engleski

Control of baker"s yeast production by neural network model

Based on PCA analysis developed are NN models for control of baker"s yeast production in 40 m3 industrial jet bioreactor. By PCA analysis ( linear projection ) the state space of 9 process variables is reduced to a space of 3 variables which account for 95 % of total data variance ( information contained in measurements). Based on PCA analysis NN structure is predetermined, i.e. the number of neurones on the hidden layer. To optimize NN predic-tivity applied are disjunct sets of patterns for training and testing. The ratio of number of patterns in the two sets is 3 :1 in favour of the testing. NN parameters are selected for minimum variance over the testing set. Applied is conjugate gradient algorithm for variance minimisation yielding very fast convergence NN models of MISO direct for controlled, and inverse structure for manipulative variable are developed based on computer simulation and plant data. Process dynamics is accounted by ARMA patterns. Optimal NN structure reveals that non-linear projections to lower dimension of 2 is sufficient to account for better than 95 % of in-formation. Accuracy of the NN model for simulation data is on av-erage of 1-2 % while for industrial data is 3-5 %. The NN for di-rect prediction is more accurate than for the inverse model. The NN models give accurate predictions ( extrapolation ) outside the space of training patterns. Accurate prediction outside training sets indicate that rather than mere interpolation between training data properly structured and trained NN can "learn" association rules between input and output. Discussed are NN internal model control ( NN IMC ) and ideal inverse model feedback control of bioreactor.

yeast production; deep jet bioreactor; adaptive control; neural networks

nije evidentirano