Documente Academic
Documente Profesional
Documente Cultură
A97-37112
AIAA-97-3641
1080
Copyright© 1997, American Institute of Aeronautics and Astronautics, Inc.
well over the scalings, we have to resort to a D-K- together with (2), where
like iteration .
The theory is illustrated on a missile benchmark A(p) + B(p)Dc(p)C(p) B(p)Cc(p)
problem as studied in 14 and ^. In contrast to Bc(p)C(p) Ac(p)
, which considers the nominal performance LPV + B(p)Dc(p)D(p)
problem, we also address robust performance issues. Bc(p)D(p)
Starting from a given nonlinear model, we obtain an = [ H(p) + E(p)Dc(p)C(p) E(p)Cc(p)
uncertain LPV representation accessible for design.
Then the design specifications are translated into • E(p)Dc(p)D(p) ].
suitable weighting functions as in the //oo-approach.
We end up with an interconnection structure and
Downloaded by PRINCETON UNIV. LIBRARY on September 30, 2014 | http://arc.aiaa.org | DOI: 10.2514/6.1997-3641
perform the iteration that is comparable to the D- Analysis of uncertain LPV systems
K procedure in //-synthesis 4. Finally, we validate
In this section we will provide an analysis result
the robust LPV controller by nonlinear simulations.
that characterizes robust stability and robust per-
formance for the uncertain LPV system (4), (2).
LPV design For that purpose we need to introduce scalings that
characterize the nature of the uncertainties Aj af-
fecting the plant in terms of integral quadratic con-
LPV systems straints (IQCs) 13. Hence, for each channel, we de-
The uncertain LPV system is described by fine a collection of matrices Qi, Si, Ri such that the
IQC
x = A(p)x + G(p)w + B(p)u
z = H(p)x + F(p)w + E(p)u (1)
y = C(p)x + D(p)w dt>0 (5)
Si
where, with a suitable partition of the signals w —
holds for T > 0 and for all signals Wi, Zi of finite
\WQ7" ,Wi
Fni i,!r ,..., w
1H T T
and z
k j1 ana [Z0T , Zj
— \y
y — VT , , . . , z
rkT!
j T , flip
tne
energy that are related by Wi — A^. As examples,
uncertainty enters as
we mention
• time- varying parametric uncertainties Wi(t) =
and Wo "-> Zo is the channel to describe the perfor- Zi(t), \8i(t)\ < 1:
mance specification (Figure 1). The parameter p(t)
and its rate of variation p(t) are assumed to be con- < 0 Ri = -Qi, = 0.
tained in the a priori given compact sets P and Pr
respectively. • dynamic uncertainty A./ : L2[0, oo) i-> L2[0,oo)
For controller design the parameter p(t) is assumed with gain not larger than 1:
to be on-line measurable. Hence, LPV controllers
take the form
= Ac(p)xc •Bc(p)y
= Cc(p)xc Dc(p)y (3) We take the I/2-gain of the channel WQ -4 ZQ as a
measure for performance. The I/2-gain of this chan-
such that the resulting closed-loop system is de- nel is bounded by the value 7 if the IQC (5) holds
scribed by with the fixed scaling
1081
Copyright© 1997, American Institute of Aeronautics and Astronautics, Inc.
Finally, we collect the scalings into block-diagonal of points in P and Pr. If Pr is described as a con-
matrices as Q = diag(<2o,Qi,-- - > Q f e ) > R — vex combination of finitely many vertices, it suffices
d\a,g(R0,Ri,.. . ,/Z fc ), and S = diag(Sb,5i,. . . ,Sk). to choose the extreme points since the parameter p
Now we are ready to provide the characterization appears linearly in (6) 19> *.
