An method of multilevel Fractional Fourier Transform filtering with self-adaptive segmentation based on rotate speed signal is proposed and applied to extract fault features of gearbox under transient-conditions. Firstly, segmenting the time-frequency curve of gearbox in Target-gear adaptively according to the threshold curvature which was set in advance, so the frequency change in each segment presents linear approximation. Then, obtaining the best order of FRFT for each segment by least square fitting of the curve and analyzing the vibration signal by FRFT to filter the interference components and noise. The experimental results show that the best order of multilevel FRFT calculated by self-adaptive segmentation of rotate speed signal is accurate, quick and adaptive; The target order component under transient–conditions can be extracted by multilevel FRFT with best order effectively and isolates other interference; Fault features can be obtained by envelope demodulation of the extracted sole order component.