基于CNN-LSTM网络的频率域井地电磁法深度学习反演研究OA

Faced with the demands of resource and energy exploration,along with the precise interpretation of complex geological structures,there is a heightened need for improved anti-interference capabilities and resolution in electromagnetic methods.The Borehole-to-Surface Electromagnetic Method combines the advantages of conventional electromagnetic techniques and borehole geophysics.In this approach,the transmitter is deployed downhole near deep target bodies to emit electromagnetic signals,while areal observations are conducted on the surface.This configuration exhibits enhanced sensitivity to resistivity variations in target bodies and produces more distinct anomalous responses.However,the Borehole-to-Surface Electromagnetic Method requires extensive areal data acquisition,resulting in large datasets that slow down inversion computations,consume significant resources,and exhibit accuracy sensitivity to initial model selection.Deep learning offers a solution by leveraging massive datasets to autonomously learn features.It excels in handling complex spatial-temporal patterns in borehole electric field data and enables efficient temporal modeling,thereby enhancing both the efficiency and accuracy of inversion.This study generated a large training dataset through forward modeling.The acquired frequency-domain data inherently exhibit sequential characteristics.The proposed methodology utilizes Convolutional Neural Networks to extract complex spatial features from borehole electric field data and employs Long Short-Term Memory networks to capture dependencies across different frequency sequences.This dual approach effectively addresses both local patterns in multidimensional complex data and long-term dependencies in sequential data.The inversion results demonstrate that the method achieves rapid inversion with high accuracy,significantly advancing the application of deep learning in borehole-to-surface electromagnetic inversion.