Research On Defect Identification Method Of Leaky Cable Fixture In High-speed Railway Tunnel

Leaky cables are an important part of the railway communication system.Leaky cables are used to achieve wireless coverage in special areas such as railway tunnels.In order to prevent the leaky cable from falling off,which will eventually affect the normal operation of the railway,it is important to monitor the status of its fixed fixtures.Traditional manual inspection,which uses the naked eye to observe whether the fixture is damaged or loose,is inefficient and can not meet the requirements of real-time monitoring of the railway system.Therefore,a new method of identifying defects of the fixture is needed.This paper selects the target detection algorithm based on deep learning to carry out the research on the defect recognition of the fixture.By training different detection algorithms on the leaky cable fixture data set and analyzing the experimental results,the practicability and reliability of the automatic defect recognition are verified.At the same time,according to the characteristics of the data set,the algorithm is continuously improved,so that the model has a higher accuracy in the recognition of fixture defects.The research content can be divided into two stages.The first part is mainly to analyze and compare the performance of different target detection models,and select representative models in the first-stage algorithm and the second-stage algorithm for training.This article first clarifies the difficulties faced by the experiment and draws up a complete technical route based on the characteristics of the leaky cable fixture data set such as the original image quality,the size of the target object,and the fixture failure situation.Then,specific detection models were selected to identify all the fixtures in the data set,including YOLOv3 and Retina Net based on the one-stage algorithm,and Faster-Rcnn-FPN and Libra-Rcnn based on the two-stage algorithm.Experimental results show that the one-stage algorithm has a higher model inference speed,but there is a problem of redundant prediction frames.The model inference speed of the two-stage algorithm is slightly slower,but the number of predicted frames generated is closer to the number of real frames,which avoids the problem of falsely high detection accuracy.After a series of experiments at this stage,the feasibility of the target detection algorithm in the defect detection task was verified,and the foundation for subsequent improvement experiments was laid.The second part is mainly to improve the detection algorithm.From the previously trained models,select the Faster-Rcnn-FPN model that best meets the requirements of accuracy and efficiency to improve the algorithm,and try to give corresponding answers to the many problems faced by the data set.The solution ultimately improves the performance of the algorithm on the task of identifying defects in fixtures.Specific improvement measures include: combining the characteristics of the data set,using preprocessing and sample expansion methods such as equalization、sharpening 、data enhancement and using smaller anchor boxes to match the size of the fixture,reducing the effects of dim images and small fixtures;at the same time,using online hard example mining and focal loss techniques to explore detection enhancement methods for difficultto-classify fixtures;finally increasing the weight of defect categories in the classification loss function and reducing the number of fixture categories to reflect the importance of defective fixtures.Through these improved experiments,the average recall rate of defective fixtures reached 96.8%,and the average accuracy rate reached 90.4%.Compared with the basic model used in the article,the two values have increased by 6.2%and 5.9%.The performance of the defect recognition system has been significantly optimized,which can also provide better protection for the communication security of high-speed railways.