Sunteți pe pagina 1din 35

F INAL YEAR B ENG P ROJECT R EPORT ME30227

Jason Newton, June 9, 2015

Project title: Genetic algorithm optimisation of a UAV using vortex lattice method
Supervisor: Michael Carley
Assessor: David Cleaver
I certify that I have read and understood the entry in the Student Handbook
for the Department of Mechanical Engineering on Cheating and Plagiarism
and that all material in this assignment is my own work, except where I have
indicated with appropriate references. I agree that, in line with Regulation
15.3(e), if requested I will submit an electronic copy of this work for submission to a Plagiarism Detection Service for quality assurance purposes.
Authors signature: .

Abstract
The aim of this project is to optimise a UAV designed to deliver aid for maximum range. These objectives were acheived by using a multidisciplinary approach of combining GA optimisation with aerodynamics, structural weight
estimations and stability. Sensitivity analysis and comparison to the existing
UAV is performed to explain why the best UAV found performs better and
what implications this may have for UAV design in general.
It was found that the effect of the lift distribution on the span efficiency is too
small for this to be of significant concern when trying to minimise drag hence
the requirement of an inboard loaded wing must either be forced or come from
accurate structural estimates. The lowest drag wings produced tend to have a
very low CP over the upper surface throughout the entire chord length with
a low taper ratio on the outer partitions. The best solutions are founds when
the population size per generation is increased. The maximum increased range
of the UAV was 55% although with an increase in wing area for TO and landing purposes a 50% increase is still acheivable. Small UAV wings can acheive
much higher AR than large aircraft due to the low bending moments on the
wing and the high strength of CFRP. It was from this that it was concluded that
long range UAV flight is best acheived with a small amount of batteries and a
large AR low drag wing.

Contents
1
2

6
7

Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Literature review . . . . . . . . . . . . . . . . . . . . . . . . . . .
2.1
Boundary layer transition at low reynolds numbers . . .
2.2
Aerodynamic GA optimisation . . . . . . . . . . . . . . .
2.3
Airfoil parameterisation . . . . . . . . . . . . . . . . . . .
Computational methods . . . . . . . . . . . . . . . . . . . . . . .
3.1
Airfoil .dat files . . . . . . . . . . . . . . . . . . . . . . . .
3.2
Determining number of Chord and Span wise panels to
use . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
3.3
Re-estimating viscous drag . . . . . . . . . . . . . . . . .
3.4
Initial simplified Flying Wing Problem . . . . . . . . . .
3.5
Revised flying wing . . . . . . . . . . . . . . . . . . . . .
3.6
Estimating fuselage drag . . . . . . . . . . . . . . . . . .
3.7
Full model . . . . . . . . . . . . . . . . . . . . . . . . . . .
Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
4.1
Number of panels . . . . . . . . . . . . . . . . . . . . . .
4.2
Initial flying wing experiment . . . . . . . . . . . . . . .
4.3
Improved flying wing experiment . . . . . . . . . . . . .
4.4
Full model results . . . . . . . . . . . . . . . . . . . . . . .
Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
5.1
Initial flying wing experiment . . . . . . . . . . . . . . .
5.2
Improved flying wing experiment . . . . . . . . . . . . .
5.3
Full model . . . . . . . . . . . . . . . . . . . . . . . . . . .
Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Future work . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Appendix . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

3
3
3
4
4
5
5
6
7
7
9
10
12
16
16
17
19
22
26
26
27
28
31
31
33
34

Glossary and notation


AR - Aspect ratio
b - Span
C - Chord length
Cr - Root chord length
CD - Drag coefficient
CM - Pitching moment coefficient
CFD - Computational fluid dynamics
C.G - Centre of Gravity
CP - Pressure coefficient
D - Drag
h0 - Aerodynamic center
iw - Wing setting
GA - Genetic Algorithm
Kn - Stick fixed static margin
Hm - Stick fixed maneuvre margin
MAC - Mean Aerodynamic Chord
MTOW - Maximum take off weight
nbatt - Number of Batteries
pop - Population size
PPLR - Project Plan and Literature Review
Re - Reynolds number
S - Wing area
TE - Trailing edge
TO - Take off
UAV - Unmanned Aerial Vehicle
V - Velocity
VS - Stall velocity
VSf - Stall velocity with flaps
V - Tail volume coefficient
VLM - Vortex Lattice Method
W = Weight
Wwing - Wing Weight
XFLR5 - XFOIL program designed for aerodynamic performance estimates at
low reynolds numbers
- Angle of Attack
t - Geometric twist angle
- Taper ratio
- Dihedral angle

Introduction

There is increasing interest within Aerospace Engineering in regards to UAVs


and their applications. One such application is using drones for delivery. A
UAV being designed by an MEng project team is intended to deliver aid to
areas afflicted by a humanitarian disaster. A typical mission scenario will involve:
-Launch from a pneumatic cannon from a position as close as possible to the
affected area at a MTOW of 7kg.
- Drop two 1kg payloads (with 0.3kg parachute) in the target area.
-Return home at a weight of approximately 4.4kg
-Perform a net landing. (1)
Strict laws encapsulated in CAP 722 and the ANO (2) mean that it is desired
for the UAV to remain under 7kg. In order to maximise the business potential of this project it is desired that the UAV is optimised for maximum range
while still meeting basic criteria such as TO, landing and maneuvre. Since there
is a high degree of non-linearity in the problem due to aircraft being multidisciplinary it is difficult to find a true global optimum for such a problem.
Historically these kinds of non-linear problems have been solved with great
success using genetic algorithms hence it will be useful to see if they can be
applied to this scenario.
The aim of this report is hence to use a GA to optimise a UAV for maximum
range. The objectives to acheive these aims are:
Combine an aerodynamic model with structural and stability estimates
to create a full model of any reasonably sized UAV.
Perform sensitivity analysis to make sure that a global optimum is being
acheived
Comparison with the existing UAV to see why it performs better.
The layout of the report will consist of reviewing revelant literature to the
study, the computational methods and design assumptions assumed, the salient
results from the intermediary and main experiments and then finally a discussion on the interesting features of the results and how they have improved
upon the existing UAV.

