You are on page 1of 20

Aerospace Science and Technology 84 (2019) 464–483

Contents lists available at ScienceDirect

Aerospace Science and Technology


www.elsevier.com/locate/aescte

UAV aerodynamic design involving genetic algorithm and artificial


neural network for wing preliminary computation
Abdelwahid Boutemedjet a,∗ , Marija Samardžić b , Lamine Rebhi a , Zoran Rajić b ,
Takieddine Mouada a
a
Military Academy, University of Defence, Belgrade, Serbia
b
Military Technical Institute, Belgrade, Serbia

a r t i c l e i n f o a b s t r a c t

Article history: In this paper, the aerodynamic design procedure of a mini unmanned aerial vehicle, intended to perform
Received 1 February 2018 aerial reconnaissance at low altitude and low Reynolds number, was summarized. Design process was
Received in revised form 21 September divided into two major parts: conceptual design phase and preliminary design phase. During the
2018
conceptual design, classical procedures and data from similar unmanned aerial vehicles already designed
Accepted 27 September 2018
were employed to define the requirements related to unmanned aerial vehicle design. The preliminary
Available online 24 October 2018
design was performed using panel method where the emphasis was given on the design of the UAV
Keywords: wing, fuselage design and empennage. The wing planform parameters were determined through an
Unmanned aerial vehicle aerodynamic optimization process using both genetic algorithms and artificial neural networks. Finally
Preliminary aerodynamic design an aerodynamic analysis using panel method, Computational Fluid Dynamic simulations and wind tunnel
Wing design testing was carried out. Unmanned aerial vehicle full configuration design process was consistent, where
Genetic algorithm optimization a comparison between the final obtained results was carried out and showed an agreement in terms of
Computational Fluid Dynamic
lift, drag and pitch moment coefficients.
Wind tunnel testing
© 2018 Elsevier Masson SAS. All rights reserved.

1. Introduction Depending on the UAVs’ mission specifications and require-


ments, the design methodology may vary significantly and each
An unmanned aerial vehicle (UAV) refers to an aircraft that can UAV concept requires a different approach [3,4]. The UAVs design
be operated autonomously or remotely, and also they can carry process was the subject of many studies. For example, the main
a variety of payloads depending on their type, functionality, op- geometric, aerodynamic, stability and performance parameters of
erational characteristics and mission objectives. The UAV segment the URCUNINA-UAV were calculated and assessed through analyti-
experienced a great growth, driven primarily by the military com- cal presizing and computer-aided design methods, by P. Mosquera
munity and to a lesser extent by border security, environmental et al. [5]. Employing linear aerodynamic performance and Com-
monitoring and other application domains. This picture seems to putational Fluid Dynamic (CFD) optimization, S. Kontogiannis and
change, where UAVs are dominating civilian and public domain ap- J. Ekaterinaris [6] performed a complete study of a small-size UAV.
plications [1]. Nowadays, UAVs offer several advantages, generally P. Panagiotou et al. [3] carried out an aerodynamic design study
related to the absence of crew-on board and the rapid develop- for a MALE UAV involving both analytical/semi-empirical presizing
ments in electronic and video equipment, such as accomplishing methods and CFD simulations. S.G. Kontogiannis et al. [7] applied
critical tasks, reducing operational cost and increasing flight en- the linear strip theory to perform the wing aerodynamic design
durance due to weight reduction. Hence, they cover a wide spec- of the ATLAS IV-UAV. Using a combined approach of traditional
trum of applications, such as monitoring hazardous environment sizing methods and CFD simulations, a layout design study was
(nuclear/chemical contamination, fire detection etc.), search and conducted by P. Panagiotou et al. [8] to carry out a complete de-
rescue, border or security surveillance and reconnaissance mis- sign and aerodynamic analysis of Blended Wing Body MALE UAV.
sions [2]. Theoretical analytical methods were employed with CFD analy-
sis to develop conceptual methodology of a hybrid solar MALE
UAV, where an aerodynamic analysis and a solar power estimation
* Corresponding author at: Military Academy, University of Defence, Pavla Jurišića
technique were detailed by P. Panagiotou et al. [9] aiming to de-
Šturma 33, Belgrade, Serbia.
E-mail addresses: boutemedjet.abdelwahid.empfse@gmail.com (A. Boutemedjet), termine a compromised selection between fuel amount and panel
marija.samardzic@vti.vs.rs (M. Samardžić). area sizing. G. Romeo et al. [10] designed a high altitude very-long

https://doi.org/10.1016/j.ast.2018.09.043
1270-9638/© 2018 Elsevier Masson SAS. All rights reserved.
A. Boutemedjet et al. / Aerospace Science and Technology 84 (2019) 464–483 465

Table 1
UAVs comparison.

UAV Wing span Length Empty weight Total weight Payload Maximum speed Endurance Altitude Mission radios
(m) (m) (kg) (kg) (kg) (km/h) (h) (m) (km)
Azimut 2 2.9 1.82 7 9 2 120 2 300 10
Azimut 3 2.9 1.82 7 8 1 120 2 300 10
Bayraktar 2 1.2 3.5 5 1.5 – 1–1.5 1200 15
Bird Eye 400 2.2 0.8 4.4 5.6 1.2 – 1 300 –
Carolo t200 2 1.8 4.6 5.6 1 – 0.75 –
RAVEN RQ-11 1.4 0.9 1.9 2.2 0.3 95 1.5 30–300 10
EMT Aladin 1.46 1.4 2.6 3.2 0.6 90 0.5–1 3000 5
Fly eye 3.6 1.9 4 – – 170 2–3 1000 10
POINTER FQ M-151 2.74 1.8 3.4 4.3 0.9 80 1 300 5
Casper 250 1.8 1.38 4.45 5.7 1.25 80 2 2000 10
Puma 2.8 1.8 3.4 4.6 1.2 96 4 300 15
SpyLite 2.75 1.35 7.7 9 1.3 120 4 9100 50

endurance solar-powered UAV developing a computer program to 2. Research relevance


design the main UAV platform taking into consideration the ef-
ficiency of solar panels, aerodynamic, structural, flight mechanics Designing a low cost homemade mini UAV was one among the
and aeroelastic performances. G. Frulla and E. Cestino [11] per- tasks entrusted to the Military Technical Institute (VTI) exploiting
formed a set of static structural tests for a solar powered high the available capacities. The designed UAV intended to accomplish
altitude long-endurance UAV where the applied loads take into ac- reconnaissance missions at low altitude and low speed depending
count the effective aerodynamic and mass distributions estimated on camera capacities and filming photos qualities limits. The asso-
during the design of the UAV. The results of this study may help ciated challenges with such mission are related to design a small
in the design of such UAVs. size and light structure UAV that operates at low Reynolds num-
In the present study the aerodynamic design of a short range ber. Hence stall aerodynamic characteristics degradations, stability
mini UAV was presented and discussed. Based on the classical and propulsive system vibrations present a serious problem to deal
with. Furthermore, to minimize the time of UAV readiness during
methods, presented in [12–14], aerodynamic optimization CFD
the takeoff phase and makes landing possible even in the absence
simulations, the design process involved the design of the UAV
of flat areas, the UAV has to be hand launched and accomplishes
wing, fuselage and empennage. The first step in designing the UAV
landing with a parachute.
was to perform a market analysis by identifying existing UAVs
To deal with these challenges a high lift low Reynolds number
intended to accomplish aerial reconnaissance at low altitudes sim-
airfoil (SD7062) was selected to produce high lift coefficients guar-
ilarly to the UAV presented in this paper, Table 1 [15–21]. The data
anteeing boundary layer transition before the separation even in
presented in Table 1 are important to highlight some guidelines the case of high turbulence flow. The used airfoil provides also
in the estimation of the UAV specifications, for example empty excellent performance at low speed flight, including a high en-
weight, maximum speed, optimization design space and other pa- durance ratio [25]. In addition, a high wing configuration with a
rameters. large aspect ratio (AR) was chosen serving to decrease induced
During the preliminary sizing, emphasis was given on the wing drag, to improve stall characteristics and at same time to ensure
design where an aerodynamic shape optimization was evaluated a sufficient space for parachute.
using genetic algorithms (GAs) in order to determine wing plan- Classical aerodynamic configuration, with a T-tail and whose
form parameters. GAs are a popular revolutionary method widely wing has a dihedral angle, was selected on purpose to provide
used in aerodynamic optimization since they are very efficient in sufficient stability characteristic against free stream disturbances.
finding global optimum. One of the key features of GAs is that Compared to the chosen classical configuration, Blended Wing
they are tolerant of noise in the objective function and have no Body (BWB) configuration presents more favorable aerodynamic
difficulty with categorical variables or topology changes. In addi- efficiency which provides better endurance and capacity of carry-
tion, the development cost of GAs is minimal since they use only ing payloads and also characterized by a lighter structure. However
objective function and do not require its derivatives. The key disad- BWB configuration is more demanding in terms of longitudinal sta-
vantage associated with GAs is that they are computationally time bility and requires a launching from catapult. Besides that, such
consuming especially in aerodynamic optimization where a CFD configuration requires many wind tunnel tests during the devel-
solver is used for fitness function evaluation [22,23]. This makes opment phase which makes design cost expensive. This justifies
classical aerodynamic configuration choice.
the use of artificial neural networks (ANNs) an efficient way to re-
The UAV was equipped with an electric motor which reduce
duce computational time. ANN is one of reliable and fast methods
significantly vibration and noise signature for better filming qual-
of predicting aerodynamic coefficients. The use of such method re-
ity. Aiming to make the UAV as lighter as possible, composite ma-
duces the time and the cost of computation since ANN decouples
terials was used to manufacture the UAV intending to make hand
the aerodynamic solver from the optimization process where the
launching feasible by a single person.
GA operates simultaneously with ANN [24]. Besides the goal of designing a mini UAV intended to accom-
In the following sections, a brief summary of the UAV design plish aerial reconnaissance, the use of GA optimization coupled
methodology is provided. In section 2, conceptual design phase with ANN to perform design computation, can be considered as a
based on the initial mission requirements and data from existing contribution in minimizing UAV design process time with respect
UAVs is shown. In section 3, the methodologies for preliminary to other design methodology. In other side, performing aerody-
design are manifested, including wing, fuselage and empennage namic optimization with different objective functions provides an
design. Aerodynamic analysis involving panel method (PM), CFD extended list of optimal solutions and makes easy to choose the
and experimental tests are presented in section 4 besides the ob- adequate one. In addition, the reliability of the proposed approach
tained results. was well checked with CFD and experimental results.
466 A. Boutemedjet et al. / Aerospace Science and Technology 84 (2019) 464–483

