Você está na página 1de 6

Spacecraft Inertia Estimation Via Constrained Least Squares

Jason A. Keim, A. Behcet Aqikmese and Joel F. Shields


Jet Propulsion Laboratory
California Institute of Technology
4800 Oak Grove Dr. M/S 198-326
Pasadena, CA 91109, USA
(818) 354-4321

Jason.A.Keim(jpl.nasa.gov

Abstract This paper presents a new formulation for spacecraft inertia estimation from flight data. Specifically, the in
ertia estimation problem is formulated as a constrained least
squares minimization problem with explicit bounds on the inertia matrix incorporated as LMIs (linear matrix inequalities).
The resulting minimization problem is a semidefinite optimization problem that can be solved efficiently with guaranteed convergence to the global optimum by readily available
algorithms. This method is applied to test data collected from
a robotic testbed consisting of a free rotating body. The results show that the constrained least squares approach produces more accurate estimates of the inertia matrix than standard unconstrained least squares estimation methods.

1. INTRODUCTION
This paper presents a new formulation for spacecraft inertia

estimation from data. The data are composed of the spacecraft quaternion and speeds of the reaction wheels. Initially,
a filtered version of the spacecraft attitude dynamics equation
of motion is used to set up a least squares parameter estimation problem. The filtered version of the differential equations eliminates numerical differentiation of noisy data to obtain the angular acceleration of the spacecraft and reaction
wheels. Then, explicit bounds on the inertia matrix are presented and incorporated into the least squares optimization
problem as LMIs (linear matrix inequalities). The resulting
minimization problem is a semidefinite optimization problem
that is solved efficiently with a guaranteed convergence to the
global optimum by readily available algorithms [1].

P r
its axis, i.e.,
tig andenatilt,the
horizotalaaxs,
anda about itSabout
vertical
spin.i.e.,
While spin
tip

is unconstrained, physical stops limit tip and tilt to a range of

sixty degrees.
Gravity influences rotation of the AP when the platform

This research is motivated by NASA-JPL's Formation Control Testbed (FCT). FCT is a hardware testbed developed
to validate the technology for future Autonomous Formation Flight (AFF) missions, such as TPF-I (Terrestrial Planet
Finder Interferometer) [2] and Stellar Imager. Three robots
will comprise the FCT when completed in 2006. Each robot
is made up of two bodies: the Translation Platform (TP) and
the Attitude Platform (AP). See Figure 1. The motion of the
robot must accurately emulate a spacecraft in deep space.

Hence grav1tahonal d1sturbnces must be m1nlmlzed. T

center-of-gravity (CG) does not coincide with the spherical


air bearing center-of-rotation (CR). The CG-CR offset induces a gravitational torque and deteriorates closed loop control performance. For this reason, the CG-CR offset must be
minimized. Additionally, in order to achieve precision control of the AP attitude, engineers must accurately identify the
APEner1ta. The
elpmyialion of the CG-CR offset andAthe den-

AY

BdH/dt as follows
co/

0 y stem

--- CG
,

:sz or.

co

BdH
dt +JixH.
Then, the following equation of motion for the attitude dy'dH

ydt

jth RFeaction

namics of AP follows

Wheel

CR:

~~~~~~~d~
-x=r

~~~~~~dtxH

CR

H+W x H=T,

(3)

where v) =B dvJ/dt for any vector vJ with coordinates given


by the column vector v e IR3. Letting matrix Ja C R3X 3 be
the representation of Ja and Jj be the representation of Jj in
the body frame, and defining

bined with saturation of the reaction wheels due to CG-CR


offset result in small data sets. In order to make accurate estimates based on these data sets, a new least squares approach,
constrained least squares, is developed.

S () =

In the following sections, we derive the equation of motion


that will be the basis for the inertia and CG-CR imbalance
estimation. The equation of motion is then reformulated in
a standard least squares form. To improve estimation accuracy, constraints are introduced in the form of LMIs. Then,
the experimental setup is presented followed by a discussion
on data processing. Finally, we estimate the inertia matrix
and the CG-CR offset using constrained least squares, standard least squares [6], and the total least squares methods [7].
The comparison of the methods indicate that the constrained
least squares approach produces significantly more accurate
inertia estimates with respect to an independent estimate of
the inertia matrix from a CAD-based model.

