You are on page 1of 234

https://ntrs.nasa.gov/search.jsp?

R=19780023164 2018-02-02T17:05:11+00:00Z

General Disclaimer

One or more of the Following Statements may affect this Document

This document has been reproduced from the best copy furnished by the
organizational source. It is being released in the interest of making available as
much information as possible.

This document may contain data, which exceeds the sheet parameters. It was
furnished in this condition by the organizational source and is the best copy
available.

This document may contain tone-on-tone or color graphs, charts and/or pictures,
which have been reproduced in black and white.

This document is paginated as submitted by the original source.

Portions of this document are not fully legible due to the historical nature of some
of the material. However, it is the best reproduction available from the original
submission.

Produced by the NASA Center for Aerospace Information (CASI)


. .
II

(N ASli-C y -157578) ALTERNATIVES FnP JET N78-31107


ENGINE CCN-TRCL Final Technical Feport, 1
Mar. 1977 - 2P Feb. 1578 (Nctre Tame "niv.)
233 p F?C X11 /MF A01 CSCL 21E Unclas
G3/07 30222
I

jj

i
^' 3

Final Technical Report

to the
r

National Aeronautics and Space Administration

on -.
NASA Grant NSG•-3048, Supplement No. 2

ALTERNATIVES FOR JET ENGINE C0NTROL

t k^ (march 1, 1977 -- February 28, 1978)

f
1

3
i

t
Under the Direction of

Dr. R.J. Leake


Dr. M.K..Sain
ti = Department of Electrical Engineering
University of Notre Dame {
Totze llama ' Sndxarra . 655
a ;,

-= 3

p̀ . - .. - • ,. ._ __ ?%va iY.y.S ,h .i. waiws -_.^. n.


t ALTERRATIVES FOR JET ENGINE CONTROL

NASA Grant NSG-3048

Supplement No. 2

ABSTROT

I This report deals with progress made on the Grant NSG--3048 during the

TI I^ calendar year beginning March 1, 1977 and ending February 28, 1978. This

1 year coincides with Supplement No. 2 of the award, which originated on


March 1, 1975. The NASA Technical Officer for this period was Dr. Bruce

Lehtinen of Lewis Research Center. The directors of the research at the


University of Notre Dame were Dr. R. Jeffrey Leake and Dr. Michael K. lain.

General goals of the research have been classified into two cate-

gories. The first category involves the use of modern multivariable. frets

quency domain methods for control of engine models in the neighborhood of

r a quiescent point. The second category involves the use of nonlinear mod-

elling and optimization techniques for control of engine models over a


more extensive part of the flight envelope.

Su"vstantial progress has been made in both categories.

In the frequency domain category, works have been published in the


areas of low-- interaction design, polynomial, design, the CARDIAD* method,
and multiple setpoint studies. A number of these ideas have progressed

to the point at which they are starting to attract practical interest.


Further effort is yet required, however, to Carry the ideas to maturity

*The acronym stands for Complex Acceptability Region for IA gonal Dom-
inance. See report for details. r

i
_
a

and to ensure their adequate dissemination. A highlight of the year was

the incorporation of realistic jet engine data as a theme problem into

the International Forum on Alternatives for Linear 'blultivariable Control.



r Dr. Sain was Program Chairman for this meeting, which attracted nearly

two hundred persons from industry, laboratories, and universities to

hear thirty papers focused in the general subject area of this grant.

r In. the nonlinear category, advances have been made both in engine

modelling and in the details associated with software for deter ^xa
i tion
of time optimal controls. Nonlinear model s fo r a two spool turbofan
i' engine have. been. , expanded and refined; and a promising new approach to
t`
automatic model generation: has been placed under study. A two time scale
scheme has been developed to do two-dimensional dynamic programming, and
{< a
i; an outward spiral sweep technique has greatly speeded convergence times
in time optim" al calculations.
The details of these and other aspects of the yearns.investigations
^.` may be found in the body of the report, which covers the most active
,rant period . to date., _
;. A

i
f ` ,.
3

.. .. y^
.; :.^ ix ''--ma- :.: .-x-,• f.
m..i. .D^ s ti afsmd ,: dxw,;-e. • r
t TABLE OF CONTENTS

3 ABSTRACT. ....
Page

S 1. INTRODUCTION ................................. ................... ^.


a - :

II. HIGHLIGHTS OF THE- RESEARCH ...................................... 7

III. PUBLICATIONS... * ............ ;oas000 ............... ............ 10 r^


TV
. T.V. LOCAL MULTIVARIABLE FREQUENCY DOMAIN METHODS...........
V. GLOBAL NONLINEAR OPTIMAL METHODS....,
r t^ 3

r `^ VI. SPECIAL INITIATIVES RELATED TO GRANT WORK ....................... 27

VII. REFERENCES ...................................................... 31

APPENDICES ^
1 1

A. GRANT BIBLIOGRAPHY, INCEPTION TO PRESENT ^3

B. HIERARCHY OF SlxMuIATION MODELS FOR A TURBOFAN GAS ENGINE


^i 1

C. FREQUENCY DOMAIN COMPENSATION OF A DYNGEN TURBOFAN ENGINE


_ MODEL

D. APPLICATION OF POLYNOMIAL TECHNIQUES TO HULTIVARIABLE CON-


TROL OF JET ENGINES

r` E. A COMPARISON OF ',-U;QUENCY DOMAIN TECHNIQUES FOR JET ENGINE


. CONTROL SYSTEMS _:SIGN ^' l
'- F. TIME OPTIMAL CONTROL OF A TWO—SPOOL TURBOFAN JET ENGINE

G. DIRECT METHOD FOR OBTAINING NONLINEAR ANALYTICAL MODELS OF A }7


JET ENGINE .?
.-
.Y I.

H. SOME FEATURES OF CARDIAD PLOTS FOR SYSTEM DOMINANCE


i
I. THE THEME PROBLEM x
r':,.

J. INPUT COMPENSATION FOR DOMINANCE OF TURBOFAN MODELS


3
4

i
{

Li. .1•-..^^_
F'-e, _ ... . >-.- _^.,.siZL'f'l,waiw^+1%-', {:'^ a..i L.:._sn{a,GMvcf .:^.. ! - ^ -
l . INTRODUCTION
.ts

l E r
The purpose of this section is to provide some of the broad back-

ground which underlies and clarifies the general nature of the Research

Ui&hlights, which are stated in the section following.

Initiation of Grant NSG-3048 in March 1975 was timed with develop-

ments in the engine industry, which was be ginning to experience some lim
is
nations in the application of classical hydromechanical control tech-

pique as the primary base technology for modern engines with ever in-

creasing sophistication. At the same time, milestone developments in

digital hardware began to open realistic possibilities for onboard comp--

utation to an extent not heretofore possible. This confluence of events

. led directly to the concept of increasing the role of electronics in


t
engine control. In turn, the availability of digital electronics itself

created a wide variety of opportunity for application of new control de-

sign philosophy and technigpe. Among the earliest of such studies is


t
the P100 Multivariable Control Synthesis Program [l] sponsored by the
r:
National Aeronautics and Space Administration, Lewis Research Center
and the Air Force Aero-Propulsion Laboratory, Wright--Patterson Air Force

M Base. This program is currently in the test phase.

The advent of digital technology on the engine scene offers not j

r; only the opportunity'to Control more engine variables but also the peas-

sibility of integrating. engine and. airframe control. Studies of`this


r:
type have also begun.

Primary tools in the FIOO Mtltivariabl.e Control Synthesis Program.


t

were linear quadratic regulator (LQR) theory in the linear case.. For the

global control, nonlinear optimal methods were not directly applied.

The purpose of Grant N5G-3048 is to evaluate alternatives to LQR in

the linear case and to examine nonlinear modelling and optimization ap-

proaches for global control.

Context for the studies is set by the DYNGEN digital simulator [2].

!k Based upon earlier computer codes GENENG [3] and GENENG 11 [4], DYNGEN

has the combined capabilities of [3] and [4], for calculating steady-

state performance, together with the further capability for calculating

transient performance. DYNGEN uses a modified Euler method to solve the

differential equations which model the dynamics of the engine. This mod--
UY
ified Euler method permits the user to specify large time steps, for ex-

ample a tenth of a second; and this can result in considerable savings of

execution time. On the other hand, convergence problems are sometimes

encountered with DYNGEN when small time steps are used.


Tit,
The DYNGEN digital simulation is particularized to a given situation

by a process of loading data for the various maps associated with a given

1 engine. The maps for the Grant NSG-3048 have been provided by engineer-

ing personnel. at Lewis Research Center. These maps correspond toga

paper engine, which is not closely identified with any current engine.

But the data do correspond in a broad., general sense to realistic two

' spool turbofan engines. The simulation provides for two essential con-

trols, main burner fuel flow and jet exhaust area. Portions of the en-
s
velope which can be used for linear or nonlinear experimentation are

limited by the convergence capabilities of the available engine data on


i! With respect to multivariable frequency domain work, the basic ap-

proaches may be classified into two groups. These two groups are often

called "direct" and "indirect".

' The direct approach can usually be recognized by its attention to

achieving a completely specified dynamic performance. Such ideas have

been discussed from the early days of organized control study. See,

for example, [5] and [5]. In fact, some of the earliest attempts to ex-

pand the direct approach to the multi-input, multi-output case involved


rt work lwxth jet engines [7,8]'. Direct approaches in multivariable appli-
cations typically involve matrices of transfer functions. In the 1950's,

there were some nontrivial difficulties with such methods in cases of

more than one input and output. Among these difficulties may be mentioned
a
(1) s he meaning and extent of cancellations of various types

in the transfer functions,

3 (2)-the question of loop stability,

(3) the problem of specifica.tiono and


Fs

i'

t^.
^y

to the classic works of Nyquist or Evans, involving, respectively, fre-

quency response plots in the complex plane or versions of the root locus.

It is a relatively easy matter to describe the focus of generalized

Nyquist methods. The key constituent ideas are related to three poly-

nomials:
it
(1), P C (s)- the closed loop characteristic polynomial (CLCP),

WT (2) po (s) - the open loop characteristic polynomial (OLCP), and

(3) 111(s)i- the determinant of return difference.


r

The CLCP is a polynomial whose zeros characterize the exponential im-

pulse response of the closed loop control system; the OLCP serves the same

purpose for the open loop system. M(s) is a matrix of transfer functions

associated with the following experiment. Break the control loops at a

convenient point and inject impulses. The difference between the trans-

forge of the signal injected and that which returns at the other end of

the loop is established by the columns of M(s). The quantities pc(s),

po (s) and M(s) derive their importance from the fact that they are re-

lated to each other by the equation

P (s) = IM(s) I POW.


c

Typically, po (s) is known; and M(s) is partly given and partly designed,

in such a way that p c (s) becomes desirable.

ti Generally speaking, a Nyquist plot of IM(s)1 tends to contain the

t same types of information which proved so successful in classical designs.

A great deal of the design effort centers upon the way in which dynamical

t compensation affects the determinant which acts on M(s). There are -three 't

well recognized ways to study this effect. These are

am
4

7
(1) direct construction of IM(s)J by any of the known methods

for determinant calculation,

,{. (2) construction of the eigenvalues of JM(s)] as a function of

s, and use of the idea that the determinant is equal to

the product of its eigenvalues [9], and


f

(3)' design of compensation so that M(s) is approximately dia-

gonal, and establishment of a relation between the plot.

`' of JM(s)J and plots of the diagonal elements of M(s) [10].

It is believed that work on this grant has advanced the application of

all these methods to jet engine design, but particularly method (3),

where a special technique has been developed to design compensation so

that M(s) is approximately diagonal. This technique is called the CARDIAD

Plot, where the acronym stands for Complex Acceptability Pegion for DT-A-

gonal Dominance, the latter term referring to a specific definition of

"approximately diagonal."

^w F With respect
to nonlinear modellingg and o F timi.zationa the emphasis
F

has been twofold: to develop good analytical nonlinear models of the

jet engine and to use these models in conjunction with techniques of

mathematical programming in order to develop advances in global control


}.
over significant reaches of the flight envelope.

J ` In general, there are several aspects to this part of the investi-

gation. First, it is possible to conceive the basic differential equa-

tions from fundamental. principles. In this case, there are usually about

sixteen nonlinear differential equations, as.well as a large number of

nonlinear static functions which serve as part of the coupling between

t r%

,.....^ .....vn-^-mi
Fr ..-y,_ Y—._5+: ; t
.+"r, v^cM. .r'-*-^iTi'?.^jT.. O-e rt4^1LYC 9`.""t^^.::..[3^kwCVLY'rXY ^3 -
the equations. These functions often have more than one argument. If

the equations arise in this fashion, than there is a significant need

to identify the parameters. This must normally be done from the DYNGEN

digital simulation. Second, it. is possible to assume a general form for

the nonlinear differential equations in such a way that fundamental prin-

ciples are not ignored but that added emphasis is placed upon general

mathematical form. If this general form is chosen according to a scheme

designed' to make maximum use of the type of data which is directly avail-

able from the digital simulation, then a type of "automatic" nonlinear

model generation becomes possible. Third, whether the first or second


ia modelling procedure is employed, there is almost always a need to con-

sider the problem of reducin g the order of the models. Though order
ti
reduction can often be highly mathematical in nature, it is almost al-
j,
ways the case that the reduced order model depends upon the scaling of

the equations. As a result, the final reduced models often depend in a A

nontrivial way upon physical insight, as well as mathematical method.

Work on this grant has focused especially upon the first and second

aspects of the modelling problem, with a gradual specialization toward

automatic model generation.

Insofar as optimization is concerned, the stress has been placed

upon time optimal control, and considerable effort has been invested in

specialized programming methodology designed to take maximum advantage

of the particular features of jet engine models...

In the next section, the highlights of activities carried out during

the calendar year correspandingtothis report are presented.

_1777
t:

1 II. HIGHLIGHTS OF THE RESEARCH


.t

This section is a brief statement of the main achievements under

, Grant NSG-3048 during the period from March 1, 1977 to February 28, 1978.
5

There are two major subdivisions, according to the main thrusts of the

investigation. The first of these is Local Multivariable frequency Domain

Methods; and the second is Global. Nonlinear Optimal. Methods.


;; r
For the most part, the wording of these paragraphs has been con-

strained so as to be as nontechnical as possible. Nonetheless, some

readers may find it useful to review the basic introduction provided in

Section T.

a.: A. Local Multivariable frequency Domain Methods

T During the calendar year ending on February 28, the following results

were achieved in the area of modern, .frequency domain control of turbofan

engine models.

(1) The first formal documentation of the CARDIAD method


(Complex Acceptability Region for DlAgonal Dominance)
was completed. See (1), Section III. Though supported
principally under a theory grant from the National
Science foundation, this technique had its origin in
class studies of older methods for approximate de-
coupling of jet engine models in the frequency domain.
Alinost all of the examples in this thesis were taken
from F100-like engine data.

(2) The first documented studies of direct Nyquist plots


of return difference determinants for jet engine
models were completed. See (3), Section Ill. This
thesis has been a helpful ancillary tool in general
frequency domain design.

(3) The first frequency domain closed loop compensation


and simulation of a DYNGEN turbofan engine model was
achieved. See (5), Section III. As explained in

7
yy.r

Section I, the DYNGEN simulation supplies two control


inputs.

(4) The first study of polynomial techniques for exact


model matching control of jet engine models in the
frequency domain was reported. See (6), Section III.
This paper has been pivotal in promoting the numer-
ical advance of such techniques for applications.
More will be reported in the subsequent semi-annual
status report..

r,
(5) About a given design point, linear models of the
standard type are obtained from the DYNGEN simulation
by the DYGABCD routine [11]. In order to use DYGABCD
at off-design points, however, modifications to DYGABCD
necessary to the research had to be accomplished. See
r: (7), Section III.

(6) The CARDIAD plot jaas applied to a series of DYNGEN off-


z' design point models in order to determine its utility
as a method for global classification of interaction
characteristics of jet engines. See (14), Section III.
_..'? The results were positive.

(7) An entire conference was convened from industry, lab-


'` oratories, and universities to hear speakers from sev-
eral countries apply their theories to a theme problem
developed from jet engine data. See (15), Section III.
F This meeting resulted in a book publication [12].
1 j
(S) The CARDIAD methodology was extended to the three--con-
'`' trot.-input case and applied successfully to Pratt-
: Whitney data for the F100 engine. See (16), Section

(9) A joint seminar series was established between the


Department of Electrical Engineering at Notre Dame and
the Energy Controls Division of the Bendix Corporation
at South Bend, in areas of mutual interest. This has
resulted in published work. See (17), Section III.

' B. Global Nonlinear Optimal Methods

The major advances and results achieved during the past year in the

jarea of global nonlinear optimal methods are the following.

(l) The hierarchy ofanalytical nonlinear models for the


two spool turbofan jet engine has been expanded and
refined. See (2), Section III. This effort is in

^,irT _.. ^'^^ ... :5.. .._ .-., .. _1 •__ - 17.t1.^1ti-^1f.:e6r..^^{.iZ•i::v-iTiK^.1 +t ^.]i'^cx`z ^., ti,. s.-. _, _ _
P . keeping with general interest in the industry concern-
ing improvement of compact general models.

(2) A comparison has been achieved between the use of a


linear affine model and a nonlinear model for time
optimal control studies of a single spool engine. See
(4), Section III. The results again support the search
ror reasonably simple nonlinear models, in the sense
that they argue in favor of models whose nonlinearity
t^ 1 is not excessively complicated_

(3)' A method using linear quadratic regulator methods to


obtain decoupled control has been tested on various
l engine models. See (8), Section III. This is also
an outgrowth of the joint Notre Dame -- Bendix seminar
series mentioned in (9), Section IIA, above.

(4) A two time- - scale scheme has been developed in order to


do two-dimensional dynamic programming on a fifth
order model of a jet engine. See (10), Section III.
This is part of a continuing study of time - optimal
control methods applied to nonlinear engine models.

(S) Convergence times in time - optimal successive approx-


imation dynamic programming have been dramatically
impraved through development of a scheme for a spiral
out sweep from the target. See (1 0 ), Section. III.

(6) A completely automatic method for obtaining nonlinear


analytical models for engine simulations has bean de-
veloped and tested numerically. See (11), Section III.
n!is approach offers considerable promise for improve-
m(-jnt over previous methods.

(7) A discrete maximum principle has been developed for


nonlinear systems haying the property that the con-
trol is constrained by the present state. See (12),
Section III.
(S) A family of optimal feedback control laws has been
developed ar d =.imulated for a variety of models.

Further details concerning these highlights may be found in Sections IV

and V. Also, as described in Section III following, a number of the doc-

4 uments have been included as appendices.

The next section contains a list of publications completed during

the current year of the grant.

B
Illy . PUBLICATIONS

This section provides a list of the nineteen documents completed

during the year March 1, 1977 through February 28, 1978. The works are

ordered chronologically.

-? Some of the listings are followed by an alphabetical code consisting

i±• of one or more of the letters A, M, and R. The letter A signifies that

the document or an abstract thereof appears as one of the appendices to


r _;
this report; the letter M signifies that the document comprises a thesis

i . for the degree of Master of Science in Electrical Engineering; and the

letter R declares that the document summarizes an effort which was closely

integrated with, but not directly supported by, the activities of this

grant.
i:
Completed publications from earlier years are not included in this

list; but a total listing of all the grant documents has been provided

x' as an appendix to the report. See the Table of Contents.

(1) R.M. Schafer, "A Graphical Approach to System Dominance",


Technical Report EE 772, University of Notre Dame, April
}. 1, 1977. (M,R)
E (2) W.E. Longenbaker and R.J. Leake, "Hierarchy of Simula-
tion Models for a Turbofan Gas Engine", Proceedings
Eighth Annual Pittsburgh Conference on Modeling and
Simulation, April 1977. (A)

(3) P.W. Hoppner, "The Direct Approach to Compensation of


Multivariable Jet Engine Models", Technical Report EE
774, University of Notre Dame, May 1977, (M,R)

(4) R.R. Gej j i and R.J. Leake, -"Time-Optimal Control of a


' Single Spool Turbojet Engine Using a Linear Affine
Model", Technical Report BE 7711, University of Notre
Dame, June 1977.

t 10
E.

R.M. Schafer, R.R. Gejji, P.W. Hoppner, W.E. Longenbakex


and M.K. Sain, "Frequency Domain Compensation of a DYNGEN
Turbofan Engine Model", Proceedings Sixteenth Joint Auto-
matic Control Conference, pp. 1013-1018, June 1977. (A)

.. (6) R.R. Gejji and M.K. Sain, "Application of Polynomial


Techniques to Multivariable Control of Jet Engines", Pro-
ceedings Fourth IFAC Symposium on Multivariable Tech-
nological Systems, pp. 421.429, July 1977. (A)

' (7) R.R. Gejji, "Use of DYGABCD Program at Off-Design Points",


Technical. Report EE 7703, University of Notre game, July
1977.
(8) E.A. Sheridan and R.J. Leake, "Non-Interactive State Re-
quest Jet Engine Control with Non-Singular B Matrix",
Proceedings Twentieth Midwest Symposium on Circuits and
Systems, pp. 539--543, August 1977. (R)

(9) R. Gejji, R.M. Schafer, M.K. Sain, and P. Hoppner, A Com-


parison of Frequency Domain Techniques for Jet Engine Con-
trol System Design", Proceedings Twentieth Midwest Sympo-
sium on Circuirs and Systems, pp. 680-685, August 1977. (A)

(10) W.E. Longeubaker and R.J. Leake, "Time Optimal Control of


§ a Two-Spool. Turbofau Jet Engine", Technical Report EE 7714, 4

University of Notre Dame, September 1977. (A,M)

(11) R.J. Leake and J.G. Comiskey, "A Direct Method for Obtain-
ing Nonlinear Analytical Models of a Jet Engine", Proceed-
ings International Forum on Alternatives for Linear Multi-
variable Control., National Electronics Conference, Chicago,
pp. 203--212, October 1977. (A) i
(12) J.A. Ortega and R.J. Leake, "Discrete Maximum Principle
with State Constrained Control", SIAM Journal on Control
and Optimization, Vol. 15, No. 6, pp. 109-115, November
1977. (R)

(13) Michael K. Sain and V. Seshadri, "Pole Assignment and a


Theorem from Exterior Algebra", Proceedings IEEE Conference
can Decision and Control, pp. 291-295, December 1977. (R)

(14) R. Michael Schafer and Michael K. Sain, "Some Features of


CARDIAD Plots for System Dominance", Proceedings IEEE Con-
ference on Decision and Control, pp. 801-806, December
1977. (A)

' (15) M.K. Sain, "The Theme Problem", in Alternatives for Linear
Maltivariable Control., M.K. Sain, J.L. Peczkowski and J.L.
Mel.sa, Editors. Chicago: National Engineering Consortium,
1978, pp. 20-30. (A)

• 1.3
s -
z
{ (16) R.M. Schafer
Sc er and M M.K.Sain "Input Comp
"ICompensation
np for
Dominance of Turbofan Models", in Alternatives for
Linear Multivariable Control, M.K. Sain, J.L.
Petzkowski, and J.L. Melsa, Editors. Chicago: Na-
tional Engineering Consortium, 1978, pp. 156-169. (A)

(17) J.L. Peczkowski and M.K. Sain, "Linear Multivariable


Synthesis with Transfer Functions", in Alternatives
for Linear Multivariable Control, M.K. Sain, J.L.
Peczkowski, and J.L. Ifelsa, Editors. Chicago: Na-
`^I.^ tional Engineering Consortium, 1978, pp. 71-87. (R)

(18) R.J. Leake and M.K. Sain, "Semi-Annual Status Report,


NASA Grant NSG-3048, 'Alternatives for Jet Engine Con-
t7rol', Supplement No. 2", March 1, 1977--August 31, 1977.
(19) R.J. Leake and M.K. Sain, "Final Technical Report, NASA
Grant NSG-3048, 'Alternatives for Jet Engine Control',
Supplement No. 2", March 1, 1977-February 28, 1978.

t
t

Al
F ;'

i'

x `

12

rr _'

1
IV. LOCAL 11ULT1VARIABLE FREQUENCY DOMAIN NETHODS

`- Progress on local multivariable frequency domain methods has been

achieved during this grant period in the areas of Low Interaction Design,

Polynomial Design, Extension of CARDIAD Method., and Multiple Setpoint

studies.

Low Interaction Design

As mentioned in the Final Report for NASA Grant NSG-3048, Supplement

No. 1, a promising new technique for designing dynamical compensation be-


This
gan to develop in the Fall of 1975. methodology, built upon what

are currently being called CARDIAD plots, was only being tentatively con-

sidered in October, 1976 when the continuation proposal for NASA Grant

USG-3048, Supplement No. 2, was being written. Based upon favorable pre-

liminary reaction by personnel from NASA Lewis Research Center, a decision

was made to investigate further the use of CARDIAD plots as a design aid
4.

for turbofan engine control in the frequency domain. In essence, this

study proved to be successful enough that it really dominated the re-

maining time period of Supplement No. 1 and has continued through Sup-

plement No. 2.

A great deal of the power of the CARDIAD plot arises from its sim-

plicity. For each frequency, a circle is constructed on a planar plot.

Data for the center and radius of this circle is obtained from the com-

plex: transfer function matrix of the plant. The circle may be solid or

dashed. If solid, the inside of the circle defines the acceptable com-

plex region for the value of a frequency dependent compensator element

in order to achieve dominance. If dashed, the outside of the circle de-

3.3
i
••.-- - „ter-.^;^ _.,_.-_.. ....._^..- --- _^. __-----------'^..__......e..:..E^..^.,._-^
fines the acceptable region. As the frequency follows a standard Nyquist A
J^
pattern, these circles result in a CARDIAD plot. (Complex Acceptability
Eegion for DIAgonal Dominance). This plot has been shown to speak con-
,
structively to the issue of compensator choice to reduce interaction.

As an example of the CARDIAD plot application to the turbofan engine


control problem, a linear model obtained using DYGABCD on the DYNGEN dig-
ital engine simulator was used to illustrate control design at the 1977
Joint Automatic Control Conference. The paper based upon this effort,
which may be seen in Appendix C, utilizes a two-input, five-state' , two-

output engine model in which the inputs are fuel flow and nozzle area,
the states are compressor speed, fan speed, burner exit pressure, after-
burner exit pressure, and high turbine inlet energy, and the outputs are
thrust and high turbine inlet temperature.

Typical examples of CARDIAD plots for such engine models may be seen
in Figures 2-5 of Appendix r. The investigators involved in this study
have seen the same type of plots arising from a variety of engine data.
This has raised the interesting question of whether there may be a-mean-
ingful concept of "engine interaction footprint" in the sense of the
CARDIAD plot.

Of particular interest is the plot shown in Figure 3 of Appendix C.


Students of classical control theory will immediately recognize the near
semicircular nature of this plot. Such semicircular behavior has been
observed frequently and serves to specify a sort of essential lead-lag
classical compensati Selement
t which can achieve diagonal dominance.

Using the CARDIAD approach, it has been possible to achieve diagonal

14
dominance at all frequencies on typical engine models. Moreover, only

simple compensators have been required to do this. While it is not pos-

sible to apply the same degree of credibility to the model itself at all

'. frequencies, it is nonetheless of considerable theoretical interest to

be able to make this accomplishment, especially for (A, B, C, D) type

t plant models which have the D matrix present to approximate modelling

errors at high frequencies. Further insight into the significance of

these steps can be obtained by examining Appendix B of the Final Report

for Supplement No. 1.

Appendix C of this report also contains evidence of two other


rt
facets of the applications researches conducted under this grant. Fig-

ures 6-8 are characteristic locus plots for the plant, after it was com-

pensated by the CARDIAD methodology. This combination of 1•deas, namely,

y the CARDIAD plot and the characteristic locus, has been quite helpful

in studies conducted up to this time. Softwares were developed and ex-

perience gained with the characteristic locus on the original grant NSG-

3048, as well as on Supplement No. 1.

Of additional interest also are Figures 9-11 of Appendix C. These

figures deal with an aspect of frequency domain control researcli which

may be called a "direct" approach. The term "direct" refers to a direct

construction of the Acterminant of return difference as frequencies fol-

low a standard Nyquist pattern. This tool really underlies all the mod-

' ern frequency domain ideas; but it is Usually handled obliquely, as for

example by the CARDIAD idea or by the characteristic locus. Studies of

the direct approach to determinant of return difference have confirmed

15
It

^^'."^•^.^^' :'. ^.. ..'..- :. _.•: i'. :... _. ..., _._:...._^'.^e^).....^...._... w• . , .. .: h s•a^!Si.HL•zN.l4)»La?ikM•F•^l.Ma.=..1::'M•Wdiw--)Y.^.^.. h-..0 _ _ ___ __ ^_^^ ^ , ^
^ g -

that it is at the very least a revealing analysis method. Figure 11 of

Appendix C, for example, reveals a condition of gain, limitation which is

understandable in a global way not so easily visualized by separate char-

acteristic locus plots. It should be emphasized, moreover, that diagrams

FYI of the type of Figure 11 can be drawn without any regard of plant size
in terms of inputs, states, and outputs.

Efforts to use the direct approach for design, as well as for anal-
u.
ysis, have posed nontrivial algebraic questions. Some insight has been
w gained, but no breakthroughs have occurred as yet.

V Pol-_momial- Design

The principal efforts and results obtained in applying polynomial

I@ design techniques to the turbofan engine control problem have been re-

ported in the Final Report on Supplement No. 1.

During the present grant period, a paper on this work was presented
qa

at the International Federation of Automatic Control's Fourth Symposium

.. on Multivariable Technological Systems at the University of New Brunswick

in Fredericton. See Appendix D.

There is now little doubt that the control area is experiencing a

resurgence of interest in transfer function methods. As part of this

resurgence, grant work an polynomial design has pointed out the necessity

of increased attention to numerical method. The investigators also be-

lieve that it has stimulated other workers to begin numerical studies.

The transfer function has a number of key properties which have long

made it popular with control practitioners. For example, the transfer

function is unique relative to similarity transformations on the state

16
apace.

Mach work remains: to be done, however, on computational, aspects of

transfer function design.


z
A presentation comparing the design experiences of the investigators

under this grant, in the frequency domain, was made at the 1977 Midwest

Symposium on Circuits and Systems, Lubbock, Texas, in August. A copy of

this brief manuscript may be seen in Appendix F.

Extension of CARDIAD Method


^a

All the work so far mentioned in regard to the CARDIAD plot was

} a carried out for plants having two inputs and two outputs. In this sit-

uation, it is certainly true that the plots have many interesting prop-

erties.

On the basks of this e -)erience, which was in its final. stages in


11ay, 1977, it was decided to extend the CARDIAD theory to the three-input,

"" three-output case. Also, the National Engineering Consortium's Inter-

national Forum on Alternatives for Linear Multivariable Control, which


L1.
took place in October, 1977, offered a prune Opportunity to apply the

theoretical extension, inasmuch as the Forum contains a Theme,Problem

rt
e+ based on a linearized model of a modern turbofan engine. See Appendix
i^-
r..
I. Assistance in the theoretical extension of CARDIAD to the three-in-

put, three-output case was provided by support extended to Dr. Sain by

the National Science Foundation under Grant ENG 75-22322. NASA support

under Supplement No. 2 was focused on the turbofan application.

Appendix J, "Input Compensation for Dominance of Turbofan Models",

1 17 I
t
T

provides a complete description of the successful work on this extension

and its application.

Technically, the extension of CARDIAD plot methods to the three-in-

put, three-output case involved the use of a bound which provides suf-

ficient conditions for diagonal dominance. Many possibilities exist for


U
the selection of such a bound, and there remains considerable opportunity

for further research along these lines. After examination of a number

of basic bounding possibilities, an initial selection was made in such a


^R

way that the bound will extend to the case of a plant having p inputs

and p outputs, where p is any positive integer greater than or equal to


x}
two, and that the bound will be tight in the place where it matters the

mast---where the plant is close to failing the dominance test.

From an engineering point of view, it was necessary to develop soft-

ware to extend the CAEDIAD idea and to establish viewpoints for studying

the plots in more complicated cases. The CARDIAD analysis was divided

into two phases. The first phase assumed one off--diagonal compensator

ti
element to he zero. An advantage of such a phase lies in its conceptual

reduction to the situation of Appendix C, where greater design experience

is available. The first--phase approach was adequate for about half of


the application to the National Engineering Consortium Thecae Problem.

The secondhase g
P assumed both off-diagonal P
compensator elements to be

nonzero, but drew the plot in such a way that the designer could as-

- certain what would happen when one of those elements was zero. This sec-- '=

and phase approach was successful in completing the aesign for the Theme

Problem.

L,. f:

u f
k.

.cis. .-p.......»..x.,.^.•...^.-j.^_.,.,^.^ wv.-- .•.,..^..,......^.^..,__,—._...-.-...,-.........._...»..---_.^..._._..........` __^.._............•«....,_ - — -- ^ ^ ..


.- ^:= _^

^a - ' - , -- -=-..,•...-.. ,.: :..- -a:^u^i-+aiyet'.-.a.:•xo:XY,ba:,.s:;^..a.:s .:.+..,:,^ ^ _ __ _ _ _


^yr+^

A number of noteworthy points should be brought out concerning this

extension research.

(1) These CARDIAD methods have been successful without assuming a fixed
i'
form of the compensator. The importance of this fact can scarcely
} j[n be overemphasized. If a fixed fora is assumed, it-may happen that

the-form is inadequate to fit the essential plant characteristics-,

and, as a result, it can well be the case that essential insight

is ,lost,

(2) The CARDIAD approach, applicable to design localized to just one

side of the plant, as . for example the input, can be used to affect

outputs that are not measurable. Other methods that use compensation

both at plant input and plant output often depend upon moving the

output compensator around the loop---an operation which is not pos-


^.: UM

,.a sible unless those outputs drive the loop. It would seem that this
F
T^ 1^.
could be quite important_ in the case of •key outputs such as high
SirJ

f.
turbine inlet temperature and thrust.

(3) In cases studied so far, and there have been over a dozen of them,
f
t^
Jm
the CARDIAD plot has achieved dominance over all frequencies, even

when the plant has (A, D, C, D) form and transmission does not roll
t.
}
off to zero at high frequencies. This design power has been ac-
f'

^.i

hP companied by a need for only relatively simple dynamical compensation.


a„

3- (4) In practice, gain selection in compensators has to be done with some

care, so as not to invalidate the accuracy of the linearized model.


t.

The CARDIAD approach convey=s considerable direct insight into the

gains available.; and does not leave the choice indirectly to an op--

19

f
a

,. .. Ye..ut 'Vri.ilY.a1.1w•?^ .. ^;.^.:i.e`..2J.! w.W . ''! S - ^ _.^ -^K ^ -


^^ .. .. .... ^ x. F.ZI —
}f^

os.
timization program.

Because of these features, the investigators feel that the CARDIAD plot

is helping to push back the research frontier in frequency domain ap-

proaches to approximate decoupling.

3 Multiple Setpoints
rr

Frcim the outset, the CARDIAD plot has offered much promise for the
c general control problem which involves linearization at multiple set-

points, design at each setpoint, and a piecing together of these designs


^A
to achieve global. effects. Such technique is certainly the norm bo`b in

present--day practice and in current research for the turbofan engine.

0 Basically, the idea is to construct CARDIAD plots to each setpoint


L! k.
and to use these to study the interaction features of the nonlinear en-

gins model over a more globsl operating regime. The investigators be-

^n
lieve that such studies can be helpful in selecting setpoints for design

and in constructing compensation which works toward global dominance.

W6ek has been proceeding along these lines, with the aid of set-
. points involving two inputs and two outputs from NASA's DYNGBN engine

simulator. The first documentation can be found in Appendix U.

In Appendix H, the setpoints are determined by fuel flows of-2.145,

2.31., 2.x'75, 2.64, and 2.75 LBM/SBC. Figures 1-10 of the appendix con-

tain the corresponding CARDIAD plots. Consider Figures 1, 3, 5, 7 and 9,

which focus on the first column. rota their clear similarity. Next con-

sider Figures 2, 4, 5, 8 and 10. Again note their clear similarity. Be-

cause of this similarity, it was possible to design one sample compensa-


for to achieve diagonal dominance at all five setpoints. This compensa-

LL

20 ,f

y,-

for is indicated in Section Four of Appendix C.

- Research is continuing on putting together a global compensation

based.upon: these..analyses on the DYNGEN simulator.

The technique is promising and is receiving maximal attention on the

project.

^Y

^t

° -
5? -

^Yr

^^
21
IT
V. GLOBAL NONLINEAR OPTIMAL METHODS

This section is concerned with some of the details of the thirdY ear of

effort on the global nonlinear optimal part of the research. As in the

previous year, this part is primarily concerned with the control of a

two-spool engine.

There are three main aspects of the work:

DYNGEN Simulator Operation

Turbine Engine Modelling

Nonlinear Optimal Control

DYNGEN Simulator Operations

e
The DYNGEN simulator, equipped with DYGABCD, has been useful in

nearly all studies related to the grant, as it provides a "real world"


R„• testing ground for the various control methods under investigation. How-

? ever, it is costly and has limitations. Two such limitations are the fact

that it is difficult to get convergence at low rotor speeds, and that only

two controls (IM and A$ ) are readily available to the user.

The automatic generation of ABCD matrices enabled by DYGABCD has

44 been invaluable. It is felt that our work has contributed to the overall

development of the simulator through feedback provided, as for example,

the simple modification suggested in Mr. Gejji's memorandum. See (7),

^•_ Section 111.

In the early stages of the work, DYNGEN was of primary concern, but

it is now fairly routine and attention has turned to other areas.


a
I', -
III

rt

^s
^..
F-^
.22 E

s
—..._-._-- ._.... _.....
.r^.^:..—sue,.
^^^'•^^.._a;.c-: — -._ . --
Turbine rigine Modelling
This phase of the work began with analog computer studies of a single

spool engine. Then a considerable effort was spent to obtain a good ana-

lytical model for a two sD- ,) l model using fundamental physical. consider-
ati.ons. This study resulted in a hierarchy of models as reported in W.B.

Lougenbaker's M.S. Thesis, Appendix F, and the paper by Longenbaker and

Leake presented at the Pittsburgh Conference on Modelling and Simulation,


Appendix B. As indicated in the results of Longenbaker's Thr'sis, the mod-

els obtained were disappointing. Even linear affine models appeared to

fare better. See (4) , Section. III.

As a result of ; ais experience, the main. emphasis in the work has


now turned toward automatic generation of models by computer methods. The

first effort in this direction is reported in the Chicago Taternational


Forum paper by Leake and Comiskey, Appendix G. The basic approach is to

use an approximation of

x = f (x,u)

which is of the form

y = A(xM (x -' g(u))


a
This form seems to work very well for jet engine models. In the first

place, there is always a unique equilibrium point for a given fixed con-

trol u, so

x = g(u)

is the equilibrium equation which is all important for steady state anai-

ysis. In the second place, jet engine A matrices rarely have poles

(eigenvalues) at the origin and hence they are invertible..


23
i
This is a great help in computing such models. Suppose then that
L
f(x,u) A(x,u) (x-g(u)),

r. T and let

(X
e
u e)

be an equilibrium pair satisfying

S r X g(u
e e

Then
af
(X u A(x u
ax e e e e

and

Rf
(X e s a A(x u
e e au(ue)
e

N ow it is well known thatcorresponds to the approximate system A


ax
matrix in the steady state, so in our model, A(x,u) is a running approx-

imation of the system A matrix and it can thus be approxi-nated by meas-

uring the A matrix at equilibrium points of interest and interpolating.

af
A key point is, however, that corresponds to the system B matrix
DU

in the steady state, and hence

3- B(x u
e A(X u e Du (ue}
e
e
or

(u A73- N u B (x U
au e e e e e

This is where the invertibility of A comes in.

Thus, if there is an automatic method of finding A and B Tfiatrices

(as we have in DYMCD) then we have an easy way to get measurements of

24
A(x,u) 9 g (u), and (u).
a

The ^ term is very important because it is the DC gain of the linear

model from control to state. To see this, consider the transfer func-

tion relation

X(s) = (sI--A) -1B U(s) .

Then the s = 0 DC gain relation is

-AIB.
It follows from the above discussion that use of the model form pre-

scribed permits one to key in on

Authentic Equilibrium Values

Authentic A Matrix Values

Authentic DC Gain Values

for a Zpb al nonlinear model by measuring only

Equilibrium Values

A Matrices

B Matrices..

These measurements can usually be made by automated methods.

"he Chicago paper, Appendix G, was a first attempt to use the ap-

proach. J.G. Comiskey's M.S. Thesis is to use Hermite polynomials which

are well matched to au derivative requirements.

Nonlinear Optimal_ Control

It is felt that W.B. Longenbaker did a comprehensive job of refin-

ing our basic successive approximation Dynamic Programming scheme and

r,

j' 25

F1 orwsmax..wwnwwu^ra^ - _._ .^.


applying it to the models he studied. Details can be found in Tech-
nical Report Igo. EE--7714, Appendix F, and in our Semi.-Annual Status
Report for the period march 1, 1977 - August 31, 1977.

26

} x.
T
Vl. SPECIAL INITIATIVES RELATED TO GRANT 14ORK

Tvo special initiatives were carried out during this year. The first

was a special session at the 1977 .joint Automatic Control Conference, and
-
the second was an entire meeting, the International Forum on Alternatives
R-

for Linear Multivariable Control.


A "'

Joint Automatic Control Conference

A session "Turbofan Engine Control" was put together for this con-

ference. Co--Chairmen and Organizers were Drs. Michael K. Sain and H.

Austin Spang. The papers are listed below.

1. System Identification Principles Applied to Multivariable Control.


Synthesis of the F100 Turbofan Engine
R.L. DeHoff and W.E. Hall, Jr.
Systems Control, Inc. (Vt.)

2. Failure Detection and Correction for Turbofan Engines


H.A. Spang, III and R.C. Corley
Ganer.,l Electric Company
,x 3. Frequency Domain Compensation of a DYNGEN Turbofan Engine Model
Q_:- R.M. Schafer, R.R. Gejji, P.W. Hoppner, W.E. Longenbaker, and M.K. Sain
University of Notre Dame

See Appendix C.

4. The Application of the Routh Approximation Method to Turbofan Engine


Models
W. Merrill
NASA Lewis Research Center

5. Minimum-Time Acceleration of Aircraft Turbofan Engines


F. Teren
NASA Lewis Research Centex

b. Optimal. Controls for an Advanced Turbofan Engine


G.L. Slater
University of Cincinnati

KIP

1.
27
`l

f ` International Forum on Alternatives for Linear Multivariable Control

In October, 1977, Dr. Sain was Program Chairman for an entire meeting
_
s^G

fo-used in the general subject area of this grant. Approximately two hur-
E
LI persons attended from industry, laboratories, and universities. About

thirty papers were presented, many by invited authorities of international

stature. Nearly two--thirds of these addressed themselves to a Theme Prob-

lem,Appendix I, which was derived from researches on this grant. Two

publications resulted. The Proceedings contained contributed papers and


}
abstracts of invited papers. The book [121 contained invited papers and

those contributed papers which best fit in with the Forum Theme.

The Forum Program appears on the next two pages. More information

can be found. in 1121.

^n

i•:
i^

28
Thursday, October 13
SESSION 1. Origins of the There Problem
Chairman: M. K. Sain, University of Notre Dame

8:00 Engine Criteria and Models for Multivariable Control System Design
R. D. Hackney and R. J. Miller, Pratt-Whitney Aircraft Group, and L. L. Small, Air Force Aero--Propulsion
Laboratory F`

8:30 A Practical Approach to Linear Model Analysis for Multivariable Turbine Engine Control Design i.
C. A. Skira, Air Force A e ro-Propulsion Laboratory, and R. L. DeHoff, Systems Control, Inc. (Vt.)

„ SESSION 2. The-meSession A: Inverse Nyquist Array ?^


Chairman: B. Lehtinen, NASA Lewis Research Center

9:00 The Inverse Nyquist Array Method


't H. HC. Rosenbrock and N. Munro. University of Manchester, England
10:00 Insight into the Application of the Inverse Nyquist Array Method to Turbofan Engine Control
H. A. Spang, III, General Electric Research and Development Center,

10:30 BREAK

SESSION 3 - 1. Transfer Functions I SESSION 3-2. Alternate Methods


Chairman: S. Kahne, Case Western Reserve Univ. Chairman: J. Gibson, Texas A A M University

11:00 Multivariable Design Problem Reduction to Scalar A New Frequency Method for Multivariable Systems
} Design Problems R. DeSantis, Universite de'Montreal
B. D. 0. Anderson and N. T. Flung, University of
x Newcastle, Australia
' 11 :30 The Multivariable Nyquist Array: The Concept of Performance Analysis of Stochastic Linear Control
Dominance Sharing Systems: A New Viewpoint
G. G. Leininger, University of Toledo S. R. Liberty, Texas Tech University
12:00 Input Compensation for Dominance of An Automatic Depth and Pitch Control System for ?^
Turbofan Models Submarines ? ;{
R. P!. Schafer and M. K. Sain, University of V. Nitsche, K. Luessow, and G. J. Thaler, Naval r ':''
e Notre Dame Post - Graduate School i
i
^- 12:30 LUNCH

SESSION 4. Theme Session B: Complex Variable Methods a


Chairman: N. B. Nichols, Aerospace Corporation

2:00 Complex Variable Methods for Piultivariable Feedback Systems Analysis and Design
A. G. J. MacFarlane, B. Kouvaritakis, and J. H. Edmunds, Cambridge University, England
3:00 The Characteristic Frequency and Characteristic Gain Design Method for Multivariable Feedback Systems
_ B. Kouvaritakis and J. M. Edmunds, Cambridge University, England

3:30 BREAK

SESSION 5 •-1. Transfer Functions IT SESSION 5 - 2. Spectral Afethads


Chairman: B. Doolin, NASA Ames Research Center Chairman: R. DeSantis, Universite de Montreal

4:00 Linear Multivariable Control--A Problem of Stability and Homotopy


Specifications R. Sacks, Texas Tech University and R. DeCarlo,
Z. V. Rekasius, Northwestern University Purdue University
4:30 Linear Multivariable Synthesis with Transfer A Compensation Procedure for the Desensitization :>
Functions csf Piultivariable Regulator Eigenvalues 3
j ,. J. L. Peczkowski, Bendix Energy Controls Division P. J. M. Martin, R. K. Cavin, III and J. W. Houze,
and M. K. Sain, University of Notre Dame Texas A It University
5:00 Application of Frequency Domain Multivariable Linear Multivariable Synthesis by Eigenvalue/
- Control. Synthesis Techniques to an Illustrative Eigenvector Assignment
. Problem in Jet Engine Control S. Srinathkumar, NASA Langley Research Center and
L. G. ilofmanu, G. L. Teper, and R. F. Whitbeck, R. P. Rhoten, Oklahoma State University
Systems Technology, Inc.

5:30 NO HOST COCKTAIL PARTY y


ORIGINAL PAGE IS
OF POOR QUALITY.
29

s -

Friday, October lk ?^
Px
SESSION 6. Theme Session C: RegteZator etfothods gjdG11
Chairman: S. Brods".y, Office of Naval Research 'RQQg V AT .

8:30 Alternatives for Linear Multivariable Control


N. Munro and S. Hirbod, University of Manchester, England
9:00 The Systematic Design of Linear Multivariable Control Systems for the Servomeahanism Problem
S. J. Davison and W. S. Gesing, University of Toronto
10:00 Linear Multivariable Control Design Based on Asymptotic Regulator Properties
C. A. Harvey and G. Stein, Honeywell Systems and Research Center
E

10:30 BREAK

SESSION 7-1. Modalling SESSION 7-2. Modet FoZZowing SESSION 7-3. Comparisons
Chairman: II. Wozny, Chairman: W. R. Perkins, Chairman: T. E. McDonald,
National Science Foundation University of Illinois Las AL mos Scientific Laboratory

11:00 A Direct Method for Obtaining Active Maneuver Load Control Quasi.-Upper Triangular Decomposi-
Nonlinear Analytical Models of for a Control Configured tion Applied to the Linearized
a Jet Engine Airplane Control of a Turbo-fan Engine--
R. J. Leake and J. G. Comiskey, N. C. Weingarten and E. G. Preliminary Results
University of Notre Dame Rynaski, Calspan Corporation W. E. Holley, Oregon State
University

11:30 A Multi-Time-Scale Design A Parameter Optimization Method Reliability Considerations in
Approach for Jet Engine Control Applied to Engine Control Decentrally Controlled Multi-
Systems System Design variable Systems
A. J. Calise and B. Sridhar, Y. Cheng, NASA Dryden Research F. N. Bailey, E. B. Lee, and M. K.
Dynamics Research Corporation Center 4undareshan, University of Minnesota

12:00 BLS as an Alternative to Linear Model Algorithmic Control On Alternative Methodologies for
Control Systems A. Rault, J. Richalet- and J. the Design of Robust Linear
R. R. Mohler and V. R. Karanam, Papon, ADERSA/GERBIOS, France, Multivariable Regulators
Oregon State University and R. Mehra and W. C. Kessel, H. G. Kwatny and K. C. Kaluitsky,
Scientific Systems Drexel University and TASC

12:30 LUNCH

SESSION 8-1. Output FeetMack SESSION 8 -2. Addittional SESSION 8-3. Design:
Chairman: E. M. Cliff Approaches Chairman: E. C. Tacker
Virginia Polytechnic Institute Chairman: I. Rhodes, Wash. Univ. University of Houston
2:00 Output Feedback Regulator Optimal- Open Laop Compensator A Conceptual. Design Approach
Design for Jet Engine Control Combined with Riecati Feed- Using Feedforward Plus Forward
Systems back Compensator Control . Compensation
Wi, C;. Merrill, NASA Lewis R. Froriep, D. Joas, G. Kreis- N. H. McClamroch, University of
Research Center selmeier, DFVLR, West Germany Michigan
2:30 A Classical Root Lov.:3 Design Observing Partial States for Computer Aided Design of Control
4 Method for Multivariable Systems with Unmensurable System via Optimization
Systems in State Space Form Disturbances David Q. Mayne, Imperial College,
G. K. Lee, University of S. H. Wang and E. J. Davison London, England
Connecticut, M. Sohrwardy, University of Colorado and
Ruhr-Universitat, and D. University of Toronto Design of Linear Multivariable
Jordan, University of Control Systems by CIP
Connecticut L. L. Gresham, J. R. Mitchell,
3:00 Output Control via Matrix On the Design of Accurate and W. L. McDaniel, Jr.
Generalized Inverse Observers Mississippi State University
R. J. Miller, D. L. Powers, S. P. Bhattacharyya and
and V. Lovass-Nagy, ?. G. Trindade, Universidade
Clarkson College Federal de Rio de Janeiro
and Universidade Federal
Fluminense, Brazil
3:30

c-'
VII. REFERENCES

I. R.L. De Hoff, W. Earl Nall, Jr., R. J. Adams, and N.K. Gupta, "F100
Multivari.able Control Synthesis Program", Vols. I and II, Technical
Report AFAPL-TR-77-35, June 1977.

2. J.F. Sellers and C.J. Daniele, "DYNGEN-A Program for Calculating


Steady-State and Transient Performance of Turbojet and Turbofan
Engines", NASA Technical Note TN D--7901, April 1975.

3. R.W. Koenig and L.H. Fishbach, "GENENG-A Program for Calculating De-
sign and Off-Design Performance for Turbojet and Turbofan Engines",
NASA Technical Note TN D-6552, February 1972.

4. L.H. Fishbach and R.W. Koenig, "GENENG II-A 'P rogram for Calculating
Design and Off-Design Performance of T<,ro and Three Spool Turbofans
with as Many as Three Nozzles", NASA Technical Note TN D-6553, Feb-
a_
' ruary 1912.

5. J.G. Truxal, Automatic Feedback Control System Synthesis. New York:


r McGraw-Hill., 1955.

6. I.M. Horowitz, Synthesis of Feedback_ Systems. New York: Academic


Press, 1963.

7. A.S. Boksenbom and R. Hood, "General Algebraic Method Applied to Con-


* trol Analysis of Complex Engine Types", NACA Technical Note 1908,
July 1949.

8. M.S. Feder and R. Hood, "Analysis for Control Application of Dynamic


Characteristics of a Turbojet Engine with Tail-Pipe Burning", NACA
Technical Note 2183, September 1950.

9. A.G.J. MacFarlane and I. Postlethwai.te, "Characteristic Frequency


• Functions and Characteristic Gain Functions", International Journal
of Control, Vol. 26, pp. 265-278, 1977.
^v

ti 10. H.H. Rosenbrock, Computer-Aided Control System Design. New York:


Academic Press, 1974.

_ 11. "DYGABCD -- A 'Program for Calculating A, B, C, and D Matrices for


Linear Descriptions of Simulated Turbojet and Turbofan Engines",
NASA Lewis Research Centex, Preliminary Report.

3 12. M.K. Sain, J.L. Peczkows-i, and J.L. Melsa, Editors, Alternatives for
Linear Multivariable Contr ol. Chicago: National Engineering Consor-
tium, 1978.

_k

J
1.'

33_
Y
1.
Yi
?s,
i

APPENDIX A

7
L̀ GRANT BIBLIOGRAPHY, INCEPTION TO PRESENT

(1) R.J. Leake, J.L. Melsa, and M.K. Sain, "Preliminary Proposal on Mod-
ern Approaches to Jet Engine Control", November 1, 1974.
TA
`° (2) R.J. Leake, M.K. Sain, and J.L. Melsa, "Proposal to NASA for Support
of •a Fork Entitled 'Alternatives for Jet Engine Control'', November
IT 27, 1974.

