The bogie system of a railway train is one of the key subsystems ensuring operational safety. It is significantly affected by factors such as track conditions, time periods, and climate, leading to complex and variable working conditions, while fault-labeled data under specific conditions are scarce. Traditional deep learning-based fault diagnosis methods typically rely on large-scale labeled datasets and often exhibit poor generalization in cross-domain tasks. Existing transfer learning methods for bogie diagnosis generally lack effective alignment of internal feature distributions within the source domain, increasing the risk of negative transfer and limiting the model′s transfer performance. To address these issues, this article proposes an innovative multi-domain adversarial transfer method for fault diagnosis of railway bogies, aiming to align fault features of railway bogies under multiple operating conditions and enhance diagnostic performance across domains. First, pretraining is conducted on a source domain that includes multiple operating conditions. During pretraining, a domain-adversarial fusion method is used to align features across different conditions and to learn generalizable features. Then, the pretrained model is partially frozen and fine-tuned on 20% and 5% of the target domain data, respectively, to adapt to the target task. Testing on the test set achieves average classification recall rates of 98.48% and 93.00%, and average false alarm rates of 0.13% and 0.58%, respectively. Across all evaluation metrics—including recall, false alarm rate, precision, and F1-score—both the average and worst-case values outperform comparison methods. Experimental results show that the proposed method achieves higher diagnostic accuracy for faults in the power transmission chain based solely on easily accessible threephase motor current and rotational speed data. It performs particularly well for confusing fault types and maintains effective diagnostic capabilities even with minimal target domain data by leveraging diverse working condition data from the source domain.