A novel modeling method integrated KPCA with RVM was proposed． The kernel primary component analysis （KPCA） was employed to identify the principal components from the nonlinear transform data of independent variables，which were regarded as character variables．Regression between character variables and dependent variables was done based on RVM，and the optimal number of the character variables was adaptively determined according to the generalization performance of the regression model．Thus ，KPCA－RVM method could eliminate the disturbance of redundant information and achieve the best nonlinear model with good generalization performance．The method of KPCA－RVM was demonstrated by a 4－CBA’s content soft －sensing of PTA．Simulation results show this method is effective and the performance is better than those of PCA－RVM and RVM．

For the high dimensional chemical engineering processing data, an outlier detection method based on self-organizing map (SOM)networks and its visualization methods was proposed. Practically, it was applied for the observed data of preflash tower and the satisfactory result was obtained. Firstly, SOM was applied to obtain the topology-preserving plane for the high dimensional data. Then, based on the mapping plane and its visualization methods, the outliers were visualized clearly and easily. The results show that the proposed method does not need complex calculation, and the outliers in high dimensional data are effectively detected and eliminated.