Professional Documents
Culture Documents
In this paper, the modeling of flexible multi-body systems (typically a satellite), the
design of a reduced-order robust controller and the closed-loop system robustness analysis
are considered. First, we present a generic tool developed to build the model of a multi-
body dynamic system taking into account parametric uncertainties for robust control and
Downloaded by COLUMBIA UNIVERSITY on April 23, 2013 | http://arc.aiaa.org | DOI: 10.2514/6.2012-4845
analysis purpose. Satellites are typically composed of a main body (hub of the satellite)
and cantilevered appendages (flexible or not). Based on the description of each element
(mechanical and geometrical properties, uncertain parameters ...), a linear model between
torques applied to the hub (control or perturbation inputs) and angular accelerations is
built using the Linear Fractional Representation framework. Second, based on this model,
a control design method is proposed: from a basic standard scheme and the latest H∞
solver, a low order H∞ controller can be derived. Roll-off requirements and worst-case
parametric configurations can also be taken into account in a multi-model design approach.
Third, robustness analysis of the design on the full order model is considered. A numerical
example illustrate the relevance and effectiveness of the proposed approach.
Nomenclature
christelle.cumer}@onera.fr)
† Professor, University of Toulouse-ISAE, 10, Av. Edouard Belin, Toulouse- 31055 FRANCE. (Email: daniel.alazard@isae.fr)
1 of 18
Copyright © 2012 by ONERA - The French Aerospace Lab. Published by the American Institute of Aeronautics and Astronautics, Inc., with permission.
American Institute of Aeronautics and Astronautics
∆ matrix of uncertainties
∆wc matrix of worst case values of uncertainties
In identity matrix of order n
0n×m null n × m matrix
IG inertia matrix at point G
Ixx static gain of the system (angular model) around X axis
diag(A, B) matrix with elements A and B on diagonal, 0 otherwise
Subscript
i flexible mode number
Acronym
LFT Linear Fractional Transformation
dof degrees of freedom
AOCS Attitude Orbit Control System
Downloaded by COLUMBIA UNIVERSITY on April 23, 2013 | http://arc.aiaa.org | DOI: 10.2514/6.2012-4845
I. Introduction
Spacecrafts are very complex mechanical multi-body systems including flexible and/or rotating ap-
pendages. Due to a lack of tools to deal with nonlinear spacecraft models, the design of the AOCS requires
a linear model taking into account all the rigid and flexible couplings between the hub (where the AOCS
acts) and the various appendages.
Note that the linear assumption is quite realistic for such systems since perturbations and motions are very
small (except for very agile observation satellites). This linear assumption is particularly valid in the context
of future missions for deep space exploration involving formation flying of several spacecrafts.
In this framework, engineers need a tool to develop quickly the dynamic model of the spacecraft including
the flexible modes, and eventually parametric uncertainties. In1 authors have proposed some tools developed
with MATLAB/SIMULINK to efficiently build the linear dynamic model of any open mechanical chain. The
method proposed in this paper is an extension of these tools in order to better study the attitude of the
satellite angular deflection around a pointing direction of the payload.
It must be emphasized that the multi-body modeling, presented in,1 first distinguishes the main body
dynamics from the flexible appendage dynamics and secondly connects them together. The resulting model
directly highlights the characteristic parameters of the main body and of the appendage. As a result, their
uncertainties can be explicitly introduced. Then a minimal LFR model of the satellite can be derived
and allows for robust stability to be analysed by classical methods (phase-gain margins, Bode plots and
µ-analysis2 ).
