The enhancement of the sample set quality and the improvement of the deep learning model are aimed to be achieved through a comprehensive approach, designed to increase diagnostic accuracy. This is necessitated by the challenges posed by gear fault vibration signals, characterized by nonlinearity, non-stationarity, sample imbalance, and variable operating conditions. The signals were processed initially using VMD (variational mode decomposition) , whereby energy entropy dimensionless indicators of each IMF (intrinsic mode function) component were extracted as the sample set. This approach is intended to counteract the detrimental effects of sample imbalance and operational variations. Subsequently, the learning rate of CNN(convolutional neural network) was optimized autonomously using PSO (particle swarm optimization) algorithm, aiming to minimize the risk of model overfitting. Additionally, a multi-branch global average pooling network, incorporating the concept of Inception modules, was designed for the purpose of feature fusion, thereby seeking to enhance the fault diagnostic accuracy of the model. The effectiveness of the method is validated through experimental data, demonstrating that fault diagnosis accuracy of up to 0.99 is achievable with the proposed method. Compared with other methods, the effectiveness and stability of this approach are highlighted.