3
-X2

-X3

X2

(4)

0 -s
X1

for any vector x C JR3, (3) can be written as


M

Jan + an x Jgo + ,: J3
V
w'

r
T

Oj

j=

M
E
j=

Jjabj,
w

where J is the inertia of the system relative to CR give by

s22

J = Ja-maSa(ra) +

2. SYSTEM MODEL

_Jj
-mjS2(rj)).

j=1

(5)

In the equation above, it is assumed that the reaction wheels


rotate around a principal axis of inertia, ?j, that is assumed
to be an axis of symmetry for the reaction wheel fixed in the
AP body frame, i.e.,

In this section, we present the equations of motion describing


the attitude dynamics of the AP with reaction wheels. The
total angular momentum of the AP with the reaction wheels
H relative to CR is given by,

Ja:O + ra x (:O x maia) +


Zj=X (Jj( + 7j ) + [ij x ( x mM

(2)

where T is the net external torque applied to AP. Once a coordinate frame attached to the body of AP is chosen, referred to
as the AP body frame, we expressed all vectors in this frame
and (2) in matrix form as

Figure 2. Diagram of the AP and Reaction Wheels. System


CG is the center of gravity of the AP and the wheels. CR is
the center of rotation.

Jw

(6)

where Ij C IR+ is the corresponding principle moment of inertia, Wj = wa i for the jth reaction wheel, and by symmetry
of the wheel
Jj = O.
(7)
By using (6) and (7), (3) can now be expressed as

(1)

i)])

Ja is the AP inertia tensor relative to its center of mass, Jj is


the inertia tensor for the jth reaction wheel about its center of
mass (assumed to be coincident with its center of rotation), G
is the angular velocity vector of the AP relative to an inertial
frame, Jj is the angular velocity vector of the jth reaction
wheel relative to the AP, a is the position vector of the AP
relative to the CR, r is position vector of the jth wheel relative to the CR, ma is the mass of the AP and m~is the mass
ofthe jth wheel. See Figure 2. The notation 'dH/dt denotes
an inertial time derivative of the angular momentum. It can
be expressed in terms of the time derivative in the body frame

Jw + W x (Jw + E iiwiiii)

T - 3 I3wji. (8)

to aevr,teol xenltru sdet h GC


ofst .eni iue2a
n ersne eea'.
J3.
The gravitational torque Tg is given by
Tg =mr x C(q)g,
2

(9)

Equation (14) is written more compactly as following,


(17)
H(t)o =rT(t),

where the total mass of the system is


m
m

ma +

Tni ,

where
g is the gravity vector expressed in inertial frame, i.e., g=
[0, 0, g9.807]T (M/S2), and C(q) is the coordinate transformation from the inertial to the AP body frame given as a
function of the quatemion q e JR4 describing the orientation
of the AP frame relative to the inertial frame. Summarizing
the discussion above and considering the kinematics of the
AP rotation via a quatemion description, we have the following differential equations that are the basis for the inertia and
CG-CR imbalance estimation,
M
Jc + w x Jw + z x C(q)g S [wjIjnj x w - I1Wfji]
j=1

Q(u)
[ 1

2(q2

Tj(t) = 3 wj (t)Ij?j x w - I wj (t)nj .


j=1

When a sampled data set is provided for wo, zb, q, wj, j


1, .... M, equation (17) directly implies a least squares problem to estimate b. However, AP sensors directly measure
f w, q, wj from which U and Ijwjhj must be derived. Since
w; and Ij is known very accurately, the above discussion implies that U and w; must be obtained numerically when (17)
is used to set up the least squares estimation problem.'

(10)

(11l)

=-Q0(wJ)q
2

2(q3ql - q2q4)
2(q2q3 + qlq4)
I - 2(q2 + q2)

2(qlq2 + q3q4)
1 -2(q2 + q2)
2(q2q3 - qlq4)

q3)

2(qlq2 -q3q4)
2(q3ql + q2q4)

and q

S(C(q(t))g)

In this paper, we use a filtered version of the equation (14) to


avoid the numerical differentiation needed to obtain angular
accelerations. The second-order causal filter,
zhmre(12)
z~~~~~~~~~~~
=s nr (12)~0
<1 (8
W)2
W)3
o -W)I
) s2 + 2wn + w2
|W3
W2
<w
<
1, (18)
0
-W2
WI
W3n
applied to Equation (14) gives rise to the following filtered
L -wi -W2 -W3 0

where

C(q)

[ Q(cZ(t)) + S(wo(t))Q(w(t))

H(t)
and

form
[ Q(9f (t)-5(t)w(0)) + bf (t)

