AIAA JOURNAL

Modified Extremum-Seeking Closed-Loop System

for Jet Mixing Enhancement

D. W. Fan,∗ Z. Wu,† H. Yang,∗ J. D. Li,‡ and Y. Zhou§

Harbin Institute of Technology, 518055 Shenzhen, People’s Republic of China
DOI: 10.2514/1.J055644
A modified extremum-seeking closed-loop control system is proposed for jet mixing enhancement. The system
introduces a variable perturbation component that may act to adjust adaptively the searching range of the optimal
control signal. The actuator is a single pulsed minijet injected radially into the main jet before its issue. The excitation
frequency f e for the minjet is controlled by an electromagnetic valve. The open-loop control measurements indicate
that, given the duty cycle and the mass flow rate ratio of the minijet to the main jet, the decay rate K of jet centerline
mean velocity, which provides a measure for jet mixing, displays a maximum at the optimum frequency ratio
f e ∕f 0
0.5, where f 0 is the preferred-mode frequency of the main jet. Comparison is made between the modified
Downloaded by UNIVERSITY OF CAMBRIDGE on July 17, 2017 | http://arc.aiaa.org | DOI: 10.2514/1.J055644

and classical systems. Although both may find automatically the optimal f e ∕f 0 and hence a maximum K, the modified
system achieves a contraction in the fluctuating f e , when the control target is reached, by 70%, and subsequently an
increase in K by 12%. The new system further exhibits a better robustness than the classical.

Nomenclature U = streamwise velocity, m∕s

a = amplitude of sinusoidal perturbation signal, Hz Uc = jet centerline velocity, m∕s
Cm = mass flow rate ratio of minijet to that of the Ue = jet centerline velocity at the jet exit, m∕s
main jet Urms = rms values of streamwise velocity U, m∕s
D = diameter of the jet exit, mm U5D = jet centerline velocity at x∕D  5, m∕s
fe = excitation frequency of the minijet, Hz u = streamwise fluctuating velocity, m∕s
fe;i = initial excitation frequency of the minijet, Hz x, y, z = coordinates; x and z are along the streamwise
fe;optimal = optimal excitation frequency, Hz direction and the direction opposite to the minijet
fp = frequency of sinusoidal perturbation signal, Hz injection, respectively, and y is normal to both
f0 = preferred-mode frequency in the uncontrolled x and z, following the right-hand system
jet, Hz Δ = increment
GHP = magnitude of the high-pass filter transfer ε = dissipation rate
function η = Kolmogorov length scale, m
HP = high-pass filter λT = Taylor microscale, m
j = imaginary unit v = kinematic viscosity, m2 ∕s
K = jet centerline mean velocity decay rate, Φ = phase spectrum

Ue − U5D ∕Ue φHP = phase of the sinusoidal perturbation signal
KDC = constant component of the output K ω = angular frequency of sinusoidal perturbation
KHP = output of the high-pass filter signal, rad∕s
KLP = output of the low-pass filter ωHP and ωLP = cutoff frequencies of the high-pass filter and
k = gain low-pass filter, respectively, rad∕s
L = integral length scale, m – = time-averaged quantity
LP = low-pass filter
l = distance between the minijet center and the jet
exit, mm I. Introduction
_
m = mass flow rate of the minijet, g∕ min
R = radial distance of contraction from nozzle axis,
mm
A S ONE of the typical shear flows, jet mixing is widely used in
engineering applications (e.g., in aero and automobile engines,
combustion, heat transfer, and chemical reactors). Jet mixing control
ReD = Reynolds number based on the diameter of the
has drawn significant attention in the past few decades from scientists
jet exit
and engineers [1–5].
t = time, s
The jet mixing control can be passive and active. The passive
techniques such as tabs [6] and noncircular jet exits [2] may have
impressive performances. However, once mounted, tabs are difficult
Received 31 August 2016; revision received 24 April 2017; accepted for to be removed or rearranged; likewise, it would be unpractical for any
publication 1 May 2017; published online 6 July 2017. Copyright © 2017 by
the American Institute of Aeronautics and Astronautics, Inc. All rights
engineering application to implement frequently noncircular nozzle
reserved. All requests for copying and permission to reprint should be geometry alteration due to cost and physical constraints. In practice,
submitted to CCC at www.copyright.com; employ the ISSN 0001-1452 the jet operation conditions may vary; for example, an optimally
(print) or 1533-385X (online) to initiate your request. See also AIAA Rights performed nozzle under a certain operation condition may not work
and Permissions www.aiaa.org/randp. optimally under another one. In aircraft, the highly intense mixing is
*Ph.D. Student, Institute for Turbulence-Noise-Vibration Interactions and required only for specific flight phases such as taking off or
Control. combating. In contrast, the active control of jets has potential to
†
Ph.D. Student, Institute for Turbulence-Noise-Vibration Interactions and achieve more flexible and drastic flow modifications. Naturally,
Control; [email protected] (Corresponding Author).
‡
Professor, Institute for Turbulence-Noise-Vibration Interactions and
many active techniques were proposed and investigated with success
Control; College of Engineering and Science, Victoria University, Melbourne, to certain different degrees, including acoustic excitation [1], plasma
Australia. actuators [5], synthetic jet actuators [4], oscillating boundaries based
§
Professor, Institute for Turbulence-Noise-Vibration Interactions and on piezoelectric actuators [7], continuous control jets [8,9], and
Control; [email protected]. Senior Member AIAA (Corresponding Author). pulsed minijets [10–13].
Article in Advance / 1
 2 Article in Advance / FAN ET AL.

Active methods are further divided into open-loop and closed-loop performance of the system becomes more effective with increasing a,
controls. The former accounts for most of the active control but the static performance goes to the opposite, and vice versa [25].
investigations previously reported. However, the turbulent shear flow However, for many flow control applications, it is highly desirable for
control can be significantly more effective if a closed-loop control is the controller to have not only a good dynamic performance during ES
applied [14–16]. A control must interact with turbulent fluctuations scheme but also a well-behaved static performance after convergence.
in a jet because the random aspects of the fluctuations reduce the On the basis of the classical methods, Wang et al. [25] proposed
effectiveness of an open-loop control. The sensor-feedback loop may a modified ES scheme, which determines a through an estimated
treat effectively the random-phase problem in turbulence dynamics gradient, to reduce the steady-state error and response time. They
[17] and allow jet mixing to be flexibly controlled [18]. further validated the scheme based on the simulation data, though no
Various control algorithms have been proposed and investigated attempt was made to develop a physical system. This scheme may have
for the closed-loop control, as discussed in a recent compendium potential to improve immensely the dynamic and static performance of
by Brunton and Noack [19]. Extremum seeking (ES) is one of the the ES scheme.
model-free closed-loop control schemes for dynamic problems, which This work aims to apply Wang et al.’s modified ES control scheme
is based on real-time output signals where a limited knowledge of the [25] for jet mixing enhancement and to overcome the drawbacks in Wu
internal structure of a system is available. The increasing complicacy et al.’s control system [16]. A single pulsed minijet, injected radially
of engineering systems has led to many challenges in the optimal into the main jet before its issue, is deployed as the actuator. The decay
control because of the difficulty in obtaining analytic solutions to rate of the jet centerline velocity and the minijet pulsation frequency fe
the optimization problems of multivariable, nonlinear, and are used as the feedback and the control signals, respectively. The
multidimensional systems. Furthermore, a model-based solution for modified ES scheme is compared with the classical one in terms of the
Downloaded by UNIVERSITY OF CAMBRIDGE on July 17, 2017 | http://arc.aiaa.org | DOI: 10.2514/1.J055644

