Process control utilizing data-based models: Applications of statistical techniques and neural networks

There is an increasing demand in chemical industry to produce high quality products at low cost. Such an objective can be achieved through optimal operation of chemical plants. Optimal operation primarily depends upon reliable control schemes and good understanding of chemical processes. It has been demonstrated by many industrial applications that statistical techniques and neural networks are useful tools in using readily available plant data to model chemical processes. This dissertation investigates some new approaches to the applications of statistical techniques and neural networks.;Based on multivariate statistical methods, namely multi-way principal component analysis (MPCA) and multi-block principal component analysis, a dynamic monitoring approach has been developed for continuous processes. By arranging the dynamic data into a three-dimensional array and projecting the array into a low dimensional space defined by principal components, dynamic processes can be easily monitored by tracking their progress in the low dimensional space. A promising feature of the monitoring approach is its ability to predict faults. The application results show that the dynamic monitoring approach has many advantages over existing approaches. A multi-block monitoring approach for large continuous processes is also discussed. The application results show that the multi-block monitoring approach has a timing advantage over single block approaches and it is helpful in locating the process faults.;The statistical concept of representing processes by latent variables has been applied in process control. The goal is to decrease the variations in product quality without on line quality measurements. The controlled variables are defined by the variations embedded in the process data using a PCA technique. The control objective is defined as maintaining the latent variables within a certain acceptable region defined from historical data based on the assumption of an implicit correlation between measurements and quality variables. A steady state controller is designed using static PCA models. For dynamic processes, MPCA is used to model the performance of continuous processes. This controller is usually developed from and implemented on top of an existing conventional PID control system. Limited experimental testing is required to develop the controller. Model predictive control (MPC) is used to formulate the control algorithm. Examples show excellent results for both the steady state and dynamic cases.