There are two reasons for applying on-line identi?cation.
First, a too simplistic
(e.g. linear) model may be used, which is only capable of describing the process behaviour within a small operating regime. The need for on-line adaptation then emerges from the process nonlinearities that are not represented by the model. This strategy is employed in classical linear adaptive control [19]. However, the second reason for the requirement of on-line adaptation is time-variant behaviour of the process. This problem, addressed here, equally exists for both linear and nonlinear models. A TakagiSugeno fuzzy model possesses nonlinear para-meters which determine the rule premises (centres and standard deviations of the validity functions) and linear parameters which determine the rule consequents. For on-line adaptation, the following strategy is pursued. It is assumed that at least a rough model has been identi?ed off-line with the GustafsonKessel fuzzy clustering algorithm. In the on-line phase the rule premises are kept ?xed and only the rule consequents are adapted. The advantage of this approach is the exploitation of mature, computationally effective and numerically robust linear recursive algorithms. A drawback lies in the requirement of an off-line model and in its limited structural ?exibility. Since many time-variant processes such as the thermal processes do not signi?cantly change their nonlinear structure over time but only their gains, time constants or zeros the latter drawback might not be very severe. In general however, if one does not have any prior knowledge of the process, it will not be known to what extent the nonlinear structure of the time-variant process changes. But even if the structural properties of the process chan