Automatic controller tuning via unfalsified control

In this paper, we present a new algorithm for the automatic adaptation of linear single-loop controllers. The algorithm is based on the idea of unfalsified control, where a set of controllers is defined and using measured data, the controller from the set is chosen online which provides the best performance. This scheme is extended here by an adaptation of the set of controllers after a change of the operating point of the plant. The original concept of unfalsified control relies on the computation of the performance of controllers that are not in the loop using so-called fictitious signals. However, the proposed cost function leads to erroneous results, in particular, it is not possible to detect closed-loop instability that may result from inserting a controller. Therefore a different cost function and a new method for online evaluation of controllers that are not in the loop without an explicit plant model are proposed. With this new cost function, the set of controllers can be adapted as well which is performed using an evolutionary algorithm. The method is demonstrated for a well-known example, the nonlinear non-minimum phase CSTR model with van der Vusse reaction scheme and PID controllers.

► We present an interesting new method for the control of process without a sophisticated process model.
► The new method is a combination between unfalsified control and an evolutionary algorithm.
► A new fictitious error signal that can detect instability of a non-active controller is presented.
► That signal leads to a new cost function which is a key for any optimization technique.
► As an example for a nonlinear process, a non-minimum phase CSTR is used to show the performance of the approach via an adaptation of PID controller parameters.