You are on page 1of 20

Control Engineering Practice 10 (2002) 951–970

Mixed H2/HN state-feedback design for microsatellite


attitude control
Ciann-Dong Yanga,*, Yun-Ping Sunb
a
Department of Aeronautics and Astronautics, National Cheng Kung University, Tainan 701, Taiwan
b
Department of Mechanical Engineering, Chengshiu Institute of Technology, Kaohsiung 804, Taiwan
Received 20 September 2000; accepted 28 January 2002

Abstract

Owing to a great decrease in size and weight, the pointing accuracy of microsatellite is vulnerable to space environmental
disturbances and the internal uncertainty of moment-of-inertia variation. Mixed H2 =HN control, giving consideration to both
stability robustness and root-mean-square (rms) performance, is particularly attractive for attitude controller design of
microsatellites. By using linear matrix inequality method, the numerical solution of mixed H2 =HN state-feedback controller can
be efficiently solved. The performance differences between mixed H2 =HN controller and its two extremes—pure H2 controller and
pure HN controller—are discussed in detail. Mixed H2 =HN controller shows the remarkable capability of achieving a balanced
compromise between H2 and HN performances. r 2002 Published by Elsevier Science Ltd.

Keywords: Microsatellite; Attitude control; Linear matrix inequality method; Mixed H2/HN control

1. Introduction 1998), Earth observation (Grassi, Vetrella, & Moccia,


1995), atmospheric data collection (Sahraoui, Boudje-
The earliest satellites are very small and simple mai, Mohammed, & Abderrahmane, 1998), space
because of the limitations of satellite technology and science (Wolfe, Alfriend, & Leonard, 1995), and
launch vehicle capacity. After the late 1960s, more and communication (Muncheberg, Krischke, & Lemke,
more large satellites were built owing to gradually 1996).
maturing technology and growing demand. However, In order to meet high-accuracy performance on
the following reasons change this trend and reenergize pointing requirement, three-axis attitude control is
the development of small satellites. First, in the foresee- usually applied to microsatellite, leading to a typical
able future no launch technology breakthroughs appear multi-input–multi-output (MIMO) control system. Two
to be on the horizon to afford larger satellites. Second, major problems in attitude control design of micro-
small satellite cluster is a new application suitable for satellite are the disturbances from space environment
many space missions. Third, from risk-sharing view- and the perturbation of spacecraft’s moment-of-inertia
point, a number of smaller satellites have a significant variation. For low-Earth orbit (LEO) spacecraft below
reliability advantage over a bigger one. Fourth, small 1000-km altitude, gravity-gradient torque, aerodynamic
satellite would be an economical choice due to the torque, and Earth magnetic torque are the major
limited budget. Therefore, microsatellite, a kind of disturbances. On the other hand, the movement of
modern small satellites, with mass o100 kg, becomes a payload and appendages, such as telescope, camera and
promising space vehicle for its advantages of reducing solar array, usually causes the change of moment of
cost, risk, and manufacturing time. It plays a more and inertia. For a microsatellite with mass below 100 kg, its
more important role in various space missions, such as moment of inertia is usually not more than 20 kg m2
position location (Alonso, Anigstein, & Sanchez-Pena, (Wertz & Larson, 1999, pp. 336–337), and thus its
attitude is very sensitive to disturbances and the
*Corresponding author. Tel.: +886-6-2757575; fax: +886-6-
moment-of-inertia variation.
2389940. In the face of disturbance and uncertainty, robust
E-mail address: cdyang@mail.ncku.edu.tw (C.-D. Yang). control theory provides designers a systematic approach

0967-0661/02/$ - see front matter r 2002 Published by Elsevier Science Ltd.


PII: S 0 9 6 7 - 0 6 6 1 ( 0 2 ) 0 0 0 4 9 - 7
952 C.-D. Yang, Y.-P. Sun / Control Engineering Practice 10 (2002) 951–970

for analysis and synthesis of attitude controller. It is well instruments that could be used to measure the attitude
known that H2 control and HN control are two main and angular rates of spacecraft accurately. For example,
streams of robust control theory (Zhou, Doyle, & star tracker provides the attitude information and rate
Glover, 1996; Skogestad & Postlethwaite, 1996; Burl, gyro provides rate information (Merhav 1996; Cruise,
1999). The time domain interpretation of H2 norm is a Bowles, Patrick, & Goodall, 1998; Helvajian, 1999).
measure of the root-mean-square (rms) value of the Thus measuring the states of microsatellite including
performance output driven by unit impulse or unit attitude angle and angular rates is indeed feasible. On
intensity white noise, and HN norm is a measure of the other hand, there are some essential problems in
robust stability with respect to uncertainty established in output-feedback design. For instance, a higher-order
frequency domain that represents the worst-case at- controller resulted from output-feedback design would
tenuation for sinusoidal input. Thus, a combination of lead to difficulties in practical implementation, and Keel
H2 control and HN control, called mixed H2 =HN and Bhattacharyya (1997) pointed out the controller
control, was first introduced by Bernstein and Haddad fragile problem of high-order controllers from dynamic
(1989) that minimized the H2 norm of some closed-loop output-feedback design. Therefore, state-feedback de-
function subject to the HN norm constraint of another sign would be a reasonable choice than output-feedback
closed-loop function. Khargonekar and Rotea (1991) design, if all states can be measured.
considered state- and output-feedback problems of Two issues are worked out in this paper: first, to
mixed H2 =HN control and gave efficient convex complete a mixed H2 =HN state-feedback design in
optimization approach to solve the coupled nonlinear microsatellite attitude control, and second, to explore
matrix Riccati equations. Scherer (1995) utilized the the performance difference between mixed H2 =HN
Youla parameterization to express the performance controller and its two extremes—pure H2 controller
requirements measured by H2 and HN norms and and pure HN controller.
reduced the H2 =HN control problem to sequences of To the first issue, two practical problems for micro-
finite dimensional convex optimization problem. Whor- satellite attitude control are moment-of-inertia variation
ton (1997) proposed a homotopy algorithm for fixed- and space environmental disturbances. Owing to the
order mixed H2 =HN controller synthesis and verified the tiny mass and volume of microsatellite, the attitude of
feasibility of the proposed method via the experimental microsatellite is more sensitive to disturbances and
flexible space structure at NASA/Marshall Space Flight moment-of-inertia variation. The key to obtain a good
Center. attitude controller is to find out a systematic method
With the development of numerical algorithms for that can make a compromise between stability robust-
solving linear matrix inequality (LMI) problems in the ness and rms performance effectively. In this paper a
last 5 years, the LMI approach have emerged as a useful thorough design procedure including attitude kine-
tool for solving a wide variety of control problems matics and dynamics derivation, disturbance estimation,
especially for multi-criterion problems (Boyd, El uncertainty modeling, and controller synthesis is com-
Ghaoui, Feron, & Balakrishnan, 1994; El Ghaoui & pleted. An integration test including reaction wheel
Niculescu, 2000). Chilali and Gahinet (1996) addressed dynamics is finally completed. Mixed H2 =HN controller
the state-feedback H2 =HN design with regional pole is proved to achieve balanced performance in micro-
placement. Scherer, Gahinet, and Chilali (1997) satellite attitude control.
presented an overview of the LMI approach to the To the second issue, someone might ask that is it a
multi-objective control synthesis of both linear state- good choice to make the comparison based on state-
feedback and output-feedback controllers. In this paper, feedback design instead of output-feedback design?
mixed H2 =HN state-feedback attitude controller of an State-feedback design leads to the simpler controller
Earth-oriented microsatellite subject to environmental structure than output-feedback design, and furthermore
disturbances and moment-of-inertia uncertainty is de- it reveals the ideal performance that a dynamical
signed by using the method of Chilali and Gahinet feedback control system could achieve. Two common
(1996). reasons make output-feedback design indispensable: the
In the literature state-feedback design for attitude sensors may be too expensive to afford, or some states
control is not very unusual. The following papers are may be physically impossible to measure. The first point
some cases of state-feedback design: Sheen and Bishop might be a good reason to derive output-feedback
(1994); Wie, Liu, and Sunkel (1995); Grassi et al. (1995); design. Unfortunately, output-feedback design degrades
Kissel (1995); Haley et al. (1997); Wiemer (1998); Won the performance because of the more complex controller
(1999); Somov et al. (1999); and Chen, Wu, and Jan structure as well as the inevitable deviation between
(2000). There is some limitation for the application of actual state values and estimated state values. As for the
state-feedback design in comparison with output-feed- second point, in microsatellite the states can be
back design because state-feedback design needs all effectively and directly measured by contemporary
states to be measured. However, there are many technology. To sum up, state-feedback design provides
C.-D. Yang, Y.-P. Sun / Control Engineering Practice 10 (2002) 951–970 953

