To solve the problems of real-time online measurement for the freezing point of diesel fuel products, a soft sensing method by near-infrared (NIR) spectrometry was proposed. Firstly, the information of diesel fuel samples in the spectral region of 750-1 550 nm was extracted by spectrum analyzer, and the polynomial convolution algorithm was also applied in spectrogram smoothness, baseline correction and standardization. Principal component analysis (PCA) was then used to extract the features of NIR spectrum data sets, which not only reduced the number of input dimension, but increased their sensitivity to output. Finally the soft sensing model for freezing point was built using SVR algorithm. One hundred fifty diesel fuel samples were used as experimental materials, 100 of which were used as training (calibrating) samples and the others as testing samples. Four hundred and one dimensional original NIR absorption spectrum data sets, through PCA, were reduced to 6 dimensions. To investigate the measuring effect, the freezing points of the testing samples were estimated by four different soft sensing models, BP, SVR, PCA-BP and PCA+SVR. Experimental results show that (1) the soft sensing models using PCA to extract features are generally better than those used directly in spectrum wavelength domain; (2) SVR based model outperforms its main competitors-BP model in the limited training data, the error of which is only half of the latter; (3) The MSE between the estimated values by the presented method and the standard chemical values of freezing point by condensing method are less than 4.2. The research suggests that the proposed method can be used in fast measurement of the freezing point of diesel fuel products by NIRS.