Reduced-order Model-based Feedback Control of

Subsonic Cavity Flows

M. Samimy1, M. Debiasi1, E. Caraballo1, A. Serrani2, X. Yuan2, J. Little1, and J. H.
Myatt3
Collaborative Center for Control Science
The Ohio State University, Columbus, Ohio 43235 USA
1

Gas Dynamics and Turbulence Laboratory; Department of Mechanical

Engineering
2
Department of Electrical and Computer Engineering
2
Air Force Research Laboratory Air Vehicle Directorate, Wright-Patterson AFB

Summary
The latest results of our ongoing research activities in the development of reducedorder models based feedback control of subsonic cavity flows are presented and
discussed in this paper. Particle image velocimetry data and the proper orthogonal
decomposition technique are used to extract the most energetic flow features or
POD eigenmodes. The Galerkin projection of the Navier-Stokes equations onto
these modes is used to derive a set of ordinary differential equations, which govern
the time evolution of the modes, for the controller design. Stochastic estimation is
used to correlate surface pressure data with flow field data and dynamic surface
pressure measurements are used for real-time update of the flow model. Three sets
of PIV snapshots of a Mach 0.3 cavity flow were used to derive three reducedorder models for controller design: (1) snapshots from the baseline (no control)
flow, (2) snapshots from an open-loop forced flow, and (3) combined snapshots
from the cases 1 and 2. Linear-quadratic optimal controllers based on all three
models were designed and tested experimentally. Real-time implementation shows
a remarkable attenuation of the resonant tone and a redistribution of the energy into
various modes with much lower energy levels.

Introduction

Flow control can be divided into two general categories: passive and active. In the
former, which is much easier to implement and has wide-spread applications,
control is accomplished by geometrical modifications to the flow system. In the
latter, mass, momentum, and/or energy are added to the flow. Active control is
divided into open-loop and closed-loop. In open-loop control, actuation takes place
based on an operators command or a predetermined input. In the closed-loop case,
information from one or more sensors in the flow along with a flow model guides
the actuation process.

Successful application of feedback control is widespread in areas such as robotics,

aerospace, telecommunication, transportation systems, manufacturing systems, and
chemical processes. Only in recent years has feedback control of aerodynamic
flows received focused attention, [1]-[11]. Open-loop flow control, which can be
quite useful in many applications, lacks the responsiveness and flexibility needed
for application in dynamic flight environments. In contrast, closed-loop flow
control is well-suited to the successful management of these flows since it allows
adaptability to variable conditions. In addition, closed-loop control shows the
potential to significantly reduce power requirements in comparison to open-loop
control strategies [1]. Unfortunately, the tools of classical control systems theory
are not directly applicable to aerodynamics flows since such systems display
spatial continuity and nonlinear behavior while also posing formidable modeling
challenges due to their infinite dimensionality, a complexity introduced by the
Navier-Stokes equations. In order to design and successfully implement a closedloop control strategy, it is necessary to obtain a reduced-order dynamical model of
the system, which can capture the important dynamic characteristics of the flow
and actuation while remaining sufficiently simple to allow its use in model-based
feedback control design.
The flow over a shallow cavity - a configuration present in many practical
applications that has been extensively studied in the literature was selected for
the present study. This flow is characterized by a strong coupling between the flow
dynamics and the flow-generated acoustic field that can produce self-sustained
resonance known to cause, among other problems, store damage and airframe
structural fatigue in weapons bays. A comprehensive review of this phenomenon
and of various control and actuation strategies developed for its suppression is
given in [6].
Rossiter [12] first developed an empirical formula, which was later modified and
improved by Heller and Bliss [13], for predicting the frequencies of cavity-flow
resonance, today referred to as Rossiter frequencies or modes. Rossiter also
investigated the concept of a dominant mode of oscillation that was later observed
by others to coincide with the natural longitudinal cavity acoustic mode [14]. In
such a condition a strong single-mode resonance occurs [3]; otherwise multiple
modes exist in the flow. A similar interaction could also occur between Rossiter
modes and the natural transversal cavity acoustic modes, [15] and [16]. Recent
theoretical models of the cavity acoustic resonance based on edge scattering
processes, [17] and [18], explain these behaviors. Rapid switching between modes
has been observed in multi-mode conditions [19], [3] and [20]. The random
switching between multiple modes on a rapid time scale places large bandwidth
and fast time response requirements on the actuation scheme and feedback control
algorithm.
Extensive work has been carried out to control the flow over a cavity. Different
open-loop control strategies have been used in recent years with varying degrees of
success, [16], [21]-[23]. There have also been significant efforts to investigate
closed-loop control approaches, [1], [3], [4], [24]-[30]. The results of these closed-

loop endeavors are encouraging, but also indicate that many issues remain to be
solved and numerous opportunities for further advancement of the technology
exist.
While we have examined other control approaches in recent years ([20], [31] and
[32]), our primary objective from the onset has been the development of control
techniques based on a reduced-order model of the cavity flow [33], [34], [29] and
[30]. The approach we have followed in the development of this model is based on
the proper orthogonal decomposition (POD). This technique relies on the energycontaining eddies in the flow, which can be extracted using the spatial correlation
tensor of the velocity field, in the form of spatial eigenmodes called POD modes.
These structures are the most dominant features in the flow and are the only
entities that can effectively be controlled. The dynamics of the flow are obtained
when these modes are modulated by time coefficients obtained by projecting the
Navier-Stoke equations governing the flow onto the POD basis. This results in a
set of non-linear, ordinary differential equations, which we use for controller
design. The equations are autonomous and not useful for controller design
purposes since the controller input is not explicit. Consequently, they must be
recast in a form expressing the control input explicitly so that a feedback controller
can be designed using the tools of control theory, [35] and [29].
In the next section we will introduce the flow facility used in this study. Section 4
will present the POD and Galerkin methods adopted for deriving the reduced-order
model, and the stochastic estimation approach used for real-time estimation of the
flow model variables directly from dynamic surface pressure measurements. This
is followed in Section 5 by a discussion of the flow characteristics and of the
reduced-order model results, and in Section 6 by the design and implementation of
the linear-quadratic controller. We will present and discuss the experimental results
in Section 7, followed by concluding remarks in Section 8.

