In the light of the difficulty for the cepstrum method to draw early weak fault, a new feature extraction method based on minimum entropy deconvolution (MED) and envelope cepstrum is proposed. First, the fault signal is de-noised and the meantime periodic impact components are enhanced by MED method to get the envelope signal. Then, the envelope cepstrum is presented by the cepstrum analysis of the envelope signal. Finally, the fault features are extracted through the envelope cepstrum. The results of the experiments show that the proposed method extract the feature frequency information of incipient fault with higher efficiency than the traditional methods do and can be used to prevent major faults from occurring．