Overhead Line Insulator Type Classification Using YOLOv3 Architectures

Overhead line insulators are among key elements in any power transmission and distribution systems. They are used for line insulation and fastening. Insulators support the electrical safety and isolation integrity of the overhead lines as well as substation equipment. Thorough inspection of insulators is extremely important as a premature failure potentially causes power outage leading to significant expenses to electricity providers. Over the past years, visual inspection of the insulators via unmanned aerial vehicle (UAV) imaging became quite popular since this method enables early identification of insulator faults in a shorter time period with least safety threats to personnel. In this regard, solutions based on artificial intelligence (AI) are highly beneficial in reducing costs and increasing efficiency in overhead power line inspections. In this work, a state-of-the-art object detection algorithm, namely, You Only Look Once (YOLO), is trained for insulator detection and type classification. In particular, YOLOv3 model and its variants, namely, YOLOv3-SPP and YOLOv3-tiny, are trained using 1832 images of different post- and suspension-type insulators. The results demonstrate that YOLOv3 and its variants can successfully detect and identify the insulator type despite external challenges such as crowded image, low brightness, and partially imaged test object. In particular, YOLOv3 showed a slightly better predictive performance than its variants. In applications where computational efficiency is desired, YOLOv3-tiny is an attractive alternative due to its lighter architecture with ~7 times fewer parameters than YOLOv3. Effective implementation of robust insulator localization and type identification technique are essential steps in development of an automated insulator condition diagnostic tool which is the primary motivation behind this study.