The time-frequency coupling relation between useful non-stationary components and noises bring great difficulties to the realization of de-noising for non-stationary signals，which can not be solved by classic de-noising method in time or frequency domain. The principles of short time Fourier transform(STFT)，Wigner-Ville transform，Chirplet adaptive decomposition are analyzed，and then a novel de-noising method for non-stationary based on joint time-frequency distribution is proposed. In this method，the analyzed signal WVD is seen as the combination of auto-WVD and cross-term WVD. Firstly，STFT energy spectrum of the analyzed signal is used as template to cross-correlate with its corresponding WVD in order to obtain the satisfactory time-frequency distribution with high time-frequency resolution and without cross-term interferences. Secondly，the useful components are decomposed as Chirplet function using the two-dimension least square fitting method，and then are extracted out to reconstruct for noise suppression. Finally，the computer simulation results verify the effectiveness of this proposed method. Its application in gearbox fault diagnosis indicates that with the method the extracted cycle of the gearbox vibration impulses has a good consistency with the corresponding fault frequency.