In order to improve the accuracy of fault feature extraction, the stochastic resonance (SR) method is proposed in the pretreatment of vibration signals, then the fault feature is extracted based on the method. First, the de-noising principle of SR is presented, and the mutable scale SR, which is suitable for large parameter signal， is analyzed. Then a fast optimization method of frequency compression ratio R is put forward. The vibration fault feature sets based on time domain, frequency domain, time-frequency domain are extracted respectively to test the proposed feature extraction method. Finally, the discrete degree index based on between-class and within-class is applied to analyze the classification performance of feature set. The analysis result shows that the classification indexes of the feature sets extracted from SR output signal are obviously superior to those from the original signal, the feature extraction accuracy is improved notably.