Taming Real World Flow Control Experiments with MLC
In this chapter, we present applications of machine learning control (MLC) to flow control experiments. Examples range from mixing enhancement of laminar flow to separation mitigation of a turbulent boundary layer. The discussion highlights the physical a
- PDF / 1,274,303 Bytes
- 32 Pages / 439.37 x 666.142 pts Page_size
- 64 Downloads / 178 Views
Taming Real World Flow Control Experiments with MLC
An approximate answer to the right problem is worth a good deal more than an exact answer to an approximate problem. John Tukey
In Chap. 2, MLC was introduced as a generic method to identify optimal control laws for arbitrary dynamics. In Chaps. 4 and 5, MLC has been applied to the control of low-dimensional dynamical systems. In these examples, we have shown (1) that it is comparable to optimal linear control design for linear dynamics, (2) that it outperforms linear control methods in the case of a weak nonlinearity, and (3) that it can even identify the enabling strongly nonlinear actuation mechanism in the case that the linear dynamics are uncontrollable. In this chapter, we describe and exemplify the application of MLC to real-world turbulence control experiments. These come with the additional challenges of highdimensional dynamics, long time delays, high-frequency noise, low-frequency drifts and, last but not least, with the non-trivial implementation of the algorithm in the experimental hardware. In experiments, MLC is executed in the same way as for the dynamical system plants in Chaps. 4 and 5: 1. MLC provides a generation of control laws to be evaluated by the plant. 2. The plant is used to evaluate and grade these individuals with respect to the given cost function. 3. MLC evolves the population. 4. The process stops when a pre-determined criterion is met. 5. After this learning phase, the best control law can be used. The only difference between a simulation and an experiment is the need to interrogate an experimental plant. This is a technical challenge but not a conceptual point of departure from MLC. Running MLC on an experiment using an existing code is a matter of a few days to a week of work, if the experimental hardware and software is ready for closed-loop control. We choose three configurations: a laminar flow over a backward-facing step in a water tunnel (Sect. 6.1), separating turbulent boundary layers in wind tunnels © Springer International Publishing Switzerland 2017 T. Duriez et al., Machine Learning Control – Taming Nonlinear Dynamics and Turbulence, Fluid Mechanics and Its Applications 116, DOI 10.1007/978-3-319-40624-4_6
121
122
6 Taming Real World Flow Control Experiments with MLC
(Sect. 6.2), and a turbulent mixing layer (Sect. 6.3). These examples encompass key phenomena encountered in most flow control problems: boundary layers, separation, mixing layers and a recirculation zone. Different kind of sensors, actuators and time scales are used and illustrate the versatility of MLC. Section 6.4 highlights the limitations of model-based linear control for the turbulent mixing layer. In Sect. 6.5, we focus on implementation issues with respect to software and hardware. Section 6.6 suggests reading on a spectrum of flow control aspects. Our interview (Sect. 6.7) addresses past and future developments in experimental closed-loop turbulence control with Professor Williams, a pioneer and leading scholar of this field.
6.1 Separatio
Data Loading...
