Professional Documents
Culture Documents
Campo \lontenegro
São José dos Campos, SP - Brazil
2017
Cataloging-in Publication Data
Documentation and Information Division
Pereira, Mateus de Freitas Virgı́lio
Polynomial Chaos-Based Control Design for an Aircraft at High Angles of Attack with
Uncertain Aerodynamics / Mateus de Freitas Virgı́lio Pereira.
São José dos Campos, 2017.
115f.
BIBLIOGRAPHIC REFERENCE
PEREIRA, Mateus de Freitas Virgı́lio. Polynomial Chaos-Based Control Design
for an Aircraft at High Angles of Attack with Uncertain Aerodynamics. 2017.
115f. Dissertation of Master of Science – Instituto Tecnológico de Aeronáutica, São José
dos Campos.
CESSION OF RIGHTS
AUTHOR’S NAME: Mateus de Freitas Virgı́lio Pereira
PUBLICATION TITLE: Polynomial Chaos-Based Control Design for an Aircraft at High
Angles of Attack with Uncertain Aerodynamics.
PUBLICATION KIND/YEAR: Dissertation / 2017
ITA
To Davi Castro and Igor Prado for our
everlasting friendship.
Acknowledgments
First of all, I want to thank all professors that I had throughout my academic life for
sharing their knowledge.
I would like to thank my advisor, Prof. José Manoel Balthazar, for welcoming me into
his research group and giving me support and guidance during my master’s degree.
I want to thank Prof. Davi Antônio dos Santos for his expertise in estimation theory
and control that helped me in the development of this work.
I also want to thank the FADEMO research team for all data and invaluable inputs
given.
My sincere gratitude to my friends Davi Castro and Igor Prado for all support that
kept me motivated during this journey.
My heartfelt thanks to my mother, Margareth de Freitas, and my brother, Lucas de
Freitas, for all love.
Finally, I acknowledge the scholarship offered by the Conselho Nacional de Desenvolvi-
mento Cientı́fico e Tecnológico (CNPq).
The road to wisdom? - Well, it’s plain
and simple to express:
Err
and err
and err again,
but less
and less
and less.
— Piet Hein, The Road to Wisdom.
Resumo
In this work we consider the flight dynamics of fighter aircraft at high angles of attack
with uncertain aerodynamic coefficients. Stochastic parametric uncertainty is dealt with
by employing spectral decomposition of the random variables by means of the generalized
polynomial chaos expansion. The projection of uncertainty onto the orthogonal polyno-
mial basis, prescribed by the Wiener-Askey scheme, provides a deterministic model from
which the control laws are designed. We propose optimal linear and nonlinear feedback
control strategies for the automatic pilot system to recover the aircraft from stall and
provide acceptable dynamic response. Optimality of the proposed control laws is proved
by solving the Hamilton-Jacobi-Bellman equation and asymptotic weak stability in prob-
ability of the controlled nonlinear system is guaranteed in a deterministic way.
List of Figures
FIGURE 5.1 – Angle of attack of the aircraft with linear control. In black, 3rd
order gPC-based design. In red, MC-based design. . . . . . . . . . . 70
FIGURE 5.2 – Pitch angle of the aircraft with linear control. In black, 3rd order
gPC-based design. In red, MC-based design. . . . . . . . . . . . . . 71
FIGURE 5.3 – Pitch rate of the aircraft with linear control. In black, 3rd order
gPC-based design. In red, MC-based design. . . . . . . . . . . . . . 71
FIGURE 5.4 – Error of gPC expansion in comparison with MC simulations. . . . . 72
FIGURE 5.5 – Angle of attack of the aircraft with different control strategies. Ini-
tial condition α = 25 deg. . . . . . . . . . . . . . . . . . . . . . . . . 73
FIGURE 5.6 – Pitch angle and pitch rate of the aircraft with different control strate-
gies. Initial condition α = 25 deg. . . . . . . . . . . . . . . . . . . . 74
FIGURE 5.7 – Elevon deflection of the aircraft with different control strategies.
Initial condition α = 25 deg . . . . . . . . . . . . . . . . . . . . . . 75
FIGURE 5.8 – State estimation error for the aircraft with different control strate-
gies. Initial condition α = 25 deg . . . . . . . . . . . . . . . . . . . 76
FIGURE 5.9 – Parameter estimation for the aircraft with different control strate-
gies. Initial condition α = 25 deg . . . . . . . . . . . . . . . . . . . 76
LIST OF FIGURES x
FIGURE 5.10 –Stability criterion verification for the aircraft with different control
strategies. Initial condition α = 25 deg . . . . . . . . . . . . . . . . 77
FIGURE 5.11 –Angle of attack of the aircraft with different control strategies. Ini-
tial condition α = 30 deg . . . . . . . . . . . . . . . . . . . . . . . . 78
FIGURE 5.12 –Pitch angle and pitch rate of the aircraft with different control strate-
gies. Initial condition α = 30 deg . . . . . . . . . . . . . . . . . . . 79
FIGURE 5.13 –Elevon deflection of the aircraft with different control strategies.
Initial condition α = 30 deg . . . . . . . . . . . . . . . . . . . . . . 79
FIGURE 5.14 –State estimation error for the aircraft with different control strate-
gies. Initial condition α = 30 deg . . . . . . . . . . . . . . . . . . . 80
FIGURE 5.15 –Parameter estimation for the aircraft with different control strate-
gies. Initial condition α = 30 deg . . . . . . . . . . . . . . . . . . . 80
FIGURE 5.16 –Stability criterion verification for the aircraft with different control
strategies. Initial condition α = 30 deg . . . . . . . . . . . . . . . . 81
FIGURE 5.17 –Angle of attack of the aircraft with different control strategies. Ini-
tial condition α = 40 deg . . . . . . . . . . . . . . . . . . . . . . . . 82
FIGURE 5.18 –Pitch angle and pitch rate of the aircraft with different control strate-
gies. Initial condition α = 40 deg . . . . . . . . . . . . . . . . . . . 83
FIGURE 5.19 –Elevon deflection of the aircraft with different control strategies.
Initial condition α = 40 deg . . . . . . . . . . . . . . . . . . . . . . 83
FIGURE 5.20 –State estimation error for the aircraft with different control strate-
gies. Initial condition α = 40 deg . . . . . . . . . . . . . . . . . . . 84
FIGURE 5.21 –Parameter estimation for the aircraft with different control strate-
gies. Initial condition α = 40 deg . . . . . . . . . . . . . . . . . . . 84
FIGURE 5.22 –Stability criterion verification for the aircraft with different control
strategies. Initial condition α = 40 deg . . . . . . . . . . . . . . . . 85
List of Tables
Q Weight on states
R, R Weight on controls
P Solution of the algebraic Riccati equation
L Function for stability verification
V [·] Variance
E[·] Expectation
cov[·, ·] Covariance
μ Mean
σ Standard deviation
μp p-th moment
f (·) Probability density function
w(·) Weight function
I, Ii Support
Q ω , Qω Covariance matrix - dynamics
Rν Covariance matrix - outputs
P State covariance matrix
Contents
1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18
1.1 Motivation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18
1.2 Literature Review . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19
1.3 Objectives . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22
1.4 Contributions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22
1.5 Text Organization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23
2 Aircraft Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24
2.1 ITA-BWB Aircraft . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24
2.2 Flight Dynamics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25
2.2.1 Preliminary Definitions . . . . . . . . . . . . . . . . . . . . . . . . . . 26
2.2.2 Equations of Motion . . . . . . . . . . . . . . . . . . . . . . . . . . . 27
2.3 Aerodynamics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31
2.4 Simulation Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 36
4 Control Design . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 49
4.1 Stability of Stochastic Dynamical Systems . . . . . . . . . . . . . . 49
4.2 Optimal Control Design . . . . . . . . . . . . . . . . . . . . . . . . . . 54
CONTENTS xvii
5 Numerical Simulations . . . . . . . . . . . . . . . . . . . . . . . . 68
5.1 Simulation Parameters . . . . . . . . . . . . . . . . . . . . . . . . . . . 68
5.2 Accuracy of gPC expansion . . . . . . . . . . . . . . . . . . . . . . . . 70
5.3 Simulation Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 72
5.3.1 Initial Condition α = 25 deg . . . . . . . . . . . . . . . . . . . . . . . 73
5.3.2 Initial Condition α = 30 deg . . . . . . . . . . . . . . . . . . . . . . . 77
5.3.3 Initial Condition α = 40 deg . . . . . . . . . . . . . . . . . . . . . . . 82
Bibliography . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 89
This dissertation presents the longitudinal control design of a high performance aircraft
flying at high angles of attack with uncertain aerodynamics. Additionally, an optimal
control design technique based on polynomial chaos expansions for nonlinear systems
is introduced. This chapter is organized as follows: Section 1.1 presents the rationale
behind the development of this work; Section 1.2 summarizes the major results found
in the scientific literature on this topic; Section 1.3 states the main objectives of this
dissertation; Section 1.4 highlights the chief contributions of this work; and Section 1.5
outlines the organization of the remaining chapters.
1.1 Motivation
Aircraft Design Study and Demonstration (FADEMO) project, a joint research program
between the Instituto Tecnológico de Aeronáutica (ITA), in Brazil, and the Linköping
University, in Sweden. This project concerns new technologies in the design of future
combat aircraft and aims at developing new technologies that will reach a high technology
level readiness in the next two or three decades. In addition to flight controls, other
areas of research include aerodynamics, structures and manufacturing. The work in the
FADEMO project focus on the study of subscale demonstrators through testing of new
aircraft design concepts. The initial design results consider a blended-wing-body (BWB)
configuration, named ITA-BWB, from which systems design, wind tunnel tests and flight
tests have been carried out.
Many design challenges arise in flights at high incidence owing to nonlinear phenomena
that might strike the aircraft performance. For instance, bifurcation behavior regarding
the elevator control input may be observed, which may cause the jump phenomenon or lead
the vehicle to instability (LIAW; SONG, 2001; HUI; TOBAK, 1984; JAHNKE; CULICK, 1994).
Moreover, in an aerodynamic standpoint, the usual potential flow approach can no longer
predict the forces and moments acting on the aircraft due to complicated phenomena
such as boundary layer separation and vortex breakdown, to not mention compressibility
(ROM, 1992). Incidentally, many aircraft models are unable to predict completely and
satisfactorily the dynamics at high angles of attack. Therefore, for a reasonable model,
nonlinearities and uncertainties present in the real operation of these vehicles have to be
considered in order to avoid unstable regions of operation.
We are especially interested in the control strategy to recover the aircraft from the stall
region while guaranteeing appropriate stability and performance. Many control schemes
for the automatic pilot system are found in the literature, such as nonlinear quadratic
regulator (GARRARD et al., 1992), washout filters for bifurcation control (LEE; ABED, 1991),
dynamic inversion (BUGAJSKI; ENNS, 1992), probabilistic robust control (WANG; STENGEL,
2005), adaptive neural networks (CALISE; RYSDYK, 1998), to cite a few. Even though such
control strategies presented good performance at controlling high incidence flights, they
may have a difficult implementation and, in some cases, do not guarantee stability.
Rafikov and Balthazar (RAFIKOV; BALTHAZAR, 2008; RAFIKOV et al., 2008) introduced
an optimal linear feedback control to handle the task of controlling nonlinear dynamical
systems. Asymptotic stability of the closed-loop nonlinear system was proved by select-
ing a Lyapunov function, which could be seen to be the solution of the Hamilton-Jacobi-
Bellman equation thus guaranteeing both stability and optimality. Due to the simple
CHAPTER 1. INTRODUCTION 20
configuration and easy implementation, the use of linear control is preferred over the non-
linear design. Such control law showed to be effective at controlling high-angle-of-attack
flights with apparent bifurcation behavior (PEREIRA et al., 2008). This linear feedback
control strategy was also applied in different tasks, such as the control of a double pen-
dulum arm excited by a nonlinear shaker (TUSSET et al., 2016), in a nonlinear automotive
active suspension system (RAFIKOV et al., 2008; TUSSET et al., 2009), or in suppression of
chaotic behavior in a nonlinear ideal and nonideal vibrating system (TUSSET et al., 2012).
Notwithstanding the success of such applications, the control method showed to be very
sensitive to parametric uncertainty which may impinge the stability characteristics of the
dynamical system (BALTHAZAR et al., 2014b; NOZAKI et al., 2013; TUSSET et al., 2016;
BALTHAZAR et al., 2014a; PERUZZI et al., 2016; TUSSET et al., 2015). As a matter of fact,
such method was not designed to handle uncertain systems.
In some applications, however, nonlinear control laws outperform the linear controllers.
For instance, Abed e Fu (1986) demonstrated that only third or higher order control laws
can avoid instability due to Hopf bifurcation, a phenomenon that frequently occurs in
fixed-wing aircraft (SINHA; ANANTHKRISHNAN, 2013). Furthermore, as shown by Gar-
rard e Jordan (1977), nonlinear controls can lead to significant improvements in aircraft
performance when the aircraft is near the stall condition. In this region, the ability of the
nonlinear control to recover the vehicle from stall and quickly reduce the angle of attack
is superior than the linear counterpart. In addition, the nonlinear controller is capable of
postponing the non-recoverable stall. Garrard and Jordan’s method for deriving subopti-
mal nonlinear control policies is based on perturbational procedures for approximating the
solution of the Hamilton-Jacobi-Bellman Equation (GARRARD, 1969; GARRARD; JORDAN,
1977; GARRARD et al., 1992). Similar to the Rafikov and Balthazar’s method, this nonlin-
ear control design does not take into the consideration the robustness of the controller to
uncertain dynamics.
Adaptive and robust control are typical approaches for designing control structures
for systems affected by uncertainties. Applications of such techniques are widespread in
the aerospace engineering literature. Recent works include nonlinear dynamic inversion
with multivariate spline-based adaptive control allocation to compensate for aerodynamic
uncertainties (TOL et al., 2016), autopilot design using H∞ loop shaping for missiles at high
angles of attack (MAHMOOD et al., 2014), adaptively augmented LQR controller for agile
aircraft (ZOLLISTCH et al., 2015), L1 adaptive controller for a BWB aircraft (ZHENG; XU,
2013), H2 /H∞ robust control for hypersonic vehicles (HUANG; ZHANG, 2015), and Lypunov
function-based adaptive control for longitudinal dynamics (GAVILAN et al., 2015). Even
though such techniques are well established, they generally still present some unresolved
issues. For instance, the mathematical proof of asymptotic stability in adaptive structures
is usually fairly involved and, in some cases, not even possible. Robust control is too
CHAPTER 1. INTRODUCTION 21
cautious for considering only the worst-case scenario for control design, thus being blind
to the actual behavior of uncertainty in the system and possibly degrading performance
and cost of the feedback scheme. Furthermore, some methods, such as the sliding mode
control, are not concerned with the optimization of a certain performance index, i.e.,
optimality gives way to robustness. In view of these problems, a less conservative method
for controlling nonlinear systems is sought, in a way that asymptotic stability, optimality
and robustness are all guaranteed.
The robustness of the controller to stochastic inputs can be assessed by conventional
approaches such as Monte Carlo sampling methods, which might be highly computation-
ally costly and not practical for a large number of random variables. Polynomial chaos
expansion is an alternative to Monte Carlo methods, since it provides a simpler framework
to deal with the propagation of uncertainty in the dynamical model. It is the spectral de-
composition of a stochastic process in the random dimension in which it is parametrized.
The random trial basis is the Askey-scheme based orthogonal polynomials. Introduced
first by Wiener (1938) for Gaussian random variables and then generalized by Xiu e Kar-
niadakis (2002) for other probability distributions, such decomposition is based on the
Cameron and Martin Theorem (CAMERON; MARTIN, 1947), which guarantees the conver-
gence in L2 sense for stochastic processes with finite second moment. Many researches
have attested that spectral methods based on polynomial chaos expansions can be a com-
putationally efficient alternative to expensive conventional approaches based on Monte
Carlo sampling (XIU; KARNIADAKIS, 2002; FISHER; BHATTACHARYA, 2009; PRABHAKAR
et al., 2010; XIU; KARNIADAKIS, 2003). The polynomial chaos expansion framework has
been applied in sensitivity analysis (ABDELKEFI et al., 2012; SUDRET, 2008), uncertainty
quantification (PRABHAKAR et al., 2010; SEPAHVAND et al., 2010), structural dynamics
(GHANEM; SPANOS, 1991; JACQUELIN et al., 2015), robust stability analysis (FISHER; BHAT-
TACHARYA, 2008b; NECHAK et al., 2013), collision avoidance (KIM; BRAATZ, 2012), fluid
dynamics (HOSDER et al., 2006; KO et al., 2011), to cite a few.
The use of polynomial chaos expansions in control theory is recent. Hover e Tri-
antafyllou (2006) demonstrated that this method applies to the stability study of nonlin-
ear systems, especially when the methods of Lyapunov are inconclusive and Monte Carlo
simulations are expensive. Fisher and Bhattacharya (FISHER; BHATTACHARYA, 2009;
FISHER; BHATTACHARYA, 2008b; FISHER; BHATTACHARYA, 2008a) presented a general-
ized procedure for the stability analysis of linear and polynomial systems and a systematic
framework for designing linear quadratic regulators (LQR) for stochastic linear systems
using polynomial chaos expansions. Other applications in control theory are found in
optimal control (PENG et al., 2010; SRIRAM; JAMESON, 2011), model predictive control
(FAGIANO; KHAMMASH, 2012; MESBAH et al., 2014) and robust control (TEMPLETON et
al., 2012).
CHAPTER 1. INTRODUCTION 22
1.3 Objectives
This work has three fundamental objectives. First, to derive the equations of motion
of the ITA-BWB aircraft, accounting for the nonlinear aerodynamic forces and moments
at high angles of attack, especially in the longitudinal dynamics, and incorporating the
uncertain parameters that are encountered in real applications. Second, to design a con-
troller that is capable of recovering the aircraft from the stall condition with adequate
dynamic response even when the aerodynamic model is not fully known. The latter task
evolves into a more theoretical work by being mindful of the key goal of the FADEMO
project in exploring novel techniques for the design of high performance aircraft. Hence,
the third objective is to investigate the application of the polynomial chaos expansion
framework in the design of control laws for nonlinear systems with uncertain stochastic
parameters and, perforce, proofs of stability, robustness and optimality are sought.
1.4 Contributions
The foremost contribution of this dissertation lies in the novel formulation of non-
linear optimal control problems with stochastic parametric uncertainty in a determinis-
tic framework whilst assuring stability and optimality, ergo addressing two of the most
critical issues encountered in traditional approaches. Most importantly, the proposed
method takes into consideration the a priori knowledge of the underlying probability of
uncertainty during the control design, thus being less conservative than the conventional
robust control methods. For assuming that the uncertainty affecting the control system
has a stochastic nature and using such information for design purposes, the proposed
control strategy can be classified under the generic name of probabilistic robust control
(CALAFIORE; DABBENE, 2007). To the best knowledge of the author, this work is the first
to present a general deterministic framework for designing feedback nonlinear optimal
control laws for longitudinal attitude control of a nonlinear aircraft model with uncertain
aerodynamic data by means of intrusive polynomial chaos expansions.
In the context of the FADEMO project, this dissertation presents the full nonlinear
model of the ITA-BWB aircraft, some of the main features of the vehicle, and the aero-
dynamic data for longitudinal motion. Furthermore, this work shows that it is possible
to explore novel control design techniques to recover the aircraft from the stall condition
with adequate dynamic response, thus allowing high-angle-of-attack maneuvers, which are
important tasks in the mission of high performance fighters.
