Original Article

Proc IMechE Part G:

J Aerospace Engineering
Flight control clearance of the Cessna 2017, Vol. 231(3) 510–532
! IMechE 2016

Citation X using evolutionary algorithms Reprints and permissions:

sagepub.co.uk/journalsPermissions.nav
DOI: 10.1177/0954410016640821
journals.sagepub.com/home/pig

Yamina Boughari, Ruxandra Mihaela Botez,

Georges Ghazi and Florian Theel

Abstract
In this paper, an Aircraft Research Flight Simulator equipped with Flight Dynamics Level D (highest level) was used to
collect flight test data and develop new controller methodologies. The changes in the aircraft’s mass and center of gravity
position are affected by the fuel burn, leading to uncertainties in the aircraft dynamics. A robust controller was designed
and optimized using the H1 method and two different metaheuristic algorithms; in order to ensure acceptable flying
qualities within the specified flight envelope despite the presence of uncertainties. The H1 weighting functions were
optimized by using both the genetic algorithm, and the differential evolution algorithm. The differential evolution algo-
rithm revealed high efficiency and gave excellent results in a short time with respect to the genetic algorithm. Good
dynamic characteristics for the longitudinal and lateral stability control augmentation systems with a good level of flying
qualities were achieved. The optimal controller was used on the Cessna Citation X aircraft linear model for several flight
conditions that covered the whole aircraft’s flight envelope. The novelty of the new objective function used in this
research is that it combined both time-domain performance criteria and frequency-domain robustness criterion, which
led to good level aircraft flying qualities specifications. The use of this new objective function helps to reduce consid-
erably the calculation time of both algorithms, and avoided the use of other computationally more complicated methods.
The same fitness function was used in both evolutionary algorithms (differential evolution and genetic algorithm), then
their results for the validation of the linear model in the flight points were compared. Finally, robustness analysis was
performed to the nonlinear model by varying mass and gravity center position. New tools were developed to validate the
results obtained for both linear and nonlinear aircraft models. It was concluded that very good performance of the
business Cessna Citation X aircraft was achieved in this research.

Keywords
Optimal controller, flight envelope, Cessna Citation X, flight dynamics, aircraft flying qualities, evolutionary algorithms

Date received: 23 April 2015; accepted: 22 February 2016

Introduction The aircraft controller determination is very com-

The aircraft’s safety relays importantly on its control-
ler, and the clearance authorities need to ensure that
the controller operates properly through the specified
flight envelope even in the presence of uncertainties
such as mass, center of gravity positions, and inertia
variations. The control clearance process is a fastidi-
ous and expensive task, especially for modern air-
crafts that need to achieve high performance.1 This
process aims to prove that the selected stability,
robustness, and handling requirements are satisfied
against any possible uncertainty.
During the industrial clearance process, the selec-
tion of the appropriate control laws with sufficient
robustness involves: the investigation of the closed-
loop eigenvalues, the stability margins and the per-
formance indices, in the presence of uncertainties.
The resulting controller is used further for the
design of the flight control system (FCS).
plex. Nonlinear methods such as fuzzy logic and
neural network methods have been applied for air-
craft identification and control.2,3 The nonlinear
hybrid fuzzy logic control on a morphing wing was
explored in Grigorie et al.4 and Popov et al.5). Due to
its complexity in the aerospace industry, the determin-
ation of the robust FCS is usually carried out using
linear methods applied on linear models, and it is fur-
ther validated using nonlinear models. In the litera-
ture, many linear control methods were used to obtain
ETS, Laboratory of Active Controls, Avionics and AeroServoElasticity
LARCASE, Montreal, Quebec, Canada
Corresponding author:
Ruxandra Mihaela Botez, ETS, Laboratory of Active Controls, Avionics
and AeroServoElasticity LARCASE, 1100 Notre Dame West, Montreal,
Quebec H3C-1K3, Canada.
Email: ruxandra.botez@etsmtl.ca
Boughari et al.
an FCS such as the linear matrix inequality (LMI)
approach, which has been used to achieve a robust
control design of an uncertain aircraft system,6 adap-
tive controls have been used for disturbance rejec-
tion,7–9 other optimal algorithms were investigated
for gust load alleviation and further tested on differ-
ent aircrafts.10–13 Then, online parameter estimations
and identifications methods were used to improve the
flight control capabilities14–16 by its recovering in the
presence of disturbances.
To obtain a flight control system by taking disturb-
ances into account, the H1 linear method proposed
by Zames and Francis17 in 1983 had gained popular-
ity as a way to guarantee robustness in the presence of
uncertainties. The H1 method has been used in the
industry to develop controllers to meet the required
specifications and needs. One of the most important
aspects of this controller is the determination of the
weighting functions (W1 and W2), which are very
important in the gains determination. There is no spe-
cific methodology to determine these weighting func-
tions. The literature points out that the weighting
functions are determined using a trial-and-error meth-
odology, or by pure experience-based methods.
Several applications of this control method have
been incorporated in the aeronautical domain,
mostly for fighter jets, where a scheduled H1 control-
ler was used on VSTOL longitudinal control,18 and it
has as well been used on the lateral control of an
F-14.19 An H1 controller design with gain scheduling
approach was successfully used on a flexible aircraft
where the weighting functions were not optimized
but were determined using engineering intuition.20
To overcome this lack of reference formulas, some
guidelines were given in Hu et al.21 and Yang et al.22
to determine these weighting functions. However, due
to its trial-and-error nature the guidelines procedure
may take many iterations to find acceptable results.
Besides, the guidelines do not guarantee fulfillment of
the required conditions. For this reason, a method-
ology to tune the weighting functions to meet the man-
datory requirements is necessary.
There exists several weighting optimization meth-
ods based on the mathematical algorithms, in which
trade-offs were arranged between maximizing the sta-
bility margin and minimizing the H1 norm of the
final closed-loop transfer function.23 These algorithms
were often performed on frequency-dependent opti-
mizations, in which the iteration process demanded
a considerable amount of memory allocation. To
overcome this frequency-dependent optimization
memory, a state space weight optimization was devel-
oped in Perhinschi et al.24 However, that algorithm
does not guarantee a global minimum convergence,
which could lead to a poor stability margin, especially
important in a system operating in a large envelope,
such as an aircraft. This paper proposes a new and
innovative methodology by taking advantage of both
the GA and DE algorithms to optimize the H1 weight
511
functions to develop a controller that satisfies the
imposed dynamic specifications and the industrial
needs. This new approach can solve the clearance
problem by reducing the complexity of calculation
and validation. However, this research aims to con-
firm that the DE algorithm optimization is more effi-
cient and accurate than the GA optimization. Storn
and Price25 have also shown the efficiency of the DE
algorithm by its comparison with genetic algorithm.
Many global optimizations based on evolutionary
principles have been used in the control engineering
field. In the aeronautical field, aircraft trajectory opti-
mizations based stochastic search, such as the GA
were performed on several civil aircrafts26,27 as well
as parameters estimation performed on autonomous
air vehicle and flight testing for intelligent flight con-
trols.28,29 These new methodologies to estimation and
control different parameters will be applied in future
for the flight dynamics and control of the business
aircraft Cessna Citation X.
All of these methods were developed with the aim of
reducing the computational complexity and time of con-
vergence while achieving satisfactory results. For this
study, the GA and the DE algorithms were selected to
optimize the weighting function parameters.
The following section presents a brief description of
the Cessna Citation X, then the description of the non-
linear and linear aircraft models, followed by the Cessna
Citation X’s flying qualities. Next section contains a
short presentation of the H1 theory. Subsequently,
the weight-selection methods are mentioned, namely
the DE and the GA, followed by their application
to the H1 problem. Our simulation and results’ analysis
are presented later, followed by the conclusions.
Preliminary results are presented in Boughari et al.30
Cessna Citation X business aircraft
The Cessna Citation X is the fastest business aircraft
in the world; it operates at a Mach number of 0.935.
The longitudinal and lateral motions of the business
aircraft, its flight envelope, and its flying qualities
requirements are described below.
The Cessna Citation X aircraft was selected for this
work because the Aircraft Flight Research Simulator
Level D was available at LARCASE laboratory.
The Level D is the highest Flight Dynamics certifica-
tion level. The availability of this flight simulator
makes possible the validation of the numerical results
with real flight test data.
In order to analyze the stability of an aircraft, its
model must be first identified. The model identification
can be done by using a combination of fuzzy logic and
neural networks methods as performed on the F/A–18
aircraft in Boely et al.31 A new system identification for
the business Cessna Citation X aircraft has been devel-
oped in 2013 at LARCASE laboratory;32,33 this system
was compared with a linearized Cessna Citation X
model obtained using aircraft simulator data.
512 Proc IMechE Part G: J Aerospace Engineering 231(3)

Figure 1. Representation of Cessna Citation X aircraft’s rotation axes.

