Prediction of bond strength of reinforced concrete structures based on feature selection and GWO-SVR model

Abstract

Bond strength, as a mechanical property of reinforced concrete (RC) structures, is a crucial factor affecting the force characteristics of RC structures. In order to assess the load-bearing and deformation capacity of RC structures, it is essential to develop a model that can accurately predict the bond strength of RC structures. Therefore, a prediction model based on Random Forest (RF) feature selection and Grey Wolf algorithm optimized (GWO) support vector regression (GWO-SVR) is synthesized in this paper. An extensive database containing 1008 bonded slip test samples was first collected, and various feature parameters were finally preprocessed to filter 935 sets of test data. Next, the optimized support vector machine was trained using the training set data using the GWO-optimized SVR method. Meanwhile, the prediction accuracy of the model was evaluated using the integrated absolute error (IAE), the coefficient of determination (R2), and the mean absolute error (MAE). The model with a good prediction effect can predict the bond strength of unknown tests. When the sample size is small and the number of features is extensive, redundant features must be filtered out before the model prediction. The model prediction after feature selection is significantly better than that before training. The results show that the GWO-SVR model proposed in this study has a higher prediction accuracy than the existing bond strength model. The prediction results of the GWO-SVR model (Test set: R2 = 0.9547, MAE = 1.2487 MPa, IAE = 12.761%) outperform benchmark models such as the multilayer perceptron (MLP) and linear regression (LR). According to the training accuracy of the regression model, it can be found that the GWO-SVR model has good generalization performance (Training set: R2 = 0.9562, MAE = 1.1203 MPa, IAE = 12.182%). In addition, different optimization-seeking algorithms, kernel function types, and machine learning models have significant effects on the prediction results of the ultimate bond strength. In addition, the compressive strength of concrete was verified to be the most sensitive variable for the bond strength of RC structures using RF and Pearson correlation coefficients. The model presented in this study can also be extended to other regression issues.