an equal basis to judge the performance of different with respect to uncertainty). Second, the mixed H2 =HN
types of controllers designed by various design methods. controller with stricter HN constraint has better robust
It is also recognized that the observer or filter, such as stability with respect to inertia matrix uncertainty, and
the Kalman filter, plays an important role in attitude superior attenuation capability to sinusoid disturbance,
determination and control system. Although the mea- whereas it sacrifices the rms performance of impulse
surement of attitude and angular rate is available for response.
most satellites by using the integrated star tracker/ The paper outline is as follows. In Section 2, the
inertial reference unit (gyrometer/accelerometer pack- modeling of microsatellite attitude dynamics is pre-
age) sensors (O’Donnell & Mangus, 1998), there are sented and the environmental disturbances are analyzed
indeed practical problems in processing these measure- and calculated. In Section 3, the moment-of-inertia
ment data from various types of sensors. For example, uncertainties are pulled out from the attitude dynamics
the gyrometers and star sensors give their measurements model to form the standard synthesis model including
at different frequencies, with different delays, different perturbation blocks and nominal plant. The synthesis
accuracy, and different drawbacks. One of the best model, in turn, is used in Section 4 to derive the LMI
solutions is to use some Kalman filters to predict and formulations of pure H2 control, pure HN control, and
correct measurement data from sensors (Bank, 1995; mixed H2 =HN control. Section 5 performs attitude
Haley et al., 1997; Wiemer, 1998; Zimbelman & Watzin, controller synthesis, makes a detailed comparison of
1999). different types of controllers, and completes an inte-
For Earth orbiting satellites, Bank (1995) uses the grated test in which the reaction wheel dynamics are
CT-633 star tracker, a three-axis magnetometer (TAM), included. Finally, conclusions are presented in Section 6.
and a Kalman filter to provide accurate attitude
information. In the transition region and coronal
explorer (TRACE) satellite, a Kalman filter is employed 2. Microsatellite attitude dynamics
for attitude determination that uses the sun sensor and
TAM measurements to update a gyro propagated Define a body-fixed reference frame B with its origin
attitude solution (Zimbelman & Watzin, 1999). The located at the center of mass of a microsatellite and with
, , ,
attitude determination system (ADS) of the midcourse basis unit vectors fb 1 ; b 2 ; b 3 g being along the principal
space experiment (MSX) satellite (Haley et al. (1997)) axes. The Euler’s equations of motion (Chobotov, 1991)
and the global imaging system 2000 satellite (Wiemer, for the microsatellite are then given by
1998) contains star trackers for attitude information,
’ 1  ðI2  I3 Þo2 o3 ¼ T1 ;
I1 o ð1Þ
ring-laser-gyros for rate information and other coarse
sensors. The measurement data are processed within
’ 2  ðI3  I1 Þo3 o1 ¼ T2 ;
I2 o ð2Þ
ADS in a double filtering process, a measurement
mixing filter and a Kalman filter, to minimize the long- ’ 3  ðI1  I2 Þo1 o2 ¼ T3 ;
I3 o ð3Þ
and short-term noise errors. The attitude rate measure-
ment from the rate gyros is used at higher frequencies where I1 ; I2 ; I3 are the principal moments of inertia,
while the attitude measurement from the star trackers is o1 ; o2 ; o3 are the body-axis components of angular
used at lower frequencies. The measurement-mixing velocity, and Ti ¼ ui þ Tdi are the body-axis compo-
filter provides the optimal mixture of attitude and rate nents of the external torques acting on the microsatellite,
measurements. The Kalman filter provides an optimal which contain control torques ui and environmental
estimate of attitude/rate states from measurement and disturbance torques Tdi :
rigid body dynamics model prediction, where the The attitude control problem for microsatellites is
dynamics model is used at higher frequencies than the unlike that for large satellites. For large satellites the
measurement used at lower frequencies. The detailed flexible, lightly damped modes should be the major
discussions and solutions to the problems of integrating concern. For microsatellites without flexible appendages
measurement data from various types of sensors are it is natural to consider the rigid body dynamics only;
referred to Haley et al. (1997) and Wiemer (1998). however, if microsatellites contain flexible appendages,
After understanding the practical problems and the it will be necessary to take the flexible modes into
remedy of using star trackers and gyrometers, this paper consideration. In this paper, the microsatellite is small
brings focus on the control law design assuming that the and compact, without any flexible part.
states are available. The observer (filter) design problem
is not addressed in this paper. 2.1. Coordinate transformation
Two highlights can be summarized. First, the mixed
H2 =HN control design provides an impressive flexibility In addition to the body-fixed reference frame B;
to tune controller to compromise between H2 cost (rms consider a local vertical local horizontal (LVLH)
value of impulse response) and HN cost (robust stability reference frame A with its origin at the center of mass
954 C.-D. Yang, Y.-P. Sun / Control Engineering Practice 10 (2002) 951–970

of the microsatellite. The LVLH frame has a set of unit where


, , , ,
vectors fa 1 ; a 2 ; a 3 g; with a 1 along the velocity direction 2 3
of microsatellite in orbit plane, , a 3 toward the Earth, and C11 C12 C13
,
a 2 being the direction according to the right-handed 6 7
4 C21 C22 C23 5
Cartesian frame. The spatial directions of frames A and C31 C32 C33
B are described in Fig. 1. To describe the orientation of 2 3
the body-fixed reference frame B with respect to the cycc cysc sy
6 7
LVLH reference frame A in terms of three Euler angles, ¼ 4 sfsycc  cfsc sfsysc þ cfcc sfcy 5; ð5Þ
the following successive rotations of Euler angles are cfsycc þ sfsc cfsysc  sfcc cfcy
applied:
and sf sin f; cf cos f; etc. As a result, the angular
, , , ,
1. Rotate the frame fa 1 ; a 2 ; a 3 g about a 3 through the velocity of the body-fixed reference frame B relative to
,0 ,0 ,0
yaw angle c to the frame fa 1 ; a 2 ; a 3 g: the LVLH frame A becomes
,0 ,0 ,0 ,0
2. Rotate the frame fa 1 ; a 2 ; a 3 g about a 2 through the
,00 ,00 ,00 , B=A B=A , B=A , B=A ,
pitch angle y to the frame fa 1 ; a 2 ; a 3 g: o ¼ o1 b1 þ o2 b2 þ o3 b 3;
,00 ,00 ,00 ,00
3. Rotate the frame fa 1 ; a 2 ; a 3 g about a 1 through the
, , ,
roll angle f to the frame fb 1 ; b 2 ; b 3 g: where
2 B=A 3 2 32 ’ 3
The resulting complete coordinate transformation can o1 1 0 sy f
6 B=A 7 6 76 7
be expressed as 6o
4 2 5
7 ¼ 40 cf sfcy 54 y’ 7
6
5: ð6Þ
2, 3 ’
2 32 , 3 o
B=A 0 sf cfcy c
b C11 C12 C13 a1 3
6 17
6, 7 6 76
6 , 7
7
6 b 2 7 ¼ 4 C21 C22 C23 54 a 2 5 The angular velocity of the body-fixed reference frame
4 5
, C31 C32 C33 , B relative to the inertia reference frame N fixed in the
b3 a3
Earth center becomes
or , , B=N , B=A , A=N , B=A ,
o o ¼o þo ¼o  na 2; ð7Þ
2 3 2 3
, 2 3 ,
a1 C11 C21 C31 6 b 1 7 where n is the orbital rate of the microsatellite. Thus,
6, 7 6 76 , 7 substituting Eqs. (4)–(6) into Eq. (7) leads to
6 a 2 7 ¼ 4 C12 C22 C32 56 b 2 7; ð4Þ
4 5 4 5
, C13 C23 C33 , , , ,
a3 b3 ,
o ¼ o1 b 1 þ o2 b 2 þ o3 b 3 ;

b1 a1

a3

Earth
Microsatellite

b3

b2

Orbital Path a2

Fig. 1. Coordinate axes of microsatellite in a circular orbit.


C.-D. Yang, Y.-P. Sun / Control Engineering Practice 10 (2002) 951–970 955

where about. On the assumption that the microsatellite center


2 3 2 32 ’ 3 2 3 of mass is in a Keplerian circular orbit and the Earth is
o1 1 0 sy f cysc
6 7 6 6
7 ’7 7 6 7 spherical, the gravity-gradient torques along the body
4 o2 5 ¼ 4 0 cf sfcy 564 y 5  n4 sfsysc þ cfcc 5: axes become (Wie, 1998, pp. 365–369)
o3 0 sf cfcy ’
c cfsysc  sfcc
Tg1 ¼ 3n2 ðI3  I2 Þcos2 y cos f sin f; ð12Þ
ð8Þ
Tg2 ¼ 3n2 ðI3  I1 Þcos y sin y cos f; ð13Þ

For small attitude deviation from LVLH orientation, Tg3 ¼ 3n2 ðI1  I2 Þcos y sin y sin f: ð14Þ
the following linearized attitude kinematics can be
obtained:
2.2.2. Magnetic torque
o1 ¼ f’  nc; ð9Þ The interaction of the spacecraft’s magnetic materials
with Earth’s magnetic field results in the magnetic
o2 ¼ y’  n; ð10Þ torque expressed as
, , ,
’ þ nf:
o3 ¼ c ð11Þ T ¼M B;
,
where M is the residual magnetic moments of the
2.2. Environmental disturbances analysis spacecraft caused by permanent and induced magnetism
,
and the spacecraft-generated current loops, and B is the
Gravity-gradient torque, aerodynamic torque, and geocentric magnetic flux density. Since the Earth’s
Earth magnetic torque are the primary environmental magnetic field is roughly dipolar, its magnitude can be
disturbances (Wertz, 1978) for spacecrafts in LEO approximated as
within 1000 km. The following estimations of the m
B ¼ 3E ð1 þ 3 sin2 YÞ1=2 ;
disturbances are based on the values of microsatellite r
parameters given in Table 1. where r is the distance from dipole (Earth) center to
spacecraft, mE ¼ 8:1 1015 T m3 is the magnitude of the
2.2.1. Gravity-gradient torque Earth’s magnetic moment vector along the magnet axial
Due to the gravity differences over an extended object direction, and Y is the magnetic latitude measured from
located in Earth’s gravitational field, gravity-gradient the geomagnetic equator (Chobotov, 1991, p. 84). In a
torque applied at center of mass of the object comes 500 km orbit, r ¼ ð6378 þ 500Þ ¼ 6878 km; the worst-
case magnetic field turns out to be B ¼ 5 105 T; and
assuming that the residual magnetic moment of the
Table 1
microsatellite is M ¼ 1 A m2 ; the worst-case magnetic
Microsatellite main parameters
disturbance torque is
Mission Imaging the earth
Mass (kg) o60 Tm ¼ MB ¼ 5 105 N m: ð15Þ
Inertia moments (kg m2)
Principal moments of inertia I1 ¼ 18:4; I2 ¼ 18:2; I3 ¼ 6:8
Products of inertia o1 (can be neglected) 2.2.3. Aerodynamic torque
Variation range 10% The rapid spacecraft motion through the tenuous
Size (cm) 50 50 60 upper atmosphere causes aerodynamic torque. The
Orbit torque can be expressed as
Type Circular
, 1 , ,
Altitude (km) 500 T ¼ rV 2 Cd Aðu s cp Þ;
Orbital rate n (rad/s) 0.0011 2
Attitude control type Three-axis control by three where r is the atmospheric density, Cd is the drag
reaction wheels
Reaction wheel ITHACO TW-2A40 coefficient, V is the velocity of the spacecraft, , u is the
Speed range (rpm) 72500 unit vector along velocity direction, A is the spacecraft’s
Momentum capacity (N m s) >2 area perpendicular to , u;, s cp is the vector from the center
Reaction torque (N m) >40 103 of mass to the center of pressure (Wertz & Larson, 1999,
Mass (kg) pp. 366–367). According to the values of microsatellite
Reaction wheel o2.55
Motor driver o0.91 parameters listed in Table 1, the velocity is V ¼
Dimension (cm) 7613 m=s and the maximum atmospheric density is r ¼
Wheel diameter o20.5 3 1011 kg=m3 (Wertz, 1978, p. 820); the area of
Wheel height o6.4 microsatellite cross-section is A ¼ 0:5 0:6 ¼ 0:3 m2 ;
Motor driver o15 19 32 the offset from the center of mass to the center of
Moment of inertia (kg m2) 5 103
pressure and the drag coefficient are taken to be scp ¼
956 C.-D. Yang, Y.-P. Sun / Control Engineering Practice 10 (2002) 951–970

