Scaling laws and similarity models for the preliminary

design of multirotor drones

Marc Budinger, Aurélien Reysset, Aitor Ochotorena, Scott Delbecq

To cite this version:

Marc Budinger, Aurélien Reysset, Aitor Ochotorena, Scott Delbecq. Scaling laws and similarity models
for the preliminary design of multirotor drones. Aerospace Science and Technology, 2020, 98, pp.1-15.
�10.1016/j.ast.2019.105658�. �hal-02997598�

HAL Id: hal-02997598

https://hal.science/hal-02997598
Submitted on 10 Nov 2020

HAL is a multi-disciplinary open access L’archive ouverte pluridisciplinaire HAL, est

archive for the deposit and dissemination of sci- destinée au dépôt et à la diffusion de documents
entific research documents, whether they are pub- scientifiques de niveau recherche, publiés ou non,
lished or not. The documents may come from émanant des établissements d’enseignement et de
teaching and research institutions in France or recherche français ou étrangers, des laboratoires
abroad, or from public or private research centers. publics ou privés.
 an author's https://oatao.univ-toulouse.fr/26688

https://doi.org/10.1016/j.ast.2019.105658

Budinger, Marc and Reysset, Aurélien and Ochotorena, Aitor and Delbecq, Scott Scaling laws and similarity models
for the preliminary design of multirotor drones. (2020) Aerospace Science and Technology, 98. 1-15. ISSN 1270-9638
Scaling laws and similarity models for the preliminary design of
multirotor drones
M. Budinger a,∗ , A. Reysset a , A. Ochotorena a , S. Delbecq b
a
ICA, Université de Toulouse, UPS, INSA, ISAE-SUPAERO, MINES-ALBI, CNRS, 3 rue Caroline Aigle, 31400 Toulouse, France
b
ISAE-SUPAERO, Université de Toulouse, 10 Avenue Edouard Belin, 31400 Toulouse, France

a b s t r a c t

Multirotor drones modelling and parameter estimation have gained great interest because of their vast
application for civil, industrial, military and agricultural purposes. At the preliminary design level the
challenge is to develop lightweight models which remain representative of the physical laws and the
system interdependencies. Based on the dimensional analysis, this paper presents a variety of modelling
approaches for the estimation of the functional parameters and characteristics of the key components
Keywords: of the system. Through this work a solid framework is presented for helping bridge the gaps between
Multirotor drones optimizing idealized models and selecting existing components from a database. Special interest is given
Scaling laws to the models in terms of reliability and error. The results are compared for various existing drone
Dimensional analysis platforms with different requirements and their differences discussed.
Surrogate models
Propulsion system
Landing gear

1. Context which can be integrated into any optimization tool for rapid pre-
liminary design of multirotor drones [8,9].
During the last decade, technological innovations [1–3] have Section 2 places emphasis on the sizing scenarios that influ-
significantly contributed to the development of smart and power- ence the design of the components and introduces the concept of
ful multirotor UAVs: miniaturization and microelectronic with the dimensional analysis for the creation of estimation models. Mod-
integration of inertial sensors, increased computational power of els based on similarities, also called scaling laws, are described in
control processors, new battery technologies with higher energy section 3 and applied in section 4 for the electrical components:
density, permanent magnet motors with higher torque density and motor, battery, ESC and cables.
high power densities electric-converters. Regression models are described in section 5 and can adopt
The expansion of drones market has led to the decrease of different forms: polynomial forms exemplified in section 6 for pro-
drone sizes and the increase of the availability of drone compo- pellers and variable power law models in section 7 for structural
nents at very competitive prices. This has facilitated the exper- components.
imentation and optimization of drone designs based on succes- Finally, in section 8 several examples of existing drone plat-
sive physical tests. Most of recent designs address very specific forms are compared with the results obtained using the models.
application-cases with particular performances/needs for which a
design optimization process is needed. The design optimization 2. Needs for prediction models during preliminary design
process becomes mandatory when the payload scale factor dif-
fers from market trend. Particularly in research projects of large Within the whole product development process, the purpose of
multirotor UAVs designed for the transport of commercial loads or preliminary design phase is the evaluation of architecture feasibil-
passengers [4–7] where expensive prototypes are used, it is sug- ity, technology selection and components high-level specifications
gested to perform advanced design studies before manufacturing, definition based on product requirements and operational scenar-
integration and testing. To accelerate the design process, a general ios. Yet, to do so, the critical scenarios should be identified and the
trend is to extend the role of modelling in design and specification. associated main components characteristics identified.
The present paper proposes a set of efficient prediction models
2.1. Multirotor main components

* Corresponding author. Fig. 1 shows a typical mass distribution for the different compo-
E-mail address: marc.budinger@insa-toulouse.fr (M. Budinger). nents of a drone according to several examples [10–12]. This paper

https://doi.org/10.1016/j.ast.2019.105658
2

Nomenclature

N
ABS Acrylonitrile Butadiene Styrene keq Equivalent stiffness . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . m2
DC Direct Current KT Torque constant . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Nm
A
DoE Design of Experiments KV Velocity constant . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
rpm
V
EMF ElectroMotive Force L Characteristic length . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . m
ESC Electronic Speed Controller larm Arm length . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . m
FEM Finite Element Method M Mach number
IGBT Insulated Gate Bipolar Transistor M Mass . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . kg
LiPo Lithium-Polymer np Parallel branches
LG Landing Gear ns Series branches
MTOW Maximum Takeoff Weight nT Coil turns
MOSFET Metal Oxide Semiconductor Field Effect Transistor P iron Iron losses . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . W
PBS Product Breakdown Structure P on Conduction losses . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . W
RSM Response Surface Model P switch Commutation losses . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . W
UAV Unmanned Aerial Vehicle r Diameter ratio . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . m
VPLM Variable Power Law Metamodel R Resistance . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 
Latin formula symbols U Voltage . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . V
U DC DC voltage . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . V
B Flux density . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . T Tf Friction torque . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . N m
Br Magnetic remanence . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . T T nom Nominal torque . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . N m
C Capacity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . A h V Flight speed . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . m
C rate C-rate . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . h−1 V impact Impact speed . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . m
s
s
CP Power coefficient
CT Thrust coefficient Greek formula symbols
D Blade diameter . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . m β Ratio pitch-to-diameter
E No-load voltage . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . V δ Rotational speed . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . rad s
h Convection coefficient . . . . . . . . . . . . . . . . . . . . . . . . . . . . . mW 2K  Average relative error . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . %
H arm Arm height . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . m ω Rotational speed . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . s rad
IO No-load current . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . A π Dimensionless number
J Advance ratio ρ Resistivity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .  m
J Winding current density . . . . . . . . . . . . . . . . . . . . . . . . . . . mA2 σ Standard deviation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . %
N
K Air compressibility . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . m2
θ Maximum admissible temperature . . . . . . . . . . . . . . . . . . . K

edge of the technology or physical-domain. The so-called compo-

nents design drivers are summarized in Table 1 with the corre-
sponding scenarios for which they should be assessed as well as
literature references.

