Comparisons of Different Data-Driven Modeling Techniques for Predicting Tensile Strength of X70 Pipeline Steels

Abstract

Mechanical property prediction for X70 pipeline steels attracts people’s attention because of maintaining high process stability and controlling production quality. Data-driven model is widely used and has the advantage of little professional knowledge requirement compared with phenomenological model. This paper introduced two new modeling techniques, namely ridge regression (RR) and random forest (RF). As a case, tensile strength prediction model of X70 pipeline steels was established and comparisons of different data-driven models, including the two new techniques and the already extensively used stepwise regression (SR), Bayesian regularization neural network (BRNN), radial-basis function neural network (RBFNN) and support vector machine (SVM), were made. The results show that all the models have reached good accuracies with relative error of ± 7%. On account of the excellent nonlinear fitting capability, models established by using intelligent algorithms (BRNN, RBFNN, SVM and RF) obtain better performance than multiple linear regression (SR and RR). Among the six models, RR provides a visualizing approach of the variable selection for multiple linear regression and RF achieves the best performance (R = 0.95 and MSE = 278.7 MPa2) on this data set.