Based on Improved YOLOv11 Transmission Line Bird Hazard Detection
Fu Chen, Song Jingliang, Li Yongliang, Yang Jia, Liu Lin, She Yubin, Zhao Lei, Pei Shaotong · Journal of Sensors · 2026
Bird hazards are a major natural cause of transmission line tripping, posing significant risks to power supply security. Current deep learning‐based bird hazard detection models have prominent limitations. Most studies only focus on single bird species identification and lack the ability to distinguish different species, with overall detection accuracy remaining relatively low. In addition, relevant models suffer from excessively large parameter counts and high model weight, thus failing to adapt to edge device deployment in transmission line scenarios. To address these issues, this article proposes a scene‐adaptive enhanced YOLOv11 algorithm for high‐precision bird hazard detection with three targeted improvements. The CIoU loss is replaced with DIoU loss to eliminate regression bias induced by aspect‐ratio penalties. The ADown module is adopted for lightweight subsampling while preserving the semantic features of small targets. The C3K2_EMA module is designed to enhance target focus and background suppression by fusing the EMA attention mechanism. Experiments on a self‐constructed dataset show the proposed model achieves 75.1% mAP50 and 56.5% mAP50–95, with respective increases of 3.9% and 5.3% compared with the baseline YOLOv11. The model has only 2.43M parameters and a frame rate of 32.9 frames per second (FPS), meeting the requirements for real‐time edge deployment. Robustness tests across weather, distance, and occlusion scenarios confirm the model outperforms YOLOv8, YOLOv9, YOLOv10, and YOLOv11. This study’s core contribution lies in the synergistic module design that balances accuracy, lightweighting, and real‐time performance, providing a reliable edge detection solution for transmission line bird hazard prevention.