The aircraft’s lateral motion dynamics are given by

Aircraft dynamics the state space equation, using the aileron and the
The Cessna Citation X rotation axes are represented rudder deflections as inputs
in Figure 1, the aircraft nonlinear model is given in
Nelson.34 To design a controller, a linearization of the x_ lat ¼ Alat xlat þ Blat ulat ð5Þ
aircraft nonlinear model is required, for flight condi-
tions within the flight envelope given by the designer. 0 1
Following the decoupling of the linearized aircraft in Yb =V Yp =V ð1  Yr =VÞ gcos0 =V
two longitudinal and lateral motions, the equations B L Lp Lr 0 C
B b C
are represented in the form of the following state
Alat ¼ B C,
@ Nb Np Nr 0 A
space system
0 0 0 0
0 1
x_ ¼ Ax þ Bu ð1Þ Ya =V Yr =V
B L Lr C
B a C
Blat ¼ B C ð6Þ
@ Na Nr A
This system is decomposed into two sub-systems 0 0
representing the aircraft’s longitudinal and lateral
motions. The aircraft’s longitudinal motion dynamics where the state vector xlat ðtÞ and control vector ulat ðtÞ
are given by the state space equation, using the eleva- are given by
tor deflection as input
 T
xlat ðtÞ ¼ b p r  and ulat ðtÞ ¼ ða r ÞT ð7Þ
x_ long ¼ Along xlong þ Blong ulong ð2Þ

0 1
Xu Xw Xq gcos
B Zu Zw Zq 0 C
B C
Along ¼ B C, . . .
@ Mu þ Mw_ Zu Mw þ Mw_ Zw Mq þ Mw_ V 0 A
0 0 1 0
0 1 ð3Þ
Xe
B Ze C
B C
Blong ¼ B C
@ Me þ Mw_ Ze A
0

where matrix A contains the stability derivatives The linear model of the Cessna Citation X was
related to the states, denoted by xlong ðtÞ, and matrix obtained for 36 flight conditions using the Cessna
B contains the derivatives related to controls, denoted Citation X Aircraft Flight Research Simulator tests
by ulong (t) as follows performed at LARCASE.35,36
The linearized model is further decomposed in
 T linear fractional representation (LFR) models37
xlong ðtÞ ¼ u w q  and ulong ðtÞ ¼ e ð4Þ using the bilinear interpolation method. Thus, these
Boughari et al.
Figure 2. Cessna Citation X flight envelope.
models are obtained for 72 flight points, and 12
weight conditions described in the following section.
LFR models design by flight point’s interpolation
The linear models interpolation using the linear frac-
tional transformation (LFT) facilitates the calculation
of the state space matrices variation with the altitude
and the TAS.37,38 Given the data extracted from the
Aircraft Flight Research Simulator provided by CAE
Inc., the aircraft flight dynamics can be described for
any flight condition in the flight envelope.
Figure 2 shows the 36 flight points chosen inside
flight envelope limits. These aircraft models are
obtained at each 5000 ft. in altitude for four different
speeds.
Before carrying out the interpolation, two steps
need to be performed. The first step is regarding the
definition of the region for an altitude and a range of
TAS, where the interpolation will be performed, and
for which the four corners of the region form the
vertices.
Each of these ranges has lower and upper
values which are the bounds. The second step regards
the normalization of these bounds in order to attri-
bute each coordinate of the vertices a value equal to 1
or 1.
To optimize the level of accuracy, the smallest pos-
sible regions have been defined, containing only three
or four flight points to use as reference points for the
interpolation. This definition only allows a bilinear
interpolation, for which four coefficients have to be
found, using the following equations
Aðh, TASÞ ¼ A04,4 þ A14,4 h þ A24,4 TAS
þ A34,4 TAS  h ð8Þ
Blong ðh, TASÞ ¼ B04,1 þ B14,1 h þ B24,1 TAS
þ B34,1 TAS  h ð9Þ
513
Table 1. Maximum relative error.
Longitudinal mode Lateral mode
A 1:04e11 % 1:97e11 %
B 3:05e12 % 3:97e11 %
Figure 3. Definition of 26 regions.
Blat ðh, TASÞ ¼ B04,2 þ B14,2 h þ B24,2 TAS
þ B34,2 TAS  h ð10Þ
where A is a matrix of 4 rows and 4 columns, Blong is a
matrix of 4 rows and 1 column, and Blat is a matrix of 4
rows and 2 columns. The least square (LS) method is
employed to minimize the relative error in these refer-
ence points.39 From Table 1, it can be observed that the
maximum errors found for the state space matrices A
and B are negligible, therefore results are good.
From these results, 26 regions denoted by rect-
angles in Figure 3 are obtained that cover a large
part of the flight envelope. The mesh is valid for all
of the weight and XCG locations presented in
Figure 4. It can be noticed from Figure 3 that some
of the regions superimpose others (darker zones) due
to the common reference points, and in some cases
there is not only the interpolation considered, but
also the extrapolation.
All of these 26 regions’ vertices lead to 72 different
flight points that can be controlled. Figure 5 shows
these 72 conditions obtained by means of the LFR
models, which makes it possible a close approxima-
tion of the flight envelope limits.
Flying quality’s level 1
The flying qualities are provided by the U.S ‘‘Military
Specification for the Flying Qualities of Piloted
Airplanes MIL-STD-1797A’’. For the aircraft longi-
tudinal motion, two modes are perceived: (1) short
514 Proc IMechE Part G: J Aerospace Engineering 231(3)

4
x 10 Cessna Citation X centering

3.6

3.4

3.2

3
Weights (lb)

2.8

2.6

2.4

2.2

2
10 15 20 25 30 35 40
x cg (%)

Figure 4. Cessna Citation X Weight/XCG conditions.

4
x 10
6

Flight conditions

4
Altitude (ft)

0
100 150 200 250 300 350 400 450 500 550
TAS (kts)

Figure 5. Flight points obtained by LFR models.

period and (2) phugoı̈d mode. Three modes are per- criteria, which are required for satisfactory flight per-
ceived for the lateral aircraft motion: (1) the Dutch formance, and are expressed in terms of damping, and
roll mode, (2) the roll mode, and (3) the spiral mode. time constant as shown in Table 2. These flying qua-
These modes need to respect some of the desired lities are given for the cruise phase or phase B, and for
Boughari et al. 515

Table 2. Aircraft flying qualities level 1.40,41

z1
Criterion Type Limits
z2
Short period damping Modal 0.3 4 sp 4 2
r + e u y
Phugoı̈d damping Modal 0.04 4 ph K(s) G(s)
-
Dutch roll damping Modal 0.3 4 dr 4 2
Roll time constant Temporal Tr < 1.4 s

w z Figure 7. Mixed sensitivity H1 configuration.

