eSmart Systems: Grid Vision® with Ultralytics YOLO Inspects 100,000+ km of Power Lines, 50% Faster

eSmart Systems, a Norway-based provider of computer vision and analytics for utility asset management, has integrated more than 30 Ultralytics YOLO models into its Grid Vision platform, which is deployed across over 70 utilities worldwide and has been used to inspect more than 100,000 kilometers of power lines. The integration halved inspection time, enabled faster defect detection, and shifted utilities from reactive repairs to condition-based maintenance. The company is headquartered in Halden, Norway, is ISO 27001 certified for information security management, and complies with Netcode Article 7.8 for secure data exchange in European electricity grid operations. Grid Vision analyzes aerial imagery captured by drones and helicopters, using YOLO models for object detection of components such as towers, crossarms, insulators, and conductors, and for defect detection covering vegetation encroachment, damage, and wear. The platform processes detected components and defects through cloud-based pipelines that flag potential issues, evaluate risk levels, and prioritize maintenance based on asset condition. The YOLO models are deployed using PyTorch for training and ONNX for CPU inference when GPU resources are limited, with the Ultralytics Python package supporting 15 export formats for environment-specific deployment. In Switzerland, a major energy company managing thousands of pylons in mountainous terrain reduced inspection time by 50% by transitioning from manual climbing to drone-based inspections using Grid Vision. A large U.S. utility digitized 1,400 transmission structures in three months, replacing manual photo reviews with AI-driven image analysis to support data-driven capital planning. In Finland, a transmission system operator reduced field visits and minimized outages by switching to drone-assisted assessments. Looking ahead, eSmart Systems is scaling Grid Vision to handle varying infrastructure, image capture methods, and data drift across regions, with MLOps pipelines enabling automated model retraining and dataset expansion.