2
2.1

Literature review
Boundary layer transition at low reynolds numbers

It is very difficult to predict when boundary layer transition will occur and this
is a topic with a lot of literature surrounding it. The UAVs of the size being
discussed are well into the transition region so it is an important consideration for the estimation of viscous drag. Since it is not a principal aim at the
moment only a small amount of research was conducted however a highly relevant piece of research was found in (3). An airfoil similar to the RG15 being
3

used was tested to find the transition point at various Reynolds numbers and
angles of attack. Based on an extrapolation of the data given in Table 1 of the
study, transition will occur at around an Re of 55000-75000 on both the upper
and lower surface at chord reynolds numbers of 150000-350000 for an AOA of
5 degrees, about the same as the wing incidence of the UAV in cruise based off
the current design.

2.2

Aerodynamic GA optimisation

As discussed in the PPLR (1), GAs have been shown to provide better solutions
than the gradient method and simulated annealing (4) which is the reason why
it was chosen to use this method for the UAV optimisation.
Some studies that have been on aerodynamic optimisation using GAs include
(5-8). The flaws with these studies when compared to the work undertaken is
threefold. Firstly the aircraft are typically much larger so many of the variables
as will be shown e.g. the effect of aspect ratio on wing weight are not applicable
to this study. Secondly, they are very constrained in that these studies specify
a cruise speed and AOA and sometimes even wing area. These are unknowns
in the work undertaken since the UAV works towards a weight limit to determine maximum range i.e. what is the optimal number of batteries, wing size
and tailplane configuration so that the UAV does not exceed a MTOW of 7kg?
Finally they typically only consider the wing on its own without any consideration for the effect on the whole design of the aircraft.
A more relevant study was performed in (9) which is similar to the full model
optimisation intended to be performed where the wholistic effects such as tail
plane sizing and wing position are considered. This study still suffers from the
same problem where variables such as cruise AOA are fixed and that the UAV
is on the order of four times bigger than the one being considered.
The only directly comparable study performed was found in (10) which attempts to do a full configuration optimisation of a similarly designed UAV
that weighs about 1kg. The only major difference is that this UAV is optimised
for endurance and maneuverability rather than range however some of the parameters are comparable to the model produced by the GA in this report.
GAs have also been used with airfoil design (11-13), however this requires
complex parameterisation as discussed in Section 2.3.

2.3

Airfoil parameterisation

The following is from the PPLR written for this project (1). It is included to
explain some of the decisions for the future work on airfoil optimisation which
was unable to be undertaken due to the time constraints of the project.
One of the main problems with optimising the airfoil is how to parameterise
the profile. Using vertices to define the profile means that the algorithm will
have too many degrees of freedom, mostly likely on the order of 100, and will
consider many unfeasible designs meaning that the solution will take too long
4

to converge. Using polynomials to define the airfoil surface will reduce the
degrees of freedom (6) but also has the inherent flaw of considering too many
unfeasible designs and will also most likely not converge since the parameters
used to define the surface will not be based on the aerodynamics and hence the
surface will be varied all at once rather than trying to refine specific qualities
of the airfoil such as leading edge radius or trailing edge slope.
This can be resolved by using Bezier curves and B-Splines (14). With an 8th
degree Bezier curve the approximate relative error between the original airfoil
and the airfoil represented with Bezier curves was around 0.3% and airfoils
such as NACA 0012 were succesfully optimised using this parameterisation by
reducing the drag at all CL s by generally at least 1% with up to 35% in one case.
The B-Spline method was advised against due to having a case to case positive or negative effect compared to Bezier curves while having higher computational burden. Bezier curves still have the problem where the optimiser gets
stuck in a local minimum due to modifying the whole curve for any change in
coefficients.
The Bezier-PARSEC method (15) improves upon this fundemental flaw and
was shown to have accelerated convergence for aerodynamic optimization
using Differential Evolution. It is proposed that the BP-3333 method from the
paper is best for the GA airfoil optimisation of the UAV.

Computational methods

Most of the work performed is done by creating a specific batch script for the
problem. This batch script is based off an example given in the program called
studentbatch.m (16).

3.1

Airfoil .dat files

The airfoils used for the project have their geometry defined by a .dat file which
can be obtained from (insert coord database link). These need to be modified
for the program so that they can be read correctly as shown in Figure 1:

Figure 1: An example with the NACA 0008 airfoil. Any non-geometry data
needs to be commented out, the number of upper and lower surface points
needs to be stated in the beginning of the file and the coordinates must be in
ascending order.
This was performed for the RG15, NACA0008 and S8025 airfoils used in
the original UAVs wing, fin and tail.

3.2

Determining number of Chord and Span wise panels to


use

It is necessary to determine the number of chord and span wise panels to use
since the optimisation will be useless if the answer is inaccurate and extremely
time consuming for an answer that is too accurate. This will be determined
by performing a fully factorial analysis where each combination of chord and
span wise panels from one to ten each will be used to find the induced drag of
the wing and tailplane of the current UAV during trimmed cruise flight as:
Drag is the most important output since this determines the potential
range of the UAV
The skin friction drag and fuselage drag estimates are not based on the
number of panels
These calculations take longer to compute
A 3D scatter plot of induced drag - no. of chord wise panels - no. span wise
panels will graphically show where a good value will lie.

