LEARNING TO CONTROL THE PERFORMANCE OF BATCH PROCESSES

The current shift in chemical industry interest away from high through put production towards small amounts of high value added products has increased the industry's awareness of the issues associated with produ ct and process performance. This fact poses a unique set of challengin g control problems characterized by their 'ill-definedness', high proc ess nonlinearities and imperfect modelling. In this work, a novel meth odology is proposed for incremental learning of a control policy that can continuously improve product quality and operating performance. Th e new concept introduced here is the notion of a performance function that implicitly includes end-product quality constraints as a process goal and operational preferences which describe different modes of ope ration. Since plant information is often scarce and expensive to obtai n, it is proposed that the performance function be learned by integrat ing together batch-to-batch data, intra-run measurements and a predict ive process model. A new framework for this integration, called coarse code generalization, is proposed which revolves around the generation of an artificial set of batch runs using a hybrid process model. In t his model, learning biases are incorporated through background knowled ge that expresses run outcome sensitivities with respect to states and actions. Artificial batch runs provide an augmented data set which is used for inductive learning of the performance function. With a minim um amount of information on process performance available, a first app roximation to the performance function is constructed and an optimizat ion program is used to define a near-optimal control policy. As more p lant data become available, the performance function refinement proced ure permits also an increasing refinement of the learned control polic y. Recipe changes that increase process performance can then be implem ented on-line. A semi-batch reactor where an autocatalytic reaction ta kes place is used to present and test the methodology. Results obtaine d demonstrate that the methodology can cope successfully with the prob lems of both imperfect modelling and scarce information which are typi cal of the industrial environment. Also, coarse code generalization fo r performance control proves robe an ideal means of integrating induct ive learning with first-principles models and other types of domain kn owledge.