Machine learning-based and interpretable models for predicting the resistance and probability of progressive collapse of steel frame-composite floor structures

Abstract

Steel frame-composite floor (SF-CF) structures are particularly susceptible to progressive collapse due to their complex load distribution and potential vulnerabilities at critical joints, making them a high-risk system in the event of localized failure. This study aims to investigate the methods for predicting and analyzing the progressive collapse behavior of SF-CF structures. A theoretical and predictive model was developed based on the Minimal Substructure Mechanics Model and the Collapse Probability Prediction Method Based on the Randomness of Structural Load (CPPM-RSL). Machine learning (ML) techniques from artificial intelligence (AI) were employed to predict and analyze the progressive collapse behavior probabilistically. Initially, a dataset comprising various scenarios of column failure was established. The performance of five typical ML models was assessed and compared based on this dataset, focusing on the robustness and generalization capability. Results revealed that the XGBoost model demonstrated inferior performance compared to the other four models, particularly regarding generalization capability. More refined hyperparameter optimization was performed to attempt to improve the performance of the XGBoost model. Grid search was employed to obtain the optimum hyperparameter configuration. K-Fold cross-validation was employed to further examine the generalization capability of the model under the optimal hyperparameter configuration. Additionally, reducing model complexity by decreasing the number of features was also attempted in response to the overfitting observed in the XGBoost model. The results indicated that these approaches did not significantly enhance the performance of the XGBoost model. Moreover, interpretable models based on the Shapley Additive exPlanations (SHAP) method and five ML models were developed. The importance of features in predicting structural collapse and their impact mechanisms were elucidated, providing recommendations to boost the structural resistance to collapse. It was also indicated that the employment of weak learners in XGBoost yields predictive values less than true values despite the obtained values being raised as much as possible in the SHAP values of the XGBoost model. Optimizing the structural geometric characteristics proved to be more effective than improving the properties of material in increasing the collapse resistance of structures. In conclusion, this study attempts to apply ML methods for predicting structural collapse behavior and provides a potential approach for predicting progressive collapse.