In industries for gas turbine rotor, there are a large number of normal operation vibration signal sample data and few fault data,which caused fault diagnosis accuracy lower. A gas turbine rotor deep transfer learning fault diagnosis method is proposed. First, a firstlayer wide convolutional kernel deep convolutional neural network (WDCNN) model is pretrained with a typical industry sample dataset, obtained the model initial weights. Second, in the source domain, the weights of the WDCNN model are updated using a large number of normal operation samples obtained from the test drive of a certain type of gas turbine; In the target domain, the normal and fault data sample characteristics of the gas turbine are extracted by using the convolutional layer trained in the source domain, and then the support vector machines (SVM) are used for classification identification, so as to achieve the gas turbine fault identification. The experimental results of the test data show that the method identification accuracy is 96%, which verifies the feasibility of migrating the pretrained deep learning model of the bearing dataset to the field of gas turbine rotor for fault diagnosis.