3.3

Re-estimating viscous drag

The current viscous drag calculation used by tornado implicitly assumes boundary layer transition at 10% chord. Given that the Re of the flow over the wings
of a UAV this size is around 200000-300000 this would suggest a pessimistic
transition at Re = 20000-30000. zeroliftdragpred.m was changed so that boundary layer transition occurs according to the research conducted in Section 2.1.
Figure 2 shows the changes.

Figure 2: Changes to the Tornado viscous drag estimation for the purposes of
this analysis.
An if statement was used to preserve the programs original function for Re
outside of the ones being examined.

3.4

Initial simplified Flying Wing Problem

In order to verify and tweak the program before creating a full model, a simplified flying wing problem is contrived that removes much of the complexity
and contraints of the full model. The goal is simply to have maximum cruise
range with the following constraints, assumptions and limitations:
The wing is producing maximum lift of (or close to) 68.67 N i.e. this
ignores the payload drop element to the problem
The wing size does not have any effect on the amount of batteries that
can be carried.
The UAV is electrically powered and the combination of wing and batteries weighs no more than 2.5kg. This is equal to the weight of the wing
and batteries in the current UAV and allowing anymore weight would
mean an unfeasible design since the other weight is required for fuselage, electronics, payload etc.
The choice of battery and motor is not affected so much by the design
changes that the correlations used are inadequate.
Stability is ignored for now
The weight of the wing and batteries does not significantly affect the required fuselage weight in order to support the shear and bending moments. This is reasonable since the fuselage is already overdesigned with
a minimum number of composite layers (17)
TO, landing and maneuvre requirements are ignored
Trim drag is ignored for now and so solverloop5.m, a simple steady
state analysis, is used instead of fTrimCLconst.m for the drag estimation. This will change when a Tailplane is added in the full model.
7

The Cm Re relationship is ignored for now since it is very hard to


predict unless a solver such as XFOIL is used and is not a core aim of the
study. An average CM 0 of -0.05 will be used for the RG15 airfoil for the
analysis (18)
For structural estimates an overall average taper ratio will be used in
order to avoid spending time modifying existing programs. The structure
estimates also assume negligible dihedral and no sweep although in this
case the sweep will not be optimised due to the propeller positioning.
The wing weight also estimates that the lift in each chordwise step is
proportional to the wing weight of that strip.
The fitness function was determined as follows:
Range = Energy/Drag since this is equivalent to Distance = Work done/Force.
Based on the propulsion expert in the UAV team the LiPo batteries provide
555000J/kg of energy (19). The electrical efficiency of these batteries is approximately 0.6 and the combined efficiency of the other components (e.g. motor,
propeller, ESC etc) was approximated at about 0.56. This gives a usable energy
of 186460J/kg and given that the batteries used are 252g each, this means the
maximum range is given by:
R=

E
D

0.252nbatt 186460
D

(=

(2.5Wwing )186460
)
D

Where 0 < nbatt < 10 and takes integer values and 2.5 0.252 nbatt + Wwing .
The problem with this fitness function is that keeping the lift at 68.67N (or
more accurately equal to the weight) will need to be a included as a constraint.
The lift depends on many variables and will cause very slow convergence as
the program tries to find correct values that satisfy it. This was found by trial
and error when trying to use the lift as an inequality constraint as the program
took about 90 minutes to just go through 10 generations with only 3 variables.
This can be overcome by incorporating a term into the fitness function that
gravitates the solutions to the correct lift. The term (1 + (W L)2 ) was multiplied to D so that a penalty is applied to solutions which deviate away from
having L = W . Since MATLABs GA attempts to minimise the fitness function,
the final equation is given by:
F itness =

D(1+(W L))2
46988nbatt

1
R)

The code for the wing weight comes from the structures leader (17) in which
a specific MATLAB script is written that accurately calculates the wing weight
for this specific UAV. This needs to be looped so that the optimum number of
ribs is found. For the design decisions such as the material, number of composite layers, load factor etc it is recommended that this document is read to
explain the reasoning behind the code. For simplicity the code is left unaltered
and so the program will be assuming a straight overall taper.
The variables to be optimised are shown in Table 1 with the bounds.
The program was run with a population size of 200 and was ended at 26 generations.
8

V (m/s)
Cr (m)
1,2,3
(Rad)
b1,2,3 (m)
t1,2,3,4 (Rad)
1,2,3 (Rad)

Lower bound
10
0.04
0.5
0
0
-0.1047
0

Upper Bound
36
0.6
1
0.174
2
0
0.785

Notes
70kt is CAP 722 maximum speed

0.192 Rad is theoretical stall angle of airfoil


t1 refers to root chord

Table 1: Variables optimised in the problem. Other variable bounds were just
estimated starting points. Numbered subscripts refer to the partition of the
variable.

3.5

Revised flying wing

Based on the results of the initial flying wing model there were two changes to
be investigated:
Should twist be kept in as a variable but be forced to change linearly?
Revamping the fitness function to acheive an elliptical lift distribution.
This is needed so that the ailerons dont stall at high angles of attack, the
downwash is constant over the span, and so that the wing spar design
and weight estimations are much easier.
It is proposed to do this in a similar way to keeping the lift constant by applying a term which puts a penalty against values which deviate from the ideal.
Since squaring the residuals worked well for lift, this method was used for this
task.
The program evaluates the local CL at typically up to 9 points along each semispan. These points can be used to calculate how close the curve is to elliptical by
finding out the ratio of their local CL at specific coordinates as shown in Figure
3:

Figure 3: Ideal ellipse geometry shown with example values. This will be performed at all points in the program.
The fitness function will be given by:
F itness =

KD(1+(W L)2 ))
46988nbatt

Pn=ystationlength
Where K =
(1 + 10 (CLn CLidealn )2 ) and CLidealn =
n=1
CLroot Cos(yn /yystationlength )
The extra factor of 10 on the residuals of the local lift coefficients is to make
the factor K of roughly the same magnitude as the lift, meaning the program
will put equal weight on this criteria as keeping L = W . This fitness function
has the advantage of retaining the main goal of range when the lift and lift distribution criteria are met while accelerating the convergence when they are not.
The finalised batch script used is given in the appendix under flyingwing.m.
This has the executable command written in the description. The experiment
was run with 200 population size linear twist, 300 population linear twist and
300 population no twist each for 50 generations.

