Vibration signal is widely used in deep learning (DL) fault diagnosis, but it is difficult to obtain and susceptible to interference. Meanwhile, DL fault diagnosis models often suffer from inconsistent domain distribution, and insufficient labeled data. To solve the above problems, this article proposes a transfer-learning model, the domain straight mapping method (DSMM), in which the vibration signal is replaced with an easily accessible current signal. The original current signal is straightly mapped to achieve domain alignment. Then graph generation and a multireceptive field graph convolution network (GCN) are used to obtain features. Finally, the sample shortage is solved by classifier and domain confrontations, and the classification result is obtained by training. The proposed method has higher convergence speed and accuracy than comparison models. In addition to mechanical fault diagnosis, it has potential in other transfer learning fields.