Aiming at the characteristic that the fault signal of the rotating parts in the gear box is a periodic impact signal, a composite fault feature extraction method of gear box based on multipoint kurtosis(Mkurt) and multipoint optimal minimum entropy deconvolution adjusted(MOMEDA) is proposed, namely MKurt-MOMEDA. The use of MKurt can effectively extract the cycle of impact vibration signal submerged by the noise in gear box to achieve the vibration signal vibration source tracking. Then, according to the period of the fault, a reasonable interval is set, and the original signal is denoised by MOMEDA to further extract the periodic impulse of the original signal.Through the analysis and verification of the simulation signal and the measured data, it is proved that the MKurt-MOMEDA method can accurately and effectively diagnose the compound fault characteristics of the gearbox.