OPTIMAL CONTROL APPLICATIONS & METHODS,VOL.

7, 1-17 (1986)

SINGULAR PERTURBATION MODELLING AND DESIGN

TECHNIQUES APPLIED TO JET ENGINE CONTROL

GEORGE P. SYRCOS
Wilkes College, Engineering Department, Wilkes Barre. Pennsylvania, U.S.
A.

AND

P. SANNUTI
Rutgers State University, Department of Electrical Engineering, P. 0. Box 909, Piscataway, New Jersey, U.S.A.

SUMMARY
This paper presents an application of recent singular perturbation modelling techniques for design and
control of the F-100 jet engine. The required feedback gains are calculated using a Chandrasekhar type
of algorithm.
K E Y WORDS Singular perturbations Chandrasekhar algorithm Linear-quadratic regulators
Turbofan jet engine control

INTRODUCTION
This paper presents an application of singularly perturbed linear-quadratic-regulator (LQR)
techniques for the control design of the Pratt and Whitney F-100-PW-100 afterburning
turbofan engine. '-*Both the LQR and singular perturbation issues will now be briefly
addressed. In order to apply multivariable control design tools such as LQR theory, one must
first resolve two aspects: (a) obtain a linearized model which includes all significant engine
dynamics about a particular operating point (the complete set of these operating points should
cover the whole flight envelope), and (b) determine the performance requirements of the engine,
including state and control constraints by appropriate choice of the performance index
weighting matrices Q and R. Once these aspects have been resolved, the system can be
compactly written as
2 = ~ j + iBU, R(O) = j i ,
J=

B=CZ
2
^x(nTF2(T)
2 sT
+ - (BTQ%+ UTRU) d t
0

where 2,U and are n-,m-and p-dimensional state, control and output vectors, respectively.
Q, F are positive semi-definite, and R is positive definite. As is well known, the optimal control
for system (1) is given by the linear feedback law
U(t) = - R-'BTP(t)Z(t) = -K ( t ) X ( t ) (2)

0143-2087/86/01 O001- 17s08.50 Received 21 December 1983

0 1986 by John Wiley & Sons, Ltd. Revised 24 September 1984
2 C . P. SYRCOS AND P. SANNUTI

where
P = -ATP - PA + PBR-'BTP - Q, P(7') = F (3)
Even though equations (2), (3) apply to any order model, their application is limited to low-
order models owing to computational and/or dimensionality constraints. Whenever there is an
interaction between slow and fast dynamic phenomena, the problem is further compounded by
stiffness of the model equations. Singular perturbation theory3 provides a method to alleviate
both dimensionality and stiffness difficulties. This is accomplished by separately analysing and
controlling the slow (reduced) and fast (boundary layer) phenomena in a ~ y s t e m . By
~ - reducing
~
the model order, singular perturbation methods reduce the required computation for analysis
as well as synthesis of systems.
Another method, 7-9 other than reducing the model order, has been introduced to simplify
the feedback gain calculation in an LQR problem. In this method the matrix Riccati differential
equation (3) is replaced by a set of coupled non-linear ordinary differential equations of lower
dimension. These are commonly referred to as generalized X- Y functions or Chandrasekhar
equations. The optimal feedback gain (2) is given by the Chandrasekhar system of equations
(CW
Y = - (A - B K ) ~ YY, (7') = L~
(4)
K = R-'BTYSYT,K(T) = R-'BTP(7')
where the symmetric matrix P(7') is factorized as
P(T) = L~SLOT
S is the s x s signature matrix of P( 7') with s = rank [P(T)] , and LOis an n x s matrix. The solu-
tion of CSE (4) required for the feedback gain calculations of the control problem (1) involves
n(m + s) equations. Consequently, the CSE (4) represents a smaller number of equations than
+
(3) if (m s) < (n + 1)/2.
For systems with both fast and slow dynamics that can be represented in a singularly
perturbed form (singularly perturbed LQR problem), it may be possible to exploit the rank
properties of P(7') and generate a potentially fast algorithm of Chandrasekhar type for the
feedback gain calculations. lo The singularly perturbed LQR problem is defined in (1)-(3),
where

Here, G > 0 is a small parameter and X, Z are nl-, nz-dimensional state vectors. Then the
optimal control (2) is given by
U ( t , E ) = Ki(t, &)X(f)+ Kz(f, E ) Z ( ~ ) (6)
The corresponding CSE (4) is transformed t o a classical non-linear singularly perturbed
problem by appropriate scaling of the variable Y

Y = (YT, EY?)T (7)

Under the following standard hypothesis, a valid asymptotic expansion in powers of E can be
 JET ENGINE CONTROL 3

constructed for the feedback gains K l ( t , c ) , Kz(1, 8 ) .

H I : The matrices Q, F are positive semi-definite, and R is positive definite.
H2: The pair [ Az2, Bz] is controllable (boundary-layer controllability).
H3: The pair [ A22, CZ] is observable (boundary-layer observability).
The asymptotic expansions of the variables take the form
Yl(t, E ) =Yl(t, c ) + cm(7, c )
Yz(t, c ) =Y2(t, c ) n(T, E ) +
Kl(t, c ) = R l ( t , & ) + d ( T ,C)
Kz(f, c) =R2(t, c) +~ ( 7c ),

where T is the stretched-time variable:

7- = ( t - r>/c
-
Y1,P2, K Iand K2 are the outer expansions throughout the open interval 0 c t < T and have
asymptotic power series expansions

The boundary layer correction terms n, cm, cl and cq have asymptotic power series expansions
of the form

The coefficients m, n, I and q are chosen so that (8) will formally satisfy the CSE (4)-(7) and
so that they tend to zero as c tends to minus infinity. The differential equations of the
boundary-layer terms can be easily obtained from

dn
-d7
=c [dY2
dt
dY
d:]"=cT+T

by assuming temporarily that the outer expansion terms (9a) are known and then, as usual,
substituting (9a, 9b) into (10) and equating like powers of c. The differential equations
defining the outer expansion terms can be similarly obtained by substituting (9a) into (4)-(7)
and equating equal powers of c. Although an infinite number of expansion terms (boundary
layer and outer) can be given, for obvious practical limitations only the first two terms are
customarily considered.
In the sequel, we briefly summarize two alternative design procedures. In one procedure, we
first construct the boundary layer correction depending on T and then find the outer expansion
depending on t. In the other procedure, the more conventional reverse sequence is adopted.
4 G . P. SYRCOS AND P. SANNUTI

That is, we first solve the reduced problem and later obtain the boundary layer correction terms
if deemed necessary. Following standard singular-perturbation procedures, we obtain the
lowest-order boundary layer terms of the Y-K variables from

d--
a - 1 - R-'B:n- I ST- I ~
dr
where
a-l(r)=Blo(T)+I-~(r)
8-1(7) =K20(T) + q-1(7)
r-l(r) + m-l(r)
=PIo(T)
along with the initial conditions
n-l(O)=Lz, a-I(O)=Kl(T), 8-1(0)=Kz(T), T - I ( O ) = L I (13)
The system (1 la, b) can be recognized immediately as a Chandrasekhar system for calcula-
tion of the feedback gains of a boundary layer type regulator problem. Thus, under boundary-
layer observability-controllability hypotheses (H and H3), n- 1 and 8- I are exponentially
decreasing functions of r as r-+ - a sand
Iim In- l(r), 8- I ( T ) I = 10,Bzo(t)I
T- --OD

In fact, if [ I + NT(T)]-' exists, an explicit solution of ( l l a , b) is

n - l ( r ) = e - M T T I I +NT(r)]-lL2
and
q - l ( r ) = R-lBfe-MTTII+ NT(r)] -lNe-MT
where
M = AZZ- B~Kzo(T),
N = p30(T) -p30(T)

T(r) = i' 0
-
e-M rSze-M dt and SZ = BZR-IB]

Also, T(r) is alternatively given by

~ ( 7 =) w - e-MTWe-MT
where W satisfies
MTW + WM + S2 = O
Since M has eigenvalues with negative real parts, the equation has a unique solution for W. l 1
 JET ENClNE CONTROL 5

Once the solution of (12,13) is known, the terminal conditions for the reduced system are

The reduced-order CSE is

*I0 = - (A - BRlo)r5!lo
-
K ~ =o R-'BqloSPTo

where
E = [ A Z I- Bzk~o(7")I- T [ A -~ BKZO(T)I
~
A -E ~ B=
A B~ -~ E ~~ B~ ~
Equations (11)-(15) outline a design procedure in which the first terms of the boundary layers
are constructed first and then the first terms of the outer expansion are calculated. However,
most often a designer would like to implement boundary layer corrections to a reduced design
which does not seem to meet performance requirements. In such cases the reduced design
already exists, and the terminal conditions for the variables PIO, Rzo and YIOare
x Z O ( T ) = R-'B;F30(T)
RIo(n
= R - ~ B : I I ~+ R 2 0 ( ~
PlO(T) = L,
where
D,4$1(T) = L,LT and L, is nl x s,, s, Inl

Then, the boundary layer Chandrasekhar equations (1 1) can be solved with initial conditions
I- I (0)= K I (T) - Rio(T), q - l(0) = Kz(T)- KO
(77
n- I (0) = Lz, m -I (0) = LI - L,
where

and Lz is an nz x s b matrix, Sb 5 nz. Thus, we see that the solution of a singularly perturbed
system using a Chandrasekhar algorithm is transformed to three lower-dimensional decoupled
sets of equations ( l l a , b), ( l l c , d) and (15). The dimensionality of each set of equations is
n ( a + rn) for the reduced-order gains, n l ( s , + m ) for the first and ~ Z ( S I+, m ) for the second
boundary layer gains, respectively. Higher-order expansion terms can be similarly obtained if
desired.
Before one can apply any of the available singular perturbation model order reduction
techniques for design and control, one must transform the given problem to the required
singularly perturbed form. Several methods are available. l 2 - I 4 The algorithm l4 used here for
slow/fast state variable decomposition will be summarized next.
The physical state variables of the system are classified as slow or fast states according to
an index. This index, Ri = Rsi/Rri, is the ratio of slow to fast dynamics contained in state X i ,
6 G . P. SYRCOS AND P. SANNUTI

where Rsi is the Euclidean norm of the ith row of M I and Rfi is the Euclidean norm of the ith
row of M2. M I and M2 are the modal matrices corresponding to the slow (small) and fast (large)
eigenvalues, respectively.
Knowledge of the index Ri can be used to reorder the state variables. The first nl state
variables are slow, and the remaining n2 are fast. This slow-fast separation of state variables
allows one to rewrite the given equations in a singularly perturbed form. l4 It is essential to
retain as many physical state variables as possible in the new state description. Experience with
singular pertubation methods has shown that most often a design based on a properly chosen
slow or reduced system is adequate. That is, such designs require only slow variables for feed-
back. This implies that a proper characterization of available physical states into slow and fast
groups is essential in order to minimize the need for estimating fast variables.

APPLICATION
This section contains an application of the singular perturbation modelling algorithm and CSE
to the F-100 jet engine. The model of the engine used here is a 16th-order version linearized
about the operating point of zero altitude, zero Mach number and power lever angle of 83",
which is near maximum non-afterburning power. The motivation for choosing this point comes
from the fact that every engine has to pass through this condition, as for example on take-off.
The engine state variables, control vector, state matrix A, and control matrix B are given in
Appendix I. The eigenvalues of this system are
[ - 577.0, - 175.6, - 59.1, - 50.6, - 47.0, - 38-9, -24.2 - 21 -2, - 17.6 fj4.5,
-15.9, - 6 . 8 k j 1 . 4 , -2.6, -1.8, -0.65)
The eigenvalue locations indicate that it may be possible to separate the physical states into two
groups: the first consists of five slow states and the second of eleven fast states. The selection
indices for n = 5 and n2 = 11 areI4
Rsi = ( 1 *753 1.213 0.078 0.017 0.01 1 0.09 0.098 0.208 0.284 0.125 0.310 0.206 0.126
0*2760.1520-191)
R f i = (0.802 0.594 0.697 0.770 0.004 0.064 0.225 0.205 1.128 0.014 0.504 1.176 0.005
1.055 1.071 1.912)
Ri= (2.1862.043 0.112 0.022 2.852 1.40.435 1.01 0.252 8'620.616 0.175 2.62
0.262 0.142 0.099)
The state selection algorithm results in the following state selection:
slow states = 1x1,X2,XS,
X I OXI31
,
X4, X6,X7,X8, X9,XI', X12, X I 4 9 XIS, X16)
fast states = (X3, -
The indicated choices of slow and fast states are in agreement with what is known from prac-
tical data.' In order to satisfy a sufficient condition that the indicated partition leads to a
singularly perturbed form, Is the states are scaled by a diagonal matrix SX,
S x = (0.1861 0.15628.80.9297 1.096 1.707 0.3703 0 ~ 1 4 0 8 0 ~ 3 1 9 0 * 0 5 9 6 0 ~0-1233
131
0.1145 0,9333 0.0345 4.063 )
Letting E = 0, we obtain the reduced system
% = (Ai I - A12A22A21)X+ (BI - AnA22B2)U
 JET ENGINE CONTROL 7

The eigenvalues of the reduced system state matrix

-A = A11 - A I z A ~ z A ~ I
are
(-6.2445 +j0.7065, -3.0441, - 1.1988, -0.6478)
These closely approximate the five smallest eigenvalues of the full-order system. This fact
further supports the validity of the modelling algorithm. Thus, the 16th-order linear engine
model is transformed to singularly perturbed form.
One way to design the engine controller is by using the known results provided by LQR
theory. The main difficulty lies in choosing appropriate weighting matrices Q and R. One
approachI6 chooses a diagonal Q and R with elements equal to the reciprocals of the squares
of the maximum allowable states and controls. Another approach computes the quadratic
performance index matrices in such a way that the transient system response meets the perfor-
mance requirements. We will use the Q weighting matrix presented by Harvey and Steinla (see
Appendix 11). This choice is based on the design goal of achieving small values for the fan and
compressor stall margins in steady state. Next, the original system is augmented with the
actuator dynamics. This increases the order of the system by three slow states to nineteen. The
actuator dynamics (see Appendix I) are described by
u = LU + MV
Incorporating this in (1)-(5) results in a singularly perturbed LQR problem

or
(4) = (A3
A I A2 XI
&) (2) (ill)
+

Xa = AaXa + BaV
with

where
J=
s: (X,TQXa + VTRV) dt

XT = (XT, UT), A: = (AT2, 0)

The full-order CSE is

Y = - (A, - BaK)TY, Y ( T ) = (LT, Lz)'
K = -R-'B,TYY~, K(T) = o
The dimensions of Y and K are 19 x 3 and 3 x 19, respectively.
Since P(7') = - Q , rank(P(7')) = 3 and Q is by construction positive definite, it is possible to
decompose the problem so that P( T) = - LoL& where L o = ( L I ,Lz)', LI is 8 x 3 and L2 is
11 x 3. The reduced CSE is
-
Y10 = - (A - BalR1O)TP1O,Plo(T) = L1- AJ&-TL2
K 10-- - R -1 T--T - (18)
BalY1OY10, KIO(T)=O
8 G. P. SYRCOS AND P. SANNUTI

where
A = A1 - A2Ai1A3
The solution of system (1 8) provides the zeroth-order approximation of the outer expansions
of the feedback gains. In order to correct singular perturbation effects, boundar-layer correc-
tion terms can now be obtained if deemed necessary. When the transient response of a system
is crucial, e.g. for the calibration phase of a navigation system, the boundary-layer correction
terms of the feedback gains are significant and can be calculated as follows. Let a = - 7 . Then

dm-1
-- - ABn- I
d7

Solving the above equations, we obtain

n-,(a) =e Ak2
m - l ( a ) = A:Ailn-l(a) = A:Ai1eA:"L2

qo(o) = j" +W
R-' B:l [(Ylo(T) + m -I ) n: I 1 da

Further,

and
mo(a) = 1"
m
(ABno + A:m- I ) da

We select l0(0), qo(0) and rno(0) such that lo(oo), qo(w) and mo(.o) are equal to zero. The
stretched-time variables 10 and qo provide the first order of E ( O ( Eboundary-layer
)) correction
terms of the feedback gains K I and K2, respectively. The O(z) outer expansions are

$11 = - [A- BalKlo] -11 + [BalKll -Ba1XrlAi'A3] TIO

+ A:Ai
 JET ENGINE CONTROL 9

with the terminal conditions given by

DISCUSSION
The F-100 turbofan engine linear model was simulated on an IBM-370 digital computer using
the IMSL software package. Several simulation runs were made, and pertinent results are
included in Figures 1-8. Figure 1 shows a representative case where the feedback gains
calculated via the full-order Chandrasekhar algorithm closely approximate the gains obtained
from the reduced Chandrasekhar algorithm. However, in cases where the reduced gains turn
out to be grossly different from the full-order feedback gains, the first-order correction terms
significantly improve the approximations as shown in Figures 2 and 3. The resulting closed-loop
responses for several variables are shown in Figures 4-8. These Figures illustrate the
significance of including the first-order correction terms in the feedback gains from the point
of view of the overall closed-loop response.
It should be noted here, that the time-varying gains are not customarily used in. control
problems; instead the asymptotic gains are used. However, in special applications, time-varying

0.5
-KIF(2,2)
0.25 - ----- K 10(2,2)

0 -

- 0.25 -
w
i
4
-0.5-
(I

-0.75 -
- I -

- 1.25 -- - - - - - - - - - - - - - - - - - - - _ - - - _ _ _ _ _ _ _ _ _ _ .
- #

- I .5 I I I
10 G. P. SYRCOS A N D P. SANNUTI

-5

cr, -10
L
4
(I
-IS

-20

I -25 ,

0. .5 1. IS 2.
TIME IN SECONDS
Figure 2. Feedback gains calculated via full-order Chandrasekhar algorithm (Klr), reduced order (KIo) and first-order
correction term

0. .5 1. 1.5 2.
TINE IN SECONDS
Figure 3. Feedback gains calculated via full-order Chandrasekhar algorithm (Klr), reduced order ( K l o ) and first-order
correction term
 JET ENGINE CONTROL 11

- X I 3 F U L L ORDER
----- X I 3 REDUCED
--- X I 3 FIRST ORDER APPROX.

0 I 2 3 4 5 6 7 8 9 10
TIME I N SECONDS
Figure 4. Fan turbine inlet temperature closed-loop response

-X I 0 FULL ORDER
----- X I 0 REDUCED
- : ---X I 0 FIRST ORDER APPROX.
...
.-a,
’\
8
I
I
I
I 8
I \

1 I I I I 1 1 I I I
0 I 2 3 4 5 6 7 8 9 13
TIME IN SECONDS
Figure 5. Burner exit slow temperature closed-loop response
12 G . P. SYRCOS AND P. SANNUTI

80

70

60
8 --- X2 FIRST ORDER APPROX.

2
2 50
Cr,
40
8
Cr, 30
Y
2- 20
x
0
u 10

0
0 I 2 3 4 5 6 7 8 9 I0
TIME IN SECONDS
Figure 6. Compressor speed closed-loop response

160

- XI
- - -- XI
FULL ORDER
REDUCED
--- XI FIRST ORDER APPROX

0 I 2 3 4 5 6 7 8 9 10
TIME I N SECONDS
Figure 7. Fan speed closed-loop response
 JET ENGINE CONTROL 13

lu
5w -5.0
2
ft -100
a
2
z I - X S FULL ORDER
-150
CU
z ----- X S REDUCED
(
3
I
-_ - X5 FIRST ORDER APPROX.
4
-200
-i
qi ;)111
.'.. I
L I I I I I I I 1
O ' - ' I 2 3 4 5. 6 7 8 9 I0
TIME I N SECONDS
Figure 8. Augmentor pressure closed-loop response

gains are of primary importance. Examples include the calibration of navigation equipment and
the transient response of an engine for a vertical take-off aircraft. Another type of problem
where time-varying gains are important is the estimation problem. This is the same type as the
control problem with Q and R being the covariances of the input and observation noise, respec-
tively. These problems require time-varying feedback gains and are among the most likely
candidates for the Chandrasekar algorithm.
The work of Calise and Sridhar'' should be mentioned since they also use singular pertur-
bation methods for jet engine control. However, their work is concerned with non-linear
minimum-time control laws. In contrast, we deal with controller order reduction via the LQR
problem. In this paper, we consider only fuel flow, nozzle area, and inlet guide vane positions
'
as important control variables. Additional variables could be augmented for more effective
control. Further, the output of an FlOO engine, in general, contains a feedforward part from
the input. This was not considered in our work.

CONCLUSIONS
The thrust of this paper is to show that singular perturbation methods can be used to reduce
the computational cost in obtaining feedback gains for LQR problems. An FlOO turbofan
engine which has 19 states has been modelled as a singularly perturbed system with 8 slow states
and 11 fast states. This modelling enables one to calculate an approximation to the feedback
gains by solving an 8th-order LQR instead of a 19th-order LQR. Instead of solving matrix
Riccati equations, a Chandrasekhar algorithm is used to reduce further the computational
complexity. The combination of singular perturbation methods with Chandrasekhar algorithm
enables us to obtain approximate feedback gains for a 19th-order turbo-engine system by
solving only 48 coupled non-linear equations of Chandrasekhar type. A direct method would
require the solution of 190 non-linear coupled equations in order to obtain the feedback gains.
Our numerical simulations show clearly the validity of our approach. Computational costs are
reduced considerably.
14 G . P. SYRCOS AND P. SANNUTI

APPENDIX I: STATE EQUATIONS OF T H E F-100 MODEL (ALT = 0, PLA = 83)

The 16 x I6 Matrix A
-4.328 0.1714 5.376 401.6 - 724.6 - 1.933 1.020 - * 9820
-0.4402 -5.643 127-5 -233.5 -434.3 26.59 2.040 -2.592
1.038 6.073 - 165.0 - 4.483 1049- - 82.45 - 5.314 5.097
0.5304 -0.1086 131.3 - 578.3 102.0 - 9.240 -1.146 - 2.408
0.0085 -0.0156 0.0560 1 * 573 - 10.05 0.1952 - 0.0088 - 0.021 1
0.8350 -0.0125 - 0.0357 -0.6074 37.65 - 19.79 -0,1813 -0.0300
0.6786 -0.0126 - 0.0968 -0.3567 80.24 - 0.0823 - 20.47 -0.0393
-0.0970 0.8666 16-87 1.051 - 102.3 29.66 0.5943 - 19.97
-0.0088 -0.0164 0.1847 0,2169 - 8.420 0.7003 0.0566 6.623
-0*00013 -0.00024 0.0027 0.0032 -0.1246 0.0104 0.00084 0,0981
- 1.207 -6.717 26.26 12.49 -- 1269. 103.0 7,480 36-84
-0.0273 -0.4539 - 52.72 198.8 - 28.09 2-243 0.1794 9.750
-0.0012 -0.0202 2.343 8.835 - 1 *248 0 * 0997 0 * 0080 0.4333
-0.1613 -0.2469 - 24.05 23.38 146.3 1.638 0.1385 4.486
-0.0124 0.0302 -0.1198 - 0.0482 5.675 - 0.4525 19-81 0.1249
-1.653 1.831 -3.822 113.4 341.4 -27.34 - 2.040 - *6166