these complicated optimization problems is likely to be a challenge due steady-state error and convergence time. The robustness of the two
to the system complexity and random variations of the initial and schemes is also examined by changing the ReD suddenly from 8000 to
boundary conditions. Therefore, a model-free ES scheme is attractive. 15,000, and their adaptabilities at different initial excitation frequency
The objective of the ES scheme is to seek maximizing or minimizing fe;i are also compared. Experimental details are given in Sec. II, and
some physical quantities via an online operation. The ES controller can the performances of the open-loop and closed-loop controls are
be implemented to achieve the optimal operation when the steady-state presented in Sec. III. This work is concluded in Sec. IV.
of the map has a well-defined extremum [20]. When the control
performance is characterized by a saturated plateau in the input–output
map, the slope-seeking controller is more suitable than the ES II. Experimental Details
controller [21]. The model-free ES system has found numerous A. Jet Facility and Actuator System
applications for fluid mechanics control [22,23]. Becker et al. used the Figure 1a shows the schematic diagram of the air jet facility.
ES controller to adjust the pulsed jets near the leading edge of the Experiments based on one radial minijet were performed in an air jet
single-slotted flap for suppressing flow separation over the flap, facility, which consists of a main round jet with a minijet assembly
achieving an improved lift gain compared to their open-loop studies that includes a stationary disk and an electromagnetic-valve. The air
[24]. Using a classical ES control algorithm, Wu et al. developed for both main jet and minijet came from a compressed air supply with
a closed-loop system to search automatically the optimal excitation 12 bar gauge pressure. The compressed air passed through a mixing
frequency of two pulsed minijets and hence to maximize the box, a plenum chamber, a 300-mm-long diffuser of 15 deg in half-
entrainment of a turbulent round jet [16]. However, the system suffers angle, two fine screens (7 mesh∕cm), and a cylindrical settling
from a relatively large oscillation in both input and output signals after chamber of 400 mm in length and 114 mm in inner diameter. The
convergence due to the limitation of the classical algorithm. As will be nozzle contraction followed a contour specified by the equation
demonstrated in this work, this oscillation has adversely affected to R  57 − 47sin1.5
90 − 9x∕8 mm, as used by [9,16]. The nozzle
a certain extent the performance of the control system. was extended with a 47-mm-long smooth tube of diameter
In the classical ES scheme, the dynamic and static performances of D  20 mm. The ReD  Ue D∕v of the main jet was fixed at 8000,
the scheme are directly related to the amplitude a of the perturbation except for the investigation of the robustness of the control schemes,
sinusoidal signal, i.e., a sin
ωt, where ω and t are the perturbation where Ue is the jet centerline velocity measured at the exit of the
frequency and time of the sinusoidal signal, respectively. The dynamic nozzle extension, and v is the kinematic viscosity.

Fig. 1 Schematic of experimental setup: a) main-jet assembly, and b) minijet assembly.

 Article in Advance / FAN ET AL. 3

Figure 1b shows the schematic diagram of the pulsed minijet for acquiring the signal from the hot wires, and the output was used to
actuator assembly. The stationary disk was drilled with six orifices generate the control signal for driving the electromagnetic valve
of 4 mm in diameter along the radial direction. The orifices were and hence the minijet. The control signal was a square wave with
connected via short plastic hose to the electromagnetic valves. A a frequency of fe and a duty cycle of 15%. A LabVIEW program
contraction nozzle of 1 mm in diameter was drilled for each orifice was written to establish the link between the computer and the
and is located at l  17 mm upstream of the jet exit. This work will NI USB-6361 hardware. The real-time information of several
be focusing on developing a new control scheme. Therefore, only one variables from the closed-loop control system was displayed using
single minijet injection was investigated to avoid the complexity the LabVIEW program.
involved when using multiple minijets (such as synchronizing
the injection of the minijets). The minijet was controlled by an C. Flow Visualization Setup
electromagnetic valve (Koganei K2-100SF-09-LL) with a maximum A planar particle image velocimetry (PIV) system (i.e., Dantec
frequency of 1 kHz. The electromagnetic valve was driven by a dc SpeedSence90C10) was deployed for flow visualization and PIV
power supply with 0–5 V square-wave signal that may act to adjust measurements in the injection (x, z) plane and the orthogonal
the frequency of air injection with a maximum of 500 Hz, exceeding noninjection (x, y) plane. ATSI oil droplet generator (TSI MCM-30)
3f0 at ReD  8000, where f0 is the preferred-mode frequency of the was used to generate fog for seeding the flow. The seeding particles,
uncontrolled jet. The mass flow rates of the main jet and minijet were which have a mean diameter of about 1μm, were supplied into the
changed and measured, respectively, by two flow meters whose mixing chamber (Fig. 1a), which was mixed with air and were fully
experimental uncertainty was no more than 1%. spread throughout the main jet. Flow illumination was provided by
The center of the jet exit is taken as the origin of the coordinate
Downloaded by UNIVERSITY OF CAMBRIDGE on July 17, 2017 | http://arc.aiaa.org | DOI: 10.2514/1.J055644