Fig. 1. Statistical data of empty weight versus total weight.


Fig. 2. SD 7062 (14%) airfoil shape.

3. Conceptual design
process. Statistically, using Eqs (1)–(2), in the case of a fixed pay-
load weight of 1.5 kg, the total weight and empty weight were
The conceptual design is a procedure, in which the first esti-
estimated respectively 8.2 kg, 6.7 kg, however for a fixed total
mates regarding aerodynamics, configuration, size, weight and per-
weight of 7 kg, the empty weight and payload weight were found
formance are calculated. In the following sections, a brief summary
respectively 5.65 kg, 1.35 kg. The empty weight of 5.65 kg was fit-
of UAV conceptual design is presented.
ted to 5.5 kg; by minimizing the structure weight, where a small
part of wing bottom was immersed with UAV body and at same
3.1. Design requirements and objectives
time using honeycomb composite material which significantly con-
tributes to reduce the weight.
In the present study a mini UAV was selected to accomplish
reconnaissance missions. The camera technical limits required the 3.3. Airfoil selection
altitude to be 300 m. The total mass of the vehicle was limited
to 7 kg and the stall speed was kept down to low level, less than For the simplicity of construction no slats or flaps were used to
12 m/s. These limits were based on typical weight that can be car- produce high lift. However a high lift, low Reynolds number airfoil
ried and hand launched by a single person and at same time allow- section was selected to satisfy the main design requirements. Since
ing an easy transportation, ensuring good quality filming photos composite materials were selected to manufacture the wing, the
and reducing take off distance. As the most used UAVs for recon- choice of the adequate airfoil was not limited by airfoil thickness
naissance purpose, the selected UAV for this study was equipped constraints. Hence, regardless the airfoil thickness, the selection
with electric motor since it contributes to the simplicity of instal- criteria were the aerodynamic efficiency, maximum lift coefficient,
lation and the low costs. The power of motor was estimated about and minimum drag coefficient. Three airfoil sections were consid-
800 W. Rate of climb after takeoff was taken into consideration as ered during the choice of the airfoil: NACA 643-418, Wortmann
a requirement, equals to 5 m/s, which is related to climbing time FX63-137 and SD7062. The airfoils were simulated using PM and
to operative altitude as well as the safe of nearby people, espe- changes in maximum lift coefficient, maximum efficiency and drag
cially in a city environment. A maximum speed of 120 km/h and a coefficient at zero angle of attack with increasing Reynolds number
long endurance flying were required during the sizing of the UAV. are resumed in Table 3.
Analyzing results from Table 3, it is clear that airfoil NACA
3.2. Weight estimation presents small lift coefficient values thus justifying the exclusion
of such airfoil from the wing design. In other side Wortmann
Total weight W estimation was based on that the total weight FX63-137 airfoil presents greater lift coefficient and higher effi-
of the designed UAV was divided into payload weight W p and ciency compared to SD7062, however the last mentioned airfoil
empty weight W e . The empty weight included the structure, mo- produces lower drag with acceptable efficiency and lift coefficient.
tor, propeller, batteries, fixed equipment, and anything else not Hence SD7062 airfoil section (Fig. 2) was selected to be used for
considered a part of payload. The payload weight was consid- wing design process. The chosen airfoil is characterized by a max-
ered the weight of the reconnaissance equipment estimated about imum thickness of 13.98% at position of 27.22% and with a maxi-
mum camber of 3.97% at position of 38.82%.
1.5 kg. The total weight is expressed as:
The airfoil was analyzed at Reynolds numbers (Re) 2 × 105 ,
3 × 105 and 4 × 105 . The obtained results were compared to ex-
W = We + W p (1)
perimental measurements [26]. Fig. 3 shows that the results from
The empty weight was statistically estimated from historical X-foil analysis were generally well consistent with those from ex-
trends of existing UAV with similar requirements and configuration perimental measurements.
(Table 1). Fig. 1 presents the curve fit equation of empty weight Table 4 resumes the main aerodynamic characteristics of the
W e versus total weight W , the fit equation was given as follows: chosen airfoil that were derived from lift coefficient versus angle
of attack (α ) curve and drag polar.
W e = 0.85W − 0.27 (2) The comparison of experimental results and those computed
using PM revealed some differences between the two cases. For
In this study, an UAV configuration with a total weight of 7 kg C L max , airfoil presented experimental greater values which can be
and a payload of 1.5 kg was the objective of the weight estimation considered as a positive contribution to obtain greater lift, however
A. Boutemedjet et al. / Aerospace Science and Technology 84 (2019) 464–483 467

Table 2
UAV design requirements.

Flight specifications Equipment Aerodynamic configuration and characteristics Structure and motor
Total weight (W ) 7 kg Camera High lift Composite material
Payloads (W p ) 1.5 kg Avionic Low drag Operation temp −20 to 50 ◦ C
Altitude 300 m Battery Low Re Electric motor
Endurance 1h Parachute High wing
Stalling speed (V s ) 12 m/s T-tail
Motor power 800 W
Maximum speed (V M ) 120 km/h
Rate of climb (ROC) 5 m/s

Table 3
Airfoils comparison.

Re NACA 643-418 Wortmann FX63-137 SD7062


CL max (C L /C D )max C D α =0o CL max (C L /C D )max C D α =0o CL max (C L /C D )max C D α =0o
0,2 × 106 1.246 64.56 0.00868 1.704 77.61 0.01311 1.562 70.25 0.00832
0,3 × 106 1.247 77.87 0.00739 1.686 92.74 0.01042 1.561 83.13 0.00745
0,4 × 106 1.261 88.99 0.00678 1.693 104.11 0.00925 1.560 92.28 0.00698

Fig. 3. Aerodynamic characteristics of SD 7062 airfoil.

Table 4 C D = C D0 + C Di (3)
Experimental SD7062 Airfoil aerodynamic characteristics.
The parasite drag is given by the following equation:
Re CL C D α =0 (C L /C D )max
max

2 × 105
1.62 0.0114 63.4 C f i K i S wet
3 × 105 1.65 0.0104 78.2 C D0 = (4)
S ref
4 × 105 1.65 0.0095 87.8
Parasite drag contribution of each evaluated aerodynamic sur-
face is given in Fig. 4. From the diagram it is noticed that the wing
in the case of C D , experimental results were greater, and hence most contributes to generating aerodynamic drag. Such result was
it affected airfoil efficiency by decreasing its values. Despite that, expected since the UAV operates at low Reynolds numbers and low
it can be considered that the selected airfoil section SD7062 pre- altitudes where strong turbulence takes place.
sented acceptable results at different Reynolds number to be used The induced drag is given by Eq. (5)
for UAV wing design.
C Di = kC L2 (5)
3.4. Drag estimation where k = 1/(π ARe 0 ), AR is the aspect ratio equals to 11, e 0 is
the interference factor related to the aerodynamic interactions be-
UAV drag coefficient was computed from parasite drag (C D0 ) tween the wing/the empennage and the fuselage and it is equal to
and induced drag (C Di ). The drag was computed from main surface 1.78 × (1 − 0.045 × AR0.68 ) − 0.64. The obtained drag polar equation
(wing, fuselage and empennage), where the corresponding form is:
factor (K i ), reference area (S ref ) wetted area (S wet ) and skin fric-
tion coefficients (C f i ) of each surface were evaluated. C D = 0.0371 + 0.0396C L2 (6)
468 A. Boutemedjet et al. / Aerospace Science and Technology 84 (2019) 464–483

Fig. 4. Pie diagram of the parasite drag contribution of UAV surfaces.

