Abstract: A signal processing method was proposed for fault diagnosis of rotary mechanisms based on empirical wavelet transform (EWT). The specific way is to divide the Fourier spectrum of relevant signals, and then to filter the divided spectrum with a set of newly built wavelet filter bank to get a group of single component ingredients, and finally to transform every single component ingredient by Hilbert method so as to obtain the instantaneous frequency and amplitude. To verify the effectiveness of the method, a comparison was made to the performance of EWT and empirical mode decomposition method in diagnosing fault signals coming from both simulation and typical experimental rotors. Results show that the EWT method can accurately analyze the mechanical fault signals and the failure eigenvalues are evident, which may be applied in fault diagnosis of rotary mechanisms.