Вы находитесь на странице: 1из 696

Modem Navigation,

Guidance, and
Control Processing
b y Ching-Fang L i n

Modeling, Design, Analysis, Simulation, and Evaluation ( M D A S E )

Modern Navigation, Guidance, and Control Processing
Advanced Control Systems Design
Integrated, Adaptive, and Intelligent Navigation, Guidance, and Control Systems Design
Digital Navigation, Guidance, and Control Systems Design
Modem Navigation,
Guidance, and
Control Processing

Ching-Fang Lin
American Gh'C Corporation

gig Prentice Hall, Englewood Cliffs, New Jersey 07632

L i b r a r y of C o n g r e s s Catalo~lng-ln-Publlcatlon D a t a

Lln. Chlng-Fang.
M o d e r n n a v l g a t l o n , guldance. and c o n t r o l processing / by C h l n g
-Fang Lln.
p. CI. -- ( S e r l e s In a d v a n c e d n a v l g a t l o n , guldance. a n d
c o n t ~ o l . a n d t h e l r a p p l l c a t l o n s : b k . 2)
I n c l u d e s b l b l l o g r a p h t c a l r e f e r e n c e s and index.
I S B N 0-13-596230-7
1. F l l g h t c o n t r o l . 2. G u i d e d ntsslles--Control s y s t e m s .
I. T l t l e . 11. S e r l e s .
TL589.4.L55 1991
629.1--0C20 90-43009

Editoriallproduction supervision
and interior design: Brendan M . Stewart
Cover design: Bruce Kenselaar
Manufacturing buyers: Kelly Behr and Susan Brunke
Acquisitions editors: Bernard Goodwin and Michael Hays

O 1991 by Prentice-Hall, Inc.

A Simpn & Schuster Company
Englewood Cliffs. New Jersey 07632

Thii book can be made available to businesses and

organizations at a special discount when ordered in large
quantities. For more information, please eontad: Prentice-
Hall, Inc., Special Sales and Markets, College Division,
Englewood Cliffs, NJ 07632

All rights reserved. N o part of this book may be

reproduced, in any form or by any means,
without permission in writing from the publisher

Printed in the United States of America

1 0 9 8 7 6 5 4 3 2 1

ISBN 0-13-59b230-7

Prenticc-Hall International (UK) Limited, Lotidon

Prentice-Hall of Australia Pty. Limited, Sydney
Prentice-Hall Canada Inc., Toronto
Prenticc-Hall Hispanoamericana. S.A.. Mexico
Prentice-Hall of India Private Limited, ,Vru, Drlhi
Prentice-Ha11 ofJapan. Inc.. Tokyo
Simon & Schuster Asia Pte. Ltd.. Singapore
Editora Prenticc-Hall do Brasil. Ltda.. Rio de Jatreiro
To my family
for their love, understanding,
and support throtrghout.
Series Foreword
1 Introduction
1.1 Overview, 1
1.2 Outline/Scope, 8

2 Modeling-Design-Analysis-Simulation-
Evaluation (MDASE) of NGC
2.1 LinearlNonlinear Intercept NGC
System, 13
2.1.1 LinearlNonlinear Intercept N G C
Processing, 14

2.1.2 Modeling and Simulation, 18

2.1.3 Guidance System Classification, 21
2.1.4 Flight Control System (FCS) and FCS
Sensing, 22
2.2 Target Signal Processing, 23
2.2.1 Targeting, 30
2.2.2 Kinematic/Relative Geometry, 34
2.2.3 Targeting Sensor Dynamics, 37
2.3 N G C System Design and Analysis, 45
2.3.1 Guidance Filtering/Processing, 45
2.3.2 N G C Stability and Pevfovmance
Analysis, 52
2.4 Target Tracking State Modeling, 68
2.4.1 Target Noise and Target Maneuver
Modeling, 68
2.4.2 Two-Dimensional Target Tracking
State Modeling, 72
2.4.3 Three-Dimensional lntercept State
Modeling, 74

3 Modem Multivariable Control Analysis

Singular Value Analysis, 78
Sensitivity and Complementary Sensitivity
Functions, 82
Design Requirements, 84
Structured Singular Value, 86
General Robustness Analysis, 89
Robustness of Real Perturbations, 90
3.6.1 State-Space Model for Additive
Uncertainty, 91
3.6.2 Slate-Spare Sing~rlarI/allres, 92
3.6.3 Root Loctrs, 92
3.6.4 A Monte Carlo Comnputation, 94
3.6.5 A Monte Carlo Analysis.for Second-
Order Systetns, 95
3.6.6 Stability Margin Comttp~rtatiot~,
Monte Carlo Analysis, 100
Covariance Analysis, 102
Adjoint Method, 106
3.9.1 Adjoint P/~ilosophy,106
3.9.2 Applications, 108

3.10 Statistical Lzncarization, 11 1

3.10.1 l'c~clrtriq~rc~sand Tools for Statistical
L~trcartzation, 116
3.10.2 Statistical Litlearization with Adjoint
hlrtlrod, 1 18
3.10.3 Applicatiot~s, 120
3.11 Qualitativc Comparison, 121
3.12 Other l'erformancc Analysts, 124

4 Modern Filtering and Estimation

4.1 Linear Minimum Variance Estimation:
Kalman Filter, 128
4.1.1 Discrete Kalman Filter, 129
4.1.2 Contirl~rortsKalmatr Filter, 129
4.2 Nonlinear Filtering, 133
4.2.1 Estetrded Kalmaw Filter (EKF), 133
4.2.2 Sratisrical Lit~earization
Teclzrtiqrre, 135
4.2.3 ~tfrrltipleModel Estirnadon, 137
4.3 Prediction and Smoothing, 138
4.3.1 Prediction, 139
4.3.2 Strroothirrg, 140
4 Kalman Filter Design and Performance
Analysis, 140
4.4.1 Kalrnar~Filter Desigtl Process, 142
4.4.2 Selectiort of Noise Itltensity
Matrices, 142
4.4.3 Error Analysis, 146
4.5 Operational Considerations, 148
4.5.1 Filter Pevfarmance, 149
4.5.2 Computational Reqtrirements, 149
4.5.3 Absence o f A Priori Information, 153
4.5.4 Filter Divergence, 153
4.5.5 &lodeling Process Noise and
Biases, 156
4.6 Combined Complementary/Kalman Filter
Approach to Estimator Design, 158
4.6.1 First-Order Complert~entary/Kalman
Filter, 158
4.6.2 Second-Order Complementary/Kalman
Filter, 163
4.6.3 Third-Order Complernentary/Kalman
Filter, 170

5 Inertial Navigation 176

5.1 Introduction, 176
5.2 Common Requirements for Inertial
Navigators, 181
5.3 Navigation Computation and Error
Modeling, 184
5.3.1 Coordinate Systems, 184
5.3.2 Position and Velocity Generation, 186
5.3.3 Data Processing, 188
5.3.4 IA'S Error Analysis and
Modeling, 188
5.4 Gimballed Inertial Navigation System
(INS), 192
5.4.1 Ginzbal Mechanism, 192
'5.4.2 Iitertial Setisors on the Stable
PlatJorrn, 193
5.4.3 Plarfornt Aligrlriient Modes, 193
5.4.4 nhvigatioit Mode, 194
5.4.5 Systein Ititernal arid External
Integaces, 195
5.4.6 Navigatioii Mechanizatiorz and Error
Model, 195
5.5 Strapdown Inertial Navigation System
(INS), 199
5.5.1 Generalized Strapdowti
Mechai~izatiori, 199
5.5.2 Strapdouwi Navixatiori ~ o r n ~ u t e 201r,
5.5.3 Strapdown Coniputer
Requireineiits, 201
5.5.4 Strapdowrl Error Model and
Attalysis, 204
5.5.5 Operation Flow Diagram, 208
5.6 Comparison and Analysis of Gimbal Versus
Strapdown, 209
5.6.1 Fcatitrc Coiiiparisoii, 209
5.6.2 h'ai~igatiori Errors Cori~parisoil,209
5.7 External Navigation Aids, 213
5.7.1 Aided liierti~l,Vavkatioir
Meckatiizatior?, 213
5.7.2 Global Positioi~iiigSysterri ( G P S ) , 215
5.7.3 Tactical A i r hlavigatiorl
( T A C A N ) , 277
5.7.4 L o r ~ gRange Navigation
( L O R A N ) , 220

5.7.5 Terrain Contour Matching

(TEHCO.M), 220
5.7.6 Doppler Radar, 225
5.7.7 Star Trackers, 225
5.7.8 Kaln~anFiltering, 226
5.7.9 Kaltnon Filtering Performance, 228
5.8 Integrated Inertial Navigation System
(IINS). 229
5.8.1 I t ~ t e ~ ~ r a rSensittcq/Flight
ed Control
Reference Systrtn (ISFCRS), 231
5.8.2 it~te~qrated Sensory Subsystetn
( I S S ) , 231
5.8.3 Itltecqrared Inertial Sensing Assembly
( I I S A ) , 232
5.8.4 Helicopter Integrated Inertial
h'avigation Systern (HIINS), 241
5.8.5 integrated Missile Guidance
Systerns, 245

6 Guidance Processing
6.1 Guidance Processors, 252
6.2 Guidance Mission and Performance, 268
6.2.1 Guidance Pevformance, 269
6.2.2 Pltases of Flight, 279
6.2.3 Operation, 280
6.3 Multiple Mode Guidance Modeling, 298
6.3.1 M~dcourseGuidance, 298
6.3.2 Terminal Guidance, 302
6.3.3 Error Analysis Model
Development, 308
6.4 Guidance Algorithm, 310
6.4.1 Preset Guidance, 310
6.4.2 Direct Guidance Methods, 312
6.5 Guidance Law, 347
6.5.1 LOS An3le Guidance, 348
6.5.2 LOS Rate Guidance, 348
6.5.3 Sensitivity and Cornparison of
G~ridarrceLaws, 373
6.6 Single-Mode, Dual-Mode, and Multimode
Guidance, 379
6.6.1 Single-Mode Guidance, 379
6.6.2 Dual-Mode Guidance, 380
6.6.3 Multitnode Guidance Applications, 384

6.7 Defense and Offense Systems, 392

6.7.1 Performance Paratneters, 392
6.7.2 Low-Altitude Air Defense
Systems, 397
6.8 Future Guidance Processing, 404
6.8.1 Areas for Teclztiological Advartces in
Signal Processitig, 405
6.8.2 Future Signal Processing for Missile
Guidatlce, 407

7 Navigation and Guidance Filtering


7.1 Target State Estimation, 414

7.1.1 Target Tracking Filter Suiizmary, 416
7.1.2 The Wiener Filter, 420
7.1.3 Kalmatl Filter, 420
7.1.4 The Sitiiplified Kalnlari Filter, 420
7.1.5 Alpha-Beta-Canima Filter, 421
7.1.6 Modified Maxitn~rm-Likelihood
Filter, 424
7.1.7 Two-Point Extrapolator, 425
7.1.8 Coriiparisori of Targct Trackirrg
Filters, 425
7.2 Practical Navigation and Guidance Filtcr
Design, 427
7.2.1 Guidance Tracking Filtcr, 427
7.2.2 Navigation and Guidance Filteringfor
Position Estiriiatc, 430
7.2.3 Navigatiorr arid Cuidatice Filteritlgfor
Position and Velocity Estitilate, 431
7.2.4 Navigation atrd Cuidarrcc Filterit~gfor
Position, I/clociry, and Accrleratiot~
Estiriiatc, 434
7.2.5 Advatrccd Guidance Filter, 440
7.3 Radar Tracking, 442
7.4 Spacccraft Attitudc Estimation, 444
7.5 Advanccd Navigation System Design, 445
7.5.1 Global Positioning Systetir (GPS)
7.q.2 Integrated CPSIIR'S, 457

8 Advanced Guidance System Design

8.1 Advanced Guidancc Laws, 468

8.1.1 O~itirttnlCrridatrce Law Survey, 469
8.1.2 Atrnlyticnl Solrrtiotr of Optirnal Filters
and Optitnal G~ridatrceLaw, 474
8.2 Complenicntary/Kalman Filtered
Proportional Navigation: Biased PNG and
Complcmcntary I'NG, 481
8.2.1 Corttliitted Seeker-Czridatrce Filtering it1
a Cottrplettret~taryFilter, 483
8.2.2 Biased PAYC ( B P N G ) Algorithm, 486
8.2.3 Cort~plettret~tary PNG (CPNG)
Alyoritl~trr, 492
8.2.4 Terrtr inal G~ridatlceSystem
Analysis, 501
8.3 Other Terminal Guidance Laws, 510
8.4 Radome Error Calibration and
Compensation, 517
8.4.1 Radorne Error Cotnpensntion, 51 7
8.4.2 Gtridowce Perforrnat~ceAnalysis with
In-Fl!qht Radotne Error
Calibratiotr, 520
8.4.3 Des[gtt Equations for P N C with
Parasitic Feedback, 528
8.5 Command Versus Semiactive Homing
Guidance System Design and Analysis, 544
8.5.1 Modeling, 544
8.5.2 Miss-Distance Analysis for Command
Guidance, 550
8.5.3 Analysis of Optimal Command
Ccridance Versus Optimal Semiactive
Homirlg C~ridance,557
8.6 Analytical S o l ~ ~ t i oonf Optimal Trajectory
Shaping for Combined Midcourse and
Terminal Guidance, 562
8.6.1 Optitnal Trajectory Sllaping
Guidance, 562
8.6.2 Problem Formulation, 563
8.6.3 Attalytic Optimal Gtridance Law, 566
8.6.4 Real-Time lmplemetztation and
Pevformance, 578
8.6.5 Discussiotls, 583
Series Foreword
The role played by modern navigation, guidance, and control (NGC) in the de-
velopment and advancement of such areas as commercial and military aviation, to
name only two, has continually expanded since its earliest inception in the 1950's.
As this field began to grow and take on added importance, many books were written
dealing mainly with the theoretical aspects of NGC, but most of these were confined
to the earlier years of N G C development. Currently, although N G C system ap-
plications continue to take on an ever-increasing importance, the availability of
reference books, especially textbooks suitable for graduate level and advanced un-
dergraduate students as well as those who practice in the field, has not kept pace.
This series emphasizing NGC systems and their applications is long overdue; in
fact, it has been 30 years since such a series dealing with N G C systems has been
written and available to the academic and ~rofessionalcommunities. Moreover. it

is the first ever such series to thoroughly discuss the advanced control system design
(modern multivariable control analysis; robust control; estimation; adaptive control;
nonlinear control; intelligent control; etc.). It comes at a time when concern over
issues such as the status of education and the decreasing number of trained, qualified
professionals in this country is a t an all-time high. It is against such a background
that the present series was conceived to assess state-of-the-art systems and control
theories, and engineering applications of advanced N G C systems. Another purpose
of the series is to develop future research agenda and at the same time encourage
xuiii Series Foreword

discussions in those areas that do not always find the systems and control community
in complete agreement.
The series provides a comprehensive coverage of the latest N G C technology
as follows. The first book begins by introducing the various applications of N G C
systems, after which it provides a thorough, fundamental treatment of what is con-
sidered the five most important stages in N G C system development: modeling,
design, analysis, simulation, and evaluation (MDASE). The second book in the
series takes up the subject of advanced estimation and guidance systems design, as
well as N G C processing. The third book is concerned with the subject of advanced
control system design, with particular emphasis placed on the topic of flight control
system (FCS) design. The topics that constitute the fourth book include integrated,
adaptive, and intelligent N G C systems design, while the fifth book is devoted com-
pletely to digital N G C systems design. Although most of the material in these five
books is self-contained, there is a natural progression in the series as a whole toward
more advanced topics. For example, much of the material in the second book actually
serves as a prelude to the third, fourth, and fifth books.
These books are the result of several years of experience gained on the part of
the authorleditor both as a professor at the university level and as a practitioner in
the field. It is believed that this series will provide invaluable insight and instruction
to students, mainly at the graduate level but also to advanced urldcrgraduate students,
as well as to those engineers who work directly or indirectly in the field of N G C
system design and applications. In addition, this series is intended to provide both
engineers and managers with the advanced N G C knowledge and concepts necessary
to make correct decisions concerning the best N G C system design in a particular
It is very likely that few people who labor in any scientific discipline are unaware
of the contributions of advanced navigation, guidance, and control (NGC) theory
to aerospace-related programs. It is, however, equally unlikely that many are aware
of the dramatic impact of this field on such diverse areas as medicine, industrial
manufacturing, energy management, and chemical engineering. While its broad
range of applications would at first appear to indicate that advanced N G C theory
is enjoying an immense popularity in scientific and academic settings in general,
this unfortunately turns out not to be the case. It is felt by many experts that many
of those in the aerospace field in particular either are content to rest on the laurels
surrounding the success of NGC theory developed in the 1950's and 1960's or have
become so conservative in their design philosophies as to be unduly apprehensive
about using advanced N G C theory. The latter appears to be especially true in the
area of aviation.
The author is quick to point out, however, that the NGC field itself is some-
what responsible for many of the misperceptions on the part of those w h o are not
convinced of the usefulness of advanced N G C theory. More than a mere shadow
of doubt has been cast on the usefulness of this theory as a result of its having taken
xx Preface

a much too mathematically-oriented turn almost immediately after the theory was
first applied in the solution of practical problems. It is the author's opinion that,
while N G C theorv is built around a rather beautiful framework of mathematics. its
primary emphasis'must nonetheless always bk placed on solving engineering prob-
lems of great practical importance. N G C technology has always been the focal point
of aerospace engineering and automation research and development. A combination
of theoretical concepts, the rapid evolution of computer and microelectronics tech-
nology, and the continued refinement of sensor and actuator technology has con-
tributed to its advances. This book examines the role of modern N G C processing
in the design of advanced N G C systems.
This volume places major emphasis on the practical applications of advanced
NGC systems, treating the subject more from an engineering than a mathematical
perspective. Nevertheless, theoretical and mathematical concepts are introduced and
adequately developed to make the book a self-sufficient source of instruction for
readers. The intent of this book is to enable readers to achieve a level of competence
that will permit their participation in the practical applications of modeling, design,
analysis, simulation, and evaluation (MDASE) to advanced N G C systems. The book
presents basic as well as advanced algorithms. A wide range of examples culled from
various applications are provided to meet the needs of the different levels and types
of readers, extending from issues requiring only a rudimentary knowledge to those
involving avant-garde research. Morever, problems in the text span from those that
concern only N G C to those that are interdisciplinary, and ultimately to those that
encompass the entire systems and control field.
An outline of the topics presented in this book is given in Fig. 1. Following
the Introduction (Chap. I ) , the text is organized according to five principal topics:
MDASE of N G C Processing (Chap. 2), Modern Multivariablc Control Analysis
(Chap. 3), Design Algorithms for Advanced NGC Systems Design (Chap. 4). Fun-
damentals of N G C Processing (Chaps. 5 and 6), and Advanced N G C Systems De-
sign (Chaps. 7 and 8). Each of these is in turn divided into a number of subtopics
that discuss the relevant theories, algorithms, and computing tools related to their
applications in advanced NGC systems. Referring to the outline organizing the five
principal topics covered in this book, only those subtopics that are connected by
solid lines are treated in this book. Those subtopics that are connected by broken
lines are treated specifically in the books referenced under them.
The numerous examples that are included in this book are supplemented by
a liberal use of rcferenccs, thus making it easier for the reader to get access to a
tremendous body of literature it] this field. As noted by one rcvicwer, there are n o
comparable books currently on the market that prcscnt in a usable format such a
complete collection of practical tools that are applicable to real world problems.
Moreover, the organization of thc material coupled with comprchensive examples
make this book well suited to self-teaching, bridging the gap betwccn thc theoretical
and the practical.


(MDASE) of NGC Processing

----- 1
Hodern Multivariable Performance
Control Analysis
Modern Filtering
Advanced Adaptive NGC
Design and Estimation Multivariable I Systems: Detection and Neural-Network
for Advanced NGC Techniques Control Design I and Identification Approaches to
Systems Design (Chap. 4) Techniques (Book 4) Systems Control
(Book 4)
- - I- - - - 1
Fundamentals of Guidance Processing Flight Control
NGC Processing Processing
(Bogk 3)

Navigation and Guidance Filtering Design

(Chap. 7)
Advanced Guidance Systems Design

Design +
Advanced Flight Control System Design
(Book 3)

r 7 - - -C - - - 1
Integrated, Adaptive, and Digital NGC Systems Design
Intelligent NGC Systems Design (Book 5)
(Book 4)

Figure 1. Modern Navigation, Guidance, and Control (NGC) Processing

The number of people who have lent assistance in one way or another to this book
is too great to allow me to thank each one individually. 1 am indebted to each person
who has contributed to making this project a success, but I feel that I must ac-
knowledge a few individuals by name for their support and effort during the count-
less hours spent on this project. My first acknowledgement goes out to my staff at
American GNC Corporation (AGNC) who have assisted me as follows: Chun Yang
and Jerry Juang have provided important technical input, while Jim Wright has
spent many hours editing the manuscript. Special thanks are due to Kylie Hsu and
Janet Young for their involvement in designing, planning, coordinating, and pre-
paring the entire manuscript, including word processing and numerous original
pieces of computer and manual artwork. I am also grateful to William R. Yueh for
his technical advice.
Several of my colleagues and friends have provided invaluable technical as-
sistance. I specifically want to acknowledge William H. Gilbert of Martin Marietta
Electronic Systems who, through his ardent support and encouragement of this
project, made this impossible task appear less so at times. I am deeply touched by his
unselfish giving of his time in the overall guidance and input to this project. I am
xxiv Acknowledgments

am also indebted to J. Stanley Ausman of Litton Guidance and Control Systems

Division who has greatly improved the quality of Chapter 5. Dr. Ausman is an
authority on the subject of this chapter and has co-authored a book on Inertial
Guidance (Wiley, 1962). I would also like to thank him for granting me permission
to adopt his lecture notes in Tactical Aircraft Weapon Delivery Systems.
I also want to express my gratitude to Gary Hewer of Naval Weapons Center,
Hsi-Han Yeh of Wright-Patterson Flight Dynamics Lab., and Keqin Gu of Southern
Illinois University for their painstaking effort in reviewing Chapter 3. T. Sen Lee
of MIT Lincoln Lab., is also appreciated for his effort in reviewing Chapters 4
and 5.
A very special note of thanks goes out to Bernard Goodwin, Vice-President,
Professional and Technical Reference Publishing of Prentice Hall, who has taken
on the task of publishing this book and has continued to support enthusiastically
throughout this project. I would also like to extend my appreciation to Michael
Hays, Executive Editor and Assistant Vice-President, Professional and Technical
Reference Publishing of Prentice Hall, for his constructive advice and his special
interest in this project. Also deserving special mention is Brendan Stewart, Pro-
duction Editor of Prentice Hall, for his excellent job throughout the production
process. I am equally obliged to other Prentice Hall personnel who have participated
behind the scenes in the different stages of this book project.
Finally, I would like to acknowledge those individuals and organizations that
have permitted me to reprintladapt portions of their outstanding work in order to
make this book a well-rounded source of information. They include (in alphabetical
order): AACC, AGARD, AIAA, Arthur Gelb of TASC, Frederick W. Hardy of
Hughes Missile Systems Group, Robert J. Heaston of GACIAC, Kaz Hiroshige
formerly of General Dynamics Convair Division, IEEE, Johns Hopkins APL Tcch-
nical Digest, Robert J. Kelly of Allied-Signal Aerospace Co., Litton. James A.
McLean of the U.S. Army Missile Command, Pergamon I'rcss, Rockwell Inter-
national Collins Avionics Division, and SCS International.
Modern Navigation,
Guidance, and
Control Processing
This chapter provides an overview of advanced navigation, guidance, and control
(NGC) design. The text is oriented to the applied rather than the theoretical aspects
of the subject matter. Although advanced techniques are discussed, the contents are
presented in such a manner as to also provide a simple and interesting picture of
the central issues underlying both classical and advanced control theory and the
practice of the NGC modeling, design, analysis, simulation, and evaluation
(MDASE) process.


The objective of this book is to provide both engineers and managers with the
advanced NGC knowledge and concepts necessary to make sensible decisions con-
cerning the best NGC system design in a particular situation. It will hopefully serve
as a useful source of information for NGC systems designers by providing them
with ideas for the solutions of current problems and future designs, as well as in-
formation about the problems encountered with microprocessor-based systems, mi-
croelectronics, aerodynamics, structures, propulsion, sensing, actuation, target ac-
quisition, and weapon systems. In an attempt to achieve such an objective, this book
demonstrates practical tools that are realistically applicable in the work area.
2 Introduction Chap. 1

The development and applications of present-day systems and control theory

were spurred on by the challenge of unsolved aerospace problems, especially by the
series of events that has occurred since the late 1950s. Therefore, it is beneficial to
review the development of systems and control theory since that time. The jolting
success of the Soviet Union's satellite technology during that time inspired the
United States to excel in aerospace technology, thus giving birth to an entire new
generation of support for the field of systems and control. This in turn led to success
with regard to solving urgent aerospace problems. The emergence of the Apollo
program in the 1960s restored confidence in systems and control research and pro-
vided opportunities for conceptualized systems and control theories to be trans-
formed into actual practical NGC system designs. Among the more well-established
concepts that found great applicability to solving real engineering/control problems
in the 1960s are the recursive minimum variance estimator, often referred to as the
Kalman-Bucy filter, the LQG technique, the time domain concepts related to the
fields of linear algebra and probability, and the Kelley-Bryson variational optimi-
zation procedure. Research into this last area proved particularly fruitful, as it re-
sulted in the successful design of optimal trajectories for several space missions. In
particular, the Apollo program and space shuttle flights that followed benefitted
greatly from the application of optimal control theory. The contributions of Kalnlan
filtering to the Apollo program represent yet another nlilestone for the emerging
control theory of the late 1950s and early 1960s. The Kalman filter was implemented
using a square-root algorithm in the measurement of stars taken by the astronauts
with the aid of a sextant. Both optimal control theory and Kalman filtering provide
tangible examples of the cruc~alrole played by control science in important programs
such as the Apollo project.
As a result of its early success in solving mainly aerospace-related problems,
NGC theory soon found application in such diverse areas as mcdicine, industrial
manufacturing, and energy management. For example, two French researchers werc
able to formulate the problem of treating certain cerebral edemas or malignant brain
tumors by the simultaneous administration of vasopressin and cortisone into a non-
linear rnultivariable control problem. The treatment resulting from their work is
currently used in the neurosurgery clinic of the Hbpital de la Pitie in Paris [Fleming,
19881. Although examples of this type demonstrate the profound impact of control
theory in general, and are not altogether rare, it remains true that the majority of
progress in N G C research and development continues to find its greatest application
in aerospace.
The preceding historical account is provided to illustrate the emergence of
advanced NGC systems and their applications. The emergence of the various N G C
methodologies contributed significantly to progress achieved in the development of
state-of-the-art systems and control theories in the 1960s. Unfortunately, it was not
long before the development of the NGC system became too mathematically ori-
ented in spite of the pressing need to solve many remaining practical problems. It
would appear reasonable that the obligation of the NGC engineer should be first
to commit to solving practical NGC problems [Ho, 19871. In the course of carrying
Sec. 1.1 Overview 3

out such an obligation, thc cngincer will oftcn discover relevant theorics that surface
spontaneously. While theoretical discovcries of this type often allow thc engineer -
to conduct furthcr systcm research, it remains impcrativc for professional engineers
to commit themselves to solving what are considered to bc rcal-world problems.
Very oftcn, practicing engineers encounter difficulty in just understanding the
theorics presented at conferences and in journals, not to mention applying them.
What is at issue here is not a debate of the merits of theory versus those of applied
technology. Rathcr, thc issue is onc of determining an effcctivc way to build a
practical N G C systcm; understanding an csotcric N G C theory is of sccondary im-
portance. While there is currently an abundance of N G C problems waiting to be
solved, along with a whole assortment of tools capable of attacking these problems,
diminished funding in this area hinders further development of N G C technology.
In order to recapture the strong financial support for NGC research of the 1960s,
the N G C engineering community must first be seriously committed to attacking
those immediate ~racticalNGC ~ r o b l e m sthat remain outstanding. " In the absence
of such a first step, any practical system and control theory is not likely to be
developed, as evidenced by the slow progress of N G C technology in aerospace. T o
highlight this fact, N G C tcchnologv used in the Apollo project was later directly
applicd to the space shuttle with hardly any new development. As a result of the
diminishing research effort in the NGC discipline since the Apollo era, scarcely any
systems or control theories have proven valuable enough for practical use in N G C
The extent to which system and control theories have facilitated the clarifi-
cation of various N G C problems and issues cannot be overemphasized. However,
the N G C engineering field still faces several practical problems, one of which in-
volves the question of identifying the basic reason for the gap between theory and
practice and recommending a means to close this gap for the sake of advancement
in the N G C field. In light of this, the impetus behind this book is an attempt to
bridge the gap between theoretical arguments and the practical needs of the N G C
community by providing detailed discussions and practical examples of MDASE
in this area. A more detailed treatment of the MDASE cycle is given in Book 1 of
the series. Because experimental and theoretical aspects of NGC system research
both represent integral parts of systems and control science, it is essential to consider
N G C theories and physical systems together. If an N G C system of a flight vehicle
or robotic system is carefully modeled, it usually serves as a strong foundation for
the design process which, in turn, leads to analysis and simulation for validation.
The final step of the procedure involves evaluation. Thus, the MDASE cycle is an
important process in yielding practical results. Through demonstrations of the var-
ious techniques and examples in this book, the author hopes to shed light on and
unravel complicated systems and control theories, and to translate them into practice.
Before they can be brought into the operational stage, advanced N G C algo-
rithms must pass through a developmental stage involving several years of studies
and experiments. Through organizing sessions, panel discussions [Lin, 1983-1985;
Lin et al., 1986; Lin and Speyer, 19851, and workshops [Lin, 1987; Lin and Franklin,
4 Introduction Chap. 1

19901 on aerospace vehicle N G C systems for the Amcrican Control Conference and
the IEEE Conference on Decision and Control over a period of several years, the
author has been able to keep abreast of the current trends in N G C technology and
to understand its practical needs in different situations. Thus, the author is able to
provide a diverse range of advanced NGC techniques in the book which the reader
can then apply to various areas of NGC with competence, developing the best
intuitive design. Since all flight vehicles and robotic systems share the same N G C
techniques, the techniques presented in the book are applicable to these complex
dynamic systems. Applications presented throughout the book and techniques for
these applications are summarized in Book 1 of the series. In addition, appropriate
applications are embedded throughout the text to enhance the reader's understanding
of the techniques presented.

Importance of advanced NGC concepts and their impact. Ad-

vanced N G C systems, although designed so that machines might interact effectively
with the elements of nature, rely on human elements for their continued progress
and success. The beginning of this chapter highlights the role played by control
theory in the success of the Apollo project. The successful development of modern
fly-by-wire aircraft such as the F-16 fighter jet can also be credited largely to control
theory. And while modern manufacturing is becoming increasingly dependcnt on
highly accurate process and machine control, it is an unfortunate circumstance that
the continuous contributions that are made in N G C technology and the resulting
achievements go largely unappreciated by the broader scientific and engil~ccring
communities. The way it1 which N G C technology is isolated from the othcr tech-
nological disciplines in governmental, industrial, and academic settings is vcry ap-
parent [Speyer, 19871. Consequently, the acrospacc industry has not fully rcalizcd
the potctltial of N G C tech11010gy it1 acrospace systems, and progrcss in this area
especially continues to be delayed. This situation will continue unabated so long as
those involved fail to grasp both the importance of making progress in the area of
advanced N G C technology and the relation of this tcchnology to vital functions in
certain acrospace projects. For example, in certain commercial aircraft projects,
advocates of new methodologies for enhancing safety and ride quality performances
are for the most part ignored. While this can be partially attributed to thosc who
prefer conventional dcsign approaches such as the root locus or othcr classical tech-
niques, some of the blame must be laid on inadequately trained pcrsonncl, many
of whom lack a complcte understanding of thc basic principles of modcrn and ad-
vanced control tcchniqucs. At~othcrreason for this is that thc designers of these
commercial transports dcpcnd entirely on the aerodynamic dcsign to improve ve-
hicle stability and efficiency, and totally ignorc the merits of advanccd NGC systems
Those who are usually quick to criticize techt~ologicaladvance have always
been skeptical about the complex and unreliabic nature of new technology, doubtirlg
the wisdom of investing so much trust in the capabilities of ncw tcchnology. How-
ever, competition and survival dcpend on technological advatlcc. It xvould bc a grave
Sec. 1.1 Overview 5

error to halt technological advance just becausc some inconvcnicnces arise at the
beginning of each new era. Despite the fact that any criticisms can generally be
found to be true in the short term, the long-term positive contributions that advanced
technology has to offer should be emphasized. After a new era of technology has
set in, new systems perform more complex tasks and become less expensive than
their archaic counterparts. The reliability factor, that is, the mean time between
failures, also improves several times over that of older technology of similar com-
plexity. The cost does not increase as much as the cr~ticsargue, as evidenced in the
recent shift between the technologies of the 1970s and 1980s. Spccifically, in the
weapon systems field, the gains have measured up to the cost. Therefore, critics
can discard the notion that new technological concepts are not important [Deitch-
man, 19871.
In the realm of aerospace, unlike other technologies such as aerodynamics,
structures, and propulsion, advanced NGC technology has yet to distinguish itself
as an essential discipline. There are two main factors that account for this predica-
ment. First, N G C technology in aircraft development has been traditionally ac-
corded a secondary role. This is because many of those who direct large aerospace
programs are unacquainted with the important role of advanced NGC technology
~ ~

and its impact on aerospace. Second, no hardware or other tangible entities are
produced by N G C technology; whereas, shapes are formed by aerodynamics, de-
livered by structures, and driven by propulsion [Speyer, 19871. Therefore, NGC
technology has barely been utilized in any of the aerospace vehicle design processes
with the exception of the Apollo project. According to the author's knowledge,
only a few newly developed European commercial airplanes employ advanced con-
trol laws. Faulty instrumentation and inadequate computational capabilities rein-
forced the negative view of NGC technology in the past, but even with the rapid
improvements existing in the present this view still persists.
T o reap the benefits of advanced technology, timely investments in the pursuit
thereof are crucial. Most people cite only the recent progress in military technology.
However, new technology must be applied to the commercial arena as well if eco-
nomic dominance is to be possible. For example, airlines tend to purchase ultra-
modern airplanes from companies employing more advanced technical sophistica-
tion. From this viewpoint, Airbus, the European company, is a commercial success
story. It is actively involved in investigating N G C issues and incorporating new
N G C technology into its products as well as its human resources. Thus, research
and development (R&D) of advanced N G C concepts cannot be stressed enough.
It is important that advanced NGC design concepts enter early in the design
phase. The earlier engineering errors are detected, the less severe future costs and
schedule delays will be. An example of this type of situation is brought into focus
in the now famous article entitled, "Probing Boeing's Crossed Connections" by
Karen Fitzgerald in the May 1989 issue of IEEE Spectrum, in which the author writes:
"By the end of Uanuary], the [FAA] had received 12 more reports of crossed wires,
not only on 757s, but on 737s, and not only on cargo fire-extinguishers, but on
engine fire extinguishers, and one case of crossed wires from engine temperature
sensors. . . . As of March 17, the FAA had received 66 more reports of bungled
6 Introduction Chap. 1

wiring and plumbing in fire-protection systems on all four classes U737, J747, J757,
J7671 of aircraft."' If it is possible to make such errors in those systems which are
by n o means as complicated as an NGC system, then errors in more complicated
systems will certainly result unless advanced system concepts coupled with the
proper quality control techniques are utilized.
As shown in Figure 1-la, the overall cost resulting from engineering error
increases dramatically as time progresses. Therefore, errors detected early can dras-
tically reduce the overall cost. Although engineering R&D design and analysis con-
stitutes only approximately 15 percent of the entire project cost, as high as 90 percent
of the overall cost of the project may rely on R&D results, as illustrated in Figure
1-lb. Therefore, inadequacy or errors in the R&D design and analysis due to in-

a) Cost

Detection of
)Engineering Error
1 2 (Normalized Year)
Cost Resulting From Engineering Error Increases
Dramatically As Time Progresses
Therefore Errors Detected Early Can Drastically
Cut Down The Overall Cost

b, Cost
OPS: Operating Cost
MAN: Manpower cost
R&D: Research 8: Development Cost

Life Cycle CDst = OPS t MAN t R&D

Figure 1-1 Cost lnvolvemcnt

Karen Fitzgerald. "Probing Bocing's Crosscd Connections." IEEE Spccfrurn. May 1989. p 33.
experience or other conservative factors can often lead to a drastic increase in the
overall time and cost. Quoting from the same article, the author writes, "A Boeing
engineer who asked that he not be identified said a too ambitious schedule for the
new 747-400 aircraft has caused wiring errors so extensive that a prototype had to
be completely rewired last year, a $1 million job. . . . In a technology that can
tolerate few errors, the crossed connections raise the specter of undetected errors in
other parts of aircraft. "'
Each new II&D assignment presents additional challenges which will tax the
skill and ingenuity of ~ ~ C - N G C engineer and will only be solved through
hard work, experimentation, and tests. Hence, each new assignment must be ana-
lyzed on its own merits, and past practices and methods must not be allowed to
stifle new ideas and concepts. Both the past and the present must never forget the
common denominator for all designs, which is the human element. Unfortunately,
in advanced technology, undue conservatism often hinders progress. Ironically, the
objective of any technology should be to go beyond the current state o f design
The NGC field, although relatively new, is currently in an excellent state. I t
is responsible or substantial contributions to engineering, science, and economics
as well as the standard of living in the United States and other countries. One of
the main contributions of this text is a new perspective of the N G C system whereby
the NGC system's interaction with other disciplines will establish the basis for truly
innovative problem formulations and methodologies. Methodologies used through-
out the book are practical and have been employed in the designing a n d testing of
NGC systems. In the past, classical designs were generally used for NGC of unaug-
mented dynamic systems. Thus, some designs covered in the book are inevitably
related to the classical approach. However, modern state-of-the-art N G C system
design for highly augmented dynamic systems, which is rapidly gaining popularity,
plays the principal role in the design methodologies used in the book. Integration
of N G C designs with various engineering automation and signal processing systems
are the ultimate goals of this advanced technology. The intention, therefore, of this
book and Books 3-5 of the series is to meet this challenge.

Advanced NGC systems. It is apparent that the usefulness of advanced

NGC systems has infiltrated the modern world. In the modern aerospace field, NGC
designers must be thoroughly equipped to handle N G C systems problems and able
to draw on a rather sophisticated knowledge base encompassing many diverse fields;
in particular, they must draw on their in-depth knowledge of modern dynamics
and MDASE techniques (see Book 1 of the series). They must also understand the
finer interactions between systems and components in their nonlinear operational
range. Thus, it is natural for them to appreciate the fact that an NGC system should
be designed as an integrated system. Additionally, they must consider the wide
range of applications of modern digital, analog, and hybrid computers. Very often,

* Ibid.. p 34-35.
8 Introduction Chap. 1

more sophisticated analytical and computational tools are required to properly model
N G C systems. The advantage of new technologies can be utilized to the fullest
extent through the advances of new mathematics, analysis, and computation. For
example, N G C system scientists and engineers are now more actively involved with
the designing and manufacturing of microprocessor/microelectronic chips and com-
puters in addition to interfacing with t h m d y n a m i c s , structure, propulsion, sens-
ing, actuation, target acquisition, and weapon systems.


As shown in the Preface, in Figure 1 labeled Modern Navigation, Guidance, and

Control (NGC) Processing, the five primary topics that make up this book are: (1)
MDASE of N G C processing (Chapter 2); (2) modern multivariable control analysis
(Chapter 3); (3) design algorithms for advanced N G C system design (Chapter 4);
(4) fundamentals of N G C processing (Chapters 5 and 6); and (5) advanced N G C
systems design (Chapters 7 and 8). O f the subtopics that appear alongside each of
these five main categories in the figure, only those that are connected by solid lines
are treated in this book. Those subtopics that are connected by broken lines are
treated specifically in the books referenced after them.

MDASE of NGC processing. For the purpose of continuity, the first of

the five main topics, as shown in Figure 1, deals with MDASE of N G C processing
in Chapter 2, highlighting some ofthe key features of Book 1. In particular, Chaprer
2 summarizes the major stages in the evolution of the design and development

Modem multivariable control analysis. Modern multivariable control

analysis, which is the second of the five primary topics as shown in Figure 1, is
mainly concerned with the analyses of robustness, performance, dynamics, and
stability. Chapter 3 deals specifically with robustness and performance analysis.
Robustness analysis is particularly important for examining multivariable control
systems design. Performance analysis, which is taken up in the later part of Chapter
3, includes the various statistical methods used in the analysis and synthesis of any
modern NGC system. The subject of stability and dynamics analysis, which is
threaded throughout this book, is one of the principal topics in Book 1 of the series.
The performance predicted by advanced control theory is rarcly achieved when
designing and developing an N G C system. A detailed analysis of N G C systems
must therefore include a treatment of the actual hardware equipment, citing those
characteristics that tend to limit system performance. The most difficult aspects of
analyzing N G C loops (at least, in missile applications) include, but are not lirnitcd
to, estimating their stability, determining their accuracy, and finding the trajcctory
of the pursuer vehicle along with the normal and lateral accclcrations necessary to
achieve that trajcctory for varying types of targct motion. Thc theoretical and cx-
perimental techniques for treating these problems are presented in Chapters 2 and
3 in this book, and in Book 1 of the series.

Design algorithms for advanced NGC systems. The methods that

together constitute the third of the five main topics listed in Figure 1, design al-
gorithms for advanced N G C systems design, arc listed with this category in the
same figure. The first of these, modern filtering and estimation techniques, is cov-
ered in Chapter 4, while advanced multivariablc control systems design techniques
are the subject of Book 3. The last two topics in this category, adaptive N G C systems
(detection and identification) and knowledge-based and neural network approaches
to system control are given special treatmcnt in Book 4 of the series.
I t is worthwhile at this point to examine the hierarchical and interactive re-
lationships among the different design techniques mentioned previously that make
up the design algorithms for advanced NGC systems. These techniques provide a
methodology for selecting successful N G C systems and integrating them according
to the mission requirements of a particular vehicle. The primary focus of this par-
ticular treatmcnt of design algorithms is to describe present-day components, sys-
tems, and synthesis techniques from the system-integration point of view. The goal
is to demonstrate how to analyze and select an N G C system to meet a set of per-
formance requirements when the vehicle maneuver and dynamic environment are
specified. The various techniques which make up the design algorithms are arranged
in hierarchical order in Figure 1-2a. From the figure, it can be seen that modern

Modern Filering and Estimation

Advanced Multivariable
(Chap. 4)
Adaptive NGC System
Control Design Techniques Design Twhniques
(Book3) CBook 4)
v v
Knowledge-Eased and Neural-Network
Awroacha to Svstems ConUol

Figure 1-2 Design Algorithms for Advanced NGC Systems (a) Hierarchical Re-
lationship (b) Advanced NGC Systems
10 Introduction Chap. 1

filtering and estimation techniques represent the starting point in the use of design
algorithms. Results from the application of these techniques are then used for both
advanced multivariable control and adaptive N G C system designs; however, adap-
tive N G C techniques are also used to iterate or improve upon the multivariable
control design. The results from the application of these last two techniques serve
as input to the design of a knowledge-based and neural network intelligent system.
The interactive relationships between components and techniques are depicted in
Figure 1-2b. This figure illustrates how the design algorithms are made more in-
telligent by incorporating a knowledge-based and neural network approach to N G C
system design. Ultimately, an N G C system is designed to accurately control the
outputs of a system whose dynamics contain significant uncertainties. This involves
the following fundamental processes and their associated algorithms: (1) modeling
ofthe system based on physical laws (see Book 1 ofthe series); (2) systern identification
based on experimental data (presented in Book 4 of the series); (3) signal processing
of the output by filtering, prediction, state estimation, and detection (modern fil-
tering and state estimation are presented in Chapter 4, while adaptive N G C systems
detection is presented in Book 4 of the series, and digital processing is presented in
Book 5 of the series); and (4) synthesizing the control input and applying it to the
system (advanced control theory is presented in Book 3).

NGC processing. Referring again to Figure 1 of the Preface, the fourth

of the five main subjects which comprise this book deals with N G C processing. As
indicated in the figure, N G C processing includes as subtopics navigation. guidance,
and flight control processing. The first of these, navigation processing, is covered
in Chapter 5 . Guidance processing is treated in Chapter 6, while flight control
processing is presented in Book 3 of the series. In the course of the last three decades,
N G C systems have evolved from a state in which they existed only in the imagi-
nation to their current state in which they are very much a part of reality. Both
military and commercial aviation, as well as other fields including automation and
manufacturing, owe a good deal of their success and technical advancement to the
parallel advancement in the area of N G C systems. It can readily be said that the
development of N G C systems and techniques, which in their ,--- infancy
- often led to
unpredictable accuracy and reliability performance, has been most successful, es-
pecially when examined in light of the often extreme environmental conditions to
which the components of these systems are typically subjected. Figure 1-3 gives a
block diagram representation of all the signal processing elements that are needed
to perform the functions listed in the previous paragraphs. Each block in the diagram
additionally lists the corresponding chapter number where the individual elements
are treated in more detail. It can be seen from the figure that N G C proccssing
involves several processors, each of which must function individually as well as in
unison with others.

Advanced NGC systems design. This constitutes the last of the five
main topics covered in this book. Looking once again at Figure 1 of the Preface, it
Feedback Sensing Information
Tracking Sensing Information


Control Signal State
Target Tracking Algorithms Reconstruction/Estimation Estimate
(Chap. 7) (Chap. 6) (Book 1 )
--- ------

I Advanced Vehicle Motion
Navigation Advanced
Guidance Estimator
System 4 System (Chap. 4; Books 1, 3,4, 5)
(Chap. 5) Control
(Chap. 8)
System Control Surface
(Books 1, 3,4, 5)
Guidance Commands Deflection Commands

Steering Signal and Attitude Information

Information Data Man-in-the-Loop Commands

(Controls & Displays)

Figure 1-3 NGC Processing

12 Inhroductlon Chap. 1

can be seen that this topic is concerned with the following subjects: navigation and
guidance filter design (Chapter 7); advanced guidance systems design (Chapter 8);
and advanced flight control system design (Book 3). These subjects prepare the
reader for the actual design of integrated, adaptive, and intelligent N G C systems
design, presented in Book 4 of the series, and of digital NGC systems design,
presented in Book 5 of the series.
Evaluation (MDASE)
of NGC Processing
It is brought to the reader's attention at this point that a very broad picture of the
modeling-design-analysis-simulation-evaluation (MDASE) cycle of the NGC sys-
tem has already been put forth previously in Book 1 of the series. In order to allow
for smooth discussions of the modern NGC processing tasks and for highlighting
the way in which NGC processing gets involved in the MDASE cycle, advanced
NGC concepts must first be examined. The examples chosen in the following dis-
cussions serve to introduce various parameters and functions of NGC processing
with respect to particular applications.


The conventional approach taken by design engineers to the avionics has been to
break up the avionics into three distinct and independent systems: the navigation,
guidance (midcourse andlor terminal), and flight control systems.

Navigation system. The navigation system functions to provide position,

velocity, and attitude of the vehicle with respect to a reference coordinate frame.
Using high-accuracy gyros and accelerometers, it is conventionally configured as
an inertial system in either a gimballed or strapdown mode.
14 MDASE of NGC Processing Chap. 2

Guidance. From the perspective of a control system, guidance is a matter

of finding the appropriate compensation network to place in series with the plant
in order to accomplish an intercept. In order for the pursuer to impact a maneu-
verable target with little miss distance, guidance uses the principles of feedback
control. The purpose of the guidance system is to determine appropriate pursuer
flight path dynamics such that some pursuer objective might be achieved efficiently.
The guidance system decides the best trajectory (physical action) for the pursuer
based on its knowledge of the pursucr's capability, target capability, and desired
objectives. In many applications, the guidance system is designed so that it makes
use of an inertially stabilized tracker (for example, seeker) that directly measures
the angular rates between the pursuer and its target in a fixed coordinate frame. The
function of the guidance computer is to mathematically integrate the separate func-
tions of navigation and the flight control system (FCS).

FCS. The function of the FCS is to control the pursucr in pitch, yaw, and
roll motion. The FCS executes the guidance commands and stabilizes the pursuer
in flight. The FCS, upon receiving commands from the guidance law, then issues
its own commands to the appropriate aerodynamic andlor thrust controls of the
pursuer so that the guidance command can be properly executed. It is usually con-
figured as a system equipped with low-accuracy inertial components (gyros and
accelerometers). Small tactical missiles and most large transport vehicles often have
open-loop control instead of the more complex FCS control.