2.3. Dimensional analysis and Buckingham’s theorem

Some references [19] or tools [20] directly use databases to

evaluate the performance resulting from the association of defined
components. The explosion of combinatorial solutions which in-
crease with the databases size may lead to very high computation
time. This is why, most of the design methodologies, despite the
wide variety of implementations [9,16,8], are based on analytical
Fig. 1. Multirotor drone average mass distribution based on [10–12].
expressions or catalogue data regressions and benefit, as most con-
tinuous problems, of numerical optimization capabilities and high
deals with propulsion, energy, power conversion and storage com- convergence performance.
ponents generally treated in a non-homogeneous way in different Dimensional analysis and Buckingham’s theorem [21,22] are ex-
papers but also addresses structural parts. All these parts have a tensively used in aerodynamics and fluid mechanics to provide a
significant impact on global mass which is strongly coupled with more physical and unified framework for all the components and
autonomy and dynamic performances. The control aspects and as- their corresponding characteristics. This paper shows how it can
sociated sensors are not considered here. be extended to other domains by addressing all the components of
a drone.
2.2. Design drivers and sizing scenarios The fundamental step is to express one component character-
istic y as an algebraic function f of geometrical dimensions and
The main performance requirements for drones are (regard- physical/material properties:
less of the control-laws): the payload mass, forward speed, hover
autonomy or range [13]. In addition to these system level power-
energy considerations, the components must withstand some tran- y = f ( L , d1 , d2 , ..., p 1 , p 2 , ...) (1)
     
sitory extreme or endurance criteria that request specific knowl- geometric physical
 3

Table 1
Design drivers and sizing scenarios.

Components Design drivers Parameters or Sizing scenario Ref.

characteristics segment
Propeller Max thrust Diameter, tip Takeoff and high [14]
speed power mission
segment
Efficiency Pitch Hover and long [14]
duration mission
segment
Motor Temperature Nominal torque, Hover and long [15]
rise Resistance duration mission
segment
Max voltage Torque constant, Takeoff and high [16]
Resistance power mission
segment
ESC Temperature Max power, max Takeoff and high [16]
rise current power mission
segment
Battery Energy Voltage, capacity Hover flight [14]
(autonomy) and
forward mission
segment (range)
Power Max discharge Takeoff and high [16] Fig. 2. Estimation models of drone components based on dimensional analysis.
rate power mission
segment
detailed reference component. These laws are established assum-
Frame Stress Mechanical Takeoff and [17]
ing:
strength landing
Vibration Resonance Takeoff and high [18]
frequency power mission • Geometric similarity: all the length ratios between the current
segment component and its corresponding reference are constant. With
x
this assumption the aspect ratios πdi = Li are constant,
• Uniqueness of design driver: only one main dominant physical
with L as the characteristic length and di and p i as the rest of
phenomenon drives the evolution of the secondary character-
geometric and physical parameters.
istic y. With this assumption, the number of π pi is equal to
The Buckingham’s theorem states that the original equation (1),
zero,
with n parameters and u physical units (time, mass, length...), can
• Material similarity: all material and physical properties are as-
be rewritten with a reduced set of n − u dimensionless parameters sumed to be identical to those of the component as the refer-
called π numbers: ence. If the number of π pi is not equal to zero, the remaining
dimensionless numbers can often be expressed through con-
π y = f  (πd1 , πd2 , ..., π p1 , π p2 , ...) (2)
stant ratios of material properties with similar units.
where:
 a
If these assumptions are satisfied, the resulting πdi and π pi di-
π y = ya y · La L pi i (3) mensionless numbers are constant and consequently to π y :
i
di 
m
a
πdi = (4) π y = yLa L p i i = constant (6)
L
 a pi , j
i =1
π pi = La pi,0 pi (5)
which gives the typical power law form of scaling laws:
j
 a
y L
Therefore, the dimensional analysis and the Buckingham theo- y ∝ L a ⇔ y ∗ = L ∗a ⇔ = (7)
rem can permit to reduce the number of input parameters of a y re f L re f
numerical model and hence its complexity and computational cost.
where a is a constant representative of the problem. This paper
The estimation model used will depend on the character of the
also uses the notation proposed by M. Jufer in [23] where a scal-
dimensionless number. Scaling laws are commonly used when πdi
ing ratio x∗ of a given parameter is calculated as x∗ = x x . The
and π pi stay constant in the range of a series, for example for the re f

sizing of the electrical components. When the components do not scaling laws present the asset of requiring only one reference com-
comply with a geometrical or material similarity, then alternative ponent to extrapolate component characteristics (written .re f ) on a
forms will be applied such as regression models for the study of wide range of exploration. More information on the construction,
the propeller parameters or the structural parts. validation and use of scaling laws for engineering purposes can be
An overview of the process is shown in Fig. 2. found in [24]. Scaling laws are used in this paper for the modelling
of electrical components: brushless motors, inverters - part of ESCs
3. Scaling laws and the battery.

The scaling laws, also called similarity laws or allometric mod- 3.1. Uncertainty of scaling models
els, are based on the dimensional analysis. These simplification
laws are really interesting in a preliminary design process because This article will provide the set of standard deviations of es-
they limit drastically the number of design parameters for each timation errors for each proposed model. In order to be able to
component while having good estimation properties based on a better use these quantities, especially for uncertainty propagation
4

Fig. 3. Comparison of error distribution to a normal law. (For interpretation of the colours in the figure(s), the reader is referred to the web version of this article.)

or robust optimization, we will show here that errors distributions

for scaling laws can be modelled as independent normal laws.
Assuming the scaling law is defined as:

y = k · xa
where x is the definition variable and y is the estimated param-
eter of the studied component. The following relationship is used
to predict the estimated error between the real variable and the
estimated one:

y =
y (1 + ε )
ε is modelled as a random variable. Possible error sources may be
related to multiple factors, such as the assumption of the geomet-
ric similarity for much larger dimensions, unfeasibility regarding
Fig. 4. Normal probability plot.
some design constraints, changing materials or physical properties.
The normal distribution is the most common method used to
model phenomena with multiple random events. For this method,
the probability density follows a Gaussian distribution:
1 1 ε −μ 2
f (ε ) = √ e− 2 ( σ )
σ 2π
where μ is the mean and σ the standard deviation. An example of
such distribution is illustrated in Fig. 3. The histogram shows the
distribution of the relative error for the estimated battery mass for
different references of Lithium-Polymer batteries.
A graphical technique used to assess the adequacy of the ob-
served distribution to a Gaussian distribution is the normal prob-
ability plot. This method fits a set of observations to a straight
line and those deviations found from the straight line are varia-
tions from the normal law. The mean and the standard deviation
are read on the abscissa from the crossing of the interpolation line
with Y = 0 and Y = 1 respectively. The collected data series fit
Fig. 5. Effect of the reference point choice on the mean error (motor example).
properly to a normal distribution as depicted in Fig. 4 with a mean
value μ = 0.05% and standard deviation value σ = 6.48%.
In order to minimize the error of estimation when applying • Correlation between errors
scaling laws, it is recommended to take into account some factors. Uncertainties from the different estimation models may be
The main factors or good practices are: linked to further random variables or independent to its own
random variable. The dependency is studied using a correla-
• Choice of a mid-range reference point
tion analysis. An example is applied for the following motor
The choice of a good reference value will have a strong ef-
equations:
fect on the scaling law as it will keep the mean error as low
as possible. Scaling laws follow, as discussed before, a normal  3/3.5
distribution where choosing a reference of the extremes en- T mot
M mot = M mot ,re f Motor mass
tails a difference increase for the other values. The next figure T mot ,re f
shows the evolution of the relative error according to the dif-  3/3.5
ferent motor torque (Fig. 5). It is observed that the error tends T mot
T mot , f r = T mot , f r ,re f Motor
to decrease for references which are close to the geometric T mot ,re f
mean of minimum and maximum values of the full range. friction torque
 5

Fig. 6. Correlation of errors between different motor models.

