Borehole Correction For Three-Dimensional Induction Logging Using Deep Learning

Since traditional logging tools cannot characterize the anisotropic formation conductivity well in thin sand-shale interlayers,the 3D induction logging tool,as a new type of logging tool,can not only characterize the anisotropic The conductivity value of the heterogeneous formation can also measure information such as the formation dip angle and the azimuth angle of the instrument.Due to the complexity of the coil system of the 3D logging tool,the coplanar coil is greatly affected by wellbore environmental parameters such as wellbore radius,tool eccentricity,formation dip angle,and mud conductivity,while the coplanar coil mainly characterizes the vertical conductivity of the formation.Therefore,it is very necessary to correct the wellbore environmental factors for the 3D induction logging data.The main purpose of this research is to eliminate the influence of instrument eccentricity and mud conductivity on the logging response.Traditional wellbore correction methods require a lot of labor costs and are difficult to achieve high-precision correction results.In order to accurately present the nonlinear relationship between formation parameters and logging response,this thesis uses a deep learning algorithm to correct the logging response.Firstly,the correctness of the three-dimensional numerical simulation software is verified by comparing the results of the three-dimensional numerical simulation software and the COMSOL modeling software of the three groups of models.Secondly,multiple sets of appropriate model parameters are selected,and the electromagnetic field response values of each set of model parameters are calculated using 3D numerical simulation software,which completes the establishment of the wellbore calibration library.A large number of diverse sample data sets can be obtained from the wellbore calibration library,and the logging component with a high correlation with formation parameters is analyzed by using the Pearson correlation coefficient,and the convolutional neural network is trained using the logging component and its hyperparameters are adjusted.Until the loss function value reaches the minimum,the model training process is completed.Finally,the wellbore environmental parameters can be predicted by using the logging component in the observation data that has a high correlation with formation parameters as the input of the convolutional neural network.The difference between the values is used as the wellbore correction value,which can eliminate the adverse effects of the wellbore environmental parameters on the logging response.According to the experimental results,the correction accuracy of wellbore environment parameters using convolutional neural network can be as high as 90% for both simple and complex formation models,which shows that the network can better eliminate the influence of wellbore environment.Moreover,the network runs fast in the process of wellbore correction,and has good robustness and generalization ability.Compared with traditional correction methods,it has the advantages of automatic feature extraction,high efficiency,strong adaptability and high prediction accuracy.