Sei sulla pagina 1di 7

CONTROL LAW COMPARISON FOR ATTITUDE

CONTROL OF A SPACECRAFT
Hossein Seif Zadeh and Craig Scott
Doctoral Students, Sir Lawrence Wackett Centre for Aerospace Design Technology
Royal Melbourne Institute of Technology
Lincoln A. Wood
(Adjunct professor, Department of Aerospace Engineering, RMIT)
Research Leader, Ship Systems and Survivability, DSTO

Abstract: This work is primarily an investigation of attitude control of a linear model of the Optus class
B satellite with solar panels using three modern control techniques, namely: ‘fuzzy logic’, ‘ℓ1 ’, and ‘H∞ ’.
For each of the methods a design process is presented. A linear model was used to simulate each of the
control methods. Responses of the system as well as the actuator command are presented. Advantages
and disadvantages of each control method are listed, and future research directions are outlines.
Keywords: Control Law, Fuzzy Logic, l1 , H∞

1 Introduction
The last few decades have seen a rapid growth in the
use of space-based communication satellites. Re-
cent satellites use progressively higher transmission
power in order to reduce ground segment costs.
But increased transmission power also brings about
higher power consumption and therefore larger solar
panels. The increased size of the solar panels in con-
junction with very low force working environments Figure 1: Schematic of the Model and definition of
have resulted in structures which are extremely flex- its degrees of freedom.
ible and have very low frequency structural vibra-
tion modes [7]. These structures also exhibit very
as the central hub. These energy terms are then
low damping, especially in the absence of aerody-
substituted into Lagrange’s equations and then ar-
namic drag (ie. aerodynamic damping).
ranged in state space form. For detailed derivation
This work is primarily an investigation of atti-
of the equations see [7, 8]. The attitude problem is
tude control of such a structure using three modern
defined as rotating the central hub along the axis
control techniques, namely: ‘fuzzy logic’, ‘ℓ1 ’, and
passing through the central hub and parallel to the
‘H∞ ’. For each of the methods a design process
solar panels’ joints. The rotation should be car-
is presented. A linear model of the Optus class B
ried out in the shortest possible time with minimal
satellite was used to simulate each control method
vibration induced in the solar panels. If possible,
for a step input of π2 rad. Responses of the system
over shoot should be avoided. The maximum torque
as well as the actuator command are presented. Ad-
available is about 1 N.m.
vantages and disadvantages of each control method
are listed and future research direction is outlines.
3 Fuzzy Logic Controller
2 Attitude Control Problem Figure 2 depicts the block diagram for the system
Definition in conjunction with the fuzzy controller. In order to
eliminate instability and to give the fuzzy controller
The configuration modelled is similar to the Optus a ‘prediction’ capability, a PD fuzzy controller is de-
class B satellite. The model has three solar pan- signed [2]. The two input signals to the controller
els on each side of a central hub (figure 1). The are the error in angle of rotation (between the ref-
equations of motion are developed by deriving the erence input and the system output) and the angu-
kinetic and potential energy of each panel as well lar speed of the main bus. These are then used to
Figure 2: Block diagram of the controlled plant us-
ing fuzzy logic controller. Figure 3: Fuzzy membership functions for input 1