3.6

Estimating fuselage drag

An approximate value for the fuselage drag was based off (20) as follows:
CD of = Cf fLD fM

Swetf
S

Where fLD is a function of the fineness ratio defined as the fuselage length,
L, over the diameter, D, and fM is a function of the mach number, neglected
since M < 0.1. The fuselage length and max diameter is 0.714m and 0.258m
respectively (Figure 4). The approximate reynolds number over the fuselage
was calculated as:

10

Re =

V L

V 0.714
0.0000146

The transition point from laminar to turbulent flow was estimated as simply
being at Re = 200000. This may be inaccurate but it should suffice for the purposes of the program.
Cf = Cf,lam

x
ltr

+ Cf,turb (1

x
ltr )

Where
Cf,lam =

1.327

Re

Cf,turb =

0.455
(log10 (Re))2.58

60
60
fLD = 1 + (L/D)
3 + 0.0025 (L/D) = 1 + (0.714/0.258)3 + 0.0025 (0.714/0.258) =
3.84

The estimated wetted area of the fuselage came from ME30219 performance
notes (21).
SW et = 0.8 Dmax L = 0.8 3.1415 0.258 0.714 = 0.463m2

11

Figure 4: Fuselage geometry (22).

3.7

Full model

The full model includes the effects of the drag from the fuselage and the drag
and weight of the tailplane and boom. The tailplane and boom size will be
based off stability criteria to find the minimum weight necessary. The effect of
the payload drop and trim drag will also be included when analysing cruise.
There are several assumptions and constraints that are made to try to simplify
the problem:
The propeller positioning does not significantly affect the aerodynamics
The fuselage shape will not be optimised.
Since iw is fixed and changing the angle of attack to increase/reduce lift
will cause extra drag from the fuselage
q not being horizontal, the speed
after payload drop can be inferred as

4.4
7

Vf ullpayload .

Any amount of batteries over six is placed in f4 (Figure 5) which is located


at 554mm from the nose. See the appendix for weights and C.G positions
12

of various components of the current UAV.


The TE of the wing is always in front of the propeller (Figure 5). It was
also decided that sweep shouldnt be investigated since the stability optimisation is already quite computationally intensive.
The downwash is constant across the wing span. This is to simplify the
problem since the downwash calculated by the program varies in accordance with the lift distribution and creates a downwash matrix that is
difficult to integrate with simple stability calculations. The downwash
35a
d
= 180AR
(23)
was estimated as d
It was calculated that a minimum wing area of 0.455 m2 was needed and
that with flaps spanning 60% of the span with 20% of the chord CM 0 at
full deflection was -0.19 (24). Plain flaps were estimated to give a CL of
1.0 at full deflection and. CLmax was conservatively used as 1.05 which
is slightly below the CLmax of the airfoil data.
The aspect ratio and taper ratio of the vertical and horizontal tails is 1.8
and 32 ARW , 1 and 0.7 respectively based on recommendations used in
(25). The fin volume coefficient is assumed to be the same as the current
UAV based on the design considerations from the same reference.
The fin and boom weights are simply correlated to volume since they are
overdesigned with the current minimum amount of fibre layers used e.g.
boom has safety factor of 6.5.

Figure 5: Sideview of aircraft geometry (17). The propeller is mounted on the


back of the fuselage hence the wing is limited to being in front of it.
To reduce the complexity of the problem it was decided that the weight estimation for the aerodynamic solver will be approximated rather than exact. Since
the program inevitably gravitates towards having the most weight possible the
weight was estimated as just being 68.67N and 43.2N for full and no payload
respectively. This is since it is preferred that the tailplane is a consequence of
the aircraft rather than simply another input variable as this would mean that
the tailplane weight would need to be found simultaneously with the wing
and battery weight meaning many unfeasible designs will be considered in the
process.
13

Another problem is that the tailplane and boom move the C.G position aftwards due to the extra weight so that in effect h0 h becomes smaller as the
tailplane effectiveness increases. It was decided the best way to deal with this is
to assume that the program gravitates towards the original tailplane and boom
sizes (since the program prefers a main wing with minimum area) to give an
initial C.G position and then loop to refine the value.
V was defined by the minimum value which satisfies Kn = 0.05, Hm = 0.08
(25) as well as the cruise, climb and descent maneuvre conditions which are
based off the trim equation CM = CM 0 (h0 h)CL V CLT .
The lift curve slope of the tailplane is estimated to be around 3.7 based on
the current UAV. In further work it is suggested that this be more accurately
predicted using ESDU sheets 89029 and 70011 (26-27).
The total weight of the combined wing, batteries, fin, tail and boom was updated to 2837g. This is like the same as the 2.5kg used for the wing and battery weight in the flying wing experiment except now since stability is taken
account of, the program should now gravitate towards a small tailplane for reduced drag and weight.
The initial full model experiment will have no constraint on wing area since
it is desired to find out what the true optimum is if TO and landing had no effect. Three experiments consisting of a population size of 400, population size
of 400 with the current UAV as the initial population, and a population size
of 500 were performed. The program was altered such that fTrimCLconst.m
the program converges when CL = 0.01 rather than 0.001 as otherwise the
number of iterations needed makes the program far too slow. A block diagram
of the program is given in Figure 6:

14

Figure 6: Simplified block diagram of the program.


