Abstract: As for the problem of softsensor modeling estimation precision in chemical processing, a method of combination SVM softsensor modeling is proposed based on a theory of multiple knowledge bases， Supervised Locality Preserving Projection(SLPP). The dimension of input data is reduced by SLPP between clusters, and then different transformed matrices and different multiple knowledge bases are gotten for every category. At last, a combinationmodel is constructed by support vector machine adaptively. Applied this method to a softsensor modeling of bisphenol A content, the simulation results show that every submodel weighted is more reasonable as compared with traditional multimodel, the estimated accuracy of model is improved, and the generalization ability is better.