Research On Quantitative Detection Method Of Rail Fastener Defects Based On Deep Convolutional Network

In recent years,with the continuous expansion of rail coverage,rail technology has been upgraded accordingly.Railroad line condition monitoring is one of the important tasks to ensure the safety of railroad transportation system,and the effective detection of rail fastener defects is an important means to ensure the safe operation of rail transportation.Due to the complex diversity of rail fastener defect features,traditional machine vision methods are difficult to obtain satisfactory detection results.The existing methods based on deep learning have problems such as large model size,too many parameters,low accuracy,slow speed and difficult quantitative detection.In this paper,in view of the above problems and the requirement of fast and accurate detection,we adopt deep learning technology,use supervised data enhancement technique to enhance the image of rail fastener defects,and establish a rail fastener defect detection model,on the basis of which we introduce lightweight network and quantitative detection method to detect the state of rail fasteners,which achieves the detection goals of high accuracy,fast speed and small model,effectively improving the detection efficiency of fastener defect detection is effectively improved,and the goal of quantitative detection of rail fasteners is achieved,which meets the requirements of rail fastener defect detection.The main research contents of this paper are as follows:(1)For the problem of less rail fastener fault information and unbalanced data categories.First,supervised image enhancement of fault information based on mosaic,feature fusion and random data enhancement is used to enhance the number of fault samples in the data set and make the data categories more balanced.Secondly,a rail fastener fault diagnosis dataset is constructed.Then,the Swin-Transformer based rail fastener fault diagnosis network model is built.Finally,the analysis is carried out through experimental validation to verify the effectiveness of the rail fastener fault diagnosis method based on data enhancement and SwinTransformer.(2)For the current defects such as large size of rail defect detection model and slow detection speed,it is difficult to obtain satisfactory results by traditional machine vision.Based on the study of YOLOv5 network,firstly,the depth separable convolution is used to replace the C3 module(three standard convolution and Bottleneck module)in YOLOv5,which can reduce the number of model parameters.Second,the convolutional block attention model(CBAM)is integrated into the network to find the attention region and improve the performance of feature representation in the network.Third,the original YOLOv5 loss function is changed to optimize the loss function calculation method and the network structure.Finally,the data set constructed in this chapter is able to locate and identify the detection of rail fasteners effectively,and a comparison test is conducted,which not only has higher detection accuracy as well as can ensure faster detection speed compared with other deep learning methods.(3)For the traditional deep learning-based fastener state detection method for offset fasteners detection difficulties and difficult to quantify the detection of defects and other problems,a railroad fastener state detection method based on improved mask region convolutional neural network(Mask R-CNN)is proposed.First,the image dataset is established through field test and the images are labeled.Then,the Mask R-CNN-based fastener state detection method is established using the dataset to extract the fastener location and segmentation.For normal,broken and missing fasteners,quantitative detection is carried out by calculating the mask area.Finally,for offset fasteners,the output layer of Mask R-CNN is improved using the minimum outer rectangle method to obtain fastener angle information of fasteners for quantitative detection and improve the accuracy of offset fastener detection.After the field test,the adopted method can effectively detect the offset fasteners,and the detection rate and accuracy rate are high,which can realize the quantitative detection.This paper adopts deep learning algorithm to detect track defects,and the proposed method of quantitative detection of track fastener defects based on deep convolutional network effectively solves the problems of large size of traditional track defect detection model,slow detection speed,difficult quantitative detection and existence of leakage and misdetection,which provides a new idea for track defect detection research.