the built DPCA model so as to calculate the estimates of corresponding variable.Since the estimates could approximate the original measured data to a large extent with only one variable was missing,the inconsistency between the two(i.e,estimation error)no longer existed obvious auto-correlation.Moreover,the variation of the estimation error could directly reflect the abnormal variation in the sampled data,the estimation error could thus be used for dynamic process monitoring purposes.Finally,through comparisons in two dynamic examples,a dynamic numerical process and the Tennessee Eastman benchmark process,the superiority of the proposed DPCA-IM approach over other dynamic process monitoring methods are validated.

关 键 词：动态主成分分析 估计误差 自相关性 缺失数据 

分 类 号：TP277[自动化与计算机技术—检测技术与自动化装置]