LPMVP算法及其在故障检测中的应用  

LPMVP Algorithm and Its Application to Fault Detection

作　　者：张沐光[1] 宋执环[1] 

机构地区：[1]浙江大学工业控制研究所工业控制技术国家重点实验室,杭州310027

出　　处：《自动化学报》2009年第6期766-772,共7页Acta Automatica Sinica

基　　金：国家自然科学基金(60774067;60736021)资助~~

摘　　要：针对数据信息的特征提取和降维问题,提出一种局部保持最大方差投影(Locality preserving maximum variance projections,LPMVP)新算法.该算法综合考虑了主元分析(Principal component analysis,PCA)和局部保持投影(Locality preserving projections,LPP)算法的优点和不足,提出了新的优化目标,使投影得到的低维空间不仅和原始变量空间有相似的局部近邻结构,而且有相似的整体结构,因而可以包含更多的特征信息.在此基础上,本文使用LPMVP算法把原始变量空间划分为特征空间和残差空间,分别构造了T2和SPE统计量对过程进行监测,建立了一种新的故障检测方法.通过数值例子以及TE过程的仿真研究,表明了LPMVP算法可以有效地提取数据信息,同时也体现了较强的故障检测能力.In order to handle the feature extraction and dimensionality reduction problem, a new method named as locality preserving maximum variance projections （LPMVP） is developed. This algorithm can be considered as a linear approach with a new optimizing target, which takes the excellence and limitation of principal component analysis （PCA） and locality preserving projections （LPP） into account. Comparing to original variable space, this low-dimension projection space enjoys similar locality neighborhood structure and global one. As a result, more feature information can be extracted. Moreover, a new fault detection method is also proposed. The LPMVP algorithm is used to divide the original variable space into two parts： feature space and residual space. Then, T^2 and SPE statistics can be built to monitor the process. Case studies of a numerical example and Tennessee-Eastman （TE） process illustrate the efficiency of the LPMVP algorithm on information extraction. Besides, the new method also shows its fault detection ability.

关 键 词：主元分析 局部保持投影 流形学习 故障检测 

分 类 号：TP277[自动化与计算机技术—检测技术与自动化装置]