A prediction model of surface roughness based on the Least Square Support Vector Machine(LS-SVM) in cylindrical longitudinal grinding is proposed. By converting the inequality constraints into equality constraints, the model transformes solving the SVM from a Quadratic Programming (QP) problem to a group of linear equations, which simplifies the learning process and improves the calculating efficiency. Experimental results indicate that the construction speed of the prediction model based on LS-SVM is more faster , and the measurement error(MSE) is less 4% and 1.3% than those of the BP neutrol algorithm and BP+GA algorithm,respectively.The method has been used in a intelligent system for cylindrical longitudinal grinding to predict the surface roughness of a workpiece in real time. By calculating the differences of predicating values and giving values and by directing the correct of grinding parameters,it completes a closed loop and intelligent control and obtains good grinding results.