An automation system of an industrial process records hundreds of signals. Simultaneous study of all these signals is not possible: the analyst must focus onto the essential-find the variables of interest. The method presented in this paper supports the search. The user defines what is interesting by labeling the data points, for example, using 'A' to mark the data points from a period of a given process control problem, and 'N' to mark a period of normal operation. The method of this paper uses a cluster reparability measure and an optimization algorithm to find the set of variables that best explain the differences between these two states. The method is illustrated with the case of the concentrator of the Outokumpu Hitura mine, where the changes in the mineralogy of the concentrator feed require major changes in the process control strategy. In Hitura the problem is that the process operator has no direct measurement of the mineralogy of the feed. The case of this paper is about finding the on-line variables that contain information of the feed type. (C) 2002 Elsevier Science Ltd. All rights reserved.