In critical working conditions, the feature extraction of the composite faults in the gearbox is difficult to be realized. Generally, it is easy for the fault characteristics to escape diagnosis or be misdiagnosed by the improper selection of the method. Due to improper selection of white noise, there will be energy leakage situation when the signal is decomposed by ensemble empirical mode decomposition (EEMD). By calculating the multi-point kurtosis (MKurt) can extract the impact fault cycles, but the tracking effect is not ideal in the strong noise environment. Considering that the accuracy of using multipoint optimal minimum entropy deconvolution adjusted (MOMEDA) to extract fault components is affected by the range of the fault cycle, a self adaptive fault feature extraction method based on CMF-MOMEDA is proposed in this paper. Firstly, the EEMD is used to decompose the signal into a series of intrinsic mode functions (IMFs) at high and low frequencies. Secondly, the original signal is decomposed into high and low frequency bands Ch and CL by the combined use of IMFs with strong a correlation with the original signal and the combined mode function (CMF). Finally, the MKurt spectrums of Ch and CL are obtained to extract the fault cycle components, then the appropriate cycle ranges are set and the fault characteristics are extracted by MOMEDA. The feasibility of this method is verified by the application of the fault feature extraction of gearbox.