As initial fault occurs in rolling bearing of low-speed and heavy-duty machinery, the impulse component, reflecting the fault feature in vibration signal, is difficult to extract for it is relatively weak and easily corrupted by strong background noise. The authors attempt to extract the impulse component from vibration signal with sparse representation method. However, it is difficult to construct the accurate dictionary which matches the impulse component since operating conditions of bearing is not stable. Hence, a method of extracting the initial fault feature, which is based on dictionary learning, is proposed here. Firstly, an adaptive dictionary is obtained by the developed K-SVD dictionary learning algorithm. Then, Orthogonal Matching Pursuit (OMP) algorithm is utilized for sparse decomposition of the vibration signal, and all kurtosis values of approximation signal of iterations are calculated .Finally, the corresponding approximation signal of maximal kurtosis value will be reconstructed and analyzed with envelope spectrum to diagnose the fault type. The test results of simulate data and bearing vibration signal demonstrate that the proposed method, which can extract the feature component more accurately than other methods, meets the demand of real-time bearing condition monitor.