Microbial Fermentation Process Recursive Fuzzy Neural Network Soft Measurement Method Based On Dynamic Study

Microorganism fermentation engineering is a complex biochemical reaction process which has high nonlinear, time-varying and hysteretic, and which internal mechanism is very complicated. The traditional measurement methods are difficult to measure some key variables online, such as cell concentration, substrate concentration, production concentration, which makes the whole fermentation process to optimize control becomes very difficult, soft-sensor technique is one of the most effective ways to solve this problem.This article take lysine fermentation process as research object, dynamic recurrent fuzzy neural network soft-sensor model is build to predict the three important variables (cell concentration, substrate concentration, production concentration) in fermentation process on the basis of the soft-sensor theory. The simulation results show that the soft-sensor model can accurately predict the key variables, and it is with preferable stability and can predict correctly under the situation of disturbance, which provide the precondition for optimizing control in the fermentation process. The concrete work is as following.Firstly, on the basis of reading literatures and studying lysine fermentation experiment, according to a curve of microbe metabolizing in the factual fermenting process, the soft-sensor models of based on Fuzzy Neural Network and dynamic fuzzy neural network are respectively established. Then the variables in the fermenting process of lysine such as cell concentration, substrate concentration, production concentration are predicted, and the performance of two models were studied and compared.Secondly, according to the research of assistant variable selection and data pretreatment method in soft sensor, the kernel principle component analysis is adopted to fix assistant variable and experimental data are processed with the modified median minimum distance algorithm(MMMD)based on Mahalanobis distance.Thirdly, with the deficiency of initial value that is extremely sensitive, easy to fall into local minimum, a fuzzy C means clustering algorithm based on data field is proposed, thus conquer the blindness of random selection of initial centers so as to quicken the speed of the operation of fuzzy C means clustering algorithm.Fourthly, on the basis of dynamic recurrent fuzzy neural network, using the improved fuzzy C means clustering algorithm to identify the model structure, applying the immune genetic algorithm to optimize the structure and parameters of model at the same time, in accordance with the simulation results, the improved soft sensor model can more effectively and quickly approach the true value, significantly improve the prediction accuracy.