2.1.1 LinearlNonlinear Intercept NGC Processing

Guided weapons or missilcs are normally guided from shortly after launch until
target intcrccption. The NGC systcrn supplics stccring commands to aerodynamic
control surfaces or corrccting clemcnts of the thrust vector subsystem to maneuver
the weapon to its targct and to make it possible for the weapon to intercept moving
targets. The guidance process, which is ruled by signal proccssing algorithms im-
plcmented in the N G C system, a prcset flight program, or both, is essentially a
feedback control systcrn where the pursuer-target engagcmcnt is considered part of
the N G C loop. Thc N G C loop consists of the guidance systcnl together with dy-
namic controls. Elc~ncntsof this loop include an information subsystem, control
clcmcnts, an opcrator, and pursucr dynamics. The components for the N G C systcrn
arc shown in Figurc 2-1 in which the ovcrall control of thc pursucr is dividcd into
two or lnorc loops. The main control loop in thc diagram is thc guidancc loop,
which is thc outcr loop that controls translational dcgrccs of frccdom, whilc the
inner controlIFCS loop controls pursucr attitude. ,The guidancc loop contains guid-
ancc scnsors for sensing pursucr motion, targct motion, or rclativc motion of thc
targct with rcspcct to the pursucr. This information is uscd in thc guidance computer
or the corrccting nctworks, togcthcr with information conccrning thc intended flight
profilc, to gcncratc guidance (latcral acceleration) commands for the FCS. Thc FCS
and actuation in turn direct control surface dcflcctions to altcr thc pursucr's trajcc-
tory. The body ratcs and accclcrations arc fed back to thc incrtial sensors to closc
the FCS loop.
16 MDASE of NGC Processing Chap. 2

The guidance and control laws used in current tactical missiles are based largely
on classical control design techniques. These control laws took birth in the 1950s
and have evolved into fairly standard design procedures. Proportional feedback is
generally used to correct missile course in the guidance loop, which is commonly
referred to as proportional navigation guidance (PNG), and is quite successful against
nonmaneuvering targets. The controller for a homing missile, in general, is a closed-
loop system known as an autopilot, which is a minor loop inside the main guidance
loop. In addition to the control surface and servon~echanism,the autopilot consists
of mainly the acceleron~etersand/or (rate) gyros to provide additional feedback into
the missile servos for missile motion modifications. Advanced sensors may measure
other variables. N o explicit state estimators are used and the signals are filtered to
reject high-frequency noise. Broadly speaking, autopilots control either the motion
in the pitch and yaw planes (lateral autopilots), or the motion around the missile
axis (roll autopilots). In general, the roll: pitch, and yaw channels are uncoupled
and are typically controlled independently of each other. All commands are am-
plitude or torque constrained to ensure autopilot and n~issilcstability. Classical
controllers have two major advantages, simplicity in design and simplicity in im-
plementation, but they also have several problems.
A pursuer is designed to complete the basic homing loop, which requires a
sensor (a seeker in the case of a guided missile) to track the targct, a noise filter to
reduce the effect of noise, a guidance law to generate thc dcsircd guidance (acccl-
eration) commands to home on the target. an FCS to rcccibc thc dcsircd accelcration
commandfrom the guidance system and to generate the required acceleration ca-
pability to insure interception of the maneuvering targct, and, finally, a good un-
derstanding of the physics of the homing engagement itsclf. Figure 7-2 depicts a
miss-distance analysis model conlposed of elemenrs rcprcscnting the intercept kin-
ematics plus clclncnts affecting the intcrccpt guidancc dynamics (that is. pursuer
dynamics and the NGC system), and illustrates the intcrdcpcndcncc bct~veensystcnl
elements. Inasmuch as thc total system perfornlancc is affcctcd by the individual
characteristics of evcry elcmcnt, the system cnginecr is grcatly conccrned with this
system interaction. The kinematic part inside the lowcr left dottcd box generates
the LOS angle in terms of pursucr and target motion. Based on the LOS anglc, the
N G C system generates a latcral acceleration for the pursuer, whilc trying to bring
the projectcd miss to zero. In the model, a rclativc position ) I , is added to a glint
noisc y,, thc result of which is multiplied by the inverse of thc range R to produce
a true LOS anglc corruptcd by glint noise. u, in inertial coordinates. A mcasured
LOS angle u is thcn obtaincd by adding ro (5. a radomc crror coupling u,and the
following angular noisc tcrms: rangc-depcndcnt (thermal) noisc o,,,, rangc-indc-
pcdcnt (fading) noisc a,, and cluttcr noisc o,.It thcn scrvcs as input to thc scckcr
dynamics to producc a mcasured LOS ratc 6 that is itsclf corruptcd by bias u l , and
g sensitivity. The output from thc scckcr u is the input to thc guidancc filter. Many
signal processing tcchniqucs arc uscd to discrin~inatcthctargct from its background,
other targcts, and decoys. The guidancc law functions to gcncratc midcoursc and/or
18 MDASE of NGC Processing Chap. 2

terminal steering commands u, based on the guidance filter output. Before being
sent to the autopilot, u, is limited. In general (except for coordinated turn vehicle),
the methods used to generate the guidance command signals u, for guiding a pursuer
in each of two mutually perpendicular planes are identical. Consequently, the guid-
ance algorithm needs to be determined forwnly one of the planes. In the treatment
presented here, this plane will be the missile s horizontal plane. In Figure 2-2, u, is
normally taken to be the lateral acceleration command. As long as saturation effects
are ignored, the model shown in Figure 2-2 is a linear time-varying system driven
by stochastic inputs. The inclusion of acceleration saturation effects causes the model
to become nonlinear.

2.1.2 Modeling and Simulation

The digital simulation technique in which the equations of motion are represented
by' a set of first-order nonlinear differential equations is widely recognized. The
modeling vector form is given as (see Figure 2-3):
x(t) = f(x(t), u(t), t) with initial conditions x(tO) = xg (2- 1a)
where the function f(x, u ) includes all of the model equations and NGC algorithms.
Equation (2-la) corresponds to Figure 2-3, in the absence of any noise inputs.
Depending on the specific application, the initial conditions are given, and the ter-
minal time tf is to be determined. For intercept flight simulation, rf is determined
by either the point of closest approach of the two vehicles or the time at which the
range rate passes through zero for the first time. Numerical integration techniques
are then used to integrate these equations with respect to time so that the time history
of the state from an initial condition x(t0) to a terminal state x(tf) can be obtained.
The output of the simulation is typically the time history of the state, from which
a performance criterion can be formulated as a function of the terminal state. The
flight vehicle to be designed determines the performance criteria in the simulation.
In the case of a guided missile, good performance is represented by short miss
distance or short range between the two vehicles at t,. The overall simulation block
diagram, shown in Figure 2-4, defines all of the state equations. Referring again to

Nonlinear Function -

Figure 2-3 Nonlinear Dynamic Systems

20 MDASE of N G C Processing Chap. 2

Figure 2-3, the modeling vector form, when a noise input vector is considered as
shown in the figure, can be written
k(t) = f(x(t), 4 t h t ) + w(t) (2-1 b)
where w(t) usually, although not necessarily, .denotes a white-noise input vector.
A more general case is considered in Figure 2-4, in which the model includes certain
error sources resulting from imperfect measurements or unmodeled dynamics. In-
corporating these error sources in Equation (2-la) gives
k = f(x, U) + b + n ~ ( x t), + m(t) (2- lc)
where b is the constant bias term; n l (x, t) is a state-dependent random vector; and
nz(t) is a state-independent random vector. Equation (2-lc) could also apply to
Figure 2-3, in which case w ( t ) is simply the sum of 6, nl(x, t), and n2(t). If M I or
n2 represents a colored-noise process, a shaping-filter technique can be used to model
either or both so that the new state vector can be augmented to Equation (2-lc).
The resulting augmented state equation is driven by the white-noise characteristics
only. A Monte Carlo technique, covariance analysis method, or an adjoint method
can be used to properly evaluate the selected performance criteria. These techniques
are described in Chapter 3.
In closed-loop control system analysis, it must be decided what level of detail
will be used to represent each element in the guidance loop. If it is desired to study
high-frequency system instabilities, then one must employ a representation that is
accurate at those frequencies. The elements that may be required to construct a
sufficiently accurate representation in this case include modeling the dynamics of
the nonrigid system, knowledge of the full nonlinear aerodynamic characteristics,
and detailed modeling of the seeker track loop, guidance signal processing, autopilot,
and noise sources. Conversely, a complicated model of this type would not be
required in the case of a preliminary parametric study of homing trends since the
main effccts of interest are found at low frequencies. Rathcr, a simplified represen-
tation of the more complex system capable of incorporating the available trim aero-
dynamic data and the low-frequency approximations to the different subsystcms
that make up the guidance kinematic loop would be more appropriate. After being
validated, this simplified representation can be used to study the relative performance
of various missile configurations as well as the result of varying system paramctcrs
such as time constants and limits [Reichert, 19811. In problems involving dynamic
modeling and controls, a state model is usually specified in addition to thc niea-
suremcnt equation, relating the state and measurement vcctors as shown in Table
2-1. Bcfore implctnenting thc preceding N G C simulation, it is important to un-
derstand the roles of thc pursucr system, the targets, and thc cr~virot~ment so that
appropriate modcls in Figure 2-4 can be devclopcd for a particular pursuer-targct
engagcmcnt scenario. Essential guidance and control softwarc modulcs arc as fol-
lows: intcrfacc, flight path planning, strapdown device, navigation, guidance, sta-
bilization and control, fin actuation, target tracking, and self-test. Their relationships
are dcscribcd in Figure 2-5.
Sec. 2.1 LinearlNonlinear Intercept NGC System 21


D t f ~ n xg ( m l ) state vcclor. y - (pxl) system output vcclor. z = (1x1) measured output vcetor.
u = (mxl) mnlml vmor, I, i(qil) cxogenram inpu v m o r
.?(I)w d a i n s all thc variables that can be measured and arc fed back to thc controller. Thc enuies of y(t) arc the variables lo be
conmllcd. up) is h c output of ihs mnuollsr. lc(l] wnsisu of thc reference inputs, the disturbance inpuu (e.g.. process noise
w(t)), and the rcnror M i s s ~(1).Note that y(t) and z(t) may have some wmponenu in common. The wnlinuous-lime nonlinear
system is dcscribcd by
i t =x I I t I , X I =x I ~ 4q1)*TI
Q ~ I= 0. I = at)b(t-r)
Z ( I ) s h(x(1). I) t "(I), Qv(t)] = 0. Qv(l) "(1) ] = R(t) b(1-z)
The continuous-time nonlinear system with d*crete.time nonlinear mcasuremcnls is repnscnled by:
i(1) = f(X(1). Nt), 11 + *I)
q k ) = h[x(tkh kl + v(k); qv(kl1 = 0. Qv(k) v(k) I = Q
w h e n the parameter k is considered to be a timc t and the measurements an sampled at discrete timc intcwak. The discrete form
of the state equation (la) is as follows:
k =k l1 k l Qqk)l = 0, q W k ) wT(k)l = or
h applying many MDASE uchniqucs. the system quations (1)need to be linearized as:

where u, w, and v arc defined in Eqs. (1). Eq.(3) is equivalently expressed as

i(t) = &(I) + Bqf) t Br i(1) and y = Cx + h
~ ) f(l) is a
The state equation in Egs. (4) is generally a nonhomogenwm differential quation with forcing term ~ .t f. (where
command vcctor anUor an anticipated forcing function mulling from a system nonlinearity andior an estimated dinurbance. In
the output quation of Eq. (4), h consisls of the measurement nonlinearilia and uncenainties. The discrete form of the system
quations (3) is:
x(k) = Wk.1) x(k-1) + W(k-1) qk.1) + wk-1) qk-1)

2.1.3 Guidance System Classification

There are several solutions that can be applied to the problem of guiding a missile
to an eventual near miss or collision with a target using only the observed motion
of the missile-target line of sight (LOS). In one of these, the missile is constantly
. - directly toward the current target location. This method, termed simple pur-
suit, suffers from the fact that the missile must generally undergo rather severe
maneuvers as intercept is approached. A second solution concerns a constant-bearing
trajectory wherein the missile leads the target, much like the way in which a pro-
jectile is traditionally fired at a moving object. A collision invariably occurs as long
as the missile flies a course that keeps the relative missile-target velocity aligned
with the LOS. As shown in Figure 2-2, the missile trajectory becomes straight line
when the missile speed, target speed, and course are constant. PNG is a steering
law that is designed to produce a constant-bearing course even in the case of a
maneuvering target. T o accomplish this, one makes the rate of change of the missile
heading directly proportional to the LOS rate. While this technique presupposes
that the missile is able to respond instantaneously to changes in the LOS direction,
22 MDASE of NGC Processing Chap. 2

o initialization Navigation Guidance Autopilot
o Euler o State Estimation o Midcourse & Terminal o Rate Stabilization
4 Transformation _* o Position + o Power On Mode o Acceleration Control
o Bias Compensation o Velocity o Power Off Mode o Aerodynamic Coupling
o Coordinate
Transformation o
o Acceleration
Estimate E m r s
Vertical Launch
Separation Mode

o Aim Point
Self Test
o Test
Right Path Planning
o Stored Mission
o Fin Rate Limitation
4- Sequence 4-W
o Initialization Profile o Fin Angle Limitation
I 0 Status
- o Change Profile

Relative RF Environment & Clutter

ECM: Target Kinematics Sensors o Equations of Motion
o Rate Gyro , o Engagement
o Accelerometer o Environment

Figure 2-5 Pursuer Target Tracking Processing Software Modules

there is in practice some delay in returning to a constant-bearing trajectory subse-

quent to a target maneuver as a result of finite missile responsiveness and filtering
introduced to diminish the effects of noise. Figure 2-6 defines the conventional
PNG configuration, in which the measured LOS rate u is filtered through a noise
filter to produce a smoothed LOS rate estimate u, and u, is made proportional to
u A similar idea is used in modern guidance approaches based on optimization
techniques. If measurement noise is to be suppressed, then consideration must be
given to a low-pass noise filter. PNG systems generally employ a first-order low-
pass filter, which has a transfer function as shown in Figure
- 2-6. The missile guid-
ante computer, using the preceding relationships, computes steering commands.
Besides noise filters, com~ensationsare introduced to offset such factors as variations
in the missile's velocity and radome effects. Also, in order to maintain aerodynamic
stability along with structural integrity, steering command limits are imposed [Witte
and McDonald, 19811.

2.1.4 Flight Control System (FCS) and FCS Sensing

That part o f the system comprising the FCS also includes the inertial reference unit
(IRU)to measure A,,,. A,,, is the acceleration of the pursuer in response to the
guidance command, which is applied to the FCS and results in operation of the
pursuer's control surfaces. The control surfaces can be caused by aerodynamic cross
coupling, dynamics nonlinearity, and structural vibration to deflect at different ori-
entations from the desired guidance input, thus contributing to miss distance. Au-
Set. 2.2 Target Signal Processing 23

velocity MiSc --
without ,
, R ~ V target
~ velocity vector


The basic proportional navlgation equation

Seeker Guidance Autopilot
Am a V,,,? = AV,b Gs(S) GF(S) GA(S)
Missile/ / 1\ ~uidance.res~onse
maneuver Navigation clos/nng Line sight
velocity = Velocity I line of sight
g(S) Range
k = Ik, Noise Ntcr bandwiith
V, Closing vebcily
s+t A Effcaive navigalioo mtw

Figure 2-6 Proportional Navigation (Courtesy o j K . Hiroshige, 1984)

topilot, actuator, structural aeroelasticity, sensors, and nonlinear airframe dynamics

are part of the actual mode of an FCS. Details of the FCS are presented in Book 3
of the series. The useful models for NGC design are summarized in Figure 2-7,
Table 2-2. and Table 2-3.


A mathematical model is used to embody pursueritarget kinematics in the form of

parameters of the LOS as shown in Figure 2-8, which shows what type of processing
is required. A complete simulated engagement of pursuer target tracking begins
with the launching of the missile, is followed by a midcourse and a terminal phase,
and ends in an intercept and the scoring of missile-to-target miss distance. In active
homing, after launch the missile seeker transmits radio-frequency (RF) energy to
the target. The target converts this to a synthesized return and radiates it back to
the seeker radar. Multiple target returns, which include different ranges, velocities,
amplitudes, angular noise, and environments, can be generated simultaneously with
or without ECM. The returns are tracked by the seeker. As shown in the relative
geometry box in Figures 2-2 and 2-8, the homing loop is closed by calculating the
missile-to-target geometry as well as using the RF array to update the relative angular
. MDASE of NGC Processing Chap. 2




(A) XCG - center of gravity lacation

XACC - arcelermeter location



- .30




& = . 3 0 = 6MAX
6 = .20
6 = 00 b = -100

Figure 2-7 (a) Horning Missile Axis System and Forces and Moments (b) Forces Due t o
Angle of Attack a and Control Fin Uetlection S in a Tail-Controlled Missile (c) Normal Force
Coefficient CN VS. Angle of Attack a (d) Total Moment Coefficicnr CM 1,s. Angle of Attack
a (Froin [n'erlincs, 1984(a)] u,irk prnnisrio~tfi.otnAACC)

position. As shown in the endgame computation and miss-distance computation

boxes, immediatcly before interception, the characteristics of the endgame including
the miss distance are determined by extrapolation. Finally, the LOS rates in the
pitch and yaw axes are computed and sent to the missile seeker as pitch and yaw-
head radar errors to fully close the guidance loop. E C M J E C C M modeling involves
digital computer simulations of an operating radarlsensing system including the
effects from an enen~v'sE C M . This allows the sensine " unit to be assessed under
field conditions as to whether it would perform satisfactorily so as to support the
tactical control system nlission. A sensing detection model is required to obtain
signal-to-noise ratio data as a function of target range. The model must be able to
allow these data to be obtained in both clear and noise-jamming environn~cnts.
Jamming can be produced by a self-protecting or stand-offjammer operating against
the radar signal in either the main lobe or the side lobe. Enemy threats may be
taking placc in ascending, level, or descending flight profiles. Multiparh and terrain
(terrain-masking and clutter) effects may also need to be considered.

Hjrmal Auxlemlion: A, = v,?, ? = Aa + I36 (1)

- Aoxlcrdlion: 0 = -(h-1%
I1ilch Anxllrr - lib; u = 0 - y (2)
G ( ~- I i~ -C HI^ - 1*i2) v m - I SkI(~+~lls+~12s2)
'liansfcr l'maions: -= - c') -= - (4)
6 s2 + (A+#>&+ <: + All s2 t (A+l>)s + C + N) (6)

(W. - HC + 1%) k ( I + )&I'

6 Dl (8)
- kl
w m lag
-0= .
6 s2 + (A+l))s + (: + N)
- 3
"1 (")

Ik,r 1;nil mnlml or c;tn:~rrlnirfr;mw with negligihlc liII from ctmlq,l s u t t ~ i.c..
~ , 1% 0. -
knrdynamic lift: A,,, = Vm? - Vm Atz = A Vm (0-Y); O =(I + s/A)Y (8)

- - -

A = -7ff - -at CN, aftcr boas1 (thrwl = 0 ) D = -Mi, = -ao (<Y(2Vm)) CMb AIRFRAME IS T H E HISULT O F TRADE-OFFS
C = -Ma = -a,CM, a, = q ~ a l y y ,a I = -qS/(mVm)
6 = fin &fleaion or thnnt
BO kn &n!s%h W c w
Wnpw-5253. I
A, =achieved normal ameleralion a1 missile Iznfer o f gravity prrprndicllar lo missile body
A,,," = auzleration measured al the amelemmeler
6 rate
k, = acceleration gain umstant = k 2 = (J((!+/U)), P = Vm ( H ( z - a l
2 ( 8 1 ~
k 3 = L I P m = MaI(l&u)d)
Ul = 2 W w d = (A+I))/(C+AD)
&- h m p i n g ralio of airframe 4 -
if M a < ( )
o --natual frequrncy of airframe =
B., = 11o:" = lllC+r\o,
. .
Ma > 0 and 2Sdwd i s small

Ta = 4 missile l m i n g rate time mnslanl = -P/(BC-AE) = 111% lIA(if B 0 ) - -

u b = low f r e w n c y lead

Courtesy of [ H i r o s h i g e , 19861 and from [Neslines, 1')84(a)l w i t h pernlissioll f r o t n AACC
TABLE 2-3 - s [Neslines. 19851 with permission from AACC
HIOH ~ R E Q ~ E~ N ~~ ~1 l From

Aaua10r: --
(1) -
Rate Gyro: ~ G ( S )-
0 )

"A "1 "G 02

1t- s2
4 ~ s ) 1 6 (s) "sf
Acaleromaer: -= (3) Suuclural Filter: &= (4)
%(S) 2t*m s2 US) 1 +2 hsf + - s2
"xc w&~~~ ~
"sf "fF
Nominal Values for a Typical Missile:
UG = UW) &% LU, = 220 radkec, = 0.65. fA= 0.65,

The mode shape cnvclopes a n shown below.


MASS t l b / ~/in) 0.0517 0.0373
0.4 i Gym Location
0.2 )X

a) Fint and E m n d Mode Shapes
bl Normalized Bodv Bendinn Sham

The detailed flcxiblc body transfer Function from b LO body rae is

- 24
(av% & ( I v l s2

- ~ ~ I Y Y &

&B -
Xffi XBL

,1 2Ei
w1' Wi

-- -
Nominal values typically used in thc above equalion are
k damping of i t h v i b r a t i o n a l mode I n m i s s i l e molaent of i n e r t i a 11.75
9 t h e v i b r a t i o n a l frequency of t h e i t h

6 -
n o r n a l i z e d .lop a t r a t s g y r o a t a t i o n
l o c a t i o n of missile ffi (5.885 f t )
l o c a t i o n of missile hinge l i n e

- --
10.00503 rad/in., -0.01986 rad/in.) (11.3108 f t l
% n o r u l i a e d airvane deflection % mass of one f i n (0.0489 slupm)
<- 10.315 i n . / i n . , -0.1485 i n . / i n . ) %
d i s t a n c e between ffi of on8 f i n and

normalized a i r v a n e s l o p e hinge l i n e 10.084 I t )
1-0.02 r a d / i n . , 0.03308 r a d / i n . ) I, m e n t of i n e r t i a of m e f i n (0.00683
lt. e f f e c t i v e mass 40.0514 l b - s e c 2 / i n . , lb-ft-s.c2)
-0.03131 lb-sec / i n . )
Klsi g a i n of s t r u c t u r a l p a t h of the
I t h moda

%mi - qu*dIatlC a s r o Of the i t h .Dde a t r u c -

t u r s l path
the general transfer funaion from b lo body rate for each mode is given by 404% - 24 41"
Sec. 2.2 Target Signal Processing

Target Dynamicl&

+ Relative ~ c o r n c n y t
-Target Stale
Target Starc
bra 4 Sellsor Models -el Filtering &A Estimation

rare Esnmator lnrduaoo

Figure 2-8 Target Tracking Processing

The most common measurements of target tracking systems consist of the

range R, range rate R, azimuth u, or elevation u, angles (see Figure 2-9). The three
coordinates to specify the position of the target are:

1. The x , y , z components of the tracking sensor-target range R

2. The azimuth u, elevation u,, and range R.
3. Direction cosines of the vector R and its length.

A tracking sensor or seeker mounted on a homing missile or an aircraft is a device

to detectltrack the relative position vector of a target with respect to the missile or
the aircraft. The function of the tracking sensor is to receive external signals that
are sent to the pursuer vehicle for the purpose of directing its course. Because of
its role in providing an essential communication link to the external environment,
the sensor becomes the key element of the entire vehicle system. In fact, a vehicle
is given a particular identity by the choice ofsensor. The source ofenergy responsible
for the signal that is picked up by the sensor may be energy reflected from the target,
energy emitted from the target, or energy from the target's own environment.
Assuming at this point that the sensor is capable of recognizing hot targets, or targets
that emit infrared energy, then the sensor will be selected on the basis of a certain
frequency or frequency band emitted by the target and will be optimized to respond
to the target signature in question. Given the relation of the sensor to the successful
operation of a vehicle, it becomes apparent that the aforementioned optimization
to a great extent affects the design of all smart weapons. Although there are other
important components besides the sensor'in the vehiclelmissile, the ability to neu-
tralize the sensor reduces the vehiclelmissile to something less than a smart weapon
[Heaston and Smoots, 19831.

Tracking sensor/guidance comparisons. One of the requirements of

a sensorltracker is to acquire and track all modes of target day or night. T o do this,
an air-to-air tracker uses a laser-auadrant tracker and a centroid track. whereas an

air-to-ground tracker uses a scene track, target track, and track adjust. A sen-
sorltracker must accurately update target position relative to IP, OAP, and VRP,
MDASE of NGC PrOCesSlng Chap. 2




Figure 2-9 (a) Visual Line-of-Sight (LOS) Angle Measurements Determine T w o Coordi-
nates of Aircraft-to-Target Position (b) Determination of Slant Range (c) External Grid Ref-
erence Targeting (Courfesy 0jJ.S.Ausman, 1986)

permitting target identification day or night. Low-altitude ranging sensors for NGC
algorithms include redundant range measurement and accurate range measurement
during terminal maneuver.
Sensors used in the capacity of seeker, depending on their operating wave-
lengths, are typically characterized, depending on the sensor's operating wavelength,
SIX. 2.2 Target Signal Proceulng 29





/ \


Figure 2-9 (Continued)

as optical sensor, infrared (IR) sensor, synthetic aperture radar (SAR), millimeter
wave (MMW), laser radar, television (TV)ivideo, microwave, antiradiation homing
(ARH), acoustic sensors, and multiple sensors. While .
.,~. of good .quality
. . . . .and high

resolution, an optical sensor nonetheless can suffer degraded-performance due to

weather conditions or time-of-day effects. An IR sensor is designed for use at day
or night, in conditions of rain and/or smoke, and is capable of hot-spot detection.
I t is also well suited for use against armor. and high-value and ship targets, and has
reasonably good resistance to ECM. However, this all comes at a high relative cost.
Like the SAR, the MMW sensor is designed for all-weather operation with the
capability of cloud penetration, and is particularly well suited to bad weather con-
ditions. MMW sensors can be used against armor and high-value targets, but are
not designed tor use against ship targets. They have a moderate resistance to ECM,
and are not unduly expensive to build. The laser radar, which utilizes three-di-
mensional distance information and surface reflectance, is desigaedfouse.atday
or night against armor and high-value targets. While not designed for use against
ship targets, it has a high resistance to ECM and comes at a relatively low cost.
Acoustic sensors are designed to function in underwater-erivironments. A T V sensor
cannot function at night and is designed primarily tor use against armor targets. It
30 MDASE of NGC Processing Chap. 2

is not very useful against high-value or ship --- targets.

-.- Although TV sensors have a
low resistance to E C M , - & ~are~ relitiVdy inexpensive. While a forward-looking
IR (FLIR) imaging sensor provides a different mode of target detection and rec-
ognition, a laser provides a different mode of target range and track. A video tracker
microprocessor FLlR consists of sensors and ancillary electronics as well as video
processing. Microwave sensors can function in all types of weather conditions and
are designed for use against high-value and ship targets, but not against armor
targets. They have low resistance to ECM and are relatively expensive to build.
ARH sensors are designed for all types of weather conditions and are only mod-
erately expensive to build [Hardy, 19861. Finally, multiple sensors act to process
multisensor information.
The interceptor/missile uses sensor data that are processed by the seeker, which
also acts to ensure that the sensor is receiving as much target information as possible.
Related to this last function, the seeker actually orients the sensor so that it can
survey, acquire, and then lock-on and track the target. The important components
making up the seeker include [Heaston and Smoots, 19831:

energy-gathering system, that is, antenna for radio frequency or lens/mirror

for IR, visible, and ultraviolet spectral regions, including radome o r IR-
stable platform and its associated control system (body-fixed seekers employ
inertial sensors for an equivalent function)
sensor to convert the received energy into a more usable signal for processing
signal processing system used to produce a signal to point the antennalmirror
at the target and/or to provide signals to ultimately control the missile's tra-

Rather than focusing too much attention on any particular tracking sensor or seeker,
it is more advantageous to consider their fundamental properties. When, however,
the problem is that of conducting a precise system analysis, then a specific dynamic
model is assumed for a particular seeker, and this is integrated with that o f the
pursuer. Three homing configurations are assumed: passive, active, and semiactive.
In the passive case, only the LOS angle is available through an 1R seeker. In the
active homing case, both range and closing-velocity radar information are available
through measurements, but in the semiactive mode, these measurements are bistatic.

2.2.1 Targeting

The presentation on targeting in this section follows an excellent lecture by [Ausman,

19861. Position sensor data must be available for targeting so as to be able to locate
the pursuer with respect to the target. One or more coordinates of pursuer-to-target
relative position is measured with a targeting sensor. This device may also be used
Sec. 2.2 Target Signal Processing 31

to locate the pursuer with respect to a pseudo-target o r offset aim point, the position
of which with respect to the real target is known beforehand. Various types of
targeting sensors with examples of each are listed in Table 2-4. A particularly com-
mon typc of targeting sensor is one that measures the azimuth angle and depression
angle of the pursuer-to-target LOS. Examples of this typc of sensor system include
gunsights and headup displays (HUD) in combination with an attitude reference.
As shown in Figure 2-9a, these sensor systems measure just two angular position
coordinatcs. A slant-range sensor such as an air-to-ground ranging radar or a laser
range tinder can be used to directly measure the third coordinate of pursuer-to-
target position. This is illustrated in Figure 2-9b, which also shows two additional
more indirect techniques for determining the third (slant-range) coordinate.
Pursuer altitude h can be measured using a barometric altimeter. Subtracting
from this the target's altitude hT, assuming knowledge of the latter, results in the
quantity h - h ~ A. radar altimeter, however, can measure h - hT directly in a
target overfly. If it is not possible to realize a preparatory target overfly, the pursuer
must accept thc difference in terrain elevation between the target and the spot below
the aircraft at release or pickle as an additional measurement error. Slant range can
also be obtained indirectly from LOS angle ratc measurements. One example of an
angle ratc measuring device is the Norden bomb sight used in World War 11, which
relied on manual tracking during a long, straight, and level bombing approach. By
using EO trackers, modern angle rate sensors avoid the long, straight and level run-
in. The last entry in Table 2-4 corresponds to the external reference type of targeting
sensor, which is illustrated in Figure 2-9c. The typical application of this sensor
type is a blind bombing of known coordinates. As will be shown in Chapter 5,
examples of external references that can locate the pursuer with respect to the target,
assuming the same reference system is used to specify target coordinates, include
ground-based radar, LORAN, TACAN, DME, and GPS.


Generic Tyve
unsight, T
Line-of-Sight ( L a ) Angles kpping ~ a d a r , ' ~
Azimuth o and Elevation o, Optical, Video Tracker
Radar, Laser, Sonar, Optical Instruments,
FCS, Passive Ranging Processor,
Range R Onboard Active Sensor
Altitude h Baro-Altimeter, Radar-Altimeter
LOS Rate b, bc EO Tracker
Range Rate k = -Vc Doppler Radar
External Reference Ground Radar Director
Courtesy 0fJ.S. Ausman, 1986
MDASE of NGC Processing Chap. 2

Line-of-sight (U)S)sensors. Common to all LOS sensors is that some

type of image of the target area is displayed to the weapon system operator. The
operator then selects the target by positioning a marker or pair of crossed reticles
over the target and depresses a target designating button. The computer responds
by recording the azimuth and depression angles of the target marker (pipper). What
the operator sees may be the real visual scene, or it may be an electronic reproduction
using TV, IR, or radar sensor subsystems. The resolution and accuracy ofthe shorter
wavelength systems (optical) are generally better than that of other systems. Ac-
curacy may be improved by using scene magnification, which can also aid in target

Range sensors. A range sensor, as shown in Table 2-4, provides t w o

means of measuring slant range from the pursuer to the target. For weapon delivery,
it is not necessary to have extremely accurate range measurement since the bomb
miss distance for typical dive deliveries is on the order of one fourth of the range
error responsible for the miss. That is, a range measurement error of 100 feet causes
a miss distance of only 25 feet, typically. What may at first appear to be an incon-
sistency is actually a consequence of the fact that the range error causes correlated
errors in ground range and in altitude that partially compensate each other. As an
example, a very large slant-range measurement causes the computer to estimate the
aircraft to be further from the target (short impact) but at a higher altitude (long
impact). It should be noted that the range R, the quality of information concerning
which would be very important if it were being directly utilized in the solution of
the NGC problem, is not an easily measured quantity. For missile applications, the
explicit range information may be obtained from either the launching platform or
launching shiplground defense site, as when the missile and target are illuminated
by the launching apparatus, or from the target sensors listed in Table 2-4. It still
may happen, however, that range information is not readily available, or that the
price in terms of space andlor dollars associated with the tracking and guidance
hardware to acquire reliable range information is prohibitively great. Such can be
the case with smaller missiles. Thus, it is advantageous to derive guidance algorithms
that do not require range information.

Altitude sensors. Inasmuch as it is a component of the pursuer's flight

instrument package, the barometric altimeter provides a convenient targeting sensor.
However, by reason of its rather doubtful accuracy, it usually serves as a backup
to be used in the event that the slant range sensor ceases to function correctly.
Assuming that local barometric pressure is accounted for, the largcst altimetry errors
are typically those arising from uncompensated stat~cpressure defect, which is a
bias type error that is on the order of a few pcrccnt of the dynamic pressure, and
to uncertainty in lapse rate, which is a scale factor typc error that is on the order
of a few percent of the pursuer's altitude above mean sea level. Besides altimetry
errors, any unccrtainty in the target altitude will result in an error in the computed
value of the pursucr's altitude above the target. While this unccrtainty is removed
Sec. 2.2 Target Signal Processing 33

on a bombing range, it can be significant in actual operational usage. Consequently,

one expects a certain amount of degradation in accuracy between bombing range
results and operational results with this type of bombing system. It becomes un-
necessary to know target altitude with a radar altimeter since this measures altitude
above terrain directly. Given that the pursuer is not usually situated directly above
the target at the instant of release, the flatness of the tcrrain between the release
point and the target to a large extent determines the bombing accuracy associated
with a radar altimeter. For this reason, it is advantagcous to make a dry pass over
the target for a relative altitude update before proceeding to the attack. This, how-
ever, is only possible if the threat environment permits.

LOS angular rate sensors. In an angular rate targeting scheme, the

relative accuracy of the slant range computation is directly proportional to the rela-
tive accuracy of the angular rate measurement. A 1 percent error in measuring
angular rate therefore results in a 1 percent error in slant range. For a typical depres-
sion angular rate of 20 milliradians per sec, this accuracy translates to 0.2 milliradians
per sec, a level not easily achieved with current state-of-the-art trackers. The great
majority of angular rate sensing systems that have been built so far are of the T V
contrast tracker type. In addition to requiring three-axis stabilization, these sensors
have a response time that is scan rate limited. Owing to this last limitation, these
devices can measure angular rates only so accurately. The T V contrast tracker is
also characterized by the requirement of a single, high-contrast point to track.

Range rate (closing velocity). Determination of the closing velocity V ,

is done with the aid of a Doppler tracking device. Doppler frequencies, which are
one of the principal outputs of the radar sensor, are proportional to V,. Continuous-
wave-illumination homing systems can exploit the Doppler frequency of the target
return to generate a good estimate of V,.

External references. Each of these targeting sensors is a navigation aid,

which locates the pursuer in an electronically generated grid or coordinate system.
The rest of the targeting task centers around locating the target in this same electronic
grid or set o f coordinates. One way in which to implement this type of targeting
is to use radar ground control to direct the pursuer's flight, which requires that the
pursuer be within LOS of the ground radar. The pursuer, in order to maintain this
LOS, must fly at relatively high altitude, particularly when it is very far from the
radar. Moreover, accuracy degrades in proportion to the distance from the radar.
Other LOS positioning systems include TACAN and DME, which generally rely
on at least two ground stations and use triangulation and/or multilateration to es-
tablish the pursuer's position. T o a large extent, accuracy is determined by the
relative geometry between the pursuer and the several ground stations, also known
as geometric dilution of position (GDOP). G D O P has a greater influence on the
altitude coordinate than on the horizontal coordinates, unless the pursuer is situated
directly overhead a ground station. This is not likely to occur, however, during a
34 MDASE of NGC Processing Chap. 2

bombing run. The LORAN system is capable of obtaining coverage beyond the
LOS and at low altitudes because it uses ground-wave transmission oflow-frequency
radio waves. However, since it is only a two-dimensional positioning system,
LORAN must rely on another method such as a baro-altimeter to obtain vertical
targeting information. When GPS is operational, it will be able to position low-
flying pursuers in all three dimensions since its reference stations are satellites. With
the GPS system, each satellite can maintain LOS to almost half the earth.

2.2.2 Kinematic/Relative Geometry

Because only kinematic equations are considered in the kinematic method, the pur-
suer may be treated as a geometric point as illustrated in Figure 2-10. The corre-
sponding guidance is therefore ideal, and the character of the pursuer trajectory can
be determined approximately. Moreover, the required accelerations may be esti-
mated. The kinematics of a pursueritarget scenario are contained in the change in
time of the vector from the pursuer to the target R. In other practical applications,
the kinematics are characterized by two pairs of kinematic equations, one of which
determines the relative motion of the centers of mass of the pursuer and target in
the horizontal plane, and the other of which determines this motion in the vertical
plane. The geometry of the homing guidance in the horizontal plane is described
by the following relations:

where, for convenience of discussion in this section, the subscript true in u,,, is
omitted in this section from this point on. The nomenclature is displayed in Figure
2-10 and the measurable variables are
2 I12
R = (x? + y,) , o = tan-' y , / x , (2-3)
The closing velocity V , = - R of the pursuer and target is determined by Equation
(2-2a), while the LOS rate u is related to the parameters of motion of the pursuer
and target by Equation (2-2b). A set of analogous equations is used for the vertical
plane. Differentiation of Equation (2-2b) and substitution of Equation (2-2a) yield
R6 + 2db = - VT sin(u - yT) + V T cos(u - ~T)+T
+ i',,, - y) - V,,, COS(U - y)+
Because kinematic Equations (2-2) and (2-4) are nonlinear differential equations,
they render it a difficult process to study and analyze guidance loops. The analysis
can be facilitated by linearizing the kinematic equations relative to the reference
trajectory of the pursuer, using approximation methods. With initial values given
by YT, and yo, theangles y~ and y are perturbed during the engagement by y, and
y, where y, and y,,, are very small compared to yr,, and yo, respectively. Conse-
quently, the instantaneous angles of the velocity vectors are y r = y ~ + , y, and
36 MDASE of NGC Processing Chap. 2

y = yo + yn,. A l y simde -case, which is of course linear as well, occurs

when sin a zz u , cos u = 1, s i n y,, cos y, = 1, sin y, = y,,, and cos y,, = 1.
The L O S angle in Figure 2-10 can be approximated as:

t, = tf - t = time to go and t f =- flight time
For analysis purposes, assuming the intercept geometry is simplified to a head-on
engagement with constant closing velocity, the range is determined from
R = Ro - VCtZZ V,(tf - t) = I/,!, (2- 5b)
where R, = V C t fThe relative velocity and acceleration are
j/r = YT - jrn = VT Y I cos Y T ~- V,n ~m cos YO (2-SC)
y, = yT- y,,, = AT" - A m y = V T j T - V,,,? (2-Sd)
AT^ = y ~ =. actual target lateral acceleration = V T jT cos y ~ , =
! VT jT
A,, = , = actual pursuer acceleration = V,,,j cos yo -- V,, j
Equations (2-5) are linear equations representing the dynamics of the engagement.
The signal that results from subtracting A,,,,, from the simulated AT, is integrated
twice to generate y,. With this information, Equation (2-Sa) can then be used to
obtain the true L O S angle a,,,,. Assuming VT and V,,, are constant, and hence V,
is constant, differentiating Equation (2-Sa) leads to the following formula for the
L O S rate u:

Rewriting the equation with the help of the geometry yields

yr = Vr&u - Vru = ~ J -
T jm = VT y, cos y ~ "- V,,, y,,, cos yo (2-7)
or in block diagram form shown in Figure 2-lla. Differentiating Equation (2-7)
y, = V,t,U - 21/,u = AT,. - A,., = VT jr cos y n , - V,,,j cos yo (2-8a)
or, in a more general form,
y , = RU - ~ V , U= AT" - A"," (2-8b)
By applying the assumption made so far, Equation (2-4) can be reduced to Equation
(2-8b), while Equation (2-7) can be derived from Equation (2-2). The component
of the system for y , to I?, based on Equation (2-8b),can be modeled with an in-
Figure 2-11 (a) Kinematic Eqi~ations
(b) Guidance Loop

tegrator contained in the positive feedback loop, as shown in Figure 2-11b. The
kinematic element, however, becomes unstable on account of the presence of posi-
tive feedback. That is, the output signal u grows continuously for a constant input
y,. A negative feedback (A,,,) loop is consequently added to the model in order to
counter the effects of the positive feedback. The guidance law, FCS, and the pursuer
dynamics and propulsion are all contained in this loop. From Equation (2-8),
AT, is given by
AT, = (Rs - 2V,)u + A ,,,, = - 2)u
V<(t,?s + A,, (2-9)

2.2.3 Targeting Sensor Dynamics

For the active RF seeker and the passive IR seeker, two gimbals (that is, azimuth
gimbal and elevation gimbal) are required. In order to describe the generalized seeker
tracking dynamics, the RF seeker with an antenna dish is used as an example. Figure
2-10 shows the essential seeker angles and geometrical quantities. As shown in
Figure 2-12, due to the presence of lags in the seeker tracking loop and radome
refraction effects, the seeker is not pointing directly at the actual target [Murray,*
19841. Figure 2-13a illustrates the tracking loop that describes the dish dynamics
in the absence of radome effects and having the seeker bandwidth w, = 117,. In the
case of small track loop time constant T , , the seeker dish rate command u; and the
LOS rate are approximately equal. Care must be taken to ensure that T, is adequately
small to be capable of tracking highly maneuvering targets, but not so small as to
induce excessive noise transmission or a stability problem. The stabilization rep-
resents a closed-loop system encompassing the dish rate feedback loop.
MDASE of NQC Processing Chap. 2


Figure 2-12 Radome Error Distorts

Boresight Error (From [Munay, 19841
with permission . t o m AACC)

Radome coupling and compensation loops. The increased miss dis-

tance associated with radar-guided air defense missiles can be blamed on radome
refraction error. Miss-distance predictions for nonlinear radome error are historically
less than that predicted for a linear radome error with constant slope [Yost et al.,
19801. As shown in Figure 2-12, the effect of radome error, a,,is to distort boresight
error measurement E. a,,which is the amount by which the direction of the apparent
Total LOS Angle
Noise v,

Gyro Acceleration
(B) Sensitivity
Figure 2-13 Seeker Block Diagram (a) Without Kadomc Effect (b) With Radomc
Sec. 2.2 Target Signal Processing 39

target and that of the true target differ, is a nonlinear function of the gimbal angle
q,. Differentiating u, with respect to u, yields the radome slope r, which is not a
constant but varies inp predictably with angle-off and from dome to dome. u, itself
depends on both pitch and yaw gimbal angle [Murray et al., 19761. Here, the effect
of u, on miss distance can be shown quite well using only a single-plane guidance
model [Murray, 19841. The radome error a, arises because RF energy passing
through the radorne to the seeker antenna is refracted by the dome material. The
magnitude of refraction during homing depends on many factors including radome
shape, fiticness ratio, thickness, material, tcnipcrature, operating frequency, and
polarization of the target 'echo signal.
~~ . Other sources of error such as standing waves
inside the dome, reflection from gimbals, receiver phasing, and processing errors
also contribute to the radome error. All of these factors differ with each flight and
therefore cannot be specifically accounted for in the computer algorithm. As a result,
the radome error magnitude can neither be precisely measured nor predicted. This
is not to say that it cannot be modeled. The most serious aspect is that the magnitude
of refraction error is also a function of where the target energy impinges the radome.
This causes in-plane and cross-plane radomc errors to become look-angle dependent
with the slopes also depending on the roll attitude of the in~erceptor.Specification

of radome characteristics involves the radome slope r, which is ~-

. ~. shown to be-the
main parameter in the homing loop. The most optimistic situation is o i i in which
the designer would be able to specify the manufacturing tolerances and the limits
on the allowable variations of r.
As shown in Figure 2-2, if the nonlinear kinematic guidance loop is to include
the radome error coupling loop, then the missile body pitch angle 0 must be gen-
erated by Equation (7) of Table 2-2. In addicion to being used to generate 0 in
Figure 2-2, A,,,l is also applied to both the guidance filtering and guidance algo-
rithms. In this study of targeting sensor dynamics, a PNG system is assumed where
the missile's antenna provides a measure of the target bearing u as follows. Looking
at Figure 2-10, there is an angle a d called seeker dish made by the seeker antenna's
center line with the inertial reference. The geometry tracking error without radome
error is ~ , , r f e ~ ~= ulrue- ud (see Figure 2-13a). The radar converts this error into a
voltage, which is then filtered. This is done in order to achieve the proper driving
signal that will precess the seeker antenna in the direction that acts to reduce ~,,rfe,,
to zero. The seeker dish angle and the LOS angle are equal in the case the seeker
tracks the target precisely. However, the actual angle tracking error E = ~ ~ ~ +f i . ~
u, + v, (see Figure 2-13b, v, = measurement noise) is the angle between the
apparent target LOS and the seeker center line measured by the gimballed seeker
dish. The seeker center line is positioned by the executive computer system and
periodically updated based on the reconstructed LOS angle 6. The pitch angle 0,
measured by the IRU, has added to it the actual gimbal angle u2measured by the
gimballed pickoff. This provides UJ (that is, ud = 0 + a,), which is then added to
E to obtain the measured LOS angle u as follows:

u = e + ud = brmr+ u, + = LOS angle measurement (2-10)

40 MDASE of NGC Processing Chap. 2

where the apparent LOS angle u as seen by the homing eye is equal to the true LOS
angle plus the aberration angle u,, as shown in Figure 2-10. In order to simplify
the analysis here, it is assumed that u, is a linear function of the gimbal angle u,
such that

Making a correction to include radome effects in Equation (2-lo), the tracking error
measurement (see Figure 2-10) then becomes

- ~ r , e - e - ~ g + ~ r + V , = ~ ~ - u , + ~ , + ~ ,

Equation (2-11) is substituted into Equation (2-12) to yield

The angular relationship of Equation (2-13) and the dish dynamics from Figure 2-
13a are employed to develop a seeker block diagram with radome effects shown in
Figure 2-13b. Since the radome-induced error here depends not only on the LOS
angle but also on 0 as well, 0 is used in the radome compensation algorithms for
stabilizing the (negative) radome error slope. T o enlarge the stability region for the
positive error slope and for cross-plane coupling slopes, the radome compensation
algorithm is inserted at the output rather than the input side of the guidance pro-
cessing algorithms [Yueh, 1983(a); Yueh and Lin, 1984, 1985(a)]. Integrating 0 to
get 0, the latter quantity is then multiplied by the radome error r and used in the
radome coupling block of Figure 2-2 to compute the LOS angle measurement.
Combining Equations (2-10) and (2-ll), another simplified equation that can be
used to calculate the LOS angle measurement is

where a,,,,,is given by Equation (2-5a). The look angle is defined as

Assuming radome aberration error u, is explicitly dependent on the look angle U L ,

then the time derivative of the measured LOS [Equation (2-lo)] to the first-order
linearization approximation can be written as

where the radome local slope is

The importance of the seeker as a missile subsystem becomes evident when one
considers what it must be able to do. The seeker must be able to track a target,
remain spatially stabilized or go into a search mode if target track is lost, and measure
Sec. 2.2 Target Signal Processing 41

a guidance signal. Most seekers have a stabilization loop (see Figure 2-13b) to isolate
body rotations and track loop to maintain the seeker axis along the missile-target
direction. These loops, in conjunction with the appropriate signal processing, permit
the measurement of the guidance signal. The knowledge of the seeker dynamics
together with information about the radome characteristics and noise sources would
be required to develop a detailed model of this measurement process. In order for
the tracking and stabilization functions to work properly, the seeker slew-rate ca-
pability for a BTT missile must be compatible with thc missile's banking motion,
being ablc to deal with potctltially large look angles. It is important also to assess
the impact of banking maneuvers on track loop requirements. A detailed seeker
analysis must include modeling of the inertial coupling between the two seeker
channels and the effects of friction. When analyzing missile banking maneuvers, one
must also consider the effects of polarization. In the case of a semiactive, linearly
polarized seeker with linearly polarized radiation, the seeker gain becomes zero when
the polarizations are 90 deg out of phase. This situation can arise, for example, when
the missile undergoes a 90-dcg bank. Also, as the relative polarization of the seeker
and incoming radiation change, there can be serious problems having to do with
changes in radome aberration error. That is, if the radome is designed for a particular
polarization, a change in this angle may produce larger aberration errors. An active
seeker may also be subject to polarization effects when used against a large target
such as a ship, for example. If the target's reflectivity is a function of the signal's
polarization, the aim point may be altered as the missile banks. Should the guidance
system mistakenly interpret this rapid change in aim point as a spike in the LOS
rate, it may cause the missile to miss the target [Riedel, 19801

