Modeling blast furnace productivity using support vector machines

Productivity of a modern generation blast furnace was modeled with the help of a leading supervised learning tool viz. Support Vector Machines in the form of (1) minimum error, maximum margin classification function in binary setting of productivity classes (low/high) and (2) the class-specific regression functions for real values of productivity based on epsilon sensitive loss function and minimum regulated risk. The SVMs were trained with large number data-points each of which consisted of a setting of 21 critical input parameters of blast furnace, corresponding productivity value observed, and the productivity class (low/high) attributed. During the training session of the SVMs, the vectors of critical input parameters were required to be mapped into high-dimensional feature space via Radial basis kernel as function and the optimum SVM-RBF classifying function with chosen setting of its hyperparameters that had good generalization property was found using quadratic optimization. The SVM-RBF classifying function could be used to predict the class of productivity (low/high) for any given setting of the critical input parameters. Class-specific SVM-RBF regression models were also developed for both low as well as high-productivity classes and these models could be used to predict real value of productivity for any given setting of the critical input parameters. The SVM-RBF regression model fitted to the high-productivity class was subjected to constrained nonlinear optimization treatment to find the optimum setting of the critical input parameters that gave maximum productivity. The optimum setting of the critical parameters could be used as the target setting obtaining high productivity in the blast furnace.