Research On Fine Detection Of Ballastless Track Slab Cracks Based On Deep Learning

Ballastless track slab is widely used in China’s high-speed railway construction.With the increase of service time,the cracks of ballastless track slab show an increasing trend under the coupling effect of train high-frequency load and environmental factors.In recent years,using machine learning method to realize intelligent and fast crack detection technology has become a research hotspot,but it still has limitations in detection accuracy,speed and robustness.Therefore,this thesis integrates deep learning and feature measurement technology,and puts forward a set of fine detection technology for ballastless track slab cracks around “image acquisition—crack identification—crack segmentation—quantitative processing”.The detailed contents are as follows:(1)The Inception-ResNet-v2 network model is introduced to establish the crack image recognition system of ballastless track slab.Data enhancement and histogram equalization are used for image preprocessing.The crack image recognition effect of the network model is tested and evaluated from three aspects: data,method and environment.Experiments show that the recognition accuracy of crack image is 99%,which is better than four common machine learning methods: support vector machine,artificial neural network,naive Bayesian classifier and k-nearest neighbor,and has good robustness.(2)The DeepLabV3+ network model is introduced to establish a semantic segmentation system for cracked images,which can perform semantic segmentation on images with cracks.The optimal hyperparameters of the model are obtained by grid search method,and are compared with Fast-SCNN and U-Net semantic segmentation network models.Experiments show that the semantic segmentation accuracy of this network model reaches 89.2%,which is better than Fast-SCNN and U-Net semantic segmentation methods.(3)An automatic crack quantification system based on improved Laplace algorithm is established to realize continuous width measurement along the direction of crack skeleton.For the problems of crack width quantification standard and automatic measurement,the capacitance model is used to define the crack width,and the edge operator is replaced by the DeepLabV3+ network model to improve the general Laplace algorithm for calculating the continuous width of cracks,and obtain the crack indexes including average width,maximum width,minimum width and standard deviation.Compared with the crack width gauge and the general Laplace algorithm,the crack width measured by the improved Laplace algorithm is more scientific and accurate.(4)A fine detection platform for ballastless track slab cracks has been developed from the perspective of system software and hardware.The platform includes three modules: crack image acquisition,crack image identification and crack width measurement.The selection of camera,auxiliary light source and photoelectric encoder and the design of software system structure are completed.