[ql, q2, q3, q4]T

j=1

The components of

gf (t) =

Bquation 19
Lo

pf,j(t) ={ww}(H),

3. ESTIMATION OF INERTIA AND MASS


IMBALANCE

Of (t)
T1f (t)

In this section, we derive the equation that leads to the least


squares parameter estimation problem. Letting

F{ }

=
=

9)

are

Wf, (t) = w}(,


Tf{wI} (t),

wheel.wf(t

} )
{S() (w)(LO

o ....
oc.
A4
oo,. . 4,

F{C ( (H)) } } (H)

is the filtering operator with transfer function


f{ } is the filtering operator with transfer function
=
sF(s) and 0(t) is the impulse response of F(s).
F(s)
Note that we use the fact that S and Q are linear operators
in their arguments in the derivation of the equation above.
The filtered version of the dynamics (14) is then written in
the following compact form,
G(t) X g(t),
(20)
where
F(s),

Jv =[ Jll J22 J33 J12 J13 J23 ]T


(13)
the equation describing the dynamics of AP (10) can be

rewritten as

[Q(L) + S(w)Q(w) S(C(q)g)] X

[wjIffi x w -

] 4

N
,Ij [nj X pf,j (t) - wfj (t) + 6(t)wj (0)]
Z'

Remark 1: Note that J and z are the unknown constant parameters that will be identified as a part of inertia identification. Here w is obtained from the quatemion measured by
the celestial sensor [8], and wj are measured by tachometers.
Numerical differentiation of wj obtains wj for each reaction

j=

S(Trf (t))

(14)

where

where
G(t) =[ Q(6f (t)-6(t)w(0)) + bf (t)

S is given by (4), and


an 12

QQv)

=
L

2
0

13

1
0

13

11 12

S(r1]f(t)) ]

~ ~ ~ ~ ~ ~ conservation of momentum equation rather than (2) to obtain an equation that

1Note that, if the external torque due to gravity does not exist, one can use

(16)

does not contain the derivatives ofthe angular rates [4]. This is not applicable
in our case.

and
N

g(t) = 3 Ij
j=1

[nj X pf j (t) Wfj (t) + 5(t)wj (0)]


-

0.8

0.6-

The actual experimental data is obtained at discrete time instances, 0, t1,... tN, i.e., the data is sampled. Therefore (20)
describes an equality at each sampled time instance tk. Consequently, we have the following set of equalities

G(tk)

g(tk),

O,1, ...,N.

X 04
c

<

(21)

Additionally the filtering required to obtain G(tk) and g(tk)


is performed digitally by using a discrete approximation of
the filter (18), that is, for any signal v(t)

-0.4

{r}(LHj)

g(0)
g(t1)

b_

g (tN)

[ G(tN)

(0)

04

[0 1
I

01

000]

600

400

800

Time (s)

1000

1200

P5

[0

0 O O
[ l o

6
J

[ 0
O

1
1

Equation (24) introduces the LMI constraints. It is a semidefinite programming problem [9], [10] that can be solved very
efficiently with guaranteed convergence to the unique optimum solution by existing algorithms [1]. Additional constraints can be accommodated in this framework to obtain
a good inertia estimate with small number of data samples,
such as bounds quantifying the diagonal dominance ofthe inertia matrix, i.e.,

(0)

I|| *

-<-41q1 <

054

< 7)41q51,

-74202 < 04 < -)4202,


42
.
-751a1 . 5 . 7511:

At this point one can use standard least squares solution techniques [6] to obtain a solution for q. However, such a formulation fails to capture known constraints of the estimated
parameters. One obvious constraint is J = jT > 0, that is,
the inertia matrix is symmetric positive definite. Indeed more
specific bounds are generally known about the inertia matrix,
such as
aI < J </31,
(23)

(5 74q1,

--5343 < 4b5 < <-)/53q3,

-762X2

where '741,

where a and /3 are positive scalars bounding the maximum


and minimum eigenvalues of J. Then the overall constraint

least squares problem can be written as

min0(A

200

Figure 3. Time History of the AP Attitude

where FT is the filtering operator for the discrete approximationofF, andTF[{LI,, LrI}I is the value of the filtered
signal at kth sample obtained by filtering the data.vo, ... ,VN.
Then, the least squares problem parameter estimation is given
by
=byb,(22)
where A C IR3(N+l) x9 and b C ]R3(N+l) given by

G(0)
G(ti)

0.2

<

06 < -76202,
7)63q53,

. 6<
<_

'742, '751, <)752, <)762, <63 are positive scalars.

