Research On Insulator Self-explosion Defect Detection Based On Improved Faster R-CNN

As my continues to expand the erection area of large capacity,trans regional and ultra-high voltage transmission lines,the increasing area not only promotes the rapid development of the power system,but also adds difficulties to the work of patrol inspectors.How to timely and accurately detect potential safety hazards in transmission lines is a new challenge.Insulators,as an extremely important part of the transmission line,may affect the power transmission,even cause a large area blackout,and lead to economic losses that are difficult to estimate when a fault occurs and is not resolved in time.In this regard,the traditional detection methods have been unable to adapt to the rapidly developing power system.With the development of intelligence,it has become the mainstream trend to use UAV to patrol insulators in transmission lines.Therefore,for the insulator images obtained by UAVs,how to detect the defective parts intelligently,efficiently and accurately has become a new research hotspot.In this paper,the insulator image in the transmission line is detected,and the depth learning method is introduced to build a VOC2007 format data set.Aiming at the problem that the proportion of insulator target self explosion defects is small,a detection scheme of cascade model is proposed,that is,first locate the insulator string from the UAV aerial image,then cut it,and then use depth learning to detect the insulator self explosion points.The detailed work content of this paper is as follows:In view of the problem of insufficient insulator image data,this paper expanded the insulator image by rotating,flipping,adjusting brightness,blurring and other methods,enriched the insulator data set,and avoided the problem of low precision of the detection model caused by insufficient data set.Aiming at the problem of uneven illumination of aerial image of transmission line insulator as input image,this paper proposes an image enhancement method using adaptive genetic algorithm to optimize linear transformation.This method uses adaptive genetic algorithm to optimize the position parameters of two inflection points of three segment linear transformation.The experiment shows that the image exposure obtained by this method is significantly enhanced,the contrast is strong,the primary and secondary are clear,and the image quality has been significantly improved.In order to locate insulator targets from aerial images for subsequent clipping and detection of self explosive defects,the two-stage detection algorithm Faster R-CNN is selected in this paper.At the same time,comparative tests based on feature extraction networks are also conducted.The test results show that Resnet-101 has higher recognition rate and faster detection speed.At the same time,the image of the identified insulator string is cut interactively,so that the image can meet the size conditions of the input image for subsequent detection.In order to improve the detection effect of the Faster R-CNN detection model,this paper modifies the structure of the Faster R-CNN network,mainly aiming at the generation process of candidate regions,designs a sequential search network,uses its strong decision-making ability to screen candidate regions,which can greatly reduce the number of candidate regions and enhance the quality of candidate regions.The experiment shows that this method can effectively improve the detection speed and accuracy of Faster R-CNN,and effectively make up for the slow detection speed of Faster R-CNN.Figure [43] Table[7] Reference[69]...