(3) R.J. Leake and M.K. Sain, "Seim- Annual Status Report- on NASA Grant
NSG-3048, 'Alternatives for Jet Engine Control.'", March 1, 1975 --
u August 31, 1975.

(4) R.J. Leake, M.K. Sain, and J.Z. Melsa, "Proposal for Continuation
of NASA Grant NSG--31048, 'Alternatives for Jet Engine Control", Oc-
tober 1,. 1975.
(5) J.C. Shearer, "An IBM 370/158 Installation and User's Guide for the
r". DYNGEN Jet Engine Simulator", M.S. Thesis, November 1975.
(6) T.C. Brennan and R.J. Leake, "Simplified Simulation Models for Con-
trol Studies of Turbojet Engines", Technical Report BE 757, Univer-
sity of Notre Dame, November 1975.
(7) R.J. Leake, -M.K. Sain, and J.L. Melsa, "Final Technical Report on
NASA Grant NSG--3048, 'Alternatives for Jet Engine Control'", March 1,
1975 - February 29, 1976.
a (8) R.R. Gejji and M.K. Sain., "Polynomial Techniques Applied to Multi-
variable Control of Jet Engines", Technical Report EE 761, March 1976.

(9) V. Seshadri and M.K. Sain, "Multivariable System Compensation in--


eluding Characteristic Methods for Jet Engine Control", Technical
Report BE 763, May 1976.
(10) A.J. Maloney ITI, "Graphics Analysis of Dominance in Jet Engine Con-
trol. Models", Technical Report EE 765, June 1976.

(11) M.K. Sain, R.J. Leake, R. Basso, R. Gejji, A. Maloney, and V. Seshadri,
"Alternative Methods for the Design of Jet Engine Control. Systems",
Proceedings Fifteenth Joint Automatic Control Conference, pp. 133-
142 9 July 1976.

(12) J.C. Shearer, W.E. Longenbaker, Jr., and R.J. Leake, "An IBM 3701158
Installation and User's Guide to the DYNGEN Jet Engine Simulator",
Technical. Report BE 766, July 1976.

111

r
(13) R.R. Ge jji and M.K. Sain, "A Jet Engine Control Problem for Eval-
uating Minimal Design Sof tware", Proceedings Midwest Symposium on
Circuits and Systems, pp. 238--243, August 1976.

(14) V. Seshadri and M.K. Sain, "Interaction Studies on a Jet Engine


Model by Characteristic Methodologies", Proceedings Midwest Sympo--
:; sium on Circuits and Systems, pp. 232-237, August 1976.

(15) R.J. Leake and M.K. Sain, "Semi-Annual Status Report on NASA Grant
IdSG-3048, 'Alternatives for Jet Engine Contr'ol's Supplement No. I",
March 1, 1976—August 31, 1976.

(16) R. Basso and R.J. Leake, "Computational Alternatives to Obtain Time-•-


r Optimal Jet Engine Control", Technical Report EE 767, September 1976.

(17) R. Basso and R.J. Leake, "Computational Methods to Obtain Time-Op-


timal Jet Engine Control", Proceedings Fourteenth Allerton Conference
on Circuit and System Theory, pp. 652--661, September 1976.

(18) M.K. Sain and V. Seshadri, "A Result on Pole Assignment by Exterior
Algebra", Proceedings Fourteenth Allerton Conference on Circuit and
System Theory, pp. 399-407, September 1976.

(19) R.R. Gejji, "A Computer Program to rind the Kernel of a Polynomial
Operator", Proceedings Fourteenth Allerton Conference on Circuit
and System Theory, pp. 1091--1100, September 1976.

(20) R.J. Leake and M.K. Sain, "Proposal for Continuation of NASA Grant
NSG-3048, 'Alternatives for Jet Engine Control'", October 25, 1976.
a
(21) R.M. Schafer, W.E. Longenbaker, and M.K. Sain, "System Dominance:
A Preliminary Report on an Alternate Approach", Technical Report,
U-aiversity of Notre Dame, November 25, 1976.
d
(22) R.J. Leake and M.K. Sain, "Final Technical Report, NASA Grant NSG-
3048, 'Alternatives for Jet Engine Control.', Supplement No. 1",
IT March 1, 1976 - February 28, 1977.

(23) R.J. Leake, J.G. Allen, and R.S. Schlun,t, "Optimal Regulators and
Follow-Up Systems Determined from Input--Output Signal Monitoring",
` Proceedings IEEE International Symposium on Circuits and Systems,
April 1977.

(24) R.M. Schafer, "A Graphical Approach to System DominanceTechnical.


",
Report EE 772, University of Notre Dame, April, 1, 1977.

(25) W.E. Longenbaker and R.J. Leake, "Hierarchy of S;,mulation Models


for a Turbofan Gas Engine", Proceedings Eighth Annual Pittsburgh
Conference on Modeling and Simulation, April 1977.

(26) P.W. Hoppner, "The Direct Approach to Compensation o£ MultivariECble

ua
I

Jet Engine Models", Technical Report EE 774, University of Notre Dame,


May 1977.
(27) R.R. Gejji and R.J. Leake, "Time-Optimal Control of a Single Spool
f Turbojet Engine Using a Linear Affine Model", Technical Report EE
7711, University of Notre Dame, June 1977.

(28) R.M. Schafer, R.R. Cejji, P.W. Hoppner, W.E. Longenbaker, and M.K.
}t, Sain, "Frequency Domain Compensation of a DYNGEN Turbofan Engine
Model", Proceedings Sixteenth Joint Automatic Control Conference,
` pp. 1013-1018, June 1977.
(29) R.R. Gejji and M.K. Sain, "Application of Polynomial Techniques to
a• Multivariable Control of Jet Engines", Proceedings Fourth TFAC Svm-
posium on Multivariable Technological S ystems, pp. 421--429, July
1977.

(30) R.R. Gejji, "Use of DYGABCD Program at Off--Design Points", Technical


Report EE 7703, University of Notre Dame, July 1977.

(31) E.A. Sheridan and R.J. Leake, "Non-Interactive State Request Jet
Engine Control with Non-Singular B Matrix", Proceedings Twentieth
Midwest Symposium on Circuits and SIstems, pp. 539-543, August
1977.

(32) R. Gejji, R.M. Schafer, M.K. Sain, and P. Hoppner, "A Comparison of
Frequency Domain Techniques for Jet Engine Control System Design",
Proceedings Twentieth Midwest Symposium on Circuits and Systems,
pp. 680-685, August 1977.

0. (33) W.E. Longenbaker and R.J. Leake, "Time Optimal Control of a Tao-Spool
Turbofan Jet Engine", Technical Report ER 7714, University of Notre
F. Dame, September 1977.

(34) R.J. Leake and J.G. Comiskey, "A Direct Method for Obtaining Non-
linear Analytical Models of a Jet Engine", Proceedings International_
Forum on Alternatives for Linear Multivariable Control, National
Electronics. Conference, Chicago, pp. 203-212, October 1977.

(35) M.K. Sain and R.J. Leake, "Proposal for Continuation. of NASA Grant
NSG--3048, 'Alternatives for Jet Engine Control'", October 25, 1977.

(36) J.A. Ortega and R.J. Leake, "Discrete Maximum Principle with State
Constrained Control", SIAM Journal on Control and Optization,
im Vol.
15, No. 6, pp. 109-115, November 1977.
Ia

(37) Michael K. Sain and V. Seshadri, "Pole Assignment and a Theorem from
Exterior Algebra", Proceedings IEEE Conference on Decision and Con-
trot, pp. 291-295, December 1977. -- - -
° (38) R. Michael Schafer and Michael K. Sain, "Some Features of CARDIAD
Plots for System Dominance", Proceedings IEEE Conference on Decision,
1e

is
And Control, pp. 801-806, December 1977.

(39) M.K. Sain, "The Theme Problem", in Alternatives for Linear Multivar-
iable Control, M.K. Sain, J.L. Peczkowski and J.L. Melsa, Editors.
Chicago: National Engineering Consortium, 1978, pp. 20-30.

(40) R.M. Schafer and M.K. Sain, "Input Compensation for Dominance of
Turbofan Models", in Alternatives for Linear Multivariable Control,
M.K. Sain, J.L. Peczkotqski, and J.L. Melsa, Editors. Chicago: Na-
tional Engineering Consortium, 1978, pp. 156-169.

(41) J.L. Peczkowski and M.K. Sain, "Linear Multivariable Synthesis with
r. Transfer Functions", in Alternatives for Linear Multivariable Con-
trol., M.K. Sain, J.L. Peczkox qski, and J.L. Melsa, Editors. Chicago:
National Engineering Consortium, 1978, pp. 71-87.

(42) R.J. Leake and M.K. Sain, "Semi-Annual Status Report, NASA Grant
NSG-3048, 'Alternatives for Jet Engine Control', Supplement No. 2",
March 1, 1977 - August 31, 1977.

(43) R.J. Leake and M.K. Sain, "Final Technical Report, NASA Grant NSG-
3048, 'Alternatives for Jet Engine Control', Supplement No. 2",
March 1, 1.977 - February 28, 1978.

F,
TI-

as
as
it

"HIERARCHY OF SIMULATION MODELS FOR

A TURBOFAN GAS ENGINE"

^ ^-, 1.

W. E. Longanbaker
R. J. Leake
i ,;

t'

"pWFL Q,UAIJT% s
i
EIGHTH ANNUAL
PITTSBURGH CONFERENCE
i ON
MODELING AND SIMULATION

i APRIL 21-22,1977

Senedum Engineering Hall


Lln[vers4 of Pittsburgh

Ahore Than 220 Papers On:

e SOCIO-ECONOMIC • REGIONAL PLANNING


SYSTEMS 9 BIOMEDICAL SYSTEMS
e ENVIRONMENTAL o ECOSYSTEMS
SIMULATION • URBAN MODELING
e SIMULATION IN • ENERGY RESOURCES
EDUCATION a TRANSPORTATION
a ECONOMIC MODELING u AND MANY OTHER
e MODELING AND SESSIONS
ESTIMATION

Sponsored by
I
Department of Electri :a1 Engineering
i School of Engineering
U n vers,ty
tfP o a sburgh
Pittsburgh, Pennsylvania

In Cooperation With

The Pittsburgh Sections


of the
Institute of Electrical and Electronic Engineers
and
The Instrument Society of America
and
The Systems, Man and Cybernetics Society
The Society for Computer Simulation
The International Associu •.'on for Mathematics
and Computers in Simulation (formerly RICA)

Conference Committee

Michael A. Shanhlatt Martin Savo[ C. P. Fong


Leone Monticane Chin Shlek Jang G. Lee
Chang-Eun Kim Khe-Tai I'.ta .;uk H. Lee

1 Aly Abullell

3
Conference Co-Chairmen

t William G. Vogt Marlin H. Mickle
t
x

7 r s rrir: %rf:re For S- 17 j 31 Fa--%

° T i
: LTERARCUY OF SItNL °1TION MODELS FOP.
- .111.VMFAN GAS ENGINE a:

d. .. ^.". .. iJ E-Iongenbaker- and R.J. Leake n 4t;t'.i:n


. .=Tepartment of Electrical l ineering
! -:Uaiversity.of Notre Dxz
' 23otre.Dame, .Tndiana 46.556
FOOR QUALITY 1
y

"' ST GT +Model-IL3. This is a normalized 3rd order linear


­ =del obtained by a hand calculated order reduction f'
• .SLh-is wort_ is•a comoarison of successively more _-of Yodel IL5.
.•_^comDrPhensive simulation models of an F-100-li.ke
= rlofan jet engine. :L large and elaborate com- -.Eodel. 2L5. This is a •normalized 5th order linear
_.^ ^puter program called DYNCEN , developed over a .model obtained by taking partial derivata:vea of the
. _nu aer of years at 1:15. Les.-is Research .Center, is --malytical tfadel Z..
S .Aoemplcyed as the mast comprehensive model for an
• —alyzing steady-state and tranaicnt performance for Model 2L3. This 3s'.: normalized 3rd orderl y ear
-.,=ntrol studies. This model employs =ny block •model obtainai&'by a hand calculated order reduction
_ .,,-rdata maps and includes about 25 states. 1a order of Model ZL5.
to perform oatimal control studies, low order nor-­^ a
Blear analytical, and linear models have been ,..WALY11Ca XODtL
_ - •;s
r eveloped. This paper reports on the details of
' ..these models and pre_ents e%perimental data on .-In this section we discuss and present Model 2., a
.=^z relative performance. ,. :-s3supl.ified nonlinear analytical model of the jet
' :-TNiTiODUG1IOk1 •- -eng•ine. There are several reasons why it is desir•- s
--,+ ' .abl eto have such a model. First of all, one likes .1.
36 .LI+is paper we-consider the determination of a • =to -see the basic nonlinear relationsUzz between
--s;mplified monlinear analytical model for a two the engine variables in order to gain insight to
'Z. -spool turba:aa jet engine. A large and elaborate their dynamical and static behavior. Secs.dly, 1
. . t (about 4000 statements) generalized engine such a model is invaluable in the application of
ea ,., ..:simulatar called D!%GZN [1,2] ceded w:Lmh represent- _optimal control techniques [4) to engine control s
-nt{ve block data =aps, dasign parameters, and two systam desi g n. In the third z1ace, i f the model isj
_,,_spool operation is taken as the principal object I reasonably ace rate, it can. be employed as a east ^a
-to be approximated. First we present the various inexpens.va nonlinear engine simulator for the
---,i dels and then perfc=arce cemuariscros are made, evaluation of linear and ronlinear con_rol strategies.
^5he models considered have been en+aaerated as fol- Finally, linear models obtained by partial di ffer-
entiation o= this model tend to have more st_ucture
' ' (zero entries in the ABCD matrices) then r-*nose. ob-
'_fodeZ 0. Thy actual jet engine (hypothetical.) Lamed nst=erically, which gibes the linear designer {
more insight. rnese linearizations then serve as a 1
.Model 1. The DY.1GM simulator, coded kith data -pre- E back up to compare with the numerical linearizations.
I st—d to have been t6-en from experimental measure- ! 3
-°-.->n!°nks on Model . 0.. This model solves nore than 16 1 1
nonli near differential equations and uses data maps Mm the detarminatson of Model 2. as an approxi--ation
nud thee-odynami.c tables which cannot be expressed of the DYNGPN Nadal 1., theoretical rzlatimiships
•an.!yt:ically. .,developed in [53 J6], and [7] were employed as a
- starting point. Certain sim p lifications suggested
!Model 2. This is an analytically expressed set of in 18) were used; and linear, least squares, ex-
'3 Annlinear differential equations plus about 20 ponential, and polynomial fits to the Hadel 1. data
• ,ata:tic equations expressing the relationship be- about the chosen de31gn point were bade.
.c een various engine var?ables. The =,aifi task is t
;tbis ' proj ear. was to decermine this model. 3 Il 2 input, 5 state, 2 output aodel bras developed
with the following variable designations.
-'Y3ode1 ILS. Thie is a Normalized 5th order linear
I.
u ^, fuel flow (4,.F8) iF
- •• ' • ^•'rmodel-shich is cbtai ned r. :merically frcm Model lr_ Z
. -casing an experi.mantal versica of a prosraa [3) U tussle area (A8)
;-being developed at NASA Lewis Research Center. ..ir+ ^1-+R.compresaar rotor speed (u

-


1 r^' Si. i'^ e Y 9 Print $.-fj.,e For 8•:{? Y '

fan rotor speed (; F ) .- . '(2) 2 + 214.2732 Cti'F_ - 48.0(A 8 - 2.948255)


.173 ^• burner exit pressure (P4)
=11 `_T

...(3) CNC
HC
-
HC
-
-A,4 after burner exit pressure (P7)
]iCDESIC\ 1 1 21 /T 2 u 10070 /T21/516. b6S
:% -.high inlet energy (U4)
T1 + thrust (FG) ;.(4) 7, - T21 + 743.2122 C.:C 2 - 68 (AS - 2.94S255)
"72- high turoine inlet tesperuture (T4)
.__1(5) T4 - U4 /cn
.-.=e model is completely determined by thQ -folio -ding
^(6) T50 - .727 T4 :.y
•.-mpecifications.
_-Constants - -,.(7).P 3 - 1.05944P4
_J - AJ - 778.26 H = 20.713.75 -.{8)
P
21. - -6.20568 + _0129774 '21 - _0185376 P3
- ,L$ J 2.948255 CPC .24 {y) WMX - 3.516739 C%'F - 63.916
..0252CPF 4 .24
-(10) P rL^x - 3.516739 CNF - .23561
.20279
^i
1.4 CVB s
P2 1.
CpgT - .22589 _ (]1) WAF - WF + 28.502 11 •- 2. 3 132681 pF`kx t 21 ) )
- 2 518.656 CPLT .27938
(1) W(
i.AX - 137.54 - 457.987 C::C + 564.317 5 CNC2
' 3C - PHTXa - 3.8 - ^ X 6 .16 - PCBLC
1E8.113 LNC3
V
ar - Pri -12 - 4.5 a - - rC6LDU - .208 -.(13) db: L1X - 6.492 - 4.9747 C,NC
[ *f -.77 - 1.65 g - PCBL:m - .726
(14) 26.43184 -- 89.0484 C".0 + 109.7234
I P01AX
49.77 -7=- PCBLLP -.066
VdFli^t ' C,NC 2 - 36.5756 C.NC3
_.:XC -I1£SIGN - Y:.'HPOS - 10070
:31F DESIGN - MC LP05 - 9651 _.) LAC - _P?1 {w -151.' {1-e-.3662
...121/518.668 C'L1\ CpAX
^" Resign Values
(P CSf - 3) )^
;_ibB - 2.?5 VC - 1899.2 FC - 1:1431.02 P21
,.__ dB - -2.948255 NF - 9873.94 T4 - 2982.04 (16) FA3 (1- p^l'A.0 - .84 WAC
`A P4 -.23.9299 --(17) WG50 - 301.957 P4 W-7-
4-
-J4 -'586.467
--(18) WG4 - ?:C50 - 86WAC p WC50 -- .316164 WAC
Y7 - 2.55007
(34) WG55 - VG50 + y^WAC - WC50 . + .01056 WAC
state £quatfons
(20) 155 - 106.002 + .85154T 50 - .10458C'rC rT2
-(s 0 ) 2 -
(1) d!
`C
[CPC1AC(T21-T3)
.j dt C1'C (21) T, - (T55 + 414.592P7)
+ CRHTWG50(T4 - T50)) 2.01365
WiC7 - 1i7.1.784Y7A8
'(22)
(2) dNF 30 2 J
. ' dt u ) T^v F [C PF W(T 2 - T21)
+ CPLTU-u55(T50 - T55)) (23) VC - .02951 i.'G? 193 + 685.8.315
+2116.217A 8 (.539'r5?•7-1)
[T4(WM + WFB - WG4))
':(3) dt yYa
. CO'iB
^/r^wwh LNw`
dP
RyeT MN-
•--(;) 7 T--- 7 - J C'r - WA3 + WAF i:G7 ;
'
dt VAMN
dU4
(5) 1' BP'T4 [T (WG4 - WEB - WA3)
•.^€
dt VCOHBF4
•^' + -1*(T 3WA3 - T4WG4 + T4 (1 + t1)WFB))
:.

} Nonlinear Funcriont Required for State Eoueticns


'^+ v and Outnuts Figure 1 - Two-spool, two-streaw
hF -WF_ • -turbofan ermine

f•: r' • J^1-i^

• r ^r!r.: .:^!::. fir 1!2 x 11 Pa-


_ i _- _... ^.^. _ .._.. , .
.,,, • I MEET. aa: „MT► .,.i • ni .•,. n.,7 ^^ r-8.42?5 -1.:54 __.^ -.047955,
_ .:-
?!A Mr flow --- A- 2.3957 -5.9"244 .0048499
=VC Gas flow L-1.11.51 -97.919 -74,905
F Pressure -
T ..2esperature 3.4058 ...79723 ,c1
-V . Internal Energy .-.S+ 2.1242 ,,56146 O
200.75 - -39.261
vp
=r.Humbers refer to the location within the engine 0 1.461 j ^j
_.-(figure 1). Ali variable na=es correspond to those -•63248 -,97908 -.01486 G
•-vbieh are used in DY4Gr_V
-^ Q ..1.2138
.;^ MODELS 1.072 -.15978
igesvalues: - 74.976, -7.1367^1.17171
•:2be following linear y odels were obtained,
•jvith thrust, (FC), as the output if there is only
.tine output. .=PFFJ_X 4'7AL CO":PARISONS

• } ._ Model 1L5 - Figures 2 and 3 represent the nonlinear steady


-state operating lines of the fan and compressor
( -3.80 -1.277 2.067 -1.152 1.448 respectively, as both nozzle area and _f uel flow are
2.748 -5.39 1.585 -1.991 1.071 changed. This is a detailed map in the vicinity o=
377.9 49.51 -264.9 86.807 78.91 our nominal engine design configuration. The design
31.26 139.39 -6.269 -88.69 27.33 point for model 2 differs from that of model 1 by
_23.91 -10.27 -37.40 -246.7 ,less than 1/2S.

}
.U259 :355326
.2.116 -.316176 Hodel 2
_ 12.54 -13.7828 -Vary nozzle
- --.6201 --94.388 r area
157.783 6.84396

^^ .8594 -.1397 .6672 1.167 .1236 pressure


.055591 .00656034 -.0018374 .0135393 .853911 ?atio• Yodel I
..-
-.102766 -:900938 (F21/P2^; I ry nczz_e
^ -.013839 -.020856
Yodel 1
vary fuel flow
^e a
: ­ AHZenvelues: -251 + 23j. - 96,.-5 + O.8j ^ //lSodel 2
vary fuel flaw
'Mod- 1 1L3
A- 3- .
2 : 43Q7 -.70897 -.81149
[, 1.395 .348751
8261 -4,9579 -1.7_35 1._585 .279331
" 2.4466 140.5 -94.982 15.434 -)8.209) xs4 218 z
2: 2=6
Pan Air Flow (WAF) (ibm/sec)
C_ D, ?lgvre 2. Steady State Fan Map
J-.1543 .013382 1.333] .2346 ,87572]
Eigenvalues: -92, -5 + 5j -
Model 2L5 ^fT Model 1 .
i \ ^ary nozzle
-12.5487 -1.59279 3.58369 .237960 1.70112
.833048 -5.51346 1.64496 0.143505 1.30396 I ^" area f
A. 671 .604 337.711 -392.675 -26.1602 150.5'_ &6
-104.155 21.1351 64.9255 -67.6031 2.63144 Model 2
+ 50.9527 -55.8546 -81.2047 -7.47450-105.743 Pressure war; fuel flow
Rat3a
0.0 1.40762
• 0.0 .758167 3 21
('%
1.28129 -122.314
Model 2
0.0 -48.9280
vary nozzle
t 149.210 -3.09196
^1 area
lo- !`adel I
0.0 010 - 0.01.46046 O.O^
0.0 0.0 0.0 0.0 1. Q vary fuel flow
0.0 1.21183 air
0.0 .se
1 0.0 'Compressor Air Flow ('-AC) (Ibn/sec)
i
W genvalues: -343, -154. -73, -7.2-+ 1.4j ""719ure 3. Steady State Compressor Hip
E
i#

t^
i
t

f
T,
^.

the time responses of vnrious states due to S.


-step inputs is f ue?. f.ov and nczzle area arc shown G D no del 2
In Figures 4 and 5 respectively. Only the nonlinear
models are represented. Although the deviations of ` t
.the states from their design point differ by as Hodrl •rnd^1 1T5
•uch as a factor of rco, the actual values (design
-.much Ao
_.-Plus deviation) of model 2 remain within 1^ of the T riodel 1
_w del 1. values. For steps in fuel of nit:us 201, Ch;^t,gn
--=the-states remain within o: of each other. ;Off Lo I,
• ._^^.Denign

T - U4 Y_ociei 2
-1 .4
ORIGINA p Time (sec)
L
AGE IS 23-Sure 6. Thrust Time Response -
-^2Ygure
'—"aa Fuel Step Input
-A Comparison of the linear system frequency
C= , Yodel 1 s-•respcnses of nozzle area to thrust, ::nd fuel _`low
.fhangc ^^ -to thrust, is sho p-: in Figure 7. :todels 11.5 znd
-.2L5 hatch extremely well for the fuel input, but
:off
E^..Dealgn
so ,^ _ not so well for the nozzle area input. Even the
' - lower order riodel 11.3 is closer in this case.
N c Model 2 --
1e N liodel 1
c,

,z .4
-Tize (sec)
-;ligure 4. State TIm`. Responses-Fual Input
1L3 i

.•.ta P 4 Model 2
1L5 i
` L* Model 1
4
1

+
., ..

Change 1L5
f
Off I
$.5
Design '4

Nozzle Area
Fuel flow---o+.

1
,
3
P 7 Model 2 j
• ^, n" e ^r o i
,^ ^, ro z io
-figure 7. Frequency Responses-Inputs
to Thrust
P 7 Model 1 -
-+ - Si K4ARY
•2 .♦
Time (set) To the best of th e cutho-a = knowledge, no non-
f^f .fi ure 5. St a r e Tice Fesponses-Nozzle
- ga- ^,odels of a two-spool,
Lir.ear analytical dN-nic
Area' Input Step two-'stream jet engine have ever appeared in l itera-
ture. 'Indeed, it is the value of the develo:.ment of
such a model which is the most iaaortant cens_dera-
'The effects on thrust by a 51 step in fuel flow
tion in the eva].unt:on of our vor:, i.e., that a
- ore shown in Figure 5. Model 2L5 yields results
-good nonlinear anal y tical dynamic r.ocel %;ill provide
%^whieh are quite close to =oriel 2, although, both
a flexibility and cse'• ulness wnic'i is non-existent
mire 61Eniflcantly di iferent from eodel 1.
do present non-analvc:C-11 jet engine siaulaticns.
.11thouf:,, som-, sigoi_icznt discrepancies exist, our
•_^=adcl yields results waich_are . accurate to within 11
C

..tiear the design point, and which degrade to an ac-


curac y of cpproxiz-atel y &:: %:ith a drop in full flow
Of 20- the frequency response for fuel inputs of
linearized models is also in close agreement.

• t _Xn eouclusion, we are encouraged by our overall


--progress towards the dc% , cJep:zent of the analytical
_..model, however, we feel that mo re wort: is needed
,._to JurLher improve the accuracy of our model.
Y-:

This research was uupported by NASA grant


x^SCr-3048.
::J EFEREt:CE1

y .[1] R.W. Koenig and L.H. Fishbach, "CENENG - A


7!7ogram for Calculating Design and 0=F-design Per- .
-fomance for Turbojet and Tur:,ofan 1--igines", 1:ASA ^, P
.--M X6552, 1972 'BOOR Qr^^GE xS
'-12] J.r. Sellers and C.J. Daniele, "DYNGEN - A . ITi,
Program for Calculating Sceady-state and Transient
--Perforrance of Turbojet and Turbofan Engines,"
.,- TN D-7901, 1975

"j3] L. Geyser, "D'_G?.JCD - .1 Program for Calculat-


.-. ng Linear A,Z,C and D T:atrices from a ::onli-car
• Xy-uamic Engine S'__ula;.or", Technical =eaort an
!I ...preparation at NSA Lewis iesearca Center, Cleve-
•.]..and, Ohio.
14] R.T. Sasso and Z.J. Leake, "Computational
-Methods to Obtain Ti-e Opt"- 1 Jet Engine Control,"
bI3erton Conference on Circuit and Systca Theory
1976.
.15] F ScIdner, L.C. Geyser, H. Gold, D. Walker,
.,=Jmd G. Burgner, "Perfornnnce and Control Study of
t
I -.is Low-pressure-ratio Turbojet Engine for a Drone
--Aircraft", NASA TY Y-2537, 1972 i
r
[6] -K. Seldner, J.R. 'iihalcev, and R.J. Blaha,
r ..".Generalized Simulation Tccnique for Turbojet
Yngine System 3.nalysis", N SSA rN D-6610, 1972
...1
,[7] Zucrva, "1.3., Aircra;t and `!fissile Propulsion,
New York, John k'iley and Sons, Inc., 19:8 d
[8] T.C. Brennan and R. J. Lcake, "Simplified
Simulation 1!3dels for Control Studies of Turbojet 1
E Engines", Technical Report No. E=-757, Department }
of Electrical Engineering, University of Notre
Dame, Notre Daze, Indiana 46.556

f
..
n 4m

in

' Appendix C

+Y "FREQUENCY DOMAIN CO1IPENSATION

OF A

DYNCEN TURBOF,L-N ENGINE MODEL"

A
nv

^ _
T
R. 24. Schafer
" L., R. Criii
P. W. Hoopczcr

M. K. '•: a t n

a,

,.2

I
3977 Joint Automatic C.Ontl:ol .0011fc1rence
San Francisco, California
June 22--24, 1 977

FRFQUFNCY DOMIN COM1PF.NSIMON' OF A DYNCE-N TURBOFAN ENGINE MODEL*

i R. f1. SCRtFE:R, R. R. GFJJI, P. W. 11OPPNFR,


W. E. I.O:,GF.;:L• 11;ER and 'E. K. SAXN
Department of Flectrical Fnginneering
Univcrrity of NOLre 'Dante
Notre llama, Indiana 46556

.Abstract elements of the return difference and it's deter--

r- kollouing Fosenbrock's ideas regarding the advan-


rainant. Typically-, when it achieves user satis-
faction, this method does so with greater speed,
and fewer concepts, than it's competitors. The jet
tages of dominance in linear multivariahle control
systems, a now graphical technique is used for the engine model is introduced in Section 2, which also
design of compensators that achieve dominance. The establishes the notation for succeeding sections.
technique is illustrated with an ap?lication to the Finally, in Section 6, we give results of simula-
problem or designing compensators for a linear tur- tions to evaluate the performance of the system.
L^ bofan--engine model. The resulting design is put in-
to perspective by examining it in the light of two 2. Jet Engine Model and Return Difference
other multivariable frequency-domain methods. One, Determinant
Mac1'arlane's method of characteristic loci, is used The linear model used for the study is based upon:
is to realize a final design for stability and low
Interaction. The other is a direct technique based
data obtained from a DYNGEN sKulation. It is
specified by the equations
upon the alE;ebraic expansion of the det a i-minant of
the return difference in terms of it's elenoi.`s. sc -.AX -1• -BU (1)
Results from simulations carried out on the 13AK Y - ox + Du (2)
DYNGEN software are included.
Where x, u, y denote the state, input and output
vectors respectively. The inputs are fuel flow
and nozzle area; the five states are compressor
Introduction rotor speed, fan rotor speed, burner exit pressure,
afterburner exit pressure and high pressure turbine
Xeccnt years have witnessed a renewal of interest in
inlet energy; while thrust and Nigh p ressure tur-
i • frequency detrain design methods for linear multi-
r bine inlet temperature constitute the two outputs.
f variable control systems. The p reponderance of
these ideas are closely related to classical Nyquist We nest consider the problem of designing cos tro1-
constructions on the determinant of return differ- lers for the plant. The underlying feedback co:v-
- once. In this paper, we use three such methods to trol schepa is shown in Fig. 1. G(s), the plant,
design a compensator for a two-input, five--state,
two-output linear model of a rodern two-spool Lurbo- 4+ c u Plant v
fan jet engine obtained fro gs the DYNGEN digital jet g 1:(s) --9 G(s)
r
engine simulation.
_ L

Rosenbrock 11) has related the classical Nyquist


construction on the determinant of return differ-
ence to corresponding classical constructions on the
diagonal elements of the return difference--pro-
vidcd these diagonal elements "dominate" their rows Rig. 1 Fccdback Control Scheme . s
or columns in an appropriate manner. Focussing the
design interest an achieving dominance in this
represents the jet engine y odel, that is,
sense, Se tion 3 presents a new graphical technique G(s) - A& - A) -1 B + D. (3)
to help with this aspect of design. Next, Sectlon
1:(s) represents the rational compensator to be de-
4 utilizes the generalized Nyquist plots to obttin
signed. G(s) can be computed as:
an acceptable compensator design. The ideas of
r,cneralizeJ Nyquist pluts ware introduced by
--502s 5-3.64E5s 4 t 4101,s5+1.93E6s4
McFarlauo [2], who related the 0 0tcrminant of the
• return difference to it's spectrum when regarded as -1.78E706.15E9s 2 1 0.69M -1.44E10s'
an appropriate linear operator.
+1.26rUls+b125cl1 A.7700_ +6.59E10
In Section 5, we utilize a direct technique which
-14.554s 5 +1.33F5s 4 20.46s3,1.69£4s4-
emphasizes the` algebraic relationship between the
^ 14, Phis n;ork^Ts supported in part by the National
+4.96E7s3+3.7£92 j +7.03E6s3+1.12E.9s2
Science foundation under Grant LNG 75-22322 and in -}•202M OV4. 95F10 10503E0+5.41 FR
Part by the N.iLional N-.• r 0naULiCS and Space s +609.57s t 1.164Ei5s +7.31.6u
Administration under Grant VSG 3048. • +602E7s+1.62E8

• l .

• E
.VRIGINAL pA ,F IS

Ceatral to the application of 11yqulst type ideas to xi and y i lie inside and dashed circles if the
rultivarlable sy5tvml is time return difference ma- acceptnhlc region is outside. 4
trix, which in this, ease hecoi-rs (1 G(•)f;(s}]. if the above procedure is repeated over a range of
It's: ptincipnl use arises from the rclatlnn of the
ces of each row the s yste m, and the
closed loop characteristic p•olynonial (CLC?) to the
circles
ircluosof intersectiono drawn,
wn, a plot
out
open Hop characteristic polynomial (00?) which
number x i +-
the acceptable valu.^s of the coreplex nr.:ber
can be stated, in the stammer of (I], as
jy
i for each frequency results. In this way, the
CI.CP Acceptable Cat;re of the function 5 (s) such that
det[SG(s)K(s)], (5)
0lCl' the i tlt row of k:(s)C(s;) is do tina 1s doscribad.
where equality is understood up-to a real constant. The analysis of the CARDIAD plot for a given ro v
Of primary concern Here is the behavior of dcL(IAGK) of G(s) proceeds as follows. If the origin of the
for values of s on the standard Kyquist contour plot is contained inside all solid circles and is
(S`:C), which encircles the open right half plane excluded by all dashed circles, the row of G(s) is
clockwise with indentations into the left half plane dominant uncompensated. If the row of C(s) is not
around poles and zeros on the imacinary axis. In dominant uncompensated, the CA`:DIAD plot is next
practice, plots are made for values of s on the checked to sce if there is a constant entry S ; that
positive imaginary axis. Srahility can then be de- will rake Qs)G(s) dominant at all frequencies.
termined from plots of det& GQ in conjunction with For this to he the case, there must be a point on
knowledge of the open loop characteristic polynomial. the real axis that is included in all sella circles
Also interestin g,, of course, is the use of such plots and excluded by all dashed circles.
to aid in the choice of a suitable K(s). th
If therS exists no constant n. such that the i
CARDIAD Plots and Dominance [3,41 row of K(s)G(s) is dominant at all frequencies, the
3.
CA11DIAD plot is used zs a guide to design a fre-
The CARDIAD (Compensator Acceptability Pegion for quency dependent g .(s) that rill achieve dominance.
11
D JA gonal llo^finance} plot is a graphical approach to This Is accomplished by realizing a function 5. (r)
the problem of choosing a compensator that will whose value at jw lies inside the circle asnoci-
achieve systems dominance. A system is said to be ated with the sam" fre q uency in the CARDIAD plot if
row (eoluc_t) dominant (11 if the magnitude of each that circle is solid, or outside if that circle is
diagonal element of the open loop transfer function dashed. This approach is illustrated by consider-
taatrix is greater than the sum of the magnitudes of ing the DYSGES problem.
the off diagonal clexents of the row (colu=n) at all f
frequencies In the 2x2 case being con s idered in 5j
this paper, the dominance condition reduces to the
•, pagnitude of the diagonal ele-ment bein, greater
than the tta.^nitude of the off diagonal eler%o n.t of
the row (column). Consonant with the Rosenbrock 3j
approach, the CARDIAD plot analysis is applied to
the inverse of the punt G(s). As a notational
point, the inverse plant transfer functici_ matrix
will be dcnoted . by G(s) and the inverse pre- Oj
compensator by 1(s).
The specific application to the jet engine design
problem involves trying to find a co:Tcnsator i(s) —3j
such that K(s)G(s) is row dominant. Without loss
of generality, the fora of 1:(s) will be restricted
r
to
—5J
L
^'(s)
1 . 1!2(s) 1
where
s i ( s ) - x i (s) •1 jy (e), 1 r 1,2. (7)
J. —4 —2 —4 0 1
If As) and 9(s) are each evaluated at a frequency Fig. 2 CARDIAD Plot Row 1
w , the equation for domunance of the ith rota of Uncompensated
o
K( ::)C(Q becomes a function f.(x ,v.)•.;hich describes
a paraboloid in three-space. t The itersection of The initial CARDIAD plots of Us) indicate that
this paraboloid and the com p lex plane is a circle row 2 of As) is dominant uncompensated since the
which is the locus of the values of x and y { such plot consists only of dashed circles, which all
:j that the ragnitule of the diagonal element of the exclude the origin. the plot for row 1, however,
i th row of Q s)G(s) is equal to the magnitude of shows that this row is not dominant uncompensated
thin off diagonal element of the row. Minima and and also that there is no constant entr y in the off
Maxima analysis of the function f. rcvcalr that diagonal cic;:cnt of row 1 of ,(s) that will make
values ofx I and y on one side of the circle will the row dominant at all frequencies. This is
e make the system dominant, whereas valuos which lie easily seen since all the circics in this plot are
on the other -ids of thr circle will not. Ia the solid and there is no point on the x axi, that is
CARDIAD plots, this diffcrcntkntion is wide by Included in all the circles. Moreover, the plot
F^ draviag ,i solid circle if the acceptable values of hints that Kett will be difficulty finding a dl(s)

G1 ..!i . to :, f':?

f._... _. - -- .... .. _...._..^ ..^._ .^_ . t ... .may.._«. ^---^._ ^.^. ..^..........__._ -..r.^...^
3.3 j

Inc IT ..N!

0j ^.- T:..' Oj

l \\\Ile
i
-3.3j

-.12j
-1 U 3.4 6.7 10.
-•.12 -.07 0 •07 .12
"Jr.. 3 CARDIAD Not Row 2 CARDIAD Plot Row 1
Fig. 14
Uncompensated
Compensated

that will nal_e this row of R(s)C(s) dominant because 16j


Of the coziplcxity of the plot and the small radii
of the low frequency circles which necessita ,-e a

eery clo;:e fit. 9.6j f

To facilitate the process of finding a compensator


that will n3ke K(s)G(s) dominant, the system was
first precumpensated with

K, a
11, 01 . of
Space limitations do not allow the CAr,DIAD plots of
Kl G(s) to be included, but the new lots are the
same share as the C.'RDIzD plots of G(s) with two
ttajor chnngcs. The row 1 plot of IK 1 (s) is the
same shape as the row 2 plot of G(s) with dashed
circles cliangcd to solid circles. Similarly, the -9.6j
row 2 plot of K C(s) is the sane shape as the row 1
plot C(::) wi ) h the solid circles changed to • n9

dashed circles. -16j


The problem of finding a R,,(s) such that K,,(s)1,1G(s)
-16 9.6 .15
Is do.A nant is now simplified. Since row 2 of
K1 G(s) is now dominant uncompensated, the off diago- Fig. 5 CARDIAD Plot Row 2
nal term in the second row of K 2 (s) is left a zero, Compensated
with the provision- that if it later proves helpful
't.
t_-[
at w-10, is regained as v--90,. and is lost al^ain at
In corpensation, the entry may be chosen to he any w=700. It is perhaps possible to.find a better
constant that lies outside all of the circles. To

F
choice of K (s) that will make row I duzir.ant at
make row 1 of f a ((s) dominant, the off diagonal all frCn. U0T1CiCS, but the de^. ,.inancc achieved by th-.
entry in row I of K 2 (s) must follow the semicircular above K2 (s) proved to be sufficient.
path throttFh the Coc:-vlcx- plane described by the
CAPM AD plot for this roa. A fit was made to this An interesting, feature of the CAP,DIAD plot is

r
I
shape and the resulting; V 2 (s) was l lustrated in the fir.al plot for row 2. Close
analysis o - this plot sl=ows that there are three
l 9.4798 + 0.'49St occurrences of solid circles changing to dashed
1 1 1. - 1.2359s circles or dashed circles changing to solid. then
(s) (B)
2 0 1 these transitions occur, the paraboloid is invert-
Y^-
ing and the circle of intersection degenerates to
The CARDIAD plots of X 7 (s)K I C(s) are considerably a line. These lines occur when the otter row
more complex tlinl the previous plots. Tllc plot for changes from being doninnnt to not dominant or vice
row 2 shows that the row is dominant at all fre- versa. Thus, each cl'anpc in dominance of row 1
quencies since tha origin is included by all solid causes a change in the type of circle being drawn
r
circles an.i excluded by all dashed circles. Ilie in the plot of xow 2. ,
CAkl))Al) plut for row 1 shows that the row is clearly
riot dominant at all frequcncics. DocainancC is lost ..1.

4. Design Vsing Characteristic 1.nci high frequency r-%fnitude of q 11 (:;) while simultano-
ously boosting the Iew frequency ntal •,nIttide, 1h I;
Another approach to design, due to A.G.J.'1.icFarlane,
Is in accordance with the freedom specified fcr K2
uses lltc IOCnS of the elgrnvalnce: of C(_:)!;(s), call-
prcvlously, the riodified 1:(s) gives rise to the
ed the chatactc • ristic Ioci (C.L.), for V.IIUe:; of s plot . of Fig. 6 and 7.
an the S':C. 'this method is barccd on the rcl.tlion
of the detert.i(nant of the return difference to Since dot.d nance is not affected by diagonal cer?en-
cigenvalues of G(::)K(s). In order to assess st..-- satovs, ;he problen becomes that of inc:cpettduntly
bility from the C.L. plots, for s-^L • , w pof;itive, shaping q 1l and q2, by means of single loc•p tech-
one vas t count the clockwise encircicr:.nts of the n 4 1ues, n order to reduce high freluency Fait: in
critical point (-1,0) :ztde by the C.L. plot_ and q22 without appreciabl y affecting; low frequency be-
sum all these up. Me closed loop system is stable havior, we use lag conpennation. A little bit of
If tills suet equals -po, :here pa is the number of cut and try led finally to K
terns of OLCP enclused by the SNC. As an a,prosi- (-s-0.1)/(2000s +70)). Tile pots corresponding , to
Mte pleasure of interaction, we compare the eif,cn- K=K1*12*K3AK, are not sho-- ru. The 1. 1 and e ll1 plot
value plots with plots of the diagonal elements of is es.enttally that of a in Fig. 6 scaled 6y a
1. Q(s}=G(s)h:(s}. For a noninreractirg system with factor of 0.440-4. SiMdarly, the q,,,, plot in-
Q(s) a diagonal sratrix, these would be identical. verted and scaled by 0.0005. (By inersicn we
In our design cxanmple, Q(s) is a 2x2 matrix, and mean reflection throur,h both axes.) The plot for
therefore ve trill be looking at plots of two cigen- A 2 is shown in Fig. 8.
velues a (s), (s) and the two diagonal elements
g ll (s ), 472(').2 5. The Direct Kethod of Analysis
First, an examination of the C.L. plots of the un- Direct methods of Wultivariable Nyquist analysis
compensated system revcr.led th.'tt, without compCn- concern themselves with the a4ebr^.i.e rclaeien-
sation, the closed loop system is unstable. The ship bct •.;ecn the elements of return difference and
plots are not includcd due to lack of space, but its determinant. For an NxN return difference,
conclusions drawn front then are given. Control pro- the most basic of these relationships is
blems for the uncoc,tensated model were complicated
by the cxlsteuce of considerable interaction, and det(Ig GK) - 1+ {F ixi principal ninors of C1:).
large gains at high frequencies. An additional S^Z (10)
difficulty was that one of the eigenvalues was riega- For the example of this paper, (10) takes the forai
tive at zero frequency. This tended to limit the 1+((G + det(C:,) . (11)
response speed o r the closed loop system. It ap- 11 K 11+6 12 1'21 )+(G 21 k 12 +G 22 K 22 ) )
poared on the C.L. plots that, from a stability In (10) and (11.) we note the advantage of minute
vicv.po,nt, the frequency range of interest is in the detail and the disadvantage of nonrecursive con-
:a vicinity of to rps. This gives justification for struction. Csesiderable interest attaches to the
use of the compensator given in the previous section. removal of this disadv;.ntat e, which can be accom-
As a practical cotter, our goal is to achieve as plished by methods drawn from the results of exte-
rapid a response as possible to a step input, with- rior algebra (5;. Consider the recursion (where y
out suffering, any overshoot. Heavy emphasis is tr denotes trace)
placed also on steady state acer•racy.
a 0 " 1 (17a)
To remove the right half plane pole in k" we choose r•-1
K 3 arbitrarily as diag (1/s,(-1+1.2359s)7s). The
a r - a1^tr (GK)r-p (12b)
resulting K(s)=Kl*K2*K3 becomes p=0
0 -1+1.2359s for 1<r<N. It can be shown that
It(e) 0 9
1 9.4796+0.2494: ( )
s
The diagonal nature of K (s) does not affect domi-
3
Hance. t,orcover, an examination of the (1,1) and
t
(1,2) cicrcnes of G(s) reveals that if the 0 in K(s)
2j
is changed to 9, we can significantly reduce the
-6000 -4000 -2000 0 200 6011)
^aa ^
lj
J375
w> 0 ^) 1 2 100
v-1
j250 -+-- —^
qll -j2E4 W=10
J125
q22
11 ^•=1 .Zj I
lU V
` -I -j4E4 1000 i

100 .2j
.-i t a2
-j6F4
t
Fig,. 6 First C.L. plot for Fig. i Second C.L. plot Fig. 8 Second C.L. plot- for

n CKtK2KZ. for GKIK2K3* Ct.1K2F:3K4. --- --- - -

I ..1

ri
H`.
i i .
d:t(im) - L (--1)a (13)
i-o
^K
The direct approaches differ appreciably f:on meth-
ods described in preccJ: , T :sections, in that they .3j
a
address thenselvOS directly to the !male of
G12^2
tdct "100) on the SNC, without any particular con-
cern for such issnos as dominance or interaction.
G 22 1 22 i
Alternate in:iphts accrue from such plots, which •J
we Illustrate here for the engine design example.
All plots are drawn for Lite finnl return differ- -.3
. 3 - —+
ence as developed in the section:: preceding.

Fig. 9 indicates the fivr constituents of a G21


det(O CE) plot as developed in (11), while Fie. 10 -.lj
tr presents the corresponding two constituents accord-
ing to (13). Fig. 11 contains the total Nyquist -.2j
plot, which is obtained by adding, the individual
curves in either of the two prior figures. Re- G 11K 11 -'3j
vealed in ibis. plot, NO 11 is a feature not so
readily noticeable in the earlier plots, nn-mely con-
ditional stability. It appears, therefore, that the
availability of a variety of graphical tools is in
the multivariable case every bit as valuable as in
Fig. 9 The Nyquist Plot of Fig. 10 The Nyquist
the more cl:ssical, one-input, one-output situation.
the Elements of Det(liCK), Plot of the Elements
It is readily seen in Fie. 11 that the plot. encir- According to Expansion Eq. of Det(I+CK) Accord-
cles -1+j0 twice in a counterclockwise direction. (11). ing to Expansion, Eq.
Therefore, the system is shown to be stable because (23),
the open loop characteristic polynonial has a double
Zero at the origin.
Further exploratory studies of direct methods as
design aids are available elsewhere [G].

6. Si- • ulation Results


loj
Closed-loop time responses were obtained both by
using the linear model simulation and by Implement- 5j^
ing the compensator on DYNGE:,, a jet engine simula-
vrO+
tion program developed by the NASA Lewis Research
Center [7). V-0
DYNGEN is a versatile digital program which ana- -l. -5j
lyzes steady-statz and transient performance of i
turbojet and turbofan engines. It uses a sixteenth l
order system to model this example, and solves
-1Oj
!^
the state differential equations by a modified Euler
method. The use- need only supply appropriate com-
ponent performnnce r_nps and design-point informa-
tion, nod then write thn control subroutines. Im-
plementation of the compensator required first order Fig. 11 Nyquist Plot of Det(I+CK).
functions to perform integration and lead-lag com-
pensation. seconds, and the turbine temperature reaches a
7-
The linear model used in this study was also obtain- maximum increase of 0.097.
ed from DYNGEN. By utilioing a special control sub- The linear and nonlinear responses were not in such
routine writt_a by NASA, called DYGABCD [s], models close agreement for a step input in the second
can be derived using whatever states the user de- channel. 1he linear model shows turbine tempera-
sires. DYNGLN thus possesses the capability to de- ture slowly ramping up (Fig. 13) as the change in
tcrvino linear models for the engine with any order thrust is held to a ninimun. DYNGEN produces simi-
up to sixteen. lar results for the turbine temperature response;
irig. 12 shows the response of the linear model to a however thrust experiences a strong decrease bef^or?
step input in the first channal. Thrust ban a rise rising to zero. At this writing, it is believed'
tim (10:;-90%) of 1.04 second, with no overshoot that the five states chosen for the DYGAN D model
occurring. HiF,h pressure tuO ine inlet tcryerature do not adequately describe local engine behavior
increases tc a c_txir..un ci 0.105 at approximately for the second channel equipped with the present
0.9 seeo a ds, then gradually decreases. controller.
_ I
MUM, and even better, results occur when the
I
compensator is employed in the DYNGEN simulation i
w using a one percent step. Thrust rise time is Woo

Septcnber 1973.
].0 (3) It. 1f. Schafer, W. E. Longenbakor and M. K.
1'h rust Snin, "SysL. • ri Domin.uicr," Preliminary Report
Department of Electrical Engj neerir-g, Univer-
0.8 .
city of Kotr( , 1)ame., Notre Dame, Indiana,
Dy71i;e1
T November 1976.
0.6 l.lncar 141 1C. M. Schafer, "A Graphical Approach to System -4

Dor;ir.ance," M. S. Thesis, Department of Elec-


0.4 i trical Enf,ine,, ring, University of Notre Dame,
Tt:r};inr. Fotre Dame, Indiana, !!ay 1977.
0.2 / Linear Temperature 151 V. IL Grcub, `tultilincar Alc.chra. Neu York:
Springer, 1967.
(61 P. W. 11oppner, "The Direct Approach to Com-
0.4 0.8 1.2 1.6 2.0 po[a atiLill of :;alLivariable Jet Engine Xodcls,"
Tire (sec)
H. S. Thesis, Department of Electrical Engi-
Fig. 12 Response to Step in First Channel
necrinF, Notre Dame, Indiana, May 1977.
171 J. F. Sellers and C. J. Danicle, "DYNGF.N-A
Frog-am for Calculating Steamy-State and
Transient Performance of Turbojet and Turbo-
fan Engincs," NASA Technical Note, NASA LV
0.4 I>-7901, Washington, D.C., April 1975.
i^ I [8) L. C. Geyser, "';ASA Preliminary Report on
0,2 DYCABCD," April 1976.
i

0.0
f
i
:-0.2 i.
t
:•-0.4 QRIGINAI;
PA GE IS
OF POOR QUALITY.
0.6 1.2 1.8 2.4 3.0 3.6 4.2 4.8
Time (sec)
Fig. 13 Response to Step in Second Channel

7.' Conclusions
t
This paper has demonstrated the usefulness of the
new CAJUTAD plot approach to designing compensators
for complex plants. The DYNGEN simulation for a
step in channel 1 has shoa-n +:hat acce p table re-
sponses can be obtained using linear co:rycusators.
j
An ordered collection of these my rare global con-
trol feasible. For steps in channel 2, conclusive
evidence was not obtained. We suspect that this is
due to inadequacy of the linear model in describing
the plant. This important factor of selecting an
appropriate linear model is often overlooked. But,
as we have seen, it turns out to be crucial in
practical applications.

The method of CAI.DIAD plots can be pencralized to


plants with more than two inputs and outputs b y con-
sidering a far.[ily of compensators with 1's on Lhe
diagonal and onl y one non-zero off-diagonal term.
As stated in (1), except for changes in the order-
ing of inputs or outputs, such a study is exhaus- i
tive.