P(s)
The transfer function between z and w can be writ-
ten as follows42
u y  
z ¼ P11 þ P12 KðI  P22 KÞ1 P21 w ð12Þ
K(s) z ¼ Fl ðP, KÞw ð13Þ

where Fl ðP, KÞ is the lower linear fractional transform-

Figure 6. Standard H1 configuration. ation of P and K. The H1 optimization problem
design is then formulated as given in43
the flight level 1 which corresponds to very good
flying qualities.40 Thus, the aircraft responses have minKstabilizing kFl ðP, KÞk1 ð14Þ
to meet the criteria given in Table 2 for the aircraft
certification.

H-infinity theory Definition of the mixed sensitivity H-infinity problem

H1 represents a modern approach to characterize The mixed sensitivity H1 optimization is one of several
closed-loop performance, by measuring the size of practical optimization problems in industry, where its
the closed-loop transfer function matrices, and the cost function is a combination of other two cost func-
way in which the control objectives can be fixed to tions, such as a control signal’s energy limitation and a
minimize closed-loop transfer functions.42 good tracking reference as shown in Figure 7. In order
to keep the system internally stable, these cost func-
Definition of the standard H-infinity tions will be optimized for a set of stabilizing control-
lers using the state space gain defined in the following
robust control problem
equation and in Basar and Bernhard44
The LFT technique is a configuration that describes " #
 ðI þ GKÞ1 
how the relationship between the input and the output  
min   ð15Þ
is affected by uncertainty as shown in Figure 6. The Kstabilizing KðI þ GKÞ1 
1
LFT is used to formulate the standard H1 configur-
ation, P(s) denotes the generalized plant which con-
tains the disturbance model and the interconnection Weighting functions are often used to respect the
structure between the plant and the controller K(s). w design specifications in the closed-loop system (con-
denotes all the external (disturbances, noise, and com- trol input signal limitations and good tracking). Thus,
mand) inputs and z denotes all of the external outputs equation (15) can be rewritten as function of both
expressed in terms of error signals, to be minimized, weighting functions W1 and W2 as follows
including both the performance and the robustness " #
 W ðI þ GKÞ1 
measures. The control input is denoted by y, and u  1 
min   ð16Þ
denotes the control signal’s vectors. The objective is to Kstabilizing W2 KðI þ GKÞ1 
1
find a stabilizing controller that minimizes the output
z, which means that it minimizes the H1 norm of the In practice, a stabilizing controller is found by iter-
closed-loop transfer function from w to z. ations using the lowest achievable value . Equation
The generalized plant P(s) can be written as (16) then becomes
  " #
P11 ðsÞ P12 ðsÞ  W ðI þ GKÞ1 
PðsÞ ¼ ð11Þ  1 
min   5 ð17Þ
P21 ðsÞ P22 ðsÞ Kstabilizing W2 KðI þ GKÞ1 
1
516
where  is the robustness criterion given as the
maximum value of the 1  norm of the system’s
closed-loop transfer function.43 In equation (17), W1
is used to shape the sensitivity function ðI  GKÞ1
and W2 to shape the complementary sensitivity
function given as KðI  GKÞ1 , which characterize
the disturbances and controls. In addition, K is
the state space gain calculated from the H1
method, while W1 and W2 represent the weighting
functions, appropriately chosen using guidelines
given by45
as þ b
W1 ¼ ð18Þ
cs þ d
1 
W2 ¼ ð19Þ
a
So the H1 problem will be then reduced to define
W1 and W2 .
Differential evolution and genetic
algorithms
This research aims to optimize the determination of
weighting functions given by equations (18) and (19),
using both the genetic and DE algorithms, in which
one fitness function will be defined and used for the
optimization process.
Objective function for DE algorithm and GA
The same objective function to be minimized is used
in both the GA and DE optimization methods, in
order to obtain the optimal solution calculated by
the H1 algorithm. In these algorithms, the objective
function gives the designer specifications for the
desired closed-loop time response of the system
using overshoot ðOSÞ, integral square error ðISEÞ,
and the frequency domain robustness criterion ðÞ as
shown in the next equation
fitness ¼ 10ðISE40:002Þ þ 10ðOS435 %Þ
þ 10ð41Þ ð20Þ
Differential evolution algorithm
The DE algorithm has been developed in 1995 by
Price and Storn,46,47 and has been used in global
optimization in many disciplines as shown in various
studies.48–51 The DE algorithm is a heuristic optimiza-
tion algorithm that uses real values, thus there is no
need for coding and decoding operations to represent
problem parameters. The key concept of the DE is its
use of a differential operator to diversify the popula-
tion. This section gives a detailed presentation of the
DE algorithm along with the GA.
Proc IMechE Part G: J Aerospace Engineering 231(3)
phase. In this phase, the number of gen-
De-initialization
erations is selected as one of the termination criteria.
The problem dimension is set according to the
number of parameters forming the fitness function.
Next, the parameters to be optimized are represented
in a vector form; at each generation, the ith vector is
described46 as
 
X~iG ¼ x1,iG , x2,iG , x3,iG , . . . :, xD, iG ð21Þ
The population is initialized randomly within the
search space constrained by the lower and higher
boundaries for each parameter. These boundaries are
represented in vectors given by the following equations

X~imin ¼ x1,imin , x2,imin , x3,imin , . . . :, xD, imin ð22Þ
 
X~imax ¼ x1,imax , x2,imax , x3,imax , . . . :, xD, imax ð23Þ
The jth component of the ith vector is initialized as
 
xj, i, 0 ¼ xj, min þ randi, j ½0, 1: xj, max  xj, min ð24Þ
where 04randi, j ½0, 141:
The next step after the finalization of the initializa-
tion step is the mutation operation.
De-mutation. ‘‘Mutation’’ is the operation of changing
parameters between different vectors. In the DE algo-
rithm, a random choice of three different parameter
vectors X~ri1 ,G , X~ri2 ,G , X~ri3 ,G is performed in the current
population; for each ith ‘‘target’’ vector X~ri1 ,G , a cor-
responding ‘‘donor’’ vector is created, which results
from the combination of the ‘‘target’’ vector and a
‘‘weighted difference’’ between two parameter vectors
X~ri2 ,G , X~ri3 ,G by a randomly chosen scalar F, where
F 2 ½0, 2. The ‘‘mutant’’ vector V~iG so called the
‘‘donor’’ vectors is defined in equation (24)45 as follows
VIG ¼ X~ri , G þ F  X~ri , G  X~ri G ð25Þ
1 2 3
De-crossover.To improve the diversity of the popula-
tion, a ‘‘crossover’’ operation is performed, from
which the ‘‘mutant’’ and the ‘‘target’’ vectors exchange
!
their components to create the ‘‘trial’’ vector U IG
!  
U IG ¼ u1,iG , u2,iG , u3,iG , . . . , uD, iG ð26Þ
There are two types of crossover operation: the
exponential (two points modulo) and the binomial
(uniform).
In the exponential crossover the trial vector is given
as follows
uj,iG ¼ vj,iG for j ¼ hniD , hn þ 1iD , . . . :, hn þ L  1iD
ð27Þ
Boughari et al.
Otherwise
uj,iG ¼ xj,iG for j 2 ½1, D
where h i denotes the modulo function with modulus D.
The ‘‘modulus operator’’ is the remainder after the
arithmetic division that is used as a function in the pro-
gram to reduce a generated number to a random one in
a smaller range of values. D refers to the number of
parameters to be optimized or the parameters dimen-
sions range, in which two integers L and n where ran-
domly generated from the range ½1, D.
The trial vector in the binomial crossover is given
as
uj,iG ¼ vj,iG , if randi,j ½0, 14Cr or j ¼ jrand
Otherwise
uj,iG ¼ xj,iG
where the crossover rate Cr 2 ½01, randi, j ½0, 1 is a
random number distributed uniformly, and
jrand 2 ½1, 2, . . . , D is an index randomly chosen to
ensure that the resultant trial vector U ~ iG , considers
in its expression at least one component from the
donor vector V~iG ¼ ½v1,iG , v2,iG , . . . . . . ::, vD, iG .
At the end of the population diversity step, a selec-
tion operation is performed as detailed in the next
phase.
Using the ‘‘selection’’ operation, we can
De-selection.
determine if the ‘‘trial’’ or ‘‘target’’ vectors survive
in the next generation or not and, thus, a constant
population size is kept. The selection operation is out-
lined as
which each iteration represents a generation; where
~ i, Gþ1 ¼ U
X ~ i, G if f U~ i, G 4f X
~ i, G
which the ‘‘individual’’ defines a string of real param-
Else X ~ i, G if f U
~ i, Gþ1 ¼ X ~ i, G 4 f X~ i, G ð32Þ
‘‘decoding’’. The different steps of the GA are
where f X ~ i, G is the objective function or the ‘‘fit-
ness’’ to be converged using iteration process.
De-iteration. The operations listed above (initialization,
mutation, crossover, and selection) are repeated until
the termination criteria are met, which consists of:
1. The maximum number of generations required by
the user; or
2. The convergence of the fitness function given in
‘‘Objective function for DE algorithm and GA’’
section.
The H1 weighting functions optimization for lon-
gitudinal and lateral control using the DE algorithm is
summarized as follows:
517
Set a population number as NP formed by the
weighting functions W1 and W2. The parameters
ð28Þ from the initial vector are
 
