Research And Application Of The Learning Method Of Insulator Detection And Location For High-speed Railway Line Inspection

Insulator plays a role in supporting overhead transmission line of high-speed railway and preventing current from returning to ground.At present,ceramic insulator is mostly used in overhead transmission line of high-speed railway.Ceramic insulator will be damaged or become faulty insulator when exposed to damp or lightning,which seriously affects the safety of high-speed railway transportation.Therefore,faults detection of insulator is particularly important in the high-speed railway line inspection.At present,the mainstream object detection algorithms are divided into two-stage object detection algorithm and one-stage object detection algorithm.Object detection algorithms based on deep learning has become a hot technology because of its powerful feature representation ability,the accumulation of data and the progress of hardware computing power in recent years.Both two-stage and one-stage object detection algorithms are widely used in many fields,such as unmanned driving,visual search and image understanding and so on.Aiming at the problem of insulator detection and location,the object detection algorithm based on deep learning is used to study insulator detection and location in this thesis.Aiming at the accuracy of insulator detection,in this thesis,we first improve the Faster R-CNN feature extraction network VGG16.On the basis of maintaining the first seven feature extraction layers of VGG16,the asymmetric convolution kernel and convolution kernel combination are introduced to form a high-level feature extraction block to replace VGG16’s deep feature extraction layer.Finally,the depth of feature extraction network is reduced and the width is widened.The test result shows that the new feature extraction network improves the accuracy of insulator detection in complex scenes.Secondly,we further use K-means++algorithm to count the cluster results of insulator label boxes.Through setting different values for the cluster centre of K-means++,and generating the initial cluster centre randomly for many times under different cluster centres.Finally,three cluster centres are selected through comparative analysis,and the size and proportion of the anchor box are set reasonably.The comparative test shows that the position of insulator can be more accurately identified after resetting the anchor box,and the detection accuracy of insulator is improved.At last we add the Ro I Transformer to the network.The rotation parameters are obtained through training,and the horizontal Ro I is transformed into the rotated Ro I.Finally,the rotated rectangular bounding box is used to identify the inclined insulator.On the basis of increasing less calculation,the positioning accuracy of the inclined insulator is improved.Aiming at the detection speed of insulator,in this thesis we improve the Faster R-CNN feature extraction network VGG16,and changes the original VGG16 feature extraction network into VGG7-Inception network with fewer layers.Compared with VGG16 network,VGG7-Inception network has fewer parameters and faster feature extraction speed.In addition,asymmetric convolution kernel,cascaded convolution kernel and 1 × 1 convolution kernel are used in VGG7-Inception to reduce the network computation and shorten the insulator detection time.This thesis verifies the improved algorithm in four insulator detection scenarios.Compared with the original Faster R-CNN,the average relative detection time of our algorithm is 61%,the detection accuracy and recall rate are 98.7% and 98.6% respectively.The results show that the improved algorithm based on Faster R-CNN finally improves the detection accuracy and reduces the detection time of insulator.