IDD-DETR: Insulator Defect Detection Model and Low-Carbon Operation and Maintenance Application Based on Bidirectional Cross-Scale Fusion and Dynamic Histogram Attention.

Abstract

Against the background of the "dual carbon" goal and the construction of a new power system, the intelligent operation and maintenance of insulators for ultra-high voltage transmission lines face challenges such as difficulty in detecting small-scale defects and strong interference from complex backgrounds. This paper proposes an improved network IDD-DETR to address the problems of inefficient one-way feature fusion and low-contrast defects that are easily overwhelmed in existing RT-DETR models. The enhanced network IDD-DETR replaces PAFPN with a Feature-Focused Diffusion Network (FFDN) and improves multi-scale fusion efficiency through bidirectional cross-scale interaction and designs Dynamic-Range Histogram Self-Attention (DHSA) to enhance defect response in low brightness areas. The experiment showed that its mAP⁵⁰ reached 81.7% (an increase of 3.8% percentage points compared to RT-DETR), the flashover defect AP⁵⁰ reached 74.6% (+6.1% percentage points), and it maintained 76 FPS on NVIDIA RTX3060, with an average decrease of 1.65% in mAP⁵⁰ under complex environments. This model reduces the comprehensive missed detection rate from 26.7% to 23.3%, reduces 45.6 GWh of power loss annually (corresponding to 283,000 tons of CO₂ emission reductions, with 64.3% of the reduction contributed by flashover defect detection), improves inspection efficiency by 60%, reduces manual pole climbing frequency by 37%, and reduces 28 high-altitude risk events annually, providing support for low-carbon operation and maintenance of transmission lines.