Crack fault of gears not only affects the overall performance of mechanical system, but also causes machine damage. Therefore, the fault diagnosis method of crack length of gears is researched. The multi-sensor vibration signal was taken as the research object, and the energy ratio coefficients of each frequency band of wavelet packet were taken as the fault features of the gear crack, and the fault diagnosis of the gear crack length was realized by classifying features of improved neural network model. The results show that the fault diagnosis method proposed has high recognition rate (97.5%) and good versatility, which can effectively identify gear faults under different working conditions.