Multivariable learning control of an extractor

Abstract

This paper deals with the application of the learning approach to the multivariable control of an extraction pilot plant. This separation process presents a highly nonlinear behaviour and time-varying dynamics. A learning system is composed of three distinct parts. The first part consists of an automation with variable structure whose actions are selected according to a probability distribution associated to the set of actions. An updating scheme (or reinforcement scheme) adapts this probability distribution according to the “good” or “bad” behaviour of the process. The quality of the behaviour is defined on the basis of heuristic rules contained in a performance evaluation unit through information collected in the process (measurements). The pilot plant to be controlled is a pulsed liquid-liquid extraction column. The control strategy involves both the control of the column in its optimal-behaviour zone and the minimization of the solvent flow rate needed to obtain a specific product quality. Previous works have shown that the column could be maintained in its optimal behaviour by means of the regulation of conductivity by action on the pulse frequency. The obtaining of a given product specification can be achieved by the control of the product concentration in the outlet stream by acting on the solvent feed flow rate. Owing to interactions between one variable and another it seems promising to use a two input-two output control scheme. Experimental results are presented which show the feasibility of such an approach for on-line control of a chemical plant.