Aiming at the problem that the convolutional neural network （CNN） is insufficiently mined on the information of vibration data structure, which leads to the low accuracy of fault diagnosis, a CNN-GraphSAGE dual-branch feature fusion method for gearbox fault diagnosis is proposed. Firstly, the vibration data of the gearbox is subjected to wavelet packet decomposition, and the wavelet packet coefficients are constructed into graph structured data containing nodes and edges. Then a dualbranch feature extraction network is established, with the CNN branch using a dilated convolutional network to extract global features of the data, and the GraphSAGE branch using a multi-layer feature fusion strategy to mine the implicit correlation information in the data structure. Finally, an attention fusion module based on the SKNet attention mechanism is constructed to fuse the dual-branch extracted features, and then the fused features are input into the fully connected layer to realize the fault diagnosis of gearbox. In order to verify the excellent performance of the proposed method in gearbox fault diagnosis, the ablation experiments were conducted first, and then comparative experiments were carried out under the condition of no added noise and adding 1 dB noise. The experimental results show that even under the condition of 1 dB noise, the average diagnostic accuracy of the proposed method is 92.07%, which is higher than the comparison models. The proposed method can effectively recognize various types of faults in gearboxes.