Abstract: Aiming at the problem that the impact of early mechanical failure is weak and it is difficult to extract due to strong background signal and noise interference，a weak impact feature extraction method combining singular value decomposition（SVD）differential spectrum and S?transform is proposed. The original signal is constructed into a Hankel matrix，and the reconstruction matrix is decomposed by SVD. The noise reduction order is determined by singular value difference spectrum for noise reduction. The S-transformation is used to analyze the time-frequency of the denoised signal，the weak impact characteristic information of the signal is extracted. The comparison between numerical simulation and actual bearing fault data shows that the method can effectively distinguish the early weak shock characteristics caused by faults in the bearing vibration signal. It can provide a priori information for the bearing fault diagnosis.