produce a control signal (torque) with fuzzy logic


relations (rules) [10, 5].
The universe of discourse (domain of all possible
input parameter values) of both inputs was divided
using five membership functions. In the construc-
tion of the fuzzy membership functions inappropri-
ate overlap was avoided. Inappropriate member-
ship functions can result in a fuzzy controller with Figure 4: Fuzzy membership functions for output
haphazard behaviour. ‘Appropriate’ overlap can be
achieved employing the following classical rules [1]:
• Every point in the universe of discourse should rotation IS positive small AND speed of rotation
belong to at least one membership function; IS positive small’ means that the bus is rotating
with a small angular velocity towards the set posi-
• no point can belong to the domain of more than tion and the current angular position is close to the
two membership functions; set point. In that case, one may decide that there
• each membership function must have a dis- is no need for any action to be taken at this time
tinct point of maximum meaningfulness (no and therefore the statement would be finished with
two should overlap); ‘THEN torque IS zero’. Another designer might ar-
gue, for the same input, that because the angular
• no membership function’s domain can go be- position is close to the set point there should be a
yond the point of maximum meaningfulness of negative torque applied to the bus to decrease its
neighbouring membership functions. angular speed, thereby reducing over-shoot. Thus
Membership functions for the first input (error constructing fuzzy logic rules boils down to a high
between the set point and the actual angular posi- degree of physical insight to the problem as well
tion), created using the above rules, are shown in as a set of priorities the designer gives to different
figure 3. The second input (angular velocity) mem- aspects of the response of the system.
bership functions are the same as those for the first For the output, seven membership functions were
input, except their domain is different. It is as- used (figure 4). The universe of discourse was set to
sumed that the first input can vary between ‘±2π’. ‘±1N.m’ which is the maximum available torque.
It was found during the simulation that the second The final step in Fuzzy logic control is known
input varies between ‘±0.25 rad/sec’. In order to as ‘defuzzification’. Because fuzzy sets can assume
prevent input saturation of the fuzzy controller, the more than just ‘0’ and ‘1’, more than one fuzzy
universe of discourse of the second input was chosen output variable may assume a non-zero degree of
to cover twice as much variation, ie. ‘±0.5 rad/sec’. truth. The defuzzification method used was ‘centre
Membership functions are named: ‘N V L’, ‘N L’, of gravity’. In this method, the ‘centre of gravity’ of
‘N M ’, ‘N S’, ‘ZR’, ‘P S’, ‘P M ’, ‘P L’ and ‘P V L’ to the non-zero membership functions are calculated;
represent input and/or output values linguistically a weight equal to the degree of truth is given to
as ‘negative very large’, ‘negative medium’, ‘zero’ each membership function during the calculation.
and so on. The defuzzification process produces the controller
Fuzzy logic rules are expressed in the form of if- output which in this case is fed to the plant (see
then statements [5]. The if-then expressions must figure 2).
be constructed with physical insight to the prob- The response of the system and the actuator com-
lem at hand [4]. For example, for the satellite at- mands to a step input of ‘ π2 rad’ are shown in fig-
titude problem, the statement ‘IF error in angle of ures 5 and 6 respectively. As is evident from the

2
1.6

1.4

1.2
Angle of rotation (rad)

0.8

0.6

0.4

Figure 7: Block diagram of the ℓ1 controller in con-


0.2
junction with the plant
0
0 5 10 15 20 25 30 35 40 45 50
Time (sec)
ify various performance constraints such as actua-
Figure 5: System response to step input (fuzzy logic tor limits, maximum tracking error, etc. that the
controller) controller should maintain. Finding the controller
requires the optimisation of two separate linear pro-
0.8
grams - one for the optimal tracking behaviour and
another for robustness considerations. The two
0.7
problems can be represented as:
0.6

0.5
TRACKING
Torque (N.m)

0.4 min {κ1 kφk1 + κ2 kuk1 }


0.3
dw1 q − u = 0
subject to
0.2
nw1 q + φ = w1
0.1
ROBUSTNESS
0  
κ1 kW2 nF k1 +κ2 kW3 dF k1 +κ3 kW4 nF k1
min
−0.1
0 5 10 15 20 25 30 35 40 45 50 +κ4 kW2 nGk1 +κ5 kW3 nF k1 +κ6 kW4 nGk1
Time (sec)

nF + dG = 1
Figure 6: Actuator command to get system re- subject to kW4 nF k1 ≤ 1
sponse of figure 5 F0 = 0
where
figures, the system does not have any overshoot and φ = tracking error
settles in about 20 seconds. Note that the actuator u = actuator
command is well within the boundaries of the avail- w1 = the input that the plant output is
able torque (±1N.m). This ensures that saturation to track
of the actuator does not happen. Note also that n, d = plant numerator and denominator
higher modes of vibration of the structure may be polynomials
excited more than normal if the actuator saturates. q, F, G = controller polynomials
Also note that no attempt to optimise the controller κi = scalar weightings on the various
was made (see section 8). norms
W2 , W3 = f requency weightings on the
actuator disturbances and
4 ℓ1 Controller sensor noise to ref lect their
f requency content
The ℓ1 controller was found using the method ex- W4 = f requency − dependent uncertainty
plained in [9] (but is somewhat different to the con- prof ile f or the plant model
ventional ℓ1 approach). It is a discrete-time tech-
nique that uses ℓ1 norms and linear programming Each term is a sequence in the z-transform vari-
methods to generate controllers of the form shown able. For example, the tracking error would be rep-
in figure 7. The method allows the designer to spec- resented by φ = φ0 + φ1 z −1 + · · · + φk z −k . The

3
plant polynomials (n and d) are the numerator 1.6

and denominator of the discrete transfer function 1.4


of the nominal plant model. The tracking optimi-
1.2
sation finds the optimal tracking error and actua-
tor sequence for the nominal plant in a noise- and 1

