Considering the nonlinear and non-stationary characteristics of rolling bearing vibration signals， a method is put forward based on improved ensemble empirical mode decomposition （IEEMD） and modulation signal bispectrum （MSB）. The vibration signals are decomposed into intrinsic mode functions （IMFs） at different frequencies by EEMD. The mean of the standardized accumulated modes （MSAM） is taken as a novel criterion to divide IMFs into low-frequency and high-frequency IMFs. Subsequently， the wavelet threshold denoising algorithm is applied to the high-frequency IMFs， which is then combined with the low-frequency IMFs to generate the reconstructed signal. The MSB is used to extract modulation features by further suppressing residual random noise and deterministic interference components. The analysis results demonstrate that the method has high accuracy in fault feature extraction by comparing with Spectral kurtosis （SK） and WEEMD-MSB.