Target tracker (seeker) modeling. The LOS rate u is corrupted by

bias ub and gyro sensitivity as shown in Figure 2-2. It is convenient at this point
to consider the tracking loop time constant T,, the stabilization loop gain K,, and
the gyro sensitivity gain K?,as shown in Figure 2-13b. In general, if the tracking
loop time constant T, is too large, the seeker will be unable to track highly maneu-
vering targets. If, on the other hand, this constant is too small, there will be problems
associated with excessive noise transmission and, hence, a stability problem. T h e
tracking loop time constant must therefore be chosen somewhere between these
two limiting values. he stabilization loop gain K, which is the loop crossover
frequency should be large (that is, perfect seeker dynamics) so as to speed up the
guidance system, with a practical upper bound placed on it so as to stabilize the
seeker well. Because it will have almost no influence on the seeker track loop dy-
namics if K, is large, the seeker dynamics can be approximated by
Without radome effect (see Fig. 2-13a) u'dlu = W,S/(S w,) + (2- 16)
With radome effect (see Fig. 2-13b) u'd = [(I + r)u,,, - re] w,/(s + w,)
that is, a,,,, is passed through a target tracker (for example, seeker), composed of
a differentiator followed by a first-order lag, to produce a seeker dish rate command
which is approximately the LOS rate I?.
42 MDASE of NGC Processing Chap. 2

Tracking error guidance system seeker model. The transfer func-

tion of the seeker as it appears in Equation (2-16) is more representative of a tracking
error guidance system shown in Figure 2-13b. Figure 2-14a illustrates a typical
tracking error guidance seeker loop in which an error signal is produced by the
angle error detector and is proportional to the misalignment of measured target
angular tracking error E. This error signal'is then input to a proportional plus integral
amplifier whose transfer function is

A rate gyro feedback must be included in the design in order to improve the stability
margin in the Nyquist sense and to help reject outside disturbances [Garnell, 19801.
The output of gimbal dynamics plus pitch rate 8 is the antenna (dish) rate ed in
body coordinates, and, since the feedback is defined as dish angle and not rate, the

Total LOS
Angle Noise
v- Anele Error Amalifier Power

Rate Gym 1

Figure 2-14 (a) Typical Tracking Error Guidance Seeker Loop (b) LOS Rate Guid-
ance System Seeker Model
Sec. 2.2 Target Slgnal Rocesslng 43

loop is closed by the implied integration of u d . An evaluation is now made of the

quantitics T I and K I K2K3 which fall under the designer's control. The generalized
tracking guidancc system seeker model is illustrated in Figure 2-13b in which the
tracking loop time constant T, is proportional to the torque loop gain TI of the
seeker. The PNG guidance system transfer function is approximated by
u = V [ ( + T , s ) ( ~ + TJ)] (3-17b)

LOS rate guidance system seeker model. An alternative to the track-

ing error guidance system sccker model is the LOS rate guidancc system seeker
model which measures LOS rate u. The target is tracked by the seeker antenna
assembly, which is stabilized with respect to missile body motion. The measured
LOS rate is given by the sum of the time derivative of the angular tracking error
E and the antenna rotation rate u d throughout Equation (2-10). That is,

Thus, to obtain a measurement of u, the signal processor provides i ( s u,) and

seeker-mounted rate gyros provide u d . The measured LOS rate u is actually very
close to the geometric u,,,,. There are roll-stabilized radar seekers that generate the
LOS rate u for each axis using Equation (2-183). While it can be shown that the
measurement of u becomes independent of the seeker track loop dynamics under
this operation, it is also true that this method may lead to noise amplification owing
to the process of differentiation. Moreover, the differentiation is electronic and is
performed on a signal measured in seeker coordinates. The process of generating
u using Equation (2-18a) together with all seeker low-frequency signal processing
must be examined in the context of being suitable when the missile and seeker cease
to be roll stabilized, but can bank at a rate of a few hundred degrees per second. A
LOS rate guidance system seeker model is presented in Figure 2-14b. The PNG
system transfer function is approximated by

Parasitic feedbacks. As illustrated in Figures 2-2 and 2-13b, body-

rotation rates always have the effect of corrupting LOS rate measurements. Imperfect
sensor stabilization or unmatched signal processing gains are only two factors that
may act to produce undesirable perturbations on the LOS rate. A parasitic loop may
develop in which body-rate induced perturbations on the guidance signal cause a
perturbation in the commanded and achieved accelerations and, hence, additional
body rate. The result of this situation is that the system can become unstable and
may suffer degraded performance [Riedel, 19801. Both the body rate and body
acceleration are parasitic feedbacks, owing to the necessity of an aerodynamic missile
to pitch to an angle of attack to be able to maneuver. Because of the radome refraction
effects, the autopilot and seeker dynamics are coupled through the missile body-
rate feedback signal. This creates an attitude feedback loop in which the missile
responds to a target LOS change by maneuvering. The resulting rotation of the
44 MDASE of NGC Processing Chap. 2

missile body causes an apparent additional change in the LOS angle, which closes
the loop. This feedback increases as the altitude increases. When the values o f the
radome slope I. are large, they represent the worst effect of an undesirable feedback
path through the gimbal angle as shown in Equation (2-13). Beside the radome
effects, the stabilization loop is fed with a false angular rate due to missile maneuvers.
The gyro acceleration sensitivity drift produces a direct unwanted feedback path
with a gain K, from the achieved acceleration to the measured antenna rate in space.
If the gyro mechanism is not perfectly balanced about the output axis, any linear
acceleration normal to that axis to which the gyro is subjected will cause an erroneous
rate signal to be generated. Such a signal will result in a tracking error that in turn
commands an additional lateral acceleration to close the loop. This effect was first
brought to light during a flight test in which the missile flew a helical path to an
excessive miss distance. After conducting an analysis, improved mass balance pro-
cedures were put into effect. In addition, the gyro was mounted in a direction that
minimized the miss-distance impact [Fossier, 19841. The preceding two parasitic
feedbacks become important at high closing velocities occurring in short-range or
high-altitude flights. Thus, the design of the seeker must take into account the
seeker's own performance characteristics and provide sufficient stability margins to
handle unwanted feedback paths.
As will be shown later in Examples 3-3 and 2-4, the effect of radome error
feedback on the control system is determined by the sign of the radome error. The
effect of a degenerative error (positive), which is one that tends to reduce the input
LOS rate, is to slow the system response and to reduce the autopilot's stability
margin. A regenerative (negative) error may bring about guidance instability at very
low frequency. Although both can result in increased miss distance, a regenerative
error is of greater concern inasmuch as low-frequency oscillations are characterized
by large displacements of thc missile. In a situation somewhat similar to what occurs
with the radome error, a feedback loop is formed by in~perfectionsin the antenna
stabilization system (sec Figure 2-13b). This causes a change in the antenna direction
as the missile body attitude, or missile acceleration, changes. In early systems. the
stabilization loop consisted of rate gyros mounted on thc back of the antenna, with
their outputs electronically integrated to drive a hydraulically actuated gimbal sys-
tem. Owing to the servo's finitc frcqucncy response, the antenna is unable to kccp
perfectly stationary for body motions a t frcqucncies of primary concern, and the
resulting motion gcncratcs a borcsight error cquivalcnt to that caused by positivc
radon~eslopc. One aspcct of the solution to this problem involved improving thc
hydraulic valve rcsponsc so that the head stabilization loop could be closed to a high
enough frcqucncy to rcducc this effect to acceptable icvcls.
Another aspect of thc solution was clcctronic in nature and centered around
the closed-out frequcncy of the stabilization loop. It was realized that this frequency
mattcrcd only to the extent that it affects thc gain in the stabilization loop at fre-
quencics of interest, typically 1 to 3 Hz, and that it could bc increased by changing
the elcctronic integrator to provide a - 2 slopc instcad of a - 1 slopc, keeping in
mind loop stability considerations. The resemblance to radome slope was brought
into grcatcr focus when it was discovcred that purposely limiting the DC gain of
Sec. 2.3 NGC System Deslgn and Analysk 45

the stabilization loop, as opposed to allowing perfect integration of the rate gyro,
created the same cffect as a positive radome slope. Thus, one could compensate for
the negative radome slope at the very low frequencies that produced large miss
distances. A rather serious accuracy problem was alleviated by the implementation
of this simple compensation scheme. Chapter 8.4.3 presents the design cquations
for PNG with thc previously-mentioned parasitic feedback. Thc cffect of body
bending is yet another type of parasitic feedback. In contrast to bcing a separate
parasitic loop, this cffect is simply a high-frequency autopilot instability in which
body bcnding is detected by the autopilot as a motion ofthe missile. In early systems,
the autopilot rate gyros were mounted as closely as possible to the nonrotational
point for the first bending mode. This in conjunction with the electronic filtering
employed made it possible to avoid autopilot instability [Fossier, 19841. Details of
this structural vibration control are presented in Book 3.


The N G C analyst works with all members of the design team. He or she must be
able to work from simple to very complex analyses and simulations, and to provide
preliminary and firm requirements and designs.
2.3.1 Guidance FilteringlProcessing

Simplified pursuer dynamics. A functional block diagram of the PNG

kinematic loop is shown in Figure 2-13a with the guidance computer defined pre-
viously in Figure 2-6. Seeker dynamics and corruption of the LOS rates due to
body motion coupling through the radome are modeled in Figures 2-lja(iii) and
Guidance Jdnematic loop. Consider the block diagram corresponding
to Figure 2-15b which is simplified from Figure 2-2. The target position mea-
surement is
"' = Yr + Yg (2- 19)
where y , corresponds to the target maneuver, which is generally considered as a
random variable, and y, is the target glint/scintillation noise (generated by the input
equivalent radar noise), which is a property of target radar return signal and arises
from sources that physically enter the system.
Noise inputs. The two target noise inputs in Figure 2-15b are denoted
by y , and U ~ (angle
N noise of the tracker). These noises, which corrupt the obser-
vation, induce miss distance, and impact pursuer acceleration requirements, are as-
sumed to be uncorrelated white-noise processes. In Figures 2-15b and 2-2, u ~ ~ ( t )
is equal to u,,,(t) + u f ( t ) + uc(t).Represented as a linear displacement at the target,
y , has a PSD of +,, and, like miss distance, is divided by rhe range to become an
angular glint noise u,,i,,, with PSD
MDASE of NGC Processing Chap. 2

a) Simplified P u r s ~ Dynamics
TNE U S Guidance & N D ~ Guidance Conlpuvr FCS
Angle Gs ('1 , ($1
G Air+c & Pmpukion
i) I Total LOS Angle N o i s
I *kc, Dirh I
qne I Ratc Command
I b;l-a I

Pursuer mnamicr

Punuer D y o a k
(See F i e a)

Figure 2-15 (a) Sirnplificd I'ursuer Ilynarnics (h) Trajectory Ilynarnics Model (c)
Equivalent Trajectory Dynamics Model

to be input to the scckcr. It is pointed out here that, evcn under the assumption o f
stationary scintillation noisc at the target, this becomes nonstationary at the seeker
by virtue of division by R ( t ) . Target scintillation is commonly assumed to be aptly
represented as stationary band-limited noise, thc amplitude and bandwidth being
functions of the physical dimensions and motion of thc target. Inasmuch as band-
Sec. 2.3 NGC System DesIgn and Analysis 47

limited noise and filtered white noise possess the same statistical properties, the
simulation can be performed by introducing a simple lag filter with ~ C L T= l/wg
as a correlation time constant between w, and y , as shown in Figure 2-2. The PSD
of y , is
+,\. = 27CLTrCLT, ~ C L T= glint noise variance (2-2Ob)

The filter was excluded in Figure 2-15b as a matter of simplicity because it falls
outside thc homing loop and has no effect on the system adjoint [Peterson, 19611.
The net return signal from a target with many radar reflectors can be modeled as
a movement of the apparent radar target position. It is typically modeled as a Gaus-
sian random variable only with zero mean and variance rcLT. that is,

where (r~gET)is 112 to 118 the target's dimension perpendicular to the LOS. That
IS, rg;T = WJa, where a = 5 is the rule o f thumb and can vary between 2 a n d 8
according to data, and W , is the wing span of the target [Alpert, 19881. Range-
independent noise (KIN) uf is the noise inherent in the receiver. It is independent
of target return power and is caused by many sources, only some of which include
signal processing effects, cross polarization of returns, quantization effects, and gim-
bal servo drive inaccuracies. uf is modeled as

of - iV(0, rf) with equivalent white RIN PSD +f = 2?frf (2-20d)

where ~f is the RIN correlation time constant. Thermal noise a,,,is the receiver
noise, which is a function of the strength of the target return, and hence the signal-
to-noise ratio. u,,, is modeled as
u,,,- N(0, r,,) with equivalent white receiver noise PSD +,,, = 27,,r,,, (2-20e)
where T,,, is the correlation time constant of u,,,and r,, is the range-dependent noise
variance governed by
(Semiactive) ~,,,sA = r,,,~(RIRO)', (Active) r , , ~ = r,,~ ( R I R o ) ~ (2-200
where Ro is referred to as the target-to-pursuer reference range. The clutter noise
u, is modeled as

u, - N(0, r,) with equivalent white clutter noise PSD +, = 2~,r, (2-20g)
where 7, is the correlation time constant of a,. Glint noise, receiver noise, and RIN
are the important contributions to tracking accuracy when radar is used. Decreasing
the range to go increases the glint noise but decreases the receiver noise. The variance
~ T Nand the spectral input + T N for the angle noise arNare determined by the
characteristics of the seeker and are of the following forms, respectively:
48 MDASE of NGC Processing Chap. 2

The total LOS angle measurement noise is then

Vm = UTN + uglinr = urn + uf f uc + Ug*llittr (2-22)
The angle noise variance and spectral input are of the following forms:
v, = v,,, + rf + r, + rCLTIRZ

In addition, the tracker noise, or.., can be added to target position noise, y,, so
that the measurement equation is modified from Equation (2-19) as:
z(t) = y T +n = target position measurement (2-24)
where n = target position noise = o , R = y, + R U T N , and the variance and PSD
of n are of the following forms, respectively: '

r,, = YTN R2 + YCLT = [rm + r, + v , ] R 2 + ICLT

+tt= ~ T , v R+~ 4 s = [@r,z + 4f f +c]R2 +
Thus, Figure 2-15b has been modified to become equivalent to Figure 2-15c.
Nonnoise inputs. The initial condition at the start of homing (launch,
transient, and so on) is the nonnoise inputs to the system. The initial seeker error
and the initial turning rate of the missile are both assumed to be constant random
variables. In certain instances, either of these two errors may be zero. For example,
the seeker initial error input becomes zero if the seeker is lockcd on and settled out
at the start of homing. Similarly, the initial turning rate input becomes zero if the
missile is not maneuvering when homing begins. Other quantities assumed to be
constant random variables include initial heading error and detcrn~inistictarget ma-
neuvering inputs [Peterson, 19611.
NGC system design requirements and design considerations.
Consider the possibility of choosing a guidance and control law such that the system
within the dotted box of Figure 2-1% looks like a guidance filter. Then, pursuer
lateral position output y,,, is controlled by the guidance computer and FCS to min-
imize E [ ( ~ -T y , ) T ( y T - Y , ) ] , that is, minimum miss. The guidance and control
system must get the missile close enough to kill the target when the warhead ex-
plodes. Therefore, an appropriate measure of guidance performance is miss distance.
which is defined as the minimum distance between the missile and target. Besides
minimizing miss, other goals that the NGC system must attempt to accomplish
include maximizing the following: the range of operating conditions, the allowable
environmental disturbances, the reliability, and the effectiveness. At the same time,
the N G C system must be designed so that both cost and sensitivity to target tactics
are minimized. When viewed from the standpoint of meeting these objectives, the
guidance and control problem becomcs a matter of designing a compensation filter
that optimizes system performance with respect to accuracy, stability, speed of
response, and external or environmental disturbances. A high relative stability is
Sec. 2.3 NGC System Deslgn and Analysls 49

measured in terms of good gain and phase margins. Concerning speed o f response,
the system must be suitably fast so as to wipe out the nonnoise inputs with small
guidance time constant T,,. As for environmental disturbances, the filter must be
able to filter whatever noise inputs arc present to acceptable levels [Axelband, 19861.

Advanced multivariable control system design. T o begin with, con-

sider Figure 2-16 which describes a standard procedure for designing a robust,
integrated control law (see Books 3, 4, and 5 of the series). After a nonlincar simu-
lation modcl is developed and the requirements for robustness and performance are
determined, the results are used to conduct an open-loop analysis. As shown in
Book 1 of the series, the purpose of the open-loop analysis is to determine the
significant dynamic characteristics and interactions of the airframe, propulsion, and
control surfaces. It consists of nonlinear and linear simulation, as well as stability,
covariance, observability, controllability, and frequency response analyses. The re-
sults of these analyses are used to define the design model. Once the design model
is defined, control law synthesis can be performed to develop control law algorithms.
Control law synthesis consists of: (1) feedforward controller design to enhance com-
mand response; (2) feedback controller design to satisfy stability and robustness
requirements; and (3) controller order reduction to minimize implementation com-
plexity. These results are then used to conduct closed-loop analysis which, as Figure
2-16 shows, also considers directly robustness and performance requirements. The
closed-loop analysis, which makes use of the control law combined with the open-
loop simulation model, consists of nonliner and linear simulation, as well as stability,
covariance, and robustness analyses. This is done to ensure that the control law

Model helupmnent
' Wind tunnel test
* Physical laws
Design Specifieatim .
Military speciliatiow
* Flyingqualities
Stability requirements 4--
SyPtem identification M i u distance 4-
Model validillim Performance
Open L m p Analysh
Design model
c Model reduction
* System limitation
Conlml Syslem D e i g n

Advanced guidanceand control laws
'Design knowledge h w

Simulation and Performance Evaluation

l i m e and frequency responses
I.inearand nonlinear Simulatim

Fast prototype Figure 2-16 Robust Integrated
Control Law Design Procedure
50 MDASE of NGC Processing Chap. 2

design satisfies control functional requirements. If these requirements are not ade-
quately met, it may happen that further open-loop analysis is required. However,
it may be that the algorithms developed in the control law synthesis stage simply
need to be refined, in which case the design model does not have to be altered.
Some of the methods available for advanced multivariable control law design are
shown in Book 3 of the series.

Classical versus modem guidance and control. As shown in Figs.

2-6 and 2-17a, low-pass filters attenuate the noise component of the sensor signal,
given the frequency characteristics of target signal and noise. Also, as can be seen
from Figure 2-17a, separation principle allows one to design the optimal estimator
and the advanced guidance law independently. Optimal estimators such as Kalman
filters ideally separate the target signal from the noise by using information about
target (and pursuer) dynamics and noise covariances. The objective of an optimal
estimator is to provide accurate estimates of all states and model parameters required
for the advanced guidance law without significantly increasing the sensor require-
ments (and therefore cost). The advanced guidance law provides the optimal steering
commands assuming a noise-free environment. Figure 2-17b is a functional com-
parison of how the current modern approach differs from the classical approach to
missile guidance and control. Classical approaches have become firmly entrenched
in guided missile designs because they worked well in the rather benign environment
of past target engagements and were easily implemented with analog circuitry.
Because of such precedence, missile designers have often tried to satisfy the increased
performance requirements of modern missiles by increasing the complexity of as-
sociated hardware such as seekers, gyros, accelerometers, airframes, and engines.
Improving performance nearly always involves a trade-off between more sophis-
ticated hardware or more sophisticated software. While such approaches in many
cases have improved performance, the increased hardware sophistication almost
always results in increased costs. In fact, the resulting cost has often been so high
that the systems either were never developed for operational use or were purchased
in small quantities [Gonzalez, 1979(a), (b)].
In spite of many classical terminal guidance and control laws which are still
being used for tactical missiles, each characterized by varying degrees of perfor-
mance, complexity, and seeker-sensor requirements, the increascd accuracy re-
quirements and morc dynamic tactics of modern warfare render contemporary guid-
ance and control laws unsatisfactory in many applications. Whilc many modern
guidance and control system des~gntechniques have yiclded better solutions than
those obta~nedfrom a classical approach, these remain nevertheless overly complex
and are not well undcrstood by industry as a whole. As a result, advanced guidance
and control sysfetn design, as extensively discussed in this book, is an ideal approach
which combines the most attractive features of both classical and modern guidance
and control system design techniques. This advanced approach, coupled with the
advent of new theoretical methods and low-costlhigh-speed microprocessing tech-
niques, makes it possible to provide better effectiveness, reliability, tremendous
increases in missile performance, and greater leeway in other subsystem performance
for the same or less power, weight, and cost.
See. 2.3 NOC System Design and Analysls


C Low C A N
LlOn M T Z


nus rson~orr~~


L1 u









Figure 2-17 Classical vs. Modern Guidance and Control (a) Guidance Filtering and Pro-
cessing (b) Functional Comparison ( ( a ) j o r n [Conzalez, 1979(b)] withpennissionjom AGARD
( b j j o m [Conzalez, 1979(a)], CC 1979 IEEE)
52 MDASE of NGC Processing Chap. 2

2.3.2 NGC Stability and Performance Analysis

The guidance analysis model shown in all parts of Figure 2-15 is composed of
elements representing the homing kinematics plus other systems that affect the hom-
ing guidance dynamics. The root-locus method and frequency response analysis can
be applied to guidance system analysis. Performance analysis techniques are used
to evaluate the NGC performance. The guidance system transfer function from a,,,,
to A,,, is shown in Figure 2-15a. The stability analysis pertains only to the guidance
system transfer function, whereas the miss-distance analysis will apply to the entire
homing loop of Figure 2-15b. The homing loop shown in Figure 2-15b can be
used to obtain a clear understanding of the requirements of an analysis model, and
to obtain quantitative comparisons of miss distance as a function of the significant
guidance system paran~eters.Such information is then used to evaluate the effects
of N G C parameters on missile performance. Miss distance is brought about by the

dynamics of the guidance system transfer function, the simplest representation for
which is shown to be a fifth-order binomial [Nesline and Zarchan, 1984(a)]. The
desir.ed/actual effective navigation ratio A, which varies due to changes in the flight
condition, is one major determinant of homing system performance. Another major
determinant of homing performance is the guidance system tinte constant T,, which is
the sum of system constants. The ejfertivegrridance systenl titne constant, T;, is defined
as the guidance system time constant with radome effect. Examples 2-3 through
2-5 later will discuss radome error and its severity on the guidance and control
responsiveness of homing missiles. As will be shown in Example 2-4, the overall
system stability needs to be studied before computing the time constant of the FCS
because the timc constant is the major factor that affects the miss distance. The trend
of the effect of cach feedback loop on the linear, time-\.arying gain systcm is clarified
by the preceding discussion using root locus. The adjoint simulation is required
because all the transient behavior in the miss-distancc response cannot bc predicted
by only the simple root structure. The most efficient method of parametric study
can be realized through a combination of the root-locus and the adjoint simulation
techniques. The former is initially used to idcntify the region of parameters for
system stability, while the lattcr is used to check thc conclusions dcrived from the
root-locus study, concentrating on the optimum choice of N G C gains.

Example 2-1: Guidance kinematic loop stability analysis. A vcry

good guidancc stability analysis has already bccn put forth prcviously in Book 1 of
thc scrics. Thc opcn-loop phase and gain frcqucncy rcsponsc charactcristics arc gcn-
eratcd so that the stability charactcristics of thc guidance kincrnatic loop can be
examincd. Once specific values arc assigned to cach paramctcr, the contribution to
this responsc by the linear clcmcnts in this loop arc easily evaluatcd. The opcn-loop
transfer function in Figurc 2-l5b with the simplified pursucr dynamics defined in
Figurc 2- 15a(i) is

Substituting jo for s in Equation (2-26) for sinusoidal input yields for magnitudc
Sec. 2.3 NGC System Deslgn and Analysls 53

and phase

where GA(w) is the autopilot gain and L$A(o) is the autopilot phase response.
Although the linear mode of the autopilot can easily be solvcd, it is difficult to
analytically obtain the contribution to this response by the nonlinear autopilot. Thus,
the nonlinear autopilot phase and gain characteristics are obtained through computer

Example 2-2: Comparison of statistical digital simulation meth-

ods. This example begins with the simplified model of an intercept homing loop
shown in Figures 2-15a(ii) and 2-15b. The FCS and the airframe and propulsion
dynamics are approximately represented by a single tag (7, = llw,) transfer function
CA(S) = A,,lu, = 1/(7,s + 1) (although a second-order representation seems more
reasonable) to include their effects in the guidance gains. The guidance system trans-
fer function without radome effect is A,,lu = AV, s/(l + ~ ~ where ~ 5 T) , = T, +
7, + T,. Included in this simplified model are: (1) effective navigation ratio A =
3; (2) closing velocity V, = 3,000 ftlsec; (3) pursuer velocity V,, = 2,300 ftlsec; (4)
7, = l / w y = 0.1 sec; (5) seeker tracking time constant 7, = l/w, = 0.05 sec; (6)
noise filter bandwidth k = l / i , = 10 radlsec; and (7) flight time t f = 3 sec [Zarchan,
1979(a) & 19881. The target maneuver model is assumed to be a Poisson jinking
maneuver shown later in Figure 2-21e where B = RMS target acceleration = 161
ft/sec2, and A, = the target maneuver bandwidth = 0.2 sec-'. The PSD of the
white glint noise u, and the white range-independent fading noise uy are denoted
by +, = 4 (ft2/Hz) and +f = 1 x 10-16 (rad2/Hz), respectively. This example is
used in Chapter 3 to demonstrate the utility of each of the performance analysis
techniques via digital simulation.

Example 2-3: Radome error-induced miss-distance predictions.

Following Example 2-2 and Figures 2-15a(iii) and 2-15b, included in this simplified
model are: (1) effective navigation ratio A = 3.5; (2) closing velocity V, = 4,500
ftlsec; (3) pursuer velocity V, = 2,500 ft/sec; (4) T, = l / o , = 0.45 sec; (5) seeker
tracking time constant T, = l/w, = 0.1 sec; (6) T, = l l k = 0 sec infinity bandwidth;
(7) aerodynamic turning rate constant T , = 2 sec; (8) flight time tf = 5 sec. The
model developed by Murray [I9841 and presented here begins with the fact that
system disturbances that cause miss distance result also in gimbal angle motion. The
effective slopes r are reduced by this motion, while the effective noise q is increased.
This example, largely fdllowing the excellent article by Murray [1984], looks at a
linear radome error model that spans the gap between nonlinear and linear radome
error miss-distance predictions. Included in this model are an effective slope and an
effective noise variance. Gimbal angle motion, which is a function of guidance
system disturbances such as target maneuver and range-independent noise (RIN),
determines the effective slope and effective noise. In the treatment presented here,
54 MDASE of NGC Processing Chap. 2

analytical expressions for effective slope and effective noise are derived for a sinu-
soidal radome error. The guidance system transfer function with radome effect is:
7; = effecti4e guidance system constant
4 rn =
A' Vcs =1 + ~ A V / V , ) ' ( T T , ) I I (+~ A V V (2-28)
U l + Tis ' A! = A
= effective navigation ratio
1 + rAV,/V,
where T~ = T, + +
T, T~ Figure 2-15a(iii) illustrates how the radome slope changes
the missile system dynamics. The radome slope r appears twice in A,,,/u. A' is
altered in terms of r effect by a factor of 1 over the quantity 1 + rAV,/V,,. For
positive radome slope, A' is reduced and the response time of the system is increased,
the effect of which is to increase the miss distance due to target maneuver. T h e
opposite occurs for negative radome slope, the result of which is to increase the
effect of A' and thereby increase the miss distance due to RIN [Fossier and Hall,
19671. System damping is reduced in the case of a negative radome error when
higher-order terms are concerned. Figure 2-18a shows a typical miss-distance sen-
sitivity to r. The figure shows that there exists a negative threshold value of r, below
which miss rapidly escalates. This value depends on the flight condition through
a/-$ and on the control system design through the value of T,. Given the variation
of a/-$with altitude, the value of T, must be increased appropriately so as to maintain
acceptable performance [Fossier, 19841. The radome slope r also enters in the de-
nominator of Arn,lu, where the time constant (s coefficient) T; is modified by a
factor of r T,A V,/ V,,,. The effect of a positive r is to increase 7i, thereby increasing
the miss distance due to target maneuver. A negative r., if it is sufficiently large,
will cause T: to go negative, resulting in an unstable guidance system. The cffect of
r on 7; is of greater concern at high altitude where T , is large. Given the parameter
values, the missile system in Figures 2-lja(iii) and 2-15b gocs unstable for r =
-0.045. This second-order missile system remains stable for all values of positivc
r, a fact whlch is not significant for the following reason. By adding to the model
a typical third-order autopilot as will be given later in Examples 2-4 and 2-5,
together with a first-order noise filter, r = 0.095 is required for missile system
instability. However, fcr positive slope values much less than 0.095, the miss dis-
tance due to target maneuver becomes unduly high.

Miss due to radome slope. Target maneuver and RIN are thc two dis-
turbanccs evaluated. As Figure 2-18b shows, the miss distance increases with ra-

Figure 2-18 (a) Representative Effect of Radome Boresight Error Slope on Miss Distance
(b) Miss Increases with Radonie Slope (c) Miss Is Less with l'eriodic Radoinc (d) Target
Maneuver Miss Is Less with Periodic lladome (e) Gimbal Angle Motion Deterniines Effective
Slope and Effective Noise (f) Target Maneuver and Noise Cause Gimbal Angle Motion (g)
Effective Slopes and Effective Noise Are Derived by Mininiizing Mean Square Error ((aJ,fmitr
/Fossirr, 1984/, 8 1984 AIAA (h)-(g) jam IMuwoy, 19841 urith permissiotr j o m A A C C )
.- ~



8 2.


1 2 3 4 d


.O.W .0.02 0 0.02 0.04 0.06 0.08

SINUSOIDAL q = r, sin (+(up *))


- -

- ...- - . .-
!0.08.o.m - 0 . ' ~ .o.b2 o 0.b2 0.04 0.05 0.08 NOISE:
(C) t

Effective Slope for

Target Maneuver Effective
Gimbal Angle Due to
Target Manewer

Gimbal Angle Due to

Range Independent Noise Oggf
Effec~iveSlope for
Range Independent Noise
56 MDASE of NGC Processing Chap. 2

dome slope. This miss due to target maneuver increases with positive slope, whereas
the miss due to RIN increases with negative slope. For this example, the initiation
time of the 3-g step target maneuver is uniformly distributed with 25 samples over
the 5-sec terminal flight time. The amplitude standard deviation of the RIN is 2
mrad within a bandwidth of 1.8 radfsec (7 = 0.55 sec). In Figure 2-18b, the RMS
value for zero radome slope is 1.0 because all RMS miss distances in this example
are normalized by the same factor.

Miss i s less with periodic radome. Actual radome is periodic with

gimbal angle [Murray, 19841. As shown in Figure 2-18c, periodic radome error
results in smaller miss distance than occurs with constant maximum radome slope.
The figure shows that miss is significantly reduced for a sinusoidal radome error
with a 30-deg period. Specifically, the miss distance due to target maneuver is less
for a periodic radome error. Miss distance is plotted against target maneuver time
in Figure 2-18d for three types of radome error. The first corresponds to r = 0,
the second to r = - 0.045, and the third to a 30-deg period sinusoidal radome with
a maximum slope of 1 0.045 1. At the initiation time of the target maneuver, the
gimbal angle is such that the radome slope has the maximum negative value of
-0.045. While the RMS target-induced miss is 0.7 for r = 0, it is only slightly
higher with a value of 1.1 for the case of the 30-deg period sinusoidal radome with
a 10.045 1 maximum slope. These ~ a l u e scan be compared with the RMS target-
induced miss for the case of constant -0.045 radorne slope, which is much higher
at 3.4. Actually, this adjoint continues to increase with increasing flight time owing
to the fact that the missile system is marginally stable for r = -0.043.

Effective slope and effective noise. As demonstrated in Figure 2-18e,

the effective slope and effective noise of a radonle error are determined by gimbal
angle motion. In this figure, the effective slope r due to gimbal angle motion is less
than the maximum slope. The difference between the nonlinear radome error, u,,
and the effective slope approximation, r u,, gives the effective noise, q. Gimbal
angle motion is brought about as a result of missile system disturbances such as
target maneuver and RIN. Gimbal angle motion, particularly that component which
owes itself to target maneuver, can be a substantial fraction of a radome error. This
example looks at a sinusoidal radome error, target maneuver, and RIN disturbances
as shown in Figure 2-18f. The sinusoidal radome is characterizcd by amplitude,
period, and phase of ro, P, and @, respectively. The gimbal angle is incremented
as a rcsult of target maneuver by the value u,l over the flight time. A uniform
distribution with a (1/3)"2u,,/2 standard deviation is assumed. Figure 2-18g shows
the two effective slopes, one for each of the two nlissile system disturbances. Min-
imizing the mean-square error, q, yields
r~ = ur(aE)* u ~ ~ l ( u , ~ ~ l l 2r, ) ,= u,(ux) * ~,~f14. (1-29)
02 = u,Z - r+u:I/12 - r34, b =i
Ssc. 2.3 NGC System Design and Analysls 57

Evaluating Equations (2-29) for the sinusoidal radome given in Figure 2-18f results
in thc following normalized effective slopcs and effective noisc variance:

~.WAX -0s( ) [[sin (y)

- rf- - COS

where r ~ f . 4= ~2 ~ r , ~ and

- - sin
2 ~ 4 ~
(Y) sin
ro P
Because it is not significant for long flight times, the bias is neglected. Equations
(2-30b and 2-30c) are plotted in Figure 2-19a for zero phase a. As the target
maneuver and RINs increase. the effective sloves decrease, whereas the effective
noise increases. The effective slopes are much less than the maximum slope, M MAX,
as the gimbal angle increment, r,~, due to target maneuver approaches and exceeds
the radome error period, P. As this occurs, the effective noise approaches 0.7 of the
radome error amplitudes. The greater gimbal angle oscillation, + ' I 2 , due to RIN
accelerates this effect of reduced effective slopes and increased effective noise.

Linear radome miss prediction. Equation (2-31) represents the linear

radome miss-distance prediction, which is the RSS of the components due to system
disturbances and the component due to effective noise

RMS Miss RMS Miss

RMS Miss
Target Maneuver RMS Miss: fM(r = rT); RIN RMS Miss: fN(r = rf) (2-31)
Effective Noise RMS Miss: fN(r = 0)
MDASE of NGC Processing Chap. 2

D E V I A T I O N ldegl

2 mrad

7.5 deg 1---

.0.06 -0.04 .0.02 0 0.02 0.04

0 --180d.p


. o . m h ,

0 5 10 15 0 5 10 15


RAWME 'J 4- rmr



1 0
.o.a . 0 . b .o.b2 o 0.b2 0.k o.'a 0.h

Figure 2-19 (a) Effective Slopes Decrease and Effective Noise Increases with Target Ma-
neuver and Range Independent Noise (b) Effective Slopes are Reduced with Range Independent
Noise (c) Effective Slope and Noise are Defined by Gimbal Angle Standard Deviation (d)
Effective Slope and Effective Noise Predict Range Independent Noise Induced Miss (e) Target
Maneuver Causes Gimbal Angle Motion (1) Effective Slope Due to Noise Is Reduced with
Sec. 2.3 NGC System Design and Analyski




1 1


0 5





I I I I I I 1
- 0 . E -0.@4 -0.02 0 0.02 0.04 0.06
1 0 0.02 0.04 0.06 0.08


0.038 MISS

6 0 3 6 RADOME


(91 191

Target Maneuver and Range Independent Noise (g) Effective Slope Due to Target Maneuver
Is Reduced and Effective Noise Increased with Target Maneuver and Range Independent Noise
(h) Effective Slopes and Effective Noise Predict Miss (i) Effective Slopes and Effective Noise
Predict Target Maneuver Miss (From [Murray, 19841 with permissionfrom AACC)
60 MDASE of NGC Processing Chap. 2

Range-independent npise (RIN) miss. The result of RIN is gimbal

angle motion, and the oscillation amplitude is greater for a negative radome slope.
here fore, in addition to the miss distance being greater with negative slope ( ~ i ~ i r e
2-18b), the gimbal angle standard deviation is larger as well. The result is a lower
effective slope and thus a reduction in miss with a periodic radome from that with
a constant slope radome as demonstrated later. Figure 2-19b shows the gimbal angle
standard deviation. The figure also shows the effective radome slope versus gimbal
angle standard deviation for a 1-0.045 1 maximum slope radome. These effective
slopes for periods of 7.5, 30, and 120 deg are obtained from Figure 2-19a. The value
of the effective slope due to RIN is given by the intersection of these two sets of
curves. For @ = 180 deg, corresponding to a -0.045 maximum slope (Figure
2-18f), the effective slope, rf, is -0.028 for a 30-deg period sinusoidal radome as
shown in Figure 2-19c. The resultant effective noise standard deviation is 1.0 mrad.
The RIN induced miss with a sinusoidal radome is predicted by the effective slope
and effective noise, as shown in Figure 2-19d. When Equation (2-31) and the ef-
fective slope and effective noise are used to compute miss, this correlates with the
miss for the sinusoidal radome. The reason for which the miss computed this way
agrees so much better than the unrealistically large miss with a constant slope radome
is not difficult to understand. In addition to increasing RIN miss, negative radome
slope also increases the gimbal angle standard deviation. The result of this is to lower
the magnitude of the effective negative slope, which in turn reduces the miss. Con-
sider a linear radome miss vrediction model for RIN-induced miss as follows: a 30-
deg period sinusoidal radome, a -0.045 maximum slope at 0-deg gimbal angle,
and a 2-mrad RIN. Based on this model, the RMS RIN-induced miss is 1.5, while
the RMS miss component due to effective noise is 0.3. The RSS for this model is
therefore 1.5. This can be compared to a value of 1.2 for the RSS of the sinusoidal
radome error model. In this examvle the RIN comvonent for a -0.028 effective
slope is dominant. The effective noise component is relatively small.

Noise and target maneuver miss. Gimbal angle motion is induced as

a result of target maneuver as shown in Figure 2-19e. Here, both simulation and
analytical results are displayed for target maneuvers initiated at 3 and 5 sec to go.
The gimbal angle increment, a,,, due to target maneuver can be expressed analyt-
icall y as

An arialytical expression for a missile acceleration, A,,, due to target maneuver.

AT,, for a zero time lag missile system can be found in Fossier and Hall [I9671 and
Jerger [1960]. Equation (2-32) is obtained for computing 8, given A,,,, and assumirlg
perfect seeker stabilization and therefore that q, = 8. This equation is not highly
Sec. 2.3 NGC System Design and Analysls 61

sensitive to radome slope, r. There is a variation of 12 deg, from 22 deg to 34 deg,

in u , ~for
, r varying from -0.045 to +0.045 with a value of 27 deg for r equal to
0.0. The results that follow therefore neglect r in Equation (2-32). That is, A, =
A. As can be seen in Figure 2-19f, the effective slope due to RIN is dramatically
reduced with target maneuver and RIN. This reduction is primarily a consequence
of the 3-g target maneuver, which results in a 27-deg gimbal angle increment,
u,~,. This caused the effective slope, r,, to decrease from -0.028 (no maneuver) to
-0.004 for the 30-deg period sinusoidal radome. As shown in Figure 2-19g. there
is also a big falloff in the effective slope due to target maneuver as target maneuver
is increased. Thc 27-deg gimbal angle increment, q,,, owing to a 3-g target ma-
neuver causes the effective slope, r ~ . to, fall off from a value of 0.038 to 0.014 for
the 30-dcg period sinusoidal radome. The effective slopes and effective noise predict
the miss with a sinusoidal radome as shown in Figure 2-19h. Using Equation
(2-31) together with the effective slopes and effective noise to compute miss gives
values that correlate well with the miss with the sinusoidal radome. These miss
distances are the RMS of four sets of 25 flights each, with the sets having the initial
gimbal anglc at the radome phase angle, 0,of 6.0, 13.5, 21.0, and 28.5 deg. Thus,
these miss distances are the RMS over RIN, uniformly distributed target maneuver,
and gimbal angle. Consider a linear radome miss-distance prediction model as fol-
lows: a 30-deg period sinusoidal radome; a 1 0.043 / maximum slope; a 3-g target
maneuver; a 2-mrad RIN; and an RMS miss of four radome phases. Using this
linear radome miss-distance prediction model, the RMS miss-distance component
due to target maneuver is 0.8, while that due to RIN is 0.8 and that due to effective
noise is 0.8. With these values, the RSS is 1.4, which can be compared with a value
of 1.5 for the sinusoidal radome. In this example, the effective noise component is
equal to the target maneuver and RIN components. The effective slopes and effective
noise predict the miss with increasing target maneuver as shown in Figure 2-19i.

Empirical radome slope calculation. It is not difficult to find in the

literature empirical relationships for radome slope [Donatelli and Fleeman, 1982;
Giragosian, 1979; Travers, 1982; Youngren, 19611. That found in Travers [I9821 is
examined here as an example to illustrate certain key points.

where p is the radome fineness ratio (radome length divided by radome diameter),
E is the dielectric constant, f~ is the deviation from design frequency, A is the signal
wavelength, K , is the performance measure, and Dd is the dish diameter. The
dielectric constant is a function of both temperature and radome material. Walton
[I9701 gives the dielectric constant for different radome materials at various tem-
peratures. Generally, radome slope error decreases with increasing seeker dish di-
ameter, with decreasing fineness ratio, and with decreasing signal wavelength. The
thickness of the radome is usually constructed to be one half of the wavelength to
avoid multiple reflections. The performance measure k, reflects how sophisticated
the manufacturing process is, higher numbers indicating an advanced radome grind-
62 MDASE of NGC Processing Chap. 2

ing process. While a hemisphere ( p = 0.5) possesses ideal electromagnetic prop-

erties, it is possible to achieve lower drag with shapes having a higher fineness ratio.
Therefore, a compromise must be struck between these opposing factors. Radome
slope is related to the total radome swing by r = + r ~ / 2which is treated as a
maximum expected value for 90 percent of the slopes for a given radome [Nesline
and Zarchan, 1984(b)].

Higher-order system analysis. In Figure 2-15a(iii), Nesline and Zar-

chan [1984(b)] consider

The guidance system transfer functions are

The effective guidance time constant, T;, can be approximated as the coefficient of
the s term in the denominator of ~ ~ u a t i o(2-35)
n anh is the same as that in Equation
(2-28). Here, as in the case of the first-order system, the effective navigation ratio,
A', is modified by radome slope according to Equation (2-28). It is easily seen that
rand T, have an effect on the guidance time constant and effective navigation ratio
the same as occurs in the first-order system of this example. It is possible to achieve
larger stability regions if one can keep small any or all of the following: the navigation
ratio, the ratio of the turning rate to guidance time constant, the ratio of closing to
missile velocity, and the radome slope [Nesline and Zarchan, 1984(b); Peterson,
Because for a given radome, increasing A acts to destabilize the system, the
designer must strike a balance between having A to be sufficiently small so as to
satisfy stability requirements and at the same time sufficiently large so as to ensure
effective homing. For a given higher Ta, radome effects were responsible for sta-
bility problems as well as an increase in miss distance. Higher T, is found at higher
altitude and lower missile velocity. While T, is usually increased at higher altitudes
to improve stability, this brings about a corresponding increase in miss distance.
Consequently, the designer is faced with attempting to make 7 ,as small as possible
to reduce the miss distancc without sacrificing too much stability. O f course thosc
components of miss distancc caused by saturation and radomc/parasitic effects will
not be affected by lowering 7,. The minimum achievable guidance time constant,
(~~),i,,,given T, and the expected r, can be determined on the basis of a useful
stability criterion originally developed by Nesline and Zarchan [1984(b)] as - 0.79
< r h,(VJ V,,,)(Ta/rg) < 2.07. The left-hand side of the preceding inequality gives
the lower limit of the allowable negative radome slope, while the right-hand side
establishes the positive upper limit of radome slope. If the guidance time constant
is increased, there is a slight degradation in miss accompanied by a decrease in the
system scnsitivity to radome swing. If the guidance time constant is decreased below
Sec. 2.3 NGC System Design and Analysis 63

the minimum predicted by the stability analysis, it becomcs impossible to attain

satisfactory performancc over the expected radomc swing range. When thc resultant
performance estimatc is not compatible with the allowable warhcad size, eithcr r or
T, must be reduccd. Reducing thc former may involvc blunting the nose, using
compensation, or somc other techniquc, while reducing the latter can be done by
increasing the lifting surface area, reducing the maximum operational altitudc, in-
creasing missile velocity, reducing missilc weight, and so on [Ncslinc and Zarchan,
1984(b); l'etcrson. 19611.

Example 2-4: Homing missile guidance and control analysis.

Trimming thc missile at high altitudc may require that the fin be deflected through
large angles. If the fin angle is limitcd, the missile's acceleration capability will also
be limited. Determining the fin dcflection required for deterministic inputs can be
aided through the use o f zero-lag guidance system results. Neglecting guidance
dynamics in Equation (4) o f Table 2-2, fin angle can be expressed according to
Nesline and Zarchan [1984(a)] as

Achieving in a rapid fashion the large control fin angles necessary to trim the missile
at high altitudc requires that the control fin move at high rates. The result of an
inability on the part of the control fin actuator to attain these rates can be instability.
Combining Equation (2-35) and Equation (4) of Table 2-2, the fin rate response
due to LOS rate for the fifth-order binominal guidance system becomes

As was true in the case of fin angle, zero-lag guidance system results are very useful
for determining fin rate response due to deterministic inputs. When guidance system
dynamics are neglected, Equation (2-37) becomes [Nesline and Zarchan, 1984(a)]

Following Examples 2-2 and 2-3, if the structural filter, actuator, accelerometer,
and gyro dynamics are relatively fast, the FCS can be represented by the ideal third-
order transfer function described by

where T, 5, and o are the FCS desired time constant, desired damping, and desired
natural frequency, respectively. The values for the zeros are determined by the
aerodynamic configuration (the way in which these values are determined i~ pre-
64 MDASE of NGC Processing Chap. 2

sented in Book 1 of the series). The FCS gain is represented by K.This third-order
model of the autopilot dynamics is useful in generating a nominal system design.
The model supposes only two homing loop disturbances with white glint noise
with PSD +, in addition to a uniformly distributed sinusoidal target maneuver
shown later in Figure 2-21d. These are two important factors that contribute to
high miss distance in a radar-guided PNG system. An adjoint model similar to that
to be presented in Chapter 3 for Example 2-2 is constructed based on the homing
loop shown in Figures 2-15a(iii) and 2-ljb, which follows the adjoint rules pre-
sented in Chapter 3. From such an adjoint model, an adjoint simulation is created
using representative values for error sources and system parameters, the results of
which are shown in Figure 2-20. Neglecting parasitic effects, the guidance system
transfer function is given by



0 0.2 0.4 0.6 0.8 1.0


o 2 i 6


Figure 2-20 (a) Effective Navigation Ratio Influences System Performance (b) Flight Control
System Time Constant Influences System Performance (c) Decreasing Flight Control System
Time Constant Makes System Performance More Sensitive to Radome Slope (d) Decreasing
Flight Control System Damping Makes System Performance More Sensitive to Radomc Slope
(From /Neslines, 1984(a)] and /Nesline G Zarchan, 19821 utith permissions from AACC and
Sec. 2.3 NaC System Design and Analysls 65

where A,, = KA is the desiredlactual effective navigation ratio. Since the gain K is
varied around unity, the actual and the desired navigation ratios are almost identical.
140, which varies due to changes in the flight condition, is one major determinant
of homing system performance. Typical performance results displayed in Figure 2-
20a show that if A,, is too high, miss distance increases due to noise and, if it is too
low, miss distance increases due to target maneuver. Therefore, A. has an optimal
value based on the error sources and system parameters, and in no case should exceed
certain bounds. This imposes limitations on the allowable variation in K,as a varia-
tion in gain is equivalently a variation in A,]. For example, a variation in from
unity along with the corresponding variation in A,, makes it impossible to attain
optimal homing performance. Neglecting parasitic effects, T , ~can be approximated
as the coefficient of the first-order term in the denominator of Equation (2-40a) as

Therefore, increasing the FCS time constant T increases T,,, while decreasing T de-
creases T , [Nesline
~ and Zarchan, 19821. The system performance is also substantially
affected by T.

Autopilot response characteristics necessary to achieve homing.

