UTILIZATION OF NEURAL NETWORKS AS SOFT SENSORS TO MONITOR

To properly control emulsion polymerization reactions in a closed loop procedure, it is necessary to monitor important process variables, as conversion and average particle diameter. Conversion can be measured or inferred on-line by densimetry, gas chromatography, or calorimetry, but many industrial plants do not have the necessary equipment. Monitoring particle size on-line during emulsion polymerization reactions is still much more complex. Up to date, there are very few equipment available to measure this variable on-line, but only for lab scale application. Due to monitoring difficulties and the requirement of quite complex mathematical models presenting high computational costs that are inadequate for on-line simulation and control, artificial neural networks are winning more importance in the polymer reaction engineering area. Experimental data of vinyl acetate and Veova10 emulsion copolymerization reactions with 55 % solids content carried out in a continuous loop reactor were used to train and to validate these networks. In order to improve robustness stacked neural networks were used. The networks were trained with experimental data of several reactions capturing a large range of operational conditions and were validated with experimental data of other reactions presenting a quite complex dynamic behavior as strong oscillations in average particle diameter and conversion. The results obtained by the neural networks demonstrated a good agreement with experimental data. 1