(203d) An Expert System for Decision Making in Fermentation Process Systems Using Bayesian Network

Decision making under uncertainty could not be handled with usual deterministic methods such as neural network. Therefore, Bayesian network as a probabilistic approach is the best option for expert system design under uncertainty. Unlike deterministic methods, Bayesian network represent each parameter by probability distribution, which is more beneficial in decision making. In the present study an expert system is proposed for the process of fermentation which three kinetic equations are occurring in a stirred tank bioreactor system using Bayesian Network. The main variables are substrate volumetric flow rate, biological cell concentration (dry mass), substrate concentration, and product concentration. Some variables were obtained by sensor signals and others were determined by solving the equation of material balances for a well-stirred tank fermenter.  Incomplete dataset is available for decision making due to operator negligence or error in sensors. In order to encounter incomplete dataset, Bayesian network is superior to other approaches. Bayesian network structure learning is done with K2 method. In parameter learning stage, while completed data is available maximum likelihood estimation (MLE) method is used; when partial of data are missed, expectation maximization (EM) method is applied. In inference stage, the junction tree and variable elimination algorithm are used which make different trade-offs between speed, accuracy, complexity and generality. Junction tree algorithm is too slow and variable eliminations algorithm is one of the simplest inference algorithms.