Based on the fact that the fault roller bearing signal is affected easily by Gauss noise, the fault diagnosis method based on the high-order statistics is proposed. This method uses higher-order spectra to reconstruct power spectra and extracts fault feature information with the reconstructed power spectra. A model is established using higher-order spectra to reconstruct power spectra. Meanwhile, analysis is conducted using the man-made data and recorded MT data. The results show that the presented method is superior to the traditional power spectral method in suppressing Gaussian noise and can extract more useful information. Compared with the traditional method, the analysis results from roller bearing signals with inner-race, out-race or bearing ball faults show that the diagnosis approach could extract fault characteristics effectively and its resolution is higher.