The second problem of interest considered in this paper addresses the control design problem. Indeed,
based on recent development of H∞ methods,34 and their implementation in MATLAB, structured and so
reduced-order H∞ controllers can be designed. Such new solvers (hinfstruct) present several advantages
with respect to classical full-order solvers already available
• the controller order is a design parameter. This is an advantage with respect to the classical H∞
approach where the controller order equals the augmented plant order, thus resulting in a complex
controller. As a result, all flexible modes can be taken into account in the design process, without the
need for a high order controller. Thus, in comparison with classical full order H∞ design, the model
and/or controller reduction steps are skipped,
• the design of a low order controller is a key point for implementation purpose
• beyond the controller order, the structured approach allows us to define one controller per axis thus
maintaining the axis decoupling property,
2 of 18
II.A. Background
II.A.1. Definition and transport of the dynamic model of a rigid body
Let us consider a rigid body S, of mass m and inertia IG at its center of gravity G. The linear (direct)
dynamic model of the system S at point G is characterized by a block diagonal matrix:
" #
S mI3 03×3
DG = (1)
03×3 IG
The dynamic model of S can be easily transported from its center of gravity G to a point P by the
relation:
DPS = τPT G DG
S
τP G (3)
−y x 0 GF F
3 of 18
If we consider now that an appendage A is cantilevered to the base MB at point P, the reaction force
→
− →
−
F M B/A and torque T M B/A,P at point P between the base and the appendage must be taken into account
Downloaded by COLUMBIA UNIVERSITY on April 23, 2013 | http://arc.aiaa.org | DOI: 10.2514/6.2012-4845
→
−
" #
aP
η̈ + diag(2ξi ωi )ω̇ + diag(ωi2 )η = −LTP →
−̇ (10)
ω
This formulation depends on the appendage flexible modes ωi through three matrices: the stiffness matrix
K = diag(ωi2 ), the damping matrix D = diag(2ξi ωi ) and the modal participation factor matrix LP at point
P.
4 of 18
5 of 18
z
A
F
P
y
A
SA
R
B GFF
Downloaded by COLUMBIA UNIVERSITY on April 23, 2013 | http://arc.aiaa.org | DOI: 10.2514/6.2012-4845
# "
~ F,SAR
δa
First, one can compute the 6 d.o.f. relative acceleration vector of the point F w.r.t. antenna
δ~ω̇F,SAR
reference frame SAR using the (6 × N ), N being the number of flexible modes of the antenna) modal shape
matrix φF between modal coordinates of the antenna η and the 6 displacements at point F . Considering
the state-space hybrid-cantilevered model MPA (s) of the antenna at point P :
η̇ 0k×k IN η 0k×6 ~aP
= +
η̈ −Kk×k −D η̇ −LP ~˙
ω
F~B/A h i η ~aP
= −LTP K −LTP D + (DPA − LTP LP ) (11)
T~B/A,P η̇ ~˙
ω
Then: " #
~ F,SAR
δa η ~aP
= φF η̈ = [−φF K − φF D] − φF Lp (12)
δ~ω̇F,SAR η̇ ~˙
ω
Then one has to take into account the acceleration of the SAR frame w.r.t. inertial frame at the point F :
" # " #
~F
δa ~ F,SAR
δa ~aF
= + (13)
δ~ω̇F δ~ω̇F,SAR ~˙
ω
with
~aF ~aP
= τF P . (14)
~˙
ω ~˙
ω
Finally:
" #
~F
δa η ~aP
= [−φF K − φF D] + (τF P − φF Lp ) . (15)
δ~ω̇F η̇ ~˙
ω
Lastly:
6 of 18
This way the dynamic model MpA (s) of the antenna is augmented with two additional outputs θ̈e and θ̈e
and is renamed MpA (s). Projecting in SAR this model reads:
2 3 2 32 3 2 32 3
η̇ 0k×k Ik η 0k×6 ~
aP
4 5=4 54 5+4 54 5
η̈ −Kk×k −Dk×k η̇ −LPk×6 ~˙
ω SAR
2 2 3 3
F~B/A
Downloaded by COLUMBIA UNIVERSITY on April 23, 2013 | http://arc.aiaa.org | DOI: 10.2514/6.2012-4845
6 4 5 7 " # 2 3
6 7
LT
h i η
6 ~B/A,P
T 7
= − P (19)
SAR K D 4 5
kerT ([z~A ]T
6 7
SAR )φF (4 : 6, :) η̇
6 7
4 θ̈e 5
θ̈a
" #2 3
A
(DP − LT ~
aP
+ P LP ) 4 5 (20)
kerT ([z~A ]T
SAR )(I 3 − φF (4 : 6, :)Lp ) ~˙
ω SAR
The connection of this model M~pA (s) to the main body taking into account the (6 × 6) rotation matrix
T between SAR and GFF is depicted in Figure 5. Angular deflection can be determined by the double
integration of θ̈e and θ̈a . Note that these integrators are also non minimal and can be removed by computing
a minimal realization of the whole model between the 6 d.o.f force vector and the 6 + 2 position vector.
..
θe
..
θa
a
A P
MP (s) .
ω
SAR
T T
T
τT τ
PB PB
− MB −1 1 1
+ [ DB ] s s
F a
ext B
.
Text,B ω
Figure 5. Block Diagram of the Main body + antenna dynamic model with augmented outputs for pointing angular
deviation studies.