of robust stability and robust performance in terms Controller synthesis
of the solvability of a so-called scaled differential
Bounded Real Lemma whose proof is straightfor- The synthesis problem consists of designing a con-
ward **» 17> 15> ^. For notational convenience we troller (3) that minimizes the robust performance
define sy(M) := M + MT. level 7 as characterized in Theorem 1. However, the
inequalities (6) are not linear in all the unknowns,
Theorem 1 Suppose there exist smooth and the Lyapunov function, the scalings, and the con-
Downloaded by PRINCETON UNIV. LIBRARY on September 30, 2014 | http://arc.aiaa.org | DOI: 10.2514/6.1997-3641
bounded functions X(p), Q(p), R(p), S(p) on P troller parameters. It has been shown in 10> 15) 16
such that how the inequalities can be linearized, for fixed scal-
ings, by a suitable nonlinear transformation of the
x(P) > o controller parameters as follows: Denote the first
* ] block rows of X(p) and X(p)~l as [ X(p) U(p) }
* <0 and [ Y(p) V(p) ] respectively, and introduce the
R(p)H(p) R(p)\ new controller parameters
(6)
K(p,p) = X(p)[A(p) + B(p)Dc(p)C(p)]Y(p)+
holds for all p €. P and p 6 Pr . Then, for all param- +U(p)Bc(p)C(p)Y(p) + X(p)B(p)Cc(p)VT(p)+
eter curves (p(t),p(t)) £ P x Pr and for all uncer-
+U(p)Ac(p)VT(p) + X'(p,p)Y(p) + U'(p,p)VT(p)
tainties (2), the system (4) remains stable and the
of the performance channel is bounded by 7. L(p) = X(p)B(p)Dc(p) + U(p)Bc(p)
Here X'(p,p) is defined as M(p) - Dc(p)C(p)Y(p] + Cc(p)VT(p)
N(p) = Dc(p).
(7)
3=1 Then synthesis inequalities are simply obtained by
performing the substitutions
Hence we have to find a parameter dependent Lya-
punov function and parameter dependent scalings to XA + LC K
XA AY + BM
satisfy a differential linear matrix inequality > . A + BNC
This generalizes the idea of using a constant Lya-
punov function 5> 2 and constant scalings for arbi- XG + LD (H + ENC)T
trarily fast varying parameters. G + BND (HY + EM)T
To solve the inequalities numerically we choose con- X I X' 0
tinuously differentiable functions fi(p) ••• fi (p) and X
I Y 0 -Y'
search for the coefficients in the expansion
in (6) where we dropped the dependence on p and p
X(p) for notational simplicity.
Q(P) Qj As explained for analysis, we can introduce basis
R(p) Rj functions and grid the parameter set to end up with
3=1 finitely many inequalities. However, the resulting in-
equalities are still nonlinear in the new variables and
The resulting infinitely many LMIs are reduced to the scalings together. Hence we have to resort to a
finitely many inequalities by picking a finite number D-K-\\ke iteration scheme that proceeds as follows:
1082
Copyright© 1997, American Institute of Aeronautics and Astronautics, Inc.
Start with the uncertainty scalings Qi(p) = — /, To further speed up the computation, we finally re-
Ri(p) = I, Si(p) = 0 and iterate the following mark that we performed all calculations after elimi-
two steps until the performance level cannot be im- nating the transformed controller parameters along
proved: standard lines by using the projection lemma, and
by using scalings with S(p) = 0 1. If columns
1. Fix the scalings and minimize 7 over of Ky and Ku form bases of the the nullspaces of
X (p), Y(p) and the transformed controller [ B^ 0 ET ] and [ C D 0 ] respectively, the
parameters K(p,p), L(p), M(p), N(p). reduced synthesis inequalities then read as
2. Fix the controller parameters K(p,p), L(p),
X I
M(p), AT(p) and minimize the performance level (10)
I Y
7 over X(p) and the uncertainty scalings Qi(p),
Downloaded by PRINCETON UNIV. LIBRARY on September 30, 2014 | http://arc.aiaa.org | DOI: 10.2514/6.1997-3641
• If X(p) is parameter dependent and Y is con- 3. For fixed controller parameters, minimize 7 over
stant, choose U(p) = 7 — X(p)Y and V = I. the scalings Q and R using (6). Go to step 2.
