On-line identification of computationally undemanding evolving fuzzy models

This paper describes an on-line evolving fuzzy Model (efM) approach to modelling non-linear dynamic systems in which an incremental learning method is used to build up the rule-base. The rule-base evolves when “new” information becomes available by creating a new rule or deleting an old rule depended upon the proximity and potential of the rules, and the maximum number of rules to be used in the rule-base. An efM based on a T–S fuzzy model is a very good candidate for modelling complex non-linear systems, when the period of time required to collect a complete set of training data is too long for the model to be identified off-line. The proposed learning scheme is computationally undemanding and is suitable for use in model-based self-learning controllers. Three example applications of the efM are given: the first involves the modelling of a simple non-linear dynamic system, the second example is a cooling coil in a real air-conditioning system; the last example shows how the efM can be used in a Model-based Predictive Control (MbPC) scheme. The results demonstrate the ability of the efM to evolve the rule-base efficiently so as to account for the behaviour of the system in new regions of the operating space. In all given cases, the proposed efM approach generates an accurate model with relatively few rules in a computationally undemanding manner, even if the data are noisy and incomplete.