3.5. Aerodynamic main parameters and performances

During the analysis of different UAV constraint design such as


the rate of climb, maximum and stall speeds, the first parame-
ters to design the wing and to select the propulsion system were
determined. It is essential that a credible estimation of the wing
loading (WL) and power loading (PL) is made. The wing loading
has a strong effect upon sized UAV weight since reducing wing
loading value, makes the wing larger which significantly increases
UAV drag. Besides that WL affects directly stall speed, climb rate,
take off distance and performances. The design point of the UAV
which minimizes both wing loading and power loading was found
WL = 83.12 N/m2 and PL = 0.19 N/W. At this stage, applying con-
dition for minimum thrust required for level flight allows finding
velocity at which thrust or drag is minimum and efficiency ratio
C L /C D is maximum. Such velocity was computed as follows:
 
Fig. 5. Wing preliminary design process.

 2W
 k
V min thrust = (7) computation of the UAV external shape to satisfy the performance
ρS C D0 requirements. In this paper a gross design of the airfoil section and
the main surfaces were conducted in X-Foil and XFLR5 [27]. Fig. 5
where ρ is air density at the operating altitude, the minimum ve-
summarizes the process of wing design. The process starts with
locity for minimum required thrust was found equal to 12.5 m/s. defining a work space based on data delivered from classical aero-
The lift coefficient for minimum drag in level flight was deter- dynamic formulations, data from existing UAV and experience from
mined by the following expression: previous researches related to UAV design domain. Once the work-
 ing space was defined, aerodynamic solver was decoupled from op-
C D0
C L min thrust = (8) timization through ANN helping to minimize time of computation.
k Defining propulsion system characteristics in order to determine
The value of lift coefficient for minimum drag in level flight performances is one of important tasks to be added to the op-
was found equal to 0.9. This optimal lift coefficient depends only timization. Combined to constraints, objective function definition
on aerodynamic parameters, so in case of changing weight, UAV leads to complete wing design using aerodynamic optimization.
can be flown at this optimal lift coefficient by varying just velocity After wing planform determination, preliminary computation was
or altitude (air density). finished by fuselage and empennage design to obtain the full con-
Flying at the velocity V min thrust corresponds to maximum aero- figuration of the designed UAV.
dynamic efficiency C L /C D which maximizes the range from oper-
4.1. Optimization process definition
ating altitude. The resulting maximum C L /C D ratio is determined
from the following expression:
After airfoil process selection, the first objective of the UAV

CL 1 preliminary design was to determine the optimum combination
= √ (9) of the wing planform geometry and the specified aerodynamic
CD max 2 C D0k
requirements. UAV wing was designed using an aerodynamic op-
The obtained maximum glide ratio C L /C D is 13.5. The corre- timization including single-point and multiobjective optimization,
sponding velocity and lift coefficient to maximum glide ratio are Fig. 6. The GA coupled with ANN was used to perform the opti-
the same velocity and lift coefficient related to minimum required mization, where the training data for the ANN was derived from
thrust numerical simulations using XFLR5. The optimization problem was
formulated in term of design variables, constraints and objective
4. Preliminary design functions.

The preliminary design is the process that advances the de- 4.1.1. Design variables
sign concepts by individually designing and sizing the major com- Trapezoidal planform geometry was chosen in this study in or-
ponents of UAV. The aerodynamic design starts with preliminary der to simplify the computation. The parameters related to the
A. Boutemedjet et al. / Aerospace Science and Technology 84 (2019) 464–483 469

respect to the root airfoil, decreases across the span and becoming
lowest at the wing tip. This is referred to as wing tip twisted down,
where wings are typically twisted between zero and five degrees.

4.1.2. Optimization constraints


Constraints were imposed on the rate of climb (ROC) at the re-
quired altitude (ROC), stall speed (V S ) and maximum speed (VM)
at sea level conditions, which arise from the requirements given
in Table 2. These constraints were added to the objective function
as penalty term. The penalty term was formulated as a weighted
quadratic loss function:

2
2
T = ω1 max 0, ROC ∗ − ROC (t ) + ω2 max 0, V S (t ) − V S∗


2
+ ω3 max 0, V M − V M (t ) (10)

where t is the design vector, ROC ∗ , V S∗ , V M


∗ are the required per-

formances, ωi =1,2,3 are weights applied to the penalty term. More


details about performances evaluation are presented in Section 4.2.
The use of loss function permits to aggregates all constraints
into one term by multiplying the value of each constraint by the
appropriate weighting coefficient. During the optimization, it was
found that there was difference between the objective function
and penalty term; therefore, scaling was performed by applying
Fig. 6. Wing shape optimization process using GA.
weights ωi .

4.1.3. Objective function definition


Objective function choice in any optimization problem is dic-
tated by aircraft design requirements. As an illustration, life cy-
cle, cost and profit are common used objective functions. In most
cases, it is very hard to connect design variables via objective func-
tions, thus, lower level objectives may be used. For UAV, achieving
Fig. 7. Wing shape. maximum endurance and minimum drag are required since they
are ideal lower level optimization objectives for the aerodynamic
Table 5
designer [19]. The choice of the objective function is a crucial
Wing planform parameters vs design variables. task to perform designing. In this study three different objective
Wing surface S = (C r + C t )b/2
functions were provided in order to explore the different possible
Aspect ratio AR = b2 / S optimal solutions at Reynolds number about 4.5 × 105 which was
Wing loading WL = W / S computed based on mean aerodynamic chord, for cruise speed of
Wing taper ratio TR = C t /C r
17 m/s. At first, the objective was to maximize endurance ratio
Wing tip twist angle θ 3/ 2
(C L /C D ), hence the objective function to be maximized is:

3
wing planform were considered as the design variables. Fig. 7 f = C L2 /C D − T (11)
shows the wing shape, where C r is the root chord,C t is the tip
chord, b is the wing span, θ is the wing tip twist angle and S where T is the penalty term from the constraints, in this case
is the wing surface. The aspect ratio AR, wing loading WL, wing ωi=1,2,3 = 10 was applied to the penalty term.
taper ratio TR, wing tip twist angle θ were chosen to reflect the The second objective was to minimize the drag coefficient at
effect of aerodynamic discipline during the 3D wing optimization angle of attack α = 0◦ , hence the objective function to be mini-
process. Table 5 resumes mathematical relationship between geo- mized is:
metrical wing planform parameters and chosen design variables to
 2
perform the optimization. Details of the geometry parametrization CD
f = 1− +T (12)
can be found in [28,29]. C ∗D
The optimization process requires a design space determination
basing on existing data from mini UAV already designed (Table 1) where C ∗D is the target drag coefficient, which was assumed to be
and from design variable first estimation. Designing a wing with unattainable within the specified constraints and its value is 0.008.
high aspect ratio values leads to maximum lift coefficient close to Weights applied to the penalty term were ωi =1,2,3 = 1.
those of infinite wing which justify the chosen values of the aspect Finally a multiobjective optimization was carried out where the
ratio. Basing on the existing values of wing span, a first estimation two competing objectives were given by:
of wing surface can be found; hence searching interval of wing
loading can be set. Since elliptical shape is ideal for minimizing  3
2
C L2 /C d
induced air resistance and reducing the minimum flight speed, the f1 = 1 − 3
+T (13)
taper ratio interval was set to ensure that the obtained wing shape (C L2 /C d )∗
would be similar to elliptical wing. Aiming to prevent tip stall and  2
to revise the lift distribution to approximate an ellipse, the wing Cd
f2 = 1 − ∗ +T (14)
is designed so that the airfoil angle of incidence, measured with Cd
470 A. Boutemedjet et al. / Aerospace Science and Technology 84 (2019) 464–483

4.2. Performance computation

“If the drag polar of an aircraft is accurately determined and


the thrust, or thrust horsepower, available is known, then all of
the performance characteristics of the subject airplane may be cal-
culated” [32]. Based on this approach, the thrust generated by the
Fig. 8. Aerodynamic coefficient approximation with ANN. propulsion system was accurately determined; hence it makes the
process of wing designing using the GA optimization more reliable
and efficient. At the electronics’ laboratory of VTI, the propulsion
The two function f 1 and f 2 are in competition because it is
system model was developed to ensure that it achieves the mis-
generally easier to produce low drag at low lift coefficient or en-
sion requirements.
durance ratio values. Consequently, the optimal candidates gener-
Generally, propellers for UAV must operate in small range of
ated using this optimization present the key advantage to produce
Reynolds numbers, which causes degradation in their efficiency co-
at same time maximum endurance and low drag, where the drag
efficient since at these Reynolds numbers viscous flow is dominant.
is minimized at the angles of attack corresponding to the maxi-
The maximum efficiency of such propellers may vary from 0.28 to
mum endurance ratio. The target endurance ratio and the target
0.6 which is far from the efficiency coefficient of classical aircraft
drag coefficient C ∗D are equal to 24 and 0.023 respectively.
which is about 0.85 [33,34]. In the aim to determine the optimal
The two objectives were combined using the weighted sum ap-
propeller parameters, a nonlinear optimization study was carried
proach as follows: out and followed by experimentally determination of the available
thrust (T A ) generated by the propulsion system [35]. Independence
f = ω f 1 + (1 − ω) f 2 (15) variables were propeller diameter (D) and pitch (H ). The objec-
tive function was maximum endurance (T prof ). Constraints were
where ω ∈ [0, 1]. applied on rate of climb (ROC) and static thrust force (T A ). A Li–Po
At this case ω presented the relative weight of the two objec- battery with nominal voltage of 29.6 V and nominal capacity of
tives. Such method for multiobjective function was used to provide 8 Ah, motor with maximum power of 800 W were used. The objec-
multiple solution points by varying the weight ω from 0 to 1. tive function defined depending on propeller diameter and pitch.
Hence, minimizing objective function f provides a sufficient con- In addition, the following constraints were entered to problem op-
dition for Pareto optimality, which means that the minimum of f timization: the rate of climb at zero altitude must be greater than
is always Pareto optimal [22]. 5 m/s with a static thrust force of 40 N.
The wing 3D design process can be expressed as a classical op- The propeller optimization problem was written as follows:
timization problem as follows:
⎡ ⎤ max T prof ( D , H ) (20)
3
f = −C L2 /C D + T
⎢ ⎥ Subject to:
Minimize f (t ) = Minimize ⎣ f = (1 − CC ∗D )2 + T ⎦ (16)
D
f = ω f 1 + (1 − ω) f 2 ROC ≥ 6 m/s (21)

