Benchmarking Generations of You Only Look Once Architectures for Detection of Defective and Normal Long Rod Insulators

Abstract

Effective infrastructure monitoring is a priority in all technical fields in this century. In high-voltage transmission networks, line inspection is one such task. Fault detection of insulators is crucial, and object detection algorithms can handle this problem. This work presents a comparison of You Only Look Once architectures. The different subtypes of the last three generations (v3, v4, and v5) are compared in terms of losses, precision, recall, and mean average precision on an open-source, augmented dataset of normal and defective insulators from the State Grid Corporation of China. The primary focus of this work is a comprehensive subtype analysis, providing a useful resource for academics and industry professionals involved in insulator detection and surveillance projects. This study aims to enhance the monitoring of insulator health and maintenance for industries relying on power grid stability. YOLOv5 subtypes are found to be the most suitable for this computer vision task, considering their mean average precision, which ranges between 98.1 and 99.0%, and a frame per second rate between 27.1 and 212.8, depending on the architecture size. While their predecessors are faster, they are less accurate. It is also discovered that, for all generations, normal-sized and large architectures generally demonstrate better accuracy. However, small architectures are noted for their significantly faster processing speeds.