Magnetotelluric Deep Learning Inversion Based On Convolutional Neural Network

On the basis of previous studies,this dissertation introduces the related technologies and methods in the field of computer vision,and realizes the whole process from theoretical derivation to practical data application based on convolutional neural network deep learning inversion.Firstly,the 2D magnetotelluric(MT) method forward modeling and the traditional regularized inversion method are taken as the breakthrough points.The 2D MT forward calculation and regularized inversion calculation are realized by using the finite element method.The experiments show that there are some differences in resistivity and geometry between the regularized 2D MT inversion and the real model.The inversion results are seriously affected by static effects.In this dissertation,the structure of deep learning convolutional neural network for MT inversion is designed by combining residual network module,multi-scale pooling module,feature fusion module,segmentation and classification module.We adopt the improved loss function of classification problem.Based on the 2D MT forward program,a large number of theoretical models are generated as training and learning samples.The experiments show that the designed neural network is suitable for MT data.The results of deep learning inversion can well recover the resistivity and geometry of abnormal body.The results of deep learning inversion are less affected by static effects.By training and predicting the samples of MT theoretical models,the L2 norm loss function and Focal loss function improved cross entropy loss are compared and analyzed.In this paper,the loss function combining Focal loss with L2 norm is used to speed up the training process and ensure the stability of training.The advantages and disadvantages of RMSProp and Adam adaptive gradient optimization algorithm are compared and analyzed.Finally,we adopt the combination optimization strategy of Adam and SGD.In this dissertation,the open source MT data from southern Africa are used for deep learning MT inversion experiments.Deep learning inversion results show that there are many redundant structures and noises in the inversion results.The main reasons for this phenomenon are the discontinuity of classification and the sparsity of the collected data,which leads to the stretching of noise in the inversion results.In order to solve the above problems,a solution is proposed to change the convolution kernel in the process of dimensionality reduction of multi-scale pooling layer,so as to alleviate the redundancy of MT deep learning convolution neural network inversion.Finally,we take the measured data of Qihe low resistivity coverage area in Shandong Province as an example,the deep learning inversion research is carried out.Through the comparison and analysis of the results with the traditional regularization inversion,this paper considers that the MT inversion method based on deep learning convolution neural network can be applied to the actual electromagnetic data,but it needs to be combined with the regularized inversion results for comprehensive comparison and analysis.