Discovery of fuzzy models from observation data

Methods for automatic identification of fuzzy models for the purposes of realtime industrial process monitoring are studied and tested. Theoretical fuzzy and neurofuzzy approaches to identification of Mamdani and Takagi-Sugeno inference models are summarized. Support of fuzzy inference in the active database system RapidBase is discussed. Commercial products Matlab and fuzzyTech are tested in a case study of predicting abnormal states in a waste water treatment plant. The most challenging part of the modeling turns out to be the structural identification, that is the derivation of the rule format and the selected variables. Best results are produced, in Matlab, by using neurofuzzy learning (ANFIS) together with the sequential forward search and, in fuzzyTech, with the learning of degree of rule support. Applicability of the studied methods to automatic extraction of RapidBase fuzzy monitoring models from measurement data is discussed.