An edge intelligence framework for rapid ice identification on transmission lines based on IARDNet

The deployment of vision-based ice detection models on computationally constrained edge devices presents a significant challenge for real-time grid monitoring. This paper presents a holistic framework integrating a novel Ice Accumulation Rapid Detection Network (IARDNet) with edge computing. IARDNet advances the YOLOv8 architecture through three principal modifications: first, the integration of DCNv3 deformable convolutions in the backbone for robust multi-scale feature extraction; second, the adoption of a Slim-Neck structure to harmonize Depthwise Separable and Standard Convolutions for an optimal accuracy-speed trade-off; and third, the replacement of ReLU with Swish activations to enhance non-linearity and prevent neuron necrosis. Validated experimentally, our framework demonstrates that IARDNet achieves high-precision, real-time identification on low-power edge hardware, thereby establishing a viable solution for rapid on-site ice accretion detection.