CHAPTER 1. INTRODUCTION 23
The remaining chapters of this dissertation delve into the aircraft modeling, uncer-
tainty propagation and control design. Chapter 2 presents the equations of motion of the
ITA-BWB aircraft, as well as the aerodynamic data and the simulation model. Chapter
3 reviews the polynomial chaos theory for propagating uncertainties through nonlinear
systems. Chapter 4 derives the linear and nonlinear optimal control strategies and the
proofs of stability, and proposes an implementation scheme. Chapter 5 presents the results
obtained from numerical simulations of the controlled aircraft model. Finally, Chapter
6 sums up the major conclusions from this dissertation and suggests topics for further
investigation.
Some results of this work were published in Pereira et al. (2017) and Pereira et al.
(2016).
2 Aircraft Model
Solutions to problems
are easy to find:
the problem’s a great
contribution.
What’s truly an art
is to wring from your mind
a problem to fit
a solution.
— Piet Hein, Last Things First.
This chapter presents the dynamical model of the ITA-BWB aircraft. Section 2.1
presents the salient features of the aircraft concept. Section 2.2 derives the full nonlinear
dynamical model of the aircraft. Section 2.3 shows the nonlinear aerodynamic coefficients
for longitudinal motion at high angles of attack. Finally, in Section 2.4 we present the
longitudinal model in the state space form used in simulations.
are usual control surfaces for tailless aircraft that combine the functions of elevators and
ailerons. When moved in the same direction, elevons cause a pitching moment to the
airframe. When moved differentially, they cause a rolling moment. The ITA-BWB has
six elevons, three in each semi-wing, as shown in Fig. 2.2. In that figure, the elevons are
identified by numbers according to their function during flight: elevons 1, which are close
to the fuselage, are responsible mostly for the pitch control; elevons 2 and 3, located near
the mid-wing and wingtip, work mostly as ailerons for roll control. We consider that the
elevons can be deflected up or down within the range of ± 30 deg.
To derive the equations of motion, first some notation is defined. To begin with, an
arbitrary scalar variable is denoted by r, while an arbitrary geometric vector is denoted
by r , and an arbitrary matrix is represented by R . Let S A = {x xA , y A , z A } represent
an arbitrary cartesian coordinate system A. In aeronautics, four coordinate systems are
defined (STEVENS; LEWIS, 2003):
Finally, the time derivative of a variable is denoted by superscript dot, that is, drr A /dt =
ṙr A .
Considering that ECEF is an inertial frame, then Newton’s second law states that
(STEVENS; LEWIS, 2003):
d
F = (mV V )e , (2.4)
dt
and
d
M = (H H )e , (2.5)
dt
where F and M are, respectively, the sum of external forces and moments acting on the
aircraft center of mass, m is the mass of the vehicle, V is the velocity in the ECEF frame
CHAPTER 2. AIRCRAFT MODEL 28
and H is the angular momentum of the aircraft. For the next derivations, it is expedient
that we write Eq. 2.4 and 2.5 in the body reference frame:
d
F = V )b + ω × mV
(mV V, (2.6)
dt
and
d
M = H )b + ω × H ,
(H (2.7)
dt
where ω is the angular velocity of the body frame with respect to the ECEF frame, defined
in the body-fixed frame as:
T
ω= p q r . (2.8)
In aeronautics, p, q and r are referred to as roll, pitch and yaw rate, respectively (ETKIN,
2012). The total force F , moment M , velocity V and angular momentum H cas also be
expressed in the body-fixed frame as:
T
F = Fx F y F z , (2.9)
T
M = L M N , (2.10)
T
V = u v w , (2.11)
H = Iω, (2.12)
Note that the zero entries in the matrix of inertia are due to the symmetry plane in the
x bz b -plane. The substitution of Eq. 2.8-2.12 into Eq. 2.6 and 2.7 yields the aircraft
equations of motion:
Fx = m(u̇ + qw − rv), (2.14)
Equations 2.14-2.19 represent the dynamic equations of a general aircraft with six
degrees of freedom. The external forces and moments acting on the center of the mass of
the vehicle are primarily due to the weight, propulsion and aerodynamic effects (ETKIN,
2012; STEVENS; LEWIS, 2003). To expresses these forces and moments in the body-frame
coordinate system, we analyze each contribution separately:
M = MA +MT, (2.21)
where superscripts A stands for aerodynamics, T stands for thrust and W stands for
weight. The weight of the aircraft acts on the center of gravity of the vehicle, thus not
contributing to the external moments. It is expressed as:
F W = mgg , (2.22)
Considering that the power plant is symmetric with respect to the x bz b -plane, then the
thrust is given by: ⎡ ⎤
F cos(αF )
⎢ ⎥
FT = ⎣ 0 ⎦, (2.24)
−F sin αF
where F is the net thrust and αF is the angle between the thrust vector and x b . If the
thrust vector is at a vertical position zF from the center of mass of the aircraft, and the
power plant has rotating parts with angular momentum IF and angular velocity ωF , then
CHAPTER 2. AIRCRAFT MODEL 30
the external moments due to the power plant are modeled as:
⎡ ⎤
0
⎢ ⎥
M T = ⎣ F zF − IF ωF r ⎦ , (2.25)
0
where the term IF ωF r accounts for the gyroscopic moment due to the rotating mass. The
aerodynamic forces are usually expressed in the wind-axes reference frame by means of
dimensionless coefficients: ⎡ ⎤
−CD
⎢ ⎥
F A = L bw q∞ S ⎣ CY ⎦ , (2.26)
−CL
where q∞ = (1/2)ρVt2 is the dynamic pressure, S is a reference area (usually the wing
√
area), ρ is the density of air and Vt = u2 + v2 + w2 is the total velocity. The aerodynamic
coefficients CL , CD and CY are the lift, drag and side force coefficients, respectively, and
they depend on the geometry of the airflow in the vicinity of the vehicle. A thorough
discussion about these coefficients is presented in Section 2.3. Similarly, the aerodynamic
moments are also represented by dimensionless coefficients:
⎡ ⎤
bCl
⎢ ⎥
M A = q∞ S ⎣ cCM ⎦ , (2.27)
bCN
where b is the wing span, c is the means aerodynamic chord and Cl , CM and CN are,
respectively, the roll, pitch and yaw moment coefficients. These coefficients are also dis-
cussed in Section 2.3.
Note that the forces acting on the aircraft are dependent on the position and orienta-
tion of the vehicle. Consequently, kinematic equations that relate position and orientation
with the translation and rotation velocities are necessary. Using Euler angles {ψ, θ, φ} for
attitude parameterization, we have:
Finally, the position of the aircraft with respect the coordinate systems NED and
CHAPTER 2. AIRCRAFT MODEL 31
ẋN
μ̇ ≈ , (2.32)
R
ẋE
λ̇ ≈ , (2.33)
R cos(μ)
where R is the radius of Earth. Equations 2.32 and 2.33 are approximate forms
considering that Earth is perfectly spherical.
In sum, Eq. 2.14-2.33 represent the full six-degree-of-freedom nonlinear dynamical
model of a general rigid body aircraft. This mathematical model can be used for guidance,
navigation and control design.
2.3 Aerodynamics
In Section 2.2, the dynamical model derived is general in the sense that it may describe
any fixed wing aircraft. Incidentally, aerodynamics is the feature that gives the real flavor
of the aircraft motion since it depends greatly on the vehicle geometry (ANDERSON, 2010;
STEVENS; LEWIS, 2003). In the aerospace industry, flight control engineers have to work
closely to aerodynamicists and flight test pilots to understand the aerodynamic model and
flight specifications that have to be taken into account in the control system design.
Two different physical mechanisms contribute to producing the aerodynamic forces of a
vehicle in a steady airflow (STEVENS; LEWIS, 2003). First, the pressure distribution, when
integrated over the aircraft surface, results in a nonzero force component normal to that
surface. Second, shear stress, when integrated over the same surface, results in the skin
friction, a force tangencial to the surface and that is proportional to the wetted area. The
combination of the pressure force and the skin friction force is the resultant aerodynamic
force in the vehicle, which is usually resolved into two perpendicular components, the
lift and drag components, the former being perpendicular to the relative airflow velocity.
The mathematical model that describes both mechanisms becomes more complex as the
airflow transitions from laminar to turbelent flow (STEVENS; LEWIS, 2003).
As presented in Eq. 2.26 and 2.27, the aerodynamic forces and moments are propor-
tional to the dynamic pressure q∞ , the referece area S and dimensionless coefficients. Such
coefficients are usually a nonlinear function of the angle of attack α, the sideslip angle β,
the Mach M and Reynolds Re numbers, as well as of the aircraft geometry (ANDERSON,
CHAPTER 2. AIRCRAFT MODEL 32
where ΔC( ) accounts for increments due to control surfaces deflections, such as rudder and
elevons, and due to non-steady flight. For a rigid body aircraft with fixed shape, except
for the movable control surfaces, if the Mach number, the altitude and the atmospheric
conditions are specified, then the aerodynamic coefficients are, in practice, a function
of the angle of attack and the sideslip angle only. In this work, we do not consider
compressibility effects nor the sideslip angle deflection, since we study the longitudinal
dynamics with leveled wings and no wind perturbations.
In Eq. 2.26, the lift coefficient CL typically varies linearly with the angle of attack at
low to moderate α. In this region the airflow moves smoothly over the aircraft surface
and is attached over most part of it. At high angles of attack, however, the flow tends
to separate from the top surface of the airfoil due to viscous effects creating a large wake
behind the airfoil and a condition of reversed flow. As a consequence, the lift force is
decreased while the drag grows appreciably (ANDERSON, 2010). Under such conditions,
the aircraft is said to be stalled, and the lift coefficient is no longer a linear function of α.
To fully characterize the behavior of the aerodynamic coefficients as a nonlinear func-
tion of the angle of attack, data are usually obtained experimentally in wind-tunnel tests.
In these tests, sub-scale models are placed in steady airflow in the same Mach and Reynolds
conditions of the aircraft operation and the resultant forces are measured by load cells.
According to the Buckingham Pi Theorem (ANDERSON, 2010), the dimensionless coeffi-
cients obtained for the sub-scale model are identical to the full scale model coefficients
since the similarity parameters, M and Re, are the same and therefore the different flows
are dynamically similar. In order to study the ITA-BWB aircraft, wind-tunnel tests were
run for the sub-scale model and the experimental curves for the lift, drag and moment
coefficients were constructed. The tests were performed in the facilities of the Aeronau-
tics Institute of Technology (ITA) by the FADEMO team, at the following conditions:
M = 0.119, Re = 0.562 × 106 , T = 20.6◦ C, ρ = 93.9kPa. The results are shown in Fig.
2.3, 2.4 and 2.5.
To incorporate these coefficients to the mathematical model derived in Section 2.2, a
third order polynomial is fitted to the experimental data. This approximation is usual in
the literature of dynamics and control of high-angle-of-attack flights, such as in the works
of Garrard e Jordan (1977), Liaw e Song (2001), Pereira et al. (2008). Such polynomials
have the following standard form:
C( ) = C( )0 + C( )1 α + C( )2 α2 + C( )3 α3 + ΔC( ) , (2.35)
CHAPTER 2. AIRCRAFT MODEL 33
0.8
0.6
0.4
CL
0.2
0
Fitted curve
Experimental data
-0.2
-0.4
-10 -5 0 5 10 15 20 25 30
α [deg]
FIGURE 2.3 – ITA-BWB lift coefficient.
where C( )0 , C( )1 , C( )2 and C( )3 are the polynomial coefficients, and ΔC( ) represent terms
not directly related to α. The fitted curves for the lift, drag and moment coefficients
are plotted in Fig. 2.3, 2.4 and 2.5, respectively. One can see that these polynomials
match with the experimental points and therefore the third order polynomial provides an
accurate model for the nonlinear dependence of the coefficients with the angle of attack.
The polynomial coefficients for CL , CD and CM are shown in Tab. 2.1, as well as the
coefficient of determination R2 of the fitted curves.
Coefficient C( )0 C( )1 C( )2 C( )3 R2
CL 0.0179 3.2569 0.5450 −9.6098 0.9979
CD 0.0355 −0.1171 1.6552 0.8908 0.9930
CM −0.0332 −2.8543 0.8669 2.3927 0.9980
Even though the third order polynomials provide a reasonable approximation for the
aerodynamic coefficients at low to high angles of attack, there may still be nonlinear phe-
nomena associated with the stall condition that were not fully captured in the wind-tunnel
tests. As described by Rom (1992), the stall evolution is a complicated process. At low
angles of attack, the viscous shear layers in the wing airfoil are confined to separation
CHAPTER 2. AIRCRAFT MODEL 34
0.45
0.4
Fitted curve
Experimental data
0.35
0.3
0.25
CD
0.2
0.15
0.1
0.05
0
-10 -5 0 5 10 15 20 25 30
α [deg]
FIGURE 2.4 – ITA-BWB drag coefficient.
bubbles and have minor effects on the aerodynamic characteristics of the aircraft. As the
angle of attack is increased to moderate and high angles, the separated layers detach from
the airfoil surface into the flow as vortex sheets, rolling up into concentrated vortices. At
still higher angles of attack, the cores of the concentrated vortices may burst at a certain
point on the lifting surface, in a phenomenon refered to as vortex breakdown, leading
eventually to the flow braking up into a nonsteady turbulent wake. Besides the compli-
cated stall process that might not be fully characterized by the function approximation,
the data obtained in the wind-tunnel tests may not be completely accurate since there
are other similarity parameters, in addition to M and Re, not accounted for (ANDERSON,
2010). Moreover, the final validation of the mathematical model is performed in flight
tests, which are usually not completed before the first control design routine. For these
reasons, there may be a mismatch between the real aerodynamic coefficient curves and the
mathematical model, due to unmodeled phenomena, especially in high angles of attack,
where the nonlinearites are more significant. To incorporate such issue in the aircraft
model, we add a parameter of uncertainty in the lift coefficient, such that Eq. 2.35 for
this coefficient is now written as:
0.4
0.2
Fitted curve
Experimental data
0
-0.2
CM
-0.4
-0.6
-0.8
-1
-10 -5 0 5 10 15 20 25 30
α [deg]
FIGURE 2.5 – ITA-BWB moment coefficient.
where ξ ∼ N(0, σC2 L ) is a normal random variable. Note that the choice of a Gaussian
distribution for CL relies on the fact that the uncertain aerodynamics at high angles of
attack originates from various statistically independent random phenomena with unknown
distribution which, when combined in the light of the Central Limit Theorem, should
eventually converge to a normal distribution (PAPOULIS; PILLAI, 2002). In this work, we
consider that σC2 L = 1.1, so that the outcome of the maximum lift coefficient is within
the range usually observed for high performance aircraft. Figure 2.6 shows a Monte Carlo
simulation of Eq. 2.36 with 1000 outcomes, as well as the mean and standard deviation.
To complete the longitudinal aerodynamic model, the increments ΔC( ) in Eq. 2.34 are
modeled as linear functions of the control surface deflections and the linear and angular
velocities and accelerations (STEVENS; LEWIS, 2003). For lift and moment increments we
have:
c Vt
ΔCL = CLq θ̇ + CLu + CLδe δe , (2.37)
2Vc Vc
c Vt
ΔCM = CMq θ̇ + CMu + CMδe δe . (2.38)
2Vc Vc
√
where Vt = u2 + v2 + w2 is the total velocity and Vc is the cruise speed. The elevon
deflection is represented by δe . Increments in the drag coefficient are not significant and
CHAPTER 2. AIRCRAFT MODEL 36
0.8
0.6
0.4
CL
0.2
Mean
-0.2 Mean ± std
-5 0 5 10 15 20 25
α [deg]
FIGURE 2.6 – Lift coefficient with uncertainty.
therefore are not considered in this work. The coefficients in Eq. 2.37 and 2.38 for the
ITA-BWB aircraft are shown in Tab. 2.2 (PEE, 2015).
Coefficient C ( )q C ( )u C ( )δe
CL 4.649 0.0919 −0.2532
CM −0.9064 0 −4.599
Sections 2.2 and 2.3 presented the full nonlinear model for a six-degree-of-freedom
aircraft. However, this work is concerned with the longitudinal dynamics only, since
we focus on control design for high-angle-of-attack flights. Such flight regime is critical
for high performance combat aircraft in rapid maneuvers such as evasion, pursuit, and
nose pointing to obtain the first opportunity of firing the weapons (ATESOGLU, 2007).
Examples are the Cobra and Herbst maneuvers, in which the aircraft for short periods
has to attain high angular velocities at extreme angles of attack. If we consider level flight
CHAPTER 2. AIRCRAFT MODEL 37
θ̇ = q. (2.42)
For the ITA-BWB aircraft, we consider that moments due to the power plant are not
significant and therefore can be neglected. Equations 2.39-2.42 can be written in terms
of α, which is more convenient when studying flights at high angles of attack. By using
w = u tan(α) and ẇ = u̇ tan(α) + uα̇ sec2 (α), and assuming that the aircraft flies at
constant velocity, i.e. u̇ = 0, we can combine Eq. 2.35 and 2.39-2.42 to determine a
nonlinear model for the longitudinal motion.
1 cCLq θ̇
α̇ = − q∞ S cos3 (α) (CL3 + ξ)α3 + CL2 α2 + CL1 α + + CL0 + CLδe δe
mu 2V0
CLu V t cCLq θ̇
+ − q∞ S sin2 (α) cos(α) (CL3 + ξ)α3 + CL2 α2 + CL1 α + + CL0
V0 2V0 (2.43)
C Lu V t
+CLδe δe + − 2q∞ S sin(α) cos2 (α) CD3 α3 + CD2 α2 + CD1 α + CD0
V0
2 2 2
−mg cos (α) cos(θ) − mg sin(α) cos(α) sin(θ) − muθ̇ sin (α) − muθ̇ cos (α) ,
θ̇ = q, (2.44)
1 cCMq θ̇ C Mu V t
q̇ = cq∞ S CM 3 α 3 + CM 2 α 2 + CM 1 α + + C M 0 + C M δe δ e + . (2.45)
Iyy 2V0 V0
Equations 2.43-2.45 represent the full nonlinear longitudinal model for the ITA-BWB
aircraft with uncertain aerodynamics. Note that the elevons deflection δe are related to
CHAPTER 2. AIRCRAFT MODEL 38
pitch control and thus are commands for elevons 1 in Fig. 2.2, as discussed in Section
2.1. For control design purposes, as it will become clear in Section 4, these equation are
conveniently rewritten in the state space form as follows:
x(t, ξ) = Ax (t, ξ) + H (x
ẋ x, ξ) + B u(t) (2.46)
where ⎡ ⎤ ⎡ ⎤
x1 α
⎢ ⎥ ⎢ ⎥
x = ⎣ x 2 ⎦ ⎣ θ ⎦ ∈ R3 and u(t) δe ∈ R, (2.47)
x3 q
⎡ ⎤
q∞ S
mu
(CD0 − CL1 ) 0 (1 − 2muV
cq∞ S
0
)C L q
⎢ ⎥ 3×3
A⎣ 0 0 1 ⎦∈R , (2.48)
cq∞ S 2
c q∞ S
Iyy
CM 1 0 C
2Iyy V0 Mq
⎡ q SC ⎤
∞ Lδ
e
⎢ mu ⎥
B ⎢
⎣ 0 ⎥ ∈ R3 ,
⎦ (2.49)
cq∞ SCMδ
e
Iyy
T
and H (x
x, ξ) = h1 h2 h3 ∈ R3 , where
1 cCLq x3
h1 − q∞ S cos (x1 ) (CL3 + ξ)x31 + CL2 x21 + CL1 x1 +
3
+ CL0 + CLδe u(t)
mu 2V0
C Lu V t cCLq x3
+ − q∞ S sin (x1 ) cos(x1 ) (CL3 + ξ)x31 + CL2 x21 + CL1 x1 +
2
+ CL0
V0 2V0
C Lu V t
+CLδe u(t) + − 2q∞ S sin(x1 ) cos2 (x1 ) CD3 x31 + CD2 x21 + CD1 x1 + CD0
V0
−mg cos2 (x1 ) cos(x2 ) − mg sin(x1 ) cos(x1 ) sin(x2 ) − mux3 sin2 (x1 ) − mux3 cos2 (x1 )
cq∞ S
−q∞ SCLδe u(t) − q∞ S(CD0 − CL1 )x1 − 1 − C Lq x3 ,
2V0
(2.50)
h2 0, (2.51)
1 3 2 C Mu V t
h3 cq∞ S CM 3 x1 + CM 2 x1 + CM 0 + . (2.52)
Iyy V0
The aircraft parameters are shown in Tab. 2.3 (PEE, 2015). Additionally, we con-
sider that the flight operation takes place at conditions of the International Standard
Atmosphere (ISA) and the local acceleration of gravity is g = 9.80655 m/s2 .
