a robust modification of a predictive adaptive-optic ...sgordeye/papers/aiaa-2016-3529.pdfamerican...
TRANSCRIPT
American Institute of Aeronautics and Astronautics
1
A Robust Modification of a Predictive Adaptive-Optic
Control Method for Aero-Optics
Robert Burns1, Eric Jumper
2, Stanislav Gordeyev
3
University of Notre Dame, Notre Dame, IN, 46556
Abstract
A modification of a previous predictive adaptive-optic controller is presented in this
paper. Conventional adaptive-optic controllers suffer from bandwidth limitations caused by
latency in their control loops. This latency severely limits their capabilities in aero-optic
applications that cannot be overcome with conventional feedback techniques. Our method
uses prior knowledge of flow behavior to predict future behavior, and thus overcome
deadtime. We have modified our previous neural network controller to use a linearized
predictor, which we demonstrate to be more accurate, more robust to noise and flow
disturbances, and less computationally expensive. Our previous neural network method
showed disturbance rejection in the range of 35-55% in simulation over our test conditions
in the most optically-active regions, while the improved method shows disturbance rejection
between 45-75% over the same range. Additionally, we demonstrate that the predictive
control method is stable, even in the presence of latency uncertainty.
Nomenclature
α = viewing angle (rad)
β = modified elevation angle (rad) or integrator gain (-)
ε = vector of wavefront prediction residuals
λ = laser wavelength (m)
ρ = air density (kg/m3)
Σ = diagonal matrix of singular values
Φ = matrix of POD modes (-)
φ = eigenmode vector (-)
A = linear prediction matrix
Az = azimuth (rad)
a = modal coefficients
C = compensator transfer function (-)
D = aperture diameter (m)
d = single-aperture aero-optic disturbance (m)
E = root-mean-square residual error
El = elevation (rad)
f = disturbance frequency (Hz)
1 Graduate Student, Department of Aerospace and Mechanical Engineering, Hessert Laboratory for Aerospace
Research, Notre Dame, IN 46556, AIAA Student Member.
2 Professor, Department of Aerospace and Mechanical Engineering, Fitzpatrick Hall of Engineering, Notre
Dame, IN 46556, AIAA Fellow.
3 Research Associate Professor, Department of Aerospace and Mechanical Engineering, Hessert Laboratory for
Aerospace Research, Notre Dame, IN 46556, AIAA Associate Fellow.
Dow
nloa
ded
by S
tani
slav
Gor
deye
v on
Jun
e 20
, 201
6 | h
ttp://
arc.
aiaa
.org
| D
OI:
10.
2514
/6.2
016-
3529
47th AIAA Plasmadynamics and Lasers Conference
13-17 June 2016, Washington, D.C.
AIAA 2016-3529
Copyright © 2016 by William Robert Burns, Eric Jumper, Stanislav Gordeyev. Published by the American Institute of Aeronautics and Astronautics, Inc., with permission.
AIAA Aviation
American Institute of Aeronautics and Astronautics
2
f = vector-valued nonlinear prediction function
g = function describing flow evolution
h = observer function of special flow properties
I = on-target beam intensity (W/m2)
I0 = unaberrated beam intensity (W/m2)
KGD = Gladstone Dale constant (m3/kg)
M = Mach number (-) or embedding dimension (-)
Nd = number of timesteps of delay in the feedback loop (-)
n = index of refraction (-)
OPD = Optical Path Difference (m)
OPL = Optical Path Length (m)
S = Strehl ratio (-)
St = Strouhal number (-)
U = coefficient matrix
V = freestream velocity (m/s)
v = wavefront vector (m)
W = wavefront vector (m)
x = vector of coefficients (-)
y = wavefront residual output (m)
z = vector of flow properties
I. Introduction
ne significant challenge to the implementation of airborne laser systems is the aero-optics problem1. Near-field
turbulence over a turret aperture causes density fluctuations that are “imprinted” on the laser wavefronts
through the relationship between index of refraction and density. These effects can greatly reduce a beam
control system’s performance in the far-field, especially when pointing through the wake region of a turret. Turret
configurations are desirable in beam control systems for their large fields of regard; however, this comes at the
expense of significant aero-optic aberrations, especially in the wake region. A schematic of the flow structures that
impact aero-optics is shown2 in Figure 1. Pressure and temperature fluctuations contribute to significant density
fluctuations as the boundary layer thickens downstream. Flow separation over the top of the turret results in large
separation regions that roll up into “horn” and secondary vortices. So-called “necklace” vortices form at the base of
the turrets and can be encountered when pointing the laser near the surface of the aircraft. Shocks can form at the top
of the turret at Mach numbers as low as 0.55. All of these flow features cause density fluctuations that adversely
impact optical performance.
The aero-optics problem in the wake region of a turret
(without compensation) is depicted schematically in
Figure 2. The beam’s unaberrated, planar wavefronts
encounter turbulence near the turret’s aperture. The
turbulence-related density variations are effectively
“imprinted” on the wavefront. As the beam propagates to
the far field, these density fluctuations will tend to scatter
the beam. The strength of these aberrations scales with the
squared Mach number of the emitting aircraft, the
freestream density, and the diameter of the turret.
The index of refraction is related to the density of the
medium through the Gladstone-Dale relationship3,
( , , ) 1 ( ) ( , , )GDn x y z K x y z , where n is the index of
refraction, λ is the wavelength, and KGD is a constant. This
relationship can be used to compute the exact aberration
along a path by first computing the optical path length
O
Figure 1: Optically-aberrating flow over a
turret1.
Dow
nloa
ded
by S
tani
slav
Gor
deye
v on
Jun
e 20
, 201
6 | h
ttp://
arc.
aiaa
.org
| D
OI:
10.
2514
/6.2
016-
3529
American Institute of Aeronautics and Astronautics
3
from some point in the aperture,1
0( , , ) ( , , , )
z
OPL x y t n x y z t dz , and then subtracting the aperture-averaged mean to
compute the optical path difference,
( , , ) ( , , ) ( , , )OPD x y t OPL x y t OPL x y t [1]
Figure 2: Schematic of the aero-optics problem for a beam control system1.
The RMS optical path difference (OPD) is a convenient metric for estimating the performance of aero-optic
systems since it enables the estimation of the ratio of the Strehl ratio: aberrated on-axis beam intensity to the
diffraction-limited on-axis beam intensity. Strehl ratios can be estimated for each aberration source in a beam
control system: for example, near-field aero-optics, far-field atmospherics, aero-mechanical jitter, optical heating,
etc. These Strehl ratios can be multiplied together to obtain the total loss of intensity at the target. In this paper, we
focus on the aero-optic aberrations only. For instantaneous aero-optic aberrations following a Gaussian spatial
distribution, the Strehl ratio can be estimated through the Marèchal large aperture approximation4,5,6
, shown in Eq. 2.