X~IG ¼ aq, iG , bq, iG , cq, iG , dq, iG , ae, iG ð33Þ
where a, b, c, d, are the coefficients of w functions
defined in equations (18) and (19). Each of these par-
ameters belongs to an interval defined by a lower and
an upper bound; for example, aq, iG 2 aq, iG , aq, iG
belongs to an interval in which aq, iG represents the
lower bound and aq, iG represents the higher bound.
The optimal gain is obtained by choosing the appro-
priate W1 and W2 parameters and then by simulating
the control system in the time domain to obtain sat-
isfactory characteristics of the system’s response. If
ð29Þ the satisfactory characteristics are not reached, the
iteration process continues, until one of the two ter-
mination criteria is achieved.
Weighting functions parameters for lateral control
ð30Þ are calculated by the same way. All the weighting
functions used in longitudinal and lateral controls
are determined using the DE algorithms once, and
another time with the GA.
Genetic algorithm applied to the H1 method
The GA is an evolutionary computation and a power-
ful stochastic search and optimization technique that
has become the most-recognized and used technique
in the last few years based on the genetic principles.
This algorithm has been successfully applied to aero-
nautical problems, such as control,52–56 optimal tra-
jectory research,57,58 and others field.59,60 The GA is a
stochastic search algorithm that finds solutions using
Darwin’s theory of natural selection; it is an iterative
process done until the desired solution is found, in
ð31Þ the best individual is examined according to its fitness.
In this research, the real-coded GA is considered, in
eters without performing ‘‘binary coding’’ or
1. Initial population: To start the evolution process a
population is randomly created. For an optimal
control problem, a population Pp of N individuals
is created, where N was chosen to be equal to 50,
and each individual is represented by two chromo-
somes corresponding to the coefficients of the two
weighting functions (W1, W2). For the optimiza-
tion of the pitch rate weighting functions Wq, the
individual is the following:
 
Individual ¼ ae, 1 a1 b1 c1 d1 ð34Þ
where aq, 1 , bq, 1 , cq, 1 , dq, 1 , e, 1 are weighting functions
coefficients of the pitch rate.
518
Figure 8. Example of uniform crossover.
Figure 9. Example of crossover by section.
2. Individuals evaluation: To quantify the adapta-
tion degree of an individual, a fitness function
evaluates the robustness and the performance of
the resulting controller using the weighting func-
tions (W1, W2) estimated for the individual. The
population is further sorted from the best-fitted
individual to the worst.
3. Crossover: To perform a crossover, an operator
randomly chooses two individuals in the current
population (parents) and crosses their chromo-
somes to create new individual (children). Two
different types of crossover methods are used to
improve the diversity of individuals and their
genes to obtain diverse results. The first type of
crossover methods regards the uniform crossover.
This method creates a random binary mask that
decides if two chromosomes can be crossed (see
Figure 8).
The second method consists in dividing the parents
into two or three sections, and each section is crossed to
obtain two individuals. Figure 9 shows two examples.
4. Mutations: A ‘‘mutation’’ is performed by chan-
ging the chromosome structure. To create a muta-
tion in an individual, two genes are randomly
selected and permuted, as shown in Figure 10.
5. Elitist selection: The process of natural selection
promotes the most fitted individuals according to
Darwin’s theory. The elitist selection can be
defined as follows: from one generation to
another, a portion of the most fitted individuals
will be guaranteed to always survive preserving its
genetic information. These individuals can be dis-
carded only if a better individual emerges from a
Proc IMechE Part G: J Aerospace Engineering 231(3)
given generation. This method has a fast conver-
gence, which can penalize the diversity of individ-
uals. To overcome the diversity problem, the
crossover is done by considering all the population,
but with more ‘‘chances’’ given to the better-fit
individual. Thus, even the less fitted individuals
can contribute to the creation of the new
generation.
To illustrate the iteration process of a genetic algo-
rithm search, the following steps are considered in a
GA as shown below:
Step 1: The weighting functions are represented as an
individual of fixed length; some parameters
are defined, such as the size of the individual
population, the crossover and the mutation
probabilities.
Step 2: The performance of an individual is quantified
by defining a fitness function, which selects
chromosomes that will be mated.
Step 3: The random initial population is set.
Step 4: The H1 norm of each individual is computed
and the gain control K is found.
Step 5: The fitness of each individual is evaluated.
Step 6: A pair of individuals is selected according to
the probability of their fitness.
Step 7: The next generation is reproduced by creating
a pair of offspring individuals.
Step 8: The best individuals were preserved from the
initial population, and with the new individ-
uals are inserted into the new population.
Step 9: Starting at step 6, the process is repeated until
the sizes of the new and the initial populations
are equal.
Boughari et al. 519

Figure 10. Example of mutation.

Start

Randomly generate initial

population

Mutation

Crossover

Gen = Gen + 1

Fitness

Selection

No
Termination criteria met
Fitness value
(gamma+ISE+OS)

Control
Yes signalU(s)
+ Actual
Hinfinity G
Optimal solution _ output Y(s)
-
H_infinity Plant
Stop controller

Figure 11. H1 optimization using the differential evolution algorithm.

Step 10: The initial population is replaced with the

new population.
Step 11: If the termination criteria has been satisfied,
obtain the solution, if not return to step 4
and repeat the process.
Two flowcharts summarize the DE and GA algo-
rithms, presented in Figures 11 and 12.
Presentation of results
The open loop of the Cessna Citation X business air-
craft is composed of aircraft dynamics, actuators, and
sensors, while in the aircraft’s closed loop, the actu-
ators’ limits, and the performances weighting functions
are considered in the control augmentation system
(CAS), as in the simulation shown in Figure 13.
520
Start
Randomly generate initial
population
Evaluate the fitness function
Selection
Gen = Gen + 1
Crossover
Mutation
No
Termination criteria met
Yes
Optimal solution
Stop
Figure 12. H1 optimization using the real-valued genetic algorithm.
The business aircraft Cessna Citation X is repre-
sented in the state space form for its longitudinal and
lateral motions. Robust control using the H1 design
is then applied on the Cessna Citation X to improve
its stability and its time response.
Closed-loop simulations of the Cessna Citation X lon-
gitudinal and lateral aircraft mode were performed for
the whole flight envelope. The results presented below
were obtained for 12 centering configurations, using 72
flight conditions obtained from both the Cessna citation
X Flight simulator and the interpolation method.
GA and DE algorithm optimization results
The genetic and DE algorithms best fitness and the
mean fitness functions during generations are
Proc IMechE Part G: J Aerospace Engineering 231(3)
Fitness value
(gamma+ISE+OS)
Control
signalU(s)
+
_ Hinfinity G
Actual
- output Y(s)
H_infinity Plant
controller
presented in Figures 14 and 15. These figures show
that the DE algorithm best fitness value converge
faster than the GA best fitness value (with running
time of 91.63 s and the solution given at the 5th gen-
eration) with respect to the fitness function of the GA
(running time of 131.18 s and the solution given at 8th
generation). The mean fitness value approximates the
best fitness value at 3rd generation in the DE, and the
4th generation in the GA, and it can be noticed
that the mean fitness varied from maximum value
of 2000 until a minimum value of 0.12856 which dem-
onstrate the diversity of the population. Mostly, the
convergence of the two algorithms is obtained
before reaching the 20th generation, which confirms
the performance and the efficiency of the two
algorithms.
Boughari et al. 521

Outer loop

Autopilot

Inner loop1

CAS Actuators Sensors Outputs

Pilot Inputs

Cessna Citation X
dynamics

Figure 13. Closed-loop representation of Cessna Citation X business aircraft.

