Foreign Object Detection For Ballastless Track In High-Speed Railway

In daily running of high-speed railway,the strong airflow brought up by 300km/h running speed will exert great suction on the foreign objects on ballastless track,thus causing hidden danger to running safety.Detection of foreign objects on ballastless track is of great significance to safety and track cleaning.It is costly and inefficient to detect foreign objects by ways of manual track inspection or manual track image review.How to use computer technology to detect foreign objects in track image intelligently and achieve efficiency and cost reduction is an urgent problem we need to solve at present.GANomaly is a semi-supervised anomaly detection model based on image reconstruction proposed in the past two years.It can complete the training of the model without abnormal samples and is suitable for the detection of foreign objects in ballastless track.Firstly,based on GANomaly,a set of detection methods for foreign objects in ballastless track is proposed in this paper.Then,by improving the testing method of the model,the separability of normal and abnormal samples is enhanced,so as to enhance the abnormal detection ability of the model.The main achievements of this paper are summarized as follows:(1)A method for detecting foreign objects in ballastless track based on GANomaly is proposed.The method is divided into two phases.The first stage: the target detection model is used to locate the foreground in the ballastless track image,and then the foreground is removed and the background is divided into 64 x 64 image blocks.The second stage: determine whether each image block of ballastless track image is abnormal by GANomaly,and then determine whether there is foreign objects.For abnormal image block,mark its position information.(2)A method for detecting foreign objects in ballastless track based on nearest-neighbor reconstruction and implicit space interpolation is proposed.On the basis of the GANomaly model,in order to increase the difference between the reconstruction errors of normal and abnormal image blocks and make them more separable,we improved the testing method of the model.The set of encoding vectors of all image blocks in the training set is called implicit vector set.In the test phase,the encoding vector of the test image is not decoded to obtain its reconstructed image,but the nearest neighbor vector of the encoding vector in the implicit vector set is decoded to obtain its reconstructed image,which effectively improves the reconstruction error of abnormal image block.At the same time,we also interpolate the areas with sparse implicit vectors in the implicit space to reduce the reconstruction error of normal image blocks.Compared with the original method,the new method takes the 2 norm between the encoding vectors of the test image and the reconstructed image,and the mean square error between them as the abnormal score.The AUC of the model on the image block data set is improved by 0.01 and 0.07,and the AUC of the model on the ballastless track data set is improved by 0.08 and 0.1,respectively.