two standard pulsed laser (Beam Tech 500-10) sources with 532 nm
system, with the x axis along the streamwise direction, the z axis wavelength and a maximum energy output of 120 mJ per pulse.
along the radial minijet, and the y axis along the direction normal to Particle images were captured at a sampling rate of 10 Hz. The
the (x, z) plane, following the right-hand rule. The (x, z) and (x, y) synchronization of flow illumination and image capturing was
planes are referred to as the injection and noninjection planes, controlled by Dynamic Studio v3.41. The captured images covered
respectively. Following [9,16], the jet centerline decay rate an area of x∕D  0 ∼ 6 and y∕D or z∕D  −3 ∼ 5 in the (x, y) and
K 
Ue − U5D ∕Ue is used to evaluate the mixing effect of the (x, z) planes. The uncertainty in the position of the traversing system
main jet under the control of the minijet, where Ue and U5D are the jet was 0.1 mm.
centerline mean velocities at x∕D ∼ 0 and x∕D  5, respectively. For each PIV experiment, 1200 pairs of images were recorded.
The overbar indicates time-averaging quantity. Background noises were estimated as the minimum of grayscale
values in a time series and subtracted from each of the images in the
B. Hot-Wire Anemometry Setup and Closed-Loop Control Facilities ensemble. Each single image pair was computed using the image
Hot wires were used to measure the streamwise velocity at balancing algorithm included in the PIV package software by
x∕D < 6. The hot wire was operated on a constant-temperature circuit Dynamic Studio v3.41, which has its origin in an adaptive PIV
(Dantec Streamline) at an overheat ratio of 0.6. The output voltage algorithm. The adaptive PIV analysis of each image pair was
signal from the constant-temperature circuit anemometer (CTA) was conducted with a grid step size of 16 × 16 pixels, a minimum
filtered at a cutoff frequency of 3 kHz and sampled with a sampling interrogation area of 32 × 32 pixels, and a maximum interrogation
frequency of 6 kHz using a 16-bit A/D board (NI USB-6361). The area of 64 × 64 pixels to minimize the measurement uncertainty. The
uncertainty of the hot-wire measurement is estimated to be less uncertainty is 2% in the estimate of the mean velocity, and the
than 2%. uncertainty in positioning the laser sheet is about 1%.
Two single hot wires were used in the closed-loop control
(i.e., wire 1 and wire 2), as indicated in Fig. 2. Wire 1 was used to
monitor Ue. To minimize the effects of the minijet on the main jet exit III. Results and Discussion
velocity signal, wire 1 was located in the noninjection plane, at A. Open-Loop Control
x∕D ∼ 0, y∕D  −0.3, z∕D  0. Wire 2 was used to monitor U5D at Figures 3a–3c show the power spectral density function of streamwise
x∕D  5, y  z  0. Both Ue and U5D were used to calculate K, velocity U measured on the centerline at ReD  8000; 9333, and
which was used in the closed-loop control system. Note that the 16,000 at x∕D  1–5. There is a pronounced peak at f  135; 155,
average time determines the loop time (i.e., 0.4 s) of the closed-loop and 250 Hz at ReD of 8000, 9333, and 16,000, respectively, implying the
control system. The choice of 0.4 s is a compromise between the occurrence of the predominant preferred-mode vortices [26]. The effect
converge time and the measurement accuracy. of fe on K was investigated for the three ReD with a constant mass
An NI USB-6361 system, connected to a computer, was used in the flow rate of the minijet (i.e., m _  3.2 g∕ min), corresponding to
closed-loop control experiment. The system contains 16 analog Cm  2.3; 1.9, and 1.2% for ReD  8000; 9333, and 16,000,
inputs (16 bit) and two analog outputs (16 bit). The inputs were used respectively, where Cm is the ratio between the mass flow rate of the
minijet to that of the main jet. One objective of the open-loop control
experiments was to provide the steady-state input–output map between
fe and K to evaluate the feasibility of the closed-loop control. Figure 3d
presents the measured K at different fe for ReD  8000. Apparently, K
depends strongly on fe , showing a twin-peak pattern, with the global
maximum (K  0.425) at fe ∕f0 ≈ 0.5 and a minor local maximum at
fe ∕f0 ≈ 1.0. The former is referred to as the optimal excitation
frequency fe;optimal at ReD  8000, leading to a factor of 8 increase in K
compared with the uncontrolled jet. In other words, fe;optimal is one-half
of the preferred-mode vortex frequency f0. A similar observation is also
made at ReD  9333 and 16,000, except that K has dropped
significantly in comparison with that at ReD  8000 (Fig. 3d). Two
possible reasons may account for this drop. First, the minijet mass flow
rate, optimized at ReD  8000, was unchanged for the higher ReD and
may not correspond to the optimum mass flow rate for ReD  9333 and
16,000. Second, the control is expected to be less effective at higher ReD.
Figure 4 illustrates the time series of the streamwise velocity U
at different downstream locations between the controlled and
uncontrolled jets. To capture the characteristics of the shear layer,
Fig. 2 Experimental arrangement for the closed-loop control. a hot wire was traversed along the line of z∕D  −0.3 in the injection
 4 Article in Advance / FAN ET AL.
Downloaded by UNIVERSITY OF CAMBRIDGE on July 17, 2017 | http://arc.aiaa.org | DOI: 10.2514/1.J055644

Fig. 3 a–c) Power spectral density function of streamwise velocity U measured on the centerline at different ReD and d) steady-state input–output map of
K
Ff e ∕f 0  at different ReD .

start to show a quasi periodicity at x∕D  2–4 and random

fluctuations at x∕D  5 in the uncontrolled jet. In contrast, random
fluctuations exist at x∕D  3–5 in the controlled jet. Furthermore,
the signals from the controlled jet at x∕D  2 are similar to that for the
uncontrolled jet at x∕D  4. Beyond these respective positions, the
signals become irregular, implying the flows undergoing shear-layer
rollup and vortex pairing. Another mechanism for enhancing the
mixing is that the controlled jet becomes turbulent earlier than the
uncontrolled jet, resulting in a reduced potential core length.
To illustrate the profound impact of the actuation at fe;optimal on the
main jet, we show typical photographs (Fig. 5a) in both the injection
(x, z) and noninjection (x, y) planes from flow visualization. There is
a distinct difference in the flow structure with and without control.
The uncontrolled jet remains laminar near the nozzle exit, as is
evident from Fig. 5a. In contrast, the potential core under control
recedes significantly with early shear-layer rollup, and the vortex
paring also takes place early in both planes. The controlled jet is
characterized by extensive spreading (smoke-marked or white-
colored fluid in the noninjection plane, Fig. 5a), compared with that
in the uncontrolled jet. On the other hand, as shown in Fig. 5a, the
flow structure in the injection plane appears flapping [11,12],
Fig. 4 Time series of streamwise velocity U at different x∕D locations.
suggesting that the flapping motion be responsible for the greatly
Hot wire is placed at z∕D
−0.3 and y∕D
0. Duration is 0.05 s. The increased jet spread and the rapid decay of the centerline mean
same scale is used for all signals to facilitate comparison. velocity.
The time-averaged flowfield obtained from the PIV measurements
may provide a quantitative measure for jet entrainment and spreading.
plane. All the U signals displayed use identical vertical and horizontal The streamwise and lateral mean velocities are calculated from 1200
scales, respectively. Comparing with the uncontrolled jet, the U signals independent pairs of PIV images (about 60 s). Figure 5b presents the
from the controlled jet exhibit some stronger fluctuations. The signals contours of the time-averaged streamwise velocity normalized by the
 Article in Advance / FAN ET AL. 5
Downloaded by UNIVERSITY OF CAMBRIDGE on July 17, 2017 | http://arc.aiaa.org | DOI: 10.2514/1.J055644