CHAPTER 2. AIRCRAFT MODEL 39
Finally, to complete the aircraft simulation model, we should take account of uncer-
tainties in the environment and the model, other than those discussed in Section 2.3.
For instance, such uncertainties could represent the effects of wind gusts or unmodeled
high-frequency plant dynamics (STEVENS; LEWIS, 2003). Moreover, we also should con-
sider noise that impairs the measurements from sensors. Equipment in the aircraft usu-
ally provides direct discrete measurements of the angle of attack, pitch angle and pitch
rate(STEVENS; LEWIS, 2003), but the states are partially observed due to sensor noise.
Therefore, for an outcome of random variable ξ in Eq. 2.46, the aircraft longitudinal
dynamics is modeled as follows:
x(t) = Ax (t) + H (x
ẋ x, t) + B u(t) + G ω(t), (2.53)
where ⎡ ⎤
0
⎢ ⎥
G ⎣ 0 ⎦ ∈ R3 , (2.55)
1
C I 3 ∈ R3×3 , (2.56)
and the signal {ω(t)} ∈ R and the sequence {νν k } ∈ R3 are noise terms, which are assumed
to be realizations of zero mean white Gaussian processes {W (t)} and {N N k }, respectively,
such that W (t) ∼ (0, Qω ) and N k ∼ (00, Rν ), where Qω ∈ R and Rν ∈ R3×3 are known
covariance matrices. The initial condition of 2.53, x 0 , is an outcome of the random variable
X 0 ∼ (x̄
x, P̄ x ∈ R3 and P̄
P ), where x̄ P ∈ R3×3 are known. By assumption, X 0 , {W (t)} and
N k } are mutually uncorrelated.
{N
3 Polynomial Chaos Expansion
In this section we introduce the polynomial chaos expansion for representing a stochas-
tic process according to the Wiener’s theory of homogeneous chaos (WIENER, 1938). We
also describe the use of such framework to transform stochastic nonlinear dynamical sys-
tems into an augmented deterministic system for the decomposition coefficients.
∞
Ψα (ξξ ) = Hαi (ξi ), (3.2)
i=0
CHAPTER 3. POLYNOMIAL CHAOS EXPANSION 41
where Hn (·) is the normalized n-th order multivariate orthogonal Hermite polynomial (see
Appendix A). The random functions Ψα form a complete orthonormal basis in L2 on the
probability space with respect to the Gaussian measure generated by ξ (BRANICKI; MA-
JDA, 2012). According to the Cameron-Martin theorem (CAMERON; MARTIN, 1947) (see
Appendix B), a finite-variance random variable g(τ, ξ ), where τ denotes any deterministic
parameter (e.g. the time instant), has the following decomposition in Wick polynomials:
g(τ, ξ ) = gα (τ )Ψα (ξξ ), (3.3)
α ∈J
g(τ, ξ ), Ψα (ξξ )
gα (τ ) = . (3.4)
Ψα2 (ξξ )
Notation · represents the multivariate inner product in L2 (Dξ ) space and Dξ is the
domain of ξ . The spectral expansion in Eq. 3.3 is referred to as the Polynomial Chaos
Expansion (PCE) of the random variable g. Given a g and a polynomial basis Ψα (ξξ ), the
expansion in Eq. 3.3 is unique. A fundamental property of this expansion is that it con-
verges in the L2 sense. In practice, the PCE will be truncated to a finite number of terms
by limiting the germ ξ to r normal random variables and the order of the multivariate poly-
nomials Ψα to |α α| ≤ d. The resulting number of terms in the truncated expansion is given
by P = (r+d)!
r!d!
− 1. The parameter P has to be chosen large enough for the approximation
of Eq. 3.3 to be accurate. The rate of convergence of the truncated PCE depends on the
the smoothness of g(τ, ξ ) as a functional in a Hilbert measure space. Roughly speaking,
for a PCE of order P , denoted by gP (τ, ξ ), the approximation error
g(τ, ξ ) − gP (τ, ξ )
is
O(P −p ), where p is the differentiability of the function g(τ, ξ ). For an analytic function
g(τ, ξ ), the convergence rate is exponential, i.e.,
g(τ, ξ ) − gP (τ, ξ )
= O(e−P ) for some
constant > 0 (KIM; BRAATZ, 2012).
An interesting result of the expansion presented in Eq. 3.3 is that the first two statis-
tical moments of the random variable g are readily computed from the terms of the PCE,
with no need of Monte Carlo sampling methods. Considering the truncated decomposi-
tion, we have: P
E[g(τ, ξ )] ≈ E gk (τ )Ψk (ξξ ) = g0 , (3.5)
k=0
P
P
V [g(τ, ξ )] ≈ V gk (τ )Ψk (ξξ ) = gk2 (τ )V [Ψk (ξξ )]. (3.6)
k=0 k=0
Xiu e Karniadakis (2002) extended the PCE framework to non-Gaussian random in-
puts by elaborating a tree that maps a certain distribution to the corresponding orthogonal
CHAPTER 3. POLYNOMIAL CHAOS EXPANSION 42
polynomial that guarantees the optimal convergence of the expansion. Table 3.1 shows the
pairing for continuous and discrete distributed random variables. This broad framework
is called Generalized Polynomial Chaos (gPC) expansion and it associates the probability
density function of the germ ξ with polynomials within the Askey-scheme that have a
similar weight function. Such polynomials form a complete basis in the Hilbert space
determined by their corresponding support (FISHER; BHATTACHARYA, 2009; ASKEY; WIL-
SON, 1985). For arbitrary uncertain input distributions outside the Askey scheme, the
Gram-Schmidt orthogonalization algorithm can be used to generate a polynomial basis to
achieve exponential convergence of the expansion (WITTEVEEN; BIJL, 2006). Note that,
in the gPC framework, the random variables in ξ do not need to have the same prob-
ability distribution function, thus we can propagate simultaneously through the model
uncertainties with different distributions, as long as we built the appropriate multivariate
tensor product polynomial for the decomposition basis (see Appendix A).
cess, in a very similar framework as shown by Peng et al. (2010). If the process covariance
is known, then such process can be approximated using Karhunen-Loève (KL) expansion
functions (FISHER, 2008) and then incorporated to the gPC expansion. However, the KL
expansion introduces new terms to the germ, which can make the gPC expansion com-
putationally intractable, since the number of terms in the expansion depends in factorial
order on the size of the germ. In the case of white noise processes, the number of terms
of the expansion will be infinite. Therefore, the application of gPC to regular stochas-
tic processes are usually limited to processes with non-zero correlation window (FISHER,
2008). Regular processes are recurrent in dynamical systems with stochastic forcing (e.g.:
an aircraft flight with wind gusts), which are not covered in this work.
where x ∈ Rn is the state vector, u ∈ Rm is the control input, A ∈ Rn×n and B ∈ Rn×m
are matrices and h ∈ Rn is a vector whose components are continuous nonlinear functions
of x . The random variable ξ = (ξ1 , ξ2 , . . . , ξr ) represents uncertainties, with a known
stationary distribution, by which the parameters of the model or the initial conditions are
expressed. Due to the randomness of the system parameters, one can see that the state
trajectories will also be stochastic, i.e., {x
x(t)} is a parametric stochastic process. Here we
assume that the control input is computed by some state feedback strategy and therefore
will also present a stochastic behavior.
Denote the components of x (t, ξ ), u (t, ξ ) and h (t, x , ξ ) by, respectively, xi (t, ξ ), ui (t, ξ )
and hi (t, x , ξ ). Also, denote the elements of A (t, ξ ) and B (ξξ ) by, respectively, Aij (t, ξ )
and Bij (ξξ ). We can represent xi (t, ξ ), ui (t, ξ ), Aij (t, ξ ), Bij (ξξ ) and hi (t, x , ξ ) by means of
the gPC expansion of order P in the orthogonal polynomial basis Ψk (ξξ ) according to the
Wiener-Askey scheme (see Table 3.1):
P
xi (t, ξ ) ≈ xi,k (t)Ψk (ξξ ), (3.8)
k=0
P
ui (t, ξ ) ≈ ui,k (t)Ψk (ξξ ), (3.9)
k=0
CHAPTER 3. POLYNOMIAL CHAOS EXPANSION 44
P
Aij (t, ξ ) ≈ aij,k (t)Ψk (ξξ ), (3.10)
k=0
P
Bij (t, ξ ) ≈ bij,k (t)Ψk (ξξ ), (3.11)
k=0
P
hi (t, x , ξ ) ≈ hi,k (t)Ψk (ξξ ). (3.12)
k=0
Note that, as long as the relation between h and x is known, the expansion for hi (t, x , ξ ),
Eq. 3.12, may be reformulated in terms of the expansions xi (t, ξ ) (see Section 3.2.2).
Let us define the following vector notation for the mode strengths appearing in Eq.
3.8-3.12:
T
x i xi,0 (t) xi,1 (t) . . . xi,P (t) ∈ RP +1 , (3.13)
T
ui ui,0 (t) ui,1 (t) . . . ui,P (t) ∈ RP +1 , (3.14)
T
hi hi,0 (t) hi,1 (t) . . . hi,P (t) ∈ RP +1 . (3.15)
Define also the matrices A k ∈ R(n×n) with {Ak }ij = aij,k (t) and B k ∈ R(n×m) with {Bk }ij =
bij,k for k = 0, . . . , P . Using the intrusive approach in Xiu e Karniadakis (2002), the co-
efficients aij,k (t) and bij,k are computed by the Galerkin projection of each mode onto
the polynomial basis {Ψq }Pq=0 , in order to ensure the error is orthogonal to the functional
space spanned by the finite-dimensional basis Ψk :
x = Ax + Bu
Consider the linear part of Eq. 3.7, that is, ẋ Bu. The gPC expansion and
subsequent Galerkin projection onto {Ψq }q=0 result in:
P
n
P
P
m
P
P
ẋi,q = aij,k xj,q êqkr + bij,k uj,q êqkr , (3.18)
j=1 k=0 r=0 j=1 k=0 r=0
where
Ψk Ψr , Ψq
êqkr = . (3.19)
Ψ2q
Equation 3.18 can be put in a more suitable form (FISHER; BHATTACHARYA, 2009; FISHER;
BHATTACHARYA, 2008b; FISHER; BHATTACHARYA, 2008a):
Ẋ = AX + BU, (3.20)
T
U= u T1 u T2 . . . u Tm . (3.22)
Matrices A ∈ Rn(P +1)×n(P +1) and B ∈ Rn(P +1)×m(P +1) are defined as:
P
A= Ak ⊗ E k , (3.23)
k=0
P
B= B k ⊗ E k, (3.24)
k=0
where A k and B k are defined below Eq. 3.15, ⊗ denotes the Kronecker product and
E k ∈ R(P +1)×(P +1) is a symmetric matrix given by:
⎡ ⎤
ê1k1 ê1k2 . . . ê1kP
⎢ ⎥
⎢ ê1k2 ê2k2 . . . ê2kP ⎥
Ek = ⎢
⎢ .. .. ... .. ⎥
⎥. (3.25)
⎣ . . . ⎦
ê1kP ê2kP . . . êP kP
CHAPTER 3. POLYNOMIAL CHAOS EXPANSION 46
Consider the nonlinear part of Eq. 3.7, that is, ẋ = h . We will first assume that h has
a polynomial form in x (FISHER; BHATTACHARYA, 2008b).
s
hi (t, x , ξ ) = xβ j (t, ξ ),
ηj (t, ξ )x (3.26)
j=1
where ηj ∈ R are uncertain coefficients of the polynomial of x , which has s terms, and
β j ∈ Rn , whose elements are non-negative integers, is a vector containing the order of
each term in the monomial. We may represent ηj (t, ξ ) in gPC expansion as:
P
ηj (t, ξ ) ≈ ηj,k (t)Ψk (ξξ ). (3.27)
k=0
Similar to aij,k and bij,k , the coefficients ηj,k are obtained through the projection onto
{Ψq }Pq=0
ηj (t, ξ ), Ψk (ξξ )
ηj,k (t) = . (3.28)
Ψ2k (ξξ )
The substitution of the expansions for xi and ηj , Eq. 3.12 and 3.27 respectively, in Eq.
3.26 yields:
P
P
P
P
xβ j (t, ξ ) =
ηj (t, ξ )x ··· ···
kη =0 k1,1 =0 k1,βj1 =0 k2,1 =0
P
(3.29)
··· [ηj,kη x1,k1,1 · · · x1,k1,βj1 x2,k2,1 · · ·
kn,βjn =0
where
Ψk · · · Ψkn,βjn , Ψq
êq,kη ,k1,1 ,...,kn,βjn = . (3.31)
Ψ2q
CHAPTER 3. POLYNOMIAL CHAOS EXPANSION 47
The inner products present in both Eq. 3.19 and 3.31 can be computed numerically.
Equation 3.30 may be rewritten in a more compact form as
ŝ
Ẋ = H = η̂j Xβ̂ j , (3.32)
j=1
where η̂j represents the coefficients of polynomials of X , which has ŝ terms, and the vector
β̂ j contains the order of each monomial. The vector H ∈ Rn(P +1) may be represented as:
T
H= h T1 h T2 ... h Tn . (3.33)
The representation of the nonlinearities of the dynamics in gPC expansions may be a chal-
lenging task when h is not a polynomial function, since the projection onto the polynomial
basis may not have a closed form. A possible approach to representing such functions is
to expand them in Taylor series around the mean of the argument. For instance, consider
the case where hi (t, x , ξ ) is a non-polynomial function of the single variable xj (ξξ ). The
Taylor series expansion around the mean of hi , that is, the first term of the gPC expansion
according to Eq. 3.5, gives:
p
N (p)
h (xj,0 )
P
hi (xj (ξξ )) ≈ i
xj,k Ψk (ξξ ) . (3.34)
p=0
p! k=1
Using the Multinomial theorem (ABRAMOWITZ; STEGUN, 1965), we can rewrite Eq. 3.34
as:
N (p)
hi (xj,0 ) p! P
hi (xj (ξξ )) = xkj,rr Ψkr r (ξξ ) . (3.35)
p=0
p! k +k +···+k =p
k 1 !k 2 ! · · · k P ! r=1
1 2 P
Equation 3.35 shows that the expansion of hi in Taylor series has a similar form as in Eq.
3.26 and therefore the resulting Galerkin projection may be written as in Eq. 3.32. As an
CHAPTER 3. POLYNOMIAL CHAOS EXPANSION 48
example, consider the expansion of the following random variable with N = 3 and P = 2:
Ẋ = AX + H + BU X(t0 ) = X0 . (3.36)
Remark 1: Note that this intrusive approach increased the dimension of the system
by a factor equal to the order of the gPC expansion. On the other hand, the behavior
of the dynamical system is now represented in a deterministic framework. Equations
3.8 and 3.9 provide the interface between the states and controls in the stochastic and
deterministic domains. The augmented system captures the behavior of the deterministic
part of the original system in Eq. 3.7, which includes the time-dependent terms, thus
preserving stability characteristics. Given that the system in Eq. 3.36 converges to the
origin in finite time, the stochastic system in Eq. 3.7 also converges in probability to such
point (see Section 4.1), as it can been seen in Eq. 3.8. Hence, the goal of the control
design presented in Section 4 is to find U that makes the origin a stable fixed point of the
augmented system.
4 Control Design
In this chapter we present the incorporation of the polynomial chaos framework for de-
signing control laws for nonlinear systems with stochastic parametric uncertainty. Section
4.1 presents some definitions about stability of stochastic dynamical systems. In Section
4.2, we introduce two optimal control strategies for regulation of uncertain nonlinear sys-
tems using polynomial chaos expansions. In Section 4.3, an implementation scheme is
proposed.
x(t, ξ ) = f (t, x , ξ )
ẋ x (t0 , ξ ) = x 0 (4.1)
where x ∈ Rn is the state vector and ξ ∈ Rd is the vector of random variables. Fol-
CHAPTER 4. CONTROL DESIGN 50
lowing the usual procedure of introducing new variables, equal to the deviations of the
corresponding coordinates of the perturbed motion from their unperturbed values (KHAS-
MINSKII, 2010), only the stability of the solution x (t, ξ ) ≡ 0 has to be considered as long
as the following condition is satisfied:
Consider the following definitions of stochastic stability of the zero equilibrium of the
system in Eq. 4.1 (CHEN et al., 1995; KUSHNER, 1967; KHASMINSKII, 2010):
In addition, the solution is said to be asymptotically stable with probability one (or almost
surely stable) if it is stable in probability and
lim Pr lim x(t) = 0 = 1. (4.4)
0
x 0 →0 t→∞
Definition: p-stability: The solution x (t, ξ ) is said to be p-stable for p > 0 and t ≥ t0
if, for every
> 0, there exists a δ > 0 such that
x(t)|p ] ≤
,
sup E[|x for all x 0 : |x
x0 | ≤ δ. (4.5)
t≥t0
x(t)|p ] = 0,
lim E[|x for all x 0 in the neighborhood of 0. (4.6)
t→∞
CHAPTER 4. CONTROL DESIGN 51
E[|x x0 |p e−ζ2 t
x(t)|p ] ≤ ζ1 |x for all x 0 : |x
x0 | ≤ δ. (4.7)
Note that p-stability of the solution for any p1 > 0 implies p-stability of every smaller
value of p2 > 0, since the Lp1 -norm dominates the Lp2 -norm if 0 < p2 < p1 (KHASMINSKII,
2010). From the Chebyshev’s inequality, we also have that (CHEN et al., 1995)
1
sup Pr {|x
x|(t) ≥
} ≤ p
x(t)|p ]
sup E[|x (4.8)
t≥t0
t≥t0
Hence,
lim sup Pr {|x(t)
x(t)| ≥
} = 0
x(t) for all
> 0. (4.9)
0 t≥t0
x 0 →0
Note that, even though similar, the definition in Eq. 4.9 is weaker than the definition in
4.3, since
sup |x
x(t)| ≥ |x
x(t)| (4.10)
⇒ {sup |x
x(t)| ≥
} ⊇ {|x
x(t)| ≥
} (4.11)
⇒ Pr {sup |x
x(t)| ≥
} ≥ Pr {|x
x(t)| ≥
} (4.12)
⇒ Pr {sup |x
x(t)| ≥
} ≥ sup Pr {|x
x(t)| ≥
} (4.13)
The definition in Eq. 4.9 is usually referred to as weak stochastic stability (CHEN et al.,
1995). Therefore, p-stability implies weak stochastic stability in probability, but does
not necessarily entail stochastic stability in probability, let alone almost sure stability.
In fact, the entire concept of Lp convergence is weaker than the almost sure counterpart
for random numbers, since, while both take conclusions from large samples due to the
smoothing effects, the former derives from the weak law of large numbers, whereas the
latter is based on the strong law of large numbers (GUT, 2013). In the particular case of
linear stochastic differential equations, the two definitions are equivalent (KHASMINSKII,
2010).
