Fourier-enhanced prototype contrastive learning for partial discharge diagnosis and localization in gas-insulated substations

Accurate partial discharge (PD) condition assessment in gas-insulated substations (GIS) is often hindered by strong background noise, limited labeled data, and unknown defect types. Existing few-shot learning approaches show promise but struggle to generalize under unknown defect types and noisy field conditions. To address these challenges, this paper proposes a Fourier-enhanced prototype contrastive learning (FEPCL) framework for GIS PD diagnosis and localization. First, a Fourier feature refiner is first designed to extract amplitude and phase spectra, enhancing feature robustness against noise interference. Then, a prototype contrastive learning strategy is designed to align intra-class features and separate inter-class representations, thereby capturing domain-invariant and discriminative embeddings crucial for reliable PD condition assessment. To further recognize unseen defects, an open-set adaptation module integrating a secondary confidence rule and progressive self-training dynamically adjusts the classifier. Experiments on real GIS datasets demonstrate that FEPCL achieves over 95% accuracy in both diagnosis and localization tasks, outperforming state-of-the-art baselines. The results confirm that FEPCL provides a noise-resilient and open-set-aware solution for practical GIS PD condition assessment.