4. Scaling laws applied to electrical components

4.1. Brushless motor

Many papers [9,16,8] use direct regressions upon manufacturer

datasheets. Here, we outline how scaling laws can represent the
same information from a single reference and with similar fidelity
while enhancing understanding of major physical and/or techno-
logical effects.

4.1.1. Design drivers and main parameters

The multirotor UAVs integrate permanent magnet synchronous
motors with an internal (in-runner) or external (out-runner) rotor.
As encountered on the current drone market, most solutions use
Fig. 7. Standard deviation associated to geometrical parameters.
out-runner type which directly drives the propeller. Thus, only a
model adapted to such type is presented hereafter.
 2  −5/3.5 Manufacturers provide technical information in the form of
KT T mot
R max,mot = R max,mot ,re f Motor equivalent DC motor such as:
K T ,re f T mot ,re f
resistance
• The

rpm constant velocity K v which is generally expressed in
After a correlation analysis for different motor models, it is . This constant is related to the speed or the torque
clear that there is no clear relationship found between the
V  
constant K T = 2π60 · K v (equivalent units
Nm
or radV/s from
different variables. Hence, it is assumed that the errors are in- A
dependent (Fig. 6). which it is possible to express the torque T = K T · I with re-
• Form of the equation spect to the current I or the no load voltage E = K T · ω with
For scaling laws linked directly to some geometric parameters, respect to the speed ω,
it can be noted that uncertainty of a dimension may grow • The continuous current related to the continuous torque T nom
with increasing power. Fig. 7 outlines the distribution of the which is limited by motor losses and heat dissipation,
standard deviation for all references of the BH range from the • A maximum transient current (or torque T max ) which is linked
manufacturer Kollmorgen. to the maximum losses and potentially maintained for a few
In the absence of more information, if the scaling laws are re- seconds or a few tens of seconds,
lated to geometric parameters, the estimation of uncertainties • The no-load current I 0 which represents the motor no-load
is directly related to the exponent of the scaling law as for losses. For high speed and high poles number motors, these
the present example: 10% for the diameter, 30% mass and 50% losses are mainly iron losses that increase with rotational
inertia. speed,
6

 
ρ J2L d1 d2
Table 2
Dimensional analysis on Equation (8).
= f , (10)
hθ L L
Parameter [Voltage] [Current] [Length] [Temperature]

 Similar analysis conducts to Equation (9) simplification:
J A
2 0 1 −2 0  

m  T d1 d2
−1 = f

ρ m
1 1 0 , (11)
θ [K] 0 0 0 1 J Br L4 L L

W 
h 1 1 −2 −1 di
m2 K Considering material and geometric similarities ( L
= constant
L [m] 0 0 1 0
hθ
d1 [m] 0 0 1 0 and ρ = constant) the following scaling laws derive from Equa-
d2 [m] 0 0 1 0 tions (10) and (11):

J ∝ L −0.5 ⇔ J ∗ = L ∗−0.5 (12)

Table 3
4 ∗ ∗ ∗4
Dimensional analysis on Equation (9). T ∝ JL ⇔ T = J · L (13)
Parameter [Force] [Length] [Current] The combination of both previous equations enables the torque
T [N m] 1 1 0 to be expressed as a function of the motor size:

A

J 0 −2 1
T ∝ L 3.5 ⇔ T ∗ = L ∗3.5
m2
B r [Wb] 1 −1 1 (14)
L [m] 0 1 0
d1 [m] 0 1 0 Still considering material/geometrical similarity, the motor mass
d2 [m] 0 1 0 can be estimated with respect to its torque:
3 3
M ∝ L3 ∝ T 3.5 ⇔ M ∗ = T ∗ 3.5 (15)
• The resistance R which causes the ohmic voltage drop and has
an effect on the copper losses. 4.1.3. Scaling laws for the electrical parameters
Another approach presented in [23] consists in setting up the
4.1.2. Dimensional analysis and main relations scaling laws directly using the analytical expression of the param-
In this sub-section the construction of scaling laws with dimen-
eter under study. As an example, the electrical circuits including
sional analysis is illustrated on brushless motors. The considered
windings should be designed to make their operating voltage com-
assumptions are that the main design criterion for the motor is
patible with the power source or the power electronics voltage. An
the maximum winding temperature and that natural convection is
analysis [24] of the no-load voltage E gives the back EMF constant
the dominant thermal phenomenon. We use the following nota-
tions for the further dimensional analysis: K T as a function of the coil turns number nt :
E k · nt ·  · B · S f
KT = = ⇒ K T ∝ nt · L 2 (16)
⎧ ⎫  ω
⎪ J: Winding current density B r : Remanent induction ⎪
⎪
⎪ ⎪
⎪
⎪ di : other motor’s dimensions permanent magnet ⎪ ⎪
⎪
K T does not vary while considering similarity and is a function
⎪
⎪ ⎪
⎪ of motor constants: surface coefficient, filling ratio, winding coeffi-
⎪
⎪ L: Motor length : Rotational speed ⎪
⎪
⎪
⎪ ⎪
⎨ ρ: Cooper Resistivity S f : Magnetic flux motor ⎪
⎬ cient, coils number per group, coils number per group and phase,
surface total notches number, notches number per group.
⎪
⎪ ⎪
⎪ The number of turns nt is also involved in the resistance R ex-
⎪
⎪ θ : Max. Admissible temperature ⎪
⎪
⎪
⎪ ⎪
⎪ pression:
⎪
⎪ rise for winding insulation B: Flux density ⎪
⎪
⎪
⎪ ⎪
⎪
⎪
⎩ h: Convection coefficient T : Electromagnetic ⎪
⎭
torque R ∝ nt2 · L −1 (17)
this resistance can be used to express copper losses.
Considering thermal properties, the current density J can be Another type of losses can be observed in synchronous motor
linked to the motor dimensions and additional material properties: and are due to induction variation: iron losses. The quantity of iron
losses can be expressed according to the Steinmetz-like formula
J = f ( L , d1 , d2 , ρ , θ, h ) (8)
      [25] as:
3 3
P iron ∝ f b · L 3 (18)
Considering magnetic properties, the torque T can be linked to di-
mensions and current density, where f is the electrical frequency linked to the pole number and
mechanical rotational speed : leading to friction torque T f for-
T = f ( L , d1 , d2 , J , B r ) (9)
      mula:
3 2 P iron
Tf = ∝ b−1 · L 3 (19)
Table 2 and Table 3 summarize the dimensional analysis of 
these equations. Equation (35) given thereafter shows that the torque is linked
The problem represented by Equation (8) contains 7 parameters to the rotational speed and the propeller diameter. Considering
and 4 dimensions and can be reduced to 3 dimensionless param- that the propeller is operating at constant tip speed then:
eters. We focus on π J dimensionless parameter which is of the
form: ∗ = D ∗−1 =⇒ T ∗ = D ∗3 = ∗−3

π J = J a J · La L · ρ aρ · θ aθ · hah = J 2 · L 1 · ρ 1 · θ −1 · h−1 This means, if an average value of 1.5 is considered for b, Equa-
tion (19) can thus be simplified to:
Equation (8) can then be reduced, using Buckingham’s Theorem 1.5−1 3
T ∗f = T ∗− 3 · T ∗ 3.5 ≈ T ∗0.69 (20)
to:
 7

Table 4
Scaling laws applied to out-runner brushless motor.

