Bayesian Methods And Its Applications In Chemical Soft-sensor

Bayesian Learning Theory represents various knowledge and uncertainty with probability. The learning and inference are realized by probabilistic rules. Therefore, it is a strong tool dealing with uncertain information. This thesis mainly studies the basic point,background and status quo of Bayesian Learning Theory. Several Bayesian classifying model are also discussed in detail in the paper. Na?ve Bayesian classifier and its improvement,Bayesian regression support vector machine and its applications in soft sensing of polyacryl0nitri process are mainly studied as the key problem.Classification is a very important task of data mining, its purpose is to find out classifying function or classifying model. This thesis proposes NBC based on mutual information. A developed attribute importance measure method for attributes reduction is defined from the viewpoint of information theory. It can find a near independent subset so as to weaken the dependent relationship between attributes. Then the subset is trained by the NBC, improving the performance of NBC.The generalization ability of support vector machine is poor in small sample. This thesis brings Bayesian Evidence framework into support vector regression problems so as to tune the regulation and kernel parameter approaching neal-optimal. Thus, the generalization ability of the model is improved in small samples.Furthermore, considering the polyacrylonitri production process is a complex,nonlinear,time-varying process. This paper establishes a multi-class soft sensor model to evaluate the quality figure of the polyacrylonitri producing process. This model is based on the processing mechanism model. The SVM model and regression and identification algorithms are used to estimate the process parameter of the sub-model or mechanism model. Meanwhile, various prior information as the equation or in-equation restriction is introduced into the mixing model. The validity of the method is demonstrated by the experiment results.