The speed of response of an FCS can be measured to first-order accuracy by 7. The
maximum system time constant permissible to achieve successful homing depends
primarily on the maneuverability of the targets the missile is expected to engage
and the reauired miss distance necessarv to intercept these targets.
" Seven-tenths of
a second would be an acceptable maximum for a moderately difficult homing en-
gagement. Figure 2-20b shows the effect of T on miss distance [Neslines, 1984(a)].
As shown in this figure, in which it is assumed that the time constants T,, T,, and
2510 are all known, if 7 is decreased, the miss due to target maneuver will also
decrease, but that due to noise will increase. Increasing 7 produces the opposite
result. For the set of inputs considered it appears from Figure 2-20b that T can be
made arbitrarily small to optimize the miss distance, but this cannot be done because
of radome refraction effects. When the parasitic radome loop is considered, the
guidance transfer function becomes
A*,, KAV,(l + +
Ails AI2s2)s
+ ?,$)(I + T , s ) ( ~ + +

cr (1 rs)(l 2[s/w + s2/w2) + (i7i\vCr/v,,,)(l + T,s)

where 7; and A' are the same as those in Equation (2-28) except that T, is given by
Equation (2-40b). The effect of radome slope error on system stability is itself
strongly dependent on T,. At high altitudes, where T, is large, the guidance transfer
function can become unstable due to either excessively positive or negative radome
slopes. Thus, small T yields a large guidance system sensitivity to radome slopes as
shown in Figure 2-20c [Nesline and Zarchan, 19821. Therefore, the optimal value
66 MDASE of NGC Processing Chap. 2

of 7 is bounded from above by target maneuver (especially at low altitude) consid-

erations, and from below by radome refraction effects (especially at high altitude).
The minimum amount of damping required in order to maintain performance, given
a reasonable radome. should be determined usingu a detailed complete missile model:
however, a damping ratio of 0.3 is a reasonable estimate of the minimum value that
a radar-guided missile could tolerate. Figure 2-20d shows the effect of reduced
damping and radome slope on miss distance [Neslines, 1984(a)]. The figure also
shows typical high-altitude performance for a FCS with high and low damping. As
shown in this figure, the FCS damping becomes more crucial when radome effects
are considered. A system with low damping is far more sensitive to radome effects
than one with higher damping. This means that if the unaugmented airframe is to
be used in a radar homing application, a FCS must improve the low damping.

Example 2-5: Radome error compensation. Following Example

2-4 and Figures 2-15a(iv) and 2-15b, assuming 7, = Ilk = 0 sec infinity bandwidth,
and assuming

7y - effective FCS time constant ;--- 27A

Combining the guidance equation (2-16) and autopilot Equation (2-42a) then gives

A,, -
- A V<u;
1 + h A s + 275s2 + 7is3

Substituting for 6 yields the guidance system transfer function

where 7; and A' are the same as those in Equation (2-28) except that 7, = 2 7 + ~
7,. With AV,/V,,, = 6, r = 5.02, and T , = 10, A V,rT,I V,,, can be & 1.2 sec and
hence the guidance system becomes unstable (that is, 7,;<0 ). The radome crror
slope can be the dominant factor in the guidance response time [Hiroshige, 19861.
A guidance and control system is designed to compensate for the stability problcm
caused by the radome error. To alleviate the severe radome stability problem for a
high-altitude engagement scenario, one of the commonly used engineering "fixes"
is to introduce pitch rate compensation G,0 with gain G , feedback to either before
or after the guidance filtering network as shown in Figure 2-2. It can somewhat
symmetrize the positive and negative slope stability region at the cost of slowing
down the autopilot response [Yueh and Lin, 1984, 1985(a)]. Adding the pitch ratc
compensation and the guidance filtering (1/(1 +
7.5)) to the guidance cquatiorl
Sec. 2.3 NOC System Design and Analysis 67

(2-42c). The guidance system transfer function bccomcs

-A,,,, (I + r)AV<s
urn,<. s(l + T ~ ~ ) [ ~ / G , ,+( SG,AV,(I
) + Tas)lV,,,llGs(s) + r A V 4 + ?,s)/Vm

The transfcr function for thc FCS, G.d(s) can be modeled as a third-order transfcr
function as described in Equation (2-3'9) or (3-42a). If thc nonlinear elcmcnts are
not included in the analysis, Equation (2-43a) based on the linear FCS model is
easily solved. Howevcr, if it bccomcs necessary to employ the nonlinear FCS model,
it is difficult to obtain an analytical solution using the classical frequency-response
approach. Assuming the input to the saturated nonlinear elements is a sinusoid, then
the nonlinearity can be represented by a variable gain N obtained from the describing
function of the nonlinear element to be presented in Chap. 3. The transfer function
for the FCS would then be given by G . 4 ( ~ = ) N(ao + als + azr2)l(bo + bls + bzs2
+ b3s3) where the parameters are specified according to a particular condition. If
G,(s) = G , = constant, C s ( s ) = 1/(1 7,s) have been assumed, then from Equa-
tion (2-43a).

(1 + G,A V,i C',,,)(T, + T,) + 7, + (G, + r) T,A V,I V,,
1 + (G, + r)AV,/V,,,

A' = (I + r)Al[l + (G, + r)AVclV,,,] (2-43c)

where the effective autopilot time constant 7, -- ~ T Athe, guidance filter constant
T,, and the radome compensation gain G , can be selected to compensate for the
stability problem caused by the radome error. In comparing Equations (2-28) and
(2-43), one can see that the radome compensation G,0 has reduced A' and increased
7;, both of which tend to stabilize the system of a given radome but at the expense
of decreasing the guidance response time. The root-locus technique is applicable
only to a linear, time-invariant system. The root locus for a linear, time-varying
system of Equation (2-43a) can be obtained with the roots at different t,. The
evaluation can be performed on the transfer function given in Equation (2-43a) for
T,, T,, and G,, which vary with respect to t,.

Example 2-6: Gyro sensitivity gain design. In this section, a method

of approximating aeroelastic effects on the dynamic behavior of a rate gyro is de-
veloped. More detailed derivations of a flexible body transfer function from control
surface deflections to body rotation rates have been presented in Book 1 of the series.
A gyro sensitivity gain expression of body-mounted rate gyro coming from body
acceleration due to the aeroelastic vibration is considered here. The following as-
sumptions are made in this derivation: (1) the first mode of vibration considered
here is excited by body accelerations, (2) only longitudinal and lateral vibrations are
considered, and (3) superposition of the aeroelastic angular rate at any given point
68 MDASE of NGC Processing Chap. 2

on the rigid-body angular rates is valid. The transfer function from the generalized
acceleration f to the normalized body-bending curve can be written as

where 5 1 is a structural damping rate, w l is the natural frequency, and Z is the

reference displacement to which the mode shape 1 shown in Table 2-3 is normalized.
The location of the body-mounted rate gyro on the normalized body bending mode
shape 1 shown in Table 2-3 is arbitrary, and it is desired to derive an expression
for the aeroelastic body angular rate at this station. T o this end, the slope of thc
normalized bending curve at the rate gyro station is determined first. Then, the
time rate of change of this slope, due to acceleration, is approximately equal to the
aeroelastic body angular rate at the rate gyro station. The product of the normalized
displacement Z N and the reference displacement Z located at the rate gyro body
station give the instantaneous reflection Z D at this station. That is, Z D = ZAjZ.
The slope of the normalized bending curve h at the rate gyro station is h =
ZN/XD and the actual slope at the same station is K = Z D / X D= hZ. The rate of
change of the actual instantaneous slope is K = h i . For small-amplitude oscillations.
the slope of the tangent to the curve at the gyro station can be approximated by the
angle between this tangent and the reference x-axis. In this case, the aeroelastic
angular body rates 8 , at the rate gyro station is
8, = K = / z ~ (2-43)
Thus, the transfer function from the generalized accelerations f to the aeroelastic
angular body ratcs a t the gyro station is found by substituting Equation (2-45) into
Equation (2-44) giving
8,/f = s / ~ / ( +s ~251wls +~ f ) (2-46)
Because the first-ordcr bending frcqucncy is typically very Iargc, thc gyro sensitivity
gain to the step accelcration can be expressed to an approximation as
K, = hlw: (2-47)
For missile applications, the gyro sensitivity gain is a very .small value since the
normalized slope h is a small value and the bending frequency w l is a large value
of first bcnding modc.


2.4.1 Target N o i s e and Target Maneuver Modeling

Targct noise has been presented in Section 2.3. I , and is applied hcrc to gencratc the
target mancuver signal-to-noise ratio. Figure 2-21 shows thc cquivalcnt signal noisc
diagram, where the signal, representing a target maneuver, is in turn represcntcd
by a shaping filter. As was done in Section 2.3.1, it is possiblc to convert thc total
Sec. 2.4 Target Tracklng State Modeling

. Position Noise
~ ~n ~ ,
Target Maneuver Model T ~position
Spectral Density 7 Signal yT
Target Position
Measurement z
Shaping Filter H(s)

a) De~ermineManeuver: b(l)

b) Brownian Target Position: u,@


C) Zero-Mean White Gaussian Targel Acceleration: ,,

d) Uniformly Distributed Targel Maneuver: us

e) Poisson Jinking Maneuver: u,


Note: u, = white noisc with PSD 9,

Figure 2-21 Target Maneuver Model

seeker angular noise to a positlon noise by multiplying the angle noise by the range
to go. In general, it is possible to generate a number of random processes by passing
a white noise w through a suitable filter. If it is assumed that the target noise n in
Figures 2-1% and 2-21 is a first-order Gauss-Markov process with correlation time
constant T,, and white noise input w , then the corresponding equivalent white noise
PSD of the position noise is

where r , and 4, are given by Equation (2-25). Figure 2-21a shows a deterministic
model, which is the simplest case. In most applications, target position or velocity
or acceleration is assumed to be filtered white noise (see Figures 2-21b through
Uniformly distributed target maneuver. For a target acceleration of
step form with r a g o m starting time varying uniformly over the flight time, the
autocorrelation function of this model is identical to that of the integrated white
noise (see Figure 2-21d). The uncertainty in target position y + is described by a
third-order integrator with a constant PSD 4, white-noise input us as:

where n T is the magnitude of the step maneuver, and tr is the flight time over which
the maneuver is uniformly distributed. Two integrators are used in tandem in order
to convert the acceleration to the target position y T , as shown in Figure 2-21d.
70 MDASE of NGC Processing Chap. 2

With regard to the uniformly distributed target maneuver, it can be seen by ex-
amining Figure 2-21d that the noise spectral density level is constant, but the signal
spectral density passes through three integrators. As a consequence, the target signal
is frequency dependent while the target noise is not at a particular range to go.
Moose et al. [I9791 give a short exposition on the evolution of the relationship of
target acceleration modeling to stochastic processes. This evolutionary development
is graphically depicted in Figur62-22 [Cloutier et al., 19881. Initially, as shown in
Figures 2-22a and 2-21c, target acceleration was simply accounted for with white
noise (depicted by a correlated process whose spectrum is flat over the system band-
width). Consequently, tracking could be maintained only for those maneuvers re-
maining within the envelope of the white noise, and even then performance suffered
due to the incorrect assumption of uncorrelated acceleration.

Correlated acceleration process. A correlated acceleration process,

Figure 2-22b, can be achieved using Gauss-Markov models [Fitts, 1974; Pearson

Figure 2-22 Historical Development

of Maneuvering Target Acceleration
Models (a) Zero-Mean White Gaussian
Process (b) Zero-Mean Correlated
Gaussian Process (c) White Gaussian
Process with llandomly Switching Mean
(d) Correlated Gaussian Process with
Randomly Switching Mean (Frotn
/Clourirr et a!., 19881 with prmisrion porn
Sec. 2.4 Target 'Itacking State Modeling 71

and Stcar, 1974; Singer, 1970; Vcrgez and Liefer. 19841. The well-known Singer
model [Singcr, 1970) has been used with various Kalman-type filters to achieve
excellent tracking characteristics over a broad class of large scale maneuvers. Its
primary drawbacks arc the need to specify apriori both the acceleration time constant
and power of the driving noise and the inability of the model-based filter to track
target motion resulting from abrupt changes (jumps) in the acceleration process.
The velocity a t which the target is moving is constant, but thc target's lateral ac-
celeration is described as a Poisson jinking maneuver. Such a maneuver, which is
one whosc magnitude B (RMS target acceleration) is constant but whosc sign is
randomly switching, can be modeled as white noise through a single-polc filter.
This is because the acceleration autocorrelation function (RA.,.)I(7) = ~ ~ e - ~ .' f' ' =
average frequency) corresponding to this maneuver process is identical to the Poisson
process. Thus, the target maneuver models are as shown in Figure 2-21e
AT, = -2fAr, + 2B d? w , = -A,AT, + h,u, (2-50)
where u, is a zero-mean Gaussian noise, that is,
~v, with PSD = - (2-51)

w , is a white noise with unit PSL), and A, = 2f = the target maneuver bandwidth.
Gholson and Moose [I9771 modeled rapid, major changes in target motion by a
semi-Markov process. The mean of their process, Figure 2-22c, randomly switched
to a finite number of states according to a Markov transition probability matrix,
with the time duration in each state itself being a random variable. Successf~~ltrack-
ing with this method necessitates a large number of preselected states, and the er-
roneous assumption of uncorrelated acceleration is still retained. Moose et al. [I9791
extended this work by employing the Singer model to represent correlated accel-
eration within each of the states, Figure 2-22d. Larimore and Lebow [I9871 de-
veloped a similar model based on a parameterized Gauss-Markov process with the
parameters changing according to a point process [Cloutier et al., 19881. Lin and
haf froth [1983(a)] concluded that oneof the most versatile representations of target
acceleration is that of the sum of a continuous-time Gaussian process and a finite
basisljump process. The continuous-time process can be obtained by means of a
classical shaping filter, while the jump process which is characterized by its basis,
jump size distribution, and interjump time distribution, can be obtained by driving
a shaping filter with a white sequence.

Other target acceleration modeling. As stated in Cloutier et al. [1988,

19891, another aspect of target acceleration modeling is a consideration of the aero-
dynamic characteristics of the target. A winged aircraft accelerates mostly ortho-
gonal to its velocity vector, in particular, orthogonal to the plane of its wings. The
magnitude of the acceleration also has asymmetric bounds (positive or negative g)
which are set by pilot and/or aircraft limitations. Modeling based solely on target
point-mass motion fails to take many of these characteristics into account. Kendrick
72 MDASE of NGC Processing Chap. 2

et al. [I9811 modeled normal load acceleration with a random variable whose prob-
ability density function (PDF) was asymmetrically distributed between small neg-
ative g and large positive g. Nonnormal accelerations were modeled as first-order
Gauss-Markov. It was assumed that the orientation of the three-dimensional target
could be obtained by means of an EO seeker in conjunction with pattern recognition.
Bullock and Sangsuk-Iam [I9841 developed a nonlinear Cartesian model of target
dynamics by using polar coordinates to model planar, circular turns under the as-
sumption of constant velocity magnitude. Similarly, Hull et al. [I9831 used polar
coordinates to model both acceleration magnitude and angle as random processes.
This led to a Cartesian model characterized by state-dependent noise. Finally, the
parameterized Gauss-Markov model developed by Larimore and Lebow [I9871 took
into account aerodynamic parameters such as bank angle, lift force, and thrust minus
drag. A particular maneuver has a given set of parameters, while any abrupt change
in the parameters corresponds to the initiation of a new maneuver. Merging a point
process, or jump process, with a continuous-time Gaussian process, either addi-
tively, as recommended in Lin and Shafroth [1983(a)], or through parametric embed-
ding as performed in Larimore and Lebow [1987], is an excellent way to model the
acceleration of a highly maneuverable target. However, of the models of this type
developed to date, only one [Larimore and Lebow, 19871 has given any consideration
to the aerodynamic characteristics of the target. The natural next step in this tech-
nology area is to merge the concepts of those models based on aircraft flying char-
acteristics with those symbolized by Figure 2-22d.

2.4.2 Two-Dimensional Target Tracking State


Consider the two-dimensional target tracking problem illustrated in Figure 2-10.

For simplicity, let velocity VT be in the x-direction while V,, = 0 and y = 0 for
a fixed tracking station. From Equations (2-2),
R = v7-cos u, 6 = - (VTIR) sin a (2-52)
Differentiating these equations yields the nonlinear state equations
R = VT cos u - l,/6 sin g ,ii = - [ ( v ~ R- v ~ R ) / R sin
~ ] u - (VT/R)6 cos cr
Following the treatment of Lewis [1986], assume that the target is not accelerating,
that is, VT = 0. Then, assuming also very small changes in R and u during each
1 to 10-scc radar scan, one has that R ;= 0 and & = 0. Hencc, random disturbances
are introduced to account for the small changes in R and b during
O R ( [ ) and ow(()
the scan time T, giving
Sec. 2.4 Target nacklng State Modellng 73

The processes w R ( t )and w,(t) are known as maneuver noise. Because the Kalman
filter is robust in the presence of process noise, R and & can be tracked. The addition
of process noise is done to offset unmodeled dynamics introduced as a result of
excluding nonlinear terms. The state equations then become linear as

The measurement model, given that the radar measures range R and azimuth a is
defined in Equation (2-3) and is given by

Suppose independent disturbance accelerations are uniformly distributed between

t n T,then w R ( t )is a radial acceleration with variance n$/3. Since w,(t) is an angular
acceleration, w,(t)R is distributed according to the maneuver noise PDF and w,(t)
has variance n$/(3R2). The covariances of process noise w ( t ) and the measurement
noise v ( t ) are given respectively by

where :u and a : denore the variances of the range and azimuth measurements, and
: <a
a : is assumed. For each scan, the scan time for which is denoted by T, the
measurements z ( t ) are available. Given that the square of the system matrix A is
zero, the series representations for the sampled matrices are easily used, giving

74 MDASE of ROC Processing Chap. 2

2.4.3 Three-Dimensional Intercept State Modeling

Twelve-state guidance law design modeling equation. The linear

state model, consisting of relative position, relative velocity, target acceleration, and
the FCS and the airframe and propulsion dynamics, is given by Equation (3) of
Table 2-1 with N = I and x, A, B, u, w defined as follows:
x = [x,, y,, z,, xr, jl?, i,,AT,, AT^, A T = ,Am,, A m , , ,Am,I T

where subscripts T and m denote target and pursuer, respectively; x,, y,, and z ,
are, respectively, the x , y, and z components of the relative position in inertial axes;
ir,jl,, and 2 , are, respectively, the x, y, and z components of the relative velocity
in inertial axes; and AT,, ATl, AT, are target accelerations which are modeled as
a Poisson jinking maneuver shown in Figure 2-21d with target maneuver bandwidth
A and with w T as the white process noise with covariance matrix Q. Definitions of
A and Q are as follows:

where A , and 4, are the same parameters as those in the horizontal plane case defined
in Equation (2-51). The states A,,, A,,, A,,, represent the pursuer accelerations,
and the first-order model defined in Example 2-2 is assumed here with its bandwidth
o defined as

o = (2-62)
0 0 0,

The nonlinear measurement equations for a sensor measuring range, range rate, and
two LOS angles can be represented as the true values corrupted by additive noise

when /I is a nonlinear function of the states and v is a white measurement noise with
covariance r. The target tracking sensor is assumed to provide measurements of
LOS angles u,,and a, slant range R, and range rate R (see Figure 2-9). Thus, the
z vector is given by

[;][ =
R,,,,. = /I,
R,,,, = 11, = (xf + yt + z ,2) 112
= (x,i, + y , j , + z,i,)l(x: + y t + xf)'12
u,,,, = h, = tan-'[z,l(x? + y,) ]
urmr= h4 = tan-' (y,lx,)
2 112

I + v (2-63b)
Sec. 2.4 Target nacklng State Madeling 75

where the subscript true denotes the geometric relationship, and the subscript k
denotes time index. The measurement noise v is defined as
R m + Rf + R, + R,r,,,
ut, + u , + a,, + 0,,,,
urn + Of + u, + ~gl,,,#

where, as in the case of two-dimensional engagement, the subscript rn denotes the

receiver noise, f the fading noise, c the clutter noise, and glint the glint noise. Spe-
cifically, the LOS angle measurement noise v , is the same as that in the horizontal
plane case defined in Equation (2-22).

Nine-state guidance filtering equation. The nine-element state vector

of the linear kinematic target-pursuer engagement model is simplified from the
twelve-state one, and is written as

where the measurement equations are the same as Equations (2-63), and x, u, A.
B, and w are defined as follows:

where the target bandwidth A and the process noise covariance matrix Q are defined
in Equation (2-61). The achieved pursuer acceleration u is assumed to be measured
very accurately by accelerometers so that an estimate of u is not needed. For a sensor
with a constant data rate, that is, sampling state every T sec, the discrete form of
the system equation in Equation (2-64) is
x ( k ) = O ( T ,A)x(k - 1 ) + W u ( k - 1) + w ( k - 1 ) (2-65a)
where the state matrix @ is
D(A, A)
A-'(13 - e K A A )
e -AA I (2-65b)

where D ( s , h ) = ( e - ' h - I + s h ) h K 2 ,the deterministic forcing term. T u ( -~ 1)

depends on pursuer acceleration, which is assumed to be known exactly and is given
76 MDASE of NGC frocesslng Chap. 2

If U(T)is assumed to be u(k - 1) between the time intervals (k - 1)T and kT, then
V( k-1 = ( -I ,
-T I 0
u(k - 1) (2-65c)

and the discrete form of the process noise w(k - 1) is a white-noise sequence

D ( k T - T, A)
- e-(kT-~)A]
E[w(k - I)] = 0, E[w(k - l)w(k - = a
and the discrete process noise coGariance matrix a is

where, as shown in Singer [1970],

Modern Multivariable
Control Analysis
Modern multivariable control analysis, which is the second ofthe five primary topics
as shown in Figure 1 of the Preface, is mainly concerned with the analyses of ro-
bustness, performance, dynamics, and stability. This chapter deals specifically with
the first two analyses, while stability and dynamics analysis is covered in Book 1
of the series. Robustness analysis (Sections 3.1 to 3.6) is particularly important for
examining multivariable control systems design. Performance analysis (Sections 3.7
to 3.12) includes the various statistical methods used in the analysis and synthesis
of any modern NGC system. The material of this chapter is based in large part on
AGNC [1986(a), (b)].

Robustness analysis. The background material for this discussion on

robustness analysis, such as the use of matrix theory and the property of singular
value, is presented in Book 1 of the series. Robustness means that the closed-loop
stability and performance are insensitive to uncertainties which include modeling
errors, nonlinearities, unmodeled dynamics, measurement errors, and unknown
external disturbances. The more the system can preserve its performance and sta-
bility under uncertainties, the more robust it is. Robustness analysis is essential since
all models of physical processes are to some extent approximations of the true plant
or state. The existence of uncertainties due to incomplete system knowledge, sim-
78 Modem MultluarfableControl Analysis Chap. 3

plified assumptions in the model, external disturbances such as air turbulence around
a flight vehicle, or measurement noises are the primary reason for employing feed-
back. Feedback aims to compensate for the uncertainty, thus preserving system
stability and performance. T6 characterize how robust a system is, a reliable and
nonconservative measure of the robustness margin is essential. For a single-input-
single-output (SISO) system, this is generally relegated to the gain and phase mar-
gins. For a multivariable system, the singular value, structured singular value, and
other robustness margins have to be used. All these robustness measures will be
discussed in Sections 3.1 to 3.6.

Performance analysis. The rest ofthis chapter is devoted to a comparison

of various computerized statistical analysis methods employed in the synthesis and
analysis of guidance systems. Through examples, the advantages and computer costs
of each method are examined in detail. The design that emerges in the synthesis of
a guidance system must specify allowable measurement noise, determine subsystem
bandwidths that guarantee adequate stability margins, prevent component satura-
tion, and evaluate total system performance (see Books 1 and 4 of the series and
Yueh and Lin [1985(a) & (b)]). Several statistical analysis techniques can be applied
to both linear and nonlinear noise-driven systems for the purpose of studying systcm
performance. Linear analysis has as exact methods the adjoint technique and co-
variance analysis. Approximate methods that exist for nonlinear analysis includc
Monte Carlo techniques, CADET, and SLAM. These performance analysis rncthods
are all examincd with the help of a unifying example involving missile homing
guidance and digital simulation. The unif>;ing example from inissilc homing guid-
ance prcscntcd in Example 3-2 is uscd to demonstrate thc utility of each of thc
pcrformancc analysis tcchniqucs via digital simulation IZarchan, 1979(a), 19881.
When trcating thc complctc nonlinear dynamics of thc vchiclc's n ~ o t i o n anal-
ysis must be pcrformcd on thc systcm of equations govcrning thc complctc N G C
loop. Thc analysis n ~ u s tconsider both detcrininistic inputs such as impulses and
steps, and random or stochastic inputs such as white noise and turbulence. In prin-
ciple, any type of problcm involving N G C systems can be solvcd using the mcthods
of pcrformance analysis. However, theoretical analyses of N G C loops almost always
involve the use of cornputcrs owing to the prcscncc of nonlincar differential cqua-
tions with variablc cocfficicnts, and thc stochastic nature of thc actual N G C pro-
ccsscs. Whilc thc importance of numerical solutions obtaincd from thc application
of computers cannot bc ovcrcn~phasizcd,analytic analyscs of N G C loops arc cqunlly
~ ~ tpurposes of estimating thc gcncral laws governing the N G C proccss.
i ~ n p o r t a for


The idca of gain and phase margins has bccn uscd in spccifying and evaluating the
robustness of an SISO systcm. Consider thc fccdback systcln dcpictcd in Fig~lrc
3-1. Thc transfcr function C(s) incorporatcs both opcn-loop plant dynamics and
Sec. 3.1 Singular Value Analysis

??3-+ Figure 3-1 Feedback System

any con~pensationemployed. The Nyquist stability theorem states that the closed-
loop system is stable if and only if the graph of Cfjo), for all o,encircles the point
- 1 + j O as many times counterclockwise as C(s) has right half-plane poles. The gain
and phasc margins are derived by perturbing C by KG and eimC,respectively. More
precisely, the system has a gain margin k,,if PC, for all p < I*(,, satisfies the Nyquist
stability theorem. Likewise, the system has a phasc margin 40 if the perturbation
P'+, for all 4 < 40, does not destabilize the system. This concept of gain and phase
margins has been extended to multivariable systems by Rosenbrock 119741 and
MacFarlane and Postlethwaite [I9741 in which the locus of the determinant of the
return difference matrix, that is det[l +G(s)], is evaluated.
It is, however, not possible to identify the design robustness under plant un-
certainties by examining solely the Nyquist plot of det[l + G(s)]. As demonstrated
in Dovle and Stein 119781, the effects of cross coupling in the internal system model
are not dircctlv rcflcctcd in the stability margins obtained from the polar plots of
det[! + C(s)]. Generally, one finds that multivariable generalization of the Nyquist
stability criterion does not directly relate the conditions of multivariable robustness
to tolerance of open-loop modeling errors except in cases where nominal systems
and perturbations thereof are diagonally dominant. M3thematically, this can be ar-
tributed to the inadequacy of the eigenvalues as indicators of the singularity of the
matrix (there is a pole crossing the jo-axis if I + G ( j w ) is singular). Any measure
of robustness or relative stability should predict how far I + G ( j o ) is from being
singular or what is the magnitude of the perturbation that can be tolerated without
destabilizing the system. This observation leads to the singular value analysis in
feedback system designs.
The singular values of a complex p x m matrix M, denoted by ui(M), are the
nonnegative square roots of the eigenvalues of M*M, that is,
a i ( M ) = ~ A ; ( M * M()i = 1, 2, . . . , k) where k = min ( p , m )
Ordering u i such that u , ( M ) 2 u,(M) 2 a 3 ( M ) 2 .., 2 a k ( M ) ,the maximum and
minimum singular values can be alternatively defined by
a ( M ) = U I (M) = max [ 11 MX 112/11 X llr] = II M 112 and
I .Y 11220 (3-1)

In the casep = m and M is nonsingular, the smallest singular value can be computed
by g(M) = uk(M) = \\ ML1\\ZJ-' which measures how near the matrix M is t o
being singular or rank deficient. As is well known, the matrix M + E is singular
if there exists a nonzero vector .Y sach that ( M + E ) x = 0. In robustness analysis,
80 Modem Multiuariable Control Analysis Chap. 3

at each frequency w, E represents a certain stable perturbation while M stands for

some nominal closed-loop transfer matrix. If the minimum singular value of M is
greater than the maximum singular value of E for all frequencies, then M +
E is
nonsingular, which implies that the system remains stable if it is nominally stable.
In fact, the singular value is the exact measure of robustness if E is unstructured
(norm-bounded only). In the case that E is structured (subject to certain constraints),
a singular value may be conservative.
Consider the following two possible representations of uncertainties.

1. Additive. The plant is perturbed as G + E for some stable E. The return

difference matrix becomes (I + G + E)-I. If the system is nominally stable,
(I + G ) is nonsingular for all frequencies, then it remains stable as long as I
+ G + E is nonsingular. A sufficient condition for this to be true is z ( I +
C) > II E 112.
2. Multiplicative. The plant is perturbed as C(I + E) for some stable E. The
system is stable if [ I + C(I + E)] or [ I + C + GE] is nonsingular or if [I
+ G + E ] is nonsingular or IT(I+ G -') > (1 E ((7.

Singular value analysis is a powerful tool for checking whether the system is robustly
stable. This notion leads directly to the generalization of gain and phase margins
for multiloop systems. The multiloop gain margin is the pair of real numbers c1
and c2 defining the largest interval ( c l , c2) such that for all Ci(s) = C i E R in Figure
3-2 satisfying the inequalities, ci < C , < c 2 (i= 1, 2, . . . , m), the closed-loop
system remains stable. On the other hand, the multiloop phase margin is the number
c defining the largest interval (-c, c) such that for all C,(s) = e-'"('") in Figure 3-2
satisfying the inequalities - c < &(o)< c ( i = 1 , . . . , I N ) , the closed-loop systcm
remains stable.
The following arc obtained using the preceding definitions and the concept of
singular values: (1) If the original system is stable and ~ ( +l C) > a , then c l 5 111
(1 - a)], c2 2 [1/(1 a)], and c r 2 sin-'[a/2]; and (2) If the nominal system is
stable and ~ ( +1 G - ' ) > a , then 5 1 - a , ( 2 2 1 + a, and c 2 2 sin-'[a/2].
Clearly, the singular value plays a crucial role in determining the robustness of the
system. Depending on thc typcs and characteristics of the uncertainty, therc are
diffcrcnt robustness mcasures for diffcrcnt sources of unccrtaintics such as plant
uncertainty or cxternal disturbanccs that have to bc considcrcd. In the next scctiotl,

Figure 3-2 Configuration for

Multiloop C;aiti/Phasc Margin l)cfi~iiti~)~l
Sec. 3.1 Slngular Value Analysb 81

suitable robustness measures are identified and more general robustness analyses are
T o show the necessity of using singular values instead of eigenvalucs, consider
the following 2 x 2 matrix:

While the determinant of this matrix is always unity for any E, an additive permr-

produces a singular matrix which cannot be predicted by the eigenvalues. The sin-

gular values of M 6 defined by Equation (3-2) are the square roots of the eigenvalues
[(E' + 2) t E c2 + 4112 of the matrix M*(E)M(E).Note that the singular values
depend on E, while the eigenvalues and determinant do not. Note also that

clearly shows that ~ ( M ( E -

) 0 as c -t m. A more interesting example is discussed
in Laub [1985]. The Ostrowski matrix is an n x n matrix whose upper triangle
elements are all + 1, whose diagonal elements are all - I , and whose lower triangle
elements are all zero. That is,

This matrix is nonsingular and its eigenvalues are all equal to - 1. Its determinant
is ( - 1)" for any n. However, the Ostrowski matrix becomes increasingly nearer to
a singular matrix as the dimension n gets larger [Laub, 19833.

In fact, the smallest singular value of the matrix P is 2-"+I and approaches zero as
n + cc. Let AP be the perturbation matrix
82 Modem Multfuariable Control Analysls Chap. 3

that is, P - h P is a singular matrix. g ( P ) is in the order of a(AP). Looking at

Equation (3-j), the disturbance A D has therefore caused the closed-loop system P
- AP to have zero eigenvalues even though the eigenvalues of the nominal system
P are all equal to - 1 [Hewer et al., 1988(a)].


Consider the system depicted in Figure 3-3. In control system design, typical re-
quirements are (1) closed-loop stability, (2) disturbance rejection, (3) tracking ca-
pability, (4) noise insensitivity, (5) robustness against plant uncertainty, and (6)
functionality under component failures. It will be shown in this section that, under
mild assumptions, the sensitivity and the complementary sensitivity function play
crucial roles in characterizing the feedback performance. That is, the singular values
of the sensitivity and the complementary sensitivity function can be measured in
order to assess the robustness.
First, consider the disturbance rejection property; the transmission from the
disturbance source d to the output y is the sensitivity function So:

The smaller the sensitivity function, the better the disturbance rejection prop-
erty. Similarly, the tracking error as a function of the reference input r is e = (1 +
GK)-' r. Again, the smaller the value of (I + GK)-' is, the smaller the tracking
error will be. Next, consider that the plant is perturbed as (I A)-'G for some
uncertainty A. The uncertainty block matrix A can be used to represent low-fre-
quency parameter errors or the changing number of right half-plane poles of the
plant [Doyle et al., 19821. After some manipulations, it can be shown that the system
+ + '
remains stable if A is a stable transfer function and [I (I GK) - A] is nonsingular
for all w. This implies that the sensitivity matrix (1 GK)-' must be small such
that g (1 + GK) > 5 (A), which in turn means that I + G K + A is nonsingular.
Since I + GK + A is nonsingular and det(1 + (1 + GK)-'A) = det(I + GK)-'
det(1 + GK + A), the matrix I + (I + GK)-'A will be nonsingular. Another
function that is useful in characterizing the robust stability is the complementary
sensitivity function To(s) = G K ( l + GK)-'. Let the plant transfer function be
disturbed as ( I + A(s))C(s) for some stable A(s) representing sensor error or ne-


Figure 3-3 Feedback System

Sec. 3.2 Sensitivity and Complementary Sensitivity Functions 83

glected high-frequency dynamics. Assume that the maximum singular value of A(s)
is bounded by the inequality -d[A(jw)] < e m ( j w ) for all w 5 0. Then the robust
stability in the high-frequency region is guaranteed if the nominal system is asymp-
totically stable and the following condition is satisfied [Doyle and Stein, 19811:

-d[GK(;o){l + GK(;~)}-'1 < l/ern(;w) for a l l o such that e m ( o ) B 1 (3-7)

The system remains robustly stable if Equation (3-7) is true. That is, the smaller
u [ G K ( I + GK)-'(jw)] is, or the smaller the maximum singular value of the com-
plementary sensitivity function is, the better the robust stability will be. Also, the
noise effect at the output node y is y = - GK(1 GK)-'r. Hence, the smaller
GK(1 + G K ) - ' is, the better the noise immunity will be. Similar arguments can
be made for the sensitivity function (I KG)-' and the complementary sensitivity
function K G ( 1 K G ) - ' at the plant input node. More interpretations of the sen-
sitivity and complementary sensitivity functions can be found in Cruz and Perkins
[I9641 and Kwakernaak [1985]. Hence, many feedback properties can be represented
in terms of these two functions, and the robustness analysis can be done by inves-
tigating the size (singular values) of the sensitivity and the complementary sensitivity
For a particular class of problems, the control design technique that reduces
the size of these two functions or their combinations directly leads to a robust design.
There are, however, important fundamental limitations that restrict the achievable
performance. The most important one is due to the fact that So + To = (I +
GK)-' + G K ( 1 GK)-' = I and (I KG)-' + + +
KG(1 KG)-' = I. Thus,
it is impossible to make both functions small at the same time, and trade-offs have
to be made. In general one requires the sensitivity function to be small in the low-
frequency range (to reject disturbance) and the complementary sensitivity function
to be small in the high-frequency range (to be robust against high-frequency un-
certainties). Since at low frequencies the loop gain is high, the sensitivity function
in the low-frequency range can be approximated as S o ( j o ) = (I + GK)-' ==
( G K ) - ' ( j o ) . Also, since at high frequencies the loop gain is low, the complemen-
tary sensitivity function in the high-frequency range can be approximated as To(jo)
= GK(1 + GK)-' = GK(jw). These two requirements simply show that the loop
gain G K must be large at low frequencies for good tracking performance and small
at high frequencies for robustness.
o n e ' drawback of using the singular value in measuring robustness is its con-
servativeness. Conservativeness results from the structure and the nature of the
uncertainty. For a repeated, block-diagonal, or real uncertainty, a less conservative
measure has to be adopted. The structured singular value (to be discussed in Section
3.4) represents one attempt in reducing the conservativeness of the singular value.
In the following section, another feature of singular value will be discussed,
that is, how the singular value analysis of the complementary sensitivity and sen-
sitivity functions is translated into feedback design requirements.
84 Modem Plultivariable Control Analusis Chap. 3


Design requirements which the control system must satisfy in general are: (1) low
sensitivity property, a low-frequency requirement; (2) fast response characteristic,
an intermediate-frequency requirement; and (3) large stability margins, a high-fre-
quency requirement. One of the design methods which can handle these require-
ments explicitly is the frequency response method of classical control theory. How-
ever, these classical techniques cannot be directly extended to multivariable control
systems in general. Performance specifications and stability robustness requirements
for multivariable control systems can be examined by singular value plots in the
frequency domain. The plots corresponding to the maximum and minimum singular
values over a given frequency range can be established as the design criteria for a
multivariable control system design.

Low-frequency requirement. From Equation (3-6), the largest singular

value of the sensitivity matrix can be written as

where, as defined in Equation (3-I), E and g represent the maximum and minimum
singular values, respectively. In general, the upper bound p ( o ) must satisfy 0 <
p ( o ) 1 to obtain a sensitivity reduction property. Using the well-known inequality
- + A] 2 g [ A ] - 1 in Equation (3-8), a sufficient condition of Equation (3-8)
can be given as
dcK(jw)l> dw)+ 1 f o r o < o,

Since g [ G K ( j o ) ]rolls off with the gradient of -20 dB1dec above the maximum
frequency of the open-loop poles, the upper-bound frequency in Equation (3-
9) has to be lower than the maximum open-loop pole frequency. Following Ohta
and Fujimori [1988],the bound p ( o ) is studied more carefully as follows. T o restrict
p ( o ) to values less than unity means that the sensitivity and the response to dis-
turbances are reduced to at least k ( ~x ) 100%. This also means that the maximum
steady-state error will be decreased by a factor of p(0) as follows. When r(s) in
Figure 3-3 is a step function given by r(s) = [I, . . . , 1IT/s = Hls, following
Figure 3-3, the steady-statc crror e, can be written as
e , = lim SSO H = So(0)H (3- 10)

The maximum singular value of e, becomes E[e,] = if[So(O)H]S F [ S o ( O ) ] ] E [ H ] .

Therefore, E[e,] 5 p(O)if[H].

High-frequency requirement. The rcquircmcnt in thc high-fr~qucnc)~

region is to make a [ G K ( j w ) ]as small as possiblc. As shown in Ohta and Fujimori
[1988],the upper bound of a multiplicative uncertainty of the model is assumcd to
be denoted as t',,(w). Then, the condition of stability robustness can be expressed
as F[GK(jw)] < l/t,.(w) for t',,,(w) > 1. Thus, the singular value of the comple-
mentary sensitivity matrix has to be less than a given upper
- - bound. This condition
is effective for suppressing the effect of measurement noises as well.

Intermediate-frequency requirement or crossover frequency.

The crossover frequency w, is related to the response speed of the system. The larger
w, is, the faster the achievable response speed can be. However, a trade-off between
the requirement in the high-frequency region and the response speed must be taken
into acount. This consideration restricts the maximum crossover frequency w,,,,
below the crossover frequency of t',,(w), and the minimum crossover frequency
w,,,, will be determined from the settling time T,,which is an alternative measure
of the response speed. Both measures ofw, and T, are closely related with the closed-
loop poles. As shown in Ohta and Fujimori [1988], a unity feedback system is
considered with the transfer function G(s) = o:/[s(s +
25w,)]. When 0 < 5 < 1,
the crossover frequency and the settling time within 5 percent error of the command
input are given as w, = w,,[(4i4 + 1)'12 - 2(2]1'2 and

T , is inversely proportional to the real part of the closed-loop pole. The smaller 5
is and the larger o,, is, the larger w, becomes. In addition, the value of w,,,, that
gives a desirable settling time will be determined according to Equation (3-11). A
typical requirement is shown in Figure 3-4. Design engineers then shape the loop
transfer function to meet the desired requirements. There exists a fundamental lim-
itation here due to the right half-plane poles and zeros. For the system to be stable,
when the plant (or controller) has right half-plane zeros, the sensitivity function

Roll-Off C.baracIcristiu

Robust Stability

Figure 3-4 Singular Value Plots and Constraints for the Loop Transfer Function
86 Modem Multluariable Control Analysis Chap. 3

must be greater than 1 over some frequency range. O n the other hand, for a plant
having right half-plane poles, its complementary sensitivity function must be greater
than 1 over some frequency range. These fundamental limitations set up the bottom
line for design engineers in judging the achievable performance.


It is a fact in engineering practice that any model used in the design process is at
best an approximation tb the actual system. The difference between the physical
svstem and its mathematical model is called plant uncertaintv. This inevitable mod-
eling error will cause performance degradation and stability problems if plant un-
certainties are not accounted for during the design process. For single-input single-
output (SISO) systems, the standard frequency domain techniques such as Nyquist
diagram, Bode plot, root locus, inverse Nyquist, and Nichol's chart can all be used
effectively to ensure robust stability and certain desired performance characteristics.
However, these techniques do not readily extend to multivariable systems in general.
Singular value analysis discussed in Section 3.1 was introduced as a candidate
for a multivariable robustness measure in the late 1970s. In principle, the stability
margin of a multivariable system can be measured in terms of the maximum singular
value of a certain transfer function matrix if the plant uncertainty is unstructured
[Chen and Desoer, 1981; Doyle and Stein, 1981; Kimura, 1984; Lehtomaki, 19811.
Unfortunately, using singular value analysis for testing stability robustness is usually
too conservative because the actual plant uncertainties are more or less structured
in some way. This motivates the definition of the structured singular value.
In Doyle 119821, it is shown that any norm-bounded perturbation problem,
regardless of structure, can be rewritten as a structured block diagonal perturbation
problem. Consider, for example, a linear multivariablc plant G(3) with two mu]-
tiplicative perturbations appearing simultaneously a t the inputs and outputs as shown
in Figure 3-5. Hcre, K(s) is a stabilizing controller for the nominal plant G(s). In
a straightforward way Figure 3-5 can be equivalently redrawn as shown in Figure
3-6. In Figure 3-6, the combination of G(s) and K(s) inside the square box can be
denoted by a nominally stable closed-loop system M(s), a transfer function matrix
from [ d r d' vT] to [,I, T r12T e T 1. It is clear that the perturbations can then be rcp-
rescntcd by a block diagonal feedback matrix, that is.

Note that the pcrturbed systcln in Figurc 3-6 is just a special casc of the gcncral
structured block diagonal perturbation problem as shown in Figurc 3-7 with 11 =
2, A = block diag { A , , A,), 2 = [ d l , d2IT. Z = [n,, n2IT, 1,. = r, and = E . 11,
Figure 3-7, M(s) has the following form:
Sec. 3.4 Structured Singular Value 87

I+A,(s) G(s) t-+l I+Az(s) Figure 3-5 T w o Multiplicative

Perturbations Appear Simultaneously a t
the Plant Input and Output

The issue of robust stability becomes one of characterizing the set of allowable
perturbations such that the closed-loop system remains stable. Note also the ap-
pearances of the sensitivity and the complementary sensitivity functions in the matrix
It is well known that the systcm in Figure 3-7 is stable if and only if the
polynomial 'PIlnr(s)'PA(s) det [I - ibl(s)A(s)]has roots with strictly negative real
parts, where 'Picl(s)and 'PA($)are the characteristic polynomials of M ( s ) and A(s),
respectively. Since M ( s ) is stable, all the roots of 9 m ( s ) have strictly negative real
parts. Thus, if A(s) is also stable, the system is stable if and only if det [I - M ( j w )
A ( j o ) ] 0 for all A and o E R. A set of structured block diagonal perturbations
is defined as
A(s) = block diag [ A ,( s ) , Az(s), . . .
(3- 13)
Ak(s) stable, E I A k ( j o ) ]5 6(w), for all k
Then the structured singular value (SSV) [Doyle, 19821 is defined as

The following theorem is just a direct result of the definition (3-14)

Theorem 3-1. The system in Figure 3-7 is stable for all A(s) E X ( 6 ( o ) ) if
an only if
p [ M ( j w ) ] < 1/6(w) for all u E R (3- 15)
Note that the perturbation 6(w) and the structured singular value p [ M ( j w ) ] are
frequency dependent rather than constants. From the preceding theorem it can be
seen that the stability margin of the system M(s) can be defined by the structured
singular value p [ M ( j o ) ] .
While the p function is conceptually simple, its computation poses a challenge.

e +
Figure 3-6 Equivalent System Diagram of the System Shown in Figure
88 Modem Multluarlable Control Analusis Chap. 3

a = [dl, 4- .... %? n = [nl,
nz, .... bl Figure 3-7 Structured Block Diagonal
Perturbation with A = Block Diag { A , ,
I . . . , Am)

When complex perturbations are allowed, some earlier results by Doyle [I9821prove
to be very useful:

where p(M) is the maximum modulus of the eigenvalue of M [Horn and Johnson,
19851. Moreover, let it = { U E Q I U*U = I ) and CT) = { D I D A = A D for A
E Q), then

p(M) = k(UM) = p(DMD-') for all U E and D E 9

It follows that
max p(MU) I k(M) 5 inf ~ ( D M D - ' ) (3-1 7 )

It turns out that the left part of the preceding inequality is actually an equality, and
this fact provides an approach to compute the p function involving multiextremum
optimization. Under some restrictive conditions, namely the number of blocks
m a 3, the right part of the inequality in Equation (3-17) also becomes an equality
[Doyle, 19821. It should be noted, however, that when only real perturbations are
allowed, the preceding relation no longer holds, and the other methods hare to be
pursued. At present, there are several algorithms to compute structured singular
values [Doyle, 1982; Fan and Tits, 19861. However, no single algorithm is univcr-
sally superior. The situation is worse where there are mixed (real and dynamic),
repeated uncertainties.
It must also be noted that the structured singular value is merely an analytical
tool, offering no assistance for controller design if the robust stability test (3-15)
fails. Recent progress in robust controller synthesis using a combination of the H"
optimization technique and structured singular value has been reported in Doyle
11983, 1984). Although this so-called synthesis approach does not provide an exact
solution for the robust perforn~anceproblem, it often yields a satisfactory (sub-
optimal) solution. Further research is thus necdcd in this area.
Auothcr alternative measure of stability margin is called the block-limiting
norm, proposed by Fan and Fu [1989], who extcndcd thc notion of a limiting norm
[Pokrovskii, 19791 by associating with it the block structure that forms thc basis of
the structurcd singular value. It is shown that the block-limiting norm is no ICSS
than the structured singular
- value and no grcatcr than the infimum of the scalcd
maximum singular value. More interestingly, where thcrc are no morc than thrcc
blocks, the block-limiting norm is cqual to thc structurcd singular value. Unfor-
tunatcly, a computational mcthod for the block-limiting norm is also unavailable
to date.
See. 3.5 General Robustness Analysis 8Q


As described in Section 3.4, the general framework of robustness analysis can be

represented in Figure 3-7. Any system may be rearranged to fit in this general
framework, although the interconnection structure can become quite complicated
for complex systems. The uncertainty considered here may be modeled in two ways,
either as external inputs or as perturbations to the nominal model. The performance
of a system is measured in terms of the behavior of the outputs or errors. Robustness
analysis is heavily related to the assumptions made concerning the input signal, the
uncertainty, and the performance requirement. Assume that the transfer matrix from
I, to can be represented in the linear fractional map

where M , , are real, rational, proper, stable transfer functions. This assumption can
be made without loss of generality because nonlinear or nonrational parts can be
regarded as uncertainties, and the system has to be stable before robustness analysis
is carried out. Depending on the assumptions made on the input, uncertainty, and
performance, diffcrent robustness measures are used as shown in Table 3-1, which
is similar to Table 1 of Doyle [1983]. Thus, robustness analysis boils down to the
computation of a certain "norm" of a particular transfer function. In some cases,
this transfer function is a combination of the sensitivity and the complementary
sensitivity functions. Also, the weighting functions used in design process are ab-
sorbed into the interconnection structure. The definitions of each "norm" follow.

where u i is the ith singular value, E is the maximum singular value, and p,(M),
defined in Equation (3-14), is a measure of the amount of structured disturbance
that can be tolerated without destabilizing the system.
T o conclude this subsection, two important theorems from Table 3-1 are
stated. Let

BA = {A E Q 1 ll A 112 5 1 )
This leads to the following theorems.

Theorem R S (Robust Stability). F,,(M, A) is stable for all A E a if and

only if II M22 1Iw = SUP ~ [ M 2 2 ( j w )<
I 1.
90 Modern Multivarlable Control Analysk Chap. 3


w Perturbatiqn P e r f o m Robustness
YIc(t) I ~ ) =I I b(t-II) A =0 weT@)e(t)l < 1 -
2z 11%<12 <1

Ie(t) = Ieo6(t)
T 1
qIcoIco) = I A=O w T ( t ) e(t)l < 1 -
2x ~~~~~2 1
!\lc\12 1 A=O Ilel12 < 1 lly7J=< 1

IIIcllz 5 1 IlAllm < 1 Ilel12 < 11M1111= < 1

II1cll.c 5 1 A=O llella < 1 11%111 5 1

llAll.c < 1
~ ~ I 5e 1~ ~ 2 A = Diag (A, ... A$ JlelJ2c oc l l ~ ~ l <l 1

llAll= < 1
-I I I c 5~ ~1 ~ A = a g ( A A Ilel12<1 IIMllr < 1

Theorem R P (Robust Peuformance). F,(M, A) is stable and sup II F.(jo) 112


5 1 for all A E &$ if and only if 1) M 1, < 1.


