5th International Conference on Advances in Control and

5th International Conference on Advances in Control and

Optimization of Dynamical
5th International Systems
Conference on Advances in Control
Optimization
February
of Dynamical
18-22, 2018.
Systems
Hyderabad,
Available
India
onlineand
at www.sciencedirect.com
5th International
Optimization of Conference
Dynamical on Advances
Systems in Control and
February 18-22, 2018. Hyderabad, India
Optimization of Dynamical
February 18-22, Systems
2018. Hyderabad, India
February 18-22, 2018. Hyderabad, India
ScienceDirect
IFAC PapersOnLine 51-1 (2018) 219–224
Inertia
Inertia and Center of Mass Estimation of a
Inertia3and
and
DoF
Center
Center
Air
of
of Mass
Mass
Bearing
Estimation
Estimation 

of
of a
a
Inertia3and
DoF Center
Air of MassPlatform
Bearing Estimation
Platform  of a
3 DoF Air Bearing Platform 
3 DoF Air Bearing Platform
Athrey Ranjith Krishnanunni ∗∗ Sreeranj Jayadevan ∗∗
∗∗
Athrey Ranjith Krishnanunni ∗∗∗ ∗ Sreeranj Jayadevan
∗∗∗∗ ∗∗
Athrey Abhilash
Ranjith Mony
Abhilash Mony ∗∗∗ Gireesh
Krishnanunni Gireesh Sharma
Sreeranj
Sharma N
Jayadevan
N ∗∗∗∗
∗ ∗∗
Athrey Ranjith Mony
Abhilash Krishnanunni
∗∗∗
Gireesh Sreeranj
Sharma Jayadevan
N ∗∗∗∗
∗∗∗ ∗∗∗∗
∗ Abhilash Mony Gireesh Sharma N
∗ Department of Aerospace Engineering, Indian Institute of Technology
∗ Department
Kharagpur,
of Aerospace Engineering, Indian Institute of Technology
West Bengal, India
∗
Department
Kharagpur, of Aerospace
West Bengal, India (e-mail:
Engineering, Indian
(e-mail: athreyr@iitkgp.ac.in).
Institute of Technology
athreyr@iitkgp.ac.in).
∗∗Department of Aerospace Engineering, Indian Institute of Technology
∗∗
Department
Kharagpur, of Aerospace
West Bengal, Engineering,
India Indian
(e-mail: Institute
Institute of
athreyr@iitkgp.ac.in).of Technology
∗∗ Department
Kharagpur,
Department
Kharagpur,
ofWest
of
Aerospace
Aerospace
West Bengal,
Bengal,
Engineering,
India
Engineering,
India (e-mail:
Indian
(e-mail:
Indianathreyr@iitkgp.ac.in).
Institute
Technology
of Technology
sreeranjj@iitkgp.ac.in).
Kharagpur,
∗∗ ∗∗∗
Department West Bengal,
ofMechanisms
Aerospace India (e-mail: sreeranjj@iitkgp.ac.in).
Engineer,
Kharagpur,
∗∗∗
Engineer,West Bengal,Engineering,
Mechanisms and
IndiaAdvanced
and (e-mail:Indian Institute
Actuators of Technology
Section,
Section, ISRO
sreeranjj@iitkgp.ac.in).
Kharagpur,
∗∗∗ Inertial
Engineer, West
Systems Bengal,
Mechanisms Unit, IndiaAdvanced
Trivandrum,
and Advanced
Actuators
(e-mail:Kerala,
sreeranjj@iitkgp.ac.in).
Actuators India (e-mail:
Section,
ISRO
ISRO
∗∗∗ Inertial Systems Unit, Trivandrum, Kerala, India (e-mail:
Engineer,
Inertial SystemsMechanisms and AdvancedKerala,
abhilash.mony@gmail.com).
Unit, Trivandrum, Actuators India Section,
(e-mail: ISRO
∗∗∗∗ Inertial Systems abhilash.mony@gmail.com).
Unit, Trivandrum, Kerala, India (e-mail:
∗∗∗∗ Head, Advanced abhilash.mony@gmail.com).
Sensors and Actuators Division, ISRO Inertial
∗∗∗∗ Head, Advancedabhilash.mony@gmail.com).
Sensors and Actuators Division, ISRO Inertial
Head,Systems
Advanced Unit, Trivandrum,
Sensors Kerala,
and Actuators India
India (e-mail:
Kerala,Division, ISRO Inertial
∗∗∗∗
Head,Systems
Advanced
Systems Unit,
Unit,
gireesh
Trivandrum,
Sensors and Actuators
sharma@vssc.gov.in).
Trivandrum, Kerala, Division,
India
(e-mail:
ISRO Inertial
(e-mail:
Systems Unit, gireesh sharma@vssc.gov.in).
Trivandrum, Kerala, India (e-mail:
gireesh sharma@vssc.gov.in).
gireesh sharma@vssc.gov.in).
Abstract:
Abstract: RigorousRigorous ground
ground testing
testing of of spacecraft
spacecraft dynamics
dynamics will will significantly
significantly reduce reduce various
various risksrisks
to
to a
a mission.
Abstract:
mission. A
A spherical-air-bearing-based
Rigorous ground testing
spherical-air-bearing-based of testing
spacecraft
testing platform
dynamics
platform provides
will nearly
significantly
provides nearly frictionless
reduce
frictionless rotations
various risks
rotations
Abstract:
about
to the
a mission. Rigorous
three axes ground
and testing
can
A spherical-air-bearing-based be of spacecraft
used to dynamics
simulate
testing will
spacecraft
platform significantly
provides attitude reduce various
nearly dynamics
frictionless on risks
Earth.
rotations
about
to the
a mission. three axes and can
A spherical-air-bearing-based be used to simulate
testing spacecraft
platform provides attitude dynamics
nearly dynamics
frictionless on Earth.
rotations
However,
about
However, the gravity-induced
three axes
gravity-induced and unbalance
can be torque
used to causes
simulate undesirable
spacecraft disturbances
attitude on the onplatform
Earth.
about
during
However, the threeThis
testing.
gravity-induced and unbalance
axespaper can be used
explores
unbalance
torque
two to
torque
causes undesirable
simulate
techniques spacecraft
for
causes undesirableestimating
disturbances
attitude
the
disturbancesinertia
on theonplatform
dynamics
onproperties Earth.
the platform and
during
However, testing. This
gravity-induced paper explores
unbalance two
torquetechniques
causes for estimating
undesirable the
disturbancesinertia on properties
the platformand
center
during
center of
of mass
testing.
mass offset
This
offset values
paper
values of
of a
explores
a spacecraft
two
spacecraft simulator
techniques
simulator with
for
with three
estimating
three rotational
the
rotational degrees
inertia
degrees of
properties
of freedom.
freedom.and
during
Numerical
center of testing.
mass This values
simulation
offset paper
results explores
ofofa these two
spacecrafttwo techniques
simulator for
estimation estimating
techniques:
with the inertia
aa batch
three rotational method
degreesproperties
using and
linear
of freedom.
Numerical
center simulation
of mass offset results
values of these
ofofa these