Now consider the truncated polynomial chaos expansion of the solution x (t) of the
nonlinear stochastic differential equation in Eq. 4.1. For the sake of simplicity and with
no loss of generality, we consider the one-dimensional case, i.e., n = 1 and d = 1:
x(t, ξ) = X T Ψ (4.14)
such that
X(t) = f (t, X )
Ẋ X (t0 ) = X 0 , (4.15)
CHAPTER 4. CONTROL DESIGN 52
T
where X = x0 (t) x1 (t) . . . xP (t) ∈ RP +1 is the vector that concatenates the
T
gPC mode strengths, and Ψ = Ψ0 (ξ) Ψ1 (ξ) . . . ΨP (ξ) ∈ RP +1 is the vector with
the polynomial bases. The following proposition presents, using the polynomial chaos
framework, the conditions for weak stochastic stability of the zero equilibrium of Eq. 4.1.
Theorem 4.1: The zero equilibrium solution of the nonlinear dynamical system with
stochastic parametric uncertainty in Eq. 4.1 is uniformly weakly asymptotically stable
in probability in the large if, for the augmented deterministic system in Eq. 4.15 for the
mode strengths of the polynomial chaos expansion of Eq. 4.1, a scalar function V(t, X)
with continuous first partial derivatives with respect to X and t exists such that:
(i) V(t, 0 ) = 0;
(ii) V(t, X ) is positive-definite, i.e., there exists a continuous nondecreasing scalar func-
X|) such that κ1 (00) = 0 and V(t, X ) ≥ κ1 (|X
tion κ1 (|X X|) > 0 for all t and all X = 0 ;
(iii) V(t, X ) is decrescent, i.e., there exists a continuous nondecreasing scalar function
X|) such that κ2 (00) = 0 and κ2 (|X
κ2 (|X X|) ≥ V(t, X ) for all t;
∂V
V̇(t, X) = + (V)T f (t, X) ≤ −κ3 (|X
X|) < 0,
∂t
For any ε > 0, there exists a δ > 0 such that κ2 (δ) < κ1 (ε). Let the initial condition be
CHAPTER 4. CONTROL DESIGN 53
X| < δ. Then
chosen such that |X
X(t)| < ε
|X for all t ≥ t0 ,
and hence the zero equilibrium is uniformly stable. To prove uniform asymptotic stability,
we have to show that for any σ > 0, there exist ε1 and g(σ) > 0 such that
|X X(t)| < δ
X(t0 )| < ε1 ⇒ |X for all t > t0 + g(σ).
Let δ and ε1 be such that κ2 (δ) < κ1 (ε1 ) and, from the result, if |X X(t0 )| < δ, then
X(t)| < ε1 for all t ≥ 0. Choose σ < δ and ν > 0 such that κ2 (ν) < κ1 (σ), then
|X
X(t)| < ν for some time t ≥ t0 + g(σ), then |X
if |X X(t)| < σ for all t ≥ t0 + g(σ). By
contradiction, we show that such g(σ) exists. Let |X X(t)| be in the interval [σ, ε1 ] for all
t ≥ t0 . From condition (iv), the function κ3 (|XX|) has a minimum κ30 in this interval.
Defining g(σ) = κ2 /κ30 , then there exists a time t2 = t0 + g(σ) at which
which is a contradiction, since κ1 (σ) > 0. Hence, the solution X ≡ 0 is uniformly asymp-
X|) is radially unbounded, an ε can be found
totically stable. Since by condition (v), κ1 (|X
such that κ2 (δ) < κ1 (ε) for any δ, which can be made arbitrarily large. Therefore, the
zero equilibrium of Eq. 4.15 is uniformly asymptotically stable in the large (NARENDRA;
ANNASWAMY, 2005).
(b)p-stability of the stochastic system: This part uses the derivations presented by
Fisher e Bhattacharya (2009) in the analysis of p-stability. Recall that the vector X (t)
contains the gPC mode strengths of the random variable x(t, ξ) in a time t, where ξ is
a random variable with known continuous probability density distribution f (ξ) over the
support I. In face of Eq. 4.14, the p-th moment of x(t, ξ) is given by:
P
P
μp (t) = ··· xi1 (t) · · · xip (t) Ψi1 (ξ) · · · Ψip (ξ)f (ξ)dξ, (4.16)
i1 =0 ip =0 I
CHAPTER 4. CONTROL DESIGN 54
for p > 0. Note that the integral in Eq. 4.16 is finite for any p and any set of orthogonal
polynomials. Therefore, if conditions (i)-(v) are true, then X (t) is bounded, which implies
that xi (t) is also bounded and consequently, all moments are bounded. Hence, the zero
equilibrium of Eq. 4.1 is p-stable for all p > 0 and t ≥ 0. Moreover, if conditions (i)-(v)
are true, then the solution X (t) ≡ 0 is asymptotically stable and X (t) → 0 as t → ∞,
that is, xi (t) → 0 and consequently μp (t) → 0 as t → ∞ for all p > 0 and t ≥ 0. Finally,
since the convergence of X holds for all t0 and all X 0 ∈ RP +1 , then the convergence of
xi (t) to the origin holds for all t0 and all xi (t0 ) ∈ R and thus the convergence of μp (t) is
uniform and the system is p-stable in the large for all p > 0 and t ≥ 0. In sum, the zero
equilibrium of Eq. 4.1 is uniformly asymptotically p-stable in the large for all p > 0. Using
the Chebyshev Inequality (Eq. 4.8) and the squeezing principle, we have that the zero
equilibrium of the nonlinear dynamical system with stochastic parametric uncertainty in
Eq. 4.1 is uniformly weakly asymptotically stable in probability in the large.
In light of the Theorem 4.1, the goal of the control design is to find a control law that
will drive the gPC mode strengths of Eq. 4.1 to the origin, thus guaranteeing the weak
stochastic stability of the equilibria of the nonlinear dynamical system with stochastic
parametric uncertainties.
The objective is to design a feedback control system that drives the trajectories of Eq.
3.7 from a certain initial condition to a small neighborhood around the origin, i.e. x (tf ) =
0, while minimizing a performance index. For convenience, the stochastic dynamical
system is restated here, but now specifying the feedback structure of the control law and
introducing
, a dimensionless parameter used for notation purposes:
x(t, ξ ) +
h
ẋ (t, ξ ) = A (t, ξ )x h(t, x , ξ ) + B (ξξ )u
u(x
x). (4.17)
where
x) = x T Qx +
N (x
l(x x), (4.19)
Q ∈ Rn×n is a real symmetric and positive definite matrix and R ∈ Rm×m is a real positive
definite matrix. N (xx) is a term that will be chosen later to account for the nonlinearities
CHAPTER 4. CONTROL DESIGN 55
of the model. Using the generalized polynomial chaos expansion, the stochastic dynamical
system can be represented as in Eq. 3.36:
H + BU
Ẋ = AX +
H (4.20)
Likewise, note that we can write the functional in Eq. 4.18 in the gPC framework, just
as presented in Section 3. Using the orthogonal polynomials property, one can show that
(FISHER; BHATTACHARYA, 2009):
E[x X,
x T Qx ] = X T Q
QX (4.21)
E[u U,
u T Ru ] = U T R
RU (4.22)
where Q Q ⊗W, R R ⊗ W and W ∈ R(P +1)×(P +1) is a matrix whose elements are
W }ij = Ψi (ξξ ), Ψj (ξξ ). In general, the expectation of N (x
{W x) can be written as a function
of the gPC coefficients for x :
x)] = N(X
E[N (x X). (4.23)
Hence, the cost function in Eq. 4.18 can be represented in a deterministic form as:
tf
min J (X
X, U ) = min U dt,
L + UT R
RU (4.24)
U U t0
where
X +
N(X
L = XT Q
QX X). (4.25)
The general solution of this optimal control problem is of the form (GARRARD et al., 1967):
X)
∂V (X
U = f X, . (4.27)
∂XX
CHAPTER 4. CONTROL DESIGN 56
Thus, it is necessary to solve a first order nonlinear partial differential equation. Here, we
present two formulations to obtain a linear and a nonlinear optimal feedback control law.
Proposition 1: Consider the augmented system presented in Eq. 4.20. Let us choose
X) = −H
N(X HT PX − XT P
PX H , so that
PH
X − HT P
L = XT Q
QX X − XT P
PX H.
PH (4.28)
The optimal control law that transfers the augmented dynamical states from a given initial
condition to the origin while minimizing the cost function in Eq. 4.24, as long as L is
positive definite, has the form
U = −R X,
−1B T P
PX (4.29)
˙ −1B T P + Q
−P = P A + A T P − P BR P (tf ) = 0 . (4.30)
V = X T P X , (4.31)
where P is a symmetric positive definite matrix and it satisfies the differential Riccati
equation, Eq. 4.30. Note that P has these attributes given that R and Q are positive
definite matrices, which is true if R and Q are also positive, since W is always a positive
definite matrix. For the infinite time horizon problem, matrix P is the solution of the
nonlinear matrix algebraic Riccati equation:
−1B T P + Q
P A + A T P − P BR = 0. (4.32)
CHAPTER 4. CONTROL DESIGN 57
The derivative of the function V , evaluated in the optimal trajectory with control given
by Eq. 4.29, is:
˙
V̇ = X T P + P A + A T P − P BR −1B T P −
(4.33)
−1 )T B T P X + H T P
P B (R X + XT P
PX H.
PH
Let us now consider the Hamilton-Jacobi-Bellman (HJB) Equation. Substituting Eq. 4.29
and 4.33 into Eq. 4.26 , we obtain
X − HT P
L = XT Q
QX X − XT P
PX H.
PH (4.34)
X) has to be equal to −H
Comparing Eq. 4.25 and 4.34, we see that N(X X − XT P
HT P
PX H.
PH
Now consider a Lyapunov function candidate of same form as Eq. 4.31. For positive
definite function L and positive definite matrix R , the derivative of the function evaluated
in the optimal trajectory is given by V̇ = −L − U T R U and it is negative definite. Then
RU
V is a Lyapunov function and the controlled system is locally asymptotically stable. If
Γ = Rn(P +1) , the controlled system is asymptotically stable in the large due to the radial
unboundness condition for the Lyapuvov function V(X X) → ∞ as |X X| → ∞.
Remark 2 : The augmented deterministic system with control given by Eq. 4.29 is
asymptotically stable and therefore the original stochastic system in Eq. 4.17 is asymptot-
ically p-stable for all p > 0 and, consequently, weakly asymptotically stable in probability,
as discussed in Section 4.1.
Remark 3 : The choice of matrix A (t, ξ ) in Eq. 4.17 is not unique for a determined
system thus influencing the control design. By the same token, the gain matrices Q and
R can be chosen in order to satisfy some performance specification.
implement the control input corresponding to U , one can use Eq. 3.9 for recovering the
stochastic variable.
Here we follow the formulation presented by Garrard et al. (1967) to design a nonlinear
suboptimal control to drive the states of the augmented system to the origin in an infinite
time horizon. From the HBJ equation (Eq. 4.26), it is straightforward to see that the
optimal control law that minimizes the cost function in Eq. 4.24 is of form:
X)
1 −1 T ∂V (X
U=− R B (4.35)
2 ∂XX
The substitution of the control law in Eq. 4.35 into Eq. 4.26 yields:
∂V ∂V T 1 ∂V T −1 T ∂V X +
N(X
+ AX +
H
(A H) − BR B + XT Q
QX X) = 0 (4.36)
∂t X
∂X X
4 ∂X X
∂X
An analytical solution for the HJB equation in Eq. 4.36, in general, does not exist.
However, it is possible to obtain a control close to the optimal if it approximately satis-
fies the conditions specified in the optimal control problem. In a practical standpoint, a
suboptimal solution can be chosen so that it satisfies criteria such as ease of implemen-
tation and reliability. Garrard et al. (GARRARD et al., 1967; GARRARD; JORDAN, 1977)
proposes a pertubational procedure to obtain the approximate solution of the optimal
control problem. Assume a formal power series expansion of the cost-to-go function:
∞
X) =
V (X X)
i−2 Vi (X (4.37)
i=2
Substituting Eq. 4.37 into Eq. 4.36 and equating the powers of
to zero, we obtain
(GARRARD et al., 1967):
X,
L = XT Q
QX (4.41)
∈ Rn(P +1)×n(P +1) is a real symmetric positive definite matrix. Assuming also
where Q
that H has a polynomial form in X , such that
N
H= X)
f n+1 (X (4.42)
n=1
where fn+1 is of order n + 1 in X . Then the Vn ’s are given by the following equations
(GARRARD; JORDAN, 1977).
and
1 −1 T ∂Vn
∞
U=− R B (4.45)
2 n=0
X
∂X
X,
V0 = X T P
PX (4.46)
where P is the solution of the algebraic Riccati equation in 4.32. Equation 4.44 is linear,
but in most cases cannot be solved analytically. Considering the augmented dynamical
system with n(P + 1) states, the general procedure for obtaining approximate solutions
is as follows(GARRARD; JORDAN, 1977):
CHAPTER 4. CONTROL DESIGN 60
1. Assume
n+2
n+2
i
Vn = ··· ani1 ,...,in(P +1) xi11 · · · xn(P
n(p+1)
+1) ,
i1 =0 in(P +1) =0 (4.47)
s.t. i1 + · · · + in(P +1) ≤ n + 2;
∂Vn
2. Calculate X;
∂X
∂Vn
3. Substitute X
∂X
into Eq. 4.44;
5. Solve the resulting simultaneous linear algebraic equations for ani1 ,...,in(P +1) .
After Vn is obtained, ∂V n
X can be calculated and substituted into Eq. 4.45 to obtain
∂X
U n+1 . The convergence of this pertubational procedure and an estimate of the degradation
of performance resulting from truncation is discussed in Garrard (1969) and Lukes (1969).
Remark 6: Optimality does not imply stability, unless the infinite horizon optimal con-
trol law U(X
X) is stabilizing, i.e., U(00) = 0 and the zero solution of the closed-loop system
in Eq. 4.20 is asymptotically stable in some neighborhood. Therefore, an assumption to
apply the suboptimal control law derived in this section is that Eq. 4.20 is stabilizable
(LUKES, 1969; LEE; MARKUS, 1967), that is, the Jacobian matrix
∂
J0 = AX + H + BU ]X =00
[A (4.48)
X
∂X
is a stability matrix. That is the case if max[Re(λ(JJ0 ))] < 0, where λ(JJ0 ) denotes the
eigenvalues of the Jacobian matrix. Considering the control law in Eq. 4.45, it is straight-
forward to verify that it is stabilizing under similar conditions for stability presented in
Section 4.2.1. The closed-loop dynamics is given by:
1 −1 T ∂Vn
∞
Ẋ = AX + H − BR B , (4.49)
2 n=0
X
∂X
CHAPTER 4. CONTROL DESIGN 61
and therefore the associated Jacobian matrix at the origin can be calculated by:
∂ 1 −1 T
∞
∂Vn
J0 = AX + H − BR B (4.50)
X
∂X 2 n=0
X
∂X
0
X =0
∂ 1 −1 T ∂V0
= AX + H − BR B , (4.51)
X
∂X 2 X X =00
∂X
since ∂VX has a polynomial form in X of order 2 or greater for n ≥ 1. Note that the first
n
∂X
order control term left in Eq. 4.51 is identical in form to the control law in Eq. 4.29.
From Section 4.2.1, we know that the zero equilibrium point in the closed loop dynamical
system in Eq. 4.20 with control given by Eq. 4.29 is asymptotically stable if the function
in Eq. 4.28 is positive definite. Consequently, under such condition, the Jacobian matrix
in Eq. 4.51 must be a stability matrix. Hence, the nonlinear control in Eq. 4.45 is
stabilizing and the closed-loop trajectories initiated within a neighborhood of the origin
asymptotically converge to the equilibrium point.
4.3 Implementation
In Section 4.2, a linear and a nonlinear optimal control strategies were presented. Both
strategies were derived for the augmented dynamical system for the gPC mode strengths,
that is, the control laws map X to U . In an actual application, however, the mode strengths
are not physical quantities, thus they cannot be measured. Therefore, the gPC coefficients
have to be estimated in real time, as well as the uncertain parameters, in order to allow
the computation of the optimal control U and the subsequent determination of the actual
control input u (t, ξ ) for the dynamical system, using Eq. 3.9. From Section 3, we know
that the gPC mode strengths of the decomposition of x(t, ξ ) ∈ Rn are given by:
Since E[xi (t, ξ ), Ψk (ξξ )] = E[xi (t, ξ )]E[Ψk (ξξ )]+cov[xi (t, ξ ), Ψk (ξξ )], E[Ψk (ξξ )2 ] = V [Ψk (ξξ )]+
E[Ψk (ξξ )]2 and E[Ψk (ξξ )] = 0 for k ≥ 1(see Appendix A), then this implies
and
Note that, in Eq. 4.54, the variance V [Ψk (ξξ )] can be computed off-line since the polyno-
mial basis is known for a given set of uncertain parameters (see Appendix A for examples).
Therefore, the determination of the gPC mode strengths of the actual state vector boils
down to the computation of the covariance between the states and the polynomial basis
at a certain time instant. The estimation of the set of uncertain parameters ξ ∈ Rd can
be performed through different on-line parameter estimation algorithms, such as the re-
cursive least square method or the maximum likelihood technique. Ljung e Soderstrom
(1983) present several methods for recursive parameter estimation. We denote the esti-
mated parameters by ξ̂ξ . Each parameter in the vector ξ is assumed to be time-invariant,
that is, each overcome of the random variables are constant over time. Therefore, the
parameter estimation algorithm has to run only until all parameters have been estimated,
when the gain of the algorithm has ultimately decreased to zero (GOODWIN; SIN, 1984).
The convergence of the parameter estimation, however, is dependent on some properties
of the input signals in the dynamical system. As discussed by Narendra e Annaswamy
(2005), this signal property is referred to as persistent excitation and it is pivotal for
on-line parameter estimation. An input signal u(t) is said to be persistently exciting (PE)
CHAPTER 4. CONTROL DESIGN 63
x).
y = c (x (4.57)
Note that the proposed approach bears a resemblance to traditional adaptive control
strategies, since the model is not completely specified and we combine on-line parameter
estimation with on-line control, providing a self-learning nature to the controller (GOOD-
WIN; SIN, 1984). The remarkable difference between the traditional adaptive control and
the polynomial chaos-based control is that the latter takes the uncertain parameters as
random variables and incorporates the probability density function of the state in the
feedback law.
In this work we suggest Kalman filtering for adaptive state estimation, that is, the
states and parameters are estimated simultaneously. The Extended Kalman Filter (EKF)
is the most common application of the Kalman Filter to nonlinear systems. The EKF
uses the assumption that all transformations are quasi-linear and then simply linearizes
about the current mean and covariance all nonlinear transformations and substitutes
Jacobian matrices for the linear transformations in the Kalman Filter equations (JULIER;
CHAPTER 4. CONTROL DESIGN 64
Ψk (ξξ )
Estimation algorithm
ξ̂ξ y
Controller Ψk (·) Estimation of ξ c(·)
X
X̂ Estimation of X
where
x(t, ξ ) + h (t, x , ξ ) + B (ξξ )u
A (t, ξ )x u(t, ξ )
x(t), u (t))
f (x ∈ Rn+d , (4.61)
¯ ¯ 0 d×1
G ∈ Rn×nw , and the signal {ω ω (t)} ∈ Rnw and the sequence {νν k } ∈ Rs are noise terms,
¯
which are assumed to be realizations of zero mean white Gaussian processes {W W (t)} and
N k }, respectively, such that W (t) ∼ (00, Qω (t)) and N k ∼ (00, Rν k ), where Qω (t) ∈ Rnw ×nw
{N
and Rν k ∈ Rs×s are known covariance matrices. The initial condition of 4.59, x 0 , is an
¯
outcome of the random variable X 0 ∼ (¯ x, P¯ ), where x¯ ∈ Rn and P¯ ∈ Rn×n are known. By
x
¯ ¯ ¯ ¯ ¯
assumption, X 0 , {WW (t)} and {NN k } are mutually uncorrelated. The addition of the noise
¯
signals accounts for the unmodeled disturbances and uncertainties in the environment and
model in a real application. Note that we consider that the uncertain parameters ξ are
constant over time, that is, we estimate an output of the random variables. For instance,
considering the aircraft model presented in Section 2, we consider that for each high
angle of attack maneuver the lift curve is different due to the outcome of the uncertain
CHAPTER 4. CONTROL DESIGN 65
parameter. However, while a certain maneuver is being performed, the lift curve does not
modify.