7 of 18
The framework is based on isolating uncertainties in the matrix ∆ from the nominal system M (s), and then
connecting them through a feedback (see Figure 6).
To obtain the LFT model, one MATLAB LFR object per subsystem is created. This step is very simplified
Downloaded by COLUMBIA UNIVERSITY on April 23, 2013 | http://arc.aiaa.org | DOI: 10.2514/6.2012-4845
by the fact that each parameter of the multi-body system have been introduced independently during the
modeling process and in a minimal way. Then, each uncertain parameter can be represented by an LFT
object in the structured model and the the LFT representation of the whole uncertain model can be easily
built by the interconnection of the various subsystems (see Figure 7 and also,1011 ).
• design model (DM): corresponding to the nominal inverse dynamic model (without uncertainty). This
model will be used to design the controller.
• validation model (VM): LFR model representing the actual system (with uncertainties) which will be
interconnected to the designed controller in order to perform the stability analysis.
8 of 18
angle approximation.
The disturbance w acts on the acceleration vector, the controlled output z weights the acceleration with
the diagonal weighting matrix.
s2 +2ξ ω s+ω2
1 1
s2
1
0 ... ... 0
.. ..
0 . 0 .
Wq = .
.. s2
+2ξ ω
i i s+ω 2 .
..
(21)
0 i
0
s 2
.
.. . ..
0 0
2 2
s +2ξn ωn s+ωn
0 ... ... 0 s2
Wq (s) is the inverse of the desired acceleration sensitivity. For a degree of freedom i, the desired accel-
eration sensitivity function specifies that low frequency disturbances on the acceleration must be perfectly
rejected while high frequency disturbances (beyond the bandwidth ωi of this degree of freedom) are not re-
jected. The multi-variable weight Wq (s) is diagonal and depends only on the specifications ωi , ξi , i = 1 . . . n
and not on the system dynamic parameters.
9 of 18
where K{Ixx,Iyy,IzzZ} are the static gains needed for each coordinate to satisfy the correct disturbance
rejection. Note that these gains could correspond to those computed√ for a classical PD controller. The
pulsation ωcut is the roll-off filter cutting frequency. The weight 2 is introduced to add a 3dB margin. The
effect of theses templates on the frequency response of each controller is highlighted in Figures 15 to 17.
Note that this controller frequency-domain shaping is a quite original feature of this methodology, brought
by the structured H∞ optimization software. In particular, one must note that, on top of these achievements,
this methodology also ensures that no pole is added to cancel the system cantilevered zeros. Indeed, the
frequency shape ensures that there is no high peak in the controller frequency response, thus forbidding such
additional badly damped dynamics in the controller. This is a very desirable feature since such poles-zeros
cancellation is a well-known source of non-robustness in the closed-loop system.
Downloaded by COLUMBIA UNIVERSITY on April 23, 2013 | http://arc.aiaa.org | DOI: 10.2514/6.2012-4845
IV. Illustration
This section is dedicated to a numerical application of the proposed methodology. At his aim, we consider
a spacecraft composed of three main elements :
10 of 18
The standard problem in Figure 11 includes 88 states. Classical methods to design an H∞ controller
(Riccati-based or LMI-based H∞ optimization approach for example) lead to a full order controller. Beyond
the large computational time to obtain such a 88th order controller, a reduction is often needed, before
the synthesis if some flexible modes are supposed negligible or after the synthesis to obtain implementable
controller. Structured H∞ controller solvers allow a low order controller to be designed directly (in one
shot).
11 of 18
In this section, we consider the nominal specifications. From the disturbance rejection requirements, the
weighting matrix (Eq. 21) can be easily tuned (see Table 3). The desired bandwidths of the closed-loop
system are very low w.r.t. the system first flexible mode frequencies. Then, this control problem is not
very challenging and a second order controller per axis is sufficient enough to meet requirement. Indeed, the
closed-loop performances are summarized Table 4. The corresponding Nichols plots are depicted Figure 12
and highlight good stability margins.
Regarding the time-domain simulation on the full order model, Figure 13 plots system responses to a
1deg initial conditions. The correct decoupling, provided by the block diagonal structure of the controller,
between all axes is illustrated by adding a torque perturbation of amplitude 5N m during 1s on X axis (at
time t = 80 s). The other axes are weakly perturbed by this impulsion.
corresponding H∞ norm is too large). A solution consists in increasing the order of the controller. But in
such a case, we loose noise rejection benefits due to high frequency behavior of first or second order controller.