spacecrafttwo estimation
simulator techniques:
with three are batch
rotational method
degrees using linear
oftechniques
freedom.
least
least squares,
Numerical
squares, and
simulation an online
results
and an online method
method using
usingtwo recursive
estimation least squares,
techniques: a reported.
batch methodThese using linear
Numerical
are
leastthen
squares, simulation
compared
and anin results
terms
online of
of
method these
accuracy
usingtworecursive
estimation
of the
recursive
least
estimated
least
squares, are
techniques:
parameters,
squares, are
reported.
a batch method
noise
reported.
Theseusingtechniques
sensitivity,
These linear
techniques and
are
leastthen
squares,compared
and an in terms
online of
method accuracy
using of the
recursive estimated
least parameters,
squares, are noise
reported. sensitivity,
These techniques and
are then
computational
computational compared load.
load. in terms of accuracy of the estimated parameters, noise sensitivity, and
are then compared
computational load. in terms of accuracy of the estimated parameters, noise sensitivity, and
© 2018, IFAC (International
computational load. Federation of Automatic Control) Hosting by Elsevier Ltd. All rights reserved.
Keywords:
Keywords: Parameter
Parameter estimation,
estimation, least least squares,
squares, recursive
recursive estimation,
estimation, air air bearing
bearing platform,
platform,
Keywords:
gravity
gravity Parameter
torque,
torque, estimation,
spacecraft
spacecraft attitude
attitude least squares,
dynamics,
dynamics, recursive
spacecraft
spacecraft estimation,
ground
ground testing.
testing. air bearing platform,
Keywords:
gravity Parameter
torque, estimation,
spacecraft attitudeleast squares,
dynamics, recursiveground
spacecraft estimation,testing. air bearing platform,
gravity torque, spacecraft attitude dynamics, spacecraft ground testing.
1. INTRODUCTION
1. INTRODUCTION of
of the
the test
test platform
platform can can limit
limit perfect
perfect reproduction
reproduction of of the
the
1. INTRODUCTION of the
orbital
orbital test platform
environment
environment can limit
(Prado
(Prado et
et perfect
al.
al. reproduction
(2005),
(2005), Schwartz
Schwartz of
et
et the
al.
al.
Simulation and 1.
testingINTRODUCTION
of spacecraft dynamics on the of the
(2003)).
orbital test platform
environment can limit
(Prado et perfect
al. reproduction
(2005), Schwartz of
et the
al.
Simulation and testing of spacecraft dynamics on the orbital (2003)).
Simulation
ground is and
highly testing
desirable of spacecraft
because it is dynamics
extremely on the
difficult (2003)).environment (Prado et al. (2005), Schwartz et al.
ground is highly desirable because it is extremely difficult If
If the
the center
center of of mass
mass (CM) (CM) of of the
the test
test bed
bed lies
lies vertically
Simulation
ground
to andthedesirable
is highly
reconfigure testing
systemofbecausespacecraft
once theitvehicle dynamics
is extremely
vehicle is in on the
difficult
in space.
space. A (2003)).
below
If its
the center center
of mass of rotation
(CM) of(CR), the testthe test
bed bed
lies
vertically
behaves
to
to
reconfigure
ground is highly
reconfigure
spherical air
the system
desirable
the system
bearing can
once
because
once
be used
the it is
thetovehicle
create
is
extremely difficult
isfrictionless
in space. 3-
A
A If below
the its
centercenter
of of
mass rotation
(CM) (CR),
of the the
test test
bed bedvertically
lies behaves
vertically
spherical
to air
reconfigure bearing
the can
system be
onceusedtheto create
vehicle frictionless
is in space. 3-
A like
below
like a
a pendulum
its center
pendulum system.
of rotation
system. Since
(CR),
Since a
a pendulum
the
pendulumtest system
bed
system has
behaves
has
spherical air bearing
degrees-of-freedom can be used
rotational to create
motion, whichfrictionless
can simulate
simulate 3- below
a stable its center of
equilibrium rotation
point (CR),
when its the
CM test
is bed
located behaves
along
degrees-of-freedom
spherical air bearing rotational
can be used motion, which
to create can
frictionless like
3- alike a
stable pendulum
equilibrium system.
point Since
when aitspendulum
CM is system
located has
along
degrees-of-freedom
spacecraft
spacecraft attitude
attitude rotational
dynamics
dynamics motion,
on
on Earth.
Earth. which
This
This can
is
is simulate
especially
especially a
the a pendulum
stable
gravityequilibrium
vector, system.
thepoint Since
when
balance aitspendulum
of theCM is
system system
located
cannot has
along
be
degrees-of-freedom
spacecraft
useful
useful in
attitude
in the
the rotational
dynamics
development
development of
motion,
on
of new
new Earth.
actuators
actuators
which
Thisorcan
or
simulate
ispayloads
especially
payloads as the
as athe gravity
stable
guaranteed
gravity
vector, the
equilibrium
just
vector, by the
balance
point
observing whenthe
balance
ofitsthe
of CMsystem
test
the
cannot
is located
bed
system cannot
be
along
maintaining be
spacecraft
useful
their attitude
in the
performance dynamics
development
can be on Earth.
of new
evaluated actuators
a This
priori. or ispayloads
especially as guaranteed
the gravity just by the
vector, observing
balance the
of testsystem
the bed maintaining
cannot be
their performance can be evaluated a priori. one fixed
guaranteed attitude
just by(Kim and
observing Agrawal
the test (2009)).
bed Therefore,
maintaining
usefulperformance
their in the development a priori.or payloads as one
of new actuators
can be evaluated fixed attitude
guaranteed
the effects ofjust by(Kim and Agrawal
observing the test (2009)).
bed Therefore,
maintaining
The spherical
their spherical
The performance segment
can beair
segment airevaluated
bearing aallows
bearing allows approximately one
priori.approximately the
one effects
fixed of gravity
