Incremental Support Vector Regression for Steering Hot Rolling Mills

Machine learning concepts, such as the support vector machine (SVM), have been theoretically discussed in great detail, but in industrial applications still rather old-fashioned solutions are preferred. To show that learning concepts improve industrial planning, scheduling, and steering processes, a cooperation with a steel manufacturer was initiated to build a prototype pass schedule calculator for a hot rolling mill. The core of this scheduler consists of MinOver for regression estimation, an SVM training algorithm to learn physical dependencies for which no exact equations exist — such as the computation of rolling force or thickness reduction. To incorporate different types of prior knowledge, MinOver is enhanced to deal with weighted data points and two approximation layers are introduced. Furthermore, the runtime of MinOver for regression is improved by kernel caching and an optimsed evaluation of the regression function. Using common parameter selection and validation methods results in error rates close to the theoretically expectable minimum. The prototype was implemented, tested and installed at a hot rolling mill of Buderus Edelstahl GmbH as a permanent substitution for an old pass schedule database. First experiences document the promised improvements and simplifications due to the usage of machine learning concepts.