Traditional intelligent fault diagnosis methods assume sufficient training data sharing the same distribution as test data. However, in actual industrial scenarios, rotating machinery often operates under different working conditions. In addition, data for some special fault types of rotating machinery are extremely scarce, which leads to an imbalance in fault samples. To address the above issues, this paper proposes the Label Distribution Domain Adversarial Network (LDDAN). This network integrates nonlinear dynamic weight adjustment, domain adversarial transfer learning, and semi-supervised strategies. By dynamically adjusting weights to enhance the focus on minority class samples, it mitigates imbalance bias. It also reduces cross-condition distribution differences through domain adversarial learning and improves the utilization of unlabeled data with semi-supervised learning, thereby enhancing the generalization ability and accuracy of the fault diagnosis model. Finally, through the comparative experiments using the bearing dataset and the rail transit dataset from Case Western Reserve University, the effectiveness of the proposed method was verified.

Deep learning,sample imbalance,transfer learning,domain adaptation