| Recent years have witnessed the rapid development of the national economy and technology,the implementation of the "the Belt and Road Initiative" policy and the development of major projects including the Sichuan-Tibet Railway,and China’s tunnel construction has thus entered a high-speed growth path.Due to the complexity of geological conditions,various geological disasters occur frequently in the process of tunnel construction,including collapse of surrounding rock,ground subsidence and water inrush,which often result in heavy property losses and casualties in tunnel construction,and drastically impede the construction progress.Therefore,it is recommended to carry out advanced geological prediction in the process of tunnel construction and identify the type,location and scale of unfavorable geological bodies in the surrounding rock ahead,which is of great significance to the safe construction of tunnels.Ground Penetrating Radar(GPR),owing to its superiorities such as fast detection speed,high resolution and non-destructive property,is widely used in tunnel engineering.However,the detection data obtained by GPR in tunnels belongs to massive data,and its recognition of unfavorable geological bodies still largely depends on manual interpretation,leading to problems such as big errors,heavy workload and low efficiency.Also,the current research on recognition algorithm for GPR images mainly focuses on artificially designed features,support vector machines,shallow neural networks and other traditional machine learning methods,making it difficult to achieve ideal recognition results due to their notorious subjectivity,and incompetence in recognizing complex objects and extracting essential characteristics of unfavorable geological bodies.In addition,the complex structure of the underground medium,divergent physical parameters,and the existence of noise interference from the system and the environment complicate the electromagnetic wave as it propagates through the strata,and the obtained radar signals consequently contain various clutter and noise,which reduces the quality of radar signal and has a negative impact on the acquisition of real underground geological conditions.To address the aforementioned issues,this paper adopted the Faster R-CNN target detection method to carry out a study on the intelligent recognition of GPR images of major unfavorable geological bodies in tunnels.The main results of this paper are as follows:(1)Theoretical analysis of the characteristics of GPR images for major unfavorable geological bodies in tunnels.The Finite-Difference Time-Domain method was adopted to establish three typical simulation models for unfavorable geological bodies including karst cave,fractured rock mass and water-rich zone.The characteristics of GPR images were theoretically analyzed from the perspective of kinematic characteristics such as electromagnetic wave trajectory and reflection characteristics as well as dynamic characteristics such as amplitude,frequency and phase,with a view to summarizing the characteristics and distribution law of GPR images,which provides a theoretical basis for the subsequent intelligent recognition efforts.(2)Research on the high signal-to-noise ratio imaging method of GPR.This paper proposed a threshold denoising method based on Empirical Mode Decomposition and Permutation Entropy,which solves the contradiction between effective signal retention and noise rejection in traditional empirical mode decomposition methods,and can adaptively remove noise to a certain extent.Through this method,the response features were extracted from radar data,the useful information was highlighted,the noise and clutter were removed,and the GPR images with high signal-to-noise ratio were obtained.By establishing multiple groups of simulation models,this paper expounded the implementation process of this method,verified the feasibility and advantages of it,and laid a solid foundation for subsequent intelligent recognition.(3)The Faster R-CNN algorithm was introduced in the advanced geological prediction of tunnels,and was comprehensively optimized through strategies such as data enhancement,transfer learning and deeper residual network ResNet,in an effort to improve the recognition accuracy of unfavorable geological bodies.Through the comparison experiment with other target detection models,the detection performance of the model.was verified by using the evaluation indicators such as Precision,Recall,Average Precision,and mean Average Precision;Through GPR detection in the project site,the detection data of the actual project was obtained and recognized to verify the feasibility and practicability of this method in the project.The results showed that the method proposed in this paper had achieved satisfied results in recognizing unfavorable geological bodies in the GPR images obtained from tunnel advanced geological prediction. |
