Abstract: In order to extract useful features from vibration signals, a fault feature separation and extraction method using the new time-frequency decomposition method, the frequency slice wavelet transform（FSWT）, was proposed. The vibration signal firstly was processed with the FSWT to get its time-frequency distribution. Then zoom analyses were adopted with higher resolutions to reveal the submerged time-frequency features for the selected time-frequency regions by investigation of the time-frequency distribution characteristics.On the basis of the interested regions, the time-frequency features were extracted with making segmentation on the distribution map. Thereafter, after filtering, the effective fault features were separated from the signal through the FSWT reconstruction. The proposed method is shown to be efficient by simulations and engineering applications. It has the ability to isolate the desired components from noisy signals. It achieves an ideal effect on feature extraction for a generator unit fault diagnosis.