The Research And Application Of Deep Learning In Tunnel Measured Ground Penetrating Radar Date Processing

Because of its ground penetrating radar’s good resolution,high efficiency, easy operation and no damage detection, Ground Penetrating Radar is widely used in various engineering fields. Deep learning is a new research area in machine learning, Deep learning have a good expression of the features, and it can automatically extract the features needed for classification. More important, it is able to achieve very good results in large scale data processing. Deep learning can be well used in GPR data processing, and can effectively study of ground penetrating radar image features. Deep learning applied to ground penetrating radar image processing, to unsupervised learning exploration of radar image, and to image classification in a supervised way.In this paper, firstly, the electromagnetic theory, data acquisition principle and detection performance of GPR are studied. Secondly, the theory of deep learning, the method of deep learning and Limit Boltzmann machine model are studied. The RBM model of the deeplearning algorithm is relatively easy to learn, and the algorithm of this model overcomes the efficiency of the direct multi-layer network training. Therefore, the follow-up experiments of this paper are based on the RBM constructed depth belief network model(DBN) and used to GPR data processing. Then, the establishment of the sample library, the setting of some experimental parameters and the construction of the RBM model are presented. As the ground penetrating radar image is formed by the reflection of electromagnetic waves, to make the processde ground-penetrating radar image more intuitive to be expressd, using the softmax algorithm to classify the learning of GPR image ication.Finally, through the experiment of good sample database has been established, using the DBN deep learning algorithm of GPR images of unsupervised learning, and used with labeled data to classify samples. The experimental results show that can unsupervised learning GPR data, and can well classify by this method.