Similar, to the Kalman Filter, the EKF is divided into two parts:
Prediction:
ˆ˙
x
x(t) = f (ˆx
x(t), u (t)) Predicted estimate (4.62)
¯ ¯ ¯
P˙ (t) = F (t)P (t) + P (t)F (t)T + G G(t)Q G(t)T
Qω (t)G Cond. cov. predicted estimate (4.63)
¯ ¯ ¯ ¯ ¯ ¯ ¯
yˆk+1|k = c k+1 (ˆ
x ) Predicted measurement (4.64)
¯ ¯ k+1|k
P Yk+1|k = C k+1P k+1|kC Tk+1 + Rν k+1 Cond. cov. pred. measurement(4.65)
¯ ¯ ¯ ¯
P XY = P k+1|kC Tk+1 Conditional cross covariance (4.66)
¯ k+1|k ¯ ¯
Update:
where
∂f (ˆ
x
x(t), u (t))
F (t) = ¯ ¯ ∈ Rn+d×n+d , (4.70)
¯ ∂x
¯
and
dcck+1 (ˆ
x k+1|k )
C k+1 = ¯ ∈ Rs×n+d . (4.71)
¯ dx
¯
Note that, in Eq. 4.54, the covariance between the state xi (t, ξ ) and the polynomial
basis Ψk (ξξ ) can be rewritten as a higher order moment between the state and the vector of
parameters ξ . In general, to compute the gPC mode strength xi,k , it is necessary moments
of order k + 1 between the state and ξ . The adaptive state estimation using EKF only
provides the first and second moments. Therefore, for a polynomial chaos expansion with
P > 1, numerical approximations of high-order moments are required. This is a major
drawback of the implementation of the polynomial chaos-based control using adaptive
estimation with EKF, since the computational cost to estimate such moments might be
significant. There are several methods to estimate high-order moments in the literature
(DUTTA; BHATTACHARYA, 2010; LIU et al., 2014; PONOMAREVA et al., 2010; GROTHE,
2013). For instance, Majji et al. (2008) present the so called Jth Moment Extended
Kalman Filter, a estimation framework for high order moments of nonlinear dynamical
systems using the state transition tensor approach. The third and fourth order moment
CHAPTER 4. CONTROL DESIGN 66
(3) (3)
K 2k+1P XY
P k+1|k+1 = P k+1|k + 3K Y
− 3K
K k+1P XXY − K 3k+1P Yk+1|k
YY
(4.72)
¯ ¯ ¯ k+1|k ¯ k+1|k ¯
and
(4) (4)
P k+1|k+1 = P k+1|k − 4K
K k+1P XXXY K 2k+1P XXY
+ 6K Y
K 3k+1P XY
− 4K YY
+
¯ ¯ ¯ k+1|k ¯ k+1|k ¯ k+1|k (4.73)
+4KK 4k+1P Yk+1|k
YYY
.
¯
The fifth and higher order moments are computed in a similar fashion. The tensors in
the right hand side of Eq. 4.72 and 4.73 are obtained through the state transition tensor
propagation of the statistical moments (refer to Majji et al. (2008) for derivations). The
implementation of this method for vector models is usually quite laborious due to the
algebraic transformations involved resulting in a high computation expenditure.
In this work, due to the computational cost, we will only consider the implementation
of control laws in an adaptive structure, as shown in Fig. 4.1, with gPC expansions up
to P = 1. Therefore, the adaptive state estimation with the EKF, or any other formula-
tion of the nonlinear Kalman filter, can be used to obtain on-line estimates of the gPC
mode strengths. In view of the highly nonlinear dynamical model for the aircraft consid-
ered in Section 2, we suggest the Unscented Kalman filter (UKF) in lieu of EKF. Even
though the EKF is the standard estimation algorithm for nonlinear dynamical systems,
it presents some issues with the reliability of the linearization along the state trajectory.
The UKF addresses possible deficiencies of linearization by providing a more direct and
explicit mechanism for transforming mean and covariance information by the unscented
transformation (JULIER; UHLMANN, 2004). Appendix C presents the formulation of the
UKF algorithm.
Next we present a summary for the control design and implementation of the method
proposed in this work:
• Step 1: Given a dynamical system of form as in Eq. 3.7, with r uncertain parameters
with known distribution, determine the expansion basis according to Wiener-Askey
scheme in Tab. 3.1 and multivariate orthogonal polynomials as presented in Ap-
pendix A.
• Step 2: Choose the polynomial order d so that the expansion error is small. The
gPC will then have P = (r+d)!
r!d!
− 1 terms.
• Step 3: Construct the augmented system as shown in Section 3.2. One should obtain
a deterministic system as in Eq. 3.36. Once the augmented system is determined,
one can analyze the statistical moments of the model by using Eq. 3.5 and 3.6.
Such results can be compared with Monte Carlo (MC) simulations of the original
CHAPTER 4. CONTROL DESIGN 67
stochastic system. The error between the results obtained through gPC and the
results from MC should be small. Otherwise, one should return to Step 2 and
increase d.
• Step 5: Implement the closed loop system as shown in Fig. 4.1. An estimation algo-
rithm should be chosen to provide online estimates of the states and the uncertain
parameters. The gPC mode strengths can be computed using Eq. 4.53 and 4.54.
For P = 1, an Extended Kalman Filter or any other formulation of the Kalman
Filter for nonlinear estimation should suffice. For P > 1, one should implement an
estimation algorithm that provides information about higher order moments.
5 Numerical Simulations
Problems worthy
of attack
prove their worth
by hitting back.
— Piet Hein, Problems.
This Chapter presents the numerical simulations of the control method proposed in
Chapter 4 applied to the aircraft model shown in Chapter 2. The algebraic calculations
were performed in Mathematica, while the numerical calculations were performed in MAT-
LAB. Section 5.1 shows the parameters used in the simulations. Section 5.2 discusses the
accuracy of the gPC expansion. Finally, Section 5.3 presents the simulation results. The
objective is to show that the controller can recover the aircraft from the stall condition.
The gains used in the control derivations, as shown in Sections 4.2.1 and 4.2.2, were
CHAPTER 5. NUMERICAL SIMULATIONS 69
selected as:
⎡ ⎤
1 0 0
⎢ ⎥
Q = ⎣ 0 1 0 ⎦, (5.3)
0 0 1
R = 1. (5.4)
Note that the gains were kept the same for all control designs in order to make a fair
comparison among the performance of the controllers. While we acknowledge that the
comparison between two different control laws in terms of performance might be ambigu-
ous due the peculiarities in the design and tuning of each controller, a comparison between
the linear and nonlinear strategies presented in Chapter 4 is reasonable since all control
laws are an approximation of the same optimal control problem. Due to space constraints,
the control laws derived are shown in Appendix D.
Considering the full simulation model in Eq. 2.53 and 2.54, which includes the output
model and the white noise terms to account for uncertainties in the environment and in
the measurements, the following parameters were used in all simulations:
Qω = 0.1745, (5.5)
⎡ ⎤
0.81 0 0
⎢ ⎥
Rν = ⎣ 0 0.81 0 ⎦ × 10−4 , (5.6)
0 0 0.81
⎡ ⎤
3 0 0
⎢ ⎥
P = ⎣ 0 3 0 ⎦,
P̄ (5.7)
0 0 3
⎡ ⎤
α0 − 0.0087 rad
⎢ ⎥
x = ⎣
x̄ 0 rad ⎦. (5.8)
0.0261 rad/s
(5.9)
As for the augmented system for the adaptive Kalman Filter algorithm presented in
Section 4.3, the following parameters were used:
P
P̄ 0 3×1
P¯ = , (5.10)
¯ 0 1×3 30
x
x̄
x¯ = . (5.11)
¯ −9.6098
(5.12)
CHAPTER 5. NUMERICAL SIMULATIONS 70
where α0 denotes the initial condition for the angle of attack. Three initial conditions
were simulated for each control law: x = [ 25o 0 0 ]T , x = [ 30o 0 0 ]T and x =
[ 40o 0 0 ]T . All initial conditions correspond to angles of attack within the stall region,
i.e., greater than 20 degrees. All simulations were run for 0 ≤ t ≤ 15 s. The integrations
were performed using a 4th order Runge-Kutta method.
A question that arises naturally is what order of the gPC expansion to use in order to
have a satisfactory accuracy in the uncertainty propagation. To answer this question, we
compare the gPC of the aircraft with linear control given by the law derived in Section
4.2.1 with Monte Carlo (MC) simulations in a statistical standpoint. For P = 3, the
mean, μ, and the standard deviation, σ, obtained through the gPC expansion are shown
as dashed and solid black lines, respectively, in Fig. 5.1 for the angle of attack. Similar
plots for the pitch angle and pitch rate are shown in Fig. 5.2 and 5.3, respectively. For
comparison, in Fig. 5.1-5.2, we also present in red the trajectories mean and standard
deviation computed by Monte Carlo simulations for a large sample (2000 realizations) of
normally distributed ξ. We observe a good agreement between the proposed gPC-based
design and the traditional MC. The computational cost for generating the gPC expansion
results, however, is much smaller than the cost to run the MC simulation.
30
gPC μ
gPC μ ± σ
25 MC μ
MC μ ± σ
20
α(t) [deg]
15 8
6
10 3.2 3.4 3.6 3.8
0
0 5 10 15
Time [s]
FIGURE 5.1 – Angle of attack of the aircraft with linear control. In black, 3rd order
gPC-based design. In red, MC-based design.
CHAPTER 5. NUMERICAL SIMULATIONS 71
-0.5
-3.2
-1
-3.6
-1.5 -4
-2.5
-3
-3.5
-4 gPC μ
gPC μ ± σ
-4.5 MC μ
MC μ ± σ
-5
0 5 10 15
Time [s]
FIGURE 5.2 – Pitch angle of the aircraft with linear control. In black, 3rd order gPC-
based design. In red, MC-based design.
-1
0.7
θ̇(t) [deg/s]
-2
0.6
-4
gPC μ
gPC μ ± σ
-5
MC μ
MC μ ± σ
-6
0 5 10 15
Time [s]
FIGURE 5.3 – Pitch rate of the aircraft with linear control. In black, 3rd order gPC-based
design. In red, MC-based design.
To compare the truncated gPC expansions with the MC results, Fig. 5.4 shows the
error of the mean and standard deviation between these two methods. The comparison was
carried out for 1st, 2nd, 3rd and 4th order gPC expansions. We can infer that the higher
the expansion, the more accurate the approximation is in a statistical viewpoint. However,
even for low order expansions such the 1st and 2nd order, the approximation is pretty
reasonable. As a matter of fact, all error curves in Fig. 5.4 have similar order of magnitude.
CHAPTER 5. NUMERICAL SIMULATIONS 72
On the other hand, the computational cost for propagating the uncertainty through 3rd
and 4th order expansions is significantly higher than for lower order gPC approximations.
Taking into consideration the trade-off between computational expenditure and accuracy,
we chose to work with 1st order (P = 1) gPC expansions in the following simulations.
Approximation of α
0.06 0.4
Error mean α [deg]
P = 1
-0.02 -0.2
0 5 10 15 0 5 10 15
Time [s] Time [s]
×10−3 Approximation of θ
5 0.1
Error mean θ [deg]
0 0.05
-5 0
-10 -0.05
0 5 10 15 0 5 10 15
Time [s] Time [s]
Approximation of θ̇
Error mean θ̇ [deg/s]
0.01 0.06
Error std θ̇ [deg/s]
0 0.04
-0.01 0.02
-0.02 0
-0.03 -0.02
0 5 10 15 0 5 10 15
Time [s] Time [s]
Next, we show the stall recovering simulations for each initial condition. For each
condition, the simulations were run for the linear, 2nd, 3rd and 5th order controllers
with the adaptive structure presented in Section 4.3. The UKF was used for state and
parameter estimation.
CHAPTER 5. NUMERICAL SIMULATIONS 73
20 20
α α
Stall Stall
α(t) [deg]
α(t) [deg]
15 Trim 15 Trim
10 10
5 5
0 0
0 5 10 15 0 5 10 15
Time [s] Time [s]
20 20
α α
Stall Stall
α(t) [deg]
α(t) [deg]
15 Trim 15 Trim
10 10
5 5
0 0
0 5 10 15 0 5 10 15
Time [s] Time [s]
FIGURE 5.5 – Angle of attack of the aircraft with different control strategies. Initial
condition α = 25 deg.
CHAPTER 5. NUMERICAL SIMULATIONS 74
Linear control
θ̇(t) [deg/s]
0 5
θ(t) [deg]
-2 0
-5
-4
-10
0 5 10 15 0 5 10 15
Time [s] Time [s]
2nd order control
0 20
θ̇(t) [deg/s]
θ(t) [deg]
-2 0
-4 -20
0 5 10 15 -40
0 5 10 15
Time [s] Time [s]
3rd order control
0 20
θ̇(t) [deg/s]
θ(t) [deg]
0
-5
-20
-10 -40
0 5 10 15 0 5 10 15
Time [s] Time [s]
5th order control
0 20
θ̇(t) [deg/s]
θ(t) [deg]
0
-5
-20
-10 -40
0 5 10 15 0 5 10 15
Time [s] Time [s]
FIGURE 5.6 – Pitch angle and pitch rate of the aircraft with different control strategies.
Initial condition α = 25 deg.
Figure 5.7 shows the control inputs computed for each controller. It is evident that
the linear control requires a smaller elevon deflection to recover the aircraft, especially in
the first second of the simulation. The nonlinear controllers are more aggressive, having
a peak of control input right in the beginning of simulation, and the 3rd and 5th order
controls reach the saturation at ±30 deg. Even so, all control inputs were adequate
for a high performance aircraft. After the initial peak, all signals eventually reach the
trim condition for the elevon deflection. We should point out, however, that we did not
consider the dynamics of the actuators and that the control signals are noisy due to the
additive white noise terms in the dynamics and in the measurements. In a real application,
the actuators responsible for the elevon deflection may not support such high frequency
command and therefore the control signal would have to be smoothed.
CHAPTER 5. NUMERICAL SIMULATIONS 75
-2
10
-4
u(t) [deg]
u(t) [deg]
0
-6
-8
-10
-10
-20
-12 u(t) u(t)
Trim Trim
-14 -30
0 5 10 15 0 5 10 15
Time [s] Time [s]
20
20
10
u(t) [deg]
u(t) [deg]
10
0
0
-10
-10
-20
u(t) u(t)
Trim Trim
-30 -20
0 5 10 15 0 5 10 15
Time [s] Time [s]
FIGURE 5.7 – Elevon deflection of the aircraft with different control strategies. Initial
condition α = 25 deg
Figures 5.8 and 5.9 show, respectively, the state estimation error and the parameter
estimation provided by the adaptive UKF. We see that for all states the estimation error
quickly approaches zero, indicating that the estimation algorithm is working well. The
estimated states are used in the control law, since the first gPC coefficient is equal to
the expected value of the random variable, as discussed in Section 4.3. In the case of
the parameter estimation ξ, which is crucial for computing the actual control signal, we
see that the estimation did not converge to the real value, but nevertheless it provided a
reasonable approximation. The parameter estimation for the system with the 2nd order
control had the better performance, whereas the estimation for the system with linear
control had the worst performance. Such issue can be explained in two ways: either
the control input was not sufficiently PE or the parameter had minor influence in the
dynamics. In the first case, as discussed in Section 4.3, the convergence of parameter
estimation is strongly dependent on whether the control signal is PE. In the second case,
since the uncertainty is more intense at high angles of attack, as shown in Fig. 2.6 an
initial condition at α0 = 25 deg may not have been high enough to provide sufficient
information for estimating ξ accurately.
CHAPTER 5. NUMERICAL SIMULATIONS 76
Linear control
1 0.5 1
Error θ̇ [deg/s]
Error α [deg]
Error θ [deg]
0.5 0 0
0 -0.5 -1
-0.5 -1 -2
0 5 10 15 0 5 10 15 0 5 10 15
Time [s] Time [s] Time [s]
2nd order control
1 1 2
Error θ̇ [deg/s]
Error α [deg]
Error θ [deg]
0.5 0.5
0
0 0
-0.5 -0.5 -2
0 5 10 15 0 5 10 15 0 5 10 15
Time [s] Time [s] Time [s]
3rd order control
1 1 2
Error θ̇ [deg/s]
Error α [deg]
Error θ [deg]
0.5 0.5
0
0 0
-0.5 -0.5 -2
0 5 10 15 0 5 10 15 0 5 10 15
Time [s] Time [s] Time [s]
5th order control
1 1 1
Error θ̇ [deg/s]
Error α [deg]
Error θ [deg]
0.5 0.5 0
0 0 -1
-0.5 -0.5 -2
0 5 10 15 0 5 10 15 0 5 10 15
Time [s] Time [s] Time [s]
FIGURE 5.8 – State estimation error for the aircraft with different control strategies.
Initial condition α = 25 deg
FIGURE 5.9 – Parameter estimation for the aircraft with different control strategies.
Initial condition α = 25 deg
CHAPTER 5. NUMERICAL SIMULATIONS 77
Lastly, we check the the stability criterion for the closed-loop system with each con-
troller. As discussed in Sections 4.2.1 and 4.2.2, the origin of the system is asymptotically
stable in probability if the function L, in Eq. 4.28 is positive definite. Figure 5.10 shows
the function for each control strategy. In all cases, L is positive during all time instants
and therefore we infer the stability in probability of the equilibrium point in the neighbor-
hood that contains the initial condition. One should note that we verify the stability for
the closed-loop stochastic system by analyzing a deterministic function for the gPC mode
strengths with low computational cost when compared to traditional methods based on
sampling.
0.03 0.03
0.025 0.025
0.02 0.02
L
0.015 0.015
0.01 0.01
0.005 0.005
0 0
0 5 10 15 0 5 10 15
Time [s] Time [s]
0.06
0.06
0.05
0.04
L
0.04
0.03
0.02
0.02
0.01
0 0
0 5 10 15 0 5 10 15
Time [s] Time [s]
FIGURE 5.10 – Stability criterion verification for the aircraft with different control strate-
gies. Initial condition α = 25 deg
controller. In contrast to the previous case, now the aircraft was successfully recovered
from stall and reached the trim condition only for the 3rd and 5th order controllers. The
5th order controller provided a slight faster response in bringing the vehicle away from the
stall condition. On the other hand, the linear and 2nd order controllers were not capable
of driving the states to the trim condition. Apparently, for these control laws, the states
approached another equilibrium point, but still within the stall region, at a very high
angle of attack, thus not being feasible in an actual aircraft mission. From the pitch angle
graphs, we again see that the aircraft loses altitude during the stall recover maneuver.
The pitch rate obtained for the nonlinear controllers was significantly higher than for the
low order control laws, reaching a peak of approximately -40 deg/s at the very beginning
of the simulation. We conclude, therefore, that the linear and 2nd order controllers are
not suitable for stall recovering maneuvers at very high angles of attack.
