Fuzzy and neuro-fuzzy modeling of a fermentation process

Neuro-fuzzy modeling may be qualified as a grey-box technique, since it combines the transparency of rule-based fuzzy systems with the learning capability of neural networks. The main problem in the identification of non-linear processes is the lack of complete information. Certain variables are, either immeasurable or difficult to measure, the soft sensors are the necessary tools to solve the problem. Those latter can be used via online estimation, and then they will be implemented in fedbatch fermentation processes for optimal production and online monitoring. The process parameters are estimated through a fuzzy logic system. The fuzzy models of takagi-sugeno type suffer of the problem of poor initialization, which can be solved by the trial-and error method Trial-and-error method is used to solve the poor initialization problem of TS models, this deals with identifying the structure of the model, such structure consists on finding the optimum number of rules, which enters in the model cost reduction. The fuzzy model might not capture the process non-linearity, especially if the number of rules is overoptimized. Bioreactors exhibit a wide range of dynamic behaviours and offer many challenges to modeling, as a result of the presence of living micro-organisms whose growth rate is described by complex equations. We will illustrate the fuzzy and the neuro-fuzzy modeling on the identification of such a system. In order to compare the NF model outputs, we use another fuzzy model that does not incorporate the neural network learning capability, to identify the parameters of the same process. Even though, the two models were trained using levenberg-marquardt algorithm, the corresponding simulation results show that a better modeling is achieved using NF technique, especially that we did not employ any involved optimization procedure to identify the NF structure.