22 ( )
0
( )( )
( )
rmsOPD tI t
S t eI t
. [2]
It is desirable to maximize the Strehl ratio by mitigating aero-optical aberrations. It is conventional to quantify
reductions in aero-optical aberration using the following definition,
,
10
,
10logrms corrected
rms uncorrected
OPDdB
OPD . [3]
Since laser wavelength is application-dependent, it is useful to consider the impact of wavefront aberration
mitigation in non-dimensional terms by rearranging the equation in terms of a corrected and uncorrected Strehl ratio,
as discussed in our previous work7. A modest -3dB improvement in aperture-averaged disturbance rejection can
improve Strehl ratio from 0.2 to 0.67, for example.
A large body of work has been performed to characterize aero-optical effects on the Airborne Aero-Optics
Laboratory8 (AAOL), which is shown in Figure 3, and its transonic successor AAOL-T
9. AAOL consists of a
pointing and a tracking aircraft. The pointing aircraft emits a diverging 532 nm beam toward the tracking aircraft’s
turret, so that the beam width at the exit aperture is small relative to the flow structures. This means that the optical
aberrations at the exit aperture are primarily tip/tilt and do not significantly impact the wavefronts. Aero-optical
aberrations near the aperture of the tracking turret are imprinted on the laser wavefronts and measured by a Shack-
Hartmann wavefront sensor in the tracking aircraft.
Dow
nloa
ded
by S
tani
slav
Gor
deye
v on
Jun
e 20
, 201
6 | h
ttp://
arc.
aiaa
.org
| D
OI:
10.
2514
/6.2
016-
3529
American Institute of Aeronautics and Astronautics
4
Figure 3: Picture of the Airborne Aero-Optics Laboratory (AAOL). The pointing aircraft is depicted on
the left and the tracking aircraft on the right.8
Normalized aero-optical aberrations tend to collapse onto a curve that is a function of two quantities: a lookback
angle, α, and a modified elevation angle, β. These two quantities can be expressed8 in terms of azimuth and
elevation as shown in Figure 4,
1cos cos( )cos( )Az El , and [4]
1 tan( )tan
sin( )
El
Az
. [5]
It has been demonstrated8 that aero-optic aberrations can be normalized using the following equation,
2
0/
rmsnorm
OPDOPD
M D , [6]
where M is the freestream Mach number, ρ/ρ0 is the ratio
of freestream density to sea-level density, and D is the
turret diameter. Figure 5 shows OPDnorm as a function of
viewing angle and modified elevation angle from AAOL.
Notably, the optical aberrations are highest in the
separated flow regime, corresponding to viewing angles
of greater than 110 degrees. Laser-based systems are
often tolerant of small aberrations encountered at
forward-looking angles, but not at aft-looking viewing
angles.
Aero-optical aberrations must be mitigated using
some technique for many applications involving either
communications or directed energy. There are two
primary ways of achieving this: flow control and adaptive
optics. Flow control devices are generally very
application-dependent. Active flow control can be used to mitigate aero-optic distortions either by suppressing
aberrations or relocating them away from an optical aperture. Some of these methods are open-loop (operating on
pre-determined, feedback-free control), while others are closed-loop (sensing and reacting to disturbances to
mitigate them).
Active flow control has been employed in transonic flows to manipulate shock formation over a cylinder10
using
pulsed jets in open-loop configurations. This type of manipulation could be used to relocate optically-active flow
features away from an optical aperture.
High-frequency synthetic jets11
have been used to reduce the optical aberrations present in the wake of a
hemisphere. Later, this group12
used a hybrid of passive and active flow control: by using a forward-protruding plate
to decouple the turret wake from the necklace vortices, the effectiveness of the oscillating jets upstream of the
optical aperture to control the turret wake was enhanced. Vukasinovic et al10
have also tested control jets in the
transonic regime to control separation both upstream and downstream of the shock to reduce the sharp velocity and
density gradients present in the shear layer. This type of flow control was studied in more detail13
using Particle
Image Velocimetry (PIV) and Schlieren imagery.
Figure 4: Definition of α and β in terms of
azimuth and elevation7
Dow
nloa
ded
by S
tani
slav
Gor
deye
v on
Jun
e 20
, 201
6 | h
ttp://
arc.
aiaa
.org
| D
OI:
10.
2514
/6.2
016-
3529
American Institute of Aeronautics and Astronautics
5
Figure 5: Collapse of OPDrms as a function of viewing angle from AAOL flight tests.8
A modification of proper orthogonal decomposition14
has been developed to model closed-loop flow control.
This method attempts to model both the controlled and natural states of the flow, as well as the transition between
these states. The research group demonstrated their temporal POD (TPOD) approach using dielectric discharge
barrier (DBD) plasma actuators to control the first TPOD mode of flow over a circular cylinder. In practice it would
be necessary to control more than the first mode, but nonetheless there are several ideas from this work that could be
incorporated into a model for the physical behavior of various turbulent flows.
An explosion of research in the field of machine learning has led to an effort in fluids research15
to improve
flow control using these new methods. One method16
uses optical sensors for feedback, slotted jets for control
actuators, and a genetic algorithm to implement a control law to delay flow separation. Genetic algorithms are used
in control systems to search for control actions that minimize a cost function for a given set of observations. The
results of their experiment were an 80% reduction in the recirculation zone behind a backward-facing step. Genetic
algorithms are often referred to as model-free control. A similar approach based on genetic programming17
was used
to control a mixing layer flow and compared to open-loop forcing. In addition to achieving better performance than
the open-loop forcing, it is also adaptable and automatically tunable.
