According to the non-stationary signal of rolling bearing, a feature extraction method of extracting energy ratio of intrinsic mode function(IMF) of normal signal and fault signal to construct characteristic vectors is proposed. Because of the modal aliasing by EMD, the decomposition results are uncertain, and the fault recognition rate of traditional energy feature extraction is low. By introducing normal signal under the same working conditions, the energy characteristics of various mode vectors are changed into energy features relative to normal signal. The experimental results show that the proposed method can effectively extract the fault feature of roller bearing and achieve its fault diagnosis.