Parameter & equation AXIa SCORPIONb KDEc

Value σ [%] Value σ [%] Value σ [%]
∗
Nominal torque T nom [N m] 0.102 1.84 1.794

Nm
Torque constant K T∗ A
0.005 0.029 0.028
∗
Maximum torque T max [N m] 0.137 1.88 1.823
∗ = T∗
T max N/A N/A N/A
nom

Friction torque T ∗f [N m] 0.006 0.041 0.021

T ∗ = b−1 · L 3 ≈ T ∗0.69
f nom 15.41 26.72 33.4

Mass M ∗ [g] 57 435 305

∗ 33.5
M∗ = T nom 11.51 10.07 9.49

Resistance R ∗ [] 0.045 0.031 0.044

∗ −5
R ∗ = K ∗2 · T nom
3.5
13.21 N/A N/A
a
AXI Motors: https://www.modelmotors.cz/e-shop/.
b
SCORPION: https://www.scorpionsystem.com/catalog/aeroplane/.
c
KDE Direct: https://www.kdedirect.com/collections/uas-multi-rotor-brushless-motors.

4.2.1. Design drivers and main parameters

The ESC main components and corresponding main design pa-
rameters are listed hereafter. The main criteria influencing the size
of the components are related to thermal aspects. The remainder
of this section aims at expressing the effect of scaling on the losses
and heat transfer properties.

• MOSFET transistors: maximum junction temperature is func-

tion of conduction P on , commutation losses P switch and ther-
mal resistance Rth jc .
• MOSFET driver: Peak gate drive currents influence the commu-
tation losses of MOSFET transistors.
• DC capacitor: stabilizes the voltage but the high-frequency cur-
rent charge/discharge generates losses.
• Case: Mechanically protects the components but also acts as a
heatsink for all internal losses.
Fig. 8. Mass vs nominal Torque for different motor families.
4.2.2. Scaling laws approach
MOSFET have two main types of losses [26]:
4.1.4. Scaling laws summary and validation
Table 4 synthesises the previous developed scaling laws for the • Conduction losses: P on = R ds,on · I 2R M S with I R M S the RMS cur-
motors and gives an example of reference data. rent and R ds,on the ON resistance,
In the following Fig. 8, mass data are compared with their cor- QG QG
• Switching losses: P switch ∝ U DC · I · IG
· f switch with
the IG
responding scaling laws. In order to select the best reference point
commutation time (fraction of gate capacitance and driver gate
with the lowest possible error with respect to the mass, the av-
current).
erage relative error was compared by testing all available data
points.
For low-power transistors and relative low switching-frequency
The standard deviation for the mass estimation is between
(as this is the case for UAVs application) switching losses can be
9.49% and 11.51%. A figure corresponding to the friction torque es-
neglected compared to conduction losses. In addition, we assume
timation is not shown here, but the standard deviation is higher
that to limit conduction losses, the number of parallel elementary
(between 15.4% and 33.41%). However, it remains acceptable for a
cells used to realize a MOSFET transistor is proportional to the cur-
preliminary model.
rent rating and thus R ds,on ∝ I −1 .
Since the scaling law for the electrical resistance depends on
more than one parameter, the use of a X = Y chart could permit to This leads to:
visualize the estimated values against the database. The standard
deviation relative error is 13% for the AXI database. P loss, M O S F E T ≈ P O N ∝ I (21)

In turn, considering that transistor design is driven by maxi-

4.2. Power electronics
mum thermal criterion and power dissipation is mainly due to
For drones, the Electronic Speed Controllers (ESC) distributing forced convection (caused by propeller air flow, ESC being inte-
power to the synchronous motors are mainly DC/AC converters (in- grated nearby), we get at equilibrium:
verters) made of power electronic switches (MOSFET transistors).
1
For drone application cases, power MOSFET are preferred to IGBT θ = · P loss, M O S F E T =⇒ θ ∗ = 1 = l∗−2 · I ∗
because of their high commutation frequency and great efficiency h·S (22)
∗ ∗3 ∗ 32
even under low voltages. =⇒ M = l =I
8

 
P mot
P max = U max · I max = U bat · max (25)
U mot

4.2.3. Scaling laws synthesis and validation

The following Table 5 lists the main scaling laws deduced for
the ESC and compares the validity of these models with several
families of components. More details on the use of scaling laws for
sizing and optimization of static converters can be found in [27,
28].

4.3. Battery pack

An essential step of the preliminary design process for multi-

rotor drones is to select a technology and size the battery pack
considering missions performances needs. Lemon et al. [29] high-
light huge differences between the electrochemical couples but
Fig. 9. ESC voltage vs. power regression. also with the cell design itself (Li-Ion specific energy can be mul-
tiplied by 4).
with h the convection coefficient (considered constant), S the con- A high discharge rate (or high power density) means high-
vection exchange surface, θ the temperature drop constant for a dynamics capabilities while high energy density stands for high
given material or technology and l the converter typical dimension. autonomy (long stationary flight). One technology stands out for its
The idea is now to link nominal power to the mass instead high performance capabilities: Lithium-Ion. From this family, the
Lithium-Ion Polymer with slightly similar performances but higher
of current but to do so, the relationship between MOSFET nom-
life cycle duration and safety (better resistance to over-charge) is
inal power and voltage has to be determined. Fig. 9 represents
taken as a reference.
the power and voltage of 75 different drone ESCs from Jive, Kos-
mik, YGE, Spin and Turnigy suppliers. Voltage level increases with
4.3.1. Design drivers and main parameters
power and a direct power-law linking voltage to power is consid-
In this sub-section scaling laws based on the analytical expres-
ered here:
sion of the main sizing parameters of a battery are constructed.
V re f The laws are valid despite the cell technology chosen:
V ∗ = P ∗a =⇒ V = a · P a = K · P a
P re f (23)
• Voltage (U in V): The batteries are usually a set of cells wired
=⇒ log( V ) = a · log( P ) + log( K ) in series to produce the needed voltage (cell voltage being
around 3.7 V). When combining ns batteries in series, to-
The constant K represents a reference component and a the
tal voltage U is the sum of all cells unitary voltage U cell
power to be determined. A linear regression is performed into log
1 (U = ns · U cell ).
space and leads to the approximation V ∗ = P ∗ 3 .
• Capacity (C in A h): By connecting batteries in series, capacity
Equation (22) can then be modified to: does not increase (but stored E energy does: E = C · U ). When
 ∗ 3 the current capacity needs to be increased n p branch should
∗ ∗ 32 P 2
be connected in parallel: C = n p · C i
M =I = = P∗ (24)
V The battery mass being almost linearly connected to its cells
number (packaging put aside), this leads to the main scaling
This is the main result for modelling the ESC. Yet during design,
law for a given cell technology:
another representative value extracted from mission profile/sizing
scenario (linked to motor/battery performances) is the maximum
apparent power: M = ns · n p · M cell =⇒ M ∗ = U ∗ · C ∗ = E ∗ (26)

Table 5
Scaling laws applied to different families of electronic speed controllers (ESC).

