Pillar Number Identification And Few-shot Foreign Matter Detection In High-speed Rail Catenary Based On Deep Learning

The high-speed rail catenary pillar number is the only identification for the pillar under the "one pillar,one file" system,and its automatic and accurate identification is the basis for catenary system prognostics and health management.Meanwhile,as the high-speed rail catenary network is exposed to outdoor work,it is vulnerable to be invaded by foreign objects,which threatens the power supply safety.At present,the main method of foreign objects inspection is manual inspections by inspectors,which are costly and inefficient.In addition,the position of the pillar where the foreign objects appear needs to be manually recorded afterwards,and cannot be automatically detected and located.The current research is based on the images collected by the high-speed rail catenary 2C system,using image processing methods to automatically detect foreign objects.After studying the existing foreign object detection methods,we found the following problems to be solved:(1)The size of the detected image is large and the background is complicated,but the traditional image processing method which is used to detect foreign objects is poor effect and low efficiency.(2)Compared to normal images,the foreign object image has a very small number of samples compared to normal images,and does not have the condition to use conventional deep learning detection methods.(3)Foreign objects detection only detects the position of the foreign object in the image,but not automatically identify the position of the pillar where the foreign object appears in the line,which is not convenient for guiding the inspectors to repair;(4)The detected images are continuously,thus processing all the images will lead to detection repeated.In view of the above problems,we research a few-shot problem of high-speed rail catenary foreign objects detection.The images collected by the catenary safety inspection device(2C device)is used as the research objects.Computer vision-based technology detects the foreign objects and locates the positions of the pillar where the foreign objects appear.Main tasks as follows:First,for the small size of the pillar number in the 2C detection image,after comparing commonly used convolutional neural networks and target detection algorithms,the YOLOv2 algorithm framework is used.Deepen the network on the basis of the Darknet-19 backbone network and introduce residual structures to avoid degradation.Improve the network structure for multi-scale feature fusion,and combine specific a priori anchor boxes with dynamically adjusted input methods to achieve precise positioning of a single small pillar number image;Then,based on the catenary pillar number image,an image sequence identification model is proposed to identify pillar number characters of indefinite length,and the result is corrected by using the continuous characteristics of the front and back frame pillar number.Meanwhile,the abnormal plates are detected.Next,in view of the small number of foreign objects in the catenary system,the idea of "foreign objects potential regions positioning and regional image classification" is used to locate the foreign object Regions of Interest(Ro I),and then a meta-learning method based on similarity measure learning is used to determine whether the image has foreign objects.Furthermore,a meta-image-enhancement network is used to supplement foreign object samples to improve classification accuracy.Finally,due to the bottleneck of the accuracy rate of a single classification method,the feature design classifier of machine learning is used to supplement.The line features are extracted from the Ro I and classified using support vector machines(SVM).Then,the metric learning classifier and features-design classifier are stacked.Heterogeneous integration of classifiers is designed and factor screening is performed to obtain an integrated classification scheme suitable for detection of foreign objects in catenary system.The experimental analysis of the actual images verify the effectiveness and feasibility of the method proposed in this paper.