Abstract
Interfacial discharge constitutes one of the most detrimental factors in multilayer composite insulation, severely accelerating aging, causing unexpected failures, and potentially triggering catastrophic accidents. To address challenges in temporal feature extraction, noise sensitivity, and limited labeled samples for interfacial discharge detection, this study proposes a physics-informed hypergraph enhancement and recognition framework. First, we develop a physics-informed hypergraph-guided denoising diffusion implicit model (PIH-DDIM) to mitigate noise interference inherent in data-driven approaches, enabling high-fidelity data augmentation. Subsequently, the HyperSwin multiscale recognition network is designed. Specifically, a hierarchical hypergraph patch merging mechanism is incorporated to reduce feature resolution while expanding receptive fields progressively. Window-based hypergraph multi-head attention is constructed to facilitate cross-region information interaction. Furthermore, masked hypergraph modeling (MHM) self-supervised pretraining is introduced to autonomously learn intrinsic hypergraph associations and topology, effectively overcoming label scarcity and poor model generalizability. Experiments demonstrate a discharge recognition accuracy of 96.32% on real-world EMU cable terminal datasets—a 2.49 percentage-point improvement over baseline methods.
Funding Information
  • National Natural Science Foundation of China (52377161)