Abstract: To address the issue of catenary dropper slackness and fractures in high-speed railway contact networks, which severely disrupt train operations, a multi-scale dynamic rotation YOLOv8n(MDR-YOLOv8n)algorithm was proposed to detect the abnormal states of the catenary dropper. High-resolution dropper images were acquired through the high-speed railway contact network 4C inspection system and enhanced via data augmentation. A convolutional local attention version 2(CloAttV2)was designed and integrated into the cross stage partial fusion(C2f)backbone network. Through collaborative axial adaptive pooling and dynamic sparse attention gating, the effectiveness of global-local feature fusion was boosted while the capture capability of key dropper features was enhanced. A lightweight multi-scale dynamic upsampling module with self-calibration mechanism was designed to adaptively adjust sampling weights of the feature maps, effectively utilizing contextual semantic information while reducing model parameters and enhancing anti-interference capability. An oriented bounding box-task align dynamic detection head(OBB-TADDH)was designed, which leveraged task-aligned optimization to enhance rotated object localization, suppress feature redundancy, and improve detection sensitivity for small targets. Experimental results demonstrated that MDR-YOLOv8n achieved 3.7 percentage points improvement in mean average precision at a confidence threshold of 0.5 and 2.3 percentage points increase in inference speed compared to the YOLOv8n model, while maintaining high detection performance under complex environmental conditions. MDR-YOLOv8n optimized the balance among detection accuracy, inference speed, and lightweight design, providing a novel solution for the intelligent upgrade of 4C inspection system.

Key words: YOLOv8n, overhead contact system dropper, dynamic upsample, attention mechanism, rotating target detection