Scholarlink

<p>The Cone Penetration Test (CPT) is one of the most reliable and efficient methods for in situ site characterization, assessing ground improvement, and evaluating the potential for liquefaction. Despite CPT’s primary limitation being its inability to retrieve physical soil samples, its real-time data continuous geo-data presents a unique opportunity. With that in mind, manual interpretation of CPT data for soil behavior type (SBT<sub>n</sub>) classification can be time-consuming and subjective. The lack of physical samples makes instantaneous soil behavior type (SBT<sub>n</sub>) classification at the site even more critical. To address these gaps, this study develops and rigorously evaluates two automation methods: (i) a cloud‑based computational framework (MapCPT) that automates the Robertson‑based normalization and SBT‑chart workflow to deliver auditable SBT<sub>n</sub> profiles and 2D sections at the site location of acquisition; and (ii) supervised machine learning (ML) models (ANN, SVM, RF) trained to predict SBT<sub>n</sub> directly from raw CPT/CPTu inputs. Two high‑quality datasets were used: 100 CPTs (37,821 points) from multiple U.S. sites and 20 CPTu soundings (13,205 points) from a single site, for development, testing, and validation. The comparative analysis and validation revealed that MapCPT successfully reproduced manual calculations for <i>q</i><sub><i>t</i></sub><i>, R</i><sub><i>f</i></sub><i>, σ</i><sub><i>v0</i></sub>, <InlineEquation ID="IEq8"><EquationSource Format="TEX">\({\sigma}^{\prime}_{v0'}\)</EquationSource></InlineEquation><i> Q</i><sub><i>tn</i></sub><i>, F</i><sub><i>r</i></sub><i>, B</i><sub><i>q</i></sub><i>, I</i><sub><i>c</i></sub><i>,</i> and SBT<sub>n</sub> along depth and generated 2D soil stratigraphic sections in layer‑scale agreement with borehole‑interpreted geology. ML results showed the RF was the superior ML model, achieving accuracy of 93.22% and 94.15% for the CPT and CPTu datasets, respectively; however, generalization was stronger for the geologically coherent, single‑site CPTu dataset than for the heterogeneous, multi‑site CPT dataset. Results indicate that a transparent, physics‑based, cloud‑native workflow presently offers the most reliable path for timely automated SBT<sub>n</sub> interpretation across sites. The algorithms developed were made publicly available to facilitate verification of this work.</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

An automated cloud-based system for in-situ geotechnical site characterization using cone penetration test (CPT/CPTu)

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年6月2日

查看全文
  • Mohamed Ezzat Al-Atroush,
  • Ezz El-Din Hemdan,
  • Christopher Woyeya

摘要

The Cone Penetration Test (CPT) is one of the most reliable and efficient methods for in situ site characterization, assessing ground improvement, and evaluating the potential for liquefaction. Despite CPT’s primary limitation being its inability to retrieve physical soil samples, its real-time data continuous geo-data presents a unique opportunity. With that in mind, manual interpretation of CPT data for soil behavior type (SBTn) classification can be time-consuming and subjective. The lack of physical samples makes instantaneous soil behavior type (SBTn) classification at the site even more critical. To address these gaps, this study develops and rigorously evaluates two automation methods: (i) a cloud‑based computational framework (MapCPT) that automates the Robertson‑based normalization and SBT‑chart workflow to deliver auditable SBTn profiles and 2D sections at the site location of acquisition; and (ii) supervised machine learning (ML) models (ANN, SVM, RF) trained to predict SBTn directly from raw CPT/CPTu inputs. Two high‑quality datasets were used: 100 CPTs (37,821 points) from multiple U.S. sites and 20 CPTu soundings (13,205 points) from a single site, for development, testing, and validation. The comparative analysis and validation revealed that MapCPT successfully reproduced manual calculations for qt, Rf, σv0, \({\sigma}^{\prime}_{v0'}\) Qtn, Fr, Bq, Ic, and SBTn along depth and generated 2D soil stratigraphic sections in layer‑scale agreement with borehole‑interpreted geology. ML results showed the RF was the superior ML model, achieving accuracy of 93.22% and 94.15% for the CPT and CPTu datasets, respectively; however, generalization was stronger for the geologically coherent, single‑site CPTu dataset than for the heterogeneous, multi‑site CPT dataset. Results indicate that a transparent, physics‑based, cloud‑native workflow presently offers the most reliable path for timely automated SBTn interpretation across sites. The algorithms developed were made publicly available to facilitate verification of this work.