The rolling bearing fault vibration signals are nonlinear and non- stationary and have strong noise background,in order to extract the fault feature effectively,a feature extraction technology which combines wavelet denoising and resonance- based sparse signal decomposition（ RSSD） is proposed. The resonance sparse decomposition is a new method for frequency division based on the tunable quality factor wavelet transform and morphological component analysis,it is different from conventional signal decomposition method based on frequency band partition,it is based on different oscillation forms of the signal components. Firstly,the signal is processed by wavelet threshold denoising,then,the signal is decomposed into two parts with different resonance characteristics by the resonance- based sparse signal decomposition method,The one is high- resonance component which has sustained oscillation characteristics,and the other is low- resonance component which has instantaneous impact characteristics. Finally,the impact fault feature is extracted from the low- resonance component by Hilbert envelope demodulation method. This method is applied to simulation signal and failure examples of impact on bearing test bench,the effectiveness of the proposed method is verified.