Gearbox is widely used in petrochemical, shipbuilding, electric power and construction machinery, etc., and its fault monitoring and diagnosis technology is always a hot issue. A gearbox fault diagnosis method based on DS evidence theory, which integrates oil and vibration with multiple parameters is proposed. Aiming at the problem that the traditional single vibration signal analysis method is not sensitive to gear wear in the early stage, radial basis function（RBF） neural network is used to establish the classification model of the fusion of vibration time domain features and oil features. Vibration frequency-domain features are extracted adaptively based on high-dimensional variational autoencoder and fault classification is completed. Different evidence bodies obtained by RBF neural network and high dimensional variational autoencoder are fused by weighted DS evidence theory to obtain the final diagnosis result. Compared with different fusion methods, the effectiveness of the proposed method in gear fault diagnosis is verified.