Adversarial domain adaptation based on contrastive learning for bearings fault diagnosis

Accurate fault diagnosis of machines is crucial for increasing efficiency, reducing maintenance costs, and preventing catastrophic consequences.While data-driven methods have shown promise in fault diagnosis, most existing models frequently encounter challenges in achieving satisfactory results in industrial fault diagnosis due to varying working conditions. Studies on domain adaptation have made significant contributions to addressing this problem. However, most of these methods concentrate on aligning the inter-domain distributions, whereas the degradation of intra-domain classification performance is overlooked, resulting in confusion at the class boundaries of the target domain during cross-domain diagnosis. To address this issue, a self-supervised domain contrastive discrimination network (SDCDN) is proposed for bearing fault diagnosis under variable working conditions. The proposed method takes the data from both the source and target domains as input for contrastive learning training. Through self-supervised learning, the feature enhancer is trained to capture domain-contrastive features and effectively distinguish the target category. By aligning the distribution of the source and target domains through adversarial learning, the cross-domain diagnosis is achieved without supervision. To validate the effectiveness of the proposed method, six cross-conditional diagnostic tasks are performed on each dataset, utilizing two bearing datasets containing different damage types and a gearbox dataset. The evaluation indicators employed are diagnostic accuracy and computational efficiency. Furthermore, an ablation study is conducted to evaluate the contribution of the domain contrast and discrimination modules. The results demonstrate that the average accuracy of the proposed method is markedly superior to that of the comparison methods for all six cross-domain diagnostic tasks in each of the three datasets, highlighting the superiority of the proposed method.