fixed attitudegravity
attitude
(Kim
(Kim
should
andbe
should
andbe
minimized
Agrawal
minimized
Agrawal
by
by aligning
(2009)).
(2009)). aligning the
Therefore,
the
Therefore,
The spherical
frictionless segment
angular air bearing
rotation of the allows
equipmentapproximately
platform the effects
center
center of
of of gravity
mass
mass of
of the
the should
platform
platform be minimized
as
as close
close as
as by aligning
possible
possible to
to the
its
its
frictionless
The sphericalangular
segment rotation
air of
bearingthe equipment
allows platform
approximately the effects
center of of gravity
rotation.
mass of In
the should
order
platform be
to dominimized
as so, the
close by
value
as aligning
of
possiblethe the
CM-
to its
frictionless
about the angular
pitch, yaw, rotation
and of
roll the
axes. equipment
A thin film platform
of com- center of rotation. In order to do so, the value of the CM-
about the pitch,
frictionless angularyaw, and rollofaxes.
rotation A thin filmplatform
the equipment of com- center CR of rotation.
offset
center of mass
needs oftothebe
In platform
estimated
order to doas close
first.
so, the as possible
value of the to
CM-its
about
pressed
pressed the
gas
gaspitch,
(or
(or yaw,
air),
air), is
isand roll
injected
injected axes.
between
between A thin
a
a film
spherical
spherical of com-
seg-
seg- CR offset needs to In beorder
estimated first.
about (ball)
pressed
ment the
gaspitch,
(or yaw,
andair), isand roll
injected axes.
between A thin
aThis filmthin
spherical of com-
film CR
seg- center of rotation.
offset needs to be estimated first. to do so, the value of the CM-
ment
pressed(ball) and aa mating
mating spherical
spherical cup.
cup. This thin seg-film The inertia
CR offset properties
needs of
of the
to be estimated the simulator
first. are are necessary
necessary for
ment
creates gasessentially
(ball)
an (or
andair), is injected
a mating between
spherical
frictionless cup.aThis
lubrication spherical
layer thin
betweenfilm The inertia
developing
properties
and implementing
simulator
the outer-loop
for
controllers
creates
ment an
(ball)essentially
and a frictionless
mating lubrication
spherical cup. layer
This between
thin film The inertia
developing properties
and of
implementing the simulator
the are
outer-loop necessary for
controllers
creates
the ball an
ball and essentially
and cup
cup (Schwartz frictionless
(Schwartz et et al. lubrication
al. (2003),
(2003), Meissner layer
Meissner (2009)). between
(2009)). The for inertia
the properties
platform. of the the
Therefore, simulator
values are
of necessary
moments for
and
the developing and implementing the outer-loop controllers
creates
the
When
When
an
balltest
and
test
essentially
cup (Schwartz
articles
articles
frictionless
mounted
mounted et al.to
to
lubrication
(2003),
the
the Meissner
spherical
spherical (2009)). for
layer segment
between
segment
the platform.
developing
products of and
inertia
Therefore, the
implementing
of the thevalues
simulator
of moments
outer-loop
have to be
and
controllers
estimated
for the platform. Therefore, the values of moments and
the ball
When
are and cup
test
balanced, (Schwartz
articles
such mounted
that et al.to
their (2003),
aggregate Meissner
the sphericalcenter segmentof mass products
(2009)). for
as
of inertiaTherefore,
the platform.
well.
products Since
of inertia the
of the simulator
ofinertia the CM
and
the simulator
haveoftomoments
values
haveoffset
be estimated
to beare and
coupled
are
When
are
balanced,
test
balanced,
coincides
such
articles
with such
that
mounted their
that their center
the rotational
rotational
aggregate
to the
aggregate
center
spherical
of thecenter
the
of mass
segment
of mass
sphere, as
the products well. Since
of inertia the inertia
ofinertia
the and
simulator CM haveoffset
to beareestimated
coupled
estimated
coincides with the center of sphere, the through
as well. the
Sinceequations
the of motion
and CM(as shown
offset later),
are they
coupled
are balanced,
coincides
rotational with
motionsuch that
the their
theofrotational aggregate
center
combination of center
of the
the air of
sphere,bearingthe through
mass as
are well. the
estimatedSinceequations
the inertia
together.
of motion
and CM (as shown
offset are later), they
coupled
rotational
coincides
rotational
and test
motion
with
motion
articlesthe ofrotational
the combination
of the those
match center
combination
of an
of
of the
of the
objectthewith
air
sphere,bearing
air bearing the through
similar are the
estimated equations
together. of motion (as shown later), they
and test articles
rotational motion match
of the those of an object
combination of the with
air similar through
bearing are estimatedthe equations
together. of motion (as shown later), they
and testproperties
inertial articles match freelythose of an
rotating in object
space. with However,similara In In this
are paper,
estimated
this aa batch
batch method
paper, together. method using using linear
linear least
least squares,
squares,
inertial
and testproperties
articles freely
match rotating in space. However,similaraa In
inertial properties
gravitational freelythose
disturbance ofcreated
rotating
torque an in object
space.
by the with
However,
unbalance and
and this
an
an paper,
online
online a batch
method
method method
using
using using
recursive
recursive linear
least
least least squares,
squares,
squares, are
are
gravitational disturbance
inertial properties torque
freely torque
rotating created by
in space. the unbalance In this paper, amethod
batchthe method usingof linear least squares,
gravitational disturbance created by theHowever,unbalancea and employed
an
employed to
online
to estimate
estimate moments
using
the recursive
moments of inertia,
least
inertia, products
squares,
products of
are
of

