Aiming to solve the problem that the standard training set in deep learning cannot fully cover the fault characteristics in actual working conditions, resulting in a sharp decline in the fault recognition accuracy, a domain adversarial neural network (DANN) transfer learning method, named Transformer-DANN, is proposed, which fuses an one-dimensional convolutional neural network (1DCNN) and Transformer layers. The 1DCNN and Transformer layers are used to improve the ability of the feature extractor to extract features and reduce the computational complexity. Aiming at the issue of different characteristics of fault data under different loads, a domain adversarial neural network (DANN) adversarial approach is adopted to process the fault data. The Transformer-DANN method has an average recognition accuracy of 98.13% and a maximum accuracy of 99.42% under varying motor operating conditions. The experimental results show that the proposed method can effectively improve recognition accuracy and meet the task requirements of fault diagnosis under variable operating conditions.