Intelligent Diagnosis Method for Bearing Condition Changes Based on Domain Adaptation with Unlabeled Samples

In the intelligent operation and maintenance of industrial equipment, labeling failure data remains a challenging task due to its high cost and low efficiency. Although incorporating a large amount of unlabeled data alongside limited labeled samples can partially alleviate this “labeling bottleneck,” the performance and robustness of models still heavily depend on the scale and quality of annotated data, which often leads to generalization issues in real industrial scenarios. To address these challenges, this paper proposes an unsupervised fault diagnosis method based on an efficient domain adaptation model named E-DANNMK. This approach reduces reliance on manually labeled fault data, thereby mitigating annotation-related issues such as high cost and potential bias. The E-DANNMK model integrates residual networks, an efficient channel attention mechanism, and domain adversarial neural networks to improve both feature discriminability and cross-domain adaptability. To validate its effectiveness, experiments were conducted on two major bearing fault datasets. The results demonstrate that the proposed E-DANNMK model achieves an average diagnostic accuracy of 94.21%, outperforming mainstream domain adaptation methods—including CDAN, CORAL, DANN, CNN-Transformer, DMT and DANN-MK—by a margin ranging from 3.12% to 7.15%.