Date of Award
Doctor of Philosophy
Charles F. Moore
John Collier, Belle Upadhyaya, J. Weslet Hines, Duane Bruns
Batch processes are widely used in the chemical industry. Recently, much attention has been given to the monitoring and analysis of batch measurement data, or profiles, with an emphasis on the detection of problems. Similarly, methods to improve the final product quality in batch processes have multiplied in the literature. However, an area that is virtually unexplored is the utilization of the data mining techniques for monitoring and analysis of batch profiles for better understanding batch processes, rather than identifying problems in batches, in order to improve the process. The thrust of this work is to apply a systematic method to increase batch process understanding by sifting through the existing historical database of past batches, to discern directions for process improvement from the increased understanding, and to subsequently demonstrate better quality control through the use of online recipe adjustments.
A database of past batches is generated from a simulated nylon-6, 6 process, with the main quality variable of interest being the number average molecular weight. The time and measurement variability in raw batch measurement profiles is characterized through scale parameters. These scale parameters are subjected to a standard principal component analysis (PCA) to understand the principal sources of variation present in a historical database of past batches. Directions for process improvement are discovered from the data mining study and appropriate manipulated variables to implement recipe adjustments are identified. Online predictions of the molecular weight are demonstrated which indicate off-target quality batches well before the end of the batch. A split-range linear molecular weight-based controller is developed that is able to reduce the variability in the quality around the target. Further process improvement is accomplished by reducing the cycle time in addition to tightly controlling the final quality.
The approach for systematically analyzing batch process data is general and can be applied to any batch system, including non-reactive systems.
Johnson, Mark, "Application of data mining techniques of batch profiles for process understanding and improvement. " PhD diss., University of Tennessee, 2001.