A Feature Selection Framework for Ground Motion Intensity Measures in the Rapid Seismic Damage Prediction of Structures Based on Ensemble Learning

In traditional fragility analysis for regional damage assessment, low-dimensional intensity measures (IMs), such as peak ground acceleration (PGA), are typically chosen to characterize ground motion (GM) and predict structural damage based on the assumption of a logarithmic linear relationship. However, previous studies have demonstrated that it is very difficult to construct an accurate mathematical model between the potential IMs and the engineering demand parameter (EDP). In recent years, machine learning methods have provided a completely new approach to accurately predicting structural damage. Although all potential IMs can be inputted directly into the appropriate machine learning algorithm to generate a seismic damage predictor, the focus is usually only on the effectiveness of input parameters (i.e., accuracy), with little consideration of efficiency (e.g., time cost on calculating input IMs, training model, and prediction). Additionally, the models trained by traditional machine learning algorithms tend to exhibit preferences for specific input variables and data, particularly when dealing with the inherent imbalance in the GM database. To solve the presented problems, a feature selection framework of IMs is proposed to predict structural damage based on ensemble learning. Within the proposed methodology of the framework, the redundant and irrelevant IMs are removed one by one without sacrificing accuracy by evaluating the effectiveness of the potential IMs, thus reducing the model complexity and improving the efficiency of the trained model. Meanwhile, an approach from the perspective of sufficiency is adopted to verify the selected IMs by quantifying the information obtained about the EDP by an IM combination relative to another IM combination. Based on the above method, the optimal IM combination that can make the trained models achieve high efficiency and accuracy is determined. Additionally, to solve the problem of trained models tending to exhibit preferences for specific input variables and data, the ensemble learning framework and 10-fold cross-validation techniques are introduced by combining multiple basic machine learning algorithms. Finally, to illustrate the proposed methodology, 20 alternative IMs are selected and the feature selection method is tested to eliminate the redundant and irrelevant IMs for five types of structure. The results all show that the method established in this study can effectively eliminate redundant and irrelevant IMs without sacrificing accuracy, achieving an average accuracy rate above 92%. Additionally, the efficiency of the optimal models in the time cost on calculating input IMs, training model, and prediction is improved by about three, 10, and two times compared to models trained based on all alternative IMs. Therefore, the method proposed in this paper has a promising potential in predicting structural damage during emergencies.