As a traditional frequency domain modal decomposition method, the empirical wavelet transform has complete theoretical support and excellent adaptability. For signals with high sampling frequency and strong noise, the empirical wavelet transform will obtain unexplained components and mode aliasing. To solve the above problems, this paper proposes an optimized empirical wavelet transform method based on Cepstrum guidance. The trend spectrum is obtained through cepstral editing, which is considered to be related to the information. The minimum of the trend spectrum is defined as the boundary between modes. Filter banks are constructed based on the scaling function and empirical wavelet to extract different modal components. Finally, the unbiased autocorrelation kurtosis of each component is calculated to screen fault information. The method is verified to be effective through simulation signals and bearing experimental signals.