Another means of aero-optic mitigation is to use adaptive optics. These systems imprint the conjugate of the
currently-present aero-optic aberration onto a deformable mirror, and then send the pre-corrected laser beam out
from the turret. When the corrected beam then encounters the aero-optic aberration, it emerges from the flow as
nearly planar. These techniques are currently employed on ground-based telescopes to correct for atmospheric
optical effects. However, these aberrations occur on the order of 1-10 Hz and thus the computational and system
requirements are not extremely demanding for currently-available technology. However, sub-sonic aero-optical
phenomena consisting of turbulent boundary layers, shear layers, attached vortices, etc. occur on the order of
kilohertz on full-scale turrets at realistic Mach numbers. In particular, the dominant frequency content of separated
flow on a turret in the wake region typically occurs near a Strouhal ratio of about 1:
fDSt
V , [7]
where f is the frequency of the disturbance in hertz, D is the turret diameter, and V is the freestream velocity. For the
Airborne Aero-Optics Laboratory, the disturbance frequencies observed in the wake region were concentrated near 1
kHz. Thus, a real-time adaptive optics system appropriate for AAOL must be able to perform corrections at a
minimum of 10kHz, assuming 10 discrete corrections per disturbance cycle. This is a demanding computational
requirement given that wavefronts must be represented by matrices that are at least 10x10, and sometimes as large as
30x30. A significant obstacle is cumulative latency in the control system. Time spent on digital calculations,
physical sensor response, analog-to-digital conversion of sensor signals, electronic amplifiers, and
electromechanical response of deformable mirrors will all contribute to a significant degree of delay18
in a high-
speed adaptive optic system.
Dow
nloa
ded
by S
tani
slav
Gor
deye
v on
Jun
e 20
, 201
6 | h
ttp://
arc.
aiaa
.org
| D
OI:
10.
2514
/6.2
016-
3529
American Institute of Aeronautics and Astronautics
6
This latency problem has previously been
addressed using various prediction methods such as
dynamic mode decomposition19
, receding-horizon
adaptive control20
, and simple “frozen-flow” advection
prediction21
. We recently developed a method7 for
overcoming this latency problem using a predictive
neural network controller in conjunction with a
conventional feedback loop as shown in Figure 6. The
neural network controller uses prior knowledge of
flow behavior to predict future behavior over a short
temporal horizon equal to the known delay in the
feedback loop. We simulated an adaptive-optic control
loop operating at 25kHz with varying amounts of
latency using aero-optic wavefront disturbance data
from the AAOL flight tests. In the most optically-
active fully separated flow regime, disturbance
rejection between 35% and 55% was achieved in our
simulations. As discussed previously, even 35%
reduction in RMS wavefront aberration can have a
large effect on a real system’s performance. The
presumption for this method was that the nonlinear
character of the flow would require a nonlinear
predictive element such as a neural network.
In this work, we develop a modification for the neural network method that uses a linearized prediction model.
We demonstrate that not only is this an acceptable assumption over the amounts of latency considered, but it is
actually scalable to a larger number of predictable modes and leads to better prediction accuracy. Additionally, we
investigate its favorable robustness to perturbed flow conditions and stability. For further details on dimensionality
reduction using POD and the impact of latency on controller bandwidth, refer to our previous work7.
II. Predictive Control System Architecture
A block diagram of the predictive controller is shown in Figure 7. The controller is denoted by C, the mirror
plant model by G, the aero-optic disturbance by d, the residual wavefront error y, the prediction model P, the lagged
wavefront sensor H1, the lagged mirror sensor by H2, and the unlagged mirror sensor by H3.
For consistency and to avoid excessive tedium, we assume that all of the latency in the control system occurs in
wavefront sensing, reconstruction, and control law computation and is accounted for in the model H3. The wavefront
predictor itself is sensitive only to cumulative latency and its predictions are independent of the individual sources of
latencies. Lagged and unlagged mirror sensors are required to reconstruct an estimate of the total wavefront
disturbance from the sensed residual wavefront and the deformable mirror position.
For the purposes of our analysis, we will assume that the conventional controller is a simple integrator and the
deformable mirror has unity transfer function. In the absence of latency this is a trivial control problem, but the
bandwidth limitations imposed by pure latency are severe enough to investigate in isolation of mirror mechanics,
amplifier electrical dynamics, etc. (see our previous paper for more details7).
H1 may be decomposed as 1 1
s T tH H e
, where ΔT is the cumulative latency in the feedback loop, δt is the
uncertainty in this latency, and 1H is the latency-free sensor. Practically, the lagged and unlagged deformable
mirror sensors may be decomposed as 2 3
s TH H e since H2 is simply a digitally-delayed copy of H3. An
important design goal is to match the latency between the lagged deformable mirror sensor and the wavefront
sensor. The latency in H2 is considered to be a tunable parameter matched to the nominal value of the latency in the
wavefront sensor, while the latency in the wavefront sensor, ΔT, is measured a priori with some uncertainty, δt. Due
to simple measurement uncertainty as well as digital jitter, this uncertainty will always be non-zero; however, the δt
= 0 case is considered first.
Figure 6: Conventional AO loop augmented
by wavefront predictor1.
Dow
nloa
ded
by S
tani
slav
Gor
deye
v on
Jun
e 20
, 201
6 | h
ttp://
arc.
aiaa
.org
| D
OI:
10.
2514
/6.2
016-
3529
American Institute of Aeronautics and Astronautics
7
Figure 7: Predictive controller block diagram.
Consider the single-input-single-output (SISO) version of this system. This system gives the following general
transfer function from the output to the reference signal,
1 2 31
C s G sys
r C s G s P s H s C s G s P s H s C s G s H s
. [8]
The rationale for matching the delays between H2 and H1 is now clear: in the absence of uncertainty, the
complex dynamics of the predictor vanish and the closed loop transfer function becomes simply,
31
C s G sys
r C s G s H s
, [9]
which can typically be stabilized by a P-I controller in an adaptive-optics control loop. Given the simplicity of our
idealized component models, the stability of this loop is trivial to assess in the absence of latency uncertainty.
Now, consider the multidimensional case. Define the transfer function matrix T(s),
1 1 1
1 2
2 2 2
1 2
1 2
N
N
N N N
N
y y ys s s
r r r
y y ys s s
r r rs
y y ys s s
r r r
T
, [10]
which in the multidimensional case can be solved for as follows if the latency uncertainty is zero,
1
3s s s s s s
T C G I C G H . [11]
As in the single-dimensional case, this system can easily be controlled (with varying performance) with
classical methods. Recall that adaptive optics system components are in general open-loop stable.
Similarly, the disturbance rejection function,
Dow
nloa
ded
by S
tani
slav
Gor
deye
v on
Jun
e 20
, 201
6 | h
ttp://
arc.
aiaa
.org
| D
OI:
10.
2514
/6.2
016-
3529
American Institute of Aeronautics and Astronautics
8
1 1 1
1 2
2 2 2
1 2
1 2
N
N
N N N
N
y y ys s s
d d d
y y ys s s
d d ds
y y ys s s
d d d
S, [12]
becomes (in the absence of delay uncertainty),
1
2 3 3s s s s s s s s s s s
S I C G P H C G H I C G H . [13]
Hence, the prediction filter must be open-loop stable to ensure closed-loop stability. The interpretation for this
necessity is essentially that the predictor acts as a feed-forward element.
