Because of uncertain working conditions and the lack of labels for most conditions, transfer learning has great potential in fault diagnosis for high-speed train bogies. This article proposed a novel domain adaptation model for intelligent diagnosis under varying working conditions, namely, deep adversarial hybrid domain-adaptation network (DAHAN). In the proposed method, to further improve the transfer learning performance at a low signal-to-noise (SNR) ratio, the local domain adaptation by Wasserstein distance is developed to align category-level features which are combined with global domain adaptation to hybrid domain adaptation. In addition, an optimized training procedure is constructed to improve the stability and training speed of the proposed adversarial domain adaptation model to enhance the efficiency and real-time performance of the diagnosis method. The effectiveness and advantage of DAHAN are verified by the Case Western Reserve University (CWRU)-bearing dataset. Finally, trained by train bogie gearbox fault dataset, the proposed method shows a great performance improvement and avoids the mismatch of weak fault features, proving its applicability on train bogie fault diagnosis.