Recognition Of GPR Shallow Surface Target Based On Attention Model

With the increasing demand of underground pipeline detection and other applications,the detection of underground targets has gradually attracted people's attention.At present,the commonly used methods for underground target detection mainly include electromagnetic method,acoustic method,infrared detection,ground penetrating radar and other methods.Among them,GPR is an efficient geophysical exploration method,which has higher accuracy than electromagnetic method and wider application range than infrared method,so it has become the most widely used underground target detection method at present.Ground penetrating radar can detect materials with different permittivity by transmitting and receiving electromagnetic pulse signals to the ground.Based on this,it can detect the internal structure of the medium and the change of material properties.It is an effective means of nondestructive detection of underground targets.Because radar wave is easy to be interfered in the process of propagation,a lot of interference information is generated in the image data,which greatly improves the threshold of data discrimination,increases the cost of radar data reading,and slows down the application and development of ground penetrating radar.In view of the low accuracy of GPR data interpretation,combined with the characteristics that GPR data can be represented as two-dimensional images,this thesis applies artificial intelligence method to the discrimination of GPR data to realize the automatic discrimination of underground targets.The main work is as follows:1.Combined with attention network to construct neural network,this thesis studies the target characteristic curve in GPR data,forms the underground target intelligent recognition method,and improves the accuracy of GPR B-scan image target recognition,especially the pipeline leakage data recognition.2.Based on the Finite difference time domain method(FDTD)numerical simulation theory and the existing soil dielectric hybrid model,this thesis constructs a layered pseudo real soil model,uses the model to generate pseudo real GPR simulation data in batches,and applies the data to the learning of neural network,and verifies the network's ability to recognize underground targets in different types of soil at the simulation data level.3.The neural network model constructed in this thesis is trained and tested in the simulation and measured data respectively,which can identify the horizontal position,depth and cross-sectional area of the underground target,and accurately judge the pipeline leakage,which makes up for the deficiency of the existing artificial intelligence methods for the identification of pipeline leakage related content in the GPR data,The accuracy of geometric position recognition of underground objects is improved.In the test set,the judgment results are compared with the tags.It is verified that the accuracy of the network structure is 94.5% and 84.6% respectively in the longitudinal pipeline leakage identification results and the horizontal target data of the simulation data,and95.1% and 96.4% respectively in the pipeline depth and leakage area identification;At the same time,the accuracy of the network is 92.3% and 84.4% respectively in the longitudinal pipeline leakage data and the transverse target data.