In the traditional transformer winding mechanical fault diagnosis method, only the winding axial vibration is considered, and the feature parameter extraction is complex and the recognition accuracy is low. This paper presents a mechanical fault diagnosis method for transformer windings based on two-axis vibration and multi-sensor fusion. Firstly, the two-axis vibration relationship graph is proposed as the feature image from the perspective of the axial and radial vibration correlation of the winding. Then the lightweight convolutional neural network MobileNet V2 is used to train the image data obtained by different sensors. Finally, the D-S evidence theory is used to fuse the multi-dimensional information source recognition results and make the final decision. The experimental results show that the fault diagnosis accuracy of the proposed method can reach 99.4%. Compared with the traditional fault classification method, the feature extraction step is simplified, and the diagnostic accuracy is improved by more than 6.2%, which provides a feasible scheme for mechanical fault diagnosis of transformer winding.