Learning a qualitative model of a complex system via data analysis

International audienceA great number of complex industrial systems produce goods whose quality is very sensitive to the variations of some production parameters. It is often the case that no model exists, which relates the "quality variables" to the "production variables" . However, the existence of such models would be very appreciable for process supervision. Knowledge based models can be constructed using the experience of the process operators, and used for diagnosis purposes. Another approach leads to use data analysis methods in order to construct statistical models, since the variables which define the product's quality and the operation parameters can often be evaluated. This is a learning situation, in which knowledge acquisition (model building) is made through the consideration of a set of examples (the data). In such a context, a precise numerical model has no meaning, and the search of a qualitative model seems to be a reasonable approach. We use results from structural analysis of systems by means of information theory in order to construct a model via rules generation. We propose some algorithms for rules simplification, in both the cases where the explanatory variables take their values in non ordered as well as in ordered sets