Experimental Facility and Techniques

Details of the experimental facility and the experimental techniques used can be
found in [16] and [36]. The facility, located at the Gas Dynamics and Turbulence
Laboratory (GDTL) of The Ohio State University (OSU), is a small scale blowdown wind tunnel with a variable-depth cavity recessed in the floor and spanning
the width of the test section. The air is conditioned in a settling chamber designed
to minimize free stream turbulence and directed to the 50.8 mm (2 in) by 50.8 mm
(2 in) test section, Figure 1, through a smoothly contoured converging nozzle
before exhausting into a large exhaust pipe taking it to the atmosphere. The facility
can be operated in the Mach number range 0.20 to 0.70. The focus of this work is
on a shallow cavity with length 50.8 mm and depth of 12.7 mm (0.5 in)
corresponding to a cavity aspect ratio, L/D, of 4. For Mach 0.30 flow, the Reynolds
number based on this cavity depth is approximately 105. Optical quality windows
surround the test section and allow laser based flow diagnostics from 15 mm
upstream to 25 mm downstream of the cavity.

The output of a Selenium D3300Ti compression driver is channeled to the cavity

leading edge where it exits at an angle of 30o with respect to the main flow through
a 2-D slot of 1 mm height that spans the cavity width. This arrangement provides
zero net mass, non-zero net momentum flow for actuation, similar to that of a
synthetic jet. Actuation can be achieved in the frequency range of 1-20 kHz. The
actuator to main flow momentum ratio is in the range of 10-4 to 10-6.
The planar snapshots of the velocity field, required for the development of the
reduced-order model, are acquired using a two-component LaVision particle image
velocimetry (PIV) system. The flow is seeded with oil particles by using a 4-jet
atomizer upstream of the stagnation chamber. This location allows homogenous
dispersion of the particles throughout the test section. The PIV provides a velocity
vector grid of 128 by 128 over the approximate measurement domain of 50.8 mm
(2 in) by 50.8 mm (2 in), which translates to velocity vectors separated by
approximately 0.4 mm.

Flow

Actuator output
L

Kulite pressure
transducer
Compression
driver

Figure 1 Scaled drawing of the

experimental set up

Figure 2 Location and numbering of Kulite

pressure transducers in the cavity flow

Flush-mounted Kulite transducers are placed at various locations on the walls of

the test section for dynamic pressure measurements. Figure 2 shows the locations
of the six transducers used in this study. The transducers have a flat frequency
response up to about 50 kHz, are powered by a dedicated signal conditioner, and
their signals are filtered between 100 Hz and 10 kHz to remove spurious frequency
components. Recordings of 32 blocks of 8192 points (262,144 samples) are
acquired at 200 kHz and converted to non-dimensional pressure referenced to the
commonly used value of 20 Pa. Short-time Fourier transform (STFT) is utilized
to provide information on the time evolution of the frequency content of the
unsteady pressure signal and spectra are obtained by averaging the corresponding
spectrograms, [36].
For state estimation, dynamic pressure measurements are recorded simultaneously
with the PIV measurements. In the current study, 1000 PIV snapshots are recorded
for each flow/actuation condition explored. For each PIV snapshot, 128 samples
from the laser Q-switch signal and from each of the pressure transducers of Figure
2 are acquired at 50 kHz. The simultaneous sampling of the laser Q-switch signal

and the pressure signals allows, for each snapshot, the identification of the section
of pressure time traces corresponding to the instantaneous velocity field.
For closed-loop control of the flow, a dSPACE 1103 DSP board connected to the
Dell Precision Workstation 650 is used. This system utilizes four independent, 16bit A/D converters each with 4 multiplexed input channels and allows simultaneous
acquisition and control processing of 4 signals and almost simultaneous, due to
multiplexing, acquisition and processing of additional signals at a rate up to 50 kHz
per channel to produce at the same rate a control signal from a 14-bit output
channel.

Reduced-order Modeling Procedure

Development of tools and procedures for feedback control based on reduced-order

models has been our primary goal from the onset of this research program [33],
[37] and [34]. Our recent work has focused on deriving the reduced-order models
of the cavity flow from PIV and surface pressure measurements [29] and [30]. The
overall technique combines three separate tools and procedures to obtain and
implement a controller. First, the POD method is used to obtain spatial eigenmodes
or POD modes of the flow. Second, the Navier-Stokes equations tailored for the
flow are projected onto the POD modes using the Galerkin projection method to
obtain the flow model, which consists of a set of ordinary non-linear differential
equations. These equations govern time evolution of the POD modes. In the third
and final step, stochastic estimation is used to correlate the flow velocity field to
surface pressure data and to real-time update the state of the model derived in step
2. Each one of these three steps will briefly be discussed in this section.
4.1

Proper Orthogonal Decomposition Technique

The POD method was introduced to the fluid dynamics community by Lumley
[38] as an objective tool to extract energy-containing large scale structures in a
turbulent flow. Implementation of POD technique requires detailed flow data,
which nowadays could readily be obtained using either numerical simulations or
laser based planar flow measurements. The original derivation, however, favored
time-resolved data over a long time period at a few spatial locations (hot-wire type
data, e.g. [39]). More details of the method can be found in [40] and [41]. Two
decades later, Sirovich [42] extended the POD approach and developed the
snapshot method, which favors spatially-resolved, but time-uncorrelated snapshots
of the flow field. Such data can be easily obtained using advanced laser-based
planar flow diagnostics (such as particle image velocimetry, PIV, or planar
Doppler velocimetry, PDV) or numerical simulations. We are currently using PIV
data with the snapshot method. Details of the snapshot method and its application
in the present context are given in [33] and [37].
The POD method uses M snapshots of the flow and casts the fluctuations in the
flow in terms of N spatial orthonormal modes (N<M) or POD modes, i ( x ) , and

time coefficients for these modes, ai(t). Equation (4.1) is for streamwise velocity
fluctuations that contain major portion of the kinetic energy in the flow.
u ' (x , t )

(t ) i ( x ) .

