基于互信息的变量加权型ICA故障检测方法  

Fault detection using mutual information based variable-weighted ICA method

作　　者：石立康 童楚东[1] SHI Likang;TONG Chudong(Faculty of Electrical Engineering and Computer Science, Ningbo University, Ningbo 315211, Zhejiang, China)

机构地区：[1]宁波大学信息科学与工程学院,浙江省宁波市315211

出　　处：《计算机与应用化学》2018年第10期799-808,共10页Computers and Applied Chemistry

基　　金：国家自然科学基金资助项目(61503204;61773225);浙江省自然科学基金(LY16F030001)

摘　　要：在现代大型工业过程中,其过程数据往往不服从高斯分布,独立成分分析(ICA)算法已广泛应用于非高斯过程监测。然而,各变量间存在着不同的相关性,为了挖掘各变量间相关性大小的差异,本文提出一种基于互信息的变量加权型ICA过程监测方法 (MI-WICA)。MI-WICA通过对计算数据矩阵中各变量间的互信息来定义相关性,并以此为依据对原数据矩阵进行加权处理,由此可得多个不同的变量加权型数据集。然后利用ICA算法对各加权矩阵分别建立故障检测模型,以用于工业过程监测。该方法突出了各变量间相关性大小的差异,对数据特征的描述更加全面。最后,通过数值仿真与TE过程仿真实验来验证该方法的优越性。In modern large-scale industrial processes,the sampled data usually does not follow Gaussian distribution,the independent component analysis(ICA) algorithm has been widely applied in non-Gaussian process monitoring. However,the correlated relations vary among different measured variables,to characterize the difference in the correlated relations,a mutual information based variable weighted ICA(MI-WICA) is proposed for process monitoring. The proposed MI-WICA method utilizes the mutual information to define the correlated relations among variables,and then weights the training dataset accordingly. As a result,multiple variable weighted datasets can be obtained. The ICA algorithm is then applied to the weighted datasets individually,so as to build multiple fault detection models for process monitoring purposes. The presented approach highlights the difference in the correlated relations,could be more appropriate in characterizing the underlying features of the sampled data. Finally,the superiority of the proposed method is validated through a numerical example and the well-known TE process.

关 键 词：独立成分分析 过程系统 互信息 故障检测 统计过程监测 

分 类 号：TP277[自动化与计算机技术—检测技术与自动化装置]