Answers are grounded in this paper and cite their sources.
Deep neural networks have proven to be highly efficient in fault detection and classification using partial discharges (PDs) in gas-insulated switchgear (GIS). However, previous studies have not fully addressed the issue of limited labeled training data, which significantly impacts the performance of these models. Existing semi-supervised learning (SSL) approaches typically discard low-confidence pseudo-labels from unlabeled PD samples, because these unreliable labels can mislead the model. This gap in current research overlooks the potential value of low-confidence samples, which, despite their uncertain class, are unlikely to belong to the classes with the lowest probabilities. In this study, we aim to overcome this limitation by proposing a novel semi-supervised contrastive complementary learning (SCCL) method. Our SCCL approach generates a larger set of reliable negative pairs using complementary labels, allowing us to utilize the entire set of unlabeled PD samples effectively. We validate the feasibility of SCCL using phase-resolved PDs (PRPDs) and onsite noise data collected through an ultrahigh-frequency (UHF) PD measurement system. Experimental results indicate that the SCCL method achieves an impressive accuracy of 96.23% by effectively leveraging low-confidence unlabeled PD samples to improve classification performance in GIS under restricted labeling conditions.
No takes yet. Share an insight, caveat, or question.
Tai et al. (2025) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: