Sunteți pe pagina 1din 5

Procssdlngr of the

American Control Conlennce


Seanle, Washington June ?OS

An LFT Approach to Autopilot Design for Missiles


Yun Huang Wei-Min Lu
Electrical Engineering 116-81
California Institute of Technology
Pasadena, CA 91125
yunhuang@hot.caltech.edu wmlu@hot.caltech.edu

Abstract sponding LMI’s.


In this paper, a tailfin-controlled missile system is described in
terms of a linear fractional transformation (LFT) on an uncer- 2. Background Theory
tainty/frequency block. The aerodynamical nonlinearities are In this section, some material on analysis and synthesis of uncer-
taken as deviations to the nominal linear model and are treated tain linear systems under LFT framework is reviewed. Refer to
as uncertainties. A state feedback robust controller is obtained 18, 3, 61 for more details. The notationused here is fairly standard
by solving some linear matrix inequalities (LMI’s). The simula and is taken from 131.
tion results show that the designed autopilot is able to track the
acceleration command, hence achieving the robust performance 2.1. LFT’s a n d Uncertain Linear Systems
requirements. LFTs are a class of general linear feedback loops and can be pic-

mu
torially depicted as follows:
1. Introduction
Realistic models of physical systems are often nonlinear. The dy-
.
namics of such systems change with the operating conditions. A
common approach to this situation is gain scheduling. The main
idea is to break the design process into two steps. First, design
local linear controllers at several operating conditions. Second, Y G U
interpolate the gains of the local controllers at intermediate op-
erating conditions to get a global controller. There have been Figure 1: Upper LFT
several papers, e.g. [13, 111, which describe this gain scheduling
framework for nonlinear systems. Since it has the advantage of
incorporating linear robust control methodologies into nonlinear The resulting input/output map can be represented as y =
control design, gain scheduling has been practiced in a variety of Fu(G, A)u, where the (upper) LFT Fu(G, A) is defined as:
engineering applications, such as flight control and process con-
trol. Despite its popularity, gain scheduling remains an ad hoc Fu(G,A) :=Gzz + G z i A ( I - G i i ) A - ’ G i z (1)
method because of its local nature. The overall design comes with
no guarantee on the robustness, performance, or even the nominal
stability. Such potential hazards are examined in [14].
Recently, an LFT approach was proposed to handle the class of
with G partitioned conformably as
[ z:: z:; 1.
A large class of linear systems can be described in terms of LFTs
linear systems whose uncertainties enter in an LFT fashion [3, 71. on some specifiedfrequency/uncertainty structures. An uncertain
For such a system, the plant is described as an LFT on some block discrete-time linear system is discussed in this section. Let P x t
structure A which contains both transform variables and uncer- denote the time-varying causal operators: i?i(Zt) + !;(Z+) and
tainties, and the controller is allowed to depend on the same block A p be the uncertainty structure defined as:
structure. Unlike p synthesis which gives a single linear time in-
variant controller using D-K iteration, this LFT method would
analytically yield an output feedback controller which, in general,
is dynamically dependent on the uncertainty block. In this sense,
it can be thought of as dynamically scheduled. However, in many Then the discrete-time system G with the structured uncertainty
practical cases, the controllers have no access to the information A p E &p can be put into the standard LFT form
about the uncertainties during their operations. So, it is of in-
terest to study the condition under which the state feedback can
achieve the same performance level as the “full information” (FI)
controller. It turns out that we only need to check the feasibility
of a pair of LMI’s, which falls into the scope of tractable convex
I . I

optimization problems. Moreover, a state feedback controller can


W
be analytically constructed from the solutions to the LMI’s.
In this paper, the LFT machinery is employed to deal with the ro-
bust performancesynthesis problem for a tailfin-controlledmissile.
There have been several papers focusing on this autopilot design
problem based on p synthesis [9] and gain scheduling 115, 10, 51.
We propose a different approach here: the nonlinear missile air- Figure 2: General LFT 5 stem
frame is treated as a quasi-linear parameter varying (LPV) system
whose state-space matrices depend on the endogenous variables. where z-l is the delay operator and
The dynamic deviationsfrom the nominal linear model are consid-
ered as uncertainties in our design. After the robust performance
problem is converted to the stabilization in the LFT setting, a B
static state feedback controller is obtained by solving the come- D

2990
where(A,B,C,D) E W n x n ~ W " x P ~ R q x n ~ W * withnXP = no+m. Note that in the above theorem, the feedback actually employed
Now, if we define A := D i a g [ z - l l , A p ] , the system G can be the "full information", which means both the real state of the
represented as an LFT on A: system, I,and the signal 20 associated with the uncertainty block
A p are fed back through the input U. In what follows, we shall
(3) refer to 5 = [ I w ] as the "extended state" of the LFT model
shown in Figure 2.
2.2. Control of Uncertain Linear Systems
To examine the robust stability and robust performanceof the sys- 3. State Feedback vs FI Feedback
tems with time-varying uncertainties, the admissible uncertainty In this section, we will derive the condition under which state
set is defined as: feedbeckcan achieve the same robust performance level as the full
information control given by Theorem 2.4.
BAp := {Ap E A P : ~IAPIIL,5 1)
where I .I L, is the e2-gain of an operator. The scaling matrix set Lemma 3.1 (Parrott's Theorem) Assume A, B , c and x are
2, with respect to A is defined by real matrices of appropriate dimensions, then
v .= ( D E cmXm: ID AD-^ [Ila l l ~> ,O , AE A)
= ~ l ~ ,I)

Definition 2.1 The system (3) is said to be robustly stable if Suppose the above infimum i s yo. T h e n f o r any y > yo, the
the system is stable f o r each fixed Ap E BAP. I t has robust
performance if the system is stable and has &-gain < 1 for each
A p E BAP.
solutions X such that @(
X B A [ 1)
< 7 are characterized b y

X = -YA*Z+r(Z- YY')+W(Z- Z*Z)+ (10)


The following result can be derived from [12,41, which gives the
the robust stability tests.
where Y and Z S O h e the matrix equations Y(y2Z - A*A)+ = B
Theorem 2.2 Consider system (3), the following are equivalent. -
and (7'1 AA*)*Z = C, with w
being an arbitrary contractdon
( i ) The system is robust1y stable.
(@(W) < 1).
(ii) inf @(DAD-') <1 Lemma 3.2 Assume (A,B) E Wnx" x WnxP and A and B are
DEQ
partitioned as in (2). Define K := {[ K OpXm ] : K E
(iii) There exists a P E V such that
R p X n ~ }then
,
APA* - P < 0. (4)
inf @(A BF) + < 1 + max{@(BfA),@[ ]} < 1
The test for robust performance, whichrequiresIIFu(M,A)llc, < 1 FEK
for all Ap E BAP, can be conducted similarly to that for the
robust stability. Without loss of generality, it is assumed that the and F = [ F1 0 ] with
number of inputs and the number of outputs are the same, i.e.
p = q. Define a new augmented block structure A N :

AN := {Diag[A, A,] : A E 6,A, E @ P x P }


and let V N denotes the scaling matrix set with respect to AN,
then we have the following result.

Theorem 2.3 System (3) has robust performance if and only if +


makes @ ( A B F ) < 1.
there exists a P E 'DN which satisfies
MPM* - P < 0. (5) Proof: Let B A = [ i:i ] be the orthogonal complement of

[:
:]
Next, we discuss the synthesis problem. Consider system (3),and B which is defined before and partitioned conformably with B in
further assume that B has full column rank, i.e. rank(B) = p 5 (2), and Bo = E WnXP be such that [ Bo BA ] is
n. Let BA E Wnx("-P) be the orthogonal complement of B, i.e.
unitary. Then
B*BI = 0 and [ B BI ] is invertible. The following theorem
is from [3]. inf a(A+BF)
FEK
Theorem 2.4 FOT system (3), the following statements are
equivalent.
( i ) There exists a static feedback matrix F such that A + B F is
robustly stable with respect to the block structure A.
(ii) There exists a P E 'D such that
B;(APA* - P)B* < 0.
(iii) There exists a P E 2, such that
APA* - P- BB* < 0. (7)
Moreover, if P E V satisfies any of the above LMI's, then a sta-
bilizing static feedback matrix can be chosen as
The last two equationsare obtainedby applying Parrott's theorem.
F = -(B*P-~B)-~B*P-~A. (8) +
Moreover, the F1 that achieves @(A BF) < 1 should satisfy

2991
( 1 4 ) . The expression for the feedback matrix K can be derived
when applying (11).
with
The above theorem conveys the idea that the extra online knowl-
edge about the uncertaintiessometimes would not help achieve the
robust s t a b i l i t y / p e r f o r e goal. This is not surprising since the
controllersought aims to handle all the uncertaintiesin the admis-
sible set, not just the particular one. If a FI feedback which uses
the extended state can be constructed by Theorem 2.4, and any
solution to LMI (6) or (7) satisfies LMI ( 1 4 ) , then a robustly sta-
If Bo is chosen as Bo = B(B*B)-*,
then Fl given in the lemma bilizing state feedback can also be found. It should also be noted
can be solved from ( 1 2 ) . that the condition under which the existence of an FI feedback is
equivalent to that of a state feedback in the LFT setting reduces
This lemma is the key for dealing with the state feedback prob- to the feasibility problem for LMIs. With the recently released
lem, which only uses part of the information employed by the FI software package LMI-Lab [ 2 ] , the computation that is needed in
controller. Theorem 3.3 can easily be implemented.

Theorem 3.3 Consider system (3), there exists a state feedback 4. Missile Dynamics Description
U = K x such that the closed loop system is Tobwtly stable if and
The missile-aifframedynamics consideredhere are taken from [91.
only if there ezists a positive definite matrix P = [2 2: ]E These dynamics are representativeof a missile travelling at Mach 3
at an altitude of 20,000 feet. The nonlineardynamics are described
V which is partitioned conformably with A an ( 2 ) satisfying the in the following state equations:
following LMZ's:

fj MY
= -f
IY
where
Moreover, the feedback matrix K can be chosen as cy = angle of attack ( d e g )
q = pitch rate ( d e g / s e c )
= -wlp-~( [ -412
-422
] p2!i(1- [
p2+ -412
A22
] Ip-1 [ A12
A22
] pi)-ipj g
M
=
=
accelerationof gravity ( 3 2 . 2 f t / s e c 2 )
mass (13.98sZugs)
V = speed (3109.3f t l s e c )

1 ,Z
f
Fz
=
=
=
pitch moment of inertia (182.5s1ug.f t 2 )
radians-to-degrees conversion ( 1 8 0 / n )
CZ(cu,q)QS = normal force (16s)
with W1 and W2 given below My = Cy(Cu,q)Q§d = pitch moment ( f t .Ibs)
Q = dynamic pressure (6132.81bs/ft2)
Wl = ( B * P - I B ) - * B * Wz = ( B ; P - l B ~ ) - * B ; S = reference area ( 0 . 4 4 f t 2 )
d = diameter (0.75ft)
Proof: By Theorem 2 . 2 , the closedloop system A+B [ K 0 ] The n o d force and pitch moment aerodynamicscoefficientsare
is robustly stable iff approximated by

inf Z ( D ( A + B [ K o Cz 4z(a)+bz6
'
=
DE'D
inf Z ( D A D - ~+ DB [ K
1p-I)

o ] 0-l)
CY = 4Y(CY)+bY6
D€D where
6 = fin deflection ( d e s )
Let VI = ( B ; ( D * D ) - l B , ) ' * B ; O - l . It is easy to check that ~Z(CY) =
0.000103~~~ - 0 . 0 0 9 4 5 ~ ~ 1 ~0.17001
~1-
V;VA = I and VI(DB) = 0. By Lemma 3.2, we have &-(a) 0 . 0 0 0 2 1 5 ~-~0.0195alal-
= ~ 0 . 0 5 1 ~ ~ These
bZ
-
- -0.034
[K
inf ~ ( D ( A + B o Ip-1) <1 by -
- -0.206
DE?) approximations are assumed to be valid for cu in the range from
e inf (max{Z(ViDAD-'),B(D
D€Q
[ 2:; ] DT1)}) < 1 -20 to +20 degrees.
In addition, the missile tailfin actuator is modeled by a second
where D = [2 Dz ].Take
order linear transfer function:

-= 4
P = (D*D)-' = [ 0
0
Ga(s) =
&,(a) s2 +1 +
. 4 ~w8~ ~
(16)

then P E V and P = P' > 0. It can be shown that where


6, = commanded fin deflection (deg)
B ( V ; D A D - ~ )< 1 a B;(APA* - P)B, < o wa = actuator bandwidth (150rad/sec)

a(r, [ 2; ] ~ ; l ) <1 0 [ 2:: ] p2 [ A ; ~ ] -p <0 The autopilot is required to control the acceleration (expressedin
g's)

+ ]
So, infDED a ( D ( A B [ K 0 ) D - l ) < 1 iff we can find some Fz
matrix P = P* > 0 defined above which satisfies LMIs (13) and r)z =-
Mg

P. 3
2992
via the commanded fin deflection,6,. The performance objec- 5.2. Missile Model Treatment
tive is to track the accelerationstep command with a steady state The missile airframe dynamics fall into the class (18). Its state
accuracy better than 0.5% and a time constant no greater than 0.2 equations (15) can be rewritten into the form (18:
seconds.

5. Autopilot Design
5.1. General Considerations
Consider a special class of affine nonlinear systems:
j. = F(t)z+G(t)u where fi , fi, g1 and h are functions solely of a,and continuous
y = H(Z)X+L(Z)U (18) and bounded on X = { la1 5 20, a E W ), while g2 and 1 are some
constants.
where t E W" is the state vector, U E RP and y E R* are input For the angle of attack of 0,flQ,and f 2 0 degrees respectively,
and output vectors, respectively. Suppose the system evolves on a the state space matrices are as follows:
subset of the the state space, say X C W", and the matrix-valued
functions F ,GI H and L are continuousand bounded on X. We
then choose matrices A0 ,Bo, CO, DO which well approximate the
matrix-valuedfunctionsF(z), G(t), H ( z ) ,L ( t ) on X ;for example,
we can choose a fixed state xo E X and define H(0) [ -1.01 0 ], L(0) = -0.203;

A0 = F ( z 0 ) Bo = G(t0) CO = H(to) Do = L ( z o ) (19)

Let f?"'" denote the set of time-varying causal operators from


t ; [ O , c o ) to t ; [ Q , o o ) , and Il.llLCl be the &-gain. The norm-
bounded set B A C is defined below:

BAC := {A E LnX": IIAllLc,5: 1).


H(f20) = [ -1.91 0 1, K(f20) = -0.203.
We then can rewrite the state space representationof (18)into the
following form: By examining the state space matrices over the angle of attack
range from -20 to +20 degrees, we see that the dynamics at a =
j. +
= (Ao AIAAAz(z))z+ (Bo +BIAB(z)Bz)u f10 can serve as a reasonable nominal model. Moreover, it is
Y +
= (CO ciAc(z)Cz)a:+ (Do DIAD(Z)DZ)U + (20) observed that F(a)(2,1)varies the most with a among all the
varying terms F(a)(l,l),F(a)(2,1), G(a)(l) and H(a)(2,1). On
where AA, AB, Ac, and AD all belong to BAC, and A,'s, B,'s,
Cl's, and Dl's are some real scaling matrices. If those A's are the other hand, from the physical considerations, F ( 2 , l ) is far
more significant than other terms because it relates directly to

dynamics [MI,
viewed as the time varying deviations from the nominal linear
(18) can be considered as a quasi-LPV
model because those time-variations (A's) are unknown a priori
the rotational stability of the rigid-body airframe. Under these
considerations,we will only model the variationin term F(a)(2,1)
in our autopilot design.
but may be measured/estimated upon operation. It is remarked The approximatequasi-LPV model of the missile airframeis shown
that the quasi-LPV model ( 2 0 ) is very general. In addition to below:
nonlinearities,uncertainties and some of the unmodeled dynamics
can also be included in this model at the same time. [p] = [ -0.89
-142.6(1+ k i A i ) ] [ 9] [+
-0.119
]
By pulling out those uncertainty blocks A's, system (20) can be
put into the LFT form F u ( M , [ fp ])with
qz = [ -1.52 0 ] [;I+ - 0.2036
The multiplicativefactor (1 IC1 A i ) captures the aerodynamical
uncertainty, with A1 allowed to vary on [-1, +1] and the normal-
izing scale factor k l = 1.25.
The diagram below illustrates the design model for the actua-
tor where A2 represents the uncertainty in the actuator gain and
phase characteristics with ((A2(IL2< 1. For a normalizing scale
and factor of IC2 = 0.6, this structure accommodates as much as 35
degrees phase uncertainty and gain deviation from approximately
AP =D~~~[AA,AB,AC,ADI. 0.6 to 2.5.
Note that A p belongs to some norm-bounded set
BA: := (Diag[Al, A2, A3, A,} : A; E BAC]
and the matrix M is not unique.
We then discretize the linear time invariant system ~ . , ( M , S - ~ I )
with an appropriate sampling rate, which yields F , ( M d , z-l I)
for some r e d matrix Md, the corresponding discretized system of
( 2 0 ) can then be described as F u ( M d , [ ]),andA$
Figure 3: The Actuator Model

belongs to the following uncertainty set To accomplish the performance specifications, i.e. to track the
step command qc with an accuracy of 0.5% and a time constant of
0.2 seconds, a frequency dependent sensitivity weighting function
W s ( s )is applied to the tracking error signal qc - qz:
Thus, the results presented in the previous section can be applied
to cope with the robust stability and performance problem for the
WS(S)= -.15s + 200
system (20). 40s + 1

P. 4
2993
Our performance objective becomes t o keep the &-gain of the
mapping from t h e commanded acceleration qc to the weighted
error e less than 1, i.e. llGvcweIILc, < 1.
The missile dynamics are first cascaded with the actuator dynam-
ics. T h e n t h e sensitivity weighting is augmented onto it t o form
the final design model whose states a r e z = (a q x a ka e)= with 0 0.1 0.2 0.3 0.4 0.5
nme(rec)
0.6 0.7 0.8 08 1

za and k, being the actuator s t a t e variables, inputs a r e ( m6 ~ ) ~


and o u t p u t is e. Let t h e sampling period be T,= 0.001 seconds. 1.6, 1

T h e L F T form 3 u ( M d r A )of the design model can b e obtained


by t h e similar treatment on the general affine nonlinear system in
t h e previous subsection, where

A = Diag[z-l15,Al,Az].
0 0.1 02 0.3 04 05
n m e (sec)
Angle-4-eltack

06 0.7 08 0.0
i
1

5.3. Robust P e r f o r m a n c e Controller Design


Our design objective is t o keep llGvccellt, < 1 for all A p = Figure 4: S t e p Responses of T h e Closed-loop System W i t h A1 =
[ 2 i2] E B a p . As mentioned in Section 2, this perfor- A2 = 0
mance test c a n b e converted into t h e robust stability test by in-
troducing an auxiliary scalar block A, E C, which results in t h e Daahsd llne Command
Solld llne Response
augmented block structure

8 -20
\
v
Partition Md into [ A B ]
where A is a square matrix corre-
4 0
0 0.5 1 1.6 2 25
n m e (eec)
3 3.5 4 4.5 5

sponding t o A, and B is a column vector corresponding t o the


1
control input 6c. T h e n this missile system has the requiredperfor- 1
mance if and only if A is robustly stable with respect t o t h e aug-
mented block structure AN. So, we want t o build a s t a t e feedback
+
controller 6c = K z such that A B K is robustly stable with re-
spect t o A N . By Theorem 3.3, there exists a static s t a t e feedback
such t h a t t h e closed loop system has t h e robust performance if and -1
2

