Research On Filtering Algorithm For Surface Diseases Of High-speed Rail Tunnel Based On Lightweight CNN

With the rapid construction of high-speed railway tunnels and the increase of service life,tunnel diseases are gradually increasing.High-speed railway requirements for smoothness make the line bridge-tunnel ratio increasing,there are more and more long tunnels.In the next5-10 years,the completed tunnel will gradually enter the peak of maintenance and repair period.It is difficult to complete huge workload of daily monitoring and maintenance with traditional " huge-crowd strategy ".Because of the short overhaul time and higher safety requirements of high-speed railway,a fast detection system for tunnel diseases emerges as the times require.Vehicle detection system needs to collect tunnel surface information in highspeed environment.Subsequently,the lining image is detected by computer indoors.However,most of the lining images are disease-free(accounting for more than 90%).Eliminating disease-free images can reduce data storage and hardware requirements for high-speed mass storage.Based on the vehicle-borne tunnel lining disease detection system,this thesis focuses on high-speed disease-free image filtering algorithm and lightweight model building during image acquisition.The main contents are as follows:1.The forms of surface diseases of high-speed railway tunnels are investigated and their proportion in high-speed railway tunnels is analyzed.This thesis describes the standard of detection speed in long tunnel and the mass data storage of lining image acquisition under high speed driving.The high-speed rail tunnel lining images are divided into 4 categories.Because of the complex environment of tunnel,an algorithm based on deep learning is proposed to filter lining disease-free images.2.The images of subway tunnel and highway tunnel lining are collected by cameras,and the tunnel surface disease data set TSDD is established by window sliding algorithm and data enlargement method.According to the requirements of real-time filtering of massive images,the commonly used backbone network is compared,and Resnet-18,which exhibits more balanced performance in accuracy and reasoning speed,is finally selected as the backbone network of the lightweight model.3.In order to compress the model parameters,depthwise separable convolution is proposed instead of standard convolution,and global average pooling is used instead of the traditional fully connected layer.Using batch normalization to improve the data mining ability of convolution layer.The built lightweight convolutional model has an accuracy of 99.91%.The inference speed is 11.91 ms per picture under GPU(RTX 2060 Super 8G)and 24.95 ms per picture under CPU environment,which is a significant speed improvement compared to the ResNet native network.4.A weighted loss function is proposed to calculate the distance between the true value and the predicted value.This method improves the accuracy of disease-free samples and improves the filtering performance.For trained ResNet-DS,static offline quantification is proposed to optimize the deployment and inference speed of the model.5.The inference speed of the final model is 21.20 ms per picture and 10.86 ms per picture in CPU and GPU environments,respectively,which is 178% and 22% faster compared to ResNet-18.The proposed lightweight model is easy to deploy,achieves automatic filtering and storage,and has strong engineering application value.