Study On Detection Algorithm Of Metro Sleeper Defects Based On Deep Learning

Sleepers are a crucial part of metro rail structure,which maintain the geometric position of rail,bear the load from all directions of rail and transmit it to ballast.During the operation of metro,sleepers will be constantly subjected to schock,together with the influence of construction quality faults and environmental factors,then that may lead to cracks and fal ing blocks of ones.These are harmful to the safe operation of metro.With the development of rail state inspection systems,the image-based noncontact detection method gradually replaces the traditional manual inspection as the main means of rail maintenance.However,the existing detection systems still suffer some shortcomings such as poor image resolution ability and low automation level.The recognition of defects is still mainly human assistance and the efficiency is still not satisfactory.In order to solve these problems,the thesis studies the detection algorithm of metro sleeper defects based on deep learning technology.The main work is as follows:1.In order to extract the sleeper defect areas from global metro rail images,a recognition and location algorithm based on improved TINY-YOLO v2 network is studied.Firstly,the input size of the original network is adjusted based on the size of sleeper defect areas.Then the rail image data set is clustered using K-means algorithm.Finally,the complexity of the original network is reduced by compressing the number(width)of convolution kernels at various levels in the one.Experimental results show that the improved TINY-YOLO v2 network has higher efficiency in sleeper defect areas localization.2.In order to detect the subtle cracks on the surface of metro sleepers,a detection algorithm based on improved YOLO v3 network is developed.Firstly,referring to the dense connection theory of Dense Net network,the first two residual modules in YOLO v3 network are replaced by dense connection modules.Secondly,inspired by the improvement of prediction module in DSSD network,the one of YOLO v3 network is optimized,and then the deep part of YOLO v3 network is compressed.Finally,data augmentation technology is adopted to augment the quantity of crack images,and the model detection effect is analyzed based on data visualization technology.Experimental results show that the algorithm can significantly improve accuracy and speed of crack detection.3.A detection algorithm based on the improved Darknet-53 classification network is studied for fal ing blocks detection of metro sleepers.Considering the difficulty of detecting sleeper fal ing blocks and the complexity of the model,the number(depth)of residual units of each residual block in the original Darknet-53 network is adjusted to prevent over-fitting,and the feature dimensions of its full connection layer is matched according to the amount of detection categories.Experimental results show that the algorithm can achieve high detection accuracy and speed.Experimental results show that the sleeper defect detection algorithms studied in this thesis can realize rapid and accurate recognition of metro sleeper defects,solve the problems of low robustness and low detection efficiency of traditional image recognition algorithms,and meet the real-time requirements of rail inspection system under the top speed of 160 km/h.