The Research On Intelligent Prediction Of Logging Parameters And Intelligent Interpretation Of Imaging Logging Data

People can use Petrophysics and logging data of reservoirs to facilitate reservoir characterization,helping engineers to identify subsurface fluid properties and reservoir petrophysics.With the development of electronic information technology,the availability of data analysis,machine learning,and deep learning methods in the logging field has significantly increased.The intelligent interpreting method provides a new solution for logging data interpretation.These methods can automatically use the accumulated core analysis or expert interpretation to interpret the latest data.Most datadriven strategies in conventional logging data interpreting are based on supervised learning,which is helpful for the characterization of lithofacies and subsurface parameters.For imaging logging,people can use a data-driven method to identify features such as fractures and faults,estimate the porosity and permeability of formation under conventional logging data's guidance,and make formation interpretations.The research on well logs prediction mainly focuses on univariate regression methods,which use the difference between the prediction and target information as the only evaluation criterion,ignoring the analysis of uncertainties and the basis of the model prediction.Our DTC\DTS prediction uses a variety of Ensemble Learning models to solve the above problems.Based on NGBoost,the uncertainty of the prediction is given.78% of the observed DTC values are within the 80% confidence interval of the model predictions,and 62% of the observed DTS values are within the80% confidence interval.Moreover,the feature importance and dependency of the input curves are interpreted based on the SHAP explanation model.The explanations are consistent with the objective facts.We design a CNN-based imaging logging pattern extraction to better apply the supervised learning model to the interpretation of imaging logging data.The method can automatically pick up imaging logging patterns with manual interpretation results as label sets.We have evaluated the model's performance on the test set and slices of unlabeled imaging data.Models achieved more than 94% accuracy and an F1 score of more than 0.95 on the testing set.Methods confirm that reliable feature labeling methods,selections,and the dataset's quality are more important than designing the refined model structure.Finally,an intelligent clustering method is adopted for imaging logging data based on semi-supervised deep clustering.Combining the reconstruction loss of imaging slices and the clustering loss of references,the method clusters log data slices without label information.The enhanced clustering method combined with clustering reference information can obtain about 60% clustering accuracy and 0.8 F1 scores,significantly higher than the traditional clustering method.Methods may have better applicability and feasibility than supervised learning methods.