An approach to the on-line design of Takagi-Sugeno type fuzzy models is presented in the paper. It combines supervised and unsupervised learning and recursively updates both the model structure and parameters. The rule-base gradually evolves increasing its summarization power. This approach leads to the concept of the evolving Takagi-Sugeno model. Due to the gradual update of the rule structure and parameters, it adapts to the changing data pattern. The requirement for update of the rule-base is based on the spatial proximity and is a quite strong one. As a result, the model evolves to a compact set of fuzzy rules, which adds to the interpretability, a property especially useful in fault detection. Other possible areas of application are adaptive non-linear control, time series forecasting, knowledge extraction, robotics, behavior modeling. The results of application to the on-line modeling the fermentation of Kluyveromyces lactis illustrate the efficiency of the approach. (c) Springer