Subject to: T A ≥ 40 N (22)

where 0.4 ≤ D ≤ 0.44 and 0.2 ≤ H ≤ 0.26


ROC (t ) ≥ 5 m/s (17) The determined parameters of the optimal propeller after per-
forming the optimization were: D = 0.4064 m and H = 0.2032 m
V s (t ) ≤ 12 m/s (18) with a maximum efficiency coefficient of 0.625 as shown in
Fig. 9.a.
V M (t ) ≥ 120 km/h (19)
The available thrust was determined at laboratory conditions
where design vector is t = [AR WS TR θ], design variable range is which considered as similar to sea level conditions. The maximum
defined in Table 6. thrust generated by the chosen propulsion system was found equal
to 44 N at air speed (V ) of 5 m/s (Fig. 9.b). The experimental data
4.1.4. Decoupling the aerodynamic solver from GA of available thrust in relation with flight velocity were fitted as
To make the process of optimization more simple and rapid, the follows:
flow solver was decoupled from the process, and the programmed
genetic algorithm was implemented with a trained ANN that gave T A = −0.0295V 2 − 0.0218V + 46.025 (23)
an approximation of aerodynamic coefficient, Fig. 8. These aerody- After accurately determining the available thrust, the performances
namic coefficients were used to evaluate the mentioned objective during the optimization process were evaluated basing on the
functions. The data base used to train ANN was obtained from a methods presented in [36].
computation set, evaluated using XFLR5 free code. Such code em-
Assuming steady level flight, the required thrust (T R ) was com-
ploys Vortex Lattice Method (VLM) to analyze the wing. The wing
puted from the equilibrium equations of the static movement of
is defined as a set of panels, that determined by the following pa-
the aircraft, and the obtained linear estimation is given as follows:
rameters: span, twist angle, root and tip chords. The principle of a
VLM is to model the perturbation generated by the wing by a sum W
of vortices distributed over the wing’s planform. The strength of TR = (24)
C L /C D
each vortex is calculated to meet the appropriate boundary condi-
tions, i.e. non-penetration conditions on the surface of the panels. where W is UAV weight expressed in N.
The wing viscous drag is estimated by drag values resulting from The available power ( P A ) and required power ( P R ) for steady
the X-Foil regenerated polar analysis. This code is well validated level flight can be found by multiplying both sides of Eq. (23) and
with tunnel test experiments and other CFD codes [30,31]. Eq. (24) by velocity, which becomes:
A. Boutemedjet et al. / Aerospace Science and Technology 84 (2019) 464–483 471

Fig. 9. Propulsion system.

PA = TAV (25)
PR = TRV (26)
While the endurance ratio is useful in characterizing an UAV be-
cause the characteristics associated with the maximum value of
endurance ratio are valid for the minimum value of required
power, which leads to maximum endurance. The required power
expressed in function of endurance ratio is given by the equation:

2 W3 1
PR = 3
(27) Fig. 10. Artificial neural network structure.
ρ S 2
C L /C D
basic aerodynamic coefficients were set as ANN output, Fig. 10.
Since the propulsion system produces smaller thrust increasing Among the different models of ANN, the Back Propagation algo-
flight altitude due to air density reduction, then the available rithm is one of the most known and widely used. This algorithm
power at an altitude h ( P Ah ) was approximated as follows: was the first practical method for training a multiple layer net-
ρh work.
P Ah = TAV (28) The input and output data generated from 256 simulations us-
ρ0
ing aerodynamic solver were used to train the network until it
where ρh and ρ0 are respectively the densities at the required al- approximated the aerodynamic coefficient functions. Each pair of
titude and at the sea level. data (the input and the corresponding output) went through two
The difference between available power and the required power phases. The forward pass involved presenting a sample input to the
represents the excess power. Rate of climb is equal to excess power network and letting the activations flow until they reach the out-
divided by the weight of the UAV and is given by Eq. (29). put layer. During the backward pass, the network’s actual output
(from the forward pass) was compared with the target output and
PA − PR
ROC = (29) errors were computed for the output units [37]. Both coefficients
W were interpolated with the same architecture of NN respecting the
The maximum value of lift coefficient C L max can be used to ex- corresponding data set.
tract UAV minimum velocity, this minimum velocity corresponds Defining the architecture of ANN depends on the number of
to stalling speed described by the equation: layers, the number of hidden neurons, the type of transfer func-
 tion, the training algorithm and the validation of ANN. For inter-
2 W 1 polating the basic aerodynamic coefficients, a three layer neural
VS = (30) network with input layer, hidden layer and output layer was cho-
ρ0 S C L max
sen. This type of ANN was considered in this case because of the
Since XFLR5 was not able to predict the exact value of C L max , it limited available data. Concerning the number of neurons, the in-
was approximated as the last value interpolated from the regener- put layer in the case of a single angle of attack was set with four
ated polar in order to simplify computation. input neurons (WL, AR, TR and θ ) while the output layer had just
UAV maximum speed was obtained as the maximum speed re- one neuron (C L or C D ). The appropriate number of hidden neu-
lated to zero excess power between the available and the required rons was determined through testing several networks where the
powers. number of hidden neurons was increased from one to five times
the number of input neurons. A number of 20 hidden neurons
4.3. Artificial neural network provided acceptable efficient fitting results. The use of 20 hid-
den neurons agrees with the assumption proven by T. Rajkumar
In this study, an ANN was introduced to define the process or and J. Bardina [38] which postulates that the number of hidden
the relation that associates each element of an input set to a single neurons equals five times the number of input neurons provides
element of the output set. During the ANN building the optimiza- good prediction of aerodynamic coefficients. The transfer function
tion design variables were considered as ANN input, however the is a continuous real-valued function that determines the neuron
472 A. Boutemedjet et al. / Aerospace Science and Technology 84 (2019) 464–483

Fig. 11. Network generalization regression analysis of C L (a), and C D (b).

Fig. 12. Histogram of percentage error.

output values from the inputs. For this study, Tanh transfer func- Fig. 12. The average of percentage error for C L network was 11.6%,
tion was selected because it yields fast training since it produces which was considered as an acceptable range; however the aver-
both positive and negative values. Due to its guarantied conver- age percentage error reached the value of 15.57% in the case of
gence and its numerical robustness, Levenberg–Marquardt method drag coefficient prediction, this due to small variations of the drag
was selected to train the network. For more efficiency, the inputs coefficient from data base input. Most simulated candidates were
and outputs were scaled to range [−1, 1]. Network training error distributed in the interval that represents error of 0%–10%. His-
was performed using mean-square error between theoretical re- tograms showed also that 70% of tested candidates within error
sults and those computed by the network. More details about the between 0% and 20% concerning the C D coefficient compared to
architecture of the used NN are presented in [24,38] 77% concerning C L coefficient, which is considered acceptable.
The response of the modeled networks concerning the lift and According to the diagrams of error distribution (Fig. 13), the
drag coefficients was presented by performing a regression analy- maximum achieved errors for C L and C D are 290% and 272% re-
sis between networks’ outputs and the associated theoretical data. spectively. And it was noted that the error is regularly spaced in
The outputs of the ANN were plotted versus the theoretical values the whole test domain. According to ANN error percentage and
with open circle points, the dashed line presented the ideal fit, the distribution analysis, the trained ANN presents acceptable charac-
correlation between network and theoretical data was given by the teristics; however, the ANN requires design modifications to fur-
solid line, Fig. 11. The trained networks showed acceptable linear ther reduce the percentage error and ameliorate data scatter.
fit and exhibited values of R-correlation near to 1.
In spite the R-correlation values showed an excellent fit be- 4.4. Genetic algorithm
tween the outputs of the trained networks and the theoretical
data, the analysis of the networks performances based on pre- In the present GA optimization, design candidates, design vari-
diction errors is more suitable. The error was performed as the ables and objective function are corresponding respectively to the
percentage error between network simulation and theoretical data, chromosomes, gens and fitness. During solution initialization, the
A. Boutemedjet et al. / Aerospace Science and Technology 84 (2019) 464–483 473

Fig. 13. Error distribution in test domain.

Table 6
Test domain.

