Abstract：Vibration signals produced from rolling element bearings in use are both periodic and random. Periodicity comes from almost periodic shock vibrations in nature for its inherent periodic mode of operation, and randomness comes from uncertainty factors such as slip of the balls, manufacturing errors, and so on. The fault model of rolling element bearing is theoretically depicted in cyclostationary model better than in periodic model due to the above reasons. A method of fault feature extraction of rolling element bearing based on cyclic autocorrelation was presented. It proved that cyclic frequency was capable of reflecting the feature frequency of the faulty rolling element bearing by theoretical analysis, computation simulations and experiments. What's more, the cyclic autocorrelation function method could suppress the noise better than the conventional envelope spectrum method could do when extracting fault features of the rolling element bearings. The proposed method was of great significance for the fault fine diagnosis of rolling element bearings.