Due to the harsh working environment, mechanical equipment is prone to malfunctions after long-term operation. Therefore, effective fault diagnosis of key components of mechanical equipment is of great significance. To enhance the diagnostic performance in the absence of sample label information, a distribution guided adversarial transfer network is presented to enhance the domain adaptation ability of the diagnostic model. Firstly, the latent distributions of the source and new samples are aligned under the guidance of the specific function, reducing the difference in discriminative features between the two domains. Secondly, to enable the feature distribution in the target domain to approach a specific distribution function, a new type of feature matching distance is designed to reduce the distribution discrepancy between samples. Finally, an adversarial training strategy is proposed to train transfer learning models, enabling the adversarial trained model to better recognize target samples in the absence of label information. The effectiveness of the suggested approach in unlabeled cross-domain diagnostic tasks is verified through experiments on the bearing dataset.