(4.1)

i=1

For each one of the several flow conditions (i.e. flow Mach number) explored in
this work, 1000 PIV snapshots of the flow field were acquired [36]. Each snapshot
contains two-component of instantaneous velocity on a grid of 128 by 128 over an
approximate measurement domain of 50.8 mm (2 in) by 50.8 mm (2 in) on the x-y
plane passing through center of the cavity. The results indicate that the mean
turbulence kinetic energy is converged by using approximately 700 snapshots [30].
Consequently all 1000 snapshots were used to obtain the POD modes.
4.2

Galerkin Projection

The second step in the process of deriving a reduced-order model is the

projection of the Navier-Stokes equations governing the cavity flow onto the POD
modes, i ( x ) , using the Galerkin projection method. The result of this procedure
is a set of non-linear ordinary differential equations for the time coefficients in
equation (4.1), ai(t)=[a1(t) a2 (t) aN (t)]. The compressible Navier-Stokes
equations used are those derived in [43] and details of the procedure are given in
[33] and [37].
In the Galerkin projection method, each flow variable is decomposed, using
Reynolds decomposition, into its mean and fluctuating components. Then, the POD
expansion in equation (4.1) is written for each of the fluctuating components.
Finally, the flow variables in the Navier-Stokes equations are replaced by the
expanded expressions of mean and fluctuating components. The new form of the
governing equations is then projected onto the POD modes by taking the inner
product of each term with the POD modes according to the vector norm defined in
[43]. This procedure yields a set of ordinary differential equations for the time
coefficients in equation (4.1). These equations are autonomous (i.e. the control
input is implicit in the equations and is not clearly identified). These equations are
not useful for controller design. In order to derive a model where the control input
appears explicitly in the equations, a few methods are currently being explored.
The one used in the present work is based on spatial sub-domain separation, [35]
and [32], which yields the following set of ordinary differential equations:
a T (t )H 1 a (t )

a& (t ) = F + Ga (t ) +
M
+ B (t ) +
T
N
a (t )H a (t )

B 1 (t ) T a (t )

B N (t ) T a (t )

(4.2)

where the matrices of constant coefficients F, G, Hi, B and B i , i=1,N, are

obtained from the Galerkin projection, and (t) is the control input applied at the
forcing location, [34]. Equation (4.2) represents a model of the cavity flow in terms

of the time coefficients ai(t) obtained with the POD method from M time
uncorrelated PIV snapshots.
By using a finite number (N) of modes to describe the flow, one not only filters out
smaller flow structures, but also fails to account for the energy transfer process
between the N retained modes and the neglected ones. Therefore, an additional
viscous term, the modal eddy viscosity [44] was added to the model to maintain the
overall energy balance and to compensate for the truncated modes. This additional
term is added to the viscous term in the Navier-Stokes equations, and is obtained
by a modal energy balance [44].
4.3

Stochastic Estimation

Design of a controller based on the reduced-order model of equation (4.2)

will be presented and discussed in Section 5. In implementing the controller in the
experiment, the variables involved in the reduced-order model must be linked to
the variables that can be measured experimentally in real-time. A similar situation
will also arise in any practical application. The real-time experimental data could
only be obtained via surface measurements (e.g. surface pressure or surface shear
stress measurements) in any realistic setting. We used surface pressure
measurements and stochastic estimation (SE) to correlate these measurements with
the flow velocity data obtained via PIV. Stochastic estimation was originally
proposed and used by Adrian [45] as a means to extract coherent structures from a
turbulent flow field. The technique estimates flow variables at any location by
using statistical information about the flow at a limited number (L) of locations.
In the current work, quadratic stochastic estimation was employed to estimate the
time coefficients of the flow model, equation (4.2), directly from real-time
measurements of surface pressure fluctuations at a small number (L) of locations.
The estimates of the time coefficients can be written in the following form:
a i ( t ) = C ik p k ( t ) + D ikl p k ( t ) p l ( t )

i=1N, k, l=1L

(4.3)

where C, D are the matrices of the estimation coefficients obtained by

minimizing the average mean square error ei between the values of ai (tr) obtained
with equation (4.2) from the snapshot and the estimated ones a i (t r ) at the same
time.
To calculate the matrices in equation (4.3), surface pressure measurements at L
locations taken simultaneously with the PIV snapshots were used. A total of 1000
simultaneous PIV-surface pressure measurements were acquired. The procedure to
obtain the estimation matrices is described in more detail in [37]. In our
experimental setup, real-time measurements of the surface pressure were obtained
at L = 6 locations in the cavity test section shown in Figure 2. While linear
stochastic estimation has often been used in the literature, in [37] was observed that
retaining both the linear and the quadratic terms in equation (4.3) significantly
improved the results. Both terms retained in the current work as well.

5
5.1

Flow Characteristics and Reduced-order Model Results

Flow Characteristics

The Mach number of the flow varied between Mach 0.20 and 0.70 using Mach
number increments of 0.01 to explore the cavity flow characteristics. For each of
these flows, surface dynamic pressure and the sound pressure level (SPL) spectra
were obtained at the center of the cavity floor, as detailed in Section 3 by using a
flush-mounted Kulite pressure transducer. Figure 3 summarizes the results by
showing the SPL intensity level and frequency as a function of the Mach number.
In the same figure are also shown the lines corresponding to the first four Rossiter
modes (R1-R4) predicted by the modified Rossiter formula [13], the 1st
longitudinal acoustic mode based on the cavity length (L1), and the 1st and 2nd
transversal acoustic modes (T1, T2). Strong resonant tones are observed near the
intersections of the predicted Rossiter modes with both the transversal and the
longitudinal acoustic modes. The observation of the interaction between Rossiter
and transversal acoustic modes is similar to that of [15] who explored low subsonic
cavity flow. The work of [12], [14] and [3] all examined cavity flows for which the
tunnel vertical dimension was significantly larger than the cavity length. Therefore,
they observed the interaction of Rossiter modes with the longitudinal cavity
acoustic mode, but not with the transverse acoustic modes, as the transverse
acoustics mode had much lower frequency in their cases.