4. DATA COLLECTION AND POSTPROCESSING

-b)TW(A -b) subjectto

For comparison of the standard least squares, total least


squares, and constrained least squares approaches, we have
al < i=l OiPi < /31
(24)
acquired test data from the FCT. The FCT avionics closely
emulate those of a spacecraft. Torques produced by three
Ci . cj < di, i = 1, . . ., 9
orthogonal mounted reaction wheels and sixteen cold gas
(Jet
=. . budthrusters. rotate the AP. Optical gyroscopes and a JPL
whrcndd, comS
where ci anddI i
, T9
bound the individual
1,T...
.... ,gq5] , W WT > 0 is a weight
Prop ulsion Laboratory) developed celestial sensor (CS) [8]
ponents of Xb
measure the rotation of the AP. A closed loop controller romatrix, and P1 ... ., P6 are symmetric matrices that are used
tated the AP without violating the physical constraints imto impose minimum and maximum eigenvalue constraints on
posed on tip and tilt. The controller oscillated the platform
the inertia matrix. They are given by
i 15 degrees in tip and tilt and 20 degrees in spin. Figure 3
l
2P]
=
illustrates the rotation as measured by the CS. Here, a, /3 and
P1 =[0 ],P2~[
L o o o jL

L o

a~are the Euler angles of azyx rotation.

1 j

Our filtered estimation approach, Equation (20), requires


measurement of the AP angular velocity and the reaction
wheels speeds relative to the AP body. Wheel speeds are directly measured by tachometers. Rotation of the platform is
measured by two devices: the CS and gyroscopes. The gyroscopes of the FCT run in accumulated angle mode producing
measurements as integrated angles. The celestial senor measures the AP attitude in the form of a quaternion. Processing
either the integrated angles or the quaternions is necessary to
resolve the body angular velocity. Both process entail comparable amounts of work and complexity [11]. Here, we derive the AP angular velocity from the measured quaternions,
because the CS measurement frame is the AP frame and CS
noise is stationary. As part of the preprocessing the data, the
quaterniontime history is smoothened to reduce noise affects
via methods presented in [9]. From the smoothed quaternion data set, we determine the spin rate of the AP by methods found in [11]. Finally, standard least squares, total least
squares and constrained least squares are applied to the derived body angular velocity and the measured reaction wheel
spin rates to estimate the AP inertia and mass imbalance.

Using the SolidWorks estimate as the basis for comparison,


the relative errors of the least squares estimates are defined
as:
= 0.2036,
els = Js Jsj1/1Jsj
sw
=
=

eCil

Jtis =
Jcis=

0.1988:

0.1259.

here most closely agrees with the preexisting, independentlyderived estimate of the AP inertia. Future work will be to
develop techniques for error estimation via simulation.

Although the characteristics of the FCT robot necessitated


this work, the algorithm developed here can produce more
accurate estimates of the inertia matrix than standard least

k m2
0
94
10.8692-0.7 670.7670m9
kg
13

squares formulations with smaller data sets. Consequently, it


will useful in other engineering applications beyond FCT. Especially, this method can provide accurate inertia estimates in
space missions with a minimal consumption of the valuable
resources such as fuel and science observation time. Additionally, the constrained least squares approach is a natural
extension ofthe standard least squares methods for inertia estimation via the inclusion of the matrix constraints as LMIs.

0l6

0974

= [ -0.5515 -0.6313 0.4282 ] x 10-4 kg. m


x 10-4 kg m
Ztis = [ -0.5515 -0.6313 0.4283
-0.6071 -0.6786 0.0413 XT 10-4 kg m

ACKNOWLEDGEMENTS

Zis

We wish to gratefully acknowledge Dr. David S. Bayard


for his very valuable comments and suggestions. This research was performed at the Jet Propulsion Laboratory, California Institute of Technology, under the contract with National Aeronautics and Space Administration.

Zcl

An independent estimate of the AP inertia relative to the CR


exists as the result of a comprehensive SolidWorks model.