8. References

111 11. H. Rosenhrock, Com p uter-Aided Co ntrol _S vsrem


Desifn .London: Acade sic Press, 1574.
(2) A. G. J. MacFarlane, and J. J. Relletrutti, t

"Ihe Characteristic Locus p _sir..a ':cthod,"


Auto p s.[ticn, Vol. 9, No.5, pp. 575-588,

, . , , .. .. i } 1 . .
E«f
i Appendix D

APPLICATION OF POLYNOMIAL TECHNIQUES TO

MULTIVARIABLE CONTROL OF JET ENGINES

R. R. Gejji
M. K. Sain

i

.,., •• , r;itro rKINI SurrALt fUR tl 1/Z x 11 YAL. ^i

APPLICATION OF POLYNOMIAL. TFC)MOUFS TO MULTIVARIABLE; CONTROL


OF JET F. NCI NES * ^ y

R. R. Gejji and M. K. Sain


Department of Electrical Engineering
University of Notre Dame
Notre Dame, Indiana
U.S.A.
46556

. 1
SLMMARY
This paper describes a complete case study of the application of the theory of minimal de-

I sign to multivariable control of jet engines. The minimal design problem is appro;.ched from the
viewpoint of polynomial modules, and computational experience with PL/I and FO .`LiC-PL/I software
is discussed. The complete minimal design solution exhibits flexibilities not apparent in early
industry studies, and a new ap p roach to pole assignment can be used to advanta g e in this situation.
1 INTRODUCTION !control system design in the near term.
^ne way to approach Lhe design of linear multivari- 2. TILE MINIu4I. Dr_SIC.; PROBLEM
able control systems is to express system specifications
rn terms of a desired closed loop transfer function ma- Suppose that F is a given field. For the jet
frix. A question which is often raised about such an engine control problem, F is taken to be R, the field
approach is the practicality of making such a specifica- of real numbers; however, a great deal of the algo-
Eion. Another, related, question concerns the possibili- rithmic nature of the discussion is more general than
ty of determining the existence of realizable compensa- that, and is so ,tated. The sec of polynomials which
tors to achieve the specification. When such compensa- are of interest is Fis], namely those polynomials in
tors do exist, there are the very practical issues of the variable s with coefficients in the field F. The
giving a finite enumeration of them, of determining fact that F[s] is a principal ideal domain ringis well
whether they have fired poles, and of assigning one or known, as is the e q ually pertinent fact that F[s] has a
more of the non-fixed poles. Of s p ecial interest, as it quotient field F(s). More intuitively, F(s) is often
turns out, for the issue of pole assignment is the idea described as the field of rational functions in s hav-
of miniruality, in the state-space sense, of a proposed ing coefficients in F.
solution in the context of all possible solutions.
The design problems of interest in the sequel are
This paper provides a thorough case stud y of such a convectionally stated in terms of r(s); however,
design approach wizen applied Lo realistic num ,^rical mod- Section 4 explains how such problems may be re-converted
els associated with an F-100-like turbofan engine. back to a corresponding F[s] form.
$pecifications are accomplished by means of the methods
Principal interest renters upon the minimal design
of linear optimal control theory, according to proce-
problem (IMP), which can be described as follows.
dures already worked out in the jet engine industry.
Let G: V +V be a linear operator for finite-di;aensional
e remainin g tasks are addressed by regarding the design 1 2 ,
F(s)-vector spaces V ]. and V,. G is regarded as
s a problem in free polynomial modules. A special fea-
realizable if its matrix is proper. Now let G 1 : V,^V
^ure of the application Iles in its attention to campen-- and G
2 : VhV l be given linear operators, where V, is
kators of simple structure, with a view [o the use of a
p ace. M,,3 con-
^radcd collection of them for the purpose of global
sists in determining whether there are realizable linear
ngine control.
i operators G which make the dia g ram in fib. I co.-unute
Section 2 describes the basic design problem, once also
and,a if
finite-dimensional F(s)-vector
so, to find one whose minimal srealization is of
1pecifications are made. Section. 3 provides the dis- least dimension among all such realizable operators.

ussion of the jet engine application, with particular Intuitively, -
ttention paid to the manner of making tha specifications
lid to the formulation of the main design problem for
lie jet engine application. Section 4 explains hoti, to V1 - - - - - V2

fast the design problem in terms of free polynomial ^:^d-


t^les, and Section 5 describes floating polo*_ compu[ ation-
a1 experience gained in applying extended precision PL/I i
t
^oftwate [o solve the jet engine problem in the free mod-
ule context. Section G outlines the correspur.ding G 2 GI
2xperienca associated with an exact rational calculation
do with FOR.`tAC-PL/I spftware.

I The results of Section 6 show that considerably


^reacer compensator design freedom is available [bar. had V3
bean apparent from early industry studies. Using these
results, a new pole placement design procedure based on
Fig. 1. M imal Des inig
n Problew.
4lternaLing multiIinear algebra achieves in Section 7 a
mialmal pole placement solution not possible by those ti-e operators G and G„ derive from the given plant and
earlier industry methods. ! from the specifications, while G represents the com-
pensators to be. desigr.cd.
o .LAa 9 u citJae^ wicu rer atrtis aesigneu co ptaee me
Work in historical perspective, to reference [Fro litera- Beyond the basic FDP, several additional issues are
• tore, and to assess the :ni rits of _polynomial methods for of practical importance. Anionr these should be included

This work was supported in part by the National Science Foundation under Crnnts GK-37285 anti
ENG 75-22322 and in part by the National Aeronautics and Space Administration under Grant NSG 3048.

._.. --
Pape
FA,C Ni l. Is 7 118 x 9 15 la p ;,It:T SL12i ,c:f f ),' a 1; 2 x 11 PAGE ^--^---I•
--^
1 (1) a finite enumeration of all possible solutions, (ii)
1 fT •T •T
determination of any fixedpoles in the matrix of C, and J - (.y Q6y + 6u R6u + 6y S6y)dt
(3)
(III) methods for assigning the poles of G which are not 0
I
^f ,fixed. The answers to these questiuns resolve such
)issues as the availability of solutions with varying where superscript T denotes matrix t-ram;posiEion. The
.degrees of integration. From a Conceptual viewpOint, wtit;liting matrices Q,R, and 5 are.. listed in table 2.
these ideas are develo
developed
p ed further in Sections 4 and 7,
2
,whereas the computational issues are discussed in
Sections 5 and 6. I I freighting matrices i
with Optimal Integral Control Solution
Next in order of presentation, however, is the
,statement of a minimal design problem for jet engine K-1trIX Matrix Elements
control. I 50,000 - 0
Q 0 10,000
3. JET ESGISC APPLICATION)
--- 550 0
R
In this section, we demonstrate tLe practicality of 0 175
;the minimal design approach in the context of jet. engine S 0 0
,control. The basic plant is a version of the. l'-100 __0_ _ 20,000
turbofan engine. Inputs are Set exhaust area and main L 0.509 0.268 1.979 2.171 2.x98
Turner fuel flow; statrw are fan inlet temperature, train -_2.137 -0.377 - 0.223 - 0.776 - 0.227
burner pressure, fail speed, high compressor speed, and 8.329 - 1.126 i
1.
afterburner pressure; and outputs are thru.t anu high- -2.511 - 1.842
turbine inlet temperature. The linearized .nodol approxi-
At this pint, a minimal design problem can be
mates the small signal behavior of these engine variables brouelkt into play. The control scheme of fi -t •,. 3 is
An a neighborhood of 47° Power Lever Angle (PI.A). This
seen to he more desirable because it incorporates out-
jcorresponds to a point approximatel y midway between
put feedback and enjoys the concomitant advzntare of
^Cngine idle and maximum nonafterburning power. The plant zero steady state error, even in the presence of plant
Is specified by the four matrices A ,BC ,D in (1) and
p p p p parameter variation.
(2).
6x - A p dx + B e' 6c (1) j i 5y
p lant
dy - C dx + D 6u (2)
P P dz l:
I i 1: r (`gyp' B p' ^^►
Zable 1 lists these matrices for our example. The
-
+
S + Cp, D p)
attempt to design simple compensators for linear control
over a specified region is part of a strategy for global S^
control of the engine using; a graded collection of these. GIN AL. PAGE IS
ITable 1 POOR QUALITY t
State De s cription Matrices for Jet Engine (H.A=717°)
-
Katrix Matrix Elements
Fig. 3. Optimal Integral Control
1 -51.1196 3.613 -10.211 - 5.451 - 2.715
19.832 -72.34 30.295 40.972 15.327 One relates the per.forman,:e of the two control schemes
A 0.66 4.496 - 3.601 - 0.011 - 2.808 by equating, in both, the Laplace transform of the
p 1.376 2.313 - 0.809 - 3.032 - O.F21 variable du, as written in terms of the respective state
O.fiR 2 0 .703 2. 9 22 1. 79 - 4.596 variables. This leads to the followin •{ equations, which
1.017 39.792 may be solved for L and 11, the values of which have been
- 0.125 4.181 listed in table 2. I
Bp - 0.077 -- 0.382
(4) 1
- 0.088 - 0.565
3.563 IL:1 111A : B
C - 0.037 0.031 - 0.016 - 0.042 1.368 - (5)
11 c p : DP IG 1 C2^
1.031 0.149 - 0.057 _0.001 - 0.086 P• p
D 0.546 0.016
That this is nothing but a form of the minimal
p P.Oi3 - 0.096
design problem can be seen by evaluating the 2x2 closed
We neat examine how engine. control s-pecifications loop transfer function matrices T(s) and V (s) for the
-I-- ...-. w '­ i I 1- .Y . - two systems in figs. 2 and 3. In fig. 29
ib.- 2, the compensators.
T(s) - P I (s) l sl - P 2 (t)1 -1 M (6)
where
P I (s) - C p (sI-A p ) -1 B p + D (7)
d zn
i- ! p (5) - G (sl-A } -1 B + C (e)
i 2 1 p p 2
i
In fig. 3, cn the Other hand,
! T'(s% - -P 1 (s) IsI- p 3 (e)1 -1 11 (9)
Where
F 3 (s) - li p I (s) + sL(st-A p ) -1 B p . (10)

^} Now, rewrite (6) using (4) as


1 1
T(s) - -1, 1 (s) I sI - f' 2 (s);-
11. (11) 1
Fig. 2. Linear Optimal Control I
i
specified by gain matrices C and (1, are chosen with the The relationship between (9) and (Il) now depends upon
1 that between (9) and (10).
objective of minimizing the per)ormance index of (3).

67.2
-.Pogo -

f.1C MV TS 1, 8 * 9 15/Ib Ph17JT SU:ZI 1C[ IOR 8 1/2 • II I'AGE ---'-^

Comparison ofP 3 (s) and P.,(s), with the aid of (5) We remark here on a special feature of (20), in
and (7), establishes the equality of the two transfer that it Implies a common set of dynamics for both rom-
unctions T(s) and V (s). pensators. Suppose that a suitable Soluti011 of (22)
]cads us to matrices A g , Bg , Cg , DS and AL , R k , C k , ad
n
We can then pose questions regarding the existence D such that
Of compensators other than 11 and t, to :achieve the same
performance as attained in fig. 3, and wh. , t., if anv, C(s) - C 9 (sI-A 6 ) -l B 9 + D 9 (23)^ i
advantages such compensators would have over that scheme.
To do this, we consider fig. 4, which is a more Eeneral K(s) C k (sl-A k ) -1 B k + Dk. (24)I
scheme of control based on fig. 3. ! I
In a typical situation, D g 0; and it can be shown that
Plant ,
E A - AC k
8 k
Fig. 5 then shcws how such a control scheme can be rea -
lized.

6z

Fig. 4. Generalized Compensation Scheme


Dur objective is to design compensators G(s) and K(s) to
tchieve exactl y the transfer function T(s) - I' (s)
between 6z and oy. This means that we must have
I
(I + P 4 (s)) -1 P 4 (s) = T(s) (12)^

where we have introduced Fig. S. Realization of Compensation Scheme

P (s)- 1 1 (s) (I+K(s)P5(s))-1 G(s) (13} Finally, we make the observation that (22) can b2
ana
4 completely solved, and a mini...al solution com p uted by
P5(s) - (sl-A p ) -1 Dp. (14).
algorithms given in the Appendix. The next section
deals with the theoretical foundation of these algori-
rom (12), we obtain the equivalent condition thms; and subsequent sections describe their application
to (22). i
P4 (s) _ ( I+P 4 ( s )) T(s),
4. FREE M O DULAR APPR O ACH TO 1•WP
phich c.•a be restated in the manner (15)ti
In matrix form, the mir-imal design problem of fig.l
P I (l+KPS ) -I G = [I+P (f+KP ) -1 GIT. (16) reduces to solving all equation i
1 5
I i
^rom table 1, D is clearly invertible; and so the linear G 1 (s)G(s) = G 2 (s) (25)^
dynamical systA P has a unique linear d y namical inverse for the various realizab l e G(s), where G,(s) and G,,(s)
system P I -1 1 which l we designate 1 I . Thus (16) is are given. Section 3 provided a nontrivial illustiation
equivalent to f (25) in (19), where i
G(T-T) - K 1 1 5 P I T = P IT (17) G 1 (s)[1'1(s)(TT(s)-1} - PS(s
) ] (26),
which in turn can be written I
I ! G (s) = I, (27)i
[(I- TT ) -T T PIP51fG P 1= T 1 Pi. (18). 2
i3nd where the field F was R, the real numbers.
T Ji
ome simplification can be achieved at this paint if we The free modular approach to MDP is based upon the
t ake advantage of the fact that the matrix T 1 (s) has an recognition that, as a set,
i nverse T T (s). Then (18) can be cleared in its right F(s)c-- F[sj x F[sl,
ember so that i
which, in turn, suggests that it ma y be possible to ex-
[P T (T T -I) - P S ] ( GT l I (19)
;•r, • ss (25) in terms of F[s]. A convenient way to bring
KT this about, as illustrated in (20), is to write
N ow compare (19) with fig. 1, from which it becomes clear
hat G ? is the identity man, or that our control system C(s) h(s) D -i (s), (28)'
minimal design problem turns out to be aversion of the where N(s) and D(s) have their elements in Ffs]. It is
minimal inverse system problem. Writing easy to see that every C(s) has re p resentation in the
GT (s}^ o I form (28). Similar representarions could be adopted for
N(s)D - 1 (s), (20) G ii (s), i - 1,2, but the presentation can be simplified
lKT(s)J ii the ChoiceMi(s)
nd
Gi(s) - i - 1, 2 (29) ,
[P _1)- P5(s)]= d(5), (21), d i (s)

there N(s), D(s), and M (s) are matrices ever R[s], and is made, with d ii (s) t F[s], i 1,2 and 1l (s) havin g elc-
here d(s) c R[s], we can nut (2i) in the poly p aaial forte menu. in F(s], i 1, 2. Equation (25) is thenclearly
(H(s) -d(s)1] f the sameas
['Ill ( s)}J a C (22); i
Hl(s) _1 !T2 (s)

I „ ! a N(S)l) (s) (30)
Equation (22) is a polynomial "kernel problem, dI(s) d2(s),
equivalent to the design problem of fip,. 4. By compar-
which, in turn, is equivalent to
Ing figs. 3 and 4, we can trivially establish that a Sol-

r
Ution to the problem does Indeed exist. Our goal in the 2 I
Id(s)M(s) d(s)M2(s)]
l `(s)] =
liequel is to give a finite enutreration of all possible
U. (31)
^D(s)j
p olutlons and to study their t.ule assignment possibili-
an equation written over F[s). For the jet engine
tics relative to the structure of f ig. 4. __ . __ 1 u---•- --- - --- ----_—•--- -- --
47,31
Pone
F . : r^ r
j
[FA(' hiVT$ 7 li0 1 9 IS;'ib PRINT SUfiFAC:C FOR S 1/2 " 11 PAGE :^^•-•^

rroblem, (22) corresponds to (31). Let Wit-h, rite-se notions, the MDP Algorithm in the
n (s) ppendix Solves MDP. The issue of pole assignment is
k ti(s) 7 i ---^ (32) taken up in Section 7.
dI00
' 5. FLOATI`G POINT FSPFRIKECF.
denote the ith column of
In view of the material presented in the previous
I!"(s)
,section, we are ready to take a closer look at (22).
ID(s)IThe matrix [`1(s): - d(s)IJ turns nut to be n 2 x 9 matrix
bf polynomials iii R[s]. Lack of space prevents us from
(Then
(33) reproducing all the numbers here, but fig. 6 shows that
(AM 2 (s)J i i (s) - 0,
[d2(s)M10): -d 1 the typical element is a thirteenth degree polynomial.
and every candidate tc construct a solutior. (79) can be We also note the large variation in the magnitudes of
t raced to such ri(s). the coefficients of the polynomials.
Thus ftDP is quite closely related to the homogeneous
iequation t

[d2(s)M1(s): -dl(s)M^(s)] t(s) - 0. (34) .1.17 E l 0 s 3 -1.37 E101s2


i
The purpose of this section is to explain briefly an +4.40 E10 s 4 -4.99 E10 s3
i
,appropriate algebraic interpretation u. (34). This in- 44.06 E10 s 5 -4.21 F10 s4
terpr.^tation is based upon generalizing the notion of the
n-dimensional F(s)-vector space F(s)" to that of a +1.18 EIO s 6 -1.07 E10 s5
I
frank-n F[s]-module F[s] n . As a vector space, F(s)' -1.24 E9 s7 +1.49 E9 s6
satisfies the usual axioms, with scalars taken from the
f i el d F(s). As a module, F[sV t satisfies exactly the
-1.26 E9 s8 +1.99 E9 s7
same set of axioms, but with scalars taken from the -2.07 E8 s9 +1.99 ES s8
principal ideal domain rip, F[s]. Despite this close
similarity, F[s]-modules do not behave in exactly the -5.55 E6 s 10 +6.93 E6 s9
s10
same way as vector spaces. But there is a class of them, -4.80 E4 s 11 +8.33 Fly
known as finite-rank free modules, which have a great ell
-7.09 El s 12 +2.38 E2
[similarity to finite-dimensional vector spaces in that
they have a basis, which can be defined in the usual way s12
+4.5 E-1 s 13 -1.00 EO
sing concepts of span and independence. F[s] n , for
[example, is said to be free on the basis Fig. 6. Folynomial Matrix
{(0, ..., 0, 1, 0, ..., 0); i - 1-,2,...,n). (35)^ In this section., we report on FORTR AN and PL/I
t
softvares developed to implement the '.DV :Algerithm on a
ith position digital computer, and our experience in the application
, of the software to the jet engine control problem de-
korrphisms of F[s]- modules are defined analogously to
linear operators on vector spaces; and, when domain and scribed earlier in the paper. Both the programs use
codomain are finite-rank free modules, the basis concept floating point arithmetic to implement the MDP Algorithm,
^s used in the usual way to define a matrix for the considered over the field of real numbers. The FORTRAN
morpli`.sm. This, then, is the interpretation to be given version, using double precision arithmetic, affords 15
^o the p x q matrix digits of precision (decimal) on an IBM 370/153 com-
i puter. The FL/I version, using extended precision
[d2(s)M1(s): -d1(s)M2(s)] (36) arithmetic, carries 33 significant digits. Our jet
^n (34), namely the interpretation of a morphism engine minimal design problem comes down to the question
of determining the rank-seven kernel of a module morphism
X1. F[s) q F[s] N (37) whose domain has rank nine, and whose matrix representa-
f finite - rank free Ffs]-modules. As a submodule of the tion in the usual basis contains thirteenth degree. poly-
finite-rank free module Ffs] `t over the principal ideal nomials. In our experience, the principal difficulties
(3omain F[s], the kernel of !d is also free, and thus the arise from roundoff error occurring as a result I of finite
olution to (34) is tantamount to finding a basis for representation of real numbers in the computer.
his kernel. The process for cslcuiatinr such a basis There are two noteworthy features of the floating
L e provided by Algorithm 1 in the Appendix. ipoint KEF.PO (KERnel of a Polynomial Operatnr) software.
If a basis First, it provides the user some control over the number
of digits considered significant during internal computer
t l (s), C2 (s), (S)(38)i arithmetic. In actual problems, this appeared as the
t4
or per M has been coc,puted, IMP solution then depends critical factor in obtaining acceptable solutions from
4on a determination of whether these basis elements can the computer. Second, it p p r{orms a verification of the
e used, through (32) and (28), to construct realizable computed results up to four significant digits. Any
1 (s) matrices---and, if they can, to find C(s) whose discrepancy so pointed up, one attempts to rectify
minimal realizations are smallest and to assign poles by varying, the number of digits considered significant.
wherever possible. It turns out to be convenient to In the case of the jet engine problett, after making
answer these questions in terms of a redu ced basis, whose several runs, we obtained an (apparentl y ) acceptable
definition is as follows. Let solution from the PL/I version by setting the threshold
for Joss of significance near eleven digits. We can
i compare this solution with the known solution to the
ti(s) (39)I
ti,jsj problem, represented by fig. 3. To d, this, we proceed
-0 as follows.

r Where t i.j c F q , t i,k


i 0, and i - 1,2,...,R. Then the
I The complete solution to the kernel problem appears
in the form of seven elements In a rack nine moduli•,
basis (35) Is said to be reduced if the matrix which are the required reduced basis for the kernel.
t.,kt. Represented in the us" al manner, five of these contained
[ t l,kI : t 2 ^ k2 :
polynomials of degree k i one or less. It is interesting
`40'I to note that tho existence of such elements can be pre-
T, as rank 4. Algorithm 2 in the Appendix reduces a basis. dicted by the following; argument. Wv interpret fit;. 3
------- ------_._---- ------------ -
67.E
Page

1 FA C IIIVIS 7 1/B 1, 9 15!15 nmr SURFACE FOR 8 1;'? "< 11 PAGE

(to yield a solution to the kernel problem, of the form Appendix, on an IBM 370/158 computer.
x(41).

r h
-11T
D-, a -s1.T
(41)
Starting cith the numbers of table 1, together
.with the 1, and It matrices of table 2, we go through the
Calculations outlined in Section 3, and arrive at an
exact rational-coefficient version of the kernel prob-
sI- lem of (22). By applying the p1DP Algorithm, conceived

C Or. the assumption that all solutions can be generated


ifrom the kernel basis, the logical conclusion is that
the two columns of (41) can be represented as a linear
row over the field Q of rational numbers, we are led
finally to an exact reduced basis for the correspond-
ing exact 2 x 9 matrix over Q[s). The seven basis
elements turn out to contain polynomials of degree ki
;combination of the five first decree elements in the re-
equal to one and no polvnomials of higher degree. Note
a duced kernel basis. Interestingly enourh, the question
that this means the floating point software missed at
of determinin g, this transformation can itself he repre-
least two elements of first degree in the reduced .basis.
iscnted as another kernel problem in polynomial modules.
Rounded to fit in the available space, the seven basis
However, attempts to generate such a transformation
,elements obtained from exact software are indicated in
turned out to be unsatisfactory, j
'.table 4.
As an alternative approach to verifying the KERPO
We would now like to compare the computer resources
Llution. we used two of the five first degree basis
needed for the floating point calculation with those
elements to realize a second order dynamical control
required for the exact calculation. E
scheme for the jet engine, along the lines of fig. 5.
From fig. 5, we could then obtain a state description In the floating point software, a sort of trial
for the overall closed Icon sysLOm, which we then com- and error process was used to optimize the calculation
pared with the corresponding optimal integral control by varying the threshold for loss of significance.
scheme system of fig. 3. This comparison was based on Though this software did not reach a satisfactory an-
the first few Piarkov parameters. Table 3 shows this com- swer for the jet engine problem, we have allotted from
parison for two of these parameters. i .our experience about seven runs of two CPU minutes cacti
Ito this calculation. Each run occupied 400K bytes of
Table 3
!memory.
Comparison of KERPO Results with i
Next consider the exact calculation. This soft-
Optimal Integral Control
ware occupied 300K bytes of memory and executed the
lb rkov Optimal Integral et engine calculation in 135 minutes CPU time. How-
KlatlO Solution
(Parameter Cont r ol ever, the great majority of this time turns out to be
I -4.4966 0.64815 -4.4969 0.64314 consumed in Algorithm 2, which computes a reduced
CB
-0.13348 -0.17304 -0.13349 -0.17305 basis. This suggests strongly that more research on
{ the reduction process---a common one in the literature
2631.6 1348 1 7 0.451 3.544
CAB ----could have a corstderai,ly greater than average effect
20693 10390 112,'.15 95.573
on practical applications of the method. Except for
We note that our solution appears to nave identified the the reduction, the remaining part of the calculation is
^ and C matrices correctly, while it is in error so far just about an order of magnitude away from being very
as the A matrix: is concerned. On 0,e basis of this reasonable; and improvements of that order can be ex-
evidence, we conjecture that. reunaoff error ; ncurred in pected to occur in the near term, either through hard-
implementing; the Euclidean di:•r sion algorithm has the ware or software advances. !
most serious impact on the correctness of the Solution.
e1 comparison is made in table .5. Here it is seen
This is because, intuitively, the effect of the A matrix
in the stale space corresponds to multiplication b y 's' Table 5
in the module. Since, in our case, the factors by which
Comparison of Floating Point and Exact Solutions
the matrix columns are multiplied are computed via the
division algorithm, we hypothesize this to be the source
Floa tin g E'oint Exact
r f the error. f Resource—
Algorithms 1 and 2 Algaritirm 1 .t l ,orithm 2
In order to solve the jet engine minimal design '
1
problem, then, one has the option of developing floating-Memory 400`: bytes 300K bytes BOOK bytes
point software which tau:. increased sophistication or of —
twitching to softwares which permit exact rational cal- CPU 14 minures(averare) 18 minutes 117 minutes
ulations. The next section reports on the latter method. --
i ;that, on the average, the difference between floating
6. EXACT RATIONAL SOLUTION I 'point and exact softwares was about an order of magni-
^ One vas to avoid the difficulties of finite machine tude in computing time.
#eprrsentatlon of real numbers is to consider the numbers For the exact solution, it is of interest also to
If table 1 as being rational instead of real numbers. examine integer sizes at various stages in Lhe calcula-
tt is then possible to get an exa c t solution to the jet tion. Such a surrmary has been made in table 6. Note
engine problem, using softwares such as FORKEIC or ALTRAN.
These have the capability of rational and s y mbolic manip-
Table 6
ttlation with an essentially unlimited degree of precision. Integer Size During Exact Solution
19aturally, as the calculation proceeds, one would expect
Stage of Computation No. of Decimal Digits
the integer size to increase quite a bit. As a con-
In Ty picalInteger
sequence, the storage requirements and computer time need-
ed to manipulate these would also be substantial. In 1. State Matricea For Plant 4 i
this section we give evidence as to the magnitude of 2. Plant Transfer Function 14
these, e.snecially to contrast with the requirements for 3. Irn"erse of Closed Loop `'ystcm
the floating; point calculations. This yields valuable (T) 33
Insight into the tradeoffs involved in terms of computer 4. Kernel Problem (2 x 9 matrix) 45
usage needed to solve typical realistic jet en g ine con- 5. After Algorithm 1 150
trol problems from the polynomial approach. the results 6. 207 through Algorithm 2 270 {
reported here are based upon FORMAC software written to 7. 607. through Algorithm 2 250
I m plement, in rational arithmetic, the procedure of the S. Final Rech ed Misi:; _ 160

47.5

P ati o
n FAC MVTS 7 1/8 1^ 9 15/16 PRINT SURFACE FOR

Table 4
8 1/2 x It PAGE r «MC

(Rounded) Reduced Basis from Exact Solution


r! 2.443E-3 0.0 0.0 0.0 0.0 0.0 0.0
1.601E-3 0.0 0.0 0.0 i I 0.0 0.0 6.565E-4 j
IT -3.824E-2 -7.199E-5 4.494E-4 4.146E-4 4.146E-4 4.146F.-4 0.0
1.125E-3s -6.649E-63 1.288F.-5s 1.18C.E-5s 5.151E-6s 5.834E-6s 5.806E-4s i

-7.898E-3 -6.41 E-4 0.0 0.0 I 0.0 0.0 0.0 I


1.87 E-4s -1.331E-5s 3.653E-6s 3.322E-6s -2.57I£-6s -1-763E-6s 8.467E-5s

-1.526F.-3 9.685E-5 2.893E-6 0.0 0.0 0.0 0.0


6.54 E-4s -6.621E-5s -4.642E-Ss -4.423E-5s -2..416E-5s -1.847E-5s -1.314F.-4s

r
0.0 5.081E-4 4.794E-4 4.479E-4 0.0 0.0 0.0 i
1.843E-4s 1.713E-4s -4.245E-5s -3.07GE-5s 1.279E-5s
0.0 0.0 0.0 -2.217E-6 I -3.491F-5 0.0 0.0
+3.131E -6s 7.438E-6s -1.421.E-4s I
I
0.0 0.0 0.0 0.0 0.0 1.622E-5 0.0
9.972E-5s j
2.935E-2 2.949E-4 -4.416E-4 -4.068E-4 -2.25 E-4 -2.434E--4 0.0

r x.69 E-4s

that integer size before and after Algorithm 2 is about


the same, while it nearly doubles during Algorithm 2.
; I

in a minimal realization of N(s)0 -1 (s) can be viewed as


a linear combination of the determinants Idk(s):dj(s)I.
2.955E-4s

This also suggests that improvements in the efficiency


of Algorithm 2 may be possible.
Table , 4 makes it clear that IDI must have dugree at
;least two; and so, since
Finally, we summarize by commenting that he float-
ng point software used on the order of 4 x 10 ` byte
Id l .d 6 I = -1.4092974E-Ss -4.7599915E-7 (47a)
2
seconds of computing power, but eventually did not yield 'd 1 :d 7 I - -8.6656773E-Ss -2.9268875E-6s (47b)
an acceptable solution. On the other hand, the exact
rational software required on the order of 2.4 x 10
Id6 :d 7 I = 2.9060023E-8s. (47c)
i
byte sccon(lS of computing resources and led to an exact
solution.
:witli the polynomials in (47) serving as a basis for
1

r 7. Mq'ENISATOF. POLE ASSIG%ME`T

The exact rational software discussed in Section 6


obtained the reduced basis, t i (s), 1 < i < 7, with
R [s], the R-subspace of his] consisting of polynomials
of degree two or less, it is possible to construct an
acbicraLy polynomial
JD(s)I - b I s 2 + b s + b 3 ,
i

(48)
ni(s)l 2

i t i (s) =

of table 4. From (20), where G T (s) is 2 x 2 and K T (s)


is 5 x 2, we see that the matrix N(s) must be 7 x 2 while
(42) i i
for b i c R, i = 1,2,3 by forming an appropriate linear
combination
I + 6 2 ^d 1 .
1

6 7 I + a 3i d 1; (49) i
D(s) is 2 x 2. Accordingly, S1Id1. 6. 7

[N(s)] sii e R, i = 1,2, 3. The B i t s are uniquely determined by


D(s) the b 's. To complete a minim
ign al(48),des it is
^ only necessary
to calculate f 11 and - f for k and
is a 9 x 2 matrix, which means from (32) that two kernel j = 1,6,7. But certain results fr...,Ne exterior
elements ialgebra, referenced in Section 8, permit the calcula-
n (s 11
L Lion of f ) and (f , f f ) as the basis
t i (s) _ ^-Z ---j,i = 1, 2(
43) (f il l 16
f 17 21 26' 27
d i(s) of the kernel of the matrix [ 8 3 , - 8 2 , 8 1 ]. Space pre-
i cludes a complete treatment of the theory, so we turn
must be chosen to effect a design. These elements (43) to the jet engine example.
vill be linear combinations of the reduced basis elements
02). If
We make the selection
I
ID(s)I r s 2 + 2s + 2, (50)
[rl(s)[2(s)]

f as a linear dynamical interpretation as described in


the Appendix, then N(s)D -1 (s) has a minimal realization
i
'not so much because these dynamics are most desirable,
but rather because the industry methods described in
whose state matrix has a characteristic polynomial
Section 3 could not be used to achieve (50) in a
minimal 'design. Thus, by solving this case, we estab-
ID(s)j - Id l (s) : d 2 (s)I. (44) I lish potential superiority for IWI' over existing in-
Now let 7 dustry techniques.

fik t R, i - 1,2. (45) r Starting then, with (50) and working backwards, we
d 1 (s) - fik dk (s), can calculate e i , i - 1,2,3 and thence (I il
k-1 l f 16 , f17)
I: Then
7 7
las well as
I, (f 21' f
26' f27). These calculations were
I D ` s )1 I kY l dk (s); jSlf2jdj(s)I performed using exact arithmetic again. The results are
flk
presented here after rounding. First we obtain,

F 7
L
k^1
7

j^ l
flkf2j Idk (s) : dj (5)I ' (46) r
I
B l r -4.202E6
6 2 -1.154£7
(51a);

(51b)I
by elementary properties of determinants. This shows] -1.095£9. (51c)^
that the characteristic polynomial of the state matrix a3
t_
47.6

.^ Page

L-

FA C: \R IS 7 1,0 8 -. 9 IS to PRIN1 SU;F ACE FOR 8 I. ? * 11 P-3E

Text, ^thc needed f l 's arti: obtained from the basis for
,the kernel of [6 3 , l 6 2 , 3 1. A workable set of f ij 's is, Output
0 (52a) 1 Output 2 i
F. f1 82/'3
f 21 - -6 1 A 3 (52b)
1

Output
1

.5
(520 ^G
Output
L 1
f 16 f27 ' 0 I
f 26 - f 1 " 1 (52d)
0 CI 4
V:: 2
,he f 's were, in our application, further scaled by a
ij (a) (b) I
factor of 10 to obtain the compensator rains as reason-
hble numbers. This can be done without upsetting the
compensator pole placement to be achieved. Eqs. (45)
and (42) c-in then be solved for t and t 2 of (43). The
1 Fig. 7 Unit Step Responses. (a) step
x 2 matrix [[ 1 .t 2 J, which represents our solution, is
seen to be, i on In p ut 1 (b) Step on Input 2.
i i
1.573 -0.937 ! As a final note in this section, it can be pointed
1.687 65.039
`out that the fixed poles in a compensator solution are
1.769s +1.183 57.634s +14.665 ! the zeros of the greatest common divisor of the poly-
0.0202s -8.12 8.395s 0.029 nomialslld i :dd , i, j = 1,2,...,7]. It is clear from
-1.158s -1.607 -13.396s 1 0.585 (53) the pairings (1,6), (1,7), (6,7) of our example that
-3.076s 1.2799 1 this CCD is 1, and thus that there are no fixed poles
0.744s -14.214s j (in the jet engine application. 1
1.622 9.972s I
0.915s +6.582 29.213s -11.258. R. RE1141TS
A number of procedures exist which lead directly 8.1 Conclu s ions
L m the matrix [C : t ] to a state-space realization for
the compensators G^1) 2 and K(s). Referring then to fig. Considerable work has been done in the control
5 and eys. (23) and (24), we find the matrices A , B °vstems area on polynomial design methods. Regardless
C , B and D for a final solution of the problet c3. g which viewpoint one takes toward the definition of
Ifl. are listed in Table 7, after rounding. such problems, their solution is usually assumed to
i !proceed according to algorithms of the type described
Table 7 [ din the Appendix. Conceptually, this theor y has achieved
Compensator Realizations eonsiderabie maturity, and so it seems appropriate to
conduct an extensive case stud y of its a pp lication to a
1 Ma trix Elements realistic problem. This is the reason for the jet
-2.0 -0.1626 ieng,ine control analyses carried out in t-is paper.
i
Ag
12.298 0 I The conclusions are generally positive in nature,
! though with some temporary limitations. On the positive
2.5732 1.686 j
Iside, Sections 3 and 7 show that MDP is a probl_m rele-
B I ivant to the jet engine control industry and that the
8 -0.9369 65.039 1MDP Algorithm offers a s'_guificanc improvement in flex-
i
ibility of design over existing algorithms in that in-
-3.2 0.1 ! dustry, The application problem detailed herein pro-
C vides a realistic and nontrivial test case for workers
S 1.092 0 IN the area of computer solution of polynomial problems.
A first limitation clearly occurs in Algorithm 2, which
-11.727 -9.725 2.888 5.944 0.824 is a popular and well known theoretical algorithm.
Both in terms of integer growth and relative CPU time,
B 36.416 3.28 -13.66 -37.83 9.147 this reduction a:gorithm points to a need for further
research. Following such an improvement, it would
D 0.119 0.777 2.363 9.973 -3.806 appoar that the second limitation is overall CPU time
k for an exact solution. Though the cost of such time
1.932 0.022 -1.265 -3.36 0.813
would be a small part of overall design cost, it appears
tThe solution given in table 7 was verified by com- desirable to reduce this time by an order of magnitude.
Since such a reduction seems to be a near-term poFsibil-
Closed loop systems of figs. 5 and 3. An exact calcula- ity by hardware or software advances, it would seem that
T* tion comparing the first two Markov parameters, showed polynomial methods may soon be ready to play a greater
these to b._ identical for both systems. Another, non- role in everyday practical. design. 1
4xact calculation, which verified the first six '4arkov 8.2 Historical Remarks
parameters, shoved agreement to tour digits. The first
two of these were listed in the second column of table 3. The original stimulus for this work was the paper
of Wang and Davison (1] in 1973, in which a minimal
lStep responses obtained from the closed loop system inverse system problem was solved. That work subse-
Of fig. 5, using the numbers of table 7, are shown in quently led to the algorithm of Forney (2] phrased in
lg. 7. ! rational vector spaces. Together, these works then led
A visual comparison of fig. 7,vitn similar plots to the free-modular IMP Algorithm (3] which has been
obtained for the optimal integral control System of applied here. The jet engine application has been
fig. 3 showed them to be identical. Hence the latter motivated by Michael and Farrar (41, whence arose our
oet is not included here. It Might he interesting; to numerical data. A report on KERPO in double-precision
examine the distribution of closed loop poles, which is FORTRAN has been presented (51, as has a more complete
given below. treatment of the pole assignment approach [61 in Section
-138.43 -4.47 + 0,986 1 g V. Background reading on the algebraic aspects of the
- 78.38 -1.678 + 0.238 1 paper is available in [7]; and the exact proposition
-_ -0.136. peeded in Section 7 can be found in Chapter XV Seetfon
, Proposition 15 of (8 J ' Further references to related
-- -----^^ L

Pone
o
•. I
.^c art ^1 7 1/8 1, 915116 PR!NT SUPfACE 10R B 172 , II PAGE
polynomial works h^+ve been cited in [3), j _S tep 2. By elementary column aperations, reduce
If (A.4) to the form
17 REFER ENCES
i I•:11 ^ 0
1. S. U. Wang and E. J. Davison, "A MinimizaL'on (A.5
Algorithm for the Design of Linear Multivariable I
1'21 E22

c
Systems," I EEE Trans act io n : on Automatic Control_,
Vol. AC-18, pp. 220 -225, June 1973. i where E l) has p rows, has no zero columns and is in an i
2. G. D. Forney, Jr., ":finimal Bases of Rational Vector echelon form.
Spaces, with Applications to Nultivzriahle. Linear st!f 3. Then the columns of E11 are a basis for
Systems," ST AM Journal on Control, Vol. 13, pp. 493- the image of M, and the columns ol I E 22 are a basis for
520, May 1975. the kernel of M (her .1A).
I
3. M. K. Sain, "A Free-Modular Algorithm for Minimal Now, let b., i = 1,2,...,a be the colu pos of E
Design of Linear Nultivariable Systems," Proceodinr-s 22
obtained from Algorithm 1 as a basis for Ker M. Then
IFAC 6 th World Congress, Part I8, pp. 9.1-1--9.1-7,
by application of further unimcdular transformations,
August 1975.
we can get an equivalent basis for Ker M which is re-
G. J. Michael and F. A. Farrar, "An Analytical duced in the sense of Section 4. Notice that we have
Method for the S y nthesis of Nonlinear ,tultivariable introduced the notation b i for elements of the basis
Feedback Control," Report M941338-2, United aircraft before reduction, to avoid confusion with ti, i = 1,2,
Research Laboratories, June 1973. ...n, which was assumed to be a reduced basis in Sec-
tion 4. The algorithm below is used to reduce the ker-
^. R. R. Gejji, "A Computer Program to Find the Kernel
nel basis. However, the procedure is more general in
of a Polynomial operator," Proceedings Fourtcenth_
nature and can be used to reduce A linearly independent
Allerton Conference on Circuit and System Theorv,
elements in F[sl q regardless of their origin. This ono
September 1976.
is typical of procedures described in the literature
. V. Seshadri and M. Sain, "An Approach to Pole for doing these kinds of calculations. However, as has
Assignment by Exterior Algebra," Proceedings_ been pointed out in the p aper, it is this part of the
Fourteenth Allerton Conference on Circuit and computation that consumes tho major portion of computer
System Theory, September 1976. time. Any research aimed at achieving efficiency in the
i
reduction process is, therefore, the most likely to have
M. K. Sain, "The Growing Algebraic Presence in
a significant payoff in terms of making the L-Ml' method
Systems Engineering: An Introducti.on," IEEE.
of control system design tractable in the near term.
Proceedings, Vol. 64, pp. 96-111, January 1975.
Algorithm 2
S. S. MacLane and G. Birkhoff, Algebra. London: The 1

Macmillan Co., 1967. Write each b i , i = 1,2,..., It in the manner


APPENDIX ei
Let
b
i j=0 w
bsj (A.6)I

I M : F[s] q _ F[sl p (A.1),


I p here b C Fq and b. a,^i # 0. WeT shall say the list
i.]
be a morphism of free modules. In the appendix, we
b l , b 2 ,.. , bA is reduced if the matrix
describe how a reduced basis for the kernel of ,M can be
obtained and used to solve *Ill P. For a more complete ... b^ r ) (A.7)
[b l £
b2
discussion, the reader is referred to [3]. Solutions are ^ 2 ' 1 ^ h
pbtained in the form of the q x Z matrix i has rank n. Then, perform: I
N (s) Step 1. If the list b l , b,,,..., b t of linearly
( (A.2). independent elements is reduced, -stop; otherwise, con-
LL(s)
^inue.
1
ver F[s).
I _Step 2. Determine field elements f. in F, 1 < i
t We change notation slightly by letting M be the < A, which are not all zero and which satisfy
h x q matrix representing the morphism M. The technique it
i
s to choose q x q u nimodula r F[s]-matrices to p ost- iIl fi bi'Ei 0. (A.8)
ultiply M. A matrix is unimcdular if it has a non-
cro determinant that is an element of F. Mathemati- Step 3. For the set of integers i having f
ally,
ion-zero, determine an i, denoted by i max , for which li
U : F[s] q F[s] q (A.3)f is a maximum, denoted by £i
max'
s unimodular if IU1 # 0, a F. Such an operation is Step 4. Perform the elementary column operation:
equivalent to a change of basis in F[s] q and Leads to a Replace b by
representation of M in the new basis. The following; max
elementary column operations are examples of such trans- r
E his max
t1)
formations. The column operations are, (1) interchang- f
ing two columns of M; (2) adding an F[s]-multiple of
One column of M to another; (3) multiplication of a
Return to Step 1.
Column of M by a non-zero element of F.
The question that remains is how the reduced basis
Given the p x q matrix M, the following; algorithm
may be used to obtain linear dynamical solutions to MDP.
leads to a basis for tier 1•!. The basis elements are re-
Let MllP take the form (A.9) when stated over F[s].
^rescnted in the usual manner. I
M
ithmIn
Algor_h h 1 0.
1. To the p x q matrix M, adjoin a g x q
dentity matrix to form Wow, in any solution

M N
(A.4)
JD
4 7.8
Pape

F!1C 111YTT i 1/8 `K 9 1516 f::Ihli SURF-ACL 1`0^ B 1/2 A 11 PAGE tau-±w

P CA. 9), each of the L columns will be contained in the T^ (A.15)


^T bi , 1 < i < A^
kernel of M. All solutions pairs (N,D) can, thus, be r
for Ker At.
built tip a+. linr,tr combinations of elements in a basis
for Ker M. Under what conditions will a SoJutlon pair St Apply Algorithm 2 to the elements of (A.15),
I (N,D) yield a minimal solution? i to form a reduced basis,
_ I
Without loss of generalit y , we may assume that for ti, 1 < i < h, (A.16}
^ny candidate pair (N,D) the L columns of
I N I Step 3. Express the reduced basis (A.16) in this
{mnner
J I ki
I D -J
f 0 (A.17
I.re reduced, becausq a they are not, a unimodulnr trans- ti k i, j s ' `i,ki
j=0
formation V on F[s] -L , chosen according to Algorithm 2,
,till produce an equivalent pair (I, D) such that the col- for i 1,2,...,rt. Form the matrix
mrs of
I [tl k t2,k ... t&A
'" _N- I. 1 2 K j
f ll
( D J If the rant; of the matrix formed from the last L rows I
f (A.18) is ncr equal to L, stop; YDP has no solution;
re reduced and - therwise, contir..:e.

ND 1 - NV ( DV ) -1 Nll 1 (A.10) ! Step 4. From the elements of (A.16) in the reduced
Then, we make the following comments, offered without basis select L elements
►roof. -1 t i ,..., ti
I I ti
,(1) K(s)D (s) can be realized by a linear dynamical 1 2 L
system if ND-1 is a matrix of proper rational func- with the properties
Lions. In such a situation, there exists a reali-
zation A,B,C,E, all matrices over F, such that (i) the rank of the matrix formed from the last L
1(s) rows of
G(s) R N(s)D _
(A.11)
I It ilik1 t i21ki tiL1kl ]
C(sl-A) -1 B + E.
1 2 L
Equivalently, we also sa y that a pair (N,D) has a !
inear dynamical interpretation if the L columns of is equal to L; and
[-N_] I L.
{I (ii)ki is a minimum.

[ j°0 j

are reduced and furthenaore, letting the i tr` column be As a matter of fact, more solutions to MDP may be
n i- possible. Any elements t 1 , L ,...,t L in Ker t1,
t i is = 1,2,..., L, (A.12) which admit a linear dynamical interpretation and achieve
J - d i Fhe minimum order dynamics predicted in (ii) above, are
LL
solution to ".DP throug t the equations
these L columns, when expressed as _ - !
( N- f jt t .. t ] (A.19)
Ei i I 1 2
l D
[,
1=0 i,j
sj, t i,mi # 0 (A.13) and

( G(s) N(s)D^1(s). (A.20)


re such that the last Z rows of the matrix
- - I Iow, in the jet engine problem, his 7 and L is 2. The
[t I'm 1 t2'm2 tL^m1t (A.14)
ro needed columns of

N_]
eve full rank. I
D
Being concerned with finding a realization with Lhe were generated in Section 7 to satisfy the pets place-
east order of dvnamics, we state two more properties.
went requirement.
2) If the roperty in (1) is satisfied, then the deter-
minant ^D(s)j is related to the minimal realization,
i being an F-multiple of the corresponding chase-
I teristic polynomial IsI-A)I. Also,
I
I(3) Tile columns of
p AGE IS
OR1 'Poo?QIJ wvi
D
when expressed as in (A.13), yield the number of C
dynamical elements in the minimal realization as [
L
mi
1L 1
With these notions, let t i a 1,2,...,R be the
educed basis obtained from Algorithm 2. Then, MDP re-
duces to generating a elements in Ker M which have a {
linear dynamical interpretation, with mini-wm order dyne- I
ica. For this, we can use the MDP Algorithm.
i
FMP Algorit hm i
L __. Step 1. Apply Algorithm 1 to obtain a basis J {{!____ — __.__--- _A
47 .9
Facie
Appendix E

"A COMPARISON OF FREQUENCY DOIMAIN TECHNIQUES

FOR
i :ET ENGINE CONTROL SYSTFMS DESIGN"

i^

c
M. K. Sain
R. M. Schafer
R. R. Gejji
P. 14. lloppner

r.
r

PUGINAL PAGE IS
pg BOOR QUALITY,

A CWIPARISON OF FREQUFNCY NW? IN TECHNIQUES


FOR JET ENGINE COtiI'ROL SYSTLI DESIGN

R. Cejji, R. M. Schafer, M. K. Sain, 6 r. Hoppner


Department of Electrical Engineering
University of Notre Dame
Notre Dame, Indiana 46556

Abstract_

Present research efforts in the area of linear multivariable control systems in-
clude activities which will probably reestablish frequency do -tin methods as fre-
quently used tools for design. Two notable branches of this activity are polyno-
mial methods and return-difference-determinant methods. This paper sketches some
features of these approaches, in the context of a numerical exauple from turbofan
engine control.

1. 1WRODUCTION _ - 5.481
57X95 3,613 - 10.2-11 2015
State variable methods for the design of lir,ar ,,0 15,327
29•f332 - 72'3~ 3O.f95 -,0•tJ72
^
multivari.able control systems are well established
as a ; ojor tool in the arplications. Variants of
A - .66o 4.116 - 3.601 - .011
the linear quadratic regulator theory are probably
the most successful, with a variety cf other tech-
t 1.315 2.313 - .809 - 3.032 .821
ni q. •y es _t t :cl i •r pu ' E p ...
-ant , d ecoun 1; ion, ,, n d
geo:etric regulator theory also available. Even
.02 2.922 1.471 - 4.516
today, however, linear quadratic regulator theory '703
still requires a somewhat indirect thought process,
a fea w re it shares with many optimization methods;
and much of the remaining technique is synthesis 39.792 1.017
oriented instead of design oriented.
4.181 - .125
Accordingly, some modern re-emergence of frequency .018 -51+4 ]
domain thought has occurred---especially for design. 1 = - •3 82 - -077 D =
broadly depicted, this work involves polynomial .066 .013
methods and return-difference-determinant methods. - .555 - .088
This paper records certain studies of these ideas,
c
on a cnra
a^n illustration from turbofan er.i;ine con- .78 -5 - 3•.63
trol. Brevity precludes in-depth treatment; we re-
ly instead on the illustrations and the references.
r - 07 .031 - .016 - .042 1.368
2. ILLUSTRATIVE PROBLEM c.
i The turbofan engine model chosen for the illustra-
L 1.051 .149 - ,057 .002 - .036

tions has two control inputs--fuel flow and exhaust G(s), in a loop under unity negative feedback of the
area, five states--fan turbine inlet temperature, plant outputs. Fast step responses with small over-
train burner pressure, fan speed, high compressor shoot are of interest.
speed, and augmentoz pressure, and two outputs--
thrust and high turbine inlet temperature. In 3. POLYNOMIAL METnODS 12 1
traditional (A,S,C,D) fotm, the state description
(1] is given by the matrices at the top of the Polynomial rcthods take advantage of the fact that
following column, at a power lever angle of 47'. action of the A-matrix an3 the s-variable are close-
For the sequel the corresponding matrix G(s), name- ly related in a nodule theoretic sense (3]. Not yet
well advanced computatiunally, polynomial methods
ly C(sl-A) -1 B+D, is recorded.
nonetheles •a offer conviderable insir4c into system
Structure.' As is to he e-petted, they resemble the
The design problem is to select compensators for
geometric methods in this regard.

W
the dorcain of the map represented
to determine seven
i'y this rutrix
(.016r 5 +.1LSs 4 - 92,05x 3 (.545sS + 71.95 4 t 121^s3 "reduced basis" elements, shown be-
low which serve to describe the
- 396.9;; 2 + 29;o1s + 9;1 ;91) -19`3... 2 - 16E55s -121 ,95) kernel. From these, construction
C(_) . of K I (s) and K,(s) involves two
linear tom.5lnalionE of these seven
- <O^Ss S + 31,,,54 + 3321,5x 3 (-,013c 5 - •it 37s4 t 68.x ,_ 3 module elements, and standard reali-
zation rethodology. Using first,
+ 255JQt; 2 + 761,4-8z;+ 762;7) + 1703.3s 2 + 17L2.9s -3532• sixth, and seventh elements, and
_ J the assumptions
B5 + 14,0.7s i' T 5337.6::, 3 + N-, / ^ t 115690s + 1333' CK ` DK 0,
CK
1 2 1 2 1
As an example, consider the selection of K (s) and
K 2 (s) in Figure 1 in order to achieve a specified a
realizations can be found in the rnncr
closed loop performance T(s). Such a r-pecification Compensator Fealizations
IS, of Course, a nontrivial. issue in its own right.
A complete treatment of such a specification can Matrix Eletaeats
be found in (2]. Relying upon the algebraic inter-
pretation of a transfer function as a pair of poly- p -2.0 - 0 .1626
nomials, such a design problem can be ccnvertcd to hl 12,298 0
a kernel calculation in V(s)- modules, where R(s]
denotes polyno-mials in s with coefficients in the 2.5732 1.056
real number field R. Considerable manipulation D
must be carried out to set up this kernel problem, k1 -0.9369 65.039
which turns out to involve a 2x9 matrix of polyno- -3.2 0.1
finials up to the thirteenth degree, as shown below. C
K1 1.092 t'

+ -11.727 -9.725 2.888 5.944 0.824


B
1.17 M O s 3 -1.37 EIO s 2 K2 3G.416 3.28 -13.6b -37.f.^ 9.147
44.40 E10 s 4 -4.>9 F10 s3
0.119 0.777 2.363 9.471 -3.806
• X4.06 EIG s 5 -4.21 810 s4
DK2 1.932 Q.022 -1.265 -3.36 0.613
i^l. l.8 l:l0 s •-1.07 810 s5
Responses to unit steps in the tiro reference than-
-1.24 C9 s 7 +1.49 F9 .. 6 nel are shown in Figures 2 and 3.
-1.26 E9 s 8 +1.99 1:9 s7
Solution of a problem by polynomial rh--thods involves
j• -2.07 F.8 s g +1.99 E8 s8 at this time nontrivial coeputational overhead,
X10 which is discussed in greater detail in (2J. It is
-5.55 E6 +6.93 E6 s9
11. likely, however, that advances ir, software and hard-
10
-4.60 E4 s +8.33 L:4 s ware will soon reduce this overhead. Advantages of
s ll the riethnd include a finite enu^eration of all soiu-
-7.09 El s 12 +2.38 E2
2 Lions for a given T(s), and perhaps eventually a
s 13
r. ^•
4-4.5 E-1 - 1.0(6 EO s l finite description of all possible performances.
Solution involves automorphiC transformations
_ on l^ RFI1 =Ln Ff^t__e r^_ n r-cn`I'L1.`:1'S 141
2.443F•-3 - 0.0 0.0 0.0 0.0 0.0 U.0

0.0 0.0 0.0 0.0 0.0 6.565E-4 The present connuta-


1.601k-3
tional.situ,tion for
-824E-2 -7.199E-5 4.494E-4 4.146E-4 4.146E-4 4.146F.-4 0.0 polynomial =ethcds
).125F.-3s -6.649E-6s 1.288F.-5s 1.18SF.-5s 5.151E-6s 5.83 s E-6s 5.806E - 4s zakcs alternate arc-
0.0 0,0 quency domain approach-
-7.6981:-3 -8.41 E-4 0.0 0.0 0.0
1.37 E-4s -1.331E-5s 3.653-6s 3.322E-6s -.2.571E-6s -1.7C5E-6s. 8.467E-5ses of interest.