R4
R3
T2
R2
L1

SPL
(dB)

T1

R1

Figure 3 Shallow cavity floor pressure fluctuations frequency and

amplitude for various Mach numbers

Based on these observations and similar results in the past [16], we have used the
Mach 0.30 flow as our reference baseline case because it shows a single tone at
about 2900 Hz, which is near with the 3rd Rossiter mode and the 1st transversal
acoustic mode. At this Mach number the actuator has enough authority and realtime feedback control is practical. In the current work we have explored this
baseline flow along with several controlled cases.
Figure 4 presents the SPL spectrum of the surface pressure measured at the test
section side wall at the center of the cavity (sensor # 5 in Figure 2) for the baseline
Mach 0.3 cavity flow resonating at the third Rossiter mode, Figure 3. At other
Mach numbers where the cavity oscillates in more than one mode, there is rapid

switching between the modes and the energy gets distributed among the modes
[36].

Figure 4 Spectrum of baseline Mach 0.30 cavity flow calculated from

surface pressure time traces.

Figure 5 presents the instantaneous planar smoke flow visualization images of the
baseline Mach 0.3 cavity flow and two open-loop forced cases. Three coherent
large scale structures are clearly visible in the baseline case, consistent with the
spectrum in Figure 4 and for a flow resonating at the 3rd Rossiter mode. Figure 5(b)
is for the same flow excited at 1830 Hz (i.e. at the second Rossiter mode). As a
result, two large coherent structures are clearly visible in the shear layer. This
confirms that forcing at this frequency weakens or eliminates the natural feedback
mechanism for the third Rossiter mode, but excites the flow at this lower mode.
Figures 5(c) presents the Mach 0.30 flow forced at 3920 Hz. This frequency is
close to the fourth Rossiter mode and as a result four large-scale structures are
visible in the instantaneous image. The phase-locked images of these three flows
also clearly show the existence of these coherent structures [36]. These results
confirm the control authority of the actuators and also the capabilities of the
experimental set up to control the flow to resonate at various Rossiter modes and to
significantly affect the flow field. Interestingly, detailed PIV results show no
significant changes in the ensemble-averaged flow characteristics for these three
quite different flows [36]. Three reduced-order models using PIV snapshots of the
baseline flow, the same forced flow at 3920 Hz, and from both cases are derived
and used for controller design.

a) Baseline
b) forced at 1830 H
c) forced at 3920 Hz
Figure 5 Instantaneous planar images of the baseline Mach 0.30 cavity flow and two
controlled flows. Flow is from left to right.

10
5.2

Reduced-order Modeling Results

Figure 6 shows the first two POD modes for the v-component of velocity
fluctuations of the baseline Mach 0.30 flow. There are three structures in the flow
consistent with the flow visualizations results shown in Figure 5 (a) and with the
surface pressure spectrum of Figures 3 and 4, which shows that the cavity
resonates at the third Rossiter mode. The energy contained in these four modes is
about 40%. These four modes were used in the design of the controller. Phaseaverage v-component of velocity fluctuations from PIV data show a structure
pattern very much similar to the 2nd or 3rd POD modes shown in Figure 6 [36].

Figure 6 First two POD modes for the v-component of velocity fluctuations
of the baseline flow.

In an earlier work, we used a logic based open-loop control and showed that at
certain forcing frequencies the cavity fluctuations are significantly reduced [16].
One such frequency is 3920 Hz (StD~0.5), which is close to the 4th Rossiter mode.
Forcing at this frequency changes the resonance to multi-mode regime and reduces
the peak at the third Rossiter modes by about 20 dB, but adds a smaller peak (10 to
20 dB lower depending on the sensor location) at the forcing frequency. It is
interesting to note, that one of the multiple modes is a sub-harmonic of the forcing
frequency, which is close to the second Rossiter mode. This mode seems to
dominate over the other modes. The first two POD modes for this case are shown
in Figure 7. It is obvious that the forcing has disrupted the natural resonance at the
3rd Rossiter mode, but established a resonance at the 2nd Rossiter mode with two
clearly defined structures in the flow. The energy contained in these four modes is
about 37%, lower than that of the baseline case.

Figure 7 First two POD modes for the v-component of velocity fluctuations
of the flow forced at 3920 Hz.

The POD modes from the combined PIV snapshots of the two cases discussed
above with a total of 2000 snapshots (not shown) take some characteristics from
each. For example, the first and second modes with two dominant structures and

11

the third and fourth modes with three structures resemble those of the baseline and
the forced cases, respectively.
The set of non-linear ordinary differential equations in equation (4.2), obtained by
Galerkin projection of the Navier-Stokes equations onto the POD modes, was
solved to check the evolution and convergence of the time coefficients. The time
coefficients for the baseline cavity flow converged using a number of modes N
between 4 and 10. After an initial transient period, the coefficient oscillates close to
zero, as expected.
The time coefficients for the baseline case from the PIV snapshots were obtained
using equation (5.1):
a i (t ) =

u ' ( x , t ) i (x ) d x .
D

(5.1)

The qualitative comparison between the results from the solution of equation (4.2)
and values obtained with equation (5.1) from the PIV measurements is quite good.
Using FFT analysis the frequency of oscillation of the time coefficient for the first
mode was found to be about 2417 Hz, a somewhat lower value than the
experimental one (about 2840 Hz). It was observed that the system trajectories
converged to the same behavior, irrespective of the initial condition of the time
coefficient used for the solution of equation (4.2), showing the existence of a stable
limit cycle.
Similar results were obtained for the time coefficients of the other three modes of
the baseline Mach 0.3 flow, and the forced case, and the combined forced and
baseline cases. Interestingly, the spectra of time coefficients for POD modes 1 and
2 show a frequency of 1632 Hz for the forced case at 3920 Hz, which is somewhat
lower than the lowest peak in the experimental surface pressure spectrum, which
will be shown later. Also, in evaluating the stochastic estimation technique results,
discussed in Section 4.3, the time coefficients obtained via quadratic stochastic
estimation, equation (4.3), were compared with those obtained via PIV snapshots
and equation (5.1). Again, the qualitative comparison was good, and the results
from the solution of equation (4.3) fell within the range obtained experimentally
from PIV results and equation (5.1) [29] and [30].

