Research On Accurate Extraction And Defect Recognition For Catenary Components Based On Deep Learning

Pantograph-catenary system is an important part of the high-speed railway traction power supply system,whose function is to transmit electric power to the Electric Multiple Units(EMUs).The practical project operations showed that defects such as “looseness,detachment,fracture,and crack” may be raised on the components of the catenary support device due to the continuous impact caused by the interaction between pantograph and catenary,and the influence of the external environment.These defects can lead to the decline of catenary structural reliability,and seriously affect the stable operation of pantograph and catenary system.Therefore,it is of great significance to find the defects in time and provide an early alert in advance to ensure the safe operation of the high-speed railway and improve the catenary maintenance strategy.In 2012,the former Ministry of China Railways implemented the “General Technical Specification for High-speed Railway Power Supply safety Inspection and Monitoring System(6C)”,and put forward the requirements of using image processing techniques to realize the intelligent inspection for catenary components.The computer visionbased approach for component inspection mainly includes two steps: component extraction and defect recognition.For the component extraction,the existing methods have the problems of missing extraction and low extraction accuracy.For defect recognition,lack of defect samples is always an issue that brings difficulty to develop effective defect recognition methods.Aiming at addressing the above problems,this dissertation takes “ positioning the main components simultaneously--> refining the positioning results(including refine the boxes that are produced by horizontal box positioning,refine the boxes that are produced by oblique box positioning,and refine the boundaries that are produced by segmentation)--> general method for defect recognition when lacking defect samples” as the research route,and proposed corresponding solutions.The main works of this dissertation are as follows:(1)Firstly,aiming at addressing the problem of simultaneous positioning of main components,a horizontal box positioning model that integrated DCNN(Deep Convolutional Neural Networks)and transformer is adopted.There are many kinds of components in the global image of the catenary.As the scale difference of these components is large,and some components are overlapped,the existing methods can not effectively locate all kinds of components at the same time in such complex images,especially for the small components and the overlapped components.To address this issue,the proposed model uses DCNN to extract global image features first,then the information of each position in the feature map is associated through the encoder of the transformer,which strengthened the position relationship between different components.In addition,the optimization strategy of the one-to-one pairing of the component target and the predicted box is adopted to reduce the missing extractions of the components.The experimental results show that the proposed method can realize the simultaneous positioning of 12 types of catenary components,and significantly reduce the missing extraction cases for the components such as Screw and Isoelectric＿line.(2)Secondly,the issue of improving the extraction accuracies of components is addressed.As the shapes of components are very different,for example,the Insulator is inclined and the Isoelectric＿line is linear.Thus,using the horizontal box extraction approach will bring serious background interference.According to the characteristics of the components,three extraction means namely horizontal box positioning,rotating box positioning,and segmentation,are adopted in this dissertation.Furthermore,three box-refinement(or boundary-refinement)methods are proposed accordingly.1)For the horizontal box positioning,a deep reinforcement learning-based method is proposed to refine the horizontal box.The Diagonal＿brace＿sleeve component is selected as the analysis object.First,the reinforcement learning system is constructed.By setting the scale changes,position changes,and aspect ratio changes of the horizontal boxes as the Action elements,as well as the deep convolution neural network as the Agent element,the key elements of the reinforcement learning system are built.Through reinforcement learning training,the agent learned the action selection strategy.It means the agent can select the appropriate actions that the horizontal box would take,according to the changes of image semantic information inside the horizontal box.As a result,the horizontal box can be dynamically adjusted to an accurate position by taking continuous actions,Experimental results show that the proposed method can effectively refine the horizontal positioning box of the diagonal brace sleeve,and improve the accuracy of the horizontal box positioning.2)For the rotating box positioning,the generative adversarial network(GAN)based method is proposed to refine the rotating positioning box of the component.The inclined Insulator component is taken as the analysis object.First,the RRPN(Rotation Region Proposal Networks)network with a cascaded regression module is proposed as the preamble basic subnetwork to obtain the initial rotation positioning box.Then,a series of candidate rotated boxes are generated around this initial positioning box.Furthermore,reconstruction errors of the candidates could be calculated by the GAN model which was trained by standard Insulator components,and the candidate box with the minimum reconstruction error is selected as the final rotation positioning result,which realized the accurate extraction of the inclined catenary component.The experimental results show that this method can effectively improve the accuracy of inclined Insulator positioning.3)For the segmentation,the Cascade PSP(Cascade Pyramid Scene Parsing Network)based method is proposed to refine the segmentation boundary.The Isoelectric＿line component is taken as the analysis object.First,two representative deep learning segmentation networks,namely Mask R-CNN and YOLACT(You Only Look At Coefficien Ts)are used to segment the Isoelectric＿line in the local image,and the YOLACT model is selected as the basic segmentation network by comparing their segmentation accuracies.Then,concatenate the coarse mask and the original local image into one input,and build a model called PSPNet that can map the input to a better mask.By cascading several PSPNet,an accurate mask could be obtained gradually.Experimental results show that this method can effectively improve the segmentation accuracy of isoelectric＿line.(3)Finally,the problem of component defect recognition under the condition of lacking defect samples is addressed.The existing methods based on image reconstruction error have some issues,such as only one type defect being focused,unreasonable anomaly measurement,and the reconstructed image is blurred.From the perspective of similarity matching,combined with the idea of “anomaly detection based on reconstruction error”,this paper proposes a component defect recognition method based on feature similarity matching and reconstruction error.Firstly,a coreset that only includes normal image features is constructed to represent the normal mode,and the similarity distance between the test feature and the coreset is used as the anomaly score of the image area where the feature is located.Then,the Teacher-Student network is constructed,and the student network is trained with normal component images to make it only learn the ability to reconstruct “normal features”;During the testing,the student network will not be able to reconstruct the abnormal features,which makes the reconstruction error of the abnormal area relatively large;As the reconstruction error cannot strictly measure the degree of abnormality,it is not suitable to be used as the anomaly score,but it can be used as an auxiliary basis for component defect identification;According to the characteristics of different defect types,this paper selectively uses “feature reconstruction error” to enhance the contrast of abnormal score obtained by similarity matching.Finally,the anomaly score is used to judge the status of components.Experiments show that the proposed method can effectively recognize the defects,such as the looseness of Nut,the damage or dirt accumulation of Insulator,and the dispersion of isoelectric＿line.This dissertation established a preliminary framework for catenary component detection system of “simultaneous extraction of components--> extraction refinement--> defect recognition under the condition of lack of samples”,and proposed a series of accurate extraction methods and an effective defect recognition method for catenary components,which can provide technical reference for power supply safety detection and monitoring system of high-speed railway.