15

4
4.1

Results
Number of panels

It was found that in order to do a trimmed analysis, some of the combinations


of chord and span wise did not give a converged solution for this wing within
the default 10 iterations in fTrimCLconst.m. For the purposes of analysis this
was increased to twenty.
The results of the batch script are shown in Figure 7. The aircraft was assumed
to be at max payload for reference. The batch script used with the reference
data inputted is given in the appendix.

Figure 7: Accuracy of analysis on the current UAV with varying numbers of


chord and span wise panels.
It was seen that the accuracy of the analysis did not depend much on the
number of span wise panels in comparison to the number of chord wise panels.
Based on Figure 7 an appropriate number of chord and span wise panels to
acheive sufficient accuracy are eight and three respectively.

16

4.2

Initial flying wing experiment

Figure 8: Geometric results of the analysis

Figure 9: Lift distribution in cruise.


The geometry of the initial flying wing optimisation is shown in Figure 8.
It was seen that the (1 + (W L)2 ) factor performed as intended and caused
the algorithm to incline towards solutions where L = W (Table 2) without being as time consuming as including this as a constraint in the program. Table
2 also shows that the program used geometric twist in such a manner to make

17

V (m/s)
Cr (m)
1,2,3
(Rad)
b1,2,3 (m)
t1,2,3,4 (Rad)
1,2,3 (Rad)
Weight (N)
Lift (N)
CL
D (N)
Re
S (m2 )
b (m)
Wing Weight (kg)
Number of batteries
Fitness

34.7
0.103
0.52, 0.78, 0.54
0.162
0.25, 0.63, 0.40
-0.0959, -0.0743, -0.0844, -0.0479
0.187, 0.213, 0.0075
67.63
67.04
0.721
1.457
130650
0.1241
2.56
0.561
8
3.876 10 6

Table 2: The optimum variables decided by the algorithm along with the salient
outputs.
the wing difficult if not impossible to manufacture. It was proposed to try either forcing a linear twist or to remove this as a variable not just for program
and manufacturing simplicity but also since it may not be necessary if the other
variables are optimised correctly as shown in (28).
It was originally assumed that the program would tend towards inboard loaded
wings for minimum induced drag, however it not perform as intended as the
lift distribution turned out to be no where near elliptical suggesting that a new
approach will be needed when creating the fitness function (Figure 9).

18

4.3

Improved flying wing experiment

Figure 10: Pressure plots for each of the wings produced in the second iteration.

19

Figure 11: Lift distribution in cruise for each wing produced in the second
iteration.
20

Experiment
V (m/s)
Cr (m)
1,2,3
(Rad)
b1,2,3 (m)
1,2,3 (Rad)
t
Weight (N)
Lift (N)
CL
D (N)
Re
S (m2 )
b (m)
Wing Weight (kg)
Number of batteries
Fitness

200 pop no twist


25.5
0.128
0.82, 0.85, 0.76
0.069
0.65, 0.28, 0.84
0.20, 0.51, 0.75
67.50
67.65
0.57
1.7
162500
0.3
3.54
0.869
6
6.433 105

300 pop no twist


32.6
0.066
0.91, 0.58, 0.47
0.096
0.50, 0.59, 0.45
0.18, 0.47, 0.75
68.60
68.55
0.74
1.42
116000
0.143
3.07
0.477
8
3.99 105

300 pop linear twist


28.3
0.119
0.60, 0.79, 0.52
0.112
0.53, 0.68, 0.42
0.15, 0.52, 0.74
-0.0455
67.96
67.93
0.62
1.71
147000
0.225
3.27
0.663
7
4.79 105

Table 3: The optimum variables decided by the algorithm along with the salient
outputs.
Although the algorithm was heavily weighted to produce a perfect elliptical lift distribution it only produced a rough approximation as seen with the
sharp rise in CL midspan (Figure 11). This is sufficient for the aircraft since it is
mostly about avoiding tip stall at high angles of attack. The program was also
seen to prefer much more dihedral than the original run.
It was seen in Figure 10 that the lowest drag wing had a very low value of
CP across the entire area which would explain why the GA chose this wing
to have a high cruise speed in replacement of wing area since this wing is designed more for high speed flight. This also explains the wing had a much
better value of L/D than the other wings.
It was found that introducing a bigger population size led to a much better
end design with having enough weight for 2 extra batteries, a 20% reduction
in cruise drag while still having a greater cruise speed in the 300 pop no twist
case although this likely due to the fact that the wing is around 2.5 times as
small. The lift distribution shape of the 300 pop no twist case (11) was interestingly similar to the 200 population experiment with a slight improvement
in the midspan. It was seen however when linear twist was introduced that an
inferior design was produced which was in between the 200 population and
the 300 population with no twist as demonstrated in Table 3.
However when looking at the lift distributions it is clear that introducing twist
has significantly improved this aspect of the design. From this it was concluded that it is worth keeping twist to improve the aerodynamic performance
however it will be necessary to have at least a population size of 400-500 to
have a likely chance of finding a global optimum in the full model. Unfortu21

nately this meant that few analyses on the full model could be performed since
this requires tremendous computational effort.

4.4

Full model results

Figure 12: Full geometry for the three full model runs.

22

Figure 13: Convergence plots for the three full model runs.

23

Figure 14: Lift distribution in cruise for the three full model runs.

