Study On Deep Learning-based State Detection Method Study For Catenary Support Devices

High-speed railway pantograph-catenary(PC)systems,as the core part of high-speed railway traction power supply system,are directly responsible for the transmission and power supply of electric multiple units(EMUs).During the interaction process of the PC system,the direct mechanical contact will easily cause the catenary support and suspension devices to vibrate violently,resulting in various harm degrees of mechanical failures such as “loosebreakage-break” of the catenary components.At the same time,the stability of the catenary structure will be reduced,threatening the safe operation of the entire railway system.Therefore,research on advanced detection and monitoring methods of catenary components in service status for realizing early warning of the failure status of catenary components,is of great significance to the safe operation and maintenance of high-speed railways.In 2012,the former Ministry of Railways of China(now China Railway Corporation)formally promulgated the“General Technical Specifications for High-speed Railway Power Supply Safety Inspection and Monitoring System(6C)”,marking the official transition from the traditional manual inspection-based methods to the new stage of non-contact intelligence catenary inspection and maintenance based on computer vision technology.To this end,this article focuses on solving the actual problems of the detection system from the perspective of the integrity of the detection system’s functions.an in-depth study of the critical links in the entire detection process “Image Enhancement-Component Localization-Component Defect DetectionStructure Parameter Detection" has been carried out,and the following work results have been achieved:1)First of all,the catenary detection system mainly collects images of catenary facilities during the night line skylight period.Due to the problems of the on-site environment and its own supplementary light system,the collected images often suffer from uneven illumination and low illuminance,causing catenary components not to be effectively detected and accurate failure analysis for the components not to be performed.Therefore,a catenary image enhancement method based on traditional Retinex illumination decomposition theory and unsupervised learning is proposed.The main idea of the method is that an Retinex-based image decomposition network is built to decompose the illuminance map and the reflection map.A self-attention mechanism is introduced to build an unsupervised auto-encoder denoising network to reduce the noise in the reflection map for solving the problem of noise interference generated during the image enhancement process;An illumination enhancement network based on an unsupervised generative adversarial networks(GANs)is built to achieve illumination enhancement of low-illuminance images without paired samples,solving the problem that the paired bright and dark illumination images cannot be obtained in actual detection.The experimental results show that this method can well realize the low-illumination image enhancement of the catenary,which is convenient for the identification and localization of components.2)Secondly,to realize the defect detection of various catenary components,it is first necessary to accurately detect each catenary component in the catenary image.However,the complex structure of catenary support devices,the variety of components,the large difference in scale,the existing deep learning detection network framework cannot realize the effective detection of all types of components,especially the detection of small-scale components.Therefore,an improved Faster R-CNN(Region-based Convolutional Neural Network)object detection network for catenary components based on cascading ideas is proposed.The main idea of the method is that the feature extraction network of the original Faster R-CNN is optimized to filter out the optimal feature expression layer of catenary components by analyzing the intermediate feature maps,so as to achieve the precise localization of the largescale components;and by utilizing the installation structure relationship between the largescale components and the small-scale components and sharing the shallow semantic feature layer extracted by the Faster R-CNN feature extraction network,a small-scale component detection network is constructed and integrated into the large-scale component detection network,forming a multi-scale catenary component object detection network under a unified framework.The experimental results show that this method can accurately and quickly realize the detection of all kinds of catenary components with multiple scales.3)Then,in order to accurately detect the state of catenary components,it is necessary to collect the fault samples of each type of components to meet the research of fault detection algorithms.Ideally,when fault samples are enough,fault detection algorithms based on supervised learning are usually simpler and more efficient.However,it is often impossible to obtain enough fault samples in actual field.Therefore,in view of the lack of defective samples,component fault status cannot be directly diagnosed by the supervised learning-based methods.a detection method of the catenary insulator contamination defect based on unsupervised autoencoder networks is proposed.The main idea of the method is that an image segmentation network is adopted to segment the catenary components to avoid the background interference of non-target components;and the functional characteristics of the autoencoder network,that the target reconstruction and the target feature vector extraction,are utilized to construct a component classification and reconstruction network based the unsupervised autoencoder for achieving accurate classification of the catenary components and fault extraction;then the density-based spatial clustering of application with noise(DBSCAN)method is used to evaluate the fault level of insulator contamination.The experimental results show that this method can accurately and effectively realize the detection and evaluation of the fault state of insulators.4)Finally,because the catenary support device is frequently subjected to the vibration shock from the pantograph,the catenary components are prone to loosening,resulting in deformation of the cantilever system structure of the catenary support device,which in turn leads to the change of the steady arm slope and affects the safety of train power supply.However,it is difficult to effectively detect the structural parameters of the catenary cantilever system under the background of the complex catenary detection line.A method for detecting the structural parameters of the cantilever system of catenary support devices based on a three-dimensional(3D)point cloud is proposed.The main idea of the method is that using the depth information of the 3D point cloud,the 3D convolutional neural network is applied to achieve efficient and robust segmentation of the catenary component regions,avoiding the problem of occlusion in the non-physical connection area of the components in the two-dimensional(2D)image of the catenary;and according to the segmentation result of the components,a rapid random sample consensus(RANSAC)plane detection method is established to realize the plane detection of the catenary cantilever systems.Based on the plane detection results,a RANSAC space straight-line detection method based on the projection plane is built to achieve the axis extraction of the skeleton components of the catenary cantilever system and the precise detection of the structural parameters.The experimental results show that this method can overcome the background interference of the detection line and achieve the high-precision detection of the structural parameters of the cantilever system of catenary support devices.Aiming at the actual problems of the detection system of the catenary support device,the thesis systematically conducts in-depth research on key issues and puts forward feasible theories and technical methods which provides important support and reference for practical engineering applications and also provides the necessary theoretical foundation and research ideas for the researchers in the industry to conduct further in-depth exploration and research.