It is important to extract fault features in fault diagnosis of machinery. The traditional signal analysis can not settle for non-stationary vibration signal whose statistic properties are variant. To deal with non-stationary signal, time-frequency analysis techniques are widely used. The experiment data of typical vibration fault signal ware analyzed by different methods, such as short time Fourier transform (STFT), Wigner-Ville distribution (WVD), Wavelet transform (WT) and Hilbert-Huang Transform (HHT). Compared with these methods, it is demonstrated that the time-frequency resolutions of STFT and WVD were inconsistent, which were easy to cross or make signal lower. WT had distinct time-frequency distribution, but it brought redundant component. HHT time-frequency analysis can detect components of low energy, and displayed true and distinct time-frequency distribution. Therefore,it is a very effective tool to diagnose the early faults of rotating machinery, and it provides new means for state detection and fault diagnosis.