Process Model

The object of research is a production of antibiotics by fed batch fermentation. An intelligent system with a structure of neural network and a genetic algorithm are used in nonparametric modeling of the fermentation process. The goal is to build a model for the prediction of fermentation eeciency. A priori knowledge of experts, who are controlling the production of antibiotics in industry, is utilized to deene the set of features from measured process variables. An optimal subset of features for the process description is determined during the optimization by genetic algorithm together with formation of the model. A linear model, a radial basis function neural network and a hybrid model are applied for the prediction of the fermentation eeciency and the best results are observed when using hybrid linear{neural model. Fermentation is used to produce a wide range of substances in pharmaceutical, chemical and food industry. Antibiotic production by fed batch fermentation is investigated in this paper. The antibiotic is produced as a secondary metabolite by the microorganisms. Secondary metabolic fermentation is a complex nonlinear process. Very high costs, inconvenient environment load and high energy consumption are characteristic for the production. Consequently, an optimization of the production is required and the rst step toward the optimal production is formation of process model. An appropriate model can then be used for the process eeciency prediction and for the production optimization. As secondary metabolic production is not well understood yet, there aren't any analytical models available at the present state of knowledge. A possible solution to the problem of fermentation modeling can be found in the area of empirical modeling. There are several references reporting successful application of neural networks to the fermentation research. Applications include biomass estimation in the penicillin industrial fermentation Willis 1990, Massimo 1992], on-line prediction of fermentation variables Thibault 1990] and a successful neu-roidentiication of an industrial secondary metabolic fermentation Tsaptsinos 1993, Tsaptsinos 1995]. Very promising approach to the fermentation modeling is a hybrid modeling, which combines diierent forms of knowledge with neural networks Thomson 1994, Fu 1996]. The goal of our work is to investigate the possibilities of nonparametric modeling of the considered fermentation process. The aim is to build a process model for the fermentation eeciency prediction. A general scheme of fermentation modeling is presented in Figure I{1. The rst step toward fermentation modeling is to collect available process data and to create a fermentation sample …