Best: 0.20042 Mean: 0.20205

1500
Best fitness
Mean fitness

1000
Fitness value

500

0
0 2 4 6 8 10 12 14 16 18 20
Generation
Best: 0.20042
0.45

0.4
Fitness value

0.35

0.3

0.25

0.2
2 4 6 8 10 12 14 16 18 20
Generation

Figure 14. The mean fitness versus the best fitness and the best fitness value for the GA.
522 Proc IMechE Part G: J Aerospace Engineering 231(3)

Best: 0.12855 Mean: 0.12856

2000
Best fitness
1800 Mean fitness

1600

1400
Fitness value

1200

1000

800

600

400

200

0
0 2 4 6 8 10 12 14 16 18 20
Generation
Best: 0.12855
0.18

0.17

0.16
Fitness value

0.15

0.14

0.13

0.12
2 4 6 8 10 12 14 16 18 20
Generation

Figure 15. The mean fitness versus the best fitness and the best fitness value for the DE.

Table 3. Weighting function optimization results.

Weight function DE algorithm GA

2 7
Range of q weighting function coefficients ½510 ; 10  ½5102 ; 107 
1715sþ221476 800sþ641200
Wq and We weighting function solutions Wq ¼ 1:6610 6 sþ223992 Wq ¼ 1:639810 6 sþ1:1476106
1 1
W e ¼ 1421997 W e ¼ 4:402910 6

Range of ’ weighting function coefficients ½0; 150 ½0; 150

W’ and Wa weighting functions solutions W’ ¼ 0:0037sþ70:4
8:89sþ0:0009 W’ ¼ 0:7932sþ66:8216
0:1416sþ0:0028
1 1
W a ¼ 129:27 W a ¼ 137:2602
Generation number 420 420

Table 3 represents the weighting functions given by

Results for 72 flight conditions
the optimization using the DE algorithm. The opti-
The simulation was performed for 72 flight conditions mization is performed for a population size of 50, and
for each XCG location whose results are presented the search range for each longitudinal weighting par-
below. ameter ðaq , bq , cq , dq , e Þ and lateral weighting
Boughari et al. 523

Table 4. Mean g values.

Centring conditions 1 2 3 4 5 6 7 8 9 10 11 12

q 0.99 0.99 0.99 0.99 0.99 0.99 0.99 0.99 0.99 0.99 0.99 0.99
’ 0.96 0.96 0.96 0.97 0.97 0.99 1.00 1.04 1.04 1.09 1.09 1.09

(a) Reponse at point [-1;-1] for the 26 regions (b)

3.5 Pole-Zero Map
60
0.66 0.52 0.4 0.28 0.18 0.09
3 50

40
40 0.82
Pitch rate control q(deg/sec)

2.5 System: L 30
Pole : -20 + 39.7i
20
2 20 0.94 Damping: 0.451
Overshoot (%): 20.4
10

Imaginary Axis
Frequency (rad/sec): 44.4
1.5
0

1 10
-20 0.94
20
0.5
30
-40 0.82
0 40

50
0.66 0.52 0.4 0.28 0.18 0.09
-0.5 -60
0 2 4 6 8 10 -60 -50 -40 -30 -20 -10 0
time (sec) Real Axis

Figure 16. Responses for pitch rate presenting good handling qualities for first XCG position (22,000 lb/33%): (a) pitch rate
responses q; (b) pole zero maps.

parameter ða’ , b’ , c’ , d’ , a Þ is defined as ½5102 ; 107 . are nine points, because of the fact that there are 72
The results are shown for the 20th generation. flight points in the flight envelope, that are tested for
Table 4 shows that the robustness criteria is not 12 XCG locations; some of these points (as the ones
strictly less than or equal to one for higher weights shown in Figure 19(b) with the same colored arrow)
than 30,000 lb which means that the system is robust present the same flight points with different XCG loca-
for a certain range of uncertainties, and beyond this tions; these flight points present the same vertex as
range the system may not be robust. indicated by the red arrow or a shared vertex between
In addition, during the flight tests performed on the two adjacent regions as indicated by the black arrow
Cessna X Flight Research Simulator at the and shown in Figure 3.
LARCASE Laboratory, the aircraft has shown a lat- In the same way, for the roll angle ’ control,
eral dynamics sensibility and high altitudes coupling. Figure 20 shows time responses and pole zero map
Furthermore, in practice it is not evident for a heavy with handling qualities superimposed. Where some
weight aircraft to fly at high altitudes, where the air- responses show an overshoot (OS) of a maximum of
craft shows a robustness index gamma slightly greater 18%, Figure 21(a) and (b) presents the same flight
than 1 for lateral control at high altitudes and weights points positions, but there are differences in number;
as shown in Table 4. they are given in Table 5 for both GA and DE algo-
As shown in Figure 16, time responses and pole rithms, where we can see that the flight points number
zero map with handling qualities superimposed are are different. Finally the responses presenting the
given for pitch rate q. Some responses show an over- worst handling qualities are shown in Figure 22.
shoot (OS) of a maximum of 20%. On the other hand, Globally, the aircraft longitudinal and lateral
Figure 17 shows responses presenting the worst hand- motions are stabilized with the H1 controller.
ling qualities for pitch controller, where new tools are For both the controls the pitch angle rate q, and
presented in Figures 18, 19(a) and (b), which consists the roll angle ’, the resulting response satisfies the
of the flight points positions in the flight envelope for handling qualities level 1 with damping ration and
which the worst handling qualities are visualized, and natural frequency within the limits according to
their coordinates are given in the text file generated by Table 2 for both the lateral and longitudinal motions,
the Matlab code. and the imposed time domain performance, the inte-
It can be noticed that in Figure 18(b) there are six gral square error (ISE) less than 2%, OS of less
flight points, and in the listing in Figure 19(b) there than 30%.
524 Proc IMechE Part G: J Aerospace Engineering 231(3)

(a) 3 (b) Pole-Zero Map

0.32 0.22 0.14 0.07

1
1
0.44
2 0.8
0.8

0.58 0.6
0.6 System: L
Zero : -0.0212 + 0.169i
1 0.74 Damping: 0.124 0.4
Pitch rate control q(deg/sec)

0.4
Overshoot (%): 67.4
Frequency (rad/sec): 0.17 0.2

Imaginary Axis
0.2 0.92

0 0

-0.2 0.92
0.2

-1 -0.4 0.74
0.4

-0.6
0.58 0.6

-2 -0.8
0.8
0.44
-1
1
0.32 0.22 0.14 0.07
-3 -0.4 -0.3 -0.2 -0.1 0 0.1 0.2 0.3 0.4
0 1 2 3 4 5 6 7 8 9 10
Real Axis
time (sec)

Figure 17. Response for pitch rate q presenting the worst handling qualities for the entire envelope: (a) pitch rate responses q;
(b) pole zero maps.

