A machine learning approach for predicting the outer diameter of seamless steel pipe in hot rolling

Abstract

Seamless steel pipes are widely used in industries such as petroleum, natural gas and chemical engineering, where stringent requirements are placed on the precision of the final outer diameter. Due to the complexity of the three deformation stages in hot rolling – piercing, rolling and reducing – and the strong coupling and nonlinearity among numerous process parameters, traditional analytical models have difficulty accurately predicting the final outer diameter, which increases the challenge of process control. To improve outer-diameter prediction under frequent specification changes, this study adopts a deviation-based learning strategy in which the model learns the deviation between the actual and target outer-diameter values. In this framework, the final outer diameter is obtained by adding the predicted deviation to the target outer diameter. Parameters from piercing, rolling and reducing are comprehensively considered, and two proxy variables – historical rolling length and interval time – are introduced to reflect the effects of roll wear and thermal deformation. Based on these inputs, a support vector regression (SVR) model with a hybrid kernel combining polynomial and radial basis function kernels is developed. Experiments on industrial data show that the proposed model achieves an R² of 0.9742 and a mean absolute error of 0.1304 mm for deviation prediction, significantly outperforming single-kernel SVR and other benchmark algorithms. The proposed approach supports online outer-diameter prediction and process-parameter optimisation in hot-rolled seamless pipe production.