YOLOv8-eRFD-AP: A Novel Domain Generalization Model for UAV-Based Insulator Inspection Under Adverse Weather Conditions

The deployment of artificial intelligence (AI)-powered uncrewed aerial vehicles (UAVs) for high-voltage power line inspection has become a crucial advancement in ensuring the stability and reliability of electrical transmission networks. Among various deep learning architectures, You Only Look Once (YOLO)-based models are widely used due to their real-time object detection efficiency. However, these models frequently struggle to generalize when exposed to diverse environmental conditions. To address this, we present YOLOv8-eRFD-AP, an enhanced YOLOv8 variant designed to improve generalization ability across varied weather scenarios for insulator anomaly detection. The model is trained under clear-weather conditions and evaluated on the IDID_Weather dataset, which includes rain, fog, and snow, to assess its performance under real-world challenges. Our model introduces a novel version of Robust Feature Downsampling (RFD) method, enhanced with Convolutional Block Attention Modules (CBAM) and Normalization Perturbation (NP) to preserve key spatial information and improve feature discrimination across different weather conditions. Experimental results confirm that YOLOv8-eRFD-AP outperforms state-of-the-art models, including YOLOv11, YOLOv12, and Real-Time Detection Transformer (RT-DETR), in both clear and adverse weather settings. The model achieves a mean average precision at 0.5 intersection over union ([email protected]) of 92.0%, exceeding the second-best model by 2.7% under clear weather. Under rain and snow, it attains 85.2% and 73.1%, reflecting improvements of 10.3% and 13.3%, respectively, over prior best-performing models. Despite the improved generalization ability, the proposed model remains real-time capable, achieving an inference speed of 12.5 ms per image (~80 frames per second, FPS) on a Tesla T4 graphics processing unit (GPU). This work demonstrates the potential of YOLOv8-eRFD-AP to enhance UAV-based power line inspections, particularly by improving generalization to previously unseen scenarios. Future research will focus on refining NP strategies for more efficient handling of both uniform and non-uniform distributions, as well as developing a lightweight version optimized for low-power edge computing environments.