To solve the problem that gearbox fault diagnosis requires a lot of expert experience and it is difficult to extract features manually,a fault diagnosis method based on feature difference learning convolutional neural network (FDLCNN) was proposed.Multi-scale networks with different depths were constructed and residual modules were introduced to improve the feature extraction capability of the network.Fault features of different scales and depths were extracted from 1D time-series signals,and then feature fusion was carried out through adaptive average pooling layer processing to enrich intelligent diagnosis and decision information.Finally,feature dimension reduction was implemented in the full connection layer,and the diagnostic results were output by the Softmax classifier.The experimental data of ten kinds of gearbox fault states were compared with the existing three methods.The results show that FDLCNN has higher fault identification accuracy,stronger robustness and faster convergence speed.