1.5261:-3 9.6a5F.-5 2.6931:-G 0.0 0.0 0.0 0.0 If we set K,, to zero
6.54 E-4s -6.621E-5s -4.642F.-5s -4.423E-5s -2.416E-5s -1.84:8-5s - 1 . 314E - 4sand denote K by K,
l.).0 0.0 Ve have the arch
etypal
0<0 5.OS1E-4 4.7941:-4 4.4794F.-4 0.0
-3.076E-5s 1.279E-5'nity negative feed-
1.643E-4s 1.713F.-4s -4.24SE-5s
back precompensation
0.0 0.0 0.0 -2.2171-6 -3.491F-5 0.0 0.0 Frohlcm. IC K is
y 13.131E-6n 7.436E-6s - 1.421E - 4 a^ssumcd to have :rate
1.622E-5 C.0 description ("K ,1,,
0.0 0.0 0.0 0.0 0.0
( 9.972E-5 -IC K .D K ), then a corn

7.9351:-2 2.949E-4 -4.416x:-4 -4.068E-4 -2.25 E-4 -2.434 E-4 (1.0 Lined state descrip-
8.69 F-4s 2.9551:-4-lion (AC ,1t C ,CCI DC ) for

681

i
F
E 711k St
l
1

Temperature
a.s

6
2 4
Seconds

Fig. 2
Closed Loop Response to Unit Step in
Commanded Thrust; Folynccial Design
., r Fig. 1 ,•.556 - .371 - .185 0 .165 .371
is Compensation Scheme for Polynomial Design
pp
F, J.23-
2

ust
-j.232

0
^econ^s
-J.464
R
Fig. 3
Closed Loop Res p onse to unit Step in
I Coaalanded Temperature; Polynomial Desigi;

f -.343 -.172 0 .172


a FiF. 4
j.214
Nyquist Plots for Compensated System;
Direct Expansion

jo j.l
k
j'
i
-fill

- j .2 i
-j.424

-j.5

3 --j.643
_-; -j•7 '3

Ffg. 5 -j.1
Nyquist Plots for Compensated System;
ii Alpha Expansion
-1.037 -.622 -.207 0 .207
Fig. 6
Nyquist Plot for det(I+CK)
Ij

6V
Seconds
2 3 4
4j -^ It;^ tl J
0 1

Thrust

.R
Oj

.6

.4 —2j
tII 11:C W
Temperature
.2

;; ' 1 ,^`` I _1
0

-4 -2 0 2 4
Fig. 9
Fig. 7 C.SRDIAD .Plot, Column 1, Uncompensated
Same as Fig. 2 for Direct Approach Design

Seconds
1 2 3 4
2j 1 Inc to

1 Tempera 1j

Oj
O T'hrus t

to 0

-1

InC to

-2j

-2 -1 1 2
Fit; 8 Fig. 4o
CARDIAD Plot, Column 2, Uncompensated
Sax as F!h, :: for Direct Approach Design
the Ic:,p can be obtained by an isomor p hism on the Important relationship that
product of Ow :Late spaCtrs X andassociated
114DD K I Isl-AC 1-1liGKI IsI-Al IsI-AKi,
j with t1w plant and compensator, r0speetieely, pro-
vided th.tt the {'ain !.0 has no nel ,,itive unit
eigenvaIucs. For this s l tuatLn, one has the upon which a Kyquist study can be based. We refer

683
qU P,J Jrl expansion chosen for the return-difference-deter-
minant. The obvious expansiun, shown for the pre-
cent illustrative ease, is

I
!lam _ / ^ f
1+G K +G 1:
li 11- 12 Z1
+G K1 2 +C K
cl 1^ 22 22
+ ;cK^;
fir
and a less obvious, more recursive expansi.oa in an
HxN rase is N
4.8./x`
Where WO
a 0 1.
and
i-1
ni - ( - 1/i) ap trace(CK)VP,
P=O
pj for i > 1.

Design based upon Nyquist plots of 11-f-GKI is tide


challenging by the intricate way in which t:he com-
pensator K relates to the determinant. At present,
Dniy introductory design studies based upon the
expansions above have been made (4). An illustra-
tion Is the compensator
1
0
s
K(e)
100 0( s +1)
0
s(s+200)
-4.$ rr . $
Which was chosen by a cut-and-try method to increase
Fig. 11 the speed of res p onse of the second output_ Figures
CAVDIAD Plot, Column 1, Compensated System 4 and 5 show the terms in the "obvious" ani _ ex-
pansions for the comp ensated system, with Fi g ure 6
indicating the sum, exclusive of the unit texts in
each expansion. Closed loop responses to reference
steps in each channel are shown in Figures 7 and 8.
Though the temperature response in Figure 8 is ac=
ceptable, the thrust response in Figure 7 exisibits
overshoot; and considerable interaction is evident.

.151 In current practice, plots such as Figure 6 tend to


be the most useful. Design technique tends to focus
upon reducing the interaction evident in these re-
sults, which brings us to the next topic,

5. CARDIAD--A DOMINANCE APPROACH (51

of In making a Nyquist plot of the determinant of re-


turn difference, H. H. Rosenbrock, 171 has estab-
lished that JOC KI encirclements can be counted as
the al£cbraic sum of the encirclements of t e diago-
nal elements of return difference (1+GK)--p_ovic_d
that a condition of "dominance" holds on (OW).
This means, in our case, that the off-diafeaal ele-
ment in a column is smaller in magnitude t%an t e
diagonal element, as a function of frequent:- (a=jw).
Related to this stability oriented usage of the dom-
inance idea is a corresponding; requirement on the
loop transmission C1:, which is used to help with de-
coupling closed loop performance.
-.15 0 .15
Selection of 1:(s) for this latter purpose, so that
Fig. 12 G(jw)K(jo) is dominant on its columns, has been

r
CAfcDIAD Plot, Column 2, Com p ensated System widely :studied for the case in which K(s) is rqo^,_
to such studies as rct"rn-difference-determinant stricted to be a constant matrix. .Such less has
methods. because of the presence of J14GKI in a key been accomplished relative to the choice of a
dynamic K(r).
role.

Carstruction of a Nyquist plot is related to the A new technique for this purpose is the CARDIAD

684

.-.+.....^....._. ^_.. - - --- - --••----•-w.•=rn^-^r-a,:ew+^rrr.. -^ran^ -^^•rYtv :- r'.r.""^+..^'y.^.^t",s ,^++^r.T^r^►^-^--^"^"""".^


pF
E
^I
i

it, acronymed for Compensator Acceptability 3. S. Kacl.anc and G. Hirkhoff, A_1 !L ra. London:
- P ion for Uf.^^ n.il Dominance.

Compensators 71kc• t:.-icmillan Co., 1967,
xz(J +jy2.(j^,)^
-^ r(j^) .r 4. P. W. 11oppner, "The Direct Approach to Cowpen-

L L S (j-) j y l (j^) 1

sr;. assumed, without loss of generality for ))rc-


satiun of Multivariahlc Jer Engine ';oriels,"
Tech. Rcpt. EE-779, Notre Dame, Indiana, April
25, 1977.
-ompcns.^tion. A CARDIAD plot for colu,-n cue of
the uncor,pensated system is shown in Fii;ure 9. 5. R. M. Schafer, "A Graphical Approach to Syste q
Encli circle corresponds to a particular frequency Dominance," Tech. Ro,it. EE--77S, Notre Dame,
a), and acceptable (x y ) pairs must be outside Indiana, March 26, 1977.
cy in question. Note
dashed circles at the frequen
that y = 0 and x, suitably negative will be ac- 6. ' R. 11. Schafer, R. R. Gejji, F. U. 1loppner, W.
ceptable for all frequencies. Figure 10 show ;, a E. Longenhaker, and M. K. Sain, "Frequency
CARl1IAD plot far column two. Acceptable (x,,,y,,) Dormin Compensation of a IMN GEN Turbofan
pairs must be inside solid circles at the f^cgiien- Eni•,lue Model," Prot. 197" Joint Automatic Con-
cy in question. trol Conference, pp. 1013-1018.

Ilie simple compensator 7. 1i. 11. Rosenbrock, Cornuter-Aid ed Con t rol System
1 .7s .44 Design. New York: Academic Press, 1974.
K(s) _ .05s + 1

-10 1
achieves dominance at all frequencies in both col-
umns, as can be seen in Figures 11 and 12, which
consist only of solid circles cacti of which inclu-
des the origin.

More detailed information about an application of


this method to dvsic,n and simulation of a turbofan
engii: control can be found in 16].

6. DISCUSSION

Recent activities in frequency domain analysis and


design of linear r..ultivariable control systems sug-
gest a certain resurgence of this viewpoint in use-
ful new ways. Though somewhat limited by space
constraints, we have tried to give a glimpse of
snore of these methods in the context of a numeri-

i
cal model from the turbofan engine area. Focus
has been on polynomial methods, which hear close
resemblance to geometric control methods in an
abstract algebraic sense, and upon methods re-
lated to the deternir.ant of return difference.
The CARDIAD plot, a new dynamical ap roach to
dominance, has been illustrated.

ACKNOWLEDGMENTS

This work vas supported in part by the National.


Science Foundation under Grant ENG 75-22322 and in

r part by the Nation.il Aeronautics and Space Admin-


istration under Grant NSG 3048.

REFERENCES

'r 1. G. J. Michael and F. A. Farrar, "An Analvtic


Method fur the S y nthesis of Nonlinear Multi-
varla;lc Cnn.rol," United 1'echnnlo g ies Research
Iabotateries. Report '1941338-2. East il.irttord,
Connecticut, June 1973.
13
2. R. R. Cejji and M. K. Sain, "Application of
F4.ilyne-l., 1 Techniques to Multtwriable Con-
trol of .let Engines," Freprints Fourth IFAC
7ntrrnatlun.F1 Symposium on Multivarlable .vch-
nological Systems, pp. 421-429, .tune 1977.
685
Y

, 1

It
T IF

Appendix F

"TIRE OPTIMAL CONTROI.

OF A WO-SP:OL TURBOrAN JET ENGINE"

W. E. Longenbaker
R. J. Leake
TIME OPTIMAL CONTROL
OF A T140-SFOOL TURBOFAN JET ENGINE

by

W. E. Longenbaker and R. J. Leake

Department of Electrical Engineering


University of Notre Dame

Technical Report No. EE-7714

September 1977

This research was supported by the National Aeronautics and


Space Administration under Grant NSG-3048.
^1

ABSTRACT

This work explores an alternative to existing methods which are

commonly used to design controls for jet engines. Who reas most modern

designs implement piecewise-linear quadratic regulators, this represents

an attempt to obtain a global nonlinear optimal control for a two-spool

turbofan jet engine.

A necessary starting point, therefore, is to have a good nonlinear

model on which to perform the control studies. Unfortunately, the only

accurate existing models of jet engines are (1) linear analytical models

valid only for small regions, or (2) massive nonlinear, non-analytical

computer programs which attempt to match experimental data. What is

needed for this study is something which lies between these two extremes,

i.e., a nonlinear, analytical model.

A fifth order nonlinear model was developed in this study which cor-

rectly models most of the qualitative behavior of the jet engine, but which

j fails to achieve strong numerical agreement with DYNGEN, a reliable non-

analytical simulator. Several linear models were derived, both from the

nonlinear analytical model, and also from DYNGEN. A time optimal control

r problem was formulated, subject to various constraints. Dynamic Program-


-
ming theory and the Successive Approximations technique were explored, and

applied to the problem of interest, while several improvements in the

numerical programming were introduced. Analytical and numerical results

were obtained for several models, both constrained and unconstrained.

Finally, these results were tested on the two principal simulators, PYNCEN

and the analytical nonlinear model.


{
I The study successfully achieved time optimal feedback control lar•,s

for various models of the two-spool turbofan jet engine. Furthermore,

valuable insight into the nature of the problem was obtained, and much
useful computer software was developed. However, an optimal control law

obtained from any model can only be as good as the model itself. For
this reason, more work is needed to develop a better nonlinear analytical
model of the two—spool turbofan jet engine.
If

TABLE OF CONTFNTS

Page

ACKNOWLEDGEMENTS. . . . . . . . . . . . . . . . . . . . . . . . . . ii

CHAPTER I: INTRODUCTION . . . . . . . . . . . . . . . . . . . . . . 1

CHAPTER IT: TWO--SPOOL TURBOFAii .SET ENGINE MODELS . . . . . . . . . . 2


2.1 Introduction . . . . . . . . . . . . . . . . . 2
2.2 Mouel 2 Development . . . . . . . . . . . . . . . . . . . . 3
2.3 Linearizations and Order Reductions of Model 2 . . . . . . 11
2.4 Linearizations and Order Reductions of Model 1 . . . . . . 13
2.5 Comparison of Model Responses . . . . . . . . . . . . . . . 17

CHAPTER III: THE TIME OPTIMAL CONTROL PROBLEM . . . . . . . . . . . 23


3.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . 23
3.2 Statement of the Problem . . . . . . . . . . . . . . . . . 23
3.3 Constraint Determination . . . . . . . . . . . . . . . . . 24

CHAPTER IV: THE DYNAMIC PROGRAMMING METHOD . . . . . . . . . . . . . 26


4.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . 26
4,2 Dynamic Programming Theory . . . . . . . . . . . . . . . . 26
4.3 Successive Approximations Technique. . . . . . . . . . . . 27
4.4 Technique Refinements. . . . . . . . . . . . . . . . . . . 29
4.5 General Program Structure . . . . . . . . . . . . . . . . . 33

CHAPTER V: OPTIMAL CONTROL LAWS . . . . . . . . . . . . . . . . . . 36


5.1 Basic Problem Considerations . . . . . . . . . . . . . . . 36
5.2 Model 2L3 Unconstrained. . . . . . . . . . . . . . . . . . 37
5.3 Optimal Control Theory . . . . . . . . . . . . . . . . . . 42
5.4 Model 2L2 Unconstrained . . . . . . . . . . . . . . . . . . 47
5.5 Model 1L2 Unconstrained . . . . . . . . . . . . . . . . . . 57
5.6 Model 1L2 Constrained . . . . . . . . . . . . . . . . . . . 65
5.7 Model 2 Unconstrained . . . . . . . . . . . . . . . . . . . 69
{ 5.8 Model 2 Constrained . . . . . . . . . . . . . . . . . . . . 73

CHAPTER VI: CONTROLLER SIMULATIONS. . . . . . . . . . . . . . . 78


6.1 Introduction . . . . . . . . . . . . . . . . . . . 78
6.2 Model 2 Simulation Utilizing
-: Model 2L2 (Unconstrained) Control. Law. . . . . . . . . . 81
6.3 Model 2 Simulation Utilizing
Model 2. (Unconstrained) Control Law. . . 81
6.4 Model 2 Simulation, Utilizing .
Model 2 (Constrained) Control Law. . . . . . . . . . . 90
6.5 Model 1 Simulation Utilizing
{ Model 2 (Constrained) Control Law. . . . . . . . . 95
6.6 Model 1 Simulation Utilizing
Model 1L2 (Constrained) Control Law. . . . . . . . . . . 95

' CHAPTER VII: SUPZTARY. . . . . . . . 105

i
iii
Page

APPENDIX A: LINEAR SYSTEM TIME RESPONSE PROGRAM . . . . . . . . . . 107

APPENDIX B: DYNAMIC PROGRAMMING PROGRAM FOR LINEAR MODELS . . . . . 109

APPENDIX C: DYNAMIC PROGRAMMING PROGRAM FOR MODEL 2 . . . . . . . . 115

APPENDIX D: MODEL 2 SIMULATOR . . . . . . . . . . . . . . . . . . . 125

APPENDIX E: INPUTS FOR DYNGEN SIMULATOR . . . . . . . . . . . . . . 131

REFERENCES. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 136

iv
CHAPTER I

INTRODUCTION

This work explores an alternative to existing methods which are

commonly used to design controls for jet engines. Whereas most modern

designs implement piecewise-linear quadratic regulators, this represents

an attempt to obtain a global nonlinear optimal control for a two-spool

turbofan jet engine.

A necessary starting point, therefore, is to have a good nonlinear

model on which to perform the control studies. Unfortunately, the only

accurate existing models of jet engines are (1) linear analytical models

valid only for small regions, or (2) massive nonlinear, non-analytical

computer programs which attempt to match experimental data. What is

needed for this study is something which lies between these two extremes,

i.e., a nonlinear, analytical model.

Finally, after a suitable model(s) of the F-100-like jet engin e is

obtained, a time-optimal control can be calculated. This control will

be determined subject to various constraints. It will be derived using

Dynamic Programming and the Successive Approximations technique.

ril

1
#.4

i
CHAS TER II

TWO SPOOL TURBOFAN JET ENGINE MODELS

2.1 Introduction

In this chapter, a hierarchy of models for a two spool turbofan jet

engine is discussed. The configuration for this engine has been speci-

fied by NASA Lewis Research Center personnel. A preliminary version of

t his work is given in reference [1] and has also bFGn reported in [2].

The models have been classified as follows:

Model 0. The actual jet engine (hypothetical).

Model 1. The DYNGEN [3] simulation program, coded with data

presumed to have been taken from experimental measure-

ments on Model 0. This model solves 1.6 nonlinear

differential. equations and uses data maps and thermo-

dynamic tables which cannot be expressed analytically.

Model 2. This model involves the primary thrust of this chapter,

and is a 5th order nonlinear analytical model. It

includes the 5 state differential equations which gov-

ern the dynamic behavior of the system, along with 20

algebraic equations which express the relationship

between various engine variables.

r In addition to these nonlinear models, several linear models have

been developed. Their original p;:rpose was to provide an indirect method

to compare Models 1 and 2. Subsequently, they also became important in

the determination of a time-optimal. control for the jet engine, when


r
comparisons showed marked differences between Models 1 and 2.

Model 1L5. This is a normalized 5th order linear model which is

obtained numerically from Model 1, using; the experi-


r
mental DYGABCD [4] program of L. Geyser.

2
3

Model 1L3. This is a normalized 3rd order linear model obtained

by means of an order reduction performed on Model 1L5.

Model 1L2. This is the corresponding 2nd order reduction of

Model 11,3.

Model. 2L5. This is a normalized 5th order linear model obtained

by taking partial derivatives of the analytical Model 2.

Model 2L3. This is a normalized 3rd order linear model obtained

by means of an order reduction performed on Model 2L5.

Model 2L2. This is the corresponding 2nd order reduction of

Model 2L3.

2.2 Model 2 Development

There are several purposes for the development of Model. 2. First,

it enables one to readily see the basic nonlinear relationships between

the engine variables. This allows one to gain insight into their static

and dynamic behavior. Second, it is fundamental that an analytical

model be available for the application of optimal control techniques.

Finally, linear models obtained by partial differentiation of this model

tend to have more structure (zero entries in the ABCD matrices) than

those obtained numerically. This in turn gives the linear control

designer more insight.

Model 2 was intended to be an approximation of Model 1, based on

the specified engine configuration. Theoretical relationships developed

in references [5], [6], and [7] were employed _s a starting point and

certain simplifications suggested in [8] were used. In various situa-

tions, least squares and exact fits were made tc: theoretical forms, and

if a theoretical form was unavailable, polynomial, linear, and expc,nen-

tial forms were used, whatever seemed to best fit the situation.

In most cases, the variables used in model. 2 correspond to those of

Model 1. A letter key provides consistency among the variable names in

the following Wanner:

P a pressure

T a temperature

U a specific energy

V a vclume

W a flow

Similarly, numbers in the variable names identify engine locations as

per figure 2.1. Table 2.1 is a list of all. variables used.

High-pressure turbine-,
Dudburner- .\ N
Comtustor-^ ^` ^, HPFXT r Low-pressu re tu Nn e
Room

28
22.21 23 ^1 < 2.1 ,s` rlr 25
h ^ ^

B
1 2 ran 22,21 Compressor 3 < d 1550 55 b< 7
<
BLf, ^
BLDU BLHP BLLP
i

Afterburner
BLF

BLOB

FIGURE 2.1. Jet Engine Diagram

r
5

Before delving into the details of the model, certain decisions had

to be made regarding the choice of state variables and the order of the

system. There is no general agreement as to what the order of a jet

engine system i- it is a physical, not mathematical, entity and thus,

every mathematical model is ^n approximation to the reality. Naturally,

the higher the order of the model, the more accurate the approximation

should be. The order that was selected (5th) was a function of the

accuracy required by the control study to follow. This contrasts with

the DYNGEN 16th order model, but is not an excessively low choice, for

even first order models could yield reasonable results.

The most obvious states to choose are the rotor speeds, (1) N and

(2) NF . The other selections were (3) the burner pressure, P 4 , a vari-

able which is strongly affected by changes in fuel., WFB; (4) the burner

internal energy, U4 , a variable which is related by a constant to the

burner temperature; and (5) the afterburner pressure, P , a variable


7
which is strongly affected by changes in the nozzle area, A8.

Table 2.2 gives a listing of the inputs, states and outputs. in

actual ex_istin control systems, inputs (1) an3 (2) are used, along wi^.h

movable guide vanes mounted throughout the compressor and fan stages.

These vanes cause . changes in the air flow in a manner similar to the

bleeds used in the model.

Tables 2.3, 2.4, and 2.5 respectively are listings of the constants

used in the model, the design values corresponding to the specified

engine configuration, and the nonlinear state equations. Note that the

state equations are formulated in teims of intermediate variables which

have a very real physical interpretation.

Table 2.6 is a listing of these nonlinear relationships existing

between the state variables and the intermediate variables. Some of


6

TABLE 2.1
SYMEJLS FOR VARIABLES

Symbol Variable Description

A8 nozzle area
CNC corrected compressor rotor speed
CNF corrected fan rotor speed
FG thrust
NC compressor rotor -need
NF fan rotor speed
compressor pressure ratio at surge
PCMA%
fan pressure ratio at surge
PFMAX
P 21 fan exit (compressor inlet) pressure
P3 compressor exit pressure
P4 combustor exit pressure
P7 afterburner exit pressure
T 21 fan exit (compressor inlet) temperature
compressor exit temperature
T3
T4 combustor exit temperature
high pressure turbine exit temperature
T50
1 55 low pressure turbine exit temperati.ire
T7 afterburner exit temperature
U, combustor internal energy
WAC compressor airflow rate
WAF fan airflow rate
WA3 airflow rate into combustor
maximum compressor airflow rate
WCMAX
correction term for maximum compressor airflow rate
AWCMAX
14B fuel flow rate into combustor
maximum fan airflow rate
WFM.AX
WC4 gaseous flow rate out of combustor
i
W050 gaseous flow rate out of high pressure turbine
_- W,55 gaseous flow rate out of low pressure turbine
I
` WG7 gaseous flow rate out of afterburner
ZC con.pressure surge margin
ZF fan surge margin
7

TABLE 2.2

INPUT, STATE, AND OUTPUT VARIABLES

Variable Description r Symbol

U fuel flow WF9


U. nozzle area A8
X1 compressor rotor speed I NC
x2 fare rotor speed I NF
x3 burner exit pressure P4
x4 afterburner exit pressure P7
x5 high pressure turbine inlet energy U4
yl thrust . FG

Y') high pressure turbine inlet temperature T4

TABLE 2.3
CONSTANTS

Y Symbol Bcscr i ption V alue


d(Aj ) mechanical equivalent of heat 7i8.20'
G force of gravity 32.174049
R(RA) gas constant I .0252
Y* ratio of sp ecific heats 1.4
P2 fan inlet pressure 518.668
IC (PMII1P) high pressure rotor polar moment of inertia 3.8
1F (PMILP) low pressure rotor polar moment of inertia 4,5
combustor volume 1.65
VCOMB
afterburner volume 49.77
VAFBN
CMNOZ nozzle thrust coefficient .9494
N DESIGN high pressure rotor design speed 10070
C
(XNHPDS)
N FDESICN low pressure rotor design speed 9651
(XNLPDS)
N combustor efficiency 20.71175
compressor specific pressure .24
CPC
CFF i fan specific pressure ,24
I
CVB combustor specific volume .20279

CPHT
i high pressure turbine specific pressure .22589
low pressure turbine specific pressure .27938
CPLT
f(PCB..C) I percent of compressor exit air bled for cooling .16
a(PCBLBU) percent of bleed air which leaks into fanduct .208
6(PCBLHP) percent of bleed air put into high pressure turbine .726
Y( PCBLI.P ) percent of bleed air put into Iota pressure turbine .066

TABLE 2.4
DESIGN EQUILIBRIUM VALUES

Variable Value Variable Value


WFB 2.75 MM 3.3624
A 8 2.948255 WAF 221.573
NC 11899.1 I 54.4151
WCMAX
NF 9873.95 1.5805
AWCMAX
P 4 23.9299 P CMAX
P„MAX 10.270

586.467 WAC 137.649


P 7 2.55142 1 WA3 115.625
CNF 1.02310 WC50 134.364
T 21 742.957 WG4 118.375
CNC .98730 WG55 135.818
T 3 1467.47 I T55 1789.15
T 4 2892.04 T7 1413.81
T 50 2103.47 WG7 224.323
25.3522 FG 13431.02
p3
P 21 2.9960 ZC .8143
W 203.123 ZF I .8333
FMAX i
1

TABLE 2.5
STATE EQUATIONS

State N State Equation

I dNC = 30 2 J _
`CPCWAC(T21 T3 ) + C WG50 ( T 4 - TS0)l
(1} dt (r, )IC NC PHT

dNK
J _`^ F
3021_N^`,pFWA_(T2 - ,
CPLTkG55(T50 - T55}]
(2) dt (n T 21 ) F
" r
4 - VRy* (T
+ WFB - WG41
(3) dt
COAT
MB
^i dP Ry*T
(4)- = V JWG4 - WFB - WA31
Arr BN
di7 C RT
(5) VVB (T {W4- WFB - WA3) + Y * (T 3WA3 - T^} WG4 + T4(l+rj)W-FB}]
dt = P
cORB 4
9

TABLE 2.6a

FUNCTIONAL RELATIONSHIPS BETWEEN VARIABLES

Eq. N ` Equation
NF NF
(1) CNF = N
FDESIGN 9651

(2) T 21 = T 2 + 214.2732 CNF 2 - 48(A 8 - 2.948255)

N % = NC
(3) CNC = —
NC DESIGN T 2 10070 /518.668

(4) 1T 3 = T 21 + 743.2722 CNC 2 - 68(A 8 - 2.948255)


;^z

(5) ` T 4 = U4AVB

(6) ^T S ^ _ .727 T4

(7) IP 3 = 1.05944 P4

(8) ,P 21 = -6.20568 + .0129774 T21 - .0185376 P3

= 261.01 CNF - 63.196


(9) 1UFMAX

(10) P FMAX = 3.516739 CNF - .23561

-2.313268 (P
FMAX-p 21)
(11) WAF = ^vFr1AX + 28.502 1 - e

(12)lW MIAX = 137.54 - 457.987 CNC + 564.325 CNC 2 - 188.113 CNC3

(13) = 6.492 - 4.9749 CNC


AWcr1AX

PCMAX = 26.43184 1 MKS CNC + 109.7243 CNC 2 - 36.5756 CNC3


(14)
P
P -.3662(PCriAX
P3)
21
(1.5) WAC = 21. — WCMAX + AW0m
(1 - 3
JT21/518.668
- - `— -....^+r.^^^^tTR^•YU.1r..ve+n^.u__ _.. ^ ^ . , .e .... t wr+..wr-.m..r.,-tirv.

10

TABLE 2.6b

FUNCTIONAL RELATIONSHIPS BETWEEN VARIABLES

Eq. #1 Equation
(16){ WA3 = (1 - m)WAC = .84 WAC

(17) 1W050 = 301.957 HI 4

(18)IWG4 = WG50 - SlWAC == WC50 - .11616 WAC

(19)IW055 = WG 50 + {¢WAC = WG50 + .01056 WAC

(20)fT 55 = 106.002 - .86154 T 50 - .10458 CNC 3


T21T50

(21) T 7 = .49661 T 55 + 205.886 P7

1121.786 P A
7 8
(22) WG7 = --

(23)IFG = .02951 WG7V!934.415 T 7 + 68558.365 + 2116.217 A 8 (.53978 P 7 - 1)

(P 3 /P 21 ) - 1
(24) ZC = _
P CNLIX 1

P 21
(25) ZF = 1 1 .
F<0 x
3.1

these can be readily observed in a DYNCEN listing, while others are far

more obscure. Equations (9) through (15) are approximations to the fan

and compressor block data maps.

2.3 Linearization and Order Reductions of Model 2

Model 2L5 was obtained through a very tedious and time-consuming

hand-calculated linearization. The partial derivative of each state

equation and nonlinear function was calculated, and then combined to-

f;ether to form a linear model. This linear model was then normalized

as follows.

Let A be n x n and let B be n x m. Then cacti state derivative may

be written

n m
Ia.. x. + b (2.3-1)
1 j=1 1J J j=1 1 J J

Let the values of x at the design point be denoted x, and denote the

normalized state variable as x

x. = x (2.3-2)
r i
x.
i

Similarly for the controls:

ui - ui
(2.3-3)
u.
i

Combinc (2.3-2) arri (2.3-3) with (2.3-1),

r•
n m
(x x_ a, x x+ b u u (2.3-4)
J=l lj J j j =1 lj j J

and simplify, resulting in


Thus, elements of the normalized matrices are -)btained by

(2.3-6)
a. ij = a.
lj _x]
x.
i

and

b = b.. u (2.3-7)
X.
1

The normalized linear Model 21.5 is given in Table 2.7. The eigenvalues c

of Model 2L5 are (1)(2) -7.2264 + 1.39L3j, (4) -73.554, (5) -153.27, and

(3) -343.11. The numbered eigenvalues can be associated with the state

of like number, as they bear a loose resemblance with the diagonal terms.

Note that all eigenvalues are negative, and the model is clearly stable.
h
It seems quite reasonable that lower order models would be almost
4F
t
as accurate, suggested by the clear difference in the m-gnitudes of the

eigenvalues. They will also be much easier to use to perform Dynamic

Programming studies, saving much storage space and c.p.u. time. As

mentioned above, the eigenvalues show that the states which will be

eliminated as the order is decreased, are P 4 (3), then U 4 (5), then P7(4).
i
t. The method used to perform the ardor reduction is to first rearrange

the states into partitions of "states to keep", Xl, and "states to

eliminate", x2.

tI
1. Xl = [ A ll A ]+Bu(2-3-8)
F ^2[X I 11

g I, X2 A21 A 2 ^ x 2 B2f

f
. Now set the derivative of X2 equal to zero, since their dynamic behavior

is to bq eliminated. Thus,
13

A21 X + A 22 X2 + B 2 u = 0 (2.3-9)

Solving for X2,

1 - B 2 u} (2.3-10)
X 2 = -A22 -1( 21 X

X2 is now replaced in equation (2.3-8) by its expression in (2.3--10),

yielding

X 1 = (A 11 X1 + (B 1 22. B2 )u (2.3-1.1)
Al2 A22^1A 21 ) - Al2 A

1 Application of this method yields Models 2L3 and 2L2 per Tables 2.8
t
and 2.9 respectively. Note that the most important eigenvalues have not

changed significantly in the model reductions.

2.4 Linearizations and Order Reductions of Model 1

As previously mentioned, linear models obtained from Model 1 will.

be useful in comparing with those obtained from Model 2. In addition, it

should yield a good model of the DYNCEN simulator in the area of the

design point.

The general method for obtaining numerical linearizations of Model ].

is outlined in reference [10], including all the necessary program inputs.

{ Additional insight into the selection of states for low order models is

provided by DYGABCD. This stems from the identification technique used

in DYGABCD, which is to perturb the inputs and states one aL a time, and

then measure the changes in each state derivative. A loose hierarchy of

states in terns of their importance in the model is obtained by measur-

ing how much each state perturbation affects the fan speed (which is

certainly one of the most important statcs). A close resemblance with

the choice of states for Model 2 occurs.


0
TABLE 2.7

MODEL 2L5

HdLrix Matrix Elements Eigenvalues

--12.549 -1.6928 3.584 .23786 1.7011 -343.11


,83305 -5.6135 1.645 .14381. 1.304 -1.53.27
A 671.66 387.71 -392.67 -26.1.6 160.53 --73,554
-104.15 21.135 64.925 -67.803 2,6314 -7.2264 + 1.3913j
50.953 -55.855 -81.205 -7.4745 -105.74

0 1.4078
0 .75817
B = 1.2813 -122.31
0 -48.928
149.21 -3.092

0 0 0 1.461 0
0 0 0 0 1

0 1.2138
0 0

TABLE 2.8

MODEL 2L3

Matrix Matrix Elements Eigenvalues

-8.4226 -1.2541 -.047955 -72.576


A 2.3957 -5.9244 .0048531 -7.1663 + 1.2441j
-11.566 56.677 -72.562

3.406 .79722
B 2.1241 .56161
31.486 -64.49

C 0 0 1.461
-.63298 -.97907 -.01486

0 1.2138
D
1.072 .1598
15

TABLE 2 .9

MODEL 2L2

Matrix Matrix Elements Eigenvalues

A -8.415 -1.2916 -7.1678 + 1.2401j


2.3949 -5.9206

B 3.3852 .83984
2.1262 .5573

C -.23288 1.1412
-.63061 -.99068

1 .63396 -.084644
D
I 1.0656 .17301

TABLE 2,10

MODEL 1L5

I'tatrlx I Matrix Elements ^- Eigenvalues

-3.8 -1.277 2.067 -1.152 1.448 -251.51 + 23.147j


2.748 -5.39 1.585 -1.991 1.071 -96.366
A 377.9 49.51 -264.9 86.807 78.91 -5.0491 + .83858j
31.26 139.39 -6.269 --88.69 27.83
-176.5 23.91 --10.27 -37.4 -246.7

-.00259 .3553
.2116 -.31618
B 12.54 -13.774
-.6201 -99.3
157.78 6.84

-.8594 -.1397 .6672 1.167 -.1236


C
.055591 .00656 -.001837 .01354 .85391

-.10277 .90094
D
-.013839 .020856

16

TABLE 2.11
MODEL 1L3
Matrix Matrix Elements Eigenvalues
-2.4307 -.70897 -.81149 -92.242
A 3.8281 -4.9579 -1.7235 -5.0644 _+ .86196j
2.4466 140.5 -94.982

1.395 .30875
B 1.2585 -.35303
15.44 -98.208

I
0 I .034897 -.0083832 1.3734
-.60014 .081353 -.12635

D - . 023854 .86846
.52351 .046147

TABLE 2.12

MODEL 1 L 2

Matrix Matrix Elements Eigenvalues

-2.4516 -1.91 -4.9795 + .91478j


A
3.7857 -7.5073

1.2631 1.1483
B
.97844 1.429

.070274 2.0232
O
-.60339 -.10555

.19932 -.55158
D
.50298 .17879
17

Table 2.10 details the results for Model 1L5 after normalization.

Inspection shows that Models 11,5 and 2L5 do not closely match on an

element-by-element basis, although there is not great disparity between

the eigenvalues of the two models.

Order reductions of Model 1L5 were performed by the same method

detailed in section 2.3, rather than direct use of DYGABCD. It was

felt that model 1L5 was a reasonable approximation to the DYNGEN simu-

lator at the design point, and it was desired not to rely heavily on an

experimental program. Tables 2.11 and 2.12 list Models 1L3 and 1L2

respectively. Again, the eigenvalues do not change appreciably after

the order reduction.

2.5 Co^ijharison of Model Responses

Concurrent with the development of Model 2 was the development of a

computer program describing Model 2 (see Appendix). It employs an

Euler integration and was used with a time step (DT) of .001, very

suitable in light of the values of the eigenvalues of the linear models.

DYNGEN was run with time steps of .01 and higher, for it employs a

modified Euler technique [3) which allows larger increments to be used.

In addition, the linear models were tested on program ABCD (see Appendix),

with a Runga-Kutta integration. Figures 2.2, 2.3, and 2.4 show

various time responses of Models 1. and 2. Unfortunately, the responses

shown are the closest Models 1 and 2 came towards agreement. The linear

r responses are evidence that the linearization and order reductions were

correctly calculated.

Comparisons of the various linear models are given in figures 2.5,

2.6, 2.7, and 2.8. There seems to be better agreement between models

for the frequency responses involving the'high pressure turbine inlet

temperature (T).
i
C 18

rt

.t O

sJ
-1

2
n 2
v
H
V U
t -2
O O
iA

V^
Z
G
Z 2
U ^-3
i^ e

o.s 0.6 0.0


7r^+i (cEC)

FIGURE 2.3. State Time Resnorises


0.2 r 0-6 < •e - Nozzle Input
• T rM E (src)

FIGURE 2.2. State Time Responses


- Fuel Input

7 0
e
v.
0
C,

C 4

w
V
Q
1
v ^i..s
r
`.^ 2

0.2. O.{ O.F O.fl


TIME (SfC)

FIGURE 2.L. Thrust Time 1


- ruel Input

to


1
zLz

ILz
=
e.1

1l.3

?L q 2L3

la 1co


FI 2.5• Magnitude Frequency response - WFB to FG

sL3,xL^
1.0
JL3, IL5

IL2

I 1


at 2L2
2

!O W TOO

FIGURE 2.6. Magnitude Frequency Response - X16 to FG


20

2L2, 2, L";
to

it2 fL3

1
BLS 2Ls

0.1

to w 100

FIGURE 2.7. Magnitude Frequency Response - WFB to T4

_,.10

fLZ 2L2
2 L.3

0.r-

%3

2.
\ f L s

2Lc

10 to Jon

FIGURE 2.8. Magnitude Fre q uency Response - Aa to Ty


^.^ VIFB
two.3
loo % N, = 11 8 99 RPM ^, 13...^ [j ^ TI w 7s
too % N F = 9874 R PM
]d0 ^^ O. S
^ el s
i MODE
L.^ ^^

MODELZ
99
^ ^ FBA ^'^ B / .

C0 0
qg , d
_ wF6
IV

,^,
97 , _ ,. `owFa

o NG WF5 j' t=6411


f
, 2A WF$ 9tv
JA
1484
f{ .. / s r
^I 2 , = °'
9^ g r

t ^'^ l• BROWS 5(GUIPY TRANSfENTS


txo.G 'o-s
a4
WF6
/ ,•''^ NO-ARROWS SIGNIFY OPERA7iNcG LIRES

a• i•

9Q
^^ Y
f......
a3 5'^ 9^ 96 97 98 9 9 Ala
.:, 22

Figure 2.9 is yet another cou-iparison of Models 1 and 2. It shows

the steady state equilibriums of both models as fuel, is varied. It is

not surprising that a change in fuel in Model. 2 produces a corresponding

change in steady state which is greater than Model 1 would produce.

This follows since Model 2L2 is known to have higher eigeava.lues than

Model 1L2. Also shown are transients for step inputs between a low A

thrust point and a high thrust point (the design point).


i
'.0

f ^

rJ
f
LL

ff
i

3 s

is

i

r

CHAPTER III

THE TIME OPTIMAL CONTROL WROBLM4

3. 1 Introduction

} In light of the disagreements between models found in Chapter II,

no single model will be relied upon to determine a time optimal control


I
law for the h et engine. The control problem is to determine a controller

which will drive a model from a low thrust equilibrium to the design

point, in minimum time, and subject to certain constraints, as yet

undetermined. In addition, it is desired that the controller be deter--


.
mined. by feedback control law, for the usual reasons: reduced sensitivity

to plant variations; control over system stability; and regarding pro-
%

gramming aspects, the ease at which a global solution can be obtained.

3.2 Statement of the Problem

The necessar y first st ep


P is to reformulate the models as discrete

f; time systems. Let

x(t + ot) = x(t) + dt • f(x(t), U(t)) (3.2--1)

represent the system with starting time k and terminal time N. It is

understood that


f(x(t), u ( t)) = Ax(t) + B u(t) . (3.2-2)
r: I
for linear models. Let x(k) be the starting state and let the terminal

time N be defined as the first instant at which the system state reaches

the designated target set S. All x(t) are eX, the state set. The per--
,
formance index

N--dt
S (x, ti) ^ Qt (3.2--3)
t=k
t = k, k + flt,...., N-et
t

i
.. . _......

L ., 24

is to be minimized with u(t) s U, the control set, and u defined as the

control sequence.

u = u(k), u(k + At),...u_(N At) (3.2-4)

Furthermore, the minimization is subj,-c:t to hard constraints of the form

g i (X(t), n(t)) < c i (3.2--5)

3.3 Constraint Determination

The final step in a complete formulation of the control problem

lies in the determination of the g i and ci of equation 0.2-5). There is

a strong intuitive need for such constraints, for the physical jet

engine has very real performance limitations. Brennan and Leake C81

have chosen turbine inlet temperature and surge margin as constraint

variables in their studies of the drone engine, and similar constraints

have been chosen for this study: (1) high pressure turbine inlet

temperature (`f4); (2) compressor surge margin (Z c); and (3) fan surge

margin (z p ). The surge margin of a compressor or fan is defined as


F'
if
- I }

Z = (p out/ p ;-n.) -- 1 (3.3--1)


P -- 1
max

if either the surge margin or the turbine temperature is too high, the

constraints will be violated. By definition, let

y. T4 = g (x(t), u ( t )) (3.3--2)

y Z c = g2 WO, u(t)) (3.3-3) a

G ^
ZF = g B (x(t) " ia(t)) (3.3--4)
s
is
The next step is to determine gi for each model.
{ Ems' Model 2 presents no difficulty whatsoever since all three constraint

s; variables are defined in the Chapter 11 development. It will be an easy

matter to incorporate these equations into subsequent control tests.

t
25

The constraints are harder to determine for the linear models, and

a starting point is needed. Control studies by Basso and Leake [12]

have used constraints which were strictly functions of the states. How-

ever, such is not the case here. Simulations of both Models l and 2

show T4 and Z to have very little steady state change over a wide range

of state space, yet step inputs elicit strong overshoots from both

variables. Clearly the constraints must be functions of both the states

and the inputs.

Once again, DYGABCD was used to obtain linear expressions for the

constraint variables. An order reduction was performed (per Chapter II)

yielding the g, for Model 1L2:

T4 = - .61459 x1 - .10759 x2 + .50292 u1 + .17689 u2 (3.3-5)

Zc -- .20154 x1 -- .45813 x2 + .20423 u + .14724 u2 (3.3-6)

ZE = -- .58229 x + .46872 x2 + .18877 u1 - .92545 u2 (3.3-7)

Constraint functions were not determined for the other linear models,

since Dynamic Programming solutions (see Chapter V) subject to constraints

were only obtained using Model. 1L2 and Model 2.

The final task remaining is to determine reasonable values for c

of the constraint equations. These c 1 will play a fundamental role in

the optimal control solutions of Chapter V, for they are hard con-

straints which will often affect the control chosen. After studying
results of DYNGPN simulations, it was decided to use the following
values:

cI = .150 (3.3--8)


c2 = .105 (3.3-9)

C3 = .080 (3.3-10)
1 ^ .

t
1
1

CHAPTER Iv

THE DYNAMIC PROGRAMMING METHOD

4. 1 Introduction

It has been pointed out in Chapter III that a feedback control law

is desired, rather than an open loop control. Furthermore, the Dynamic


Programming method has been extensively used in such situations to obtain.

numerical solutions. One of the more recent examples of its application

is found in reference (121, where Basso and Leake. have successfully ob-

tained a feedback control law for a single spool turbojet engine. Use

of Dynamic Programming methods to solve time optimal problems was shown

to involve a successive approximations technique.

4.2 Dynamic Programming Theory


The basic applications of the Dynamic Programming method are fixed

time, free right end problems. Let

x(t + dt) = x(t) + f(x(t), U(Q) (4.2-1)


with u(t) sU

The starting time k is known, and the terminal time N is known. The

target set is any x(N) s X. The object is to find

Vk W = min Jk (x, u) (4,2-2)

for a given initial state x, where


N-At
Jk (xyo) K(x(N)) + Y Z(x(t),u(t) t) (4.2-3)
t=k

Rewriting:
2T

The Principle of Optimality states that in order for the entire

state trajectory to be optimal from k to N, it has to be optimal from

k + At to N. Thus, equation (4.2-4) can be reformulated as

Jk (x,u) = L(x,u(k),k) + V + At (x(k + At)) (4.2--5)

which leads to

Vk (x) = min {L(x,u(k),k) + Vk + At (x(k + At))} (4.2-7)


u (k)

Since the minimization really only concerns u(k),(u(k + At) .... u(N - At)

are previously determined), u(k) can be defined as u, and Bellman's

Equation results;

= min {L(x,u,k) + Vk + At (x + f^(x,u))}


Vk(x) (4.2-5)
U

The boundary condition is

VN (x) = x (x(N)) (4.2-9)

These equations are necessary and sufficient for optimality.

4.3 Successive Approximations Technique

The task is to fit the time optimal problem (i.e., free time, fixed

right end), into a forms which can utilize the basic Dynamic Programming

method. This was developed by Leake, Liu, and Richardson in references

[13] and [14], and Later applied by Basso and Leake in [12].

As per [12], let Vk ( x) be any function such that Vk (x) > Vk(x)

and let vn ( x,k) be a control law which results when performing the

minimization. ]
i
min
U E:U
[L (x, u, k) + V + At (x + f (x, u)) ] (x, k) 4 S (4.3--1)
l
1

28

It is shown in [14] that if Vn ¢l (x) is the performance index

resulting from Vn (x,k), then


1

V (x) < Vk +l ( x) < Vkn (x) (4.3-2)

and further that V kn (x) converges monotonically to V k (x) in a finite

' number of steps, although each (x,k) may require a different number of

steps. Thus, it is concluded in [12] t]at

V k •^l (x) = min [L(x,u,k) + Vk (x + f(x,u))] (4.3--3)


• uEU
(x, k) f S

which very closely parallels Bellman's Equation. The only difference is

that in the solution of the fixed time, free right end problem, equation

(4.2--8) is relating two performance indices for the same state, but

separated by At in time whereas, equation (4.3-3) is relating two per-

formance indices for the same state, and the.same time k, one being a
r
better approximation than • the other.
t ,^

r It now appears that the time optimal free time free right end prob-

lem can be successfully solved, using the existing Dynamic Programming

method. indeed this is true for all practical purposes; however, there

is a slight discrepancy between the successive approximation theory and

its application to Dynamic Programming. To be specific, it is a fallacy,

to conclude that equation (4.3--3) guarantees that (4.3-2) be true. By

r definition, ilk ^l is the cost which results when applying V n (x,k), until
t
the target set S is reached, which is not equation (4.3-3). For example,
i
d let

V 'o-w - ^nax (4.3-4)


F3
(x, k)_ S.

and

^ _.... - .• _ _.. _uM.tx..•.:w•I..wli-+•+emu-rWU_wcur_^.^.s..r.^^W_,._ sr ..- _. ^.. .. _. _.


29

(x, k) E S
i
t
Then, if x is sufficiently far away from S, it is quite possible that

there exists no v(x,k) E U which will enable the equation

Vko (x + f (x, u)) = 0 (4.3-6)


to be true, i.e., the control- could not cause the system to reach the

target set in a time of At. Since equation (4.3--4) is true for all

(x,k) E S, then

Vk I (x) = min [L(x,u,k)] + Vk o (x) (4.3.7


uEU

and equation (4.3-2) is no longer valid. In practical situations, how-

ever, the method used in [12] and also used in Chapter V of this

study, using equation (4.3-3), will still converge.

A further simplification can be made when the control problem is

time-independent, which is the case in this study (see section 3.2).

Equation (4.3--3) simplifies to

Vn+lW min JL(x,u) + V n (x + f(x,u))] (4.3-8)


uEU (x,k) E S

1 4.4 'technique Refinements

One way of assu*ring that equation (4.3-2) will always be true is to

replace (4.3-8) with


Vn (x)
^. n+1
V (x) = min
min
uEU [L(x s u) + Vn (x + f(x,u))] (4.3-9)

Rewriting this in terms of the problem as described in Chapter III,

V n (x)
Vta+l (x) = min

lV i min [Qt + Vn (x + f(x,u))] (4.3-10) rv.


^uEU

r
v_
A successive- approximation problem allows still another departure .
from the basic Dynamic Programming problem (fixed time, free right end).

Let us examine how equation (4.2-8), describing a fixed time problem,


y
would be implemented on a computer. V N-Q (x) would be calculated for
all ^i e X' and stored •' .V (x) would be calculated and stored '' and so
N-2flt

forth,. Therefore, each iteration has a specific time associated with it.
However, in the successive approximations technique, either all app ro xi-

mations are concerned ;with the same time; or the 'problem' is time-inde-

pendent.
pendent. In Basso and Leake [121,'V (x) was calculated for all x.e X

and stored iii,aii array. Then V (x)-was..: calculated ,. and . .-after that.had .

be en . compxeted . far all. x "E X; V "(X) r epla c ed V (i) . in the array, and so
forth. it would be more :efficient , to immediately change each VI (^) to

{ the 'us calculated V2. x in a state b state manner. In. reality then

the approximations" .
for ' V changes -much more rapi dly, f o r one. d oes. not.

Wait unLil the ;completion: of the sw eet' t hrou gh state space before using

Q _. information'. derived, during that sweep . In this manner, :yew . nfoxmaton


i
becomes available at a , faster rate,; speeding. up the convergence to V(x) .
r
"Furthermore, if one.: starts . the
he:, state space s.weep : at the. target,. and
slowly moves away from the target, convergence will occur still faster:
u ,.
By start-in near the, taxg^t,: one is. t.es.ting c o ntrols for states. which
^? .,
can probably reach the target in a time of the order of. At. Since V o-(x)
:. o
for the target -is equal to zero, while guesses for v (x) at other slates

must be.made safely higher than the unknown solution, it is a benefit to


-start: at . a point where the i.ij ormation is the . best, letting the ins orma-
tion. propagate outward to other states. Figure 4. 1 showa the logic for
this.state:search. This l.agic:requires the .target to lie at the: center

Of the state space

..,.raw
32

TABLE 4.1

A COMPARISON OF COSTS FOR "FAST UPDATE" SUCCESSIVE


APPROX116ATIONS vs. THE "SLOW UPDATE" METHOD

is listed first for each state)


(V(x)slow

Iteration
#6 X1

{ o.85 0.95 1.05 1.15

1.15 0.5463 o.3825 0.2733 0.2925


. 0.468o 0.3530 0.2513 0.2678
I
1.05 0.4024 0.2222 0.1378 0.267+
0.3658 0.2156 0.1354 0.2525
X2
0.95 0.3101 0.1578 0.2190 0.3726
0.2890 0.1519 0.2126 0.3410

0.85 0.3477 0.2968 0.390 0.5184


0.3102 0.2643 0.3498 0.4480

Y f

E_

t
ii
^. 33

Asroof
P of the numerical su P eriorit y . of this "fast update"
P method,

Table 4.1-compares Dynamic Programming results for V6 (x), one obtained

through regular "do loop" sweeps through state space, the other by the

"fast: update" method.. Both started with the same V o (x). Note that the

x superscript on the..cost function no longer refers to the approximation

number,. but L;erves as a record of how many sweeps through state space :a

. ` have been made.. In the fast update of Table 4.1, there will have been

.; 6p2 approximations made, where p is the number of discrete points for

:each state in this second order. system. Of course, c.p.u. time is


;a
virtually identical for both program s.
x j

Unce th e number of actual approximations in the fast update method

is equal to the number of the sweep through state space times the number
^k
of points - ta state space, the finer the quantization, the more benefit

is derived through use of the fast.update method. An alternative expla-

nation: is that the old method (do loops) makes you wait even longer
E ;
} before obtaining new information, when you increase the quantization of
the state space.

.4.5.. General Program. Structure


One of the first considerations is c.p.u. time. This is a function

' of the number of points in..control space and state space, as well as the
ky
time increment At which is used.. (It should also be mentioned that this

3 y . refers to c.p.u. time..an ^n lBM:37Q:/158 computer.)-.


_-
In this study, indi--
vidual'programs:were limited to 14 minutes, 59 seconds,.to avoid addi-

tional j obi ,control language complications which occur for higher times.
i, Thus, 'single Dynamic Programming solutions may be the result of several

program. runs.. Near the 'end of the time limit,. each job st ores. cos.t

information on,
on.. disk to be used by a subsequent job as a: starting point, #
r

F^ _

34

LL nput quantization information and time step

Define the allowable points in X and U

DeHiiW --fLtial guess for V(x); read V(x) from disk if


program i_;^ _. a qpntinuat:.*

Loop for each successive approximation

Decide which state to test per Figure 4.1 algorithm

Loop for each u z U

st
.No re Yes
AL PAG9 IS
tj't e
constraint
s atisfd
. Determine x(t+At) 4UALIT%

T t

No
x 1 Yes
(t+At)F-X Interpolate to determine cost
?
iii ^
I t is stat
%a ii5st
4- 4-1
Yes
Save the cost and control
1

a contr6,k
No Yes
_s tisfy test
s 2 for his
all
Unco e
at
tates been
Yes J ^ispj&n arbitr ary
sted
\^teV
oi,
is s e
NO &isapp
malle
Cost
tim Smalle s
No /t
app
or
o store Cost
^._ ,_ - .^-- _ t result ?
Write es br, Yes
on disk ISate cost & contro l
r Yes
-Stop FIGURE A.2. Flow Chart of Dynamic Programming
55

0
V (x). In addition, control_ information is stored, so that the optimal

feedback control law will be easily accessed. by the simulators in

Chapter VI.
Dynamic Programming is generally best-written in a somewhat ad hoc

fashion. The number of subscripts in an array is dependent on the

order of the system, and interpolation schemes will differ according to

the dimension of state space. However, there still remains a basic

structure to the program. Figure 4.2 shows a flow chart, while the

actual program is contained in-,the Appendix Note the absence of a

redo loop" for searching the state space. Also, the target cost is set

to zero and left there, never allowing interpolation errors due to

1 quantization to occur. When controls are tested . for possible violation

of constraints, the values of the "present state" x(t) are used.Hover

ever, the 'future state" x(t + At) is used when testing whether or not a

particular control takes the state outside of the state set X. The

interpolation scheme is a standard method as used in reference (12] for

4 trao dimensions, and is analogously extended for third order models

(Model 2L3 in particular).

r.

