A simulation study of dynamic neural filtering and control of a fed-batch bioreactor under nonideal conditions

Abstract

Bioreactors utilizing genetically modified bacteria under realistic conditions are difficult to monitor and model because of imperfect mixing, disturbances and uncertain kinetics. In this work, neural filtering and control have been applied to such a nonideal fed-batch bioreactor containing recombinant Escherichia coli to produce β-galactosidase. Data simulating industrial fermentation were generated by introducing incomplete mixing in the broth and Gaussian noise in the feed stream. Based on a previous study, an Elman neural network was employed to represent the simulated data. To this, an autoassociative network was added in order to filter the noise and a feed forward network to control the fermentation. Performance of the fermentation with this system of three neural networks optimized together has been shown to be superior to sequential optimization, neural control without filtering, PID control with filtering, and also a noise-free fermentation. Thus, a suitably designed system of neural networks provides rapid on-line estimations and improves bioreactor performance under conditions simulating industrial fermentation.