Machine learning-based assessment of seismic performance in fiber-reinforced polymer (FRP) retrofitted bridges

Abstract

It proposes machine learning for the seismic performance estimation of FRP-retrofitted bridges using a Mayfly Optimization Algorithm (MOA) optimized Convolutional Neural Networks (CNN). A very impressive CNN model was proposed with an R² score of nearly 0.92 for Damage Index and 0.88 for Repair Costs by achieving RMSE values very low as 0.05 and 158.11, respectively. For the classification tasks, the model obtained an overall accuracy of 91% in predicting Failure Mode with a precision of 0.89, recall 0.87, and F1-score of 0.88. The most relevant factors for the prediction of the Damage Index, as determined from the feature importance analysis using SHAP, were PGA, Bridge Type, and FRP Layer Count. Pier Height and Span Length were critical features in determining which type of failure mode exists, such as Column Shear Failure. Model discriminative explainability techniques with LIME ratified such predictions' accuracy by improving the practicality of the model's handling in real engineering scenario applications.