Abstract: Aiming at the weak impact signal produced by rubbing fault of steam turbines, the location and duraction of the signal singularity were determined by conducting filtering process to the measured signal through wavelet transformation, extracting the transient impact component on a specific scale, carrying out Hilbert demodulation analysis to the impact component, enveloping the singularity of the reflection singal under the action of rubbing fault with the amplitude of transient impact singal, and conducting wavelet singularity detection to the amplitude-enveloped singal. An analysis was made to a nuclear power turbine unit using actual measurements of vibration signal. Results show that the feature extraction of slight rubbing fault can be effectively improved by detecting its abrupt information using wavelet singularity analysis.