We will now detail the algorithm of the predictor itself, and then re-assess the stability problem using this
framework.
III. Prediction Method and Estimation
We will first briefly summarize the key assumptions from our nonlinear neural network predictor7 and then
discuss the necessary modifications. We assume that we can approximate the evolution of an aero-optical wavefront
vk by some discrete function f that depends on prior wavefronts vk-1…vk-M+1, with some amount of truncation error
denoted by εk+1.
1 1 1 1, ,...,k k k k M k v f v v v ε [14]
A wavefront predictor should approximate f such that the mean-square of the residual error is minimized,
2
1
L
kkE
ε ,
where L is the total number of wavefront observations in a dataset. This is has the structure of a non-linear
autoregressive (NAR) problem.
For practical applications, it is desirable to reduce the order of the prediction model. Practical applications of
adaptive optics in aero-optics will have wavefront measurements on the order of 15x15 up to 30x30 subapertures,
ranging from ~200 subapertures to ~700 subapertures for a circular beam inscribed in a rectangular sensor array.
Direct wavefront prediction is computationally-intensive and prone to measurement noise. A simple linear
prediction matrix of the form vk+1=Avk alone would contain in the worst case between 105 and 10
6 elements, which
is impractical for real-time applications. This problem becomes substantially worse if more than a single prior
wavefront is needed for a prediction; which, in the general case is true due to the nonlinear nature of the flow.
A number of methods can be used for dimensionality reduction. A natural choice of basis functions for many
optical applications are the Zernike modes, since they describe physically meaningful optical aberrations such as
focus/defocus, coma, astigmatism, etc. and are orthogonal on the unit disk. However, these modes are physically
unrelated to the physics of the flow itself. A better modal decomposition for aero-optics is the Proper Orthogonal
Decomposition22
, which can be computed via a snapshot method that produces the most rapidly-converging set of
orthogonal modes possible.
These modes, also known as Karhunen-Loève (K-L) or Principle Component Analysis (PCA) modes, can be
calculated for multidimensional discrete-time data using the Singular Value Decomposition (SVD) or by solving the
eigenvalue problem on the data matrix’s autocorrelation matrix. In this case, let V be a matrix of measurement
snapshots of wavefronts organized by column vectors of samples ordered by increasing time as shown,
T
1 2 NV v v v . [15]
Dow
nloa
ded
by S
tani
slav
Gor
deye
v on
Jun
e 20
, 201
6 | h
ttp://
arc.
aiaa
.org
| D
OI:
10.
2514
/6.2
016-
3529
American Institute of Aeronautics and Astronautics
9
The data matrix, V, is decomposed using SVD,
HV UΣΦ . [16]
The spatial modes are extracted from the columns of Φ,
1 2 RΦ φ φ φ , [17]
where rank( )R V . The temporal coefficients are then calculated from a projection of the spatial modes onto the
original observations. Taking advantage of the fact that the pseudoinverse of an orthonormal matrix is its Hermitian
transpose, this can be written as
H x Φ V Φ V , [18]
where x is a matrix of temporal coefficients. This is an important fact in terms of computational efficiency since it
means that the projection of wavefronts onto POD modes is a simple matrix multiplication. The POD modes are
ranked by the importance of their contribution to the overall energy of the system. Quite often, the POD modes
converge quickly to give a good low-dimensional model. Additionally, in the case of naturally occurring fluid flow,
low order modes typically exhibit smooth behavior. If it is assumed that the POD modes do not change, then each
wavefront can be decomposed as follows,
1
( )N
k n n
n
x k
v Φ , [19]
where nΦ are the time-invariant POD modes, xn(k) are the coefficients at timestep k, and N is the desired truncation
dimension. Additionally, we assume that M snapshots are sufficient to approximate the next wavefront in the
sequence: we refer to M as the embedding dimension23
. In this case, the NAR problem becomes a function of the
modal coefficients, as shown,
1 1 1 1 1, ,...,N
k k k k M k v Φ g x x x ε , [20]
where 1
N LxNΦ R are the truncated set of POD modes and : NxM Ng R R is the nonlinear prediction function.
Thus, it is necessary to use some method to estimate the function g.
We recognize at this point that every continuous, differentiable nonlinear function can be linearized at any point
along a system’s trajectory. For the practical application of latency compensation, we assume that the flow dynamics
are approximately linear over the prediction horizon. This assumption allows us to write g as a matrix function,
1
1 1 1
1
, ,...,
k
k
k k k k M k
k M
x
xx g x x x A ε
x
, [21]
where A is referred to as the prediction matrix. The prediction matrix can be calculated by setting up the
overconstrained linear problem using a set of training input and output matrices X and Y,
1 2
2 3 1
1 2
1 1
...
L M
L M
M M L
M M L
x x x
x x xY x x x ε AX ε A ε
x x x
, [22]
Dow
nloa
ded
by S
tani
slav
Gor
deye
v on
Jun
e 20
, 201
6 | h
ttp://
arc.
aiaa
.org
| D
OI:
10.
2514
/6.2
016-
3529
American Institute of Aeronautics and Astronautics
10
and then solving for the A matrix by calculating the pseudoinverse of the right-hand data matrix or similar
technique. The resulting matrix will be optimal in the least-squares sense for the POD coefficient prediction; that is,
it is desired to solve the convex optimization problem,
1
1
1
arg min
kL M
k
k
k M
A
x
x A
x
. [23]
The pseudoinverse method to solve this problem is,
1
1H H H
Y AX AUΣV A Y UΣV YU Σ V , [24]
or equivalently using ordinary least squares,
1
H H
A YX XX. [25]
This method is known as Vector Auto-Regression24
(VAR), and combined with the POD model reduction we
refer to it as POD-VAR. It is noteworthy that this solution coincides with the maximum likelihood estimate if the
error terms are normally distributed. Additionally, it generates a stable predictor if the modeled process is stationary
with time-invariant covariance. Writing the prediction problem such the prediction matrix is square,
1
1 1
2 1 1
k k k
k k k
aug
k M k M k M
1 2 Mx x xA A A
x x xI 0 0A
x x x0 I 0
[26]
a stability analysis for this autonomous process simply requires that,
sup : 0 1z z augC I A . [27]
VAR estimators will always meet this condition for natural stationary processes, such as mean-removed, tip/tilt-
removed aero-optic disturbances resulting from an unchanging flow condition.