C.
C

}
F
F ^

5
CHAPTER V

OPTIMAL CONTROL LAWS

5.1 Basic Problem Considerations

In order to compare control studies of the linear models of

Chapter II with studies of the nonlinear Model 2, linear affiue models

must be formulated. If the linear description of the system is

and the equilibrium values at the target are designated as x and u, then

the linear affine system is

x = (x - x) = A (x - x) + B (u - u) (5.1-2)

Similarly the constraint variables become

(y -- y) = C (x - x) + D (u -- u) (5.1-3)

where y is the equilibrium of the constraint variable. Since all the

linear models found in Chapter II were normalized, x = u = y = 1.

The time increment At (henceforth known as DT) for the linear


j
models was selected based or_ the eigenvalues of each system. In all

cases, DT = .01 seemed to be an acceptable choice, and convergence of

the approximations did occur with this value.

I. Quantization of the control and state spaces must be considered

next. In general., one would lake as fine a quantization as possible,

but practical limitations on the cpu time will dictate a compromise. It


nn
is desirable that the quantization of the state apace be small enough

such that the program does not rely too heavily on interpolation. How-
1 ever., if the state quantization becomes too small, DT trust also be

decreased. In other words, the amount by which a state can change in a

' time step L'-' will also have a bearing on the state quantization. For

36

t
37

example, at a point in state space near the target, it is possible that


is
the true optimal cost V(x) could be less than DT, if the state quantiza- t
J!
tiDn is too fine.
The presence of constraints is important, for one desires small

enough quantization to ascertain when the constraints are affecting the

choice of control. If quantization is coarse, it may be much harder to

recognize that *a control is riding a constraint.

A big factor in an optimal control solution is the definition of

the control set U, not only as regards the quantization, but also the

maximum and minimum values. These, of course, are chosen to reflect a

true physical situation, and as such, it is expected that they influence

the resultant control law. In these studies, the controls were limited

such that

0.5 < WFB < 1.4 (5.1-4)

0.7 < AS < 1.2 (5.1-5).

Again, these are normalized -values. The state set, X, does not affect

the solution for the states of interest, as long as these states are

sufficiently far from the boundaries of X.

5.2 Model 2L3 Unconstrained

The basic choice of V (x) for all models was

is
a
V (x) min V .70
max
1)2 + c (x 1)2 + c
C (X
2 2 3

with the c., chosen such that V'(x) V(x)- Whether or not this condi-

tion was satisfied was easily recognized by the success or failure of

the Vn (x) to converge to a solution.

J,
^^ . .^ _

._.. _
- ^ ^ .. .. ._.. i ..r __.i ....... .. ... __ y < ... i .. _

jY

0.30 0.92 a,`J4 Oo96 o.s$ i•90 1.a7- 1.a^ 1.46 i,a8 ^,1a

.* a.^a^ v•^v a,5a a,5a a.an o•^v v,^v c,^a. a•^9 v•av 0.50 ^ 1.20
iQ^a 5.

* a•$0 0•^a 9,aa vrsv 0.$a v•5a a.5a - a,^v ^.$v o•av v,3a. * 1.va
i.tl$
1.Ob ^ o•aQ ' tl•50 a• 59 0.50 " a•5 o ^ ' Os50 0.5a^ 0.5 a ^ 0.50 ^^ 9.5 a Oraa ^ 1rG6

^ 0.59 0.5a 0.59 0.50 O,aa 0,50 .. O,SO 0.501 0•a0 9r50 0.5D * x.•94
1.s0^4
1.02 ^ 0,5,1 a.50 0.5a 0x60 4.50 0,3d• Oa80 Or80 0.50 O.aO t^.50 ^ 1.02
NF' 1r90 * 0.50 1.3a 1 •:ra 3r39 1.20 1.OD 0.80 9•Fp 0.70 1.40 1.40 ^ 1.00

a,9$ * 1.4a 2.49 1.40 1.4a i•4a 1,4a 1.49 1.4v 1.40 1.40 1.40 ^ l?,98
^ 1,4a 1+40 1.40 2.40 1.40 1.44 1.40 1.40 1.40 1.4a^ 1,4a ^ a.95
a•9G
a•94 * 1 n 4a 1,40 2.40 2.49 1.40 1.40 1.40 2.40 1.40, X1.40 1.40 ^ a.94
0..92 1•'40_ 1.90 1.40 1.40 2.40 1.40 1.40 1.4G ^ 1.40 1.40 x.42 ^ a.s2
*
A.,9D ^ 1.40 1.40 1.40 1.40 1.40 1.49 1.40 1r40 1.40 1.4E1 1,40 ^ Ar90

0.90 13.92 0.94 9056 0,98 2000 1•U2 1•a4 .1.06 1.08 1.20.

FIGURE ^.1-A. Made1 2L3 - Opta.mal Co^tral Law (1. Control)

woa

" _

NC

Or98 4.9D o..9x r^.sa oo^^ 3x09 1x92 irO^F 1rDf► x.09 1,10
* ,^ *. ^ ^ ^ ^
1.14 * 0..242.E 0,2374• 9,2325 6.2273 0.2224 4.2159 9.$091 D.291:i a.isz8 a.ls^^ 0.172T *^ X•10
S,.D& * O^Z288:' -02231 0x2172 x.2307 0:2934 9r19*^1 Dr186b a.i7^a 4.1619 0.1+{96 4.1391, ^ 1.48
1.06 * 0x2140 0.2072 '; 0,997 0.1911 4x1813 0.1593 0.1653 br13a9 0,1215 0.1139 0x114'«5 *. 1rti^a
1x94 * 0.1977 - 0.1893 • 4x1793 9.1674 0.1530 ^` 041352 0.1139 0.9915 a.asut^ 0.11990 d•1120 ^ 1x04
1.02 * 0.1753 ^ 0.1684- 0x1546 0,1374 0.1153• O.0g72 6.0574 0.9749 0.99.22- 0x1129 0.1312 * 1,D2
NF 1x00 ^ D.1567 0.•7.4'30 0.1231 4.9976 4.0621 4.0 0.0681 U.0^373 9.1229 9^1y:25 4.1559 ^ Z.00
17.43 * D-.1344 0.7.179. 0.9977 p.0Tfi5 O.06fr4 0.0x72 0.3152 0-.1371 Dri541J (1x1671 0 s 177,2 ^• 0 .98
0.36- * ,D.1210 Dr1493 0.1911 0x101$ 0.1.162 0.1353 • 0.1539 9x1673 Oo1787 b.1880 D.1959 ^ D.96
9.9E * 0.1277 ^.i250 0.1306 0.1426 0.1566 4x1709 11,1616 0.1913 n.lss4 ^ 9.2us3 0.23,26 ^ 0.94
0.92 * 6..7.512. 4.15(.9 0..1565 0.1772 Ou1B74 •0.1965 D..2044 0x2113 0 x21712 0.22,26 0e2277 ^k . 0 x92
A.90 ^. Os1810 •17.1886 9.1967 D.2044 9.2115 0.2178 6.223K 0.2286 D.^2334. 0. 2 373 9.2435 .^ 0.94
^ ^ * ^ * +M
090
1.96 Or92 Orgy 0.96 0.98. 1.60 1.02
i. b+b 1.98 1s10

FTGi7RE ^ , Z-B . Model. 21 ,3 - Cosh { 1 Cont ;rot )


,,

^ - - c^.^ ^.,^._. ^.,^. _ _ - _.. ^ _ .._ _. .. _ ._ --- -^----^s

N^
n,sn. D,9^ D.sw a.s^ n.sa 1.4a .^.^^ 1.44 1,D6 l.,aa 1.10

1.10
'^`
1«10 * Oi70 D.70 Oi70 0.70 ,^ Oi^O a r 70 4vti 70 , ,7
0x 0 ^ O iTO 4x70 o,^a
^..oa ^ 0,50 0.5D Dv54 0.50 0050 0.50 D.50 O.SD 0,50 U.50 0.50 * 1.D^
0.70 0,70 4.70 4a74 0.70 D.74 ^ 0.70 4.70 '0.74 0.70 D.70
1.t16 * O0SO 4.50 0.50 405U 0.50 0.54 0.50 De50 -, 050 4x50 0.50 * 1.06
0.70 0.70 0.70 0.70 O.T4 ^ 4,70 OPTO 4.70 0.70 0.74 0.74
^..D4 4.00 0.50 4.50 Qe50 4.54 •0.60 0.50 4.50 0.54 0.50 D.50 ^ 1.44
^ 0,70 0.70 070
0.70 J,70 4.74 0.70 0.70 0.70 0.70 0.70
3a0Z ^ O.S4 Oo50 0,60 0.60 0.70 O,bO 0.50 O.SO 4.50 0.50 0.50 * 1.02
4.74 0.70 0.70 4.7a 0,70 4.00 0.80 0.70 4.70 4.70 1,20
1x00 ^ 4.5D 1.30 1.'50 1.30 1.40 1.00 4.64 4.70 0e70 1.40 1,40 ^^ 1000
^F D.7D 0.90 0.90 1.UD ladR 1.40 0.80 1.40 1.20 1020 1.20
fl.98 ^ i.44 Z.^k4 1x40 1.40 3.40 1.40 1.40 1.x+0 1.40 1.'^i0 X,40 * 0,`38
1.Z0 1v20 1.24 1.20 1.24 5.20 1.20 1.20 1.ZD 1.20 1.2a
tf.9i ^ 1,40 1x40 1.4D 1.44 1.40 1.40 1.40 1r40 1,40 1.40 1.µ0 #• t1.9^
1.20 1.20 1.2D 1020 1.20 1.20 71..24 1.20 i.20 l.^24 1.20
Q.9^F !^ 0
1. 40 1.44 1.40
1,.20
1.41!
1.20'
1.40
1.20
1040
1.20
1.40
1.20
i.44
1.,20
1x40
1.20
^ 1.40.
1.20
1,40
1.20
# 0.9 1} '
1.2.4 1.20
^ i.40 1.140 ^ 1.40 1.40 1.44 3..40 1.40 1.40 1.4D 1.µD 1.1+0 ^* Oo42.
G,9^ 124 1.20
1.20 7^.2D 1,20 1.2D 1.20 1.20 1.20 5..24 Sr^O
s n ^C * 1„4D 1.40 1.4D 1.'40 2.40 1.40 1.40 1,40 1.40 1.40 1.44 * D.9D
loxD 1x20 1x20 1.2D 1.20 1.20 1.20 1.20 1.20 lE2D 1.2D
^ * # * * ^k * * ^ # ^k

a.9D 1► .92 0.94 O.F6 4.98 1.00 1,02 1.D4 1r0b 1.00 1.10

FIGURE 5.2--A. Moc^^1 2L3 - Optimal Control Law {2 Controls)

v {' •:. ..-;


^. .:. _ .

- , . _.

uc
0.90 0.9^ 0r9^ 0,96 Or9d 1r00 1,02 ^r4^ 1r08 1 rQ8 1r10
* *• * * q^ * * ^ * *
;Lalfl ^ Or2i6:+ p,^21d 4r^163 4v211^ Ov$0^9+ 0,2nfi^ - 0.1937 Orx861 Or17?d Or16?^S -Or13b^. * 1r14
1.rOd * Oa213^! Oa^?0?^ Ov201=F^ 4e^9'S2 Ov1 g 6^i OvidO'3 Dci71'l Ori$Da' *
Or1604 4r:47'^ Oai32+1 ^ le0d
i.D^ * Or1987 Dv1918 '- Ov18uy 6 Dv1766 Qe?.b71 4v135d Dv1^22 Or1257 4v10'^6 ,OitlJ,̀il ' Dv0944 * iv.6
1s0^F ^ 4x1$27 Dv17^F4 ^ Da1652 Ov1539 0.1401 Dv1232 OviDl9 0.075} Ov0719 0.08{8 D.494b +R 1.04
1.02 * Oelb48 OrS54^F ti ^v1414 Dv12^6 Ov103'.5 0vD764. 0.0416 4.D55D 0.0763 0v0984 D.1167 * 1v0^
^^ 1.DD * 0.1443 Dv13D2 0>11G5 0.095,2 OvD51'Z D.v4 Ov0511 OaD$49 Oa11D4 4a127fa D;1T3^ * 1v00
0.9$ * Ov1231 Ov1022 OaD$,21 Dv0617 OaO^#9'S Ov0'162 DviD31 4.1238 Dv1395 Dv151? 0.1617 * Ov48
Ov9^ * 0.-1434 4,4910 0,0822 Dv0$3fs ` 0.1029 D.i228 Ov13g3 0.1526 0,1633 4x1722 Qv18D0 * Dv96
Ga9^ # O v iD7'8 Drzo bS D.1129 0.1279 n.142$ D.is57 D. 1b6S 0.1795 4.1832 Dv1940 Ov1963 * OvY^^
Or9.^ * Or1317 0.1393 Dv1509 0rib22 $,1,722 Ua1808, D*3652 0.1947 Or2005 0.2058 Dv2111 * Cr92
Or90 * Ovi'b31 Qv1720 0,I$Ob Oei$$3 Orl'I61 Dv2010 Dv^Ob4 0.2112 De213$ Ovi^242 Dv22^F7 * Ov90
^ * * ^a * ^a s *
Dv90 Or9a^ 4.94 O.S6 Dv9d i r00 1.02 ie0 +^ 1.46 ^ 1rD8 1.10

F^GTJ,^E ^.2 B. Mode3. 2L3 -- Cost {^ Cantro^.$)

^-

,.

^2
^.4

N The program was origixially developed using only one control, Wk'^i.

Su'6sequently, the A8 coxxtrol ores added but .conis^raints were riot yet

`= considered. Figures 5,l and 5.2 present the results, showixzg a s^.rigle

cross sectioxa of the three-dimensional state space, as defined by the

p?ane x3 {P ' } = 1.

Tt is interest^trig that the control, law Wb'B (x} rama^.ns basically

unchanged Frith the addition of the second input. However, the benefit

derived by its addition is clearly evident, sznce the two-input system

reaches the target roughly lol faster than the one--input system, for an y

given state.

Another significant result is that the optimal. contra]. soluti.an


for klode;. 21,3 is z*irtualiy the same for a gi^ren x l and x2 , regardless

E ^ of the choice of x3 . 1'he largest difference occurs at (x l = 1, x2 = 1,


4

x^ ^ 1} where V(x} = 0.Olfil seconds, as opposed to V(x) = 0 at the target

(xl = 1, x^ = 1, x3 = 1_). Tfie farther that x l and x^ are from the

^ target, the smaller the difference becomes. This suggests that a second

order model. would be satisfactory to use, and, indeed, it comes as no


':
surpr^.se considering the e^.genvalue information garnered in Chapter TT:
2:. ^,
P7 (afterburner pressuze} reacts so quickly that it on^.y slightly alters

the cost when x3 is far from the target.

^.3 Optimal Control.Th:eor


':
Henceforth, it is assumed. that secoxtd order models of the jet

^ er£gi^ie are entirely suitable for this study. Considering the law order

^-
:^;
of t11e system,, 3.t was decided that analytical apcimal control theory

s ^^
,, might provide good insight into the ultimate feedback controller
i`-,

solution,_ at the same time providing a means to check the, accuracy of

the nuiiierical progra:a. The analytical approach. is eXamined heze.


^3

Zet us fizst restate the genera]. prob^.em in continuous time:

Minimize

ti

J(xo ,u ,t) = K(x(t x }) i- ^ L(x(t},u(t),t}dt {S . 3-^.)


^t^'t^ t
0

} such that

C where t o =the starting time

i
i
t = the time at which the target set S is reached `,:,

' 3

^ ^ ^ (t 33
^i O o
'^
,;
:i

^^
}
t u{t) e U, the control. set

^`
(xf^^) =t,) ^ s
I?
^, ^

^^ ;j^ u(t =the continuous control. over the interval. from t o tot 9
yJ ^
rte ^ '

;' •.
t `^
^} - Na'Ge that ^3^e canstrain^s gi (x{^) ,u(^) } axe excluded from the prabZem. in ,
phis analytical, s^iady.

1 _;
}. 1

}: } ^,
^3
-^
i
i^ -...
^ .
Such that

U = ^u(t)^ min ^ uC^i ^. max

^^ ^ ^.. 1 °
_.

Thee are. two princ^.pal artaly.t^:ca1 approaches tri so^.ve this: px'obZeta,.
^..I`

}-^ and; Iio^h e^np7.oy tl^.e Aa^ti^ 3 ^on.ian A defined as


` ^ ^

,.
^.l
^(^>pru,.t) _ ^ P,f(x,u,t) ^ + p^L(x,u,t.) (^.3-5)
;.
,: ;

where p .i:s ^-^own.. as the adjoint var^.abla, and p equals . 3; in :-this case, .
a

the first approach ^l lt^ states that iz an 3nfinum:.7 exists for

equation: (5.3'3),,.

o. o a. o
i

o
v - - .
-. ' :i -. -

^y then it ^o1v'es the Hamilton; Jacobi equat3.on;


^;;
_ * ,^

- _.
a^ a^
..
.., o 0
rf ^ .
t -.

;:

:. r; 1:^
is
=f _
.^__ ..... ^:.^. _ __
s

1
F ^^

J' *(xo,ta) = K(xo ,to ) for (xo ,ty ) E S (5.3--8}


I

j`
3 i

I
Solution of equation {5.3-?) will yield a control v (x, aJ^ T (xo ,t o ) , t).
S a^

Zf this control does carry (x(t v ),to ) tv S, the target set, then the
i

control i.s optimal. Often the control law reduces to v (x), ^..e., a

pure state feedback control law. This xs similar tv Dynamzc Programming,

and, in fact, Bellman's equation (4,2-8) is actually a discretization of

the Hamilton-,7acobi equat^.on {5.3-7) .


In this case, an easier analytical approach is through application

of the M^.nimum Principle [ ^7 at,d the Hamilton Canon^.cal Equations.


i
't
Pontryagin's Minimum Pr^.nci.^le states the Ham^,l.tonian must bE :^ minimum
E
f

4 as a necessaryr cand^.tYOn for optimality, ^.. e. ,


^
H(x ,p^,u^,t} ^ H(x,g,u,t) ^ (5.3-9)
- 1,
^uEH
r ^

:, where u is the optimal control. The Canonical. Equations state that

v ^ ^^^

1 ax
_^
.
x — BHT (5.3-11.)
a
^g

^, such that {

^" < p,dx ^ - Hdt -- p dKl = 0 (5.3-12}


^^. ti
^;
3;. y
^, where the differentials of equation (5.3--12) are consistent with the J`
Ir ,

problem constra^.nts. Equation (5.3-12) is known as the Transversala.ty

Condition L^-^^

For the problem of interest,


r

i

^' ^:::
t e

Y -:- _.i-rf
n ,-^. - -. .- ,. ^^ Asa+..-...^--_..'_^ .. ^.. ..
... -^ ., V .1..
^' ;.
I !- - ^f
.^

-
Ii : ^ . < p., Ax + . Su > + X ^ ^ A ^ ; x } + ^ :B . p ^ n >' ,'-E-.. ^. (5 .:3^-^3 ^
-
L

..% t ,.

,.^
where p,^, `and. ii are vectors. To'mnirdize (5.3^^?}, u must' lie at an
^^: extreme poi^tt of its control set, depending- on the' signs of {B^p} l ..and
-
f ^ .^

`t (B p }g i.e.,
'p - - -
1 i...
T r^
^f (^ ^) x _ ^ ^g p) ---
_
i ^ (I} u 1 max. u 2 max 2. ,
-
. ^ ^ (-^) ^,^ mi:n: u2 max ' 3.f (BTp } i = + (B^p } 2 - -
^ !
,. ^ msn ^ min
t ^.

^,. l max 2 min ^. Z. y


^ .
,.-.
Thus we obtain ttie usual 4lbang-bang''. so^.uti.on; so cha^acter^.stc. of many.
,^

tilrte optimal cont^al pro^sl.ems. It is now. certa^.n that the controls wi3:1
t ^ °._
ride the boundaries of the control se.t U.: The probxem lies in: determ^.n:-
^ ^ - '.- .

ing what control. is apg^,^,ed when,, and for haw long.. This leads:ane to a.
..
^ ^

4
.. . ' ' switching
.. point analysXS.
.... ^ ^- . - :. -
.
^ -
. - ;
^ -

^: ^ ^n reference ,CII.^, Fon^ryaga^n shows a method for-.such an ana^.yss.

^ . ^
,' .. ^gt^aton,
-.
(5.3--14) shojis that 3.f. the tra^^.ctory of p (the a^.j,oint
`.
_. system}

^ ^ is known:, then the sw^.tchings are known.:.. Thus..Pontryagin. del^res 3nt.o .an
r. .
' } analysis of the ad^oint system, 'employing various transforzrtatians: and
94
translat^.ons to obtain.
. .
^ Ci

,} ^ _;
^+7
,'F
f
^... - ^

^^ i ^ ^^ 5.k^ Model 2L2 Unconstrained


!
I'oz Model 2I,2, unconstrained, cantr ra].s {I.) and- (3) as per equation
!:
.
t {5.3- 14) have time intervals of approximately 1.27 seconds, while can-
:. # ^
trots (2.) and (4) have time intervals of O.00:L^9 seconds. From this

,. ^
^.nformatiozf,.state space trajectories can be constructed. Since the

j solutions of (5.3^-I5} and (5.3-Z6) are basically sinusoids, the sequence

of cozztrols will begin at a particular control., depending on its loca-


G`i

tio3p. in state space, and fallow the control sequence in ^xumerical order

(.asd repeating) for the specified time intervals, until the target is

reached. phis, naturally means that the first and last intervals may

be shorter than the others.

Let us first construct trajectories for a number of initial points

in state space, given ,that only a single contro, is applied. As shown


r

in Fxo yes 5.3 through 5.6, each set of traj ectories has a point of

singularity, which is the equilibrium point of the system with the given
^'
control applied. Lath "x" mark shows a time interval of 0.1 second. As

`! the fa:gures show, there exists a single trajectory for each cantro'!

which will pass th^raugh the target. Let us start at the target, and

reconstruct the trajectory for each control, going backwards in time

{.. (Figure 5.7). Considering the optimal control knowledge embodied in


^a
equation (5.3--1^), it is clear that the final stage of a^ optimal. tra-
"''
^^R
je.ctory must ne cessarily . follow part or. all of one . of these arcs, a.n

order to reach the target. ^'arther^ore, if the last stage of an optimal

^. trajectory follows the trajectory due too control i to the target, then

it necessarily was mov^.ng in accordance with control, law i-I prior to

. ^ the switching. In this way, figure 5.8 can be constructed. `there will

o^^cur at most two switchings in the optimal control law for the ar^:a of
^r

state space which is of interest in this problem.


€;
L :^
^. .
._ . _ ..
t u8

^f

'r
0
i

^ ^ y
..

j
1

U
N

t O

1
t ^_

i a
I ry 0.
x^

0
0

11!
X ^ ^ ^ ^ .
L o

a ^ ^

^.6^ 0.70 0.8^ 0.90 1400 i.t0 i.2^ I. 30 1.40


x.

FIGURE 5.3. Model ;2L2 _. Control {1) Trajectories


^r c
4.
^-
^:
^^ Y

k
f

E
^ ii

;^

1.
kg

^fi
i

^;

^ o
'.

x•

t
U.liU ll. /U t]. Of? U.SIU 1.U(] 1^. SO t^„'I! 1.3^—j.ill]
x^

FTc3^FcE 5.^, l4o^e^. 2L2 - Control (2). Tra,jeci:ories


^.

^^
50

(f
.:,
^^.
m

li
i'. ..

i 0

0
N ^^
x -^

0m
cy'

6
m
O

O
O^ e

^^.E^7 0. 7D 0.8D 0.9D ^.DD


kl
1.ID l.tu ^^su ! 4U

FIGURE 5.5. Model. 2,L2 -• Control (3) 7'r^,j ectori es

`4

.-. ..-^,. •:? is ..^ __.^._ _, '-


-, y- .. .-
'^w..r.._

^1

04
ill
X"'

O
p

0m
^:

C7
M1

a
I

^^
^ 4
-T--'-- r-- ----,r-- C?.90 1 QQ 1.10 1.213 1.3 q 1 .4G^
^. 6D 0. 70 U. Bfl
X

F3GU^'tE 5.6. Model 2L2 -• Control (^ ► ) Trajectories



k .;g ^..,

• Trajectory due
to Control ^

^.+E

. ^ - ^^
- ^
T1^GET
tt^
Trajectory due to ^.,.,,^ Tra^jecto'ry due to 1
X2 l Control ^ ^c^r ^. Control. 2 for I
ma:^im^n ti^ze maximum time }
^.nterval. interval }

f ^. - E ^ -
Tra^ectbry due
to Control. ^

Xi
^'IGi1RE 5.7. Model 2L2 ^ Opti.ms.Z Control Synthesis

_^
l.-^e

:L -•2c 1 - E 1 I^- e 1.^2e


X
5=8. Model. 2^2 - 4iatimal Control. Synthesis
^„

^-

S3

0
x

0
M
Control 3

. 40
^1

FIGL^iE 5. g . Model 2L2 - Optim^.l Control Synt:^esis


-.... ..

Irt

13y plotting the four trajectory systems cn one graph, a comprsite

picture is obtained, showing the optimal trajectories for the entire

state space. Tltis is'shown iii Figure 5.9. Due to the relativel; short

time inters*al for which controls (2) and (4) are applied, they have a

negligible effect on the solution. In fact, for the scales used in


i

figure 5.9, these control regions do hot appear.

As an example, suppose the initial state is (1.05, 0.75). Control

^^.} would be applied for approximately 0.2 seconds, then control (2}

for 0.00149 seconds, and finally control (^} for apgroxi^nately 0.06

seconds. Control (2) could have been eliminated for all practical pur-

poses, In fact, that is precisely what occurs in the Dynamic Programming

for the problem. '

The Dynamic Programming results for Model 2L2, unconstrained, are

shown in Figure 5.10. In general, there seems to be goad agreement be--

r tween the analytical study and the numerical results iahen consider^.ng

the amount of time necessary to reach the target, V{^:}. The control

laws, however, are not exactly the same. While the Dynamic Programming

results shozr controls (l) and (3) to be optimum in the same areas (for

the most part} as the analytical results, the boundary areas between the

control regions do not agree as well as had been desired

This is accounted for by a simple explanation. For some states,

there exist control la^+^s (in the Dynamic Programming results} which are

not theoretically optimal. But when these control, laws are applied,
p^
^.: they yield costs which are so close to the optimal cost float, in a

^- nut^terical study subject to interpolation error, a non-optimal cost with

a carresponding non-optimal control lass may be chosen over an optimal

cost c,rith a carresponding optimal control law. Thus, a].tliough the

^...
f
I
^i
i
Q,90 A.9,2 0.94 0.95 Oe98 ^eQO 102 1.04 1.06 x•00 1.+!}

1,IC ^e ^ 4s5t} 0.50 D^5A 4.5D 4.50 4.54 D.5D 0,5D D^°5U Gr50 0.50 .d^ x.•10
0.70 4x70 0.74 0,70 0.70 Oe70 0.7U Us70 0.7t! 0.74 4 n 70
1e0^ to 0.54 0.54 0.50 0.5U 4.50 4.50 0.50 Ds5D D.50 a,54 0.50 # i.08
q ,70 0,70 4.70 0.70 0e70 0.70 -, D.70 D.70 0.70 0,70 0,70
3..06 ^c D.SD 0.84 0.5D D.50 0.50 '0,50 0,50 0.50 0.5C 4.50 0,50 ^ g.06
', 0,70 0.74 0.7D 4,74 D.70 4.70 0.70 0.70 0.74 D.70 0.70
71.O^F ^ 0.50 0.5D D,50 0,50 0.55 0.55 D,54 U,50 0.5D 4,50 0,50 .s E • A+}
a.70 0.74 0.70 0.70 U.70 t1.7a Oa70 0,7D 0.T0 0.70 - 0.70
5.02 ^k 0.50 D.50 U.SD 0.fi5 q.55 0.7U 4,50 4.50 0,54 q .50 0.50 ^ x.02
0.70 Or7D 0.70 0.70 4.70 0.70 0.70 0.70 0.70 O.7a 1,20
N^, x . q 0 * d,50 7 . 35 5„35 5.40 x . 4a x.UO 4,60 O.fi5 O.fiS 1 . ^FU x.4 0 * 1.dD
0,7'0 0,74 U,74 U.70 0.7a x.00 7.2G x.2U 3..20 ^.2U 3•,:20
4. •58 * x.40 x.^ D 1.40 x. 44 x. +^0 5.35 1,38 7..40 x. 4D x.^ 0 A.iFD ^ 0•`^8
1, •r_n x.za l,^u l.zo 5.24 x.ao 1.20 1.2 a 7.. 2 0 1.2a 1.2a
4.9b * 1,4 0 1.4 u x. 40 x. ^1U x. 40 1.4 a 5, 4 D x.^ca S.^a x. 4U x e^+D ^x ue^^
x,zo z,2u S,z4 x.zD x,?^ S.za x.2a 7..^a x.za x.xn a..aa
o.s^ ^ 7.4D x o44 x.^+a 7..4 0 1.44 x.^+a 1.4 0 1.4 0 1.^: D 1.44 x.4 a ^ o.g^
1.2D 1.20 1.20 1,20 1.20 x,20 x.20 ^, 1.20 ir20 1.20 x.20
O.S2 ^ 1.4U 31.4U la^4a 7..+kU x,+^a 1.4a 1.30 1.40 1.44 x.44 1,40 ^ fi,Y2
1.20 x.20 x,2a 1s20 x.20 Y.2a 1,2D 1.20 3..20 1.20 1.20
0,90 ^ i,4D 1,4a 1.44• 1.40 1,4a •1•.41 x,4D 1.40 1. ►^o I-.40 5.44 ^ Oe90
x.2D 3.•20 x.20 7.,20 x.20 7.. 2p x ► 20 x.20 1.2D ^ x.20 x.20

0,90 a. g z U.S^f 0.9f^ Oe96 . x. 00 3. . 02 1.04 I . D6 1.08 3.,10


.t

x'IGUR^ 5.^0 -A. Mad.e^. 2T^2 ^Uneons^xaa.ned} - .E?p^^.ma1 Can^xaZ haw


s

^,:, -
^n
^^.- ^ .n

f -

^ . ,, ..
^C
0,9fl Os92 0.94 p096 0098 1004 1s02 9•s04 1.05 1r08 1s 10

1r3.0 * Os^21'S fls2170 Os2124 0.2076 4ai!g24 401967 001902 001 ©27 Oo374 "s 0.164°5 0.1530 *^ 10?4
1.08 * 0s20$T 0s2034 0.3.578 0.1917 0s1.849 O 0 i77t 401678 0e1570 0s1jk43 0.1294• 0s117^k ^ 1.,0$ •
• 1 0 06 ^ Os1^46 0s1802 Os1E312 0.3.732 u'st637 0 0 152^k 0013Q9 O.i2^5 001027 Os p 922. Os0916 ^% 10g6
1a A4 * 0s179U Us1710 0.1618 U 0 1bU5 Os1.f57 4 0 1199 0.0989 0.0726 O.u59i Or07$1 0.0920 ^ a*.04
1 ;0r * Os16i4 0s1510 D01379 0.x213 O,i004 Oa0734 Os03Q5 0.0522 0 0 07^k3 0.0958 40i1^{0 * 1.02
Ids, 3.s GA :^ 011411 0s12b7 Os1fl7i 800 $ 22 000485 004 Os44a8 0 . 4[324 Os1073 001250 Os138'S * 1000
Us98 ^r 4s2-1^s9 4rU992 asi1794 U0U592.' Os447^F g 0 g T37 403U46 4.?212 Osi36'1 Q•a491 4.1591 ^ Us9$
0.96 ^ 0 0 1407 1100884 0.4797 fls4811 4;i0Q4 U,i202 0.1367 401540 0.1.E+06 4s1699 0 0 1773 ^ Os4r
Os9^k * 001g52^ 401041 As1i48 Oa1254 001402 0 0 9.531, p .163$ 0.7.729 0.1.805 0.1873 0s1935 rr U.9q
092 ^k 0.11~$1 Os13Fi7 (7.1484 9s19^}7 001G95 4si7$1 001$55 00192U 0 0 1978 UO2431 Os24$Zf' +ic flo`32
11090 ^ 0.160'.5 001695 0a178D 4 0 9.867 #]si924 0,1984 0s^0'..37 _ 4.2065 Oe213$ Os2174 0.2219 ^ 0a9Q
^ ^ ^ x^ ^* ^ ^ ^ ^ ^ ^
t1v9A 0v9Z Q094 0.96 0098 1e40 7;vfl2 1s0^t 1.0€1 1,48 1,1p

FIGURE 5.].0-^3. Mod^1 2L2 ^Unconstra^ ned.^ -Cast

,,,

rn

• ^.
.:^.,.....^.^:..^.^...,...J.,..^.._...^..._.,__.1...^.W ^^.^
.. .^_,^ ^^f ._.. _ _ __ _ _ ___
n ,;
^.
^,

;;
,.,
^a 57

3.
r:
;^
"' . f eren a 'bet een controls . ? and 3 is tar e ^aliera one would be
^; - ..
optimal and the other r^on aptxmal) , applicati.o.n of each at the part-i cu-
,^';
Iar state (in the v^.cini-^y of a boundary) ^riL'1 yield costs which. are
^^

^' ^ nearly equal.


.. ^, .

5.5 Made1 1L2 Unconstrained

^ Sur3:tch3.ng point analys^.s (per Sectibra 5.3}^foxl^iodel 1T;B sFiows.that.,


^; cantro3^s (1) and (3} have max^,mum time intervals of 1.22 seconds, wTriile
:,,:^ controls (2) and (4) have maximum ^.ntervaJ.s of 0.49 seconds. In this
_, ^ ^ case, a maximum of one sF .-^,tch^.ng may occur far the state space areas of
.. ^ _
interest. Figures 5.11 tt.rough 5..14 show the various txajectories for
1..^
-'.^ each control. 1n this cases, the singularity paints are not as close

^.^_^ together as they were for rlodel 2L2, and thus the different controls
r
- cause much different trajectari^s to occur. Recall that in Section 5.4,
t.
t ^^^ the close proximity of t^a'o s^:ngularities revealed that two^contrals Moth

^ had the. same ge^te^a^ effect on ttie stake trajectoxy, but one central •
^^
is

^' was altaays slightly better for a much larger area of the state space.
E

S The composite effect is shown iYl Figure 5.15. These results are
^, . ^.^
i

dramatically different from those. of Model 1Z2, which is not surprising.

i since the relative magnitudes of elements in the B matrices of the two

models is .very different.


r

Al^:haugh the results axe not included in th3.s 'ta'ork, it is a


relatively easy job to devise a controller based strictly on these
^,
,.^.' .^ analytical results. 5imp^.e knowledge of several paints along each of

^ the four swain trajectories (wha:ch co^►pletely define the control regions)
s:
will allodr one to fit a curve ^o each. bounciaxy..^-. Tn -:this -way:, the
- ,.
cantra^ is app^.ied accardisig to the xeg^.on of state space, cantinuo^lsly

^ .testing., as the trajectvrg-moves throughout .state space. • This wou1^, be

€' ^ '

..^

^^
_ _
,,
i
^_ ^ s 5a
F.

of

+;
f

[4
x

^,
^^
, o

'o
c^

... o

N
ID

- 6

N
M1 Y
i

X1.75 :• C.84 0.85


. . xi

FIGURE 5 . ].1. Mader 1L2 :- Cantra^. (]. } bra j ectaxi^s


;a

_'
^.
,.^

. ,,

:; ;,

^i .,_^ __ -- -..._...... o
{' ^ !

',, ^

va
N

.^

i
a
'^=-M.

62

FTGU7iF 5.x.5, Model 1L2 - Opt.ima^. Control Synthesis

^ AGE ^3

,^^^^ ^ au^^^^
^,
,r: .

1^I^

^1s90 De92 tI.9+3 0.46 0.98 8x00 1,42 8.04 ]L•05 1.08 2.10

+s1D ^k 1•^ 0 Is^F4 1s40 1 sL1'17 ? •40 1 s 30 0.70 Or 54 0•a0 O s $0 O s 5G rit 1,rQ
0.70 0.79 4.70 4.70 I1.74 0.70 Os7l"+ a.74 0.74 Os70 0.74
1x021 * 2s40 1.40 1n^F0 ls^i4 2•x}4 1•lU 0.50 U•50 - 0.50 4.5U 0.50 * 3.•4$
0,70 Oa7U IS•74 Us7U U•74 Os70 U•7U U•70 4x70 4r7U 0.7U
^.s4f^ * 1x40 1x 44 1 x4 0 1 siF 0 As 40 1•I4 U•50 0'50 Os 50 OrSU D•54 * 3.sOfa
0,74 U97U Uo70 Us7U U•70 ^ Os74 4s70 Us7U Uo70 Os7U 0,70
1r04 ^K 1sRF0 1.4U 3r.s40 Z.•.4U 7.•44 1,f15 Uo54 Os54 U454 U•54 U.50 * 3 ► 04
' 0,,70 Os70 U•70 Us7U Us70 0 ► 70 4`70 0.74 Q•%0 0070 4s70
1s0 `^ ^ z,40 1•+^U %•44 1.40 1x40 ? 0 0a 1],50 4,a0 Os5U OsSD 0.50 # 7.x02
0,70 Os74 Os74 4e70 4s7D 0.70 Us74 0.70 0.70 O,iO 0.70
N^ l o OD ^• 1.40 Ae44 1.44 101F0 1.40 1.00 0.50 Os50 0.50 0.50 4x54 +K 2sOD
1x 2 0 3.20 1s14 1.013 4x90 3.00 1.20 1x20 1x20 1x15 Is20
Do98 * 3.4Q ]..E}0 1.40 Ss$0 1.40 3*#0 1x35 1x34 4x50 4050 Os50 # 0.98
1e24 1.20 1.24 1x20 1.24 1.20 7..20 1.20 1.20 1s2U 1020
Ds9^ * Ys90 1s40 1.40 Ss4U 1•x}0 1x30 I.30 1x35 ls'-SU Ov'3U 0.50 ^ 41x96
I,24 1.20 102A 1.2U 1x20 1se'0 f•2D 1s2U 1s2U 71,,20 1.20
Ds9^f' # 7.•40 1.40 1x40 1.40 1x40 1.30 1,30 1x35 1.35 1e30 0.50 * q.91^
3sZ0 1n20 1.20 1.24 1r70 1,2U 1sZ0 Is24 1a20 1x20 1s20
41 ' 9?. ^ 1x44 a.rr3D 1,40 3.4U 1.40 1035 1035 1040 1.4U 1s4U 1.25 * t1s98
1x24 3.s 20 1x24 1x20 1.20 1x2`3_ 1020 1020 1s2U Is2U IO20
Ds9b * 1s t F0 1.44 1.44 1r^FD 3e^F0 7.040 1stF4 7..40 1x44 1.40 1.40 * Qe3C
1.20 1e24 1x24 : ls2D Is24 Zs20 I.20 1.20 .y.. 20 ko?. 4 1,20

' : 4x30 G•92 Gs9^? D^96 41.9$ 3,.40 1.42 a.s04 2x05 2.4D I.IQ

FIGtJR^ ^ . 3b -A. Model. 1.x,2 (Unconstrained } -- Oo ta .r^21 ion yrol Law

rn
w

l
^^: .., _^ _ ., ,a^+.^.:u:3,^,a..:...^x ^ ._,..^ - s,a...^.^=-_- , ^ u^,,.aad
^,::

i ^C
i i
F's

0x90 fl n 92 Oy94 Oe9Er Or9A Fs00 1y02 1x0!1


2x04 1x08 1x10
* * * # # * * # * ><
2.10 * 0x2928 Ua27'7 ra 0x21;16 Os2450 Or22$1 tDr2102 0.1510 Q.173.13 Os15Cs2 Or1►i8b 0s1 ri0!"i * 1,2p
1r0^ * Os2786 Or2G1$ 0.2439 Os2251 4.2053 Oy1842 0.1620 4e2^-21 0.111 0.1324 0.1439 * 3•.08
Yy05 # 0x2634 0x2447 Or 2 ,244 Os2U2b [1.1791 u s1535 0sll $ 2 0.1117 0x1118 ps1,2^9 Oa1k64 # 1 ► pb
2•Q4 # Os24Pa9 p r22b0 0. .2!127 0x17'69 Os148.:+ Us11,''i9 p r0401 Oe[}f313 Dr1058 Os1319 fls3.579 # 1e0^F
lr p 2 # Oa22SU Ua^054 0.17$5 0x1475 ,• 0.1111 E1^f157 a
r Dr0^52 psp835 0x1 4[31
O s1ATl7 (1e1756 # I.. 0,2
^^ 1, p D y5 Os209fi [1.1629 Os1514 Dr1134 " p s065'43 4e0 'pe050S Os1D52 0.14176 0.170,2 p p1955 * 1sDU
0's 913 * Os1^08 Ua7`fs25 4.3,308 Os0969 Oro&74 Us0705 0x113«.1 0.1470 0 e 174]. 0 r 3.970 9.2174 # u,98
Os3fi * 0.7,770 0.3,509 Os12w3 Os1U`^4 0.1018 (3.1222 0.1542 Om ],804 Or+^'024 - fls221,3 0s23^32 * 0.9b
0..94 * A y 1706 0,1501 0x1344 0sA291 0.1391 0.1818 Orl^i69 Or20$2 4.2265 Or?427 {3.2572 * 4.94
0,92• * OiR172U Uy3^a6$ 0r1a29. 0.157b Qr1T29 0.1 x.'37 Oa2142 Qr2319 0y247S 00261; 0.2743 # U r`12
iJ^30" * Q^1801 Q..17 j40 Or17S7 5.1858 TJra^.021 0 .2204 0.2374 Oe2S25 Os26f^D Os2784 0.2898 # 4 x9{3
* * * * * * * # * '
0.913 1!x92 Do94 0.96 Oy9& 1rOC 1.02 1x04 1s0fi 1.x08 2x10

FIGURE S.^.G•-$. Model ^.L2 ^Jncons^raa.nel3) - Cosh


z

. c1
^-

rem .u. nl ne-` ^W^w`e^s^-^.'iaa^^^ ^ tsiwbMgwl i . aV̀Sdi^E


' / ^ ,^

the method utilized in a controller simu:l.ation, if const^:aints did not

! need to be considered. ^^

t
The Dynamic Programming results for T`[adel 1L2, unconstrained, are ,_
shoran in Figure 5. J.G. Again, while there is excel lent agreement otx V (x)
bettaeen tlne analytical and the numerical results, the optimal control

latas do not precisely agree, for the same reasons as mentioned in

Section 5. 4.

5. G Model 11;2 Constrained


Analytical results are not possible when the constraints, as

developed in Chaptex IXI, are considered. Basically, these state-

Control constraints 3nay he interpreted as control constraints tahiclx

vary as a function of the state. Furthermore, it is not clear taltich

constraints affect the control later at a given state merely by studying

the trajectar_v of an optimal solution. If the constraints tt+ere functions


of the state only {as was the case in reference [12^},the optimal tra-

jectory would easily reveal wlxen tlxe control was riding a state consiraxnt.

The Dynamic Prograntm^.ng results for *Mdel 11;2, constrained, are


presented in T'igux-e 5.17, and the effects of the Constraints are seen in

liigure 5.15. Note that each of the constraints has an effect on the

I , optimal feedbacic control J.atti^ for some area of state space. In fact,
^.

there is only a small region of state space tahere the constraints do not

affect the solution. The main impact of these constraints is tlxat the

control taro no longer et*en resembJ.es a bang-bang Controller, but instead

^^ is a continuously chaixging function. The more finely quantized that the

control set U becomes, the smoother the control law tti^ill be.

.. .
:...:. _ .: _ . .5
^..;^• . . W. _
_. ,

^^

D.9D {'1v9^2 0.9t}, 4.96 R.35 -1.00 .1.02 1.4# 1•D6 1.48 1,14 .
* * ^ * # ^ ^ * m xr

1s10 :^ D o 64 Ds6'.S 0.50 Oe55 ^ 4,fan 0•^5 4.50 4varD n,50 Dr54 Ov50 * 1s10
0.95 4.95 0.94 4.90 4.90 0.90 ,•^. U,85 0•H5 0.85 ." Ov85 ny80
1.OQ ^ D•65 D.70 .•0.55 0.6D 0.65 0,70 Da55 0.8D D•5D 0.54 0a50 i# 1•.4$
4,95 Ds^S n.90 ^ 0.90 n.90 0.90 4aAb ns85 Do$5 n.85. 4 .sn
1 4 06 ^ 7.,15 O.aO 4.55 0.65 D.70 0.50 0.60 0,65 4.54 4ai'iD .. 0.54 * 1,06
1x45 Oe90 0.50 Oo90 4.94 4v0rJ U,05• D a85 0.85 4v80. 4 o0Q
1eDr} ^ ^ 0vfi4 D•70 0.75 Oo80 0•G^ 0.5D ^ Ov9D 0.5n 0,5n * l., nt}
^.US ^^ © ^ U.9D 4.90 4 . 94 Oo90 t3e85 0 v8'J- 0v&4 9.80 Or^iO
1a42 1.15 1s10 ts15 0.75 DeBD 4v85 4.50 ^ D•54 D•5D, D.5D 4v.50 ^ 1r D2
1,45 1000 1 •nn 0.90 9.94 4.94 4^•a5 4.8D• 1.10 1.Z4 1.2D
loD p :x• 1.16 1•;:6 ^ 7.•ZO 1a2D ^-25 1aDf1 0.5D 0a5U 4o5U 0y50' 0a.54 ^ ^. r.D 0
2v U 5 2rU4 1•UD .1•D0 I;pO 1oD0 1., 2D 1020 1.20 162D 1.20
• Da90 ^ z .15 •• 1.7.5 1o2R x.015 1,3.5 1.15 ^ t.15 .1.15 a;zn - ^ 0,65 - n .SD * 0..:'S
I0 D5 1.10 1vD5 1v^0 1.20 i ► 20 1.?D 1,^n 7.v2D 7i.s^U 1^2U
Os46 la^r.8 1.14 ~ . 1 a 15 ^ 1.15 1.15 1.2D 1.2D 1o2D ^ 1e23 1.3 0 ^ 1,30 #• Oa96
1•n5 •• 1,20 .i^.15 102U .. 1•ZD 1.20 1.24 1;2D 1.20 1•ZD 10.24
O s 9^3 ^!s 1,1b 1,10 1s25 1,15 ^ 1v1^a 1,24 1.211 Io25 1.2a 1.30 2.30 # 4v94
1.15 1«24 ".lal$ 1•ZO 7:•24 1.24 1.24 1x24 7.02D 8020 .1•.24
40}2 ^ 1v10 1.1D .101!] 1:1'-5 1,7.8 1.7.5 1.2Q 1.24 1a2'•i 1.25 I.25 •^ 0.`12.
1,15 1.24 ,3.•20 .7.•24 1,2D 1.2D 7..20 .1.20 1.24 1v2D 1.20
0.30 sx 1v1D 10D5 ^ .^05 1e7.D 1014 1.10 1015 1.15 1.24 7..20 :L.,20 *. 0 ' ^ 4
1.10 1.20 ^.:-4 1v^2R 1a20 1•Zn 1084 1.24 2.20 1.20 1v20

i],90 C,9,2 4r9^• Or96 n.s$ x044 1.02 7.v4f} 1005 7,00$ , l •7. p

^`^OTI^E 5•^-7-^• Nadel 2^,2 (Cons-^xaa.ned) -- O^S^ima^ Control. Law

ca
rn

.. . ,^ " _,. `. _
,.
^..: , .
f-

n,..., ,.. __,_.... ...


- ^,, ..

-1

Ne

4,90 0.92 Ds94 0.96 0,98 .3.40 3,42 1,04' 1^Dfi •. 3.08 .3,10
* ^ * * ^ xc ^: # ^ # x^
1 3 10 # Os4z19 Os40iCs. Oe3519 4.3750 0.3567 0.33SD 0.3107 Ds281a^. 0,247'4 4,2095 Ds172$ # 1,1D
1.08 # O.k045 9,3685 Os3706 8,3513 0,3296 0,3D37 Ds27'19 .0,2345 D,1909 0.1443 0,13$4 * 1, D8 '
1sOfi # 0s2^8',51 Oe3567 0.3463 0.3235 0,2964 0.2b28 Ds2215 Dr1712. 0,1212 0,1899 ^ 0x1471 * 1,06
1^D4 ^ 0.3E3^. 0s34^7 0.3178 4.285'5 x.2543 x,208° 0.150a 0,4919 0.1022 0,1374 U.i't03 * 1,44
3..42 # 4 . 3391=' 0.3226 0,2$31 ^ 0,2^bh5 Or1 '3$5 4,1316 O.D564 0 , 0860 0x1288 4s1641. Os1936 # 1, OZ
^F 1s00 ^k 0.31'S3 0.2759 4.2^F02 O.i897 U.12.8fJ U,4 0.0735 4.1232 4v1GU3 0.3.944 U.2159 * 1^UU
0,98 * 0.2882 4,8490'-0.2017 0.1435 O.D843 Q,0$99 0*1392. 0.169+} Os1969 4,21yg D.,2',5gg # Us98
D,96 # 4e2G48 0,2225 Os1T4$ Os1309 4+11.52 0,1481 R.Y79li,:. Os2053 D.286$ 0 , 2454 D.2623 ^ 0, 9b
0.54 * Os8453 0.2050 411681 O,1^D(3 Qo16^+?1 Os1945 4s2142, D ► 2341 D.25?.6 0,2673 Os281b # U,9^¢
Ds98 ^ 0.2'.528 Vr2004 4s1808 Oel$33 0,2025 Os2237 0.2420 _ Qs2582 0.272° 4.2x64 Ds29El9 ^ 0,`32
4x`30 ^ Ds2315 O,CD`39 0x2454 4.2313 Os^338 (!.2505 0,?&54 0x2790 Da251b Os30'Sk 0x3144 # Ds9R

0,90 0.92 0.94 8.96 Qs98 1. U0 1,02 2s0W a..0fi 1sD8 2x7.4 '

^'IG[T^ 5.1.7-^. Model. ^^2 {Constl"ain^d) -Cost

a ^.^_ .. ^_ --^- -^ -- _. ; .
... ,.
.._- -
^^:^..-.._...
r
^']

O^
{?
^:

.^
't`
!.

'^ ^j
}

^_
4

:-:^
{
I

Lf^
+-y
3 ^^
i

6
'^
r::a
9

aj
:\(i
"+^
1

L a
1
,— a

,: 3
RAZN1'S

^^
^; ^ ^ ^ ^
^ F
Y
^^^ ^ -. _;
X77

f g

1 ^ '

}, ^ -

f` ^,,,,,,o_,,,,,..,..--""ter
k ^ ^ 3

o^^,...

_ ^/7
`fMA

^"

g
g9

^^^
J
JJ `^. 8S Q. 91 q. ^7 1 . t13 1 . C19 1 . 1 ^ ,,^

;. ^C ^"^
^a '^
FZGURF 5.18. effects of Constraints on hfodel ^.L2 O^^t,s^na1 Control. T,s1k=
M . .. _.

f'V

^:

^.

J

b9
^^^^ ^ ^

E
S.7 1`iodel 2 Unconstrai ned
__
:^
^. After sufficiently analyzing the linear models, it now remains to

'_ ` study the nonlinear r[odel 2. No attempt will be made here at a non-
^^
!' linear system analysis, which would be mare complicated than the linear

analysis outlined in Section 5.3. )aurthermore, tahile the I3ynamic


i
i
^', Programming theory remains unchanged, the actual progranm^ing gets much

more complicated,

'^ If I~Iodel 2 could be embodied iii S state equations, each a function

of state variables and inputs, there still would be no significant


^_
^' -difference in programming difficulty from the linear models. Unfor-

tuF^^tely, riodel 2 consists of ^^any equations involving many intermediate

variables, and the number of operations required for each paint in state

space is dramatically increased. In fact, a fifth order solution would

^'^
1i .
require a prohibitive amount of c.p.u. time, and is automatically ruled

^^= ^ out. A method must nar^r be found to reduce the order of the system,
i
preferably to second order. Ideally, one desires to set the derivatives
^}

^' of the unwanted state variables to zero, just as was done in SecL-ions 2.3
,:.
and 2.4 with tha linear models. However, it i.s impossible to obtain a

;•' closed form solution for all the intermediate variables, with state

^^' variables 3,4, and 5 eliminated. This leads us to consider iterative

4, s.alutions.

