Batch Control and Diagnosis

Abstract: Batch processes are becoming more and more important in the chemical process industry, where they are used in the manufacture of specialty materials, which often are highly profitable. Some examples where batch processes are important are the manufacturing of pharmaceuticals, polymers, and semiconductors. The focus of this thesis is exception handling and fault detection in batch control. In the first part an internal model approach for exception handling is proposed where each equipment object in the control system is extended with a state-machine based model that is used on-line to structure and implement the safety interlock logic. The thesis treats exception handling both at the unit supervision level and at the recipe level. The goal is to provide a structure, which makes the implementation of exception handling in batch processes easier. The exception handling approach has been implemented in JGrafchart and tested on the batch pilot plant Procel at Universitat Politècnica de Catalunya in Barcelona, Spain. The second part of the thesis is focused on fault detection in batch processes. A process fault can be any kind of malfunction in a dynamic system or plant, which leads to unacceptable performance such as personnel injuries or bad product quality. Fault detection in dynamic processes is a large area of research where several different categories of methods exist, e.g., model-based and process history-based methods. The finite duration and non-linear behavior of batch processes where the variables change significantly over time and the quality variables are only measured at the end of the batch lead to that the monitoring of batch processes is quite different from the monitoring of continuous processes. A benchmark batch process simulation model is used for comparison of several fault detection methods. A survey of multivariate statistical methods for batch process monitoring is performed and new algorithms for two of the methods are developed. It is also shown that by combining model-based estimation and multivariate methods fault detection can be improved even though the process is not fully observable.