IV. Stability of the Closed-Loop System with Delay Uncertainty
The stability problem of Section 2 becomes more difficult to address if there is a mismatch between the
prediction horizon and the true latency in the system; that is, if 0t . This mismatch will always be present in a
real system – while digital systems can be very good at measuring this sort of latency and matching it closely, the
effect of digital electronics jitter and some small uncertainty cannot be ignored. The remainder of this section will
address the effect of this mismatch.
A general framework for investigating the effect of delay uncertainty is described in Chapter 11 of Michiels and
Niculescu25
. The key results of their analysis will be applied here. First, make the simplifying assumptions that
s G I and 1 2 1s T s ts s s e e H H H , where s H I . In simple terms, the assumptions are that the
deformable mirror responds instantly while the wavefront sensor responds with some pure time delay and some
uncertainty in that delay. The resulting transfer function becomes,
1
1s T s ts s s s s e e
T C I C C P [28]
Dow
nloa
ded
by S
tani
slav
Gor
deye
v on
Jun
e 20
, 201
6 | h
ttp://
arc.
aiaa
.org
| D
OI:
10.
2514
/6.2
016-
3529
American Institute of Aeronautics and Astronautics
11
The integrator compensator will have the form, ss
C I , where β is the feedback gain. The transfer function
reduces to,
1
1 1s T s tss s e e
T I P . [29]
Hence, the characteristic equation that must be studied for stability is,
det 1 1 0s T s tss e e
I P . [30]
Now, choose a proper matrix fractional description 1
P Ps s s
P N D such that NP(s) and DP(s) are co-
prime polynomial matrices and DP(s) is Hurwitz (allowing roots on the imaginary axis). This fractional matrix
description will exist if the system is both controllable and observable26
. The stability of this characteristic equation
can be treated using the theory of delay differential equations25
. If the above characteristic equation is written in the
following form,
det / 1 0s T s t
P Ps s s e e D N , [31]
then its stability can be analyzed by studying the neutral equation25
,
1 1
0 0 0 0( ) 0t t T t t T x N D x N D x , [32]
where N0 and D0 are the column-reduced coefficient matrices of P sN and /P s s D , respectively. The
details of this calculation are not important in this case for reasons that will be shortly addressed, but refer to
Michiels and Niculescu25
for additional information.
This neutral equation is strongly stable if25
,
1 21 1
0 0 0 0sup : 0,2 , 1,2 1i i
ir e e i
N D N D .
The stability condition with respect to infinitesimal delay uncertainty may be equivalently stated as,
1
0 0
1
2r
N D . [33]
However, because 1
/P Ps s s
N D is a strictly proper transfer function, the spectral radius above is 0 (as
discussed in Michiels and Niculescu25
) and practical stability is preserved for infinitesimal delay mismatches.
In summary, the delay compensation system outlined in this section is practically stable for infinitesimal delay
mismatch if:
1. P(s) is proper.
2. P(s) is open-loop stable.
3. P(s) is observable and controllable.
These conditions will be validated later in this section.
While Eq. 26 is convenient for practical estimation, it is helpful to put the predictor into state-space form for
application and further analysis. One such state space model is,
1k k k
k k k
P P
P P
x A x B u
y C x D u. [34]
Dow
nloa
ded
by S
tani
slav
Gor
deye
v on
Jun
e 20
, 201
6 | h
ttp://
arc.
aiaa
.org
| D
OI:
10.
2514
/6.2
016-
3529
American Institute of Aeronautics and Astronautics
12
In this model, the input vector uk is the most recent measurement of POD coefficients, the output vector yk is the
POD coefficient prediction at timestep k, and xk are a set of “stored” POD coefficient vectors that are used in the
prediction. Decompose the prediction matrix, N NMA R as, 1 MA A A where
N N
k
A R . The state
space matrices themselves have dimension ( 1) ( 1)N M N M PA R ,
( 1)N M N PB R , ( 1)N N M PC R , and
N NPD R and are chosen to be,
N N N N N N
N N N N N N
N N N N N N N N
N N N N
N N N N N N N N
P
0 0 0
I 0 0
A 0 I 0 0
0 0
0 0 I 0
, [35]
N N
N N
N N
P
I
0B
0
, [36]
2 3 MPC A A A , [37]
and,
1PD A . [38]
This form of the predictor allows a convenient description of the transfer function matrix in discrete form, P(z),
as follows:
1
( )z z
P P P PP C I A B D . [39]
As discussed previously in this section, this system must be controllable and observable for stability in the
presence of delay uncertainty. Otherwise, internally-cancelling modes will exist in P(z) that appear in both the poles
and zeros, and hence the resulting fractional decomposition P(z) = N(z)D(z)-1
will not be coprime. Controllability is
tested with the following condition26
,
( 1) 1rank[ ] ( 1)N M N M P P P P PA A B A B , [40]
and observability is tested with,
1 1
rank 1
N M
N M
P
P P
P P
C
C A
C A
. [41]
For this system,
( 1)
kN N
k
N N
M k N N
P P
0
A B I
0
for k > 2, [42]
Dow
nloa
ded
by S
tani
slav
Gor
deye
v on
Jun
e 20
, 201
6 | h
ttp://
arc.
aiaa
.org
| D
OI:
10.
2514
/6.2
016-
3529
American Institute of Aeronautics and Astronautics
13
so controllability is guaranteed analytically through the structure of the system matrices (in particular, 2rank[ ] ( 1)M N M P P P P PA A B A B ). The observability matrix will also be full-rank via a similar
argument if CP is full-rank. Since CP is estimated from a natural process, Cp will in general be full-rank.
Now it is desired to assess the stability of this discrete predictor as it interacts with a real adaptive optics
system. The approach taken here will use conformal mapping to transform discrete space to continuous space in
such a manner that conventional s-plane analysis tools can be applied. A better (but far more complex) approach
would be to use the theory of sampled data systems to accurately model the interaction of the digital predictor with
an analog system. Future research should investigate this.
The exact mapping between the z-plane and the s-plane is Tsz e . The difficulty with this mapping is that
conventional analysis tools cannot be used with non-polynomial expressions. There are several useful
approximations to this mapping that could be chosen to map between the s-plane and the z-plane. One
approximation that is closely related to this is the Padé approximation27
,
2
1 0
2
1 0
1
1
nTs n
n
n
k s k s k se
k s k s k s
, [43]
which is used to approximate time delay in continuous systems. This approximation is derived through a Taylor
series expansion of the exponential.
A similar approximation is the Tustin transformation (or bilinear transformation), defined as28
,
1 / 2
1 / 2
Ts Tsz e
Ts
. [44]
An important characteristic of this mapping is that the stability boundary (the imaginary axis in the s-plane and
the unit circle in the z-plane) is mapped exactly.
