Short Term Prediction of Aluminium Strip Thickness via Support Vector Machines

Abstract

The fundamental principle of cold rolling process is the tension produced by the coiling and uncoiling motors of the rolling machine. If the tension is not properly regulated, the strip thickness will not be homogenous over the surface and even ruptures may occur. Therefore, short-term prediction of the aluminium strip thickness is important to control the tension. In this study, nonlinear time series analysis methods were applied to the recorded thickness data in order to obtain the embedding vectors with appropriate embedding dimension and time delay. For various prediction horizons, the embedding vector and corresponding thickness value pairs were used as the data set to assess the prediction performance of Support Vector Machines (SVM) with k-fold cross validation. The comparison results were given for Polynomial kernel with different exponent values, RBF kernel and Universal Pearson VII function (PUK) kernel. The SVM model with PUK kernel gave the most accurate results. The closest accuracy levels to PUK were belonging to Polynomial kernel of exponent p=3, but the time taken to build the SVM model with Polynomial kernel was very longer than the SVM model with PUK. The RBF kernel had the shortest SVM model building time with the worst accuracy levels.