Scholarlink

<p>The primary objective of this work is to mitigate the impact of irregular mix designs on the prediction of compressive strength in Ultra-High-Performance Concrete (UHPC) by employing a hybrid strategy of anomaly detection and ensemble learning. Initially, six ensemble algorithms, CatBoost, Random Forest, AdaBoost, Gradient Boosting, XGBoost, and LightGBM were assessed on an 810-record UHPC dataset obtained from various literature sources. CatBoost attained the maximum accuracy, with R² = 0.98, MAE = 4.59, and RMSE = 6.55. Feature relevance analysis employing SHAP values and Partial Dependence Plots indicated curing age, fiber content, and cement content as the most influential characteristics. Subsequently, unsupervised anomaly detection techniques, such as Isolation Forest, One-Class SVM, and DBSCAN, were employed to find abnormal mixture compositions. DBSCAN was chosen for its exceptional capacity to identify low-density, irregular data points, uncovering 69 abnormalities. Eliminating these anomalies and reassessing CatBoost enhanced the model’s performance to R² = 0.99, MAE = 4.39, and RMSE = 6.47, significantly improving predictions in both the lower and upper strength ranges. The results indicate that focused anomaly identification markedly enhances the accuracy, reliability, and resilience of predictions about the compressive strength of UHPC. The suggested anomaly detection with tuned ensembles, using DBSCAN–CatBoost to robustly model UHPC strength across heterogeneous mixes, significantly outperforming conventional models.</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

Mitigating mix design anomalies in ultra-high performance concrete strength prediction through anomaly detection and ensemble learning

  • 期刊论文

  • 发表日期：

    2025年10月8日

查看全文
  • Ganapathy Ganesh Prabhu,
  • Lakshmi Keshav,
  • B. Narendra Kumar,
  • Kolli Ramujee,
  • Mary Subaja Christo,
  • Subhi A. Ali,
  • Gasim Hayder,
  • Gokulan Ravindiran

摘要

The primary objective of this work is to mitigate the impact of irregular mix designs on the prediction of compressive strength in Ultra-High-Performance Concrete (UHPC) by employing a hybrid strategy of anomaly detection and ensemble learning. Initially, six ensemble algorithms, CatBoost, Random Forest, AdaBoost, Gradient Boosting, XGBoost, and LightGBM were assessed on an 810-record UHPC dataset obtained from various literature sources. CatBoost attained the maximum accuracy, with R² = 0.98, MAE = 4.59, and RMSE = 6.55. Feature relevance analysis employing SHAP values and Partial Dependence Plots indicated curing age, fiber content, and cement content as the most influential characteristics. Subsequently, unsupervised anomaly detection techniques, such as Isolation Forest, One-Class SVM, and DBSCAN, were employed to find abnormal mixture compositions. DBSCAN was chosen for its exceptional capacity to identify low-density, irregular data points, uncovering 69 abnormalities. Eliminating these anomalies and reassessing CatBoost enhanced the model’s performance to R² = 0.99, MAE = 4.39, and RMSE = 6.47, significantly improving predictions in both the lower and upper strength ranges. The results indicate that focused anomaly identification markedly enhances the accuracy, reliability, and resilience of predictions about the compressive strength of UHPC. The suggested anomaly detection with tuned ensembles, using DBSCAN–CatBoost to robustly model UHPC strength across heterogeneous mixes, significantly outperforming conventional models.