Research On Detection Of Major Defects In Transmission Lines Based On Deep Learning

With the rapid growth of the national economy,the transmission lines witch China required should be higly stable in operation to ensure the normal operation of social production.At present,unmanned aerial vehicle(UAV)is used to assist in manual inspection of transmission lines in China,and its level of intelligence is low.In order to improve the intelligence level of patrol inspection,the insulator self-explosion and bird’s nest in the transmission line are taken as defect detection objects,and the deep neural network with high detection accuracy and fast speed is intended to be explored and designed to detect the defects of the transmission line.The work done is as follows:Firstly,aiming at the requirements of high detection accuracy and fast speed,the MobileNetV2 network is proposed as the skeleton network of YOLOv3,which can speed up the operation of the network on the premise of ensuring the detection accuracy.In order to accelerate the detection speed and improve the detection rate of self-destructive defects,separable convolution is proposed to replace the ordinary convolution in the multi-scale feature pyramid structure to reduce the network computation and increase the network depth to improve the detection speed and accuracy.Secondly,in order to analyze the performance of several optimal bounding box regression loss functions in neural networks,simulation experiments are used to simulate the regression process.Experiments show that the convergence speed of the loss function of IoU(Intersection over Union)and GIoU(Generalized IoU)is low.And there are even some problems can result in non-convergence.The DIoU(Distance-IoU)loss function degenerates into IoU-Loss when the center points of the two frames coincide,and the DIoU and CIoU(Complete-IoU)loss functions cannot converge when the two frames are far apart.In order to solve the above problems,a new loss function ZIoU is proposed by adding the distance information between the horizontal and vertical directions of the center points of the two frames to the GIoU-Loss loss function.Through the analysis and verification of experiment,ZIoU-Loss is a measure reflecting the positional relationship between two frames,which is non-negative,positive qualitative,symmetrical and satisfies triangle inequality.The convergence speed,regression accuracy and robustness of ZIoU-Loss are better than the above loss function,which can improve the detection rate and positioning accuracy of the network.Finally,in order to enhance the generalization and anti-interference ability of the network,the original data set is expanded by using data enhancement method.After analyzing the functional images of ZIoU and IoU indexes,it is found that ZIoU’s IoU loss is compressed.According to compression mapping and ockham’s razor principle,a improved NMS(Non-Maximum Compression)algorithm is proposed to remove redundant detection frames to improve detection rate.Through test set verification,the results show that the improved YOLOv3 network has higher detection rate,more accurate location and fasterspeed for he detection of transmission line defects.The public data set verifies that the algorithms of ZIoU-Loss and ZIoU-NMS can improve the detection rate and positioning accuracy of the network,and can be applied to other detection tasks.