Research On Time-lapse Inversion Monitoring Method Based On Deep Learning

Time-lapse electromagnetic inversion technology is an effective method to observe the change of underground electrical structure with time.By collecting the observation data of the same position at different times,the change of data difference between different times is analyzed,and then the resistivity change of underground medium is inverted.The technology plays an important role in the field of earth detection and monitoring such as reservoir fracturing evaluation,carbon dioxide storage monitoring and dry and hot rock dynamic development.The magnetotelluric method is based on natural field sources.Due to the convenience of the observation device,it can repeat the observation at the same position,and has broad application prospects in the field of time-lapse inversion monitoring.Firstly,this paper introduces the theory of magnetotelluric inversion algorithm,and discusses the existing time-lapse technology according to the theory of Tikhonov regularization.Secondly,in order to improve the continuity of inversion results at different times and suppress the data noise,this paper adds the time regularization term to the traditional objective function to realize the time-lapse magnetotelluric simultaneous inversion algorithm.This algorithm inverts the observation data at multiple times together,improves the influence of noise and improves the continuity of the resistivity of the model at adjacent times.At present,the time-lapse inversion algorithm is mainly based on the regularization theory,and gradually approximates the actual model through the iterative algorithm.In each iteration,the forward operator needs to be derivatized,so a large number of forward calculations are needed,and the computational resource consumption is serious.In order to improve the time-consuming problem of inversion process,this paper combines deep learning technology,by changing the position and shape of the abnormal body,randomly generating a large number of training samples to simulate the change of dielectric resistivity under actual conditions.Two deep learning algorithms are used for inversion,one is two-dimensional inversion algorithm based on attention mechanism and convolutional neural network,and the other is one-dimensional inversion algorithm based on physical guidance.The results show that the inversion algorithm combined with deep learning can effectively reflect the change process of abnormal body in dealing with time-lapse problem,and the speed is faster.Figure 62,table 3,reference 66.