Research And Applications Of Multi-model Modeling Method Based On Support Vector Machines

In modern complicated industrial process, some variables are very hard to be measured or even cannot be measured on-line by existing instruments and sensors. Soft sensor is an effective device for on-line evaluating these variables. The Support Vector Machine has been widely used in soft sensor modeling in petrochemical and chemical process, because of its ability of approaching any linear functions or nonlinear functions at arbitrary precision. However, it is still difficult to describe the complex nonlinear production process precisely by a single soft sensing model, since there are the features of high nonlinearity and multiple working conditions in the practical production processes. According to the engineering application background, three methods for multi-model modeling are proposed in this thesis:Dividing the sample set is the prerequisite in multi-model modeling, and the clustering algorithm is the popular method for data division. But the cluster number c has to be assigned in advance, which causes great restriction to a certain extent. An improved satisfactory clustering algorithm is presented to solve the commonly problems in usual clustering algorithm. By using the novel clustering algorithm, a nonlinear system can be quickly divided into several parts, and the multi-model can be built according to the subsets.The output data can reflect the system veritably and it is significant to the study on the system variation. The Quadratic Discriminant Analysis algorithm is introduced to data classification, which can divide the sample set into several subsets according to the output data of a system. Firstly, the known sample set is divided into several classifications by taking the value interval of the object variable as the criteria of prior classification. Secondly, the QDA algorithm based on the prior classification information is used to classify the other input data. Finally, the multi-soft model is established according to the data classification.The completeness of the data is necessary to build highly accurate model. Data information includes input information and output information for an object. It is difficult to characterize the process of the object variation by dividing the sample only according to the input information or output information separately. To preserve the completeness of sample data information and classify the samples in terms of the input information and output information of the system, a multi-model modeling method is proposed based on the CAQDA.The fuzzy clustering method is used to divide the sample data set into several subsets according to the similarity of the input information. Second division is performed on the subsets with QDA according to the output information of the system, and sub-model is built for each class with SVM. Based on the novel method, a soft sensing model is established to measure the content of Bisphenol A yield in the bottom of the rearrangement reactor for the Bisphenol A production decomposition reforming unit. The new model is better than the single SVM model and multi-SVM model based on the fuzzy clustering in the model precision and generalization power.