In order to solve the disadvantage that it needs to set an experiential value of existing stop criterion of Intrinsic Mode Functionin EMD (Empirical Mode Decomposition) sifting process, the paper proposes a new criterion of Intrinsic Mode Function based on Shannon entropy. There is no need to set an experiential value of a parameter, which avoids the defects of differences in decomposition results caused by different experience value in sifting process. By comparing with other criterion, it demonstrates that the decomposition result are more accurate and have smaller index of orthogonality according to the proposed stop criterion, the IMFs can reflect the characteristic of non-stationary and nonlinear in signal. This stop criterion is applied to the fault diagnosis of rolling bearing in wind turbine. The result showes that this criterion can better retain the fault feature information such as the shock pulse, amplitude and frequency modulation, and diagnose the fault site accurately.