Abstract: The classical failure criteria for prediction of rock strength can not be accurate. This study uses the diagonal intersection least squares SVM(DILSSVM)as new method for prediction of the rock strength in a wide loading such as uniaxial and triaxial loading. For each rock type, data obtained from field experiments and laboratory experiments are divided into training and test sets. DILSSVM is employed to train with the compressive stress (c) and minor principal stress (3) and to predict the value of major principal stress (1f) at failure. The training sets are used in regression analysis for m in Hoek-Brown (H-B) equation. Then, the test sets are used to examine the accuracy of target rock strength with DILSSVM after training and these of the two H-B model(with different m). Comparison of the results of the DILSSVM with the two H-B models shows that the DILSSVM always has less root mean squared error(decreased 45% ～55%) and higher coefficient of determination(enhanced 0.055～0.085,near to 1). The DILSSVM shows better flexibility in 1f at failure in each rock type and a wide loading range.