^' ^°
^ If equations 1 through 25 of Table 2.G are to be solved {whi.ch is
,..
necessary to evaluate the state derivatives), then values for P^, U^ anti

^: P7 must somehow be determined. Recall that Dynamic Programming involves

'. a determination of x{t + fit} for a given ac and a given u. The first
,,
attempt, then, was to supply an initial. guess for the eliminated states

{P^, U4 , and P 7 ), holding 3v^, i^r , and the controls fixed, and iterate
^. `

^t

-;

----.--
f' j.,

>_€ ^ 0 r
7
3
I. sA

^. b
'
's .
until asteady--state solution raas reached.
^ ^ This method failed. for several reasons.. if one undergoes this
!_:
;'
iteration process tar each state and each control on every successive

^^ approximation, c.p.u. time is extremely high. Altex•natively, if one -:;


.:^
^ stores all the steady'state values for P 7 , P^, and V^ for each state j

and each control after the first approximation (e1.^.^ninatiug. the need-`ta -:`.:
it + -1
^ iterate on subsequent app:roxim^.tivns) a prohibitx•vely 'nigh amount of ^^
;•:
_ memory is required,. Furthermore, if an iztiti:al guess for , the e1.^.mxna:ted "^
^^
s ^ ' states i..s not close to the actual steady--state solution, ^.nstabiZii:ies

will actor and the system blows up. All of which requires us to look

# i for another solution. -


As a compromise to the problems encountered in determining values ,• '

far the eliminated states, linear approximations are obtained from _^


€ ^ ^
^ Model 2L5 and an order reduction is performed. This eliminates any i.p
^ ^
instability problems and also drastically reduces c.p.u. time. The }
j ^ 1

resulting equations are


'.^
P 4 = 1.4663 x 1 +.53032 x2 +.•40998 u l --,1$155 u^ (5.7-1) ;:

P = -.15436 x +.78107 x -^-.43343 u--.88875 u 2 {5.7-2)


7 1 2 1

,^ II4 = -.63063 x l --.99071 x 2 +1.0655 u l +.17300 u 2 (5.7 - 3)

^b ^ `

P These equations are them converted to linear affine for utilization Uy

^'- the program.


. ,
^° The Dynam3.c ^'rogramming computer program for Model 2 is found in •the
t , ':^

Appendix. Tt is divided into faun subroutines: (^.) the main program,

which is basically the Bynamic Pragramt^ting method as outlined ^.n Figure ^.:.

4.2, along tti*zth the constants used in Model^2 per Table 2,3; (2) the
^``
^, ..
.' static relations a.f rSodel 2 per Table 2..6; (3) the dynamic relations of
_ ^,-
4 ^:.
i^..
^ ;.
,.
:;

-- -: „- _ ,.
.^^...^...M..,.^.:...^.^....:...:.^...^, ^:.^
NG

0r9Er . 105 1,08 1r10 .


R, p U 0.92 Or `^^} 0,98 .1x40 1,02 .1.O^F

^c 1r3D 0.50 0.50 0.50 4.50 ^q50 4x50 ,0.54 0,50 ^. 0 * 54 ^ ^ D,SD ^ 1r7:0
7.e1G - 1, q 0 1.00 1,D0
1., D0 _ 3.1V 7.•04 1.00 1.00 .;.90 1.00 1,Ot3
1. q8 * 1r10 Or50 0.50- 0.54 Oe5 q 0.50 0.50 0.50 ^ Or54 4.50 0.50 ^k •1e08
1rU4 1•D4 i•00 1.40 3..04 1.00 ^ 1,00 1.OD 1.00 3.90 0,°4 .
' 1.40 0,50.;. 0.50 0.50 Og50 Or54 O,SO 0.5D ' . Oe50 4s54 O.SO * ^1r0$
1r9£s ^ 3..00 .1.00 1.00 0,9D OgBD
1.00 l .aD • , 1•oD 1.4V 7.•00 3,.00

4.5D .. 0.50 D, 50 0.50 .. 4+50. 0.50 0,50 ^ 1rD4


1x04 yfi 1r40 V,,raV 0.50 ^ UrSD
1.00 ^ 1,00 1.00 0,&0 D.74 0.70
D,90 1•D4 ^ 1.00 1rD0 ^ 1.00
1 n 40 0.54 ^- 0.50. Oo50 ..- 0:.50 4.50 ., 0.50 . - . 0•`•i0 :. 0.50 4•b0 D•90 ^ .1.T12
7.•02 ^ 0.70 Or70 .
4.90 1.DO ^ 1.00 1.,00 3..00 3.•00 0.50 0.74 ^ 0•TD
1rOb ^K• 1• urD 104.0 ^ 1.40 1s40 7.•40 ' ' 1.,OD D•$4 13.80 .' 0.90 0.90 x.'.SO •* •1.Op
k' 0.90 0r9D Os94 0,`30 4.9D 1.40 4,74 0.74 ^ 0.70 0.70 0.70

1.^F4 1.40 1.^}0 ^ 7:440 1.+40 1.^^J '.°..^+0 1,;F0 1'.4D 1,3D ^ ^q.98
Or°8 ^R 9..^t0
0,90 Or94 1•DO 1.20 ^ 1.20 0.70 Dr7D 4c70 0.70 4070 0.70
0,96 ^ 1r44 "da4rQ 1rGF0 1.40 1•++4 1.40 ^ 1.40 .1.40 ^ lr r+`J 1,40 1.30 ^ p.96
1.00 1a 20 1.20 1.2V 2.20 D•70 +;•70 Ur70 0,70 0.70 0,7U
D.°:; ^c. 1.40 1..x.4 ;.i^0 1 . 40 -.].,44 x.+FO 7.•^+D^ 1.^i0 1.4D 1. ►+0 1.x30 ^k 0.`^^F
1,24 1r2D 1x20 7,.24 1.20 1.24 0^7D D•70 Or70 - 0.74 0.70 ,

4ry2 4 1•;}0 1,e40 1.40 ' 1s44 1.40 ' . 1.4 q ^ 1•^+D ^ 1r ++D 1.40 3..I+V 1g3D •M^ •4t92
1,14 3..20 1.20 1r2U - 1.20 1r2D .1x20 1r20 V•70 ge7U 0.70
* 1,40 1r40 x•40 1.44 .,1.40 1x40 .1.40 .. 1.40 .1.40. 1.•40 i•20 * 4.90 _
Og94 1.20 0.70 0.70 0 ► 7D
1.20 1r20 1.24 3.24 1.20 .lr"^0 ^ 1020

• Or 90 O•`32 0.9^r Or`3E► . ^ 0.98 1r4^0 3x02- 1.Ot} Zr06 1r08 7.r 10
ti

^'^GURE 5.x.9 -A. Mo6te1 2 (Uncoxzs-Grained.} r Optimal Contxo3. Lew

Ev
NC

a.9a De92 0,94 0.95 0.98 n.00 1.02 .1.04 x.n^ 1.08 lain
^ r ^ •, ^ ^
^% ^
1.10 ^ Q.2420 8.2332 0.22116 0.2160 x•2974 a. 19as ^0.1 a95 O.i79S 0.1685 0,1559 0.1416 ^ 1.3,n
0.214$ 0.2052 0.19x6 ^ 0.1859 0.1755 0.1642 0.151.5 D.1366 0.1193 ^• 0.1n56 ^ 1.45
1,x6 * 0.20x4 0„194G 0.1839 9.1463 •
0.1'730 ^ Oa1B?3 0.133 q ^ 0.115£, O.OJ45 D.0$32 O.DB02 * 1.06
1.4^} * 0.1841 a.17^a °n.iana 4.1471 0.1322 0.13,4 1 {• Ov0929 O.n670 D.o62a
1.0+2 ^ 0.7.611 0.1475 x.133,1 O.i1b2 n.a9511 nr06$9 n.a363 4.0444 v.a^na.. 0.{1772 0.0920 *- 1.n^u
NF 1.00 * o.135v p .i196 o.loas 0.0769 O.at}53 n.0 0:0303 4.0629 Os0a33 0.0997 x.1120 ^ 1..04
0.x:3 ^ o.lvati U.4923 0.0734 0.0533 0.039s n,057z n.o7a2 O.n95u D.108'1 0.1206 0.13Sf3 ^ Q.98
x.96 * 0.0935 a.vao7 o.47n^ O,n679 0.0812 0x0960 ..0.3,092 n ► 12a5 0.1303 0.13$9 0,1.471 ^ 0.96
0.54 * 0.0451 n. 0912 O .n934 a.1a37 a, 11^16 0,1247 0,3.336 0.1416 0.1487 0.1553 0.1625 * U.91i
0.92 .#- 0.1131 0.1168 4.1246 0.1327 4.11}03 0.1473 4.1536 0.155+} 0 ^.16^-t3 0.171.►2 0.1757 ^ 4,92
0•`10 ^ •0.1383 0.14^4a 0.1500 0.1556: 0.1609 n.1G5$ 0.1704 x.1748 0,1792 0.182f^3 a.19a5 ^ Sf.90
* rt ^ ^ #
• '^
0p90 0.90 a n g4 Or96 0.99 -1.4 ►i
1rD0 1.02 1.06 1.08 ls1,u

FIGiJRE S 5.19-B. Node 2 (Unconstxai.n^d) - Cost

• ^

,. ,^i.
^ :. --^.^, _....,Y`.._^...:,__:......._.n.,. ^ , .^,.._- L.L. _:...,^^. _.^^_.^.^ .^.-_._-^___.^w .__. ,_,e^^.._... _w___.._..a...-,^ ^____. .._.__.^,^^.,^^...;^a,...^^.._..^__.^.^...,..,.,W - ^..f,
u
i
^

7 3 ^`

^, ;
^I s
Model.-2 per Table 2.5, Manus the eliminated state variables; and (4} ^.:

^-
equations for determination of P^, P 7 and U^,. per equations {5,7--I},

(5.7- 2} and (5.7-•3}. Note that subroutines ^SLaTd and riUD1;L2 utilize the

unnormalized systems while the main program and FAST utilize the

normalized system. Thus conversions from one system to the other are

made at several points an the program.

xt is also important to note ghat, even thaugh c.g.u. time has been

cut as much as possible, it still takes up to five times longer to obtain

^ 2. Dynamic programming Solution fox i^iodel 2 than for the linear naadels.

' •For this reason, it is important to use as much c.p.u. time as possible

• per jab run, but stall leaving enough time to insure that the results are
i
stored an disk before the allotted program time limit is exceeded. The

program itself insures that the results are safely stored, by measuring

how riuch c.p.u. time is reciuired for the first successive approximation,

and then using that infarmation to decide when to write the results on
- .I '
d3.slt.

{{ The first results presented in Figure 5.1.9 axe for Model 2, uncon--

strained, and are normalized values. The control law is similar to the

^^ solution fo g' Aiadel 2L2, unconstrained, but the cost is lass than the

= Model 2L2 cost for most points in state space. Zt is somewhat surprising
a
that nozzle area does not rode the limits of U at several states.

.^
^3
5. $ l^iadel 2 Constrained
^^^ ^ ^'iguxe 5.2(3 shotas the Dynamic 'Pragxanuning results far kiadel 2 faith
_
a' the constraint limits as specified in equations {3.3--8),(3.3-•9), and

(3.3--10), and using equations 23, 24, and 25 of Table 2.b. T^ Thi1e the
i.
:' cantraJ: can actually only ride one canstr'aint at a time,..there is a

^: k large area of state space where bath T^ and Z c axe very near their

1 `.
^^ _ ^.

..... ^ .,....
. ^. ^ ., .-._ .. ......, .^......^.. _.,_. ... ..._...r"._.--..,.. ,._ _ _ T. _. . ^ .-
c:._^_ _ _.. .. _ _
:T ^_._ ^.
7. -- _ ___.__ _. _ ^.
r..^ F
_. _. _.. ^. .. .. . .._. .,, .. .. _... _.

NC

` - 4o9a Oa98 0.94 0.9b 4,9A 1•DO 1,02 1,04 l,aE• 7.•D$ 1•ta -
^ * * * # * ^
X,18 +ic 4.5D 0.50 4.5D 0.50 a•54 D•50 0,5a 0 n 50 ae54 4.50 a•50 * 1•la
5•Za 1,^0 7..15 1•x"3 1.15 Sel•5 ,^ .3.,55 Zr15 1•SG lala 1.10
'^,
^^ 1r48 * 4'.'JD 4.50 U•5D D•50 U•54 .. 0•a4 U•'3D U ► 5a a•5U U•5U U•5a * Z,D6
z 1'.15 1,15 1,15 1,15 x.10 1,I0 l.la 1.1D 1.10 1•l a 1.05
`: 1.4f ^ Q.50 a•50 4.5R 0.5a 0•aD pe50 •0,5n q •50 0.50 D,5D 4.54 * 1s{t6
I,14 5•Sa 1•la 1.1D 1.14 I,14 i,a5 1,05 1x05 1.05 1.05
# 1.0$ ^ 4,50 0.54 ^ a,aD 0.50 0,50 0.50 0.5U 0.50 arrs0 0.5U 0.50 ^ 1•D^i,.
^I 1•la 1.24 1.05 1•U5 1.05 1.05 1,D5 5.45 1 .05 1.4D 1.Oa
1x02 ^k 1,UB 1•D5 4.54 R•5a 0.5D 0.55 0.50 D•54 a•50 a.l30 0•"^5 ^x 1,02
1•a0 3.a0 lra5 Sr05 S^aS 1•a5 ^ 3.•40 1.00 1.OR a•95 0.95
y N 1•DD ^^ 1,U5 5.45 ^ 3•SU 5•lU 1.14 SsUD 4,T0 U•75 U•9U 1•lU 1.2U M^^ 3,00
F 1.15 5.15 CI.95 7..Ua S,DS S.GO 5,20 5.2D 4.90 0.9D a•90' .
E'
0.98 ^ 1•a5 1.05 1. D5 1.50 1.1D 1.. D5 ^,1,SD 1.15 1,25 3.,15 1.2D * 0.9s
5.05 x•Sa 1.15 U•95 1•a5 1.20 7..3.0 ar95 U.$5 U•85 U,$5
0,96 ^R 1.05 1.05 1•D5 1•U5 1•a5 1.10 1^Sa 1.10 1 ► 15 1.25 1.1'a * 4.9G
r,E U•95 1.00 1.Oa 5.10 1.20 3.•Da 1,U5 1.05 O.B5 0.95 1.00
^;^ U,9^b ^ 1.0a 1.OS 1.05 1.05 1,05 1•la 1'10., 5.10 1,15 1.15 1.15 * 0.94
1.50 a•9a a•95 f•a5 1•a5 0.85 4.95 1.00 0•BD 0•$0 0,55
0.9s'' ak 1•UD 1.OU 1,05 1•U5 1•a5 1•U5 •^ 3.14 1.10 1.10 SoIU 1.1'3 * U•92
1,0a 1.05 0.85 0.90 3.00 1,00 U•85 ^ 4.90 a•93 D•95 a•ua
as90 5c 3,•00 3.OD 1.04 1.Oa X.05 1.05 .. 1,D5 1,7.0 1,14 1x1,0 1.10 * a,9a
0.9a a.95 1.00 3.00 0,90 0.95 -0„95 Da84 a•$5 a.90 U•30
8 xR rk ^ ^ , +R ^% # # %^ rk

^ Q,90 0.92 a•9^4 0.96 a•98 1•Qa loa2. I..gr3^ 1,06 1.08 1,1•.0

FIGURE x,20-^.. Mode. 2 (Constrained) -- Optzmal Control tiaw

.^-

i;
,. ,:
..:..

^ • ,'
,^.. ^..^.... _.^.^. M,. r ^..^ .....,. .-.
^^^ -
^
_ ^^ ^,
^.
^ ^ ^ ^ ^

f
', a

,.

:! .

.^C.
f.

^..
i _

r
a,^a a,9^ o.a^ a,^^ D.98 1,Ot7 .1. 02- 1vf1^8 l .os 1. D8 I:,Sa
' . - - * ^ * ?k :k ^ ^

1'.10 w 0.2677^.O.Z6U5 Q.2526 Os$^4^4 Oa2$55 0.22G1. - O..zJL57, q .2045 ' 0.191.5 0 ,.176 9 0,16p5 ' ^ .1',10
^.r0$ ^ 0.2509 D.2428 0.c'^'S5 OQ223$ 0c2129 0,2012. 0.1081 0417'3 0.x559 0 v 1361. 0,1223 * x.09
t
lv0b xe q .232^ 442227 0..2120 0 * ZOUt! ..1866 D,1715. 101.535 0.X330 0,10819 Os0993 0,101E ^ 1.06
f i
0.3.722 1.04 * 4.2123 U,20U6 0*1873 0s154B 9.1341 0.1484 .,0785 0,0776 . 0.4$94 0.146} * 1,04
IsfJ2 * 0,1$95 0.1753 0.1.583 0,13$5 0.a1µ5 0,4$6 ^ 0.0453 ..0609 0,406^F 0.1093 0,1275 ^ 1vD2
^^ 1400 * 4.I65U D,1455 Qr1227 0.0960 D.0585 D40
^ 4.955D 0.0921 0.1183 0.137$ x.1509 ^ 3.400
4,9b ^ Q 4 18i9 x.1613 4.1172 x.0947 .0,0869 4x11+46
O.i395 0.1.35'.,1 0.1509 0.7.636 a.17a^ * v.s$
^,
0 r 0.96 ^ a:2133
.1580 D.^93B 44i7ka ^ 0.1470 0,1452.' Oex5$6 .0.1705 0,1742 6.1$74 x,193$ * .496
^ ,.^ 0,94 ^ Or2462 x,22$6 0.2121 O.i976 0.19x4 041487 041^942^ .0+2040 0,2040 O.;^U90 v,2z2^ * u,94
a'.' 0+192 * 0.2783 gsE617 n.z^7u 4 n 2;i +F9 x.2262 0.223E 0.2z40^. ,. 0,2261. Qv2271 o,^zz^a 0,,2310 ^ u,9 2
f1..90 ^ vs3485 Rr2929 0.27$9 D.2fa75 0.2571 vr25,^5 0.250$..;; Or249^4• ^ Ov^485 0.2490. 0,2}97. +k 4.90
^ ^ • ^.
m i^. q 0
.9 0.92 0 r 9^• 6 .498 1.00 1.02- 1.04 2.06: 7..08 1.7s©

FSGUHE ^ . 20-^ . Model. 2 (Constr^ Fined} -Cost


{ ti
7^

f^
^.
G^
t;

': 'y

'_
'^

j
,: l

-^

r
c
^,

7T
i
^^

respective l^.m^.ts, and-thus they both are shown as azf pcting the ^,

solution in Figure 5.21. ^^ oily the constraint closest to its limit `^-^

is showns it would result in at least 10 smaller regions. In fact, it i

a.s unrealistic tti show these smaller reg^.ons, since the sarne^ahat coarse ^.,
^'

quantization of U used in this study' Y^ill have a strong effect on cahich i.


^:
^..::

c^onstraa.nt the controls are xising. The result i.s an almost random ^^` ,'j
1 .^

choice as to which constraint (Z c or T^) that the controls ride (in the
-_^;

T^ 2 c reg^.on.). Note also that. tie constxaints affect the solutio^t .u ►^ch ^,

differently than they did for riodel 1I.2 in Figure 5.18. ^.


This ]]ynamic Programming sal:^tian requires ^.pproximately 90 minutes

of c.p.u. time, using a 225-point state space search, and a 209-point !;,^^
!_
,._:

control space searcki. , '. ,

!:
^_
P'

^i
},
;;.
t

^^
^
;3^^

} -

^
;''
^^-.

:}^ '^'
^` .
., L • ^

^!$-
^^.

' C1iAPTlaR VJ
CONTROLL^P. 5I1+NLA7'1.ON5

^•^- Jntroductian

1. Na^ry that feedback control Jaws have been abtai.ned for several.

models, bath cansCrained and unconstxained, any state (N C , Nf,) can be

driven to the specified target in (approximately} minimum time. Thus,

initial starting states can be choser. which xeflect law-thrust candi--


`:
tioiis for bath AladeJ l and Afodel 2. The simulators road the feedback ;:
control loco from its disk storage and change the fuel. flow and the ^

nozzle area accordingly. This will necessarily involve an interpolation

scheme to decide cohere the txajectvey lies in state space, and is accom-
o i
pushed using the soma scheme as employed in the Dynamic Programming. '^

It is desirable that the quant^.zation of the control law be as fine as


,f
passible for these purposes, but limits an c.p.u. time ante again lead

^ to a compromise. '

1=nterpalation wi11 often lead to error when you axe it^terpalating


;.
;-
i
r
r

9 in a region of state space where the control laws change abruptT.y.


1
Obviously, the optimal control which is desired is either one extreme

,^ ar the other, and not sametliir^g in--between. In this case, the interpola-- `' ^
1
tion scheme could be overridden by an analytical test, similar to that
^.
which was mentioned in Secrian 5.5. Such considerations have not been ;.^
^^
implemented in this study. ^
_
The addition of constraints, and their resultant "smoothinu" effect
_ ^^

^„ on the original "bang--bang" control law, will reduce the number of abrupt

changes in the control law. Hence, the interpvlaL-ion scheme is expected


o ta be n.axc reliable tanen a pp1y in^g a control law which rasa derived sub-- y
Y
Sect to constraints.

_^_
f $.

-,--^-

. ... .. .
F

-. y4

^ 4

^' ^ 1

- LL
r

r
There is another very fundamereta-1 Consideration ^r^r .the controller
E
simulations. idhile the optimal. control laFi at some axhitrar}r state
^- '.
(^cl, x^) is quite possibly the same, or nearly the same as that for
,
^..
state (xl -i- c, x^ + d}, (s, S .small.} , this is not at all. true near the

target. The olatimal control law at tl-^e target is singly (u l = 1, uz = l)


^ ^ k:: °FCr the normalized s}Tstem. HoEtTever, this law is only valid for that

^ ^^ single point in state space tiahich defir_es the target. The state

(xI = 1 + c, x ? = 1 + S} will have an entirely dif^terent law.


- I ^,

i
^'f
In reality, we will consider the target to be sama small. region,
i
1 not an inxxnitesimally small point. As far as the simulations are con-
^
-

cerned, this small target region has already been detersuined by the

quantization used in the pynamic Programming. Whenever taoth states are

within ^x (the quantization size) of the target, the simulator will

begin to interpolate on the control lair (u l = 1, u2 ^ I) for the target.

This will cause the states to be slowly eased towards the target, which

might be considered unacceptable, depending on ha^,r large a region of

state space is xnvol^*ed. I?or this reason, these simulat^.ons det^^.^nine


approximate times at tohich the target oa^int is reached, and exact times

at which the target area is reached.

Since all the bynamic Programming solutions invo3.ve normalized

values, both simulators must convert to the unnorinalized system. A11

platted figures in this chapter. are normalized, for purposes of

comgarison.

Although the specified goal, i.s to tape a la«--th^:ust starting state

to the high-thrust target state in nu^ixmum time, the nature of ah. optimal

control study (and in particular, the Dynamic Programming method] is

that it is more concerned with state space and the time domain, As such,
^^

^^ so

^^ na plots of thrust are presented, although thrust plots would be vezy

Necessary i4x frequency domain transfer function studies of linear sys-

'^ terns. It is assumed here that satisfactory output responses are `'

,^
obtained when the system constraints are not violated.

The choice of the initial state for the fallowing simulations is


i
i somewhat affected by the peculiarities of the DYNGEN (Model 1} computer

program. Specification of the initial conditions is determined by an

"off-design point" (see reference [10]}, which is generally determined

by specification of WF'B and AS. Ideally, it would be desirable to

choose a starting state as far from the target as possible, in order to

demonstrate the usefulness of the global feedback control law. This

often would involve specif ication of an extremely low 1^PB, and is easily

accomplished. However, the contirol law at this paint will be a much

. i higher WF'B, aNd results in convergence problems once the transient simu-

lotion has begun. The success or. failure of the simulation to converge

is strongly controlled by the TOI^A.LL and DT variables (see referents


[10] and the pro gram, inputs given iN the Appendix} and amounts to much

trial.-and--error technique. Evan these variables cannot totally control.

the convergence difficulties, and further changes in the DYNGEN program

„. itself are sometimes rcauired, as e^-pinined in ^3].


EI

'^' Due tv the above considerations, a somewhat High initial condition

'^ is used in the simulations, the feeling bQing that 3otaer initial condi-

tiaras would require an unacceptable amount of tampering with the tapes

^? an which DYF^GE^I is stored. I'ar similar reasons, no control lairs which

^^' were obtained from unconstrained models are tested vn 1?Y1^TGEN.


1i w

i
^ _ _^.^._._,..,...^...._.,.T, - . - _ .- .___ ^__ _ _..,,...^. _.

r
{
^,:
sl
^:
G.Z Pfodel 2 Simulation Utilizing Mode, 2L2 (Unconstrained} Control Law
^,
The Model 2 controllers are implemented using the program shown in
^.

the Appendix. Figures G.l through G,4 show the results of a simulation

k• ti 7. _ ,', utilizing the Model 2L2 (unconstrained} control law, which was presented
^.
i in Section 5.^. The starting state chosen corresponds to a thrust q f
i,
f
approximately $Qf of design thrust. The effects of interpolation on

the control law axe readily seen in Figure G.l, showing the inputs ver-

sus time. After 0.08 seconds, the controls are slowly eased towards

their normalized design values {WI'B ^ 1, A$ = I). Figure 6.2 -reveals

an overshoot of approximately IDD^ far compressor speed (N C ) and 3DJ for

fan speed (Ids,), certainly not a desirable response for a jet engine.

Unfortunately, elimination of the constraints in the determination of


i

the corttral law has resulted in these undesirable consequences, as shown

zn Figure G.3, Turbine inlet tem p erature (T^) has skyrocketed to 5501%

of its design value in only D.02 seconds, and the surge margins have

also reached intolerable levels. The state space trajectory, as shat~=n

in Figure b.4, agrees remarkably well faith the optimal trajectories

which were analytically determined and presented in Figure 5,9 of the

^" ^ previous chapter. The time it tastes to reach the target is also in

,: agreement with: the cost results as presented in Figure 5.I0, approxi--

mately 0.12 seconds.


P^
^: G,3 I^fodel 2 Simulation, Utilizing Model 2 (Unconstrained} Control Lata
Figures 6.5 through 6,8 represent the results when utilizing the
^r

bfodel 2 (unconstrained) control law, as determined in Section 5.7. The

same initial state is used bare as far the linear controller of

Saction 6.2, and with remarkably close resu^.ts. The control laws axe

slightly different, but the resulting state and constraint variable '
.. .. ... ..^.._ .. ^.. ... _. . ^._m .... ;.

c
^-,

r_
Q

^"'^

^C
^^
C

C
C

cG,UU q .C15 U. 1Q Q.1^ 0. 2Q Q.zS Q. 30 o,^s Q.^V


TIME
^'SGTJRE 6.1, Model. 2 Sim ula^3on Utila.zin^ Model 2T^2 (Uzacons^rained) Control haw;

00
ro

^.. __^ ^.._.,...._. a .,_........^.,,.a


w
^r.<

-1 ...- < -- s r - v 4
^.

Q
^--^

^
fV!Y
f^

r^^
V J^

C
Q
C

TIME
FIGURE 6.2. P4odp1 2 Simul.a.tion Utilizing Mode3. 2L2 (Unconstrained) Control. Law; ,
lax (O,uantization) _ .02

,.

^ ....... . _ .....
^ z..._. - ,. ,_, .___ _ ..., . __ .__. _ , __ _ _. T._^....^ _. _._ ..._ ^.

.._.^ _e_ .^.,_. ^_. ._^_ --

c
^r

.:.::
,. :..
,,.,
c

U7 ^
^^
.,—;
^i
^-..
c
^c

^n
U

c
c

-c
c

^. eg o o, as a. ^ a a. is a. ^a a. ^s a. 3a ^. ^a a.^^
T^MF
FIGCIR^ 6.3. Mode. 2 53.mula^ion Uti^.a.zing ?^ode7. 2L2 (Unconstx^a7 ned) Control. Law;
Qx (Ruantization) = .02

a^

^ -. w. ..^ -: .. ^.._ - .ti. _ .^.--_ ... - - r._


r_._^..,..._^_..a..^ 1, -_. - -- - - _- _ _ _ .
.
^..^..^... ._._^_^_. .,^,.. .,.... ..._ _. ._ __
^.
......

:+^.
{ 3^

i
. ^

^,

n^
^
(
1 ^

i; . C7]
m
.
^ ^
,i
^- w„

. _r._.. _..._ ^-__-__--


X3.85 0.89 D. 98 D. 97 1 . O1 1 . OS
. N C ,,
FIGURE ^.^. Ntadel 2 Simulation IFtilizin^ Model 2T , 2 . '(llticanstrained) Control
^' i,aw; dx ( Qu3ntizat. ian^ = .02; marks indicate .al second
intervals

^f
^.
3.. ^:

a
i'

r^

^ ^-y ^
v/

Z =
F^

{^

-^.uu ^.u^ o. ^a a. i5 q .^a o. 25 ^^3^ ^. J^ L. ^U


T^ P^l E

F^G^ b,5. Mc^r^e^. 2 Simul,a^ion U^^.].izing Tdode3. 2 (Uncons-^rained) Con^ xoJ. ^,ara;

00
cr
t

T^

^4rnw - .^..^ ...,., _.......^.... .,,.....^..,.s^....w.^._ .m..,^^......_ _ ..._ .... _ x ..... _..._. _, .._ _.. ._^,..,..
r _, .. _ _.. _._ _..

j:

O
.--^

^f

. ^
L r,:

C
t:.

^'---' Et
U?^
C

C
Q
C


^. an a. a^ a. is o. is a. zo a. ^^ a. ^a a. ^s a. ^a
TAME
FZGUR^ 6.6, Model 2 Simulation Utilizing Model 2 (Unconstrained) Control Law;
ax (^uantizatzon) ^ .Q2

o,

i
... ,..
,.:. _. .: _
^• k.
^-

1 ^ ..

U. 35 a.^a
TIME
^'TGUR^ ^,7. Model 2 Sa:mulation Uti^.izin^ Model 2 (Unconstra^.ned) Control. ^a^r;
^x {Quantization} = .D2

^. _ ^
oa

,.. , .
^~ . 8g
^.

;i
,.

^^
^:

^T.

1 _..

,` (; ^
;l
i --^

-^4^
.^

I^ -
- ^
^f
m
az
c^
4-

^ ^' ,

^1
r
r^^
W
^Y^
}+^
^^ W

'1.
L;J

A.Z
i.
is ^
^^ ^

- r

^^

^. `^l.85 Q. 89 0.93 x.97 1.Cl1 1.C^5

a;
^ G a

^'^Gt]RE 6.8. Model 2 Simulation U^ilizin^ Model 2 {.LTncons^rained)


- ,^ Gontrnl Law-, Qx (Quantiza-^ion) ^ .02; marks indica-^e
`^„ ,.DZ second intervals
_ _
... . _ . ...... . ^..^.__
G. -_

F'' I
^,. f
9^

trajLctories are quite similar to the trajectories`produced with the

Model, 2L2 central law. This is certainly striking evidence that Model 2

^.s a nearly linear system.

6.4 rilodel 2 Simulation, 1Tt^.l.izin Mode? 2 (constrain ed) Control Law


The effects of constraints on the central'law can be quite clearly

demonstrated. Figures 6.R tlr,raugh 6.12 represent the results of the

Model 2 simulation, utilizing the Model 2 (constrained) control law,

as developed in Section 5.8. The starting state corresponds to a thrust

of approximately 74% of the design value. The centrals are considerably

'smoothed out, and Figure 6_10 shows that the state-time trajectozies
proceed to the target much slower and less abruptly, than was the case .

in the previous two sections. However, the constraints era now at

acceptable levels, as evidenced by Figure 6.11. Furthermore,. the central.

is riding both the turbine temperature and compressor surge margin con-

straints from the time ^ 0.02 seconds to time = 0.16 seconds. This

agrees with, the constraint analysis as shown in Figure 5.23. of the pre-

vious chapter. Unfortunately, the constraint limits, as given in

equations {3.3-8}, {3.3--9), and {3.3-10) are slightly exceeded, even in

this simulation. This i.s not entirely unexpected, wht^n ^^onsidering the

rather important fact that the effects of three state variables are not
^-
seen i.n these results. Recall that the optimal control law was derived
^^
^. (see Section 5.7),out of necessity, by employing linear approximations

for states (3), {1^}, and {5) of hfodel 2 {see Table 2.5). The simulation
.S7
!^^
^^ of the controller as presented in this section uses no such appraxima--

tiara, and hence, same variation is expected. Furthe^nare, the choice


^y
of initial conditions for states (3}, {^+), and (5) introduces yet

another consideration, and in .^ract, the particular choices for this


- ^,

simulation were soutewhat arbitrary. Regardless of these slight

^^^ !

^.,
^::

_.._ _^.__ .. _ _ .. __..__ - T-- — _


-^-- - ^ .. ^ ^._:a - --- - --_ ___ ..

C
n
.-

C
O
C
U3
}---

^C
^C
C

e
n
c

^.^. as a. a^ a. ^ Q a. 15 0. ^ 0 0. ^5 a, 3Q 0.35 ^. 43
^^M^
i F'TGURE 6.9. Model 2 SimuZa-^ian F3^^iliza,ng Model 2 (Constraa.ned) Control I,aw;
^x (Qus,ntization) _ .02

,^ --^.,...-....-.^.,^.^ .... -..^.,ti-._._ ........... ........ .^. .. _,.. ,._


,._

C)
---^

^.^
e

(!7
W
1--
CL
t" ' Lr

C
ff

u. UU ^, ^i^ [!. lU C.f. 1^ 0. ^U 0. ^5 0.30 0.3^ Q.aQ


T^^E
FIGURE 6.10. Model 2 S3.mula^ion Uri iizing Mode. 2 (Cons^x ai.ned) Con^ro^. ^ww;
^x (euan^izata.on} _ .D2

,. .,. — - - P_^ __- ^^x r^.-^-:.


^.
to

. ^
C

U!^r
^c

^c
^u
^^
U

U
^5

u. u^ u. u.^ u. 1u l!. J.5



C.E. ^^J LI. ^'.^ p. ^^ 0.35 a.^a
T^r^E
^^GURE 6.3.? . P^Ifld^^. 2 S^mula^ion ilt^.lizing Mad.e^. ' 2 { Constrained) Cnntro^. Zara;
dx {auantizs.ua.on) = . R2

w
'^

91E

z
w

^,

r
cn
n

cn

rn
ca
a

ce
C" ^— ,.^ .^
^^--
^7. $5 0. S g Q. 93 t]. S7 i . 0I i . Oa
N^
rIGLIRE &.12. Model ? Simul.ai.ion Uti1^ zing Mode]. 2 (Constrained)
Control I,aw; ax (Quuni:izatinn) = .Q2; marks indicate
• . [^^. second i nterval s
^^

' ^ 95

^, deiici.encies, the cost is in goad agreement with the Dynamic Programming

^-
>': results.
- 'r
il`;
6.5 Model l Simulation Utilizing PSadeZ 2 (Constrained) Control Law

• In view of all the evidence accumulated in Chapters 2 and 5 which

shaca the differences between Model 1 and Model 2, non--optimal results


;- i , .
r=^^ p are expected when applying an Model Z a control Zaw which was derived

^. from Model 2. Indeed, this is clearly the case, as demonstrated by


;'.
'^' figures 6,13 through 6.16. The starting state corresponds to an off-

design point of (WfB = 2.2, A$ : 2.95) on the DYNGEN simulator, and is


^;

the same starting point as was used x.n Section 6.G. While the constraa.nt

n, variables are within acceptable limits, the state-time trajectories

;• resemble very sloca ramp functions. It takes 0.3^ seconds to reach the
^^
v` target area and, while Model 1 is known to react more slowly than

Model 2, it is not expected that the cost be that high. Clearly the
i
^:
' control laY•r is not satisfactory.for use on Model 1.

!fy 6.6 Modal 1 Simulation Utilizing Model ZL2 {Constrained) ContraZ Law

Application of the Model ZL2 {constrained) control Zaw produces the

bast results far Model ^.. This is established by Figures 6.17 through

6.20, using the same starting state as the previous two simulations

(approximately 74% of design thrust). After 0.23 seconds, both rotor


..,
'^
;.. speeds are within 1.0% of their respective design values, a significantly

^, better perfar^nance than is provided by the T^Iode1 2 control l.aC,r. Zt is

^. somewhat slower, however, than the cost predicted by the Dynamic


r
Programming results of Section 5.5 (.205 seconds). This i5 not dis--

turbing, and perhaps quite satistacCory, when considering that Piodel Z

^^ is a Z6th order nonlinear simulation. It must be expec^:ed that the use

of a second order linear approximation in obtaining a control. Zaw cannot-


. ,.
,:^p

^^

i
•."__-3 F/1^4 ^'^ ^. ^.LL..i i.^.::..W-i :
L=. «a:: J [^^:+i e^^a^wuf •.i,.a, .._^ .^ _e: ^uF L.^..:.'.: 37 L.... C--_^.•.-. t'-^...:;i
1
^ i.^.^ ^ ^

'^
1..1 I

r1

G
C
CJ7
}^—

Q~
^ C
a.-_.. C_

C
r.

C
C

^Q.QQ Q. QS Q.fQ Q.IS D.?Q Q.25 Q.3C Q.3^ Q,^O


I ^ ^( c
FIGURE &.13. Model 1 Simu^.ation U^i.lizing i^Iadel 2 ^Cans-^rained} CanLrol Lair;
^^ (Quan^^. zs. pion) ^ . 02

w
a`
r,

,.. -.. ...-. _.. _ .,.-.__., ....,-Y. -

I! ^..^ _. ....

C]

E
L

W
F---

F— u
^^
C

C
G
C

^a. o0 0.05 0. in ^. 15 0. ^0 n. 25 n.3 n a,^5 n. ^n


TIME

FTGiJRE 6.3.x. Model l S^.muls.^Ei.on U^zXa.za.ng Model 2 iCons-^xe.a.ned} Con-^rnl ^,sw;


^x (Quan^iz^tion} = .02

— _ :.k
o:

r,^

t!7 ^
F-^-
^^
Q:
F--
^ C

© ^
(_}

i_r
^n
tl

`C. 00 ^. ^S 0. 1 ^ Q. 15 ^. 2d 0. c5 Q. 3Q a.3^ a.^r^


TAME
FIGURE 6.15. Model 1 Simulation Utilizing I ylodel 2 (Constrs,ined} Control I,aw;
0x (Quantization} _ .d2

ou

_ _ _ w:z.^
99

.j

`^

cn

!^

cr,
Q
c

6
cc

u
tc

^^.85 0^.$S 0.93 0^.9? 1.01 1^.U5


^C

FIGUFcIr 6.1G. Model 1 ^'imulatian utilizing Afodel. 2 (Constrained)


Control Lntr; 4x (Quantiz^.tion) = ,G2;' nEarks indicate
.G? second intervals

. u, - ............... . ^.^...-----_ .^.^_ _.. _. _ . _


IQ
^.+ '

^,

^`'^ C
V

n^
es

^ ^
H L

C
C
_ 1

`0.00 0. C^5 0.10 0 . i 5 0 . ^0 0. ^5 0. 30 0.35 0.40


rTME
FIGURE 6.17. i+^ode1 1 Simulation Utilizing Model 1L2 tConstrained) Control Law;
^x (Quantization) _ .Ql

00

_. _ LL _ .
,. ,,
^., ..
^-. . , - -

^W.,.. e
J Y. ^.-U ^_t.-.7
9 c-.. e v

L
e

G'?
LtJ
}--

1--- u
U7 G
C

C
a
C


o.^s d,^U
TIME.
SILTULYIl 6..^.$. Simulation Utilizing MOd^3. ^.L2 (Constxsa.ned) Can^xo'! Law;
Moc1e]. 7.
ex (^3uan^ization^ = .Ol

N
0
'i • F-' i

__ ^_ _
.._...^i]1vYii^Waiwlmao:
.. ..., ,.. . Y: -, .,-....,- , . _.:.., _. ,.,_.,z
__

r,^

Cl7 r
F--
^^
^-.i
. ^
fr
I--
U7^
^^
oC
U

u
c

TIME

^'TGURE b.^9. P9ode^, ^. Sxmul.atian U^iiizing Model ].L2 ( C


ax (Quantization} _ .fl1
onstrained) Cantral Lavr;

I` :..
..r^.....s.,_y_,.W..:., ..^..r. w...,^.....^.:. ^ ^._._.. ,__ u. .., ,...,..-....^.........^ ^..^...,.e....... . _............_^ ._.,...__....e.r::s,..,..^...y.^.a.....,,....^_..,....^..,,.^....,..z..^.^.,......»....i:{i:r.lua^:d.cd1

103

a'

LL
!
cr
6
C

E-
i

^. CC a

I
i

. L3 ^J

FIGURE 6.20. Model 1 Simulation Utilizing Model 1L2 {Cons-trained} ^-


Control Law; qx {Quantizatsan} _ .Ol;^m^.rks indicate '
.Ol second intervals

»^a

h='
^.. ,^: __ .. .
ti
a

10^t

possibly result in the exact prediction of the minimwn time it takes the
i
tra^ecto:V to reach the target. The control gaxte clearly rides the
f

turb^.ne . nlet temperature and campressvr surge margin constraints over

most o^ the trajectory, in agreement with the constraint analysis of

Figure 5.l$.
C^IAPTFR VII

SUP^iL^Y

The goal of this work ^ti*as to obtain a global nonlinear optimal

control for a two spool turbofan jet engine. Various mode's were

developed, pursuant to this goad.. Most important of these models was

the nonlinear analytically-expressed Model 2, which correctly models

most of the qualitative behavior of the jet engine, but trhich fails to
achieve strong nvn^erical agreement with the non-analytical Aiodel 1

simulator. The time optimal control program was then expressed in

detail, and various constraints were added to the problem. Dynamic

Progran^ning theory and the Successive Approximations technique were

explored, and applied to the problem of interest, while several improve-

ments in tine numerical programming were introduced. Analytical and

numerical results were obtained for several models, both constrained

and unconstrained. Finally, these resulL-s were tested on the Lira

principal simulators, Model 1 and Model 2.

Indeed, this study has successfully achieved time optimal feedback

control laws far various models of the two-spool turbofan jet engine.

Furthermore, valuable insight into the nature of the problem has bean

obtained, and much useful computer software has been developed. Unfor-

tunately, all enthusiasm far the results achieved 'in this study must be

tempered by the realization that an optimal control late obtained from

any model can only be as good as the model itself. For this reason,
r....

mare work is needed to develop a better nonlinear analytical model,

similar to Model 2 as presented in this study.


^^
^:
As the accuracy of these models is further improved; more consid-

'! eration should also be given to the details trhich so grearly influence

the time optimal feedback central law: the determination of the

Z05
t

i^
a N
l06

;;
"^ allowable controls, U, and the 'limits placed on L he selected consCraint

' variables. As Che entire analytical problem formulation (model, con--


- ^^.

sCraints, etc.) becomes closer and closer to the actual pltysical problem,

p ^. Wrote detailed salvtions can then be obtained in the nutilerical analysis.

In conclusion, this study should be vietred as one more step in the

^ ^ efforts to ac. ,t=e global optimal control laws for Ciao-spool turbofan
jar engines. Zt has accomplished much of its original goal, bur leaves

much more trorlt remaining.

F. ,
i

i .
8
9

^^

^ I

f
^::

3
1

is

^,

r
K
"1

t ^:
4 . --

4
1
^:—^

1
a
f-

L; . . .r..:_ . .. __ _ _ .
.. _ ... .._ ___
^,.-.
{,

;^^

et:

-'-_ ^f`

^ippez^axx G

^.

arDTR^CT ^i^'^'HOD ^'.OR O^TAZNSi^G NOIvTL^NEAR

^T^Y'^:^CAL ^I3E^,S 0^ A 3ET ENGIId^'a

+^' R. J. Eeake
^. G. Camsk^y

►^
E

^a

_ . _. ..
-

f 1^1
tit t ?.rJ- d ^ ^T ^`^L' .

-^_:^LLCTl'.^tQ?^IICS
.13^ t ^ s ^ . •^
p ^^

} car./,^cz^.^^Tc^
+K t.^l^R^,Rr r,1^/.1^P,
fask Ar p ok Cietutivf Flaza #2 1223 ^f. 22nd SL. • Oah Flta p k, lII. 6b523 JCC 311 72a S7e0 is f ^^nr^+/na

^^^ ^^

` ^ ,

'

. ^i^TERi^AT^^^t^L FoR^I^'i
^+ l

^LT^^^i^TIVFS Fob LI^I^AR ^^^ULTIV^^^^^L^ co^T^oL

.^ IIgat^ Regency O'Hare Hotel .


a a ^^.1.inois, i^SA
^^1_%ca a s
' ©ctobex 3.3-^^, 1977

.,

-"

6Pbl'd$OIt$ OF NATIONAL. CLLCt170HIC9 COHF[l1CI[CC HATlONAt COfi ►fUNICATIOk x 61 6R y 1.1 f TlIOTt5519HAL CROWYH 1N CHGINtlIIIl^IC. INST[TUTCS ^ "^

' ^
1.6 1^4EnSu£fiS ^'' ILL1N476 iH'JTIT UYC tlF TlCH NOLgGY NaftT H^VF7 TtRN UNtYrt : ^ITY UNIVCi[ 'S 17Y OF SaI [ NrrCTO tic
IOWA LiAiC UHIL'! II aIYY TNK dlll p Si ATC UN+V^ ra}[TY YNL UNiV LRSI * Y Or kISYU UHS .LT woLL4
I ^)^"`^^ }^}) LIA pq Ul'iTC UHIL' 111t1tY FilhGUL' 11Hlt' ! li!ItTY l7NIYCItGITY p l' HPTRY 17AS^C
^^' i^ a11CHICAN LTATC UI+IIVr; p Sl7Y YHS UNlYLRS17Y OF LtIC MILAN NNIYClt31TY p P LVIIICCN S IH

:. ... - j

- . ^ ^

7 R

^ .
i

A DxP^ECT rfET^IflD FOR D3T^I:II':G i^a::LI;"E^LR _

' ^ ^I^LYTICAL r1a_DELS aF A .7ET E^GItiE _

^. ^ R. J. Leaks and .T. G. Comiskey

Department o£ Electzical Engineering


^ r Jni.versity of A'otre Dame
, , .
' Y ^ Ilotre Dame, Indiana •, 4G5S6 ^.

AB^IZACT

w_ This paper describes an algorithm fox obtaining a nonlinear analytical model. oi' a
het engine from measurements of two equilibriu^^ point values and the linearized A
and S matrices at those points. The method is compared ^.^ith rote conventional proce-
._
;" dares of interconnecting individual component ^pprq^i^tions. .

^ '
..,

-^
...

^r

r =':

+ .1 i

._ ^a

^^
^i

^3`
`E ^ ^ 1 ^ t

^J c:
,ti
^^ • •
7
^— _.
.:i.
I:r
ri1 O^ r ^.

u! ; ,^ ^ •'

s_ .. .... _. .. .. i

n=^ . + •

'^^' S ^ ^

^ • ^

_^_
,... _ ..__.. .r
+ _., .. ,

. ' .

^^
i
^yy

^ DrFZr•.c^ ricTFion roR oBTt^rrrrNG Nol^LLl^Etilt

^aALY^'zc^,^. rioD^LS a^ ^^, ^^T E^c^^T^^

R. J. I.cake and J. G. Comiskey ^^


^` !
^Jepartmcn^t of Electrical Engineering
Unl.v p xsity of Notre Damn
`^^ Y7atre Dame, Indiana 46556
. ^ ^iBST%AC"^.' .
This paper describes an algorithm for obtaining a nonlinear analytical model of a
^ het engine from measurements a£ two equilibrium point values and the linearized A
:^ ^snd B matrices at those points. The method is compared cfith r=iots conventional proce•-
d^res of interconnecting individual component approximations.
rr.,
^.. YPFI'RDDUC'DYCN
' ^^ ^n L•his work we continue the study of nonlinear analytical models for a two spofll
,,.; turbofan jet engine foist reported in [1]. The model given in [l.j is refined here
and compared frith an entirely new model. which l.s the maim subject of • this paper. In
order to distinguish the various models, we make the following designations.
. ^ ^Sodel 3. -A large, flexible generalized engine simulator called DYNGr,N G.hich has
' been developed at. NASA Lewis Research Center [2,3^ and coded f^►7 a pazti-
-^ oular hypothetical t-^a spool turbofan engine.
f
"' Model 2. An analytically e:.•`pressed set of 5 nonliuear differential. equation, plus
about 20 nonlinear static equations agpraxirtating the re?atianship between
various engine variables.
^'^ ^fodel 3. A relatively simple two--input, ^f^,ve--state model ;vhich can be genQrated
_ :^ui'otraatically for any engine from data of two equilibrium paints plus A
- • ' and B matrices by the al.gor3th3n to be presented haloes.

,I. - e
^. DESCRIPTIaN OF MODEL 2
A refined and updated version of the two-input, five-state, two--output nonlinear
anal.ytica3. model. presented a.n [l^ is given in this section. The variab3.e designs-^
^ dons are:
U^ ^ fuel flow (T,TFB)
^,• Ux ^ nozzle area (A$) ^ .
Xx = compzessor rotor speed {NC)
I^2 ^ fan xotor speed {NF)
ICS ^ burner exit pressure {k 4 ) '
l^^ ^ after burner exit pressure {P7 ) .-.•
f

^^ ^ high inlet. energy {U^) .


^`^i YI m thrust (FG3 - r
,^., X^ ^ Iaa.gh turbine inlet temperature {T4)
_ ,,: ^'lxe sfstem is completely s p ecified as fol.l.ows. .'
^.
:^= '

__ _
..:
•'^
},
i Cn»st.^^its ;.^^^
^;,.

J ^_ .A.T A 778.26 ^ = 20.711,'5 _

t G ^ 32.1.74049 C,
l C
.24 ^ ^fN^ ^'AC^
^'Op$ QU ^]
R p I:A a .0252 ^ CI,F a .24 1^,r^
y^ 1.4 CAB a .,20279
.e
F2 = 518.668 , 22589
CPiIT
IC = PrilxP = 3.8 • 2793S
CPI.T
I F ^ Pr iIL r = 4 .5 BI'C
'- _ FCBLC = .16
^ z ;4^AC _

VCO;-IB ^ 1.65 u ^= PCL'LDU = .208
49.77
^AFRti = g = PCBLHP = .726
c^^rmOZ = .9494
y FCBLLF = .066
Cl^F S I' = 2116.217
' NC DESIGiT = X`i'1IPDS = 10070 -
NF DESIGN = lu^L,PDS = 9651

Des ^n Equilibrium Point (Sea Level Static)


WFB = 2.75 NC 11899,1 FG = 1343.02
A$ ^ 2.94$2558 NF = 9873.95 T^^ 2892.0'+
Fri = 23.9299
U4 586.46' '
i P^ = 2.55142

C\TF = 1.02310 P3 = 25.3522 ' WG50 = 134.364


i -121 = 742.957 P21 = 2.9960 WG4 = 118.375
CNC = , 98730 WF^S,AX = 203.123 WG55 = 135.818
T T3 1467.4'1 YCkil1X = 10.270 YF^fE"^7^; = 3.3 6 2 4
150 = 2103.47 WG7 C 224.323 Wl^F = 221.573

1 155 L 1789.15
ZC = .81430
WAC
2F
137.649
.8333
LIr*:^^ = 54.4151
AWriAX = 1.5 S 0 5
SrC = .737071 BYPLISS = .609694 WG?4 = 88.5047

1 State Equations

(1)
d'^C
d t
= ( 30 )
n
2 J [C WAr,(^^, -^T) + C. jaG50( T -T ) ]
zCNC PC ^1 3 ,'IiT 4 50
.2
30)
PFIti:AF(T 2-121 ) + CPLTI`1G55(T50-T55'J
(2) dttF _ ( I N [C
F F

f VfiY
[T4 tr'AS ^- ^v1FIl -WG4]

j.. (3) dt4 =


d P

(4) dt7 = V
COMB
i:•j =t i
7 [^:G4 -WFB - WA3]
.
. '^

AFI3N

I
v

`. ,

• dui` c^^^tr^t
(5) a [r {c^c 4 ^ i^^^Ii ^. t«t^} a- ^.^{T WA3 - T wc4 -I• T (^+r^)i,^i'i3}^
cIC ^ 4 3 4 ^
VC at i BI

Nan3.fncar fiunct^ons ^tequfrrcl I^}fit' ^tat^ Equations


^P
{1} CNP u ^Pr •
NFU^.5.Gt^f 9651
{2} T = T -f• 214.2732 C^l:a 2 — I^g(A -- 2.548255) •
21 Z g
N ^ .
C
{3] ChC — - C
^ICDESZGN ►^
7'2 10t^70r T21(5J.8.668
.2
21 + 743.2722 ChC
{4} T3 " T
— 68 (AB--...9j^8255}
^
^
^ ^
C5) T4 =
u 4^ C VB ^

{b ^ T 5Q• — .727 T
4

($^ P 21 = —6.205G8 + .0129774 T21 -- .QIS5376 P3

(g). W^^ + 261.01 CPtt' -- 63.196 ^ . .


FPiAX — 3.51&739 CNF -- .235&].
`^ ^gC} P
^ --2.313265 {P ^fE^ Y—P 21 )^
{3.^) WAF = W^^AX + 213.502 ^1—e ,,{ -
..;_ ^
I
WCifA^ W 137.54 — 4.57, 987 ChC + 5&4.325 CNC 2 -- 1B$.3.13 CISC3
^^- ^^^
a ^13^ 6.492 -- 4.9749 ChC -
°^''Cri^x '
CI~i^X ' 2&.43184 -- 89.o4s4 c..0 + 109.7243 cNC Z — 36.575& ChC3
^14 ^ P
F g^
1 --.3662 ( P CB - P ^
P
^ 27.
' {^.5^ WAC = i^GPiAx + ^WCr^x{^'—e 21
T21 j51S.66S ,

^3.b) idA3 = {1—^^idAC = .84 WAC


. (^fi^ WGSa = 301.957 P4J^

^^.$^ IvrG4 = SdG50 -- ^^[dAC = t4'G50 -- .3.163.6 [dAC

{1.9^ ^tiTG55 = WGSO + y^i^FAC A tdG50 •i- . Q105b WAC

.zo 45 s cHCfxz • . j ':


^20^
r55 ^ 106.002 — .s6154 T5^ — a x53.
, (21) T 7 = .49663. T55 + 205. S8b P 7 _ ^^ ',
.^
1121.78G ?? 7 A^ +r
{22) WG7 — .

