Sie sind auf Seite 1von 58

Feedback control techniques in fluid dynamics

Fulvio Martinelli

Laboratoire dHydrodynamique (LadHyX)


CNRS - Ecole Polytechnique

Oct 24, 2011


Why control of fluid flows

Several engineering applications:

Transportation efficiency (airplanes, cars, ships)


Mixing enhancement
Noise reduction
Combustion stabilization
Quality of chemical processes
Quality in food industry
...

Also, fun research problems!


Aeronautical examples

Passive control: winglets


Aeronautical examples

Active, feedforward control: film cooling


Aeronautical examples

Active, feedback control: aeroelasticity


Large space O(1 m) and time O(10 Hz) scales.
Aeronautical examples

Active, feedback control: maneuvering


Small space O(10 m) and time O(10 kHz) scales.
Outline

Three canonical problems in simple geometries

Delay of laminar-turbulent transition


Drag reduction in wall turbulence
Control of flow separation

Method

Definition of a linear(ized) model


Control design based on linear theory
Testing in linearized and nonlinear flows
Outline

1 Transition control

2 Turbulent drag reduction

3 Separation control
A model problem

Incompressible flow in a plane channel

Velocity components u, v , w in x, y , z
Translational invariance in x and z
Uh
Reynolds number Re =
Flow dynamics

Navier-Stokes equations

v =0
+ B. C.
v + v v = p + 1 v
t Re
B.C.: No-slip or transpiration for control

Solutions

Laminar flow: steady equilibrium, stable for Re < 5772


Turbulent flow: unsteady, chaotic motion, for Re > 1000
Transition to turbulence - mechanisms

Linearization about laminar solution and space discretization

Navier-Stokes eqs x = Ax

Asymptotic dynamics Short-time amplification

max Re{(A)} G = max ||e At ||2E


t0
Asymptotic vs short-term perturbation dynamics

Linear amplification can bring the flow state outside the attraction
basin of laminar solution (Schmid and Henningson, 2001)

70

60

50 Re=2000, oblique wave ==1


Re=8000, 2D most unstable mode
40
E(t)

30

20

10

0
0 50 100 150
t
Subcritical transition - oblique waves

Isosurfaces of streamwise perturbation velocity


Nonlinear simulation, Re = 2000.
Estimating the maximum amplification

Upper bound via Lyapunov theory

Gu = max (P)max (P1 ), P = PT > 0, PA + AT P < 0


Estimating the maximum amplification

Upper bound via Lyapunov theory

Gu = max (P)max (P1 ), P = PT > 0, PA + AT P < 0

Minimal upper bound - LMI generalized eigenvalue problem

min :
P = PT > 0, PA + AT P < 0
I < P < I > Gu
Numerically solved with established interior point algorithms.
Transition control

Input: distributed blowing/suction - zero net mass flux

x = Ax + Bu
Transition control

Input: distributed blowing/suction - zero net mass flux

x = Ax + Bu

Full-state control

u = Kx x = (A + BK)x
Transition control

Input: distributed blowing/suction - zero net mass flux

x = Ax + Bu

Full-state control

u = Kx x = (A + BK)x

Objective
Lower amplification in closed loop upper-bound minimization.
Design methodology

Closed-loop Lyapunov inequality (more d.o.f. due to K)

P(A + BK) + (A + BK)T P < 0 P = PT > 0,

Final form: again an LMI generalized eigenvalue problem

min :
Q = QT > 0, AQ + QAT + BY + YT BT < 0
I < Q < I
from which K = YQ1 .
Results - oblique waves

Linear amplification
Nonlinear transition thresholds

Thres. I.F.
Open-loop 2.39 106
u - act. 9.89 106 4.14
v - act. 3.04 105 12.72
w - act. 9.77 106 4.09
Outline

1 Transition control

2 Turbulent drag reduction

3 Separation control
Turbulent drag reduction by feedback

The flow problem

Turbulent flow in a plane channel


Statistically stationary, x and z homogeneous directions
Turbulent drag reduction by feedback

The flow problem

Turbulent flow in a plane channel


Statistically stationary, x and z homogeneous directions
Control: wall transpiration, zero-net mass flux
Sensor: wall shear stress and pressure fluctuations
Turbulent flow in closed loop

Effect of the control is small (then additive)

d n

y x u
+ C + H K
Turbulent flow in closed loop

Effect of the control is small (then additive)

d n

y x u
+ C + H K

n: turbulence H: input-state dynamics


Turbulent flow in closed loop

Effect of the control is small (then additive)

d n

y x u
+ C + H K

n: turbulence H: input-state dynamics


u: wall blowing/suction K : compensator
Turbulent flow in closed loop

Effect of the control is small (then additive)

d n

y x u
+ C + H K

n: turbulence H: input-state dynamics


u: wall blowing/suction K : compensator
y : measurements C : sensors
Measuring H

Average impulse response to wall forcing (Quadrio, Luchini and


Zuccher, 2006)

Wall forcing with a small space-time white Gaussian noise on


the wall-normal velocity at the wall vw (x, z, t)
Measuring H

Average impulse response to wall forcing (Quadrio, Luchini and


Zuccher, 2006)

Wall forcing with a small space-time white Gaussian noise on


the wall-normal velocity at the wall vw (x, z, t)
In the linear setting, the perturbed flow reads:

vtot (x, y , z, t) = v(x, y , z, t) + v(x, y , z, t)


| {z } | {z }
turbulence control effect
Measuring H

Cross-correlating the flow and the wall forcing:

E {vtot (x 0 + x,y , z 0 + z, t 0 + t)vw (x 0 , z 0 , t 0 )} = . . .


