Bond strength prediction of FRP reinforced concrete using soft computing techniques
Xinhua Xue,Ruiliang Zhang,Xuanquan Chen,Li Tao
DOI: https://doi.org/10.1007/s00500-024-09904-8
IF: 3.732
2024-07-19
Soft Computing
Abstract:Bond strength is the key factor to determine the strengthening performance of fiber-reinforced polymer (FRP) and concrete, so it is important to establish a bond strength prediction model. To solve the problem that the parameters considered are not comprehensive enough and the accuracy needs to be improved, multivariate adaptive regression splines (MARS) and wavelet neural network (WNN) methods are proposed to estimate the bond strength of concrete confined by FRP. A total of 749 specimens and ten input variables collected from the literature were used to develop the MARS and WNN models. To verify the prediction performance of the proposed models, an improved particle swarm optimization method based on the shrinkage factor method was used to calibrate ten empirical models. The recalibrated empirical models were then compared with MARS and WNN models. The root mean squared error (RMSE), mean absolute error (MAE), coefficient of determination (R 2 ), root relative squared error (RRSE), mean absolute percentage error (MAPE) and integral absolute error (IAE) values of MARS and WNN models were 3.454, 2.635, 0.917, 0.289, 0.153 and 0.127; 3.277, 2.516, 0.925, 0.274, 0.164 and 0.122, respectively, for the training datasets and 3.454, 2.701, 0.94, 0.245, 0.153 and 0.121; 3.652, 2.751, 0.933, 0.259, 0.184 and 0.123, respectively, for the testing datasets. While the RMSE, MAE, R 2 , RRSE, MAPE and IAE values of the training set and test set of the ten recalibrated empirical models are 7.081–11.293, 4.386–6.994, 0.410–0.701, 0.571–0.807, 0.215–0.306 and 0.212–0.312, respectively. The results show that the accuracy of MARS and WNN models is much higher than that of the ten recalibrated empirical models.
computer science, artificial intelligence, interdisciplinary applications