Design variables xmin xmax


Wing loading [kg/m2 ] 8 10
Aspect ratio 8 11
Taper ratio 0.55 0.83
Twist angle [◦ ] −5 0

population size was set to twenty chromosomes (NC = 20) and the
first generation was randomly initialized according to each gen up-
per and lower limits, Table 6.
After the initialization, objective function is evaluated for each
candidate. Then the population is optimized according to each
chromosome objective function value. Hence, a process of rank-
ing is used where the most fitted chromosome is given a number
Fig. 14. Convergence history for endurance ratio maximization.
ranking one until the last most fitted individual is given a number
ranking NC. The overall process consists of selection, pass through,
crossover, perturbation mutation and mutation. Selection is the 4.5. Optimization results
first operation required to determine the next generation, where
4.5.1. Endurance ratio maximization
the chromosomes are copied in mating pool according to their
The convergence history of the genetic algorithm is plotted in
rank. Pass through is an elitist strategy operator where twenty Fig. 14. The plotted function is the endurance ratio since it is ap-
percent of the fittest chromosomes are directly transferred into proximately equal to objective function except a small difference
the next generation to ensure that the fitness never drops from due to contribution from constraints. The function was increased
generation to generation. The next implemented GA operator is from 19.8 to 23.62 after 47 iterations, after 400 iterations (16000
called Crossover. This operator is used to create new genes from calls to the NN) the GA produced a wing with endurance ratio of
parent chromosomes that are randomly selected. A simple one- 23.65.
point crossover operator was used to determine twenty percent
of chromosomes for the next generation. Mutation is a GA oper- 4.5.2. Drag coefficient minimization at α = 0◦
ator that used to generate thirty percent of the next generation The corresponding convergence history of drag minimization is
shown in Fig. 15. Since the objective function related to drag min-
chromosomes where it was applied to randomly selected genes of
imization had low values, the plotted function was 1/ f . After 50
each chromosome. Perturbation mutation is the last used GA oper-
iterations, the equivalent function is increased to 28.14 for a drag
ator which is implemented in the same way as mutation; however,
coefficient of 0.00951. After 400 iterations, the produced wing gen-
gen’s value is modified as follows:
erates a drag coefficient of 0.0095 corresponding to an equivalent
  function value of 28.33.
xk+1 = xk + (xmax − xmin ) R (0, 1) − 0.5 β (31)
4.5.3. Multiobjective optimization
where xk is the chosen gene first value and β is a specified toler- In this optimization, the two-dimension Pareto front was pro-
ance. β is chosen equal to 0.001 to avoid gene’s values, modified duced by genetic algorithm using the weighted sum approach and
by this operator, to exceed its limits as given in Table 6. The num- the gradient-based technique and. In this optimization, the gra-
ber of modified chromosomes using this operator is set to thirty dient technique was used in order to verify the obtained results
percent. The used GA was built at the VTI laboratory basing on the with GA. The generated front using gradient based technique can
details presented in [39,40]. be considered as the true Pareto front, since there is a substantial
474 A. Boutemedjet et al. / Aerospace Science and Technology 84 (2019) 464–483

Fig. 15. Convergence history for drag minimization. Fig. 16. Pareto front for the two objectives of drag and endurance ratio.

Table 7
Genetic algorithm results with weighted sum approach.
3/2
Candidate No ω f1 f2 (C L /C D ) CD
1 0.01 0.02833 0.05801 19.960 0.028540
2 0.1 0.02222 0.05841 20.422 0.028559
3 0.2 0.01511 0.05997 21.050 0.028633
4 0.3 0.01306 0.06018 21.256 0.028643
5 0.4 0.01247 0.06250 21.320 0.028750
6 0.5 0.01136 0.06120 21.441 0.028690
7 0.6 0.00178 0.07954 22.988 0.029487
8 0.7 0.00120 0.08071 23.167 0.029534
9 0.8 0.00075 0.08366 23.342 0.029652
10 0.9 0.00061 0.10502 23.406 0.030454
11 0.99 0.00051 0.10913 23.458 0.030598

reduction in the gradient for each point. Furthermore, the front is


convex, so the weighted sum approach is effective [41]. For the ge-
netic algorithm, in order to simplify the process of optimal wing
selection, the weighted sum approach was realized with eleven
values of ω from 0.01 to 0.99, Table 7.
The generated fronts were plotted in Fig. 16, where drag and
endurance ration objectives were respectively plotted on y-axis
and x-axis. The obtained results presented small differences be-
tween the two methods due to convergence errors. When ω ap-
proaches the value 1, the endurance ratio objective is close to zero,
since the target endurance ratio is achievable; the produced wing Fig. 17. Stream lines and pressure coefficient distribution at.
in this case is characterized by a maximum endurance ratio of
23.46 and a drag coefficient of 0.0306. On other hand, the drag
majority of optimization constraints and design requirements were
objective does not approach zero with ω = 0.01, since the set drag
satisfied, where the maximum achieved speed and rate of climb, as
target is unattainable. The obtained wing gives a maximum ratio
shown in Fig. 18, were respectively 130 km/h and 8 m/s. On other
of endurance about 19.96 and drag coefficient of 0.0285.
side, in spite the PM is not able to determine the exact maximum
At the end of the optimization, a data set of 13 wings was
lift coefficient, the obtained stall speed of 10.22 m/s satisfied stall
formed; 11 candidates were obtained from Pareto front, Table 7,
constraint.
and two candidates from endurance and drag optimization. The
The final obtained wing (Fig. 19) and the optimal wing
adequate wing used to design the UAV of this study was selected
(Fig. 17.b) have almost the same dimensions, aerodynamic charac-
among this set of wings. The process of wing selection was car-
teristics and performances. The aerodynamic analysis of this wing
ried out by a compromise between the following considerations:
is shown in Section 4.
objective functions values, wing planform parameters and the ob-
tained performances. The data from mini UAV of the same category
4.6. Fuselage design
were also exploited in the decision of the selection of the appro-
priate wing. Finally the wing number 10 from the Pareto set was
selected to be mounted on the concerned UAV, Fig. 17.a. The UAV fuselage is designed in such a way that it ensures
The selected wing was simulated using XFLR5 and its evaluated enough room for the electro-optical equipment, electric motor and
performances and characteristics are resumed in Table 8. It was re- batteries to be placed, and at the same time minimize aerody-
markable that there was a small difference in maximum endurance namic drag. Fuselage length and maximum diameter provided the
ratio since the shown value in Table 7 was computed with ANN, basis aerodynamic design of the fuselage. Fuselage fineness ratio
however the given value in Table 8 was determined using PM, this is defined as the ratio between the fuselage length and its maxi-
is due to the error generated by ANN. It was observed also that the mum diameter. Theoretically, the subsonic drag is minimized by a
A. Boutemedjet et al. / Aerospace Science and Technology 84 (2019) 464–483 475

Table 8
Optimal wing characteristics.

W /S 9.25 S 0.75 m2 VS 10.22 m/s α for (C L3/2 /C 1D/2 )max 5.5◦


3/2 1/2
AR 10.5 b 2.818 m VM 36.2 m/s C D for (C L /C D )max 0.0296
TR 0.625 Ct 0.206 m ROC max 8 m/s C D α =0◦ 0.0115
−2.00◦
3/2 1/2
θ Cr 0.330 m (C L /C D )max 22.96 (C L /C D )max 29.26

Fig. 18. Optimal wing performances analysis.

section is not a circle, an equivalent diameter (de ) is calculated


from the cross-sectional area as follows:
l22
1.55(π 4
+ l1l2 − l22 )0.625
de = (32)
(π l2 + 2l1 − 2l2 )0.25
where l1 and l2 are the length of major and minor side respec-
tively.
At the end the equivalent diameter and fuselage length were
found equal to 0.65 m and 1.94 m respectively. These results pro-
vided a fineness ratio of 2.98.

4.7. Empennage design

In this study the conventional T-tail configuration was selected


to be connected to the UAV body via a boom. The design of the
Fig. 19. Final wing planform parameters.
tail planes was carried out using the classical methodology based
on static stability criteria. A longitudinal analysis was carried out to
allow the UAV to perform a stable flight, taking into consideration
fineness ratio of about 3.0 [12]. According to the mentioned con- the allowed static margin referenced to the UAV neutral point. As
siderations above, the fuselage maximum cross section was chosen a result from this analysis, the trim angle of the designed UAV was
similar to an oblong cross section, Fig. 20. Since the fuselage cross found 1◦ (Fig. 21). The static margin and the neutral point were

Fig. 20. Fuselage geometry.


476 A. Boutemedjet et al. / Aerospace Science and Technology 84 (2019) 464–483

Table 9
Similarities between the designed UAV and the existing UAVs from market analysis.