Taking derivatives in (8) reveals
(9)
The missile control problem
such that the terms in (7) that depend on p
indeed drop out. For the application we have chosen a missile bench-
• If X is constant and Y(p) is parameter depen- mark problem that has been extensively studied in
11
dent, choose U = I and V(p) = I - Y(p)X. > > 19 > 8 and is particularly suited for address-
This implies X'(p,p) = 0, U'(p,p) = 0 such ing gain scheduling as well as robustness issues. The
that, again, the variable p disappears in (7). problem is to design a longitudinal autopilot for a
tail-fin controlled missile providing normal accelera-
Note that this restriction to constant X(p) or Y(p) tion tracking over a large range of speed and angle-
certainly introduces conservatism, with the benefit of-attack. In order to arrive at a design model (in
of a simpler controller implementation. section missile control strategy), the exact problem
1083
Copyright© 1997, American Institute of Aeronautics and Astronautics, Inc.
specifications are first given in section performance These coefficients are valid for the missile traveling
description. These are based on the missile model between Mach = 2 and Mach = 4 at an altitude of
which is defined in section missile model. 20,000 ft. Typical maneuvers for this missile result
in angle-of-attack values ranging between —20 and
The missile model +20 degrees. Hence the approximation cos(a) « 1
The non-linear state equations of the missile are is legitimate. Then (13) simplifies to
1084
Copyright© 1997, American Institute of Aeronautics and Astronautics, Inc.
tail-fin deflection rate should not exceed 25 deg /s frequency gain of Wperf is 300 to bound the track-
per commanded g-level. ing error by 0.33 %. The high frequency gain is
As very strong simplifications in the missile mod- chosen to be 0.5 in order to limit the overshoot to
eling have been made, we take the robustness is- be less than 5 %. To reflect the tail-fin deflection
sue originating from the uncertainty in the aerody- and deflection rate limits of 20 degrees and 25 de-
namic coefficients Cn and Cm into account. The un- grees/s per g respectively, the filters Wg = ^ and
certainity levels considered are ACn = ±10 % and W$ = •%£ have been chosen accordingly. Finally,
A<7m = ±25 %. noise filters Wnl = 0.001, and Wn2 = 0.001 are used
The controller provides fin commands Sc that are to reflect the measurement imperfections in pitch
processed through second order actuator dynamics rate and normal acceleration.
2
given b y G^s = ,aa+2C , tta a > with natural fre-
quency a) = 150 rad/s and damping C = 0.7. To Results
avoid exciting unmodeled high frequency dynam-
ics, the multiplicative input uncertainty weighting
is placed at the actuator. The design of the LPV controller
Control strategy In this section we will use the synthesis LMIs (6)
as derived previously. Solving the LMIs is done
To realize the specifications over the prescribed op- via basis functions and gridding of the parameter
erating range, the missile dynamics are reformu- space. For the missile control problem, the Mach
lated into an uncertain parameter varying system number M is the remaining parameter for schedul-
representation. The LPV system with parameter ing. The angle-of-attack a, the aerodynamic coeffi-
p = (a, M) has now an uncertain part arising from cients Cn and Cm and the actuator were all assumed
the perturbations in the aerodynamic coefficients. A uncertain. As the Mach number M can be placed
further strategy, which might not be the best one is in an LFT linearly, we choose a basis function as
to view the angle-of-attack in the parameter vector p /i(M) = M (see also 1). For the function X(M),
as uncertain. One one side we chose this strategy to e.g., we thus have
illustrate the method and furthermore because the
angle of attack a. is not measurable in the problem
specification. Another approach would be to use a
as a schedule parameter and and use an observer The other functions depending on the parameter M
to implement in the schedule as was shown in . are Y(M), Q(M), K(M), L(M), M(M), N(M),
However, the resulting uncertainty model would not and have the same structure as X(M).
be anymore challengeing enough to test the effect of The grid of the parameter set M € [2, 4] consisted
scaling. Hence, the angle-of-attack a and the uncer- of five points equally spaced between Mach = 2
1085
Copyright© 1997, American Institute of Aeronautics and Astronautics, Inc.
