Aimed at the fault of roller bearing, the method of fault diagnosis of roller bearings based on correlative principles of optimum threshold plus wavelet noise elimination and Least Squares Support Vector Machine (LS-SVM) in combination was presented. Correlative principles of optimum threshold plus wavelet transformation was applied to extract the feature of fault of roller bearings at early stage, energy feature method was used for signal extraction. And then, the recognition of fault types were carried out by using LS-SVM classified algorithm. Testing and simulation results show that the fault diagnosis method based on correlative principles of optimum threshold plus wavelet transformation and LS-SVM in combination can diagnose typical faults of the roller bearing effectively.