In order to improve the accurate extraction rate of the traditional complete ensemble empirical mode decomposition with Adaptive Noise ( CEEMDAN) for motor bearing fault characteristic signals and reduce the distortion of the reconstructed signal, an improved CEEMDAN algorithm is proposed. The original signal is initially decomposed using traditional CEEMDAN to obtain several feature components (IMFs) and intrinsic modal components. Some IMFs are de-noised and extracted by entropy weight method. The filtered IMF components are secondary decomposed and secondary screened to obtain typical fault sensitive signals. Then the signal reconstruction is carried out by using SG ( Savitzky-Golay) smoothing filter and the motor bearing signal is de-noised. Finally, the performance of the improved algorithm is analyzed by using the data of Case Western Reserve University. The results show that the method can effectively reduce the signal noise of the motor bearing signal, and its SNR is improved by 2. 2 dB compared with the original signal.