Process characterization and control using multivariate statistical techniques

Fast paced developments in electronic hardware technology have resulted in heavily instrumented chemical plants. Process data from these units are frequently logged on to computers leading to data overload. To cope with these trends, data mining tools that extract useful information from the database have been proposed. These include methods based on simple visualization. multivariate statistical techniques (such as principal components analysis (PCA), partial least squares (PLS) and canonical correlations analysis (CCA)), artificial intelligence (induction or rule based) and neural networks. Recent studies indicate that a new data mining prototype is introduced every three months.; In this thesis, the use of multivariate techniques in the characterization and control of chemical processes (continuous and batch/semibatch) is explored. Utilizing the dimension reduction properties, these tools have long been used for applications related to process monitoring and fault detection in a statistical process control (SPC) framework. In certain situations (e.g. inferential model building), these methods have provided a robust alternative to the ordinary least squares regression procedure. Besides describing the theory and applications of these techniques in such traditional areas, we have investigated their suitability in the modelling and control of dynamic multivariable systems.; A powerful empirical (black-box) identification strategy that provides multivariable state space models (Canonical Variate Analysis, CVA) is reviewed. Extensive simulations are used to establish the superiority of CVA over another popular state space identification algorithm (N4SID). Extension of the CVA method to model a class of nonlinear systems, the Hammerstein structure, is provided.; Identification and control of univariate (single input single output--SISO) processes represents a relatively mature field; it is easily understood and readily implemented. We propose a novel multivariate modelling and controller synthesis strategy that is based on a combination of the PLS technique and the identification/control theory developed for SISO systems. Recognizing that, industrial plants usually operate in the regulatory mode, expressions for the design of multivariable feedforward controllers are developed. To cope with constraints on the process variables, the PLS model has been integrated into the Model Predictive Control framework. The domain of applicability extends to nonlinear systems--the Hammerstein and Wiener models provide motivating examples.; Case studies involving simulations, laboratory experiments and industrial data are included wherever appropriate.