Masters Theses

Date of Award

12-1994

Degree Type

Thesis

Degree Name

Master of Science

Major

Chemical Engineering

Major Professor

Charles F. Moore

Committee Members

Peter Cummings, Mary Leitnaker

Abstract

Today's typical chemical process has attached to it a vast, complex "nervous system" of sorts an array of sensors, valves, wires, and control computers which frequently monitor and store information regarding hundreds or thousands of measurements. In addition to these continuous measurements, key process and product streams are periodically analyzed to measure quality parameters such as composition or viscosity. The sheer volume of data available often results in "data overload" for those trying to improve or troubleshoot a process, since a human without data analysis tools is unable to interpret large sets of raw data and make analytical judgements regarding the state of the process. A number of multivariate analysis techniques, designed to use the data set as a collection of related information instead of as a group of individual measurements, have been developed to address this problem.

Over the past 15 years, a data analysis technique known as principal component analysis (PCA) has been adopted and applied to process data to perform multivariate process analysis. This method results in measurements which indicate variability in the process that is consistent with a base data set and variability that is not consistent with a base data set. These measurements can be collected and combined to make efficient use of the whole data set for diagnosis of process conditions and fault detection.

Another multivariate technique which has received recent attention in the area of process data analysis is partial least squares (PLS). PLS can be used to predict variables which might be hard to acquire on a continuous basis using available process measurements. Such inferred quality measurements can be used for process monitoring and, in some cases, for closed loop control.

In this work, the usefulness of multivariate statistical monitoring tools such as PCA and PLS is assessed through their application to a challenging process control problem. The problem examined is control of a reactive distillation column in which the use of internal composition information for closed loop control has been identified as a key need. A control strategy for the column is designed using control analysis tools. A PLS-based estimate of the internal composition is developed and used for closed loop control on a rigorous simulation of the column. PCA-based process monitoring is then applied to the column model to help identify excursions from normal operation which suggest degradation of the inferred composition measurement.

The PLS composition estimator accurately predicts the composition of interest when disturbances are applied that are contained in a base data set used for PLS regression. When disturbances are applied that are not included in the base data set, the estimator performance degrades. For those cases in which the estimate is reasonably accurate, closed loop control based on the estimate is successful in controlling product impurity levels. Closed loop control is considerably less successful in cases where the estimate is inaccurate.

The PCA process monitor is successful in differentiating between areas of process operations consistent with the base data set and areas which are not. This is accomplished through the use of the PCA residual statistic, or Q. This measure can be used as an indicator to the user of regions of operation where the estimator has the potential for gross error due to input data behavior not modeled in the PLS regression.

Download
Files over 3MB may be slow to open. For best results, right-click and select "save as..."

Share

COinS