Controller Design and Implementation

In this section, we present and discuss the design of the model-based controller.
The control design approach has been presented in details by the authors in
previous works [34] and [29], and thus it will be only outlined here. The design
procedure includes equilibrium computation, coordinates transformation, linear
approximation of the Galerkin system, and linear-quadratic state feedback control
design. Three reduced-order flow models obtained from POD methods have been
investigated in this work: (1) the baseline Mach 0.3 flow; (2) the same flow forced

12

at 3920 Hz using open-loop; (3) a flow model obtained from POD modes derived
combining PIV snapshots from cases 1 and 2. The reduced-order flow model for all
three cases for control design is the same nonlinear state space model given by
equation (4.2), with N = 4, whereas the numerical values of the model parameters
obviously varies for each case.
6.1

Equilibrium analysis and model simplification

In performing equilibrium analysis and coordinate transformation on the model,

equation (4.2), the constant term F is removed from the model, shifting the origin
of the coordinates to the equilibrium point corresponding to the mean flow. The
resulting simplified state space model in the new set of coordinates
a~ = a a 0 becomes
a~
~ ~

~
&
a = G a +
a~

H
M
H

B
a~

+ B +
4 ~
B
a

M
4

a~

a~

(6.1)

where a 0 is the equilibrium point computed for the model, equation (4.2), and

a 0T H
~

G = G +
a 0T H

+ (H
M
+ (H

)T

) , ~

)T

(B

B = B +
(B

)T a
M

)T a

(6.2)

Clearly, the modified model has an equilibrium point at the origin, which is more
convenient for controller design and stability analysis. The reader is referred to
[29] for a detailed description of the model simplification techniques.
6.2

Linear quadratic state feedback control

A linear approximation of equation (6.1) at the origin is readily obtained as
~
~
a~& = G a~ + B

(6.3)

The eigenvalues of the system matrix of the unforced system, equation (6.3) have
been computed for the three cases respectively as

( )

~
G1

1596.6 + 7023.1i
1596.6 - 7023.1i
=

- 3652

- 879.9

1567.2 + 6880.9i

- 4030

- 524.6

1567.2 - 6880.9i
~
G2 =

( )

1397 + 7061.7i ,

1397 - 7061.7i
~

G3 =

( )

- 2870.9
- 697.2

(6.4)

where the subscript i =1, 2, 3 corresponds to the model based on the baseline flow,
the model based on the same flow with open-loop sinusoidal forcing at 3920 Hz,
and the model based on the combination of these, respectively. All the fourth-order
Galerkin systems for the three cases exhibit the same qualitative features (2
unstable complex conjugate eigenvalues plus 2 stable real eigenvalues) and
quantitative similarities as well. The presence of two unstable complex conjugate
eigenvalues implies, as expected, that the mean flow (corresponding to the
equilibrium a 0 ) is an unstable solution for the Galerkin system, equation (4.2).

13
Since the pairs ( G~ , B~ ) for all cases are controllable, linear state-feedback design
based on the linearized model, equation (6.3), offers a simple approach to the
design of a controller for the nonlinear model, equation (6.1). Recall that the
stochastic estimation method provides a way to estimate the time coefficients of
the Galerkin system from real-time surface pressure measurements, equation (4.3).
The availability of real-time estimates of the state of the Galerkin model, equation
(6.1), allows the use of linear state-feedback control to globally stabilize the origin
of equation (6.3). This, in turn, yields a controller that locally stabilizes the origin
of the nonlinear system, equation (6.1). A convenient and well-established
methodology for the state-feedback controller design is offered by linear-quadratic
(LQ) optimal control. The LQ design computes the gain matrix K such that the
state-feedback law
(6.5)
( t ) = K a~ ( t )
minimizes the quadratic cost function
J

( a~ , ) =

~
(a

a~

a~ + W

) dt ,

(6.6)

where Wa~ > 0 and W > 0 are positive definite weighting functions for the state
vector and the control signal, respectively. Minimization of J c results in
asymptotic stabilization of the origin, while the control energy is kept small. In our
design, the weights have been chosen as W a~ = I 4 4 and W = 1 for all three
models, and the corresponding control gains with respect to the three flows read as
K 1 = [ 56 . 2
8 .8
417 . 2
12 . 8 ] ,
(6.7)
K 2 = [52.5
3.9
- 168
- 102 ] ,
.
K 3 = [17.6
208 . 8
11 . 6
146 . 8 ]
Applying the state feedback control, equation (6.5), to the linearized system,
equation (6.3), results in mirroring all the right-half plane eigenvalues of the matrix
~
G to the left half plane. Figure 8 shows the simulation results obtained by
applying the state feedback control, equation (6.5), to the finite-dimensional
nonlinear model, equation (4.2), which indicate that the closed-loop state
trajectories a(t ) converge to the corresponding equilibrium points in each case,
given by
a 01 = [ 0.5036 0.2788 0.1930 0.4980 ]T ,
a 0 2 = [ 0 . 3081

a 0 3 = [ 0 .3261

0 . 1483

0 . 2083

0 . 4895 ] T

0 .2158

0 .2598

0 .4753 ]

(6.8)

It can be concluded that, in principle, the LQ controller, equation (6.5), designed

for the linear approximation, equation (6.3), succeeds in stabilizing the equilibrium
of the four modes nonlinear Galerkin system, equation (4.2).

14

(a)

(b)

(c)

Figure 8 Time coefficient solutions of the closed-loop simulation results. (a) baseline
flow model, (b) open-loop forced flow model with forcing frequency of 3920 Hz, (c)
combined flow model of the above two cases.

