The nonlinear and non-stationary characteristics of the weak fault signals of gears in industrial equipment, which are prone to be obscured by strong background noise, as well as the severe imbalance typically exhibited by industrial fault data, an intelligent fault diagnosis method based on the extreme gradient boosting model optimized by the Sparrow Search Algorithm for convolutional neural networks (CNN-SSA-XGBOOST) is proposed. The one-dimensional gear vibration data is reconstructed using the Gramian angular field, retaining the original data information while including the time correlation; the sparrow search optimization algorithm is employed to iteratively optimize the structure of the encoded convolutional neural network. Experimental validation is carried out using the gearbox dataset from Southeast University, and the experimental results show that the method can adaptively generate the network structure to enhance the shock component in the weak gear faults, and effectively extract the weak gear fault features that are overwhelmed by strong noise, with an average diagnostic accuracy of 98.5%.