Parameters & equation JIVEa KOSMb YGEc SPINd TURNe

Value σ [%] Value σ [%] Value σ [%] Value σ [%] Value σ [%]
ESC voltage 44.4 51.8 51.8 44.4 22.2
V E∗ SC [V]

Maximum power 2664 10360 6216 3419 3330

∗
P max [W]

Maximum current 60 200 120 77 150

∗
I max [A]

Mass M [kg] 84 200 119 105 169

M∗ = P ∗ 13.75 14.14 22.58 19.42 25.36
M ∗ = I ∗3/2 46.07 20.11 43.98 30.52 21.17
a
JIVE ESC - http://shyau.com.tw/documents/.
b
KOSMIK - https://www.kontronik.com/en/products/speedcontroller/speedcontroller1/kosmik.html.
c
YGE - https://www.yge.de/en/home-2/.
d
SPIN OPTO - https://www.modelmotors.cz/product/jeti-spin/.
e
TURNIGY - http://www.turnigy.com/.
 9

Table 6
Scaling laws applied to batteries.

Parameters & equation GENSa PROLb RAMPc TATd KOKe

Value σ [%] Value σ [%] Value σ [%] Value σ [%] Value σ [%]
Voltage 22.2 22.2 22.2 11.1 3.7
U ∗ [V]

Capacity 5000 4400 1800 1550 200000

C ∗ [mA h]

Mass M [kg] 755 601 291 128 4180

M∗ = C ∗ · U ∗ 3.05 4.61 7.07 19.71 3.91
a
GENS ACE (Lithium-Polymer) https://www.gensace.de/.
b
PROLITEX (Lithium-Polymer) http://www.thunderpowerrc.com/.
c
RAMPAGE (Lithium-Polymer) http://www.thunderpowerrc.com/Home/Rampage-Series.
d
TATTU (Lithium-Polymer) https://www.genstattu.com/.
e
KOKAM (Lithium-Ion) http://kokam.com/.

• C-rate (C rate in h−1 ): Discharge current is often expressed as Table 7

a C-rate to normalize against battery capacity. A C-rate is a Scaling laws applied to electric conductors.

measure of how fast a battery can be discharged relative to its Parameters & Equation Wiring cablea
maximum capacity without presenting permanent damage. If Value σ [%]
C rate = 1, then the discharge current will discharge the entire Current I [A] 120
battery in one hour. Formula linking maximum current I max to
Radius r [mm] 5.2
C-rate is the following: r ∗ = I ∗2/3 5.56

M

kg 
I max = C rate · C (27) Linear mass L m
0.191
M
L
= I ∗4/3 11.08
∗ = C ∗.
For a given cell technology, C rate stays constant and I max a
PVC-insulated single wiring cables https://www.
engineeringtoolbox.com/wire-gauges-d_419.html.
Specific power and energy are assumed to remain constant over
a cell assembly even if the packaging can vary on a given series
range with no particular correlation to capacity.

4.3.2. Scaling laws synthesis and validation

Table 6 summarizes the previous scaling laws proposed for the
batteries and gives examples of reference data. 187 batteries com-
ing from Li-Po and Li-ion are used.

4.4. Electric cables

Electric conductors cross-section area is selected with nomi-

nal current. The design driver of such a component is in fact the
maximal insulator temperature. The temperature drop is linked to
power losses that are mainly due to the Joule effect:

P loss ≈ P loss, J oule = ρ · J 2 · V (28)

where ρ is the resistivity of the conductor, J the current density Fig. 10. Radius vs. Current load.
and V the conductor volume.
2
It is assumed here that because of the thin insulator thickness, I = J · π · r 2 =⇒ r ∗ = I ∗ 3 (31)
the thermal conduction resistance can be neglected compared to
Equation (31) can be used to find the relationship between the
the convection one. This means that the power loss can be linked
linear resistance RL of a wire and its nominal current caliber:
to insulator exchange area S = 2π · r · L (r cable radius and L
 ∗
length): R ρ −2 R −4
= ∝r =⇒ = I∗ 3 (32)
L S L
P loss = h · S · θ (29)
M
At the same time, the linear mass L
of the conductor can be
with h the convective heat transfer coefficient and θ the admis- expressed as a function of r:
sible temperature drop (constant for a given external temperature    ∗
and insulator material: h∗ = 1 and θ ∗ = 1). M M 4
∝ r 2 =⇒ = I∗ 3 (33)
Equations (28) and (29) lead to the following formulation of the L L
nominal current density ( J ):
4.4.1. Scaling laws summary and validation
−1
J ∗ = r∗ 2 (30) Table 7 summarizes the scaling laws:
The comparison of these laws with catalogue data yields a
Considering the current is I = J . S, current can be expressed as mean relative error of 16% for the linear mass and about 8% for
a function of the radius: the radius (Fig. 10).
10

Table 8 Therefore in this section, we first apply Buckingham theorem to

APC 10 × 4.5 inch multirotor datasheet [32]. determine the problem dimensionless parameters and then achieve
n (rpm) V (mph) C T (−) C P (−) a polynomial regression to fit the manufacturer datasheet.
2000 0.0 0.1102 0.0428
2000 5.1 0.0832 0.0436 6.1. Application of the Buckingham’s theorem
2000 13.1 0.0057 0.0114

6000 0.0 0.1126 0.0432 Let us consider the following problem:

6000 16.1 0.0822 0.0427
6000 33.5 0.0226 0.0196 T hrust = f ( D , pitch, ρair , K , n, V ) (39)
     
