Aiming at the problem that at early stage of bearing fault, the fault feature is very weak, feature components are easy to be submerged in background noise and can not be detected in time. This paper presents a novel method based on combination of harmonic decomposition and self-complementary Top-Hat transformation to enhance detection of bearing weak fault. Firstly, with the improved generalized harmonic wavelet function, the sub-band number and bandwidth range are not subject to the limitation of the binary decomposition, and the most optimal frequency band signal which contains relatively concentrated fault feature is found by kurtosis diagram of spectral kurtosis. Secondly, in order to depress strong background noise and enhance the weak fault feature of bearing, the optimum frequency band signal is handled by multi-scale self-complementary Top-Hat transformation, and the most optimal structure element (SE) scale is selected by using a novel method named fault feature energy radio (FFER). Finally, weak fault feature of bearing is detected by envelop demodulation analysis. The proposed method is applied to simulated signal and bearing full lifetime vibration datasets, the results show that this method can effectively detect weak fault of bearing, and is valuable for the engineering application.