| The airborne magnetotelluric method uses the natural electromagnetic field as the field source,receives the magnetic field signal generated due to the uneven lateral electrical structure of the underground through the vertical magnetic field acquisition coil mounted on the aircraft,collects the horizontal magnetic field signal on the ground through the base station,and obtains the tipper through conversion.Furthermore,the geophysical interpretation of the subsurface electrical conditions is carried out by taking use of the tipper data.The method is widely used in the geophysical prospecting work in complex and difficult mountainous areas due to the advantages of high exploration efficiency and low terrain limitation.The inversion of tipper is the process of inferring the subsurface electrical structure in the observation area based on the tipper response.The traditional method of tipper response uses the way of solving the partial derivatives of the objective function.It has the disadvantage of high dependence on the initial model and is easy to fall into local minima during gradient descent.Deep learning has shown excellent ability in dealing with nonlinear inverse problems.For this reason,in this paper,an artificial intelligence algorithm based on fully convolutional neural network is proposed for the inverse problem of tipper data with huge data volume.The implementation of the algorithm mainly consists of the following two parts: 1)scientific and reasonable construction of the dataset;2)the construction and optimization of the fully convolutional neural network.On the construction of sample dataset,three types of anomaly body model designed are block anomalies,ladder anomalies and irregularly shaped anomalies;two types of fault structure model designed are normal fault and reverse fault.Carry out forward calculation on the above geoelectric model to obtain its corresponding tipper value.The real and imaginary parts of the tipper are used as the feature values,and the resistivity values of the corresponding geoelectric model are used as the label values,which together form a data pair.On the construction of neural network framework,a fully convolutional network is designed,called Tip Inv-net,to achieve "end-to-end" inversion of the tipper response.The network takes the classical U-net network as the backbone,and uses the average pooling method at the input side to achieve multi-scale tipper response feature extraction to construct multi-scale input in the encoder path;at the same time,the dense jump connection method is used to further fuse the extracted multi-scale features;in order to better integrate the extracted tipper response features at each stage,Tip Inv-net adopts atrous spatial pyramid pooling to enhance the detail of the anomalies and tomographic models to improve the accuracy of the inversion.In the end,the inversion study of theoretical data,noise-added data and measured data is carried out.The inversion results show that:(1)Tip Inv-net has certain generalization ability and anti-noise interference ability,and has good inversion effect for some complex anomaly combination models;(2)When the anomalies are close to a certain extent in the subsurface space,which affects each other’s tipper response characteristics,it is not enough to train a single anomaly model training set,and a new sample dataset should be rebuilt to complete the training;(3)Compared with the traditional iterative inversion method,Tip Inv-net has more accurate localization and higher computational efficiency for the anomalies;(4)In the face of fault structure tipper response data with strong noise interference,Tip Inv-net can still achieve effective inversion results,and the inversion result of the measured data are consistent with the actual geological situation,which illustrates the effectiveness of the method. |
