High-Voltage Insulators Condition Classification Using Intelligent Techniques

Abstract

High-voltage insulators are essential components of overhead power transmission lines as they isolate live conductors from energized power towers while providing mechanical and electrical support. Since insulators are commonly located outdoors, they are inevitably exposed to severe weather conditions, leading to power outages. An interruption in power distribution affects numerous industries and national security sectors and leads to financial losses. This thesis aims to develop a high-voltage insulator condition monitoring system integrating Artificial Intelligence (AI) driven object detection and image classification frameworks. Various condition-monitoring approaches have been developed for high-voltage insulators, including on-foot inspections, electrical tests, protective coatings, and regular cleanings. However, these methods are labor-intensive, time-consuming, and prone to human error. The industry has adopted infrared (IR) thermography, ultrasonic techniques, and ultraviolet (UV) cameras for their cost-effectiveness and safety. However, these solutions are limited by environmental sensitivity, close-range measurement requirements, and the inability to detect mechanical defects, making large-scale deployment challenging. With the integration of AI, deep learning-based solutions have emerged to analyze high-voltage insulator conditions. However, a significant gap remains due to the lack of an accurate, robust, and reliable classifier model that generalizes across different fault types and insulator materials. This thesis proposes an AI-driven object detection and image classification framework for high-voltage insulator mechanical fault and surface contamination detection, along with contamination severity evaluation using laboratory-based FDS and Ansys Maxwell simulations. A large custom high-voltage insulator dataset was collected as a part of this thesis, incorporating common mechanical faults, surface contaminations, and contamination severity data. A significant contribution of this thesis is the development of a You Only Look Once (YOLO) detection framework for identifying mechanical faults and contamination types. For mechanical fault detection, the comparative evaluation demonstrated that the YOLOv5x model achieved a mean Average Precision (mAP) of 0.995 at an Intersection over Union (IOU) threshold of 0.5 (mAP@0.5). For surface contamination detection, YOLOv11m achieved a mAP@0.5:0.95 of 0.9597 and 0.7615 on unseen laboratory test and industrial data, respectively, demonstrating its generalizability.