. . . E {v(x 0 + x, y , z 0 + z, t 0 + t)vw (x 0 , z 0 , t 0 )} + . . .
| {z }
=0, statistical independence
0
. . . + E {v(x + x, y , z 0 + z, t 0 + t)vw (x 0 , z 0 , t 0 )}
The average impulse response

Cross-correlation defines an impulse response function:

Hv (x, y , z, t) = E {v(x 0 + x, y , z 0 + z, t 0 + t)vw (x 0 , z 0 , t 0 )}


The average impulse response

Cross-correlation defines an impulse response function:

Hv (x, y , z, t) = E {v(x 0 + x, y , z 0 + z, t 0 + t)vw (x 0 , z 0 , t 0 )}

The average response to wall forcing is then

Z
v(x, y , z, t) = Hv (x x, y , z z, t t)vw (x, z, t) d xd zd t
The average impulse response

v response to v
The average impulse response

w response to v
Control objective

Minimize drag minimize wall shear stress

Link to average dissipation rate:

dhUi
D = Dmean + Dfluctuations
dy w
Control objective

Minimize drag minimize wall shear stress

Link to average dissipation rate:

dhUi
D = Dmean + Dfluctuations
dy w

Dmean affected indirectly via nonlinear interactions


Control objective

Minimize drag minimize wall shear stress

Link to average dissipation rate:

dhUi
D = Dmean + Dfluctuations
dy w

Dmean affected indirectly via nonlinear interactions


Dfluctuations affected directly by zero net mass flux
blowing/suction
Control objective

Minimize drag minimize wall shear stress

Link to average dissipation rate:

dhUi
D = Dmean + Dfluctuations
dy w

Dfluctuations
D grows with Re!

Re Dfluctuations Dmean
1500 26.8% 73.2%
3000 39.5% 60.5%
Control design technique

Minimizing Dfluctuations while keeping control effort small


Natural setting: linear quadratic Gaussian control

J = E {xT Qx} + E {uT Ru}


| {z } | {z }
Dfluctuations control effort
Control design technique

Minimizing Dfluctuations while keeping control effort small


Natural setting: linear quadratic Gaussian control

J = E {xT Qx} + E {uT Ru}


| {z } | {z }
Dfluctuations control effort

Time-domain formulation not practical go to frequency


Enforcing causality of K in frequency: Wiener-Hopf problem
Drag reduction results

Overall best performance with 7.7% of net power saved.

Re x z p
1500 2% 0% 0%
3000 8% 6% 0%

Uncontrolled Controlled
Outline

1 Transition control

2 Turbulent drag reduction

3 Separation control
Interlude

Up to now we used:

High dimensional discretizations


Full-state control design
This is often not feasible in more complex geometries.
Interlude

Up to now we used:

High dimensional discretizations


Full-state control design
This is often not feasible in more complex geometries.

Realistic flow control problem

Few actuators/sensors on the boundary


High order (degrees of freedom may well be O(108 ))
Disturbance environment often unknown
Interlude

System identification and input/output control design

I/O data only reduced computational burden.


Accounts for the influence of disturbances
Avoiding modeling errors
Enabling technology applicable to experimental data.
Control of vortex shedding

Model problem: cylinder wake

Objectives
Reduce oscillatory loads, increase critical Re
Flow configuration

Identification in presence of disturbances


Formulation

Discrete-time state-space form:


(
x(k + 1) = Ax(k) + Bu(k) + Bd d(k)
y (k) = Cx(k)
Bd d(k) : scalar, periodic
Formulation

Discrete-time state-space form:


(
x(k + 1) = Ax(k) + Bu(k) + Bd d(k)
y (k) = Cx(k)
Bd d(k) : scalar, periodic

Why do we care about periodic disturbances?

Real world disturbances could be deterministic!


These can be made periodic, in practice...
Identification approach

ARX model in presence of d

y (k) = 1 y (k1)+. . .+p y (kp)+1 u(k1)+. . .+p u(kp)

If p is selected large enough:


Exact reconstruction of impulse response from i , i
Reconstruction independent of the disturbance amplitude
Feedback control

Control loop:

1
Sensitivity function S = 1+HK use K to shape it.
Minimize a weighted norm of S.
Feedback control (contd)

Design K to shape the sensitivity


Next steps: extension to nonlinear

Saturated base flow as a periodic disturbance


Not quite right! Linearization about a periodic base flow

x = A(t)x + Bu A(t) = A(t + T ), t


y = Cx

General solution
Z t
y (t) = C (t, )Bu( ) d (t, ) : State-transition matrix
0
Next steps: extension to nonlinear

Warning:
Any linear identification technique would approximate

(t, 0) e At

Example ((t, 0) is known analytically).

1 + cos( 2 sin( 2
 
T t) T t)
A(t) =
sin( 2
T t) 1 + cos( 2
T t)
B = [1, 1]T C = [1, 1]
Next steps: extension to nonlinear

Time-scale separation effect

T =1 T = 100
Conclusions

Transition and turbulence control

Interesting and diverse mathematical approaches


Limited practical applicability

Control of separation

Potential practical applications / experiments


Work in progress....
Perspectives of feedback flow control

Likely applications

Microfluidics (low Re)


Flows with external body forces (EHD, MHD)
Flows featuring a low-dimensional dynamics
Conclusions

THANK YOU!

Das könnte Ihnen auch gefallen