Florian Browne

In this article, we propose an iterative learning control (ILC) algorithm for a class of periodic processes with a variable time-delay that is greater than one iteration of the process. We propose a definition of the time delay as the sum of an integer and a residual such that the two components can be estimated separately. This structure enables the derivation of a stability law for an ILC algorithm that is a function of the delay-estimation error. Additional remarks are given regarding the steady-state performance of the combined time-delay-estimation and ILC algorithm. The proposed algorithm is then applied to a twin-roll strip casting process. We demonstrate the sensitivity of the ILC algorithm to the time-delay-estimation error through simulation results and define an upper bound for the time-delay-estimation error that can be tolerated before the ILC algorithm begins to amplify the disturbance. We then validate the combined time-delay-estimation and ILC algorithm's performance using experimental process data from an industrial twin-roll strip casting process.

During twin-roll steel strip casting, molten steel is poured onto the surface of two casting rolls where it solidifies to form a steel strip. The solidification process introduces a two-phase region of steel known as mushy steel which has a significant effect on the resulting quality of the manufactured steel strip. Therefore, an accurate model of the growth of mushy steel within the steel pool is imperative for ultimately improving strip quality. In this paper, we derive a reduced-order model of the twin-roll casting process that captures the dynamics of the mushy region of the steel pool and describes the effect that the casting roll speed and gap distance have on the solidification dynamics. We propose a switched-mode description that leverages a lumped parameter moving boundary approach, coupled with a thermal resistance network analogy, to model both the steel pool and roll dynamics. The integration of these models and simulation of the combined model are nontrivial and discussed in detail. The proposed reduced-order model accurately describes the dominant dynamics of the process while using approximately one-tenth of the number of states used in previously published models.