Research On Recognition And Classification Of Tunnel Lining Diseases Based On Deep Learning

In recent years,my country is actively building a comprehensive and efficient intelligent transportation infrastructure.As one of the important transportation facilities,the tunnel will inevitably appear various structural diseases in the tunnel lining during its long-term use,which will endanger the safe operation of the tunnel.Therefore,the efficient identification and classification of tunnel lining structural diseases is conducive to ensuring the safety of tunnel operation,and has certain engineering application value and significant social and economic benefits.Groundpenetrating Radar(GPR)is a widely used tunnel lining structure disease detection tool.However,the current interpretation of GPR data mainly relies on manual work,and its degree of automation is low.At the same time,due to the complex and diverse tunnel lining structure and the presence of radar signal interference in the detection process,it is very difficult to identify the lining disease data.In this regard,this paper takes the realization of forward simulation of tunnel lining disease models and automatic identification of lining disease types as the main objectives of research.The main tasks are as follows:(1)Forward simulation of tunnel lining diseases.In this paper,firstly,Gpr Max simulation software based on Finite Difference Time Domain(FDTD)algorithm is used to construct simulation data of tunnel lining diseases,and on this basis,the prior information of various diseases and the characteristics of disease signals in GPR images are obtained.Since noise interference is inevitable in the process of tunnel data acquisition,in order to improve the robustness of the recognition algorithm,this paper further extracts the characteristics of different noise sources based on the actual measurement image analysis,constructs a tunnel noise model,and relies on the noise model to perform simulation data on the lining disease Add noise,and further generate a pseudo-real tunnel lining disease data set on the simulation data set of the lining disease.(2)A lining disease recognition algorithm based on deep residual network and transfer learning is proposed.In order to solve the problem of insufficient learning sample size and improve the generalization ability of the network,this paper uses a 50-layer residual network(Residual Network-50,Res Net-50)as a pre-training model,and adopts a model-based transfer learning method to identify lining diseases.In order to improve the recognition rate of the algorithm,a dynamic convolution module is introduced to improve the network structure for identifying tunnel lining diseases.The experimental results show that compared with other transfer learning methods,the algorithm proposed in this paper can obtain better recognition performance,and the recognition rate on the pseudo-real tunnel lining disease data set can reach 99.7%.On the real data set,the algorithm proposed in this paper can obtain better accuracy and smaller detection error than other model algorithms,which can basically meet the requirements of actual engineering applications.