Fig. 5 Comparison in the typical flow structure between the uncontrolled jet and the controlled jet a) photographs of flow visualization, and b) contours
 U e . ReD
8000.
of time-averaged velocity U∕

 e ,), with and without control.

time-averaged jet exit velocity (i.e., U∕U Yang et al. [12] manipulated a turbulent round jet using two
Three observations can be made. First, under the excitation of the in-phase unsteady minijets and observed the flapping motion of the
minijet, the location where the contour of U∕U  e  0.95 intersects jet column (their Figs. 6 and 7). The presently captured flow structure
the jet axis moves from x∕D ≈ 5.9 in the uncontrolled jet to 0.9 in the shown in Fig. 5a is very similar to that shown in their Figs. 6ci and 7b,
noninjection plane of the controlled jet, that is, the control has reduced that is, the jet flapping also occurs in the present experiments.
the potential core length by 85%. Second, as shown in Fig. 6, the Yang et al. attributed the rapidly increased decay of the centerline
starting location where the centerline velocity decay is closer to the jet mean velocity to the flapping motion and further unveiled a close
exit in the controlled jet than in the uncontrolled jet, and the decay rate association between the flapping motion and vortex pairing in the
of the controlled jet is higher as well. Third, the controlled jet starts near field. Figure 8a shows the phase spectrum ΦU1U2 between U1
spreading earlier, with a significantly higher spreading rate, in the
injection plane than the uncontrolled jet, as indicated by a comparison
between their contours of U∕U  e  0.1 (Fig. 5b). The observation is
linked to the early rollup in the injection plane (Fig. 5a). The maximum
U along the radial direction is almost at the center of the velocity
profile, indicating that the minijet injection does not change
appreciably the symmetry of main jet. Yang et al. [11,12] deployed two
asymmetrically arranged pulsed minijets to manipulate a turbulent
round jet and ascribed the substantially enhanced jet decay to two
physical phenomena produced by the minijets. One is the greatly
enhanced vortex pairing that vigorously engulfed ambient fluid to the
jet core, and the other is the jet column flapping that greatly promotes
jet spreading [11,12]. Similar mechanisms are responsible for the
present rapid decay of the centerline mean velocity.

Fig. 6 Comparison and difference in K between controlled Fig. 7 Block diagrams of a) the classical ES control scheme, and b) the
(f e ∕f 0
0.5, Cm
2.3%) and uncontrolled jets. modified ES control scheme.
 6 Article in Advance / FAN ET AL.
Downloaded by UNIVERSITY OF CAMBRIDGE on July 17, 2017 | http://arc.aiaa.org | DOI: 10.2514/1.J055644

Fig. 8 Representations of a) phase spectrum ΦU1U2 between two streamwise velocity signals measured at x∕D
1.5 and y∕D
0.3 with different
f e ∕f 0 , and typical flow structure in the (x, z) plane: b) f e ∕f 0
0.5, and c) 1.0. Cm
2.3%.

and U2 measured at x∕D  1.5 and y∕D  0.3 at different fe ∕f0 , the organized structures, as demonstrated by Yang and Zhou [13].
where the minimum ΦU1U2  −π occurs at fe ∕f0  0.5. Note In the far field, say x∕D > 10, however, the jet is fully developed,
that both ΦU1U2  −π and π correspond to exact antiphase [12], and the coherent structures generated in the near field have mostly
suggesting the flapping motion of the controlled jet at fe ∕f0  0.5 is broken up, whether under control or not. Thus, the present control is
strongest compared with other fe ∕f0 examined. The assertion is highly effective, though largely limited to the near field.
substantiated by the flow visualization data shown in Figs. 8b and 8c The minijet has a big influence on the turbulence. Figure 9 shows
for fe ∕f0  0.5 and fe ∕f0  1, respectively. Thus, it can be the streamwise evolution of turbulence intensity Urms ∕U c on the
concluded that the jet column flapping motion is the mechanism
behind the greatly enhanced jet mixing. This motion is characterized
by vortex pairing and will naturally be strongest under excitation of
fe ∕f0  0.5, instead of fe ∕f0  1.0.
Farther downstream (i.e., x∕D > 5), K rises as x∕D increases for
both the controlled and uncontrolled flows (Fig. 6). It can be seen in
Fig. 6 that the maximum ΔK occurs near x∕D  5, that is, K
estimated based on the streamwise velocity measured at x∕D  5 is
sensitive to the change in the control parameter fe, compared with K
estimated at other locations. The difference ΔK between the K values
with and without control diminishes beyond x∕D  5, that is, the
control is more effective in the near field. This is because the
effectiveness of the control is underpinned by greatly manipulating

Fig. 10 Comparison of the turbulence scales on the centerline between

Fig. 9 Streamwise evolution of turbulence intensity Urms ∕Uc for the the controlled (f e ∕f 0
0.5, Cm
2.3%) and uncontrolled jets: a) L,
controlled (f e ∕f 0
0.5, Cm
2.3%) and uncontrolled jets. b) λT , and c) η. Calculated from hot-wire data.
 Article in Advance / FAN ET AL. 7
Downloaded by UNIVERSITY OF CAMBRIDGE on July 17, 2017 | http://arc.aiaa.org | DOI: 10.2514/1.J055644

Fig. 11 Effects of the a) a, and b) f P on the input–output map (k
0, ReD
8000, f e ∕f 0
0.33).

