Insulator Defect Detection And Recognition Based On Deep Learning

China’s transmission lines are expanding year by year,and insulators are important components on transmission lines.The automatic and rapid detection of insulator defects is of great significance for the safe operation of the lines and the stability of the power grid.Due to the unstable shooting angle and complex image background of inspection drones,there are problems in the current insulator detection task,such as incomplete device segmentation,slow detection speed,small targets,and low accuracy of irregular defect detection.Based on the summary of relevant research at home and abroad,this article improves the algorithms for insulator image segmentation and defect detection,improving the detection speed and accuracy.The main research work is as follows:(1)The high-precision instance segmentation algorithm Blend Mask and the widely used Mask R-CNN algorithm were used for image segmentation comparison experiments on aerial insulators.Blend Mask had better comprehensive indicators and achieved detection accuracy of over 90% for composite insulator segmentation and glass insulator segmentation;In response to the problem of slow segmentation and detection speed in the original algorithm,a lightweight network Mobile Net-V2 is proposed to replace the backbone network of Blend Mask from Res Net.This reduces the number of parameters and convolution operations in the feature extraction stage,thereby reducing the training time and detection time of the original algorithm by 14% and 29%,effectively improving the training and detection speed of the algorithm.(2)A comparative experiment was conducted using the one-stage object detection algorithm YOLOX-S and the two-stage object detection algorithm Faster R-CNN to detect the self explosion defects of composite insulators,ceramic insulator defects,ceramic insulator discharge traces,and glass insulator self explosion defects.YOLOX-S has relatively good comprehensive indicators,but there are problems such as inaccurate positioning,small target defect areas,and irregular defect and missed detection.This thesis proposes to add CAM attention mechanism to the PAFPN module and replace the loss function from binary cross entropy loss to Vari Focal Loss loss function,which improves the ability to capture the feature information of insulator defect parts.The results show that the improved YOLOX-S algorithm has increased the accuracy of ceramic insulator defect detection from 86.3% to 96.2%,and the accuracy of discharge trace detection has increased from 85.3% to 94.2%,effectively improving the detection accuracy,while the calculation speed has not decreased significantly.(3)A cascaded detection model is proposed to address the low accuracy of the proposed improved algorithm in detecting self exploding defects of glass insulators.Firstly,the improved Blend Mask algorithm is used for insulator segmentation,and then the summation function under the Detectron2 framework is used for background and mask removal.Finally,the improved YOLOX-S algorithm is used for defect detection,removing the interference of complex and similar colored backgrounds on defect detection of glass insulators,The results show that the accuracy of the cascade model in detecting self explosion defects of glass insulators is 95.2%.Compared with the improved YOLOX-S algorithm,which directly detects the original image,the accuracy has been improved by10.1%,and the m AP value has been improved by 10.2%.Finally,a visual display system was designed to achieve comparative display of insulator image segmentation and defect detection results.In summary,the improved algorithm and cascading model proposed in this article have significant improvement in detection accuracy and speed for different types of insulator defects.They have certain practical value and significance for rapid automatic detection of insulator defects in transmission lines and ensuring the safe and stable operation of power grid equipment.