Research On Data Processing And Density Inversion Based On Deep Learning

Deep learning theory can model relationship between data and label,a large number of practical applications show that it has great application potential.when it be applied in the data processing and interpretation of potential field,it can extract a lot of prior information from potential filed data so that it can calculate the density of the geological body and finish interpolation,filtering.Regular grid method,such as kriging method,minimum curvature method,due to treat abnormal grid for statistical hypothesis method in the abnormal area,common filtering methods such as median filtering method when it filter the random noise of abnormal,it also filter abnormal's data,it leads to lower precision in the abnormal area.The method based on full convolutional neural network don't work well in density inversion,because these models' s ability of building the nonlinear correspondence between ground observation anomaly and underground density model isn't strong.Introducing self attention neural network that can extract global potential field abnormal information,meshing method based on self-attention mechanism neural network can extract the global feature of the anomaly and reconstruct it as a regular point by using priori information learned in the potential field's training set,the method's error is 10 times lower than kriging meshing method,method's error does not change with the anomaly,the accuracy is higher.Noise recognition method based on convolutional neural network is proposed,which can effectively identify noise that don't belong to geological anomalies.Denoising autoencoder based on convolutional neural network can filter the random noise of the potential of data,extract the local features of the data through the convolution layer,and reconstruct the gravity and magnetic anomaly data.Compared with the traditional median filtering method,the accuracy of the denoising autoencoder is higher,and the error does not change with the anomaly distribution.According to the global characteristics of density inversion,a density inversion method based on selfattention neural network is proposed.By incorporating global potential field anomaly information into the calculation,the method can calculate the density division more accurately than the convolutional neural network.Applying data processing method based on deep learning to measured magnetic anomaly data from Hu Ge area Nei Meng Gu Province,result shows that method can effectively remove the banded interference,applying the density inversion based on self attention pretrained neural network to Vinton salt dome's measured gravity anomaly data,result show the method can effectively calculate the density distribution of geological body.Self-attention neural network is introduced into potential field data processing,apply convolutional neural network and self-attentional neural network in data processing and density inversion according to the characteristics of different neural networks,which provides a certain reference for the next application of deep learning applied in the potential field.The next research direction will focus on the interpretable research of deep learning theory in the potential field data processing and density inversion,and determine the source of effectiveness of deep learning,so as to continue to improve the application advantages of deep learning theory applied in the potential field.