centerline, for both the controlled (fe ∕f0  0.5, Cm  2.3%) and consistent with the occurrence of vortex paring [12] and the greatly
uncontrolled jets. Apparently, the minijet actuation reduces the increased K. In the far field (i.e., x∕D > 7.5), however, L is appreciably
potential core length. An intensity peak occurred at x∕D  3.5 reduced. It seems that the effect of control persists in the far field, at least
without control. This peak appears less pronounced and shifted to up to x∕D  20. Figures 10b and 10c display the centerline streamwise
x∕D  2 under control, resulting presumably from the breakdown of evolution of λT and η. The control results in a decreased λT but an
primary coherent structures [27]. This indicates that the breakdown increased η beyond x∕D > 4.0. η in general rises slowly in the
occurs earlier under control than that without control. The turbulence uncontrolled jet, as expected, but appreciably more rapidly under
intensities in both controlled and uncontrolled jets show a decrease control. Note that the difference in λT between the controlled and
after reaching their respective peaks and then increase farther uncontrolled jets diminishes as x∕D increases, whereas that in η rises
downstream. After x∕D  10, the turbulence intensities reach their rather significantly.
respective plateaus, indicating the fully developed turbulent jets.
However, the turbulent intensity (34%) under control is higher than B. Closed-Loop Control
that (24%) without control. Figure 7a shows the schematic diagram of the classical ES
The integral length scale L, Taylor microscale λT , and Kolmogorov algorithm. The flow control system is here considered as a block with
length scale η have been calculated with and without the control to the input fe
t and output K
t. A local maximum takes place on the
gain understanding of the control effects on the turbulent structures, curve K  F
fe  when fe  fe;optimal , where dK∕dfe  0. The
where algorithm can automatically adjust fe to approach fe;optimal following
Z the feedback law (i.e., dfe ∕dt  kdK∕dfe , where k is the gain and
U c τ0
L U
tU
t  τ dτ dK∕dfe is the local gradient on the input–output map). Thus, the
U2 0 autonomous adjustment of fe
t from the initial excitation frequency
fe;i to fe;optimal can be achieved.
(τ0 corresponds to the first zero crossing of the auto-correlation
1∕2 2 ∕
∂U∕∂t2 1∕2 , and
For computing dK∕dfe , a small-amplitude sinusoidal signal
function), λT 
2u2 ∕
∂U∕∂x2  ≈ U
2u  a sin
ωt can be superimposed on the control input fe , i.e.,
η 
v3 ∕ε1∕4
ε ≈ 15  ν
∂U∕∂x2 ≈ 15  νU −2
∂U∕∂t2 . Figure 10a fe
t  fe;i Δfe
tasin
ωt, where ω
 2πfp  is the perturba-
presents the centerline evolution of L with and without control. In the tion frequency and Δfe
t is the feedback increment (Fig. 7a). This
near field (i.e., x∕D < 7.5), L under control (fe ∕f0  0.5, Cm  2.3%) input can lead to an approximate sinusoidal output in K
t, fluctuating
is considerably larger than that without it, suggesting a great around the mean value K DC , with the amplitude governed by aF 0.
enhancement of the large-scale organized motions, which is fully That is, K
t  K DC  aF 0 sin
ωt, where F 0 is the gradient of the
 8 Article in Advance / FAN ET AL.

input–output curve. The K DC in the output can then be removed by obtain the feedback increment Δfe
t. In summary, if Δfe
t is
a first order HP, i.e., GHP
jω  jω∕
jωωHP , to produce negative for fe > fe;optimal, fe would decrease and vice versa.
KHP
t  jGHP jaF 0 sin
ωt  φHP , where jGHP j is the magnitude of Eventually, fe will converge to fe;optimal .
the HP transfer function. The filtered signal KHP
t can be demodulated Different
pfrom the classical ES, the constant amplitude a is replaced
by multiplying a perturbation sin
ωt. The demodulated signal by r jKLP j in the modified ES scheme, where r is a constant (Fig. 7b).
ξ
 jGHP jaF 0 sin
ωt  φHP  sin
ωt contains a constant, i.e., Because K LP  jGHP j
F 0 a∕2 cos
φHP  as mentioned earlier
jGHP j
F 0 a∕2 cos
φHP , and a time-varying component, i.e., (Fig. 7a), we have a  r2 jGHP
F 0 ∕2 cos
φHP j, which indicates
jGHP j
F 0 a∕2 cos
2ωt  φHP , where that the amplitude a of the perturbation signal is proportional
to jF 0 j because r2 jGHP
1∕2 cos
φHP j is a constant. Thus, a will
1 gradually decrease as jF 0 j approaches 0. Namely, the output fluctuation
jGHP j  p
1 
ωHP ∕ω2 will decrease as fe converges to fe;optimal .

and φHP  arg
GHP   arctan
ωHP ∕ω. The time-varying component C. Parameter Choice of the Extremum-Seeking Scheme
can be filtered out by a first-order low-pass (LP) filter. Thus, the output Because the flow system is nonlinear, a hysteresis may exist
KLP  jGHP j
F 0 a∕2 cos
φHP  is a steady signal proportional to the between the input and output signals in most of the active flow control
local gradient F 0 . The K LP signal is then integrated and amplified by k to applications when the input signal is oscillating. Using a sinusoidal
Downloaded by UNIVERSITY OF CAMBRIDGE on July 17, 2017 | http://arc.aiaa.org | DOI: 10.2514/1.J055644

Fig. 12 The f e responses (broken lines) of the controller at different gain k, where the solid line indicates the signal of the quasi-steady component of f e t.

a) b)

c) d)

e) f)

Fig. 13 Comparison in f e ∕f 0 , K, and a (in hertz) between the classical (left column) and modified (right column) ES schemes.
 Article in Advance / FAN ET AL. 9
Downloaded by UNIVERSITY OF CAMBRIDGE on July 17, 2017 | http://arc.aiaa.org | DOI: 10.2514/1.J055644

Fig. 14 Responses of the classical (left column) and modified (right column) ES schemes as ReD varies between 8000 and 15,000: time histories of
a–b) ReD , c–d) f e ∕f 0 , e–f) K, and g–h) a (in hertz).

signal in the ES scheme may result in a phase shift in the output, exists. When a  15 Hz, fe
t produces a distinct phase lag in
which may adversely influence the gradient estimation in the ES K  F
fe ∕f0 . Thus, we choose a  10 Hz, which yields an
scheme. However, an appropriate setting of the perturbation adequately large effect on K and produces only a minor hysteresis in the
parameters (i.e., fp , a, and r) of the controller may minimize the fe –K map.pFor the modified ES, we choose r  7.5, such that the initial

