A modified fast empirical wavelet transform (MFEWT) based on compromise threshold function was proposed in order to solve the problem of improper segmentation caused by soft threshold function in fast empirical wavelet transform (FEWT). For this method, the trend spectrum is firstly calculated by Fourier transform and inverse Fourier transform and the result of calculation is optimized by wavelet denoising with compromise threshold function. Then, filter bands are built with optimized trend spectrum and the reconstruction of EWT empirical modes are made according to filter bands. With the fusion of kurtosis and Pearson correlation coefficient, characteristic components are selected. With minimum entropy deconvolution (MED), characteristic frequency of signal reconstructed by characteristic components can be calculated. Fault diagnosis of rolling bearing is finished with the comparison between characteristic frequency in experiment and theory at last. Results of experiment demonstrated that MFEWT performed better than FEWT in signal decomposition. For MFEWT, peaks of characteristic frequency in envelope spectra are clearer. The MFEWT improves the performance of signal decomposition of EWT and the reliability of rolling bearing fault diagnosis.