Thickness prediction of thin strip cold rolling based on VBGM-RBF

Abstract

The thickness error of the thin strip determines the product quality of thin strip. As a typical dynamic continuous non-Gaussian process, the thin strip has complex irregular noise, and the general prediction method will produce significant errors. According to the rolling characteristics of thin strip, a radial basis function neural network based on variational Bayesian Gaussian mixture clustering algorithm (VBGM-RBF) is proposed to predict the thickness of cold-rolled thin strip. The production data of 160000 sets of cold-rolled strips are obtained by removing outliers according to the data characteristics by the combination of manual selection, isolated forest algorithm, and Bessel formula. Considering the mean square error (MSE) and correlation coefficient (R$^{2}$), the parameter settings of RBF neural network and optimization algorithm are studied to obtain the optimal model. This paper compares the prediction performance of RBF neural network optimized by different clustering algorithms, tests the prediction error of adding noise samples in the neural network, and analyzes the influencing factors of strip thickness. The results show that VBGM-RBF has the highest prediction accuracy compared with some traditional RBF neural networks. The MSE of the VBGM-RBF model is $0.5053$ $\mu$m$^{2}$, the maximum error percentage (MPE) is 1.842$\%$, and the absolute error of 99.40$\%$ of the predicted data is less than 2$\mu$m. In the analysis of the factors affecting the thickness, the model is consistent with the physical law of thin strip rolling. VBGM-RBF model has strong learning ability and generalization performance and can be well-applied to the production of thin strip steel.