Classification And Fault Detection Of High-speed Railway Catenary Supporting Components Based On Computer Vision

Catenary system is a crucial part of the traction power system in high-speed railways,which transmits the electricity from traction substations to electric multiple units(EMUs).Complex mechanical and electrical interactions between the pantograph of the train and the catenary may lead to faults,breakage and loosening of catenary supporting device components.These are harmful to the safe operation of high-speed trains.Therefore,accurate and efficient inspections of the working status of catenary device components are of great inportance.With the development of ‘6C' high-speed railway detection systems,the image-based non-contact defect detection has replaced traditional manual inspection as the main means of catenary maintenance.However,the currently used catenary inspection systems suffer various drawbacks such as poor understanding capacity of image contents and low degree of automation.The recognition of faults is still mainly human aided and the efficiency is still not satisfying.In order to solve these problems,a series of recognition and fault detection methods based on computer vision for high-speed railway catenary supporting devices are developed,the intelligence level of the inspection systems can be improved.1)In order to extract the catenary supporting devices from global catenary images,a recognition and localization method based on Histogram of Oriented Gradient features and modified cascade classifier is proposed.Support vector machine are trained to classify the hard samples generated in the early training stage and are used to replace the last few layers of a traditional cascade adaboost classifier.Making use of the advantages of support vector machines in dealing with high dimensional classification problems in small sample size situations,the problem of slow performance improvement of the classification model caused by the degradation of sample quality can be solved.Experiment results show that the modified cascade classifier has higher accuracy in catenary supporting device localization.2)Secondly,adopting deep learning techniques,a convolutional neural network-based object detection model is trained for catenary supporting device recognition and localization.Devices belonging to different categories can be recognized and localzed at the same time.The performances of existing object detection frameworks,such as faster R-CNN,YOLO and SSD in catenary supporting device recognition and localization are compared and analyzed.Based on the comparison results,faster R-CNN is chosen as the basic structure of our network.Modification are made to the faster R-CNN network to further improve the recognition and localization accuracy.The RoI pooling of multiple scales is performed on multiple layers and the context information of RoIs is utilized as auxiliary information in device classification.Experimental result show that the modified network can effectively improve the localization accuracy of devices of small size,such as bracing wire hooks,windproof wire rings,insulator bases and brace sleeve screws.3)A detection scheme based on local periodic anomaly is proposed for broken ceramic disc detection of insulators.This scheme can also be used to detect foreign objects clamped between insulator discs.Firstly,the local period of the texture information on the surface of an insulator is evaluated using signal processing techniques.Then the local period intensity image of the insulator is generated based on the local period evaluation result of each pixel.The diagnosis of defects such as broken ceramic discs and foreign objects is achieved by detecting the singular value on the local period intensity image.In order to detect the fractures of rotary double ears,a detection method based on 2D Gabor wavelet and distance transform is proposed.The skeleton of the extremum area of the 2D Gabor waelet intensity image is firstly extracted.Then the skeleton is projected to the standard rotary double ear image using image calibration techniques based on shape context.Finally,distance transform is used to calculate the distance between the projected skeleton and the contour of the rotary double ear and the fracture can then be recognized.4)Finally,in order to detect very fine surface cracks of rotary double ears,a detection scheme based on both image processing techniques and deep learning is developed.The segmentation of ear pieces is first achieved based on Mask R-CNN.The ear piece area is then divided into single-scale grids and a lightweight convolutional neural network is used to recoginze grids that contain cracks.Candidates of crack areas can be generated based on the recognition results.In order to eliminate the false alarms,the crack areas are finally reidentified using the coherency sensitive hashing algorithm.Compared to detection methods that only involve deep learning,this detection scheme is more efficient,when the same level of accuracy can be achieved.