Research On Dynamic Real-time Detection Algorithm And System Integration Of Rail Fastener Defects

With the rapid development of rail transit in China,more and more attention has been paid to the routine maintenance and repair of rail lines to ensure public safety.With the development of rail vehicle technology and the maintenance pressure brought by the increase of speed and mileage,the rail fastener is becoming more and more important in maintaining the stability of the rail line.Rail fastenings play an important role in the connection and fixation of rail and track bed,and in the maintenance of track spacing.At present,the detection method based on artificial detection in the daily maintenance of track fastener has been unable to meet the needs of intelligent detection of track line.The human eye can only conduct visual inspection for the damaged fastener,but can not evaluate the buckle pressure of the fastener.The real-time detection system of rail fastener defects researched in this paper provides a reliable and stable detection method for intelligent detection of rail lines Firstly,this paper discusses the method of fastener defect detection,and designs the detection method and process.According to the detection target and demand,the advantages and disadvantages of different shooting schemes are analyzed,and the reasonable hardware selection and design of image acquisition system are carried out.This paper expounds the related technologies of image processing in fastener damage detection task,and puts forward a target detection method based on deep learning technology which is suitable for rail fastener detection task.In order to improve the real-time performance of the detection system,a method combining image classification and image target detection is proposed.In the task of image classification,the fastener image sequence is analyzed and the data set making,model network structure,model parameters and optimization methods are analyzed The data set is expanded by using the method of image data augmentation,and the robustness of the model is greatly improved by introducing the natural image.In the aspect of model optimization,the reasonable convolution kernel size,activation function and optimization method are discussed,and the parameters that meet the needs of detection task are selected through targeted design experiments.Finally,the output of the model is optimized to avoid possible misjudgment.In the task of target recognition,the types of fastener damage and image features are analyzed to mark the fastener image and train an efficient target detection network to identify the marked target.The damage of fastener is summarized as four features of fastener image,and a detection method which can not only detect the missing damage of fastener,but also analyze and predict the withdrawal of fastener is designed.At the same time,the robustness of the detection model is tested and the output judgment conditions of the detection model are optimizedFinally,according to a certain process and method,the above fastener detection model is deployed to the hardware platform.In different environments,the fracture,missing and demoulding of the fastener strip are tested.The analysis of the test results shows that the real-time detection system of the rail fastener studied in this paper meets the design requirements.