24

Figure 15: Pressure distribution plots in cruise for the three full model runs. It
was found that there is a bug within Tornado VLM which means sometimes
the tailplane is not plotted in the pressure plot graphs.
The geometry of each design is given in Figure 12. All designs tended to converge within about 25 generations so they were stopped here (Figure 13) since
they take 2-3 days to run. The 400 pop group had the best lift distribution shape

25

Experiment
V (m/s) Full & no payload
Cr (m)
1,2,3
(Rad)
b1,2,3 (m)
1,2,3 (Rad)
t
Wing TE position (m)
Boom length (m)
CL
Average Drag (N)
S (m2 )
Sf in (m2 )
Stail (m2 )
V
b (m)
Wing Weight (kg)
Number of batteries
Range (m)

400 pop
18.1, 14.4
0.175
0.85, 0.95, 0.5
0.152
0.31, 0.74, 0.25
0.16, 0.22, 0.53
-0.056
0.691
1.618
0.91
2.59
0.37
0.017
0.0947
3.19
2.60
0.88
6
109000

500 pop
18.8, 14.9
0.172
0.63, 0.77, 0.66
0.1458
0.73, 0.53, 0.39
0.10, 0.48, 0.47
-0.069
0.522
1.29
0.88
2.36
0.36
0.0272
0.0302
0.946
3.30
0.96
6
119500

400 pop initial pop


23.2, 18.4
0.187
0.97, 0.69, 0.50
0.13
0.35, 0.26, 0.31
0.20, 0.24, 0.66
-0.044
0.568
1.75
0.76
3.53
0.27
0.0081
0.0232
1.04
1.83
0.53
8
106500

Current
21, 16.65
0.26
0.928, 0.928, 0.928
0.096
0.327, 0.327, 0.305
0, 0, 0.17
-0.0279
0.6
0.95
0.57
3.65
0.455
0.045
0.046
0.4
1.94
1.03
6
77300

Table 4: The optimum variables decided by the algorithm along with the salient
outputs.
(Figure 14). The 500 pop group had the lowest drag and was shown in Figure
15 that CP was typically much lower towards the TE of the wing planform as
well as the LE suggesting that the wing is good at producing lift and avoiding
boundary layer separation. This aircraft had around 55% extra range than the
current UAV. It was found that the 400 and 500 pop experiments were converging towards a similar point suggesting that a global optimum may have been
found (Table 4). The main difference between the two is that the 500 pop size
found a much better stability solution than the 400 pop experiment meaning
that a larger aspect ratio wing could be afforded leading to reduced drag and
extra range. The 400 pop experiment with the initial population interestingly
converged towards a completely different solution albeit one that has nearly
the same performance as the other 400 pop experiment. This utilised having
2 extra batteries at the expense of a low weight low aspect ratio wing. In theory this could be the best design since it will have a similar range to the others
except it will be much more maneuvarable and faster however all of these designs ignore the wing area component to the problem (for TO and landing)
hence it will be very unlikely that the 8 battery design could be feasible.

5
5.1

Discussion
Initial flying wing experiment

There are a couple of interesting points to note about the results of the initial
flying wing experiment.
26

Firstly due to the low stresses in the wings, much higher aspect ratios can be
acheived compared to large aircraft such as passenger aircraft as the weight is
more affected by the total volume of the wing rather than the aspect ratio causing the design for a wing that can withstand high bending moments. A wing
aspect ratio of around 52.5 was produced whilst only weighing 0.581kg with
a reference area of 0.1241 m2 . This is in contrast to the wing that the current
UAV has where an aspect ratio of 8.25 has a 1.03kg wing with a reference area
of 0.455 m2 . This means the current UAV wing has around 3.7 times the area of
the flying wing whilst only weighing 1.84 times more however the aspect ratio
is far lower which suggests that low Re UAVs should be designed to have as
small wing area as possible with a larger aspect ratio than might be expected.
According to Raymer the correlation of wing weight to surface area and aspect
ratio is S 0.758 and AR0.6 and 0.04 (29) which would suggest that:
Current UAV wing weight =k 0.4550.758 8.250.6 = 1.95k, initial flywing wing
weight =k 0.12410.758 52.50.6 = 2.21k i.e. the wing should weigh more due to
the bending moment that it needs to be designed to resist (Assuming 0.04 1).
Although theoretically an elliptical lift distribution will give the lowest induced drag this was not even close to being seen despite the main target of
the program being to reduce drag. There are several possible explanations for
this result:
The algorithm didnt have enough variation or enough generations to
converge to an elliptical distribution and got stuck in local minima.
Due to the finite number of ways the wing was allowed to be changed
the algorithm had difficulty finding correlations which would lead to an
elliptical distribution
The effect of an elliptical lift distribution on drag is small and hence
the traits of a high AR and lots of batteries on range outweighed this
consideration.(30-31)
It was also interesting to note the similarities in the results of this experiment
with the one performed by (5). The program also used geometric twist in such
a way to make difficult shapes (t1,2,3,4 = 0 , 0.4691 , 3 , 4.9685 ). Another
was that the program also preferred a variety of solutions rather than a single
region of wing geometries as was shown in the improved wing experiments.
This has been suggested as being due to the fact that nature prefers diverse
solutions to a solitary optimum. Similar types of results were found in (32)
where a diverse range was the outcome of the experiment.

5.2

Improved flying wing experiment

It was seen that the high amounts of dihedral compared to the initial wing
experiment could be explained by observing the wings without any dihedral.
Despite the increase in lift the drag remains at a similar value of 1.78N and Figure 16 shows that the lift distribution has changed dramatically suggesting that
dihedral is being used to change the lift distribution whilst having a negative
impact on L/D.
27

Figure 16: Lift distribution of the wing without dihedral.

