A hybrid simulator for improved filtering of noise from oscillating microbial fermentations

Many microbial processes exhibit sustained oscillations under practical conditions. Often these oscillations are corrupted by noise from the environment. It is then important to filter out the noise and recover ‘true’ oscillations for subsequent studies. Previous studies have used algorithmic or neural or hybrid filters in conjunction with mathematical equations for the kinetics. However, this approach is limited because in real situations it is difficult to measure and model some of the variables. Therefore, a hybrid neural simulator (HNS) has been developed here and tested with the continuous fermentation with Saccharomyces cerevisiae. The HNS combines a hybrid neural filter (HNF) for the noise, a hybrid description of the fermentation kinetics and macroscopic balance equations for the bioreactor. The HNS achieved 96% recovery of noise-free oscillations, compared to 91% with an HNF and lower efficiencies with pure neural and algorithmic filters. The commonly employed extended Kalman filter was ineffective as a stand-alone device but contributed to good filtering by the HNF and the HNS, thus indicating that a proper distribution of variables between the mathematical and neural components can significantly improve the performances of both.