0:04 m and Cd ¼ 1:5; respectively. Thus, the worst-case following matrix form:
aerodynamic disturbance becomes
M q. þ D’q þ Kq ¼ G d d þ G u u; ð20Þ
11 2
Ta ¼ ð3 10 Þð7613Þ ð1:5Þð0:3Þð0:04Þ=2
where
5
¼ 1:56 10 N m: ð16Þ 2 3
I1 0 0
Notably, the atmospheric density r is highly depen- 6 7
M ¼40 I2 0 5;
dent on time of day and the level of solar activity. At
0 0 I3
500 km altitude, the atmospheric density variations 2 3
between daytime solar maximum and nighttime solar 0 0 nðI1 þ I3  I2 Þ
minimum are approximately 2 orders of magnitude 6 7
D¼4 0 0 0 5;
(Chobotov, 1991, p. 90). Besides, as a result of the nðI1 þ I3  I2 Þ 0 0
complex electromagnetic interaction between solar wind
and the geomagnetic field that deforms the simple dipole 2 3
4n2 ðI2  I3 Þ 0 0
model of magnetic field, the precise model of Earth 6 7
magnetic field is also hard to find (Campbell & K ¼4 0 3n2 ðI1  I3 Þ 0 5;
2
McCandless, 1996, p. 94). Therefore the estimated 0 0 n ðI2  I1 Þ
worst-case aerodynamic torque and magnetic torque in 2 3
1 0 0
Eqs. (15)–(16) would be useful in the analysis and design 6 7
G d ¼ G u ¼ 4 0 1 0 5;
of attitude controller.
0 0 1
 T
2.3. Linear model of attitude dynamics and q ¼ f y c is the state vector of Euler angles;
M; D; and K are the mass matrix, damping matrix, and
Substituting the disturbance torques Eqs. (12)–(16) stiffness matrix, respectively; G d is the disturbance input
into Euler’s Eqs. (1)–(3) and using the linearized relation distribution matrix; G u is the control input distribution
(9)–(11) gives an important linear model for micro- T
matrix; d ¼ Td1 Td2 Td3 is the disturbance input
satellite attitude dynamics as follows: in Section 2.2, and u ¼
vector as obtained
T
I1 f ’ þ 4n2 ðI2  I3 Þf ¼ u1 þ Td1 ;
.  nðI1  I2 þ I3 Þc ð17Þ u1 u2 u3 is the control input vector.
The principal moment of inertia can be expressed as a
nominal value Ii plus a perturbation:
I2 y. þ 3n2 ðI1  I3 Þy ¼ u2 þ Td2 ; ð18Þ
Ii ¼ Ii þ DIi di ; jdi jp1; i ¼ 1; 2; 3;
I3 c ’ þ n2 ðI2  I1 Þc ¼ u3 þ Td3 ;
. þ nðI1  I2 þ I3 Þf ð19Þ where DIi ’s are the variation envelopes of Ii ’s and di ’s
are the normalized uncertainties. The perturbations in
where ui is the control torque and Tdi ¼ Tm þ Ta
matrices M; D; and K caused by the variations of the
represents the sum of worst-case aerodynamic and
principal moment-of-inertia DIi can be written as
magnetic torques. Since the Earth observation micro-
satellite rarely performs a large angle maneuver, the M ¼ M 0 þ XM DM PM ; ð21Þ
linear model is reasonably accurate and acceptable to be
used in attitude controller design. It is also noticed that D ¼ D0 þ XD DD PD ; ð22Þ
the pitch dynamics decouples from the roll/yaw dy-
namics in linear model on the small angles assumption, K ¼ K 0 þ XK DK PK ; ð23Þ
so the control laws of pitch channel and roll/yaw
channel can be designed separately. where M 0 ; D0 ; and K 0 represent the nominal matrices
and DM ; DD ; and DK are the diagonal matrices with the
normalized uncertainties di ’s in the diagonal entries. The
3. Modeling of inertia matrix uncertainty detailed results are shown as follows:
2 3
I1 0 0
Moment-of-inertia variation is the major source of 6 7
uncertainty for microsatellite attitude control design. In M 0 ¼ 4 0 I2 0 5 ;
order to perform controller synthesis, at first a standard 0 0 I3
synthesis model containing perturbation blocks and
2 3
nominal plant is constructed by pulling out the inertia 0 0 nðI1 þ I3  I2 Þ
matrix uncertainty from the attitude dynamics model. 6 0 0 0 7
D0 ¼ 4 5;
The following procedure is motivated by the works of
Wie, Liu, & Bauer (1993). Rewrite Eqs. (17)–(19) as the nðI1 þ I3  I2 Þ 0 0
C.-D. Yang, Y.-P. Sun / Control Engineering Practice 10 (2002) 951–970 957

2 3 where G d*; the perturbation input distribution matrix, is


4n2 ðI2  I3 Þ 0 0
6 7 defined as
K0 ¼ 4 0 3n2 ðI1  I3 Þ 0 5;  
0 0 n2 ðI2  I1 Þ G d* ¼ XM XD XK : ð28Þ

and It is noted that the moment-of-inertia uncertainties


2 3 are lumped together in the fictitious disturbance input d*
DI1 0 0 and there are no moment-of-inertia uncertainties di
6 7
XM ¼ 4 0 DI2 0 5; appears in the coefficient matrices M 0 ; D0 ; and K 0 : The
0 0 DI3 interconnection of the nominal system and the un-
2 3 certainties is shown in Fig. 2 wherein D is the normalized
d1 0 0
6 7 gain matrix of moment-of-inertia uncertainties, z* is the
DM ¼ 4 0 d2 0 5; output from the nominal system to the uncertainty
0 0 d3 block D; and d* is the input from the uncertainty block D
2 3 to the nominal system.
1 0 0
6 7 The robust stability requirement with respect to the
PM ¼ 40 1 0 5; uncertainty D is represented by z;* i.e.,
0 0 1
*
zN ¼ z: ð29Þ
2 3
0 DI1 0 DI3 0 DI2 The effect of moment-of-inertia uncertainties appears
6 7 in the perturbation d: * Eq. (26) indicates the relation of
XD ¼ n4 0 0 0 0 0 0 5;
the HN constraint (measure of stability robustness) zN ;
DI1 0 DI3 0 DI2 0 * and uncertainty block D: Taking N-
2 3T perturbation d;
1 0 1 0 1 0 norm on both sides of Eq. (26), it leads to
6
PD ¼ 4 0 0 0 0 0
7
0 5 ; zN =d* N ¼ 1=jjDjjN : According to small gain theorem
(Zames, 1966a, b; Desoer & Vidyasagar, 1975; Zhou
0 1 0 1 0 1
et al., 1996, pp. 217–221), if there is a controller
DD ¼ diagðd1 ; d1 ; d3 ; d3 ; d2 ; d2 Þ; satisfying zN =d* N og; then the closed-loop system
will be robust stable for any uncertainty j Dj N o1=g:
2 3
DI2 0 0 0 4DI3 0 Specifically, if a g smaller is obtained, the closed-loop
26 7 system can remain robust stable for a larger uncertainty,
XK ¼ n 4 0 0 3DI1 3DI3 0 0 5;
i.e., a larger range of moment-of-inertia variation in this
0 DI2 0 0 0 DI1 paper.
2 3T
4 0 0 0 1 0 On the other hand, in a viewpoint of H2 control, the
6 7 exogenous disturbance can usually be modeled as an
PK ¼ 4 0 0 1 1 0 05 ;
impulse input with random direction, and the mini-
0 1 0 0 0 1 mization of rms values of impulse response (calculated
by H2 norm) leads to optimal H2 controller. In addition
DK ¼ diagðd2 ; d2 ; d1 ; d3 ; d3 ; d1 Þ:
Define
2 3
DM 0 0 ∆
6 7
D 4 0 DD 0 5; ð24Þ ~
d
~z
0 0 DK
2 3 2 3 z∞ = ~
z
z*M PM q. Nominal System
d z2
6 7 6 7
z* 4 z*D 5 ¼ 4 PD q’ 5; ð25Þ P0
z*K PK q
u x
d* *
Dz: ð26Þ
Substituting Eqs. (21)–(23) into Eq. (20), and rewriting K
the results in terms of the notations defined in Eqs. (24)–
(26), yields
Synthesis Model
M 0 q. þ D0 q’ þ K 0 q ¼ G d*d* þ G d d þ G u u; ð27Þ Fig. 2. General configuration of mixed H2 =HN control.
958 C.-D. Yang, Y.-P. Sun / Control Engineering Practice 10 (2002) 951–970

