An ultrasonic-AI hybrid approach for predicting void defects in concrete-filled steel tubes via enhanced XGBoost with Bayesian optimization

This study proposes an Ultrasonic-AI Hybrid Approach for predicting void defects in concrete-filled steel tubes (CFST). Based on 3600 ultrasonic measurement samples, an Extreme Gradient Boosting (XGBoost) model was enhanced through oversampling and hyperparameter optimization via Bayesian optimization (BO-XGBoost). The BO-XGBoost model demonstrated superior performance compared to baseline models (Random Forest, AdaBoost, and Gradient Boosting Decision Tree), achieving an overall prediction accuracy of 0.92, precision and recall of 0.90, and an AUC of 0.98. SHAP (SHapley Additive exPlanations) analysis revealed that sound velocity, sound time, acoustic amplitude, concrete strength, and fly ash content were the most influential features for model predictions. This hybrid approach offers high efficiency and accuracy for void defect detection in CFST, providing a novel solution that leverages the strengths of both traditional ultrasonic methods and artificial intelligence algorithms. The method not only detects the presence of void defects but also quantifies their extent, advancing CFST inspection from qualitative analysis to quantitative assessment. • Use the ultrasonic method combined with AI to detect voids in CFST. • Employ oversampling and Bayesian optimization techniques to significantly improve model prediction accuracy. • Analyze the impact of different feature variables on prediction results using the SHAP tool. • Compared to the baseline model, the proposed model exhibits better prediction performance and interpretability.