Input variable selection for multivariate statistical process monitoring

Rahman, Md. Musfiqur (2013) Input variable selection for multivariate statistical process monitoring. Masters thesis, Memorial University of Newfoundland.

[English] PDF - Accepted Version Available under License - The author retains copyright ownership and moral rights in this thesis. Neither the thesis nor substantial extracts from it may be printed or otherwise reproduced without the author's permission. Download (7MB)

Abstract

Multivariate statistical methods are widely used in process operations for predicting unmeasured quality, and detection and diagnosis of faults. Performance of these monitoring tools greatly depends on selecting the right set of variables as input to the tools. In a typical chemical process on average 1500 variables arc logged. Selection of appropriate input variables from these large set of variables is a daunting task. This thesis investigates the application of retrospective Taguchi method in selecting input variables for multivariate statistical monitoring tools. Taguchi's design of experiment (DoE) approach has been widely used in industrial process design, primarily in manufacturing industries for optimizing process parameters. Instead of relying on an arbitrary selection of levels, experiments are conducted following an orthogonal array as determined by the Taguchi method. In the current research, the method is adapted for selecting important input variables for process monitoring tools, namely, support vector regression (SVR) and principal component analysis (PCA). Taguchi's DoE assumes t hat variables are uncorrelated which is contrary to process data. Process variables are highly correlated and show dynamic variations due to the frequent change made in the set points causing difficulty to select data to match the orthogonal array of the Taguchi method. These implementation difficulties were addressed in the proposed methodology. Retrospective Taguchi method was adapted for dealing with process data. Additional data preprocessing and correlation analysis steps were proposed to condition process data for Taguchi method. Detailed methodologies to apply Taguchi method to select input variables for SVR and PCA are described in the thesis. The methodologies were demonstrated using industrial data from a petrochemical process and a hydrometallurgy process respectively. The performance of the proposed Taguchi based method was compared with variable importance in projection (VIP) method. The industrial case studies clearly show that the proposed methodology can minimize the computational efforts in variable selection and it can improve the performance of the monitoring tools.

Actions (login required)

View Item