Scholarlink

<p>This study proposed a knowledge-guided sequential learning framework for soil stratification based on cone penetration test (CPT) data. Empirical knowledge of CPT interpretation was distilled into curve-morphology indicators that quantified single-curve fluctuations and peak prominence, as well as inter-curve relative strength and crossing frequency. These indicators were encoded into a differentiable knowledge-based rule layer, where learnable gating functions transformed heuristic curve–soil relationships into knowledge-augmented soft labels. A multi-scale feature extraction scheme and a bidirectional long short-term memory network were used to learn depth-dependent CPT representations from local transition-sensitive features and mid-scale morphological features, which enabled joint prediction of soil-layer boundaries and soil types. The framework was optimized using a composite loss function that combined supervised learning with soft-label regularization to guide predictions toward stratigraphically consistent results. The method was validated using a cross-project dataset with 265 CPT soundings and 102016 records. It achieved F1-scores of 0.762 for boundary identification and 0.829 for soil-type classification and outperformed the variant without the knowledge-based rule layer, the random forest baseline, and the deep neural network baseline. The results demonstrated that incorporating differentiable geotechnical knowledge into sequential CPT learning improved stratification accuracy, profile-level stability, and cross-project robustness.</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

Knowledge-guided Sequential Learning for CPT-based Soil Stratification in Site Investigation

  • 期刊论文

  • 发表日期：

    2026年7月14日

查看全文
  • Deming Xu,
  • Peng Li,
  • Aidong Li,
  • Guohe Li,
  • Chunyu Qi,
  • Ran Wang,
  • Dou Zhao,
  • Dazhong Huang,
  • Zelian Chen

摘要

This study proposed a knowledge-guided sequential learning framework for soil stratification based on cone penetration test (CPT) data. Empirical knowledge of CPT interpretation was distilled into curve-morphology indicators that quantified single-curve fluctuations and peak prominence, as well as inter-curve relative strength and crossing frequency. These indicators were encoded into a differentiable knowledge-based rule layer, where learnable gating functions transformed heuristic curve–soil relationships into knowledge-augmented soft labels. A multi-scale feature extraction scheme and a bidirectional long short-term memory network were used to learn depth-dependent CPT representations from local transition-sensitive features and mid-scale morphological features, which enabled joint prediction of soil-layer boundaries and soil types. The framework was optimized using a composite loss function that combined supervised learning with soft-label regularization to guide predictions toward stratigraphically consistent results. The method was validated using a cross-project dataset with 265 CPT soundings and 102016 records. It achieved F1-scores of 0.762 for boundary identification and 0.829 for soil-type classification and outperformed the variant without the knowledge-based rule layer, the random forest baseline, and the deep neural network baseline. The results demonstrated that incorporating differentiable geotechnical knowledge into sequential CPT learning improved stratification accuracy, profile-level stability, and cross-project robustness.