Aimed at gearbox incipient fault features are very weak and these features are difficult to be distinguished, a diagnosis model based on DSmT theory and wavelet neural network for gearbox incipient fault is proposed. Firstly, multiple vibration sensors are reasonably arranged on gearbox critical position to collect multi-source vibration information for feature extraction. Several shunt-wound wavelet networks are used to carry on primary fault diagnosis and acquire independent evidences each other. Then, DSmT theory is used to combine different independent evidences and got the final decision result. The DSmT theory overcome the shortcoming of the traditional DST theory, the wavelet neural network realized the objectivity of multi-source evidence belief assignment. The diagnostic tests show that this method can effectively improve the identification accuracy of gearbox incipient fault features and reduce diagnostic uncertainty.