4
(a) x 10 (b) x 10 4
6 6

5 5

4 4
Altitude (ft)
Altitude (ft)

3 3

2 2

1 1

0 0
100 150 200 250 300 350 400 450 500 550 100 150 200 250 300 350 400 450 500 550
TAS (kts) TAS (kts)

Figure 18. Flight points where the handling qualities for the pitch rate q control are the worst: (a) the DE algorithm; (b) the GA.

Table 5 shows the flight points (within the flight overshoot OS higher than 30% and a long settling
envelope) for which the worst handling qualities time Ts which reduces the controllability (handling
were obtained by using optimized weighting functions quality level); excess weights can affect the structural
for both the GA and DE algorithms. The resulting limits given by the designer; high speeds can affect the
weighting functions have been validated using a aerodynamic forces and can lead to aircraft failure
linear model for almost all of the flight conditions, (loss of control surfaces). Moreover, if we compare
except for a few flight conditions, which are shown the two algorithms, it can be deduced that the results
by numbers and percentages for pitch angle rate q obtained by using the DE algorithm are more accur-
and roll angle ’. These few flight conditions belong ate than those obtained by the GA optimization.
to either very high loads or high speeds, or to low The H1 method thus gives a controller that
loads and low speed cases, which means flight points approximates the good handling qualities level 1 for
at the limit of the flight envelope. They present an both longitudinal and lateral motions as given in
Boughari et al. 525

Figure 19. Flight points coordinates: (a) the DE algorithm; (b) the GA.

(a) Reponse at point [-1;-1] for the 26 regions (b)

1.4 Pole-Zero Map
40
0.8 0.66 0.54 0.4 0.26 0.12
System: H
Pole : -19.8 + 38.7i
1.2 30 Damping: 0.456
0.9
Overshoot (%): 20
Frequency (rad/sec): 43.5

20
1
0.97
Roll angle phi (deg)

10
Imaginary Axis

0.8

0 50 40 30 20 10
0

0.6
-10

0.97
0.4
-20

0.2 -30
0.9

0.8 0.66 0.54 0.4 0.26 0.12

0 -40
-60 -50 -40 -30 -20 -10 0
0 2 4 6 8 10
Real Axis
time(sec)

Figure 20. Responses of the roll angle ’ for the entire envelope presenting good handling qualities: (a) roll angle responses’;
(b) pole zero maps.
526 Proc IMechE Part G: J Aerospace Engineering 231(3)

4 4
(a) x 10 (b) x 10
6 6

5 5

4 4
Altitude (ft)

Altitude (ft)
3 3

2 2

1 1

0 0
100 150 200 250 300 350 400 450 500 550 100 150 200 250 300 350 400 450 500 550
TAS (kts) TAS (kts)

Figure 21. Flight points where the handling qualities for the roll angle ’ control are the worst: (a) the differential evolution algo-
rithm; (b) the genetic algorithm.

(a) 20 (b) Pole-Zero Map

3
0.81 0.7 0.56 0.38 0.2
15 0.89

2
10 0.95

1
5
0.988
Roll control phi (deg)

Imaginary Axis

0 5 4 3 2 1
0

-5
0.988
-1

-10
0.95 System: L
Zero : -0.703 - 2.76i
-2 Damping: 0.247
-15 Overshoot (%): 44.9
0.89 Frequency (rad/sec): 2.85
0.81 0.7 0.56 0.38 0.2
-20 -3
0 1 2 3 4 5 6 7 8 9 10 -5 -4 -3 -2 -1 0 1 2
time (sec) Real Axis

Figure 22. Flight points presenting the worst handling qualities for the roll angle control ’: (a) roll angle ’; (b) pole zero maps.

Table 5. Flight points with the worst handling qualities. validation. In this research different flight conditions
were used to cover the entire flight envelope to valid-
Flight points with the Flight points with
ate the H1 controllers. In the previous researches per-
worst handling qualities the worst handling
Controls for the DE algorithm qualities for the GA formed in the aeronautical field, only one controller
(XCG location) was optimized for unmanned aircrafts
Pitch rate q 4/864 (0.5%) 9/864 (1.04%) or helicopters.
Roll angle ’ 13/864 (1.5%) 18/864 (2%)

Nonlinear validation
Finally, to prove the efficiency of the optimized con-
Table 2, improves the aircraft’s stability, and its troller and its robustness against uncertainties, a non-
dynamic performance. This is the first time that linear validation was performed using the Cessna
such research was performed on the flight control Citation X aircraft’s nonlinear model developed to
clearance using a real business aircraft model for its simulate a real aircraft dynamics, using the Cessna
Boughari et al. 527

Citation X Research Simulator data. A simulation of It can be seen that the pitch angle rate q and roll
a pitch angle rate q and roll angle ’ control responses angle ’ responses remain stable during the simulation,
were performed, with the results as shown respectively and that all the performance criteria are reached;
in Figures 23 and 25 for the altitude of 2000 ft, TAS of therefore the system is robust as desired.
230 knots and load of 26,000 lb, and varying mass
shown in Figures 24 and 26.
Robustness analysis of H1 controller
Figures 27 and 28 show robustness results for the H1
controller. The tests were performed on the nonlinear
1
model of the Cessna Citation X that takes into account
the nonlinear dynamics, actuators, sensors, saturations,
0.8
and signal processing times. A total of 160 tests were
performed by generating uncertainty of 5% on the
0.6
mass and the center (position of center of gravity) with
respect to a nominal condition for which the controller
q deg/sec

0.4
was obtained. The selection of the nominal flight con-
dition and uncertainties were random. The results
0.2 obtained on pitch rate control and roll angle control
are presented in Figures 27 and 28.
0 The results reveal that the pitch rate, and roll angle
controls are stable with respect to the mass, and center
-0.2 of gravity position variations, where their variations
0 1 2 3 4 5 6 7 8 9 10
time sec are stable and included in the acceptable range.

Figure 23. Pitch angle rate q hold control responses using

nonlinear.

4
x 10 Altitude (ft)
2.04

2.02

1.98
0 1 2 3 4 5 6 7 8 9 10

True airspeed (kts)

231

230.5

230

229.5
0 1 2 3 4 5 6 7 8 9 10

Heading(deg)
118

117.5

117
0 1 2 3 4 5 6 7 8 9 10
4
x 10 Mass(lb)
2.6

2.595

2.59
0 1 2 3 4 5 6 7 8 9 10
Time(sec)

Figure 24. Altitude, true airspeed, heading, and mass variation responses using nonlinear aircraft model.
528 Proc IMechE Part G: J Aerospace Engineering 231(3)

1.4

1.2

0.8
phi (deg)

0.6

0.4

0.2

0
0 1 2 3 4 5 6 7 8 9 10
Time(sec)

Figure 25. Roll angle ’ control responses on a nonlinear aircraft model.

x 10
4 Altitude (ft)
2

2
0 1 2 3 4 5 6 7 8 9 10

True airspeed (kts)

230.01

230

229.99

229.98
0 1 2 3 4 5 6 7 8 9 10

Heading(deg)
118

117.5

117
0 1 2 3 4 5 6 7 8 9 10
4
x 10 Mass(lb)
2.6

2.598

2.596
0 1 2 3 4 5 6 7 8 9 10
Time(sec)

Figure 26. Altitude, true airspeed, heading, and mass variation responses on a nonlinear aircraft model.
Boughari et al. 529

1.2

0.8
Pitch rate [deg/s]

0.6

0.4

0.2

0
Reference
Aircraft
-0.2
0 0.5 1 1.5 2 2.5 3 3.5 4 4.5 5
Time [s]

Figure 27. Pitch rate q response using mass and the center variation.

1.4

1.2

1
Roll angle φ [deg]

0.8

0.6

0.4

0.2
Reference
Aircraft
0
0 0.5 1 1.5 2 2.5 3 3.5 4 4.5 5
Time [s]

Figure 28. Roll angle ’ response using mass and the center variation.

