Design Of Real-time Detectionsystem For Rail Fasteners Based Ondeep Learning

Track fasteners,as the connecting parts between sleepers and rails,are very important in ensuring track stability and reliability,and are the focus of track line maintenance.With the development of computer technology,machine vision and artificial intelligence technology are gradually being applied to track maintenance work.In order to further improve the efficiency of track fastener maintenance,this paper designs a track fastener detection system based on deep learning,which can realize the real-time and accurate detection of fasteners at the train’s maximum speed of 350km/h,and completes the construction and detection of the system hardware platform.The design of the algorithm and the realization of the program,the main contents are as follows:(1)Determine the overall structure,various components and workflow of the track fastener detection system.Carry out demand analysis and parameter calculation for the camera,lens,light source lighting and mechanical structure to provide a basis for the construction of the image acquisition platform.According to the characteristics of fastener status detection,the detection task is divided into two parts: abnormal fastener detection and abnormal fastener classification.The applicability of the main algorithms is analyzed,the algorithm research route is determined,and a parallel processing is designed.method.In addition,to solve the problem that abnormal fasteners cannot be positioned in fastener abnormality detection,a sleeper-assisted positioning method is proposed.(2)Starting from the theoretical knowledge of deep learning,summarize the commonly used improvement methods of neural network models.Several well-performing deep learning target detection algorithms have been constructed,and fastener anomaly detection experiments have been carried out.Aiming at the problems of a slightly lower recognition rate of existing fasteners,detection speeds that cannot meet real-time requirements,and general positioning accuracy,yolov4 Based on the tiny algorithm,improvements have been made in target box clustering,network structure and loss function.The experimental results show that the various indicators of the improved algorithm are significantly improved and can be applied to the system.(3)A classification algorithm for abnormal fasteners based on morphological operations is proposed,which specifically includes: expansion of fastener edge information,contour extraction,contour area calculation and sorting operations,establishment of area threshold judgment relations to realize fastener missing state classification;center-based The partition method of bolts is used to extract the contours and area calculations of part of the spring bar images,and establish the area threshold discriminant relationship to realize the classification of the damage state of the fasteners;use the rectangular fitting method of the left and right spring bars to obtain the angle information of the left and right spring bars,and establish the angle Threshold judgment relational expression realizes the judgment of the twisted state of the fastener.Through confirmatory experiments,the effectiveness of the algorithm is proved.(4)Choose the camera,lens,light source lighting and data processing equipment suitable for this system,and complete the construction of the mechanical structure and the configuration of the software environment.In addition,the system program is modularized design,including image acquisition module,fastener detection module and output module,and the code realization of the overall program is completed.(5)System verification experiments are carried out,and the results show that the camera,lens,light source lighting and data processing platform of this system can meet the working requirements,and the system program can meet the real-time and recognition rate requirements of fastener detection under the train running state of up to 350km/h.