Abstract: The gearbox is an important component in the transmission system, and its failure rate is high and it is difficult to directly identify the fault condition. Aiming at the problem that the gearbox fault vibration signal often contains a large amount of noise and it is difficult to extract accurate and comprehensive fault characteristics, a gearbox fault diagnosis method based on adaptive wavelet noise reduction and Inception network is proposed. Firstly, the acquired vibration signal is adaptive wavelet noise reduction, and then the noise reduction signal is fed into the Inception network for fault feature extraction and classification. The Inception module has multi-scale abstract feature extraction performance, which can extract comprehensive fault feature information from the signal, including the weak fault signal of the gearbox. The results show that the fault identification accuracy of this method still reaches 92.65% in the environment where the signal-to-noise ratio SNR is −4 dB, and the fault identification accuracy of the signal processed by noise reduction processing is 6% higher than that of the signal directly entered into the Inception network for fault identification. Therefore, using the method proposed in this study, real-time monitoring of the gearbox and timely detection of safety hazards are of great significance to ensure the stable operation of the gearbox and prevent property damage.