In this study, the weighting functions were deter-

Conclusion
This paper presented a new application of evolutionary
robust design controller, which aims to develop con-
trollers providing robustness against disturbances and
uncertainties that are present in the real environment.
mined using two different evolutionary algorithms,
the DE and the GA; one of their greatest advantages
is that no prior knowledge is required about the con-
trol method. Using these evolutionary algorithms
over conventional optimizations improves the
530
reliability, and effectiveness of the clearance process
owing to their flexibility and adaptability.
Combined with the H1 method, the proposed
objective function helps to reduce considerably the
calculation time of both the algorithms DE and GA,
rather than the use of more complicated methods.
However, the solution given by the DE algorithm
optimization is more efficient and accurate than
the GA optimization for the clearance process.
The efficiency of the optimization-based clearance is
due to its flexibility concerning the formulation of the
clearance criteria (as the handling qualities, robust-
ness, and time performance criteria).
Simulations were performed using the resulting
optimal gains for the Cessna Citation X aircraft’s lon-
gitudinal and lateral motions, for 12 centering
(XCG positions) and 72 flight conditions, selected to
cover the whole flight envelope. The optimized feed-
back gains enhanced the aircraft’s closed-loop per-
formances, according to handling qualities level 1,
and designer specifications, while the DE and GAs
demonstrated a high efficiency in global optimization
with minimum time convergence. In this research, new
tools have been developed to validate the results of
linear and nonlinear models, which provide a clear
and accurate analysis for the user, and to facilitate
the controller’s certification process.
Acknowledgments
This work was performed at the LARCASE (Laboratory of
active controls, avionics and aeroservoelasticity research).
We would like to thank to the project leader Mr Ken
Dustin and his team at CAE Inc. for their support in the
development of the Aircraft Research Flight Simulator at
the LARCASE laboratory.
This simulator was obtained thanks to research grants that
were approved by the Canada Foundation for Innovation
(CFI) and the Ministère de Développement de l’Économie,
de l’Innovation et de l’Exportation MDEIE. Dr Ruxandra
Botez is Canada Research Chair Holder in Aircraft
Modeling and Simulation Technologies. Thanks are also
due to Mrs Odette Lacasse and Mr Oscar Carranza at ETS
for their continuing enthusiasm and support.
Declaration of Conflicting Interests
The author(s) declared no potential conflicts of interest with
respect to the research, authorship, and/or publication of
this article.
Funding
The author(s) disclosed receipt of the following financial
support for the research, authorship, and/or publication
of this article: This research was funded by the Natural
Sciences and Engineering Research Council of Canada
NSERC in the frame of Canada Research Chair Programs.
References
1. Fielding C and Varga A. Advanced techniques for clear-
ance of flight control laws. New York: Springer Science &
Business Media, 2002.
Proc IMechE Part G: J Aerospace Engineering 231(3)
2. Kouba G, Botez RM and Boely N. Identification of
F/A-18 model from flight tests using the fuzzy
logic method. In: Proceedings of 47th AIAA aerospace
sciences meeting including new horizons forum aerospace
exposition, Orlando, FL, USA, 5 January 2009,
pp.1–28.
3. Boely N, Botez RM, Kouba G, et al. Identification
of a nonlinear model between control and structural
deflections of an F/A-18 aircraft. In: 47th AIAA aero-
space sciences meeting including the new horizons forum
and aerospace exposition, Orlando, FL, 5–8 January
2009.
4. Grigorie TL, Botez RM, Popov AV, et al. A hybrid
fuzzy logic proportional-integral-derivative and conven-
tional on-off controller for morphing wing actuation
using shape memory alloy-Part 1: Morphing system
mechanisms and controller architecture design.
Aeronaut J 2012; 116: 433.
5. Popov AV, Grigorie LT, Botez RM, et al. Real time
morphing wing optimization validation using wind-
tunnel tests. J Aircraft 2010; 47: 1346–1355.
6. Ibrir S and Botez R. Robust stabilization of uncer-
tain aircraft active systems. J Vib Control 2005; 11:
187–200.
7. Balas MJ and Frost SA. Direct adaptive control for
infinite-dimensional symmetric hyperbolic systems.
Procedia Comput Sci 2014; 36: 549–555.
8. Balas MJ and Frost S. Robust adaptive model tracking
for distributed parameter control of linear infinite-
dimensional systems in Hilbert space. IEEE/CAA J
Automat Sin 2014; 1: 294–301.
9. Balas MJ and Frost S. Discrete-time infinite-dimen-
sional adaptive control and rejection of persistent dis-
turbances: To D or not to D?. In: 21st Mediterranean
conference on control & automation (MED), 25 June
2013, pp.1042–1049. New York: IEEE.
10. Frost SA, Bodson M, Burken JJ, et al. Flight control
with optimal control allocation incorporating structural
load feedback. J Aerosp Inform Syst 2015; DOI:
10.2514/1.I010278.
11. Frost SA, Taylor BR and Bodson M. Investigation of
optimal control allocation for gust load alleviation in
flight control. In: AIAA atmospheric flight mechanics
conference, Minneapolis, Minnesota, 2012, p. 4858..
12. Frost S and Balas M. Evolving systems: An outcome of
fondest hopes and wildest dreams. In: Guidance,
Navigation, and Control Conference, ser. Guidance,
Navigation, and Control and Co-located Conferences.
AIAA 2012.
13. Aouf N, Boulet B and Botez R. H 2 and H1-optimal
gust load alleviation for a flexible aircraft. In:
Proceedings of the American control conference, vol. 3,
2000, pp.1872–1876. New York: IEEE.
14. Perhinschi MG, Campa G, Napolitano MR, et al.
A simulation tool for on-line real time parameter iden-
tification. In: Proceedings of the 2002 AIAA modeling
and simulation conference, Monterey, California, 2002.
15. Campa G, Napolitano M, Seanor B, et al. Online par-
ameter estimation techniques comparison within a fault
tolerant flight control system. J Guid Control Dyn 2002;
25: 528–537.
16. Perhinschi MG, Lando M, Massotti L, et al. Online
parameter estimation issues for the NASA IFCS F-15
fault tolerant systems. In: Proceedings of American
Boughari et al.
control conference, 2002, vol. 1, 2002, pp.191–196. New
York: IEEE.
17. Zames G and Francis B. Feedback, minimax sensitivity,
and optimal robustness. IEEE Trans Autom Control
1983; 28: 585–601.
18. Hyde R and Glover K. The application of scheduled
H1 controllers to a VSTOL aircraft. IEEE Trans
Autom Control 1993; 38: 1021–1039.
19. Balas GJ, Packard AK, Renfrow J, et al. Control
of the F-14 aircraft lateral-directional axis during
powered approach. J Guid Control Dyn 1998; 21:
899–908.
20. Aouf N, Boulet B and Botez R. A gain scheduling
approach for a flexible aircraft. In: Proceedings of the
American control conference, 2002, vol. 6, pp.4439–
4442. New York: IEEE.
21. Hu J, Bohn C and Wu HR. Practical H1 weighting
functions and their application to real-time control of
a pilot plant. In: Proceedings of the American con-
trol conference, 2 June 1999, pp.920–924. New York:
IEEE.
22. Yang CD, Ju HS and Liu SW. Experimental design of
H-infinity weighting functions for flight control sys-
tems. J Guid Control Dyn 1994; 17: 544–552.
23. Lanzon A. Weight optimisation in H1 loop-shaping.
Automatica 2005; 41: 1201–1208.
24. Perhinschi MG, Napolitano MR, Campa G, et al.
Design and flight testing of intelligent flight control
laws for the WVU YF-22 model aircraft. 2005 AIAA
Guidance, Navigation, and Control Conference and
Exhibit, San Francisco, California, pp.1–12.
25. Storn R and Price K. Differential evolution–a sim-
ple and efficient heuristic for global optimization
over continuous spaces. J Global Optim 1997; 11:
341–359.
26. Murrieta-Mendoza A and Botez R. Aircraft vertical
route optimization deterministic algorithm for a flight
management system. SAE Technical Paper, 15 Sep
2015.
27. Patrón RS and Botez RM. Flight trajectory optimiza-
tion through genetic algorithms for lateral and vertical
integrated navigation. J Aerosp Inform Syst 2015; 12:
533–544.
28. Mario GP. Parameter optimization via genetic
algorithm of fuzzy controller for autonomous air
vehicle. In: AIAA guidance, navigation, and control
conference and exhibit, Portland, OR, USA, 1999
August.
29. Osinuga M, Patra S and Lanzon A. State-space solution
to weight optimization problem in H1 loop-shaping
control. Automatica 2012; 48: 505–513.
30. Boughari Y, Botez R, Ghazi G, et al. Evolutionary
algorithms for robust Cessna citation X flight control.
SAE Technical Paper, 16 September 2014.
31. Boely N, Botez RM and Kouba G. Identification of a
non-linear F/A-18 model by the use of fuzzy logic and
neural network methods. Proc IMechE, Part G: J
Aerospace Engineering 2011; 225: 559–574.
32. Hamel C, Sassi A, Botez R, et al. Cessna citation x
aircraft global model identification from flight tests.
SAE Technical Paper, 17 September 2013.
33. Hamel C, Botez R and Ruby M. Cessna citation x air-
plane grey-box model identification without prelimin-
ary data. SAE Technical Paper, 16 September 2014.
531
34. Nelson RC. Flight stability and automatic control. New
York: WCB/McGraw Hill, 1998.
35. Ghazi GE. De´veloppement d’une plateforme de simula-
tion et d’un pilote automatique-application aux
Cessna Citation X et Hawker 800XP. Doctoral
Dissertation, Master’s Thesis, University of Quebec-
École Polytechnique de Montréal, Canada, 2014.
36. Ghazi G and Botez R. Development of a high-fidelity
simulation model for a research environment. SAE
Technical Paper, 15 September 2015.
37. Poussot-Vassal C and Roos C. Flexible aircraft reduced-
order LPV model generation from a set of large-scale
LTI models. In: American control conference (ACC),
29 June 2011, pp.745–750. New York: IEEE.
38. Hecker S, Varga A and Magni JF. Enhanced LFR-tool-
box for MATLAB. In: IEEE international symposium
on computer aided control systems design, 4 September
2004, pp.25–29. New York: IEEE.
39. Biskri DE, Botez RM, Stathopoulos N, et al. New
mixed method for unsteady aerodynamic force approxi-
mations for aeroservoelasticity studies. J Aircraft 2006;
43: 1538–1542.
40. Jackson EB, Bilimoria KD, Mueller ER, et al. Cooper-
Harper experience report for spacecraft handling quali-
ties applications. National Aeronautics and Space
Administration, Langley Research Center, 2009.
http://aviationsystemsdivision.arc.nasa.gov/publication
s/shaq/NASA_Pub_2009_Bailey.pdf
41. Roskam J. Determination of stability, control and
performance characteristics: Far and military
requirements. Roskam Aviation and Engineering
Corporation, 1988.
42. Zames G and Wang LY. Local-global double algebras
for slow H1 adaptation. I. Inversion and stability.
IEEE Trans Autom Control 1991; 36: 130–142.
43. Gu DW, Petkov PH and Konstantinov MM. Robust
control design with MATLABÕ . New York: Springer
Science & Business Media, 2005.
44. Başar T and Bernhard P. H-infinity optimal control and
related minimax design problems: A dynamic game
approach. New York: Springer Science & Business
Media, 2008.
45. Beaven RW, Wright MT and Seaward DR. Weighting
function selection in the H1 design process. Control
Eng Pract 1996; 4: 625–633.
46. Price KV. Differential evolution: A fast and simple
numerical optimizer. In: 1996 biennial conference of
the North American fuzzy information processing society,
19 June 1996, pp.524–527. New York: IEEE.
47. Storn R and Price KV. Minimizing the real functions of
the ICEC’96 contest by differential evolution. In:
Proceedings of IEEE International Conference on
Evolutionary Computation, Nagoya, Japan, pp.842–844.
48. Botez RM, Boughari Y, Theel F, et al. Optimal Flight
Control on Cessna X Aircraft using Differential
Evolution. International Association of Science and
Technology for Development IASTED Modelling,
Identification and Control (MIC 2014), Innsbruck,
Austria, 17 Februay 2014.
49. Rogalsky T, Kocabiyik S and Derksen RW.
Differential evolution in aerodynamic optimization.
Can Aeronaut Space J 2000; 46: 183–190.
50. Wu CY and Tseng KY. Stress-based binary differential
evolution for topology optimization of structures.
532 Proc IMechE Part G: J Aerospace Engineering 231(3)

