As the scale and frequency of overhead transmission line inspections continue to increase, the massive volume of aerial imagery poses significant challenges to cloud computing resources and network bandwidth. Meanwhile, existing bird-related outage detection methods still struggle to balance accuracy and real-time performance in scenarios involving complex backgrounds and densely distributed small objects. To address this issue, this paper proposes a lightweight object detection method for bird-related fault detection on edge devices, termed BDEA-VFNet. First, based on the VFNet architecture, a fully convolutional masked autoencoder pre-training strategy is introduced together with a global response normalization layer, which accelerates model convergence and reduces model complexity, and enhances the feature extraction capability for small objects. Second, a feature pyramid structure optimized via neural architecture search is employed to improve the efficiency of multi-scale feature fusion and reduce cross-layer information loss. Finally, a dynamic soft label assignment strategy based on SimOTA is designed, and an improved FCOS anchor-free mechanism is proposed to generate candidate targets, thereby enhancing hard example learning and reducing the risk of missed detection of bird-related fault in complex backgrounds. Experimental results on an overhead transmission line bird-related outage dataset and UAV edge devices demonstrate that BDEA-VFNet can rapidly and effectively detect bird-related outages on edge devices. Compared with the original VFNet, BDEA-VFNet improves mAP by 5.20%, while reducing the number of parameters by 77.37% and the computational cost by 83.19%, effectively balancing detection accuracy and the lightweight requirements of edge deployment.