Abstract: In order to apply the advantages of deep learning to recognize 2D images for the fault diagnosis of planetary gearboxes, a fault diagnosis model of planetary gearboxes based on gram angle field-convolution neural network (GAF-CNN) was proposed. Using two methods, gram angle difference field (GADF) and Gram angle sum field (GASF) in the gram angle field (GAF), the planetary gearbox vibration signal was converted into a 2D image, and the image features were extracted and input into the optimized CNN model, and the ideal recognition accuracy was finally obtained. Analyzing the influence of network parameters and different network layers on the fault diagnosis model, an optimal model combination was carried out. The test and comparison analysis results show that GADF-CNN can provide higher recognition accuracy than GASF-CNN; GADF-CNN is superior to other intelligent algorithms in the faults diagnosis of planetary gearbox.