Structure of Principal Component Based Neural Network Models of Dynamic Systems

Abstract

A new structure of neural network based systems for modeling and control of dynamic industrial processes is developed. The structure is composed of three serially connected subsystems: 1) ARMA - auto regression moving averages to account for system dynamics; 2) PCA - principal component analysis as statistical preprocessor for noise reduction and pattern compression 3) ANN - artificial neural network with static neurons and feed forward pattern propagation for nonlinear mapping of input/output interaction. Training of neural networks is performed with Ribiera-Polack-Powell conjugate gradient method for minimization of the variance in output patterns between a real and a model system. The proposed modeling procedure is applied to data from a fed batch operation of an industrial deep jet bioreactor. Predictive power of the model is based on the analysis of responses in pseudosteady and oscillatory mode of operation with the trained and untrained patterns. The aim of the work is to develop a general neural network structure and analyze its applicability in the process control in biotechnology.