A new method based on simulated annealing algorithm (SA) and least-squares support vector machine (LS-SVM) (SA-LS-SVM) was proposed to select the characteristic wavelength for visible-near infrared (Vis/NIR) spectroscopy discrimination. In order to find suitable numbers of characteristic wavelength and corresponding characteristic wavelength, discriminating rate was used as object function for SA, and LS-SVM was adopted as discrimination model. The Vis/NIR spectroscopy characteristic wavelengths of three categories of lubricant were processed by SA-LS-SVM, principal component analysis (PCA) and partial least squares (PLS) respectively, and then predicted by back-propagation artificial neural network (BP-ANN). The results of experiment showed that discriminating rate by using combination of SA-LS-SVM with BP-ANN reaches 100% only using 4 characteristic wavelengths from total of 751 wavelengths, while discriminating rate did not reach 100% by other methods. The proposed algorithm not only reduced the number of spectral variables, but also improved the discriminating rate.