AbstractA realistic, efficient and robust technique for the control of amplifier flows has been investigated. Since this type of fluid system is extremely sensitive to upstream environmental noise, an accurate model capturing the influence of these perturbations is needed. A subspace identification algorithm is not only a convenient and effective way of constructing this model, it is also realistic in the sense that it is based on input and output data measurements only and does not require other information from the detailed dynamics of the fluid system. This data-based control design has been tested on an amplifier model derived from the Ginzburg–Landau equation, and no significant loss of efficiency has been observed when using the identified instead of the exact model. Even though system identification leads to a realistic control design, other issues such as state estimation, have to be addressed to achieve full control efficiency. In particular, placing a sensor too far downstream is detrimental, since it does not provide an estimate of incoming perturbations. This has been made clear and quantitative by considering the relative estimation error and, more appropriately, the concept of a visibility length, a measure of how far upstream a sensor is able to accurately estimate the flow state. It has been demonstrated that a strongly convective system is characterized by a correspondingly small visibility length. In fact, in the latter case the optimal sensor placement has been found upstream of the actuators, and only this configuration was found to yield an efficient control performance. This upstream sensor placement suggests the use of a feed-forward approach for fluid systems with strong convection. Furthermore, treating upstream sensors as inputs in the identification procedure results in a very efficient and robust control. When validated on the Ginzburg–Landau model this technique is effective, and it is comparable to the optimal upper bound, given by full-state control, when the amplifier behaviour becomes convection-dominated. These concepts and findings have been extended and verified for flow over a backward-facing step at a Reynolds number $\mathit{Re}= 350$. Environmental noise has been introduced by three independent, localized sources. A very satisfactory control of the Kelvin–Helmholtz instability has been obtained with a one-order-of-magnitude reduction in the averaged perturbation norm. The above observations have been further confirmed by examining a low-order model problem that mimics a convection-dominated flow but allows the explicit computation of control-relevant measures such as observability. This study casts doubts on the usefulness of the asymptotic notion of observability for convection-dominated flows, since such flows are governed by transient effects. Finally, it is shown that the feed-forward approach is equivalent to an optimal linear–quadratic–Gaussian control for spy sensors placed sufficiently far upstream or for sufficiently convective flows. The control design procedure presented in this paper, consisting of data-based subspace identification and feed-forward control, was found to be effective and robust. Its implementation in a real physical experiment may confidently be carried out.
Subspace identification algorithm for stochastic systems equipped with zeros close to unit circle