Proc IMechE, Part C: J Mechanical Engineering Science problem. In: 4th international conference on modeling,
2010; 224: 443–457. simulation and applied optimization (ICMSAO), 19
51. Yu WJ and Zhang J. Adaptive differential evolution April 2011, pp.1–5. New York: IEEE.
with optimization state estimation. In: Proceedings of 60. Tan MK, Chin YK, Tham HJ, et al. Genetic algorithm
the 14th annual conference on genetic and evolutionary based PID optimization in batch process control. In:
computation, 7 July 2012, pp.1285–1292. ACM. 2011 IEEE international conference on computer appli-
52. Ghazi G and Botez RM. Lateral controller design cations and industrial electronics (ICCAIE), 4
for the Cessna citation X with handling qualities December 2011, pp.162–167. New York: IEEE.
and robustness requirements. In: 62nd CASI aero-
nautics conference and AGM, Montreal, Quebec,
Canada, 2015.
53. Schirrer A, Westermayer C, Hemedi M, et al. Robust Appendix
H1 control design parameter optimization via genetic
algorithm for lateral control of a BWB type aircraft.
In: Intelligent control systems, 29 September 2010, Notation
pp.57–63. ESS steady-state error
54. Manocha A and Sharma A. Three axis aircraft auto-
Fl(P,K) lower linear fractional transformation
pilot control using genetic algorithms: An experimental
ISE integral square error
study. In: IEEE international advance computing confer-
ence, 6 March 2009, pp.171–174. New York: IEEE. K feedback gain
55. Wang HT, Guo SW, Guo P, et al. Application of gen- OS overshoot
etic algorithms for aerodynamic parameter estimation p, q, r angular speeds along Ox, Oy, Oz axis
of large parachute. J Astronaut 2010; 4: 007. Ts settling time
56. Shi Y, Qian W, Wang Q, et al. Aerodynamic parameter u, v, w speeds along the Ox, Oy, Oz axis
estimation using genetic algorithms. In: IEEE congress u(t) control vector
on evolutionary computation, 16 July 2006, pp.629–633. V total aircraft speed
New York: IEEE. W1 sensitivity weighting function
57. Patrón RF, Owono AO, Botez RM, et al. Speed and W2 complementary sensitivity weighting
altitude optimization on the FMS CMA-9000 for the
function
Sukhoi Superjet 100 using genetic algorithms. In: 2013
x(t) state space vector
aviation technology, integration, and operations (ATIO)
conference and international powered lift conference X, Z, Y aircraft aerodynamic forces
(IPLC), 2013. e , a , r elevator, aileron, and rudder deflections
58. Patrón RF, Kessaci A, Botez RM, et al. Flight trajec-  robustness criterion
tories optimization under the influence of winds using !n natural frequency
genetic algorithms. In: AIAA guidance, navigation, and , ,  pitch angle, sideslip angle, and roll
control conference, 2013, pp.19–22. angle
59. Zakaria MZ, Jamaluddin H, Ahmad R, et al. Effects of damping coefficient
genetic algorithm parameters on multiobjective opti-
mization algorithm applied to system identification