Data-driven multivariable ILC: enhanced performance by eliminating L and Q filters

Iterative learning control (ILC) algorithms enable high-performance control design using only approximate models of the system. To deal with severe modeling errors, a robustness filter Q is typically employed. Irrespective of the large performance enhancement, these approaches thus require a modeling effort and are subject to a performance/robustness tradeoff. The aim of this paper is to develop a fully data-driven ILC approach that does not require a modeling effort and mitigates the performance/robustness tradeoff. The main idea is to replace the use of a model by dedicated experiments on the system. Convergence conditions are developed in a finite-time framework, and insight in the convergence aspects is presented using a frequency-domain analysis. Extensions to increase the convergence speed are proposed. The developed framework is validated through experiments on a multivariable industrial flatbed printer. Both increased performance and robustness are demonstrated in a comparison with closely related model-based ILC algorithms.