Rolling bearing is the easy wearing part. In order to better and timely detect early and slight bearing fault vibration signal under low signalnoise ratio, the method of wavelet packet optimal entropy combined with EEMD was proposed. Signalnoise separation was achieved by wavelet packet optimal entropy, which highlighted the wavelet packet noise reduction effect was obvious, through EEMD decomposed the signal into a plurality of component. Finally, crosscorrelation, kurtosis criterion were used to extract fault signal component in order to avoid the blindness of component selection. The results show that: the method of bearing early fault has good noise reduction effect, which can accurately and quickly detect bearing fault, thereby to show that the method is feasible and effective.