To address the issues of inconsistent feature distributions and the influence of noise components on the transfer effect in gearbox vibration data collected under different operating conditions, this paper proposes a deep transfer learning fault diagnosis method that integrates an attention mechanism with domain adversarial transfer networks. First, labeled and unlabeled vibration signals are constructed into datasets using a fixed-length data segmentation method. Second, to reduce the negative transfer impact caused by noisy samples, a convolutional block attention module (CBAM) and a discriminative loss term are used to assist the feature extractor in extracting discriminative features and enhancing the classification decision boundary. Finally, to solve the problem of inconsistent data feature distributions, a multi-kernel maximum mean discrepancy (MK-MMD) is employed to align the global distributions of the source and target domains, and an adversarial mechanism is used to align the subdomain distributions between the two domains. Experimental validation on a publicly available variable-condition gearbox fault dataset demonstrates that the proposed method achieves an average recognition accuracy of over 96.25%. A comparison with other diagnostic methods further validates the effectiveness and superiority of the proposed approach.