Mathematical Modeling of Bioprocesses by Neural Networks
Ž. Kurtanjek
Faculty of Food Technology and Biotechnology, University of Zagreb, Pierottijeva 6, 41000 Zagreb, Croatia
Summary:
Mathematical modeling of bioprocesses by artificial neural networks (ANN) is presented. Bioprocesses are considered as complex, nonlinear and dynamic multiple input / output systems (MIMO). Proposed is a general structure of ANN model composed of three serially connected subsystems: auto regression moving averages (ARMA), module for principal component analysis (PCA), and subsystem with layers of static neuron networks (NN) with feedforward pattern progression. The ARMA subsystem accounts for approximation of process dynamics by finite differences. The PCA module has two objectives: 1) rejection of measurement noise and 2) data compression by removing of collinearity between measured process patterns, i.e. reduction of a high dimension input vector to a few principal components. The NN provides highly adaptive interconnectivity between input and output patterns, and approximation of their nonlinear functional dependence. Parameters of neurons are adapted by use of conjugate gradient technique with the Ribiére-Polak-Powell algorithm for minimization of variance between the ANN model and measured output test patterns of a bioprocess. Applicability of ANN models in biotechnology is illustrated by models for prediction of protein secondary and tertiary structures based on amino acid sequences, process identification in production of penicillin, and the study of ANN internal model control (IMC) in industrial production of baker's yeast.