gravitational disturbance
with ISROtorque Inertialcreated
Systemsby the unbalance and
inertiaan and
onlinethe method
CM-CR using
offset recursive
of a least squares,
spherical-air-bearing- are
 In collaboration Unit,
In collaboration with ISRO Inertial Systems Unit, Trivandrum,
Trivandrum, employed
inertia and to estimate
the CM-CR the moments
offset of a of inertia, products
spherical-air-bearing- of

Kerala, India.
In collaboration
Kerala, India. with ISRO Inertial Systems Unit, Trivandrum, employed
inertia and tothe
estimate
CM-CR theoffset
momentsof a of inertia, products of
spherical-air-bearing-
 In collaboration with ISRO Inertial Systems Unit, Trivandrum, inertia and the CM-CR offset of a spherical-air-bearing-
Kerala, India.
Kerala, India.
2405-8963 ©
Copyright © 2018, IFAC (International Federation of Automatic Control)
2018 IFAC 235 Hosting by Elsevier Ltd. All rights reserved.
Copyright
Peer review©under
2018 responsibility
IFAC 235Control.
of International Federation of Automatic
Copyright © 2018 IFAC 235
10.1016/j.ifacol.2018.05.047
Copyright © 2018 IFAC 235
5th International Conference on Advances in Control and
220
Optimization of Dynamical SystemsAthrey Ranjith Krishnanunni et al. / IFAC PapersOnLine 51-1 (2018) 219–224
February 18-22, 2018. Hyderabad, India

 
0 −a3 a2
[a×] = a3 0 −a1
−a2 a1 0

−ḣ−ω ×h is the input control torque τ = [τx τy τz ]T pro-

vided by the momentum exchange devices. If the actuator
dynamics is ignored and the control torque is assumed to
be instantaneously available, then (1) becomes
˙ J˜ + [ω×][Ω]J˜ + [g×]mr = τ
[Ω]
or
 