Feedback Control Results and Discussion

Before presenting the results of the experimental implementation of the controller,

it is worth summarizing the structure of the model-based controller derived in
Section 5. As depicted in Figure 9, the model-based controller includes a stochastic
estimation subsystem and a feedback from the estimated states. The estimate a~ of
the deviation from the equilibrium of the time coefficients of the Galerkin model,
required to implement the feedback law, equation (6.5), may be in principle
obtained by means of stochastic estimation by first estimating a (t ) from raw
pressure measurements using equation (4.3), and then subtracting the equilibrium
p

Stochastic
Estimation

Plant

-K

a~
Controller

Figure 9 Diagram of the closed loop system with LQ state feedback control.

value a 0 computed from the model data. However, in implementing the controller,
subtracting the equilibrium values from the estimated ones is not required, since
the DC values have been removed from the pressure measurements by means of
high-pass filtering. That is, equation (4.3) naturally produces the values of a~ from
the experimental measurements. It is important to point out that, to prevent any
damage to the actuator, the control input signal is limited to the range 10V. Since
the gains of the LQ control, equation (6.7), are quite large, constant saturations of
the actuator were observed during closed-loop experiments for all cases under
investigation. Therefore, it was necessary to introduce a scaling factor 0<<1 in the
state-feedback to keep the actuator below the saturation limit. The largest possible
scaling factors have been found to be 1 = 0.265 , 2 = 0.35 and 3 = 0.5 for

15

each of the three flow models considered, and the corresponding scaled control is
in the form
(7.1)
( t ) = K a~ ( t ) .
The scaled LQ controls, equation (7.1), have also been simulated on corresponding
nonlinear models, equation (4.2), as depicted in Figure 10. It is evident that, though
the scaled LQ control for the given values of is not able to asymptotically
stabilize the origin of the nonlinear model, equation (4.2), it nevertheless provides
a significant reduction of the amplitude of the stable limit cycle in all three cases.
This result is in agreement with a mathematical analysis carried out on the
nonlinear finite-dimensional Galerkin model, equation (6.1), which predicts a
reduction of the amplitude of limit cycle (corresponding to the fundamental cavity
tone) as the gain increases from 0 to 0.5, with complete suppression of the
oscillation only possible for >0.5.

(a)
(b)
(c)
Figure 10 Closed loop responses at P 3 with different scaling factor . (a) baseline flow
model, (b) open-loop forced flow model with forcing frequency of 3920 Hz, (c) combined
flow model of the above two cases

The performance of the scaled control law, equation (7.1), has been tested
experimentally, for different flow and forcing conditions. We now discuss the
results obtained for Mach 0.3 cavity flow. Specifically, we present the closed-loop
sound pressure level spectra from sensor 5 located on the cavity wall in Figure 2
(sensor #6 exhibit similar results) obtained in closed-loop with each LQ controller
designed on the basis of the three flow models discussed in Section 6. In addition, a
comparison is made with the results obtained using open-loop sinusoidal excitation
at 3290 Hz (optimal forcing frequency), and a combined open-loop forcing with
closed-loop LQ control.
The results for closed-loop LQ control, shown in Figure 11 show a considerable
attenuation of the resonance peak in sensor location 5 (with a similar behavior at
sensor location 6), and a redistribution of the energy into various modes, especially
lower frequency modes, with much lower energy level. It is worth noting that the
results do not differ significantly when a forced flow model (Figure 10 (b)-(c)), is
considered in place of the baseline flow (Figure 10 (a)), although a slightly more
uniform attenuation of the SPL can be noted in the third case. This is somewhat
surprising, as one would expect the presence of forcing to improve the fidelity of
the model in closed-loop conditions, and ultimately to provide a richer model,

16

capable of delivering better results. The lack of significant improvement is

probably related to the particular technique for control separation that has been
employed to render the presence of the control input explicit in the model [35]. A
better resolution of the effect of external forcing may be obtained resorting to a
method of control separation that makes use of actuation modes directly at the
level of POD modeling. This is a current area of investigation.

(a)
(b)
(c)
Figure 11 Sound pressure level spectra obtained from sensor #5 in closed-loop
experiments with LQ design based on: baseline flow model (a), forced flow model
(b), and combined flow model (c)

A comparison with the results obtained under optimal open-loop forcing shown in
Figure 12 (a) reveals that, while both control cases forced the flow to multi-mode
regime, overall the closed-loop control performs better than the open-loop control.
Note specially the presence of a significant peak at the open-loop forcing
frequency (which is as high as 120 dB for the SPL recorded by sensor #6, not
shown) and lack thereof any significant peak in the feedback control case. This
behavior, intrinsic in open-loop forcing, may be somehow alleviated combining
optimal open-loop forcing with LQ feedback control, in which the open-loop
sinusoidal forcing plays the role of a feedforward control. From the point of view

(a)
(b)
(c)
Figure 12. Sound pressure level recorded by sensor #5 under open-loop forcing (a),
and under combined open-loop and feedback control designed on the forced flow
model (b), and combined open-loop and feedback control designed on the combined
flow model (c)

of the finite-dimensional modeling and control separation methodology adopted

here, this situation is indeed preferable to using only LQ feedback control derived
on the basis of forced flow models, as the addition of sinusoidal forcing resembles
the conditions under which the forced model has been derived. Also, one may hope

17

to combine the beneficial effects of feedback control in terms of a more uniform

attenuation of the resonance with the sharper results obtained by open-loop forcing.
The results reported in Figure 12 (b)-(c) seem to validate this conjecture only in
part, and further investigation is needed to clarify the appropriateness of this
approach.

