Accurate lithofacies classification in low-permeability sandstone reservoirs remains challenging due to class imbalance in well-log data and the difficulty of the modeling vertical lithological dependencies.Traditiona...Accurate lithofacies classification in low-permeability sandstone reservoirs remains challenging due to class imbalance in well-log data and the difficulty of the modeling vertical lithological dependencies.Traditional core-based interpretation introduces subjectivity,while conventional deep learning models often fail to capture stratigraphic sequences effectively.To address these limitations,we propose a hybrid CNN–GRU framework that integrates spatial feature extraction and sequential modeling.Heat Kernel Imputation is applied to reconstruct missing log data,and Borderline SMOTE(BSMOTE)improves class balance by augmenting boundary-case minority samples.The CNN component extracts localized petrophysical features,and the GRU component captures depth-wise lithological transitions,to enable spatial-sequential feature fusion.Experiments on real-well datasets from tight sandstone reservoirs show that the proposed model achieves an average accuracy of 93.3%and a Macro F1-score of 0.934.It outperforms baseline models,including RF(87.8%),GBDT(81.8%),CNN-only(87.5%),and GRU-only(86.1%).Leave-one-well-out validation further confirms strong generalization ability.These results demonstrate that the proposed approach effectively addresses data imbalance and enhances classification robustness,offering a scalable and automated solution for lithofacies interpretation under complex geological conditions.展开更多
China’s credit bond market has rapidly expanded in recent years.However,since 2014,the number of credit bond defaults has been increasing rapidly,posing enormous potential risks to the stability of the financial mark...China’s credit bond market has rapidly expanded in recent years.However,since 2014,the number of credit bond defaults has been increasing rapidly,posing enormous potential risks to the stability of the financial market.This study proposed a deep learning approach to predict credit bond defaults in the Chinese market.A convolutional neural network(CNN)was selected as the classification model and to reduce the extreme imbalance between defaulted and non-defaulted bonds,and a generative adversarial network(GAN)was used as the oversampling model.Based on 31 financial and 20 non-financial indicators,we collected Wind data on all credit bonds issued and matured or defaulted from 2014 to 2021.The experimental results showed that our GAN+CNN approach had superior predictive performance with an area under the curve(AUC)of 0.9157 and precision of 0.8871 compared to previous research and other commonly used classification models-including the logistic regression,support vector machine,and fully connected neural network models-and oversampling techniques-including the synthetic minority oversampling technique(SMOTE)and Borderline SMOTE model.For one-year predictions,indicators of solvency,capital structure,and fundamental properties of bonds are proved to be the most important indicators.展开更多
基金supported by the Langfang Science and Technology Program with self-raised funds under the project“Application of Deep Learning-Based Joint Well-Seismic Analysis in Lithology Prediction”(Project No.2024011013)the Science and Technology Innovation Program for Postgraduate students in IDP subsidized by Fundamental Research Funds for the Central Universities,under the project“Research on CNN Algorithm Enhanced by Physical Information for Lithofacies Prediction in Tight Sandstone Reservoirs”(Project No.ZY20250328).
文摘Accurate lithofacies classification in low-permeability sandstone reservoirs remains challenging due to class imbalance in well-log data and the difficulty of the modeling vertical lithological dependencies.Traditional core-based interpretation introduces subjectivity,while conventional deep learning models often fail to capture stratigraphic sequences effectively.To address these limitations,we propose a hybrid CNN–GRU framework that integrates spatial feature extraction and sequential modeling.Heat Kernel Imputation is applied to reconstruct missing log data,and Borderline SMOTE(BSMOTE)improves class balance by augmenting boundary-case minority samples.The CNN component extracts localized petrophysical features,and the GRU component captures depth-wise lithological transitions,to enable spatial-sequential feature fusion.Experiments on real-well datasets from tight sandstone reservoirs show that the proposed model achieves an average accuracy of 93.3%and a Macro F1-score of 0.934.It outperforms baseline models,including RF(87.8%),GBDT(81.8%),CNN-only(87.5%),and GRU-only(86.1%).Leave-one-well-out validation further confirms strong generalization ability.These results demonstrate that the proposed approach effectively addresses data imbalance and enhances classification robustness,offering a scalable and automated solution for lithofacies interpretation under complex geological conditions.
基金supported in part by the Emerging Interdisciplinary Project of Central University of Finance and Economics,Beijing,China.
文摘China’s credit bond market has rapidly expanded in recent years.However,since 2014,the number of credit bond defaults has been increasing rapidly,posing enormous potential risks to the stability of the financial market.This study proposed a deep learning approach to predict credit bond defaults in the Chinese market.A convolutional neural network(CNN)was selected as the classification model and to reduce the extreme imbalance between defaulted and non-defaulted bonds,and a generative adversarial network(GAN)was used as the oversampling model.Based on 31 financial and 20 non-financial indicators,we collected Wind data on all credit bonds issued and matured or defaulted from 2014 to 2021.The experimental results showed that our GAN+CNN approach had superior predictive performance with an area under the curve(AUC)of 0.9157 and precision of 0.8871 compared to previous research and other commonly used classification models-including the logistic regression,support vector machine,and fully connected neural network models-and oversampling techniques-including the synthetic minority oversampling technique(SMOTE)and Borderline SMOTE model.For one-year predictions,indicators of solvency,capital structure,and fundamental properties of bonds are proved to be the most important indicators.