10000 0.0 0.1154 0.0485 2 4
10000 29.2 0.0789 0.0404
10000 60.7 0.0112 0.0120
where K is the air compressibility. Applying Buckingham Theorem
leads to following derived expression:
T hrust
5. Regression with dimensionless numbers CT = = f  (β, J , B ) (40)
ρair · n2 · D 4
In the case where it is not possible to consider as constant Similar analysis can be conducted replacing T hrust by P ower
the dimensionless numbers πdi and π pi , the approximation of the parameter and following dimensionless set arises:
function f can be achieved by performing data regressions [30,31].
P ower
The data can come from manufacturer product data as presented CP = = f  (β, J , B ) (41)
in section 5 for the propellers, test measurements or finite element D5 · n3 · ρair
simulations results computed on DoE as presented in section 6 for Where for both thrust and power coefficients C T and C P :
the drone structure parts. The utilization of π numbers to set up
a response surface model (RSM) or a surrogate model has several pitch
advantages: • β= is the ratio pitch-diameter.
D
V
• A decrease in the number of variables to be manipulated and • J= is the advance ratio.
n·D
therefore a drastic decrease in physical or numerical experi- K
• B= is the air compressibility indicator (similar to
ments to be carried out, ρair · n2 · D 2
• An increase of the regression robustness [30] in particular if Mach number M) and its effect is visible at high speeds.1
the RSM is built within the logarithmic space as for the VPLM
methodology (Variable Power Law Metamodel) [31] which In this section, surrogate models are built to describe the pro-
shows good results in interpolation but also in extrapolation peller’s performance according to (41) in static scenarios ( V = 0),
because of the power law form. applicable at the hovering mode and at takeoff, and in dynamic
scenarios for the rest of cases where V = 0.
6. Polynomial regression applied to propeller datasheets
6.2. Static scenarios, V = 0
The propeller represents a key component in the drone propul-
sion chain. Its performance can be expressed as a function of In the static models, the effect of the advance ratio is removed.
two coefficients C T and C P respectively expressing thrust (Equa- In addition, for given aerodynamic configurations, the propeller
tion (34)) and mechanical power (Equation (35)) equations: must adapt its performance by limiting the speed to a certain RPM
limit proposed by the manufacturer. For APC MR this operating
T hrust = C T · ρair · n2 · D 4 (34) range remains below 105000 Rinch PM
.
In Fig. 11, the variation of C T and C P for different APC MR
3 5
P ower = C P · ρair · n · D (35) Propellers are plotted against β and compressibility factor B. The
blue dots correspond to an operation range below the maximum
where ρair represents the air density, n the rotational speed [ rev ]
s RPM limit (105000 RinchPM
), while the black dots show a behaviour
and D the propeller diameter.
beyond that limit.
When an engineer is facing a new drone design problem, he
It can be hypothesized, that there is a relationship between C T
can recursively try different propeller references and test perfor-
and β (the same with C P and β ), while on the contrary it can be
mances against different mission profiles (takeoff, dynamic perfor- seen how the air compressibility indicator B has a little effect on
mances/vertical acceleration, range...). In that case, he can interpo- the propeller performance when the operation range is below the
late manufacturer datasheets such as the one presented in Table 8 established RPM limit.
to have an estimate of the performance coefficients for each work- Beyond that limit, air compressibility effects begin to have a
ing condition: negative impact on the performance of the propeller with a large
increase of the power coefficient (C P ) in comparison with the
C T = f t (n, V ) (36) thrust coefficient (C T ).
The conclusion is that, for a RPM limit lower than 105000 Rinch
PM
,
C P = f p (n, V ) (37)
the thrust and power coefficient C j generally depends on β :
where V represents the relative air speed.
Nevertheless, as explained before, the performance of the res- C j ≈ f (β, B , J ) →
C j = f (β) (42)
olution is enhanced while defining continuous expression to es-
timate the performance coefficients over the whole multirotor
K
dataset.
1
Considering that = a2 where a is the speed of sound, this yields
ρ
2
a 1
∝
C j = f (n, V , pitch, D , ...) (38) tip speed2 2
M tip
.
 11

Fig. 11. Evolution of the performance coefficients in static case.

Fig. 12. Multi variable correlation matrix using R 2 heatmap (left) and scatter matrix (right).

The estimated thrust coefficient

C T and power coefficient C P for a 6.3. Dynamic scenarios, V = 0
operating range below 105000 Rinch
PM
follow the regression models:
According to the Buckingham’s theorem, a relationship may be
found to link dimensionless parameters:

C T = 4.27 × 10−2 + 1.44 × 10−1 · β R 2 = 0.946 (43) C j ≈ f j (β, J , B ) (45)

Data are filtered in order to respect manufacturer optimal op-

C P = −1.48 × 10−3 + 9.72 × 10−2 · β R 2 = 0.891 (44)
erating conditions (n. D ≤ 105000 Rinch
PM
). Considering such an hy-
pothesis, the compressibility coefficient indicator B effect can be
with a high value of R 2 , the squared correlation coefficient, fitting neglected and Equation (45) is approximated by:
the model very well to the data.
For high values of n · D, a regression based only on the an- C j ≈ f j (β, J ) =
Cj (46)
gle is not very accurate due to the effect of air compressibility. A The use of a correlation matrix and the scatter matrix in Fig. 12
model for higher RPM should consider the effect of both variables: help us to visualize that C T and C P do not depend on a single
β and B. parameter:
12

Fig. 14. C T model performance considering 1st to 8th terms.

Fig. 13. C T model relative error  = f (μ, σ ) evolution with terms selection.

Computations have shown that a 3rd order polynomial form

can be suitable at early design phases:

Cj = (a j ,0 + a j ,1 · β + a j ,2 · J + a j ,11 · β 2 + a j ,12 · β · J
j −∈[t , p ]

+ a j ,22 · J 2 + a j ,111 · β 3 + a j ,112 · β 2 · J (47)

2 3
+ a j ,122 · β · J + a j ,222 · J )
When a linear regression is performed on normal-centred val-
ues, coefficients absolute value can be used to sort terms by order
of importance and regression model can be constructed with in-
creasing complexity (terms number).
As shown in Fig. 13 applied to C T problem, error may not be re-
duced after a certain complexity. With APC datasheet, if a 8-terms
3rd order polynomial model is considered, the relative error is lim-
ited (¯ = 0.4% and σ = 8.8%) as shown in Fig. 14:

C T = 0.02791 − 0.06543 · J + 0.11867 · β + 0.27334 · β 2
− 0.28852 · β 3 + 0.02104 · J 3 − 0.23504 · J 2 (48)
2
+ 0.18677 · β · J
Similar results can be obtained considering the C p parameter
(¯ = 0.5% and σ = 7.5%):

C P = 0.01813 − 0.06218 · β + 0.00343 · J + 0.35712 · β 2
(49)
− 0.23774 · β 3 + 0.07549 · β · J − 0.12350 · J 2

Those
CT ,
C P estimators, depicted in Fig. 15, present multiple
advantages:

• To be continuous over pitch and D parameters, which leads to

faster optimized design and may converge to specific propeller
design. Fig. 15. 3D regression models of estimated performance coefficients and reference
• To have analytic derivatives. points.
 13

Fig. 16. Simplified structure hypothesis.

• To have regular expressions that can be handled by symbolic

parser for different output/inputs alternatives (ex: pitch =
f (C T , D , n, V )).