Minimum-phase systems in the s-plane are also mapped to minimum phase systems in the z-plane, however the
converse is not guaranteed. In fact, time delay systems are in general non-minimum phase. Minimum phase is not a
requirement for this analysis, but it is useful to understand the limitations of this approximation.
Finally, some frequencies will be significantly distorted as a consequence of this transformation – particularly
those near the Nyquist frequency, 2/T. The dominant disturbance frequencies addressed in this research are between
800 and 2000 Hz, while the control loop samping frequency is 25 kHz. Since the primary disturbances occur at an
order of magnitude lower than the sampling frequency, high-frequency distortion will amplify noise but should not
significantly impact overall predictor response.
Using the bilinear transformation, P(z) can be approximated in continuous space as P(s). Because the stability
boundary is preserved through the transformation and P(z) is stable, P(s) is also stable (but not minimum phase).
Because P(s) is controllable and observable, there exists a proper and stable matrix fractional description P(s) =
NP(s)DP(s)-1
such that NP(s) and DP(s) are co-prime and DP(s) is Hurwitz. Therefore, the stability conditions outlined
in this section are valid for this predictor and the system is stable in both the nominal case and in the presence of
infinitesimal delay uncertainty.
V. Simulation Results
Our simulation software is written in C++ using LAPACK for linear algebra routines. The wavefront
disturbance data was taken from Notre Dame’s Airborne Aero-Optics Laboratory (AAOL) flight test program. Each
test condition consists of 15,000 total wavefront snapshots. The first 3,000 wavefront snapshots are used to calculate
POD modes and train the coefficient prediction matrix. This value was selected as this is sufficient many snapshots
for well-converged POD modes. Then, the next 12,000 wavefronts are used to evaluate controller performance
independently. Thus, a separate predictive controller is trained for each flow condition.
Dow
nloa
ded
by S
tani
slav
Gor
deye
v on
Jun
e 20
, 201
6 | h
ttp://
arc.
aiaa
.org
| D
OI:
10.
2514
/6.2
016-
3529
American Institute of Aeronautics and Astronautics
14
An important selection during the training phase of
the controller is to determine a priori the optimal
number of POD modes to include in the controller. The
lowest-frequency (and thus, more predictable) content in
the disturbances tends to occur in the low-order POD
modes. As the prediction window increases, the higher
frequency content becomes harder to predict and thus
diminishing returns are observed. An optimization study
was performed to determine the optimal number of POD
modes to include for varying amounts of latency. The
results are shown in Figure 8. There is not much
practical benefit to including more than about 20 modes
for if larger amounts of latency are anticipated in the
controller. For small amounts of latency, improvements
are observed up to 64 modes. Since there is no
significant performance penalty for using a larger
number of modes, we will select N=16 (to match our
prior work) and N=64. A practical application may
require fewer modes to reduce computational cost.
We may now expand the results of the previous optimization study to further investigate how the controller
behaves. The full aperture-averaged disturbance rejection results for N=16, N=64 and M=4 are shown in Figure 9.
Similar to the results obtained in our previous work7 (labeled as “POD-ANN”), we observe better disturbance
rejection in the highly-separated flow regime around a viewing angle of 120 deg. However, we now observe
disturbance rejection near -6.5 dB for one timestep of latency with a viewing angle beyond 120 deg., whereas our
previous controller achieved roughly -4dB. For a large amount of latency, we observe approximately -3dB of
disturbance rejection while the neural network controller achieved roughly -2dB.
Figure 9: Performance comparison of the linear controller with the prior neural network controller for
several amounts of latency.
Figure 8: Ensembled mean disturbance
rejection of all viewing angles for latencies
between 1 to 5 timesteps (ascending order).
0 10 20 30 40 50 60 70-4.5
-4
-3.5
-3
-2.5
-2
-1.5
-1
Num Modes (N)
Reje
ction (
dB
)
Ensembled mean rejection for all viewing angles
Dow
nloa
ded
by S
tani
slav
Gor
deye
v on
Jun
e 20
, 201
6 | h
ttp://
arc.
aiaa
.org
| D
OI:
10.
2514
/6.2
016-
3529
American Institute of Aeronautics and Astronautics
15
That a linear matrix predictor performs better than a nonlinear predictor in this highly-separated flow regime is
consistent with expectations, since a separated shear flow is dominated by linear dynamics. We may also investigate
the structure of the prediction matrices as shown in Figure 10 to gain a better understanding of the dynamic
dependencies between POD modes. Naturally, the strongest dependence of a prediction (timestep k+1) is on the
immediately preceding coefficient (timestep k) as indicated by the diagonals. The next-strongest dependence is on
the diagonals of the previous coefficient (timestep k-1). The reader will observe that this is approximately a discrete
solution of the continuous differential equation, 0ax x . Higher-order derivatives (up to the 3rd
-order) are also
automatically approximated in the matrix as evidenced by the diagonal terms. The influence matrices also indicate
that strong dynamic coupling is present in the fluid flow. This is an important feature of the predictor. Capturing the
coupling between each mode improves prediction quality. This is part of the reason that simple single-input-single-
output conventional PID controllers do not perform well in aero-optic applications. An additional observation is that
the prediction matrix is primarily lower-triangular, meaning that higher-order modes tend to depend on lower-order
modes. A possible modification to the predictive controller would be to force the prediction matrix to be purely
lower-triangular, which would cut the computational cost of the prediction roughly by half.
Figure 10: Influence matrices at various viewing angles.
VI. Robustness and Sensitivity
We may investigate the robustness of the controller by training the predictor on a data set and establishing
baseline performance as per the previous section, and then testing the controller once again using a data set from a
perturbed flight condition.
In order to make this possible, we make the assumption that the flow direction is consistent with respect to the
plane of the optical aperture. We use the spectral four-beam Malley probe technique29
to estimate the mean flow
direction and then rotate the wavefronts with bilinear interpolation such that the flow direction is in a consistent
direction. This step is important because low-order POD modes (i.e., the most energetic and predictable modes) tend
to convey information about the convective nature of coherent structures while higher-order modes contain
information about their evolution. If the flow angle changes substantially, the modes would attempt to predict an
incorrect direction for the convective portions of the disturbances. While the four-beam Malley technique is
computationally expensive, the aperture-plane flow angle could also be estimated geometrically from azimuth and
elevation.
