An advanced model based framework for improved monitoring and quality control of batch and semi-batch processes
One feature of batch processes that is typically ignored, in the formulation of monitoring and control strategies, is that significant correlation can exist in a batch-to-batch sense. The source of this correlation arises from frequent disturbances in the feed, utility sources, or from within the reactor itself. This dissertation shows that a understanding of the complete correlation structure of the process can be beneficial in terms of deriving successful quality monitoring, inferential prediction, and control formulations. ^ A data-based framework is introduced that extracts the intra-batch and inter-batch correlations from historical data of the plant. It is shown that the resulting model has utility in deriving both an improved monitoring strategy and a inferential quality prediction model. An important feature of the developed data-based model is the ability to integrate all information from the plant including both on-line process measurements and off-line quality information. A natural extension of this data-based model is its use in control, where the form of the model naturally complements both on-line and off-line control strategies. ^ The application of fundamental models to batch processes is also considered. The estimation of uncertain parameters is formulated and it is shown that the performance of the model can be unproved by modeling parameter uncertainty and model errors as the output of a dynamical-stochastic system that explains that model uncertainty can have batch-wise correlated portions as well as purely independent type changes. The inclusion of this batch-to-batch model into the estimator design allows for improved convergence on model parameters and lifts a fundamental limitation of current estimator design methodologies which have to be formed based on a finite number of measurements within a particular batch. By adding a batch-to-batch component to the model the ability to carry over information from batch-to-batch is added which allows the use of more measurements for convergence on model parameters. ^
Major Professor: Jay H. Lee, Purdue University.