^3
. .. ^
1
l.. • t

' i ^ '

_ ... .. -
;_ I.^.-... •^- ^• ♦ it ^ ^..
+ ^ 1 ! r w „ T .uv,M.^+^^"^`.'r_ .,..

.. .. _
-
1 ...- ... ... .^..
y

^:
- u
Nonlinear Equ:^tivns far OuC^uts '-

(1.) F'G ^ .02451 taG7 J.93 !^.4J.5 T7 a- G8558.3b5 + 2IlG.2l7 AS (,5397B ^^,--^.}

tz} ^^
^^ ^cr^x ^ ^'

rs^ !^•. Z1` a p'L^.


_
F
0^ 1 L ^^^ ^ ^ . .
.^1 F

rh ' 3. DSFECT 1^fHTHOD


t
We now consider a direct computer method for obtaining nonlinear mode7.s. I^et
(^^
..^,. x ^ f {x^u) -
t
Frith x an n vector and u an m vector denoting a dynamica3. system such as a jet
engine, in which the state variables and parameters u re gain positive throughout
^. the system operation and th^xe is a function g {u) such that £or each equilibrium
s
pa^.nt
,^
The steady state system analysis involves the study of the function g(u).
,i .
.g We propose to approximate the system {^.) by

max.
where A(x} is a square matrix Which varies as a function of x. Notice that if ^^,,
,^,. is an equilibrium paint of {1}, ^ = g{u D ), then a zinearization about this equa.^i-
bri.um point results in the linear sys*em

^,
`^^^ snd a linearization of the approximating system (2) at xg = g{r D) results in

r4; . . ^x = A(xD ) Sx + [-A{:^} ^u (u D ) ] dci - ^ {a)

^' ^ien.ce, the linearization of (2) wild. match the linearization of (l) if and only if
;• a
,^^^
^L7sa, if A O is invertible, as is often the case for jet engine models a equation (f)
.. ; ; yi.elds ' .

The basic idea of the proposed direct method is to use the above deve^.opments toge-
,;
Cher with function approximations far xl(:c) and g{u) to arrive at a nonlinear model.
Since equilibrium points and linearized models at those points can be obtained by
known algorithms, we shall use this fact. Our initial approach is to use just two
'.t ,; equilibrium points,.say :tp and xW . The input information is thus

^'-• wheze x^, and ^^ are design and off-design equi.libriu^n points. ^ and 1^ are she
,. system ^-matrices at these points and uD, D' gW' are the associate parameter
.,, and input matrices. Mathematically, uW' H
^: ..

^^^ 8x D ^u t^3
.^_^ =^ .
r':3.
.., ^ 1s
We eliall em 7.o a linear aPP
P Y roxii^atian £or A {
x ) ^ g i.ven b Y

1^(x) ^ ding —^-- -l- ^ diag (7.0^


x^^^ x^^- x W^ ^ .
^D^-
^I1 L•7 hich diag {•} is a diagonal matrix which causes the jth ' calumn of ^l(x) to be
interpolated linearly between the jth ca7.umns a£ ^ and ^ with x, as the interpolaw
t^.on variable. J

^'he parameter vector u is presumed to be utade up of physical control variables, and


para^aeters such as altitude and Mach nunber. The equilibrium function is to be
approximated in a manner such that both the equilibriusa values and the lineariza-
tions of the approlimating system (3} match these of system (1) at bath ^ and xW.
This requires then that
.^

and also

The method we pro p ose hers ::.s to appraxxm^te each scalar component g i (u} of g{u) by
a linear affine Hower law farm •

^^(u} ^ c aul +...•^- cmum -^-


m+l
c 2m+lul.
u2^^ .... ^ -d- c2m+2 {^.^}

for which the nth paxtial derivative is .^

agi u^^l... um^ ^ c

^u^ cj } a2m+l.orn'-j u^ ^ ^^-^^

^. •Norm ifr the variables are normalized and scaled such that
^^ rz^ ^ [7.,1,...,1} -- 7. uW - {a,a,...,a) -- a 1;1,5}

then, the conditions of (11) and (12} can bs put in the farm
2g

^ ^ •^
^g ..^
^c ^ • a:

.. k^'^ ^ aui {a) - c^ + o 2mE-l a m-1-j a' ^ (lei)


^
.•.4^

.. j^a^-1 ^ gi{x} = ^c^ -h c


zrrE-1 ^ ozm+z •
^^
i
^^:
k2m-^-2 ^ g.^ {a) = arc h + C 2m+1.a ^^ -f- c zArE^2 •
^xnd summing the first two of these over ^ yields

._ r
..
^cm^ -^.
''. ^k^^ ^ Eck + C
2m-E-la ^ ^ otn+^ '
., . 1;17?
^• - tx T r -^ n -E- r
.^ ^ _
• ,

^c ,^
e
k2Arh2 ^ a^c^ + c 7.m-h^. a ^^ -}- c2m+2
which is off' tha form
s^ ^ rl + x3r2
r —^.
2
^^ ^ xl ^. x3rza .
El8}
s^ Q r l -^- r 3 -h r^

,^, r
s^ ^ arl + x3 a ^ -^- r^
:.
which, incidentally, is the m = :l. condition also. This set of transcendental equa-
^d
Lions is solved nuzaerically fox x Z ,r 2 ,r 3 ,r^ and (16} as then used to solve for
each c^. In the event that (l8) has no solution, a best fi.t is made on the second
eguatiun by varying r^ whi.le the other conditions are satisfied exactly,

^"' ^. ALGORITHI^f OF TI•IE DIRECT riETHOD


r=• . In this sect^.on, we present an algorithm which sexves to automate the process ai
finding a nonlinear model for a system
^ ^ ^ (^'y
u} ^l)

^o be apgroxzmated from rD,uD,^^,u^^,k^,ED,A^^,D^^, by a naxYnaliaed system. The algo-


^,.
r3thm k*x.11 automatically gezform the normalization and, hence, actually appra^zmate
RT the s; a Lem
^;
^ L
A
r `

^,: where xi = ^rif^ ,


4.. A
l
^ ^ ^(^) L x^- ^( u ) ^
i u^ = 3 D
u ju
^
'T^-ie apgroximating s}*stem is of the form .

^^^
^, w n w w
. iwhere ^D "" xi xi ^ ^d
.. ^(x} _ ^ diag i- + .AD diag _
i ^4 }
,' ^f ^T^
A
^`Di ^i
E y

^^^

^^ ^ ^G^II^ ^- 7T u^^'d
^.^
c2IIr1-3 "^' e^m-t-^

wihexe u^ ^ a^u^ + ^^ . 9

'
. ^
" ' Algorithm I. .^

'- a
.^;Y 3. Input: ^,^,A^,$D,m^n^a^ c,X , ^,^' ,B ^ =`
d1 w, "Y ^+I w
.. '^ ^ r' • _^^

,^. '^. ^alculatet ^ t^

^,' ^ diag (l J. } ^L diag (x ) _ ^ • f


,.•. ^ ^i ^ Di •
x = • ^
^ +. diag(lf ^) ^ diag{xD)
,, rf a:
=^ ^ ^. ^

,^,,^ I3^ ^ diag(lf xA ) BD diag(uA ) .


^. i '
L, .
^. ,..
..
__^a_a...`._ _..._._^ ___-
L,,
•.
^ _
°,

-i

^^,F ^ diag{l1xD) diag {u^) ' _'


^j3
• ^' ^' ^R-IGF^IA^ AGE I^
^^ ^ ^^3F,LITY
3. Calculate
a^ a {l-a^ up f (uD -^^
i ^ ^ ^ ^ 1 } ...,n '

R^ ^ {au -u ) f Cu
r ilk W^ D^^uW^ - ^

'^ fit. Calculate:

^ ^ ^y w w . , ^. ^
i..^^ C—^T^^t.Tf i^
^^SIE7-L YY ^^D..[ Y W

5e Calculate ^9
i
^^ ..

- $^ r ^'2m•i-1 s4 ^ k^m^-2 ^
.j
- ' . Gw Co to Algorithm I7^.
Send: si, s^', si, si , a, s

_ 7. Cal.clalate:
i ^. i 1
^.
... ^ ^2m+1 ^ r3 e2m^-2 — r4 ^ ^
^..
ki —kid
:.
^^.^ --- c — k r^ c^
j J J '.
i r2-^

.- ' ' ^
$w. QuL'put
i 'i 3

.^.. ^,^ ^ ^`w^ . -


•^ (;
..
a
`' .^ 4^

-9..,
i
."^ f - }^ 9

-^`'.
3. input: si,s^,s^,s^:E,-a ^ ^ ^ `• ^ `^ j

• i ^. Calculate. - ^:^
s ^-54 ,;

,•.
. :^
• _ e ..

.. -

3. ^finimize by Rine search: `

'. 8x ,-- a--1


p 2 '" Px
ax--1
x _ a-1

- for-10<x<IOy x^0, x^1 '


.^ -
I^, Calculate:
P1 p^^^A^ P^^^ ^
x z x r ^ ,_:.
^:^ ^ 3 8x2-^ ' F^^ ^^1.^^ ^.
r2.. a- 1
{
x2 '

. c rl. 7r-a ^ ^ m {s1 - s 2 ^- r2r3 to ' ^' ^'


:L r2
s4 -- as 3 - r 3 (a -a) .

'^' ^ -.. r2 ^ 1-a ^ + p c

5. Return to Algorithm Z.6 ^ .:

f . ,^. NL^fERiCAL ^ESlT'LTS


^^;:
_i
The algorithm of the previous seci^.on was applied to data abtained using DY^GI;^ with
;^: -. ^ and u^ speczfi.ed as in Section 2. An off--design po^.nt was obtained us^g u j^ _
.`^' x,72727, .72727), with the resulting normalized state ^F = (.9000, .7897, ,7381, .
. _ - 09401, .9454). The normalized A and B matr^.ces are '
^!'•
{" ^• •-3,$ --2.277 2.067. -x.152 1.448 -.00259 .3553 •
_' 2.748 -5.39 1.5$5 •1.991 1.071 .2116 -.31&18
`f AA ^ 377.9 49.51. -264.9 8v _ 807 78.91 BW T 1.2.54 --13.774 (1)
31.26 139.39 -6,269 --88.69 27.83 --.6201 -99.3
' '' --176,5 23.91. --1.0.27 --37.4 -246.7 1.57.78 b.84
,^,...
u, . • - 4, 744 -1..38138 3.2468 -1.4591 1.1969 --.04546 .0013
_ - ,82186 --26.725 2.5585 -1.8609 .45548 .0086 -.0121
^^ 475.73 137.55 -328.91 27.791 97..495 Bw = 2.434 -.613 {2}
,^ -50.103 17.0.91 63.188 --116.69 8.2853 .67865 --97.467
--156.77
}. --67.682 --41,681 24.58& --243.23 203.44 .&4755

•. - ETsing the,. ..paramEter value a = .7, the c^ coefficient which specify the equilibrium
:: function g {u) as i±^ Section 4„are given by the matri^c ,f

^• e24267 --.002113 1.90082 8.0991& .02864 .73088 --"


^^ s,
^^^«' 1.01593 .85407 .89872 .6&97,9 -.81$79 -.05121
`"^^ C ^ .73445 .10133 6.905S6 3.09409 .011495 .15272 (3) ^E
`^'^+:^'^ .77234 -.35905 2. ^i g 867 2.87415 -D75198 .6&191 ~5
4:^:^ .39503 .27262 -3.44682 13.4468 .01838 .85921 ^-

^', phis matrix together with the values a = 1.1 and $ ^ 0.1 and the matrices A^ and
W ::', uanzp7.eteJ.y specify Made1 3A. :. .^
"' Another model which we wi11 call Model 3B is easily obtained by using a linear '
^' sffine appraxi^natinn to ^{u; such that g(i^) = ^D , g(uw ) = ^c^. Afodel 3B is specified
r.
;c .

. .. .. ^:

__ ',^. ...Y_.
3 h;^ ^a ^ ew9 , a m- 3. 1778, ^ ^ --^,.3^.778 and C^tc ^a^^f^yci.en.t ^:a^x^.^
^' a 1.553 , Oa2.i3 ^.. ^ ^.. ^ 0. ^ .841.8
i .. ^^
.1G19
.5351
.17 d7
—.1208
7..d
1.f3
^.r1
1..Q
U.
d.
Q667;; ^
.585'

'(^^
.5878 —, 49313 1.. d ^.. €^ 0. a 9053
.2962 --.2d99 1..^f ^.p d. .4^.3^

- In o; der to m compare the four models, a test point x T far^remo^red from x 4^ was-chosen
^ by setting uT ^ (.S,I), and calcu^.a^ing the equilibrium { A step change to u =
^ 1 s 1^ then causes an accelexatian transient back to ^. ^he results of this cotn•-
'^ parison far the rotor spee^js are ^howa in `Fable ^., Mare detailed infa^atz-an a.s '
shown £ar ^ludels ^. and 2 in Figure 1.
. Mode:L ^. ^od^i ^ ^^pdel 3A Mode1 3B
`^ ^ N NF Ne ^ Ne NF H^ NF
^ ^
0 .918 .929 .934 ,903 ,^36 .940 .92$ .925 ..
'^, R^ .. 9 6 n8 . 94 6 . gSS s ^ql^^ 7 85 ,955 .954 .945
^ S nn [ 1'1 n i^
1 G, RJ; . 9 9V. . 7-t 7 ^a LEI ^3 7 g^^ a^&^ .9T8 .983 .9.T5 •^
^ •' 43. • ^98 a 993 y
Ya Qdl} , ^^^ ,^9f .991 ,995 .990 ^',
- `• 7 z. odd . 9 98 ^,. 0 03 ^. ao^ 998 997 .998 .997
^ r ^i #

• ^ o^ p^x^^Qa egg ^^ ^^u^.i.x^i^^^tn and Acceleration Trans^i.ents ^ ^^';_ ^


-
^e 1.a

^. ^^„^^r^L FADE ^ : • ^:4 ;


... ' ^^ ^^^@^TSSIOI^ ^^ ^{ ^U^LI'TY _ ;± ^

`^^• +^A ^it^merip^^. algnri^hm fir ^bta^n^.p.g ^ar^^^^ea^ analytical models for h et engines is :^,.
1^^ •. p xesente^^ The method ^,^ to Sepa^a^e ^}^^ ^r^n^i.^n.i. A{x} and equil?brium g{u} Parts

^f the system dynax^ies aid t3ppro^^.^a^e those ^as^.ng easily accessible data, The cam-- ',
Parients of g{^,} are agpra:^^t^aCea by a linear affi.ne power law tor-r^e. The principal
numerical di.fficulry is that a1z bpundar-y earcditions may be impossible to meet. The ^'^^
^' algorithm then satisfies ^.^^, but the geeQnd of the equations {IS} in Section 3 and ,
s' fats the second as closely 3S p QSSible, ^^ie vari able y i.s zero when an exact fit ^_,
^cCUrs; but, othercaise, iii ^^.uses a ^eas^ squares fir on the derivative conditions at x.^ ^^
;;:.. ^ equations {l6}, Secti o^ 3, wh^.le mafce^ng the other conditions exactly. The 'i
;. free parameter :in the lineax ^.ff3ne ^pprox^.^at^on of Model 3B art handled similarly.

c: •^.
' ,,.,This research was supported by the ^Ia^^o^al Aexonaut^cs and Space Adinxnzstration
` ^: under Grant NASA I^SG-3048. :: ^$ =^
^^ ^ • ^1.,

•:,^T $: REFERENCES ,i. ^;


^^ 1^ W. E. Langenbaker, R. J. Leake, ^ r H^.exarchy of Simulation Models far a Turbofan z
^;2! Gas Engine," _Prot. 8th ^!nnual ^ittsbur^h^ Conference on Modeling and Simulation,
3 1^pri1, ? 977 ..

^' [2^` J. F. Sellers, C. J. Aan3.e^.e, '^T?YNGF.r1 - A Program. far Calculating Steady State - }'
^5^ and Transient Pcrfaz-manc^ o^ ^^rbpje^ and Tt^rbaf:an Engines, r` NASA TN D--7902, ,
,,^ Anr^.1, 1.975. .^ -`
_ •-
^
^ 4 ..,
,''^' ^ 3^ L Ge y se r, 1Q ll ^GE1.^3C D -^ A ^r^ag^a^a ^Qr ^,^,^ct^la^^ng LTnear A,B,C, and D Diatrices
from a I^onlittear Dynamic R^fi?^?^ ^?^.^^a^P^'a xti ^'ech. Report in Prenaratian at .
NAS11 I,ecais ;°e^earch G^ttta^ 3 ^I^^^^^^da 4^k^^. ti.,
^^i
,,0 j^^ W, E. Langenbakcr, "T^.mo ^^^^^^^^ ^^^^^^^^ Q^ ^ Ewa-- Spaa1. Turbofan Jet Engine, r'
.,;^i; ^^^E Tliesl.s, ^^^:vcrs^,ty ^^ ^a^^^ ^^^^,^ ^^ 4 ^y^^as, September, 1977. '

,^._..__ —_^.. _—_ ._— _ _^ ,___._.__s_ _. ^ __ ^_ _ _v


.:,
.,.^
.. ._. :.; _ _
,. ;,.. ,.

€'^ G.,.,-.-d. 6^--^^ C---J^ cam.. ^—.—:-; ^"_-=^ ¢.-^ tr^=-^a^7 ^=--^ ^ »-^:_) C..;:_^ E". _...^.7
^•----

^I
^r^.^•
} r . `r r r • •x4.23 ^lr•^

• _ to^^ N^ a 974"RP^i • . ^.^'^'^- ^¢a,^ ^ ,^


Ida • ^^^^ ^^ ^S(}e ^•
/^ + 0^
• ^ ^V^^^ j - ^ v^^ • '^^0.^ ^ ^^ rl

^ ^ ^ $^'C.^ f
R^
^'WFt3

• t=a,1 [^^•
^r
^ . - r ^ -^^ a.z s^ .^`e
?8 ^ ^^ ^`^ wFS f '

q^ r ,d `'^w^e ^

• . t
2^
^
B p
46 ^ / 2 .^, ^ ^ ^s^. ^
^. ^ .._...... _..

! ^ ^'
.s
4s ^'^ ^Yo.^
4 !^ 2^^^^ t^s
{.^ ^ - ' ^^ .^
• ^ ^^ =®,s
^+^^d ,^RRoWS "^"RA^SI^HT^
StGt^iF1' - o ^,^
^ ^
• ^4 ^
^ , ^^.^ ^• ^ N^. A^^ow^ s ^^ ^ ^FY ^F^'R^7t^lC^ ^^KE^ :^ `°^a
W F S ^^ ^ •^ ^ l^ ''^,

. ^ ', l^
A^pei^dix H

:'SOME FEfi1'iiYtES OF CA1t^YAD PLOTS

FOR SYSTEM DOMI^IAPICE"'

!t. M, Schafer
M. K. Sain

i^.
°R^^ ^ ^vA,,crY _^.. .. __..._ _ _.. ...._

i !^^ , •

.. ... _.. .-

SO`2F. 1'GITUKES Qt G1Fi11IASJ F[.OTS }'Ufi :1'S1E31 DO`:IZ;.ISCE^


ai t.^' ..i. S..^i.•k) dl • ,l ^

_ _. _.... _____._..^.^^ --_...— _... 1


cl.f:_ 47.., :r ^^^• .

^
I

^ _ -_.__
P,. Fitchacl SchZfrr _ •._^-- Hichacl F:. Sain
i
Dr;+artc^• nt of lacctrical EnginecrinE Drpnrtrtent cf la cctrScll Enginccring
Univc • rsi[ •a oC l:^tcc Da^•a Unlvcrsit y n^ Kutrc Dame .__ . __._^__
1{otrc i:aa:c, Indiana GG556 _ _ Notre Aae^:r, '.loon 4056_._ _.,,
I f
t
• • _ ^; 1
_ ^ ^._.__ _I__..— __._.^
,;u::r^_t^^tci e: `^ • :^n .:^ r. •r ^, ,r^ e,•t i:^^. {^ _ • on^y.
f 2... ^ E ^ •r :rx!
t

Abstract ^ 2. Specific Assumptions


1

P.eccntly, the CAFDIAD (Cot^plex Acceptability Pant modc]s used to construct Che plcrs in
r egion for DiAgon:^I D oaiirancc) plot i7as here in- the ::cc,ucl hs y e been Generated frost the sertcral
tlOdl_'CC j and applied to the pro'ulem of desiFninb purpn::c digit.tl jet tnPinc simulator DYSGE:: [6]
dynattieal precoraaensatio q to achieve colu-:n doer under a load t,hich gravid^s behavior similar to
fr.ance. This pcper i'_lustrates ..^weral basic that of the F-100 two-spool turbofan er:gine at sea
features of tl7e rrthnd while using it to desi^,n a level static cotditions. The c:odcls have two in-
si r^.]c, lot:-ord•^r dynamical compensator which a- puts, five status, :.ud two outputs, Thcr are lin-
chieves doxinancc at fire operating points of a earizations of DYXf:f:^ abtained with the aid of the
realistic tt:o-spool turbofan digital simulation. DYGAI:CIt package [ 7 ] under development at :;.ASA Lewis
....^ Rcse^ICh Cen[zr. FhVSic^1 description of the
states.caa be found in the references [3). The in-
1. Introduction puts are fuel flow and crhae:st area; the outputs
are thrust and high turbine inlet [e7aperature.
the CARIIIAU (Coaplex Acccncr.bili *_y Region for Paranet.^rication is acce^ap[ishrd tltsnt:Eh tae no.-.,-
DiAFc.^.al Dominance) plot in a graphical technique inal value of the. fuel flora °dF3, which takes the
for choosing dynamical cocpensatnrs to achieve five values 2.145, 2.31, 2.475, 2. G4, end x.75
diagonal rcw or column dc-^inancc, as do*inr_^. by Lh`i/`;EC, raniue from a ]ow thrust condition [o
Roscnbrock [1]. itiithout esGential lots of Eeneral- hi Eh thrust without aug::entation. All the aK+c!cls
i[y, the coc:;,ar.s3tor is assumed to have its di- have been normalized.
agonal elements equal to unity, and a typical
CAhD1A°J plot describes the acceptable range of tae Thus the plan[ transfer furictinn tr:.^trix has
re:l and icaginary parts of the off-diagonal el- tc;o rows and two colu^: s, and ex^tibits tra:,sfcr
emen*_s such that doninaace is achieved. 71te basic functions of degree five in bath nur,cracor and
graphical building block is the circle. F.ach denominator. Space limitations preclude their
circle represents the acceptable range at a spec- presence in this manuscript.
ific fzeq:]rncy. Solid circles are dra^nt if ac-
cEptable real and it:aginary pairs correspond tc Denote t}te plan[ by G(s). ,-ttcn the issue is
paints inside the circle, and dashed circles arc. to select a precortpcnsator E; (s) in such a s.•.Iy tear
drawn if acceptable pairs correspond to ^^ints act- C(s) ]i(s) is colu:tn dc;;inant (1J. In particular,
side the circle. Plotted as a function of fre- it is desired to select on_e K(s) so that colur.::I
quency, these circles describe the acceptable range do:ain+n^c is r.:.-tintainrd over alI five nominal fuel
of the eomnensat^r eleacnt in question, considered f]ow conditions.
as a cocplcx fuuctior, of frequency.

Recently, CARDIAD p]o[s have been shown to be 3. General C.\FJ)IAD Fca[^res


an effective desiy_a tool in dyn^::ical preeom-
pensation of cwl[ivariahlc punts to acl7leve docr- If the. origin of the G1RnI,1>7 p]ot for a given
inance [2, 3, 4, 5]. ''sots paper focuses upon an- colur..tt is inclu^'.cd l,y all solid circles and e^•-
other aspect cf the C.1CDIi,D plot, n7-cly its cluded by all dashed circles, that Column oC the
abSlity to assts[ with chc• classification of var- systcct is do ^inan[ without further cosrpensation,
ious opera[in^ paints of a nonlinear s y stem with in as much as the ori[;in represents unity cot::-
re£ard [o their drainance possibilities and to help pcnsaiion. iltus, the eventuZl g,aal of cnrper.satiott
with the desif,o of eampcnsatars which achieve dom- using the C:\r !IIAD plot method is tr arrive at a
inance of n ultip le opc:atSng points. system where all Lhc CA1'.11 G1D plats have this
feature. If there exists a point on [he rtal axis
]This rk wns
was i
pported in part by [he 5atinnal
:o such that [hr pcint i^. inclosed h^• all solid
Acron:rti_^cs and Space Adafitistration under Grant circles and c::cluded by all dashed circle:, in site
l:SC 30:S and in part by the national Science CAF.U]i^D plot fee a y,iven column, then the choice
Foundation under Grant ENG 75 - 22322. of the value of this point in [he off-diagonal
.. ^ ...

.•
1

. .

^._

..u•E.il %' ^ .. ..t:::


PAGE ^
OgIGiN AL.
JE ^R QU p,L1T `j
sntry wltieh the C:\5DL1 ;]
plat ropresents will t:^kc nominal vsluc of the fuel flow Increases. Since
the [olt.^nt dominant at all trequencies. If there there is more. mat• }:in for crrctr in the lower Hour
is ao such polo[ the CA:iUlAi1 plot dcscribcv the anal value of lu^l fl.• . r.:odels, a cog..,^.cnsator
ranFu of a Crryuency dopen.l.^nt off-diagonal entry cLich 1:. fit to a tou^tt avcrat:c of the five plots
attach viii make the cclur.,n do^tInant. and t.•h1cL tends to be c:oscr to the higher noain.tl
value of fuel fl.v models, might achieve doainance
Ct.F.^L\D p]nts fee t^•o input, two output sys- in all five models. _ '- "'-"'-' •
tems have sat»c interesting featurrs. A circle at
a specific frequency in the G1:.DIAS plat for one The avera g e value of thr eentrr of the lowest
colu:,.zt will be solid if an.i only 13 [hc c^hcr frequency circle of the five plots is -9.51. 11[15
eoluan is dominant a: that frenueney. Thus, when su£,Fcsts that des:Saing a cot--,cnsator to fit tl:e
i syscea is doairant at all frequencies, ail the nominal fuel floe of 2.75 codel tzaich has as the
clreles in the CAttLIaD plots g ill be solid and alI center of the lowest frequency circle the value of
will contain the origin. .mother interesting fea- -9.59 etight achieve doainance it. all of the _tod-
4urc is the etfert of a color.:n switch, tt:at Ss. els. The second order funtion that vas chosen is
swltirlieaticn by a m^*_ri° with roe only non zero
cletrcnts bcin^ ones or, the oiI-dia , ;onal. The - . 74?s -• 4.59
effects of such a witching of the in,uts are that .014s T- :99Ss + 1.
ell the sol±d circles becoae dashed, all. the dashed
circles becor.^e so'_id, and the shapes of the color.^ and t}^e next compersator, K2 (s), is
one and two plots arc witched. •]-!tis fact e;ill he
1 -.742s - 9.59

r escd in the next section tc achieve doainance in


the various sc[ point ttodels. K(s) -
0
.0)4s2 -.998s +1
1

4. Design Example
Titus, the overall compensation is K(s) given by
The CAF.DIAD plots of the five uncompensated
0 1
uodeis are all very 'si,^ilar in shape. This great
t¢itaflarlt^ sus_rests that one cor.aensator night be
K(s)

1 found that ti• ill cake all of the redels column dom-
inant. The urcoapensated plots also show ..:at a
colour[ switch would cake the first colors[ of each
of the models dominant at all frequencies without
1
-.742s - 9.59

.014s 7 -,998s T 1.
^
t

Figures 11 - 20 are the CrVtDIAD plats of

1 furtt:et compensation. Thus, K 1 was chosen ,o be

0
>t
1
o

G(s)K(s} for Che five c.odels. It is clear that
they are all dominant at all frequencies since
all of the circles are solid and all include the
origin. Thus, one compensator has been found

T FiE • tres 1 •- 10 ate the CP.RDIAD plots of G(s)F.^ for


the five aodels. The repetition of the general
6itapcs of the plots, which 1s unaffected by the
eol.umn switch, is very apparr_nt. T::e plot. also
which will t:•tl:e all Live of fire models considered
is this paper dominant. `

I short that the first column of each of the models


1s now dominant. I1tis can be ascertained either
by the fact that the oriSins of Che colutut oac
plots are inclosed by all solid circles and ex-
t. Conclusions

Through the use of CARDIAD plots, it has been


possible to achieve doainance over a range of
cluded by all dashed circles or by noting that all operating points of a jet engine simulation. The
of the circles in the coiner: two plots a:;e solid. eor;pensator given above also achieves dominant^_ at
all but a very narrow range of frequencies in Lire
Te ac}tieve doainance in t}te second colttr..ns of model of another operating point. The results
the models, it is clear that cone sort cf frequency stSEcst two things. rust, using the CAFZDIAD
depenten[ corapcnsation will be necessary be.ause pots as a guide, it could be possible to design
there exist no poir.rs nn the real a::es of the plots a compensator which varies with the nominal value
which lie inside all of the solid circles. A first of rite fuel flow and achieves global dominance
Choice of a function to fit the paths of the cir- over a aide range of operating points. Tats is
cles could be a simple first order Function which eurrent]y heinl, studied. Second, the repetitive
traces a seaicircle through the cocplex plane as shape of the CARDIdD plots eves the range of
the frequency varies. Houcver, it- is desired thaC operating points sug;,es[s that the CA.RDIAD plot
One such tuuctien be found [hat will work on alI might be a useful tool in the classification of
five of the models; so, a second order compensator operatin;; points with regard to inierar.tion. Such
fill be used to fit better the shape of tt:e circles a feature could be quite helpful in analysis of
at the hihher frequencies. I\:o things that should which models to use over flight envelopes varyin;;
he Hated about the slt^pes of Lhe: circles in the from sea level to hiEh altitude and frotu low
colutm two plots arc Chat the circles tend [o be through high thrust. ^r
larger for the lower values of fuel floe and that
1n Eeneral, the center of tha lowest frequency
(larsest) circles roves toward :he origin as the

^t .. . Tr..,., ., .
e .f1 1 a .a ..1..

pgIGINA^ ^ F `°`GE ^
+^ -
^ ^^ ^R QUALITY

^^ ^ ttttry .hick the C_1Ci.^IAl3 plot represent. will Hake nominal valrrc of the fuel flaw increases, Since
the calumtr dominant at ail Erequcneits. It there there is more m7rFirt for error itt rho lower nnt^-
3s no such point the Cd2llIA p plot describes tire inal value of fuel fla:^ modals, a compensator
ron};c of a Cruquettcy- dependent af[-diagonal entry which is fit Ca a rough average of the five plots
:^ which viii make the column dominant. and WhiC}t tend . to he closer to Lhc higher nominal
value of fuel flow models, mig}rt achieve dominance
CAFd}IAD p1oCS for two input, two output sys- in ail five models. -" '^ " "^"' '
toms have same interesting features. ^ circle at
a specific frequency in the CA^I,1t3 p1aC far one The average value of Che canter aF Che lowest
ealarsn viii be solid if an.3 only i£ the cHrcr Frequency circle of the five plots is -9.SI. leis
eol^tt is dominant at; that frequency. Thus, vhan - suggests that cirsig:ti.ng a compensator to fit rho
a system is dominant at all frequencies, a!I the aarai.nal fuel £low of 2.75 model which has as t}te
circles in the CAhI3Iall plots will he-solid and all canter of the Ios;est frequency circle the value of
will contain the origin. .mother interesting fea- --9.54 might achieve dominance in a1I of t}te >:sad-
ture is the effect of a colurvt s:ritch, that is, els. The second order funtion that ttas^chosen is
multiplication by a matrix faith the only non zero
elements being ones an the off-dia6onal, The -.742s - 4.59
effects of sue$ a switching of the inputs are that .dl4s - .99Ss + 1.
ail the solid circles become dashed, all the dashed
s
circles become sitlid, and the shapes of the calum;t and the next compensator, }:^(s}, is
one and two plots are switched. ?}tic fact will be
,^- aced in the next section to achieve dominance in 3 ... -.742s '- 9.59
Lhe various set point models. IC(s) - .Ql.4sz -=.998s +T
d 1
^ ^ 4. Design Example
c Thus, the overall cot^pensatian is K(s) given by
The CARIII^LIT plots of the five uncompensated a
stodels are all very 'similar in shape. This great I
aimilarit;a suggests that one compensator raigErt be _.
'" ^' found that v:ill male all qf Che trodel.s calu:rn dam- iC(s) ^ -.742s - 9_.54 i
;^ 3nant. The uncompensated plats also show tarot a 3 7_
column switch wogld t-iaake the first column of each •Q14s --.998s •i- I. i
of the models do;rrinant at all frequencies without
further cnr^pensation. Thus, IE 1 vas choseTi ^o be Figures Il -- ?d are the CAI{i)I.lLU plats of
G(s)K{s) for the five models. It is clear that_
0 1 they are aII dominant at all frequencies since
-1 all .of the circles . axe solid and all include the
Q
^ _ - Origin. Thus, one compensator has been found ^
^` Figvres 1 - 10 are the CS.RDIAD plots of G(s)t`t foz ^a•hich will make-all live of the modelsconsidered -
the five models. The repetition of the general in this paper dominant. t
_shapes of the .plots, w}tich is unaffected by the _ 1 ^'
c®ltnart switch, is very apparent_ T} ;e plo Cs also ^, Conclusions
6hov Lhat the first colu+an of each of the modals -
^,A is now dominant. T7tis can be ascertained either
#ay the fact t}rat the origins of the column one Through the use of CAfiJ]IAD glots, it has beena
- plots are included by all solid circles and ex- possible to achieve dominance aver a range of s"
eluded by all dashed circles or by noting that all. operating paints of a jet engine simulation:. The
a# the circles in. the column two plots are solid. compensator given above also achieves dominance at.
aII but a very narrow xange of frequencies in the '
To achieve dominance in the second col llten5 of model of another operating point,. The results i
sltggesC two t}rings. xirst, using the C1iRDIAD a
the models, it is clear thaC some:sorc . of frequency
depeatenG campensaeioII gi5.l be necessary because Plots as a guide, it could be possible to design ^
there exist Fla points on the rea], a:;es of Che plats a compensator which varies kith the nominal value
- vtiich lie inside all of the solid"circles. A fires. of the fuel flow and achieves global dominance .
choice of a function to fit the paths of the cir- aver a wide range of• operating points. This is
rles could be a simple fixer order function which, currently being sttdied. Second, the repetikive
.traces a semicircle through t}te cauinlex plane as s}tape of the GARUTAn plots over the range of
the fre q ue nC v operating paints suggests Chat the C^1PDlAD plot
` y aries. However ' it 's a. de s fired t ha ^
..one such function be found that will w qrk on all might be a useful tool in the classification of
.^ Sive of the models; sa, a second order compensator operating points with regard to interaction, -Such
will be used to Sit Letter .the shape of the circles... a feature could be .quits helpEul,in analysis of
a t the hig her frequencies: lt:o things that should v}rich models to use aver flig}tt envelopes varying
from sea level to°High - altitude and frasu low 3
be Hated about the shapes of t}se circles in rite
column two plats are that Che circles tend to be through high thrust."- ^^
larger for'the lover values of fuel flow and teat _.
- 5
in general, the center of the lowest frequency -- ,,
_ (largesC} circles moves toward the origin as the ^: ^.... '.
^,^ r -. . • • .....r .... .rte—^. .—.• ^. .. -- : r.... • -- — ^^' ^
y

e. •..

' ^ ,

_r.---_
..^u.r+[wr-L..v...- ...r+:^•^.^...-. .x.+_^..__-._.^.._--' - -

^ r f I:Jr\l ...':: ... '

._ ._^__ -___ _ .. _ _ i
y.`+•.
Acknn>wledgment r
-
—^
T}^e nuthor3 thank R. R. Gcj!i for hi9 assi[- - --• ! ^ I
ante in obtaining the linear c.^dals used for these ^ ^
plots.
t
. ,_
-- - ---- - - - -- ._... :.: _.

ReCerenccs ` ^
s `ter\ i

1. H, H, fiosenbrock, Cnc+:+_u t e_r Aid e d _Con tro l Svs- ^^ ^'~ \.
----.:. 1^
tea Dcsi^n. London: Academic Press. ].974. _— ^ _ ^.r ^
qT.i
\ j
— +
Y, - I
^^ 2. R. Mich :el Schafer, "A Graphical Approach to ^ ^^ ^ ^slI^^"+^•• , I 1
5y{tem Dominance," Technical Report \o. 778, - — •' ,f^rr;t4^a^ j
Dci^srtrrcnt of Electrical cnrineering„ Uni- • - ^ ^ .`' , ^ ^ ^, j
versify of Sotre Dar.:e, Piarch 24, 1977. = ^^='^^^'J ^ Inc. m II

3. Y.. M. Schafer, R. R. c:ejji, F. N. Hoppner,


i^'. F.. Longenbakcr, and `L }:. Sain, "Ftcqucncy
^^`^
! ^^`"-""^
^
- ^
Domain Co^pcnsation of a P:'\GE': Turbofan Engine . -. . ^—•^.K - .. •... . ^^ .,M ^ ,. ._. ^. ^.Y .+.:/
rSodc• 2," Pre p rints 197? .faint Aueo^liic Control -
Conference, Voluc:e 11:0, pp. 1013-1015. FiE. 2, S.'FB=2,145, G(s)K l . Coluc:n 2.

d. R. Gejji, R. M. Schafer, H. I:. Sain, and P. !


Hoppner, "A Comparison of Frequency Do.:.sn ^ ___ i
Tcchniqucs for.3^t Engine Control S} • stew c

Design," Proc. 20th ",idlest Symposiva on ^ ^ '

._
Circuits and Systems, Part 2, pp. 650-485,
Ass ;r.st 1977. ^

^^^ i ^

^^ ^
S. D. N. Schafer and "f. K. Sain, "A Dynamical
,l}- , zoach Lo Diigonal Uosinance, a,th Illustra- - ^ ^^^ ^',» ; ` 1
[ions fora TurLofasz Engine Ltodcl," Proc. 1977 =i ^fl! - 'r• '^.:^ 4 . - •t^ y i
._„ ZiEC International Forum on A.ternatives :or ^• "` }C^ ++^r J r
Linear lfultivariablc Control, October 1977. ^y :.`^ "• // i
^f
y.^; r i
6. J, P. Sellers and C. J. Daniele, "DY:CGEN - A ^ ',/i j
Program for calculatinr_, Steady State and Tran- s
scent Pcrforr^ncc of Turbojet and Turbofan ' ^ -^ ^
£nEfncs," }CASH TCi t}-7901, April 1975. ^ ^%' ^

. 7. L. Geyser, "DYGABCD - A Program for Calculating = ^ .


Linear A, B, C, and D Pfatriees from a tior_li.near . w —^ ;-: ,w ,.- ,.,, ,.w ,.w
.ar
Dynamic Engine Simul: tor," private CoLnunica- ^`^
..on. Fig,. 3. L1FB^2.°1, C(s)K 1 , Column 1.

^ i ^,` ^

' ' ^ .1==•^= / ^^^^ / Inc. r,s


t ^.7 = I ^ / ^^^
^
^ ' '

_I
t ^^N ^^^,! ^c ^ ♦ . ^ ^.w ♦ f '•YM - w .w .tY /,.1 •w a.w ..w ara .,.ry Y

^'

Ftg. 1. {1F'B=2.145, C(s)K 1 , Column 1. Fig. 4. L1FB-2.31, G{s)K^, Column 2.

- . -: ^ Y ^^
.• ^ ^ • ^ ^^^PWR QUAL^^^ ii
---- - .._ .-_-_...____-__.._._ --- . _ ... ---- ----- - ----- --- - 'ii

.•
r....-._ -- -
U^ . ^ ^xu^^^^''^''.^z'
---___-^^'~___''---'..-^-----_-..
^^^ ^^ " ^'
'
!
-~ |
/
' '« /.
- ^
'- |
.
. |
' .
.
''-''
|
^ ^`/ /
^
'''--- ^
.. ^ /
.
` |
^
i /
/ ^.
^ ^ .
.~
*un. »' p^n~z ' ^7^, o(y )r,. cvzv^n z. r^u' u ' n^n~z.^^, c(x}n~, co^^om x'
N ' ^ ' ^
''^^
^^ ^N ^ ' i

^ | |
U^ ^
^ ' | '
^
^
U
^
'
'
^
^ ^ ^
^ ^

. ^
.
^ ^ -~^^^=~-
^
^ .
^^ .
- . ^
' ^
@^ ^c^' o ' n^u~z ' 47^, c(»)u,, n^z ""m z, s«o. p , n^^~z.7s, c(s>x ^ noxu"^ z'
, ^ z
. ^

^
" .
^ .
'^ '
^
^^ .

~ ^ ~~ ~~ ~. ^~ ^^~^: .~ ^~ ~~ .~ ..
^
^^o ' r, ^^m~^.a4 ^ o(s>x, ' ovzu^n z' , _ z^o, ^ u '_ n^^~z.7s^ n(v)x `" ^oz""m_z.. . ~_ _
. ^ ^ __ _' --__- __^ __-__-_--__-_--^_--__''-_-_--^--_---_-
X ~° ~
^ . '
^' ' `''^---'^`. ' ^ ' -,^ ^ .', ^&G^ ^^
, ^
r^^^^
_ _ ' _ _ _ ____________ _~_______ ______________________________^^ _^

'
^ N^
^ ^^
~ __-____-- '
-~r

^ Y '. ^r^`^^^,r^^^
^ . [x^*\^^`^^^^^
.
^ ^' ^ _ . _- __-____-_'_---- -_'-'__-
^ .^

- -^//^^ / \ ^ .>
' -'' . |
^ - ^ /
-''

^^ '/ ^_-_-
/. ^
^ ___ |
w^ +^, i
|
.
` '
^ ^
^^
°^ ~~ ^~ ^~ ,:"^u^ ^~ .. " .~ ._
° ^ ---'
'..- - \/
^^x. zz. no^~x ' z^s. n(")x^^). c " ^"^^ z. ^^a. ^4. nro~u ' Ja. o(v^o(;), cuzuo^ z.
^ '
°^ ^ ' (.

'
N .|
^
°* '
.
^
^ .
Nv
^
^
,

^^ ^
^^ . .^
. . .

^
^ . @c '
^^ .
.' "
^^ ^^ ^^
^^°.. "~ .~ ^_ ^~ -^^^^ ~. .. '~ ^~ .~ ^
^ ^
^^ v^n ' zz ' wru~z.z4s, o(")x(">, coz"=, z ' n^o. z» ' n^^~z.^7s, n(s^u(s}, oozno^ ^'
i '^ _
/
' '

.
^^
^ ^
U '
o '
'
.

^
^
' |
i

'

^ ~~_. ~~ ~~ `. ~. .^^^° ^. .- ~ ~ '.


. .
^^ p^^ ^s. n^o~z ' o^, o(o)u(s> " co»" pn ^. ^ ^^o^ z^' n^o~e'47s,'c(s>x<"^' c^^o=^ ^^ . ~- -
, ' ^ ^ .-- -- _-- -__-_'_- '__-_-_----_--___^^-_^_--'_----_--_-__- ^. --_
^ - ~
^m^
` ^ ^^ ^ . '^ ;'^ ' ' '
'^U !^ r^ ^^^^FTFo {J^I{^^b^^ ^A/38 %^
^ _-' ''^ ^^^^^^D^^^^
- -'-- '-------- - ' ' --- ------------''-------'--'---'------ -'-----
X^
^
_
^ ,
'
{^

,^ ^.I ^1 :. . ':'1'c':.'i I'.1i^k:s It'^'i^
^
_. Y .. ... -a
J


a. • i.it `...:..^ rt i.i(! dl^a ^rw.i6t:1^i1 J ;: ^:S tip:r ^__ _, _ . ^ c t.11[ s«u:a! ; :^ulnf} SS^t1tt.E ;kid Su^C^]Cklill) tl.. ; _^ !1^ !e
1
r
i i _..^r.._.^_ ._....,»...-
^14it% U.v t-• '_ ^ ...^.^^it liw, G^f^7^!' _fl^ — I 1
:- ------. ., .. ----- i -- ------
^

^^ * ^ ^ r ' ^ J' n '

• ^ ^^`\ ^ e ^I i

^ - ^^ . ^
^ l ^^^: .
T ^ - •^ ! ` -

.
z

^iR ice '' tiCc^rFlt

. ^

Fig. 17. KFS = 2.64, G(s)EL(s), Colu^ •.n 1. ^ Fig. 14. hT f'B = 2.75, G(s)K(s),• Coluut 1.

. - i

r--- .

• ^r ^^ .^\ ,
^^
^/ ^_ \.
!. `,} ^
..
^=
^- ^r^ ^
^ „•:,^
_

4
\^^ ` ^
r • ^-^!!
\ ^^^ /

.e i

r
3 r. _ _.. _ _^..._ . ... ^ . ^[; r^ s ^
. l 1 - - _

• t J 3

.^ Fig. I$. 47FE3;2.bu. G(s)K(s}, Calamn 2. Pig. 20. 1aFB a 2.75, G(s)K(s), Column 2.

M ^ ^ ^ ^RIGINAi, P^'^(^^, ^$ ^

- .^.—..-. ... _ .^^r..^^_ ^—


^ ^__^
r ..
1 . _ , .. :Ir ! ^ — ^, ^ ^

{
APpenciix x

THE THII^IE P^^^31.^:^i

2^1, 1Z, Sa^.n

- _ ^ ^ ^
_,
_:
.i .^
i=

,^, is
^^^` ^^,1`CY
^ R

THE T^lEME PBQB^ENi


e
a

..^Tl

^. 1^.. J^

',

^'

i
IForeword

From the outset, the use of a Theme Problem has posed certain challenges. Aut>:ors
from academic b3clrgrounds tend to be in need of highly detailed information about
plant and specifications, cahile workers in industry and laboratories must often be
satisfied with indirect information and sometimes with none at all. We have tried
to arrange a reasonable compromise somewhere_ on middle ground. Our decision to
select a problem related to a realistic modern turbofan engine had special ramifi-
cations of its own, not the least of which was the fact :.h a L- certain types of addi-
tional data were precluded for proprietary or ot}^er reasons. We believed all alC^no
that the advantages of da_a realism outweigh the disadvantages of incomplete infor-
mation.

The chronology of the Theme Problem begins in late summer, 1976, during discussions
with J. L. 1•felsa. Subsequent contacts with several potential Forum participants
led to the drafting of a Tentative Theme Proble:,^ Description, which was sent out to
various•worlcers for cr.itiaue in early 1977. jti'hen evaluations orere in hand, a Theme
Problem Description was prepared on March 1, 1977 and became the warkin^ document
for authors preparing paper.• for the meeting. Corrnnunications with several addition-
al researchers established the need for minor modifications and clarifications,
which were decided at a committee meeting held during the Joint Automatic Control
Conference at San Francisco in •7une, 1977. These decisions formed the basis for an
a^3di*_ion Theme Problem 1`femorandum mailed to all participants on Juiy 1S, 1977.

A1.1 these adjustments are included in the Final Theme Problem Description, which is
included here.

Any clarity which may be present in this final problem description is due in large
part to the valued advice of many colleagues, among whom I must especially mention
$, T,. DeHoff, R. D. I}ackney, B. Lehtinen, W. C. Merrill, J. L. Peczkocaski, C. A.
Skira, and H. A. Spang, III. Credit for any and all obscurities must, of necessity,
accrue to the author.

M. K. Sain
Notre Dame, Indiana
September, 1977

ORIGINAL PAGE IS
=;
OE POOR QUALITY.
+^R^t^IrlAL PRGE 1g
^`1^; ^ DOR QJ.ALIa ^. I'INAi. TIiFAfI: PI:OHLFrf DESCFIPTIOIV

ABSTRACT

_. ^ ,e design problem should serve the Forum goals in at least three wa}*s. First;
ft r,:,-^ •_1d help to unify the presentations and, thus, make them more useful for group
^'tt^^; • zfter publication. Second, it should helg to make Lhe Foz-um relevan t to the
^^re^:ent-day design world by focusing upon a real system of considerable current zn-
_:tere:=t. Third, it should help to delineate the state of comoutaticnal readiness of
thz various design viewpoints, and so help to paint out where additional numerical
::researches would be useful.

£av^aat: Yt is • impartant to recognize the generally positive intent involved with


,_ the use of this problem. It is not intended that the theme problem usage degene~ate
^^,nto a computational contest.

T - 1. Il^TKO>;UCTIO}1
i ^- '
very important developing area for linear multivariable r.ontral has ar.iser. because
T --of r?ceazt increases in the complexity of aircraft turbine engines. Engines in use
today have, essentially, the one • control variable of fuel flow, though some matte use
of a variable nozzle area which is not unli3,e the iris da_aphragm that controls
a p erture settings i^: a camera. Engines in the p at-so-distant future can be expected
to permit control of Tones i.n the stator portions of the various compressor stages.
.- Further docan tl^E development line are engines with enough variable geometry to re-
ceive the informal designation of "rubber engines" by research engineers in the
industry.

I* is widely accepted that the alder, workhorse, hydromechanical control methods are
not equal to these new tasks and that they wall, therefore, gives way to electronic
digital control. The entrance of the digital computer opens up vast numbers of new
design possibilities, which are now beginning to receive increased attention in the
industry. 1'he central role played by the aircraft turbine engines in civil and
military aviation makes clear the economic import of these trends. It would be hand
to select a more timely theme design example for comparison of ?inear control alter-
natives than the jet engine.

In the United States, a joint study is now under-,aay on the Pratt & ^+ Thi.tney F100-Pjd-
100 afterburning turbofan, a low-bypass-ratio, twin-spool, axial--flow engine. Spon-
Bored by the Air Force and by the National Aeronautics and Space Administration,
this study focuses an the linear quadratic regulator theory, applied at nrultaplc
operating points in the control regime.

One eff?ct of the theme usage o£ such a plant in the NEC Forum should be a broaden--
ing of the design discussion to include other design viewpoints as well.

2. PLANT

The numerical model of the jet engine is supplied in (A,B,G,D) form on Attachment 1.
For the A and C matrices, note that columns 9-16 are listed below columns 1-8. This
modal is for zero altitude and for a power lever angle (P1^^) of 83 degrees, whict7 is
near maximum non-afterburning power. The motivation far choosing this operating
po:.nt comes from the fact that every engine has to pass through ti^is condition, as,
fa:: example, an takeoff. Also supplied is a list of the input, state, and outpr^t
va^_iables associated with this model. ThPSe two pages are taken from the report

3
;^?^..

R. J. 1•filler and R. D. llc^cicney, "F100


_ I^iultivariable Control System riodels/
Design Criteria," Pratt and CJhitney
;; Aircraft Group, United Technologies
Corporation, West Palm Beach, Florida,
November 1975.

Because a number of the techniques which will be discussed at the Forum have graphi-
cal aspects, i*_ is planned to facilitate the inclusion of curves in the publication
by limiting the plant to three control inputs. In consultation with numbers of our
Theme Problem Advisory Committee, we have selected U , U^, and U as these inputs.
Workers who feel an absolute necessity to use all five inputs are welcome to do so;
however, we would aslc that in such a case they pro-.^ide a comparison of the effect
of using five inputs over and above that of using only three. This request is
` designed to increase the comparability of the various design results.

.Actuator information for the three control inputs is given in Attachment 2. Also
prcwided is information associated with the actuatior. of U 4 , if that input is used
in addition to U U 2 , and U Finally, should U 5 be used in addition to (J l , U^,
.and U a servo ^ime constan^ cf 0.02 sec. can be assumed for actuation. Various
rate limits on the actuators can be noted, as in Table A. ...

Table A
Actuator Rate Limits
U1 15,800 (lb,lhr)/sec.

U 2 3.F Ft2/sec.

U 48 Deg/sec.
3
U4 40 Deg/sec. .

The actuators have some limits, also, which will be mentioned here. On LT 3 , it may
be assumed that the limit is -+- 6°. On U 2 , a limit arises because the nozzle area
is pretty well doGm to its minimum at this operating point; the limit is assumed to
be about 1 square feet in that direction.

The Theme Problem models are in absolute, unnormalized form, without any mention of
the set point values. This makes it difficult to size inputs. Tine committee worked
out a proposal to supply "ballpark" set point values so that the model could be
normalized. U^.fortunately, it was not possible to obtain even such approximate in-
formation.

A cors^ , uence of this fact is that the absolute rate limits of Table A have meaning
aaly in relationship to the size of reference commands assumed. Because we are un-
able to supply the suggested reference command, the effect of actuator rate limits
can be treated only hypothetically; and we have to leave the issue of whether to do
this, and haw to do this, in the hands of the authors.

Turning now to the sensed variables, we have available X l , X 2 , X 3 , X 5 , and (Xi2 +


X1 ^), the 1^► st of which is denoted FTIT for "far. turbine inlet temperature."
Sensor time ^_onstants in seconds are listed in Table B.

^^.:_^
..
Table B
Sensor Time Constants
P ^^^^; P AC ^ ^S X G. 0 3
Ft ^^p^T'Y 1
^^ ^ X2 0.05
X3 0.05
X5 C.05
Sensing of FTIT is a bit more elaborate and is indicated on Attachment 3.

