ELSEVIER Fuzzy Sets and Systems95 (1998) 273-293 FU|ZY sets and systems The application of fuzzy logic in automatic modelling of electromechanical systems P.J. Costa Branco*, J.A. Dente CA UTL-Laborat6rio de Mecatr6niea, Instituto Superior T~cnico, Av. Rovisco Pais, 1096, Lisbon Codex, Portugal Received April 1996; revised August 1996 Abstract Electromechanical systems are usually modelled using energy conversion theory. However, this representation is not accurate enough. The reasons are the presence of non-linear relations between the variables, changes in system parameters, and the difficulty encountered sometimes in taking into account, in a simple and precise way, physical phenomena like, friction, viscosity, and saturation. So, it is useful to automatically extract the relations that represent the system behaviour. We investigate in this paper three fuzzy learning algorithms which represent the development of our study and are used for automatic modelling of electromechanical systems. We begin with a very simple algorithm. Some problems are pointed as containing the requisites to be a good model; next, two methods which are composed of a fuzzy-cluster-based algorithm and a fuzzy-supervised-learning algorithm are employed. We explore their learning capabilities in situations like modelling in a direct and inverse way, the amount of information necessary to build a good model, and the problem of selecting the information relevant to the learning process. The algorithms are analysed in an experimental system in our laboratory. We close with a simple control application for the relationship between fuzzy modelling and electro- mechanical systems designing a feedforward learning controller for the experimental system. © 1998 Elsevier Science B.V. All rights reserved. Keywords: Linguistic modelling; Approximate reasoning; Pattern recognition; Machine learning 1. Introduction The conventional approaches for modelling electromechanical systems use differential or difference equations, partial differential equations, and so on. These model formalisms can present some drawbacks: (a) the mathematical models are not a complete system representation since there are simplifications even for the simpler models, (b) they do not allow the inclusion of heuristic information permitting to expand the systems operating range, and (c) the internal structure of electromechanical systems usually contain non- linear relations difficult to model or, when the model available is accurate enough, the system parameter values are difficult to obtain, In high-performance applications and for control purposes, there is a need for * Corresponding author. 0165-0114/98/$19.00 © 1998 Elsevier Science B.V. All rights reserved Pll S01 65-01 14(96)00265-5