We propose a residual-based approach for fault detection in rolling mills which is based on data-driven soft computing techniques. The basic idea is to transform original measurement signals into a feature space by (i) identifying multi-dimensional relationships in the system, (ii) representing the nominal fault-free case, and (iii) analyzing residuals with incremental/decremental statistical techniques. Model identification and fault detection are conducted in a completely unsupervised manner, that is, solely based on the data streams recorded online. Thus, neither annotated samples nor fault patterns/models, which are often very time-intensive and costly to obtain, must be available a priori. We use purely linear models, a new genetic variant of Box-Cox models (termed Genetic Box-Cox) that consider weak non-linearities, and Takagi-Sugeno fuzzy models, which are able to express more complex non-linearities, trained with an extended version of SparseFIS. Using three typical scenarios from rolling mill production, we compare our method to state-of-the-art approaches that are based on principal components analysis and multi scale principal components analysis. The results show that our method outperforms these state-of-the-art approaches.