Abstract: Aiming at the problem that a large amount of noise in early fault signals of rolling bearings makes it difficult to extract fault features，a new and improved wavelet threshold-based noise reduction method is proposed. This method uses the complementary set empirical mode decomposition（CEEMD）method to decompose the original fault signal to obtain the intrinsic mode function（IMF）components of each order. The key IMF components are selected to reconstruct the signal，and the reconstruction signal is filtered through the new improved wavelet threshold algorithm and fast spectral kurtosis to reduce noise. The Hilbert envelope demodulation is performed to obtain the characteristic frequency of the rolling bearing fault. The method is verified by the simulated noise signal and the experimental signal of the rolling bearing，and the new improved wavelet threshold algorithm is compared and analyzed with the traditional wavelet hard threshold and wavelet soft threshold algorithm. The results show that the method can effectively improve the reliability of the fault signal. Signal noise ratio and noise reduction effect are obvious，and the fault characteristic frequency of the rolling bearing can be effectively obtained.