The vibration signal of the mechanical rotor was decomposed by the EMD method, and the end effect produced was restrained through mirror extension. Then，several intrinsic mode functions which reflected the inherent variation characteristics and regulations of rotor fault signal were obtained. The AR models of each component were established , and the auto-regressive parameters the variance of remnant of the AR models were regarded as the fault feature vectors，and the SVM classifier was built to recognize the fault pattern. The results show that the rotor fault diagnosis based on mirror extension of EMD-SVM could obtain the accurate characteristics and status of the rotor fault system.