Scholarlink

<p>This study addresses the engineering need to predict multi-position underground responses during multi-hole polymer grouting in shallow voided soil, where coupled effects of grouting volume and void depth may lead to insufficient filling or excessive top-soil heave risk. To improve prediction accuracy—especially for the top soil response—an intelligent hybrid model integrating a Convolutional Neural Network, a Bidirectional Gated Recurrent Unit, and an Attention mechanism (CNN-BiGRU-Attention) is developed. The framework extracts coupled parameter patterns via CNN, learns pseudo-sequential dependencies constructed from the static input features via BiGRU (rather than true temporal evolution), and applies an attention mechanism to dynamically re-weight the learned representations for prediction. Based on indoor test data, grouting volume and void depth are used as inputs, and soil responses at the two-hole end, one-hole end, and top are set as outputs. Model performance is evaluated using the coefficient of determination (R<sup>2</sup>), mean absolute error (MAE), and root mean square error (RMSE), and is compared with single CNN, single BiGRU and SVR models. In addition, SHAP (SHapley Additive exPlanations) is employed to analyze feature contributions. Results indicate that the CNN-BiGRU-Attention model significantly outperforms the other models: the average test-set R<sup>2</sup> reaches 0.92982, while RMSE and MAE decrease to 0.18124&#xa0;kPa and 0.14386&#xa0;kPa, respectively, demonstrating improved accuracy and stability. SHAP analysis shows that grouting volume dominates the holeend responses (contribution &gt; 65%), whereas void depth has a stronger influence on the top response (contribution 63.3%), consistent with soil mechanics theory. Overall, the proposed model provides a practical tool for grouting parameter optimization and risk control.</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

Prediction of multi-hole polymer grouting effect in voided soil based on CNN-BiGRU-Attention model

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年3月18日

查看全文
  • Caiping Zhu,
  • Yibing Zhao,
  • Zhaoxu Yang,
  • Hui Luo,
  • Dan Wang,
  • Qian Wang,
  • Mingming Chen,
  • Hongchao He,
  • Peng Zhao

摘要

This study addresses the engineering need to predict multi-position underground responses during multi-hole polymer grouting in shallow voided soil, where coupled effects of grouting volume and void depth may lead to insufficient filling or excessive top-soil heave risk. To improve prediction accuracy—especially for the top soil response—an intelligent hybrid model integrating a Convolutional Neural Network, a Bidirectional Gated Recurrent Unit, and an Attention mechanism (CNN-BiGRU-Attention) is developed. The framework extracts coupled parameter patterns via CNN, learns pseudo-sequential dependencies constructed from the static input features via BiGRU (rather than true temporal evolution), and applies an attention mechanism to dynamically re-weight the learned representations for prediction. Based on indoor test data, grouting volume and void depth are used as inputs, and soil responses at the two-hole end, one-hole end, and top are set as outputs. Model performance is evaluated using the coefficient of determination (R2), mean absolute error (MAE), and root mean square error (RMSE), and is compared with single CNN, single BiGRU and SVR models. In addition, SHAP (SHapley Additive exPlanations) is employed to analyze feature contributions. Results indicate that the CNN-BiGRU-Attention model significantly outperforms the other models: the average test-set R2 reaches 0.92982, while RMSE and MAE decrease to 0.18124 kPa and 0.14386 kPa, respectively, demonstrating improved accuracy and stability. SHAP analysis shows that grouting volume dominates the holeend responses (contribution > 65%), whereas void depth has a stronger influence on the top response (contribution 63.3%), consistent with soil mechanics theory. Overall, the proposed model provides a practical tool for grouting parameter optimization and risk control.