Research On Semi-airborne Transient Electromagnetics Dataset Expansion And Fully Convolutional Network Noise Reduction Based On Generative Adversarial Network

With the gradual progress of national infrastructure construction,engineering safety and engineering quality have received more and more attention.In the southwestern region of China,there are many mountains and complex terrain,and the high bridge-to-tunnel ratio is also a major feature of the road engineering here.However,the tunnel construction process is complex and the risks are high,such as water and soil inrush,which is a major problem in the process of tunnel construction.Using geophysical methods to predict and warn of hidden dangers has become one of the main methods to solve such problems.The semi-airborne transient electromagnetics(SATEM)with grounded long wire source overcomes the shackles of traditional methods limited by complex terrain,and has been recognized and promoted with the advantages of high efficiency and good accuracy.However,the device form of ground transmission and air reception introduces more noise interference into the acquisition process,posing more challenges to data processing and interpretation.In recent years,with the improvement and enhancement of computer computing power,the development of artificial intelligence technology has been promoted,and the introduction of deep learning has enabled it to have stronger recognition and classification capabilities.Many geophysicists have turned their attention to the field of deep learning,hoping to solve more geophysical problems with this method.However,the traditional application fields of deep learning,such as image recognition,have low acquisition cost and large number of data sets,and high cost and small number of geophysical data acquisition,which limit the accuracy of its application to a certain extent.As far as the noise reduction problem is concerned,since it is impossible to completely separate the signal and noise of the actual data,it is obviously not the optimal solution for sample identification and learning simply through synthetic noise.Based on such questions,the research contents of this paper are as follows:(1)The SATEM method is different from the transient electromagnetics(TEM).It adopts the form of ground excitation and aerial reception.This method increases the difficulty of data correction and introduces more different types of noise into the data.,such as electrical noise,coil motion attitude noise and speed changes and related interference generated by the drone itself.Using traditional methods to remove noise requires the selection of different methods and parameters to adapt to complex and changeable interference situations.If the selection is improper,it will affect the valid data itself.As far as the deep learning noise reduction method is concerned,the supervised learning method must use the simulated pure signal and noise superposition to create a data set.At the same time,SATEM has complex noise types and strong randomness,and the artificial synthesis of noise data is very difficult.Based on this problem,a method of superimposing the actual noise and forward data to produce a high-quality SATEM synthesis dataset of long wire sources is proposed for supervised training of noise reduction networks.(2)The realization of noise reduction and recognition functions through deep learning methods is very dependent on the quality of the dataset and the number of samples.However,the amount of actual noise data used by the noise reduction network is scarce and cannot meet the needs of processing accuracy.Based on this problem,this paper proposes a sample expansion method using generative adversarial networks to sample noise data.Based on the sample distribution of sampled noise data,the two network structures of generator G and discriminator C are used to identify and distinguish between Simulate the sample distribution of the actual noise data,and achieve the purpose of data set expansion by sampling a large number of samples in the simulated distribution.It has been verified that the denoising effect of deep learning using the augmented dataset is significantly better than the denoising effect of deep learning without the augmented dataset.(3)After obtaining a sufficient number of noise samples,a large amount of pure TEM signals were obtained by SATEM one-dimensional forward modeling method.According to the signal-to-noise ratio of the actual signal and other characteristics,the pure TEM signal and the noise sample are superimposed to obtain a large number of synthetic data samples,and the noise reduction network is trained.After the training,the results are verified to know that the data processing by the noise reduction network can effectively improve the parameters such as signal-to-noise ratio and correlation coefficient,and effectively improve the quality of the detection data.Finally,relying on the SATEM projects of the Damo Tunnel,FengHuang No.1 Tunnel,and FengHuang No.2 Tunnel,this method was used to de-noise the field data,and the generalizability of the de-noising network was verified.Using the apparent resistivity imaging method,the apparent resistivity distribution of the underground in the work area before and after noise reduction is compared,which proves the reliability and effectiveness of the noise reduction method proposed in this paper in practical engineering.