5.3

Full model

The results for the full model experiments are similar to the UAV optimised in
(10) in that a large AR ratio wing was considered optimum by the program.
However the UAV optimised for endurance has an AR of 10 rather than 30.
This is likely due to the fact that the wings for this UAV are made out of just
foam hence they would not be strong enough for such high ARs. The UAV also
needed to weigh around 1kg so having an excessively large wing could not be
considered as it would weigh too much. Since the UAV needs to TO and land
on a runway the wing area was greater in proportion to the designs in section
4.4 which is the reason why the UAV from this study has a calculated cruise
CL of 0.5.
Due to time constraints on the project a full model with a fixed wing area was
not performed. However the 500 pop answer gives the most likely candidate
of being used for as the new design. This is because it has the best combination of being able to meet the wing area requirements whilst having the largest
range out of the experiments due to having a low drag wing whilst still having
a reasonable number of batteries. Parameters were changed in a trial and error method to create a UAV with S = 0.455m2 based off the 500 pop solution.
This is given in FIGURE 17:

28

Figure 17: Modified optimum UAV results.


As shown in Table 5 the UAV still gives around 50% extra range than the
current UAV due to its low drag performance although interestingly it sacrifices a battery in order to acheive this. A wing aspect ratio of about 32.5 is
produced which explains why the drag is 81% lower and how this is worth the
17% decrease in availible energy. This again demonstrates how strong CFRP

29

V (m/s) Full & no payload


Cr (m)
1,2,3
(Rad)
b1,2,3 (m)
1,2,3 (Rad)
t
Wing TE position (m)
Boom length (m)
CL
Average Drag (N)
S (m2 )
Sf in (m2 )
Stail (m2 )
V
b (m)
Wing Weight (kg)
Number of batteries
Range (m)

16.7, 13.24
0.19
0.63, 0.77, 0.66
0.1458
0.68, 0.8, 0.45
0.097, 0.48, 0.47
-0.069
0.522
1.29
0.86
2.02
0.455
0.039
0.0175
0.427
3.86
1.26
5
116300

Table 5: The revised values to increase the wing area.


is for the skin of the wing. In reality this may be lower or even significantly
higher since the current wing weight estimation fails to take into account the
fact that only multiple layers of composite are possible rather than a continous
thickness.
There are a few errors and uncertainties that havent been looked into with
this design:
The boom is 1.29m long and the stiffness has not been looked into so it
may be that this needs to be limited if the elasticity decreases the effectiveness of the tailplane during maneuvres.
Dynamic stability hasnt been looked into so it may well be that the dihedral will need to change.
Since the span is nearly 4m long it may be that this is too large for maneuverability purposes.
So far it has been seen that the GA prefers solutions with rather high
value of CLcruise . This suggests that a significant reduction in drag will
be acheivable with changing or optimising the airfoil for minimum drag
at around CL = 0.9. For the current airfoil this gives nearly double the
estimated drag coefficient at CL = 0.6 (33), the cruise CL of the current
UAV. This will also have a twofold effect since many high lift airfoils
have a high CLmax meaning that the wing area could be reduced to the
S = 0.36m2 optimum point which was found.

30

Conclusions

It was found that 8 chord wise and 3 span wise panels are needed for sufficient
accuracy in analysing a UAV. Dihedral can be used advantageously to refine
the lift distribution of a wing. The effect of span efficiency is very small in
comparison to the effect of AR and the number of batteries. The lowest drag
wings produced tend to have a very low CP over the upper surface throughout
the entire chord length. Better solutions are founds when the population size
per generation is increased. A combination of linear twist and dihedral produces the best lift distributions for the program however it appears to increase
the chances of finding local minima leading to an overall higher drag wing.
The optimal design for cruise was produced by the 500 pop group which converged to a similar yet overall best design to the 400 pop group. Introducing
an initial population appears to increase the chances of finding a local minima
with this type of problem as shown with the other 400 pop group. The maximum increased range of the UAV was 55% although with an increase in wing
area a 50% increase is still acheivable. Small UAV wings can acheive much
higher AR than large aircraft due to the low bending moments on the wing
and the high strength of CFRP. It was because of this that a lower amount of
batteries with a large AR low drag wing appears to be the optimum for long
range flight whilst being able to acheive reasonable TO and landing speeds.

Future work

Firstly, it is desired to have more sensitivity analysis performed as well as some


research into interesting areas of the experiment. Some potential areas include:
Complete the experiment with a fixed wing area of 0.455m2
If it is still necessary to force an elliptical lift distribution, what effect does
changing the weighting have on the final results?
Would squaring the number of batteries and drag in the fitness function
lead to better convergence as the program will equally favour minimum
drag from the start of each run to the other criteria?
What is the effect of including stability in the calculations? How would
the results differ if the original tailplane and boom were used?
Secondly the results should be grealty improved if the wing weight calculations to take into account lift distribution and discrete layers of thickness. This
should eliminate the need for forcing an elliptical lift distribution since wing
will weigh much more which is likely to have a much bigger effect than the
span efficiency.
Thirdly there is a lot of potential to reduce the drag further by creating a separate airfoil optimiser that runs in a loop with the configuration optimiser e.g.
shown in diagram below:

31

Figure 18: How the program should include an airfoil optimiser.


This should make the program more stable than optimising simultaneously
since it would not only mean having half of the variables but it will allow relationships to be more easily found by the program i.e. it would be undesired
if the wing airfoil is affecting the size of the tailplane. Fourthly, the current
fuselage drag estimates are approximate and assume a drag similar to that of a
well designed passenger aircraft. Considering how much the fuselage reduces
the ideal cruise velocity (by nearly a factor of 2!) it would be beneficial to have
a more accurate value. This may also have the potential to create a third optimising stage for the fuselage.
And lastly, it would be desired to have a better estimatation of the lift curve
slope of the tailplane based on (26-27) and to use the full downwash matrix
provided by the program.

