Impact feature extraction is the key to realize bearing and gear failures identification, but it is easily disturbed by noise and other unrelated components. Different from classical timefrequency analysis methods such as shorttime Fourier transform and synchrosqueezing transform, a short time Fourier transform method based on frequencydomain window function was proposed. The vibration signal was filtered by using maximum correlation kurticity deconvolution to improve the signal quality; through the window function in frequency domain, the time precise location and the impact characteristics accurate identification of in the twodimensional timefrequency plane were realized, and the timefrequency ridge of complex multicomponent signals was further sharpened. The numerical simulation signals and actual bearing fault signals were analyzed by using the proposed method, and the effectiveness of the proposed method was verified.