Development of a UAV for vision-based fault detection in power transmission systems using deep learning

Purpose Reliable operation of power transmission systems depend on timely detection of insulator faults, yet conventional manual inspection is costly, hazardous and inconsistent. This study presents an unmanned aerial vehicles (UAVs)-based vision inspection approach integrating platform development, autonomous flight validation and a controlled multi-generation You Only Look Once (YOLO) benchmark for binary insulator fault detection. Design/methodology/approach A lightweight quadrotor UAV was developed and validated through six autonomous missions along a 132 kV transmission corridor near Prayagraj, India. A dataset of 3,951 annotated insulator images combining 3,173 UAV-collected images and 778 from a publicly available source, comprising 2,847 normal and 1,104 defective instances, was assembled with an 80:10:10 split and flight-level data leakage prevention. Five YOLO variants (YOLOv8s–YOLOv12s) were trained under matched conditions from Common Objects in Context (COCO)-pretrained weights. A FLUKE TiS75+ thermal camera was used separately as complementary qualitative evidence. Findings YOLOv8s achieved the best test-set performance: precision 0.895, recall 0.810, mAP@0.50 of 0.879 and mAP@0.50–0.95 of 0.482 at 142 FPS, outperforming all four newer YOLO generations. Thermal imaging at defective locations revealed hotspot temperatures of 83.5°C–197.6°C against backgrounds below 11°C, corroborating RGB-based fault identifications. Research limitations/implications The dataset was collected under specific environmental conditions near Prayagraj and generalisation to different geographic regions, insulator types and voltage configurations requires further validation. Six autonomous missions represent a limited operational sample. Real-time onboard detection using the Jetson Nano remains future work – all results reported here represent offline evaluation. Thermal imaging was conducted separately and was not fused with the RGB detection pipeline. The binary normal/defective formulation does not distinguish between specific fault types such as contamination, cracking and flashover discolouration, limiting the granularity of fault diagnosis currently achievable by the framework. Practical implications The proposed framework reduces dependence on hazardous manual inspection of overhead transmission lines by enabling systematic, UAV-based image acquisition combined with automated deep learning fault detection. The 142 FPS offline inference speed confirms strong potential for future real-time onboard deployment. The controlled cross-generation YOLO benchmark provides practitioners with an evidence-based model selection guide for similar binary insulator datasets. The complementary thermal imaging results demonstrate that RGB-flagged defective insulators exhibit temperature anomalies consistent with active fault conditions, reinforcing the practical value of multi-modal inspection for power grid maintenance and predictive maintenance programmes. Originality/value This study combines purpose-built UAV development, real-world flight validation, a cross-generation benchmark on a field-collected dataset and complementary thermal corroboration, providing a more complete and reproducible inspection study than single-contribution approaches.