A Novel Impulse Information-Enhanced Semi-Supervised Learning for Few-Label Fault Diagnosis of Rotary Machines

Intelligent fault diagnosis based on data-driven models has shown excellent effectiveness in recent years, significantly promoting diagnosis technology. However, data-driven models require large amounts of labeled data for supervised training, and the quantity of labeled data in industrial engineering is limited. Few-label learning is still a significant challenge for engineering applications in industry. To address this issue, this article proposes an impulse information-enhanced semi-supervised learning (IESSL) model to improve the performance of few-label fault diagnosis. We design a novel encoder module based on fault-related impulse knowledge to construct unsupervised features (UFs) from large amounts of data without labels. UFs are utilized to constrain the decision space of the backbone network. Then, a few labeled data are used to fine-tune the backbone networks to make the more accurate mapping between the fault types and decision space. The experimental results on two datasets demonstrate the effectiveness of our proposed method.