Predicting tensile strength of spliced and non-spliced steel bars using machine learning- and regression-based methods

Abstract

Mechanical properties of steel reinforcement bars, which have a critical effect in the overall performance of reinforced concrete (RC) structures, should be reported and assessed before being used in structural elements. Determining bars’ properties could be time-consuming and expensive specifically in the case of incorporating splices. Therefore, this study aims to predict tensile strength of bars using machine learning-based methods including nonlinear regression, ridge regression and artificial neural network. To this end, a comprehensive database including over 200 tests on non-spliced and spliced steel bars by mechanical couplers was collected from the available peer-reviewed international publications. Bar size, splice method, steel grade, temperature and splice characteristics (length and outer diameter of couplers) were the input parameters considered for predicting tensile strength. The efficiency of the models was evaluated through Taylor diagram and common performance metrics namely coefficient of determination (R2), root mean square error (RMSE), mean absolute error (MAE) and mean absolute percentage error (MAPE). The results demonstrated that the predicted values agreed well with the actual values reported in the experimental studies used for collecting the database. A parametric study was also conducted in order to examine the influence of coupler length, coupler outer diameter and temperature on the tensile strength of spliced bars. Based on the parametric study results, three different equations were suggested for calculating tensile strength of spliced bars using the mentioned parameters. The outcomes of this study can assist practitioners to effectively and accurately estimate tensile strength of spliced and non-spliced steel bars in reinforced concrete structures without the need to carry out expensive and timely physical tests.