The general robustness analysis provides a framework in robustness margin com-

putation. However, for all practical purposes, there are other definitions or alter-
native algorithms that achieve the same goal. Robustness measures using graphic
metric [Vidyasagar, 19841, hypersphere radius [Bhattacharyya, 19871, and Khari-
tonov's theorem [Barmish and DeMarco, 19871 have been proposed. A particularly
important problem in feedback analysis is to compute the multivariable real-per-
turbed stability margin. Using p. computation to evaluate the real-perturbed stability
margin tends to be conservative. In the following, several techniques for assessing
the real-perturbed stability margin will be discussed based on Hewer et al. [1988(a)-
(c), 19891. A state-space model for real additive perturbation errors that is algebra-
ically and dynamically equivalent to Doyle's SSV is introduced here. Based on this
theory the problem is cast as a spectral assignment problem employing output feed-
Set. 3.6 Robustness of Real Perturbattons 91

back, where the variable feedback matrix is restricted to the block diagonal elements
represented in Figure 3-7. Next, a link between root locus and SSV is established.
T o bring out the utility of this theory a Monte Carlo search algorithm is used in
conjunction with some graphical convergence strategies to compute for third-
order systems. The Monte Carlo algorithm was used by Hewer et al. [1988(a)] to
compute the SSV for the autopilot example first studied by De Gaston and Safonov
[I9881 with good numerical agreement. For this example, the bounds resulting from
the Monte Carlo algorithm are sharper than those obtained with the algorithm used
in Fan and Tits [I9861 and Doyle's estimates [Doyle, 19821, which is in line with
the well-known fact that their estimates are valid for both complex and real per-
turbations. The Monte Carlo algorithm is a "brute force" approach and should only
be used in a concurrent manner with other structured singular value algorithms and

3.6.1 State-Space Model for Additive Uncertainty

Similar to a frequency-domain formulation, in state space formulation the system

can also be arranged such that it consists of a nominal forward block

k = Ax + [Bl, Bz] [:I. [:I = [S:] (3-19)

with uncertainty expressed as feedback perturbation

The matrices A, B1, B2, C1, C2 are of dimension n x n, n x [, n x k, p x n,

and k x n, respectively. The state vector x, the input vectors rl and r2, and the
output vectors yl and y2 all have compatible dimensions. Taking the Laplace trans-
form of Equation (3-19) yields the following two-port (p +
k) X ([ +
k) block
transfer matrix [Hewer et al., 1988(b)]:

with the transfer matrices G;,(s) = C,(sl - A)-'B, (i = 1, 2; j = 1,2). The triples
(A, B1, C1)and (A, B2, C2)are both minimal (that is, controllable and observable)
realizations of the respective transfer matrix G ~(s)I and G22(5). The conditions under
which the transfer matrix Gll(s) is exponentially stable (exp. st.) [Desoer and Chan,
19731 are that it is strictly proper (that is, Gll(m) = 0) and that all of its poles are
contained in the open left half-plane. The p x 4 loop transfer matrix Gll(s) in-
corporates both open-loop plant dynamics and any compensation employed. The
advantage of the state-space formulation is that it is convenient to treat parametric
uncertainty, where the perturbations are necessarily real. In effect, the matrix A is
perturbed to A + AA, where AA = BzSACz. The matrix S is a k x k real-valued
92 Modem Multiuariable Control Analusis Chap. 3

diagonal scaling matrix, and A is a k x k member of Q. The p function is defined

as (where S has been absorbed into B2 or CZ)

A E Q and det(jw1 - A

Introducing the set

- B2AC2) = 0 for some 0 5 w Ia:
48) = {A E Qs 1 det(jw1 - A - B2AC2) = 0 for some 0 I w 5 m)

The k function can be written as p,,(A, B2, C2) = SUP

Ae(-) 11 A 11'
3.6.2 State-SpaceSingular Values

The following theorem, which is proved in Hewer et al. [1988(c)] is a state-space

robust stability theorem.
Theorem. Let (A, B1, C1) be an exp. st. system, assume e(8) is nonempty,
and assume that ( A + AA, B1, C,) is a minimal system for every A E Qs. The
system in (3-19) and (3-20) is exp. st. with respect to the real perturbation set Qs
if and only if p,(A, BZ, C2).6 < 1.
The following corollary, which establishes the basic link between the fre-
quency-derived SSV and the eigenvalue-based SSV, depends on the following well-
known identity. Let Wand Z be n x k and k x n dimensional matrices, respectively;
det(lk + ZW) = det(1, + WZ) (3-21)
Corollary. For every s with real part 2 0 and every A E e(8), p.,(A, B2,
CZ) = p(Cz(s1 - A)-'Bz).
The proof of this corollary is given in Hewer et al. [1988(b)]. These results
make it evident that the additive perturbation structure laid out previously produces
a workable and rigorous model that yields a nonconservative state-space test for
real perturbations. With Equation (3-21), it is possible to construct a Monte Carlo-
based algorithm that reduces the computation of p., to an eigenvalue search. The
corollary forms the basis for the Monte Carlo search algorithm outlined in Section
3.6.4 and in the appendix of Hewer et al. [1988(b)], reducing the computation of
p. to an eigenvalue problem. The Monte Carlo eigenvalue search is proposed as a
heuristic tool to supplement other algorithms.

3.6.3 Root Locus

This section focuses on the link between root locus and SSV and is based on Hewer
et al. [1988(b)]. A comparison is made between the stability margins predicted by
root locus for second- and third-order polynomials and those predicted by SSV
For second-order polynomials they are identical, but for third-order systems they
are distinct. Let g ( s ) denote a strictly proper nth-degree transfer function, and let
( A , b2, c2) be the minimal control-canonical form realization such that g ( s ) =
c2 ($1 - A ) - ' b 2 . Using the basic matrix identity Equation (3-21) results in the
following equation which relates root locus and the spectral set e(6), for some pa-
rameter A:

for every s such that s l , - A is nonsingular. Any polynomial subject to additive

perturbations can be represented as a dynamical control system such that ( A , b ) is
in controller canonical form, and the nonzero coefficients in the row vector c r e p
resent the additive perturbations of A . Thus, as an additive robustness measure,
root locus represents a uniform affine shift of all the perturbed coefficients. Consider
the third-order stable polynomial p ( s ) = s3 + a2s2 + a l s + a. and the canonical

Let the perturbed polynomial be p p ( s ) = s3 + a2s2 a l ( l + A l ) s + ao(l + Ao).
The Hermite matrix [Lancaster and Tismenetsky, 19851for the third-order perturbed
polynomial is

Given that the original polynomial is stable, it can be concluded that the perturbed
polynomial is stable if and only if H p ( A ) is positive definite, or, equivalently, its
leading principle minors are all positive. Applying the latter test to H p ( 0 ) produces
the well-known result that p ( s ) is stable if and only if a0 > 0 , at > 0, a2 > 0, and
ala2 > ao. The perturbation matrix A A for root locus is Ab2c2, and the perturbation
matrix A A for the structured singular value is A A = B2AC2

where the diagonal matrix is A = diag(A0, A l ) . By Hermite's test p p ( s ) is unstable

when the first leading principle minor in H , ( A ) is zero. The gain margin predicted
by root locus would be unity at w = 0 . The structured singular value spectral set
e(6) = { A E Q s I det(jw1 - A ' - B 2 A C 2 ) = 0 for some 0 Iw < m)
would contain the matrix A = diag(1, I ) , which is not necessarily the minimum
element. Now, consider the second leading principle minor of H p ( A ) , which can
94 Modem Multivarlable Control Analysis Chap. 3

vanish whenever the linear equation A . = m A l +m - 1 is satisfied, where the

positive real variable m = alazlao. The matrix

A = diag 0, l i m )

is always a member of e(S). Moreover, the maximum singular value of A is always

less than 1 inasmuch as in is always greater than 1. Thus, the results of the root
locus and the structured singular value tests are not the same even for this simple
problem. Retracing these steps for the second-order polynomial a(s) = s2 + a l s +
a,,, it is easy to see that p,(A, b, c) is always equal to the root-locus value as long
as the 2-norm is used for comparison.

3.6.4 A Monte Carlo Computation

As an application of this theory, the Monte Carlo search algorithm is used t o com-
pute a pe(A, Bz, CZ) for a third-order system, following Hewer et al. [1988(b)].
The system matrices are structured as in Equation (3-22) with a" = 1, a, = 2, a 2
= 2. T o estimate the 1.1. function, Equation (3-17) is used. Standard eigenvalue
routines were employed in the computation of the spectral radius. Its upper bound
is computed using a balancing algorithm. These inequalities hold true for real and
complex perturbations, so they will be generally conservative for real perturbations.
However, they do establish the following range where the robustness margins are
smallest. Also, they serve as a useful guide to the variation in the robustness margins
as functions of frequency. The Monte Carlo robustness margins predict 11 A l l = ~
,7268 at the frequency w = ,7678. Doyle's inequality predicts p = 2.3789 at the
frequency o = .8214. To transform this to a comparable delta, first take the recip-
rocal of p. which is ,42036 and then compute 11 diag(.42036, .42036) = 0.5945.
The Monte Carlo search uses an F-norm, which represents a 2-sphere for the
Qa set that describes the robustness region. The Monte Carlo algorithm is described
as follows:

1. Generate A randomly with uniform probabilistic distribution such that

Z A ? = 1.

2. Start with known upper bound a for e(6).

3. Set a C 2 a until either (a) or (b) as follows happens:
a. A + aB2AC2has an eigenvalue with a positive real part. Go to Step 4.
b. a is greater than some cutoff value. Let a -a. Start Step 3 again.
4. Use bisection to obtain an a such that A + aBzAC2 has an cigenvalue suf-
ficiently close to the irnaginary axis.
Sec. 3.6 Robustness of Real Perturbations 95

5. Save a. Repeat Steps 1 to 4 a sufficiently large number of cirncs. The minimum

a is an estimation of robust margin I( A 1, ,.

These same Montc Carlo techniques arc used to compute the robustness margin for
the example of a typical SISO autopilot problem. These examples provide firm
support for using the Monte Carlo scarch algorithm to find y for interesting control
problems. However, as mentioned previously, the Monte Carlo scarch is only a
heuristic tool that is best utilized in conjunction with other algorithms. It must
always be kept in mind that the Montc Carlo scarch may deliver a reasonable result
but not ncccssarily the supremum.

3.6.5 A Monte Carlo Analysis for Second-Order


As a second example, consider the following second-order system:

A = [-: -1, ='[:I, BI CI = [ I , 0,. B2 = [: c2= [i:]

Let 0 be a random variable that is uniformly distributed in the interval ( - a , a ) ,
and let A = diag(r sin 0 , r cos 0) where r is a scaling parameter such that A E e(8).
The characteristic equation for this problem is the random polynomial
A' + (3 - r cos @)A + (2 - r sin 0) = 0
Now A E e(6) if and only if the following two cases occur:

These equations define a random variable r ( 0 ) which is a member of P(8) on the

interval [2, 3/cos(arc tan(2/3))]. The cumulative distribution function for the random
variable can be derived using standard transformation formulas in Papoulis [1984].
Upon integrating,

P(r(0) 5 d} = [2/a arc cos(3Id) + f(2/d)]/2 3 5 d 5 3.6056

where f(a) = 1 - 2 / a arc sin a. The density function corresponding to this random
variable has a discontinuity a t d = 2 and d = 3. In the absence of a numerically
robust and efficient algorithm to compute y, the Monte Carlo search algorithm
combined with graphical convergence tools make for a practical, working method
that can be used to augment the Doyle bounds, and the De Gaston-Safonov and
Fan-Tits algorithms.
Sec 3.6 Robwtnes of Real Pemubatlons

Frequency At A2 A3

Carlo 7.753 .I473 e.2645 -.1933

Function 7.697 ,1229 -.3 164 -.?391

102 t 5% ESTIYATES VERSbS FRESUENCY .:..-. ..- -

. .. .
.. '..'..
* '1. 4 7

S -

V ,,,,,..,.. .""
., ......
. .' ..I.

I' 10' q
E .
S *.- --.---
- ---*--C--*r---


Figure 3-8 Robustness of Real Perturbations (From [Hewer et at., 1988(a)] with
permission from S C S )
98 Modern Multiuarlable Control Analysis Chap. 3

3.6.6 Stability Margin Computation

For the example problem considered here, the original block diagram is shown in
Figure 3-8a. This example is found in De Gaston [I9851 and includes a number of
essential features of an NGC system. The purpose of this example is to demonstrate
state-space modeling and to illustrate Doyle's claim that a linear system can be
rearranged to match Figure 3-7 and include some compound bounds for IJ.. The
nominal parameter values are: K = 800, z = 2, p = 10, o2 = 6, w-, = 4, o,,=
4, o,,= 6. The block diagram in Figure 3-8a is transformed into one with scaled
perturbations, and results in the diagram shown in Figure 3-8b. The scaling is
selected such that a value of + 1 for the perturbation represents 100 percent of the
nominal parameter. The Ai (i = 1, 2, 3) represent real perturbations and o,((i =
1, 2) are the nominal poles and scaling parameters such that the perturbation rep-
resents a percentage of the nominal parameter [Hewer et al., 1988(a)]. Transforming
the block diagram in Figure 3-8a to the desired form is accomplished by rearranging
the loops at the inputs to the perturbation blocks, which is illustrated in Figure 3-

8b. Treating the inputs and outputs of these blocks as separate system inputs and
outputs, respectively, conventional block algebra can then be used with Figure 3-
8c to obtain the desired M matrix by evaluating all nine transfer functions in Figure
3-8d. The original system inputs and outputs are no longer represented in Figure
3-8c. The exact entries for M are found in Figure 3-8e. The values are identical
with De Gaston except for sign differences. Following Hewer et al. [1989], the
controllable and observable linear dvnamical svstem for Figure " 3-8a can be derived
using the following state variables x l , x2, x-,, and x4 that are identified in the same

These equations yield the following dynamical system ( A , B1, C1):

with perturbation matrices:

B2 = [ -; !]

0 -1
2 =
c [. : :]
0 0 0 1
A = diag(Al, A2, A;)
Sec. 3.6 Robustness of Real Perturbatlons B9

and scaling matrix S = diag(K, w2, ~ 3 ) The

. perturbation matrix 6A with 6K =
KAl, 6w2 = -oS,A2, 603 = - w,,A3

factors into the product 6A = BzSACz. Before the state-space equivalence of M(s)
in Figure 3-8e can be determined, the inverse of the matrix (sl - A) must be
determined. Let (sl - A);, denote the ( n - 1) x (n - 1) matrix formed by deleting
row i and column j from A, and define the cofactor matrix

Then s l - A has an inverse equal to det-'(sl - A), multiplied by the transpose

of the matrix obtained from (sl - A) by replacing each element by its cofactor
whenever r 65 R ( A ) where R(,-l) is the spectrum of A. Now for the De Gaston
example, it is easily shown that 6 = det(s1 - A ) where 6 is defined by the equation
in Figure 3-8e. The transpose of the cofactor matrix for ( s l - A ) is:

It is not difficult to verify that the matrix M(s) = SC2(sl - A ) - ' B 2 is identical to
the transfer matrix in Figure 3-8e. The transfer matrix with input scaling Ml(s) =
Cz(:l - A)-'BZS is not equal to ,W(S) with output scaling. However, by the
fundamental determinant identity the polynomials in the three variables A 1 , A2, and
113 defined by the equations det(1 - Ml(s)A) and det(1 - M(s)A) are identical.
This observation leads to the following theorem about input-output scaling.
Theorem. If a ( A ) < 0 and e(A, B2, CZ. A) is nonempty, then for any
nonsingular transformation T that coinmutes with the members of A
p(A, B2T. C r , A) = k(A, Br, TC2, A)
Using standard root-locus techniques on the nominal system in Figure 3-Sa, the
following stability margins can be obtained [Hewer et al., 1988(a)]: A1 = ,931, A2
= - ,481, A, = - ,652. However, the simultaneous use of these marginal values
results in an unstable perturbed system. The scaling bound in Figure 3-8g predicts
instability for A i > .15 ( i = 1, 2, 3) at the frequency 7.326. Note that the margin
of stability occurs at the reciprocal of the peak. Using the Monte Carlo algorithm
and one called the penalty function technique, Figure 3-8f lists two possible near-
minimum values for the joint variation in A. Figure 3-8g shows some bounds for
CL. Curve n is a plot of the minimum singular value of M, while curve b is a plot
100 Modem Multluariable Controf Analysis Chap. 3

of the spectral norm of DMD-' using a balancing algorithm in Garbow et al. [1977].
A balancing algorithm is discussed in Osborne [1960]. An algorithm due t o Fran
and Tits [I9861 was used to generate curve c, and curve d is a plot of the spectra]
radius of M. Recall that the svectral radius is the maximum absolute value over the
eigenvalues of M. Curves a an2 d can be computed using standard routines in arbo ow
et al. [1977].


The Monte Carlo method [Hammersley and Handscomb, 19641 is the most general
method available for estimating the performance of noise-driven time-varying sys-
tems, being applicable to linear as well as nonlinear systems. The method is founded
on direct and repeated simulation,'involving a large number of random error factors.
Because of this, much of the computational work revolves around looking a t a
sufficient number of test cases plus postprocessing of the resultant data in order to
determine the average system behavior. If there is a minor drawback to this method,
it is the often large number of simulations needed to achieve an adequate level of
confidence in the accuracy of the results. For this reason, it is primarily used as an
evaluation tool. Owing to its wide range of applicability and the ease with which
it can be used, however, it remains perhaps the most widely used tool for performing
nonlinear statistical analysis. In the case of systems that are driven by white-noise
disturbances, these disturbances are approximated by Gaussian random number gen-
erators. The standard deviation of these generators, which are used to approximate
white noisc, relate to thc spccial density of the white noise being simulated by:

where R k is the covariancc of the Gaussian N(0, Rk)noise gencrator, @ is the power
spectral dcnsity of the white noise in Hertz, and A t is the interval of simulation
integration (time spacing bctwccn random nunlbcr calls). Equation (3-23) allows
for the approximation of white noise since elements in the rcal covariance matrix
Rb(t) take on infinite valucs. If the bandwidth of thc noise is much greater than the
systcm bandwidth, then Equation (3-23) is accurate to thc cxtcnt that the noisc
appcars white to thc systcm. Bccausc thc integration intcrval A t is usually taken
much sn~allerthan the sn~allcstsystcm time constant, the noisc in fact looks white
as far as thc systcnl is concerned. Statistical cstialation by the Montc Carlo tccl~niquc
begins by computing thc standard dcviation of the rcsults according to

whcre y (1,) is the mcan valuc of final rcsults 2nd Aidenotes thc numbcr of simulations
in the Montc Carlo proccdurc. Bccausc only a finite numbcr of simulation runs arc
performed, thc standard dcviation dcrived in this mcthod is only approximatc and
Sec 3.7 Monte Carlo AnalysLs


0 50 100 150 200 250



Figure 3-9 (a) Theoretical Confidence Intervals for Gaussian Distributed Random
Variable (b) Acceleration Limit Study Utilizing Monte Carlo Approach (From [Zar-
chan, 19881 with permission from AGARD)
102 Modem Multluariable Control Analysls Chap. 3

must itself be treated in a statistical fashion. This gives rise to expressing confidence
in a particular estimate of the standard deviation. The confidence level rises as the
number of simulations is made larger. As an example, for the case in which these
statistics are Gaussian distributed, confidence intervals can be calculated like those
shown in Figure 3-9a. In the case of 50 simulations, the confidence level can be
shown to be around 95 percent, that is, the standard deviation is found to be between
0.85 u and 1.28 a. The uncertainty is reduced even more by increasing the number
of simulations to 200, a sample size that would give rise to a 95 percent confidence
that the standard deviation in fact lies between .91 u and 1.12 u [Zarchan, 19881.
For this reason, confidence levels are usually presented in conjunction with Monte
Carlo data.

Example 3-1. Following Equation (3-24), the miss-distance standard de-

viation is calculated from:

where jj(tf) is the miss distance from the ith run. Following Example 2-2, a 200-
run Monte Carlo simulation was performed of the homing loop of Figure 2-l5a(ii)
and 2-15b, corresponding to the nonlinear mode. In the study conducted by Zarchan
[1988], one of the parameters was the missile acceleration limit. The results of the
study are displayed in Figure 3-9b, which includes the 95 percent confidencc in-
tervals. This study highhghts the effect of the acceleration limit on miss distance.
It can be seen that results approach those of the linear analysis only in the case of
large acceleration limit. It is interesting to note that, even with the large number of
simulation runs, the 95 percent confidence limits point out the high degree of un-
certainty in the answers.


It is important to be able to calculate the vchicle responsc to random process modcls

such as atmospheric turbulence. Calculating the response of dynamic systems to
random processcs can be achievcd through the use ofthe covariance matrix equation,
which is applicable to both deterministic and random excitation functions. Using
the computer-implemented method of covariance analysis, only one computcr run
is required to do an exact analysis of linear and nonlinear, time-varying noise-driven
systems. Uy directly intcgrating a linear or nonlinear matrix differential equation,
the covariance matrix of thc system state vector is propagated in time to generate
exact time-dependent statistical performance projections of any state. Extensive use
Sec. 3.8 Covariance Analysls 103

of this nlcthod is found with problems involving optimal cstimators, as is shown

in Chapter 4. The linear differential equations rcprcsenting the vehicle's equations
of motion are first written as linear timc-invariant statc equations. The statc equation
includes a term representing the random process input function. The statc vector
is then augmented to include the components of the original statc vectors, the orig-
inal systcnl input, plus any required auxiliary variables, that is, thc auxiliary dis-
turbance statc variables, as shown in Equation (3) ofTable 2-1 where, for simplicity,
D and E are assunled zero from this point on. and where w is a white-noise vector
having PSI1 matrix Q ( t ) which gives the 1'SD function of the white-noise process
a constant value of ( l I 2 n ) Q over all possible values of the circular frequency w. For
example, the properties of the turbulence field are usually specified as PSD functions
making the preceding representation suitable for aeronautical applications. Equation
(3) of Table 2-1 can be used to analyze whole-body motion, flexible motions, and
control system dynamics. Handling unsteady aerodynamic effects can be achieved
through a quasi-steady assumption, through the use of approximate lift-growth
functions, or by the numerical matching of unsteady aerodynamic matrices. Bias
and drift, which are both types of nonwhite-noise inputs, can be handled by white
noise passing through an appropriate shaping filter. It is also interesting to take note
of the covariance matrix of the augmented state and output as follows:
P(t) = E[(x(t) - x)(x(t) - x ) ~ ] , x = E[x(t)] (3-26a)
~,(~)=E[(Y(~)-Y)(Y(~)-Y Y =) E~ []Y, ( t ) l (3-26b)
For the augmented state, the covariance takes on the following matrix form:

u u:.> ... 1
u;, is used for the variances of the
individual states .u,.

... is used to represent the covariances

between different states xi and xj.

T o take it one step further, the diagonal elements of P(t) denote state variable
variations when the expectations of the disturbance responses are zero. Furthermore,
the off-diagonal elements denote the degree of correlation between the state vari-
ables. In the proceeding discussion, both the system states x and the forcing functions
w are vectors, the elements of which are random variables. For simplicity, the
deterministic inputs u are neglected here. There are two situations (continuous and
discrete cases) to be analyzed. In the first situation, the state vector obeys the linear
relationship Equation (3) of Table 2-1 without the input term Bu. The covariance
matrix is defined as in Equation (3-26a) and the covariance matrix P satisfies

This equation is used to study the parameters that characterize the state during the
vehicle's flight. Since Equation (3-27) can be integrated, the statistical properties
104 Modem Multluarlable Control Analysis Chap. 3

of P(t) can be analyzed directly in only one computer run. The solution of Equation
(3-27) gives the augmented state covariance matrix, which can be written as

Stability of the augmented dynamic matrix A guarantees that P(W) = 0.SO that, if
A is stable, at steady state 0 = AP +
PAr + NQNr. Using Equation (3-26a),
the covariance matrix Equation (3-26b) is given by P, = CPCr. In the discrete-
state model given by Equation (5) of Table 2-1, assuming 4' ' = 0, the equation for
projecting the error covariance is given by

Furthermore, for a continuous nonlinear dynamic system described by Equation (1)

of Table 2-1, the nonlinear variance equation is approximated by
~ ( t=
) A(x(t), t)P(t) + P(t)AT(x(t), t ) + N(t)Q(t)Nr(t) (3-30)
where A(x(t), t) is the matrix with its ijth element defined by Aci(x(t),t) = [ a f , ( x ( t ) ,
t)]l[dxj]. In the nonlinear discrete case, similar results can be obtained.

Example 3-2. This example, following Zarchan [1979(a)] and Example

2-2, again uses the linear homing guidance loop of Figure 2-15a(ii) and 2-15b to
illustrate here the usefulness of covariance analysis. The system equation expressed
in matrix form is
The state covariance P(t) is obtained by intcgrating the covariance propagation
equation, Equation (3-27), with N equal to an identy matrix. With respect to Equa-
tion (3-27), the preceding equations define A ( t ) , and can be uscd to determine Q ( t )
as follows:

Fi ure 3-10a plots the RMS relative separation between the missile and target,
h, as a function of time, which war found by intcgrating Equation (3-27).
This value becomes the RMS miss distance at if, or the end of flight. That is, RMS

TlME lrecl

a) corariance Ana1y.i~ Provides

RI Trajectory Profile
TlME lucl

b) cwariuc. Malyml. Provldoa

mm Accoloretion Protile

Figure 3-10 Covariance Analysis Provides RMS Trajectory Profile (From [Zar-
than, 1979(a)], 0 1979 AIAA)
106 Modem MultIuarlable Control Analysis Chap. 3

, (,,,,. Because covariance analysis is capable of generating

miss distance = v ~ ( 3 3)
statistical information corresponding to each state, like the RMS acceleration of
Figures 3-lob, it is possible with this technique to establish the correctness of as-
sumptions concerning system linearity. That is, such an assumption would be val-
idated if there were no acceleration saturation.


The method of adjoints [Besner and Shinar, 1979; Derusso et al., 1970; EASAM,
1969; Howe, 1965; Laning and Battin, 1956; Peterson, 1961; Zarchan, 1979(a)] rep-
resents a widely used technique in guidance system design. This adjoint technique
is founded on the system impulse response, and is capable of analyzing exactly linear,
time-varying noise-driven systems such as a guidance loop in one computer run.
The exact performance projections of any quantity at any instant can be estimated
using this method. Additionally, the information that measures the contribution of
all disturbances (inputs) to the total performance projection (output) can be ex-
tracted. Although it has been used mainly in the past in missile guidance system
design and analysis, the adjoint technique can certainly be applied to many other

3.9.1 A4joint Philosophy

The impulse response of the adjoint system I* is related to the impulse response of
the original system 1 in the following manner:
I*(tf - ti, tf - to) = I(to, ti) (3-32)
where t , and to are the impulse application and observation times of the original
system, respectively. If it is desired to observe the impulse response of the original
system at the final time tf after several impulses are applied at times t, ( i = 1,
. . . , n), then the system response must be simulated for each of the impulse ap-
plication times to generate [(I/, t,), as demonstrated in Figure 3-lla. For the adjoint
system, however, if the observation time is the final time (to = t,), then only one
adjoint response must be generatcd. Equation (3-32) reduces to I*(tf - I,, 0) =
I((/, 1,). The impulse response of the adjoint system, except for being generatcd
backwards, is identical in evcry aspect to that of the original system. Figures 3-1 la
and 3-llb illustrate the way in which the two responscs are relatcd [Zarchan,
The system rcsponse y(t) at time t to any input u is given by

T o dctcrminc how the method of adjoints becomes handy in this situation, Equation
(3-32) is substituted into Equation (3-33) to obtain [Pctcrson, 19611
~ c 3c.9 Aaolnt Method

Uto, t , )

crvation Time


Impulse Application Time.t i tO=ti

Figure 3-11 (a) Generation of I(tf, ti) in Original System (b) Impulse Responses
of Original and Adjoint Systems are Related (From [Zarchan, 1979(a)], 6 1979


y(t) = ul 0

I*(tf - T . tf - t) U(T) di
Z = t - T
3 y(t)

=lI*(tf-t+z,y-t)u(r-z)dz (3-34)
108 Modem Multiuariable Control Analysis Chap. 3

If the final time t k is the time at which the response is desired, then Equation
(3-34) becomes

The concept of adjoint system can be extended to account for stochastic inputs. A
linear time-varying system that is driven by white noise has a mean square response
[Zarchan, 1979(a)]

where 4, is the spectral density of the white-noise input in units of Hertz, and is
assumed double-sided and stationary. Unfortunately, the many computer runs re-
quired to generate I(tf, t i ) render the direct simulation of Equation (3-36) imprac-
tical, as mentioned previously. Substitution of Equation (3-32) into Equation
(3-36) yields

For the case in which the final time is of interest ( t = tf), Equation (3-37) can be
expressed as

Therefore, the integral of the square of the output of the impulsively driven adjoint
system from t = O to t = tf is equixralent to the mean-square response at time t, of
a linear time-varying system driven by white noise. The RMS value of the output
of the terminal time is the square root of Equation (3-38). The RMS value of the
output of the original system in the presence of a white-noise input is then computed
by first squaring, then integrating, and finally taking the square root of the adjoint
system's impulse response.

3.9.2 Applications

From any linear system, an adjoint system can be constructed by the following stcps
[Besncr and Shinar, 1979; Pctcrson, 1961; Zarchan, 1979(a)]:

1. Substitute tf - t for t in the arguments of all variable coefficients, that is,

2. Reverse all signal flow, redefining branch points as summing junctions and
vice versa.
3. The input to the adjoint system is a unit i~npulse6(t), whose application point
corresponds to the output of interest in the original system.
Sec. 3.9 AdJolnt Method 10g

Using the systcnl adjoint together with the outputs described previously, it
is possible to plot llMS miss at interception versus range a t the start o f hotning for
each computer run. Because the RMS miss caused by each of the input quantities
can likewise be obtained, it is also possible to isolate which factors have the greatest
influence on RMS miss. Using the adjoint system technique is therefore much less
time consuming than using Monte Carlo techniques and applying noise generators
to the r c g ~ ~ l analog
ar circuit. By going to acceleration adjoints, one can also study
the RMS lateral acceleration requirements for the missile. A balanced design can
then be achieved by comparing miss and acceleration statistics. It is easily seen that
the adjoint technique is well suited to the problem of statistical atlalysis and opti-
mization of linear time-varying systems. Using this technique, quick estimates of
the required gains, natural frequencies and damping, acceleration limits, and ailow-
able radomc tolerances can be obtained for several initial geometry and other input
conditions. As a preliminary design tool, the method of adjoints can significantly
lower the amount of work involved in the ensuing analysis of a refined nonlinear
system model [Peterson, 19611.
Example 3-3. The adjoint solution resulting from only one simulation
supplies a tremendous amount of information concerning system performance and
behavior as shown in this esa~llplefrom Zarchan [1979(a)]. What has been presented
can now be used to study thc homing loop shown in Figures 2-15a(iii) and 2-15b
of Examplc 2-2. In the deterministic input, the homing loop has a step target ma-
neuver xvhich has been converted to an impulsive input through integration. The
output to be considered is the miss distance y ~ ( t f due
) to the step target maneuver.
Figure 3 - 1 3 shows an adjoint model of the homing guidance loop of Figures 2-
13a(ii) and 2-15b, constructed according to the rules laid down previously. The
idea is to apply an in~pulseto the adjoint system at the location X4, corresponding
to xvhich the output ofinterest in the original system is denoted by y,. In the actual
simulation, however, an initial condition of unity is specified at X4 as opposed to
a unit impulse o n the derivative. In the process of constructing the adjoint model,
the three inputs o f the original system, target maneuver, glint noise, and fading
noise, all become outputs. That is, the outputs in the adjoint system are now miss
sensitivities due to target maneuver (y=(tf) for deterministic, y.ar,(tf) for stochastic),
miss sensitivity due to glint noise y,, and miss sensitivity due to fading noise yf(tf).
Stochastic inputs. Because there is no dependence of the sensitivity coef-
ficients of the adjoint system on the spectral density levels of the error sources,
changes in the latter do not necessitate additional simulation runs. The total RMS
miss distance, [E(yz(tf))]1'2 applying the principle of superposition, is given by
uy,(tf) = {u;,(tf) I Tgt Mvr + u:,(tf) I Glint + uf,(tf) I RIN
+ u;,(tf) I Range Dep.)Ii2
yT(Q Step Target Maneuve~
Miss Distance Sensitivity


4r ('t)


M = 11s 1 I
2 q1( I f )
9,( 1 0

= -,/+ ~ E) IIY($)I
=~ ~ Y L ? + ~ + YI: (141




Figure 3-12 (a) Adjoint Model ofLinear Homing Loop (b) RMS Miss Distance Error Budgct
Is Automatically Generated hy Adjoint Method (c) Adjoint IJrovides information on Perfor-
mance Sensitivity l>ue to Target Mat~cuver(From /Zarchan, 197Y(a)l. Q 1979 AIAA; and
/Zarrhat~,19881 ulirh prJnnissiot#.jam ACARDJ
Sec. 3.10 Statistical Linearization 111

Figurc 3-12b plots the individual RMS contributions to miss distance together with
the total RMS miss distancc as functions of adjoint time. Hence, adjoint time implies
cithcr timc of flight or time to go at which disturbances occur. The figure shows
that the biggest contributor to miss distancc in this homing guidance loop is glint

Deterministic inputs. It is, howcvcr, neither a difficult nor expensive

task to compute othcr disturbance sensitivities as well. For example, thc miss dis-
tance scnsitivity due to a step in target acceleration y T(tf) was also calculated and
plotted against timc in Figure 3-12c, which suggests an optimal time of 0.6 sec for
the targct to maneuver in ordcr to rnaximizc miss distance before intercept. The
figure also indicates that miss distance will bc small if the target maneuvcrs too
soon, that is, if the adjoint timc is too large. As the adjoint time approaches infinity,
y T(tf) will always approach zero if the guidance system is well designed. The actual
adjoint timc required for the miss-distance scnsitivity curve to settle down is a
function of the overall guidance system time constant.


When it is desired to treat nonlinearities in a stochastic guidance system, that is,

acceleration saturation, Monte Carlo techniques must be employed to compute
mean-square miss distance. The problem lies with the large number of trials that
must be run to achieve even moderate results. Most often, system nonlinearities are
dealt with through the construction of a linear model. Using suitable conditions
(for example, assuming the actual trajectory of the vehicle is close to its nominal
trajectory), a nonlinear function can be expanded in a Taylor series about some
operating point (nominal state vector) of the nominal trajectory and only the first-
order or linear terms kept. This approximates a nonlinear function for small per-
turbations about the operating point and can be used to obtain small signal lin-
earization. When an input is no longer accurately represented by a particular lin-
earization, the input is relinearized about a new operating point. The entire trajectory
can be segmented so that the small linearization approximation is valid over each
segment to reduce the analysis errors. The perturbations of the system states around
the nominal values can be studied using the covariance method. If changes in the
variable N G C loop parameters are very slow in comparison with the rate of the
associated processes, then the errors introduced by treating the time-dependent coef-
ficients in the linearized equation as constants will be correspondingly small. More-
over, significant analysis errors are not introduced by the process of linearizing the
N G C loop equations if: (1) the quantity in question is deterministic; or (2) if the 3u
value of the random signal at the input to each of the elements does not exceed the
112 Modem Multiuarlable Control Analysis Chap. 3

linear region of the element; or (3) the linearized function is differentiated. An even
more subtle restriction is posed by the fact that the linearization procedure involves
the system state vector itself. Thus, any results obtained will require a lot of com-
putation. These types of restrictions therefore lead to Monte Carlo approaches.
However, discontinuities in the system (for example, jamming and sudden target
maneuver in interce~t" guidance) such as a limiter and two-level switches restrict
these techniques and render linearization in the ordinary sense impossible [Gelb and
Warren, 19731.
When Monte Carlo techniques are required in the linearization procedure, or
when the conditions for ordinary linearization are lacking, statistical linearization
may be used. A statistical linearization that approximates a nonlinear operation, but
depends on some properties of the input signal, is often referred to as a quasi-
linearization. This technique can produce different linear approximations to the same
nonlinear function applied to different input signal forms. Quasi-linear approxi-
mation, unlike linear approximation, can be used for any range of signal magnitudes,
and is dependent on the input signal amplitude [Gelb and Vander Velde, 19681.
What has emerged as an immensely useful tool for the analysis of nonlinear systems
having random inputs is the method of statistical linearization. When statistical
linearization is used to analyze nonlinear systems with random inputs, the nonlinear
element is replaced by an equivalent gain which depends on the form of the input
signal. A method for determining the equivalent gain was developed by Booton
[I9531 and Gelb (19741. Consider an approximation of the nonlinear function f(s)
by a linear function, in the sense suggested by Figure 3-13, using statistical lin-
earization. T h e input x in the nonlinear system can be decomposed into a mean
component m plus a zero-mean illdependent random process I., giving x = PI + r.
For this systcm, it is desired to replace the nonlinear element by an equivalent gain

Figure 3-13 Statistical Linearizatiori

Approxinlation (Frorn /Grlh, 1Y80/ufirlr
pcrrrriuiorl ,fnorn Tltc Arrillyti( Sricriic~
Corp. )
K,,. The error signal e(t) is defined as the difference bctwcen y = f ( x ) and the
equivalent gain output, and has a mean-square value defined by

Using standard calculus, the minimum value of E[e2] can be found by differentiating
this quantity with respect to K,, and setting the derivative equal to zero. When this
is done, the result is

It is assumed that the signal r(t) is a zero-mean Gaussian random process whose
probability density function is p(r) = e-'2'('"2) /(6 u) where u is the RMS value
of r(t). The equivalent gains can therefore be written as

K,, = d3
dr, K, =
1 ,.f(,,, + r)e-r21(zo2) dr
-7 utn d 2 7 u~

Now, when it is known that the signal x itself is a zero-mean random Gaussian
process, m is zero, that is, x = r and the equivalent gain reduces to

In Gelb and Vander Velde [I9681 input-sensitive gains of the preceding type which
approximate the transfer characteristics of the nonlinearity are called random input
describing functions. These functions are tabulated for several important nonlin-

Example 3-4. This example, following the work of Zarchan [1979(a)] and
Example 2-2, centers around the limiter shown in Figure 3-14a. The describing
function given by Equation (3-43) becomes

xe - ~ 2 / ( 2 ~ 3
dx + -* 1 lim
J-lim x2e-x21(2u:) dx


xe -s1(2u:) dx
Modem Multlvariable Control Analysls Chap. 3



0 lb h ;O do 50 60

Figure 3-14 (a) Input-Output Characteristics of a Limiter (b) Randoni Input Describing
Function Approximation to Limiter (c) Acceleration Limit Study Utilizing CADET Approach
(d) SLAM Model of Nonlinear Homing Loop (e) Describing Functions for Various Accel-
eration Limits vs. Forward Time (0 Acceleration Limit Study Utilizing SLAM Approach (g)
SLAM Provides Information on Performance Sensitivity Due t o Target Maneuver (From [ Z a r -
than, 1979fa)). 0 1979 AIAA; and /Zarchan, 19881 u~irhprnnirsion.fiom ACARD)


A ~ Y




116 Modem Multluarhble Control Analysis Chap. 3

which can be evaluated to give

1 -x2/(20f) dx

Finally, rewriting this last equation in terms of the probability integral yields as the
describing function -.

Abramowitz and Stegun [I9641 give the following approximation to the preceding
integral, which can also be found by table lookup, that is accurate to five decimal
K,, = 1 - (21fi)e-""'~'(~":) (0.43618360 - 0. 1201676w2 + 0.937298w3)

Figure 3-14b plots the describing function for the limiter against values of lim/u,,
the ratio of the limit to the RMS value ofthe input signal. As expected, the describing
function is determined by these two values only.

3.10.1 Techniques and Tools for Statistical Linearization

Recent techniques involving statistical linearization of nonlinear system elements

with covariance analysis allow for the direct statistical analysis of nonlinear systems
and have been used on nonlinear missile guidance systems [Pricc and Warren, 19731.
An important tool for analyzing and evaluating the statistical behavior of nonlinear
stochastic systems is an approximate computer-implemented technique known as
the covariance analysis describing function technique (CADET). It involvcs the
direct statistical linearization of nonlinear systems, combining in the process clc-
ments of covariance analysis with those of random input describing function analysis
[Booton, 1953; Gelb and Vandcr Velde, 19681 to yield statistical pcrformancc pro-
jections in one cornputcr run. Thc dcrivation of the describing functions makcs usc
of a Gaussian assumption, which may at first glance seem unduly restrictive. How-
ever, because lincar elements outnunlbcr their nonlinear countcrparts in the majority
of dynamical systems, this turns out not to be the case. Morcover, nonlinear outputs
that are not Gaussian in nature are easily converted to nearly Gaussian inputs through
the use of low-pass filtering. What has been shown is that the large nurnbcr of
Monte Carlo runs can be replaced by one computer run with this ncw method, and
that the mean-square miss distance can be computed about as accurately. For 111any
types of nonlinear systems the CADET method [Gelb and warren, 19731 can often
be used as a less expensive alternative to the Monte Carlo approach in order to
obtain approximate performance projections. CADET has proved itself to be a useful
and efficient tool in the preliminary evaluation of nonlinear missile guidance system
performance. T o apply CADET to guidance systems, the following principal steps
must be implemented.

1. Substitute the appropriate random input describing function gain for each non-
linear element in the original system, assuming a Gaussian probability density
function for the input to the nonlinearity.
2. With the linear system model that results from the preceding substitution, use
conventional covariance analysis techniques (Section 3.8) to propagate the sta-
tistics of the system state vector. Note that the describing function gains are
themselves functions of these statistics.
3. Use the elements of the system covariance matrix to calculate the mean-square
output at t,, for example, mean-square miss distance at the intercept time.

Example 3-5. This example follows the work of Zarchan [1979(a)] and
Example 2-2. It illustrates how CADET can be applied to the system of Figure 2-
lja(ii) and 2-15b when the latter is operating in the nonlinear mode with saturation
effects. Following the previous discussion, the first step is to replace the acceleration
saturation nonlinearity by a random input describing function Kl,,. The linearized
system equation is then given by
118 Modem Multluariable Control Analysls Chap. 3

At this point, P(t) is found by integrating Equation (3-27), in which A is obtained

from the preceding linearized system equation and Q is still given by Equation (3-
31). The previously derived describing function gain for the limiter is a function of
the statistics of the unlimited commanded acceleration u, and the limit level nli,,
and can be computed from Equation (3-47). T o compute the RMS level of input
signal to the nonlinearity u:,, u, is first expressed as a function of the states. The
resulting mean-square value can be written as 02, = (AV,)'P(5, 5). Example 2-2
provides the input values that were used to perform C A D E T analysis of the homing
guidance loop. As in the case of the previous Monte Carlo analysis, the parameter
that was selected for the simulation runs was the missile acceleration limit MI,,,.
Figure 3-14c displays the results of this study, which were generated from six
C A D E T runs. This can be compared with the 200 runs used to generate the Monte
Carlo results, which are superimposed on the results of this study in Figure 3-14c.
T h e results shown in the figure point to the high degree of accuracy that can be
obtained using CADET. A study undertaken by Price and Warren [I9731 showed
that CADET was capable of producing results comparable in accuracy to those
obtained by Monte Carlo analysis involving several hundred simulation runs.

3.10.2 Statistical Linearization with A4joint Method

Statistical linearization has also been used in connection with adjoint techniques to
produce successful results. The combination of the two methods allows for the
assessment of primary factors in the overall measure of performance as well as the
ability to cvaluatc the stability of the guidance system. SLAM [Zarchan, 1979(a)]
in particular, is one of the approaches that have combined statistical linearization
with an adjoint method, and is another example of an approximate, computerized
technique that is available for the complete statistical analysis of nonlinear noise-
driven systems. It basically combines CADET with an adjoint tcchnique, the result
of which is capable of generating accurate statistical performance projections and
producing an approximate error budget that quantifies the influence of each dis-
trubance on the total system performance. In addition, SLAM has been successfully
applied to preliminary analyses of guidance system performance. The following
discussions on SLAM are based on [Zarchan, 19881. The principal steps to be fol-
lowed for using SLAM are:

1. This stcp is identical to stcp 1 of the C A D E T method.

2. This step is identical to stcp 2 of thc C A D E T mcthod.
3. Store thc resulting dcscribing function gains for cach nonlinearity as a function
of time.
Sec 3.10 Statistical Linearization llD

4. Convert the linearized system modcl to an adjoint model in the following way:
(a) substitute tr - t for t i n the arguments of all variable coefficients including
the described function gains; and (b) reverse signal flow, causing the original
systcm inputs to become adjoint system outputs.
5. Propagate the systcm in adjoint time.

Steps 4 and 5 together replace the linearized system model of the original system
by its adjoint modcl. As pointed out previously, the RMS miss distance derived
from SLAM is identical to that derived from CADET. However, SLAM can ad-
ditionally generate an approximate error budget that shows how each individual
error disturbance contributes to the total RMS miss distance. The approximate
nature of this error budget, which is valid only so long as the time history of the
gain K,,,,,of the describing function is valid, renders suspect any estimates of miss
distance for different error source input levels based on extrapolations. Such an error
budget is still useful inasmuch as it serves to put in relief those error sources that
have the greatest impact on the total RMS miss distance in the nonlinear system.
It is also much less expensive to generate than its counterpart derived from the use
of Montc Carlo or C A D E T techniques. These methods, in order to examine in-
dividual contributors to the total RMS miss distance, must rely on running simu-
lations with one error source a t a time, the results of which are combined in some
fashion to compute the total RAMSmiss distance. For nonlinear systems, this method
of combining individual error sources does not always produce the same total RMS
miss distance that would have resulted from running all error inputs at once.
In addition to the aforementioned error budget, another product that arises
from the use of statistical linearization methods are sensitivity functions. Although
they cannot be used to compute miss distance, being more qualitative in nature,
these functions relate system sensitivity to a step-target maneuver, and provide
insight into the relative stability of the system. Finally, another meritorious feature
of the SLAM concept lies in its self-checking nature. Should the adjoint and CADET
portions of the program fail to yield identical RMS miss distances, then it is certain
that the program is flawed in some manner, conceptually or otherwise. While it is
not possible for any program to totally eliminate the possibility of undetected errors,
the SLAM concept makes great strides toward this goal.

Example 3-6. Following Zarchan [1979(a)], the example - presented

. here
looks again at the nonlinear stochastic guidance system of Figures 2-15a(ii) and 2-
l j b to illustrate the usefulness of SLAM. Using the inputs of Example 2-2, the
time history of the describing function Kli, is first generated from the CADET
module of the program. Substituting tf - t for t as described earlier, the SLAM
program then generates K*,,,, the gain of the reversed describing function, which
120 Modem Multiuarlable Control Analysls Chap. 3

is entered into a linearized adjoint model of the original nonlinear system. This is
shown in Figure 3-14d. An approximate error budget for the nonlinear system is
then generated by running the adjoint module of the SLAM program. The parameter
selected in this study was, as in the case of the previous examples in this chapter,
the acceleration limit nli,. Figure 3-14e plots the describing function gains resulting
from the C A D E T analysis as a function of forward time for different values of nl,,.
It can be seen from the figure that, as long as no saturation occurs, the gains are
unity, and that these decrease with increased saturation. This is, of course, as the
theory predicted. Figure 3-14f plots the total RMS miss-distance error budget ob-
tained from SLAM as a function of the missile acceleration limit. As expected, the
total computed RMS miss distance corresponding to each of the acceleration limits
is identical to that obtained from running the CADET program alone in Example
3-5 (see Figure 3-14c). When the system is slightly saturated, that is, ittin, = 60 g,
glint noise can be seen to be the greatest contributor to the total RMS miss distance,
which can also be confirmed by Figure 3-12b. This is quickly superseded by random
target maneuver as saturation increases, or as nl,, decreases. Conversely, the con-
tributions from fading and glint noise slightly decrease. This is understood by re-
alizing that saturation acts to provide additional filtering.
The sensitivity function due to a step-target maneuver is also plotted in Figure
3-149 as a function of adjoint time for various values of titi,,,. These functions
represent mathematically the impulse response of the quasi-linearized system, but
cannot, as mentioned before, be used to compute the miss distance due to step-
target maneuver. The sensitivity curve of Figure 3-14g corresponding to t~l,,,,= 5 ,
that is, the linear casc, first peaks then rapidly decreases to an asyn~ptoticvalue of
zero. This situation, which is also depicted in Figure 3-12c, would be found to
occur in a well-designed missile guidance system en~ployingI'NG. I n this casc. a
target maneuver initiated a t a time to go of not less than ten guidance time constants
should have no bearing 011 the total miss distance. Figure 3-14g demonstrates that
the miss-distance sensitivity asymptotically approaches values other than zero as
soon as an acceleration limit is introduced. Such a situation is typical of PNG systems
whose effective navigation ratio is not sufficiently large. In the case of acceleration
limits smaller than those indicated in the figure, there would be a monotonic increase
in the sensitivity function, representing a system that is stable yet incapable of guid-
ing effectively on maneuvering targets. Such information, while not of a purely
quantitative nature, noncthclcss provides the guidance system designer with insight
into thc details of system behavior.

