Improved YOLO v8: An Efficient Real-Time Fault Detection for High-Voltage Insulators for Autonomous Aerial Vehicles (AAVs) Images Using Deep Learning Techniques

Insulators are the components that offer support to the conductors in the transmission lines and also provide the insulation to these conductors. It is important for power transmission lines. Hence, conducting inspections to verify the safety of a depends extensively on the inspection of insulator systems. There are standard practices for inspecting Insulators may involve hand inspections or inspections carried out by helicopter. However, many a time, this type of inspection may turn out to be extremely expensive, and time-consuming. An alternative to performing conventional In the context of insulator inspections of the transmission line, Unmanned Aerial Vehicles (UAV), which are (HD) cameras, used to perform comprehensive aerial inspection of the insulators. Inclusion of techniques using deep learning, artificial intelligence has even furthered defect detection, identification, and classification processes. Among these, the You Only Look Once (YOLO) object detection method is well known for its speed and accuracy. The use of YOLOv8 for insulator failure identification using UAV images is covered in this work. The methodology involves precise localization of insulators followed by crack detection. The following enhancements were made to increase the detection accuracy. In order to enhance global modelling capabilities, YOLO v8 first introduces the Contextual Transformer module and secondly introduces DRConv (dynamic region-aware convolution) to enhance convolutional format. A dataset of 3700 transmission line images taken under various lighting conditions, camera angles, and intricate natural backgrounds was used to train and assess the proposed model. To guarantee the accurate localization of insulator faults in practical situations, all samples were manually annotated using the Roboflow platform with oriented bounding boxes. The experimental findings show that the image classification network proposed in this study can compete with the traditional YOLOv8n network in terms of detection accuracy. It shows an increase of 4.1% in precision,5.5% increase in recall, and 4.1% increase in mAP compared to the original YOLOv8 model.