Sensitivity-based approaches for an efficient design of learning-type controllers of a flexible high-speed rack feeder system

Abstract

In previous work, it has been shown that sensitivity-based procedures can be employed effectively for the design of predictive control strategies, for the implementation of state estimators as well as for the offline and online identification of system parameters. These procedures were used, on the one hand, for control of dynamic processes which perform a certain task only once and, on the other hand, also for the control of systems that are operated in a repetitive manner. The latter class of applications is hence closely related to the design of iterative learning control strategies. A common feature of all sensitivity-based approaches implemented so far by the authors is that the control signals are piecewise constant on an equidistant time discretization mesh. However, this assumption may make the computation of differential sensitivities inefficient if long control horizons are taken into account for learning-type controllers of processes with a fast dynamics. Therefore, this assumption is removed in the current paper, both by a control parameterization using polynomial ansatz functions and by a computation of optimal switching points for piecewise constant control signals. The adaptive discretization scheme of the latter approach allows for obeying predefined performance constraints with a minimum memory demand. These procedures are demonstrated by simulations for a prototypical high-speed rack feeder system.