Bearing fault diagnosis technology can effectively ensure that mechanical equipment can maintain good operating conditions. However, the bearing vibration signals collected on rotating machinery such as engines are often affected by a variety of noise disturbances and changes in working conditions, which makes the vibration signals have different characteristics, thus limiting the effectiveness of fault diagnosis. The common problems of widely used bearing fault diagnosis algorithms include complex structure design, large computation amount, and poor diagnostic effect in noisy environments. Therefore, we have designed a lightweight model. This model incorporates global attention on the basis of residual networks. The model firstly utilizes the global attention mechanism to enhance the feature extraction capability, and subsequently optimizes the residual neural network structure to improve the accuracy of fault detection and reduce the complexity of the algorithm. In addition, the model has good diagnostic ability and significant anti noise ability for variable load bearing faults in noisy environments, as demonstrated by our experiments.