Considering the strong background noise of rolling element bearing fault signal, a rolling element bearing fault feature extraction method is proposed based on improved EMD(empirical mode decomposition) and spectral kurtosis method. Original fault signal is decomposed by EMD to contain a finite number of stationary intrinsic mode functions (IMFs). Then grey incidence and mutual information are used together to remove pseudo-components in the traditional EMD results. Finally, real IMF component is analyzed by the spectral kurtosis and the envelope demodulation to extract the fault feature frequency. Experimental results show that the method can effectively extract the fault feature of rolling element bearing, and is more effective than the envelope demodulation in fault feature extraction.