Motor fault diagnosis based on deep learning has become a popular research focus in recent years. However, as the network deepens and the parameter scale expands, existing motor fault diagnostic models experience two issues. Firstly, existing diagnostic models based on deep learning are difficult to train and frequently encounter problems of gradient vanishing and non-convergence. Secondly, the training costs of existing diagnostic models are too expensive, such as long training time, high demand for data, and huge computational resource consumption. To solve these two issues, a pre-training method based on Siamese residual network is proposed. To solve the first issue, the residual theory is introduced and a residual network with cross-layer connections is constructed to extract motor fault data features. To solve the second issue, two constructed residual networks with shared parameters are embedded into two parallel branches of a Siamese network to construct the Siamese residual network which is used to pre-train the residual networks. Finally, a pre-trained residual network is connected with a classifier to form the diagnostic model and it is further trained on motor fault datasets for accurate diagnosis. The proposed method is verified on an experimental motor fault dataset. The results show that the pre-training based on Siamese residual network can reduce the training difficulties and costs of diagnostic models and improve diagnostic accuracy on balanced and imbalanced data.