Study Of Dielectric Properties Of Typical Targets And Their Effects On Radar Signals

Ground-penetrating radar is an efficient and non-destructive road detection technology,which can obtain the location of underground targets and other related information by observing the changes in the amplitude of the reflected waveform.However,the accuracy of this technique may be somewhat limited due to unknown factors such as dielectric properties.Therefore,it is necessary to seek a simple and effective method to interpret the ground-penetrating radar detection results in a more accurate and standardized way.The conversion of radar data into dielectric constant images can bring visual and clear results.Dielectric constant images can improve the accuracy and efficiency of engineering inspection by analyzing two-dimensional data.Conventional geophysical inversion methods take a long time and are highly dependent on the initial model.In recent years,deep learning techniques have been increasingly applied to the field of EM inversion.Compared with traditional EM inversion methods,deep learning methods are more efficient and faster,but they also have some drawbacks,such as incomplete feature extraction and difficulty to guarantee spatial correspondence.Therefore,a dielectric constant inversion method is urgently needed to improve the efficiency and accuracy of road structure disease detection.The main research work and results of the thesis are summarized as follows:(1)For the ground-penetrating radar image inversion problem,the thesis proposes a ground-penetrating radar image dielectric constant inversion network called GRINet,which uses a deep learning method and combines ground-penetrating radar image features to solve the dielectric constant inversion problem.First,we constructed a feature encoder to extract the features of the B-scan.Then,we constructed a dielectric constant decoder which is able to reconstruct the dielectric constant image from the downsampling of the encoder and is able to perform upsampling.Since clutter in the ground-penetrating radar image can mask and distort the target response,which affects the accuracy of the dielectric constant inversion,a Residual Dense Block(RDB)is integrated into the jump connection between the encoder and the decoder,whose main purpose is to adaptively retain the features associated with the target response and reduce the clutter features,thus improving the inversion The main purpose of this block is to improve the accuracy of the inversion by adaptively retaining the features related to the target response and reducing spurious features.(2)For different dielectric properties,different sizes and shapes of typical targets and different radar center frequencies,the response is different and thus the corresponding inversion effect will be different.In this paper,the dielectric constant models with different sizes and shapes,different dielectric constants and different ground-penetrating radar center frequencies are established,and the corresponding B-scan is obtained by forwarding,forming a data pair,and the expansion of the data pair is done to establish a typical target model.The data set was expanded to establish a typical target model.The data set is then used to train GRINet,and the inversion performance of GRINet under different conditions is tested and compared.The dielectric properties of typical targets and their effects on radar signals are analyzed,and the effectiveness of GRINet inversion is verified and compared with other inversion networks.The network parameters are also tuned to demonstrate the better performance and accuracy of the GRINet network.(3)For the application of ground-penetrating radar inversion to actual road damage detection projects,the authors of the thesis conducted actual road damage detection engineering experiments,obtained actual road damage data,and established the corresponding dielectric constant model according to the actual road damage conditions.Then the dielectric constant model and the actual B-sweep graph were used to build the corresponding actual road disease data set and put into the GRINet network for training.The final test results successfully inverted the location and shape size of the target object,which proved the potential of the GRINet network in practical engineering applications and its superiority in inversion performance.