Images Recognition Of Underground Cavity Target Of Ground Penetrating Radar Based On Machine Learning

In recent years,road collapse accidents caused by underground cavitation occur frequently.In order to prevent the occurrence of accidents,an efficient means of road detection is needed,and the efficient vehicle-mounted array 3D ground-penetrating radar is the most suitable technology to fulfill the work.However,for cavity target detection in reality,an individual can only differentiate data of 3 kilometers by manual interpretation per day.In the contrary,vehicle-mounted array 3D ground-penetrating radar can achieve road detection at a speed of 80km/h,producing a huge amount of data.In allusion to the problems above,the research of this essay is divided into two parts: the design and implementation of array 3D GPR system and the image recognition of GPR underground cavity based on machine learning.The GPR studied in this essay obtains the capacity of supporting high-speed image acquisition and 3D image formation.It adopts the three-stage design of upper computer + main control board + acquisition system.The acquisition system is the core of the GPR system,including acquisition board and array antenna.The hardware design and embedded software design of acquisition board are demonstrated in detail.The acquisition board adopts KINTEx-7 series FPGA as the main control chip,and adopts high-speed AD chip combined with equivalent sampling to achieve stable and accurate acquisition of underground echo signal under the high-speed condition of 80km/h.In this paper,an automatic underground cavity identification algorithm using ground-penetrating radar A-scan data and C-scan data is proposed.The A-scan data collected by ground-penetrating radar contains abundant time-frequency information,which can reflect the essential characteristics of the signal,meanwhile the data amount is small.C-scan data can reflect the detailed status of three-dimensional underground space and the information contained is more detailed.In this essay,the characteristics of A-scan data of GPR are extracted in the first place and in the case of a large detection range,the Gaussian Mixture Model-Hidden Markov Model(GMM-HMM)is adopted to quickly determine the rough location range of the detection target.Then,within the determined range,a full play is given to the advantages of the array 3D groundpenetrating radar system to extract features from the horizontal section images of Cscan data collected,and the feature vector sequence takes shapes layer by layer,which is accurately identified by GMM-HMM.Finally,the identification results are analyzed and the effectiveness of the method is verified.