UAV Wing span Length Total weight Payload


(m) (m) (kg) (kg)
√ √ √ √
Azimut 2
√ √ √
Azimut 3 –

Bayraktar – – –

Bird eye 400 – – –

Carolo t200 – – –
RAVEN RQ-11 – – – –
EMT Aladin – – – –

Fly eye – – –
√ √
POINTER FQ M-151 – –

Casper 250 – – –
√ √ √
Puma –
√ √ √
SpyLite –

the full UAV configuration analysis involved three different tech-


niques: PM simulation, CFD simulation and wind tunnel testing.
Fig. 21. UAV pitching moment (C m ). The followed methodology for the aerodynamic analysis focused on
analyzing the Reynolds number similarity. For that aim, a scale 1:1
respectively estimated at 15.81% and 47.06% of the mean aerody- 3D solid model of the UAV was designed and manufactured. Fur-
namic chord. Furthermore, the horizontal plane was formed from thermore comparisons between the results of different techniques
the symmetrical airfoil NACA0008 [42] which tilted with an an- were carried out in term of the basic aerodynamic coefficients as:
3/ 2
gle of 2o and characterized by a surface area of 0.11 m2 , a span C L max , C D α =0 , (C L /C d )max (C L /C d )max and trim position.
of 0.68 m and a chord ratio of 0.81. In similar way the vertical
plane characteristics was specified where the determined surface 5.1. CFD analysis
area was 0.07 m2 , a span of 0.62 m, a chord ratio of 0.80 and the
same symmetrical airfoil NACA0008 was used to form the vertical The CFD analysis was carried-out with the commercial code
stabilizer. Ansys Fluent [43]. The analysis based on resolving Reynolds-
Averaged–Navier–Stokes equations (RANS), coupled with the Shear
4.8. Final concept Stress Transport (k-ω SST) turbulence model. The Shear Stress
Transport turbulence model was used since it models typical low
The final UAV concept and the main dimension obtained per- Reynolds aerodynamic problems [44]. The computational domain
forming design computation are shown in Fig. 22. Compared to was selected as a parallelepiped with a 4 m × 4 m inlet section
the data from existing UAV, The designed UAV presented a well and a length of 12 m, Fig. 23. The whole domain was meshed
correspondence. In term of wing span, length, total weight and with tetrahedral elements generating approximately 2.5 million el-
payload, the designed UAV seems similar to Azimut 2/3, Pointer ements for the wing and 4 million elements in the case of the
and Puma UAV, Table 9. The UAV aerodynamic characteristics and UAV. In order to properly capturing boundary layer turbulent flow,
performances will be discussed in the following section. 15 inflation layers were implemented on the walls, the first of
which was placed at 5 × 10−5 m from the wall generating y +
5. Aerodynamic analysis approaching 1. The boundary conditions correspond to UAV flight
conditions, where the freestream velocity was set at 17 m/s, and
During the aerodynamic analysis phase, the optimal wing cho- the ambient temperature and pressure corresponded to the alti-
sen during the optimization process was analyzed using only the tude of 300 m. Symmetry condition was imposed in the half of the
3D PM implemented in XFLR5 and the CFD simulation, however longitudinal plane of the model. The walls of the computational

Fig. 22. UAV final design dimensions expressed in meter.


A. Boutemedjet et al. / Aerospace Science and Technology 84 (2019) 464–483 477

Fig. 23. Computational domains and grid generation for the numerical simulations.

domain were defined as ideal walls where free slip condition was
imposed. The wing and UAV surfaces were treated as no-slip wall.
The Reynolds number was calculated equal to 4.5 × 105 , based
on the mean aerodynamic chord. A range of angle of attack from
α = −2◦ to α = 22◦ was examined.

5.2. Experimental setup

The UAV model was tested in the T-35 wind tunnel of the VTI
[45]. The T-35 wind tunnel is large subsonic, continuous, closed
circuit pressurized wind tunnel. It is driven by a variable pitch fan,
powered by four AC motors with a maximum power of 7.2 MW.
Mach number from 0.1 to 0.5 can be achieved in the test section
and it is regulated to 0.5% of its nominal value. Stagnation pres-
sure can be between 1.0 to 1.5 bars and it is regulated to 0.3% Fig. 24. The UAV model in the T-35 wind tunnel test section.
of its nominal value. Reynolds number capability is up to 12 mil-
lion per meter. According to turbulence level in T-35 wind tunnel,
free stream turbulence is below 0.09%. The UAV model was tested
in the 3.2 × 4.4 m octagonal test section with quadrant/tail-sting A data reduction was performed using standard T-35 wind tun-
model support, Fig. 24. The model was supported by bent sting nel software package. The data acquisition system consisted of
which enabled angles of attack in the range −6.5◦ ÷ +23.5◦ . The 64-channel Neff 620/600 system controlled by VAX 8250 computer.
bent sting was mounted on a pitch and roll mechanism which en- The data from all analog channels were digitalized by a 16-bit
abled both step-by-step and continuous model movement. Position resolution A/D converter. The overall accuracy of the acquisition
accuracy of the model support was 0.1◦ in pitch and roll. During system was about 0.05% to 0.1% FS of the channel signal range. All
the tests, the stagnation pressure in the test section was measured channels were sampled with the same 200 sample/s rate.
by Mensor quartz bourdon tube absolute pressure transducer. The Aerodynamic forces and moments acting on the model were
measuring range of such transducer is 1.65 bar and its nonlinear- measured using an internal six-component monoblock strain gauge
ity and hysteresis is about 0.02% of its full scale (FS). The static balance, Fig. 25. The balance diameter is 35 mm. Table 10 lists its
pressure was measured by differential pressure Druck transducer design loads and gives a summary of achieved accuracy of the cal-
with a range of 0.07 bar. Nonlinearity and hysteresis of the Druck ibration as obtained in checkout after the calculation matrix. The
transducer is about 0.05% FS. results of the balance calibration have a 2-sigma level of confi-
478 A. Boutemedjet et al. / Aerospace Science and Technology 84 (2019) 464–483

maximum reached lift coefficient is 1.35. Hence one of the mains


aerodynamic requirements related to stall speed was achieved. The
wing is characterized with a soft stall at high angle of attacks
due to the use of high-lift airfoil, which is very adequate aero-
dynamic behavior characterizing UAVs operating at low Reynolds
number [5].
According to the wing aerodynamic efficiency, the lift to drag
coefficient for both analyses had the same tendency. The max-
imum value computed using PM and CFD analysis were 28.8 at
α = 2◦ and 24.3 at α = 4◦ respectively; such difference is mainly
justified by the fact that PM predicts low drag values. It was ob-
served from the wing drag polar that there was a similarity in the
tendency of the two curves generated by CFD and PM. The values
obtained from drag polar using PM were close to numerical results
Fig. 25. Six-component strain gauge balance. especially at the front of polar.
The UAV full configuration was obtained assembling its de-
Table 10 signed parts, and it was simulated using CFD. It was observed
Balance calibration results. that UAV’s aerodynamic coefficient was improved introducing an
Component Design Uncertainty Standard deviation adequate winglet. To show winglet’s contribution in ameliorating
load (%FS) of the errors (%FS) aerodynamic behavior, the UAV full configuration was simulated
Axial force 850 N ±0.112 0.03 with and without winglet, a comparison between the two cases
Side force 1320 N ±0.173 0.049 is resumed in Fig. 26. Compared to configuration without winglet,
Normal force 2560 N ±0.052 0.017
whose maximum lift coefficient was 1.41, the maximum lift co-
Rolling moment 75 Nm ±0.16 0.054
Pitching moment 150 Nm ±0.115 0.029 efficient value related to the configuration with winglet was 1.59
Yawing moment 75 Nm ±0.221 0.053 (Fig. 27.a). This was explained by the minimization of tip vortices
effect (Fig. 28) and the additional lifting surface of the mounted
winglet. Fig. 27.b shows the plot of C D vs C L2 , as it is noted that the
dence. More details about the selected six component strain gauge
two curves have the same slop which means that the two configu-
balance and its calibration is presented in [46].
rations produce same induced drag. This is attributed to the small
The non-dimensional aerodynamic coefficients were obtained in
size of used winglet where there is just relatively small decrease
the following forms: in total drag as shown in Fig. 27.c, consequently, regarding lift to
RD drag comparison (Fig. 27.d) the used winglet has no contribution
CD = (33) to improve maximum lift to drag ratio. Related to those results, a
qS ref
parametric winglet design procedure, as described in [47], could
RL further increase aerodynamic performance
CL = (34)
qS ref Observing the results in Fig. 29.a, from angle of attack −2◦ to

11 , the three curves present a linear behavior with similar slope
MM
Cm = (35) in spite of noting different values of C L at α = 0◦ . The obtained lift
qlS ref
coefficient was overestimated using PM since the simulation was
where R D , R L are axial and normal forces, M M , is pitching mo- carried out considering only the wing and the empennage. The
ment which are expressed in body axis coordinate system, S ref is absence of fuselage eliminates the aerodynamic flow interaction
model reference area (wing area), l is wing mean chord and q is induced by the junction with the wing, which explain the cause of
dynamic pressure. The Mach number and dynamic pressure in the obtaining high lift values using XFLR5. Additionally, as well as for
test section were calculated using the isentropic relations: the wing, the PM cannot predict the stall effect; hence there is no
variation in the slope of the curve for high angle of attacks. Con-
  γ γ−1 12
2 p0 trary to the PM curve, numerical and experimental curves exhib-
M= −1 (36) ited the same behavior even for stall region. Both curves presented
γ −1 p st
pre-stall angle of 12◦ which is a favorable aerodynamic behavior,
γ p st M 2 such characteristic allow the UAV to avoid crash risks due to stall
q= (37)
2 effect when approaching stall angle. In addition the two curves
where p 0 is stagnation pressure, p st is static pressure and γ is reached similar maximum values of lift coefficient, 1.57 for experi-
heat capacity ratio. mental and 1.59 for numerical curve. Such lift coefficient value was
the consequence of using high lift low Reynolds number airfoil and
5.3. Results at same time employing winglet at wing tips.
The drag coefficient results are shown in Fig. 29.b. Similar to
The CFD results of the optimal wing are presented in Fig. 26. isolated wing results, PM presented lower values of drag coeffi-
The obtained results was used to determine wing aerodynamic cient since, firstly; such method could not exactly predict viscous
characteristics, and at same time to be compared with those ob- drag, secondly; the results were obtained without considering the
tained from PM computation. According to the lift coefficient re- fuselage. Furthermore, for angles of attack less than 11◦ , both nu-
sults, there was a good agreement in the linear region between merical and experimental results were well consistent, where the
the two analyses; however, for high angles of attack PM showed drag coefficient values at zero angle of attack were 0.041 and
no variation in the slope lift curve since it was not able to exactly 0.038 respectively. However, for high angles of attacks, experimen-
predict viscous and stalling effects. The drag coefficient values sim- tal curve exhibited higher values. The raison of that difference was
ulated with PM were lower due to the weakness of this method deduced during the experimental test, where it was observed that,
to predict the viscous drag. The CFD analysis showed that the at high angles of attack, the UAV tail started vibrating under the
A. Boutemedjet et al. / Aerospace Science and Technology 84 (2019) 464–483 479

