Object Classification with YOLOv5 for Electric Utility Asset Inspection using UAVs

The inspection of electric utility assets is an important procedure for repair and hazard prevention such as wildfires. Traditionally, human workers inspect power lines manually which is time consuming and potentially dangerous due to high elevation and high voltage. Manual inspection requires power to be shut off during the procedure and results in inconvenient blackouts for residents and businesses which can be exacerbated by time spent diagnosing faults and repair equipment. With recent developments in computer vision and artificial intelligence (AI) using machine learning, the process of inspecting electric utility assets can be both expedited and made safer using unmanned aerial vehicles (UAVs) in conjunction with sustainable and resilient federated learning communication networks. This work aims to train object classification models for deployment on UAVs which will detect common electric utility assets during inspection processes. The models are trained on a large dataset of approximately 30,000 high resolution images capturing five class objects: crossarms, cutouts, insulators, poles, and transformers. Eight different model configurations of the YOLOv5 algorithm are trained using the dataset and scored against each other to assess performance and computational cost for deployment on UAVs.