  J˜ = τ
˙ + [ω×][Ω]
[Ω] [g×] (2)
Fig. 1. Schematic of spacecraft simulator mr
It is evident that ω̇ is needed for solving (2), which requires
based spacecraft simulator. The effectiveness of the two
numerical differentiation of measured angular rate signals.
estimation techniques (often abbreviated later as ’LLS’ or
That would lead to amplification of the noise in the signal,
’the batch method’, and ’RLS’ or ’the online method’)
which is undesirable. To avoid that, (2) is integrated from
are studied through numerical simulation of the system
the initial time t0 to time t to get
dynamics. The results are then compared on the basis of
  t  t    t
accuracy of estimated values, sensitivity to sensor noise, J˜ = τ dt
and computational load. [Ω(t)]−[Ω(t0 )]+ [ω×][Ω]dt [g×]dt
t0 t0 mr t0
(3)
2. SYSTEM DYNAMICS AND KINEMATICS The kinematic equations represent the variation of atti-
tude of the platform. In this paper, the attitude is de-
The equations of motion governing the kinematics and scribed by Euler angles φ, θ, and ψ about the roll, pitch,
dynamics of a rigid platform attached to an air bearing and yaw axes respectively. The corresponding Euler rates
is described in this section. The platform rotates under φ̇, θ̇, and ψ̇ are related to the body angular rate by
the combined action of the gravity-induced unbalance the following equation based on the 3-2-1 Euler sequence
torque and the control torque generated by a momentum (Bekir (2007)).
exchange actuator attached to it.   
φ̇  
1 sin φ tan θ cos φ tan θ ω1
Fig. 1 shows the schematic of a rigid platform having three  θ̇  = 0 cos φ −sin φ ω2 (4)
rotational degrees of freedom suspended on an air bearing. 0 sin φ/cos θ cos φ/cos θ ω3
The offset between CR and CM and the coordinate system ψ̇
attached to the platform are also shown. In such a body- Since the extent of rotation of the platform is restricted
fixed coordinate system, the Euler equations of a rigid (as shown later), θ will not increase close to 90o to cause
platform with a momentum exchange actuator under the a mathematical singularity.
influence of gravity torque is given by
Gravitational acceleration vector g is transformed into
[J]ω̇ + ω × [J]ω = −ḣ − ω × h + r × mg (1) the body-fixed coordinates using the 3-2-1 Euler angle
Here [J] is the inertia matrix of the platform (calculated transformation from Bekir (2007).
with respect to CR), ω = [ω1 ω2 ω3 ]T is the angular rate of  
the platform, h is the total momentum of the momentum −g sin θ
exchange devices, m is the mass of the platform, r is a g = g sin φ cos θ (5)
constant vector from CR to CM (i.e. the position vector g cos φ cos θ
of CM), and g is the gravitational acceleration vector
(expressed in body-fixed coordinates). The unknowns to
be estimated in (1) are the inertia matrix [J] and the CM
offset r. 3. ESTIMATION USING LINEAR LEAST SQUARES

Instead of r, it is easier to estimate the product of mass

Consider a system of linear algebraic equations given by
and CM position vector, mr = [mrx mry mrz ]T . In order
to convert (1) into the standard least squares form as given [ΨN ]x = ZN − VN (6)
in Kim and Agrawal (2009), the angular rate matrix and
the inertia vector are respectively defined as which are used to describe a set of N experiments done on
  a model, to determine n unknowns denoted by the vector
ω 1 0 0 ω2 ω 3 0 x. Each experiment may provide a set of k equations, so
[Ω] = 0 ω2 0 ω1 0 ω3 that the coefficient matrix [ΨN ] has dimensions kN × n.
0 0 ω3 0 ω1 ω 3 Likewise, kN -vector of right-hand-side values is denoted
T
J˜ = [Ix Iy Iz Ixy Ixz Iyz ] by ZN . Both [ΨN ] and ZN are obtained from measure-
ments, and the errors in them are represented by VN .
The skew-symmetric cross-product-operation representa-
tion of a vector a = [a1 a2 a3 ]T is defined as Equation (3) is of the same form as (6), where

236
5th International Conference on Advances in Control and
Optimization of Dynamical SystemsAthrey Ranjith Krishnanunni et al. / IFAC PapersOnLine 51-1 (2018) 219–224 221
February 18-22, 2018. Hyderabad, India

  t1  t1 
x̂(N ) = x̂(N − 1) + [LN ][z(N ) − [CN ]x̂(N − 1)] (11)
 [Ω(t 1 )]−[Ω(t0 )]+ [ω×][Ω]dt [g×]dt 
 t0t2 t0t2  Finally, the covariance matrix is updated.
 