0.9990 1.521 -4.062 9.567 10.08 -0.6017 -0.1312 0.0960

11.32 10.90 -4,071 - 0.0574 -0.6063 - 0.075 -0.5936 - 0,0960
- 04094 0.1352 5.638 0,0225 0.1797 0.0241 1.100 0.0274
-3.081 -4.529 5 * 707 - 0.2346 -2.111 - 0.2460 - 0.4686 -0.3223
0.0021 - 0.0525 - 0.0408 -0.0092 -0.0817 0.0343 0-0050 - 0.0126
-0.0195 -0.1622 -0.0064 - 0.0235 -0.2201 - 0.025 1 -0.0037 - 0.0336
0.0188 -0.2129 - em93 -0.0314 - 0.2919 -0.0337 0.0887 - 0.0446
0.0225 0.1791 0.0084 0.0265 0.2560 0.0283 - 0.0375 0-0363
-49.99 0.0676 39.46 0.0050 0.0890 0.0054 O.Oo0 0.0137
- 0 6666 - 0.6557 0.5847 0.0 0.0013 0.0 0.0 0.0002
0.2854 2.332 -47.65 0 * 3406 3.065 0.3624 -0.4343 0.4681
- 9.627 -9.557 38.48 - 50.01 0.101 1 0.0120 - 0.0468 0-0171
- 0.4278 - 0.4245 1.710 - 2.000 - 1 *996 0.0005 - 0.0020 0.0007
- 4.414 -4,354 17.66 -3.113 -3.018 - 19.77 - 0.0500 0.0150
-0.0011 -0.0067 0.0183 - 0-0010 -0.0135 -0*0011 - 20.00 - 0.0020
0.5004 - 0.1437 - 2.416 - 0.1073 - 1.078 30.63 19.89 - 50.16

The I6 x 3 matrix B
-0.0457 -451.6 105.8
0.1114 - 546.1 - 6.575
0.2135 1362- 13.46
0.3262 208.0 -2.888
0 * 0099 -98.39 0.5069
0.0273 71.62 9.608
0.0172 71.71 8.571
-0.7741 - 141.2 -0.8215
0.0385 -7.710 - 0.0437
O.OOO6 -0.1144 -0.0006
5-727 - 1745. - 8.940
0.1392 -24.30 -0.2736
0.0062 - 1.082 -0.0118
0.0678 16-60 0.3980
0.0019 9.147 -0.8241
0-1677 435.8 - 89.94
 JET ENGINE CONTROL 15

Engine state variables

XI = F a n Speed, rpm
XZ = Compressor speed, rpm
X3 = Compressor discharge pressure, psia
X4 = Interterbine volume pressure, psia
X5 = Augmentor pressure, psia
X6 = Fan inside diameter discharge temperature, OR
X7 =Duct temperature, "R.
XS = Compressor discharge temperature, "R.
X9 = Burner exit fast response temperature, "R
X I O= Burner exit slow response temperature, "R
X I I= Burner exit total temperature, "R
XIZ= Fan turbine inlet fast response temperature, "R
XI3 = Fan turbine inlet slow response temperature, "R
XI4 = Fan turbine exit temperature, "R
XI5 = Duct exit temperature, O R
XI6 = Duct exit temperature, "R

Engine inputs
U , = Main burner fuel flow, lb/h
UZ= Nozzle jet area, ft
U3 = Inlet guide vane position, degrees

Actuator dynamics transfer functionslSZ0

UI: [O*ls+1)(0.2s+ l)] -'= (0.12s+l)-I
Uz: [(0*01s + 1)(0.0OO7~~ + 0-03s+ l)] G (0.04s + l)-'
U3: [0.02~+ 1)(0*071~~+0*012~+ l)]-'=(O*ls+l)-'

APPENDIX 11. DERIVATION OF THE STATE WEIGHTING MATRIX Q

Given the system
X=AX+BU (19)
let
X=MZ (20)
where M is the modeal matrix. Substituting (20) in (19),we obtain
i = (M-~AM)Z+ ( M - ~ B ) u
The state vector Z can be reordered so that the magnitude of the eigenvalues of A = M-IAM)
are in increasing order. That is, if
W=EZ
then

["I]
wz
= p'[,.]: where ['p:] =EM-'
16 G. P. SYRCOS AND P. SANNUTI

Reordering and scaling the physical state vector X according to the singular perturbation
modelling algorithm, that is
-X = EN’X
results in
W1= PIX = (PIEN)%
The performance index is’*
T-+W

J=i lo (rrT + UTRU) df

where
r = HlWl+ V
If we substitute (21) into (22) so that
+V
r = (HIPIEN)%

where %= [E]
and partition the matrix product HIPIENinto (L1, Lz),then (23) becomes

r = LIX + LzZ+ V = [ L I , 13, Lz] V

I:[
Thus, the weighting matrix Q for the performance criterion used in equation (16) is a 19 X 19
matrix

