Research On High-efficiency Density Imaging Method Of Large-scale Gravity Data Based On Migration And Deep Learning

Gravity exploration can obtain the comprehensive response of uneven density distribution underground.Fine and high-efficiency density imaging of large-area data is an important requirement in gravity exploration.However,the inversion is complicated and time-consuming,and it is very dependent on the initial model and the constraints used.Considering the huge amount of calculation and calculation time,the inversion will face the limitation of calculation efficiency when the actual data is processed.Despite the rapid development of high-efficiency computing equipment and computing technology,real-time imaging of density distribution still has untapped potential.If the overall three-dimensional modeling of the survey area is carried out,it will face two main bottlenecks: First,a large amount of computer memory is required to store the sensitivity matrix.The second is that the calculation time required to solve large underdetermined equations is beyond the capacity.Unexpected discoveries increasingly come from the analysis of large data sets.Reducing memory usage and improving computing efficiency have become urgent issues to be solved.For this reason,on the basis of regularized inversion and semi-automated imaging,we are trying to find a method to reduce memory usage and fast density imaging,which can solve the memory and computational efficiency limitations of underground density modeling in large areas.Forward modeling is the basis for realizing inversion density imaging,and forward density modeling is the basis for realizing underground density inversion imaging.Normally,the sensitivity matrix will be calculated and stored in advance for subsequent iterative use in the inversion process.The larger the division scale,the larger the sensitivity matrix and the longer the calculation time.In order to reduce the memory consumed by storing the sensitivity matrix,this paper proposes a sensitivity reduction storage method.According to the translation equivalence and exchange symmetry of the sensitivity matrix,the sensitivity matrix and the index matrix of the spatial position are established,and the forward calculation is performed based on the moving sensitivity domain..We also conducted experiments on models of different scales to analyze the advantages of this method in reducing memory consumption and reducing calculation time.This method lays a foundation for rapid density imaging.For rapid density imaging of large-scale gravity data,we propose a gravity regularized focusing migration method,which uses the regularization method to realize the focusing migration density inversion.In this method,the model parameters are solved by the conjugate migration direction method,and the iteration step size is selected based on the Wolfe-Powell criterion.Model tests show that the proposed method greatly improves the computational efficiency,effectively improves the divergence of existing migration imaging,has higher horizontal and longitudinal resolution,and has a strong anti-noise performance.We have also applied this method to the interpretation of the actual gravity anomaly in Qihe skarn type iron mining area in Shandong Province,studied and analyzed the exploration potential of iron ore resources in the investigated area in Shandong Province,and put forward some suggestions for further mineral exploration in the investigated area.In order to overcome these difficulties in the initial stage of the interpretation workflow,we studied the potential of data-driven inversion based on convolutional neural networks in the rapid imaging of three-dimensional density of gravity data.We give a design scheme of a neural network inversion system for density real-time imaging.Our network is based on "Unet" which has the best effect in the field of image segmentation.The mean square error loss function(MSE)is used for optimization training,and the forward error of the test set is used as the observation.We compared the quantitative performance of the neural network-based method and the regularized smooth inversion on the synthetic data set,and proved that the neural network-based method can reconstruct the underground density structure more accurately.Compared with the traditional inversion method,it is expected to provide Faster convergence speed.This method was applied to the gravity field dataset in Luofu,Shaanxi,and the results were basically consistent with the regularized inversion results,which quantified the applicability of this method to real data.