For initial robustness test, we trained the predictive controller on 3,000 wavefront snapshots using flight test
data at α = 120.7 deg, β = 72.4 deg, M = 0.51, and ρ = .78 kg/m3 (14kft altitude). The controller was then evaluated
using the following 12,000 wavefront snapshots to establish baseline performance. We then used exactly the same
controller on a different set of wavefronts from a different flight test. The “robustness” test was performed on a data
set with α = 129.3 deg, β = 65.3 deg, M = 0.50, and ρ = .86 kg/m3 (11.5kft altitude). This is a substantially perturbed
Dow
nloa
ded
by S
tani
slav
Gor
deye
v on
Jun
e 20
, 201
6 | h
ttp://
arc.
aiaa
.org
| D
OI:
10.
2514
/6.2
016-
3529
American Institute of Aeronautics and Astronautics
16
flow condition from the training condition; however, it is still in the same highly-separated flow regime. The results
of the robustness test are shown in Figure 11 for varying amounts of feedback latency. There is essentially no
difference between the baseline and robust performance for small amounts of latency. For larger amounts of latency,
errors in the training condition build up and disturbance rejection quality drops by up to 7%.
Figure 11: Effect of perturbing the flow condition on controller performance for varying amounts of
latency.
We may further investigate what is really happening by comparing the POD modes of each flow condition. The
POD modes are “reordered” such that they are arranged by similarity rather than energetic contribution, as shown in
Figure 12. While some differences are certainly present, most of the modes from the original predictive controller
are still present in the perturbed flow condition. This observation gives insight into why the predictive controller
tends to perform well. Model reduction via POD not only reduces computational cost but also increases robustness
by reducing sensitivity to perturbed flow conditions.
Figure 12: POD mode comparison using mode matching. (Mode 1 comparison at the top-left, Mode 4
comparison at the top-right, Mode 32 comparison at the bottom-right.)
0%
10%
20%
30%
40%
50%
60%
70%
80%
90%
0 2 4 6
Dis
turb
ance
Re
ject
ion
(%
)
Number Timesteps of Feedback Delay
Baseline Rejection
Robust Rejection
Dow
nloa
ded
by S
tani
slav
Gor
deye
v on
Jun
e 20
, 201
6 | h
ttp://
arc.
aiaa
.org
| D
OI:
10.
2514
/6.2
016-
3529
American Institute of Aeronautics and Astronautics
17
We now evaluate controller sensitivity further for several combinations of baseline cases and perturbed flow
cases. The method for testing this is to train the predictive controller on a subset of wavefront data in a baseline
condition, evaluate the controller on the subsequent wavefronts from the baseline condition to establish baseline
disturbance rejection, and then evaluate the controller again using the wavefront disturbances from a perturbed flow
condition.
The sensitivity to each parameter can be determined by calculating the change in controller performance as a
function of the change in each parameter. Ideally, the sensitivity to each parameter could be measured by perturbing
one parameter per test case, and calculating the sensitivity directly. Since this data is not available, an indirect
method will be required.
The fully separated flow regime will be examined here since it is of primary interest for real applications. It is
assumed that the sensitivity is constant in this regime; i.e., that performance degradation will follow linearly with
parameter changes. For convenience, define the function f as the mean aperture-averaged RMS disturbance rejection
of the controller over L test wavefronts,
,
,1
1 L
n rms
n rmsn
yf
dL
. [45]
Then, define s as a vector describing the change in viewing angle and modified elevation angle,
s . [46]
The directional derivative of f in the direction of s will define the sensitivity of the controller to a perturbation in
that direction,
f f s
s
s
. [47]
Recognizing that the change in RMS disturbance rejection from the baseline to the perturbed condition is
approximately equal to the LHS times the “length” of the perturbation vector, the following equation can be written,
perturbed baselinef f f u s . [48]
The above quantity may be measured directly from experiment and simulation. Each test condition then gives a
new set of equations,
perturbed baseline
f ff f
, [49]
which in turn leads to the following system of equations for Ntest sensitivity tests,
1 1 1
2 22
test test
test
perturbed baseline
perturbed baseline
N Nperturbed baseline N
f ff
f f
f
f f
. [50]
The full set of test conditions used for this sensitivity analysis is listed in Table 1.
Dow
nloa
ded
by S
tani
slav
Gor
deye
v on
Jun
e 20
, 201
6 | h
ttp://
arc.
aiaa
.org
| D
OI:
10.
2514
/6.2
016-
3529
American Institute of Aeronautics and Astronautics
18
Table 1: Sensitivity test conditions (flat window).
Condition # α (deg) β (deg) ρ (kg/m3) M (-)
1 126.4
52.0 0.85 0.51
2 129.3 65.3 0.86 0.50
3 135.7 51.1 0.87 0.51
4 141 59.7 0.85 0.51
The resulting estimated sensitivity parameters are shown in Table 2. An important result of this analysis is that
the POD-VAR control algorithm is nearly 4 times more sensitive to changes in modified elevation than to changes
in viewing angle. This is a consequence of the fact that POD modes do not significantly change with α since this
direction is closely aligned with the flow direction. Changes in β can result in encountering different parts of flow
structures such as horn vortices, necklace vortices, secondary vortices, etc.
Table 2: Sensitivity of mean disturbance rejection to perturbations in viewing and modified elevation angles.
∂f/∂α 0.10 (%/deg)
∂f/∂β 0.38 (%/deg)
This has practical importance for system design. If the POD-VAR predictors were to be trained and stored as a
function of α and β rather than dynamically updated in real-time, then it would be necessary to have roughly 4 times
finer resolution in the β lookup than in the α lookup.
On the other hand, if the POD-VAR predictor is updated in real-time in an outer loop, then these values can
establish the “drift” in controller performance as a function of the predictor update loop and turret slew rates.
VII. Conclusions
Latency in adaptive-optic control systems significantly limits controller performance in aero-optic applications
due to the high-frequency nature of disturbances. We have presented a modification to our previous neural network
controller that focuses on mitigating this limitation of adaptive-optic systems using flow prediction. The new linear
POD-VAR controller improves disturbance rejection from 35%-55% in the case of the neural network controller to
45%-75% over the same range of test conditions in simulation. While a nonlinear predictor may in general be more
accurate under ideal conditions, cumulative error in multistep prediction problems tends to build up more rapidly in
a nonlinear predictor while error buildup is not as rapid in the linear model. We have demonstrated good robustness
to perturbed flow conditions for the most optically-active case and discussed some of the physical reasons for this
characteristic. We have also shown that our controller is stable in both the nominal condition and in the presence of
delay uncertainty.
Future research should include realistic mirror and amplifier models, and re-evaluate the performance of the
controller under these conditions. The problem of varying amounts of latencies for each component in the feedback
loop does not affect the performance of the predictor in the nominal condition, but the stability analysis of this paper
should be expanded to address this more realistic case.
