Research On Insulator Defect Detection Of Transmission-Line Based On YOLO-V4

Since the reform and opening up,the economic infrastructure of China has developed at top speed.At the same time,the demand for electric energy was also increasing.A stable supply of electric energy is related to the fundamental interests of the people of all ethnic groups in China.Transmission-line is the only medium responsible for power transmission.To ensure the safe,stable,continuous the power grid operation,the transmission-line must inspect regularly.However,with the constant expansion of the scale of the transmission-network,the contradiction between it and the low efficiency of manual inspection methods is gradually exposed.At the same time,the field of unmanned aerial vehicles is in the hot research field,which has made unmanned aerial vehicles develop rapidly in recent years,and the technology has become more and more mature.Therefore,under this background,the daily inspection method has changed from the traditional "Manual Inspection" mode to the "Drone Inspection" mode.For transmission-line,the insulator is an important part of it,and the detection of insulator defects is also a key work of drone inspection.This thesis takes insulator defects in drone inspection work as the research object.According to the image data obtained in inspection work,optimization is made based on the YOLO-V4 algorithm,a structure of the new network is designed,the loss function is improved,and a more suitable model,the DYOLO-V4 model,for insulator defect detection is proposed.The main research work of this thesis is as follows:(1)Firstly,manually screened and labelled the insulator image data collected during drone inspection.The image dataset was augmented with rotation,mirroring,and mirror-rotation methods to simulate insulator images captured by drones from various angles.Data augmentation technology,Gaussian-filters and Median-filters,removed image noise from image acquisition.The bilinear interpolation algorithm compressed the image to 416×416resolution,which is more suitable for the training-size of the target detection algorithm.(2)Secondly,this thesis introduced the models of the typical algorithm in the field of target detection,and built the models.CSPDarknet53 was used as the feature extraction network,PANet was used as the feature aggregation network,this thesis built an insulator defect detection model based on YOLO-V4.And trained the models using the dataset of insulators and their defects.According to the experimental results of the YOLO-V4 model with other typical algorithm models,the performance of the YOLO-V4 model was better.(3)Finally,because the background of insulator defects was complex,the embedded EC component proposed in this thesis was connected between the output layer and the feature aggregation network to make the model more suitable for detection tasks.According to the characteristics of unbalanced data samples,the confidence loss function of Focal loss was used to adjust the learning weight of positive and negative samples.According to the occlusion phenomenon of the insulator instance target,Mosaic data-augmentation technology was used to randomly cut and arbitrarily splicing the training images,which increased the network’s learning ability for occluded insulator samples to a certain extent.With the above improvement measures,this thesis designed a more suitable model for insulator defect detection called DYOLO-V4.Through the results of DYOLO-V4 ablation experiments and the visual comparison,the m AP of the DYOLO-V4 model proposed in this thesis reaches 88.57%,which is 9.42 percentage points higher than that of the original YOLO-V4 model.The inference speed is 0.0279 seconds per image.It will be the general trend that the "Drone Inspection" mode replaces the "Manual Inspection" mode.The DYOLO-V4 model proposed in this thesis has certain validity and feasibility.In the drone inspection of transmission-line,the model proposed in this thesis can be applied to insulators,and it can be extended to other components of transmission lines,thereby increasing the accuracy and efficiency of drone inspection.