Scholarlink

<p>Three-dimensional prediction of vertical settlement, mean effective stress and excess pore-water pressure beneath a geosynthetic-reinforced pile-supported (GRPS) embankment on soft alluvial ground was investigated using a slim 3D U-Net. A reference numerical model of a high-speed railway embankment with eight floating concrete piles, a basal geogrid and three soft-soil layers was discretised onto a 64 × 64 × 48 voxel grid spanning 248,610 mesh nodes across fourteen staged construction and consolidation phases up to 500 days. The 3D U-Net, with about 13&#xa0;million parameters, jointly predicted the three target fields from 50 to 500 stochastic sparse virtual sensors emulating settlement plates, piezometers and inclinometers. Three-fold spatial-block cross-validation produced mean coefficients of determination of 0.981, 0.902 and 0.983 with normalised root-mean-square error below 4%. Compared with an ordinary kriging baseline at the same sensor count, the network reduced settlement error from 2.66% to 1.93% and raised the stress-field R² from 0.07 to 0.90. A deployment scenario using 55 instruments yielded full-grid R² above 0.97 on all targets. Generalisation across embankment configurations, constitutive uncertainty and realistic sensor degradation were identified as limitations requiring further field validation.</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

Deep Learning Prediction of Settlement and Stress Fields in Geosynthetic Reinforced Pile Supported Soft Ground

  • 期刊论文

  • 发表日期：

    2026年7月30日

查看全文
  • Ha Bui Manh,
  • Tuan Nguyen Anh,
  • Duc Nguyen Van

摘要

Three-dimensional prediction of vertical settlement, mean effective stress and excess pore-water pressure beneath a geosynthetic-reinforced pile-supported (GRPS) embankment on soft alluvial ground was investigated using a slim 3D U-Net. A reference numerical model of a high-speed railway embankment with eight floating concrete piles, a basal geogrid and three soft-soil layers was discretised onto a 64 × 64 × 48 voxel grid spanning 248,610 mesh nodes across fourteen staged construction and consolidation phases up to 500 days. The 3D U-Net, with about 13 million parameters, jointly predicted the three target fields from 50 to 500 stochastic sparse virtual sensors emulating settlement plates, piezometers and inclinometers. Three-fold spatial-block cross-validation produced mean coefficients of determination of 0.981, 0.902 and 0.983 with normalised root-mean-square error below 4%. Compared with an ordinary kriging baseline at the same sensor count, the network reduced settlement error from 2.66% to 1.93% and raised the stress-field R² from 0.07 to 0.90. A deployment scenario using 55 instruments yielded full-grid R² above 0.97 on all targets. Generalisation across embankment configurations, constitutive uncertainty and realistic sensor degradation were identified as limitations requiring further field validation.