In order to extract the fault feature frequency of weak signal, an impr ove HilbertHuang transform (HHT) method combining with the wavelet packet tran s form is put forward in this study. After performing wavelet packet denoising, th e empirical mode decomposition (EMD) was used to decompose the signal into a num ber of intrinsic mode functions (IMFs). This process always generated some undes irable IMFs at lowfrequency regions. The method for removing the undesirable I M Fs was proposed by using the correlation coefficient of the IMF and the undecomp osed signal as a criterion, and the IMFs whose correlation coefficient was bigge r than the set value were chosen as the effective data for the future Hilbert ma rginal spectrum analysis. The improved HHT can not only avoid the impact of unde sirable IMFs but also save memory space and operation time. Experimental results show that the proposed method is effective in fault feature extraction.