Research On Bolt Loosening Recognition Of Medium And Low Speed Maglev F-rail Based On Deep Learning

With the continuous advancement of urbanization in China,the traditional wheel-rail contact urban transportation system can not fully meet the growing demand for passenger transport.As a new urban rail transit system,the F-rail of the medium and low speed maglev system which is set up in mid-air has large slope and small bend,and can be easily installed,disassembled and maintained.The F-rail can be set up in densely populated large and mediumsized cities,effectively improving the urban public transport.As the running track of the medium and low speed maglev trains,the bolted F-rail provides levitation force and guidance force for the train,and ensures the safe running of the train on the rail.The nut loosening fault for F-rail bolts easily occurs due to the gravity of the nuts of F-rail bolts and the vibration caused by the running of maglev trains.Therefore,it is necessary to detect and monitor the status of F-rail bolts.At present,the main methods of detecting the status of Frail bolts are to inspect the F-rail and view the images of F-rail artificially.In this paper,image processing algorithm based on deep learning is applied to the dataset of mass high-definition Frail images collected by medium and low speed maglev detection vehicle to automatically identify the nut loosening fault of F-rail bolts,which can effectively reduce the missing detection of loose F-rail bolts caused by human factors.Firstly,the characteristics of F-rail image data set are studied.The image of F-rail with the not completely occluded chassis of the detection vehicle is determined as the key frame image of F-rail.Convolutional neural network ZFNet is used to classify the F-rail images into two categories,and the key frame image of F-rail is selected.Then,The key region of F-rail bolts is determined as the region including F-rail bolts located on the front side of the same steel pillow,the convolution neural network Faster R-CNN is trained on the dataset of F-rail key frame images.and the key regions of F-rail bolts are located and segmented from the images of F-rail key frame.Finally,the convolution neural network DeepLabv3 is trained on the key region images of the F-rail bolts.The nuts and screws of the F-rail bolts are segmented semantically into two different categories,and the nuts and screws belonging to the same bolt are judged.The nut loosening fault is identified by counting the number of threads on the top screw of the fastening bolt near the camera.For connecting bolts near the camera and the fastening bolt away from the camera,Gabor features of the corresponding regions in the key region image of F-rail bolts and their template images are matched to identify their nut loosening faults.The number of connected areas in binary images of the nuts of connecting bolts away from the camera is counted to identify their nut dropping faults.