Multimodal teacher–student learning for high-voltage insulator cleaning decisions
Year: 2026
Type: conference-paper
Abstract: Progressive surface contamination degrades the dielectric performance of high-voltage (HV) insulators and substantially increase the risk of pollution flashover, making timely cleaning decisions critical for transmission-network reliability. This paper proposes a physicsinformed teacher–student deep-learning framework that converts surface-discharge images into actionable cleaning decisions for glass HV insulators. A laboratory dataset of approximately 450 night-time discharge images was acquired from an artificially contaminated planar glass model energised by a 300 kV AC source, spanning four contamination levels defined by Equivalent Salt Deposit Density (ESDD) and Non-Soluble Deposit Density (NSDD). Following the IEC site-pollution-severity classification, each sample was mapped to a binary maintenance label: No Cleaning Required or Cleaning Required. A multimodal teacher network based on MobileNetV3-Large jointly exploits discharge images and ESDD/NSDD descriptors to learn contamination-aware representations, while knowledge distillation transfers this capability to a lightweight MobileNetV3-Small student that requires only images at inference. The teacher reached 95.65% accuracy and a 0.924 ROC-AUC, and the distilled image-only student retained 89.13% accuracy, 95.83% recall and a 0.934 ROC-AUC, outperforming conventional imageonly training while preserving the safety-critical detection of insulators that need cleaning. The framework offers a practical route toward UAV-based condition assessment and predictive maintenance of HV insulators.
Authors: Ferrah, Imene, Alzahrani, Ateyah
Language: English
Subfield: Materials Chemistry
Field: Materials Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7213255693