Aiming at the problem that there is inseparable region and more than one traditional Support Vector Machine(SVM) classifiers need to be trained in the multi-class classification problem, the support vector regression machine based multi-class classification method is researched. This method utilizes regression theory to solve multi-class classification problems, in which the classification samples are served as regression input, and their class labels are served as regression output, then the relationship between samples and their class labels are fitted by support vector regression machine method. The samples are classified into the regression function, and the class labels are obtained by adding a rounding operation to the regression output. This method uses only one classifier, which significantly simplifies the classification process. In addition, the composite kernel function is introduced to improve the performance of the support vector regression machine. The datasets of multi-class classification problems selected from University of California Irvine(UCI) database are used for simulation. Compared with traditional multi-class support vector machine, both classification speed and accuracy of the proposed method have been significantly improved.