influence of the hysteresis [16]. For a fair comparison, the same value of r jKLP j is 10 Hz.
parameters (i.e., fe;i , fp , and k) are used in the two control schemes. Figure 11b presents the response of K to the variation in fp from
The r for the modified ES is determinedpbased on the classical ES 0.0625 to 0.25 Hz. First, the irregular changes in K take place more
such that the initial value of r jK LP j  a, as is used in the obviously at fp  0.0625 Hz (Fig. 11b), compared with that at
classical ES. fp  0.125 Hz (Fig. 11b) and 0.25 Hz (Fig. 11b). The reason may be
The parameters of the ES controller are determined experimentally due to that the change in fe is too slow at fp  0.0625 Hz. Thus, K is
at ReD  8000, with the mass flow rate of the minijet fixed at more easily influenced by the external flow, affecting adversely the
m_  3.2 g∕ min (i.e., Cm  2.3%). The selection of the key control accurate gradient estimation. Second, the fe –K map exhibits an
parameters is similar to that used in [16,28]. A sinusoidal perturbation extreme hysteresis at fp  0.25 Hz. Therefore, fp  0.125 Hz is
with an initial frequency fe;i
 45 Hz, i.e., fe
t  fe;i  a sin
ωt, selected for the controller because only a small phase lag occurs
is applied in the input signal to examine the impact of different a and fp between fe and K at this frequency. Furthermore, the cutoff
on K. k is set at 0 when selecting other parameters. The parameters a and frequencies for the HP and LP filters are both selected as 0.1 fp
fp are investigated independently by fixing one and testing the other. As (i.e., ωHP  ωLP  0.1 × 2πfp ).
a result, we may determine the optimal perturbation parameters of the With a  10 Hz and fp  0.125 Hz, experiments were then
controller such that the output K is in phase with fe
t. Figure 11a shows performed at ReD  8000 and m _  3.2 g∕ min
Cm  2.3% to
the comparison between the responses of K measured with a  5; 10, investigate the optimal gain k of the classical ES algorithm. Figure 12
and 15 Hz (k  0, ReD  8000, fe;i  45 Hz, fp  0.125 Hz), exhibits the time response history of the closed-loop control system
respectively. The relation between K and fe appears to be approximately for different gains k  10, 20, and 30. The initial perturbation
linear. At a  5 Hz, the range of the sinusoidal perturbation generates frequency fe;i  45 Hz is applied for the minijet at t  0, which is
a minor change and irregularity in K, compared with that at a  10 and less than fe;optimal identified in the open-loop case (Fig. 3d). The fe
15 Hz, which means that a  5 Hz is too small, and a large hysteresis increases immediately once the signal is feedback into the controller
 10 Article in Advance / FAN ET AL.

at t > 0. For k  10, it takes about 108 s for fe;s converges to near perturbation amplitude diminishes with fe approaching fe;optimal and is
fe;optimal , where fe;s  fe;i  Δfe
t is the quasi-steady component stabilized in a range of 64.5–70.5 Hz, and the process takes about 50 s
of fe . As a result, K converges to 0.375 after 108 s. As k is increased, (Fig. 13b). Correspondingly, the K is increased from 0.25 to 0.42,
the variation in Δfe is enlarged. Consequently, the convergence stabilized within a range of 0.39–0.45 (Fig. 13d). Compared with the
time is reduced to about 59 s at k  20. If the k value is further classical ES scheme, the decrease in a of the modified scheme has led
increased to 30, the enlarged Δfe would result in an overshoot in to a reduction in the steady-state error of fe by 70%. More importantly,
fe , thus increasing the fluctuation in K and weakening the the mean value of the stabilized K goes up by 12%, compared with its
system stability. In conclusion, k  20 exhibited the shortest counterpart (0.375) of the classical scheme. The mean value of the
convergence time with a good stability in the closed-loop control stabilized K is a long time-averaged value (about 50 s) of K in Figs. 13c
system. In summary, a  10 Hz, r  7.5, fp  0.25 Hz, and and 13d. Furthermore, compared with the classical scheme,
k  20 are chosen for the two control schemes in the following the convergence time also shrinks by 15.3% (from 59 to 50 s), as
experiments. shown in Figs. 13a and 13b. The reason is mainly due to the self-
adjustment of the perturbation amplitude in the modified ES,
jF 0 jr2 jGHP cos
φHP ∕2j, which is directly proportional to jF 0 j
D. Comparison in Response Between Two Extremum-Seeking Schemes because the values of r and jGHP cos
φHP j are constant. The gradient
To gain a better understanding of the modified ES scheme, it is very jF 0 j diminishes to zero in the modified ES scheme. Thus, the amplitude
important to compare carefully the results of this scheme with those of a also decreases from 10 to a low level as shown in Fig. 13f. Therefore,
the classical one. As shown in Fig. 13, the two methods are distinct in the amplitude a of the modified scheme will be much less than that of the
performance from each other under the same parameters (i.e., k  20,
Downloaded by UNIVERSITY OF CAMBRIDGE on July 17, 2017 | http://arc.aiaa.org | DOI: 10.2514/1.J055644

classical scheme when the system approaches the stabilized state. Note
fe;i  45 Hz, fp  0.125 Hz, and a  10 Hz). fe rises gradually that the amplitude a in Fig. 13f starts from 0 due to the use of the HP and
with a fixed perturbation amplitude in the classical ES scheme. After LP filters. Similar phenomena can also be found in [29].
about 59 s, the controller is stabilized, with fe fluctuating within Robustness is another important aspect in the evaluation of
58–78 Hz (Fig. 13a). Meanwhile, K increases from 0.25 to 0.375 and is a closed-loop control system and is one of the significant indicators
stabilized within 0.35 to 0.44 eventually (Fig. 13c). In contrast, fe in for control performance. The robustness is examined when ReD is
the modified ES begins with a large initial perturbation amplitude. The changed suddenly at a fixed Cm and k. For the classical scheme, when
ReD is suddenly increased from 8000 to about 15,000 at t  120 s
(Fig. 14a), the fe –K map is changed, and fe;optimal rises as shown in
Fig. 3. The gradient of the fe –K curve becomes positive, resulting in
a positive Δfe . The controller increases fe until the maximum K is
reached again, with fe;optimal  120 Hz. This process has lasted for
80 s. In contrast, as ReD is suddenly raised at 104 s, the modified
controller takes only 50 s to achieve the maximum K, indicating
a significantly improved robustness and cutting down the
convergence time by 30 s. One of the important reasons is
the parameter a, which adapts itself with the gradient variation in the
modified scheme. Figure 14h shows that a can increase from 2 to 15,
when ReD is suddenly changed at 104 s, to increase the search range
until fe;optimal for ReD  15;000 is found. However, this parameter
a is a constant and cannot change with ReD in the classical scheme
(Fig. 14g). The a would also increase when ReD is suddenly
decreased at 198 s (Fig. 14h). To investigate the adaptation of the
modified controller, three different initial excitation frequencies are
Fig. 15 The f e response of the modified ES scheme at different f e;i . applied at ReD  8000 (i.e., fe;i  45, 60, and 75 Hz, respectively).