Fig. 26. Wing aerodynamic characteristics.

effect of stall detachments. Hence the measured UAV drag was in- The static pressure distribution and the streamlines on the UAV
creased under the direct influence of the tail vibrations. surfaces are presented in Fig. 30. The study of the flow field was
The aerodynamic efficiency curves are resumed in Fig. 29.c. The carried out in order to explain the UAV aerodynamic behavior, to
corresponding angle of attack to the maximum lift to drag ratio us- get information about the location of the flight control surfaces
ing PM was α = 2◦ , in other side both numerical and experimental and to determine the aerodynamic loads to be applied on the UAV
curves presented similar values at similar angles, being 15.82 and structure analysis and sizing. For the angle of attack 0◦ it was ob-
15.31 at α = 6◦ respectively. Therefore, this angle of attack is cor- served that the streamlines are attached to the UAV surface and
responding to the minimal required thrust for accomplishing the no significant variation of the flow field was observed. However
reconnaissance mission. for the angle of attack 16◦ , fluid separation areas appeared on the
Fig. 29.d presents the variations of pitching moment coefficient wing and between the horizontal and the vertical stabilizers. This
at the center of gravity versus the angles of attack. It was noted is manly caused by the stall effect that takes place at high angles
that the three methods gave similar results especially for low an- of attack.
gles of attack where the curves presented positive moment coef- During UAV performance evaluation, the most important in-
ficient at zero angle of attack, similar trim position α = 1◦ and formation that has to be derived is the drag polar. Combined
negative slope. Thus the longitudinal stability conditions for ensur- to the propulsion system available thrust, results from PM, CFD
ing UAV balance were satisfied. and wind tunnel tests were used to calculate UAV important per-
Table 11 summarize the main results obtained during the pro- formance parameters (Fig. 31) including required power, rate of
cess of aerodynamic design, it was concluded that both numerical climb, maximum velocity and stall speed. Fig. 31.a shows the
and experimental results of the UAV final configuration presented available power and the required power evaluated using results
well concordance between them. Regarding the values evolution of from different analysis. The different methods have almost the
drag coefficient, it was observed that the coefficient experimental same stall limit which corresponds to a stall speed of 9.6 m/s.
value was greater than the value estimated in the other cases; this The obtained stall speed indicates that the designed UAV provides
was reflected directly on the UAV performance, since it increased enough maximum lift coefficient to fulfill stall constraint. Regard-
the required power Eq. (12). In spite of that, the designed UAV ing the achieved maximum speed, using PM the speed was found
satisfied the most of requirements with the adjustment of some 126 km/h which is in good agreement with V M constraint, how-
parameters. ever, the reached speed using CFD and experimental results were
480 A. Boutemedjet et al. / Aerospace Science and Technology 84 (2019) 464–483

Fig. 27. UAV with winglet comparison.

Fig. 28. Flow field and vortices near the final wingtip. α = 16◦ .

respectively 113.76 km/h and 110.16 km/h. In this case, the UAV 6. Conclusion
cannot reach the desired V M since the obtained drag is greater
than the one computed using PM. According to Fig. 31.b, the ob- The design process and the related aerodynamic analysis of
tained maximum values of ROC using PM, CFD and experimental a low cost homemade UAV designed to accomplish aerial recon-
tests were respectively: 7.81 m/s, 6.77 m/s and 6.53 m/s. It is naissance at low altitude were developed. During the conceptual
clearly that the UAV withstands with constraint related to ROC. design traditional sizing methods were used to carry out the first
As it is illustrated, there is a similarity between the performances UAV layout. The concepts of the previous phase was developed em-
computed using both of CFD and experimental results which is the ploying aerodynamic optimization to design the wing. PM and CFD
projection of similarity in aerodynamic characteristics. were also used to design the UAV wing and fuselage. Besides that,
A. Boutemedjet et al. / Aerospace Science and Technology 84 (2019) 464–483 481

Fig. 29. Results of the aerodynamic characteristics of the UAV.

Table 11
Results resume.

Aerodynamic parameters Isolated wing Full configuration without winglet Full configuration with winglet
Numerical Panel method Numerical Panel method Numerical Experimental
CL max 1.35 – 1.41 – 1.589 1.574
α for C L max 16◦ – 14◦ – 18◦ 17◦
C L α =0◦ 0.282 0.308 0.313 0.432 0.315 0.367
C D α =0◦ 0.0162 0.0128 0.032 0.0205 0.0321 0.0413
(C L /C D )max 24.29 28.83 15.54 25.96 15.31 15.81
α for (C L /C D )max 4◦ 2◦ 4◦ 2◦ 6◦ 6◦
3/2 1/2
(C L /C D )max 19.91 23.62 14.77 22.05 14.70 15.56
α for (C L3/2 /C 1D/2 )max 6◦ 6◦ 6◦ 4◦ 6◦ 6◦
C m α =0◦ – – – 0.014 0.011 0.015
αtrim – – – 1◦ 0.58◦ 0.96◦

the obtained configuration was tested in wind tunnel on purpose During the process of wing preliminary design using optimiza-
to check the reliability of the used design approach. tion coupled with GA and ANN, it was concluded that such method
Regarding the different obtained results from the different anal- was efficient for the aerodynamic UAV designing using PM, since
ysis, the UAV fulfills almost the whole design requirements and the numerical results showed that the isolated wing and the full
exhibits similarity with existing UAV from market analysis in terms configuration without winglet exhibited similar aerodynamic char-
of wing span, UAV length, total weight, payload and maximum acteristics. This strongly confirms that the wing is the main part
speed. The reached maximum speed of 110 km/h is smaller than characterizing the UAV aerodynamic behavior. The main differ-
the required speed 120 km/h. To deal with this problem in future ence between the different results was observed in drag estimation
works; a propeller efficiency optimization is suggested in order to which leaded to a deviation in results especially those of aerody-
increase the available power, and also a drag minimization also is namic efficiency. Therefore, as a perspective for future work, it is
recommended to reduce required power. very advisable to take into consideration the underestimated drag
482 A. Boutemedjet et al. / Aerospace Science and Technology 84 (2019) 464–483

Fig. 30. Static pressure distribution and streamlines on the UAV surface.

Fig. 31. UAV performances.

