Data driven variable selection, without including physical knowledge, is an important prerequisite for many applications in the field of data based modeling. This paper deals with a novel approach to optimize the dimension of the input space by a combination of common variable selection methods with multivariate correlation analysis. The results are input structures with revised pseudo correlations between input channels and a physically better interpretable structure. The presented method is successfully applied to measured data from steel industry. Some exemplary results are shown in this paper. Copyright © 2007 International Federation of Automatic Control All Rights Reserved.