Aiming at the impact feature in fault signals of the rolling bearings,The improved algorithm diagnosis method EEMD-RA-KU was proposed to capture the impact feature. Firstly,Fault signals were decomposed into a series of intrinsic mode functions（ IMFs） with ensemble empirical mode decomposition（ EEMD）. Secondly,The IMFs were selected to reconstruct signals based on relative analysis（ RA） and kurtosis（ KU）. Thirdly,The reconstructed signals were filtered by the filter that the parameters were determined with the spectral kurtosis. Finally,The fault can be diagnosed by comparing the frequency of the envelope spectrum with the fault characteristic frequency of the bearing. The result shows that the fault feature information of the signal can be extracted accurately when there was a lot of background noise.