to attitude states, it is necessary to include the control larger size of uncertainty matrix introduces conserva-
effort in the cost function to prevent saturation of tism or not is a question for argument. To answer this
actuators. Accordingly, the H2 cost is defined as question, some practical computation and discussion on
2 3 2 3 two synthesis models with different size of uncertainty
cq 0 " # 0
6 7 q 6 7 matrix D are given in Appendix A.
z2 ¼ 4 0 cq’ 5 þ 4 0 5u; ð30Þ
q’
0 0 cu
where cq ; cq’ ; and cu are weighting coefficients. In terms
of the augmented state vector x and the generalized 4. LMI formulation of mixed H2 =HN design
disturbance w defined as
" # " # After substitution of state-feedback controller u ¼
q d* Kx into Eq. (32), the closed-loop system becomes
x¼ ; w¼ ; ð31Þ
q’ d
x’ ¼ðA  B2 KÞx þ B 1 w;
Eqs. (27), (29) and (30) can be rearranged to the zN ¼ðC 1  D12 KÞx þ D11 w;
following state-space representation z2 ¼ðC 2  D22 KÞx:
x’ ¼ Ax þ B1 w þ B 2 u;
Let TzN w and Tz2 w be the closed-loop transfer matrices
zN ¼ C 1 x þ D11 w þ D12 u; from the generalized disturbance w to the performance
z2 ¼ C 2 x þ D22 u; ð32Þ outputs zN and z2 ; respectively:
where " # " #
" # A  B2 K B1 Acl B1
0 I TzN w ðsÞ ¼ ¼ ; ð36Þ
A¼ ; C 1  D12 K D11 C clN D11
M 1
0 K0 M 10 D0
" # " # " #
0 0 A  B2 K B1 Acl B1
B1 ¼ ; Tz2 w ðsÞ ¼ ¼ : ð37Þ
M 1
0 G d* M 1
0 Gd C 2  D22 K 0 C cl2 0
" #
0 The goal of mixed H2 =HN control is to find an
B2 ¼ ; ð33Þ
M 1
0 Gu
internally stabilizing controller K that minimizes the H2
2 3 performance, Tz2 w 2 ; subject to the HN constraint,
PM M 1
0 K0 PM M 1
0 D0 TzN w N og: It is noted that Tz2 w 2 can be computed
6 7 as
C1 ¼ 4 0 PD 5
2
PK 0 Tz2 w 2 ¼ TrðC cl2 S 0 C Tcl2 Þ;
2 3
PM M 1
0 G d* PM M 1
0 Gd
6 7 where S 0 > 0 is the solution of Lyapunov equation
D11 ¼4 0 0 5; Acl S 0 þ S 0 ATcl þ B1 BT1 ¼ 0: For any S satisfying Acl S þ
0 0 SATcl þ B 1 B T1 o0; it has S > S 0 and
2 3
PM M 1
0 Gu
2
Tz2 w 2 oTrðC cl2 SC Tcl2 Þ:
6 7
D12 ¼4 0 5; ð34Þ
Hence the useful H2 performance constraint can be
0 expressed by linear matrix inequalities as follows.
2 3 2 3
cq 0 0
6 7 6 7 4.1. LMI of H2 performance
C2 ¼ 4 0 cq’ 5; D22 ¼ 4 0 5: ð35Þ
0 0 cu 2
Acl is stable and Tz2 w 2 ou if and only if there exist
To this stage, the synthesis model of Eqs. (32)–(35) two symmetric matrices X 2 and Q satisfying the
consisting of nominal plant, moment-of-inertia uncer- following linear matrix inequalities (Scherer et al., 1997):
" #
tainties, environmental disturbances, and H2 and HN AX 2  B2 Y 2 þ ðAX 2  B 2 Y 2 ÞT B 1
penalties has been obtained for controller synthesis in o0; ð38Þ
B T1 I
the next section.
" #
Q C 2 X 2  D22 Y 2
Remark 1. The Eqs. (20)–(26) used to represent the > 0; ð39Þ
uncertainties of the moment of inertia are not unique. A ðC 2 X 2  D22 Y 2 ÞT X2
smaller matrix D of size 9 9 instead of 15 15 (in this
paper) would be obtained alternatively. Whether the TraceðQÞou; ð40Þ
C.-D. Yang, Y.-P. Sun / Control Engineering Practice 10 (2002) 951–970 959

where Y 2 is an auxiliary matrix, Y 2 K 2 X 2 : The there exist Lyapunov matrices X 2 ; X N ; and additional
resulting pure H2 controller K 2 is auxiliary variables Y 2 ; Y N ; and Q that satisfies the
K 2 ¼ Y 2 X 1 ð41Þ LMIs (38)–(40) and (44). For tractability, only one
2 :
single common Lyapunov matrix X > 0; i.e.,
By Parseval’s theorem, the time-domain definition of
H2 norm of Tz2 w ðsÞ is given by X2 ¼ XN ¼ X ð46Þ
Z N 1=2
T
is sought to fulfill the LMI constraints. Clearly, it brings
Tz2 w ðsÞ 2 ¼ Trace½g ðtÞgðtÞ dt ; ð42Þ about some conservatism. However, such conservatism
0
is paid by the following valuable merits. First, it makes
where gðtÞ is the impulse response of Tz2 w ðjoÞ: The the larger set of LMIs become numerically tractable.
corresponding physical interpretation of Tz2 w 2 is the Second, it exploits all degrees of freedom in choosing
rms value of the output z2 driven by an impulse or Lyapunov matrix X: Specifically, the common Lyapu-
independent, zero mean, unit intensity white signals. nov matrix X is shaped by LMI optimization until either
Thus, it is necessary to design a controller to minimize all specifications are met or all degrees of freedom for
the H2 norm of the output penalty z2 ; if the disturbance design are exhausted. Scherer et al. (1997) named such a
w has stochastic nature. design procedure the Lyapunov shaping paradigm. The
On the other hand, the HN norm of TzN w ðsÞ is defined LMIs of mixed H2 =HN control under a common
as Lyapunov matrix X and disturbance input w are
T zN w N
¼ sup smax ½TzN w ðjoÞ ; ð43Þ expressed as follows.
o

where smax denotes the maximum singular value. 4.3. LMI of mixed H2 =HN controller
TzN w N represents the peak value of steady-state
output zN if the system is driven by a unit-magnitude Minimize TraceðQÞ (an upper bound for Tz2 w 2 ) over
sinusoidal over all frequencies. The HN performance matrices X ¼ X T ; Q ¼ QT ; and Y that satisfy the
constraint can also be expressed by linear matrix following linear matrix inequalities (Chilali & Gahinet,
inequalities as follows. 1996):
2 3
4.2. LMI of HN performance AX  B 2 Y þ ðAX  B 2 YÞT B1 ðC 1 X  D12 YÞT
6 7
4 BT1 I DT11 5o0;
Acl is stable and TzN w N og if and only if there exists C 1 X  D12 Y D11 g2 I
a symmetric matrix X N > 0 satisfying the following
LMI (Sanchez-Pena & Sznaier, 1998): ð47Þ

2 3
AX N  B2 Y N þ ðAX N  B2 Y N ÞT B1 ðC 1 X N  D12 Y N ÞT
6 7
4 B T1 I DT11 5o0; ð44Þ
2
C 1 X N  D12 Y N D11 g I

where Y N is an auxiliary matrix, Y N K N X N : The " #


Q C 2 X  D22 Y
resulting pure HN controller K N is > 0: ð48Þ
ðC 2 X  D22 YÞT X
KN ¼ Y N X 1
N: ð45Þ
Assuming that the solutions of LMIs (47) and (48) are
From the small gain theorem, the constraint
X  ; Y  ; Q ; the corresponding mixed H2 =HN controller
TzN w N og can be interpreted as a disturbance
K  is expressed by
rejection performance, and what is more, this constraint
is also useful to enforce robust stability with respect to K  ¼ Y  ðX  Þ1 ; ð49Þ
uncertainty. It guarantees that the closed-loop system
which leads to
remains stable for all perturbations d* Dz; * with qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
j DjjN og1 : TzN w pg; T p TraceðQ Þ: ð50Þ
N z2 w 2
The LMI formulations of closed-loop system specifi-
cations under H2 and HN performance consideration The Lyapunov shaping paradigm for multi-objective
have been summarized as above. A key observation is design provides a greater flexibility than single-objective
that all LMI constraints involve the solutions of optimal design techniques such as optimal HN techni-
Lyapunov matrix Xj ; j ¼ 2; N: For mixed H2 =HN que or optimal H2 technique. In mixed H2 =HN
controller synthesis, the specifications on transfer synthesis, the goal is to minimize the H2 norm on
function matrices TzN w ðsÞ and Tz2 w ð joÞ are satisfied, if one transfer matrix subject to some moderate HN
960 C.-D. Yang, Y.-P. Sun / Control Engineering Practice 10 (2002) 951–970