12 of 18
Figure 13. Temporal responses to initial conditions and torque perturbation on X axis
Then roll-off specifications are required. The high frequency behavior of the structured controller can be
easily imposed by implementing the standard problem depicted Figure 9.
In this framework, the H∞ norm becomes 1.22 and the corresponding Nichols plot are proposed Figure
14.
On Figures 15, 16 and 17, frequency responses of controllers with (continuous black line) or without (con-
tinuous red line) roll-off constraints are plotted. The required frequency-domain behavior of the structured
controller can be verified axis by axis. Considering time-domain responses, the decoupling of each axes is
preserved. As expected, responses to initial conditions or torque perturbation are faster (see Figure 18).
13 of 18
14 of 18
15 of 18
Figure 18. Temporal responses to non null initial conditions and torque perturbation on X axis
Let us consider the 18th order controller designed section IV.B. The µ-lower bound analysis shows that
the system is not robustly stable with respect to uncertain parameter variations of Table 6: the maximum
µ-lower bound is greater than one on Figure 19.
This analysis provides a worst case value of the uncertain ∆-block. As proposed Figure 10, the new design
process should stabilize the systems in such a worst-case configuration, beyond the respect of constraints
of the previous design process. The synthesized controller leads to a H∞ norm of 1.24 for the closed-loop
system (very close to the one obtained section IV.B: 1.22). The robustness analysis of the interconnection of
this controller to the uncertain system is proposed Figure 20 and illustrates the effectiveness of the proposed
standard scheme: the maximal value of the µ lower bound decreased in a significant way.
V. Conclusion
A global methodology was proposed for multi-body dynamic systems -typically satellites- control and
analysis design. This method includes a generic dynamic modeling tool, including dynamic modeling for
16 of 18
Figure 19. µ-lower bound obtained with ωdes = [1 1 1]rad/s and without worst-case consideration
Figure 20. µ-lower bound obtained with ωdes = [1 1 1]rad/s and with worst-case consideration
control design and uncertainty representation for analysis purpose. The second step is a control design
method based on the latest version of the structured H∞ control design MATLAB function. Most impor-
tantly, this features low-order controller design, controller frequency shaping and worst case consideration.
Finally, more classical analysis tools are considered based on the uncertainty modeling earlier mentioned.
The whole method has been successfully applied to a generic satellite model as reported through numerical
simulation results.
Acknowledgements
The authors would like to thank Massimo Casasco from ESA and Chiara Toglia from Thales Alenia Space
Italy for their supports.
17 of 18
on-board angular momentums,” Proceeding of the 7th International ESA Conference on Guidance, Navigation and Control
Systems, Tralee (Ireland), June 2008.
2 Ferreres, G. and Biannic, J., Skew Mu Toolbox (SMT): a presentation, http://www.cert.fr/dcsd/idco/perso/Ferreres/index.html,
2003.
3 Apkarian, P. and Noll, D., “Nonsmooth H
∞ synthesis,” IEEE Transactions on Automatic Control,, Vol. 51, No. 1, jan.
2006, pp. 71 – 86.
4 Gahinet, P. and Apkarian, P., “Structured H
∞ Synthesis in Matlab,” IFAC World Congress, Università Cattolica del
Sacro Cuore, Milano, Italy, 2011.
5 Alazard, D., Fezans, N., and Imbert, N., “Mixed H /H
2 ∞ control design for mechanical systems: analytical and numerical
developments.” AIAA Guidance Navigation and Control conference, Honnolulu (Hawaı̈¿ 12 ), August 2008.
6 Guy, N., Alazard, D., Cumer, C., and Charbonnel, C., “Reduced Order H-Infinity Controller Synthesis for Flexible
Structures Control,” 7th IFAC Symposium on Robust Control Design, Aalborg, Denmark,, June 2012.
7 Charbonnel, C., “H
∞ and LMI attitude control design: towards performances and robustness enhancement,” Acta
Astronautica, Vol. 54, No. 5, 2004, pp. 307 – 314.
8 Cumer, C. and Chretien, J., “Minimal LFT form of a spacecraft built up from two bodies.” Proceedings of the 2001
AIAA Guidance, Navigation, and Control Conference and Exhibit, AIAA, 2001.
Downloaded by COLUMBIA UNIVERSITY on April 23, 2013 | http://arc.aiaa.org | DOI: 10.2514/6.2012-4845
18 of 18