In this paper, a new method based on wavelet analysis for feature extraction of gearbox vibration signals is explored. The similarity of the power spectrums between gearbox vibration signals and 1/f processes signals makes natural the use of wavelet-based fractal analysis for gearbox fault diagnosis. Then the principle of this method was discussed. To verify the feasibility and practicability of this presented method, experiments based on an automobile transmission gearbox were carried out. Vibration signals of different working stages were gathered by an acceleration sensor attached on the case body surface. After pretreatments, these vibration signals were decomposed to ten detailed signals at different wavelet scales by using the discrete wavelet transform with Daubechies wavelet. Then the variances of detailed coefficients at scales 3 to 7 were calculated and the fractal features of all acceleration signals were estimated from the slope of the detailed coefficients variance progression. The results of temporal window trials demonstrate these fractal features are significantly different for the different working stages of the gearbox and show high reproducibility, which suggests that the fractal features extracted by the method presented in this paper is convictive and the wavelet-based fractal analysis is effective for classifying the vibration signals of gearboxes.