For maintaining the safety of industry process and the quality of product, on-lineprocess monitoring and real-time fault detection and diagnosis become the major research topic in the field of process control. The statistical method based on the data is more practical, because this method is independent of process model and the required process data can be obtained easily in industrial process. This method need to collect the data both under the normal condition and fault condition. The diagnosis process includes fault detection and fault identify.This thesis particularly introduces the mainly used fault diagnosis method in industrial process, and analyzes the advantage and disadvantage. As the background of the Tennessee-Eastman Process, analyze the industrial process. Introduce the theory of principle component analysis and kernel principle component analysis, and the use of multivariate statistical fault detection method in the TEP. The KPCA introduce the Kernel function concept, and map of the data set from its original space into a hyperdimensional feature space and carry out PCA, then make the data have good separability. KPCA is the extend of PCA, it conquer the disadvantage of PCA can only reduce the dimension in the linear space.. Use the two methods in the TEP, educe that the multivariate statistical fault detection approach have a better performance than PCA.Support vector machine have theoretical maturity, global optimization, excellent generalization, also have good classify ability. The thesis improves the conventional SVM, and uses several one-versus-one classifiers to construct multi-class classification. Combine the PCA and KPCA with the SVM; input the linear PC and nonlinear PC to the SVM. Make the model used in the TEP, draw that the KPCA-SVM have a better fault diagnosis ability. |