In view of the deficiency of traditional bispectrum that has good insensitivity to independent Gaussian noise but is incapable of avoiding the disturbance of non-Gaussian noise, a new feature extraction method based on EMD and bispectrum analysis is proposed. Firstly, decompose the signal by EMD. Then, energy-related method is applied to removal the illusive intrinsic mode functions (IMFs) of EMD. Finally, bispectrum analysis is applied to the real IMFs to extract the fault feature. The results of experiment and practical signals analysis show that the proposed method is feasible and effective for extracting feature of rolling bearing, and it is more effective compared with traditional power spectral analysis and traditional bispectrum analysis.