[?
0 05 1 15 2.5 3 3.5 4 45 5
only if there exists a positive definite matrix P = nme(mc)
:
2
]
with PI E C 5 x 5 and P2 E ID = ( D i a g [ d l , d z , d s ] : d l , d 2 , d 3 E C),
which satisfies LMI's (13) and (14).
T h e software package LMI-Lab [2] was used to solve the pair of
LMI's. It is shown that they are indeed feasible, which means
that s t a t e feedback can d o just as well as FI control for achieving
t h e required robust performance in this case. And such a state _------
feedback matrix K is computed from t h e solution t o t h e LMI's. 40f! 0'5 1:5 4 2:5
T h e (sec)
4 315 415 b
To validate our design, t h e missile dynamics are simulated along
with t h e state feedback controller. Figure 4 shows the s t e p re- Figure 5: Square Wave Responses of The Closed-loop System
sponses of the closed-loop system when t h e uncertainties A1 and
A2 a r e frozen to zero. T h e nonlinear simulation results are il-
lustrated i n Figure 5 , where t h e command signal qc is selected t o
achieve t h e angle-of-attackin the range of interest and the actuator [7l Packard, A, K. Zhou, P. Pandey, J. Leonhardson, and G. Balas
(1992), Optimal Constant 1 / 0 Similarity Scaling for Full-Information
uncertainty A2 is randomly set. For both cases, t h e closed-loop and State-Feedback Problems, Systems and Control Letters, Vo1.19,
system behaves as desired and satisfies t h e s t a t e d performance pp.271-280.
goals. [8] Redheffer, R.M. (1960), On a Certain Linear Fractional Transfor-
mation, J . Math. Phys., Vo1.39, pp.269-286.
6. Acknowledgements [9] Reichert, R.T. (1990), Robust Autopilot Design Using p-Synthesis,
Proc. 1990 ACC, 1313.2368-2373.
T h e authors would like t o thank Prof. John Doyle for t h e helpful
discussions.
[lo] Reichert, R.T. (1992), Dynamic Scheduling of Modern-Robust-
Control Autopilot Designs for Missiles, IEEE Control Systems Mag-
azine, V01.12,NOS, pp.3542.
References [Ill Rugh, W.J (1991), Analytical Framework for Gain Scheduling,
[1] Doyle, J.C. (1985), Structured Uncertainty in Control System De- IEEE Control Systems Magazine, Vol.11 , No 1, pp.79-84.
sign, Proc. 1985 IEEE CDC, pp.260-265. [I21 Shamma, J.S. (1994), Robust Stability with Time-Varying Struc-
[2] Gahinet, P, A. Nemmirovskii, A.J. Laub and M. Chilali (19941, tured Uncertainty, IEEE Trans. AC, Vo1.39, pp.714-724.
The LMI Control Toolbos, The MathWorks Inc..
[13] Shamma, J.S. and M. Athans (1990), Analysis of Gain Scheduled
[3] Lu, W.M., K. Zhou and J.C. Doyle (1991), Stabilization of LFT Control for Nonlinear Plants, IEEE Trans. AC., Vo1.35, No.8, pp.898-
Systems, Proc. 1991 IEEE CDC, Brighton, England, pp.1239-1244. 907.
[4] Megretski, A. and S. Treil (1993), Power Distribution Inequali- [14] Shamma, J.S. and M. Athans (1992), Gain Scheduling: Potential
ties in Optimization and Robustness of Uncertain Systems, J . Math. Hazards and Possible Remedies, IEEE Control Systems Magazine,
Systems, Est., and Contr., Vo1.3, pp.301-319. V01.12, N0.3, pp.101-107.
[5] Nichols, R.A., R.T. Reichert, and W.J. Rugh (1993), Gain-
Scheduling for 'HOO Controllers: A Flight Control Example, IEEE [I51 Shamma, J.S. and J.R Cloutier (1992), A Linear Parameter-
Trans. on CST, Vol.1, pp.69-79. Varying Approach to Gain Scheduled Missile Autopilot Design, Proc.
[6] Packard, A. and J.C. Doyle (1993), The Complex Structured Sin- 1992 ACC, pp 1317-1321.
gular Value, Automatica, Vo1.29, pp.71-109.

2994

S-ar putea să vă placă și