Concluding Remarks

The work presented and discussed in this paper is part of our ongoing research
activities in the development of reduced-order models based feedback control of
subsonic cavity flows. The cavity is shallow with L/D of 4 and spans the width of
the wind tunnel test section. The facility can be operated continuously between
Mach 0.2 and 0.7, but the majority of the work was carried out around Mach 0.3
with a Reynolds number based on the cavity depth of approximately 105. The
output of a compression driver is channeled to the cavity leading edge where it
exits at an angle of 30o with respect to the main flow through a 2-D slot of 1 mm
height that spans the cavity width. This arrangement provides zero net mass, nonzero net momentum flow for actuation, similar to that of a synthetic jet. Actuation
can be achieved in the frequency range of 1-20 kHz. The actuator to main flow
momentum ratio is in the range of 10-4 to 10-6. With open-loop forcing, we can
force the cavity to operate in a single-mode and lock onto various Rossiter modes
or to operate in multi-mode with rapid switching between modes.
The work includes using various laser based flow diagnostics to understand flow
physics and also to obtain detailed data for the derivation of reduced-order models
for controller design. Particle image velocimetry data and the snapshot based
proper orthogonal decomposition technique are used to extract the most energetic
flow features or POD eigenmodes. For each flow case, 1000 PIV snapshots are
used (over 700 snapshots are needed for the kinetic energy to converge). The
Galerkin projection of the Navier-Stokes equations onto the POD modes is used to
derive a set of non-linear ordinary differential equations, which govern the time
evolution of the modes, and to use for the controller design. Quadratic stochastic
estimation is used to correlate PIV data to surface pressure data thus enabling a
real-time update the state of the model based on the continuously time resolved
pressure measurements.
Three sets of PIV snapshots of a Mach 0.3 cavity flow were used to derive three
reduced-order models for the controller design: (1) snapshots from the baseline (no
control) single-mode flow, (2) snapshots from the same flow open-loop forced at
3920 Hz which produces multi-mode cavity resonance, and (3) combined
snapshots from the cases 1 and 2. A linear-quadratic optimal controller based on all
three models was designed to reduce cavity flow resonance and tested in the
experiments. The results obtained for all three flow models outperform those
obtained using open-loop control, and indicate that feedback control strategies
based on reduced-order flow models represent a compelling approach to subsonic
cavity flow control. Notwithstanding the encouraging results reported and

18

discussed in this work, further investigation is needed to understand how to

incorporate more effectively the presence of actuation in reduced-order POD
models of the flow system that can capture more closely the behavior of forced
flows, as well as to clarify the interplay between feedforward and feedback control.

Acknowledgments
This work is supported by the AFRL/VA and AFOSR through the Collaborative
Center of Control Science. The authors would like to thank Hitay zbay, Chris
Camphouse, and Kihwan Kim for help and fruitful discussions.

References
[1]
[2]
[3]
[4]
[5]
[6]

[7]

[8]

[9]

[10]
[11]

[12]

[13]

[14]

Cattafesta III, L.N., Garg, S., Choudhari, M., and Li, F., Active Control of FlowInduced Cavity Resonance, AIAA Paper 97-1804, June 1997.
Gad-el-Hak, M., Flow Control Passive, Active, and Reactive Flow Management,
Cambridge University Press, New York, NY, 2000.
Williams, D., Fabris, D., and Morrow, J., Experiments on Controlling Multiple
Acoustic Modes in Cavities, AIAA Paper 2000-1903, 2000.
Kegerise, M., Cattafesta, L., and Ha, C., Adaptive Identification and Control of
Flow-Induced Cavity Oscillations, AIAA Paper 2002-3158, 2002.
Rowley, C., and Williams, D., Control of Forced and Self-Sustained Oscillations in
the Flow Past a Cavity, AIAA Paper 2003-0008, 2003.
Cattafesta III, L. N., Williams, D. R., Rowley, C. W., and Alvi, F. S., Review of
Active Control of Flow-Induced Cavity Resonance, AIAA Paper 2003-3567, June
2003.
Samimy, M., Debiasi, M., Caraballo, E., zbay, H., Efe, M.O., Yuan, X., DeBonis,
J., and Myatt, J.H., Closed-Loop Active Flow Control: A Collaborative Approach,
AIAA Paper 2003-0058, 2003.
Siegel, S., Cohen, K., Seidel, J., and McLaughlin, T., Feedback Control of a
Circular Cylinder Wake in Experiments and Simulations (Invited), AIAA Paper
2003-3569, June 2003.
Gerhard, J., Pastoor, M., King, R., Noack, B., Dillmann, A., Morzynski, M., and
Tadmor, G., Model-Based Control of Vortex Shedding using Low-Dimensional
Galerkin Models, AIAA Paper 2003-4262, June 2003
Glauser, M. N., Higuchi, H., Ausseur, J., and Pinier, J., Feedback Control of
Separated Flows (Invited), AIAA Paper 2004-2521, June 2004.
Tadmor, G., Noack, B., Morzynski, M., and Siegel, S., Low-Dimensional Models
for Feedback Flow Control. Part II: Control Design and Dynamical Estimation,
AIAA Paper 2004-2409, June 2004.
Rossiter, J.E., Wind Tunnel Experiments on the Flow Over Rectangular Cavities at
Subsonic and Transonic Speeds, RAE Tech. Rep. 64037, 1964 and Aeronautical
Research Council Reports and Memoranda No. 3438, Oct. 1964.
Heller, H. H., and Bliss, D. B., The Physical Mechanisms of Flow-Induced
Pressure Fluctuations in Cavities and Concepts for their Suppression, AIAA Paper
75-491, March 1975.
Rockwell, D., and Naudascher, E., ReviewSelf-Sustaining Oscillations of Flow
past Cavities, Journal of Fluids Engineering, Vol. 100, 1978, pp. 152-165.

