Machine learning-based models for predicting the progressive collapse resistance of truss string structures

Evaluating the progressive collapse resistance of truss string structures (TSSs) in the context of key member failure presents a significant challenge, particularly when this indicator is crucial during structural design and performance evaluation processes. Fortunately, machine learning (ML) methods can establish complex and nonlinear relationships between input and output variables during structural performance evaluation. In this study, firstly, 20 practical projects of TSSs are surveyed to determine the critical design parameters and statistical attributes of TSSs, which provide a realistic basis for the subsequent establishment of a sample database of the progressive collapse resistance of TSSs. Secondly, 464 models of TSS are generated by random sampling through MATLAB, and the progressive collapse resistance of TSSs under the failure of cable or lower chord at the support is analyzed based on the validated finite element model, followed by the establishment of the progressive collapse resistance database. Moreover, based on the results of the analysis, the effect of different parameters on the progressive collapse resistance of TSSs is discussed for the two typical failure scenarios. Subsequently, five ML models are introduced, and a framework for predicting the progressive collapse resistance is given. Next, the prediction results of different ML models were compared. The results show that the extreme gradient boosting (XGBoost) model performs best for the failure of cable or bottom chord at the support, with R2 values of 0.988, 0.920, and 0.972, 0.801 for the training and testing sets, respectively. Finally, the physical and quantitative interpretation of the progressive collapse resistance of TSSs obtained from XGBoost model-based predictions under two typical failure scenarios is conducted using the Shapley additive explanations (SHAP) method.