In recent years, the importance of industrial equipment in modern manufacturing has become increasingly prominent. Fault diagnosis, as a key technology to ensure stable operation of equipment, has become an indispensable part of modern industrial management. However, the diversity of production environments and equipment types leads to significant differences in equipment operating states and a lack of fault labels for the target operating conditions. The generalization ability of traditional fault diagnosis methods for a single working condition usually decreases substantially, making accurate diagnosis very challenging in real scenarios. To address this issue, we present a transfer learning approach for fault diagnosis based on frequency-augmented contrastive learning. It uses a frequency-enhanced recurrent generative adversarial network for data augmentation and constructs corresponding positive-negative sample pairs, and optimizes the discrimination and domain invariant between features by domain-adversarial contrastive learning. Experiments are conducted on three public datasets, and the experimental results validate the effectiveness of the proposed approach.