REFERENCES
1. Sain, M. K., ‘The theme problem’, The 1977 NEC International Forum on Alternatives for Linear Multivariable
Control, pp. 1-12.
2. Hackney, R. D., R. J. Miller and L. L. Small, ‘Engine criteria and models for multivariable control system design’,
The 1977 NEC International Forum on Alternatives for Linear Multivariable Control, pp. 14-28.
3. Kokotovic, P. K., R. E. O’Malley and P. Sannuti, ‘Singular perburbations and order reduction in control theory’,
Autornatica, 12, 123-132 (1976).
4. Chow, J. H. and P. K. Kokotovic, ‘A decomposition of near optimum regulators for systems with slow and fast
modes’, IEEE Trans. Automutic Control. AC-21, 701-705 (1976).
5. Kokotovic, P. K. and R. A. Yackel, ‘Singular perturbations of linear regulators: basic theorems’, IEEE Trans.
Automatic Control, AC-17, (I), 29-37 (1972).
6. Yackel, R. A. and P. K. Kokotovic, ‘A boundary layer method for the matrix Riccati equation’, IEEE Trans.
Automatic Control, AC-18, 17-24 (1973).
7. Kailath, T., ‘Some new algorithms for recursive estimation in constant linear systems’, IEEE Trans. Inf Theory,
IT-19. 750-760 (1973).
8. Casti, J., ‘Matrix Ricatti equations, dimensionality reductions and generalized X-Y functions’, Utilitas Math., 6,
95-110 (1974).
9. Casti. J., ‘Dynamic systems and their application: linear theory’, Math in Science and Eng., Vol. 135, Academic
Press, New York, 1977.
10. Syrcos, G . P, and P. Sannuti, ‘Asymptotic expansions of singularly perturbed Chandrasekhar type of equations’,
American Control Conference, Arlington, Virginia, 14-16 June 1982, pp. 1058-1062.
11. Anderson, B. D. 0. and J. B. Moore, Linear Optimal Control, Prentice-Hall Inc., Englewood Cliffs, NJ, 1971.
 JET ENGINE CONTROL 17

12. Avramovic, B., P. K. Kokotovic, J. R. Winkelman and J . H. Chow, ‘Area decomposition for elecromechanical
models of power systems’, Automatica, 16, 637 (1980).
13. Fossard, A. J., M. Berthelot, and J. F. Magni ‘Oncoherency-based decomposition algorithms’, Automotica, 19,
247-253 (1983).
14. Syrcos, G. P . and P. Sannuti ‘Singular perturbation modeling of continuous and discrete physical systems’, fnt.
J. Control, 37,(5), 1007-1022 (1983).
15. Chow, J . H. and P . K. Kokotovic, ‘Eigenvalue placement in two-time scale systems’, Proceedings of the IFAC
symposium, Udine, Italy, 1976.
16. Bryson, A. E. and Y . C. Ho, ‘Applied Optimal Control’, Gin and Co., Waltman, Mass, 1969.
17. Kuris, E. D. and A. I., Chernomorskii, ‘Design of linear systems with assigned dynamic properties on the basis
of a quadratic performance criterion’, Automation and Remote Control, 10 May 1982, pp. 1588-1592. Translated
from Avtomatika i Telemechanika, 42,(12), pt. 1, 11-16 (1982).
18. Harvey, C. A. and G. Stein, ‘Linear multivariable control design based on asymptotic regulator properties’, The
1977 NEC International Forum on Alternatives for Linear Multivariable Control, pp. 189-201.
19. Calise, A. J . and B. Sridhar, ‘A multi-time-scale approach for jet engine control systems’, The 1977 NEC
International Forum on Alternatives for Linear Multivarioble Control, pp. 2 13-223.
20. Peczkowski, J. L. and M. K. Sain, ‘Linear multivariable synthesis with transfer functions’, The I977 NECInterna-
tional Forum on Alternatives for Linear Multivariable Control, pp. 1 11-127.