This estimate is

REFERENCES

[ 8.0105 -0.2330 -0.50631


J =
0.0163 kg m
-0.2330 11.4476
[ -0.5063
0.0163 14.9564 J

[1] J. F. Sturm, "Using SeDuMi 1.02, a matlab toolbox


for optimization over symmetric cones," Optimization
Methods and Software, vol. 17, no. 6, pp. 1105-1154,
2002.
2001, pp.
[2] P. Lawson, "The Terrestrial Planet

From this model, the lower bound on J were chosen to be


0o50oftelargetdiagoalterm i..
[ 7.5 0.0 0.0 1
0.0 7.5 0.0 kg.m2

lls/

-sj/IJsj

three estimates, the estimate using the technique developed

5.8399 -0.7533 -0.9346 1


0.7733 kg m2
-0.7533 10.8726
0.7733
15.3611
-0.9346
J
L
F 7.5691 -0.5298 -0.0300 1
-0.5298 11.6034
0.8122 kg. m2
[ -0.0300
0.8122 16.0833 J

J>

cis

In this paper, we present an inertia estimation technique


based on a constraint least squares minimization problem
with LMIS This approach can incorporate the positive definiteness ofthe inertia matrix directly. Furthermore, additional
bounds on the inertia matrix can also be considered. These
are particularly useful in obtaining more accurate estimates
of the inertia matrix when there is a limited amount of angular motion allowed (in terms of angular velocities and displacements) as in the example case of FCT robot presented
here. This technique, along with standard least squares and
total least squares, is then applied to estimate inertia matrix
of the FCT robot AP by using the experimental data. Of the

A conventional least squares estimate, a total least squares


estimate, and constrained least squares estimate are applied
to test data described above. The estimates of the AP inertia
and CG-CR offset parameter are the following:
-0 75678

6. CONCLUSIONS

5. RESULTS FROM EXPERIMENTAL DATA

J,s--0.7568 10.8692

Iti

eCtis

Finder,"'

2005-2011.
[3] M. L. Psiaki, "Esimation of a spacecraft's attitude dy[0.0 0.0 7.5 Jnamics parameters by using flight data," Journal of

Guidance, Control, and Dynamics, vol. 28, no. 4, pp.


594-603, July-August 2005.
[4] A. Y. Lee, "In-flight estimation of the Cassini spacecraft's inertia tensor," Journal ofSpacecraft and Rockets, vol. 39, no. 1, pp. 153-155, 2002.
[5] 5. Tanygin and T. William, "Mass property estimation
using coasting maneurvers," Journal ofGuidance, Con-

Joel Shields is a member of the Guidance and Control Analysis Group at the
Jet Propulsion Laboratory. He received
his B.S. degree in Applied Mechanics
m UC San Diego in 1990 and his
M.S. and Ph.D. degree in Control Sys1997, respectively. His dissertation re-

August 1997.
C. L. Lawson, Solving least squares problems.
Philadelphia: SIAM, 1995.
S. V. Huffel and J. Vandewalle, The Total Least Squares
Problem: ComputationalAspects andAnalysis. SIAM,
1991.
J. F. Shields, "The formation control testbed celestial
sensor: Overview, modelling, and calibrated performance," in 2005 IEEE Aerospace Conference Proceedings, Big Sky, MT, March 2005.
S. Boyd and L. Vandenberghe, Convex Optimization.
Cambridge University Press, 2004.
S. Boyd, L. El Ghaoui, E. Feron, and V. Balakrishnan,
Linear Matrix Inequalities in System and Control Theory. SIAM, 1994.
J. R. Wertz, Spacecraft Attitude Determination And
Control. Kluwer Academic Publishers, 1997.

search was on control ofexercise machines. He is the holder


of a patent based on this work. Upon graduation, Dr. Shields
worked as a consultant before joining the technical staff at
JPL. Dr. Shields has extensive experience in design ofmechatronic systems, system identification, adaptive control, precision servo control, and real time applications of control systems. At JPL he has worked on precision pointingfor space
based interferometry andformationflying.

[6]
[7]

[8]

[9]
[10]

[11]

Jason A. Keim received his B.S. degree


in Biomedical Engineeringfrom the University of Southern California in 2002.
Afterward he joined the Jet Propulsion
Laboratory as a member of the Guidance and Control Analysis Group. He
expects his M.S. degree in Mechanical
Engineering from California State University, Los Angeles in 2006.
A. Behfet Av,kme,e received his MS
in Mechanical Engineering, and his Ph
in Aerospace Engineering, both degrees
from Purdue University. He was a Visiting Assistant Professor at the School of
Aeronautics and Astronautics at Purdue
University before joining JPL. His re_
search interests include linearlnonlinear
system theory, robust and nonlinear control, formation flying control, LMIs (linear matrix inequalities) in control theory, optimal control, and convex optimization and semidefinite programming. He is currently developing guidance and
control algorithms for formation flying, small body proximity operations, and Mars pinpoint landing as well as developing interior point algorithms for the numerical solution of
semidefinite programs.

Você também pode gostar