7. VPLM regression applied to structural parts

The UAV structure has a non-negligible impact on its per-

formance since it represents between 25% to 35% of the Maxi-
mum Takeoff Weight (MTOW) that includes the payload. Therefore,
it is important to have an estimation of the mass of its sub-
components: landing gears, frame arms and body.
It is difficult to address the wide range of complex geometries
which can be found over the models market. A simple geom-
etry is presented in Fig. 16 for further analysis and calculation Fig. 17. Equivalent stiffness problem decomposition with 4 landing-gears.

steps.
7.1. Crash sizing scenario
• Body is assumed to be an aluminium or composite cross with
hollow rectangle section,
The crash sizing scenario considers a maximum speed V impact
• Arms are made of aluminium hollow bars with square cross of the drone when hitting the ground. At such speed the structure
section, should resist (i.e. the maximum stress should not be exceeded)
• Landing gears are assumed to be printing ABS plastic parts. and for higher speeds, the landing gears are the parts that break
as structural fuses.
The structure is sized with respect to two sizing scenarios:
To calculate the equivalent maximum load resisted by the land-
takeoff maximum thrust (arms) and a crash with a given impact
ing gears, the energy conservation law applies the kinetic energy
speed (body, arms, landing gears).
stored in drone mass to potential energy in structural parts transi-
The first sizing scenario is relatively simple and considering
tory deformation:
that the body is rigid compared to the arms, the equivalent prob-
lem is a larm length cantilever beam with an applied load F = 1 1
Thrust. The maximum stress is estimated with safety coefficient keq · δ x2 = 2
M tot · V impact
2 2
k s as: 1
⇒ F max = (keq · δ x + M total · g ) (52)
4
H arm 12 · T hrust · larm σalloy 1

σmax = ≤ (50) = ( V impact · keq M total + M total · g )
2 4
H arm − ( H arm − 2e )4 ks 4

which can be written with dimensionless arm aspect ratio πarm = To calculate the maximum stress induced by the maximum
e
: load F max applied to one landing gear, the equivalent stiffness keq
Harm
should be determined. For this purpose, the problem is broken
 1 down into simpler structural parts, as depicted in Fig. 17, and the
6 · T hrust · larm · k s 3
equivalent stiffness keq is expressed considering the effect of each
H arm ≥ (51)
σalloy (1 − (1 − 2 · πarm )4 ) stiffness on the whole part.

The second sizing scenario is more complex to consider and is k1 ·

k2
keq = 4 · (53)
developed in the next subsection. k1 +
k2
14

The calculation of the body/arm deformation applies the fixed-

beam formulas, while for the landing gear complex geometry a
FEM simulations are used to build an analytic model.

7.2. Landing gear stiffness and induced stress estimation applying VPLM

To design landing gear which can be a specific printed ABS part,

engineer will face one problem: estimate the applied load in the
case of a crash in order to determine the induced stress. But, even
before, as the stress depends on the load applied to the land-
ing gear, an estimation of the stiffness k2 should be established.
Therefore, FEM simulations are performed on different geometries
(H 1 ∈ [5, 20] cm, H 2 ∈ [1, 10] cm, t ∈ [3, 15] mm, θ ∈ [20, 45] deg)
with constant thickness b.
If the problem is expressed in regular form:
k2
= f (H1, H2, t, E , θ ) (54) Fig. 18. πk2 model performance considering 1st to 5th terms.
b    
3 2

A dimensional analysis leads to the following problem formula- 8. Assessment of estimation models performance during
tion: preliminary design
 
k2 H1 t
πk2 = ≈f π1 = , π2 = , π3 = θ (55) The estimation models are the tools that allow us to work
b·E H2 H1
in a continuous domain, reducing the work time and eliminat-
and then: ing the limitations related to data tables. The use of analytical
models also facilitates the implementation of sizing and optimiza-
k2 = πk2 · b · E (56)
tion procedures. Reference [33] explains how to solve the main
πk2 is fitted by a 2nd order variable power-law expression: design problems. Article [34] shows in detail the implementation
of a sizing procedure for multirotor drones applying the mod-
πk2 = 10a0 · π1a1 +a11 ·log(π1 )+a12 ·log(π2 )+a13 ·log(θ ) els described in this article in an optimization routine. The fol-
a +a22 ·log(π2 )+a23 ·log(θ )
(57) lowing table gives the specifications and results of the overall
· π2 2 · θ a3 +a33 ·log(θ )
drone sizing and the resulting errors when compared to their ac-
In logarithmic space this formula can be adapted to the following tual industrial reference. It can be seen that the proposed mod-
linear form: els allow performing a relevant preliminary design of drones (Ta-
ble 9).
log(πk2 ) = a0 + a1 · log(π1 ) + a11 · log(π1 )2
+ a12 · log(π1 ) · log(π2 ) + a13 · log(π1 ) · log(θ) 9. Conclusions
2
+ a2 · log(π2 ) + a22 · log(π2 )
In this paper a set of valid estimation models for the key com-
+ a23 · log(π2 ) · log(θ) + a3 · log(θ) + a33 · log(θ)2 ponents of all-electric multirotor drones is presented. One of the
(58) main hallmarks of these models lies in its mathematical continu-
ous form, which facilitates the implementation of design and opti-
with x1 = log(π1 ), x2 = log(π1 ) ... 2
mization tools. The use of dimensional analysis, apart from provid-
After an analysis of the error distribution, it is observed that ing these physically-motivated approaches with a unified physical
a 5-terms 2nd order model achieves a sufficient fidelity level basis, gives them greater confidence and understanding compared
(Fig. 18): to other pure statistical approaches in the literature. As a result,
scaling laws show robust outcomes in terms of deviation, although
πk2 = 10−0.37053 · π1−3.11170 · π21.10205 · θ −6.61617−4.86580·log(θ ) it is recommended to follow certain guidelines to reduce such de-
viation. Based on the current available technologies, they can be
(59)
applied to the preliminary design of new systems as only a refer-
A dimensional analysis conducted on constraint σmax leads to ence value is needed. Surrogate modelling techniques are used in
following formula: order to obtain an analytic model based on data-sheet data (pro-
  pellers) and on FEM simulations data (landing gear). In these cases,
σmax · b · H 1 H1 t
πσmax = ≈f , , θ = f (π1 , π2 , θ) (60) the dimensional analysis facilitates the choice of the main inde-
F H2 H1 pendent parameters and strengthens the prediction of the mod-
Using a 5-terms 2nd order form, the model obtained is: els.
The proposition of this new multirotor drone sizing model li-
σmax = 100.14690 · π12.08982 · π2−0.98108 · θ 3.38363+2.66468·log(θ )
π brary is important to perform an overall sizing loop. This is em-
phasized in the case of high payloads, where the mass of the
(61) components increase significantly in order to maintain similar au-
Therefore, this model enables to predict the mechanical stress tonomy performance requirements.
generated on the landing gear in the case of a crash. The number
of design parameters is reasonable and the prediction is relatively Declaration of competing interest
accurate (¯ = 0.1% and σ = 4.1%) similarly to the rest of the mod-
els proposed in this paper. None declared.
 15

Table 9
Drone platforms references and sizing results.

Drone platforms a) MK-Quadroa b) Spreading Wingsb

Specifications
Payload: 1 kg 4 kg
Arms/Prop. per arm: 4/1 8/1
Climb speed: 10 m/s 6 m/s
Autonomy: 15 min 18 min
Sizing
Total mass: 1642 g (+16.5% ref.) 8513 g (−10.3% ref.)
Battery mass: 353 g (+7.29% ref.) 2133 g (+10.4% ref.)
Motor mass: 47 g (−12.9% ref.) 160 g (+1.2% ref.)
Max. Power: 173 W (−1.14% ref.) 449 W (−10.29% ref.)
Diameter propeller: 0.256 m (+2.4% ref.) 0.40 m (+5.26% ref.)
a
MK-Quadro: http://wiki.mikrokopter.de/en/MK-Quadro.
b
Spreading Wings S1000+: https://www.dji.com/fr/spreading-wings-s1000.