[Ω(t2 )]−[Ω(t0 )]+ [ω×][Ω]dt [g×]dt 
[ΨN ]=
 t0 t0

 [PN ] = [[I9 ] − [LN ][CN ]] [PN −1 ] (12)
 . . . . . . 
    According to Haykin (2009), the algorithm is to be in-
 t N t N  tialised with
[Ω(tN )]−[Ω(t0 )]+ [ω×][Ω]dt [g×]dt
t0 t0 3N ×9 [P0 ]9×9 = δ −1 [I9 ] (13)
  (7) where δ is a small positive constant. x̂(0) = 0 is assumed
˜
x= J as the initial estimate.
mr 9×1
  t1  The original RLS algorithm detailed in Haykin (2009)
involves a forgetting factor. In its most general form, the
 t τ dt  RLS with forgetting has been widely used in estimation
  0t 
 2  and tracking of time-varying parameters in various fields
 τ dt 
ZN =
 t0


of engineering. However, when excitation of the system is
 ···  poor, this scheme can lead to the covariance ”wind-up”
 t  problem. During poor excitations, old information is con-
 N 
τ dt tinuously forgotten while there is very little new dynamic
t0 3N ×1 information coming in. This might lead to the exponential
and subscript i denotes the data collected from the ith growth of the covariance matrix, and as a result the estima-
experiment, conducted at time ti . tor becomes extremely sensitive, and therefore susceptible
to numerical and computational errors (Fortescue et al.
Due to the presence of VN = [v1 v2 . . . vN ]T , (6) (1981), Vahidi et al. (2005)).
will not, in general, have an exact solution, but only an
approximation, x̂. A large number of experiments are The inertia of the platform and the unbalance values
carried out to make x̂ a better estimate, so that kN > n. remain constant with respect to the body-fixed coordinates
It can be shown that the solution x̂ (estimate of x) that throughout the estimation process, so that the parameters
minimizes the sum of the squares of the errors from each to be estimated do not vary with time. Therefore, the
experiment (i.e. ||v1 ||2 + ||v2 ||2 + · · · + ||vN ||2 = ||VN ||2 = particular RLS algorithm used in this paper does not in-
VN T VN ) is volve a forgetting factor and the issue of covariance ”wind-
up” does not occur. Nonetheless, proper and continuous
x̂ = [[ΨN ]T [ΨN ]]−1 [ΨN ]T ZN (8) excitation of the spacecraft simulator is ensured.
The values of J˜ and mr can then be extracted from x̂.

4. ESTIMATION USING RECURSIVE LEAST 5. STABILIZING CONTROLLER

SQUARES
The excitation torque τ is continuously applied to the
Haykin (2009) provided the algorithm for an online platform to keep its dynamics active so that the rate terms
method that utilises the recursive least squares technique. in (3) exists, which enables the estimation of J˜ and mr.
It has been adapted to suit the purpose, as explained Mathematically, τ causes the coefficient matrix [ΨN ] in
below. (7) to have full rank in order to aid in its inversion.
Consider the scenario mentioned in section 3. N − 1 However, a typical air-bearing structure limits the rotation
experiments have been conducted, so [ΨN −1 ] exists. From angles of the platform in the roll and pitch axes within
the N th experiment, new data collected is added to [ΨN −1 ] ±30o (Jung and Tsiotras (2003)). Therefore, a stabilizing
and ZN −1 . Let [CN ] denote the new row to be appended controller has to be turned ON in order to prevent the
to [ΨN −1 ] and z(N ) be that to ZN −1 . platform from hitting safety stops. Joshi et al. (1995)
developed a nonlinear robust stabilizing controller that
From (3) and (7) it is apparent that these variables take does not require knowledge of the unknown [J]. It is given
the values by
  tN  tN 
[CN ] = [Ω(tN )]−[Ω(t0 )]+ [ω×][Ω]dt [g×]dt 1
t0 t0 τ  = − [([α×] + α4 [I3 ])[Gp ] + γ(1 − α4 )[I3 ]] α − [Gr ]ω
 
3×9
(9) 2
tN (14)
z(N ) = τ dt where [Gp ] and [Gr ] are positive definite gain matrices
t0 3×1 and γ is a positive scalar, all chosen such that 0 <
The gain matrix [LN ] defined in Haykin (2009) becomes λmax ([Gp ]) < 2γ, and [I3 ] is the 3 × 3 identity matrix.
 −1 The vector α = [α1 α2 α3 ]T is the Euler parameter vector
[LN ]9×3 = [PN −1 ][CN ]T [I3 ] + [CN ][PN −1 ][CN ]T and satisfies the constraint α12 + α22 + α32 + α42 = 1.
(10)
Since Euler angles are used to describe the attitude of
where [I3 ] is the 3 × 3 identity matrix and [PN −1 ] is the
the platform, the following equations from Bekir (2007)
9 × 9 covariance matrix.
are employed to convert the Euler angles to the Euler
The value of x̂ is updated with each sampling. parameters.

