Neural Network Applications in Bioprocesses

摘要

Feed-forward neural networks with a backpropagation learning algorithm wereprogrammed under MS-Windows environment and implemented on an Intel i486/DX2-66Mhz based personal computer. The characteristics of neural networks such as the number of hidden neuron, type of transfer funtions, the quality of the training data files, learning speed, etc., was analyzed. Simulation studies on the use of a neural network as a filter for process noise, and for one-step ahead prediction of state variables in batch and fed-batch bioprocess were carried out based on modelling data. The ability of neural networks in handling noisy data in dynamic processes, and in the estimation of missing values was demonstrated. In a start-up phase of chemostat ethanol fermentation, a neural network with well trained data from fermentation with positive initial time derivative of glucose concentration satisfactorily predicted substrate, biomass and ethanol concentrations, even in the case when the intitial time derivation of glucose concentrations was negative or the substrate level was very low. Further, the real measurement data of state variables in the glucoamylase fermentation were successfully used for multi-step ahead neural prediction of enzyme activity and biomass concentration. A complex, efficient MIMO dynamic cooking extruder controller was constructed by integrating two neural network models, a process model and a controller.