Performance-oriented Adaptive Switching Control

This thesis addresses the problem of controlling uncertain multivariable systems by means of adaptive switching control (ASC) schemes. Indeed, in many real applications, a large number of actuator and/or sensors may be employed so as to achieve the desired control task, thus requiring to treat the process as a typical multi-input/multi-output system. In particular, the attention is directed to model-based switching schemes and the goal is to develop solutions which aim at improving transient/regime performance. The main feature of the examined architecture is that stability does not depend on model distribution and performance improvements can be achieved without increasing the number of models. Part I aims at extending a model-based control approach, so far restricted to singleinput/ single-output systems, to a general multivariable setting. The proposed scheme relies on a “high-level” unit, called the supervisor, which at any time can switch on in feedback with the process one controller from a finite family of candidate controllers. The supervisor performs routing/scheduling tasks by monitoring suitable data-based test functionals. In addition, a possible modification to the original scheme is introduced, whereby switching among fixed candidate controllers can be suitably combined with an adaptive mechanism, this idea being of interest for on-line implementation of highly performing ASC schemes. Part II addresses the problem of the control transfer in model-based ASC schemes. Indeed, the switching is a source of nonlinearity and can cause variations of closed loop dynamics yielding significant performance degradations. To cope with this event, the proposed technique aims at promptly recovering an adequate closed-loop behavior and it exploits the model distribution/uncertainty structure so as to suitably reset of the state of the switchedon controller, in accordance with the regime behavior predicted by the a-priori information. From an implementation viewpoint, the technique is flexible enough so as to allow the designer to trade off performance vs. memory and/or computational complexity, even when the process is described by a continuous distribution of models. Since simulations of adaptive control systems are often useful for performance evaluation, Part III focuses on a numerical multivariable example.