A complicated signal can be decomposed into a series of functions (IMFs) by ensemble empirical mode decomposition (EEMD). Since some IMFs are closely related to the faults while others are irrelevant, the method for selecting sensitive IMFs remains a problem. In light of this problem, a novel feature extraction method of IMFs based on fast kurtogram is proposed. Firstly, the fast computation of the kurtogram is used to calculate the kurtosis distribution of the original signal and the kurtosis distribution of each IMF in the frequency domain. Then, the reference frequency bands of the above-mentioned signals are determined by the maximum kurtosis band. By comparing the relationship between the reference frequency bands of the original signal and each IMF, the sensitive IMFs can be found. Finally, accurate information about the fault will be obtained for subsequent fault diagnosis processing. Both the simulated and real signals verify the effectiveness of the proposed method.