Research On Electromagnetic Inversion Method Based On Machine Learning

Traditional nonlinear electromagnetic inversion methods often have problems such as multiple solutions and local extremum in electromagnetic inversion,and these methods often rely on enough prior information to establish an appropriate initial model.Machine learning related algorithms provide a possibility for electromagnetic inversion.It does not need to linearize the inversion problem.It just transforms the entire physical process into a"black box" for processing,and directly obtains the inversion result.Machine learning algorithm is a valuable electromagnetic inversion method.In this thesis,firstly,Convolutional Neural Network(CNN)algorithm is used to study electromagnetic inversion.By picturizing the characteristics of electromagnetic inversion(position parameters,relative dielectric constant),pixel matrix samples are obtained.Then,these samples are input into the full convolutional neural network for training.The results show that the algorithm based on CNN is an effective and highly generalized nonlinear electromagnetic inversion algorithm.Then this thesis introduces K nearest neighbor(KNN)theory into the field of electromagnetic inversion for the first time and analyzes its advantages and disadvantages.Three elements of KNN algorithm(distance measurement,K value and average mode)are determined through simulation,and the simulation results are given.On this basis,this paper also proposes a dense-crossover grid modeling method for KNN electromagnetic inversion,and gives its algorithm idea and program implementation.Then,its inversion effects on different sample sets are verified,and the inversion accuracy and time cost of the improved KNN electromagnetic inversion algorithm are analyzed and compared.The simulation results show that the method has a great improvement in time with less training sample set and computational resources and sufficient prediction accuracy.Finally,the advantages and disadvantages of the two electromagnetic inversion algorithms are compared in terms of time cost and inversion accuracy.The results show that CNN based electromagnetic inversion algorithm is more generalized and more suitable for complex inversion problems.K nearest neighbor electromagnetic inversion algorithm has a very high inversion accuracy,which is faster and more accurate when dealing with some simple electromagnetic inversion problems with small sample space.