237
5th International Conference on Advances in Control and
Optimization of Dynamical Systems
222 Athrey Ranjith Krishnanunni et al. / IFAC PapersOnLine 51-1 (2018) 219–224
February 18-22, 2018. Hyderabad, India

Fig. 2. Generated torque

Fig. 3. Measured body rate during estimation
φ θ ψ φ θ ψ
α1 = sin cos cos − cos sin sin
2 2 2 2 2 2
φ θ ψ φ θ ψ
α2 = cos sin cos + sin cos sin
2 2 2 2 2 2 (15)
φ θ ψ φ θ ψ
α3 = cos cos sin − sin sin cos
2 2 2 2 2 2
φ θ ψ φ θ ψ
α4 = cos cos cos + sin sin sin
2 2 2 2 2 2
6. NUMERICAL SIMULATION

The dynamics of the air bearing platform under the action

of control torque and gravity disturbance torque was
numerically simulated, with the plant parameters obtained
from a CAD model. The sampling was done at a frequency
of 100 Hz and for a total simulation time of 30 seconds.
The input torque τ (in Nm) was applied as
 
 t t t T
2· sin 2π sin 2π sin 2π , if − 5o < φ, θ < 5o
τ = 5 5 5
τ  , else
(16)
with the components of τ  being Fig. 4. Observed orientation during estimation
 
25.2 0 0 For initialization of the online method using recursive least
[Gp ] = 0 25.2 0 squares technique, δ = 10−4 was used. The results of this
0 0 28.8 method are given from Fig. 5 to Fig. 8. mrx , mry and
 
20 0 0 mrz reaches steady values within 5-10 seconds (Fig. 5).
[Gr ] = 0 20 0 Similar convergence is observed in the moments (Fig. 6)
0 0 32 and products (Fig. 7) of inertia. This is confirmed by the
γ = 20 convergence of the norm of the covariance matrix in Fig. 8.
These values were obtained after performing several runs Table 1 shows the results of the simulation. The % error
of the simulation model. Fig. 2 shows the variation of input was calculated by
torque with time. Fig. 3 and Fig. 4 display the angular rate ||Jerr ||
profile and variation of Euler angles of the platform with % error = × 100
time, respectively. ||J˜||
where Jerr is the vector difference between the true value
Least squares methods require very accurate measurement J˜ (obtained from the CAD model) and its estimate. A
of independent variables. The spacecraft simulator hard- similar formula holds for mr as well.
ware in Jung and Tsiotras (2003) had its resolution level
of the rate gyro as 0.029o /sec. (It also had a sun sensor From table 1, it is apparent that the linear least squares
with a resolution level of 0.05o to measure the attitude of technique estimated inertia and CM offset more accurately
the simulator.) (with errors of 0.1263 % and 0.0631 % respectively, com-

238
5th International Conference on Advances in Control and
Optimization of Dynamical Systems
Athrey Ranjith Krishnanunni et al. / IFAC PapersOnLine 51-1 (2018) 219–224 223
February 18-22, 2018. Hyderabad, India

Fig. 5. Estimation of unbalance values using RLS Fig. 7. Estimation of products of inertia using RLS

Fig. 6. Estimation of moments of inertia using RLS

Fig. 8. Covariance plot of RLS
Table 1. Results of estimation using linear least
squares and recursive least squares techniques greater than the sampling rate of the sensor. As a result,
the samples obtained from the sensor are perturbed by
Parameter True LLS RLS a white noise sequence, which introduces a zero-mean
Ix 43.2490 43.2654 43.1190 random walk error into the integrated attitude signal,
Iy 78.6600 78.6418 78.6004
whose standard deviation grows proportionally to the
Iz 98.0600 98.2176 97.4552
Ixy 0.6530 0.6295 0.6862
square root of time. It is common for manufacturers
Ixz 0.0340 -0.0131 0.3215 to specify noise using an√angle random walk (ARW)
Iyz 0.2628 0.2637 0.1459 measurement with units o / hour (Woodman (2007)).
% error − 0.1263 0.5231
mrx 0.1722 0.1722 0.1722 White noise with standard deviation σ = 10−5 was added
mry 0.0089 0.0089 0.0089 to the angular rate. The corresponding ARW measurement
mrz -0.1516 -0.1517 -0.1517 was calculated according to Woodman (2007) using the
% error − 0.0631 0.0707 simulation time-step δt = 0.01 sec.
√ √
pared to 0.5231 % and 0.0707 % in the recursive least ARW = σ · δt · 3600 · 180o /π = 3.44 × 10−3 o / hour
squares technique). (17)
The estimates vary slightly with the experiment. The
The output of a rate gyro will be perturbed by some average values along with the standard deviations, of each
thermo-mechanical noise which fluctuates at a rate much parameter, is shown in tables 2 and 3.

