Research On Insulator Image Defect Detection Based On Deep Learning

In the power system,the insulator is a special insulation control device,which has important functions such as electrical insulation and mechanical fixation in highvoltage transmission lines.Insulators are exposed to the air for a long time,and they are more likely to be damaged or missing by wind and sun,which seriously threatens the safe operation of the power grid.Therefore,regular inspections of insulators are required.The traditional manual inspection method has potential safety hazards and low efficiency,and has been gradually replaced by drone inspection.In the process of UAV inspection,accurate detection of insulator defects in inspection images is the focus of our research.This thesis takes the insulators in the UAV inspection image as the research object,based on the deep learning algorithm,adopts the idea of "first recognition,then segmentation,and finally defect detection".The main work is as follows:(1)When recognizing the insulator string,taking into account the requirements for recognition speed,this thesis uses the one-stage object detection algorithm YOLOv3 to locate the insulator.The Io U strategy used in the YOLOv3 algorithm cannot accurately reflect the correlation between the predicted boxes and the ground truth boxes,resulting in a decrease in recognition accuracy.This thesis improves on the original algorithm and uses GIo U strategy to replace Io U strategy.After experimental comparison of the two algorithms before and after the improvement,the YOLOv3 algorithm based on GIo U has a certain improvement in m AP and recall rate.The final recognition result also verifies the effectiveness of the algorithm.(2)After completing the identification and positioning of the insulator,the U-Net network is used to segment the image.As a result of the direct segmentation of the UNet network,the edges of the insulators are not clearly segmented,which will affect the subsequent detection of insulator defects.Therefore,based on the idea of multi-task learning,a branch is added to segment the edge of the insulator on the basis of the original segmentation task,and the ECA attention mechanism module is used to strengthen the segmentation of the edge of the insulator.The experimental results show that after the improved network,the insulator segmentation effect is significantly improved compared with the existing methods.(3)After the identification and segmentation of the insulators,a pure insulator string is obtained.The outline of the insulators presents a regularly distributed ellipse.Mathematical modeling is carried out using the regularity of the insulator pixel point distribution,and the insulator string is scanned in a straight line.The detection of insulator defects is realized.Finally,a simple insulator defect detection platform was built to realize the identification,segmentation and defect detection of insulators.The research in this thesis has certain engineering application value and great significance to the intelligent development of power grids.