Key Technologies And Characteristic Mode Analysis Methods For 5G Base Station Antennas

The borehole radar is a special ground penetrating radar,which is mainly designed for the detection of deeper strata,and also has the characteristics of high efficiency and non-destructiveness.Transient pulses are emitted during logging,and the ground is detected by capturing reflected echoes.Nowadays,the processing of radar logging data is mainly the preliminary judgment of underground targets,such as the location of the target,the trend of fractures,the characteristics of abnormal bodies,etc..With the increasing demand for resource detection and anomaly detection,accurate inversion of dielectric properties has also become an urgent need.The traditional inversion is realized by solving Maxwell's equations numerically,but there are still some problems such as large amount of calculation,low precision of inversion results,and heavy dependence on the initial model.With the rapid development of deep learning,new ideas are provided for solving the inversion problem.Inversion can be regarded as the conversion from two-dimensional B-scan data to two-dimensional permittivity data,which is consistent with the image translation problem in deep learning.It is very fast to apply the trained deep learning network to inversion.This paper mainly studies the deep learning method applied to inversion.First,the transient pulse system logging radar and three logging data modes are described.On top of this,the characteristics and preprocessing methods of B-scan data are explained.Then,some basic network structures and methods in deep learning are explained.Adversarial networks are described in detail,providing a basic reference for solving inversion problems through deep learning.Both the encoder-decoder network and the generative adversarial network are commonly using deep learning methods to solve the problem of image translation.On the simulation data,the encoder-decoder network is trained using two loss functions,MSE and MAE.The generative adversarial network is divided into two parts: the generator and the discriminator.The generator has the same structure as the encoder-decoder network.And the classifier is adopted as the discriminator.In order to stabilize the training of the generative adversarial network,the Wasserstein distance with regularization is used as the loss function.The comparison results show that the generative adversarial network has better generalization ability on the inversion problem than the encoder-decoder network in the case of a small amount of training data,and the generalization error is reduced by 3%.Due to the high cost of conducting a large number of experiments with the borehole radar,the underwater experiment of the borehole radar is used to verify the method.The inversion results show that the generative adversarial network can effectively invert the position,shape and dielectric properties of the target.Generative adversarial networks with classifiers as discriminators have poor results in the inversion of details.PatchGAN is an efficient discriminator structure that pays more attention to details.According to the characteristics of B-scan data,the input mode is improved,and the output of PatchGAN is improved in the context of large background and small target.The improved method is verified by ground penetrating radar experiments.Compared with PatchGAN,although the improved structure has obvious streak noise in the background inversion,it is not suitable for the inversion of water,air and wet sand under the sand background.The performance is improved by23%,36% and 12% respectively,and the improved results have better inversion effect for small targets.This paper completes the preliminary exploration of the application of generative adversarial network to the inversion of the permittivity of radar logging,and proposes an improved PatchGAN structure for small targets,which can quickly obtain higher-quality permittivity data.