Identification And Predictive Control Based On Fuzzy On-line Support Vector Regression

Support vector machine (SVM) is a learning machine based on statistical learning theory,which can solve small sample learning problem effectively. According to minimum structurerisk principle, SVM overcomes shortcomings of neural network, such as slow convergence,local optimum, and bad generalization. Therefore, SVM has been widely applied to manyareas: pattern recognition, signal treatment, system modeling and control et al., it hasgradually become a research hotpot in machine learning area. Up to now, severial SVMtraining algorithms have been widely used, whereas these algorithms are almost offlinetraining, which can’t adapt to nonlinear system real-time change. Based on previous results,online support vector regression (OSVR) theory and application were mainly studied in thisthesis. An advantage of OSVR is online learning ability, which can adjust model parametersonline. The main contents of this thesis are:(1) Fuzzy online support vector regression algorithm (FOSVR) is proposed to solveweak anti-jamming capability and slow training speed problems of OSVR. Through settingdifferent weighted factor to each sample and optimizing learning steps, FOSVR algorithmimproves model accuracy and training speed. Simulation results verified the improvements ofFOSVR algorithm.(2) Biological concentrations of fermentation process are hardly measured online,FOSVR is proposed to model glutamate fermentation process. Through online learning abilityof FOSVR, biomass concentration and production concentration soft sensor model were built.Futhermore, combining first principle model and FOSVR soft sensor model, a hybrid modelstrategy is proposed for fermentration process.(3) To solve predictive model mismatch problem, a nonlinear modle predictive control(MPC) based on FOSVR was proposed. FOSVR can modify predictive model parametersonline, single-step MPC and multi-step MPC were studied. Due to local optimum problem ofgradient decent method, particle swarm algorithm (PSO) was applied to rolling optimizemulti-step model predictive control.(4) For predictive functional control (PFC) bad performance in strong nonlinear andtime-varying system control, a PFC based on FOSVR inverse model was proposed. FOSVRwas used to obtain the inverse model of the Wiener model nonlinear part, the nonlinear PFChas been transferred to linear control. In this way, nonlinear objective function can be solvedby linear optimum algorithm. Aiming to improve performance of PFC, wavelet basis functionwas selected as basis function of PFC. Due to compact support and multi-scale analysis ofwavelet, the optimal parameters are aggregated.(5) Low automatic level of fermentation process and difficulty in biologicalconcentration control, therefore, the predictive control was employed to glutamatefermentation process feeding optimized control. Based on fermentation distribute controlsystem, Delphi and Matlab hybrid programing was used, so based on key biologicalconcentrations FOSVR model, substrate feeding predictive control was carried out. Throughthree feeding control experiments, the higher cumulative production concentrations were obtained.Through the stuy of this thesis, it is shown that FOSVR is an effective online modelingalgorithm. The algorithm not only can be used in solving modeling problem of fed-batchfermentation process, but also provide a powerful tool in complexity biochemical processcontrol. The study of this thesis has important significance for biochemical process.