Date of Award
Doctor of Philosophy (PhD)
John F. MacGregor
This thesis focuses on developing empirical methodologies for improving process operation and product quality, in four important chemical engineering problems, using multivariate projection methods, such as Principal Component Analysis (PCA) and Projection to Latent Structures (PLS). The four problems addressed in this work are concerned with (i) improving and optimizing the trajectories of manipulated variables in batch processes; (ii) improving the identification of non-parsimonious dynamic process models using the Jackknife and the Bootstrap methods; (iii) developing meaningful specification regions for raw materials entering a consumer's plant and, (iv) improving transition policies in start-ups, re-starts and grade changeovers that are routinely performed in multi-product plants. The first problem addresses the situation where one desires to gain understanding of how and when, during the course of a batch, manipulated process variables have a significant effect on product quality. This amounts to estimating the sensitivity of product quality to manipulated process variables at various degrees of completion in a batch process. This information can be used in process development and in the optimization of already existing processes. The proposed approach involves adding designed experiments to batch policies currently used and then analyzing the resulting data bases using multi-way multi-block PLS. A new pathway PLS algorithm was developed for incorporating intermediate quality measurements collected during the course of each batch. In the second problem, the identification of non-parsimonious dynamic process models is improved through a more judicious selection of the meta parameter in regularization methods (ridge regression) and latent variable methods. These methods are often used to overcome ill-conditioning frequently encountered in the identification of such over-parameterized models. A new criterion for selecting the meta parameter (ridge parameter in regularization methods and the number of components in latent variable methods) is proposed, based on Jackknife and Bootstrap statistics. It is shown that this criterion outperforms the use of cross-validation (default criterion) and leads to the identification of models that are closer to the true process behavior. Developing an approach for defining multivariate specifications on incoming raw materials was important because there is a void in the quality control literature in this area. Specifications are usually defined in a univariate manner, based on past and often subjective experience. This work provides a sound, data-based approach for developing truly multivariate specification regions in a variety of industrial situations. The approach uses PLS methods to analyze historical data on the incoming raw materials, on the consumer's plant, and on the consumer's end product to define multivariate specification regions for the incoming raw material properties. The last problem consists of improving the performance of process transitions (start-ups, re-starts and grade changeovers), using historical process data. In particular, this work addresses two questions: (i) how to improve transition policies to minimize transition time and amount of off-grade materials, while ensuring safe operating conditions and (ii) how to ensure that at the end of a transition, steady-state process conditions are such that good quality products are obtained, and are consistent with past periods of production.
Duchesne, Carl, "Improvement of processes and product quality through multivariate data analysis" (2000). Open Access Dissertations and Theses. Paper 2668.