Multivariable control of a liquid-liquid extraction column using a probabilistic automaton

Describes an application of a probabilistic automaton to the control of a liquid-liquid extraction column which presents a highly nonlinear behaviour and time-varying dynamics. The control algorithm is based on a hierarchical structure of automata operating in the random medium corresponding to the process to be controlled. The learning system collects and processes the current data (input/output measurements) and changes its internal structure and associated probability distribution to maintain an optimal behaviour of the column. These optimal operating conditions have already been defined from hydrodynamic and mass-transfer considerations. The automata keep a probability distribution over their actions (process inputs) which is updated at each sampling period through a learning algorithm. This adaptation of the probabilities is performed through a reinforcement scheme, on the basis of information for the process behaviour delivered by a performance evaluation unit. The control variables are the pulse frequency and the solvent-feed flowrate. The conductivity measured at a specific place in the column (under the distributor) and the concentration of product in the output flow are selected as the controlled variables. No perfect or a-priori information on the system to be considered is assumed beforehand. Experimental results illustrate the successful performance of this probabilistic automation to control such a liquid-liquid extraction column.