FIGURE 5.11 – Angle of attack of the aircraft with different control strategies. Initial
condition α = 30 deg
Figure 5.13 shows the control inputs computed for each controller. Now we see that
the elevon deflection for the vehicle with the linear and 2nd order control laws does not
reach the trim condition, since the aircraft is driven to another point of equilibrium inside
the stall region. The 3rd and 5th order nonlinear controllers are slightly more aggressive
than the previous case, still reaching the saturation limits at ±30 deg. Even so, we deem
that these two controllers provided adequate dynamic response for the aircraft.
CHAPTER 5. NUMERICAL SIMULATIONS 79
Linear control
20 20
θ̇(t) [deg/s]
θ(t) [deg]
10 10
0 0
-10 -10
0 5 0 5 10 15
Time [s] 10 15
Time [s]
2nd order control
20 20
θ̇(t) [deg/s]
θ(t) [deg]
10 0
0 -20
-10 -40
0 5 10 15 0 5 10 15
Time [s]
Time [s]
3rd order control
0 20
θ̇(t) [deg/s]
θ(t) [deg]
-5 0
-10 -20
-15 -40
0 5 10 15 0 5 10 15
Time [s]
Time [s]
5th order control
0 50
θ̇(t) [deg/s]
θ(t) [deg]
-5 0
-10 -50
-15 -100
0 5 10 15 0 5 10 15
Time [s] Time [s]
FIGURE 5.12 – Pitch angle and pitch rate of the aircraft with different control strategies.
Initial condition α = 30 deg
FIGURE 5.13 – Elevon deflection of the aircraft with different control strategies. Initial
condition α = 30 deg
CHAPTER 5. NUMERICAL SIMULATIONS 80
Linear control
Error θ̇ [deg/s]
1 1 1
Error θ [deg]
Error α [deg]
0.5 0.5 0
0 0 -1
-0.5 -2
-0.5 0 5 10 15
0 5 10 15 0 5 10 15
Time [s] Time [s] Time [s]
1 2nd order control 1 2
Error θ̇ [deg/s]
Error α [deg]
Error θ [deg]
0.5 0.5
0
0 0
-0.5 -0.5 -2
0 5 10 15 0 5 10 15 0 5 10 15
Time [s] Time [s] Time [s]
1
3rd order control 1 1
Error θ̇ [deg/s]
Error θ [deg]
Error α [deg]
0.5 0.5 0
0 0 -1
-0.5 -0.5 -2
0 5 10 15 0 5 10 15 0 5 10 15
Time [s] Time [s] Time [s]
Error θ̇ [deg/s]
1 5th order control 1 1
Error θ [deg]
Error α [deg]
0.5 0.5 0
0 0 -1
-0.5 -0.5 -2
0 5 10 15 0 5 10 15 0 5 10 15
Time [s] Time [s] Time [s]
FIGURE 5.14 – State estimation error for the aircraft with different control strategies.
Initial condition α = 30 deg
FIGURE 5.15 – Parameter estimation for the aircraft with different control strategies.
Initial condition α = 30 deg
CHAPTER 5. NUMERICAL SIMULATIONS 81
Figures 5.14 and 5.15 show, respectively, the state estimation error and the parameter
estimation provided by the adaptive UKF for each control strategy. Similar to the previous
case, all state estimation errors quickly approach zero. As for the parameter estimation
ξ, we see that the estimation provided a much better approximation in comparison to
the previous case. The parameter estimation for the system with the linear, 2nd and 5th
order controllers converged to the real value of the parameter, while the estimation for
the system with the 3rd order control provided a reasonable approximation. As discussed
before, the improvement in the parameter estimation results may be due to the selection
of an initial condition for which the influence of the uncertainty in the dynamics was more
significant, thus providing information for estimating ξ more precisely.
Figure 5.16 shows the stability criterion verification for the aircraft with each con-
troller. As expected, the L function for the linear and 2nd order controls are not positive
definite, therefore the trim condition is not asymptotically stable in probability in the
neighborhood of the initial condition considered. The results for the system with 3rd and
5th order controllers are similar to the previous case.
0 0
-0.1 -0.1
L
-0.2 -0.2
-0.3 -0.3
-0.4 -0.4
0 5 10 15 0 5 10 15
Time [s] Time [s]
0.08
0.1
0.06
L
0.04
0.05
0.02
0 0
0 5 10 15 0 5 10 15
Time [s] Time [s]
FIGURE 5.16 – Stability criterion verification for the aircraft with different control strate-
gies. Initial condition α = 30 deg
CHAPTER 5. NUMERICAL SIMULATIONS 82
FIGURE 5.17 – Angle of attack of the aircraft with different control strategies. Initial
condition α = 40 deg
Figure 5.19 shows the control inputs computed for each controller. The results are
similar to the last simulation. The 2nd and 5th order nonlinear controllers are more
aggressive than the previous case, reaching the saturation limits at ±30 deg for a longer
period of time at the beginning of the simulation.
CHAPTER 5. NUMERICAL SIMULATIONS 83
Linear control
20 20
θ̇(t) [deg/s]
θ(t) [deg]
0 0
-20 -20
0 5 10 15 0 5 10 15
Time [s] Time [s]
2nd order control
20 20
θ̇(t) [deg/s]
θ(t) [deg]
0 0
-20 -20
0 5 10 15 0 5 10 15
Time [s] Time [s]
3rd order control
0 100
θ̇(t) [deg/s]
θ(t) [deg]
-20 0
-40 -100
0 5 10 15 0 5 10 15
Time [s] Time [s]
5th order control
0 200
θ̇(t) [deg/s]
θ(t) [deg]
-20 0
-40 -200
0 5 10 15 0 5 10 15
Time [s] Time [s]
FIGURE 5.18 – Pitch angle and pitch rate of the aircraft with different control strategies.
Initial condition α = 40 deg
FIGURE 5.19 – Elevon deflection of the aircraft with different control strategies. Initial
condition α = 40 deg
CHAPTER 5. NUMERICAL SIMULATIONS 84
Linear control
1 0.4 2
Error θ̇ [deg/s]
Error θ [deg]
Error α [deg]
0.5 0.2
0
0 0
-0.5 -0.2 -2
0 5 10 15 0 5 10 15 0 5 10 15
Time [s] Time [s] Time [s]
1 2nd order control 0.5 1
Error θ̇ [deg/s]
Error α [deg]
Error θ [deg]
0.5 0
0
0 -1
-0.5 -0.5 -2
0 5 10 15 0 5 10 15 0 5 10 15
Time [s] Time [s] Time [s]
Error θ̇ [deg/s]
1 3rd order control 1 1
Error θ [deg]
Error α [deg]
0.5 0
0
0 -1
-0.5 -1 -2
0 5 10 15 0 5 10 15 0 5 10 15
Time [s] Time [s] Time [s]
1 5th order control 1 2
Error θ̇ [deg/s]
Error α [deg]
Error θ [deg]
0.5 0.5
0
0 0
-0.5 -0.5 -2
0 5 10 15 0 5 10 15 0 5 10 15
Time [s] Time [s] Time [s]
FIGURE 5.20 – State estimation error for the aircraft with different control strategies.
Initial condition α = 40 deg
FIGURE 5.21 – Parameter estimation for the aircraft with different control strategies.
Initial condition α = 40 deg
CHAPTER 5. NUMERICAL SIMULATIONS 85
Figures 5.20 and 5.21 show, respectively, the state estimation error and the parameter
estimation for each control strategy. Similar to the previous cases, all state estimation
errors approach zero. Regarding the parameter estimation ξ, we see that for all control
strategies the adaptive UKF provided very good estimates.
Figure 5.22 shows the stability criterion verification for the aircraft with each con-
troller. As expected, the L function for the linear and 2nd order controls are not positive
definite. The results for the system with 3rd and 5th order controllers are similar to the
previous cases. The conclusion is that, for the control method presented in Chapter 4,
only third or higher order control laws in the closed-loop system for the aircraft longitu-
dinal dynamics guarantee that the trim position is asymptotically stable in probability.
0 0
-0.1 -0.1
L
-0.2 -0.2
-0.3 -0.3
-0.4 -0.4
-0.5 -0.5
0 5 10 15 0 5 10 15
Time [s] Time [s]
0.5
0.2
0.4
0.15
L
0.3
0.1
0.2
0.05
0.1
0 0
0 5 10 15 0 5 10 15
Time [s] Time [s]
FIGURE 5.22 – Stability criterion verification for the aircraft with different control strate-
gies. Initial condition α = 40 deg
6 Conclusions and Future Work
6.1 Conclusions
the fundamental objective of the FADEMO project in exploring new technologies for the
aeronautical engineering field.
As for the Control Theory area, this work presented a systematic method for studying
the stability and robustness of closed loop nonlinear systems with stochastic paramet-
ric uncertainty. By using the polynomial chaos theory, we can take into account the a
priori knowledge of the distribution of the random parameters and use this information
for control design purposes. The theorem presented in Section 4.1 sets up the sufficient
conditions for the weak stochastic stability of such stochastic nonlinear systems. Further-
more, we can use the polynomial chaos expansions to propagate the uncertainty through
the closed loop system and verify all possible scenarios with a low computational cost
when compared with methods based on sampling. Therefore, the proposed method has
clear advantages over classic methods for being less conservative and ensuring optimality,
stability and robustness for nonlinear systems.
As future work, we might consider the implementation and validation of the proposed
control system. Additionally, we can work to improve the computational cost of the poly-
nomial chaos-based controller, either in the order of expansion to have a better accuracy,
or in the implementation to estimate the mode strengths in real time. Some topics of
research that might be worthwhile delving into are the following:
1. The full aircraft model with coupled longitudinal and lateral dynamics. In a typ-
ical mission, fighters are required to perform advanced maneuvers in all degrees of
freedom. Therefore, the control design has to consider not only the longitudinal
dynamics, but also the lateral-directional motion in order to guarantee satisfactory
dynamic response. Furthermore, the control system has to mitigate the effect of the
aircraft dynamical modes, such as dutch roll, and bifurcations.
3. Explore other approaches for estimating the gPC mode strengths. As discussed
in Section 4.3, depending on the order of the gPC expansion, high order moment
CHAPTER 6. CONCLUSIONS AND FUTURE WORK 88
estimates are required for computing the mode strengths in real time, which can
be computationally costly. Other approaches for estimating the gPC coefficients
are possible. For example, methods based on sampling can be used in a reinforce-
ment learning fashion, that is, the repetition of a specific set of tasks improves the
performance of the controller by acquiring data about the system.
ABED, E. H.; FU, J. H. Local feedback stabilization and bifurcation control, i. hopf
bifurcation. Systems & Control Letters, v. 7, n. 1, p. 11–17, fev. 1986.
HUI, W. H.; TOBAK, M. Bifurcation analysis of aircraft pitching motions about large
mean angles of attack. Journal of Guidance, Control, and Dynamics, v. 7, n. 1, p.
113–122, 1984.
KO, J.; LUCOR, D.; SAGAUT, P. Effects of base flow uncertainty on couette flow
stability. Computers & Fluids, v. 43, n. 1, p. 82–89, 2011.
BIBLIOGRAPHY 92
LEE, H.-C.; ABED, E. H. Washout filters in the bifurcation control of high alpha flight
dynamics. In: IEEE. American Control Conference (ACC). [S.l.], 1991. p. 206–211.
LIU, J.; WANG, Y.; ZHANG, J. A linear extension of unscented kalman filter to
higher-order moment-matching. In: IEEE 53rd Annual Conference on Decision
and Control (CDC). Los Angeles, USA: IEEE, 2014.
MAHMOOD, A.; KIM, Y.; PARK, J. Robust h∞ autopilot design for agile missile with
time-varying parameters. IEEE Transactions on Aerospace and Electronic
Systems, v. 50, n. 4, p. 3082 – 3089, 2014.
MAJJI, M.; JUNKINS, J. L.; TURNER, J. D. A high order method for estiamtion of
dynamic systems. The Journal of the Astronautical Sciences, v. 56, n. 3, p.
401–440, 2008.
MESBAH, A.; STREIF, S.; FINDEISEN, R.; BRAATZ, R. Stochastic nonlinear model
predictive control with probabilistic constraints. In: IEEE (Ed.). American Control
Conference (ACC). Portland, USA: [s.n.], 2014.
NASA. Blended wing body - a potential new aircraft design. Langley Research Center.
December 2016. Disponı́vel em:
<https://www.nasa.gov/centers/langley/news/factsheets/FS-2003-11-81-LaRC.html>.
NECHAK, L.; BERGER, S.; AUBRY, E. Non-intrusive generalized polynomial chaos for
the robust stability analysis of uncertain nonlinear dynamic friction systems. Journal
of Sound and Vibration, v. 332, n. 5, p. 1204–1215, 2013.
PENG, Y. B.; GHANEM, R.; LI, J. Polynomial chaos expansions for optimal control of
nonlinear random oscillators. Journal of Sound and Vibration, v. 329, n. 18, p.
3660–3678, 2010.
PONOMAREVA, K.; DATE, P.; WANG, Z. A new unscented kalman filter with higher
order moment-matching. In: BUDAPEST, HUNGARY. 19th International
Symposium on Mathematical Theory of Networks and Systems. [S.l.], 2010.
SRIRAM; JAMESON, A. Robust optimal control using polynomial chaos and adjoints
for systems with uncertain inputs. In: AIAA (Ed.). 20th AIAA Computational
Fluid Dynamics Conference. Honolulu, USA: [s.n.], 2011.
STEVENS, B. L.; LEWIS, F. L. Aircraft Control and Simulation. 2nd. ed. [S.l.]:
John Wiley & Sons, Inc., 2003.
TOL, H. J.; VISSER, C. C. de; SUN, L. G.; KAMPEN, E. van; CHU, Q. P. Multivariate
spline-based adaptive control of high-performance aircraft with aerodynamic
uncertainties. Journal of Guidance and Control, v. 39, n. 4, p. 781–800, jan. 2016.
TUSSET, A. M.; PICCIRILLO, V.; BUENO, A. M.; BALTHAZAR, J. M.; SADO, D.;
FELIX, J. L. P.; BRASIL, R. M. L. R. F. Chaos control and sensitivity analysis of a
double pendulum arm excited by an rlc circuit based nonlinear shaker. Journal of
Vibration and Control, v. 22, n. 17, p. 3621–3637, 2016.
ZHENG, F.; XU, J. Angle of attack control of a blended wing body aircraft using state
feedback l1 adaptive controller. In: Chinese Automation Congress. Changsha,
China: IEEE, 2013.
A.1 Definitions
Here we present some definitions and facts about orthogonal polynomials. To begin
with, let us define orthogonality between functions (SZEGO, 1939).
Definition: Let γ(x) be a fixed non-decreasing function which is not constant in the
interval x ∈ [a, b]. An orthonormal set of functions {φ0 (x), φ1 (x), . . . , φn (x), . . . , φs (x)},
with s finite or infinite and φn (x) real valued and in the class L2γ (a, b), is defined by:
b
φn , φm = φn φm dγ(x) = δn,m n, m = 0, 1, 2, · · · , s. (A.1)
a
Theorem: Consider the following real-valued functions of the class L2γ (a, b) and linearly
independent.
f0 (x), f1 (x), f2 (x), · · · , fs (x) s finite or infinite
Definition: Consider again the non-decreasing function γ(x), but now with infinitely
many points of increase in the finite or infinite interval x ∈ [a, b]. Let the moments
b
cn = xn dγ(x) n = 0, 1, 2, · · ·
a
1, x, x2 , · · · , xn , · · · ,
In the case dγ(x) is a distribution of the type w(x)dx, being w(x) a non-negative function,
!b
measurable in the Lebesgue’s sense and that a w(x)dx > 0, then we call ψn (x) the
orthogonal polynomials associated with the weight function w(x).
Orthogonal polynomials usually can be written in a recurrence relation of the form
(ABRAMOWITZ; STEGUN, 1965):
where an , bn and cn are appropriate constants. We assume that, for all orthogonal,
polynomials, the degree zero polynomial ψ0 (x) is equal to one for all x ∈ [a, b]. Therefore,
b b
ψ0 (x)w(x)dx = w(x)dx (A.4)
a a
APPENDIX A. ORTHOGONAL POLYNOMIALS 98
and b
ψn (x)w(x)dx = 0 (A.5)
a
w(x)
f (x) = ! . (A.6)
I
w(x)dx
where E[·] denotes the expectation function. We can show using Eq. A.4 and A.5 that,
for the mean and variance, the first and second moments of ψn (X) respectively, we have
(BAUDIN, 2015):
1, for n = 1
μ1 = E[ψn (X)] = , (A.8)
0, otherwise
and
0, for n = 1
μ2 = V [ψn (X)] = E[ψn2 (X)] = 2 , (A.9)
||ψn || , otherwise
I w(x)dx
where ||ψn || = ψn , ψn is the L2w (I) norm. Using the orthogonality property, it is also
straightforward to verify that
p
x) =
Φ(x βα xαi i , (A.11)
|α
α|≤d i=1
for any x ∈ Rp .
As shown by Baudin (2015), the number of degree d multivariate polynomilas of p
APPENDIX A. ORTHOGONAL POLYNOMIALS 99
variables is
p+d
P =
d
Naturally, multivariate orthogonal polynomials associated with multivariate weight func-
tions can also be defined (BAUDIN, 2015).
x 1 ∈ I 1 , x2 ∈ I 2 , . . . , x p ∈ I p ,
I = I1 ⊗ I2 ⊗ . . . ⊗ Ip .
Assume that w1 (x1 ), w2 (x2 ), . . . , wp (xp ) are univariate weights on I1 , I2 , . . . , Ip , each being
!
a non-negative function, measurable in the Lebesgue’s sense and that Ii wi (xi )dxi > 0
for i = 1, 2, . . . , p. Then a weight function on I is a non-negative function of x ∈ I given
by the tensor product function
for any x ∈ Rp . Now, considering the tensor of product interval I of Rp and the associ-
ated multivariate tensor product weight function w(x x), the associated multivariate tensor
product polynomials are:
p
x) =
Ψk (x ψα(k) (xi ) (A.12)
i
i=1
where ψα(k) (xi ) are a family of univariate orthogonal polynomials for k = 1, 2, . . . , P . The
i
degree of Ψk is
P
d= αik . (A.13)
i=1
if m = n. Note, however, that even though this set is orthogonal, it is not necessarily
orthonormal.
APPENDIX A. ORTHOGONAL POLYNOMIALS 100
Now, if we consider that {Xi }i=1,2,...,p are independent random variables, then the
probability distribution function, derived from the multivariate weight function on I ∈ Rp ,
is given by
x) = f1 (x1 )f2 (x2 ) · · · fp (xp ).
f (x (A.15)
Hence, we can now define the p-th moment of Ψk (X X ), where X is a multivariate random
variable associated with the multivariate orthogonal polynomials {Ψk }k≥1 :
μp = E[Ψpk (x
x)] = Ψpk (x
x)f (x
x)dx
x. (A.16)
I
and
0, for k = 1
X )] = E[Ψ2k (X
μ2 = V [Ψk (X X )] = #p 2
. (A.18)
i=1 E[ψ α
(k) (Xi ) ], otherwise
i
X ), Ψn (X
Cov(Ψm (X X )) = 0 for m = n. (A.19)
for n ∈ N+ . Table A.1 shows the first eight Hermite polynomials. Hermite polynomials
are associated with the Gaussian weight
x2
w(x) = exp − , (A.23)
2
APPENDIX A. ORTHOGONAL POLYNOMIALS 101
for x ∈ R, thus corresponding to a Gaussian distribution with zero mean and unit variance.
