Vibration signals of rolling bearing contain deterministic components and random components, both of them reflect the failure information of bearing. For qualitative diagnosis of bearing faults using random components we need less vibration signal data that increases the computational efficiency for cyclostational analysis. In this paper we propose to use logarithmic contour maps of spectral correlation density firstly to reveal the change of weak random components caused by bearing faults, and then to take the slice at a resonance frequency to extract the fault information. An analysis example with real bearing vibration signals show that even in condition of lower frequency resolution, spectral correlation density can realize the demodulation of fault information, achieve the purpose of fault feature extraction, and improve the calculation efficiency.