基于鲁棒主元分析的故障诊断方法  

Fault diagnosis method based on robust principal component analysis

作　　者：邓晓刚[1] 田学民[1] 

机构地区：[1]中国石油大学信息与控制工程学院,山东东营257061

出　　处：《中国石油大学学报（自然科学版）》2008年第1期147-151,共5页Journal of China University of Petroleum(Edition of Natural Science)

基　　金：国家'863'计划项目(2004AA412050)

摘　　要：针对工业过程的建模数据中含有离群点的情况,提出了一种基于鲁棒主元分析(PCA)的故障诊断方法。该方法使用广义极大似然估计(M估计)代替最小二乘估计,将传统的主元分析问题转化为一个加权的重构误差优化问题,然后通过改进的非线性迭代部分最小二乘(NIPALS)算法来求得问题的最优解,在此基础上建立主元模型并构造监控统计量检测过程故障。在连续搅拌反应器(CSTR)仿真系统上的应用结果表明,鲁棒PCA方法能够消除离群点对主元模型的影响,比PCA方法分析过程数据更为准确,能更有效地诊断过程故障。A robust principal component analysis （PCA） method was proposed to analyze the model data with outliers in process monitoring. By replacing the least squares estimator with a robust M-estimator, the traditional principal component analysis problem was transformed into a weighted reconstructed error optimization problem. The problem can be solved by improved nonlinear iterative partial least squares （NIPALS） algorithm so that precise principal component model was available and monitoring statistics were used to detect faults. The simulation results on a continuous stirred tank reactor （CSTR） system show that the proposed robust PCA method can remove the influence of outliers, analyze the process data more accurately and diagnose process faults more effectively than traditional PCA method.

关 键 词：故障诊断 鲁棒主元分析 离群点 NIPALS算法 M估计 

分 类 号：TP277[自动化与计算机技术—检测技术与自动化装置]