Predicting the drift capacity of precast concrete columns using explainable machine learning approach. (1st May 2023)

• Record Type: Journal Article
• Authors: Wang, Zhen; Liu, Tongxu; Long, Zilin; Wang, Jingquan; Zhang, Jian
• Is Part Of: Engineering structures. Volume 282(2023)
• Publication Date: 2023-05-01
• Subjects: Precast concrete columns -- Drift capacity -- EXtreme Gradient Boosting (XGBoost) -- SHapley Additive exPlanations (SHAP) -- Machine learning; Structural engineering -- Periodicals; Structural analysis (Engineering) -- Periodicals; Construction, Technique de la -- Périodiques; Génie parasismique -- Périodiques; Pression du vent -- Périodiques; Earthquake engineering; Structural engineering; Wind-pressure; Periodicals; 624.105
• DOI: 10.1016/j.engstruct.2023.115771
• Languages: English
• ISSNs: 0141-0296

Abstract

Highlights: Explainable machine learning approach is firstly used to study the DC of PCCs. XGBoost model with feature selection is proposed to predict the DC of PCCs. The proposed model is experimentally verified and compared with empirical formulas. SHAP analysis is used to globally and individually interpret the proposed model. The proposed model can accurately and reliably predict the DC of PCCs. Abstract: Accurately and reliably predicting the drift capacity (DC) of concrete columns is crucial for the seismic design and damage evaluation of structures. Despite precast concrete columns (PCCs) being applied to seismic zone, the method for predicting the DC of PCCs is still scarce owing to its high complexity. This study aims to develop a machine learning-based model for predicting the DC of PCCs using eXtreme Gradient Boosting (XGBoost) algorithm. A DC database of PCCs was assembled from existing literature which involves 177 flexural-dominant specimens with 44 features. A model establishment procedure was carried out to develop XGBoost models, including data cleaning, feature selection, and hyperparameter optimization. The models with and without feature selection were then validated by test results, and the former as the proposed model was further compared with existing empirical formulas and explained by global interpretation, individual interpretation, and feature dependency using SHapley Additive exPlanations (SHAP). Results show that XGBoost algorithm can develop adequate models to predict the DC of PCCs with high accuracy and great reliability. The feature selection method is effective to identify 11 dominant features and delete the rest for the proposed model. The empirical formulas are not suitable to directly predict the DC of PCCs. Global interpretation presents the influence of the 11 dominant features on the DC of PCCs. Feature dependency proves that there are high dependencies between these features. This study firstly develops special models for predicting the DC of PCCs using a machine learning approach, as well as systematically identifies and discusses the effects of various features on the DC of PCCs.