Prediction and Reliability Evaluation of Flexural Capacity of Post-Fire Corroded RC Beams Based on Improved GMM-VSG Model and Ensemble Learning Model

Evaluating the flexural capacity of post-fire corroded RC beams is crucial for post-disaster repair and structural reinforcement. In this study, a machine learning (ML)-based prediction model is developed to estimate the flexural capacity of RC beams affected by nonlinear degradation mechanisms such as fire damage and corrosion. In order to construct the dataset required for ML model training, test data considering fire and corrosion scenarios are collected to form the initial dataset. Regarding the issue of small-sample, improvements are made to the Gaussian Mixture Model Variable Sampling Generation (GMM-VSG) to expand the dataset. The GMM-VSG model has been added with physical constraints and dynamic weight adjustments to generate data more suitable for this work. The improved GMM-VSG is considered to effectively expand the sample space of the dataset and improved its performance in ML model. Based on the expanded dataset, four single ML models and two ensemble learning models are employed to develop the predictive model. Among single ML models, the XGBoost model demonstrates the highest predictive accuracy, with a coefficient of determination (R²) of 97.44%. After XGBoost is combined with MLP to form an ensemble learning model, R² increases by 1.67% to 99.11%. The SHapley Additive exPlanations (SHAP) method is applied to interpret the XGB-MLP model's parameter impacts and to facilitate parameter analysis. Additionally, to achieve the target reliability index, capacity reduction coefficients are proposed for corroded RC beams subjected to natural cooling and water cooling conditions. Finally, an easy-to-use and interactive user interface was developed to support the practical application of the proposed model.