performance requirement on another transfer matrix. (v) Create the Pareto diagram (Maciejowski, 1989, pp.
While there exist a lot of Lyapunov matrix candidates P 346–350) to indicate the trade between H2 perfor-
that meet the specified HN performance, classical mance and HN performance.
optimal HN synthesis may find a solution with the best (vi) Evaluate the time-domain simulations of the mixed
HN performance but with poor H2 performance H2 =HN controllers obtained in Step (iv). If there is
probably. Even if the bound of HN performance is set no controller in the Pareto optimal set satisfying
looser that means a lot more HN performance is both the H2 performance requirement (rms perfor-
sacrificed, the resulting controller from optimal HN mance of impulse response) and HN performance
technique might still have unsatisfied H2 performance. requirement (stability robustness with respect to
This is because the optimal HN technique does not uncertainty), then go back to Step (ii) and adjust
have a mechanism to utilize the ‘‘freedom’’ from the weights.
sacrificing the HN performance to improve the H2
performance. The same result is also occurred in
applying optimal H2 technique to obtain a controller Remark 2. At present norm minimization is necessary
with poor HN performance. In contrast, Lyapunov for most robust controller designs that all specifications
shaping paradigm exhausts these degrees of design have to be translated into frequency domain by using
freedom to effectively optimize the H2 performance. appropriate weighting functions. The incompatibilities
The resulting controller synthesized by Lyapunov between time-domain and frequency-domain specifica-
shaping paradigm, while conservative to some extent, tions still remain. In fact, there is no universal design
does improve controller performance over the classical methodology that can produce the desired controller at
single-objective optimal synthesis techniques, and one stroke to satisfy the time domain specifications
furthermore the Lyapunov shaping paradigm provides including bandwidth, over-shoot, settling time, etc.,
a useful tool to fine-tune the performance between except the very simple single-input–single-output case
competitive specifications. (Rotstein & Sideris, 1994; Franchek & Herman, 1998).
The above set of linear matrix inequalities can be Therefore, a series of controllers in the Pareto optimal
solved efficiently by convex optimization. As for a set has to be computed. In Step (vi) a lot of time-domain
detailed description of algorithms and software to solve simulations of these controllers are performed to test
LMI problems, refer to the excellent books (Boyd et al., their performance of disturbance attenuation and
1994; El Ghaoui & Niculescu, 2000) and MATLAB robustness against moment-of-inertia variation. This is
LMI Toolbox (Gahinet, Nemirovski, Laub, & Chilali, a standard procedure for spacecraft attitude control
1995). system design (O’Donnell & Mangus, 1998). The Pareto
diagram indicates the trade-off of controller perfor-
mance between different criteria and gives insights into
controller selection.
5. Attitude control of LEO microsatellite
Thus far, in Sections 2 and 3, Steps (i) and (ii) have
5.1. Mixed H2 =HN controller design
been accomplished; Section 4 gives the solutions of Steps
(iii) and (iv). The HN penalties are derived in Eqs. (25)
The process of mixed H2 =HN attitude controller
and (29):
design for microsatellite includes the following steps:
zN pitch ¼ ½ y. y y T ; 1 3
(i) Derive linear model for microsatellite attitude
dynamics. zN roll=yaw
(ii) Formulate the synthesis model containing the . . f’ ’ f’ ’ ’
¼ ½f c c c f c’ 4f c f c T
1 12 :
nominal plant and the moment-of-inertia uncer-
tainty; define appropriate HN penalty and H2 The H2 penalties from Eq. (30) are
penalty and the corresponding weights.
z2 pitch ¼ ½ cq y cq’ y’ cu u2 T ;
(iii) Solve the LMI (44) for the pure HN optimal
controller to obtain the minimum HN bound as the ’
z2 roll=yaw ¼ ½ cq f cq c cq’ f cq’ c’ c u u1 c u u3 T
:
baseline for the next step.
(iv) Select a set of HN performance constraints that are zN is used to penalize the robust stability with respect to
larger than the minimum value obtained from Step parametric uncertainty; z2 is used to penalize the rms
(iii), and solve the LMIs (47) and (48) for every performance of Euler angle, Euler rates, and control
single HN performance constraint. Thus, a family effort. Since the designed microsatellite is planned to
of mixed H2 =HN controllers is obtained and each image Earth, minimizing the distortion of images is the
of them corresponds to a specific level of HN most important consideration, which implies that the
performance. Euler rates need to be severely limited. To reflect this
C.-D. Yang, Y.-P. Sun / Control Engineering Practice 10 (2002) 951–970 961

requirement, the weighting coefficients penalizing the uncertainty). The resulting Pareto diagram is shown in
rms performance of Euler angle, Euler rate, and control Fig. 3 that describes the trade-off between H2 perfor-
effort in Eq. (30) are selected as cq ¼ 1; cq’ ¼ 100; and mance and HN performance in pitch channel and roll/
cu ¼ 10; where the heavier weight is imposed on the yaw channel. As can be seen in Fig. 3, the pure HN
Euler rates to avoid rapid attitude variations. The controller at one extreme produces the best stability
uncertainty of the moment of inertia considered to be robustness to uncertainty but the worst rms perfor-
710% of nominal value is reflected in Eq. (28). mance of impulse response; in contrast, the pure H2
After solving the LMI (44), the minimum HN bound controller at the other extreme produces the worst
in pitch channel is TzN w N p0:1142; and in roll/yaw stability robustness to uncertainty but the best rms
channel is TzN w N p0:1778: The mixed H2 =HN con- performance of impulse response. The mixed H2 =HN
troller corresponding to each HN constraint controller in the middle of these two extremes performs
TzN w N pg; gAf0:2; 0:5; 1; 2; 3; 4; 6; 8; 10g; is obtained more like the pure H2 controller at a larger g; while more
by solving the LMIs (47) and (48). By this way, a family like the pure HN controller at a smaller g: It is up to the
of mixed H2 =HN controllers, constituting the Pareto designer to choose an appropriate g between stability
optimal set, is synthesized for which H2 performance robustness and rms performance of impulse response.
(rms performance of impulse response) is traded with The detailed gain matrices of each controller are
HN performance (stability robustness with respect to expressed in Tables 2 and 3.

pure H∞ controller
38
28 Larger maximum control effort for impulse disturbance
Faster response to nonzero initial condition 36
More Robust to inertia matrix uncertainty

H2 performance (Roll/Yaw channel)


Stronger capability to attenuate sinusoidal disturbance
34
H2 performance (Pitch channel)

26

32
24 Pitch channel
Roll/Yaw channel
30

22 28

26
20 Smaller maximum control effort for impulse disturbance
Slower response to nonzero initial condition
More sensitive to inertia matrix uncertainty 24
Weaker capability to attenuate sinusoidal disturbance
18
22
pure H2 controller
16 20

18
14
0 2 4 6 8 10
H∞ performance

Fig. 3. Trade-off plot of mixed H2 =HN controllers.

Table 2
Control gains for the model with uncertainty matrix D of size 15 15

Gain K2 K2=N (g ¼ 2) KN

Roll K1f 9.403 102 4.728 100 1.729 102


K1c 7.097 105 1.214 103v 0
K1f’ 1.016 101 1.664 101 1.211 102
K1c’ 5.011 105 8.579 104 7.760 103

Pitch K2y 9.405 102 1.650 100 5.158 101


K2y’ 1.018 101 1.267 101 5.735 101

Yaw K3f 6.569 105 2.132 103 0


K3c 8.553 102 9.871 101 2.425 101
K3f’ 1.090 105 2.211 103 7.760 103
K3c’ 1.004 101 1.068 101 3.220 101
962 C.-D. Yang, Y.-P. Sun / Control Engineering Practice 10 (2002) 951–970

Table 3 also observed that the difference between the responses


The closed-loop poles of different type of controllers (uncertainty of the nominal plant and the perturbed plant increases
matrix D of size 15 15)
by increasing g; indicating that the mixed H2 =HN
K2 K2=N ðg ¼ 2Þ KN controller with a stringent HN constraint (a smaller g)
Pitch 0.009405 0.1735 1.576+0.5933j provides a better performance robustness against inertia
0.5497 0.5225 1.5760.5933j matrix uncertainty. Also it is noticed that the pure H2
controller recovers the microsatellite to its original
Roll/yaw 0.008565 0.09856 0.9398 attitude from the initial perturbation much slowly than
0.009421 0.4523+0.2289j 2.095
its competitors—the pure HN controller and the mixed
0.5427 0.45230.2289j 3.795
1.469 1.473 4.486 H2 =HN controller—as indicated in the lower part of
Fig. 5.

Remark 3. In Figs. 4 and 5 the moment-of-inertia


5.2. Evaluation of controllers variations of perturbed plant 1 are DI1 ¼ 10%  I%1 ;
DI2 ¼ þ10%  I%2 ; DI3 ¼ 10%  I%3 ; and that of the
5.2.1. Robustness to inertia matrix uncertainty perturbed plant 2 are DI1 ¼ þ10%  I%1 ; DI2 ¼ 10% 
The moment-of-inertia variation of the microsatellite I%2 ; and DI3 ¼ þ10%  I%3 : The above combinations of
owing to the movement of instrument or payload is an moment-of-inertia variations result in the largest
important problem in attitude controller design. Atti- change in coefficients of the roll–yaw channel of linear
tude controller must be robust enough against the model.
moment-of-inertia uncertainty. Here, the uncertainties
in principal moment of inertia, DIi ; i ¼ 1; 2; 3; are Remark 4. For a normal configuration and payload
considered to be 10% perturbation with respect to their position, the moment of inertia of a satellite is
nominal values. considered to be a nominal value. The inertia matrix
In testing performance robustness of different con- uncertainty results from the inevitable difference be-
trollers with respect to inertia matrix uncertainty, the tween the nominal model and the reality during the

initial tip-off rates fð0Þ ’
¼ cð0Þ ¼ 0:01 rad/s, when the mission. For example, the solar panel rotation, the
microsatellite separates from launch vehicle, are payload movement, and the propellant consumption all
adopted as a nonzero initial perturbation (Zimbelman cause the variation of moment-of-inertia. The robust
& Watzin, 1999). From the Euler angle and Euler stability with respect to the inertia uncertainty is very
angular rate histories shown in Figs. 4 and 5, the important for attitude controller design. The effect of
performances of the nominal plant and the perturbed moment-of-inertia uncertainty on the nominal model
plants are almost the same for the pure HN controller, results in a perturbation d* (in Eq. (26)), in which
which indicates that the pure HN controller has the best the normalized inertia uncertainty D and the HN
robustness against moment-of-inertia uncertainty. It is penalties z* are introduced to provide the measurement
of stability robustness with respect to inertia uncer-
tainty. According to small gain theorem, if a controller
1.1 can satisfy the smaller constraint g of zN =d* N ; i.e.,
1 zN =d* N og; the resulting closed-loop system will be
robust stable for the larger range of uncertainty, i.e.,
0.9
j DjjN o1=g: The design parameter g in mixed H2 =HN
0.8 control design indicates the trade-off between H2
K2 controller, γ → ∞ performance and HN performance as shown in Fig. 3.
0.7
From Figs. 4 and 5 it is clear that the mixed H2 =HN
nominal plant
φ (deg)

