Modeling and monitoring of a high pressure polymerization process using multivariate statistical techniques

This thesis explores the use of multivariate statistical techniques in developing tools for property modeling and monitoring of a high pressure ethylene polymerization process. In polymer industry, many researchers have shown, mainly in simulation studies, the potential of multivariate statistical methods in identification and control of polymerization process. However, very few, if any, of these strategies have been implemented. This work was done using data collected from a commercial high pressure LDPE/EVA reactor located at AT Plastics, Edmonton. The models or methods developed in the course of this research have been validated with real data and in most cases, implemented in real time.; One main objective of this PhD project was to develop and implement a data based inferential sensor to estimate the melt flow index of LDPE and EVA resins using regularly measured process variables. Steady state PLS method was used to develop the soft sensor model. A detailed description of the data preprocessing steps are given that should be followed in the analysis of industrial data. Models developed for two of the most frequently produced polymer grades at AT Plastics have been implemented. The models were tested for many sets of data and showed acceptable performance when applied with an online bias updating scheme.; One observation from many validation exercises was that the model prediction becomes poorer with time as operators use new process conditions in the plant to produce the same resin with the same specification. During the implementation of the soft sensors, we suggested a simple bias update scheme as a remedy to this problem. An alternative and more rigorous approach is to recursively update the model with new data, which is also more suitable to handle grade transition. Two existing recursive PLS methods, one based on NIPALS algorithm and the other based on kernel algorithm were reviewed. In addition, we proposed a novel RPLS algorithm which is based on the Krylov Controllability based PLS theory. It was found that the new method is much faster, and hence is suitable for time varying system containing many variables.; Finally, we present a data based monitoring scheme to detect the onset of decomposition in a LDPE reactor. This novel method combines PCA and an energy balance around the reactor. This relatively simple method was able to detect the onset of decomposition with reasonable lead time. The method has been implemented at the plant where it is being used as an additional monitoring tool to ensure safe reactor operation.