In the actual industrial environment, the collection of gas turbine rotor fault data is challenging, leading to a scarcity of fault samples and an inability to meet the massive training requirements of fault models. Leveraging the advantages of DenseNet in image feature extraction and the Transformer structure in the visual field, an improved gas turbine rotor fault diagnosis method based on the DenseNet-ViT joint network was proposed. Firstly, the classification layer of DenseNet was abandoned, and only the feature extraction layer of DenseNet was utilized. Subsequently, the output layer of the modified DenseNet was connected to the input layer of the ViT model to form the joint network. Additionally, in response to the issue of lengthy training time for the fault model, transfer learning was employed to transfer the trained model’s weight parameters, which could expedite the training process and conserve computing resources. Simulated data of gas turbine rotor faults could be acquired through the gas turbine rotor simulation experimental platform constructed in the laboratory, and real fault data of different types of rotors in the actual environment were obtained on a certain type of gas turbine test bed. Utilizing both the simulated and real data for model testing could better verify the reliability of the proposed method. The experimental results indicate that the fault recognition accuracy rates reached 96.8% and 97.3% in the tests of two distinct rotor fault datasets, respectively, demonstrating that this method possesses a relatively high rotor fault recognition accuracy. In the subsequent comparative verification experiments, by comparing with CNN and VGG-16, etc, the fault classification accuracy of this model was also higher than those networks, thereby further validating the superiority and reliability of this model.