3.10.3 Applications

It has been shown [Zarchan, 1979(a)]that the preceding C A D E T and SLAM tncth-
ods produce the same degree of accuracy demonstrated in Monte Carlo methods
applied to missile guidance system evaluation. In gcncral, though, thc missile ac-
Five to Seven Times the Acceleration Capability of the Tar et
Is Needed by the Missile in Order to Intercept the Target sing PNG.
Two to Four Times the Acceleration Ca ability of the Tar et Is Needed
b the Missile in Order to Intercept the arget Using APr$or Advanced
duidance Law.

Linear Analysis

20 30
Missile Acceleration Limit (g's)
Figure 3-15 Missile Acceleration for Target Intercept: Nonlinear System Analysis

celeration limit greatly influences the RMS miss distance. Figure 3-15 shows that,
for 7-g maneuvering targets an acceleration capability five to seven times greater
than that of the target is required by the missile to be able to intercept the target
close to the accepted level if PNG law is used. However, the use of augmented
proportional navigation (APN) guidance or modern guidance laws requires the mis-
sile to have an acceleration capability two to four times that of the target to be able
to intercept the target. For further details, see Chapters 6 and 8.


The qualitative comparison presented in this section is based on material presented

in Zarchan [1988]. The selection of a particular computerized method for performing
statistical analysis is not always a straightforward matter. Depending on the system
being analyzed and the type of information required, there are occasions that call
122 Modem Mu(tiuariab1e Control Analysis Chap. 3

for the use of more than simply one method to gain a complete system understand-
ing. Because it requires basically a random number generator and a postprocessing
routine, in addition to a simulation program for the system equations, the Monte
Carlo method is the most general and most easily applied o f these techniques. In-
spection of a block diagram of the system is often easily translated into the system
equations. When it is desired to use the adjoint technique, the modified diagram
can be obtained from the original block diagram of the system, and these are also
easily translated into system equations. For this reason, the adjoint technique is
almost as easy to utilize as the Monte Carlo method. The same is not true of co-
variance analysis, which requires the system equations to first be expressed in state-
space form. Covariance analysis is therefore more difficult to implement, and sys-
tems composed of both analog and digital sections must be handled very carefully.
Problems that are basically linear except for the inclusion of specific nonlinearities
such as saturations or dead zone lend themselves well to analysis by application of
the CADET and SLAM techniques, both of which are more difficult than the Monte
Carlo method t o utilize yet no more so than covariance analysis.
Besides difficulty of implementation, another way in which the methods differ
from one another is their C P U usage, or computer running time. The greatcr the
number of differential equations needed to describe the system and the number of
computer runs required to perform a statistical analysis, the higher is the cost of
doing the analysis. The Monte Carlo method requires many more computer runs
than any of the other methods to obtain accurate results; whereas, in the casc of
linear systems, the adjoint and covariance methods only require one run because
they are both exact methods. For an 11-state system, n differential cquations must
be integrated for the adjoint technique compared to ti' differential equations for a
covariance method. Thcrc arc no exact methods of statistical analysis for nonlinear
system analysis. The accuracy obtained by using CADET and SLAM is close to
that obtained from several hundred Monte Carlo runs [Gelb and Warren, 1973; Price
and Warren, 1973; Zarchan, 1979(a)]. For an 11-stable system, 17 differential equations
must be integrated with CADET, while ,I' + rr cquations must be integrated with
SLAM. A comparison of the different statistical analysis techniques can therefore
be made on the basis of cost, which is defined hcre as

cost A number of equations x number of runs (3-48)

Figurcs 3-16a and 3-16b comparc the costs of using these methods. The): s l ~ o \ ~ ~
that, in spite of the usually large number of computer runs required, thc Montc
Carlo mcthod can be chcapcr to i~nplcmcntthan covariance analysis, CADET. or
SLAM. In the previous discussions of the quantitative and qualitative aspects of thc
various computerized statistical analysis mcthods, it has been dcmonstratcd for thc
case of linear systems that the mcthods of adjoints and covariance analysis yield
exact pcrformancc projections. In the casc of nonlinear system analysis, the Montc
Carlo mcthod is easily rivalcd by the CADET and SLAM techniques.

10 100 1000
a) coat Cmparlmn lor Linear Syatema



b) Comt C-11- lor Nonlinear Syatema

Figure 3-16 (a) Cost Comparison for Linear Systems (b) Cost Comparison for
Nonlinear Systems (From [Zarchat~,19881 with permission from ACARD)
124 Modem Muitiuarhble Control AnalysIs Chap. 3


Stochastic disturbances that greatly affect NGC performance must be identified and
accounted for in the NGC system synthesis. The turbulence and gust velocities
produce incremental aerodynamic forces and moments acting on the vehicle. The
turbulence and gust disturbances are applied to the aerodynamic terms in the equa-
tions of motion. but not to the inertial or acceleration terms. Similarly. ,, their effects
- ..
are included for the aerodynamic sensors but not for the inertial sensors. For ex-
ample, a 30-ftlsec horizontal crosswind hitting a flight vehicle with a 300-ftlsec
forward speed will create a 5.7-deg aerodynamic sideslip while the inertial sideslip
still remains zero. Turbulence and gust are responsible for random variations in
airspeed VtOt,r,in the angle of attack a, and in the sideslip angle P. The model
accounts for turbulence and gust through gust components of a, P, Vt,,,l as a =
ai + a,, Q = Q i + f&, and Vt,,,l = V i+ V,, respectively, where If,, ol,, and p,
are the increments of VtOtal, a, and p, respectively, due to the turbulence and gust.
The dynamic model includes three channels of independently generated Dryden or
von Karman model turbulence.
Dynamic loads due to atmospheric turbulence, discrete gusts, and discrete
control inputs must be calculated in order to analyze vehicle's flexibility effects. In
the dynamic load analysis, the following data are needed: (1) externally produced
structural data composed of mode shapes, lumped masses, generalized masses, and
generalized stiffness; (2) vehicle geometry; (3) control law; and (4) information of
flight condition, for example, velocity, altitude, Mach number, and mass distri-
bution (center-of-gravity location). The maximum expected load varies as a function
of the atmospheric turbulencc environment and the maneuver level of the vehicle.
It is the sum of the steady-state load and the predicted peak incremental value. Results
of a dynamic load analysis are rcpresentkd in terms of statistical quantities (for
example, RMS responses), PSD response, time response of the dynamic loads (for
instance, torsion, bending moment, acceleration, and both deflection and rate con-
trol surface activity), and singular value response.
Recall that spectral (PSD) analysis is used to obtain vehicle responses to random
disturbances. ~ h k r e s u l t of
s PSD analysis can then be used to estimate fatigue damage
rates, maximum expected loads, stability information, and crew compartment ac-
celeration environmcnts. In gcncral, vehicle response is constraincd by aerodynamic
and structural restrictions. Control surfacc activity is limited by actuator and hinge
momcnts. As an cxamplc, the closed-loop system RMS control surface rate must
be significantly below lld2 times the rate limit, sincc the boundary value (maxi-
mum and minimum valuc) is about ~ times the RMS valuc. Generalized harmonic
analysis is an cxtcnsion of thc spectral analysis in obtaining a more comprehensivc
vehicle's dynamic response to atmospheric turbulence. In generalized harmonic anal-
ysis, the theoretical frequency response functions are derived from the nonlincar
Sec. 3.12 Other Performance Analyses 125

flexible dynamic equations of motion. Letting the gust input be a unit impulse and
taking the Fourier transform of the flexible dynamic equations of motion yield the
frequency response equations. The response equations are then solved algebraically
to obtain complex motion frequency response functions. Summing the resulting
airload and inertial load responses allows the load and stress frequency responses to
be obtained. Consequently, the complex load and motion frequency responses are
Modem Filtering
and Estimation
As the caption of Figure 4-1 suggests, this chapter is concerned with modern filtering
and estimation techniques. As can be seen from the figure, the fundamental (purely
classical) aspects of filtering and estimation are covered extensively in Book 1 of
the series, which also includes discussions on signal reconstruction and observers.
Filtering and estimation techniques can be divided into two approaches, determin-
istic and stochastic, the latter of which is applied to systems with noisy inputs. The
deterministic approach leads to the design of the Luenberger observer, treated in
Book 1 of the series. Both deterministic and stochastic methods are related to the
classical elements of complementary filtering, also covered in Book 1 of the serics.
The bulk of the current chapter deals with optimal estimation techniques, which
make up Sections 4.1 to 4.5. These are stochastic methods including linear minimum
variance estimation, nonlinear filtering, prediction and smoothing, Kalman filtcr
analysis, and operational considerations. Finally, in Section 4.6, a combined com-
plementary1Kalman filter approach is presented to close the gap between the hcu-
ristic classical dcsign and the analytical modern estimator dcsign.
Many dynamic modeling and control applications require at some time during
operation estimates ofunknown variables in the equations and/or state of the system.
Ultimately, the values of these quantities depend on measurements and the assumcd
model structure, and on statistics related to measurements, the model, and inputs
to the system. This section develops techniques designed to produce such estimates,
Chap. 4 Modem Fllterlng and Btlmatlon Techniques

Filtering & Estimation

(Book 1)

Deterministic Stochastic

Luenberger Complementary Optimal Estimation

Observer Filtering (Kalman Filter)
( B o ~ k1) (Bopk 1) (Secs. $1-4.5)
Combined ComplementaryiKalman
Filter Approach to Estimator Desig
(Sec. 4.6)
Adaptive and Digital Applications Figure 4-1 Modern Filtering and Esti-
(Books 4 & 5 ) rnatlon Techniques

which can be applied in either a batch or a recursive mode. In a batch mode, the
estimate is determined after all of the measurements are recorded, whereas in a
recursive mode, an updated estimate is obtained after each measurement is recorded.
Recursive techniques take on special importance in the case of on-line applications
involving digital computers. Modeling of a time-varying dynamic system for which
there are uncertainties associated with inputs, outputs, or the system itself, is for-
mulated through state vectors. Modern estimation methods make use of known
relationships to compute the desired information from measured information, taking
into account measurement errors. the effects of disturbances and control actions on
the system, and prior knowledge of these variables for which the new information
is desired. Filters designed from modern estimation methods for noisy data reduce
the effects of drift with regard to the desired information. State estimates of a dy-
namic system are not confined solely to being estimates of the current state. In fact,
there are occasions where it is desirable to obtain an estimate of the state at a specified
future time. Such a problem arises in the application of intercept guidance. This and
other types of similar problems are known collectively as the prediction problem,
and estimation techniques can easily be modified to treat these. The smoothing
problem represents yet another extension of basic estimation techniques. It is desired
in this problem to determine the state at some previous time based on measurements
recorded up to the current time. Both the prediction and smoothing problems are
covered in subsequent sections.
Kalman filtering is a method for estimating the state variables of a dynamic
128 Modem Nlterlng and Estimation Techniques Chap. 4

system recursively from noise-contaminated measurements. The filter determines

the system's present state by optimally combining theoretical estimates with mea-
surement noise characteristics, based on a knowledge of the system model. In con-
trast to methods that -process measured data simultaneously and require substantial
data-storage capability, such as least-squares methods, the Kalman filter processes
measurements sesuentiallv and thus reauires much less data to be stored. Each new
measurement results in a new state estimate. In addition to the estimate itself, a
covariance matrix of estimation errors representing uncertainties in the estimate is
also generated by the Kalman filter. For linear systems, the actual, measured data
are not required to generate the covariance matrix, so that the latter can be calculated
in advance based o n different models or measurement arrangements. Consequently,
estimation errors are known before the dynamic system is actually implemented,
allowing a determination to be made of the effect of additional state variables on
estimation accuracy. This in turn permits a prediction of improvements in accuracy
that can be achieved by using additional sensors. The Apollo space program is
generally accepted as being the first known application of the Kalman filter. Since
then, Kalman filtering has proved eminently useful in such applications as the de-
- -

termination of spacecraft orbits, satellite tracking and navigation, digital image pro-
cessing, economic forecasting, industrial processes, and nuclear power-plant in-
strumentations. Some particular applications of the use of Kalman filter theory
include ascent guidance at Cape Kennedy and orbital determinations for Voyager
1 and Voyager 2. In this last problem, spacecraft-based optical observations were
combined with earth-based radiometric observations to accurately determine the
orbit during the Jupiter encounter approach phase. In the case of nuclear powcr-
plant instrumentation, on-line instrument failures are detected and fault-tolerant
computer control systems are designed, each by making use of Kalman filter theory
[Baheti, 19871.
As powerful and useful as Kalman filter theory is in so many diverse areas of
application, it still cannot circumvent unacceptable results arising from poor system
modeling and improper conditioning of the covariance equations. Modeling is pri-
marily concerned with the definition of state variables, selection of the coordinate
system to be used, the method of linearization, and the effects of bias, disturbance,
and modeling errors. Even in the same area of application, the dynamic model
developed for a particular problem may no longer be suitable if changes or new
restrictions are introduced into the problem. The following sections provide a revicw
of approaches and advances in Kalman filtering. The reader who is also intcrcsted
in a chronological account of the development of this mathematical filtering thcory
and its widespread engineering acceptance should consult Schmidt [I9811


The Kalman filter technique is summarized in this section and is shown in Figure
Sec. 4.1 Linear Minimum Variance Estimation: Kalman Filter

a) Syucm Model and Discrete Kalrnan Filter



b) Linear Continuas Kalrnan Filter

, Optimal Stale Gnimalion

Figure 4-2 (a) System Model and Discrete Kalman Filter (b) Linear Continuous Kalman

4.1.1 Discrete Kalman Nlter

Table 4-1 defines the well-known Kalman filter [Kalman, 19601. A historical treat-
ment of this subject is documented [Gelb, 19741. The Kalman filter is the optimal
state estimator that provides a systematic scheme for computing the estimator gain
matrix Kk.A flow chart is drawn in Figure 4-2a. Table 4-2 defines the multiple-
rate Kalman filter.

4.1.2 Continuous Kalman Nlter

In Gelb [I9741 and numerous other places the Kalman filter is derived for a linear
system with a continuous measurement equation given in Equations (3) of Table
2-1 by performing a limiting process involving Atk = tk+, - tk on the discrete
filter solution. The result is a natural generalization of the discrete case and only the
result will be presented here [Powers, 19781. I

1. Given the state equations (3) of Table 2-1 and the initial values E[x(to)]= 20
and E[(x(to)- io)(x(to)- 2 0 ) ~=] Po. The matrix R ( t )must exist at each t.
2. Integrate the matrix differential equation using P(to) = Po as follows
130 Modem Nltering and Estfmation Techniques Chap. 4


Step I : Problem Formulation

Given Eq. (5) of Table 2-1, no measurement is made at time to, and k(0)' is used to
represent the estimate of x at to with known covariance
P: = E[x(O) - i(0)') (x(0) - k ( ~ ) + ) ~ ]
Application of the exbectation operator to Eq. (1) produces an apriori estimate of x(1)
which is denoted by %(I)'. The associated covariance P; is obtained in a similar manner.
The notation [ ](k) - represents the value of [ ] at 1, before the measurement z(k) is
performed, while I ](k) + represents the value of [ ] at t, after the measurement. The latter is
called the apmeriori estimate. The equations for fr(k)+ and P: are functions of %(k)-PI, and..
z(k), and represent the linear unbiased minimum variance (LUMV) estimate of i(k)'. In
this recursive approach, each time a new measurement is performed, immedialely
afterwards an updated estimate of x is obtained.
Step 2: Time Update
Since E[w(k-l)]=O, then given %+(k-I),and fj , the expectation operation on Eq. (Sa) of
Table 2- 1 gives
i(k)- = Q(k-1) 8(k-1)' + Y(k-1) u(k-I) (2)
Based on the definition.
P; = E{[x(k) - i(k).] [x(k) - i(k)-lT)
The following equafon is derived:
Pi = @(k-l)~;., QT(k-1) + N(k-1) Q,., liT(k-I) (4)
Step 3: Meastrrenrerlr Update
Given $,(k)- and P: ,calculate

P; = [I - K, H(k)]PL (6)
i(kf = %(k)-+ K,[z(k) - H(k) k(ky] (7)
Step 4:
If measurements are exhausted, stop; otherwise, set k=k+l and return to step 2.
- - -

It should be noted that in many cases, Eq. (6) is replaced- by -

P; = [I - K, H(k)] P i [I - K, ~ ( k ) ] '+ K, R, K: (8)
because of its superior numerical properties. That is, Eq. (6) may result in bad subtraction,
whereas Eq. (8) involves the sum of two positive definite matrices (which will keep P:
positive definite).
Sec 4.1 Linear MInlmwn Varknce Estlmatlom Kalman Fflter Dl


S~epI : Problem Formulation

Given Eq.(5) of Table 2-1,the problem formulation is the same as that in step 1 of Table
4-1. Let the measurement vector be defined as
zf = fast measurement available every Tf sec.
= q = slow measurement available every T sec.

and zfis available 1times as often as q, i.e.. T, = Tf. Time is defined by the index (k, i),
where k refers to the basic time period T, and i refers to the short time period Tf. At time
index (k, 0) both q and z, are available giving the multi-rate measurement equation.

[-I'[=. Hp (k, 0) + v, (k, 0)

& - H,x (k, 0) t v, (k, 0)
v, (k,0) N(0, r,)
v, (k, 0) N(0, r,)
At the remainder of the time instants only q is available and the corresponding measurement
equation is

Step 2: Time Updare

i = 1 , 2 , ..., 11-1
vf (k, i) N(0, rf) -
Repeat for i = 0, 1, 2, ..., 1-1.
i(k, i+l)' = @(k, i) k(k, i) + Y(k, i) u(k, i)

Step 3: Measurement Update

For i = 0. P(k, 0)' = k(k, 0)-+ Kf[zAk, U) - H&(k, O).] + &[q(k, 0) - H,k(k, O r ]
For i = 1, 2, . -1, S(k, i)' = i(k, i)' + K,[z&k, i) - ~ $ ( k ,i)']
Step 4:
If the measurements are exhausted, stop; otherwise, set k = k + l and return to step 2.

The above results can also be applied to multiple sampling rates. Following [Lewis,
19861, let the continuous system be discretized using a period T = Ti, where Tf is much less
than the data sampling period T,,where T, = k Tt. No data arrive between kT,, so during
these intervals only the time update (step 2) should be performed. For this update, use the
system dynamics discretized with period TI. When data are received, at times kT,, the
measurement update portion (step 3 but treat K, = 0) should be performed. For this update,
the measurement noise covariance R' = R,T, is used.

and define rhe Kalman gain

~ ( t =) ~ ( t ) ~ ( t ) ~ ~ - l ( t ) (4- 1 b)
The state estimate is defined by integrating
i ( t ) = A(t)?(t) + B ( t ) u ( t )+ K ( t ) [ r ( t )- H ( t ) i ( t ) ] , $(to) = i o (4-lc)
132 Modem Filtering and Estimation Techniques Chap. q

The estimation is corrected by the residual term x - H2 through K. Figure 4-2b

presents the linear continuous Kalman filter. The Bu(t) term includes all the known
inputs to the model dynamics so that the filter can use this information to improve
its estimate. If E [ w ( t ) v ( ~ ) ~=] S(t)8(t - T) is a white-noise process, then only
Equation (4-lb) is modified to give
K(t) = [ p ( ~ ) H ( t+) ~N(t)S(t)]R- (t) (4- 1d)
In many cases the filter solution is studied in two phases: the transient phase and
the steady-state phase, with the steady-state characteristics defined by the solution
as t approaches infinity. In classical linear differential equation terminology, these
correspond exactly to the homogeneous (steady-state) and particular (transient) so-
lutions. There exists a steady-state Kalman filter gain which is the most useful for
control problem applications. Supposing a t this point that the matrices A, B, N, H,
Q, R, and S in Equations (4-1) are all constants, it can be shown that the limit of
the covariance matrix P(t) as t approaches infinity, exists, for example, P,, and
furthermore that it is defined by the algebraic Riccati equation

This last equation is nothing more than Equation (4-la) with P = 0. The resultant
Kalman filter gain is the constant matrix

and thus the steady-state estimate by Equation (4-lc) is

Given that these arc the statc equations o i thc stcady-state Kalnlan estimator, thc
stability of this last differential cquation is strictly a function of the matrix A -
K,H. A great deal of information about the bchavior of Equation (4-4) can be
obtaincd by analyzing the roots of the characteristic cquation corresponding to det[sl
- (A - K,H)] = 0. A sufficient condition for the steady-state filter estimate to
be asymptotically stable is that thc system is observable. Assuming Q 2 0 (positive
sernidcfinite) and R > 0 (positive definite'), the system is observable if and only if
H ~ i ... i ( A ~ ) " - ' H ~has
the 11 x r ~ pmatrix [ H T I A ~ i (AT)'HT ] a rank 11.

Example 4-1: First-order altitude rate Kalman filter. As in Book

1 of thc scrics, thc altitudc rate h can bc infcrrcd from two mcasurcmcnts, a bar-
omctric tncasurctncnt iln and an accclcromctcr mcasurcmcnt /;:. It is worth\vhilc
here to apply the stcady-statc Kal~nanfiltcr to thc samc problem. First, is trcntcd
as a stochastic input with known noisc charactcristic w(t), that is, ?(I) = 1 ) =
/;,(I) t u f ( t ) , whcrc ul(r) is a zero-mean whitc-noise process with variance q. Scc-
ondly, /IB is trcatcd as a nlcasurclncnt z = = 11 + v = x + v where v(r) is a
zcro-mean white-noise proccss with variancc r , and r is a constant. With rcgard to
Equations (4-1). A([) = 0, B(r) = 0, N(r) = 1 , H(r) = 1, Q = q, and R = r.
Applying the steady-state Kalman filter produces the algebraic Riccati cquation
- ~ ' l+r q = 0 whose solution is P, = d :.i The stcady-state Kalman filter gain
is K I = d& and the estimate equation is h = I;:, + g& ( i l H - i). Comparing
this solution with the complementary filter solution of Book 1 of thc series, h =

;1 + -T1 ( I.l l , - jl), shows that T = VG. If T E 12, 61, then, physically, the h B ( t )
measurement is about two to six times noisier than thc I;:, measurement, with respect
to variance [Powers, 19781.


The linear filtering problems of the previous sections are basically solved by a single
approach. In many applications, the system dynamics and the measurement models
are nonlinear. Nonlinear problems in general require more problem-oriented tcch-
niques, along with considerably more knowledge of stochastic process theory. For-
tunately, one of the most successful techniques for treating nonlinear problems
developed to date is a natural extension of the linear-oriented Kalman filter known
as the extended Kalman filter (EKF).

4.2.1 Extended Kalman Filter (Em)

The Kalman filter is a standard filter for state estimation in linear systems. For
systems with nonlinear dynamics, a natural extension is to linearize the system
around the current state estimate 2(t) and apply the Kalman filter to the resulting
linear, time-varying system. This is the EKF which has been described in detail in
Jazwinski [1970]. This filter has also been widely used for the combined (nonlinear)
problem of estimating th'e state and the parameters of a linear system. Discussions
of the latter application are presented in Book 4 of the series. The EKF is presented
in this section along with the procedure for treating the continuous dynamic model,
discrete measurement case. The continuous nonlinear system is described by Equa-
tion (1) of Table 2-1. The main idea is to use expansions of the nonlinear functions
f and h about the estimate 2 whenever linear forms are necessary in the development.
There is not a great deal of difference in the structures of the covariance and gain
computation equations for the Kalman filter and for the EKF, and these are sum-
marized as follows. The initial condition is also assumed to have a normal distri-
bution, E[x(to)] = 20, E[(x(to) - 20)(x(t0) - 20)T] = PO.
134 Modem Filtering and Estimation Techniques Chap. q

1. Define
A[i(t), t] = - H[?(t), t] = (4-5a)

2. Integrate the coupled (in general) differential equations

K[2(t), t] = p(t)HT[2(t), t ] ~ - ' ( t ) (4-5d)
The EKF yields very nearly optimal estimates if +e linearization is accurate. The
EKF involves the integration of n + [n(n + 1)/2] '(since P is symmetric) coupled
differential equations in the variables 2 and P with the initial conditions given by
Equations (4-5b and 4-5c). As mentioned previously, the idea behind the EKF is
that of linearizing the nonlinear equations about the current state estimate 2(t), a
technique which is termed relinearization. In utilizing such a procedure, the ap-
proximation remains linear because the estimated state is as often as not closer to
the actual state. The structure of the algorithm for the EKF is illustrated in Figure
4-3. Because of the linearization procedure, the EKF covariance (and hence the
gains) depends on the current state estimate and the measurement data. Conse-
quently, it is not possible to compute the EKF covariance and gain matrix as a
function of time off line as in the linear case of Section 4.1 since the state estimates
of the EKF are no longer optimal. O n the other hand, if an a priori nominal tra-
jectory, for example, x*(t) replaces 2(t) as the base trajectory in the expansion
procedure, then Equations (4-5a to 4-5d) will involve known functions of time
(for example, A[x*(t), t], H[x*(t), t ] , and so on) and the Kalman gain can be
computed a priori. However, this procedure has not met with the same success as
the EKF in actual applications. Note that the standard KF guarantees stability while
the EKF does not.

Calculation of
Eqs. (4-5a)
Initial Estimation Estimation Error Linearization of
Error Covariance -b Covariance p(t) and 4- Model About
P(t,) Kalman Gain K(t) State Estimate
State Variable Model A
of Nonlinear System K(t)
Measurement Data +Computation of State
Initial State State Estimate * +Variable
Variable Estimate Eq.(4-Sb) Estimates

Figure 4-3 Structure of an Extended Kalman Filter

Sec 4.2 Nonlinear Flltertng 135



Given Eqs. (2a&b) of Table 2-1, and E[x(to)]= 4, -

E[(x(to) $,) (~(1,) - to)T] = PO.
Step I: Defme

Step 2: Given a(!;), Pi, integrate forward to ti+,

8= qa(t),u(I), ti, = ~ i ( t11)P. + PA[$(^). 1
P 1+~Q(0. P(!,) = P(~;)
Set k = k+ 1and go to step 3.
Step 3: Given $(ti), Pi, z(k), deternine
&ti) -
= $(ti) + K,{z(k) h[!!(t;), kl)
P: = [I - KrHk[%(t;)]]P;
where = Pi ~:[ft(ti)l{Hk[a(ti)lPi H:[s(~~)]+ %I-'
Go to step 2.

Note that if to < I,, then the definitions %(t:) = ko, P; = Po are made, and step 2 is
implemented as the first step. If a measurement is given at to, i.e., I, = to, then the
definitions $(ti) = 9,. P; = Po are made and step 3 is implemented as the first step.

Several modified versions of the EKF have been proposed and implemented
to deal with nonlinearities in the model. Some of the more successful variations of
the EKF are represented by the second-order filter, the iterated EKF, and the J-
adaptive filter Uazwinski, 19701. The algorithms developed are capable of yielding
significant benefits only in certain applications and are not of a very general nature.
For this reason, the application of any of these algorithms must be carefully examined
and acccomvanied bv a substantial amount of simulation studies. The EKF filter
for the continuous-time nonlinear system with discrete-time nonlinear measure-
ments (see Equations (2a & b) of Table 2-1) is presented in Table 4-3, from which
the linear case of continuous dynamics, discrete measurements can be easily deduced.

4.2.2 Statistical Linearization Technique

A linear approximation is desired for a vector function f(x, t) of a vector random

variable x, having probability density function p(x). Following the statistical ap-
proximation technique, a statistically linearized filter is given in Table 4-4 in which
it should be observed that the structu2e is similar to the EKF in Table 4-3; however,
now K, and K h have replaced A and\ H k . The computational requirements of the
statistically linearized filter may be greater than for EKF derived by taking Taylor

series expansions of the nonlinearities because the expectations f , h[x(tk), k], and
so on, must be performed over the assumed Gaussian density for x. However, the
136 Modem Filtering and Estimation Techniques Chap. q



Given Eqs. (2aBrb) of Bble 2-1, and vx(to)] = kO, E[(x(to) - 9,) (~(t,)- 9,) T] = Po.
It is reasonable to approximate f(x,t) by the linear expression
f(x, t) m a(t) + x(t) (1)
where a(t) and K,(t) are a vector and a matrix to be determined. Defining the error
e f(x, I) - a(t) - %(I) x(t) (2)
it is desired to choose a(t) and K, such that the quantity
J = ~ [We] e ~ (3)
is minimized for some symmetric positive semi-definite matrix W. Substituting Eq. (2)
into Eq.(3)and setting the partial derivative of J with respect to the elements of a(t) and K,
to zero, gives
a({) = ?(XI - yk, K, = {[fxT],,, - ?kT} P-I
To derive approximate minimum variance filtering algorithms, Eq. (4) is substituted into
Eq. (1) yielding
f(x, 1) m kx) + &(x - 8). h[x(tk). k] h[k(tk). k] + I$,(k)[~(t,) 9(ti)l- (5)
This formulation necessitates computing f, hi, &, and Kh, each of which depends on the
probability density function for x, a quantity that is generally not readily available.
Consequently, an approximation is needed for p(x) that permits the above quantities to be
calculated. For this purpose, it is frequently assumed that x is Gaussian. Since the
probability density function for a Gaussian random variable is completely defined by its
mean k and its covariance matrix P, both of which are already computed in any filtering
algorithm, it is possible to compute all the averages in Eqs. (5). Assuming that the right
sides of Eqs. (5) are calculated by making the Gaussian approximation for x, the statistical
linearization for f and h can be used for the nonlinear filter. The step by step procedure of
the statistically linearized filter is summarized as follows:
Srep 1: Definitions
[fxT],,, ?, Et are expectations calculated assuming x -N($, P). k, k(ti) and [h xT(t;)],,, are
expectations calculated assuming ~ ( t k ) N [$(ti), pi].
Step 2: Given kcti), P;, integrate forward to ti+,
Et = REt(t), u(t), I], P = K$P + PC+ Q I ) , P(1,) = PO;)
&(I)= {[fxqC,,- ?kT} p-l(r)
Set k = k+l and no to step 3.
Step 3: Given a(&). Pi, a k ) , detlrmine
%(ti) -
= k(tk) + Kk { ~ ( k )h[k(tk), k]}
= [I $K,,(k)] PC
where Kk= Pi K L ( ~() q ( k ) Pi ~ l ( k+) RI,)-'

From [Gclb. 19801 with prrmission from The Analytic Scie~lccsCorp.

Sec. 4.2 Nonlinear Mitering W7

performance advantages offered by statistical linearization may make the additional

computation worthwhile. A detailed derivation can be found in Gelb [1974].

4.2.3 Multiple Model Estimation

A multiple model approach considered here is to assume two or more levels of

process noise. A filter is set up for each model and, based on their likelihood func-
tions, the probability for each model being correct is obtained [Bar-Shalom and
Luh, 19861. Let Ae ( e = 1, . . . , m) bc the event that model e is correct with prior
probabilities P{Ae} = pe. The likelihood function of measurements up to time k
under the assumption of model Ae is Le(k) = P ( z k 1 Ae) = n P [ v e ( i ) ]where


= { z ( l )... z ( k ) }and, under the Gaussian assumption, the probability density func-
tion of the innovation from model e is

Using Bayesian rule, the a posteriori probability that model E is correct at time k

PIAe I z k } = P { A ~ } P { ( zAe}
~ P{A,}P{zk I A,} = L e ( k ) p e
j=l j= 1

It is assumed in Equation (4-6) that the same model has been in effect from the
initial time. However, this is obviously not true if the model changes at some time
during the interval [ I , k ] . In view of this, a "sliding-window" likelihood function
has to be used. It is now assumed that the model changes at time k - N, so that
one considers only the likelihood functions of the measurements from k - N + 1
to k. Each filter starts from the same initial condition at k - N . The likelihood
function corresponding to model e is then

Le(k - N , k) = P [ z ( k - N + I), . . . , ~ ( kI Ae,

) ~ ( -k N ) , . . . z ( l ) I
probability of model
and t at time k: P{Ae I z k }
138 Modem N1terin.q and Estimation Techniques Chap. q

The state estimate is a weighted average of the model-conditioned estimates with

the preceding probabilities as weights:

E{x(k)I zk} = E { x ( k ) 1 A,, z k } P { A j 1 z k ) (4-8)


A "fading memory" approach to the computation of probabilities in a multiple

model situation is now considered. The discounted (fading-memory) average norm
of innovations with discount factor 0 5 a-,5 1 ( j = 1, . . . , m ) is
P.;(K) = - 1) + ~ ; r ( k ) s ; l ( k ) ~ , ( k=) 2 a ~ - C ~ ~ ( t ) ~ . ; l ( e ) (4-9)

where v j ( k ) is the innovation from the filter based on model j. The "effective"
memory length is (1 - a j ) - I . Let S j ( k ) = S, be the steady-state covariance of the
innovations from filter j. Then, it is possible to replace Equation (4-7) with
-- I
1 1
P_I ( k ) = ~ , e - ' ~ " ~P;(O)
' Cce-1 ,P<(k)
Pe(O), C , = I S j I ?(I - a,) (4- 10)

where Pc ( 0 )are the prior model probabilities. If these prior probabilities are equated
to the posterior probabilties, then Equation (4-10) simplifies to Pi =
c,~-$"~) cce -7,:')
- A typical multiple model estimation is drawn in Figure
4-4 in which the combination of the estimates is done according to Equation


The methods of the previous sections are concerned with the problem of estimating
the state of a system at time t based on the measurements from time io to i. This
problem is called the filtering problem. In this section, formulae for the predicted
state, that is, ? ( t ) based on nleasurements fro111to to I.,- < I , and the smoothed state,
that is, 2 ( t ) based on measurements from to to I,,., t E (to, tn) are presented. Since
prediction and s t ~ ~ o o t h i ninvolvc
g state estimates at times othcr than the measure-
ment time, the follo\ving notation is employed in this section.
2(f I t ~ ) expected
: value of x ( t ) based on measurements from to t o t N
P(t I I,.): covariance matrix at time t based on measurements from to to I~

Ssc. 4.3 Predlctlon and Smoothing



of Estimates -
$ p

ICalculation of
Probabilities I
Figure 4-4 Multiple Model Estimation

In a discrete-time system, Equation (4-11) is rewritten as

.+(k I N):cxpected value of x(k) based on measurements from k = 0 to k = N
P(k I h'):covariance matrix at time k based on measurements from k = 0 to k = N
(4- 12)

I t should be noted that many intermediate and advanced books on state estimation
employ this notation for all developments including the filtering problem. Equations
(4-11) are simply the estimates at t conditioned under the measurements up to t , ~ ,
and thus are called the conditional expectations of the state and covariance.

4.3.1 Prediction

Actually the prediction problem has already been presented as part of the filtering
problem. Optimal prediction can be thought of in terms of optimal filtering in the
absence of measurements, that is, R = 0, and hence the filter gain is zero. In par-
ticular, for the continuous dynamics, discrete measurement case, the state and co-
variance propagation equations between measurements represent the best estimates
of the state and covariance. A similar situation is true for the discrete dynamics case,
and thus the prediction equations are as follows:

Linear, discrete Kalman filter-Equations (2) and (4) of Table 4-1

i(kIN)=@(k-l)i(k-lIhq+Y(k-l)u(k-1) fork>N
~ ( ( khq = @(k - i ) ~ ( k- 1 I iv)aT(k > I ) +~ ( -k i ) ~ ~ - , -
~ 1)
~ ( k
(4-1 3)
Linear, continuous Kalman filter-Equations (4-la and 4-lc) with R-' = 0
140 Modem Filtering and Estimation Technlques Chap. q

and K = 0, that is
B(r 1 t , ) = A(t)i(t 1 tN) + B(t)~i(t) for t > tN (4- 14)
~ ( 1 t,t) = A(t)P(t 1 t,) + P(r 1 tN)A T(t) + N ( t ) ~ ( t ) ~ ~ ( t )
Nonlinear, continuous EKF-Equations (4-5b and 4-5c) with R-' = O and
i(t It ~ =) f[t, I ( t I tN), i*(t)] for t > t~ (4- 15)
P 1t ) = A[t, q t ( tN)lP(t ( t ~ +) P(t 1 t ~ ) A ~ [ t , 1 t ~ ) 1+ Q(t)

With R-' = O implying infinite variance, that is, no measurement or worthless

measurement, K = O gives no weighting to a measurement aspect in propagating 2.

4.3.2 Smoothing

The primary advantage of Kalman's solution is that the equations that specify the
optimum filter are recursive in nature, so that they can be adapted easily to the
digital computer. However, Kalman filtering does not consider the important prob-
lem of smoothing. AGNC [1986(f)] presents a simple and elegant derivation of the
principal equations for smoothing. The derivation is based on the method of max-
imum likelihood and depends primarily on the simple manipulation of the proba-
bility density functions. The resulting equations of the linear discrete smoother are
summarized in Table 4-Sa, while the linear continuous smoother can be derived
and is summarized in Table 4-5b. Finally, the stnoother of this section is called a
fixed-interval smoother since the interval [0, N] is fixed and 2(k 1 is obtained
from k E [0, S ] .T w o other types of smoothers are the fixed point (fixing k and
determining 2(k 1 N) as N increases) and the fixed lag (given h r = constant, de-
termine 2(k - A1 I k) as k increases). Equations for each of these smoothers for
continuous and discrete time are developed in Gelb [1974].


The design of a well-tuned Kalman filter is achieved only after successfully iterating
a process that combines system modeling with performance analysis and trade-off
studies. While system modeling is based mainly on physical insight with respect to
the problem a t hand, performance capabilities arc often subject to computational
constraints and requirements imposed by the actual hardware involved. Therefore,
although Kalman filter theory embodies a good deal of mathematical formalism,
much effort is still spent arriving at an acceptable design.
When the Kalrnan filter state estimates arc used as the state feedback, that is,
when the Kalman filter is inserted into the control loop, a precise design method-
ology may be as follows [Maybeck, 19791. The first step consists of developillg a
Sec. 4 . 4 Kalman Filter Design and Performance Analysls 141


a) Linear Discrete Smoother

Given Eq. (5) of Table 2-1,
Step 1: Apply the linear discrete time Kalman filter (Table 4-1) for
k E[O, N], and denote % (k I k) = he)+, P(k I k) = P;
P(k I k-1) = P; and %(k1 k-1) = %(ky
Step 2: Smoother State Estimate
%(kI N) = %(k1 k) + Ck[4(k+l I N) - k(k+l 1 k)]
where Ck = P(k I k) oT(k) P(k+ll k).' and k(N I N) = &N)+
Step 3: Error Covorinnce Matrix Propagation
P(k I N) = P(k I k) + C,[P(k+ll N) - P(k+ll k)] C:
where P(NIN)=P;
b) Linear Continuous Smoother
Given Eq. (3) of Table 2-1 and k(to) = %,, P(to) = Po
Step I:
Apply the linear continuous Kalman filter (Sec. 4.1.2) for t E [to, i], and denote the
resultant state and covariance by iF(t), PF(t).
Step 2:
The smoothed state and covariance at t E [to, i] are denoted by %(tI i) and P(t I i), and are
determined by integrating backward the coupled differential equations from i to I,
d -1 = - P ~ A - A
-P T -1 1
P ,-P;NQN T -1
P, + HT R -1 H, p B 6 j1 = O
dt B
T -1
~ = - ( A ~ + P ~ N Q N ~ ) ~ + Px, ~yB( i ~) =+o H R
and then substituting the resultant y(t), pB(tjl into:
p(t1ij1 = pB(tjl + p,(tj1
i(t Ii) = &,(I) + ~ (1 i)t [y(t) ~ i ( tQt)]

dynamic model that ideally resembles system behavior and validating the model
with experimental data. While it may not be of any benefit to develop an overly
complex physically based model, it is absolutely necessary to develop one which
lends itself well to Kalman filter design. The second, and perhaps most significant,
step is to design and tune the filter against performance requirements, comparing
estimation errors with the specifications and taking into account computational con-
straints. The complexity of the filter may be influenced by the definition of state
variables and the choice of coordinate system for certain applications. In the next
step, an attempt is made to combine and remove states where possible, and to delete
142 Modem Filtering and Estimation Techniques Chap. q

weak cross couplings, in order to obtain reduced-order filters. At this stage, possible
approximations to the optimal gain are also evaluated. Following this, a covariance
sensitivity analysis and Monte Carlo simulations-are conducted, after which it is
then possible to select a final design based o n a trade-off between the performance
and the computational requirements. Finally, the last step consists of performing a
checkout, final tuning, and an operational test of the filter. Further details of Kalman
state feedback design are presented in Book 3 of the series.

4.4.1 Kalman Filter Design Process

This section highlights in more detail the second step listed in the design procedure
outlined previously. Regarding Section 4.1.2, to reiterate, the design parameters
that can be selected to design the Kalman filter are z(r), w ( t ) (some of which may
be fictitious), Q, and R. The estimator dynamics gain and bandwidth are completely
determined by the measurement signal selected and the designer's specification of
Q and R for the design. Q and R are design parameters selected by the designer to
create the filter gain K, via the Riccati equation (4-la), and thus determine the
dynamics characteristics of the estimator. Actual physical process noise and mea-
surement noise will in general differ from these design values. For good estimator
performance, these specified noise intensities should roughly correspond to distur-
bances or uncertainties that will actually be encountered, since the estimator will
be tuned to these conditions. Sometimes, inexact inte~lsitiesare provided to the
Riccati equation in ordcr to achieve certain desirable filter characteristics, but the
engineer should be aware of the costs and benefits. Thc error covariancc P i s useful
in terms of certain limiting cases where constraints are established by the open-loop
model dynamics itself to bc placed on the effectiveness of the cstimator. Some very
rough Kalman filter design rules of thumb are given as follows. For small Q and
large R , the design tends to be such that the filter poles are close to the model
dynamic poles, the filter is a low-bandwidth filter, and there is slow state estimation
and good measurement noise rejection. For large Q and small R , the design tcnds
to be such that the filter poles approach the nlodel dynamic transmission zeros or
stable reflections thcreof (some actually move to infinity in the left half-plane), the
filter is a high-bandwidth filter, and there is fast state estimation and poor mea-
surement noise rejection. Some useful transfer functions for design follow.
From I A ) (to ~ ) ~ ( r )H(s1
: - A ) - I .Y (4- 16)
From x(r) to i ( r ) : ($1 - ('4 - K H ) ]- ' KH (4- 1 7)
From z ( t ) and rr(t) to i ( f )[ s:l - (A - KHJ- 'IBK] (4-1 8)

4.4.2 Selection of Noise Intensity Matrices

For given Q and R, specificd for a Kalman filtcr design, some propertics of the filter
follow. 111 general, the magnitude of estimator gain matrix K is proportional to the
magnitude of the Q and R - ' matrices. Multiplying Q and R by the same positive
Sec. 4.4 Kalman Filter Design and Performance Analysls 143

scalar k yields the same estimator as Q and R. The initial guess of Q and R can
easily be made using thc following rule of thumb. Since K is lowest for Q R - ' =
0 and increases as Q R - ' increases, the estimator break frequency is lowest for Q R - I
= 0 and increases as Q R - ' increases. The estimator bandwidth is high for low
measurement noisc and low for high measurement noisc. It is assumcd that Q =
diag(ql, qz, . . . q,,) and R = diag(rl, rz. . . . , r,,,). A nominal process
- with
noisc intensities 71, . . . , q,, is applied, and the response intensity ?', . . . , ?,, that
comes through at the sensors due to the nominal process noisc Q (as the nominal
sensor response intensity) is recorded. Then R = diag(arl, a?,, . . . , a?,,,) is chosen
as the measurement noise for the filter design wherc a is a scalar that will be allowed
to vary, and will serve as the scanning parameter for the filter design.

Q9X-l- 0. In the case of stable, open-loop model dynamics, as R goes to

infinity or as Q approaches zero, analysis shows that the estimator operates open
loop. Moreover, P will be zero, K approaches zero, and the Kalman filter poles
move to the positions of the open-loop poles.

Example 4-2: Stable, first-order filter. The model dynamics cqua-

tions are $(I) = ax(t) + b ~ r ( t )+ w(t), E[w(t)] = 0, E[IU(~)LV(T)] = q8(t - T), z(t)
= hx(t) + ~ ( t )E[v(t)]
, ) ] r8(t - T), whcre a < 0, q 2 0, and
= 0, and E [ v ( ~ ) v ( T =
r > 0. The model dynamics transfer function is G(s) = hl(s - a). The Riccati
equation as specified by Equation (4-la) is h'P?ru - 2aP - q = 0 and yields
P = r[a + q a ' + h2(qlr)]/ll'. The Kalman filter gain is k = [a +
a + h'(q/r)]/h and the state estimate equation is

The preceding equations s h o ~ vthat the estimator dynamics are a function of the
ratio qlr, and not of either q or r alone. T h e root locus of the Kalman filter is shown
in Figure 4-5a and is discussed here. Using Equation (4-17). the transfer function
from x(t) to 2(t) is
k(3) - kh k(0) kh
- where the DC gain is --- - (4-20)
X(r) s - (a - kh) X(0) -a+kh
The magnitude of x(s)/x(~) as a function of frequency is plotted for increasing
values of qlr in Figure 4-5b. For a given value of r, increasing the value of q in the
design results in increased values of the gain k. From Equation (4-20), this means
that X(O)/X(O)will tend to unity. Also, for a constant v, increasing q will increase
the estimator bandwidth, while for a constant q, decreasing r will have the same
effect. N o w , if the input u is zero, and w is a constant but unknown disturbance,
then at steady-state conditions the model dynamics equations imply that x = w l a
= constant, while, from the state estimate equation, i= 0 and

2 = (khx + kv)l(-a + kh) (4-21)

Modem Filtering and Estimation Tecftnlques Chap. 4


b\ a


qh- qir = 0
k- = k =0
-a a
Model ~ ~ n a m iPole
Figure 4-5 Selection of Noise Intensity Parameters q and r

Assuming n o measurement noisc, that is, v = 0, and using Equations (4-20), it is

true that
2 = kkhxl(-a + kh) = X(O)X/X(O) = k ( O ) u ) / [ ~ ( O ) a ]
As noted previously from Equation (4-20). specifying a large q means that ~ ( 0 ) l
X(0) will be close to unity, and that x wrill be close to 2 at stcady state. Under thcsc
conditions, the state estimate is a good disturbancc cstin~atc.If, on the other hand,
w = 0 but v is a constant but unknown disturbancc (for example, a sensor bias),
then Equation (4-21) at stcady state becomes 2 = kvl( - a + kh). When q > 0 and
r > 0 , then Equation (4-19) bccorncs

2(t) = -da2 + h2(q/r) :(I) + b11(t) + la + qa2 + k2(qlr)]~(t)lh

and thc transfer function fro111 the measurement inputs z ( t ) to the state estimatc
Sec. 4.4 Kaiman Filter Design and Performance Analysls 145

=)[a + g a 2
i ( t ) , given by Equation (4-18) is x ( s ) / z ( s + h2(q/r)]l[h(s +
V a * + }t*(q/r))].T w o limiting cases are now considered.
Case I : q -r 0. This case corresponds to zero process noise, implying no
uncertainty about the model dynamic statc, including x(O), exists. For this case, P
= 0, k = 0, and the statc estimate equation is ;(I) = o.?(t) + bu(r) and i ( 0 ) =
x ( 0 ) . The filter therefore generates open loop. For a constant step input at steady
state, 2 = x = hula = constant.
Case 2: r -r0. In this case, the filter bandwidth, or break frequency, ap-
proaches infinity, k also approaches infinity, but P approaches zero. Also, as shown
in Figure 4-5c, x(s)/z(s) = Ilh.

Unstable open-loopmodel dynamics. When the open-loop model dy-

namics are unstable, the estimator gains are nonzero, and the error covariance will
not go to zero even when there is little or no process noise. In addition, the estimator
poles move toward the positions of the stable poles which are the left half-plane
reflections (about thejw-axis) of the unstable open-loop poles. This is termed the
minimum measurement effort estimator, and sets a lower limit on thc estimator
gains. Even in thc Lvorst case, therefore, the Kalman filter always guarantees c l o d -
loop system stability by providing the minimum measurement effort estimator as
long as the model dynamics are completely observable. However, in classical design
approaches, when the model dynamics have right half-plane poles, the closed-loop
system is unstable for low filter gains.

Example 4-3: Unstable first-order filter. The open-loop model dy-

namics for this example are the same as the previous example, except that now a
> 0. Similarly, the model dynamics transfer function, the Kalman filter gain, and
the state estimate equation are also the same, respectively. Finally, the root locus
of the Kalman filter is sholvn in Figure 4-5d in which the estimator locus begins
at the stable reflection of the open-loop model dynamics pole. If q > 0 and r > 0,
then not only is the state estimate equation identical to that in the previous example,
but A ( s ) / ~ ( s is
) also identical. Again, it is possible to consider two limiting cases.
Case 1: q + 0 . Here, the variance and filter gain are nonzero and are given
by P = 2aih"and k = 2 a / h , and the filter equation becomes i ( t ) = a$(t) + bu(t)
+ 2az(t)lh.
Case 2: r -+ 0. The results are the same as those in the previous example.