values during the sizing process by following enrichment strategy [2] R. Austin, Unmanned Aircraft Systems: UAVs Design, Development and Deploy-
during optimization introducing results from high fidelity solver ment, vol. 54, John Wiley & Sons, 2011.
(for example CFD) to form a multi fidelity aerodynamic solver. [3] P. Panagiotou, P. Kaparos, C. Salpingidou, K. Yakinthos, Aerodynamic design of
a MALE UAV, Aerosp. Sci. Technol. 50 (2016) 127–138, https://doi.org/10.1016/
As noted that the maximum efficiency was set at low angles of j.ast.2015.12.033.
attacks, so improving drag estimation at these incidences leads [4] A.C. Watts, V.G. Ambrosia, E.A. Hinkley, Unmanned aircraft systems in remote
to approaching performances/constraints to their real values. In sensing and scientific research: classification and considerations of use, Remote
addition introducing the empennage to optimization process and Sens. 4 (2012) 1671–1692, https://doi.org/10.3390/rs4061671.
adding more parameters in objective functions relating to UAV tail [5] P.D.B. Mosquera, L.B. Bolivar, D.A. Giraldo, H.D.C. Muñoz, Aerodynamic design
analysis of a UAV for superficial research of volcanic environments, Aerosp. Sci.
sizing and at same time taking into account the aeroelastic effect
Technol. 70 (2017) 600–614, https://doi.org/10.1016/j.ast.2017.09.005.
would significantly ameliorate the results and make the process of [6] S.G. Kontogiannis, J.A. Ekaterinaris, Design, performance evaluation and opti-
optimization more reliable and efficient. The use of GA coupled mization of a UAV, Aerosp. Sci. Technol. 29 (2013) 339–350, https://doi.org/10.
with ANN benefited the preliminary wing computation process 1016/j.ast.2013.04.005.
used in this paper in term of decreasing the required computa- [7] S.G. Kontogiannis, D.E. Mazarakos, V. Kostopoulos, ATLAS IV wing aerodynamic
tion time. On the other hand, data from similar mini UAV already design: from conceptual approach to detailed optimization, Aerosp. Sci. Tech-
nol. 56 (2016) 135–147, https://doi.org/10.1016/j.ast.2016.07.002.
designed were exploited in order to decrease researching time of
[8] P. Panagiotou, S. Fotiadis-Karras, K. Yakinthos, Conceptual design of a Blended
the optimum solution. The obtained results through this paper can Wing Body MALE UAV, Aerosp. Sci. Technol. 73 (Feb. 2018) 32–47, https://doi.
serve as a contribution to enrich the data base of existing small org/10.1016/j.ast.2017.11.032.
range UAV since different results from Pareto front and other ob- [9] P. Panagiotou, I. Tsavlidis, K. Yakinthos, Conceptual design of a hybrid solar
jective functions were explored. MALE UAV, Aerosp. Sci. Technol. 53 (June 2016) 207–219, https://doi.org/10.
1016/j.ast.2016.03.023.
[10] G. Romeo, G. Frulla, E. Cestino, Design of a high-altitude long-endurance solar-
Conflict of interest statement
powered unmanned air vehicle for multi-payload and operations, Proc. Inst.
Mech. Eng., G J. Aerosp. Eng. 221 (2007) 199–216.
There is no conflict of interest. [11] G. Frulla, E. Cestino, Design, manufacturing and testing of a HALE-UAV struc-
tural demonstrator, Compos. Struct. 83 (2008) 143–153.
References [12] D.P. Raymer, Aircraft Design: A Conceptual Approach, American Institute of
Aeronautics and Astronautics, Reston, VA, 1992.
[1] K. Dalamagkidis, G.J. Vachtsevanos, K.P. Valavanis, Introduction, in: K.P. Valava- [13] J. Roskam, Airplane Design, DARcorporation, Lawrence, Kansas, 2004.
nis, G.J. Vachtsevanos (Eds.), Springer Handbook of Unmanned Aerial Vehicles, [14] J.D. Anderson, Aircraft Performance and Design, WCB/McGraw–Hill, Boston,
Springer, Dordrecht, Heidelberg, New York, London, 2015, pp. 1–99. MA, 1999.
A. Boutemedjet et al. / Aerospace Science and Technology 84 (2019) 464–483 483

[15] Alcore technologies© , France, UAV, http://www.alcoretech.com/?vehicle_type= [33] Z. Rendulić, Mehanika Leta, Vojnoizdavački i Novinski centar, Belgrade, Serbia,
uav/, 2014 (Accessed 30 January 2018). 1987.
[16] BAYKAR© , Turkey, systems-Bayraktar mini UAS, http://baykarmakina.com/en/ [34] F. Smetana, Flight Vehicle Performance and Aerodynamic, AIAA Wright-
sistemler-2/bayraktar-mini-iha/, 2015 (Accessed 30 January 2018). Patterson Air Force Base, Ohio, 2003.
[17] S. Doyle, P. Donaldson, D. Lake, I. Kemp, Unmanned Vehicles Handbook 2008, [35] K. Velimirović, L.G. Krstajić, N. Velimirović, Electrical unmanned aerial vehicle:
The Shephard Press Ltd., UK, 2007. parameters determination of optimal propeller, in: 43th Symposium on Opera-
[18] Avia.pro© , encyclopedia, Unmanned LA, http://avia-pro.net/blog/bespilotnye- tional Research, Tara, Serbia, 2016, pp. 1–6.
letatelnye-apparaty-drony-istoriya, 2018 (Accessed 30 January 2018). [36] J.N. Ostler, W.J. Bowman, D.O. Snyder, T.W. McLain, Performance flight testing
[19] BlueBird Aero systems© , UAV systems, SpyLite-Mini UAV, http://www.bluebird- of small, electric powered unmanned aerial vehicles, Int. J. Micro Air Veh. 1
uav.com/spylite/, 2015 (Accessed 30 January 2018). (2009) 155–171, https://doi.org/10.1260/175682909789996177.
[20] WB Electronics S.A© , Poland, Solutions, BSP, Fly Eye, http://wb.com.pl/fly-eye-
[37] D.E. Rumelhart, G.E. Hinton, R.J. Williams, Learning internal representations by
2/?lang=en, 2011 (Accessed 30 January 2018).
error propagation, in: D.E. Rumelhart, J.L. McClelland (Eds.), Parallel Data Pro-
[21] Department of the US Army, Army unmanned aircraft system operations, in:
cessing, vol. 1, The MIT Press, Cambridge, MA, 1986, pp. 318–362.
Field Manual Interim, FMI 3-04.155, Washington, USA, 2006.
[38] T. Rajkumar, J. Bardina, Training data requirement for a neural network to pre-
[22] D.W. Zingg, M. Nemec, T.H. Pulliam, A comparative evaluation of genetic and
dict aerodynamic coefficients, Proc. SPIE 5102 (2003) 92–103, https://doi.org/
gradient-based algorithms applied to aerodynamic optimization, Eur. J. Comput.
10.1117/12.486343.
Mech. 17 (2008) 103–126, https://doi.org/10.3166/remn.17.103-126.
[23] M. Ebrahimi, A. Jahangirian, Aerodynamic optimization of airfoils using adap- [39] C. Reeves, Genetic algorithms, in: F. Glover, G.A. Kochenberger (Eds.), Hand-
tive parameterization and genetic algorithm, J. Optim. Theory Appl. 162 (2014) book of Metaheuristics, ©2003, Kluwer Academic Publishers, Dordrecht, 2003,
257–271, https://doi.org/10.1007/s10957-013-0442-1. pp. 55–82.
[24] M.S. Khurana, H. Winarto, A.K. Sinha, Application of swarm approach and arti- [40] T.L. Holst, T.H. Pulliam, Aerodynamic shape optimization using a real-number-
ficial neural networks for airfoil shape optimization, in: 12th AIAA/ISSMO Mul- encoded genetic algorithm, in: 19th AIAA Applied Aerodynamics Conference,
tidisciplinary Analysis and Optimization Conference, Victoria, British Columbia Anaheim, CA, USA, 2001.
Canada, 2008, pp. 1–15, AIAA 2008-5954. [41] R.T. Marler, J.S. Arora, The weighted sum method for multi-objective opti-
[25] S. Worasinchai, G. Ingram, R. Dominy, A low-Reynolds-number, high-angle- mization: new insights, Struct. Multidiscip. Optim. 41 (2010) 853–862, https://
of-attack investigation of wind turbine aerofoils, J. Power Energy 255 (2011) doi.org/10.1007/s00158-009-0460-7.
748–763, https://doi.org/10.1177/0957650911405411. [42] H. Sogukpinar, Numerical simulation of 4-digit inclined NACA 00xx airfoils to
[26] C.A. Lyon, A.P. Broeren, P. Giguère, A. Gopalarathnam, M.S. Selig, Summary of find optimum angle of attack for airplane wing, Uludag Univ. J. Fac. Eng. 22
Low-Speed Airfoil Data, vol. 3, SoarTech Publications Virginia Beach, Virginia, (2017) 169–178, https://doi.org/10.17482/uumfd.309470.
1997. [43] Ansys Inc., ANSYS FLUENT 14.0 and GAMBIT 2.1 licensed to VTI, 2010.
[27] XFLR5 V6.12,2016, GNU General public licence, http://www.xflr5.com, 2016 [44] F.R. Menter, Two-equation eddy-viscosity turbulence models for engineering
(Accessed 7 June 2018). applications, AIAA J. 32 (1994) 1598–1605, https://doi.org/10.2514/3.12149.
[28] A. Sóbester, A.I.J. Forrester, Aircraft Aerodynamic Design: Geometry and Opti-
[45] G. Ocokoljić, D. Damljanović, B. Rašuo, J. Isaković, Testing of a standard model
mization, ©2015, John Wiley & Sons, Ltd, 2015.
in the VTI’s large-subsonic wind-tunnel facility to establish users’ confidence,
[29] S. Rajagopal, R. Ganguli, Multidisciplinary design optimization of long en-
FME Trans. 42 (3) (2014) 212–217.
durance unmanned aerial vehicle wing, in: CMES – Computer Modeling in
[46] A. Boutemedjet, M. Samardžić, D. Ćurčić, Z. Rajić, G. Ocokoljić, Wind tunnel
Engineering and Sciences, vol. 81, 2011, pp. 1–34.
measurement of small values of rolling moment using six-component strain
[30] XFLR5 v6.02 Guidelines, http://www.xflr5.com/xflr5.htm.
gauge balance, Measurement 116 (2018) 438–450, https://doi.org/10.1016/j.
[31] B. Maskew, Program VSAERO Theory Document, NASA CR-4023, 1987.
measurement.2017.11.043.
[32] T.R. Yechout, S.L. Morris, D.E. Bossert, W.F. Hallgren, Introduction to Aircraft
[47] J. Weierman, J. Jacob, Winglet design and optimization for UAVs, in: 28th AIAA
Flight Mechanics: Performance, Static Stability, Dynamic Stability, and Classi-
Applied Aerodynamics Conference, Chicago, Illinois, USA, 2010.
cal Feedback Control, American Institute of Aeronautics and Astronautics, Inc.,
Reston, Virginia, 2003.

You might also like