3. ENVIRO:^'^iEtdT

Measurement noise is on the order of 1%; and state noise as negligible. Therefore
it is not planned to sup p ly any noise data. Authors wishing to rake noise studies
must make their awn assumptions. This is not unrealistir_ for the present stage of
discussion. :Chough same techniqu4s may well make use of observers or dynamical out-
put feedback,' no Formal stress on filters is anticipated. The Forum, then, is
visualized primarily a.s a control meeting, although contributed papers in tl:e stoc-
hastic area ^^i11 be accepted if they contribute Co the Forum theme,

Practice in the industry involves the use of multiple linear models at various oper-
ating points from sea level to high altitude and from low to high thrust. As oper-
ation transitions from 1112 neighhorllood of one operating point to the neighborhood
of another, these models change in consonance with some physical variable. Para-
meter variation is, therefore, a!^ aspect of design.

But publicly available neighboring linear models arc not near en^^ugh to the Theme
Problem model to provide meaningful data on parametric variatiar.. This fact, com-
bined with lack of set point infarmation, led the committee to suggest a 5% change
in ei^;envalues as one, hopeful.l_y useful, measure of such variation. Because normal-
ization of the model is a similarity transformation, this characterization is 7n-
dependent of set point.

4, REDUCED OR1}ER MODELS

Approximate eigenvalues of the Theme Prob lem plant are -577, ^-17G, -59.2, -50.7,
-^^7.1, -34.7, -2I.3 + i.822, -17.3 + i4.73, -19.0, -b.71 -{- i7..31, -2.62, -1,9I,
-.648. It is the nature of the het engine control problem that these can usually
be well identified with physical. variables. Far exaE^ple, -.648 associates with X10'
-1.91 associates with X -2.62 associates with X and so forth. X • is related
13
to the eigenvalue pair -6.71 + i1.31. This type of infarmation can be deduced
from a study of the eigenvectors corresponding to a particular eigenvalue. It can
he expected that actuator modes, surd as that involved with fuel flow, will enter
into this list. Same discussion on ^.'Yis point can be found in :t. 1,. Delioff and
W. )w. 1ia11, Jr., "Design of a )!Iultivariable Controller for an Advanced Turbofan
Engine," Proceedings 1.976 IEEE Conference on Decision and Control, page 1002.

In the interest of offering some Fssistance to authors who might be having c^mputa-
tional difficulty with the full size problem, the following; reduced model has been
ode available by Dr. Delioff of Systems Control, Inc. (Vt.). It is a model which
neglects sensor dynamics, augments the plant by tha dominant actuator dynamics, and
then reduces to fifth order. Ilse resulting five states are

Xl = Fan Speed (rpm}

X 2 t Compressor S]^eed (rpm)

X3 = "Augmentor" Pressure {ps^.a)


X4 C F°ue1 Flow (lb./hr.)

3C5 = Burner Pressure (psia)

IV'ote that the "^1uo entar" Pressure X is not to be identified with X the quanti-
5
^ ties are not. defined at th° sari° physical location. Nate also that X5 was not
:,
- ^ one of thv oragina7. states.

.:, R.emark: The ti Actuator diagram shotas a Servo 5ystetr gain of 2.4.
It has come to
=;; our attention hat a more realistic number for this gain would be about 12:0. The
effect of this gain change is to taI`e the dominant CIW position actuator eignevaluF
.^; from a location of high dominance in the overall plant-actuator system to a location
of cansidc^rably less dominance. It is not necessary for authors to make this charge
if they , ha^e already completed their calculations, inasmuch as the 2.4 gain apparent-
ly is one of those "g15.tches" which crept in an uninvited manner. Some authors may
choose to comaare the effect of the gain 12.0 with the gain 2,4, if time and space
permits. Cde have included this remarl-; Here so that the reduced order model, whic}^
has the soma controls and outputs as the full. size system, may be more understand-
able.

A (5 x 5)
^-.3245E+O1 -.2158E+01 -.9355E+03 .5731E+D0 .1342E-^•D3
.1'42L+O1 -.5941E+01 --.2816E+03 .1897E+00 .5705E'+D2
,1685E-01 -.255^sE-OI -.1003E+02 .7994E-0l .5S07E•i^0
.0000 .0000 .0000 -.1000E+D2 ,0000
-.2163E+01 .6862E+O1 .7405E+03 .1195E+01 --.1715E+03

B {5x5)

.1432E-01 -:3553E+03 -.9906E+02 -.1549E+02 .2220E+05


.2871E-00 .7286:+03 .2514E+02 -.6487E+02 .8122E+D4
-,2469F.-^02 -.1030E+03 .6333E+00 -.3213E+00 -.7418E+02
.100E+02 .DODO .00OO .DODO .aoao
-,1311E+00 .3295E-t-03 -.2500E+02 .b257E+D2 --.6445E+D5

C (5x5)

.1662E+01 -.1168E+OJ. .7999E+D2 -.1890E+00 .3771E+02


.1383E-01 .3142E-05 -.1D60E-C)1 .12$9E-03 -.1$39E-^06
,1694E+00 -.1129E+00 --.4959E+DI .7386E-D1 -,1835E+00
.7590E-04 .3269E--05 -.1,477E-01 .2284E-05 .4315E-04
-.4859E-04 .1381E-03 .1140E-OI .1951E--D4 -.2688E-02

^ ( x 5)

,1302>;Y00 .1942E+03 .4802E+D2 -.1503E+f)2 .1083E-^05


.1.449E-06 .3395E-^-00 .6506E+00 .28121-^03 .3204E-03
.2967ER 01 .7927E+D2 .2567:+O1 --.7637.E+00 .2D66E+04
.1046E-05 -.772DE-02 w.5514k,-D2 .11.57E-03 ,G6D5E-01
-.8395E-05 -.7897E-02 -.6$41E-D3 -.9643E--03 -.2815E+00

^^.
5. SPECIFICATIONS

The overall viewpoint of the controller is quite simple. T}Ie pilot has one lever,
which we *.light intuitively call the throttle and :^]Iich sets I.hat is called in the
industry the "power J.ever angle." Basically, the pilot increases the levor angle
to obtain riore thrust. All the ether variables must be controlled so as to achieve
the new thrust quickly, but I,itl^out overshoot and ^.itllout violating some important
physical considerations. AIi example of one of these is the temprcature at the inlet
to the "high" turbine jest aft o£ the burner. This temperature is ordinarily
scheduled very near its maximum safe value, and temperature increases are not wel-
come because the turbine elements are thin, respond ver y fast, and can be permanent-
ly damaged or create a need for more frequent engine overhauls. elnother example of
a constraint i.s the various undesirable stall conditions in the compressor.

Thfs problem comes down_ to IIS in the following form. Ass:Imin^; a step change in
power lever angle, we want to move the engine to a slip,htly different operating
point in the above described acceptable dynamic fashion. The power lever angle
change is converted by a master engine scheduler into a reference input for our
linearized feedback model. The nature of this reference input is not highly specif-
ic. Step inputs are commonly studied. It is not likely that highly detailed infor-
mation about these references will be available, but we can try to firm up any par-
is cular issues which may be crucial to one paper or another. The exact nature of
these references gets one into the exact_ nature of the schedulers. It does not seem
too productive in a linear meeting to go very far into such "global" issues. If
greater reference variety is needed, it can probably be safely assumed. It would
he good, however, if each paper tried to discuss at least the reference step.

For purposes of design, we can group the variables into two fami:ies. Y l , Y 0 , X1,
and X 2 ar.e desired to respond fast witl.out overshoot. Y 4 should not decl-^ase more
than .05; Y 5 should not decrease more than .15.

Remark: TIIe decrease limits on Y 4 and Y are to be refiarded in the same spirit as
the U actuator gain change in the precc^ing section. If calculations are complete,
there 3 is no requirement to incorporate it. Some authors may wish to study its
effect, however.

6. VIEWPOINT

We believe that the theme problem should appear in each presentation as the major,
and prc•bahly the only, illustration of the particular design methodology being
described. Gde visualize each paper as an exposition of design viewpoint, with jet
engine illustration. tde do not vis^alizc the paper as an exposition of jet engine
design. In other words, the theme problem will be an apparent thread through the
fabric of the Forurn, but the pattern of the fabric will be set b_y the various linear
control alternatives as entities in themselves. Put in yet another way, the forum
is on "Alternatives for Linear Piultivariable Control" and is not upon "Various
Approaches to Jet Engine Control."
ORIGINAL PAGE IS
Q$ QUALITx

L">^fG1AiA.. .

1 ^ ^^^ r

1
1 '
^,;. o^,

^+•^ ^^ ^

loo MoDEZ, Az^ =o. a p^=s^


THE A MATR;rX

-4.328 .1724 5.376 401.6 -724.6 -1.933 1.020 .-.9820


-.4402 -5.043 1.27.5 -233.5 -434..'. 20.59 2.040 -2.592
1.038 6.073 -165,0 -4.483 1049. -82./+5 -5.314 ,5.097
.5304 --,1086 131.3 -578.3 102.0 --9.240 -1.146 -2.408
.8476E-02 -.1563E-01 .5602E-OI 1.573 -10.05 .1952 --.8804E-02 -.2110E-0?
.8350 --.1249E-01 -.3567E-Oi -.G074 3.7.65 --19.79 -.1313 -.2962E-01
.6768 -.1264E-01 -.9683E--01 -.3567 80.24 -.8239E-01 -20./E7 -,3928E-01
-.9696E-OI .8666 16.87 1.051 -102.3 29.b6 .5943 -19.97
-.8785E-02 -.1636E-01 .1847 .27-69 -8.420 .7003 .566GE-OI 6.623
-.1298E-03 -.2430E-03 .2718E-02 .3214E-02 --.1.246 .1039E-Oi .8395E-03 .9812E-01
-1.207 -6.717 26.26 12.49 -1269. 103.0 7.480 36.84
-.273f]E--01 -,4539 -52.72 198.8 -28.09 2.243 .1794 9.750
-.3206E--02 - ,2017E - 01 - 2.343 8,.835 - 1.248 .9975E -01 .8059E-02 .4333
-.1613 -.2469 -14.05 ?3.38 146.3 1.638 .1385 4.486
-.1244E-01 .3020E-01 -.1198 -.4821E-01 5.675 -.4525 19.81 .1249 a
H
-1.653 1.832 -3.822 113.4 341.4 27.34 -2.040 -.6166
n
^ .9990 1.521 -4.062 9.567 10.0$ -.6017 -.1312 .9602E-01
r^
11.32 1Q.90 -4.071 -.5739E--01 -.6063 -.74$5E-01 -.5936 -.9602E-01 zti
-.9389E-02 .1352 5.638 .2246E--01 .1747 .2407E-01 1.100 .2743E-01
-3.OEtl -4.529 5.707 -.2346 --2.111 -.2460 --.4686 -.3223
.2090E-02 -.525GE-OI -.4077E-01 -.918ZE-02 --.8178E-01 .3428E-01 ,4995E-02 -.1256:-01
-.1953E-01 -•.1622 -.6439E-02 -.23,6E-01 -.2201 -.2514E-01 -.3749E-02 -.3361E--01
.1878E-OI -.2129 -.9337E-02 -.3144E-OI -.2919 -.3370E-03. .8373E-01 -.4458E-01
.2253E-01 .1791 ,8371E-02 .2645E-01 .2560 .2835E-01 -.3749E--01 .3635E-01
-49.99 .6760E-01 39.46 .4991E-02 .8983E-01 .5349E-ri g .0 .1372E-01
-.6666 --.6057 .5847 .6G54E-04 .1347E-02 .7131E-04 .0 .2057E-03
,2854 2.332 --47.65 .34x6 3.OG5 .3624 -.4343 .4581
-9.627 -9.557 38.48 -50.01 .1011 .1203E-01 -.4686E-01 .1715E-01
-.4`78 -.4245 i.71U ,-2.000 -1.996 .5.149E-03 -.1999E-02 .7544E-03
-4.414 -4.354 17.66 -3.113 -3.018 -19.77 -.4999E-01 .1509E-03.
-.1127E-02 -.6760E-02 .1835E-01 -•9981E-03 -.1347E-01 -..1070E-02 -20.00 -.2057E-02
.5004 -.1437 -2.416 -.1073 -1.Or'8 30.63 19.89 -50.15

,^

® ^ ^ ^i0r ^b ^ Iorl.e ^^^ ^nrd rrrr^ t^ro^ rrs ...i^


TIDE B MATRIX

--.4570E-Di -451.6 -105.8 -1.5D6 851.5


.1114 -546.1 -6.575 -107.8 3526.
.^i53 1362. 13.46 20.14 -.6777E+05
.3262 208.0 -2.88$ -1.653 -2.69.1
.9948E-02 -98.34 .5069 -.194D -94.70
.2728E-01 71.62 9.G08 -.3160 -184.1
.1716E-01 71.71 8.571 .7989 -515.2
-.7741E-01 -141.2 -.8215 39.74 1376.
.3855E-01 -7.710 -.4371E-01 -.1024 -6684.
.5707E-03 -.1144 -.6359E-03 -.1432E-02 -99.02
5.727 -1745. -8.940 --17.96 .8898E+05
.1392 -24.30 -.2.736 -.3403 -6931.
.6172E-02 -1.082 -.1183E-01 -.1452E-01 -307.7
.6777E-01 16.60 .3980 .2311E-01 -2588.
.1880E-02 9.'_47 -.8241 .8984E-OZ -32.31
.1677 435.8 -89.94 4.900 -295.5

THE C r1ATRIX

^ .4866 -.6741 5.392 95.42 24.03 ].0.52 .8190 -.4492


.1383E-01 .2709E-05 .0 .0 -.1081E-01 -.5545E-04 .4722E-04 .0
.0 .0 .0 .0 .0 .0 .0 .0
.7418E-D4 .5496E-05 .4790E-05 .1478E-03 --.1504E-01 -.6503E-04 .8820E-04 .4999E-05
.1538E-04 .1201E-03 -.2579E-02 --.1609E-U3 .1618E-01 -.1071E-02 -.9561E-04 -.5503E-OS

j .SI95 .8437 -1.863 .5709E-01 .4815 3.428 2.161 .7681E-01


.0 .0 .0 .0 .0 .0 .0 .0
.0 .0 1.000 .0 .0 .0 .0 .0
.3434E-05 .2727E-04 .1128E-D5 .4002E-05 .3673E-04 .4290E-05 -.498E-05 .5609E-05
-.3732E-OS -.2996E-04 -.1234E-OS -.4330E-05 -.4024E-04 -.4721E-05 .5324E-OS -.6103E-05

THE D rfATR IX

-.6777E-OI --42.5 32.97 -1.824 1245.


.1282E-03 .3353 .6804 .-.5605E-04 -.1199E-01
.0 .0 .0 .0 .0
.1030E-05 -.1193E-01 -.5806E-02 .6015E-04 .4463E-01
.8109E-OS .2328E-01 .1178E-03 -.5538E-02 -.1039

^ll
I.

^.
1. Engine State Var^.ables
X1 =Fan Speeci, SNFAN (N 1 ) - rpm
X2 = Compressor Speed, SNCO:; (N 2 ) - rpm
X3 = Compressor Discharge Pressure, Pt3 - psia
X4 Tncf:rturbine Volume Pressure, P t ,^ , 5 - psia
X5 = Augmentor Pressure, P^^ m - psia
X6 =Fan Inside Diameter Lischarge Temperature, °R
Tt2.5h
X^ =Duct Temperature,
T t2.5c - °R
1 X 8 = Compressor Discl-:^rge Teperature, T t3 - °R
Xg = Burner Exit Fast Response Temperature, °R
TL-4hi
X10 = Burner Exit Slow Response Temperature, Tt410 °R
X11 = Burner Exit Total Temperature, T t ^` - °R

r X1Z ,= Fan Turbir._ Inlet Fast Response Temperature,


X13 = Fan Turbine. Inlet Slow Response Temperature,
Tr4.5hi
Tt4.510
°R
°R
Fan Turbine Exit Temperature, T t5 - °R
X14 =
Duct Exit Temperature, T t6c - °R
X15 =
X16 - Duct Exit Temperature, Tt7m °R
t 2. Engine Ir_puts

U1 = Main Burner Fuel Flog=, i^tB - lb/hr


U 2 = Nozzle Jet Area, A. - ft2
3
U3 =Inlet Guide Vane Position, CIW - deg
T U4 =High Variable Stator Position, RCW - deg
US - Customer Compressor Bleed Flow, BLC -

3. Engine Outputs
Y1 = Engine Net Thrust Level, FN - 1b
Y z =Total Engine Airflow, I^'FAN - lb/sec
Y3 = Turbine Inlet Temperature, T t4 - °R
Y 4 =Fan Stall Piargir.:, SMAF

Y S = Compressor Stall rfargin, SMHC

t
t
10

l
^.

%x^.

A^^Ar,^it^t^ta^r 2 ^:

rieterin^ V11ve Pump Controller


F'uei F'lo;a 1 1 Fuel Flow }
^-
0.025 + 1 O.1S -^- ].
Request i'o F,ngine

U I Actuator

Scra p Syste^^ Air riotor Bynamics


1 _ 6
_ 1 1 2 2^
A^ Request O.O1S + ]. S + ---5 + 1 ^`
., ^ 2 ^ A^ Position A

,: n n
4

c^ = 6Hz
n
^ a 0.5b
i

- ''

T.T 2 Actuator
';
Servo Power
StepFrr Motor System System Cylinder

Position 1 2.4 1 CIVV


0.025 + 1 - ^, O.O1S + 1 5
Request Position
.^

{:.^^

[^^ A: tuator

_ ^

Servo System Power Cylinder


Position Request RCVV Pnsitiox^
40 1
^- ^ 0.015 -+- 1 S ^^`

U Actuator
4

11

:^
__
g
ATT.ACfi^fFNT 3

FTIT Sensor

_0. i09
0.5955 + 1

.--^ + FTIT
FTIT
^" Sensed
O.C9I +
'^-
^ --
_ ^^' S.49S + 1
^

^.. ^ '.

i `

^::^:
^:.

t
.;

I2
ei Appendix J

ii
i ''INPUT COiff ENSATIO^I

FOR DO^fINA"CE Or TURBOFAN MODELS"

i
i R. M. Schafer
M, K. Sain

^l

_ - .^. -
^ ^^'^
...

j: ^.

• ^^

^ ^^ - •,

l ^.

7.v'PUT CO?1PE?^SATIO:v FOP. D0`fIi;A^'CE OF TUKItCfF:1^ PfODEI.S


^^-
R. t•i. Schafer and t1. K. Sain
T.
• bei?artment of Electrical Engineering
University of Notre Dame
Notre Dame, Indiana 4b5S6
U.S.A.

• ABSTRACT .

The determinant o` return difference establishes a crucial link bet h^een open and T

closed loop characteristic polynomials in multiv;?riable feedback control s}stems.


As a result, :^} quist constructions on this deter:+.inanr carry important design in:or-

matian. One wad to extract this information is by achieving diagonal dominance.


Ticis pG,?er presents a method c^^hich uses dynamical input compensation to achieve
Coluail dor:^ina:ic^^. Application to the Theme Probler.: is included.
'• .

^•! .

3%;

^^ 4 ^ Z

ry .,

^^ ^+ ^`
w ./

^tii

^^

♦ l'

.. `= /


t. • ••
f
• 1 r.

.^

M
. ^

^^•,,

.t

^' 4,

^.A • .
^ •

^ ^ •
.^,

INI'11'i' C^:SFF.`7SAT70': I'OR DO': ,.SCI: OF 'I'111:1:01':1,I ;•SOD1:7.S


--
t^
^ •
^_, R. I1. Schafer and i`i. K. Sain

'>;
Department of Electrical Engineering
University of Notre Damc
• Notre tame, Indiana 46556
U.S.A.

. AnSTf:ACT

The determinant of return_ difference establishes a Crucial link between open and
closed loop characteristic polynomials in multivariable feedback control sys`ems.
As a result, Nyquist constructions on this determinant carry important design infor-
mati.on. One way to extract this information is by achieving diagonal dominance.
This paper presents a method which uses dynar:^ical input. compensation to achieve
column dominance. ^lpplication to the Theme Problem is included.

^^ 1. INTP.ODUCTIO.d
Recent advances in the generalized Nyquist theory ^3r Iinear multivariable feedback
^s • control systems Y:ave brought about very substantial new op p or_tuniti.es for research
in the area of frequency domain control design. i`fost of th_•se advances are predi-
sated upon the relationship between closed loop and open loop characteristic poly-
, nomial.s--as embodied in thedeterminan.t of return difference. Features of the Nyquist
diagram of this detert,^inant are important aids to control system design.
^ It is apparent that a diagonal return difference will decompose she return differ-
ence determinant into a product of its diagonal elements, thus reducing a multi-
.
' variable problem to classical single-input, single-output form. Less apparent, but
of much greater practicalsignificance is the fact that an approximately diagonal
return difference can have essentially the same reducing effect on a multivariable
^- ^. problem, t^*hen regarded from a generalized Ny q uist viewpoint. The best known of
Y
,- these approximately diagonal con:litions has come to be described as diagonal domi-

Hance. A productive dCS1gI1 strategy can be mounted, therefore, in two ste p s. First,
b
achieve dia^onaldominance' > second a 1Y classical sin g le-in P
> PP g
ut^ sin le-out Put
'techniques [1]. ._
., Unfortunately, methods to attain diagonal dominance have been rather slaw to advance.
For the most part, they have been restricted to the selection of constant real co;^t-
^pensators, the entries of wh-ch are typically obtained by procedures of optisniza*_ion
^: , that da li.ttle to preserve some of the classical advantages, such as insight, affor-
- ded by the frequency domain approach. rluch work needs yet to be done on the theory
of attaining diagonal dominance by use of frequency dependent, dynamical compensation.
,,^ This: paper considers the application to the 'Theme Problem of a useful new desii;n aid
callr_d the C:1F.DL1D I'Iot. In its present form, this method deals ;^}ith the design of
'
s dynamic precompensator for the plant, in such away that column dr^minance is
•^ achieved. An im p ortant feature: of the approach is the enhancement of designer in-
••• sight toward the coupling present in a plant. -
Secti^rti 2 introduces the CIIRDIiD method for two-input, two-output plants, and Sec-
Clo p 3 provides an illustration of certain basic features of the method, in the con-
e text of a het engine plant related to the Tlleme Problem. Section G gives a genera-
- F
Iizatioit of the idea to tlsree inputs and three outputs, and Section 5 applies these
^^results to the Theme Problem. Conclusions appear in Section 6.
• ^

-

^
• ^ ,• ^

^.-. ...... ^_, „^ ,.,..^• stn• ey.. r— ^ r. ._^_. ,..... .^.. ..


-----
. ^ - ••-j
t ;^ ;, ,
^c^r^;e ^ .,.... — _ .,^ttati^.^ __.
?.. c l ;nrrizcnL nr Yl :oncrl f

^^'he i tli column of a matrix Z (s ) is said to be dominant if '


n
'^
J^^- ^
for all s on a Nyqu^.st contour n, k similar definiCion can be made for row domi-
nance.
For 2 two-input, two-ou~..put system, Lq. (1) can be equivalently written

IZif
cG) {2 - iZ
3 i (s)i Z > 0 i ^ j (2)
for all s on D.-
Censidcr a two-input, two-output system havinE on?y precompensation. The open loop
transfer function of the system is

Let K(s) be restricted to the forn^


1 a (s),
2
K(s) = (4)
a l (s) 1 J .

^' Since any matrix having; nonzero entries an its main diagonal may be put into this
form by multiplication with a diagonal matrix, and since . multiplicat^un by a diaFO-
nal matri^c does not affect dominance, this czn be done ^wit^^out essential loss of nen-
erality.
Let G(^=) be evaluated at a specific frequency w. Then ,
ll j r 12 + i l ^j 1 x^ + y2j
rll + i

r 21 ^ i 213 r 22 + i 22 J xi + y lJ 1
Performing tlzz indicated matrix r,^ultiplication, thr_ four entries in the matri;;
Q (s}^ - are
j^
q ll r ll -!- j
i ll + (r l ^ + it ?j) ( xl + Y lj ) , (^) .
+
ql2 ^ r12 + i 12 j {r
ll + illj) (X2 + y 2 j
), • (7 )

( ), (8)
q21 r 21 + i21j + r 22 + 122J){xl + Y lj

q 22 r 22 + i 22 j + ( r 21 + i ^ 1 J)( x 2 + Y 2 j )• (S)

From Eq. (2}, the first column of Q(s) f will be do:^iinant if


^
s=jw
(10)
^ g 11^ 2 - ^g21^2 ' 0.
^^ Performing the indicated subtraction resul[s in what will be referred to as the domi-
Hance inequality for column 1. The forni of this inequality is
f^(x l ,y l } = axe + ayi + 2bx 1 + 2cy l •I- d 0,

'where the constants arc defined as


2 1. 2 2 ^
a ^ r AZ + i 12 •- r 22 - i ?Z (11.

' .b ^ 21 i_ 22 , {12)
:r ll r l2 + i ll i l2 r21r22 - i
G e t 2i 1 -+ r r -
1 1 21 i22 1li1 2 r 22. i 21' AT p,4^^ r3 (13)
OF
u 2 2 _ 2 _ 2 ^R ^UALITy (14)
x l.l + i ll. r 21 21^
Zrotc that each constant is composed of complex field e lem_nts which came from evalu-
alien of G(s) at a specific frequency m.
The functian f l ( x l , y l ) is a para^soloid in three-space and is normal to the xl-yt

1^= olane. If this paraboloid intersects the s l- y I plane, the intersection will be 3
circle. Stanr'.ard maximum-minimum analysis gives that the maximum or minimum of the
dominance function occurs at
xx ^ -b/a yl -c/a (I5)

To dctermir.e if tl^e point that was four_d i s a minimum or a r.:aximum, the hessian is
formed. If the Hessian is negative definite, the point found is a ma:;imum. If ti7e
3 hessian i.s positive definite, the point found is a minimum. The Hessian of the dor.:i-
nance e^uatian for colur..n one is

^^
^ 2a

0 2a
0
^
(lb)

so that the second derivative tf_'st reduces to a test o:^ the sign of a. ^ _
Pror_eeding from this ar:alys^.s, there are four possible cases. T}ie point that was ,
found Gas a positive maximum, positive minimum, negative n.;z ;i-:um, or negative mi_ni-
mun:. The twa cases that are of interest are the positive maximum and t^^e negative
minimum since it has been shot„•n [2J that the other two cases cannot occur. In each
of the cases of interest, the positive maximu^^ and the negative mini,;:um, there is
an intersection of the x - y l plai.e. Recalling that the column will be dominant if
f l ( x i , y l ) is positive, t^e analysis of the two cases is as follows. In the pasitive
maximun case, the values of x l and y l which will result in solution of the dominance
inequality are those points which lie inside the intersecticn of f l (:^ l , y l ) and the
x l-}' 1 plane, that is the circle which is the solution of f l (x l' y l ) Q' In the nega-
tive minimum case, the choices of x l and yl which result in solution of the domi-
nance inequality are these poi..^.ts which lie outside the circle of intersection. .
Thus, the intersection of the dominance function fl('^^1'y1) for column one and tha
x l -y^ plan:. defines the acceptable ra.zl;e of x l and yl such that the system will be
dominant in the first column at the stecific frequency at which the analysis c^as per-
^ .formed. In like fashion, the second colunn of th4 system may be analyzed, and the
• acceptable choices of s 2 and y 2
may be determined.
If. this dominance analysis is repeated over a range of frequencies, and the result-
, i.ng circles of intersection plotted, a C:1i.DI:1D (Complex Acceptability Region fer
DlAgonal. _Dominance) Plot is produced. A solid circle is dra:rn if the acceptable
''.^ choice of x and y lie inside the circle, and a dashed circle is drawn if the accept-
, able region is outside the circle of intersection. Associated with each C:1RDI:1D plot
is a locus of centers p lot, which indica[es the centers and labels tt^e frequency of
^^^ each. Space limitation, do not allow the locus of ccr.ters plots to he included with .
tt^e CARDI:ID plots in this paper; but they will be mention4d and referenced as neces-
nary,
. 3. ILT.UST1u^TIO^I
Figs. 1 and 2 are C:1ItDIAD Plots of a two-input, five-state, two-output model of a
jet engine. The n:odcl is derived from a jet engine siriulator called D1tiCl:^I [3,'.J
^"
*^ and represents an F-1.0;1 turbofan ,jet cnl;ine with a fuel flew of 2,75 Lb:a/sec. (full

%^ `.

^^^^^ ^^^,g^"^,

;^ rea f. iY'
4^ P^O^ r

.
r %!

!!r ~
Y om,
`\` •,, ^` ♦•.
•.

l
^(i i^ .I ' i,.t SI^4l
„ ! 111 _Y^ •; _ ^
^^^^.^' _ —r ^ ^ \ ♦+♦ , •+ * \I ,^; \, \ +1
1/71
1

'^ `^,,^ 1 .+ ♦11 f+


1
r te r !
1
1. .f^ ,l^
1 • i

^O 1 i • I
L ; ; ^ 11 ^^
^^ +^ 1 I + ;
I, 1 ♦ ^I i^}t I f ^ I
/ 1 • .^ _
RN f ; ^ 1 ,1\; ., ^,^^ ^ s„1 !♦ ♦• !! ^ 1 1 ^
1 ♦ r ay i =• : ^. ^: ^ q :: ^ % ^ It i f + l ♦ ♦+♦^ f ^ '.^..
t 1 1 \r f

1 E'
r•f

1 7 •l f I I

t o f
a
1

+
r
^

• •
+tea_ ^'
1

^„ ' T^ c..^..^'
i ^ r!
r f
r +
!
^r \\`
I^R^
-. -
^^ ^a1
!
`

^I
^ ^,.
!

,+
`^: r ,• ,
f-1.M ♦.oa+ `:¢ -2.CG - LC'T -
• Y 1 • t.p
-- Es9' y
` 7.t9 , , •.r + s^
t. flEPt ltEPt. P t34I

t Fag. 1. Column 1, Uncompensated Fig. 2. Column 2, Uncompensated

,i throttle without afterburners}, The inputs axe fuel flow and exhaust area and the
outputs are thrust and high turbine inlet temperature. This model is one of a
sexl.es of such models presently being used in a set point study of an F-100 like

i ^et engine.
The analysis of CARDIELD plots proceeds as follows. Recall that, at any given fre-
quency, the acceptable region is outside the circle if the circle is dashed or

l inside if the circle is solid, The first question of interest is whether the colu;^ns
of the aysterE are dominant unCOnponsated, For this to be the case, the origin of

i
the C^iRDIAD plot must be 3.ncluded in all solid circles and e^:cludcd by all dashed
circles, since the origin xepresents identity compensation of tY:e column. This is
not the case far either of the two CARDIAD plats of this system. The next question
is ^--rhethe: the system can be made dominant by constant real. precompensation. if
this is the case, there will e^cist a point on the real e^cis which lies inside all
solid eixcles and outside all dashed circles. Fig. 1 shows that the first column
of 'the system can be made dominant at all frequencies by the choice of any constant

6 x1 which lies outside all the dashed circles of the Ct1ciAZ^'^D plot. Fig. Z shows
t.fiat there exists no constant value that will make the second column of the system
dominant at all frequencies. Thus, some form of frequency dependent precompensation
will. be necessary.

c before grocceding wz.th dominating this system, some of the features of CARDI.AD plots
should be mentioned. One property is that d circle at a specific frequency in the
plot far one column will be sl^lid ii the other column is dominant at that frequency
and will be dashed if the other column is not dominant. From this £act it follot.s
that the transition from one type of circle to the other in the CARDI^'1D plot for
one column occurs when there is a c.hangc in dominance in the other column. Once

a again considering Figs, 1 and 2, these £acts indicate that the second column is not
dominant at any frequency since all of the circles in the CRLgDIAD plot for the £zrst
column are dashed and that- the first column is dominant at low frequencies {until

i W =7} because the circles in the CAF.UTAD plot for the second colunn are solid for
this and all lower frequencies.
^ second .feature ox the CARDI^ID Plat is the effects of a column switch on the plots,

t n

w. _.. .. .. -. _. .., ,. .-
....:.... ^. .. w.-r, -:., Imo., .^ _ .-,t....,t, ^ .:... a

.:. ,.. . .
'^

^^ that is, pramultiplicatian by a matrix with the only non^.ero entries being off-die--
• gonal 1's. The effects of such a switching of fire inputs are that all solid circles

' become dashed circles, all dashed circl^.s becat^e solid, and kite shapes of the column
. _ one and twa plats are switched. Th` Cf1TtDT^1D plats o;E the system with this type of
compensation ar^^ given in Figs. 3 attd G. :dote that the first column is now dnrti-
nant at all frequencies wl.thout further compensation. This fact cren be ascertained

__ A -- - ^F ^M
O

^ •y\
t
i L
^ 17
l
^^^ ^
. ^\_t\
f `
^^^}V, .", . , .
^^ ,,, yi
f e
if1. 1 f
i I
^^ 1. ^^ n ^
,.. ,_^^^

1^^

i' ^' h:
'^; . _f
^ ^:==^r^,f ^ ^ •fir
^- %^ ^ ^[ ^ ^

^ • ^^

[3.90 -I2.'I.7 • i0.50 .10 -x.57 •• '^ -2.v0 .77 1.1: ], 1Y 4.


EEH^ ^44Z

^^.g. ^. Column 1, G{s) K^ F^.g. ^. Coiumn 2, G(s) Kl


e^th^.r from the fact that the origin in the CARDI^,D plat for column one is included
,
^^ _ oy al.l solid circles and excluded by all dashed circ,.es, ar from the fact that all
of the circles in the CA£.D IAD plot for the second column are solid.
^^nce switching the inputs makes one column dominant; uncompensated, it seems a logi--
ca3. first step in compensating for dominance at all. frequencies. Thus, K Z is chosen.
to ba ^ 1

^ ^ - (la)
^ a
1!:t 3.s still necessary to ntal:e :he secand column af- the system dominant. From the
C^DIAD plat for this column {Fig. ^), it is apparent that frequency dependant com-
pensation will. he necessary since there exists na point in the real axis which is
r ^.nol.uded in all the solid circles of this plat. To design such'a compensator, a
f^sncCion of s is fitted to the shape of the CP^^DT1D plot so that, at any given fre--
gt€enc}', the compensator lies inside the solid circle associated with the same fre-
e quancy in the CAFAI<tD plat. While at is possible to find a first order compensator
That will make this calitrtn dominant, a second order compensator has been used be--
cause this same compensator tould also achieve dominance at four other set paints
of tIte model.. is (s) is the compensator the'- achieves dominance in the secand column
- pg U ^s d x^Z , ^ -
. ^ °-. 742s ^ 9. ^9
vc^l4s^ — 0998s -t• Z.
. ^^ fs) ^^.a)
GAT ^^^^ IS
l^ 1 dr
..,1.j ^^ ^ ^^A^^L^`7^
^.
^:-^ a
e 1

- ;.t
- ^3

4 ,
^ ^ the averal.^. ^am^ensation is K3,^R2 (s) ^ K(s) given bcl.ow.
. ,
t ^'f p ^. _'
(I^1)
^"^s) ^ ^ PAGE ^3
^ --0.742s — 9.59
2 n ^ ^ +f^iJALITY, c
o ^l^+s — s^J^ s '^' 1.

^ The C^iRDTAD plots of the system with this compensator are giti^en in Figs. 5 and 6.
^ _ It 3.s obvious either from the fact that only solid circles appear in the plots or
from the fact that all. the solid circles include the origin that each column. of the
system is now dominant at all €requencies.
`^'.

.^ ^.

,..;` `^ -
F^.g. 5, Column ^, G{s)^K(s) Fig. ^. ^ol.umn 2, G(s)^1:(s)
^_t^.
^. GEt^^RALT^k^TT^J^
the CARDTAD Plot approach tv syster: dominance in the three--?nput, three-output case
r 4
^ ^s similar to the app?-aach in the two-input, two-output case.
- the actual condition for dominance in the ^ x 3 cases is the i th column of a matrix
^ Z(s^ will be dominant if
c _ ^
^ ^ -

P ,. i 3 ^i
fox a].l s ort D. Tf bath sides of this ^.a^egt^ali.ty are squared as, in the 2 x 2 case,
^: Mien an equivalent ^ condition is ^

^€.
^i
^^. Using inequality (?l.) = the condition for dominance in, say, the first column i.s
^^ 2
^^ l^x^^s^ ^ ^+ z s ^ ^^^ s
,^-2x ^ I z
1 ^^^ s^ l C^^^
^^^^^^) ^ 1 31 ( ^ i
' the czass term produced by squaring adds non--integral. power reams to-the dominance

.•,^_
^^. •..

`` 1 •

i^

`l, _ .,.
_
inequality for the 3 x 3 system. To circui^^vent this prnUlem, the last term ai ine-
•quality {2^} ^.s replaced by an upper bound. Since '

23. j j 31 I— I a l II 3l I
with equality when ^y2i1{s}I ^ Iz 31 (s)I, it is convenient to^replaca the last form
- of inequality (22} wits the Left member of inequality (23). This yields a suffi-
.^
cient condition fox dominance. For column 1, the condition is :.- ^;
-: ^z ^x (s} I^ — alz a^. {^) I^ -- zjZ ^^. (^) I Z > ^; c^^^
and the general farm is
. , : ^ ^ 2
- ^z^^^s} I . --z ^ jZ ji {s) j ^ o, ^ = x,^,^. c2s3
^-^
^^^
` From inequality (24}, the derivation of the dominance equation for the 3 x 3 case
proceeds analogously to the 2 x 2 derivation. The general form of the compensator
^ssed i.n the analysis is

^ ^^-^{s, a13{s}
^($^ - u 2^. Cs7 l x`23
{s) ' {2b 7
X31{s} a32{s} l
^' _..
where a ij = x ij + y ij 3. .
o

Cnce again,6-^the
^
open loop transfer function matrix G(s) and the general form {26) of
• the compensator are evaluated at a specific frequency and multiplied to form Q{3w},
(' Then, using inequality {2S'1, a dominance inequality for eac;i of the three columns of
• ^{jw} can be form:.d. For example, the first column of Q{jw) w^.11 be do;ninant at the
frequency w ix
2
—2 (2 ^ }
^ ^^a. I ^ °^^ I q ^i I ^ I q3l 1 ' °
and the dominance function for co7.umn 1. is

f ^.^x21' y 21' x 3l' y3l ) - cl + x21 2c 2 + y21^c2 + x 3].^ c 3 + y31^a3


-^- ^x^lc4 + ^y21c5 + 2x31c6 -+- ^ y31c7 +, 2x^].x3lc$ ^^$)
-^- ^`y^ly3^.c8 + 2x21.y31c9 - 2x Y^ ly 21 c 9 > , d
where the constants c l --c^ are functions of G{s} evaluated at the frequency cu. Simi-^
^.ar dominance functions can be derived for the other two columns. '
The maximum-minimum analysis is performed in trao different ways. zn the first
appxoach, which c,rill be referred to as the standard analysis, the variables of the
danirance inequality are first paired by file entry in the compens^-^tor which they
represent; and the maximum-mini:^^um analysis i5 gerfnrmed on each pair assuming that
the other pair rs zero. The resulting maximums nr minimums are
xa l ^ — c^^c^; yxl ^ — c$1c2,

^^^ a -c s f o3 ; y3^. ^ - c71c^.

ti...^ . ^u_..- _._. _



7.'he hessian for'each pair of variables is d^.agonal g ad tl^e second derivative test
'once al;ain reduces to a sign tcaG. •.
The dominance analysis is repented over .^ range of frequencies and C1^RDTAD plots
result. `there is one plot for each off-diagonal entry in the camnenaator and etch
entry is plotted assuming that L• he other aft-diagonal cntrti= in the: column is zero.
' Using CAPtiDTAD plots generated by the standard analysis, dominance is achieved by
^^ setti.ng one of the aft--diagonal entries to zeta while the other i.s chosen as wets the
case in the 2 x 2 design.
Them. does not alraays exist a va"^^_4 in one off-diasanal entry of a column of the com-
pensator that will mane the column of the system dominant when the other off--diagonal.
entry %n that column of the compensator is zero. FThcn this occurs, the znaximum-
tnini.ntum analysis i.s pt^;.formed by finding the full. gxadient of the do:^inance function.
^ The hessian is no longer diagonal. but the eigenvalues of the hessian are x1.1 nega-
tive in Section 5, so the point that is found is a maximum. Design which is per--
formed on ploL-s generated by the full. gxadient analysis involves bath of the off-
cl.iagonal entries of a column of the compensator, and functions must be fit to each
to achieve dominance.
A r^ew symnol appears in the plots. At any given. frequency, unless dominance can be
_' achieved at that frequency with the other entry zero, a small triangle is drawn
'`^ which shows the best that can be done towards achieving dominance. Tt shauJ-d ba
mated that the triangle can appear in plots generated by either analysis. Tn the
standard analysis CARDIt'1b plots, if triangles appear in one plot for a column but
- not the other, dominance can be achieved by keeping the entry in tahich the triangles
appeared zero and using the other entry to achieve dominance. Tn the full gradient
an.a.lysi.s plats, triangles appearing in both plots do not mean that dominance cannot
be achieved. Given that one entry in the compensator is chosen exactly on the tri-
- angle ,^t a-certain frequency, there is a radius of points around the triangle in
the ether plot that will achieve dominance; but since the size of the circle is a
ft^nctxon of how well the other entry is fit to the triangles, such a circle Could '
- easily be misleading. Both of these points wi11 be illustrated in the next section.

5. THESE PROBDEM r'^'3?iI.,YSIS .^ '.

The. following design i.s performed on the reduced order model of the theme problem
wa.th state feedback. Tha states being fed back are the two turbine speeds and the
^^._;.^ pressure P b . Dominance will be achieved using only precompensation.
The plots for the uncompensated system using the standard dominance analysis showed
_ that the first two columns of the system could be made dominant with one off-diag--
vnal entry in each of the first two columns of the compensator zero. The third
-column, however, could not be made dominant at env frequency kith either one of the
- off-diagonal entries in the third column zero. Physically, this a.ndicates that the
principal effects of all three inputs {fuel. flow, exhaust area, and ;uide vanes)
are nn the two speed states. To facilitate achieving dominance, a colunjn switch was
_ ^ done Tay choosing the first camgansator to be ,
a 1 ^

^^ x ^ ^ .
_ ^F3.gs. 7--12 are the CARDIAD plots of the system with this compensator and use the
standard domina.:^ce analysis.. The plots for the entries in the first column, Figs. '
- ^` 7 and 8, show that tlxc first column ^s damxnant wxthaut further compensation, since
the origin a£ each p1a.t is included inside x1.1 solid circles and excluded by all
slashed circles. Fags. 9 and 10 a,:e the Ce1i.I)Tr1D plats for the second cal,umn. Fig.
^.1}, the plot far the 3,2 entry, has several triangles in it, .indi.ca .ti:ng that,. at
^,
s

a^. ^ ^^ .f
! •'^.'7e^^^•`^^ ~,ate -^.^^.: '^

A;^^y t ``,. `^;, ,.^

Cn

s Ik ^! ^='

^+ `' -- ^'^:,'
t`lax ;sip: xai-ss=-5'^

e ^^^. ^ •

t Kl , 3,1. Entry
F^.g. 7. G (s) ^^: 1 , 2,1 Entry Figs $. G(s)

a
• ^:
^r

R :^..

^4
^i.^
4L
/•R

^^

My^^,+
- .

p'
S^

S^ L•
^.i ^:1
Li

F
C a ,'
4,

r
f^.
. ^;i

t' •
a^^+
a
s^'^ ^

•nl
^wS 1
•^^t t
'^' t
'^:,
rt ^

t
tt
i r.. ^

St
t
^
t
t
1
^
i ^^t r

ww
. lt{ ^
r ;
^
t
t
S
rwn 1 t
r.^.li i t

r' ^^ \\ ^
r.7^ f.71
^
a.lo
^
I•
^^•^^
^.lt 9..7 a.l,
7 .n •7.:7 -7.35 •^.ld a.71 9.69
IiEia LAST +t^
^.vi a. 19 9.IS a.l9 9.! ^: ^. ?

4 .ae -e.al 0.71 a.cz ^ •^


• F.C^i ^A'kT +k9

- ^

Fig. 9. G(s)^^^ l, 2 Entry F3.g. ^.^. G(s) hl l , 3,2 Ent^'y


r
,^ the ^rcquencies where they occur ^ t3zere is na value in the 3, 2 entry tfkat wi^.l ma^:e ^
shows that there are
the co^.umn dominant with the l,2 element zero. However, Fib. 4
no such trian^;ies i.n the ^.,2 entry; so, if a function is fit to the shape of the
circj es of tha,s p3.ot and if the 3, 2 entry ^ths a
so^.^.d
;^ ^ch^.eved. T'i^s. it and 3.2 are the CltF.^7IA â plats 1 for ` third n co^.uMn^an Th ao^.n 3han-
_- ^ txy is al.l trian^;^.es .-end the 2,3 entry has tr^.ans*7.es at 3.ok=er fre q uencies. Thus,
Chore is no way to niakc this co^.umn of the system dominant with one of the off-
.:
!, , 5..
f,^

4
-
s i ^.
M..

S.. • a

.:.< •_ ,:, .i ..^:. <^1.^,a.. ..a ., mod_.. ,r
^^

,-.
• ^

i w ,g

i

t ^
a
a

=n.^a ,op i,ra s: s crc^ • p,vs 2^:h6 x.ao sl:m ► a.oa s-.
`^ . ; ^ xE^ Fair

'^
' Fig. ^.3,. G (s) ^xZ , ^., 3 entry ^'^.g. 1. 2. G (s) ^Kl , 2, 3 Entry
diagonal entries i.n the compensator zero.

Figs. l3 and ^.^ axe the plots for the third column using the full. gradient rather
khan tine standard analysis. The solid circles which appear at high frequencies in
Fig. J.^ are very important. Recall that the circle c^^.IZ only be drawn if dominance
can be achieved while the other entry is zero. Th^.s means that by staying inside ^a
these solid circles, dominance can be achieved at the frequencies at which they :n
occur while the l,3 entry in the compensator is zero. Thus, ir: designinn the 2,3

-,a

''' ^

,^

i^^

^' i s'ntY; ,t F i^' rrxat ,^;^ that ^,s emp^.oyCd is to falla1a tl ►e tri.ang].e., at low frcquenc^.es
_ - and :.t ^^;- .its ?de :. ', solid circles at the ha filter frequcnc.ics. ^If this is done, the
dc:^il;^t crF the 1, ^`^tCry will be simplified bccausc it will only be necessary to fit
- ' the ^:^: ^:y to the ,,-^^ca frequency triangles and have the function go to zero at higher
fa _ ^ r . Lcs. '

• ; this strategy, a Iag compensator was designed to fit the 2,3 entry as
^',",...~r. ^^cd previously. The compensator entry that was chosen is
1_• (s} _ -I29.4s -I9^r0.2 s
-
- ^23 .0365s + 1.

. . , fit, i':e same Lime, another lag compensator is fit to the solid circles in I'ig. 9,
^.: ^__ }..• :;^sRDIAD plat for the 1, 2 entry. This was chosen to be
:. ^ . Q3.27 ^ ^
_ ^ ^ x`12 (s} _ .1162s + I. '
- DeLin.ing this compensator as K^(s) with all the other off--diagonal entxies zero, the
c:^ • ralZ com
P ensation thus far ^s K
^(s} = K K1 2(s) .
0 I
-- 129.4s -- 19^Q.2
^^
_ . .03b5s + 1
^^IG^TAL PAGE ^g

. . i7^.27 4
1 . i3.G2s + ^..
.-,
Figs. 15 - 20 are the CAADI_^ plats of G(s)K3(s) using; the standard dominance analysis.
i^ The pints show that the first two calursns of the system are dominant at all freauen--
' ties since in 1?igs. I5- 18 the origin of each plot is contained by all solid circles
and e^culuded by all dashed circles. Fig. 19 shows that the strategy applied in the
design for the 2,3 entry was successful. To make the third column dominant, it is
now on^.y necessary to fit a compensator to the shape of the. solid circ3.es in Fig. I4

^•

E
F-^ `\.\
S
•^ ,

^'-;
t ^ f t^
s r^ ^

^^

s.,°, i l:j 1 i
¢a 's i
!^F ac ^^
o. ^ ^ 4 ^ ' ^
^n
¢^
s^ i

Is `` •

Z ^^ ..^ ,. ''`I ,I_ ^r ^ 15

^C e
m

_ ^

P ,'1 ^
• ^k

.•^ `

t.+ 1 ^// \,^.^


... ;` ^
;; ,
+lso.w -im.[o •la.t: •Ir. [a -1o. . at e^,u: s;.oa ^s, ^i iii. '-t.vf - O.x •^.cs - .o i O,C• f.lt u.cL •
a,. e E.0
IiEPL tRSf st -O.Y MEAL ERRS
' . ^,„^

• a i
. .. f

Fag. xS. G (s) *1:3(s), 2,1 Entx'y Fig. ].6. G{s)*h^{S}, 3,^ Entry

.+

•:
, , ,.
au highe r frequ encie s. The funct ion was
ond have it
.,5321' + 16,917
' ~.u ;!~,~
l 2
.0127 s + .19865 + 1.
The only chang e this has on the overa ll compe nsa'to r
is that the ze,ro in the 3,3
• fOO1L ..,1IJiI
co11!mn is replo tted using this
.entry is replac ed by this funct ion. When the third
I
, comp ensato r and stand, ard domin ance analy sis, Figs.
show tha,t the third colum n is now d<lmin ant at all
21 and 22, the CMliilIMI plots
frequ encie s. Thus, the sys,te m is

I ,

I
I
I.
'I:. .. ,
F:l.~. 18. G(S)*K3 (S), 3,2 En~ry
G,(S)*K3 (5), 1,2 Entry
I. Fig. 17.
=ir."~----------~--------~~~
..
I,
•• ..
·l. t


I
I
I.,. '

,I '.
,J F:l.g. 19. G(s)"'K3 (S), 1,3 Entry Fig. 20. G(S)*K3 (S), 2,3 Entry

,
. 1'~ •
'
~, .. •

...
_ ''V~)

tt
:..,
""

t: g^1^ =^ r
^ ^;'•^
^ ^^^.

w.
Sri i ' ' ^ ^ ^,V^^^{

f ^ f { ±^ ^ ^

^ ;^ '.

f r4
^ ^ ^^^ r ^ 1^^ ^^ 1^^^ .

e.
t. ^
` "^^%
^i.: ' ` ^

f ^ I

t A
s.ec .,.a ^-i.:.r^ ' -r.e^ -e,ze r.::
1lE[iL lGSY
Y
111 ^

^.ea z:ac
'

im <ia a.e

' Ffg. 2I. G(s)^K{s), 1,3 Entry F^.g. 22. G(s}^K(s), 2,3 Entry

fi, CQNChUSIQNS ^^^ ^^


j o^ ^oo^ ^va^^.t^
The graphical CARDTr^,D method described ir, this paper has baen effective on the Thera
Problem, The authors' e^;perience indicates that it is an easily learned design aid
which can.be quite helpful in achiet • ing dominance for realistic plants. A special
advantage of the Cr^RDIAD approach leas xn the way in whx.ch ^t prov^.des ^nslght to _ ;^
the designer. The plots indicate whether or not it will be possible to achieve F^

dominance with sinple, lead-lag compensators. Examples up to this tir,^e suggest _that, ^.
over the useful bandwidth, simple compensators are often succassful in this regard.
:, .1
It should be noted that this paper illustrates only compensator selection for dor.^i- i',^
Nance. Completion of the design is by classical means, For an example, see [5]. ^^
'{ particular note is the fact that compensator denominators having right half plane ?^
^^
zeros do not necessazily land to unstable controllers. This may also be seen in [3]. '^;^:
;t:.
Continued research on this class of graphical., interactive methods is in progress. ^,
^i
:, ,
:^
,t.
7. AC1^Iq^dI^EDGAIENT
14

^`h$s work was supportFd i.n part by* the National Science Foundation under Grant E^tiG ^_±
'^--22322 and in part by the National Aeronautics and Space Administration under ^p
Grant NSG 3048. -
a
..
S. fiEFEr.ENCES r
. ^
:i X
1. H. i^. Rosenbrock, Com uter Aided Control S y stem Deli. n, New York: Academic ^:
Press, ].974. ^ ;^ 5

2. R. M. Schafer, "A Graphical Approach tv 5ystem Dvmxnance," M.S. Thasis, Depart- ^^,
_^.
meet of Electrical Engineering, EJnivexsity of Notre Dame, Notre Dame, Indians, y...
`'
;3
:,;' I^ay 1977.

3, J. F. Sellers and C. J. Daniele, Y 'DXNGEN--A ^'ragram for Calculating 5Ceady-State ^i

^.: .. . .. .
^•.. ^rsnd Tran.^3et*: f'erLor^Etancc of 'ittrbojet and Turba^at^ F.n^ines,' r NASA Te}tctt^.cal
• Motes, NASA ' :+" ^}-- 7901, Sdas }tinBton, D.C., April 1975.

--_ ^.. ?,. ^. Ge}r sex, `"NASA Preliminary P^cport on DYGISBCD," April 1976.

"...^'•i^•. ri. Schafer, R. R. Gcjji, F. W. ^}oppner, jd. E. LonBenba}:er, and M. K. Safn,


I . . `Frequency Da;itai,n Catarieiisatxon of D^..GEN Turbofan Fn^inc ,iodel," Prof. .Toir^t
- tutomati.c Control Conlcrencc, pp. lOJ.3-101$, ,3urte x.977.

:_

,f_ .-

__

^,

_. -

^^

^ _
'

^•

^. ^ ^ .
... ^

.^`
:^ ^,.

= `'

,; ^' ^ -
_. ; . _

'; ,^
__^ „

You might also like