Comparative study of black-box and hybrid estimation methods in fed-batch fermentation

A neural network based softsensor is proposed for a PHB fed-batch fermentation process. The softsensor is designed to estimate the biomass concentration on-line. The design is based on the following model structures: 1. a feedforward neural network, 2. a RBFN (radial basis function neural network) and 3. hybrid models composed of either feedforward or RBFN neural network and the a priori known dilution term of the mass balance equations. The different designs are experimentally implemented and compared using Alcaligenes eutrophus as a model fed-batch system. Additionally, the possibility of directly inferring the substrate (glucose) concentration from the estimated biomass was investigated by assessing the variability of the corresponding yield coefficient. The combination of the neural network model and mechanistic differential equation provided the best results. Because of the variability in the yield coefficient, substrate concentration could not be inferred directly.