Study on the relationship of design parameters and damage modes for RC slabs subjected to large-scale hard missile impacts base on task-driven approach

Abstract

This study proposes a novel hybrid ensemble learning (EL) model that integrates CatBoost and SMOTE-NC algorithms to evaluate the damage modes of reinforced concrete (RC) slabs under large-scale hard missile impacts, overcoming the limitations of conventional approaches. To develop the optimization model, a database was constructed using publicly published experimental data from 304 large-scale missile impact tests conducted on RC plates. Furthermore, the SMOTE-NC oversampling algorithm is employed to optimize the database and address the issue of class imbalance. Subsequently, four ensemble machine learning models and empirical formulas are assessed. The findings indicate that the hybrid ensemble learning model combining CatBoost and SMOTE-NC exhibits superior performance, showcasing robustness and efficacy. Furthermore, through the application of SHAP analysis, it was revealed that velocity, thickness, concrete compressive strength, missile mass, reinforcement ratio, and missile diameter are the most influential variables, and they interact with each other in significant ways. Notably, the analysis indicated that once the concrete compressive strength reaches a certain value, further improvements may have a detrimental effect on penetration resistance. Additionally, the impact of yield strength on slab performance becomes more pronounced as the slab thickness decreases. Drawing upon these findings, a proposed approach for damage assessment and protection design utilizing an interpretable hybrid EL model was introduced. However, it is crucial to acknowledge that the applicability of this model is confined to RC targets with finite thickness, specifically engineered to withstand large-scale hard missile impacts.