Controle de escoamentos não estacionários em malha fechada: abordagens não lineares via busca extremal e redes neurais

Author(s)

Tarcísio Costa Déda Oliveira

• Document type: Doctoral Thesis
• Press: Campinas, SP.
• Institution: Universidade Estadual de Campinas (UNICAMP). Faculdade de Engenharia Mecânica
• Defense date: 2023-01-06

Examining board members

William Roberto Wolf; Scott Thomas McGregor Dawson; Kunihiko Taira; Yeh Chi An; Fernando Zigunov

• Advisor: William Roberto Wolf

Abstract

The study of unsteady flows finds application in several industrial processes in mechanical and aerospace engineering. For example, the flows that develop over aircraft wings and wind turbine blades are turbulent and, therefore, unsteady. In certain configurations, transition to turbulence may also lead to important dynamical features and flow unsteadiness. For most flows, transition and turbulence are directly related to drag increase and aeroacoustic noise generation, which may be undesired flow features in some engineering devices and processes. On the other hand, heat transfer and mixing enhancement are also achieved by turbulent flows and can be desired in many applications. For some of these cases, it is important to develop flow control strategies to reduce drag and noise, besides increasing heat and mass transfer. In this thesis, we investigate techniques of active flow control well suited for applications in unsteady flows involving unsteady flows. These techniques are employed in high-fidelity numerical simulations to validate the strategies proposed. We present the application of extremum seeking control (ESC) -- a model-free closed-loop optimization control strategy -- for reducing aeroacoustic noise generated in airfoil flows. It is shown that, with an adequate cost function measured in real time and the appropriate actuation scheme, considerable sound attenuation can be obtained and, in some cases, total suppression is achieved. Furthermore, another approach is presented by leveraging machine learning with supervised training of neural networks for flow control. Through backpropagation of neural network surrogate models (NNSMs) of fluid flows, two automatic control design approaches are proposed. The first one consists of training a neural network controller (NNC) in a closed-loop approach within a finite horizon. The second one is the automatic linearization of the NNSMs to enable traditional linear control techniques. The approaches are tested with a range of dynamical systems, including a set of low order systems, the perturbed Kuramoto-Sivashinsky (KS) equation and a compressible cylinder flow with the goal of stabilization. Finally, an improvement to the training of NNSMs and NNCs is proposed through an iterative process. The approach provides improved approximations of natural equilibrium points, which are estimated automatically. As a byproduct, the data produced can be leveraged for conducting stability analyses besides optimal sensor selection -- from an initial set of candidate ones -- during the NNSM training. The technique is tested with the Lorenz system in a chaotic configuration, a modified version of the KS equation and with a confined cylinder flow and a plane channel flow. The results demonstrate the effectiveness of computationally inexpensive neural networks in modeling, controlling, and enabling stability analysis of nonlinear systems, providing insights into the system behaviour and offering potential for stabilization of complex fluid systems (AU)

• FAPESP's process: 19/19179-7 - Flow control strategies for unsteady flows involving transition and turbulence
• Grantee: Tarcísio Costa Déda Oliveira
• Support Opportunities: Scholarships in Brazil - Doctorate