Machine learning driven post-impact damage state prediction for performance-based crashworthiness design of bridge piers

This study applies machine learning (ML) methods to predict post-impact damage states of reinforced concrete (RC) bridge piers under vehicle collision. 251 datasets of various vehicle-bridge collision scenarios are synthesized for training and testing six supervised ML models, including K-Nearest Neighbors (KNN), Support Vector Machine (SVM), Decision Tree, Random Forest, eXtreme Gradient Boosting Trees (XGBoost), and Artificial Neural Network (ANN). Comparisons on confusion matrices indicate that SVM, Random Forest, XGBoost, and ANN possess superior and comparable classification capabilities. ML models also achieve a much higher level of accuracy when compared with existing empirical models in the literature. Furthermore, the Shapley additive explanations (SHAP) algorithm is utilized to interpret and explain the prediction process of ML models. In particular, the Shapley value of each feature captures its positive or negative contribution for the ML model to predict each damage state, where the most influential design variables include impact speed, truck mass, engine mass, and pier diameter. To facilitate the performance-based crashworthiness design of RC bridge piers, an endto-end interactive software is devised to automatically predict impact damage states using the top three ML models against any given design scenario. Real-time interactive illustrations are also provided to elucidate the Shapley value contribution of each design parameter for the Random Forest model to reach each damage state. Finally, the final damage state is selected to have the highest likelihood of damage among the three ML model predictions.

Automated summary

This study uses machine learning (ML) methods to predict post-impact damage states of reinforced concrete (RC) bridge piers under vehicle collision. Six supervised ML models, including K-Nearest Neighbors (KNN), Support Vector Machine (SVM), Decision Tree, Random Forest, eXtreme Gradient Boosting Trees (XGBoost), and Artificial Neural Network (ANN), are trained and tested using 251 datasets. The results show that SVM, Random Forest, XGBoost, and ANN have superior and comparable classification capabilities. The Shapley additive explanations (SHAP) algorithm is used to interpret and explain the prediction process of the ML models.