0.6 controller with a smaller g improves the robustness


perturbed plant 1
0.5 perturbed plant 2
against inertia uncertainty.
0.4
K2/∞ controller, γ = 2 5.2.2. Disturbance attenuation
0.3 To access disturbance attenuation of different con-
trollers, two types of disturbance models—sinusoid and
0.2 K∞ controller,
γ → γmin impulse—are commonly used. The sinusoid is the
0.1 prototype of some periodic disturbances. According to
0 the analysis of Section 2, the environmental disturbances
0 5 10 15 20
such as magnetic torque, and gravity-gradient torque
time (sec)
are cyclic essentially and they can be represented by
Fig. 4. Comparison of pitch angle history of tip-off rate cease. sinusoids. The impulse is an idealized model of a sudden
C.-D. Yang, Y.-P. Sun / Control Engineering Practice 10 (2002) 951–970 963

0.6

0.5 nominal plant


perturbed plant 1
perturbed plant 2
0.4

dφ/dt (deg/sec) 0.3


K2 controller, γ → ∞
0.2

0.1

-0.1 K2/∞ controller, γ = 2


K∞ controller, γ → γmin
-0.2
0 5 10 15 20
Time (sec)

Zoom of the above Plot


0.01
nominal plant
K∞ controller, γ → γmin perturbed plant 1
0.005
perturbed plant 2
dφ/dt (deg/sec)

K2/∞ controller, γ = 2
-0.005

-0.01
K2 controller, γ → ∞
-0.015

-0.02
20 40 60 80
Time (sec)
Fig. 5. Comparison of pitch rate history of tip-off rate cease.

disturbance. It is suitable to represent the load undesirable saturation effect, designers can manipulate
disturbance with short duration. the HN constraint g as a tuning parameter in mixed
In Fig. 6, considering the maximum variation of Euler H2 =HN control to effectively reduce the maximum
angle with respect to sinusoid disturbances with 105 control effort.
amplitude at different frequencies o ¼ n; 5n; 10n; 20n;
30n; 40n; 50n; it shows that the pure HN controller has Remark 5. From Eqs. (25) and (29), zN includes f; . y;
.
the best attenuation capability, the mixed H2 =HN con- . Since the expressions of f;
and c: . y;
. and c . in Eqs. (17)–
troller ranks second, and the pure H2 controller gets the (19) contain u1 ; u2 ; u3 ; control efforts are in fact involved
last place. As a result, it is evident that a controller with in HN constraint zN implicitly. On the other hand, from
the stricter HN constraint has the better attenuation Eq. (30), z2 contains explicit weighting on control effort.
capability for sinusoid disturbance. In comparison with zN ; z2 is dominant in maximum
In Fig. 7, the control effort expense for impulse control effort limitation. Designers can effectively
disturbance shows two important results. First, the prevent very large control effort by imposing a heavy
maximum control effort by using the pure HN weighting on u:
controller is about 4 times larger than that by using
the pure H2 controller, since the control signal is not 5.3. Integrated test
explicitly included in the HN penalty. Second, a
controller with the stricter HN constraint always To perform three-axis control, this microsatellite uses
spends the larger maximum control effort that would three reaction wheels mounted along the principal axes.
probably cause saturation of actuator. To avoid the A reaction wheel is used to regulate the angular rate of
964 C.-D. Yang, Y.-P. Sun / Control Engineering Practice 10 (2002) 951–970

K2 K2/∞ K∞

0.00125 0.0015

0.0001
K2 , γ → ∞

0.006
K2/∞ , γ = 8
K2/∞ , γ = 2
K∞ , γ → γmin

0.005

8E-05
0.001
0.004

6E-05
max. θ (deg)

0.00025 0.0005 0.00075


0.003

4E-05
0.002

2E-05
0.001
0

0
n 5n 10n 20n 30n 40n 50n
frequency (orbital rate n = 0.0011 rad/sec)
Fig. 6. Comparison of maximum deviation of pitch angle for sinusoidal disturbance input.

dynamics of reaction wheel have to be included to offer


0
a more practical evaluation.
The ITHACO’s Type A reaction wheel is a prototype
for design reference. The detailed specifications can be
-10 found in Bialke (1998) and http://www.ithaco.com.
Some important parameters of model number TW-
2A40 reaction wheel are: speed range 72500 rpm,
cu × u3 (N-m)

K2 , γ → ∞
-20 K∞ , γ → γmin momentum capacity >2 N m s, reaction torque
K2/∞ , γ = 0.5 >40 mN m, mass (including reaction wheel and motor
K2/∞ , γ = 2
K2/∞ , γ = 8
driver) o3.5 kg, wheel diameter o20.5 cm, wheel height
-30 o6.4 cm, and motor driver size o15 cm 19 cm
32 cm. The size of microsatellite is 50 cm 50 cm
60 cm that is enough to accommodate suitable reaction
-40
wheels. For microsatellite attitude control it is common
to use reaction wheels as actuators. For example, the
Italian Scientific Microsatellite for Advanced Research
-50
0 1 2 3 4 and Technology (SMART) and Germany microsatellite
time (sec) mission BIRD (Bi-spectral Infrared Detection) are
Fig. 7. Comparison of control torque history for impulse disturbance designed by three-axis attitude control with three and
input. four reaction wheels (Pastena & Grassi, 1998; Brieb
et al., 2000). As a matter of fact, the advances in
microsystem technology speeds the development of
spacecraft by making use of the conservation of angular microsatellite and nanosatellite (o10 kg) (Muncheberg
momentum. When a disturbance torque is applied to a et al., 1996).
spacecraft and results in angular rates change, the excess Suppose three reaction wheels are mounted along the
angular momentum of spacecraft can be transferred to three principal axes. The angular momentum of the
the angular momentum of wheels by internal torques reaction wheels can be expressed in terms of the body-
generated from a DC motor to change the rotation fixed reference frame as follows:
speeds of wheels so that the original attitude of , , , ,
spacecraft is recovered. In system-integrated test, the h W ¼ J1 O1 b 1 þ J2 O2 b 2 þ J3 O3 b 3 ;
C.-D. Yang, Y.-P. Sun / Control Engineering Practice 10 (2002) 951–970 965

where J1 ; J2 and J3 are the moments of inertia of the sinusoids with angular rates o ¼ kn; k ¼ 1; 5; 10; 20; 30;
reaction wheels, and O1 ; O2 and O3 are the correspond- 40; 50 and magnitude of 105 N m are summed up, to
ing rotation speeds. Then the time derivative of angular approximate the effect of magnetic disturbance torque.
momentum of the reaction wheels measured in the
inertial reference frame becomes On the other hand, as a result of the operation of
. (
, , , some payloads, the induced load disturbances are
hW ¼ , o h W þ hW ; ð51Þ intense but of exceedingly short duration, which is
where , o denotes the angular velocity of the micro- properly considered to be the impulse type of dis-
(
,
turbance with strength 0.001 as follows:
satellite described by Eqs. (9)–(11); h W represents the 0:001
,
time derivative of h W measured in the body-fixed frame, Tdi;i¼1;2;3 ¼ ½ðsðtÞ  sðt  eÞ ; e ¼ 103 : ð57Þ
e
(
,
which can be expressed by hW ¼ J1 O ’ 1, ’ 2,
b 1 þ J2 O b2þ The block diagram of integrated system is described
’ 3,
J3 O b 3 . The following first-order model describes a DC in Fig. 8. K denotes the designed controller, which is
motor providing the driving torque: implemented by pure H2 controller, pure HN controller
’ i ¼ uc  tv Oi ; i ¼ 1; 2; 3; and mixed H2 =HN controller in turn to perform the
Ji O i ð52Þ
integrated test. From the preliminary simulation results,
where uci is the torque command from the designed it is unexpected that there are slow oscillations of
controller within the limit of 0.04 N m and reaction wheel speeds in roll and yaw channels. These
tv ¼ 0:0384 103 N m s/rad is the viscous friction oscillations are produced by the following two terms in
constant (Bialke, 1998). Considering Eqs. (51)–(52) with Eqs. (53) and (55): nJ3 O3 in roll channel and nJ1 O1 in
, ,
Eqs. (17)–(19), the attitude dynamics including reaction yaw channel, which are the dominant effects of o hW:
wheels become In order to eliminate these slow oscillations, an
additional control term u# is required:
’ þ 4n2 ðI2  I3 Þf þ J3 O3 ðy’  nÞ
I1 f.  nðI1  I2 þ I3 Þc 2 32 3
’ þ nfÞ ¼ uc þ tv O1 þ Td1 ; 0 0 nJ3 O1
 J2 O2 ðc ð53Þ 6 76 7
1
u# ¼ 4 0 0 0 54 O2 5 ¼ KX: # ð58Þ
’ þ nfÞ  J3 O3 ðf
I2 y. þ 3n2 ðI1  I3 Þy þ J1 O1 ðc ’  ncÞ
nJ1 0 0 O3
¼ uc2 þ tv O2 þ Td2 ; ð54Þ
For the results of the worst-case disturbance test,
’ þ n2 ðI2  I1 Þc þ J2 O2 ðf
I3 c. þ nðI1  I2 þ I3 Þf ’  ncÞ Fig. 9 shows that the pure HN controller suppresses the
variation of pitch rate to a lowest extent, and Fig. 10
 J1 O1 ðy’  nÞ ¼ u þ tv O3 þ Td3 :
c
3 ð55Þ indicates that there is no remarkable difference in
As discussed in Section 2, a spacecraft in space is control effort expense among the three types of
subject to small but persistent disturbances from a controllers. Thus, in view of the attenuation perfor-
variety of sources, such as magnetic torque and mance for cyclic and secular disturbances, the pure HN
aerodynamic torque. The magnetic torque is cyclic that controller is indeed the best, the mixed H2 =HN
appears in a summation of sinusoids with different controller is fairly good, and the pure H2 controller is
frequencies; the aerodynamic torque is secular that poor. On the other hand, for the results of the impulse
accumulates with time and does not cancel out over an disturbance test, the pure HN controller causes serious
orbit (Wertz & Larson, 1999, p. 354). From the worst- saturation of actuators as shown in Fig. 11. The
case disturbance torque in Eqs. (15) and (16), let the resulting maximum transient speed of wheel by using
disturbance torque in each channel be the following the pure H2 controller is only 720 rpm, and that by using
combination of sinusoids and finite step: the mixed H2 =HN controller is below 1200 rpm, whereas
X that by using the pure HN controller is up to 1300 rpm.
Tdi;i¼1;2;3 ¼ 105 sin knt þ 3 105
k
Also it is important to point out that the speeds of
reaction wheel by using different controllers do not
½sðt  1000Þ  sðt  4000Þ ;
converge to the same value finally. On a clear inspection
k ¼ 1; 5; 10; 20; 30; 40; 50; ð56Þ after 800 s elapse, the speeds of reaction wheel by using
where n ¼ 0:0011 rad/s is the orbital rate and sðt  t0 Þ is the pure H2 controller, the pure HN controller, and the
the unit step function occurring at t0 : mixed H2 =HN controller are 711, 704, and 691rpm,
respectively. As to these results, it is a little surprising
Remark 6. The magnetic disturbance torque is a but not unreasonable. The explanation is that the
periodic disturbance. According to Eq. (15), the integrated system concerning spacecraft attitude dy-
worst-case magnitude of magnetic disturbance is namics and reaction wheel dynamics is uncontrollable
estimated to be 5 105 N m; however, the frequency (Nam, Hashimoto, & Ninomiya, 1997). Exactly to
is hard to be determined. In simulations seven say, the states of reaction wheels are uncontrollable.
966 C.-D. Yang, Y.-P. Sun / Control Engineering Practice 10 (2002) 951–970

