Research On Ground-penetrating Radar Lining Disease Identification Method Based On Deep Learning

With the rapid development of China’s infrastructure construction,as an important transportation facility,structural damage will inevitably occur during the long time use of tunnels,and how to scientifically achieve efficient identification and classification of the internal structure of tunnel lining has become an urgent problem for construction quality inspection and operation and maintenance.Ground-penetrating radar reflects the underground target scene in the echo data according to the principle of refraction and reflection of electromagnetic waves in the interface of media with different dielectric constants,and is widely used in the detection of tunnel lining structure.The classification and identification of targets in the images also cause troubles.In response to the above problems,this paper investigates the ground-penetrating radar data processing technology and automatic target type identification method,and the main research work is as follows.(1)Aiming at the differences of feature curves of different targets in ground-penetrating radar images,the research of ground-penetrating radar orthorectified simulation is completed.In this paper,Gpr Max simulation software based on time-domain finite difference method is used to construct the simulation data of tunnel lining targets under different geological conditions,and the working principle of ground-penetrating radar is combined to analyze the signal type represented by each feature curve in the signal model,and at the same time,the characteristics of noise sources are extracted from the measured data for analysis,so as to better interpret the information obtained from actual measurements.(2)For the extraction of effective signal components in ground-penetrating radar data,the data processing techniques are studied,the direct wave components are removed and the noise components are suppressed using the modal decomposition method,firstly,the simulated signal is processed using the nonlinear adaptive method,the signal is decomposed into a series of eigenmodal components and the form,the characteristics of the signal decomposition methods are analyzed,and then the signal reconstruction is carried out to verify the feasibility of the method after,the simulated data and the measured data are processed using the method for noise suppression respectively,and the effectiveness of the algorithm for noise suppression in ground-penetrating radar data is verified by comparing the original ground-penetrating radar data image with the reconstructed image.At the same time,CEEMD and permutation entropy combined denoising method is studied,by calculating the permutation entropy of the components obtained after ground-penetrating radar data decomposition,obtaining specific separation thresholds to distinguish the target signal from the noise signal,and then superimposing the signals that meet the thresholds to finally obtain the ground-penetrating radar image of the lined target.(3)Deep learning based tunnel lining target recognition method.To address the problem of low efficiency and insufficient accuracy of tunnel lining image disease recognition,an automatic detection method of lining targets based on Faster R-CNN is studied,and an improved feature extraction network structure is proposed based on the original one,which combines the depth features and shallow features in the feature extraction network to form new features to improve the accuracy of target recognition.And by comparing the experimental results of forward simulation generation and real image target recognition,it is verified that the improved network model has better detection effect and the detection accuracy of the target has been significantly improved.