Predicting bond strength of corroded reinforced concrete after high-temperature exposure: A stacking model and feature selection

Abstract

The bond strength between concrete and steel reinforcement bars is crucial for determining the ultimate load-carrying capacity and serviceability of reinforced concrete (RC) structures. However, rebar corrosion and exposure to high temperatures significantly affect this bond strength. Regrettably, research on the combined effect of these two factors on bond strength is scarce, and there is a lack of a unified, precise, and efficient predictive model. This research developed a stacking model to forecast bond strength under the combined impact of high temperature and corrosion. The initial layer of the model comprises SVR, KNN, MLP, RF, GBDT, and XGBoost as base learners, with the second layer being the linear regression model. The analysis led to the following conclusions: In a comparative study, the stacked model demonstrated superior performance compared to the six base learner models (SVR, KNN, MLP, RF, GBDT, and XGBoost). Among all combinations, Stacking Model Three showed the most robust predictive performance, achieving an R² value of 0.9439, an MAE of 0.8553, an MSE of 2.2721, and an RMSE of 1.5073. Stacking Model Three surpassed XGBoost, the most effective base learner, showing improvements of 1.78 % in R², 20.69 % in MAE, 22.66 % in MSE, and 12.05 % in RMSE. The machine learning model’s enhanced reliability was further confirmed by comparison with existing models. Stacking Model Three outshone the Australian Standard model with improvements of 639.04 % in MAE, 2568.77 % in MSE, and 416.61 % in RMSE. Similarly, XGBoost, the top base learner, exceeded the Australian Standard model with gains of 486.09 % in MAE, 1964.39 % in MSE, and 354.36 % in RMSE. The outcomes of the SHapley Additive exPlanations (SHAP) affirm the interpretability and physical validity of the stacking model employed. The SHAP analysis indicates that the corrosion level of steel bars (CL) and temperature (T) are the critical factors influencing bond strength. This research underscores the practical value of SHAP feature importance in feature selection by comparing the predictive performance of stacked models with various input feature combinations, thus confirming that feature selection significantly impacts the model’s predictive accuracy. In optimizing traditional civil engineering standards and empirical formulas, the feature selection results of machine learning can be referenced.