Defect Recognition Of Catenary Dropper Based On Deep Learning

With the rapid development of electrified high-speed railways in China,the safety and stability of catenary equipment has attracted more and more attention.As an important part of railway infrastructure,the operation of catenary equipment directly affects the normal operation of high-speed train.As one of the important components in the contact suspension,the catenary droppers installed between the catenary wire and the contact wire improve the sag and elastic uniformity of the contact line,adjust the working height between the contact line and the rail surface,and improve the current extraction quality of the electric locomotive pantograph.Due to the long-term mechanical vibration of the vehicle,the wind and the sun,the dropper components will be abnormal,which will affect the normal operation of the highspeed railway.Therefore,it is particularly important to identify the defects of catenary droppers.At present,intelligent algorithms have been applied to the fault detection of droppers,but the accuracy rate is low and needs manual review,which is time-consuming and labor-intensive,and has certain limitations.Therefore,this thesis studies the intelligent identification algorithm of catenary droppers defect detection,in order to improve the accuracy of identification,provide guidance for the maintenance of catenary,and eliminate potential safety hazards.The main research contents are as follows:First of all,this thesis takes the catenary droppers image collected by the 4C system(the catenary suspension state detection and monitoring device)as the research object,and preprocesses the dropper images.Aiming at the problem of poor picture quality caused by factors such as illumination,high-definition camera viewing angle,and reflection of metal components during the image acquisition process,the histogram equalization that limits the contrast is used to enhance the image and improve the visual effect of the image;aiming at the motion blur phenomenon in the process of image acquisition,wiener filtering is used to restore the motion blurred image;catenary droppers are outdoors all the year round,and are easily affected by rain and snow weather factors.However,few such samples are collected from the site.Since the salt and pepper noise is a discrete white or black pixel in the image,salt and pepper noise is added to simulate this environment.Because the images collected from the scene are very limited,rotation,random brightness changes,random noise addition,and image symmetry transformation are used to increase the number of samples and improve the quality of algorithm training.Then,this thesis takes Faster R-CNN as the main algorithm to deal with several defect states of droppers in the way of divide and conquer.For the defects of broken and obvious relaxation of the droppers,the algorithm is used to detect directly.For the droppers that may be normal,the algorithm is used to locate them from the original images,and then the located dropper area is cut out from the original image,to reduce the interference of other backgrounds,in preparation for further detection of the parts of the droppers.Aiming at the problem that Faster R-CNN has poor recognition effect on small objects in catenary parts,it is proposed to use Feature Pyramid Networks(FPN)and K-means algorithm to improve Faster R-CNN to complete target recognition of dropper clamp and chicken heart rings.The residual network Res Net101 is used to replace the VGG16 network in the original algorithm,so that the extracted features contain richer information.Then,the features at different scales are fused with the feature pyramid network to obtain the multi-scale features with rich semantic information.Then,the K-means algorithm is added to cluster the data,and the obtained clustering center is used to replace the anchor frame size set by experience in the original algorithm,so as to optimize the size and aspect ratio of the anchor frame,Finally,the positioning of dropper clamp and chicken heart ring is realized.Through simulation verification,compared with the original Faster R-CNN algorithm,the positioning accuracy of chord clamp is improved from 89.4% to 94.7%,and the positioning accuracy of chicken core ring is improved from 85.6% to 95.3%,which lays a foundation for the subsequent defect identification of dropper clamp and chicken core ring.It lays a foundation for defect identification of chord clamp and chicken core ring.Finally,the positioned dropper clamp is cut out,the data set is establish,and the state of the dropper clamp nut is classified and identified.Since the attention mechanism imitates the human information processing process,it focuses more attention on the region of interest in the image and ignore other parts,so this thesis proposes to embed the channel attention mechanism(SENet)network into the Inception Res Net-V2 classification network.SENet network pays more attention to the feature channels.By automatically learning the weight information of each channel,the weight coefficient of the channel related to the target is increased,and the weight coefficient of the irrelevant feature channel is weakened,so as to enhance the expression ability of the extracted feature map.It can automatically classify the missing,loose and normal status of the catenary dropper clamp nut efficiently and accurately,and achieve the effect of computer-aided inspection.The simulation results show that the classification accuracy of the three states of suspension wire clamp is high,and the average accuracy is 96.61%,which lays a necessary foundation for the defect detection task of catenary parts.Based on the above research,this thesis uses the self-made data set of catenary dropper images collected by 4C device to verify the algorithm in this thesis.Through the test of the actual images,the algorithm in this thesis can realize the accurate positioning of the chicken heart ring,and effectively identify the defects of the catenary droppers broken,obvious relaxation and the missing and loose of the dropper clamp nuts,which provides certain guidance for the maintenance of the catenary.