Research on transmission line component detection based on scenario prior knowledge and knowledge distillation

Abstract

Abstract Transmission lines for power systems are often located outdoors, which exposes them to various hazards that can seriously impact system reliability. Integrating UAV inspection and intelligent detection technology has emerged as a primary method for operational and maintenance tasks. However, challenges persist regarding detection accuracy and speed. This paper introduces a novel transmission line component detection approach, leveraging scene prior knowledge and knowledge distillation techniques. Initially, high-resolution aerial images of transmission lines are perceptually analysed. Subsequently, scene prior knowledge informs the selection of clustering algorithm parameters, enabling adaptive sub-region clustering and model optimization. Knowledge distillation is then applied to transfer comprehensive image feature knowledge from a teacher model to a student network, utilizing the obtained sub-regions as shared feature distillation masks. The efficacy of the proposed method is evaluated through extensive dataset analysis, demonstrating significant accuracy improvements ranging from 2.6% to 49.0% compared to alternative models.