A Few-Label Contrastive Transfer Fault Diagnosis Method for Rolling Element Bearings

Rolling bearings in rotating machinery are prone to failures under harsh operating conditions. While supervised fault diagnosis methods using fully annotated vibration datasets have advanced, practical industrial scenarios often involve unlabeled sensor data. Manual labeling is time-consuming, resulting in limited labeled data for each fault category. Acquiring fault data becomes challenging when operating conditions change, leading to a scarcity of fault data. To address these challenges, this paper proposes a fault diagnosis approach based on few-labeled data and contrastive transfer learning. It utilizes a contrastive predictive coding model in self-supervised representation learning to extract discriminative features from unlabeled data. The model performs contrastive learning on partially normal data from both the source and target domains, extracting shared features. The feature extractor is trained using a small amount of labeled data from the source domain, forming the fault diagnosis model. Experimental validation demonstrates the effectiveness of this approach. This fault diagnosis approach, leveraging few-labeled data and contrastive transfer learning, addresses the challenges of limited labeled data and changing operating conditions. It enhances applicability in practical industrial scenarios with limited labeled fault data. The experimental results validate its effectiveness for fault diagnosis in rolling bearings and similar machinery.