Abstract: Aiming at the problem that gear fault signal is non-linear and always accompanied by noise, a fault feature extraction mathod for gear based on variational mode decomposition (VMD) based auto regressive (AR) mode and correaltion dimension was proposed. The gear vibration signal was decomposed by VMD and a number of intrinsic mode functions (IMFs) were obtained. The sensitive IMFs were selected by calculating the frequency domain cross correlation coefficient of each IMF component. The reconstructed signal based on the sensitive IMFs, which were selected by calculating the frequency domain cross correlation coefficient of each IMF component. The AR model of the reconstructed signal were constructed. Finally, the correaltion dimensions of auto-regressive parameters in AR model were calculated and as feature vector to identify the working state and fault type of gear. Through the analysis of the measured vibration signal of gear, the effectiveness of the proposed method was verified.