He0 = 1
He1 = x
He2 = x2 − 1
He3 = x3 − 3x
He4 = x4 − 6x2 + 3
He5 = x5 − 10x3 + 15x
He6 = x6 − 15x4 + 45x2 − 15
He7 = x7 − 21x5 + 105x3 − 105x
for n ≥ 1.
P0 (x) = 1 (A.27)
P1 (x) = x (A.28)
(n + 1)Pn+1 (x) = (2n + 1)xPn (x) − nPn−1 (x) (A.29)
for n ∈ N+ . Table A.2 shows the first eight Legendre polynomials. Legendre polynomials
are associated with the unit weight
w(x) = 1, (A.30)
APPENDIX A. ORTHOGONAL POLYNOMIALS 102
1
f (x) = , (A.31)
2
for x ∈ [−1, 1], which corresponds to the uniform distribution in probability theory.
P0 = 1
P1 = x
P2 = 12 (3x2 − 1)
P3 = 12 (5x3 − 3x)
P4 = 18 (35x4 − 30x2 + 3)
P5 = 18 (63x5 − 70x3 + 15x)
1
P6 = 16 (231x6 − 315x4 + 105x2 − 5)
1
P7 = 16 (429x7 − 693x5 + 315x3 − 35x)
2
Pm , Pn = δmn . (A.32)
2n + 1
for n ≥ 1.Moreover, if the argument is a random variable uniform in interval [−1, 1], then:
1
V (Pn (X)) = , (A.33)
2n + 1
for n ≥ 1.
L0 (x) = 1 (A.34)
L1 (x) = −x + 1 (A.35)
(n + 1)Ln+1 (x) = (2n + 1 − x)Ln (x) − nLn−1 (x) (A.36)
for n ∈ N+ . Table A.3 shows the first eight Laguerre polynomials. Laguerre polynomials
are associated with the weight
w(x) = exp(−x), (A.37)
APPENDIX A. ORTHOGONAL POLYNOMIALS 103
for x ∈ [0, ∞], which corresponds to the standard exponential distribution in probability
theory.
L0 = 1
L1 = −x + 1
L2 = 12 (x2 − 4x + 2)
L3 = 16 (−x3 + 9x2 − 18x + 6)
1
L4 = 24 (x4 − 16x3 + 72x2 − 96x + 24)
1
L5 = 120 (−x5 + 25x4 − 200x3 + 600x2 − 600x + 120)
1
L6 = 720 (x6 − 36x5 + 450x4 − 2400x3 + 5400x2 − 4320x + 720)
1
L7 = 5040 (−x7 + 49x6 − 882x5 + 7350x4 − 29400x3 + 52920x2 − 35280x + 5040)
The inner product between two Laguerre polynomials is equal to (BAUDIN, 2015):
for n ≥ 1.
Appendix B - Cameron-Martin
Theorem
Polynomial chaos expansions are originally based on the work of Wiener (1938), which
was formally presented by Cameron e Martin (1947) as an explicit discretization of the
white noise process through Fourier expansion as outlined below. Here we follow the
notation in the original publication.
Let C be the space of real functions x(t) which are continuous on the interval 0 ≤ t ≤ 1
and which vanish at t = 0. Consider a partially normalized Hermite polynomial given by
the Rodrigue’s formula:
2 dn −u2
Hen (u) = (−1)n 2−n/2 (n!)−1/2 eu (e ) for n = 0, 1, 2, . . . . (B.1)
dun
2 /2
The set {π −1/4 Hen (u)e−u } is a complete orthonormal set on u ∈ (−∞, ∞). Let
be any complete orthonormal set of real functions, each belonging to L2 (0, 1). Define for
almost all x ∈ C
1
Φm,p (x) = Hen αp (t)dx(t) for p = 1, 2, 3, . . . ; m = 0, 1, 2, . . . (B.3)
0
The Fourier-Hermite series of F [x] converges in the L2 (C) sense to F [x], that is,
$ $2
w $ N $
$ $
$ F [x] − Am1 ,··· ,mN Ψm1 ,··· ,mN (x)$ dw x → 0 as N → ∞, (B.6)
C $ m ,··· ,m =0
$
1 N
Proof : See Cameron e Martin (1947) for complete proof. Outline: From the work
of (PALEY; WIENER, 1934), we know that, for each index p = 1, 2, 3, . . ., the integral
!1
0
αp (t)dx(t) exists as a generalizes Stieltjes integral for almost all functions x(·) of C
and that the equality
w 1 1
G α1 (t)dx(t), · · · , αp (t)dx(t) dw x
0 0
C
∞ ∞
−p/2 2 2
=π ··· G(u1 , . . . , up )e−u1 ···up du1 · · · dup .
−∞ −∞
holds for every function G(u1 , . . . , up ) for which the integral on the right exists as an abso-
lute convergent Lebesgue integral. Prove that the set Ψm1 ,··· ,mp is orthonormal. Consider
the Bessel inequality and the best approximation theorem for the set. Consider also the
function
2 2
f (u1 , · · · , un )e−(u1 +···+un )/2 ∈ L2 (−∞, ∞)
Then prove that, for any non-negative indices m1 , . . . , mp , the following result holds:
w
0, if n < p and mp = 0
F ∗ [x]Ψm1 ,··· ,mp [x]dw x =
C fm1 ,··· ,mn if p = n,
where fm1 ,··· ,mn is the ordinary n-dimensional Hermite coefficient given by:
∞ ∞
n
−n/2 −(u21 +···+u2n )/2 2
fm1 ,··· ,mn = π ··· fm1 ,··· ,mn e [Hemi (ui )e−ui /2 dui ].
−∞ −∞ j=1
Now, prove the aforestated Cameron-Martin theorem for F ∗ [x], with Fourier-Hermite
coefficients A∗m1 ,...,mn , by considering the functional F [x] ∈ L2 (C), with Fourier-Hermite
APPENDIX B. CAMERON-MARTIN THEOREM 106
$ $2
w $ N $
$ ∗ ∗ 2 $
$F [x] − Am1 ,··· ,mN Ψm1 ,··· ,mN (x)| dw x$ dw x <
/4,
C $ m ,··· ,m =0
$
1 N
$ $2
w $ N $
$ $
$ F [x] − A∗m1 ,··· ,mN Ψm1 ,··· ,mN (x)|2 dw x$ dw x <
.
C $ m ,··· ,m =0
$
1 N
Then by the best approximation theorem it follows that the last inequality will remain
true if we replace A∗ by A. To complete the proof, show that any functional of L2 (C)
can be approached in the L2 (C) sense by functionals satisfying the hypotheses made for
F ∗ [x], i.e., these functionals are everywhere dense in L2 (C).
Appendix C - Unscented Kalman
Filter
In this Appendix, we present the Unscented Kalman Filter for state estimation of
nonlinear systems. Even though the Extended Kalman Filter is well established in the
estimation and control community as a computationally efficient extension of Kalman
filtering for nonlinear estimation, a major downside is apparent when it comes to the
application in highly nonlinear systems that are poorly approximated by linear functions
along their trajectories. As discussed by Julier e Uhlmann (2004), when linearization
fails to capture the error propagation, the performance of the filter is undermined and
the estimates can possibly diverge altogether. In some cases, the linearization is not even
possible due to the discontinuities or singularities inherent to the dynamical system. In
addition, the linearization procedure requires the calculation of Jacobian matrices, which
can be a very difficult and error-prone process. In view of these hurdles, the UKF is
presented as an alternative to EKF. The UKF algorith computes the mean and covariance
through the so called unscented transformation (UT), which uses a set of sampling points,
named σ-points, to gather information about these moments and propagate the probability
distribution by applying the nonlinear function to each point.
Consider a general continuous nonlinear system of the form
x(t) = f (x
ẋ x(t), u (t)) + Gω (t) (C.1)
xk+1 ) + ν k+1 ,
y k+1 = c k+1 (x (C.2)
where x ∈ Rn , y ∈ Rs , G ∈ Rn×nw , and the signal {ω ω (t)} ∈ Rnw and the sequence
{νν k } ∈ Rs are noise terms, which are assumed to be realizations of the zero mean stochastic
processes {W W (t)} and {N N k }, respectively, such that W (t) ∼ (00, Qω (t)) and N k ∼ (00, Rν k ),
where Qω (t) ∈ Rnw ×nw and Rν k ∈ Rs×s are known covariance matrices. The initial
condition of C.1, x 0 , is an outcome of the random variable X 0 ∼ (x̄ x, P̄ x ∈ Rn and
P ), where x̄
APPENDIX C. UNSCENTED KALMAN FILTER 108
Similar to the Kalman Filter, the mean and covariance propagation are divided into two
parts. Note that, in contrast to the EKF, the moments in the prediction stage are obtained
by statistics of the set of transformed σ-points.
Prediction:
2na
x̂ k+1|k ≈ ρiX ik+1|k Predicted estimate (C.6)
i=0
2na
P k+1|k ≈ ρi (XXik+1|k − x̂ k+1|k )(XXik+1|k − x̂ k+1|k )T Cod. cov. pred. estimate (C.7)
i=0
2na
ŷ k+1|k ≈ ρiY ik+1|k Predicted measurement (C.8)
i=0
2na
P Yk+1|k ≈ ρi (YYik+1|k − ŷ k+1|k )(YYik+1|k − ŷ k+1|k )T Cond. cov. pred. measurement(C.9)
i=0
2n a
k+1|k ≈
P XY ρi (XXik+1|k − x̂ k+1|k )(YYik+1|k − ŷ k+1|k )T Conditional cross covariance (C.10)
i=0
Update:
Refer to Julier e Uhlmann (2004) for details of the implementation of the UKF algo-
rithm.
Appendix D - Control Laws
Here we present the control laws derived according to the methods shown in Chapter
4. All control laws were derived for a first order (P = 1) gPC expansion. Only the more
significant terms were kept in each expression.
Nonlinear control computed according to the method presented in Section 4.2.2 with
n = 1.
u1,0
U2 = (D.2)
u1,1
where
and
Nonlinear control computed according to the method presented in Section 4.2.2 with
n = 2.
u1,0
U3 = (D.5)
u1,1
where
u1,0 = 3.58x31,0 − 0.529x21,0 x1,1 − 3.32x21,0 x2,0 + 0.183x21,0 x2,1 + 0.0854x21,0 x3,0
−0.0012x21,0 x3,1 + 0.0106x21,0 + 10.8x1,0 x21,1 + 0.366x1,0 x1,1 x2,0
−6.65x1,0 x1,1 x2,1 − 0.0024x1,0 x1,1 x3,0 + 0.171x1,0 x1,1 x3,1 + 0.836x1,0 x22,0
−0.0923x1,0 x2,0 x2,1 − 0.00982x1,0 x2,0 x3,0 + 0.0766x1,0 x2,0 + 0.836x1,0 x22,1
−0.0575x1,0 x2,1 x3,1 + 0.00125x1,0 x3,0 − 0.232x1,0 − 0.529x31,1 − 3.32x21,1 x2,0
+0.55x21,1 x2,1 + 0.0854x21,1 x3,0 − 0.0036x21,1 x3,1 + 0.0106x21,1 (D.6)
−0.0735x1,1 x22,0 + 1.67x1,1 x2,0 x2,1 − 0.0575x1,1 x2,0 x3,1
−0.138x1,1 x22,1 − 0.00982x1,1 x2,1 x3,0 + 0.00168x1,1 x2,1 x3,1 + 0.0766x1,1 x2,1
+0.00125x1,1 x3,1 + 0.00757x32,0 − 0.00249x22,0 x3,0
+0.0196x22,0 + 0.0227x2,0 x22,1 + 0.0234x2,0 x2,1 x3,1 + 1.x2,0
−0.00249x22,1 x3,0 + 0.0196x22,1 + 1.01x3,0
and
u1,1 = −0.176x31,0 + 10.8x21,0 x1,1 + 0.183x21,0 x2,0 − 3.32x21,0 x2,1 − 0.0012x21,0 x3,0
+0.0854x21,0 x3,1 − 1.59x1,0 x21,1 − 6.65x1,0 x1,1 x2,0
+1.1x1,0 x1,1 x2,1 + 0.171x1,0 x1,1 x3,0 − 0.0072x1,0 x1,1 x3,1
+0.0211x1,0 x1,1 − 0.0735x1,0 x22,0 + 1.67x1,0 x2,0 x2,1 − 0.00982x1,0 x2,0 x3,1
−0.138x1,0 x22,1 − 0.0575x1,0 x2,1 x3,0 + 0.00168x1,0 x2,1 x3,1
+0.0766x1,0 x2,1 + 0.00125x1,0 x3,1 + 10.7x31,1 + 0.55x21,1 x2,0
(D.7)
−9.88x21,1 x2,1 − 0.0036x21,1 x3,0 + 0.254x21,1 x3,1 + 0.836x1,1 x22,0 − 0.277x1,1 x2,0 x2,1
−0.0575x1,1 x2,0 x3,0 + 0.00168x1,1 x2,0 x3,1 + 0.0766x1,1 x2,0
+2.49x1,1 x22,1 + 0.00168x1,1 x2,1 x3,0 − 0.171x1,1 x2,1 x3,1
+0.00125x1,1 x3,0 − 0.232x1,1 + 0.0205x32,0 + 0.0227x22,0 x2,1 − 0.00249x22,0 x3,1
+0.0234x2,0 x2,1 x3,0 + 0.0391x2,0 x2,1 + 0.00757x32,1
−0.00693x22,1 x3,1 + 1.x2,1 + 1.01x3,1
APPENDIX D. CONTROL LAWS 111
Nonlinear control computed according to the method presented in Section 4.2.2 with
n = 4.
u1,0
U5 = (D.8)
u1,1
where
u1,0 = 10.6x51,0 − 6.67x1,1 x41,0 − 21.x2,0 x41,0 + 2.5x2,1 x41,0 + 0.314x3,0 x41,0
−0.0158x3,1 x41,0 − 1.17x41,0 + 106.x21,1 x31,0 + 18.7x22,0 x31,0
+18.7x22,1 x31,0 + 0.00177x23,0 x31,0 + 0.00179x23,1 x31,0 + 0.4x1,1 x31,0
+10.1x1,1 x2,0 x31,0 + 2.15x2,0 x31,0 − 84.x1,1 x2,1 x31,0 − 4.32x2,0 x2,1 x31,0 − 0.158x2,1 x31,0
−0.165x1,1 x3,0 x31,0 − 0.425x2,0 x3,0 x31,0 + 0.0191x2,1 x3,0 x31,0 − 0.00728x3,0 x31,0
+0.0546x2,0 x3,1 x31,0 − 0.42x2,1 x3,1 x31,0 + 3.58x31,0 − 23.2x31,1 x21,0 − 8.3x32,0 x21,0
+2.27x32,1 x21,0 − 7.03x21,1 x21,0 − 6.55x1,1 x22,0 x21,0 − 1.49x22,0 x21,0 − 13.4x1,1 x22,1 x21,0
−24.9x2,0 x22,1 x21,0 − 1.49x22,1 x21,0 − 0.00187x2,0 x23,0 x21,0 − 0.00121x1,1 x23,1 x21,0
−0.00189x2,0 x23,1 x21,0 − 0.529x1,1 x21,0 − 126.x21,1 x2,0 x21,0 − 0.481x1,1 x2,0 x21,0
−3.32x2,0 x21,0 + 28.3x21,1 x2,1 x21,0 + 2.88x22,0 x2,1 x21,0 + 6.45x1,1 x2,1 x21,0
+112.x1,1 x2,0 x2,1 x21,0 + 0.2x2,0 x2,1 x21,0 + 0.183x2,1 x21,0
+1.9x21,1 x3,0 x21,0 + 0.29x22,0 x3,0 x21,0 + 0.286x22,1 x3,0 x21,0
+0.00234x1,1 x3,0 x21,0 + 0.161x1,1 x2,0 x3,0 x21,0 + 0.00431x2,0 x3,0 x21,0 (D.9)
−1.26x1,1 x2,1 x3,0 x21,0 − 0.016x2,0 x2,1 x3,0 x21,0 + 0.0854x3,0 x21,0
−0.402x21,1 x3,1 x21,0 − 0.00725x22,0 x3,1 x21,0 − 0.054x22,1 x3,1 x21,0
−0.067x1,1 x3,1 x21,0 − 1.28x1,1 x2,0 x3,1 x21,0 + 0.258x1,1 x2,1 x3,1 x21,0
+0.572x2,0 x2,1 x3,1 x21,0 + 0.00425x2,1 x3,1 x21,0 + 0.0107x1,1 x3,0 x3,1 x21,0
−0.00369x2,1 x3,0 x3,1 x21,0 − 0.0012x3,1 x21,0 + 0.0106x21,0
+55.7x41,1 x1,0 + 1.29x42,0 x1,0 + 4.12x42,1 x1,0 + 0.896x31,1 x1,0
+1.92x1,1 x32,0 x1,0 + 0.259x32,0 x1,0 − 39.8x1,1 x32,1 x1,0 − 1.29x2,0 x32,1 x1,0
−0.0138x32,1 x1,0 + 10.8x21,1 x1,0 + 56.1x21,1 x22,0 x1,0
+0.202x1,1 x22,0 x1,0 + 0.836x22,0 x1,0 + 105.x21,1 x22,1 x1,0 + 7.7x22,0 x22,1 x1,0