Acknowledgements [16] C. Ampatis, E. Papadopoulos, Parametric design and optimization of multi-rotor

aerial vehicles, in: Applications of Mathematics and Informatics in Science and
The research was carried out as part of the DroneApp project Engineering, Springer, 2014, pp. 1–25.
[17] P. Wei, Z. Yang, Q. Wang, The design of quadcopter frame based on finite ele-
funded by the RTRA STAE (Sciences et Technologies pour l’Aéro- ment analysis, in: 3rd International Conference on Mechatronics, Robotics and
nautique et l’Espace, www.fondation-stae.net). The authors grate- Automation, Atlantis Press, 2015, pp. 1353–1356.
fully acknowledge the support of François DEFAY. [18] S.K. Phang, K. Li, K.H. Yu, B.M. Chen, T.H. Lee, Systematic design and implemen-
tation of a micro unmanned quadrotor system, Unmann. Syst. 2 (02) (2014)
References 121–141.
[19] S. Bouabdallah, Design and control of quadrotors with application to au-
tonomous flying, Tech. rep., EPFL, 2007.
[1] K.J. Walchko, M.C. Nechyba, E. Schwartz, A. Arroyo, Embedded low cost inertial
[20] ecalc eCalc, the most reliable rc calculator on the web, https://www.ecalc.ch/.
navigation system, in: Florida Conference on Recent Advances in Robotics, FAU,
(Accessed 25 March 2019).
Dania Beach, FL, May 2003, pp. 8–9.
[2] M. Logan, J. Chu, M. Motter, D. Carter, M. Ol, C. Zeune, Small uav research [21] M. Holmes, Introduction to the Foundations of Applied Mathematics, 2009,
and evolution in long endurance electric powered vehicles, in: AIAA Infotech@ pp. 1–42.
Aerospace 2007 Conference and Exhibit, 2007, p. 2730. [22] E. Groesen, J. Molenaar, Continuum Modeling in the Physical Sciences, vol. 13,
[3] G. Cai, J. Dias, L. Seneviratne, A survey of small-scale unmanned aerial vehicles, 2007, pp. 1–29.
in: Recent Advances and Future Development Trends, vol. 2, World Scientific, [23] M. Jufer, Design and losses, in: Scaling Law Approach, The Nordic research
2014, pp. 175–199. Symposium on Energy Efficient Electric Motors and Drives, 1996, pp. 21–25.
[4] C. Silva, W.R. Johnson, E. Solis, M.D. Patterson, K.R. Antcliff, Vtol urban air [24] M. Budinger, J. Liscouët, F. Hospital, J. Maré, Estimation models for the prelim-
mobility concept vehicles for technology development, in: 2018 Aviation Tech- inary design of electromechanical actuators, Proc. Inst. Mech. Eng., G J. Aerosp.
nology, Integration, and Operations Conference, 2018, p. 3847. Eng. 226 (3) (2012) 243–259.
[5] R. Schiavullo, Ehang 184 world first self driving taxi car to flight autonomously [25] C.P. Steinmetz, On the law of hysteresis, Proc. IEEE 72 (2) (1984) 197–221.
at low altitude, Virtual Real. 3 (2016) 02. [26] G. Lakkas, Mosfet power losses and how they affect power-supply efficiency,
[6] Ehang184, autonomous aerial vehicle, http://www.ehang.com/ehang184/index. Analog. Appl. 10 (2016) 22–26.
(Accessed 25 March 2019). [27] X. Giraud, M. Budinger, X. Roboam, H. Piquet, M. Sartor, J. Faucher, Optimal de-
[7] Volocopter reinventing urban mobility, https://www.volocopter.com/en/. (Ac- sign of the integrated modular power electronics cabinet, Aerosp. Sci. Technol.
cessed 25 March 2019). 48 (2016) 37–52.
[8] D. Bershadsky, S. Haviland, E.N. Johnson, Electric multirotor uav propulsion [28] F. Sanchez, S. Delbecq, M. Budinger, I. Hazyuk, Modelling and design ap-
system sizing for performance prediction and design optimization, in: 57th proaches for the preliminary design of power electronic converters, in: Electri-
AIAA/ASCE/AHS/ASC Structures, Structural Dynamics, and Materials Conference, macs 2017, Toulouse, France, 2017.
2016, p. 581. [29] S. Lemon, A. Miller, Electric Vehicles in New Zealand: Technologically chal-
[9] O. Gur, A. Rosen, Optimizing electric propulsion systems for unmanned aerial lenged, University of Canterbury New, Electric Power Engineering Centre (EPE-
vehicles, J. Aircr. 46 (4) (2009) 1340–1353. Centre), 2013.
[10] Mk-quadro – mikrokopterwiki, http://wiki.mikrokopter.de/MK-Quadro/. (Ac- [30] D. Lacey, C. Steele, The use of dimensional analysis to augment design of exper-
cessed 25 March 2019). iments for optimization and robustification, J. Eng. Des. 17 (1) (2006) 55–73.
[11] The mk8-3500 standard, http://www.mikrokopter.de/en/products/nmk8stden/ [31] F. Sanchez, M. Budinger, I. Hazyuk, Dimensional analysis and surrogate mod-
nmk8ueb%eren. (Accessed 25 March 2019). els for the thermal modeling of multiphysics systems, Appl. Therm. Eng. 110
[12] DJI – the world leader in camera drones/quadcopters for aerial photogra- (2017) 758–771.
phy, https://www.dji.com/en/spreading-wings-s1000/spec. (Accessed 25 March [32] E. Motors, Performance data | apc propellers, https://www.apcprop.com/
2019). technical-information/performan%ce-data/. (Accessed 25 March 2019).
[13] N.A. Vu, D.K. Dang, T. Le Dinh, Electric propulsion system sizing methodology [33] A. Reysset, M. Budinger, J.-C. Maré, Computer-aided definition of sizing proce-
for an agriculture multicopter, Aerosp. Sci. Technol. 90 (2019) 314–326. dures and optimization problems of mechatronic systems, Concurr. Eng. 23 (4)
[14] D. Shi, X. Dai, X. Zhang, Q. Quan, A practical performance evaluation (2015) 320–332.
method for electric multicopters, IEEE/ASME Trans. Mechatron. 22 (3) (2017) [34] S. Delbecq, M. Budinger, A. Reysset, A. Ochotorena, Efficient sizing and opti-
1337–1348. mization of multirotor drones based on scaling laws and similarity models,
[15] B.-S. Jun, J. Park, J.-H. Choi, K.-D. Lee, C.-Y. Won, Temperature estimation of Aerosp. Sci. Techol., submitted for publication.
stator winding in permanent magnet synchronous motors using d-axis current
injection, Energies 11 (8) (2018) 2033.