Ice Sleeves on Overhead Power Lines: a Deep Learning Approach for Semantic Segmentation

The wet-snow accretion on power grid cables and its detachment due to external phenomena represent a serious threat to the safety and continuity of the electricity grid service provisioning. As external phenomena, for instance, we can include temperature increases, wind and gravity. This paper presents a deep learning solution, based on a U-Net hybrid architecture, to recognize the formation of ice sleeves on overhead power lines. The semantic segmentation model has been trained on a set of images collected at the called Wetsnow Ice Laboratory Detection (WILD) experimental station. The aim of this work is to remove the currently-involved manual effort in studying the images coming from WILD. Starting from the content of this paper, the inference coming from the semantic segmentation model will be enriched in the future by further information about the dimensions of the area involved and the accretion over time. Such an approach will support operators while monitoring and characterizing wet-snow accretion on conductors cables.