Abstract:

Conventional principal component analysis (PCA) selects several Principal Components (PCs) with most variance information of normal samples in process monitoring. Because fault information may not be necessarily mapped into high variance components, PC selection by variance contribution often leads to serious information loss and poor monitoring performance. ReliefF-PCA algorithm was proposed to select PCs of higher weighted and more effective components in distinguishing normal and fault conditions. The PCs selected in this way avoided subjectivity, blindness, and critical information loss as happened in conventional PCA. ReliefF-PCA algorithm in process monitoring had two advantages of better monitoring performance and more effective dimension reduction of original data. Subsequently, a weighted contribution plot of fault variables was proposed. The result of Tennessee Eastman (TE) simulation study showed that the ReliefF-PCA algorithm achieves desired outcomes.

Key words: process system, process control, principal component analysis, ReliefF-PCA algorithm, fault detection, fault location

摘要：

传统的主成分分析（principal component analysis，PCA（算法选取包含大部分方差信息的成分作为主元，并将其应用到过程监控中。但是故障信息不一定会投影到方差较大的成分上，使用方差贡献度挑选主元会导致严重的信息丢失和监控效果的恶化。因此使用ReliefF-PCA算法，其中ReliefF算法从故障角度出发，挑选出在区分正常样本和故障样本上权重更高，效果相对更好的成分作为主元。这样挑选出的主元避免了传统PCA算法在主元挑选过程中出现的主观性、盲目性以及重要信息的丢失。ReliefF-PCA算法在过程监控中主要有两个优势，第1，监控效果更好；第2，对原始数据降维效果更好。随后，基于ReliefF-PCA算法，提出一种加权的故障变量贡献图方法。最后，通过Tennessee Eastman（TE（仿真实验测试，ReliefF-PCA算法达到了预期效果。

关键词: 过程系统, 过程控制, 主元分析, ReliefF-PCA算法, 故障检测, 故障定位