32

References
(1) Jason Newton, Final year BEng project plan and literature review ME30227,
University of Bath, 2015.
(2) Jason Newton, UAV1: Phase 2 Business Report - Airworthiness and certification,University of Bath, 2015.
(3) M. Brendel and T. J. Muellerf. (1988). Boundary-Layer Measurements on an
Airfoil at Low Reynolds Numbers. J. AIRCRAFT. OL. 25, NO. 7
(4) S. Obayashi (1997). Genetic Algorithms in Engineering Systems. Sendai:
Tohoku University. p. 203-228
(5) Cervenka, Miroslav ; KReslek, Vojtech. (2009). Aerodynamic wing optimisation using SOMA evolutionary algorithm. Studies in Computational Intelligence. Vol.236 p.127-138.
(6) Quagliarella, D., and Cioppa, A. D.: Genetic algorithms applied to the aerodynamic design of transonic airfoils. AIAA paper 94-1896, June 1994
(7)Yamamoto, K, and Inoue, O.: Applications of genetic algorithm to aerodynamic shape optimization. AIAA paper 85-1650-CP, a collection of technical papers, 12th AIAA Computationalfluid dynamics conference, CP956, San
Diego, CA, June 1995, p. 43-51
(8) Luis. F. Gonzlez , K. Srinivas, Jacques Priaux, and Eric J. Whitney . (2005).
Multidisciplinary Design Optimisation of Unmanned Aerial Vehicles (UAV)
using Multi-Criteria Evolutionary Algorithms. 6 th World Congresses of Structural and Multidisciplinary Optimization.
(9) N. Van Nguyen, J.-W. Lee, Y.-D. Lee, H.-U. Park. (2014). A multidisciplinary
robust optimisation framework for UAV conceptual design. The Aeronautical
Journal. Volume 118
(10) Neo Siah Jen. (2010). Unmanned Air Vehicle (UAV) Design using Multidisciplinary Optimization. Available: http://dynamicslab.mpe.nus.edu.
sg/dynamics/thesis0910/Unmanned%20air%20vehicle%20(UAV)%20%
20design%20using%20multidisciplinary%20optimization.pdf Last
accessed 09/06/2015.
(11) Jahangirian, A. ; Shahrokhi, A.. (2011). Aerodynamic shape optimization
using efficient evolutionary algorithms and unstructured CFD solver. Computers and Fluids. 46 , p270-276.
(12) Shigeru Obayashi, Shinkyu Jeong, Abel Mata Zetina. (2013). Airfoil Aerodynamic Optimization for A HighAltitude Long-Endurance Aircraft Using MultiObjective Genetic-Algorithms . IEEE Congress on Evolutionary Computation
p. 2314-2320
(13) David W. Zingg , Marian Nemec, Thomas H. Pulliam (2008) A comparative
evaluation of genetic and gradient-based algorithms applied to aerodynamic
optimization, European Journal of Computational Mechanics, 17:1-2, p103-126
(14) Paul Dan Silistaneau, Ruxandra Mihaela Botez (2012). Two dimensional
airfoil design for low speed airfoils. Minnesota: AIAA.
(15) Derksen, RW ; Rogalsky, T. (2010 Jul-Aug). Bezier-PARSEC: An optimized
aerofoil parameterization for design. Advances In Engineering Software. Vol.41(78), p 923-930.
(16) Tomas Melin. Available: http://tornado.redhammer.se/ Last accessed 09/06/2015.
(17) Jakub Kucera, UAV1: Phase 2 Primary Report - Structures, University of
Bath, 2015
33

(18) RG-15 8.9% - Rolf Girsberger RG 15 low Reynolds number airfoil. Available: http://airfoiltools.com/airfoil/details?airfoil=rg15-il#
polars Last accessed 09/06/2015.
(19) James Walker, UAV1: Phase 2 Primary Report - Propulsion, University of
Bath, Bath 2015.
(20) Sadraey M.. (2009). Drag Force and Drag Coefficient. Available: http://
faculty.dwc.edu/sadraey/Chapter%203.%20Drag%20Force%20and%
20its%20Coefficient.pdf Last accessed 09/06/2015.
(21) ZhijinWang. (2014). Drag. In: Aircraft Performance ME30219. p10.
(22) Jakub Kucera, UAV1: Phase 3 Primary Report - Structures, University of
Bath, 2015
(23) D. Stinton. The Design of the Aeroplane. Oxford, BSP Professional Books,
1983.
(24) Jason Newton, UAV1: Phase 2 Primary Report - Aerodynamics, University
of Bath, Bath 2015.
(25) Abinesh Mohan, UAV1: Phase 2 Primary Report - Stability and Control,
University of Bath, Bath 2015.
(26) ESDU 70011. Lift-Curve Slope and Aerodynamic Centre Position of Wings
in Inviscid Subsonic Flow. ESDU International, March 2012.
(27) ESDU 89029. Installed Tailplane Lift-Curve Slope at Subsonic Speeds.
ESDU Interna- tional, July 2013.
(28) Spyridon G. Kontogiannis. (2013). Design, performance evaluation and
optimization of a UAV. Aerospace Science and Technology. 29, 339-350.
(29) Snorri Gudmundsson. Statistical weight estimation methods. General
Aviation Aircraft Design: Applied Methods and Procedures. -: ButterworthHeinemann. p142
(30) Ismet Gursul (2014). ME30032 Aerodynamics. University of Bath: p 59.
(31) Raymond E. Mineck and Paul M, Hi W. Vijgen. (1993). Wind-Tunnel Investigation of Aerodynamic Efficiency of Three Planar Elliptical Wings With
Curvature of Quarter-Chord Line. Nasa technical paper.
(32) Andre C. Marta. (2003). Parametric Study of a Genetic Algorithm using a
Aircraft Design Optimization Problem.
(33) Michael S. Selig, Christopher A. Lyon, Philippe Giguere, Cameron P. Ninham and James J. Guglielmo (1996). Summary of Low-Speed Airfoil Data .
Virginia: Soartech publications

Appendix

34

S-ar putea să vă placă și