<p>Non-destructive testing (NDT) techniques such as the rebound hammer (yielding rebound number, RN) and ultrasonic pulse velocity (UPV) are widely used to infer concrete strength without damaging specimens, yet their standalone accuracy remains limited. In this study, we propose a weighted-ensemble machine learning framework that simultaneously predicts RN and UPV based on mix design parameters (cement, aggregates, water–cement ratio, admixtures) and curing age. Six traditional regressors–ElasticNet, SVR, KNN, Random Forest, XGBoost, and LightGBM–were each tuned via Optuna hyperparameter optimization. Ensemble weights were derived from inverse-RMSE scores on out-of-fold validation. On a hold-out test set of 30 specimens, the ensemble achieved RMSE = 0.83 and <InlineEquation ID="IEq1"> <InlineMediaObject> <ImageObject Color="BlackWhite" FileRef="42107_2025_1455_Article_IEq1.gif" Format="GIF" Height="16" Rendition="HTML" Resolution="72" Type="Linedraw" Width="21" /> </InlineMediaObject> <EquationSource Format="TEX">$R^2$</EquationSource> <EquationSource Format="MATHML"><math> <msup> <mi>R</mi> <mn>2</mn> </msup> </math></EquationSource> </InlineEquation> = 0.94 for RN, and similarly strong performance for UPV, representing a 40–50% improvement over individual models. We further quantify prediction uncertainty with 95% bootstrap intervals (empirical coverage <InlineEquation ID="IEq2"> <InlineMediaObject> <ImageObject Color="BlackWhite" FileRef="42107_2025_1455_Article_IEq2.gif" Format="GIF" Height="16" Rendition="HTML" Resolution="72" Type="Linedraw" Width="49" /> </InlineMediaObject> <EquationSource Format="TEX">$> 90\%$</EquationSource> <EquationSource Format="MATHML"><math> <mrow> <mo>></mo> <mn>90</mn> <mo>%</mo> </mrow> </math></EquationSource> </InlineEquation>) and interpret model behavior via SHAP and Sobol sensitivity analyses. Feature attributions reveal that nonlinear interactions–particularly second-order terms of aggregate contents and the water–cement ratio–dominate predictions, while sensitivity indices confirm cement dosage and <InlineEquation ID="IEq3"> <InlineMediaObject> <ImageObject Color="BlackWhite" FileRef="42107_2025_1455_Article_IEq3.gif" Format="GIF" Height="19" Rendition="HTML" Resolution="72" Type="Linedraw" Width="32" /> </InlineMediaObject> <EquationSource Format="TEX">$w/c$</EquationSource> <EquationSource Format="MATHML"><math> <mrow> <mi>w</mi> <mo stretchy="false">/</mo> <mi>c</mi> </mrow> </math></EquationSource> </InlineEquation> ratio as primary drivers. This integrated approach offers highly accurate, well-calibrated predictions and actionable insights for NDT-based quality control in concrete construction.</p>

Weighted-ensemble machine learning for simultaneous non-destructive prediction of rebound number and ultrasonic pulse velocity in concrete

期刊论文
发表日期：

2025年7月30日

查看全文

Neha Sharma,
Arvind Dewangan,
Neelaz Singh,
Devjani Bhattacharya,
Sagar Paruthi,
Rupesh Kumar Tipu

摘要

Non-destructive testing (NDT) techniques such as the rebound hammer (yielding rebound number, RN) and ultrasonic pulse velocity (UPV) are widely used to infer concrete strength without damaging specimens, yet their standalone accuracy remains limited. In this study, we propose a weighted-ensemble machine learning framework that simultaneously predicts RN and UPV based on mix design parameters (cement, aggregates, water–cement ratio, admixtures) and curing age. Six traditional regressors–ElasticNet, SVR, KNN, Random Forest, XGBoost, and LightGBM–were each tuned via Optuna hyperparameter optimization. Ensemble weights were derived from inverse-RMSE scores on out-of-fold validation. On a hold-out test set of 30 specimens, the ensemble achieved RMSE = 0.83 and $R^2$ $R^{2}$ = 0.94 for RN, and similarly strong performance for UPV, representing a 40–50% improvement over individual models. We further quantify prediction uncertainty with 95% bootstrap intervals (empirical coverage $> 90\%$ $> 90 %$ ) and interpret model behavior via SHAP and Sobol sensitivity analyses. Feature attributions reveal that nonlinear interactions–particularly second-order terms of aggregate contents and the water–cement ratio–dominate predictions, while sensitivity indices confirm cement dosage and $w/c$ $w / c$ ratio as primary drivers. This integrated approach offers highly accurate, well-calibrated predictions and actionable insights for NDT-based quality control in concrete construction.