To extract effectively features of different damage levels on rolling bearings’ inner race, a method of quantitative fault feature extraction based on the combination of Volterra kernel function theory and bi-spectral analysis was proposed. Firstly, input signals and output ones of a system were used to determine a Volterra model. Secondly, Volterra kernel function of the model was solved with the improved multi-pulse excitation method. The model was identified using the generalized frequency response function (GFRF). Finally, using the means of bi-spectrum and its slices, the information of damage level features implied in the second order kernel function due to phase coupling was separated, quantized and extracted. A rolling bearing test table was used to collect faulty bearings’ data to verify the proposed analysis method. The results were compared with those using the envelope spectral analysis method. The results showed that the bi-spectral slice method can be used to intuitively and quantitatively express the information implied in Volterra second order kernel function when there are not obvious shock vibration, and effectively distinguish normal bearings and faulty bearings with different inner race damage levels.