Fig. 16 Responses of the classical (left column) and modified (right column) ES schemes as ReD varies between 8000 and 9500: a–b) time histories of ReD ,
and c–d) f e ∕f 0 .
 Article in Advance / FAN ET AL. 11

When fe;i is less than fe;optimal
 67.5 Hz, the gradient of the fe –K Acknowledgments
curve (Fig. 3d) is positive, corresponding to a positive Δfe . On the Y. Zhou wishes to acknowledge support given to him from
other hand, when fe;i is larger than fe;optimal , the gradient is negative, the National Natural Science Foundation of China through grant
and so is Δfe . As shown in Fig. 15, in all the cases, fe  51421063. The support from the National Research Foundation
fe;i  Δfe
t  a sin
ωt could converge to fe;optimal , where the for the Doctoral Program of Higher Education of China under
gradient of the fe –K curve is zero, suggesting an excellent adaptation grant 20132302110054 is acknowledged.
of both classical and modified schemes to the initial excitation
frequency.
The convergence time of 50 s is too long for some applications with References
high-frequency perturbations. However, it is expected that the
modified ES method can be used for some flow separation controls [1] Zaman, K.-B., and Hussain, A.-K., “Vortex Pairing in a
Circular Jet Under Controlled Excitation. Part 1. General Jet
[29,30], jet mixing enhancement [16], and bluff-body drag reduction Response,” Journal of Fluid Mechanics, Vol. 101, No. 3, 1980,
[31]. Note that, the convergence time depends highly on the initial pp. 449–491.
state of the variable (i.e., fe;i ). The closer the initial state of the doi:10.1017/S0022112080001760
variable is to the stable state, the faster the system will be converged. [2] Ho, C.-M., and Gutmark, E., “Vortex Induction and Mass Entrainment
As indicated in Fig. 15, the convergence time is about 30 s for in a Small-Aspect-Ratio Elliptic Jet,” Journal of Fluid Mechanics,
fe;i  60 Hz. Also, if the variation in ReD is not as sharply reduced Vol. 179, June 1987, pp. 383–405.
as in the present test, the convergence time could be reduced greatly. doi:10.1017/S0022112087001587
For instance, Figure 16 shows the time responses of the two control [3] Zaman, K.-B., Reeder, M.-F., and Samimy, M., “Control of an
Downloaded by UNIVERSITY OF CAMBRIDGE on July 17, 2017 | http://arc.aiaa.org | DOI: 10.2514/1.J055644

Axisymmetric Jet Using Vortex Generators,” Physics of Fluids, Vol. 6,

schemes when ReD was suddenly increased from 8000 to about 9500 No. 2, 1994, pp. 778–793.
(Figs. 16a and 16b), and the convergence time is only 14 s for both doi:10.1063/1.868316
methods (Figs. 16c and 16d). [4] Smith, B.-L., and Glezer, A., “Jet Vectoring Using Synthetic Jets,”
The current closed-loop control scheme can optimize only one Journal of Fluid Mechanics, Vol. 458, May 2002, pp. 1–34.
variable (i.e., fe ), not multivariables such as Cm , fe , etc., doi:10.1017/S0022112001007406
simultaneously, to achieve the maximum K. In the future, we are [5] Samimy, M., Kim, H.-J., Kastner, J., Adamovich, I., and Utkin, Y.,
going to investigate other control methods, such as machine learning “Active Control of High-Speed and High-Reynolds-Number Jets Using
control [32], which could control multiple variables for achieving the Plasma Actuators,” Journal of Fluid Mechanics, Vol. 578, May 2007,
pp. 305–330.
best jet mixing performance. This method has been successfully doi:10.1017/S0022112007004867
applied for the control of various flows, both numerically and [6] Bradbury, L.-J., and Khadem, A.-H., “The Distortion of a Jet
experimentally. by Tabs,” Journal of Fluid Mechanics, Vol. 70, No. 4, 1975, pp. 801–813.
doi:10.1017/S0022112075002352
[7] Wiltse, J.-M., and Glezer, A., “Manipulation of Free Shear Flows Using
Piezoelectric Actuators,” Journal of Fluid Mechanics, Vol. 249,
IV. Conclusions April 1993, pp. 261–285.
The control of a turbulent round jet has been experimentally doi:10.1017/S002211209300117X
investigated based on a single pulsed radial minijet. A modified [8] Davis, M.-R., “Variable Control of Jet Decay,” AIAA Journal, Vol. 20,
No. 5, 1982, pp. 606–609.
Extremum seeking (ES) scheme is applied for the first time to achieve
doi:10.2514/3.7934
the optimal excitation frequency fe;optimal and hence to maximize jet [9] Zhou, Y., Du, C., Mi, J., and Wang, X.-W., “Turbulent Round Jet Flow
mixing automatically, and the control results are compared with those Control Using Two Steady Minijets,” AIAA Journal, Vol. 50, No. 3,
of the classical ES scheme. Both deploy excitation frequency fe and 2012, pp. 736–740.
jet centerline mean velocity decay rate K as the control and the doi:10.2514/1.J050838
feedback signals, respectively. The following conclusions may be [10] Seidel, J.-F., Pappart, C., New, T.-H., and Tsai, H.-M., “Effects of
drawn from the investigation. Multiple Radial Blowing Around a Circular Jet,” 43rd AIAA
1) Under the actuation of a single pulsed radial minijet, the open-loop Aerospace Sciences Meeting and Exhibit, AIAA Paper 2005-866,
Jan. 2005.
control is highly effective in enhancing the jet mixing performance at
doi:10.2514/6.2005-866
ReD  8000, and the mass flow rate m _  3.2 g∕ min
Cm  2.3%, [11] Yang, H., Zhou, Y., Zhu, Y., and Liu, Y., “Flapping Motion of
producing an increase in K by a factor of 8 at fe;optimal  0.5f0 , a Turbulent Jet Under the Asymmetric Excitation of Two Unsteady
compared with the uncontrolled jet, where f0 is the preferred- Minijets,” Fluid-Structure-Sound Interactions and Control, Springer,
mode vortex frequency. Accordingly, there is a drastic change in Berlin, 2016, pp. 259–264.
the flow structure. At a higher ReD
 16;000, the maximum K doi:10.1007/978-3-662-48868-3_42
occurs again at fe ∕f0  0.5, even though f0 rises to 250 Hz [12] Yang, H., Zhou, Y., So, R. M., and Liu, Y., “Turbulent Jet
(cf. 135 Hz at ReD  8000). Nevertheless, the maximum K drops Manipulation Using Two Unsteady Azimuthally Separated Radial
considerably compared to its counterpart at ReD  8000, which is Minijets,” Proceedings of the Royal Society A Mathematical
at least partly due to the fixed m _ or Cm, which is not optimized Physical and Engineering Sciences, Vol. 472, No. 2191, 2016,
Paper 20160417.
for ReD  16;000. doi:10.1098/rspa.2016.0417
2) The modified model-free ES control system is developed. Given [13] Yang, H., and Zhou, Y., “Axisymmetric Jet Manipulated Using Two
a fixed m_ or Cm, the controller could search fe;optimal automatically. Unsteady Minijets,” Journal of Fluid Mechanics, Vol. 808, Dec. 2016,
Compared with the classical scheme, the modified scheme manages pp. 362–396.
to reduce the steady-state error in fe;optimal by 70% and the doi:10.1017/jfm.2016.634
convergence time by 15.3%; meanwhile, it increases the mean K [14] Lee, C., Kim, J., and Choi, H., “Suboptimal Control of Turbulent
value by 12%. The parameters of the two closed-loop control scheme Channel Flow for Drag Reduction,” Journal of Fluid Mechanics,
are all experimentally determined, corresponding to the best control Vol. 358, March 1998, pp. 245–258.
performance. doi:10.1017/S002211209700815X
3) The modified scheme is also found to be robust and [15] Zhang, M.-M., Cheng, L., and Zhou, Y., “Closed-Loop-Controlled
adaptable when ReD and initial excitation frequency fe;i are Vortex Shedding and Vibration of a Flexibly Supported Square Cylinder
Under Different Schemes,” Physics of Fluids, Vol. 16, No. 5, 2004,
separately changed. More importantly, compared with the classical pp. 1439–1448.
scheme, the modified scheme may reduce the convergence time by doi:10.1063/1.1687413
37.5% in the robust experiment because the amplitude a of the [16] Wu, Z., Zhou, Y., Cao, H.-L., and Li, W.-J., “Closed-Loop Enhancement
sinusoidal perturbation signal can be automatically adjusted as of Jet Mixing with Extremum-Seeking and Physics-Based Strategies,”
ReD is changed, thus increasing the dynamic response of the Experiments in Fluids, Vol. 57, No. 6, 2016, pp. 1–14.
controller. doi:10.1007/s00348-016-2194-9
 12 Article in Advance / FAN ET AL.

