Study Of Soft Sensing Modeling Based On EGK'M-RBF Neural Network And Reinforcement Learning Control Algorithm

Firstly, to solve the problem of on-line measurement of mooney viscosity of polybutadiene rubber, a soft sensor of mooney viscosity was proposed by using soft-sensing technology. Through analysis and researches on the process of polybutadiene rubber, an modeling method by radial basis function(RBF) network based on enhanced global K'-means algorithm(EGK'M) was presented, a soft-sensor model of mooney viscosity based on EGK'M-RBF network was established using field data of the process of polybutadiene rubberSecondly, in order to tackle the control problem of complex industrial process, a method for complex industrial process control using reinforcement learning algorithm was proposed. Through research on the process of Saccharomyces cerevisiae fermentation and reinforcement learning algorithm control theory, an improved multi-step action Q-learning control algorithm is presented. Algorithm was developed to control the ethanol concentration of the Saccharomyces cerevisiae fermentation process. Main contributions of the thesis are as follows:1. At the beginning, overviews are made both on soft sensing technology and reinforcement learning technology. Then, principle theory of soft sensing based on RBF network was introduced, including the advantages and disadvantages of RBF network center selection algorithms, the advantages and disadvantages of PCA and KPCA for extracting non-linear feature information and achieving selection of auxiliary variable.2. An enhanced global K'-means clustering algorithm is presented, and it had been developed for clustering Gaussian datasets and several actual datasets. The clustering results of the actual datasets demonstrate that the enhanced global K'-means algorithm can get better clustering results compared to the modified global K-means algorithm and K'-means algorithm, respectively. Then, the enhanced global K'-means algorithm was applied to determine the structure of the hidden layer of RBF network. A modeling method by radial basis function (RBF) network based on enhanced global K'-means algorithm (EGK'M) was presented. In the proposed method the structure of RBF network was detrmined through EGK'M algorithm, KPCA algorithm was used for non-linear feature information and secondary selection of auxiliary variable. Finally, in the thesis, a series procedure of on-line self-calibration model algorithm which based on EGK'M-RBF network was given.3. Modeling of mooney viscosity of polybutadiene rubber with EGK'M-RBF network was proposed. From the comparative analysis of the modeling results, one can see that advantages of EGK'M-RBF network lies in that the model proposed much better fitting results, stronger predictive ability, smaller absolute error. At the same time, comparative analysis on PCA and KPCA for extracting non-linear feature information shows that KPCA is more suitable for non-linear feature extraction. Finally, a brief introduction to the development and interface design of soft-sensing software package of Mooney viscosity of polybutadiene rubber was given.4. An improved multi-step action Q-learning control algorithm was proposed for the process of Saccharomyces cerevisiae fermentation, which combines multi-step action Q-learning algorithm and a fuzzy control gain parameter selector. The fuzzy control gain parameter selector was used to adaptively select the control gain parameter, it can lead to faster tracking and help to alleviate the overshoot of controller. Experiment results show that the improved multi-step action Q-learning controller has much lower overshoot, faster tracking, shorter transition, and smoother control signal and so on.