Saturation-type transfer function for backpropagation network modeling of biosystems

A saturation-type transfer function, bx÷1+|x|, with back-propagation type of neural network (BPN) was proposed for solving problems of several bioreaction systems. The biosystems include multiple components separation, batch cell culture, on-line monitored fermentation system. This saturation-type transfer function was successfully applied to the simulation/prediction, dynamic identification of these practical systems. For the separation of multiple components by adsorption, BPNs with this saturation-type transfer function were applied to the modeling of a series of multicomponent adsorption systems. The results show that the isotherms obtained from the neural network approach well correlate with the experimental data. For batch cell cultures, the initial state strongly governs the growth pattern. A 2-3-8 BPN with initial glucose and cell inoculum as the two inputs, cell densities measured at eight each hours as the eight outputs was constructed. The simulation and prediction results demonstrate again the performance of this transfer function. The ability for extrapolated prediction is also shown. For the on-line monitored fermentation, An inverse-type neural network model of 11-3-1 was designed for the identification of this fermentation. It is modified being able to predict the dynamic response of the 2,3-BDL fermentation. The one-step ahead identification/prediction of this dynamic BPN is thus performed.