Research On Bayesian-Kernel Learning Modeling For Batch Processes

In the Batch Processes, some key parameters, which couldn't be measured by the sensors directly, play a very important role in the product quality control and the improvement of process automation.It is no reliable mechanism, time-varying intensity, functuations of raw materials in Batch Processes that causes some difficulties in modeling and controlling and also leads to product quality and energy consumption out of effectively controlled. Currently, support vector regression(SVR), least squares support vector regression(LSSVR) and other similar methods based on data-driven modeling approach,are widely used to solve these problems and achieve some success.However, we are confronted with some common problems:1.the uncertainty of data, correlation and linear sensitive.2.some methods requires an estimate of the regulatization parameter and the regression also need to adjust the sensitivityparameters.3.most of the SVR,LSSVR research are still used to bulid off-line model.Litter attention is payed to the online modeling research.To solve the problems above, we conduct some research on the Bayesian Network Kernel Learning Modeling for Batch Processes.1. In order to improve performance and to promote accuracy, and make full use of prior probability information, we propose a Bayesian-Kernel Learning modeling for Batch Processes, assuming a priori given probability distribution and the distribution of the likelihood conditions, constantly revised prediction model, to improve the model prediction accuracy of the nonlinear process and promotion of performance. Through some simulating expriments in the Penicillin Fermentation,the results shows that the B-KL can effectively obtain the parameters of the model,having better performance than the traditional neural networks, LSSVR.2. We introduce a Recursice Bayesian Regression and Recursice Bayesian-Kernel Learning Modeling approach with the sparse strategy of samples based on the model prediction error. When online modeling, we propose a method to delete redundant sample based on forecasting accuracy without the key sample.Rubber Mixing Process by a formula of actual simulation data show that the R-BKL has a better forecasting performance with some on-site industrial applications than RBR,B-KL.3. R-BKL is applied in Mooney Viscosity online prediction of Rubber Mixing Process in a domestic tire manufacturing company. We design an online modeling strategy in the implementation of the rubber mixing process. Also, the approach is develop as a Mooney Viscosity online prediction System by the on-site developers. The test result by the system show that it has a good performance in forecasting some key parameters of Rubber Mixing Process.Finally, we make some conclusion and take some consideration into the further research assignments in the future.