Example 4-4: Neutrally stable first-orderfilter. In this example, the

open-loop model dynamics are again the same as those in Example 4-2, but this
time a = 0. Setting a = 0 in Equation (3) of Example 4-2 also gives the model
dynamics transfer function, while the state estimate equation becomes i ( t ) = bu(t)
+ k(z(t) - h?(t)) and k = g&. Hence, as in the previous two examples, the
estimator dynamics are a function of the ratio q l r . The important point to highlight
146 Modem Filtering and Estimation Techniques Chap. 4

in this example, however, is that, if q = 0, then k = 0 and i ( t ) = bu(t), and the

estimator is neutrally stable, which is unacceptable. If q > 0, the estimator is stable
with a pole at - kh. Therefore, if the open-loop model dynamics have a pole on
the jw-axis, it must be disturbed with a process noise in order to have a stable
estimator. For a zero input, u = 0, and a constant initial state x(O), at steady state,
x = x(0). From the state estimate equation, 2 = x + v l h . If v = 0, then 2 = x
= k(0).

QR-'-, m. In this case, all measurement covariances R go to zero (that is,

ideal sensors with excellent measurements) or the process covariance Q goes to
infinity. As R goes to zero or as Q approaches infinity, the estimator gains increase
drastically, and the bandwidth of the resulting Kalman filter increases very rapidly.
When all the transmission zeros in the model dynamics are in the left half-plane,
that is, minimum phase, the closed-loop poles asymptotically approach the trans-
mission zeros, and the residues of the slow model which is created by the poles
approaching zero have small values in 2 ( t ) = x ( t ) - 2(f). However, the residues
of these modes in K[z(t) - H<(t)] are not small. If the number of sensorslmea-
surements is greater than or equal to the number of disturbances, the error covariance
P approaches zero under the condition of ideal sensors or as Q approaches infinity.
However, in both minimum and nonminimum phase systems, if there are more
disturbances than measurements, the minimum value P can achieve for any corn-
bination of Q and R is always strictly positive. This is true even if the product QR-'
becomes infinite. Therefore, even in the event of having all ideal sensors, perfect
estimation cannot be achieved as long as there are not enough sensors.
As shown in Examples 4-2 to 4-4, as qlr approaches infinity, the estimator
poles move toward the transfer function zeros located at - x (see Figures 4-ja and
4-5b). Nonminimum phase systems impose dcfinitc restrictions on the accuracy
which the filter can achieve. When the trans~nissionzeros of the model dynamics
are in the right half-plane (no~~minimum phasc), as the measurcrnents approach
perfection the cstimator has high gains, and the closed-loop Kalman filter eigenvalucs
move towards the positions of the stable poles which are the left half-plane reflections
(about thejw-axis) of the unstable transmission zeros of G(s). Thus, unstable trans-
mission zeros are not desirable because they lead to an inefficient estimator. The
residues of these modes remain large as the estimator gains increase for the closcd-
loop system. Even if the number of measurcrnents is greatcr than or equal to the
numbcr of disturbances, the resulting error covariance docs not approach zero and
is much larger than that for minimum phasc systems. Optin~alestimation always
guarantces closcd-loop system stability evcn in thc evcnt of high cstimator gains.
Howcver, in classical design, for a nonminimuln phase systcn~,thc closed-loop
system is unstablc for high controller gains.

4.4.3 Error Analysis

,Examining thc state cstimatc equations (4-I), the structural csscnce of thc Kaln1a11
filter is that which is also used in the general case. Also, the definition for K ( t ) is
that which defines the optimal gain, and thus the optimal filter, and is shown to bc
Sec 4.4 Kalman Filter Design and Performance Anaiysls 147

a particular case of the general filter gain matrix developed in this section. For the
case of a general filter gain matrix, the state estimate error is defined as f(t) = ?(I)
- x(t) and E[.f(r)] = 0, which yields the error dynamics of the estimator

The linear error dynamic system is therefore being driven by white noise, as shown
in Figure 4-6a. As shown in Figure 4-6b, the Kalman filter is thought of as a
regulator loop in error estimate .i. coordinates. The equation defining the error co-
variance for the filter with a general filter gain matrix K(t) is given by

If the eigenvalues of A - K H are in the left half-plane, the estimator is stable and,
as r approaches infinity, the state estimate i ( t ) converges to the actual state x(t) and
the estimator error 2(r) approaches zero. The design procedure consists of selecting
the filter gain matrix K(t) that places the estimator eigenvalues in the desired lo-
cations. A wide bandwidth gives fast rate estimates but lets measurement noise
corrupt the estimates. Using the Kalman filter gain Equation (4-lb), the error dy-
namics Equation (4-22) is a stable linear system driven by white noise. Substituting
in Equation (4-23) the Kalman filter gain gives the same equation for Pas in Equation
(4-la). As t becomes large, the error covariance P(t) in Equation (4-23) satisfies
P(;) = 0 and the steady-state covariance matrix is obtained by solving the linear
Lyapunov matrix equation,

The Kalman filter gain minimizes P(t). That is, P ( t ) from Equation (4-24) using a


b I 'f+@ I

I------- _I Figure 4-6 Error Dynamic Model for

Kalman Filter
148 Mdern Filtering and Estimation Techniques Chap. q

general filter gain matrix is always greater than P(t) from Equation (4-2) using
the Kalman filter gain Equation (4--3).

minimizing minimizing trace

(by Kalman gain) P = + u:, + +
u s + ... u:,,

For the particular case in which all the states are measured, H = I and R = diag(rl,
r2, . . . , r,), and the Kalman filter gain then becomes

K = PH~R-'= (4-25)

The ith component of the state estimate is

2 2
+ ... + -
((t) - 2,))
+ ... + - @,,(I) - i t + ...
rj r,,
If the noise intensity ri is large, the gains uZ,+;lrj will be small, and the signal zi(t)
will not be weighted heavily in computing 2i(t). If, however, the error covariance
crzjr, is large, the gains ~ : , ~ , l will
r , be large, and the signals will be weighted heavily
in computing i , ( t ) .


Problem-dependent judgments must often be made to enhance filter performailce

(for example, reduce the transient time in the estimation procedure at the expense
of a greater variance), reduce computational requirements, furnish missing infor-
mation (for example, the Kaln~anfilter needs complete a priori covariance infor-
mation on the initial state, process noise, and il1easuremcnt noise), and/or establish
procedures for stabilizing diverging filter estimates. In the Kalman filter in~plc-
mentation, the computation of the transition matrix together with the propagation
of the error covariance matrix represent a substantial fraction of the overall COIII-
putation effort necessary. Fortunately, because the accuracy of these quantities nccd
not be as great as that required by the state vector, the former may be computed
at a slower rate than the state vector update. In numerous applications, essential
Set 4.5 Operational Considerations 149

statistics information concerning the process and measurement noise, in the form
of standard deviations and correlation time constants, may cither be missing alto-
gether or be poorly defined. Moreover, the analytical formulation ofthis information
rarely reflects the behavior of the actual nlcasurements, resulting in reliable data for
an extended period of time followed by a series of poor data points. Filter operation
is further affected by modeling errors, linearization approximations, and the re-
duced-state nature of practical filters. Because of all this, the covariance matrix
computed from the Riccati equation of the Kalman filter may not resemble the true
covariance matrix of errors in the cstin~atedstate. For this reason, many simulations
need to be performed to adequately examine filter performance [Baheti, 19871.

4.5.1 Filter Performance

The Kalman filter is capable of solving a certain optimization problem (minimum

variance), which is an indicator of filter accuracy. For numerous problems, however,
additional considerations in the filter design (for example, settling time to a steady-
state value, sensitivity to parameter variations) may be more significant than de-
termining the estimation with the least variance. For instance, it may be necessary
to obtain a constant gain filter that has the minimum settling time subject to an
upper-bound constraint on the trace of the covariance. Figure 4-7a shows hypo-
thetical state estimates and variances for the minimum variance and minimum set-
tling time filters [Powers, 19781. It is also possible to design a filter that minimizes
a function of both the variance and the sensitivity of the filter to modifications in
a system parameter. Figure 4-7b illustrates a hypothetical case where the Kalman
filter is the minimum variance filter for a = a" (that is, the design value); however,
it is not necessarily the minimum variance filter for off-design values. In this par-
ticular case, the minimum sensitivity filter, being less sensitive to parameter changes,
would probably be the best filter for an application in which a was known to vary
over the range [a,, az]. In defense of the ~ a l m a nor minimum variance filter, it is
possible to compute the Kalman gain without iteration. In the two problems noted
previously (minimum settling time and sensitivity), iterative parameter optimization
schemes are required to determine the filter gains; thus, they would be employed
only when the Kalman filter is not adequate for the given application. An appropriate
reduction of the number of states modeled will cause a suboptimal filter to become
a noticeably smaller computer. Figure 4-7c illustrates that this suboptimal filter may
be employed to reduce the sensitivity to uncertain parameters.

4.5.2 Computational Requirements

Kalman filtering fundamentally represents a digital computer-oriented approach to

state estimation. Thus, the analysis of a Kalman filter would consider all problems
inherent to the digital computer.
150 Modem FIlterlng and Estimation Techniques Chap. q

Minimum Variance Estimate

Minimum Settling Time Estimate

Trace P(t)

For Minimum Settling Time Filter

For Minimum Variance Filter

Trace P

Figure 4-7 (a) Minimum Variance vs. Minimum Settling Time Filter Performance
(b) Covariance as a Function o f Parameter Variation for Minimum Sensitivity and
Variance Filters (c) Conceptual Example of Minimum Sensitivity Filter
I and sensitivity trade-off)
I I n>m>P


C Expected Range of Uncertainty

Figure 4-7 (Continued)

Square-root filter. The reliability of time-varying Kalman filters has been

significantly improved by the so-called square-root or (uDu') algorithms [Bier-
man, 1977; Kaminski et al., 19711. They propagate the square root of the symmetric
covariance matrix instead of the covariance matrix; this provides almost twice the
accuracy for a certain computer word-length as Kalman's original algorithm. In
addition, this process also prevents numerical truncation from generating a covar-
iance matrix that has negative eigenvalues [Bryson, 19851. Filtering problems can
result from the finite word length of the digital computer. For instance, after nu-
merous measurements have been processed, round-off error build-up can produce
theoretically impossible filter behavior. In other words, the Kalman filter is enhanced
assuming infinite precision, and accumulated round-off errors can cause the covar-
iance matrix update to result in a nonpositive, semidefinite matrix. The main way
of dealing with this problem is to utilize the square-root filter. This filter was initially
developed [Potter, 19641 to rectify Kalman filter implementation problems in the
Apollo program. The computer employed in the Apollo spacecraft program was a
fixed-point digital computer.
The square-root filter involves a reformulation of the propagation and update
equations in terms of a matrix S , where P = S S T (that is, S represents the square
root of the covariance matrix). Utilizing S instead of P causes the precision of the
computation to essentially double. In addition, it also prevents P from being non-
positive and semidefinite. Unfortunately, the fact that the matrix S is not unique
results in a proliferation of square-root algorithms [Andrews, 1968; Bellantoni and
152 Modem Filtering and Estimation Techniques Chap. 4

Dodge, 1967; Kaminski et al., 39711. The adverse effect of employing the square-
root filter is that it requires more computer time. However, recent developments
with more efficient implementation have substantially decreased this effect. Bierman
[I9771 discusses these efficiencies and the computer program listing of the Jet Pro-
pulsion Laboratory orbit determination program square-root filter.

Suboptimal filter. Another computer-oriented problem that relates to the

Kalman filter is the potentially numerous differential equations that must be inte-
grated (that is, state and covariance differential equations). Because the major portion
of computer burden in Kalman filtering is typically represented by the solution of
the error covariance equations, an effective way to make the filter satisfy computer
hardware limitations is to precompute the error covariance, and thus, the filter gain.
As a result, the suboptimal filters for a simplified system model are usually imple-
mented instead of the optimal Kalman filter. Suboptimal filters may deal with fewer
state variables (for example, states that produce negligible effects in the Kalman filter
are ignored), fewer covariance equations, and in many cases, simplified filter gains
which approximate the optimal Kalman gains. T o avoid on-line covariance calcu-
lations, the error covariance matrix can be computed at an earlier stage and the filtcr
gain histories are obtained. While a prerecorded set of precise gain histories could
be stored and utilized, a more effective method is to approximate the optimal gain
behavior by analytical functions of time, which can be easily computed in real time.
A piecewise constant suboptimal gain and a decaying exponential function gain can
be employed. Steady-state gains referring to the limiting case of a set of pieccwise
constant gains can also be employed by choosing each gain to be constant over all
time. A Kalman filter that utilizes steady-state gains is idcntical to a Wiener filtcr.
In numerous practical systcms, the constant-gain Wiener filtcr offers estimating
accuracy that is equivalent to that of the more sophisticated Kalman filter, but a t
one third of the computational cost [Powers, 19781.

Kalman filter design software. A parameter estimation subroutine

package for the construction of efficient multipurpose estimation algorithms has
becn developed [Bicrnlan and Nead, 19781. This software was developed under
NASA contract of thc Jet Propulsion Laboratory, and has been applied to several
Kalman filter designs and simulations. More rcccnt works [Pappas et al., 1980;
Verhaegcn and Van Dooren, 19861 have rcsultcd in significant improvements in the
computational reliability obtainable in the nu111cricalsolution of the Riccati equation.
The algorithms arc applicable to problems with singular transition matrices \vhich
arisc in thc case of sampling a continuous-time systcm with time dclays. A softwarc
package has also becn devcloped [Arnold and Laub, 19831 that solves the gcncraIizcd
algcbraic Iiiccati cquation. Many research organizations have already dcvelopcd an
interactive Kalman filtcr analysis and design package. By making the softwarc user
friendly and easily adaptable, providing a broad range of graphics capabilitics, 2nd
supplying numerically cfficicnt algorithms, thcy have produccd a rathcr elegant 2nd
powerful tool for the dcsigncr's use. Such tools will invariably si~nplifythe Kalmnn
Sec. 4.5 Operational Considerations 153

filter dcsign proccss in the future and may at the same time substantially improve
performance (Bahcti, 19871.

4.5.3 Absence of A Priori Information

Much a priori information, that is, .?(I)), P, Q(r), and R ( r ) , is required by the Kalman
filter. If this information is not available, then straightforward least-squares tech-
niques (recursive or batch) should be applied. In many cases, the Kal~nanfilter
approach can still be elnployed successfully even though only a portion of the re-
quired a priori statistics is known. This occurs because of several reasons. First, the
order of magnitude estimates or relative magnitudes of the statistics can be guessed,
and then a Kalman filter can be established. Often, the resultant filter performs
adequately. Second, if the first reason given prcviously produces undesirable per-
formance, then an analysis of the simulations should indicate which missing a priori
statistics are most critical. An acceptable filter design can be achieved through a
parameter study involving the critical statistical parameters. Finally, in numerous
cases, .?,, and Po are the absent values. In some systems, these parameters mainly
affect initially stable, transient effects and have little impact on the long-term op-
eration of the filter. Note that the steady-state Kalman gain is independent of ko
and Po. Another technique of dealing with absent statistics is to develop an adaptive
filter (that is, the statistics are generated as the proccss proceeds by analyzing dif-
ferences between the model predictions and the measurements) [Powers, 19781.
Adaptive filter techniques can be found in several textbooks. Details of adaptive
filters are presented in Book 4 of the series. An identification problem occurs when
@(k) and q ( k ) in a discrete case, or A and B in a continuous case, are unknown.
This problem is the topic of Book 4 of the series. If process noise w ( k ) in a discrete
case, or w ( t ) in a continuous case, is known, then the on-line identification of the
process covariance Q h in a discrete case, o r Q ( t ) in a continuous case, is required.
A problem will occur if the initial value is not known and thus, P is unknown.

4.5.4 Filter Divergence

T w o sources of error which may cause the measured performance of a Kalman filter
to differ from what was analytically predicted are modeling errors and computer
round-off errors. As a result of these, errors in the estimates may become intolerably
large at some point during the operation of the filter. Filter divergence is often
explained in terms of the calculated covariance matrix, the assertion being that this
matrix becomes unrealistically small and results in placing unreasonable confidence
in the estimates. Consequently, the filter gain is reduced and subsequent measure-
ments are effectively ignored. In the study of Kalman filters, divergence refers to
the situation where the estimated filter error converges to a value less than the true
filter error. In some cases of divergence, the true error converges to a finite value;
this is known as an apparent divergence. This is typically caused by errors in the
mathematical model of the process (for example, neglected forces which generate
154 Modem Filtering and Estimation Techniques Chap. q

a bounded-state response). The resultant filter may perform satisfactorily for the
particular application, with the only deficiency being the estimate of the error by
the Kalman filter approach [Powers, 19781. True divergence represents the most
critical type of divergence in filter design. This type of divergence means that the
true error becomes arbitrarily large as time increases while the estimated error con-
verges to a finite value. True divergence is usually the result of unmodeled effects
which can cause an unbounded-state variable response, or in some cases, the type
of instability presented in Section 4.1.2.
" .
Several methods. of which the use of larger vrocess noise covariance and es-
ponential data weighting have been particularly successful, have been applied t o the

problem of avoiding filter divergence and improving the estimation accuracy. The
increased noise covariance offsets modeling errors and improves filter stability, while
exponentially weighting the measurements prohibits obsolete data from saturating
the filter and sets a lower bound on the filter gain [Baheti, 19871. A primary factor
in the onset of filter divergence that surfaced after many Kalman filter implemen-
tations is computer round-off errors in the covariance propagation equations. To
maintain acceptable numerical accuracy, these equations generally require double-
precision computations. As shown in Section 4.5.2, algorithms have been developed
based on matrix square-root propagation, which are capable of producing an ef-
fective precision that is double that of the conventional Kalman filter. A procedure
similar to the square-root formulation and called U-D covariance factorization, has
also been developed and tested in many applications for consistency and stability.
Studies emphasizing number of operations, actual C P U time, and data storage re-
quirements have demonstrated that optimally coded U-D factorization and the con-
ventional Kalman filter implementations are very similar. However, for processing
large sets of data with infrequent estimates, an even more efficient algorithm has
been developed called a square-root information filter [Bierman, 19771.
Most of the procedures that deal with true divergence are conservative,
straightforward methods that increase stability, and thus, degrade the performance.
Although one could obviously attempt to estimate unmodeled states, the resultant
filter size would most likely be This is especially significant in practical
applications where the process models are not always completely validated. In-
creasing the process noise in the model (for example, making Q(t) a positive, definite
matrix) represents a simplified way to approach the divergence problem. This ap-
proach implies that one does not know the model exactly, and will not allow the
state estimate propagation equation to predict 2(t) with absolute certainty (that is,
P(f) = 0). As noted in previous examples, if Q(r) = 0, then the Kalman gain may
approach zero and thus not use any additional measurements in the state propagation
equation. The effect of Q(t) > O causes P(r) > 0. In othcr words, a nonzero lower
bound implies a nonzcro Kalman gain. An increase in the error estimate and poor
performance will be thc effects of an arbitrary selection of Q(t) > 0. More sophis-
ticated techniques that produce less pcrformancc loss are discussed in Book 4 of
series. The use of finite memory filtering represents an alternate solution for pre-
venting divergence.
Kalman filter divergence and suboptimal design. If the eigenvalues
of Equation (4-24) arc in the right half-plane, the estimator is unstable and, as t
becomes large, the state estimate i ( fdiverges
) from the actual state x ( t ) and the
estimate error 2 approaches infinity. One of the principal causes of this failure is
that the systcm modcl contains states or models that are undisturbed by the model
process noise (modcl uncertainty). One cure for such problems is the periodic re-
starting of a time-varying Kalman filtcr. Other cures include minimum variance
observers with eigenvalues constraints, added noise, pole shifting, and destabili-
zation. Thc constrained minimum variance problem may be stated as follows [Bry-
son, 19781. Find a set of observer gains K that minimize J = trace (2hlP) where P
is found from Equation (4-23). ,\.I is a positive semidefinite matrix, and the gains
are C ( K ) > 0. This constraint ensures that the eigenvalues ( s i ) of A - K H always
lie in the left half-plane of the original point at - a , where u is a specified positive
number, that is, R,(si) 5 -0. The nccessary condition for minimizing J with con-
straints Equation (4-23) and C ( K ) > 0 is

The following example of estimating an undisturbed constant state demonstrates

this technique. The simplest model of a Kalman filter for a neutrally stable plant is
given in Example 4-4 where b = 0, q = 0, and h = 1 have been assumed here.
The initial state x ( 0 ) has zero-mean value and variance Po. The Kalman filter for
estimating x ( t ) is = k ( z - i )where k ( t ) = ( P o / r ) l [ l (Polr)t]instead of zero.
Clearly for t (rlPo), k ( t ) -+ 0 , and the estimator ignores the measurement z. In
practice, such a filter will almost certainly diverge for (Polr)t % 1 with drift errors.
From C ( K ) > 0, k = u where a is the eigenvalue constraint level. The variance 06
the estimator is predicted to be P ( t ) = Poe-2"' + ar(1 - e-2"')/2 so that P ap-
proaches ur12 in steady state, but the estimator will not diverge.
Schmidt [I9701 presents a method for modifying unsatisfactory Kalman filters
by introducing additional noise into the system model Equation (4-la). T h e spectral
densities of added noise may be varied until the filter has the desired stability. The
reason that this technique works can be seen by examining the steady-state error
covariance matrix equation in Equation (4-2). If certain elements of Q are zero,
corresponding elements in the steady-state value of P, and consequently K , are zero.
However, if these elements of Q are assumed to be nonzero, then the corresponding
elements of K will be nonzero, and the filter will always try to track the true system.
The choice of the appropriate level of the elements of Q is mainly heuristic, and
depends largely on what is known about the unmodeled states.
Breza and Bryson [I9741 present a pole-shifting method in which the first step
is to design the steady-state Kalman filter. If the stability is unsatisfactory, the slow
eigenvalues are shifted to specified, faster eigenvalues by changing the gains so as
to leave the other eigenvalues unchanged. The method does not give a unique so-
lution if there are two or more measurements.
156 Modem Filtering and Estimation Techniques Chap. 4

4.5.5 Modeling Process Noise and Biases

Fundamental to Kalman filter theory is the use of white Gaussian state and mea-
surement noise models. In actual practice, random inputs are likely to be time cor-
related and represented by first-order or second-order Gauss-Markov process.
Moreover, the dynamical system may be subject to certain parameters whose values
are not well known. Augmenting the original state vector to include the exponen-
tially correlated process noise and biases is a method commonly employed, but this
can greatly increase the amount of computations required for many applications
[Baheti, 19871. An algorithm for partitioning the estimation of the bias variables,
or unknown constants, from the dynamic variables was published by Friedland
[1969]. One advantage realized by the use of this algorithm is reduced data storage
requirements in the computer. A second advantage of the partitioning is that it
creates additional analytical insight to the filter behavior based on special properties
of the variables. The following reviews the basic theory behind separate bias esti-
mation. It begins by considering a discrete-time process
x(k + 1 ) = A ( k ) x ( k ) + B(k)b + w ( k ) and z ( k ) = H ( k ) x ( k ) + C ( k ) b + v ( k )
where w ( k ) and v ( k ) are white Gaussian noise processes and b is a constant but
unknown bias vector. The optimum estimates ? ( k ) and b^(k)are given by ? ( k ) =
Z(k) + ~ 6 ( k where
-x ( k ) is the ) Z ( k ) is computed as if no bias errors were present. That is,
optimal "bias-free" estimate, which is then corrected to account for the
effect of the bias. The matrix V represents the ratio of the cross covariance matrix
of the bias-free state cstimate and the covariance matrix of the bias. Such a decom-
position becomes very practical in the case of a large number of bias parameters.
The dynamical modcl parameters, noise statistics, and initial estimates of a particular
system to which the Kalman filter is applied generally do not agrce with the real
system that generates the observations. In the case of using a model of fewer di-
mensions than the original model to reduce the number of computations, such
differences are imposed deliberately. The price of using a reduced-order model may,
however, be degraded filter performance. Algorithms for analyzing the mean-square
error performance when the system model is not the model used in the filter dcsign
have been developed Uazwinski, 19701. The error analysis generated is particularly
useful for conducting sensitivity studies. That is, one can study the effect of some
known variation about a norllinal value in a ccrtain system paramctcr on the filter
pcrformancc, if the filter is dcsigncd using the nominal valuc.
Timc-corrclatcd noisc is occasionally rcfcrred to as colored noise, as opposed
to white noise. In thc discrctc modcl case, it is callcd sequentially correlated noise.
The utilization of this type of information should cnhancc filtcr performance. Fol-
lowing Powers [1978],procedures for dealing with timc-correlated noise with dif-
fcrential equation models arc discusscd here. Suppose that the measurcmerlt r~oise
is white, but the proccss noisc is colored with thc corresponding discrctc-time niodcl
where wl ( t ) represents zero-mean white noise with variance G, and @ ( t o ) and
E{[w(t,,) - @ ( t o ) ] [w(to) - @(to)lT}are assumed known. If the random process
[Equation (4-26)) is a Gauss-Markov process, then it can always be expressed by
a linear dynamic system with white noise forcing function only

where li, is zero-mean white noise with variance Q. Equation (4-27) and the mea-
surement equation can then be treated by all of the results developed in the previous
sections. Note that Q - is never required in these equations, which allows the colored-
process noise case to be treated by simply defining the augmented state vector of
Equation (4-27). The colored-measurement noise case is more complicated because it
invariably leads to a singular measurement covariance matrix, which is not allowable
in the standard Kalman filter equations since they involve R - ' . Supposing
.G = A x + Nw, Z = Hx + v, b = Fv + Gvl (4-28)
where w and v l are uncorrelated zero-mean white noise with variance Q and R ,
respectively, and Equation (4-28) defines the correlation properties of the mea-
surement noise v, in which E[v(to)]= 0 and ~ [ v ( t ~ ) v ( t =o )Vo.
~ ] Define

which implies, upon differentiation of z ,

i(t) = (H + H A - FH)x + H N w + Go,, i t ) = x + I7 (4-30)
where H = H H A - FH, I7 = HNw +
C v l , and E[i;] = H h l E [ w ] + G E [ v l ]
= 0, E[i;ijT] = H N E [ w w T ] N T H T G E [ v l v T ] G T = [ H N Q N ~ H + ~GRGT]r
x 8(t - 7 ) = ~ 8 ( -t 7 ) . Here, use has been made of the fact that E [ w ( ~ ) v I ( T ) ~ ]
= 0. It should be noted that in the new problem variables ( x , i ) , the process and
measurement white noises w and 17 are now correlated as

However, this case is easily handled by employing Equation (4-ld) instead of Equa-
tion (4-lb) with S(t) = Q ( t ) N ( t ) T ~ ( t )Thus

As shown in Bryson and H o [1975],one need not perform the differentiation of

the measurement signal z in Equation (4-29) by introducing

x* 9 i ( t ) - K ( t ) z ( t )J i* . (4-32)
= (A - KH)(X* + K z ) - ( K + KF)z, xF = i0- K(to)z(to)
158 Modem filtering and Estlmation Techniques Chap. 4

which would then replace Equations (4-31) in the filter calculations. Finally, to
determine 2(t) and P(t) from Equations (4-31), care must be taken with respect to
the initial conditions for the integration. Note that the derivation of the Kalman
filter assumes that 2o and S(to) are not conditioned upon any other information.
However, in the x, 2 system, ,?(to) is conditioned upon knowledge of z(to) and
4(to), that is, Equation (4-29). Thus, for a proper application of the Kalman filter
equations, i ( t ) and P(to) should also be conditioned upon this information. One
can view this as the given io and Po corresponding to 2(tc) and P ( t f ) , respectively,
with a measurement 2(to)at to. Thus, the initial conditions for integrating Equations
(4-31) are
i ( t o ' ) = 20 + K(t0)[4(to)- ~ ( t o ) i ~ ] ,P ( t 2 ) = [ I - K(to)H(to)]Po
where K(to) is defined by Equation (4-31) with P = Po.



This section presents, derives, and verifies the continuous form of first-, second-,
and third-order complementary/Kalman filters. As discussed in Book 1 of the series,
conventional complementary filter architectures have been used with the filter gains
derived from Kalman filter techniques. These low-order filters are shown in various
NGC applications in Book 1 of the series. It was noted previously that comple-
mentary filters are really simple Kalman filters, but this section further highlights
this fact, especially after a thorough comparison of complementary and Kalman

4.6.1 First-Order Complementary/Kalman Filter

The generalized continuous mechanization forms of the first-order complementary/

Kalman filter are illustrated in Figure 4-8. A comparison of the complementary
filter and the Kalman filter is also shown in this figure. As in the standard first-
order filters, rate limiting is achieved by inserting a limit function after the Kalman
gain correction block. The limit function ensures that the rate of change of the
innovation sequence v = z - i is limited. When the rate of change of v is less than
the imposed limit, the filter functions identically to the linear first-order filter. Ho\v-
ever, when the rate of change of 11 is equal to or exceeds the limit, the changc of
the estimated output 4 must be due to a pure integration of the input signal M.
Consider the model for the first-order dynamic system in which x = position and
i = rate, and the equation of motion in statc-space form is

where w and v arc uncorrclatcd and zero means with variances q and r, respectively
An illustration of the preceding model dynamics is shown in Figure 4-8a ~vhicll
Sec. 4.6 Fllter Approach to Estimator Deslgn
Complementa~~lKalman 159

Kalman Filter

+Rate Limit

Complementary Filter



k, t bu, 4
Z= x t v = xl+bu1+v z = xtv. v = high frequency noise

$U dt y = xtn, n = low frequency noise

k=fi lh
Figure 4-8' First-Order Complementary/Kalman Filter

includes the optimal state estimator. With regard to Kalman filter formulation,

and the state estimate equation is & I = u k [ z - +

- b u l l where k is the Kalman
filter gain. The estimate outputs are 2 = 2 , + bul. Substituting Equation (4-36)
into Equation (4-18) yields
A , ( s ) / ~ ( s=) l / ( s + k), X I (S)IUI
= - bkl(s + k), .k,(s)/Z(s) = k l ( s + k)

The characteristics e uation is \s + k = 0.The algebraic Riccati equation is -P2/

r + q = 0 or P = &r and hence the Kalman gain is k = q r . The Kalman filter
gain k is the inverse of tkcomplementary time constant and the Kalman filter
provides more insight into the filter gain characteristics.
Case 1: 40~itiot1 bias. k
the actual measurement z from Equation (4-
33) also contains a bias error xb beside the measurement noise v , that is, z = x -
160 Modem Filtering and Estimation Techniques Chap. q

xb + v , then the first approcch in implementation is that

the bias error xb must be
removed by adding a ul = $ x b correction after the i~tegrationand b = k . The
second approach is to set b = 0 and treat xb as an extra input equal to Au = kx,
to be added to the input u .

Example 4-5: Estimation of center-of-gravity (CG) lateral accel.

eration. As shown in Book 1 of the'series, the inertial lateral acceleration is
measured at the IRU, which is a t a distance L , from the center of gravity (CG), as
opposed to being measured at the actual CG location. Consequently, A, measured
at the IRU contains an error due to the yaw acceleration R, that is, xb = L , R / ~ .
Hence x = AYcC and z = A ,,,"
= x - xb + v . The first approach is to set b =
k and ul = $ xb = Rt = LxRIg. The acceleration rate measurement jc = A,,, is
not available, and hence u = 0. The second approach is to set b = 0 and u = kxb
(requiring R to be available). Therefore, the first approach is selected because it is

Example 4-6: Estimation of CG n o d a l acceleration. Similar to

the estimation of the CG lateral acceleration, from Book 1 of the series
x = Azcc, UI = J x =
~ L,QIg,
z=Azlnu=x-xh+lv, and u = O

Exampie 4-7: Turbulencelgust velocity estimator. Let z = Vl =

inertial acceleration, u l = VTAs= air data velocity, b = - k , and u = 0. Substituting
this equation into Equation (4-36) and using the definition VTAs = V, - VC yiclds
k ks k . ks sk
- Z(s) - -U , ( s ) = -v, - -VTAS= -

Z(s) = V c =
s+k s+k s+k s+k s+k
VTASand k are operated on by the filter to produce a gust rate estimate Vc. That
is, as can be seen from the preceding equation, the rate of change of the inertial
component of VTASexactly cancels V I , resulting in the production of the afore-
mentioned gust rate estimate.

Example 4-8: Turbulencelgust angle-of-attackrate estimate. Let

z = & I = inertial angle of attack rate, u , = a* = air mass angle of attack, 11 =
k , and u = 0. Substituting this equation into Equation (4-36) and using a* = ar
+ ac yields

a~ and a, are operated on by the filter to produce a turbulence/gust a rate estimate

&. The rate of change of the inertial component of a A exactly cancels the
resulting in the production of the aforementioned turbulencelgust a rate esti~nntc,
Sec 4.6 ComplementarylKalman Filter Approach to Estimator Desfgn 161

Example 4-9: Altitude rate estimator. The problem of estimating

altitude ratc is fornlulated as z = l>k = s k k , r r l = All + I/c;ycsls = hls, rr = 0,
and b = k. Applying Equation (4-36) yiclds

This equation verifies that the problem formulation is corrcct. If thcrc cxists an
acceleration bias A h in z , then z = hlk + Alak = 1;ik + sAj11k. Substituting the
new z into Equation (4-36) yiclds

As mentioned previously, the acceleration bias Al;?k can be implemented in the

position bias term by adding A hlk to u r as ul = hls + A hlk. Substituting thc new
u l into Equation (4-36) also yields Equation (4-38). This verifies that an alternative
implementation can be done.

Case 2: Rate bias, b = 1. If the rate input contains a bias error jCb, that
is, u = j: + i b , then the rate bias error jCb can be removed by a position correction
input u l = xh after the integrator as in the position bias case. Since the rate input
u has a unit forward loop gain, u l should also have a unit forward loop gain, that
is, b = 1.

Example 4-10: Mrst-ordera estimator using inertial a rate at IRU.

In this case, the C G normal acceleration estimate is not used in the inertial a rate
computation, and the problem is formulated as follows: x = a = angle of attack,
z = x + v = a~ = a1 + a~ = measured or derived air mass a, a r = inertial a,
X D = ac = turbulencelgust a, and u = &I,,L, = inertial a rate at IRU. Also,
aIIRU contains an error due to pitch acceleration, Q, from the inertial a rate at the
actual C G location iurC,, as discussed in Book 1 of the series. The correct terms are
&rcc = + dlb and & b = LxQ/LTAs One way to obtain arc, information at
the CG would be to add the error term & b to iYI, , and then integrate the sum of
these two terms. Unfortunately, Q, and hence & b , are not available. A better scheme
is to integrate &I,R,, and then add a pitch rate correction term u1 = a b = L,QI
VrAS with b = 1 after the integrator to obtain accurate arcG.

Example 4-11: First-order sideslip estimator using inertial side-

slip rate at IRU. The problem is formulated the same manner as in Example
4-10 as follows: x = P '= sideslip, z = x + v = PA = PI + PC = measured or
derived air mass P, pr = inertial P, x~ = PC = turbulence/gust P, u = pl,,, =
inertial sideslip rate at IRU, U I = P b = LxRIVTAs, and b = 1.

Case 3: NO bias, b = 01. In this case, b = 0. Hence, u l = 0 and the

estimate is 2 = u + k(z - 2 ) . If there exists a disturbance x D in the x measurement,
162 Modem Filtering and Estimation Techniques Chap. q

then z ( t ) = x ( t ) x D ( t ) and the rate input is u = i.hn the Laplace domain, Ul(s)
= 0, U ( s ) = s X ( s ) , and Z(s) = X ( s ) +I X D ( S ) .Substituting the preceding result
into Equation (4-36),

Example 4-12: Rate estimator. The problem is formulated such that

x = x = rate, z = x + v = measured/der,ived X , u = ji = measuredlderived x,
ul = 0, and 6 = 0. The estimate equation is k ( t ) = ~ ( t +) k ( X ( t ) - ~ ( t ) )Applying
Equation (4-36), ~ ( s )= [ X ( s ) + k ~ ( s ) ] l (+s k ) = ~ ( s )This
. verifies that the
formulation is correct. The first-order altitude rate estimator in Example 4-1 can
be solved by this formulation with x = h , and x = ha. The estimate equation is
the same as in Example 4-1.

Example 4-13: First-order a estimator using inertial a rate at CG.

Following Example 4-10, if the CG normal acceleration estimator is used in the
inertial a rate computation, then b = 0, 1 4 1 = 0, and ir = &I,, = inertial rate at
CG. Here, & tracks mainly the air mass a.4 a t low frequencies so that the comple-
mentary a rate A & = k ( a A - &) is zero a t low frequencies. Applying Equation
(4-39) yields &(s) = a l ( s ) + k a c ( s ) l ( s + k ) . Thus, a~ and &I,, are operated on
the filter to produce an a that contains pure c r l plus filtered &c.

Example 4-lk First-order sideslip estimator using inertial side.

slip rate at CG. Following Example 4-1 1, if the C G lateral acceleration csti-
mator is used in the inertial sidesiip rate computation, then 6 = 0, i d l = 0, and
PI,, = inertial sideslip rate at CG. fi tracks mainly the air mass sideslip P A a t low
frequencies so that complcn~entarysideslip rate A @ = k ( P A - p)
is zero a t low
frequency. Applying Equation (4-39) yields fi(s) = Pl(s) + kPc(s)l(s + k ) . Thus
P A and OI,, are operated on the filter to produce a fi that contains pure inertial
sideslip P I plus filtered fit.

Example 4-15: First-order velocity estimator. Thc problem is for-

mulated as follows:
x = I~T.AS = air data vclocity, x ~ = Vc = turbulence/gust vclocity
2 = x + v = V I - V c = air data velocity measurement, V I = inertial velocity

u = V I = inertial acceleration, ul = 0

) c l ( s ) - K V C ( S ) I ( Sf k ) . Thus, V r ~ s
Applying Equation (4-39) yields ~ T A S ( S =
and 01are operated on the filter to produce a velocity estimate 6 ' ~ that ~ s contains
pure V I plus filtered 6 ' ~ .
Sec. 4.6 ComplementarylKalman Nlter Approach to Estimator Design la3

Example 4-16: First-order altitude estimator. The problc~nis for-

mulated as follows:
x = It = altitudc, u = /I = altitude ratc via INS and FCS
z = ItR = x + v = barometric altitudc, 11, = 0
The resulting altitude estimator is = it + k(k - I;). The altitudc estimate is very
important in INS and FCS altitudc tracking.

Example 4-17: First-order Euler attitude estimator. The problem

is formulated as follows:
x = 4 = attitude, tr = r/r = attitude ratc transformed from body rate
z = IJJ, = INS attitude, rr, = 0
The resulting Euler attitude estimator is shown in Figure 4-9.

4.6.2 Second-Order ComplementarylKalman Filter

Figure 4-10 shows a generalized second-order complcmentarytKalman filter. Con-

sider the model for the second-order dynamic system. Applications for this filter
can be classified into three major categories:

(Formtrlntiorr 1) Position and Rate Estimators with Position and Acceleration

Measurements: Given measuredlderived position x and measuredtderived ac-
celeration 2 , it is desired to obtain the best estimate of the position and rate.
(Forrrrrrlntion 2) Position and Rate Estimators with Position and Rate Mea-
surements: Given measuredlderived position x and measuredlderived velocity
k, it is desired to obtain the best estimate of the position and rate.
(Formulation 3) Position and Rate Estimators with Rate and Acceleration Mea-
surements: Given rneasuredlderived rate k and measuredtderived acceleration
x, it is desired to obtain the best estimate of the position and rate.

Second-order complementarylKalman filter: formulation 1. This

formulation is used to design position and rate estimators with position and accel-

Roll Rate
Pitch Rate
Yaw Rate
- Body to
Euler Axis

Figure 4-9 Euler Attitude Estimator

164 Modern Filtering and Estimation Techniques Chap. 4

Kalman a l t e r

Complementary Filter
High Pass

4 =Fl
-nT n-v

Low Pass

+ 2 j wn s +a);;
G(s) Low Pass

el + bu2 L
z = r + v = n l + b%+v z = s+v. v = high frequency noise

JJu &dl + Jul d y = x+n, n = low frequency noise

- -

k/kI 2j!w,, - T = complementary filter time constant

Figure 4-10 Second-Order ComplementaryiKalman Filter

eration measurements. Consider the model for the second-order dynamics system
in which x l = x = position, xz = i = rate, and 14 = x = acceleration. T h e equation
o f motion in state-space form is i l ( t ) = ~ ~ ( and 1 ) i , ( t ) = . ( I ) , with position x as
an output, and z ( t ) = x, ( I ) . If the acceleration measurement ii contains a bias crror
A x , that is, ii = x + A x , then the acceleration bias error A x can be removed by a
rate correction input u l = A i after the rate integrator so that i l ( t ) = x 2 ( t ) + I I I ( ~ )
+ w l ( t ) where w l is a zero-mean white process with variance q l . ril can be a rate
correction equal to the rate difference between the actual rate measurement and the
sensor rate measurement. The measured/derived acceleration ji is treated as a control
input u with known noise characteristics such that i 2 ( 1 ) = U ( I ) +
w Z ( t ) ,wherc w z
Sec. 4.6 ComplementaryIKaiman Filter Approach to Estimator Deslgn 165

is a zero-mean white process with variance q 2 . The preceding model dynamics are
illustrated in Figure 4 - 1 0 which includes the optimal state estimator. A comparison
of a second-order complementary filter and Kalman filter is also shown in the figure.
As in the first-order filtcr formulation, the measured/dcrived position x is treated
as a measurement z with bias error bu2 and a measurement noise v , and z = xl +
bir? + v , where v is a zero-mean white process with variance r. With regard to
Kalman filter formulation,

where K is the Kalman gain matrix, and the state estimate equations are
il(t) = + u l ( t ) + k [ z ( t ) - i l ( t ) - bu2(t)]
22(t) (4-40)
i2(t) = ~ ( t+
) k I [ z ( t ) - . t l ( t ) - 6112(t)]

where i = .? = .GI + bit2 = position estimate, .t2(()= rate estimate, and i ( t ) =

.??(I) + irl(t) = calibrated rate. Applying Equation ( 4 - 1 8 ) yields

where gains k and k l have been expressed in terms of the second-order parameters.
i(s) = 2(s) = kl(s) + bU2(s) (4-42a)
= [U(S)+ s U ~ ( S +) b s 2 U r ( s ) + ( k s + k J z ( s ) ] l D ( ~ )
k ( s ) = k2(s) + UI(S) (4-42b)
= [(s + k ) U ( s ) + (s2 + k s ) U l ( s ) - bkIsU2(s) + s k I Z ( s ) ] l D ( ~ )
All the transfer functions from either the control inputs u , u l , and 242, or the mea-
surement z to the state estimates are low-pass filters. The estimator gains k and kr
166 Modem Filtering and Estimation Techniques Chap. q

can be related to the estimator poles sl and s2 by the following formulae:

The location of the estimator poles sl and s2 are chosen based on applications and
are different for different applications. The transfer functions from the measured
z ( t ) t o k and 4 are simply the last column of Equation (4-41a),

k ( s ) = (ks + kl)Z(s)lD(s), ~ ( s =) klsZ(s)lD(s) (4-44)

The steady-state values for the elements of P, and K , are found to be

Case (a). If ql = q2 = 0, then in the steady state the elements of the error
covariance matrices all vanish, that is, P, -+ 0, and the Kalman filter gain vanishes.
Since the perfect model and the filter need only to find the missing initial condition
of the state, the processing being observed is nonstationary and the estimate con-
verges on the true value of the states which is purely the double integral of the illput

Case ( b ) . If q, = 0, qz # 0, then in the steady state

k = ~ ' Z ( ~ ~ l r ) ' / k1
~ ,= (4-46)
Using Equation (4-41 b), w,, = 6= (q2/r)114and 5 = -
- 0.707. Thus, the
damping ratio is a constant and the undamped natural frequency is determined to
be a function of only the ratio of errorlnoise in the rneasuredlderived acceleration
x to the measuredlderived position x . Furthermore, w,, depends only on the relative
noise between q2 and r, not on the absolute noise levels.
Case (c). If q 2 = 0 and ql # 0, then in the steady state k = and kl
= 0. From Equations (4-40) k1 = i 2+ U , + k [ -~ 2 , - bu2] and i2= u. Since
qz = 0, a perfect accelerometer model is used and the filter needs only to find the
missing initial condition on the rate state, the process being observed is nonsta-
tionary, and the rate converges on the true value of the rate states that are
Sec. 4.6 Complementaryllblman Filter Approach to Btlmator Deslgn 167

the integral of the accelerometer input. The position estimate is the first-order com-
filter position estitr~ate.
If the Kalman gain K is assumed zero where the forcing function in the filter
design is not considered, the covariance equations (4-23) are employed here with
K = 0 since they hold for an arbitrary equation K-matrix.

The error covariance will grow according to Equation (4-47) cvcn if there is only
a small amount of white noise of spectral dcnsity q l forcing the .ul state or a small
amount of white noise of spectral dcnsity q- forcing the xz state.

Example 4-18: Navigation filtering for position and velocity es-

timate. The navigation filtering for position and velocity estimate presented in
Chapter 7 falls cxactly into the previous category, with the following definitions:
x = X, Y, Z position; z = X, Y, Z position measurements;
i= A, Y , z velocity
I! = X , V , z acceleration;
til = A X , A Y , A Z velocity corrcction; u2 = 0
Following Example 4-1 and defining x l = h = altitude and x2 = h = altitude
rate, this example can then be applied to the design of an altitude and altitude rate
estiftlator where z = ha.. = h + v = X I + v = Z vosition .. measurements. The
acceleration bias error sail is assumed to exist so that h = I;, s h h + w = u +
j t r l + w , and the state equation is obtained as $1 = x? + ul and i z = u + w.

Thus, this is the same problem as that in Case (a) wherein q l = 0. q, = q, and r
= r. The Kalman gain is given by Equation ( 4 - 4 3 , and the state estimate eqyations

are given by Equation (4-40) where i = i1 = h = altitude estimate and h = 22

= altitude rate estimate. The calibrated altitude rate estimate is h , = 22 + u l . The

transfer function from z ( t ) = h R ( t ) to h^ is h^lhR = ( k s + (a
k,)lD(s) =
+ V & ) l ( s 2 + V'2Vqlr s + d&)and it is a low-pass filter. In the Laplace domain,

from Equation (4-42) the altitude estimate and the altitude rate estimate in terms
of u = hlRS, u 1 = A h , and z = h are h^(s) = ( U ( s ) + s U l ( s ) f (o: 2<w,s)Z(s)]/
D ( s ) = [ h ( s ) s h h ( s ) + (w: + 2<w,,s)h(s)]lD(s)and i(j) = [ ( s + 2<w,,)h(s) +
(s2 + ~ < w , , s ) A / L (+x )SW: h ( s ) ] / D ( s ) The
. problem of estimating gear altitude and
altitude rate is formulated as follows:
x = gear altitude, z = gear altitude measurement,
u = inertial altitude acceleration measurement,
u l = inertial altitude rate correction, u2 = 0

Example 4-19: Position error and rate error estimator with PO-
sition error and acceleration measurements. Trajectory errors in terms
168 Modem ~iiteringand Estimation Techniques Chap. 4

of position error and rate error are important parameters for NGC and hence need
to be estimated. The problem of estimating trajectory error is formulated as follows:
x = AX = trajectory position error,
z = x +v = trajectory position error measurement
u = x = accelerometer measurement
ul = A X = trajectory rate correction

Applying EAquation(4-42) yields A ~ ( S )= [ ~ ( s +) SAX(S) + (ks + kl)AX(s)]/

D(s) and AX(^) = [(s + k ) ~ ( s + +
) (s2 ks) AX(S)+ sk~AX(s)]/D(s).In estimating
the glideslope deviation, the problem is formulated as AX = A h = altitude error,
A X = ~h = altitude rate error, and x = h = vertical acceleration. In estimating
the localizer deviation, the problem is formulated as AX = A y = lateral position
error, A X = A j = lateral velocity error, and x = j = lateral acceleration.