239
5th International Conference on Advances in Control and
224
Optimization of Dynamical Systems Athrey Ranjith Krishnanunni et al. / IFAC PapersOnLine 51-1 (2018) 219–224
February 18-22, 2018. Hyderabad, India

Table 2. Results of LLS with white noise proximately the same amount of time, irrespective of the
number of samples.
Parameter True Average S. D.
Ix 43.2490 43.2424 0.2078 7. CONCLUSION
Iy 78.6600 78.3464 0.6112
Iz 98.0600 98.3980 0.9818
Ixy 0.6530 0.5700 0.0937
The system dynamics and kinematics of a 3-degrees-of-
Ixz 0.0340 -0.0914 0.3965 freedom air bearing platform is derived and modified to
Iyz 0.2628 0.2390 0.4206 make it amenable to parameter estimation. The static
% error − 0.3653 − unbalance of the platform and its moments and prod-
mrx 0.1722 0.1720 0.0004 ucts of inertia are estimated using a batch method using
mry 0.0089 0.0089 0.0000 linear least squares, and an online method using recur-
mrz -0.1516 -0.1523 0.0012 sive least squares. The performance of the two methods
% error − 0.3234 − are compared through numerical simulation studies. It is
observed that the recursive least squares technique gives
better prediction of the platform unbalance values while
Table 3. Results of RLS with white noise
the batch method gives better prediction of the inertia.
Parameter True Average S. D. But the latter is computationally intensive when compared
Ix 43.2490 43.0336 0.0216 to the former, and is also more sensitive to sensor noise.
Iy 78.6600 78.4116 0.3087
Iz 98.0600 97.2661 0.4437 Since it gives better prediction of the unbalance values,
Ixy 0.6530 0.7039 0.0076 is less sensitive to noise, and is easier to implement on
Ixz 0.0340 0.4351 0.1200 practical systems due to lesser computational load, it
Iyz 0.2628 -0.0879 0.1764 is decided to use recursive least squares technique for
% error − 0.7614 − estimation of inertia and CM offset of the spacecraft
mrx 0.1722 0.1723 0.0000 simulator.
mry 0.0089 0.0089 0.0000
mrz -0.1516 -0.1515 0.0000 REFERENCES
% error − 0.0536 −
Bekir, E. (2007). Introduction to modern navigation
systems. World Scientific, Singapore.
From tables 2 and 3, the batch method is seen to have Fortescue, T.R., Kershenbaum, L.S., and Ydstie, B.E.
smaller error (0.3653 %) in computing inertia than the (1981). Implementation of self-tuning regulators with
online method (0.7614 %), but the situation is reversed variable forgetting factors. Automatica, 17(6), 831–835.
for the CM offset (0.3234 % and 0.0536 % respectively). Haykin, S. (2009). Adaptive Filter Theory. Pearson
An uncertainty of less than 1% in inertia values will not Education, Inc., London.
adversely affect the performance of the platform attitude Joshi, S.M., Kelkar, A.G., and Wen, J.T.Y. (1995). Ro-
controller. However, accuracy in the estimation of unbal- bust attitude stabilization of spacecraft using nonlinear
ance is critical for obtaining the desired performance for quarternion feedback. IEEE Transactions on Automatic
the air bearing test platform, and the online method is less Control, 10, 1800–1803.
sensitive to noise for that purpose. Jung, D. and Tsiotras, P. (2003). A 3-dof experimental
The bias of a rate gyro is the average output from the test-bed for integrated attitude dynamics and control
gyroscope when it is not undergoing any rotation (i.e. the research. AIAA Guidance, Navigation, and Control
offset of the output from the true value), in o /hour. When Conference and Exhibit.
a bias error is integrated, it causes an angular error which Kim, J. and Agrawal, B. (2009). Automatic mass balanc-
grows with time (Woodman (2007)). Since the estimation ing of air-bearing-based three-axis rotational spacecraft
process is conducted only for a short duration, the sensor simulator. Journal of Guidance, Control and Dynamics,
bias is ignored in this analysis. 32, 1005–1016.
Meissner, D.M. (2009). A three degrees of freedom test bed
The inertia properties may have to be re-estimated multi- for nanosatellite and CubeSat attitude dynamics, deter-
ple times during ground testing whenever there is a change mination and control. Ph.D. thesis, Naval Postgraduate
in the platform configuration, e.g. if a new harness line School.
is connected, or if a new sensor or actuator is added or Prado, J., Bissiacchi, G., et al. (2005). Three-axis air
replaced. Hence, an algorithm with less computational bearing based platform for small satellite attitude de-
requirement is preferred. The batch method involves col- termination and control simulation. Journal of Applied
lecting a large amount of data and processing it in a single Research and Technology, 3(3), 222–237.
step, while the RLS estimation technique performs the Schwartz, J.L., Peck, M.A., and Hall, C.D. (2003). Histor-
prediction at each time-step when a new set of data is ical review of air-bearing spacecraft simulators. Journal
collected. The [ΨN ] matrix in (7) has dimensions 3N × 9 of Guidance, Control, and Dynamics, 26(4), 513–522.
which takes up more memory as N becomes larger. In Vahidi, A., Stefanopoulou, A., and Peng, H. (2005). Recur-
comparison, the matrices to be evaluated in RLS have a sive least squares with forgetting for online estimation
constant maximum linear dimension of 9. Therefore, as of vehicle mass and road grade: theory and experiments.
the number of samples gets larger, any potential gain in Vehicle System Dynamics, 43(1), 31 – 55.
accuracy from using the batch method is offset by its Woodman, O.J. (2007). An introduction to inertial nav-
greater computational requirement. On the other hand, igation. Technical Report 696, Computer Laboratory,
the online method can provide accurate results in ap- University of Cambridge, United Kingdom.

240