[17] Bushnell, D.-M., and McGinley, C.-B., “Turbulence Control in Wall 2014, pp. 8687–8692.
Flows,” Annual Review of Fluid Mechanics, Vol. 21, 1989, pp. 1–20. doi:10.1109/ChiCC.2014.6896460
doi:10.1146/annurev.fl.21.010189.000245 [26] Gutmark, E., and Ho, C.-M., “Preferred Modes and the Spreading
[18] Kasagi, N., “Toward Smart Control of Turbulent Jet Mixing and Rates of Jets,” Physics of Fluids, Vol. 26, No. 10, 1983,
Combustion,” JSME International Journal Series B Fluids & Thermal pp. 2932–2938.
Engineering, Vol. 49, No. 4, 2006, pp. 941–950. doi:10.1063/1.864058
doi:10.1299/jsmeb.49.941 [27] Mi, J., Xu, M., and Zhou, T., “Reynolds Number Influence on Statistical
[19] Brunton, S.-L., and Noack, B.-R., “Closed-Loop Turbulence Control: Behaviors of Turbulence in a Circular Free Jet,” Physics of Fluids,
Progress and Challenges,” Applied Mechanics Reviews, Vol. 67, No. 5, Vol. 25, No. 7, 2013, Paper 075101.
2015, Paper 050801. doi:10.1063/1.4811403
doi:10.1115/1.4031175 [28] Maury, R., Koenig, M., Cattafesta, L., and Delville, J., “Extremum-
[20] Ariyur, K.-B., and Krstić, M., Real-Time Optimization by Extremum- Seeking Control of Jet Noise,” International Journal of Aeroacoustics,
Seeking Control, Wiley, Hoboken, NJ, 2004, pp. 3–19. Vol. 11, No. 3, 2012, pp. 459–473.
doi:10.1002/0471669784 doi:10.1260/1475-472X.11.3-4.459
[21] Benard, N., Moreau, E., Griffin, J., and Lii, L.-N., “Slope Seeking for [29] Wu, Z., Wong, C.-W., Wang, L., Lu, Z., Zhu, Y., and Zhou, Y.,
Autonomous Lift Improvement by Plasma Surface Discharge,” “A Rapidly Settled Closed-Loop Control for Airfoil Aerodynamics
Experiments in Fluids, Vol. 48, No. 5, 2010, pp. 791–808. Based on Plasma Actuation,” Experiments in Fluids, Vol. 56, No. 8,
doi:10.1007/s00348-009-0767-6 2015, pp. 1–15.
[22] Wang, H.-H., Yeung, S., and Krstic, M., “Experimental doi:10.1007/s00348-015-2032-5
Application of Extremum Seeking on an Axial-Flow Compressor,” [30] Chabert, T., Dandois, J., and Garnier, É., “Experimental Closed-Loop
IEEE Transactions on Control Systems Technology, Vol. 8, No. 2, Control of Separated-Flow over a Plain Flap Using Extremum Seeking,”
Downloaded by UNIVERSITY OF CAMBRIDGE on July 17, 2017 | http://arc.aiaa.org | DOI: 10.2514/1.J055644

2000, pp. 300–309. Experiments in Fluids, Vol. 57, No. 3, 2016, pp. 1–17.
doi:10.1109/87.826801 doi:10.1007/s00348-016-2123-y
[23] Creaby, J., Li, Y., and Seem, J.-E., “Maximizing Wind Turbine Energy [31] Brackston, R.-D., Wynn, A., and Morrison, J.-F., “Extremum
Capture Using Multivariable Extremum Seeking Control,” Wind Seeking to Control the Amplitude and Frequency of a Pulsed Jet for
Engineering, Vol. 33, No, 4, 2009, pp. 361–387. Bluff Body Drag Reduction,” Experiments in Fluids, Vol. 57,
doi:10.1260/030952409789685753 No. 10, 2016, p. 159.
[24] Becker, R., King, R., Petz, R., and Nitsche, W., “Adaptive Closed-Loop doi:10.1007/s00348-016-2243-4
Separation Control on a High-Lift Configuration Using Extremum [32] Duriez, T., Brunton, S.-L., and Noack, B.-R., Machine Learning
Seeking,” AIAA Journal, Vol. 45, No. 6, 2007, pp. 1382–1392. Control — Taming Nonlinear Dynamics and Turbulence, Springer,
doi:10.2514/1.24941 Switzerland, 2016, pp. 123–151.
[25] Wang, L., Chen, S., and Zhao, H., “A Novel Fast Extremum Seeking doi:10.1007/978-3-319-40624-4
Scheme Without Steady-State Oscillation,” Proceedings of the IEEE
33rd Chinese Control Conference (CCC), IEEE Publ., Piscataway, NJ,
A. Naguib
Associate Editor