Angle of rotation (rad)


disturbance-free environment. The ‘q’ part of the 0.8

controller is fixed by this first optimisation. The


0.6
robustness optimisation minimises the effect of the
disturbances, sensor noise and errors in the plant 0.4

model on the tracking error and actuator usage. 0.2

The ‘F ’ and ‘G’ parts of the controller depend on 0


this second optimisation. In both optimisations, the
scalar κi terms can be adjusted to specify the rela- −0.2
0 2 4 6 8 10 12 14 16 18 20
Time (sec)
tive importance of each norm in the objective. The
more important it is for a particular norm to be
Figure 8: System response to step input (ℓ1 con-
minimised, the higher its associated κi value. The
troller)
controller used in this paper was formed using the
following parameters: 1

0.8
TRACKING:
0.6
κ1 = κ2 = 1
w1 was a step input. ‘u’ and ‘f ’ were restricted 0.4
Actuator command (N.m)

to 400 terms (no improvement was observed with 0.2

more terms) and the actuator was also restricted 0

to | u |≤ 0.8, | u̇ |≤ 0.3 and | ü |≤ 0.05. The −0.2


first two conditions on the actuator arose from the −0.4
limitations of the system to be controlled, while the
−0.6
third condition was included to reduce the vibration
at the end of the maneuver. −0.8

−1
0 2 4 6 8 10 12 14 16 18 20
Time (sec)
ROBUSTNESS:
κi = 1 (i = 1 · · · 6)
Figure 9: Actuator command to get system re-
‘F ’ and ‘G’ were restricted to no more than 400
sponse of figure 8
terms. The weighting functions were

s + 100 s + 10 and the more conventional ℓ1 methods. H∞ is es-


W2 = 0.015 W3 = 0.005 W4 =
s + 10 s + 50 sentially a frequency domain technique, whereas ℓ1
control is set in the time domain. The H∞ method
The response of the system and the actuator com- of obtaining controllers requires the designer to
mand to a step input of ‘ π2 rad’ are shown in figures 8
specify three weighting functions (W1−3 ) to tailor
and 9 respectively. These results include distur- the solution to the problem. Following the usual
bances of up to 0.15 N.m and sensor noise of up to notation for this type of analysis, these weighting
0.005 rad. The plant model was also modified by functions refer to the sensitivity function, a quan-
changing the hinge stiffnesses and the level of damp- tity that has no common name and the complimen-
ing to represent modelling errors. The figures show tary sensitivity function respectively (see [3] for fur-
that the plant not only reached the set point with ther details). The weights used for the controller in
no overshoot, it also settled very quickly (about 8 this paper were
seconds). The controller was also able to keep the
effect of the noise, etc. very small.
W1 = 2 × 10−5 ss +
+ 500
0.01 W2 = 50 ss+
+ 10
500 W3 = 0

5 H∞ Controller It is worth noting that these weighting functions


were arrived at by trying various combinations until
The H∞ approach to controller design is signifi- acceptable results were obtained. This process was
cantly different to both the ℓ1 approach used above by no means trivial, as certain combinations would

4
2.5 to modelling errors.

• The more physical insight to the problem the


2
designer has, the better the fuzzy logic con-
troller will perform.
Angle of rotation (rad)

1.5

• The designer can specify the relative impor-


tance of each of the inputs as well as the track-
1
ing error.

0.5
6.1.2 Disadvantages

0
• Performance is related to the designer experi-
0 5 10 15 20 25
Time (sec)
30 35 40 45
ence.

Figure 10: System response to step input (H∞ con- • Stability and/or optimisation analysis are not
troller) yet well established.

6.2 ℓ1
1

6.2.1 Advantages
0.8
• The designer has direct control over the be-
haviour of the plant. Constraints may be
Actuator command (N.m)

0.6
placed on quantities such as the tracking er-
0.4
ror, maximum actuator level or slew rate, or a
number of other quantities. These constraints
0.2
can even be different for positive and negative
values or change with time.
0
• The method yields a controller that is opti-
mally insensitive to plant modelling errors and
−0.2
0 5 10 15 20 25
Time (sec)
30 35 40 45
guarantees stability for the specified uncertain-
ties.
Figure 11: Actuator command to get system re-
sponse of figure 10 • The controller is optimally insensitive to dis-
turbances and sensor noise which may be char-
acterised by weighting functions reflecting the
yield unacceptable actuator levels or excessive slew frequency contents.
rates which the system was incapable of achieving.
The response of the system and the actuator com- • The controllers tend to yield close to minimum-
mands to a step input of ‘ π2 rad’ are shown in fig- time / minimum-fuel maneuvers, with the
ures 10 and 11 respectively. The figures show that maximum controller order being specified to an
the system exhibits a relatively large degree of over- extent by the designer.
shoot and it settles in about 40 seconds.
• The designer can specify the relative impor-
tance of the tracking error and the actuator
6 Comparison usage in the optimisation itself.