Euler angles ,,


Disturbance Td
Integrated System . . .
Eqs. (53)-(55) Euler rates ,,
_I
Reaction
torque u

J
_1 Rates of wheels Ω1 , Ω2 , Ω 3

s

_ v

Reaction wheel
block diagram


uc
∑ K

Controller
Fig. 8. Block diagram of integrated attitude control system for microsatellite.

Mixed H2/H∞ controller , γ = 2


Pure H2 controller, γ → ∞
0.0004

Pure H∞ controller, γ → γmin


0.0002
dθ/dt (deg/sec)
0
-0.0002

0 1000 2000 3000 4000 5000 6000


time (sec)
Fig. 9. Comparison of pitch rate history of worst-case disturbance test on integrated system.

Moreover, in Fig. 12 the pure HN controller produces resulting pitch rate recovers very slowly. As a whole,
much larger oscillation in pitch rate than the pure by considering the results of worst-case disturbance
H2 controller and mixed H2 =HN controller do. It is test as well as the impulse disturbance test, the
also noticed that while the pure H2 controller mixed H2 =HN controller shows the most balanced
produces the smallest oscillation of pitch rate, the performance.
C.-D. Yang, Y.-P. Sun / Control Engineering Practice 10 (2002) 951–970 967

0
Mixed H2 /H∞ controller, γ=2

-0.0002
Pure H2 controller, γ → ∞
Pure H∞ controller, γ → γmin

torque command uc1 (N-m)

-0.0006 -0.0004
-0.0008
-0.001

0 1000 2000 3000 4000 5000 6000

250

Mixed H2 /H∞ controller, γ=2


Pure H2 controller, γ → ∞
speed of reaction wheel Ω1 (rpm)

200 Pure H∞ controller, γ → γmin

150

100

50

0
0 1000 2000 3000 4000 5000 6000
time (sec)
Fig. 10. Comparison of control effort history and wheel speed history of worst-case disturbance test on integrated system.

0.05

0.04
Mixed H2 /H∞ controller, γ=2
Pure H2 controller, γ → ∞
0.03 Pure H∞ controller, γ → γmin
torque command uc1 (N-m)

0.02

0.01

-0.01

-0.02

-0.03

-0.04

-0.05
0 50 100 150 200

1400
speed of reaction wheel Ω1 (rpm)

1200 Mixed H2 /H∞ controller, γ =2


Pure H2 controller, γ → ∞
Pure H∞ controller, γ → γmin
1000

800

600

400

200

0
0 50 100 150 200
time (sec)
Fig. 11. Comparison of control effort history and wheel speed history of impulse (strength: 0.001 N m s) test on integrated system.
968 C.-D. Yang, Y.-P. Sun / Control Engineering Practice 10 (2002) 951–970

1.2 certainty. Three conclusions are summarized as follows.


Mixed H2/H∞ controller, γ = 2 First, mixed H2 =HN control provides an impressive
1
Pure H2 controller, γ → ∞
Pure H∞ controller, γ → γmin flexibility to compromise between H2 performance (rms
0.8
performance of impulse response) and HN performance
0.6 (robust stability with respect to uncertainty). Second,
0.4 mixed H2 =HN controller with stricter HN constraint
dθ/dt (deg/sec)

0.2
has better robust performance against moment-of-
inertia uncertainty, and superior disturbance attenua-
0
tion capability with respect to sinusoid disturbance at
-0.2 the expense of spending excessive maximum control
-0.4 effort with respect to impulse disturbance. Finally, the
results of integrated test including reaction wheel
-0.6
dynamics confirm that mixed H2 =HN controller exhibits
-0.8 the moderate performance in every aspect than pure H2
-1 controller and pure HN controller, which leads to a
-1.2 most practical design.
0 50 100 150 200
As for the future expansions, developing multiple
time (sec)
trade-off parameters instead of one trade-off para-
Fig. 12. Comparison of pitch rate history of impulse (strength: meter to meet different performance requirement on
0.001 N m s) test on integrated system. specific input–output channel is a challenging
direction. In this paper, the plant is confined to having
one generalized input w and two output penalties
Table 4 zN and z2 : For more general applications, constraints
Control gains for the model with uncertainty matrix D of size 9 9 are imposed on the closed-loop transfer function
Gain K2 K2=N ðg ¼ 2Þ KN resulting from any appropriate input–output channels,
2 2 without limitations. Specifically, the performance of
Roll K1f 9.495 10 9.655 10 6.649 104
K1c 7.142 105 3.426 105 4.996 102 every specific input–output channel ought to be
K1f’ 1.016 101 1.026 101 4.507 101 tuned by an individual design parameter. Such a
K1c’ 8.736 106 6.318 104 4.861 102 philosophy may lead to a multi-channel mixed
H2 =HN control design with multiple trade-off para-
Pitch K2y 9.405 102 1.650 100 5.158 101
meters that enables designers to gain more freedom and
K2y’ 1.018 101 1.267 101 5.735 101
flexibility.
Yaw K3f 6.743 105 4.507 104 7.386 102
K3c 8.781 102 8.958 102 4.186 105
K3f’ 1.013 105 2.376 104 2.239 102 Appendix A
K3c’ 1.006 101 1.008 101 1.666 101

For the synthesis model with the larger uncertainty


matrix D of size 15 15 (see Eq. (24)), the HN penalties
Table 5 in pitch channel and roll/yaw channel are:
The closed-loop poles of different type of controllers (uncertainty
 T
matrix D of size 9 9) zN pitch ¼ y. y y 1 3 ;
K2 K2=N ðg ¼ 2Þ KN
zN roll=yaw
Pitch 0.009405 0.1735 1.576+0.5933j
.
¼ ½f . f’ ’ f’ ’ ’
f c’ T
0.5497 0.5225 1.5760.5933j c c c 4f c f c 1 12 :

Roll/yaw 0.008782 0.008943 6.234 106+0.002217j For the synthesis model with the smaller uncertainty
0.009511 0.009579 6.234 1060.002217j matrix D of size 9 9, the HN penalties in pitch channel
0.5429 0.5481 2.4496 and roll/yaw channel are:
1.470 1.473 2.4497
zN pitch ¼ ½ y. y y T
1 3;

zN roll=yaw
6. Conclusion
¼ ½ f.  nc’ c’ þ 4nf c’ þ 4nf c. þ nf’ f’  nc f’  nc T
1 6:

This paper applies mixed H2 =HN control to perform Clearly the two different models have the same HN
microsatellite attitude control design in the face of penalty in pitch channel and the different HN penalties
environmental disturbances and moment-of-inertia un- in roll/yaw channel.
C.-D. Yang, Y.-P. Sun / Control Engineering Practice 10 (2002) 951–970 969

The control torques in roll, pitch, and yaw channels Brieb, K., Barwald, W., Gerlich, T., Jahn, H., Lura, F., & Studemund,
are expressed as H. (2000). The DLR small satellite mission BIRD. Acta
Astronautica, 46(2–6), 111–120.
u1 ¼ K1f  f þ K1c  c þ K1f’  f’ þ K1c’  c;
’ Burl, J. B. (1999). Linear optimal control: H2 and HN methods. Menlo
Park, CA: Addison-Wesley Longman.
Campbell, B. A., & McCandless Jr., S. W. (1996). Introduction to space

u2 ¼ K2y  y þ K2y’  y; sciences and spacecraft applications. Houston, TX: Gulf Publishing
Company.
Chen, B.-S., Wu, C.-S., & Jan, Y.-W. (2000). Adaptive fuzzy mixed H2/
u3 ¼ K3f  f þ K3c  c þ K3f’  f’ þ K3c’  c:

HN attitude control of spacecraft. IEEE Transactions on Aerospace
and Electronic Systems, 36(4), 1343–1359.
Through pure H2 design, pure HN design and mixed Chilali, M., & Gahinet, P. (1996). HN design with pole placement
H2 =HN design in Section 4, the control gains are constraints: An LMI approach. IEEE Transactions on Automatic
obtained. For the synthesis model with uncertainty Control, 41(3), 358–367.
Chobotov, V. A. (1991). Spacecraft attitude dynamics and control.
matrix D of size 15 15, the gains and closed-loop poles
Malabar, FL: Krieger Publishing Company.
are listed in Tables 2 and 3. Similarly, for the synthesis Cruise, A. M., Bowles, J. A., Patrick, T. J., Goodall, C. V. (1998).
model with uncertainty matrix D of size 9 9, the gains Principles of space instrument design (pp. 160–179). Cambridge:
and closed-loop poles are listed in Tables 4 and 5. Cambridge University Press.
In pitch channel, both models produce the same Desoer, C. A., & Vidyasagar, M. (1975). Feedback systems: input–
control gains and closed-loop poles. In roll/yaw channel, output properties. New York: Academic Press.
El Ghaoui, L., & Niculescu, S.-l. (2000). Advances in linear matrix
the differences of both models in control gains and inequality methods in control, SIAM’s advances in design and control
closed-loop dynamics are distinguished. If two synthesis series. Philadelphia, PA: Society for Industrial and Applied
models have the same penalties, the resulting controller Mathematics.
will be the same; if not, the resulting controller will be Franchek, M. A., & Herman, P. A. (1998). Direct connection between
different. The key point is the content of design penalty time-domain performance and frequency-domain characteristics.
International Journal of Robust and Nonlinear Control, 8,
not the size of uncertainty. If the penalty is adequate, the
1021–1042.
resulting controller will lead to good performance After Gahinet, P., Nemirovski, A., Laub, A. J., & Chilali, M. (1995). LMI
checking the closed-loop poles in Tables 3 and 5, it is control toolbox. Natick: The MathWorks.
noticed that the controller obtained from the synthesis Grassi, M., Vetrella, S., & Moccia, A. (1995). Preliminary design of the
model with smaller uncertainty matrix D of size 9 9 attitude control system of a microsatellite for earth observation.
produces a very slow response. Therefore, although the Space Technology, 15(4), 223–230.
Haley, D. R., Strikwerda, T. E., Ray, J. C., Fisher, H. L., Heyler, G.
synthesis model has a smaller uncertainty matrix, the A., & Pham, R. T. (1997). Performance of the MSX guidance and
controller performance is unacceptable. As to what is a control system. Guidance and control 1997—proceedings of the
proper penalty, the suggestion is that a proper penalty annual rocky mountain guidance and control conference. Advances in
should not contain any state multiplied by a very small the astronautical sciences, Vol. 94 (pp. 311–330), American
coefficient. Astronautical Society Publication, San Diego.
Helvajian, H. (Ed.). (1999). Microengineering aerospace systems (pp.
347–387). El Segundo, CA: The Aerospace Press and Reston,
Virginia: American Institute of Aeronautics and Astronautics.
References Keel, L. H., & Bhattacharyya, S. P. (1997). Robust, fragile, or optimal?
IEEE Transactions on Automatic Control, 42(8), 1098–1105.
Alonso, R., Anigstein, P., & Sanchez-Pena, R. (1998). SAC-A attitude Khargonekar, P. P., & Rotea, M. A. (1991). Mixed H2/HN control: A
control design. Spaceflight dynamics 1998—proceedings of the AAS/ convex optimization approach. IEEE Transactions on Automatic
GSFC international symposium on space flight dynamics. Advances Control, 36(7), 824–837.
in the astronautical sciences, Vol. 100, Part I (pp. 99–100), Kissel, G. J. (1995). Precision pointing for the Pluto mission
American Astronautical Society Publication, San Diego. spacecraft. Guidance and control 1995—proceedings of the annual
Bank, T. (1995). All stellar attitude estimation using A ball CT-633 rocky mountain guidance and control conference. Advances in the
star tracker. Guidance and control 1995—proceedings of the annual astronautical sciences, Vol. 88 (pp. 421–432), American Astro-
AAS rocky mountain guidance and control conference. Advances in nautical Society Publication, San Diego.
the astronautical sciences, Vol. 88 (pp. 59–66), American Astro- Maciejowski, J. M. (1989). Multivariable feedback design. New York:
nautical Society Publication, San Diego. Addison-Wesley Publishing Company.
Bernstein, D. S., & Haddad, W. M. (1989). LQG control with an HN Merhav, S. (1996). Aerospace sensor systems and applications. New
performance bound: A Riccati equation approach. IEEE Transac- York: Springer.
tions on Automatic Control, 34(3), 293–305. Muncheberg, S., Krischke, M., & Lemke, N. (1996). Nanosatellites
Bialke, B. (1998). High fidelity mathematical modeling of reaction and micro systems technology—capabilities, limitations and
wheel performance. Guidance and control 1998—proceedings of the applications. Acta Astronautica, 39(9–12), 799–808.
annual AAS rocky mountain guidance and control conference. Nam, M.-R., Hashimoto, T., Ninomiya, K. (1997). Design of HN
Advances in the astronautical sciences, Vol. 98 (pp. 483–496), attitude controllers for spacecraft using a magnetically suspended
American Astronautical Society Publication, San Diego. momentum wheel. European Journal of Control (3), 114–124.
Boyd, S., El Ghaoui, L., Feron, E., & Balakrishnan, V. (1994). Linear O’Donnell, J. R., Jr., & Mangus, D. J. (1998). An approach to the
matrix inequalities in system and control theory, SIAM studies in design and implementation of spacecraft attitude control
applied mathematics. Philadelphia, PA: SIAM. systems. Spaceflight dynamics 1998—proceedings of the AAS/GSFC
970 C.-D. Yang, Y.-P. Sun / Control Engineering Practice 10 (2002) 951–970

international symposium on space flight dynamics. Advances in the Wie, B., Liu, Q., & Bauer, F. (1993). Classical and robust HN control
astronautical sciences, Vol. 100 (pp. 125–139), American Astro- redesign for the Hubble space telescope. Journal of Guidance,
nautical Society Publication, San Diego. Control, and Dynamics, 16(6), 1069–1077.
Pastena, M., & Grassi, M. (1998). SMART attitude acquisition and Wie, B., Liu, Q., & Sunkel, J. (1995). Robust stabilization of the space
control. The Journal of the Astronautical Sciences, 46(4), 379–393. station in the presence of inertia matrix uncertainty. Journal of
Rotstein, H., & Sideris, A. (1994). HN optimization with time-domain Guidance, Control, and Dynamics, 18(3), 611–617.
constraints. IEEE Transactions on Automatic Control, 39(4), Wiemer, D. (1998). Attitude determination and control for the
762–779. global imaging system 2000. Guidance and control 1998—
Sahraoui, N. H., Boudjemai, A., Mohammed, M. A. S., & proceedings of the annual rocky mountain guidance and
Abderrahmane, L. H. (1998). Pre-project of microsatellite for control conference. Advances in the astronautical sciences, Vol. 98
remote sensing and climate data collection. Space Technology, (pp. 139–159), American Astronautical Society Publication,
18(4-6), 143–148. San Diego.
Sanchez-Pena, R. S., & Sznaier, M. (1998). Robust systems theory and Wolfe, S. M., Alfriend, K. T., & Leonard, B. S. (1995). A magnetic
applications. New York: Wiley. attitude control system for sun pointing satellites. Astro-
Scherer, C. W. (1995). Multiobjective H2/HN control. IEEE Transac- dynamics 1995—proceedings of the AAS/AIAA astrodynamics
tions on Automatic Control, 40(6), 1054–1062. conference. Advances in the astronautical sciences, Vol. 90, Part I
Scherer, C., Gahinet, P., & Chilali, M. (1997). Multiobjective output- (pp. 1047–1064), American Astronautical Society Publication,
feedback control via LMI optimization. IEEE Transactions on San Diego.
Automatic Control, 42(7), 896–911. Won, C.-H. (1999). Comparative study of various control methods for
Sheen, J.-J., & Bishop, R. (1994). Spacecraft nonlinear control. The attitude control of a LEO satellite. Aerospace Science and
Journal of the Astronautical Sciences, 42(3), 361–377. Technology, 3(5), 323–333.
Skogestad, S., & Postlethwaite, I. (1996). Multivariable feedback Zames, G. (1966a). On the input–output stability of nonlinear time-
control: Analysis and design. Chichester: Wiley. varying feedback systems, Part I: Conditions derived using
Somov, Ye. I., Butyrin, S. A., Matrosov, V. M., Anshakov, G. P., concepts of loop gain, conicity and positivity. IEEE Transactions
Antonov, Yu. G., Makarov, V. P., Sorokin, A. V., Bashkeyev, N. on Automatic Control, 11(2), 228–238.
I., & Kondrat’yev, O. A. (1999). Ultra-precision attitude control of Zames, G. (1966b). On the input–output stability of nonlinear time-
a large low-orbital space telescope. Control Engineering Practice, 7, varying feedback systems, Part II: Conditions involving circles in
1127–1142. the frequency plane and sector nonlinearities. IEEE Transactions
Wertz, J. R. (1978). Spacecraft attitude determination and control. on Automatic Control, 11(3), 465–476.
Dordrecht: Kluwer Academic Publishers. Zhou, K., Doyle, J. C., & Glover, K. (1996). Robust and optimal
Wertz, J. R., Larson, W. J. (1999). Space mission analysis and design control. New Jersey: Prentice-Hall.
(3rd ed.). California: Microcosm Press, London: Kluwer Academic Zimbelman, D., & Watzin, J. G. (1999). On-orbit performance of the
Publishers. transition region and coronal explorer (TRACE) attitude control
Whorton, M. (1997). High performance, robust control of flexible space system. Guidance and control 1999—Proceedings of the annual AAS
structures. Ph.D. Dissertation, California Institute of Technology. rocky mountain guidance and control conference. Advances in the
Wie, B. (1998). Space vehicle dynamics and control, AIAA Education astronautical sciences, Vol. 101 (pp. 437–455), American Astro-
Series. Reston: AIAA. nautical Society Publication, San Diego.

You might also like