+0.313x1,1 x22,1 x1,0 + 11.x1,1 x2,0 x22,1 x1,0 + 0.776x2,0 x22,1 x1,0 + 0.836x22,1 x1,0
+0.00534x21,1 x23,0 x1,0 + 0.00105x22,0 x23,0 x1,0 + 0.00103x22,1 x23,0 x1,0
−0.0037x1,1 x2,1 x23,0 x1,0 + 1.26x1,1 x3,1 x31,0 + (· · · )
APPENDIX D. CONTROL LAWS 112
(· · · ) + 0.521x21,1 x2,0 x2,1 + 0.005x21,1 x23,1 x1,0 + 0.00105x22,0 x23,1 x1,0 − 0.004x1,1 x2,1 x23,1 x1,0
−1.27x21,1 x2,0 x3,0 x1,0 + 18.x31,1 x2,0 x1,0 + 6.45x21,1 x2,0 x1,0 + 0.366x1,1 x2,0 x1,0
−0.153x1,1 x22,1 x3,0 x1,0 + 0.0766x2,0 x1,0 − 119.x31,1 x2,1 x1,0 − 0.629x32,0 x2,1 x1,0
−0.946x21,1 x2,1 x1,0 − 49.7x1,1 x22,0 x2,1 x1,0 − 0.0546x22,0 x2,1 x1,0 − 6.65x1,1 x2,1 x1,0
−24.1x21,1 x2,0 x2,1 x1,0 − 5.97x1,1 x2,0 x2,1 x1,0 − 0.0923x2,0 x2,1 x1,0
−0.283x31,1 x3,0 x1,0 − 0.127x32,0 x3,0 x1,0 + 0.00116x32,1 x3,0 x1,0
−0.0218x21,1 x3,0 x1,0 − 0.0172x1,1 x22,0 x3,0 x1,0 + 0.00375x22,0 x3,0 x1,0
−0.378x2,0 x22,1 x3,0 x1,0 + 0.00377x22,1 x3,0 x1,0 − 0.0024x1,1 x3,0 x1,0
−0.00107x1,1 x2,0 x3,0 x1,0 − 0.00982x2,0 x3,0 x1,0 + 0.303x21,1 x2,1 x3,0 x1,0
−0.00123x22,0 x2,1 x3,0 x1,0 + 0.00853x1,1 x2,1 x3,0 x1,0 + 1.15x1,1 x2,0 x2,1 x3,0 x1,0
+0.00125x3,0 x1,0 + 0.846x31,1 x3,1 x1,0 − 0.0011x32,0 x3,1 x1,0 − 0.225x32,1 x3,1 x1,0
+0.00617x21,1 x3,1 x1,0 + 0.579x1,1 x22,0 x3,1 x1,0 + 0.613x1,1 x22,1 x3,1 x1,0 + 0.0248x52,0
−0.0099x52,1 − 0.0506x2,0 x2,1 x3,1 x1,0 + 0.0758x2,0 x22,1 x3,1 x1,0 + 0.171x1,1 x3,1 x1,0
+0.273x21,1 x2,0 x3,1 x1,0 + 0.0624x1,1 x2,0 x3,1 x1,0 − 0.877x21,1 x2,1 x3,1 x1,0
−0.378x22,0 x2,1 x3,1 x1,0 − 0.00299x1,1 x2,1 x3,1 x1,0 − 0.238x1,1 x2,0 x2,1 x3,1 x1,0
−0.0575x2,1 x3,1 x1,0 − 0.00222x21,1 x3,0 x3,1 x1,0 − 0.0075x1,1 x2,0 x3,0 x3,1 x1,0
+0.0015x1,1 x2,1 x3,0 x3,1 x1,0 + 0.00413x2,0 x2,1 x3,0 x3,1 x1,0 − 0.232x1,0 − 3.07x51,1
−0.788x41,1 − 0.15x1,1 x42,0 + 0.0422x42,0 − 0.628x1,1 x42,1 + 0.204x2,0 x42,1
+0.135x42,1 − 0.529x31,1 − 8.3x21,1 x32,0 − 0.0176x1,1 x32,0 + 0.00757x32,0 + 4.14x21,1 x32,1
−0.00883x22,0 x32,1 + 0.0513x1,1 x32,1 + 5.47x1,1 x2,0 x32,1 − 0.0287x2,0 x32,1 + 0.0106x21,1
−4.01x31,1 x22,0 − 1.49x21,1 x22,0 − 0.0735x1,1 x22,0 + 0.0196x22,0 − 8.93x31,1 x22,1 + 0.25x32,0 x22,1
−1.21x21,1 x22,1 − 1.71x1,1 x22,0 x22,1 + 0.253x22,0 x22,1 − 0.138x1,1 x22,1
−23.1x21,1 x2,0 x22,1 − 0.12x1,1 x2,0 x22,1 + 0.0227x2,0 x22,1 + 0.0196x22,1 − 0.00185x21,1 x2,0 x23,0
+0.00207x1,1 x2,0 x2,1 x23,0 − 0.00127x2,0 x22,1 x23,1 − 0.00269x21,1 x2,0 x23,1
−16.3x41,1 x2,0 − 0.423x31,1 x2,0 − 3.32x21,1 x2,0 + 1.x2,0 + 8.36x41,1 x2,1
−0.00137x42,0 x2,1 + 1.6x31,1 x2,1 + 5.13x1,1 x32,0 x2,1 − 0.00939x32,0 x2,1 + 0.55x21,1 x2,1
+5.32x21,1 x22,0 x2,1 + 0.776x1,1 x22,0 x2,1 + 0.0766x1,1 x2,1 + 31.2x31,1 x2,0 x2,1
+1.67x1,1 x2,0 x2,1 + 0.45x41,1 x3,0 − 0.0106x42,0 x3,0 + 0.00381x1,1 x2,0 x2,1 x23,1
+0.0594x42,1 x3,0 + 0.00157x31,1 x3,0 − 0.00341x32,0 x3,0 − 0.424x1,1 x32,1 x3,0 + (· · · )
APPENDIX D. CONTROL LAWS 113
(· · · ) + 0.00819x2,0 x32,1 x3,0 + 0.0854x21,1 x3,0 + 0.287x21,1 x22,0 x3,0 − 0.00249x22,0 x3,0
+0.00208x1,1 x22,0 x3,0 x3,1 + 0.953x21,1 x22,1 x3,0 + 0.118x22,0 x22,1 x3,0 + 0.0835x1,1 x2,0 x22,1 x3,0
−0.0102x2,0 x22,1 x3,0 − 0.00249x22,1 x3,0 + 0.0909x31,1 x2,0 x3,0 + 0.00427x21,1 x2,0 x3,0
−0.975x31,1 x2,1 x3,0 + 0.004x32,0 x2,1 x3,0 − 0.377x1,1 x22,0 x2,1 x3,0
−0.00982x1,1 x2,1 x3,0 − 0.125x21,1 x2,0 x2,1 x3,0 − 0.0506x1,1 x2,0 x2,1 x3,0
+1.01x3,0 − 0.0571x41,1 x3,1 + 0.00117x42,0 x3,1 + 0.00479x42,1 x3,1 − 0.00861x31,1 x3,1
−0.127x1,1 x32,0 x3,1 + 0.0634x1,1 x32,1 x3,1 + 0.0895x2,0 x32,1 x3,1 − 0.00542x32,1 x3,1
−0.0036x21,1 x3,1 − 0.0622x21,1 x22,0 x3,1 − 0.0253x1,1 x22,0 x3,1 − 0.124x21,1 x22,1 x3,1
−0.0238x22,0 x22,1 x3,1 − 0.0285x1,1 x22,1 x3,1 − 0.313x1,1 x2,0 x22,1 x3,1
+0.00135x2,0 x22,1 x3,1 + 0.00125x1,1 x3,1 − 0.276x31,1 x2,0 x3,1 − 0.00151x21,1 x2,0 x3,1
−0.0575x1,1 x2,0 x3,1 + 0.115x31,1 x2,1 x3,1 + 0.0787x32,0 x2,1 x3,1 + 0.00391x21,1 x2,1 x3,1
+0.079x1,1 x22,0 x2,1 x3,1 + 0.0195x22,0 x2,1 x3,1 + 0.00168x1,1 x2,1 x3,1
+0.378x21,1 x2,0 x2,1 x3,1 + 0.0234x2,0 x2,1 x3,1 + 0.00186x31,1 x3,0 x3,1 − 0.00108x32,1 x3,0 x3,1
+0.00248x1,1 x22,1 x3,0 x3,1 − 0.00214x21,1 x2,1 x3,0 x3,1 − 0.0093x22,0 x2,1 x3,0 x3,1
and
u1,1 = −1.33x51,0 + 53.x1,1 x41,0 + 2.49x2,0 x41,0 − 20.9x2,1 x41,0 − 0.0158x3,0 x41,0
+0.317x3,1 x41,0 + 0.1x41,0 − 23.2x21,1 x31,0 − 2.08x22,0 x31,0
−3.92x22,1 x31,0 − 4.69x1,1 x31,0 − 84.2x1,1 x2,0 x31,0 − 0.158x2,0 x31,0 + 17.9x1,1 x2,1 x31,0
+37.3x2,0 x2,1 x31,0 + 2.15x2,1 x31,0 + 1.26x1,1 x3,0 x31,0 + 0.0553x2,0 x3,0 x31,0
−0.421x2,1 x3,0 x31,0 − 0.283x1,1 x3,1 x31,0 − 0.428x2,0 x3,1 x31,0 + 0.0932x2,1 x3,1 x31,0
+0.00356x3,0 x3,1 x31,0 − 0.00727x3,1 x31,0 − 0.176x31,0 + 112.x31,1 x21,0 + 0.857x32,0 x21,0
+10.8x1,1 x21,0 − 14.5x32,1 x21,0 + 1.5x21,1 x21,0 + 56.3x1,1 x22,0 x21,0 (D.10)
+0.118x22,0 x21,0 + 75.1x1,1 x22,1 x21,0 + 5.14x2,0 x22,1 x21,0 + 0.212x22,1 x21,0
+0.00534x1,1 x23,0 x21,0 − 0.00185x2,1 x23,0 x21,0 + 0.005x1,1 x23,1 x21,0 − 0.002x2,1 x23,1 x21,0
+28.2x21,1 x2,0 x21,0 + 6.46x1,1 x2,0 x21,0 + 0.183x2,0 x21,0 − 153.x21,1 x2,1 x21,0
−24.9x22,0 x2,1 x21,0 − 1.17x1,1 x2,1 x21,0 − 23.9x1,1 x2,0 x2,1 x21,0 − 2.99x2,0 x2,1 x21,0
−3.32x2,1 x21,0 − 0.403x21,1 x3,0 x21,0 − 0.00862x22,0 x3,0 x21,0 − 0.0614x22,1 x3,0 x21,0
−0.067x1,1 x3,0 x21,0 − 1.28x1,1 x2,0 x3,0 x21,0 + 0.271x1,1 x2,1 x3,0 x21,0 + (· · · )
APPENDIX D. CONTROL LAWS 114
(· · · ) + 0.259x1,1 x32,0 + 0.573x2,0 x2,1 x3,0 x21,0 + 0.00426x2,1 x3,0 x21,0 − 0.0012x3,0 x21,0
−13.2x21,1 x22,0 x1,0 + 0.226x2,0 x32,1 + 1.75x21,1 x3,1 x21,0 + 0.292x22,0 x3,1 x21,0 + 0.202x22,1 x3,1 x21,0
−1.25x21,1 x2,0 x1,0 + 0.00617x1,1 x3,1 x21,0 + 0.306x1,1 x2,0 x3,1 x21,0 + 0.00429x2,0 x3,1 x21,0
−1.28x1,1 x2,1 x3,1 x21,0 − 0.128x2,0 x2,1 x3,1 x21,0 − 0.00149x2,1 x3,1 x21,0
−0.00222x1,1 x3,0 x3,1 x21,0 − 0.00375x2,0 x3,0 x3,1 x21,0 + 0.0854x3,1 x21,0
−15.8x41,1 x1,0 − 0.106x42,0 x1,0 − 0.663x42,1 x1,0 − 6.19x31,1 x1,0
−16.6x1,1 x32,0 x1,0 − 0.0149x32,0 x1,0 + 8.46x1,1 x32,1 x1,0 + 5.72x2,0 x32,1 x1,0 + 0.365x32,1 x1,0
−0.0735x22,0 x1,0 − 27.2x21,1 x22,1 x1,0 − 1.59x21,1 x1,0 − 1.57x22,0 x22,1 x1,0 − 3.8x1,1 x22,1 x1,0
−2.99x1,1 x22,0 x1,0 − 58.3x1,1 x2,0 x22,1 x1,0 − 0.152x2,0 x22,1 x1,0 − 0.138x22,1 x1,0
−0.00121x21,1 x23,0 x1,0 − 0.00371x1,1 x2,0 x23,0 x1,0 + 0.00207x2,0 x2,1 x23,0 x1,0
−0.00539x1,1 x2,0 x23,1 x1,0 + 0.00381x2,0 x2,1 x23,1 x1,0 + 0.0211x1,1 x1,0 − 120.x31,1 x2,0 x1,0
−6.65x1,1 x2,0 x1,0 + 33.9x31,1 x2,1 x1,0 + 5.14x32,0 x2,1 x1,0 + 8.88x21,1 x2,1 x1,0
+10.7x1,1 x22,0 x2,1 x1,0 + 0.776x22,0 x2,1 x1,0 + 1.1x1,1 x2,1 x1,0 + 158.x21,1 x2,0 x2,1 x1,0
+1.07x1,1 x2,0 x2,1 x1,0 + 1.67x2,0 x2,1 x1,0 + 0.0766x2,1 x1,0 + 0.84x31,1 x3,0 x1,0
−0.00301x1,1 x2,0 x3,1 x1,0 − 0.274x32,1 x3,0 x1,0 + 0.00471x21,1 x3,0 x1,0
+0.577x1,1 x22,0 x3,0 x1,0 + 0.89x1,1 x22,1 x3,0 x1,0 + 0.079x2,0 x22,1 x3,0 x1,0
+0.171x1,1 x3,0 x1,0 + 0.262x21,1 x2,0 x3,0 x1,0 + 0.0624x1,1 x2,0 x3,0 x1,0 − 1.11x21,1 x2,1 x3,0 x1,0
−0.378x22,0 x2,1 x3,0 x1,0 − 0.241x1,1 x2,0 x2,1 x3,0 x1,0 − 0.0506x2,0 x2,1 x3,0 x1,0
−0.0575x2,1 x3,0 x1,0 − 0.243x31,1 x3,1 x1,0 − 0.127x32,0 x3,1 x1,0 + 0.0634x32,1 x3,1 x1,0
−0.0844x21,1 x3,1 x1,0 − 0.154x1,1 x22,0 x3,1 x1,0 + 0.00375x22,0 x3,1 x1,0 − 0.251x1,1 x22,1 x3,1 x1,0
−0.463x2,0 x22,1 x3,1 x1,0 + 0.00854x22,1 x3,1 x1,0 − 0.0072x1,1 x3,1 x1,0 − 1.84x21,1 x2,0 x3,1 x1,0
−0.00982x2,0 x3,1 x1,0 + 0.352x21,1 x2,1 x3,1 x1,0 + 0.0846x22,0 x2,1 x3,1 x1,0
+0.097x1,1 x2,1 x3,1 x1,0 + 2.01x1,1 x2,0 x2,1 x3,1 x1,0 + 0.00168x2,1 x3,1 x1,0
+0.00559x21,1 x3,0 x3,1 x1,0 + 0.00208x22,0 x3,0 x3,1 x1,0 − 0.00429x1,1 x2,1 x3,0 x3,1 x1,0
+0.00248x22,1 x3,0 x3,1 x1,0 + 0.00152x1,1 x2,0 x3,0 x3,1 x1,0 + 1.29x1,1 x42,0 + 0.0141x42,0
+0.00125x3,1 x1,0 + 54.6x51,1 − 0.00908x52,0 + 0.0365x52,1 + 0.3x41,1
+9.77x1,1 x42,1 − 0.0144x2,0 x42,1 − 0.00696x42,1 + 10.7x31,1 + 2.16x21,1 x32,0
+0.0205x32,0 − 61.9x21,1 x32,1 + 0.304x22,0 x32,1 − 0.0527x1,1 x32,1 − 2.71x1,1 x2,0 x32,1
+43.9x31,1 x22,0 + 0.251x21,1 x22,0 + 0.836x1,1 x22,0 + 134.x31,1 x22,1 − 0.0585x32,0 x22,1
+0.296x21,1 x22,1 + 14.2x1,1 x22,0 x22,1 − 0.027x22,0 x22,1 + 2.49x1,1 x22,1 + 13.x21,1 x2,0 x22,1 + (· · · )
APPENDIX D. CONTROL LAWS 115
(· · · ) + 1.55x1,1 x2,0 x22,1 + 0.00312x31,1 x23,0 + 0.00103x1,1 x22,0 x23,0 + 0.00355x1,1 x22,1 x23,0
−0.0052x21,1 x2,1 x23,0 + 0.00667x31,1 x23,1 − 0.00219x32,1 x23,1 + 0.00223x1,1 x22,0 x23,1
+0.0066x1,1 x22,1 x23,1 − 0.00839x21,1 x2,1 x23,1 − 0.00198x22,0 x2,1 x23,1
−0.232x1,1 + 8.71x41,1 x2,0 + 4.51x31,1 x2,0 + 0.55x21,1 x2,0 + 0.0766x1,1 x2,0
−133.x41,1 x2,1 + 0.124x42,0 x2,1 − 0.474x31,1 x2,1 − 1.14x1,1 x32,0 x2,1 + 0.169x32,0 x2,1
−9.88x21,1 x2,1 − 55.6x21,1 x22,0 x2,1 − 0.135x1,1 x22,0 x2,1 + 0.0227x22,0 x2,1 − 18.5x31,1 x2,0 x2,1
−6.52x21,1 x2,0 x2,1 − 0.277x1,1 x2,0 x2,1 + 0.0391x2,0 x2,1
+1.x2,1 − 0.0552x41,1 x3,0 + 0.00495x42,1 x3,0 + 0.00523x31,1 x3,0 − 0.127x1,1 x32,0 x3,0
+0.0626x1,1 x32,1 x3,0 + 0.0883x2,0 x32,1 x3,0 − 0.00685x32,1 x3,0
−0.0036x21,1 x3,0 − 0.0568x21,1 x22,0 x3,0 − 0.0253x1,1 x22,0 x3,0 − 0.122x21,1 x22,1 x3,0
−0.0244x22,0 x22,1 x3,0 − 0.0222x1,1 x22,1 x3,0 − 0.258x1,1 x2,0 x22,1 x3,0 + 0.0015x2,0 x22,1 x3,0
+0.00125x1,1 x3,0 − 0.025x31,1 x2,0 x3,0 − 0.00158x21,1 x2,0 x3,0 − 0.0575x1,1 x2,0 x3,0
+0.112x31,1 x2,1 x3,0 + 0.0787x32,0 x2,1 x3,0 − 0.0146x21,1 x2,1 x3,0 + 0.0788x1,1 x22,0 x2,1 x3,0
+0.0195x22,0 x2,1 x3,0 + 0.00168x1,1 x2,1 x3,0 + 0.0831x21,1 x2,0 x2,1 x3,0 + 0.0234x2,0 x2,1 x3,0
+1.27x41,1 x3,1 − 0.0105x42,0 x3,1 + 0.136x42,1 x3,1 + 0.00233x31,1 x3,1 − 0.00341x32,0 x3,1
−0.924x1,1 x32,1 x3,1 − 0.0379x2,0 x32,1 x3,1 + 0.254x21,1 x3,1
+0.712x21,1 x22,0 x3,1 − 0.00249x22,0 x3,1 + 1.89x21,1 x22,1 x3,1 + 0.209x22,0 x22,1 x3,1
+0.194x1,1 x2,0 x22,1 x3,1 − 0.0158x2,0 x22,1 x3,1 − 0.00693x22,1 x3,1 + 0.123x31,1 x2,0 x3,1
+0.0146x21,1 x2,0 x3,1 + 0.00168x1,1 x2,0 x3,1 − 2.07x31,1 x2,1 x3,1 + 0.00857x32,0 x2,1 x3,1
−0.0015x21,1 x2,1 x3,1 − 0.941x1,1 x22,0 x2,1 x3,1 + 0.00145x22,0 x2,1 x3,1 − 0.171x1,1 x2,1 x3,1
−0.256x21,1 x2,0 x2,1 x3,1 − 0.115x1,1 x2,0 x2,1 x3,1 − 0.00185x2,0 x22,1 x3,0 x3,1
+0.0019x1,1 x2,0 x2,1 x3,0 x3,1 + 1.01x3,1 .
FOLHA DE REGISTRO DO DOCUMENTO
1. 2. 3. 4.
CLASSIFICAÇÃO/TIPO DATA REGISTRO N° N° DE PÁGINAS
Polynomial chaos-based control design for an aircraft at high angles of attack with uncertain
aerodynamics.
6.
AUTOR(ES):
Optimal Control, Adaptive Control, Stability, Uncertainty, Polynomial Chaos, High angle of attack
9.PALAVRAS-CHAVE RESULTANTES DE INDEXAÇÃO:
Controle ótimo; Controle adaptativo; Dinâmica de voo; Ângulos de ataque; Estabilidade; Controle.
10.
APRESENTAÇÃO: X Nacional Internacional
ITA, São José dos Campos. Curso de Mestrado. Programa de Pós-Graduação em Engenharia Aeronáutica
e Mecânica. Área de Sistemas Aeroespaciais e Mecatrônica. Orientador: Prof. José Manoel Balthazar.
Defesa em 20/06/2017. Publicada em 2017.
11.
RESUMO:
In this work we consider the flight dynamics of fighter aircraft at high angles of attack with uncertain
decomposition of the random variables by means of the generalized polynomial chaos expansion. The
projection of uncertainty onto the orthogonal polynomial basis, prescribed by the Wiener-Askey scheme,
provides a deterministic model from which the control laws are designed. We propose optimal linear and
nonlinear feedback control strategies for the automatic pilot system to recover the aircraft from stall and
provide acceptable dynamic response. Optimality of the proposed control laws is proved by solving the
12.
GRAU DE SIGILO: