Research On Automatic Recognition And Positioning Of Rail Fasters

Rail fasteners are important railway facilities that connect rails and sleepers.They are susceptible to impacts during train operation,resulting in loosening,damage,or even loss of fasteners,and safety accidents such as train derailment.Therefore,the condition of the rail fasteners must be regularly checked and maintained.In the maintenance work of rail fasteners,most of the fasteners are first loosened and then tightened to achieve fastening,and they rely on manual installation.Therefore,it is most important to develop a system that can replace manual rail fasteners for automatic disassembly and assembly.One of the key points in the research of this system is to complete the identification and positioning of rail fasteners.Combining the current development status of rail fastener identification and positioning technology,as well as the challenges in the railway environment,this study takes elastic Ⅶ rail fasteners as the research object,and adopts two-step method of rough positioning and fine positioning to realize automatic identification and precise positioning of bolts in rail fasteners.The first step adopts rough positioning to quickly identify and locate the approximate position of the rail fastener.Since there is no public data set,this study uses data enhance-ment and annotation methods to establishe the dataset of rail fasteners.The identification and positioning models of rail fasteners were established based on Fast-RCNN,SSD and YOLOV4 algorithms,and the dataset of rail fasteners was taken as the sample for compar-ative experimental analysis.The results show that the YOLOv4 algorithm performs best on the dataset of rail fastenings,with an average accuracy of 97.11% and a detection speed of 28.66 FPS.However,there is a lag phenomenon in video recognition.To this end,this study makes improvements from three aspects: firstly,the Mobile Netv3 network based on deep separable convolution is used to replace the CSPDark Net53 network based on the original YOLOv4 algorithm;then K-means Ⅱ clustering algorithm is used to select the optimal anchor frame;finally,the prediction layer of YOLOv4 was optimized,and the13*13 feature layer was removed to reduce the semantic loss of the network model in the training process.Experiments show that compared with the original YOLOv4 algorithm,the detection speed of the improved YOLOv4 algorithm is higher(11.96FPS)than that of the original YOLOv4 algorithm without loss of the original accuracy,which can meet the requirements for the identification and rough positioning of rail fasteners.The second step adopts fine positioning,which is based on rough positioning to fur-ther locate the precise position of the rail fastener.Based on the geometrical characteristics of fastener bolt in special,this study first uses histogram equalization to improve the qual-ity of the rail fastener image,and uses the improved median filter to denoise.Additionlly,we use the improved canny operator to complete the image edge extraction,and perform the hough transform on the discrete edge information to obtain the center coordinates of the bolt.In this study,the identification and positioning experiments of rail fasteners are completed on ballastless and ballasted tracks respectively.The studies have shown that for ballastless track,the average positioning error of rail fasteners is 5.28 mm,and the maximum positioning error is 6.33 mm.For ballast track,the average positioning error of rail fasteners is 6.25 mm,and the maximum positioning error is 7.42 mm.It can well meet the requirements of positioning accuracy of disassembly system.