A soft sensor modeling method is proposed by combining the kernel principal component analysis (KPCA) with the support vector machine (SVM). Via KPCA the method is able to capture the high-ordered principal components among the secondary variables, and use SVM to establish a correlated regression model between the featured principal components and the primary variable. The proposed KPCA-SVM method is used in soft sensor modeling for the freezing point of light diesel oil. Compared with the models of linear PLS, linear SVM and PCA-SVM, the result obtained by the KPCA-SVM approach shows better estimation accuracy and is more extendable