Machine Learning for Analyzing Concrete Cover Separation in Externally Bonded FRP RC Beams

Abstract

Strengthening of concrete structures with Externally Bonded (EB) Fiber Reinforced Polymer (FRP) is one the most cost effective, efficient, and sustainable rehabilitation techniques. However, usually the ultimate tensile strength of the FRP cannot be achieved due to the premature debonding of the FRP or delamination of the concrete cover. In this last case, known as concrete cover separation (CCS), high stress concentration at the free end of FRP laminate, in combination with the shear stress at that section, cause the initiation of a shear crack that, if not controlled by the shear reinforcement, propagates in concrete just below internal reinforcement. Various analytical and empirical models have been developed to predict this phenomenon but none of them could give a failure load with reasonable accuracy. On the other hand, Machine Learning (ML) has been proved very effective in predicting behaviors that are difficult to quantify using mechanics.This paper explores ML algorithms to predict the failure load of flexural elements that undergo CCS. To this end, a database has been collected that comprises 140 four-point bending tests on beams reinforced with EB CFRP that failed by CCS. K-nearest neighbor (KNN), support vector machine (SVM), random forest (RF) and extreme gradient boosting (XGBOOST) algorithms have been applied and then ranked based on their R-squared score. R-squared scores obtained from ML are the highest compared to the analytical models described in this paper. Moreover, the mean values and standard deviation of the ratio of experimental failure load and that of calculated by the ML model are the lowest compared to the analytical models. The predictions from the ML models are found more aligned with the experimental results than the analytical models. KNN is chosen as the best ML algorithm for the prediction of failure load with an R-squared score of 0.97.