A Self-Supervised Multiview Contrastive Learning Network for the Fault Diagnosis of Rotating Machinery Under Limited Annotation Information

The field of rotating machinery has long faced the challenge of limited annotated samples, which significantly affects the performance of neural network models. Most existing methods rely mainly on a few annotated samples for fault diagnosis algorithm design without fully utilizing a large amount of unannotated data in the monitoring phase. To alleviate the effect of limited annotation information, this paper proposes a self-supervised multi-view contrastive learning network (SMCLN) based on time-frequency analysis. This framework learns from a large number of unlabeled samples and leverages the rich feature information contained in time-frequency images to train a feature extractor with strong generalization capabilities. Given the significant differences between vibration signals and images, this paper designs appropriate data augmentation methods. Additionally, a multi-task mechanism is introduced by designing a novel loss function, enabling the network to simultaneously handle both matching and prediction tasks. This approach enhances the representational capacity and further improves overall performance. Finally, a simple linear classifier is added and fine-tuned on a small amount of labeled data. Experiments on both public and laboratory datasets demonstrate that our proposed SMCLN can effectively extract valuable feature information and significantly improve classification accuracy under limited annotated dataset.