Application Research Of Soft Sensor Model Based On The Bayesian Networks

The increasing complexity of modern industrial processes makes a large number of process variables in industrial processes difficult or even impossible to measure on-line. In order to solve the measurement problems of such variables, soft sensor technology has got rapid development, as an effective means for online measurement in process control,soft sensor has become a hot topic in the field of contemporary industrial control. As Bayesian network model can be graphically representation, inference unknown in the sample data information, coupled with the characteristics of its simple structure and clarity, this article focuses on how to apply Bayesian Network in Soft Sensor Modeling.The main contents of this paper include the following aspects:(1) Introducing the specific steps to establish a soft sensor model in detail, to provide a base for a right soft sensor model in this article. First, learn and research the soft sensor model based on support vector machine method, and simulated by the model to predict an industrial process PTA 4-CBA content.Verify the availability and accuracy of soft sensor technology.(2) Since the Bayesian network structure int the soft sensor model in this paper is certain, the parameters of Bayesian networks should get more study and research. To study parameters, the main method is EM algorithm, combined with the existing sample data to estimate the parameters of the network, obtain the best value, and ultimately through the network do a predictive analysis of the test sample. In Bayesian network parameter learning, cite three examples to emulate verify its feasibility, the simple linear function, simple non-linear function, as well as a series of sample data, simulation predicted results is good,makes its application in soft sensor modeling possible.(3) Since the sample data in this industry tend to have large values, and each of the input variables have different dimension, so before enstablishing model the input sample data should be normalized so that the model can be calculated easer in the back.In the Bayesian network,it is difficult to learn parameters directly through the maximum likelihood estimation method, and thus this paper introduce a Gaussian mixture model, make this model approach Bayesian network, by computing Gaussian mixture model to replace the mixing coefficients Bayesian computing network parameters. In Bayesian network, Based on the relationship of the various input variables that were analyzed to determine the Gaussian component, determine the final Bayesian network model, and finally, use an industrial process, the content of 4-CBA forecasts for this article proposed model simulation, and analyze the number of Gaussian components and the effects of different combinations of ingredients gauss predicted results and analyze the pros and cons of the Bayesian network model and SVM model.