Crack Detection Of Railway Concrete Sleeper Based On Deep Learning

Crack is one of the most important factors that characterize the safety of double-block concrete sleepers,which are important components of high-speed rail.China has the world's largest highspeed railway operating mileage.With the increase of high-speed railway operating mileage,the traditional method of relying on manual inspection and machine-assisted detection of sleeper cracks has become increasingly unable to meet the actual railway maintenance needs.Automated sleeper crack detection system.Existing crack detection algorithms are mainly based on deep learning frameworks.These methods can be divided into single-stage and dual-stage methods according to different principles.However,at present,neither of these two methods can meet the requirements of speed and accuracy of the detection task at the same time: the single-stage algorithm has high detection speed but low accuracy;the two-stage algorithm has high detection accuracy but low speed.The goal of this subject is to propose an algorithm that can meet both accuracy and speed requirements.In response to this problem,considering that the detection speed of the two-stage algorithm is difficult to increase due to principle reasons,this topic will focus on improving the performance of the single-stage algorithm.The main work contents are as follows:First of all,this paper carried out related basic work around the crack detection problem studied,including: 1)Data collection and processing.A total of 54,205 actual sleeper pictures were collected,annotated,and posted online for use by subsequent researchers;2)Proof of the effectiveness of the neural network to extract crack features.By visualizing the feature map,it is proved that the neural network can effectively distinguish the difference between the crack and non-crack picture features,which provides an explanation for the effectiveness of the deep learning method proposed in this paper;3)Analysis of the cause of crack miss detection.Through the feature map analysis,it is concluded that the maximum pooling layer should not be used in shallow network design as much as possible,which is the main cause of missed detection of small cracks,and provides guidance for the selection and design of the detection network afterwards.Secondly,this paper proposes a single-stage CF-NET algorithm with higher accuracy.The main reason for affecting the accuracy of the single-stage algorithm is that the scale change range of the crack is too large,and the divide-and-conquer labeling method can artificially reduce the crack scale change range to improve the algorithm accuracy.However,this labeling method has the problems of repeated detection of a single crack and inaccurate crack position information.The CF-NET network mentioned in this paper effectively solves these problems by improving the new labeling method,detection process,confidence fusion strategy and loss function.Experiments show that the highest accuracy rate is 99.1%,and the real-time detection rate is still 17 FPS.CF-NET effectively improves the accuracy of the single-stage algorithm while maintaining the original single-stage algorithm detection speed,and can meet the needs of the real-time detection task of the sleeper detection trolley with a maximum speed of 23.3 km / h.Afterwards,this paper further proposes a Fast Crack network with a faster detection speed for the problem that the detection speed of the CF-NET network is still too slow under certain detection scenarios.The feature maps between channels and layers of the CF-NET network have a large number of overlapping areas,and there is a large amount of parameter redundancy.To solve this problem,Fast Crack network uses pruning algorithm to cut the network parameters based on the single-stage algorithm YOLOV3,greatly reducing the amount of calculation to improve the speed of the algorithm.The pruning algorithm incorporates a new layer pruning strategy and a local attenuation sparse training strategy,and uses knowledge distillation to improve accuracy in the finetuning stage of the model.Experiments show that,under the premise of ensuring a high accuracy rate of 97.2%,the Fast Crack network has a 10 X drop in parameter quantity compared with the original model,and the detection speed has reached 126 FPS.Its comprehensive performance can meet the real-time detection task of the track inspection vehicle with a maximum speed of 170 km / h.Demand.Finally,this paper develops a practical sleeper maintenance management system based on the previously proposed algorithm.The system uses modular and distributed design,which can well meet the needs of automatic detection,maintenance and management of sleepers.