Gear′s vibration signal in varying conditions have nonlinear and non-stationary characteristics, in view of this, a method combined Vold-kalman order tracking(VKF-OT) and detrended fluctuation analysis(DFA) for fault feature extraction is proposed. The method takes rotating and meshing frequencies of the gears as the extraction frequencies of VKF-OT. Two kinds of order signal under any time-varying conditions are obtained and the frequency modulation interferences caused by speed variation are weakened. The accurate amplitude and phase are obtained by solving complex envelop for keeping the transient information of the gear status. On this basis, DFA method is introduced to deal with the original and order signals of rotating and meshing frequencies, to eliminate the jamming of amplitude modulation caused by load variation. Compared with the double logarithmic function graph from three kinds of signals, the eigenvector of gear vibration signal is selected. The experiments for 150 sets gear vibration signals under different wording conditions are conducted. The results show that the fault feature extracted by this new method can effectively distinguish the early local weak fault of gear under time-varying conditions.