Research On Prediction Of Penicillin Concentration By Using Multi-Model Approach Based On RBF Neural Network

Penicillin is a sort of very important antibiotics. The main production of penicillin is fermentation. Penicillin fermentation is an extremely complex biochemical reaction process with strong non-linearity, time-variability, and uncertainty, and its precise mechanistic model is hard to build. The essential requisition of the optimization to the fermentation is the availability of the online value of biomass including substrate concentration, cell concentration and penicillin concentration etc. At present, the biomass is mainly detected by off-line empirical testing methods through sampling. It may bring other bacteria into the fermentation tank while sampling, and make adverse effects on the fermentation, leading to a decline in yields and quality of penicillin. Penicillin concentration is the most important biomass, and this thesis will carry out a deep stuy on its online prediction.Currently, a variety of soft-sensor modeling methods are usually used to build a global model for online prediction of penicillin concentration. However, when the volatility of initial conditions and operational parameters of fermentation increases, as a result, the differences of batches’characteristics ascends. In this case, the accuracy of the global model comes down and its prediction error highlights because of the shortcomings of the single personality of the model. To be directed against this deficiency, a multi-model approach of modeling is proposed in this thesis.The subtractive clustering algorithm is used to classify the batches in the sample into several classes automatically according to its initial conditions and operational parameters of fermentation in the first stage. The result of clustering is that the characteristics of batches in the same class are as similar as possible and of a high dissimilarity when they’re in different classes. Then, RBF neural network is used to build the prediction model for a single batch which is designated to the modeling unit. Under the circumstances of an accent analysis on the mechanism and technics of fermentation, the inputs of RBF neural network are dicided to be these key vaviables:sampling time, culture volume, dissolved oxygen concentration, CO2 concentration, fementor temperature, pH, aeration rate, agitator power, substrate feed rate, and substrate feed temperature. The training algorithm of RBF neural network is K-means clustering based on subtractive clustering algorithm, and gradient descent algorithm. This training algorithm can automatically determine the number of hidden neurons of the network in terms of the characteristics of training data, and also adjust the centers and widths of RBF and the weights of hidden layer to output layer during training to the network, improving the accuracy of the batch model. Subsequently, the batch modeles of a class can be combined into a class model by means of the optimal weighting method which works out the weighting coefficients by minimizing the sum of squares of the combination of prediction error of the batch modeles. This kind of class model is one of the optimal modeles of a class. So, a library of multiple class modeles can be built for the fermentation whose differences of batches’ characteristics are remarkable. In the last stage, a calss model is selected from the model library for the current test batch in terms of the minimum distance among the distances of its initial conditions and operating parameters of fermentation and that of the centers’of all classes to get down to the work of online prediction of penicillin concentration of the test batch.At last, Pensim is taken as the simulation platform for the penicillin fermentation and orthogonal experiment is used to design the modeling sample which the batches included in are of small amount, comprehensive information, and uniform distribution. The simulation results show that the accuracy and ability of generalization of the prediction model built according to the modeling method proposed in this thesis are well guaranteed, and the performance of the prediction model is better than that of a single global model’s in the case of the fermentation with a notable volatility of batches’characteristics.