References 1 Wang, M., Mani, A., and Gordeyev, S., “Physics and Computation of Aero-Optics,” Annual Review of Fluid
Mechanics, vol. 44, Jan. 2012, pp. 299–321.
2 Gordeyev, S., and Jumper, E., “Fluid dynamics and aero-optics of turrets,” Progress in Aerospace Sciences, vol.
46, Nov. 2010, pp. 388–400.
3 Gladstone, J. H., and Dale, T. P., “Researches on the Refraction, Dispersion, and Sensitiveness of Liquids,”
Philosophical Transactions of the Royal Society of London, vol. 153, 1863, pp. 317–343.
4 Mahajan, V. N., “Strehl ratio for primary aberrations: some analytical results for circular and annular pupils,”
JOSA, vol. 72, 1982, pp. 1258–1266.
Dow
nloa
ded
by S
tani
slav
Gor
deye
v on
Jun
e 20
, 201
6 | h
ttp://
arc.
aiaa
.org
| D
OI:
10.
2514
/6.2
016-
3529
American Institute of Aeronautics and Astronautics
19
5 Ross, T. S., “Limitations and applicability of the Maréchal approximation,” Applied optics, vol. 48, 2009, pp.
1812–1818.
6 Porter, C., Gordeyev, S., and Jumper, E., “Large-aperture approximation for not-so-large apertures,” Optical
Engineering, vol. 52, 2013, pp. 071417–071417.
7 Burns, R., Jumper, E., and Gordeyev, S., “A Latency-Tolerant Architecture for Airborne Adaptive Optic
Systems,” 53rd Aerospace Sciences Meeting, 2015.
8 Jumper, E. J., Zenk, M. A., Gordeyev, S., Cavalieri, D., and Whiteley, M. R., “Airborne aero-optics laboratory,”
Optical Engineering, vol. 52, 2013, pp. 071408–071408.
9 Jumper, E. J., Gordeyev, S., Cavalieri, D., Rollins, P., Whiteley, M. R., and Krizo, M. J., “Airborne Aero-Optics
Laboratory-Transonic (AAOL-T),” AIAA Paper, vol. 675, 2015, p. 2015.
10 Vukasinovic, B., Gissen, A., Glezer, A., and Gogineni, S., “Fluidic Control of Transonic Shock-Induced
Separation,” American Institute of Aeronautics and Astronautics, 2013.
11 Vukasinovic, B., Glezer, A., Gordeyev, S., Jumper, E., and Kibens, V., “Active control and optical diagnostics of
the flow over a hemispherical turret,” AIAA Paper, vol. 598, 2008, p. 2008.
12 Vukasinovic, B., Glezer, A., Gordeyev, S., Jumper, E., and Bower, W. W., “Flow control for aero-optics
application,” Experiments in Fluids, vol. 54, Mar. 2013.
13 Gissen, A. N., Vukasinovic, B., Glezer, A., Gogineni, S., Paul, M. C., and Wittich, D. J., “Active Transonic Shock
Control,” American Institute of Aeronautics and Astronautics, 2014.
14 Gordeyev, S., and Thomas, F. O., “A temporal proper orthogonal decomposition (tpod) method for closed-loop
flow control,” AIAA Paper, vol. 359, 2010.
15 Bunton, Steven, and Noack, Bernd, “Closed-Loop Turbulence Control: Progress and Challenges,” Applied
Mechanics Reviews, vol. 67, Sep. 2015.
16 Gautier, N., Aider, J.-L., Duriez, T., Noack, B. R., Segond, M., and Abel, M., “Closed-loop separation control
using machine learning,” Journal of Fluid Mechanics, vol. 770, May 2015, pp. 442–457.
17 Parezanović, V., Duriez, T., Cordier, L., Noack, B. R., Delville, J., Bonnet, J.-P., Segond, M., Abel, M., and
Brunton, S. L., “Closed-loop control of an experimental mixing layer using machine learning control,” arXiv
preprint arXiv:1408.3259, 2014.
18 Whiteley, M. R., and Gibson, J. S., “Adaptive laser compensation for aero optics and atmospheric disturbances,”
38th AIAA Plasmadynamics and Lasers Conf, 2007.
19 Goorskey, D. J., Schmidt, J., and Whiteley, M. R., “Efficacy of predictive wavefront control for compensating
aero-optical aberrations,” Optical Engineering, vol. 52, 2013, pp. 071418–071418.
20 Tesch, J., Gibson, S., and Verhaegen, M., “Receding-horizon adaptive control of aero-optical wavefronts,” Optical
Engineering, vol. 52, 2013, pp. 071406–071406.
21 Page, K. A., “Applications of Linear Predictors in Adaptive Optics,” The University of Wyoming, 2005.
22 Berkooz, G., Holmes, P., and Lumley, J. L., “The Proper Orthogonal Decomposition in the Analysis of Turbulent
Flows,” Annual Review of Fluid Mechanics, vol. 25, 1993, pp. 539–575.
23 Kennel, M. B., Brown, R., and Abarbanel, H. D. I., “Determining embedding dimension for phase-space
reconstruction using a geometrical construction,” Physical Review A, vol. 45, 1992, pp. 3403–3411.
24 Lütkepohl, H., New Introduction to Multiple Time Series Analysis, Berlin: Springer, 2005.
25 Michiels, W., and Niculescu, S., Stability, Control, and Computation for Time-Delay Systems: An Eigenvalue-
Based Approach, Society for Industrial and Applied Mathematics, 2014.
26 Antsaklis, P. J., and Michel, A. N., A Linear Systems Primer, Boston, MA: Birkhäuser, 2007.
Dow
nloa
ded
by S
tani
slav
Gor
deye
v on
Jun
e 20
, 201
6 | h
ttp://
arc.
aiaa
.org
| D
OI:
10.
2514
/6.2
016-
3529
American Institute of Aeronautics and Astronautics
20
27 Franklin, G., Powell, D., and Emami-Naeini, A., Feedback Control of Dynamic Systems, Pearson, 2015.
28 Franklin, G., Powell, D., and Workman, M., Digital Control of Dynamic Systems, Prentice Hall, 1997.
29 Abado, S., Gordeyev, S., and Jumper, E., “AAOL wavefront data reduction approaches,” SPIE Defense, Security,
and Sensing, International Society for Optics and Photonics, 2012, p. 83950A–83950A.
Dow
nloa
ded
by S
tani
slav
Gor
deye
v on
Jun
e 20
, 201
6 | h
ttp://
arc.
aiaa
.org
| D
OI:
10.
2514
/6.2
016-
3529