Research On Multiphase Soft Sensor Modeling For Fermentation Process

The microbial fermentation process is more and more important in modern industrialprocesses, because of the complexity of the reaction mechanism, highly non-linear anduncertainty of the model, relative to other industrial processes, its difficult modeling and verylow degree of automation affects the real time optimization control and brings hugedifficulties to the industrial production. Based on the mathematical model of fermentation canachieve some measure of unpredictive parameters and optimize control system, so as toimprove the monitoring and optimizing of industrial production process. Therefore, studyingthe microbial fermentation process has important significance and practical value to theindustrial production process.Least squares support vector machine (LS-SVM) with fast modeling, better forecastingperformance and stronger generalization ability, but it is lost some robustness. Weighted LeastSquares Support Vector Machines (WLS-SVM) weighted the square of the prediction errormakes up for the lost robustness problems of the LS-SVM. The property of WLS-SVM ismainly depends on two parameters: the penalty parameter and the kernel parameter. In orderto improve the accuracy of the established model, this paper uses quantum particle swarmoptimization (QPSO) to optimize the two parameters.This paper in-depth analysis of the penicillin fermentation process, combined with themultiphase characteristics of the microbial fermentation process, research on multi-modelmodeling for it. For the penicillin fermentation process has the characteristics of highnonlinearity and uncertainty, firstly, using the Fuzzy C-Means algorithm for the datapreprocessing and gets several stages, then establish multi-stage WLS-SVM sub-modelswhich optimized by QPSO, fnally, fusing the estimation of every sub-model by computing theweighting coefficient which is the fuzzy membership of predict sample to the cluster center.Through the simulation and comparing with the global model, the results show that thismethod has higher prediction accuracy and better generalization ability, and this model withhighly reference value for similar problems.