The early weak fault of rolling bearing is annihilated in the high strength background noise, which makes the fault feature information extraction difficult. An improved resonance sparse decomposition method was proposed. Variational mode decomposition was used to denoise the collected signal. According to the kurtosis-correlation coefficient criterion, the component containing the large amount of fault feature information was selected for signal reconstruction. Then the resonance sparse decomposition of particle swarm optimization was performed on the reconstructed signal. Finally, the envelope analysis of the lowresonance components obtained was used to extract the fault feature frequency.Experimental results show that this method can effectively extract the fault feature frequency and reduce the interference component compared with traditional resonance sparse decomposition.