19

[15]

[16]
[17]
[18]
[19]
[20]

[21]

[22]
[23]
[24]
[25]
[26]

[27]

[28]

[29]

[30]

[31]

[32]
[33]

Ziada, S., Ng, H., and Blake, C., Flow Excited Resonance of a Confined Shallow
Cavity in Low Mach Number Flow and its Control, Journal of Fluids and
Structures, 18, 2003, p. 79-82.
Debiasi, M. and Samimy, M., Logic-Based Active Control of Subsonic Cavity
Flow Resonance, AIAA Journal, Vol. 42, No. 9, pp. 1901-1909, September 2004.
Kerschen, E., and Tumin, A., A Theoretical Model of Cavity Acoustic Resonances
Based on Edge Scattering Processes, AIAA Paper 2003-0175, 2003.
Alvarez, J., Kerschen, E., and Tumin, A., A Theoretical Model for Cavity Acoustic
Resonances in Subsonic Flows, AIAA Paper 2004-2845, 2004.
Cattafesta, L., Garg, S., Kegerise, M. and Jones, G., Experiments on Compressible
Flow-Induced Cavity Resonance, AIAA Paper 1998-2912, 1998.
Debiasi, M., Little, J., Malone, J., Samimy, M., Yan, P., and zbay, H., An
Experimental Study of Subsonic Cavity Flow Physical Understanding and
Control, AIAA Paper 2004-2123, June 2004.
Stanek, M.J., G., Kibens, V., Ross, J.A., Odedra, J., and Peto J.W., High Frequency
Acoustic Suppression The Mystery of the Rod-in-Crossflow Revealed, AIAA
Paper 2003-0007, January 2003.
Shaw, L., Active Control for Cavity Acoustics, AIAA Paper 98-2347, June 1998.
Grove, J., Leugers, J., and Akroyd, G., USAF/RAAF F-111 Flight Test with Active
Separation Control, AIAA Paper 2003-0009, January 2003.
Shaw, L., and Northcraft, S., Closed Loop Active Control for Cavity Resonance,
AIAA Paper 99-1902, May 1999.
Cattafesta, L., Shukla, D., Garg, S., and Ross, J., Development of an Adaptive
Weapons-Bay Suppression System, AIAA Paper 1999-1901, 1999.
Williams, D.R., Rowley, C., Colonius, T., Murray, R., MacMartin, D., Fabris, D.,
and Albertson, J., Model-Based Control of Cavity Oscillations Part I:
Experiments, AIAA Paper 2002-0971, 2002.
Rowley, C. W., Williams, D. R., Colonius, T., Murray, R. M., MacMartin, D. G.,
and Fabris, D., Model-Based Control of Cavity Oscillations Part II: System
Identification and Analysis, AIAA Paper 2002-0972, January 2002.
Cabell, R. H., Kegerise, M. A., Cox, D. E., and Gibbs, G. P., Experimental
Feedback Control of Flow Induced Cavity Tones, AIAA Paper 2002-2497, June
2002.
Caraballo, E., Yuan, X., Little, J., Debiasi, M., Yan, P. Serrani, A., Myatt, J., and
Samimy, M., Feedback Control of Cavity Flow Using Experimental Based
Reduced Order Model, AIAA Paper 2005-5269, June 2005.
Caraballo, E., Yuan, X., Little, J., Debiasi, M., Serrani, A., Myatt, J., and Samimy,
M., Further Development of Feedback Control of Cavity Flow Using Experimental
Based Reduced Order Model, AIAA Paper 2006-1405, January 2006.
Efe, M., Debiasi, M., Yan, P., zbay, H., and Samimy, M., Control of Subsonic
Cavity Flows by Neural Networks-Analytical Models and experimental Validation,
AIAA Paper 2005-0294, January 2005.
Yan, P., Debiasi, D. Yuan, X., Little, J., zbay, H., and Samimy, M., Closed-loop
Linear Control of Cavity Flow, to appear in AIAA Journal, 2006.
Samimy, M., Debiasi, M., Caraballo, E., Malone, J., Little, J., zbay, H., Efe, M. .,
Yan, P., Yuan, X., DeBonis, J., Myatt, J. H., and Camphouse, R. C., Exploring
Strategies for Closed-Loop Cavity Flow Control, AIAA Paper 2004-0576, January
2004.

20

[34]

[35]

[36]
[37]
[38]
[39]

[40]
[41]

[42]
[43]
[44]

[45]

Yuan, X. , Caraballo, E., Yan, P., zbay, H., Serrani, A., DeBonis, J., Myatt, J. H.
and Samimy, M., Reduced-Order Model-Based Feedback Controller Design for
Subsonic Cavity Flows, AIAA Paper 2005-0293, January 2005.
Efe, M.., and zbay, H., Proper Orthogonal Decomposition for Reduced Order
Modeling: 2D Heat Flow, IEEE Int. Conf. on Control Applications (CCA'2003),
June 23-25, Istanbul, Turkey, pp. 1273-1278, 2003
Little, J., Debiasi, M., and Samimy, M., Flow Structure in Controlled and Baseline
Subsonic Cavity Flows, AIAA Paper 2006-0480, January 2006.
Caraballo, E., Malone, J., Samimy, M., and DeBonis, J., A Study of Subsonic
Cavity Flows - Low Dimensional Modeling, AIAA Paper 2004-2124, June 2004.
Lumley, J. The Structure of Inhomogeneous Turbulent Flows, Atmospheric
Turbulence and wave propagation. Nauca, Moscow. 1967 166-176.
Glauser, M., Eaton, E., Taylor, J., Cole, D., Ukeiley, L., Citrinity, J., George, W and
Stokes, S., Low-Dimensional descriptions of Turbulent Flows: Experiment and
Modeling. AIAA Paper 1999-3699, June- July 1999
Holmes, P., Lumley, J.L., and Berkooz, G., Turbulence, Coherent Structures,
Dynamical System, and Symmetry, Cambridge University Press, Cambridge, 1996.
Delville, J., Cordier, L. and Bonnet, J.P., Large-Scale-Structure Identification and
Control in Turbulent Shear Flows, In Flow Control: Fundamentals and Practice,
edited by Gad-el-Hak, M., Pollard A. and Bonnet, J., Springer-Verlag, 1998, pp.
199-273.
Sirovich, L. Turbulence and the Dynamics of Coherent Structures, Quarterly of
Applied Math. Vol. XLV, N. 3, 1987, pp. 561-590.
Rowley, C. W., Modeling, Simulation and Control of Cavity flow Oscillations,
Ph.D. thesis, California Institute of Technology. 2002.
Noack, B., Tadmor, G., and Morzynski, M. ., Low-Dimensional Models for
Feedback Flow Control. Part I: Empirical Galerkin Models, AIAA Paper 20042408, June 2004.
Adrian, R. J., On the Role of Conditional Averages in Turbulent Theory,
Turbulence in Liquids, Science Press, Princeton, 1979.