3D Detection System Integration And Recognition Algorithm For High-speed Railway Fastener Defects

With the rapid development of high-speed railway in China,the total length of high-speed railway is more than 25,000 kilometers,ranking the first in the world.Track structure is an important foundation to offer the smooth and safe operation for high-speed trains.The railway fasteners are the key components to ensure the reliable connection between rail and sleeper,which can prevent the movement of the rail against the sleeper and relieve the impact of the train on the track under the action of train load.So fasteners play a critical role to guarantee the overall stability of the track system.However,with the extension of service time of railway,the initial defects of fasteners develop rapidly under the influence of high frequency train load and complex environment.These defects of railway fasteners such as elastic clip looseness and break are more common,which will make the condition of track worse and influence the operation safety of trains.The traditional track inspection is performed by workers in the night,which is high cost and low efficiency.With the rapid development of nondestructive testing technology,the non-contact detection technology based on computer vision and digital image processing is gradually applied in the railway detection field.The application of the railway fastener defects automatic detection system will greatly improve the working efficiency of railway defects inspection and provide a good tool to improve the safety of trains.On the basis of summarizing the research progress of fastener defects detection,this paper comprehensively researchs the images acquisition system,fastener location,elastic clip segmentation,the extraction of stable feature and the classification and identification of fastener images.The aim is to develop a railway fastener defects detection system with high stability and high accuracy,in order to provide a strong technical support for the railway maintenance department.The main research results of this paper are summarized as follows:(1)Aiming at the poor quality of 2D fastener images,a railway fastener images acquisition system based on 3D laser imaging was developed to collect high quality 3D images and provide a good database for the follow-up work.In order to meet the needs of storing 3D images rapidly,the improved data compression algorithm based on JPEG method was proposed,which can handle 3D images with arbitrary depth.(2)Based on the 3D images,the fastener location algorithm validated with prior knowledge was presented,then an elastic clip segmentation algorithm was put forward.Finally,the method of creating negative sample of the fastener images was used to solve the problem of unbalanced quantity of positive and negative samples for fastener detection.The experimental results show that the location algorithm can extract the fastener sub-images accurately,and the segmentation algorithm can get elastic clip sub-images correctly.(3)Combination with the characteristics of 3D images,a recognition algorithm based on Histogram of Oriented Height Gradients(HOHG)and Support Vector Machine(SVM)was proposed to classify five types of fastener images including normal elastic clip,elastic clip missing,upside of elastic clip fracture,downside of elastic clip fracture and bilateral elastic clip fracture.The length of indoor China Railway Track System(CRTS)? slab ballasteless track is short.Therefore,the data of indoor track was collected repeatedly to get more images which are efficacious verified by calculating the average characteristic distance between repetivive images.In the end,the recognition accuracy of this fastener detection algorithm was tested respectively with fastener iamges of two types of track structure.The results show the accuracy of recognition algorithm combining HOHG and SVM is about 98.5%.(4)In order to solve the problem the feature based on manually selecting the feature-descriptor couldn't work well,Deep-Learning was applied to detect railway fastener defects,then the fastener defects detection algorithm based on Convolution Neural Network was proposed.Firstly,the CNN model used to identify four types of fastener images including normal elastic clip,upside of elastic clip fracture,downside of elastic clip fracture and bilateral elastic clip fracture was developed,then the CNN model was finetuned to reduce training time.Secondly,the influence of learning rate and batchsize on the model error convergence speed were analysed to set the optimal parameter value of learning rate and batchsize.Last but not the least,80,000 elastic clip images were employed to train CNN model,and the test set was used to get the accuracy of the algorithm.It is shown that the fastener defects detection algorithm based on CNN can classify fastener images accurately,and it has a certain universal recognition ability to other kinds of hook-shaped railway fastener.