Research On Defect Detection Of High-speed Train Floor Bolt Based On Deep Learning

As the continuous progress that China has made in the field of Chinese railway technology,high-speed train has become the first choice for people to travel.By the end of the first quarter of 2019,China’s high-speed train has transported more than 10 billion passengers.For making sure the efficient and safe operation of high-speed trains,each train needs to spend up to 100 minutes at night for comprehensive maintenance.The contradiction between the maintenance capacity unable to grow rapidly and the increasing number of trains is becoming increasingly prominent,which makes the transition from "human inspection" to "machine inspection" imminent.At present,TEDS system is used in high-speed train detection,but there is high false alarm rate and missing alarm rate,so a lot of manpower is still needed to detect.Since 2012,deep learning has made breakthroughs in the field of image recognition and performed well in the Image Net competition.Compared with traditional SVM and other image recognition technologies,deep neural network does not need CV engineers to design specific image features manually,but relies on Convolutional Neural Network(CNN)to train a large number of images to extract deep image features,so both in recognition accuracy and operation speed are beyond the traditional image processing.The detection method based on deep learning proposed in this thesis can realize the accurate positioning of the bottom bolt of high-speed train,and identify whether its state is normal or not,which is of great practical significance for improving the efficiency of train maintenance.In this thesis,the defect detection of train bolts is divided into two stages,the first stage is to complete the positioning task of the bottom plate bolts,the second stage is to complete the defect detection task of the positioning bolts,and judge whether the working state is normal or missing.The main work is as follows:Firstly,in the aspect of bolt location,Labelimg is used to mark three bolts in 13,239 pieces of images of base plate,and the data set is generated.Through experiments,the detection effect of Faster R-CNN,SSD and Yolov3 algorithm in bolt positioning is compared,and yolov3 algorithm is improved: a genetic algorithm based on K-means++ is proposed to optimize the calculation of bolt positioning.The initial population of genetic algorithm is generated by K-means + +.According to the reciprocal of the distance between each anchor and the target frame as the fitness function,adaptive cross mutation rate based on fitness is used to evolve the anchor closer to the real target frame.The experimental results show that yolov3 algorithm is superior to the other two algorithms in real-time performance and detection accuracy,and the anchor generated based on genetic algorithm also effectively improves the recall rate and IOU of Yolov3.Secondly,in the aspect of bolt defect detection,aiming at the problem of too large weight and low detection efficiency of neural network,the VGG16 network is improved from two aspects.One is to use the global average pooling instead of the full connection layer,the other is to use the network pruning to eliminate the convolutional core which has a small contribution to the network output,so as to achieve network compression.In the process of network fine-tuning training,the warm-up learning rate is introduced.By comparing with other learning rates,the most suitable learning decay strategy is selected.Finally,the network compression rate is 98.38%,and the network detection efficiency is improved by 63.35%.