基于独立元的k近邻故障检测策略

Fault detection strategy of independent component–based k nearest neighbor rule

中文关键词  k近邻  独立元分析  主元分析  故障检测  间歇过程

英文关键词  k nearest neighbors  independent component analysis  principal component analysis  fault detection  batch process

中文摘要

      k近邻故障检测(fault detection based on k nearest neighbors, FD–kNN)方法能够提高具有非线性和多模态特 征过程的故障检测率. 由于系统故障通常由潜隐变量异常变化引起, 而该类型故障并不能被观测数据直观表现, 因 此直接在观测变量上执行FD–kNN方法, 其故障检测率降低. 本文旨在提高FD–kNN方法针对潜隐变量故障的检测 能力, 提出基于独立元的k近邻故障检测方法. 首先, 通过对观测数据应用独立元分析(independent component analysis, ICA)方法, 获得独立元矩阵; 接下来在独立元矩阵中应用FD–kNN方法进行故障检测. 这等同于直接监控过程潜 隐变量的变化, 可以提高过程故障检测率. 通过非线性实例仿真实验, 证明本文方法检测潜隐变量故障是有效的; 同 时, 在半导体蚀刻工艺过程的仿真实验中, 与主元分析(principal component analysis, PCA)方法、核主元分析(kernel principal component analysis, KPCA)方法、基于主元分析的k近邻故障检测(principal component–based k nearest neighbor rule for fault detection, PC–kNN)方法和FD–kNN方法进行对比, 实验结果进一步验证了本文方法的有效性.

英文摘要

      Fault detection based on k nearest neighbors (FD–kNN) method is able to improve the fault detection rate (FDR) in a process with nonlinear and multimode characteristic. Since some faults are caused by some abnormal change of latent variables and they are difficultly recognized through the observed variables, when FD–kNN is implemented directly in observed data set, its detection result is disappointed. Aiming to improve the fault detection ability of FD–kNN on abnormal change of latent variables, a k nearest neighbors fault detection strategy based on independent component analysis (ICA) is proposed in this paper. First, implement ICA in observed data set to obtain an IC matrix, in which all variables are independent. Then, the conventional FD–kNN is implemented to detect faults in the proposed IC matrix. When FD–kNN is implemented in IC matrix, it means that some latent variables are monitored by FD–kNN. Hence, the faults occurring on latent variables are able to be detected by FD–kNN. The efficiency of the proposed strategy is implemented in a simulated case and in the semiconductor manufacturing processes. The experimental results indicate that the proposed method outperforms PCA (principal component analysis), KPCA (kernel principal component analysis), k-nearest neighbor rule based on PCA (PC–kNN) and FD–kNN.