ings are set to unity. Once convergence is achieved, -15 g causes a 3 % overshoot violation. As a remedy
the large LMIs (6) are solved for the last scaling Q one could try to redesign the weightings. A possible
that was found in the iteration. choice to enhance damping of the acceleration re-
In the section on the analysis of uncertain LPV sys- sponse is to increase the weight on the fin rate filter
tems it was explained that choosing either X or Y Wg since maximum fin rate is by far not reached in
constant led to a controller that does not need a mea- the non-linear simulation. Also the performance fil-
surement of the parameter rate. For the missile con- ter could be adjusted to further punish the overshoot
trol problem, both options were tested. The scheme (increase high frequency gain of the filter). However,
using X constant and Y parameter dependent con- we left the filters W\d and WpeT{ the same as in
verged faster than the one where X was parameter in order to be able to compare the results. The
dependent and Y constant. In 14 steps a 7-value LPV controller synthesized here has a larger over-
of 2.50 was reached whereas the other option only shoot, but it is faster than the one in 19. Moreover
reached the value of 3.87 after 20 iteration steps. In the LPV controller of shows angle-of-attack ex-
figure 3 the achieved 7-value is set out against the ursion beyond —20 degrees while also commanding
number of iteration steps for both processes. four times more fin deflection rate during the —45 g
command. Figure 5 and 6 shows that our controller
The final controller synthesis was carried out on the remains within the limits which might be interpreted
choice (Xo, Y(M)) where 7 = 2.5 was achieved. The as a rather conservative design in the sense that it
scaling matrix Q of the last iteration is used to syn- doesn't use the available fin deflection to enhance
thesize the controller using the full LMIs (6). The performance. It should be noted that the missile in
achieved 7-value was 2.51 and the test on a denser 19
runs along a slightly different Mach trajectory.
parameter grid (with twice the density) gave 7 val- Finally, to demonstrate the robustness properties,
ues between 2.22 and 2.45. The achieved perfor- figure 8 shows the acceleration command responses
mance level given by the propsed scheme is com- for all combinations of the aerodynamic uncertain-
parable to results given in ^ (7 = 3.13) and 8 ties. As can be seen from the figure, overshoot in
(7 = 3.855) with the difference that in our design pa- the 45 g step is the most sensitive to uncertainties,
rameter rate boundness and uncertainties are taken while the other performance characteristics seem to
into account. Since the angle of attack was not behave well.
measured but viewed as uncertain it increases the
achievable 7 level. On the other hand the parameter
rate information reduces the achievable L2 induced Conclusions
norm, therefore compensating conservatism intro-
duced by uncertainty effects and the fact that less We have shown a controller synthesis technique for
sensor information has been used (measurements of Linear Parameter Varying (LPV) that takes the
p and a). boundedness of the parameter variation rates into
1086
Copyright© 1997, American Institute of Aeronautics and Astronautics, Inc.
account. This technique gives guaranteed stabil- [6] Becker, G. (1996). Additional results on pa-
ity and performance levels. Moreover, robustness rameter dependent controllers for LPV systems,
against uncertainties has been incorporated via the IFAC World Congress, vol. G, pp. 351-356,
use of scalings. Using basis functions and grid- 1996.
ding, the synthesis problem is reduced to an iter-
ation of solving finitely many linear matrix inequal- [7] Gahinet, P., A. Nemirovski, A. Laub, and
ities. The method was applied and tested on a mis- M. Chilali (1994). LMI Control Toolbox. The
sile benchmark problem. The non-linear simulations Mathworks Inc.
have shown that the proposed method is successful
in achieving the desired performance and robustness [8] Helmersson, A. (1995). Methods for robust gain-
goals. Further research should be directed towards scheduling. PhD thesis, Linkoping University,
Downloaded by PRINCETON UNIV. LIBRARY on September 30, 2014 | http://arc.aiaa.org | DOI: 10.2514/6.1997-3641
1087
Copyright© 1997, American Institute of Aeronautics and Astronautics, Inc.
0 2 4 6 6 10 12 14 16 18 20
number of iterations
£i
3 [lif—1
\3 Li—u
1088
Copyright© 1997, American Institute of Aeronautics and Astronautics, Inc.
4
Downloaded by PRINCETON UNIV. LIBRARY on September 30, 2014 | http://arc.aiaa.org | DOI: 10.2514/6.1997-3641
3.8
3.6
3.4
L3.2
i 3
| 28
2.6
2.4
2.2
1089
American Institute of Aeronautics and Astronautics