Research On Cross-hole Resistivity Tomography Algorithm Based On Deep Learning

Electrical Resistivity Tomography(ERT)is a significant geophysical exploration technique with wide applicability,high exploration efficiency,and low economic cost.It finds extensive application in areas such as petroleum exploration.ERT involves the placement of measurement electrodes and excitation electrodes in two or more boreholes.The excitation electrodes introduce current into the subsurface at different depths,while the measurement electrodes in neighboring boreholes detect the potential values at various depths within the boreholes.By utilizing the positions of the excitation electrodes and the measured potential values at different depths from the measurement electrodes,the technique employs resistivity inversion algorithms to reconstruct the resistivity distribution of the subsurface.Compared to surface electrical resistivity imaging techniques,the measurement devices in cross-well resistivity methods are located below the surface.This provides enhanced resolution for anomalies at certain depths,making it more suitable for exploring deep-seated geological resources such as deep-seated mineral deposits.Hence,this technique holds significant importance for the exploration of deep geological resources.Currently,the main algorithm used for cross-well resistivity tomography is the linear inversion method,such as least squares.However,the non-linear mapping relationship between the resistivity distribution in the field under investigation and the measured potential data leads to drawbacks in traditional linear inversion imaging methods,including unclear inversion shapes,the presence of false anomalies,and long inversion times.To address these issues,a neural network is employed to learn from a large amount of known resistivity distributions and measured potential data,thereby uncovering the non-linear mapping relationship between these two types of data.This enables more accurate inversion imaging results.Additionally,the trained network can be saved as a model,allowing for faster inversion of unknown potential data.This study first establishes cross-well resistivity models with different shapes and sizes of anomalies.Using the COMSOL simulation software,forward modeling is conducted to obtain potential data on the measurement electrodes within the boreholes.Under the assumption of a uniform half-space,the influence of different grid refinement methods on forward modeling accuracy is explored.It is observed that using local grid refinement effectively solves the problems associated with long calculation times when applying uniform grid refinement and large calculation errors when no grid refinement is applied.In terms of inversion research,the potential data obtained from forward modeling is used as input to a convolutional neural network(CNN)to invert and obtain the conductivity distribution image of the target area.To address issues such as false anomalies and insufficient boundary characterization in the inversion results,attention mechanism is first investigated through ablation experiments.It is found that the attention mechanism effectively suppresses false anomalies in the inversion.Furthermore,comparative analysis is performed on learning rates,optimization algorithms,and activation functions within the CNN.Dynamic learning rates are found to be more effective than fixed learning rates in suppressing the occurrence of false anomalies.Adam optimization algorithm and Leaky Re LU activation function demonstrate superior boundary characterization capabilities compared to other similar types of neural network parameters.By inverting various types of anomalies with different shapes,experimental results demonstrate that the CNN algorithm with residual modules and attention mechanism can effectively determine the position and shape of individual anomalies within the target area.Similarly,it can also characterize the positions and approximate shapes of multiple anomalies in multi-anomaly inversion.The study further analyzes the impact of anomaly thickness and resistivity contrast with the background field on inversion results,revealing that thicker anomalies and greater resistivity contrast between anomalies and the background field result in relatively better inversion outcomes.Additionally,experiments are conducted to evaluate the network’s noise resistance performance,verifying its ability to handle random noise of 2%.Finally,the proposed inversion method is compared with the least squares method,and it is found that the proposed algorithm provides better boundary characterization capabilities for anomalies compared to the least squares method.The proposed method primarily addresses the issues of relying on initial models and long computation times in traditional inversion algorithms.It offers a valuable approach and methodology for future research on crosswell electrical resistivity tomography in the field of engineering.