An identification and adaptive control scheme using fuzzy parameterized linear filters

Presents a nonlinear (fuzzy) control structure enhanced with supervised learning and/or adaption. Availability of at least a partial process model is assumed. A nonlinear process identification procedure is used to complete the partial model. Based on the identification model the system sensitivity model is derived which guides the training process to keep the training time (the number of observations) at minimum. The identification model introduced consists of a linear dynamical section (filter) and a nonlinear zero-memory section (implemented by a fuzzy mapping). Only the filter section is on the primary signal path. The nonlinear mapping delivers the filter parameters (depending on the system input and state). An adaption procedure is introduced, which tunes the nonlinear mapping (e.g. membership function parameters) to minimize identification error. The process identification and the controller tuning can run parallel, in this way the online adaption of the controller can be realized in a straightforward way. The scheme proposed can incorporate a priori knowledge on two levels (system topology and fuzzy rule set). The most distinctive features of the scheme are that it directly supports controller design and/or (online) tuning in various ways and results in a dynamical system, which is relatively easy to analyze using well established traditional analysis methods. The identification and control scheme can readily incorporate other types of sufficiently smooth parameterized nonlinearities (e.g. neural networks or domain specific specialized nonlinear mappings).