Damage Detection Method Of Steel Spring Of Floating Slab Based On The Convolutional Neural Network

As the urbanization accelerates,the space resources in urban transportation system are becoming increasingly scarce.The subway network ameliorates the urban transportation pressure but it causes serious problems of environmental vibration.Hence,the steel-spring floating-slab track(FST)is widely used in the subway construction due to the excellent vibration damping characteristics.The damage of steel-spring will adversely affect the structural integrity of FST and seriously endanger the operation safety of vehicle.However,the current researches on steel-spring damage detection method are scarce.Therefore,the study of highly automated and accurate damage detection method is of great significance to ensure the operation safety of vehicle and improve the future operation and maintenance of urban transportation.Recently,the convolutional neural network(CNN)has received widespread attention because of its outstanding capability on feature extraction,and has many application precedents in the field of damage detection and fault diagnosis.It can adaptively extract sensitive information that characterizes damage or failure from data without the prior knowledge of the damage or failure.Therefore,this thesis studies the CNN-based damage detection method of the steel-springs.Firstly,the one-dimensional CNN(1D-CNN)is established based on the deep learning theory,and the performance of CNN is studied on binary-classification tasks(to classify the vibration response under normal conditions and damage conditions).The vehicle-track coupled dynamics simulation is employed to generate a variety of data sets required for CNN training and testing.On this basis,the performance of CNN on different signal types,different operating conditions,different sensor positions,and fastener failures condition are investigated.The results of the binary-classification tasks show that: 1)The vertical vibration acceleration of floating-slab is more suitable for steel-spring damage detection than that of rail.2)Different vehicle speeds will have a certain influence on the performance of the CNN.3)Different vehicle load conditions have little effect on the performance of the CNN.4)The greater the degree of steel-springs damage is,the stronger the detection performance of the CNN is.5)The steel-spring damage in the middle of slab is more difficult to identify than that in the end of slab.6)The simultaneous damage of the steel-spring and the adjacent fasteners above it will enhance the distinction between the vibration response of slab under damage conditions and that under normal conditions,but it will be more difficult for CNN to judge whether the steel-spring is damaged when the fastener has been damaged.Furthermore,a deep residual network is established on the basis of the 1D-CNN via the idea of residual learning.Based on the idea of multi-sensor feature fusion,the feasibility of damage locating for all steel-springs on a floating-slab is studied on multi-classification tasks.Concretely,the influence of the complex positional relationship between the damaged steel-spring and the sensors of different number on the CNN performance is analyzed,the optimized sensor deployments are proposed,and the generalization performance of the CNN on the data of different conditions is investigated.In addition,a general rule about determining the number of sensors in a deployment is proposed.Finally,the method and the rule are verified on the full-scale test platform of FST.The results of the multi-classification tasks show that: 1)The multi-sensors deployments are more robust than the single-sensor deployment.2)The sensor should be neither too concentrated nor too scattered.3)The damaged steel-spring in the middle of the slab is more difficult to detect than that in the end of slab,which has nothing to do with the position and number of sensors.4)The CNN has a certain generalization performance on the data of different vehicle speeds,but the generalization performance on the data of different damage positions and load conditions is poor.5)the general rule for the number of sensors can be expressed as: the number of sensors =(the pair number of steel-springs on single slab ±1)/2.