AR predict filer based Spectral kurtosis method is proposed to research the early fault diagnosis of rolling bearings. Combining the characteristic of extracting the resonance damping component of bearing fault signal which an Auto regressive (AR) predict filer holds, the early weak fault of rolling bearings are successfully detected after that the filtered signal is processed by spectrum kurtosis method. An accelerated life test of rolling bearing is established to obtain the nature fault signal of rolling bearings which overcomes the insufficient of traditional artificial fault process. The result of experiment data analys is shows that the proposed method can not only extract the fault characteristic by eliminating the interference component, but also can rise the stability of spectral kurtosis analysis method.