6.1 Fuzzy logic 6.2.2 Disadvantages


6.1.1 Advantages • The controllers generated by the technique are
• The designer has direct control over the be- usually of quite high order.
haviour of the system.
• The optimisation problems can become very
• It does not require a mathematical model of the large for high order plants. so numerical dif-
plant. As such it is, by definition, insensitive ficulties may occur for very large problems.

5
6.3 H∞ in the process of being extended to the multi-input,
multi-output (MIMO) case. H∞ is also an active
6.3.1 Advantages
area of current research with experimental imple-
• The order of the controllers generated are usu- mentations appearing in the literature as well as
ally close to that of the plant. theoretical advancements.

• The time to calculate a solution once the design


variables have been decided is relatively short. References
• The controller can be made optimally insensi- [1] A. V. Bourmistrov, P. Riseborough, and
tive to plant modelling errors. J. McPhee. Fuzzy logic based guidance sys-
tem for autonomous flight vehicle. In Control
6.3.2 Disadvantages 95 Preprints, volume 2, pages 375–379, Mel-
bourne, Australia, October 1995.
• The controller found may exhibit relatively
poor time-domain characteristics, such as long [2] Y. Y. Chen and C. C. Yen. PD-type vs.
settling times, too much overshoot or excessive PID-type fuzzy controllers. In Proceedings of
actuator demands. TENCON’92 IEEE Region 10 Conference, vol-
ume 1, pages 341–345, Melbourne, Australia,
• The weighting functions needed for the ap- 1992.
proach are not necessarily easy to determine.
[3] Richard Y. Chiang and Michael G. Safonov.
• Optimising for both tracking performance and Robust Control Toolbox for use with Matlab.
disturbance rejection (or robustness) at the The Math Works Inc., 1992.
same time is very difficult.
[4] M. Jamshidi, N. Vadiee, and T. J. Ross. Fuzzy
logic control software and hardware applica-
7 Conclusion tions, chapter ‘Fuzzy rule-based expert sys-
tems II’, page 93. Prentice Hall, Inc., 1993.
The current research has indicated that when math-
ematical model of a system is not known in detail [5] MathWorks Inc. Fuzzy Logic Toolbox for use
fuzzy logic controller is a good choice. But one with Matlab. The MathWorks Inc., 1994.
needs to have a physical insight to the system in
order to design a fuzzy logic controller for it. Fur- [6] P. Riseborough, R. Hill, J. Vethecan, and L. A.
thermore fuzzy controller yields better time domain Wood. ℓ1 norm based optimal tracking and vi-
performance than H∞ controller. If mathematical bration control. In First Australasian Congress
model of the system is known in detail, and per- on Applied Mechanics, pages 819–824, Mel-
formance of the system is of prime concern ℓ1 con- bourne, Australia, February 1996.
troller should be used. Although it can result in
[7] Paul Riseborough. Dynamics and Control of
controllers with very high orders.
Spacecraft With Solar Array Joint Nonlinear-
ities. Ph.D. Thesis, RMIT, Melbourne, Aus-
8 Current Research tralia, 1993.

The fuzzy logic controller presented in this paper [8] Paul Riseborough, Lincoln A. Wood, and M. B.
was the result of the first attempt to design such a Pszczel. Nonlinear problem in the control of
controller for the system. Fine tuning of the con- flexible spacecrafts. In International Aerospace
troller is currently in progress and results will ap- Congress, Melbourne, Australia, 1991.
pear in the literature in the near future. Further- [9] C. N. Scott and Lincoln A. Wood. Optimal
more, noise rejection of the fuzzy controller is cur- robust tracking subject to disturbances, noise,
rently being investigated and preliminary results, at plant uncertainty and performance constraints.
this stage, look promising. Journal of Guidance, Control, and Dynamics,
The first known implementation of an ℓ1 con- April 1998.
troller was published in 1996 [6]. The work in [9]
for a robust form of an ℓ1 controller has been ex- [10] H. B. Wang, J. P. T. Mo, and N. Chen. Fuzzy
perimentally tested on the rig and the results shall position control of pneumatic cylinder with two
appear in a future paper. The theory is currently state solenoid valves. In Control 95 Preprints,

6
volume 2, pages 387–391, Melbourne, Aus-
tralia, October 1995.

Potrebbero piacerti anche