Scholarlink

<p>The uniaxial compressive strength (UCS) plays a significant role in earth sciences and engineering operations for effective drilling, hydraulic fracturing, and stability analysis in underground mining operations. This study aims to investigate the performance of predictive models, and to conduct sensitivity analysis for feature selection to assess the impact of individual variables on UCS prediction. In this study, novel deep learning techniques of convolutional neural network (CNN), and transformer are adopted to create resilient models for forecasting the UCS of sedimentary rocks using readily available predictor variables, including Schmidt hammer rebound number (SHR), point load strength index, neutron porosity, and compressional wave velocity. Additionally, support vector machine (SVM) and feed-forward neural network (FNN)-based models are adopted to obtain UCS and also outcomes are compared for reliability using coefficient of determination (R<sup>2</sup>), mean absolute error (MAE), and root mean square error (RMSE). A Friedman statistical test and Taylor diagram are also utilized to investigate the significance of the alterations in model performance. The results indicate that all predictive models performed well, with the transformer model outperforming the others. A feature importance investigation is conducted using shapley additive explanation analysis and four filter methods of correlation matrix, Fisher scores, Chi-Square test, and mutual information analysis. Based on the simulated results, the transformer model demonstrated as R<sup>2</sup> values of 0.99 and 0.95, along with MAE values of 0.73 and 1.3 and RMSE values of 1.20 and 1.58 for training and testing datasets&#xa0;respectively, indicating superior predictive accuracy with precision and minimal error, comparing with other models of CNN, FNN and SVM., It is found that the&#xa0;SHR is most significant predictor variable, whereas point load index is least contributed input variable to predict rock UCS across all other models.&#xa0;Taking novel approaches into account this study provides valuable insights into sedimentary rock behavior, offering improved UCS estimation for wellbore stability analysis and hydrocarbon exploration in the petroleum and mining industry.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Data-driven resilient model development and feature selection for rock compressive strength prediction using machine learning and transformer techniques

  • 期刊论文

  • 发表日期：

    2025年2月25日

查看全文
  • Md. Shakil Rahaman,
  • Mohammad Islam Miah

摘要

The uniaxial compressive strength (UCS) plays a significant role in earth sciences and engineering operations for effective drilling, hydraulic fracturing, and stability analysis in underground mining operations. This study aims to investigate the performance of predictive models, and to conduct sensitivity analysis for feature selection to assess the impact of individual variables on UCS prediction. In this study, novel deep learning techniques of convolutional neural network (CNN), and transformer are adopted to create resilient models for forecasting the UCS of sedimentary rocks using readily available predictor variables, including Schmidt hammer rebound number (SHR), point load strength index, neutron porosity, and compressional wave velocity. Additionally, support vector machine (SVM) and feed-forward neural network (FNN)-based models are adopted to obtain UCS and also outcomes are compared for reliability using coefficient of determination (R2), mean absolute error (MAE), and root mean square error (RMSE). A Friedman statistical test and Taylor diagram are also utilized to investigate the significance of the alterations in model performance. The results indicate that all predictive models performed well, with the transformer model outperforming the others. A feature importance investigation is conducted using shapley additive explanation analysis and four filter methods of correlation matrix, Fisher scores, Chi-Square test, and mutual information analysis. Based on the simulated results, the transformer model demonstrated as R2 values of 0.99 and 0.95, along with MAE values of 0.73 and 1.3 and RMSE values of 1.20 and 1.58 for training and testing datasets respectively, indicating superior predictive accuracy with precision and minimal error, comparing with other models of CNN, FNN and SVM., It is found that the SHR is most significant predictor variable, whereas point load index is least contributed input variable to predict rock UCS across all other models. Taking novel approaches into account this study provides valuable insights into sedimentary rock behavior, offering improved UCS estimation for wellbore stability analysis and hydrocarbon exploration in the petroleum and mining industry.