Research On Rail Surface Damage Identification Based On Deep Learning

In recent years,with the acceleration of train speed,the increasing number of train services and the increasing amount of passengers and freight volume,railway transport in China is now facing a more serious situation on aggravation of rail surface wear.Therefore,it is necessary to detect the rail surface damage accurately and maintain without delay.In the past,the detection of rail surface damage was mainly operated manually.Besides,the existing research results are still not adaptable enough to the complex and changeable damage forms in the engineering environment.Many studies are still in the exploratory and trial stage.In regard of the problems above,this thesis intends to propose a date set of rail surface damage that adapted to the engineering environment.In this thesis,accurate detection of the rail surface damage is realized by improving the Deep Leaning object detection algorithm.First of all,based on rail images collected in actual engineering environment,five common types of damages including abnormal rail gap,poor light band,corrugation,fatigue crack and stripping off block,as well as polish,are counted.Rail images are screened and classified.The small sample data(polish)is augmented by the improved Deep Convolution Generative Adversarial Networks.After manual annotation,this date set of the rail surface complex damage contains 4355 images with 6832 damage annotations.In the next step,three object detection models(YOLO v3,Faster R-CNN,Cascade RCNN)are built for rail surface damage detection.Training process and detection effect are described in detail in this thesis.After comparison,Cascade R-CNN with the best detection effect is selected as the basic model.Finally,the selected algorithm is effectively optimized according to the characteristics of the rail surface complex damage data set.The backbone is changed to Res Next101 to enhance the ability of extracting rail damage features.K-means clustering algorithm and Guided Anchor algorithm are provided to optimize the Anchor to adapt to the morphological characteristics of rail damage.Besides multi-scale training/testing is also adopted to enhance the robustness of the model.These experiments indicated that the improved Cascade R-CNN is very effective when used in the field of rail surface complex damage data set.The m AP nearly reaches 90%,which increases 13.5% comparing with original algorithm,and recall reaches 93.5% as well,which increases 10.5% comparing with original algorithm.In conclusion,the improved Cascade RCNN is suitable for rail surface complex damage detection tasks in actual engineering environments.