Second-order complementaryIKalman filter: Formulation 2. This

formulation is used to design position and rate estimators with position and rate
measurements. The following are defined:
x = position (same as before), z = x +v = measured position
u(t) = Ax(t) = acceleration bias error (if it exists) instead of x in Formulation
ul(t) = i ( t ) = rate measurement (regarded as a control input) instead of A i
in Formulation 1
xl(r) = position (same as before), x2(t) = A i = rate measurcment
disturbance instead of i in Formulation 1
2 , ( t ) = position estimate (same as before), .G2(t) = rate measurement dis-
turbance estimate
i(t) = ?(t) = + bu*(t) = position estimate (same as before)
i(t) = i2(t) + u l ( f ) = rate estimate (same as before)
The state and measurcment equations are the same as those in Formulation 1 , that
is, i 1 = x2 + U I + U ' , . i 2 = u + ul2, and z = x, buz + +
o. The results of the
second-order filter (Formulation 1) can all be applied here with a slight revision as
follows. If there exists a disturbance X D in the x measurement, then z(t) = >-(I) +
. the Laplace domain, U(s) = 0, Ul(s) = sX(s), and Z(s) = X(s) + X D ( ~ ) .
x ~ ( t )In
Applying Equation (4-42) yields
Sec. 4.6 CompiernentaryIKalman Niter Approech to Estimator Design 169

Thus, z ( r ) and r r , ( 1 ) arc operated on by the filter to produce: ( 1 ) a position estimate

; (2) a rate estimate
i ( t ) that consists of pure position .u(t) and a filtcrcd i I > ( t )arid
; ( I ) that consists of a pure ratc i ( r ) and a filtered io(r).
The preceding results assume
the acceleration bias rr = A x to be zero. If it is nonzero, then a third state needs to
be modeled. Therefore, a third-order complcmcntaryIKaln~anfilter such as the one
presented in Section 4.6.3 is required.

Example 4-20: Second-order a estimator. This example follows the

first-order a estimator problem formulation in Example 4-10.
Case 1: Using inertial a rate at ZRU. In this case, u l = k = &I = &, =
inertial a ratc at IRU, b = 1, and u = 0. A means of removing offsets in the
accelerometer used to compute the inertial a rate is to bias the inertial a rate by the
integral of the error between a.+and ti. This also removes computational errors.
As in Example 4-10, a better scheme of correcting ah is to integrate &I,,,, and then
add a pitch rate correction u 2 = a b = L,Ql VTAS after the position integrator (see
Figure 4-10) to obtain an accurate a at C G information. The estimates are given
bv & = i t + 6rr2 and & = .?? + 1 4 , . Note that & has the same formulation as that
i n the first-order estimator, while & is not provided by the first-order estimator but
is provided here by this second-order estimator.
Case 2: Using inertial a rate at C G . In this case, u l = jr = Cul = &I,, =
inertial a ratc at C G , 6 = 0, tr = 0, and 1.12 = 0.The solution of this case is simpler
than that in previous Case I . Applying Equation (4-48) yields
& ( s ) = ads) + (ks + kl)ac(s)/D(s), &(s) = &l,,(s) + klsaG(s)lD(s)
Thus, a.4 and &l,:,; are operated on by the filter to produce: (1) an & that consists

of a pure a 1and a filtered tic; and (2) an & that consists of a pure &I,, and a filtered
Example 4-21: Second-order sideslip estimator. This example fol-
lows the first-order sideslip estimator problem formulation in Example 4-11.
Case 1: Using inertial sideslip rate at IRU. In this case, U I = 2 = PI,,, =
inertial sideslip rate a t IRU, b = 1 , u = 0 , and u2 = Pb 7 LxRIVTAS.The state
estimate is given by = 2 1 + bu2 = sideslip estimate and P = i2 ul = sideslip
p +
rate estimate. *Note that p for the same formulation as that in the first-order esti-
mator, while 6 is not provided by the first-order estimator but is provided here by
this second-order estimator.
Case 2: Using inertial sideslip rate of C G . In this case, u l = jr = P I =
Dl,, = inertial p rate at CG, b = 0, u = -0, and u 2 = 0. The solution of this case
is simpler than that ip previous Case 1. The state estimates are now P = 21 =
sideslip estimate and 6 = 322 +
u1 = sideslip rate estimate. Applying Equation (4-
40) yields
fi = ?2 + 0lcC + k(PA -~fi), i 2 = k l ( P ~- Cj)
170 Modem Filtering and Estimatton Techniques Chap. g

Applying Equations (4-42) and using @I($) = s P ~ ( s and

) PA(s) = PI(s) + PC(S)
= [X@I(~)+ ( k s + k r ) P ~ ( s ) ] f D ( =s ) PI(^) + (ks + k ~ ) P c ( s ) l D ( s )
b ( 3 ) = [(s2 + k s ) b ~ ( s +
) sk~P~(s)]lD= ( s )b l ( s ) + k ~ s P c ( s ) l D ( s )
Applying Equations (4-48) also yields the same equation. Thus, P A and PI,, are
oppated on to produce: (1) a fi that consists qf pure P I and a filtered fiG; and (2)
a 3( that consists of a pure @I,, and a filtered bc.

4.6.3 Third-Order Complementary/Kalman Filter

Figure 4-11 shows a generalized third-order complementary/Kalman filter.

Third-order complementary/Kalman filter: Formulation 1. Con-

sider the model for the third-order dynamic system in which x = position, i =

U 4

a) Formulation 1


b) Formulation 2
Figure 4-11 Third-Order Complementary/Kalrnan Filter
Sec. 4.6 Comp(ementarylKa1manNlter Approach to Estimator Design 171

rate, and x = acceleration, and the equations of motion in state-space form are as
XI = X? + 111 + LVI, E \ W I ( ~ )=
] 0. )] -
E [ I v ~ ( ~ ) w ~ =( Tqlb(t T)

where x = s l + 6 1 1 2 ,i = x 2 + l r l , and .t = x.3 + rr, and the measurement equation


The preceding model dynamics are illustrated in Figure 4-lla which includes the
optimal state estimator. With regard to Kalman filter formulation,

where K is the Kalman gain matrix, and the state estimate equations are

+ b112, x = i 2 + U I , and x = 23+,,.

and the estimator outputs are i = 2 = it

The solution of the second-order filters is identical to that of the third-order filter
presented in this section, with the exception of k, being set to zero. This results in
the elimination of the last state x 3 in the third-order filter. The transfer function
from the control inputs u ( t ) , i r l ( t ) , and u 2 ( t ) , and the measurements z ( t )to the state
estimate vector ~ ( t are
Modem Nlterlng and Estimation Techniques Chap. q

where d ( s ) = O is the characteristic equation. Therefore,

Z(S) = A($)= kl(s)+ bU2(s) = [sU(s) + s 2 U l ( s )
+ bs3U2(s) + (ks2 + kls + k l ) Z ( s ) ] l d ( s )

The transfer functions from z ( t ) to ~ ( trepresent

) the last column of Equation (4-
51) as

As shown in Equation (4-52), all the transfer functions from either the control inputs
u, u l , and u2. or the measurement z , to the state estimates and the estimator outputs,
are low-pass filters. The algebraic Riccati equation is

and the steady-state Kalman gain is

Case 1. In this case, 43 = O, q l # 0, and q2 # 0, and Equation (4-55) with

kl = O has been found to be the same as the second-order filter gain Equation (4-
45). Therefore, when q3 = 0, that is, the process noise intensity on the third-state
is zero, the second-order filters employed in Section 4.6.2 are identical to the third-
order filter presented.

Case 2. In this case, ql = 0, q2 = 0, q3 # 0, and Equation (4-55) has been

found to be
k = 2(q3/r)I1", kl = 2(q3/r)'I3, k l = (q3/r)1'2 (4-56)
Sec. 4.6 ComplementaryIKaImanFilter Approach to Estimator Design 173.

Case 3. If there exists a disturbance x n in the x measurement, then r = x

+ XI,. If ~ c ( t )= 0, u l ( t ) = jc = rate, and u2(t) = 0, then in the Laplace domain,
U ( s ) = 0, U l ( s ) = s X ( s ) , and Z ( s ) = X ( s ) +
X D ( S ) .Applying Equation (4-52)
?(s) = X ( s ) + (ks2 + k r s + k l ) X n ( s ) l d ( s ) ,

X(S) = X(S) + (kls2 + k l s ) X ~ ( s ) l d ( s ) (4-57)

X(5) = X(S) + kls2X~(s)/d(s)
By setting k , = 0, Equation (4-37) is reduced to the second-order filter Equation
If a gain h is added to the measurement input, then z' = h r = h x ,
Case 4.
+ hbuz + hv and v' = hv is the new measurement noise with zero mean and variance
r' = h2r. The new state estimate equations and Kalman gains are
21 = 22 + 141 + k ' ( ~ -' h.?] - hbL42)

j3= k ; ( z ' - hG1 - hbuz), k' = klh, k ; = krllt, ki = kllh

The error covariance P is the same as that in the previous formulation except that
r is replaced by r'lhZ.

Example 4-22: Trajectoly estimation. In a three-dimensional flight,

the trajectory estimation problem involves finding the best estimate of position,
velocity, and acceleration in the X , Y, Z inertial coordinates, given the position,
velocity, andlor acceleration measurements. The problem is defined by
x = X , Y , Z position, u = X, Y , z acceleration
z = X, Y, Z position measurement, ul = AX, AY, A Z velocity correction
2 = X, Y, z velocity, u2 = AX, A Y , A Z position correction
Note that the problem formulation is similar to the navigation filtering for position
and velocity estimate presented in Example 4-18. However, uz is nonzero here and
the acceleration estimates are provided as the estimator outputs which were not
available in that section. In some cases, X, Y , Z acceleration measurements may
not be available and hence need to be estimated. Then, u = AX, A Y , A Z acceleration
correction and x j = unknown acceleration to be estimated. In altitude, altitude rate,
and altitude acceleration estimator design, the problem is formulated as
x = h = altitude
z = x + v = XI + bu2 = altitude.measurement
u = h = altitude acceleration
Modem Filtering and Estimation Techniques Chap. 4

ul = ~ h = altitude rate correction

u2 = A h = altitude correction
Terminal guidance tracking filter (presented in Chapter 7) is another example of the
trajectory estimation problem discussed in this section.

Third-order wmplementary1Kalman filter: Formulation 2. The

problem is defined the same way as in the third-order complementary/Kalman filter
(Formulation I), except x 3 here is assumed to be filtered white noise, that is, the
state Equations (4-49a) now become

and the state estimate equations are

i1= 22 + u1 + k(z - 21 - bu2)

43 = -2317 + k l ( z - 21 - bu2)
Applying Equation (4-la),

Equations (4-61) are solved nun~ericallyfor P ( t ) to generate gains k, kl, and k ~ ,

which mechanize the third-order filter depicted in Figure 4-11b. The steady-state
Kalman gain is
k = Plllr, k, = P12/r, k, = P l J r (4-62)
If a gain h is added to the nIcasurcment input, then z' = Irz = h x l + I~brrz + Ifv
and v ' = hv is the new measurement with zero mean and variance r' = h'r. The
new state estimate equations and Kalman filter gains are
Sec. 4.6 ComplementaryIKalman Filter Approach to Estimator Deslgn 175

The error covariance P i s the same as that in the previous case except that r is replaced
by r'lh2. Assuming measurement noise v in Equation (4-49b) to be a colored noite,
then following Equations (4-28) to (4-31) yields the new measurement equations
z = x +
v, ir = F v + +
~ H HA - FH = [ - F , 1,0], E [ f i ( t ) ~ ( ~=) ~ ]
G v l , =
[ H N Q N T H T GrGT]6(t - 7 ) = r,S(t - 7), andS(t) = Q ( I ) N ( ~ ) ~ HApplying
Equations (4-31) yields the new equations for state estimates, covariance matrix,
and Kalman filter gain matrix. If F and C are scalar parameters, then Equations (4-
31) yield the following simple form of new covariance and Kalman gain equations:

-~ J ' I-I - 2P12 + q l - kfr,, -PI-2 - P22 + PI3 - k,krcr,

dt dr
Inertial Navigation'
The block diagram of Figure 5-1 is provided to illustrate how the different parts
o f the subject matter presented in this chapter are related t o one another. For instance,
the discussions of common requirements, navigation con~putation,and coordinate
systems together form the framework for the design of a self-contained INS. The
t w o types o f self-contained INS are the gimballed INS and the strapdown INS.
Following discussions of each of these two types o f INS, a comparison o f the two
is presented. Combining INS with other navigation aids introduces the subject of
filtering, and finally extending INS uses to provide flight control outputs leads to
an integrated INS dcsign.


Inertial navigation is thc process of measuring accclcration onboard a vchiclr 2nd

then integrating that accclcration to dctcrminc the vehicle's vclocity and position
relative to a known starting point. Accclcrotnctcrs arc uscd to sense thc magnitude
of thc accclcration, but acceleration is a vector quantity having direction as wcll as

'This chapter has becn technically cditcd and improvcd by I l r . J. Stanley Ausrnan of Litton
Guidance and Control Systcms.

Sec 5.1 lntroductlon

Common Requiremenu

Coordinate Systems Self-contained INS Navigation Computation

(Sec. 5.3) (Secs. 5.4 & 5.5) (Sec. 5.3)

Comparison of
Gimballed INS &
Strapdown INS
I I (Sec. 5.6)

External Navigation Aids

(Sec. 5.7)
Figure 5-1 Inertial Navigation

magnitude. For this reason a set ofgyroscopes, or simply gyros, are used to maintain
the accelerometers in a known orientation with respect to a fixed, nonrotating co-
ordinate system, commonly referred to as inertial space.
This does not mean that the accelerometers themselves are kept parallel to the
axes of this nonrotating coordinate system, although some systems are mechanized
this way. Instead, the accelerometers may be kept parallel to a set of earth fixed
axes by causing the gyroscopes to precess at earth rate. O r they may be maintained
locally level by precessing the gyros at earth rate plus the angular rate of the vehicle
over the earth. In most of the modern aircraft inertial navigation systems today, the
gyros are precessed to rotate with the aircraft itself, resulting in what is known as
the strapdown system.
One characteristic inherent in accelerometers is that they do not measure grav-
itational acceleration. For example, an accelerometer in free fall will measure zero
acceleration (0 g), but it is really accelerating downward at the acceleration of gravity.
For another case in point, consider an accelerometer sitting on a stationary table. It
will measure 1 g due to the gravitational reaction force of the table pushing upward
on the accelerometer, but the accelerometer is not really going anywhere, at least
with respect to the earth.
In order to overcome this deficiency of accelerometers, it is necessary to add
in the effect of gravity based on a gravity model. A computer uses this model to
calculate gravity as a function of position, which in turn is determined from the
double integration of the measured acceleration. The most common gravity model
employed is one which accounts for both the latitudinal and altitudinal variations
in gravity. This model is symmetric about the earth's polar axis and also about the
equator. For most navigational purposes, this ellipsoidal model is sufficient, but
highly accurate systems must in addition account for gravity anomalies which are
deviations from this simpler model.
178 Inertlal Haulgation Chap. 5

It is difficult to navigate a flight vehicle at night or in bad weather, even with

the availability of radio and radar navigational aids. Similarly, smaller, faster, more
maneuverable surface vehicles armed with guided missiles, torpedos, and cannons
meet with the same difficulty. Consequently, the tremendous need for improved
navigation gave rise to the necessity for onboard systems that would supply the
position, velocity, and attitude of the vehicle on which they are mounted.
The first inertial navigation systems (INS) in operation were those in the Ger-
man V-2 rockets of World War 11. Since then, INS have become standard equipment
in ballistic missiles and in all modern fighter aircraft, airliners, and naval ships. what
makes INS particularly suitable for ballistic missile guidance is their autonomous
mode of operation during the boost phase. This feature yields a jam-proof guidance
system [Bar-Itzhack and Berman, 19881.
Much of the progress in the design and use of inertial systems can be traced
to the strategic missile programs of the 1950s and, specifically, to the work done
at the MIT Instrumentation Laboratory (currently the Draper Laboratory). It was
during this time that the need for high accuracy at ranges exceeding 1,000 k m using
autonomous systems became more and more apparent. As a result, the U.S. Air
Force requested the laboratory to develop inertial systems for the Thor and Titan
missiles, while the U.S. Navy requested the same laboratory to develop an inertial
system for the Polaris missile. The practicability of a self-contained all-inertial nav-
igation system (INS) for aircraft had just been demonstrated by the laboratory in a
series of flight tests using a system known as Spire [Draper, 19811. With the MIT
Instrumentation Lab providing the spark, several privately-owned companies took
up the challenge and developed and produced their own INS for long-range cruise
missiles (North American Aviation for the Navaho missile and Northrop Aircraft
for the Snark missile), for strategic bombers (Sperry Gyro Corp. for the B-38), and
for ballistic missiles (GE for Polaris and American Bosch Arnla for the ATLAS
missile). The remarkable succcss of those early missile and aircraft programs paved
the way for applications in aircraft, ships, missiles, and spacecraft. Today, inertial
systems are standard equipment in both commercial and military applications.
The types of gyros currently found in inertial navigation systems (INS) include
the floated rate integrating gyro, the electrostatically supported gyro (ESG), the
tuned rotor gyro (TRG), and the ring laser gyro (RLG). The first three are me-
chanical in nature, relying on a spinning mass to produce a large angular momentum
which tends to stay fixcd in inertial space. The floated and tuned rotor gyro suspend
the spinning mass in mechanical gimbals, while the ESG suspends the spinning
mass, generally a spinning ball, with electrostatic forces so as to prevent any mc-
chanical contact between the spinning ball and the outside case [Schmidt, 19871.
The RLG is a closed optical cavity containing two beams of single-frequency
light traveling in opposite directions around the closed cavity. The two beams in-
terfere with one another to form a standing wave pattern around the cavity. This
pattern of bright and dark spots tends to stay fixed in inertial space, so that an
observer or light detector fixcd to the cavity can sense rotation of the cavity as he
(or it) rotates relative to this fixed pattern. A more complete dcscription of the four
types of gyros listed earlier, together with accelerometers and error models for each
can be found in Savage 11978, 1984(a)].
Inertial navigation systems are either gimballed or strapdown. In a gimballed
system the gyros and accelerometers are isolated from the rotations of the vehicle
so that they can be maintained in a specific orientation relative to the earth or inertial
space. This greatly simplifies the computations of velocity and position and also
greatly reduces the requirements on the gyros which would otherwise have to mea-
sure very high angular rates. In a strapdown INS the gyros measure the full rotation
rates of the aircraft and keep trace of the instantaneous orientation of the acceler-
ometers in order to properly integrate the accelerations into velocity and position.
I t was the advent of small, high-speed digital computers that made strapdown INS
In general, the three basic functions of an INS are sensing, computing, and
outputting. The accelerometers and gyros perform the sensing function and send
their measurements of acceleration and angle rate to the computer. T h e computer
uses these data to generate velocity, position, attitude, attitude rate, heading, alti-
tude, range and bearing to destination. If true air speed is available from an air data
system, the INS can also compute wind speed and direction as xvell as drift angle.
The outputting function simply means sending the appropriate data to thc flight
control system, fire control system, reconnaissance sensors, or the control and dis-
play unit (CDU) as required for the particular mission.
Accordingly, the significance of the navigation system to the flight mission
becomes obvious, as the former provides dynamic parameters for flight and path
control, assists in terrain-following flight, allows avoidance of enemy defenses via
corridor flight, and reduces operator load. The ability of a navigation system to
accomplish these tasks greatly enhances mission success. Particularly desirable fea-
tures in a navigation system include: self-containment; security; worldwide capa-
bility; bounded navigation errors for position, speed, attitude and heading; no at-
titude, weather, or terrain limitations; minimum airframe structural impact;
minimum size, weight, power, and cost; and high availability.
If aided by other external navigation devices, the INS is called an aided INS.
The so-called external navigation devices can be position aids or velocity aids. The
former consists of GPS, MLS, MLS-GPS, LORAN, OMEGA, T A C A N and ba-
rometer (see the following for definition). The latter consists of Doppler, odometer,
and laser velocity meter. The following list covers several important navigation
devices and the corresponding section in which each one is covered in some detail:
global positioning system (GPS), Section 5.7.2; tactical air navigation (TACAN),
Section 5.7.3; long-range navigation (LORAN), Section 5.7.4; terrain contour
matching (TERCOM), Section 5.7.5; Doppler radar, Section 5.7.6; and star trackers,
Section 5.7.7.
An aided INS composed of the INS and external navigation devices is depicted
in Figure 5-2a. The evolution of these flight vehicle navigation devices is given in
Figure 5-2b. It is seen that a GPS is the most up-to-date navigation device, offering
the highest accuracy for flight vehicle navigation. The operational characteristics of
180 Inertial Nauigation Chap. 5

1950 1960 1970 198)


Sec. 5.2 Common Requirements for lnertlal Navigators 181



Figure 5-2 (a) Navigation Parameters Available from Sensors (b) Navigation Technology
Evaluation (c) Operational Characteristics of Navigation Sensors ( F r o ~ n[.21izrri11, 19861 w i t h
pennissio~zfrom Litton)

these navigation devices are given in Figure 5-2c. Generally speaking, the accuracy
required of a navigation system for most applications is on the order of 1 nautical
mph of flight. More stringency may be placed on the required accuracy in certain
phases of flight, such as ground attack for example, in which the accuracy must be
in the neighborhood of a few meters. In these cases, an external reference is necessary
to update the INS in order to achieve the required accuracy.


Vehicle and weapon systems. The requirements of a navigation system

for weapon systems can be categorized as follows. In the case of piloted vehicles,
medium accuracy (1 nautical mph) is mainly what is required of the navigation
182 Inertial Navigation Chap. 5

system for tactical aircraft, whereas landfall position fix, velocity, and nuclear hard-
ness are all required of the navigation system for long-range strategic bombers.
Furthermore, low accuracy or simple attitude reference is the basic navigation re-
quirement for observation aircrafts, while size and weight are critical t o helicopteri.
In the case of short-range missiles and tanks, size, weight, and low cost figure
predominantly in the requirements, whereas medium-range missiles require some-
what better navigation. Reliability is a key requirement for intercontinental ballistic
missiles and tanks. Submarines and ships require extremely high accuracy over long
time periods. For military use, the ideal navigation system should be fully auton-
omous, undetectable, unjammable, usable worldwide in all weathers, and of van-
ishingly low weight and infinitesimal cost. These aims have driven development
for the past thirty years. Systems are now sufficiently reliable and accurate for
virtually all modern fixed-wing military aircraft to be equipped with an inertial
navigator [Barnes, 19871.

Navigation requirements for weapon delivery. Navigation accuracy

is a key factor for successfully carrying out low-level attacks against targets at known
locations. First, it gets the attacking aircraft to the right position to find the target.
Once the target is identified, the INS must provide an attitude reference for the
targeting sensors (optical sight, laser range finder, radar, IR sensor, or TV sight).
After the target position is located with respect to the aircraft, that position is con-
tinually updated by the INS as the aircraft moves into its weapon launch position.
For the weapon delivery itself, the INS provides initial position, velocity, and at-
titude data for the weapon.
Stand-off weapons require the same set of funct~ons,but the missile itself takes
over the final target acquisition and terminal homing. For such weapons, the INS
in the launch aircraft simply has to provide initial conditions for the missile, which
may also have a small INS for midcourse guidance, good enough to steer the n~issile
into a target acquisition basket.

Stand-off weapon systems. The following discussion on navigation re-

quirements for precision-guided stand-off weapon (PGSOW) delivery are based on
detailed results shown in Shapiro (19861. The successful delivery of a PGSOW
depends, in general, on three key elements: knowledge of launch position param-
eters, knowledge of target position, and accuracy of navigation between these two
positions. The basic PGSOW mission scenario is shown in Figure 5-3 which depicts
the flight legs from aircraft takeoff to weapon launch point, then the weapon flight
to target arca window, and finally to the impact point.
The focus hcrc is on the choices for providing accurate navigation to the low-
cost, short-range (20-mile) RPV or powered glide bomb which cannot sustairl the
size, weight, or cost penalties of a high-precision navigation system. In order to
ascertain the accuracy requirement for the PGSOW, it is necessary to estimate the
capability of the terminal seeker which, in the final analysis, determines the sizc of
the target window or search arca. The burden placed on the terminal seeker incrcnscs
Sec. 5.2 Common Requirements for Inerthi Naulgators 183

- - ~




Figure 5-3 PGSOW Mission Scenario ( F r o ~ n[Shapiro, 19861 w~idtpennisrionjom ACARD)

with the size of the required search area imposed by uncertainty in relative weapon
and target positions. In the case of an 1R seeker, for example, the window may be
relatively narrow for a given angular field of view because of the limited range due
to atmospheric attenuation. An active millimeter wave radar can provide longer-
range operation in bad weather (several kilometers); however, the limited aperture
available for the antenna and the consequent limited angular resolution creates prob-
lems with false alarms and poor detection probability as the search area is increased.
Therefore, the PGSOW must navigate to the target zone with sufficient cross-
track accuracy to insure that the target lies within the swath described by its forward
motion and the scanning limits of the terminal seeker. Based on the foregoing types
of considerations. it can be concluded that a vractical search area width is about 1
km. The corresponding navigation cross-track error budget should be limited to
one half of this, or 0.5 km. The along-track error is dependent on the allowable
false alarm and desired detection probability criteria. Obviously, as the along-track
extent of the search decreases, these probabilities become more favorable. The along-
track navigation error can be specified, however, as being the same as the cross-
track error, namely 0.5 km. T o achieve the preceding performance objectives, the
navigation system for the PGSOW must have a heading error over the 20 miles of
better than 1 deg and an along-track error ofless than 1.5 percent of distance traveled.
The future will see an increased demand for more agile missiles for both air-
and ground-launched applications. Gun-launched missiles involving high acceler-
ations at launch will contain guidance systems. Inasmuch as strapdown guidance
systems have been included in missiles that are subject to high roll rates and high
184 Inertial Navigation Chap. 5

acceleration, the need for more rugged navigation instruments having a dynamic
range greater than what conventional gyros offer is brought to the forefront. A
fiber-optic gyro is described [Kay, 19873 that, although it is still in the laboratory
stage, has demonstrated good performance and, at the same time, ruggedness, quick
starting, and low cost.

Visual attack systems. Following Barnes [1987], in an effort to expose

the aircraft to as little ground-based radar as possible, much of the development
subsequent to World War 11 has been focused on low-altitude flight. Another goal
related to this effort was, until the development of look-down Doppler radars, to
make it more difficult for a defensive fighter aircraft to identify an incoming raid.
Considering that the total time of an attack pass may be just a few seconds, target
acquisition is rendered more difficult by low-altitude attacks. In these few seconds,
the pilot of an attacking aircraft armed with unguided weapons must maneuver on
to an attack heading, stabilize the aircraft path, and release the weapon. Precise
navigation and an accurate calculation of weapon release are two elements that are
absolutely essential for a successful attack. The former ensures that target acquisition
occurs as early as possible, while the latter minimizes delivery error.
Regarding attacks that can be conducted under visual conditions, the require-
ments for precise navigation with an acceptable workload, especially at high speed
and low altitude, have been met by the introduction of inertial platforms and moving
map displays. Although such developn~entssimplify the task of acquiring pre-
planned, fixed-position targets early, they do little for acquisition of targets of op-
portunity. Similarly, two cle~nentshave combined to meet thc rcquirements of
accurate calculation of weapon release. The first involves the ability to obtain data
on aircraft altitude, heading, and speed from the inertial platform. Thc second rclates
to the ability to digitally process in the aircraft range to target obtained from a
ranging systenl. The ability to carry out high-speed, low-altitude attacks a t night
and in poor visibility has also bcen enhanced owing to other dcvclopmcnts. One
of the first of thcsc was ground-mapping radar, which enabled pilots ro attack targets
giving discrete radar returns and targets whose position was known relative to some
observable and recognizable feature on the radar screen. However, radar resolution
is poor relative to visual wavclcngths and forward-looking infrared (FLIR) scnsors
greatly cnhancc the ability of the aircrew to identify both targets at known Iocatiol~s
as wcll as targets of opportunity.


5.3.1 Coordinate S y s t e m s

I n navigation computations, thrcc colnponct~tsof accclcration scr~scdby a n or-

thogonal triad of accclcro~nctcrshave to be transfornlcd onto a rcfcrencc coorditlntc
systcm. Thc vclocity and position of thc vchiclc arc then computcd with rcspcct to
Sec. 5.3 Naufgatlon Computation and Error Modeling

Figure 5-4 (a) Earth-Fixed Coordinate (b) Geographic Coordinate

this reference coordinate system (RCS). Because the reference coordinate system
affects both the computation time and accuracy, a suitably selected RCS can min-
imize the processing errors.
Beginning with an earth-fixed set of coordinates (Figure 5-4), we can, by a
series of two rotations first about the z-axis through the longitude angle and then
about the negative y-axis through the latitude angle plus 90 deg, transform from
the earth-fixed coordinates into a navigation coordinate system. In this case the
navigation coordinate system is defined as one in which the x-axis is north, the y-
axis is east, and the z-axis is down. The transformation is depicted graphically in
Figure 5-5. Figure 5-6 derives the resulting direction cosine matrix.
Such a transformation is known as an Euler angle transformation. A similar
transformation will take us from the navigation coordinates into the aircraft body-
axis coordinate system. The first rotation is about z (down) through the heading

Figure 5-5 (a) [Z] Transformation (b) [YI]

Transformation (c) [Yz]
186 Inertial Navigation Chap. 5

X -sin @ 0 - a s @ w s h sin h 0 X X
Y - 0 1 0 -sinhwsh 0 Y
Z ms @ 0 -sin cP 0 0 1 Z Z

- sin@ms h - sin@sin h - cosQ

-sin h msh 0 where h = longituic (East)
<= Q = latitude (North)
cos@msh cos@sin h -sin@
Figure 5-6 Earth-to-NAV Direction Cosine Matrix

angle +. The second rotation is about the new y (right-wing) axis through the pitch
angle 0. The third and final rotation is about the new x (roll) axis through the roll
angle 4. The direction cosine matrix for this transformation is derived in Figure 5-

In a gimballed system, the accelerometers would be maintained parallel to the

navigation system axes. In a strapdown system, they would be parallel to the aircraft
body axes (or in some fixed relationship to the aircraft axes) and their outputs have
to be transformed into navigation axes in order to calculate velocity and position
with respect to the earth. The inverse of the direction cosine matrix shown in Figure
5-7 could be used for the body-to-NAV transformation, but this involves com-
puting nine quantities a t a high iteration rate. More often this transformation is done
with quaternions, a four-parameter representation of the same thing. Conceptually
quaternions describe a coordinate transformation by a single rotation. Three of the
parameters are related to the direction cosines of the axis of rotation, and the fourth
parameter is related to the magnitude of the rotation about that axis.

5.3.2 Position and Velocity Generation

A simple block diagram of space-fixed navigation computation is given in Figure

- - -

5-8. In the figure and xSare inertial position and velocity vectors, respectively,
in a space-fixed coordinate system. Navigation and guidance of the Saturn launch
vehicles using this navigation computation has been implemented onboard with
negligible error contributions from the computations. For propellant economy, the

X 1 0 0 w s 8 0 -sin0 aSIV sinIV 0 X X

Y - 0 w s + sin+ 0 1 0 - s i n v cosv 0 Y
Z 0 -sin+ a s + sin0 0 w s 8 0 0 1 2 Z

cosems~, ws 0 sin v -sin 0

where Q = heading
%-- - cos 4, sin yl + sin + sin 0 cos 9 + +
cos cos + sin sin 8 sin I) sin cos 0 + 0 = pitch
- +
+ sin 4 sin Q + cos 4, sin 0 cos 9 sin cos 9 + ms 4 sin 0 sin I$ cos cos 8 + = roll
Figure 5-7 NAV-to-Body Direction Cosine Matrix
Sec 5.3 Naulgatlon Computatlon and Error Modellng

A ~ I

;I I - - - ...
* (X*,Y.,Zs)




Figure 5-8 Iqavigation Computations (Frorn [Haeusserrnar~n, 19811, O 1981 AIAA)

iterative, path adaptive guidance mode was selected. The flight vehicle is guided
along a minimum propellant trajectory to obtain the required end condition, such
as the required velocity vector and vehicle position in orbit or the velocity vector
for translunar injection. Figure 5-8 shows how the vehicle state is derived from the
velocity, measured by the accelerometer integrator, in the form of position and
velocity vectors in a space-fixed coordinate system which has its origin at the center
of the earth. The diagram further shows the calculation of the acceleration in the
space-fixed coordinate system as thrust-to-mass ratio, Flm.
The iterative guidance mode for flight- path
. adaptive guidance along a mini-
mum propellant trajectory determines the necessary vehicle attitude corrections and
the urouulsion termination command to obtain the reauired end conditions for the
i x

flight vehicle. The guidance functions for the thrust direction and for its termination
were expressed by a series development of the reference attitude and of the pro-
pulsion cutoff time in terms of position and velocity components, time, and ac-
celeration; the coefficients of the terms in the series were determined using calculus
of variations and stored in the guidance computer. Forover-the-earth flight vehicles,
the required position vector and velocity vector x are not expressed in space-
fixed coordinates but in earth-fixed coordinates, as seen in Section 5.3.1 [Haeus-
sermann, 19811.
Inertial Navigation Chap. 5

5.3.3 Data Processing

A block diagram of the processing of a generalized navigation computer is given

in Figure 5-9. In this figure, it is seen that: (1) the outputs A 8 c and A V A of the
inertial sensors, that is, gyros and accelerometers, are incremental due to digital
processing; (2) coordinate transformations are required in order to transfer the body
coordinates to earth coordinates ,(see Section 5.3.1); and (3) filtering is sometimes
required for data smoothing.

5.3.4 INS Error Analysis and Modeling

Following Bar-Itzhack and Berman [1988], the outputs of an ideal INS are the exact
position, velocity, and attitude of the host vehicle. The error analysis of INS or, in
particular, terrestrial INS brings to light many interesting features typical of INS.
It is no wonder that much of the literature on INS is concerned with its error analysis
and error propagation. In analyzing INS errors, questions arise immediately con-
cerning the behavior of the errors as a function of time, the ability to even measure
all error components, the ability to estimate those errors that cannot be directly
measured, and whether or not it is possible to control the errors. Rather than fo-
cusing, therefore, on new features of the INS, emphasis is placed on casting the
known INS characteristics in terms of linear systems and control theory concepts
and exploring hidden relationships between system parameters and states that affect
system performance.
In the examination of system observabiljty, a straightforward transformation
is used into observable and unobservable subsystems that, in turn, exposes the states
that hamper the estimation of INS errors during the initial alignment and calibration
1) A0 Proossing: (Angle Inawmt)
- - - +--
-Thnsforma~on Earth Caord.
sq - - - - -
-We, Ve

Navigation -& 6 6
Procasing +e"E
2) AV Ptocssing: (Vekxity Inaancnt)
+ xD YES%
&Ordinatc --+
Naviption +k
-%>~O,-AO~- P'occssing
& + AE
Ffltering ---4
, --b r
+,q: EulaAnglcs VB: Vclodty in Body Cmrd.
0 ~ : Body Rate in Body Coord. VE: Vdm'ty in Earth Coad.
6 ~ :Body Rate in Earth Chord. Xt YB & Pcsitions in Earth M.
r: T m Lag
Figure 5-9 Processing of a Generalizrd Navigation Computer
Sec. 5.3 Naulgatlon Computatlon and Error Modellng 189

phase o f operation. This approach was adopted successfully in the past by Kortum
[I9761 who considered the problem of INS platform alignment, in which the mea-
surements were the horizontal accelerometer outputs. A comparison of this approach
with the classical one that is presented here, as well as the discussion of uniqueness
and thc relationship between observability and quality of estimation, provide ad-
ditional insight into the observability issue. The examination of INS as a unified
system from a control theory point of view sheds more light on the systcm and
contributes additional insight into the analysis of INS [Bar-Itzhack and Bcrman,

I N S error models. There arc two approaches to the derivation of INS

error models. O n e of them is known as the perturbation (or trrre frame) approach,
and the other is known as the psi-angle (or cotnputer frame) approach [Benson, 1973).
When deriving the perturbation error model, the nomlnal nonlinear navigation equa-
tions are perturbed in the local-level north-pointing Cartesian coordinate system
that corresponds to the true geographic
- . location of the INS. The psi-angle error
model, on the other hand, is obtained when the nominal equations are perturbed
in the local-level north-pointing coordinate system that corresponds to the geo-
graphic location indicated by the INS. The development of the first model can be
found in Broxmeyer [I9641 and Britting [1971], whereas the development of the
psi-angle model can be found in Pitman [I9621 and Leondes [1963]. Benson [I9751
showed that both models are equivalent and yield, therefore, identical results. The
differential eauations that describe the error behavior of the INS are divided into
equations describing the propagation of the translatory errors and equations de-
scribing the propagation o f the attitude errors.
Bar-Itzhack and Berman [I9881 pointed out that most of the published works
on INS errors adopt the psi-angle approach and use the velocity error version of
the translatory error equation. This model is also used in the present analysis. It
possesses the advantage that the translatory error is not coupled into the attitude
error equations [Pitman, 19621. The physical attitude difference between the plat-
form and the local-level north-pointing coordinate system is calculable using the
position and attitude errors obtained from the solution of these INS error equations.
The choices made yield a complex terrestrial INS error model expressed by the
following equations [D'Appolito, 1971; Nash, 19721:

where v , r, and I) are, respectively, the velocity, position, and attitude error vectors;
0 is the earth-rate vector; w is the ang~~lar-rate vector of the true coordinate system
with respect to an inertial frame; V is the accelerometer-error vector; f is the specific-
force vector; Ag is the error in the computed gravity vector; p is the vector of the
rate of rotation o f the true frame with respect to earth; and E is the gyro-drift vector.
190 Inertial Naulgatlon Chap. 5

It can be shown [Leondes, 19631 that in the local north, east, and down coordinate

R cos L
[ 0
- R sin L
] (5-2)

where L is the local latitude. The vector w is computed as follows:

o = n + p (5-3)

.=[ icos L
- A sin L
and ); is the longitude rate. When Equations (5-1) are resolved in the true frame,

nine scalar differential equations are obtained that can be put in a state-space model.
If the expressions given in Equations (5-2) to (5-4) are used, the resulting state-
space model is shown in Figure 5-10. The INS error propagation is of interest when
the system is at rest. In most cases, INS alignment and calibration, which are a most
essential phase of the operation of any INS, take place when the system is at rest.
Also, the distinct behavior of INS errors is exposed when the system is a t rest.
When the INS is at rest, Equations (5-1) reduce to:
0 0 1 0 0 0 0 0
0 0 0 1 0 0 0 0
0 0 0 0 1 0 0 0
d V . ~ 0 0 0 2RD 0 0 A? 0
- -w2 0 -20D 0 2fiN -2 0 0
vD 0 202 0 -2nN o 0 0 0
O O O O O O RD 0
4E 0 0 0 0 o -RD o R,W
-. - 0 0 0 0 0 0 -ON 0 -

where 0' = g / R , R N = cos L is the north component of earth rate, and R D =

-a sin L is the down component of earth rate. The latter model can be writtcn as
i ' = A f x + f' (5-5b)
Sec. 5.3 Naulgatlon Computation and Error Modellng 191

where the definitions of x ' , f ', and A' arc obvious. Initial alignment and calibration
arc usually performed when the INS is resting at a location whose geographic co-
ordinates are known almost perfectly; that is, r N - - - r~ r~ 0 [Pitman, 19621.
For this reason, the first three states of the state vector can be eliminated. The
measured signals during initial alignment and calibration are v . and ~ V E , but v ~ ,
the vertical velocity error, is of no interest. Moreover, the vertical channel is only
very weakly coupled with the horizontal channels [Hutchinson, 1968; Yavnai, 19801.
Thus, v r , does not play any role in the development of the measured signals v ~ and
V E ; hence. v ~ can also be eliminated from the state vector ofthe model that describes
the INS error behavior during the initial alignment and calibration stage. In modern
systems, a Kalman filter is used to perform the initial alignment and calibration;
however, the model given in Equations (5-5) is not suitable for use in a Kalman
filter since the accelerometer error and gyro-drift rates are not white-noise processes,
as required for proper use in a Kalman filter. This obstacle is very easily overcome
since, fortunately, the statistical characteristics of realistic accelerometer and gyro
error data can be represented quite accurately by the outputs oflinear models driven
by white-noise processes [Gelb, 1974; Heller, 19751. When this is done and the
models are augmented with the basic INS error model, the driving force of the
augmented model does not contain correlated noise, and the model can then be used
in a Kalman filter. In the analysis here, it is assumed that the accelerometer errors
are basically bias errors and that the gyro errors are basically constant drifts, that

When r . ~ ,YE, r D , and v ~ are removed from the state of the INS error model of
Equations (5-5) and the resulting model is augmented with Equations (5-6). the
following is obtained:

This model can be written as

where the definitions of x and A are obvious. The latter model describes the prop-
agation of v~ and v~ when the INS is at rest and the accelerometer and gyro errors
192 Inertial Navigation Chap. 5

sL and CLdenote, rcspcaively, the sine and &e of L,and the subscripts N, E, and
D denote the n o d , cast, and down mmponents, rcspsctively.
Figure 5-10 State Equation of Terrestrial INS Error Model (From [Bar-ltzhatk et al., 19881,
0 1988 AIAA)

are constants. Although it seems that this model is not general enough, it turns out
that the conclusions derived from it are quite useful. An estimate o f the state vector
yields the estimation of $, as well as the estimate of the accelerometer and gyro
constant errors. Obtaining the former is known as alignment, and obtaining the
latter is known as calibration. An examination of the models given in Equations
(5-5) and (5-7) from the point of view of a control analyst is presented in Bar-
Itzhack and Berman [1988].


In inertial platform gimbal mechanizations, the gyroscopes mounted on a stable

element in a gimbal1 system measure angular rates, and the gimbal drive system
uses the angular rate information in a feedback manner to rotate the gimbals and
null the angular motion sensed by the gyroscopes. In this manner, the gyroscopes
and accelerometers mounted on the stable element are inertially stabilized from the
vehicle motion and the stable member physically represents an inertial reference
Alternatively, control torques to the gyros can cause them to precess a t pre-
cisely the same rate as the combination ofthe earth's rate of rotation plus the angular
ratc of thc vehicle's position with respect to the earth. In this way two of the
accelerometers can be maintained in a locally level plane, the third being parallel to
the vertical axis a t all times. Most aircraft and marine gimballed inertial navigation
systems are mechanized in this manner.

5.4.1 Gimbal Mechanism

In aircraft applications, a four-gimbal mechanism as shown in Figure 5-1 1 for the

platform unit of a gimballed INS is required to avoid gimbal lock. In surface vchicle
applications, a three-gimbal mechanism for the platform unit of a gimballed INS
suffices as shown in Figure 5-12.
Sec 5.4 Glmbalied Inertlal Navigation System (INS)



Figure 5-11 Four-Gimbal Mrchanisn~

5.4.2 Inertial Sensors on the Stable Platform

Inertial sensors on the stable platform are gyros and accelerometers. If SDF (single-
degree-of-freedom) gyros are to be used, three gyros (mutually orthogonal) are
required for the north, east, and vertical (NEV) coordinates. If 2DF (two-degree-
of-freedom) gyros are to be used, two gyros are required and installed. In Figure
5-13 three gyrolaccelerometer pairs are installed so as to measure the angular rate
and acceleration along each of three axes.
5.4.3 Platform Alignment Modes
There are two alignment modes, coarse alignment and fine alignment. In the coarse
alignment mode, accelerometers are used for leveling control of the platform, while

Figure 5-12 Three-Gimbal Mechanism

Inertial Navigation Chap. 5

Figure 5-13 Inertial Sensors on the Sta-

ble Platform: Three-GyroIAcceleromet~

the gyros sense the direction of zero earth rate (east-west). In the fine alignment
mode, A Z gyro slew gives a more accurate aiignment of the platform. During the
fine alignment or gyro compass phase, the objective is to keep the pitch, roll, and
AZ gimbal aligned with true north and the horizontal. Any AZ misalignment causes
the east gyro to sense a component of the eaith rate and thus tilts the platforn~about
the east-axis. The north accelerometer then generates a pseudo velocity due to the
platform tilting about the east-axis. The pseudo velocity is then used to drive a r
gyro torquer and its associated gimbal servo to achieve the proper alignment.

5.4.4 Navigation Mode

Gimbal systems provide a good dynamic environment for inertial instruments, par-
ticularly in vchidcs that are to perform at high maneuver Icvels, since the gimbals
isolate the gvros
", from the rotational environment. In fact, the state of the art is such
that navigation performance of gimbal systcms can approach alnlost error-free in-
strument operation, to the point where uncertainties in the knowledge ofthe gravity
field become the dominant sources of navigational error. A simplified diagram to
show the relationship between the platform and the navigation computer is given
in Figure 5-14. The two mechanizations for the platform attitude reference are the
space stable and local vertical. The platform space-stable inertial mechanization is
given in Figure 5-l5a. Figure 5-l5b illustrates the case in which the geographical-
north-pointing local-level navigation frame is the instrumented stable member frame
for a gimbal system. The local~levelframe is the most common for gimbal systems.
The local-vertical frame is usuallv a form of a .,fiec-arimirth reference frame in which
no attempt is made in a gimbal system to point one of the stable-member axes north,
as illustrated in Figure 5-15c. This mechanization eliminates difficulties in flying
over the earth's poles. T w o of the axes are maintained level but are free to rotate
in azimuth aboutthe third axis, which is maintained along the local vertical [Schmidt,
Figure 5-16 shows a block diagram of the x channel in a navigation computer.
The operations of a navigation computer include: (1) The correction rate o, is used
Sec. 5.4 Gimballed Inertial Navigation System (INS)

I Acaltr~metm 1 Navigation
Gym 'bquing Rates mpuution

P l a t f m Heading
L - - - - J
Figure 5-14 R e l ~ r ~ o n s h bcr\vecn
lp rhe Plarfi)rni a t ~ t~l i ~N L Y I ~ L ~CIoO~ Ii i~p i ~ t e r

to torque the y gyro, and thus causes gyro precession. The gyro output is then used
to drive the platform servo so as to maintain the x-axis horizontal. (2) A s , Ay, and
Az are the measured accelerations from accelerometers in the navigation coordinates.
(3) The correction at the Ax summing point is composed of coriolis, centrifugal,
and transport accelerations.

5.4.5 System Internal and External Interfaces

Itlterfilce: The gimballed INS typically consists of eight main functional
blocks. For each function block, the interface infor~nationis given as shown
in Figure 5-17.
E.urrnrnl 111teuf;lce:The I10 (inputloutput) interface between the gimballed INS
and vehicles is given as follows: (1) Inputs-In order to permit operation in
aided-inertial configurations, a standard INS accepts the following digital in-
puts in M U X serial format (MIL-STD-1553): position update (latitude and
longitude), velocity update (velocities in INS coordinates), angular update (an-
gles about INS axes), and gyro torquing update (torquing rate to INS gyro
axes); and (2) Outputs-The system provides pitch, roll, and heading in both
analog (synchro) and digital form. In addition, the following outputs are pro-
vided on a serial M U X channel (MIL-STD-1533): present position (latitude,
longitude, altitude), vehicle attitude (pitch, roll, true and magnetic heading),
vehicle velocity (horizontal and vertical), and steering information (track angle

5.4.6 Navigation Mechanization and Error Model

Mechanization. In the example of Figure 5-18 Uoos and Krogmann,

19811, the diagram consists of two portions: I M U and navigation computer. O n
the I M U platform, there are three SDF gyros and three accelerometers with their
sensitive axes pointing in (N,E,V) coordinates. In the navigation computer, the
vertical channel is aided by the baro-height. The navigation computer includes com-
pensations for earth rate, vehicle rate with respect to the earth, coriolis acceleration,
and gravity. Figure 5-19 depicts Doppler aiding where the aiding velocity reference
V.YDcomes from a Doppler radar [Streeter; 19721. With the errob e,, the original
trtod m o n
stabilized Pccekrometer-
platform measured spccif ic
force in lnertml


Init~olc o n d ~ t m s
Posttlon ~n computer
* Lotitude ( L)
Lonpitude ( I I
Alt~tude( h )


Indial Altitude
Wndtt~otn mfamoton
Velocity lv7

Figure 5-15 (a) I'latforn~ Space-Stable Inertial Mcchanization (b) I'latform Geo-
graphic Local-Vertical Mechanization (c) Platform Free-Azimuth Mechanization
(Rrprinfed with prnnissio~tjotn (Srh~nidf,Sysfons fi C o ~ ~ f r61cytlopedio),
ol Copyright
11987), Per.qafnon Press PLC)
Sec. 5.4 GImballed lnertlal Navlgatlon System (INS)

Figure 5-16 x Channel Mechanization

o i a Navigation Computer

undamped loop can be damped as follows: (1) With free inertia ( k = 0, n o Doppler
aiding), the characteristic equation is "s glR = 0 with natural frequency w, =
(glR)'12 = 2 n l T where the period T is the so-called Schuler period; and (2) With
Doppler aiding, the characteristic equation is s' + ks + glR = 0 with damping
ratio 5 = kl(2w,,). Thus, with Doppler aiding, the original undamped system will
be damped with damping ratio 5.
With altimeter nidirlg (see Figure 3-20), the altitude reference h~ is given by an
altimeter and thus generates the flight height error E,, to damp the loop. In Figure
5-20, thc feedback gains K I and K 2 can be selected so as to nicely damp the vertical

*15V DC Power llSAC 400&

Gyro Master
28V DC
26AC 400 Hz

Gym Torque DatalAddress Bus

Gyro Output IMU Control - I
-Acc. Output Platform
IMU Status
INS I Vehicle
Platform ACC. Captu: ~~~~~~~i~~ velocity pulses VO .e, I
Gimbal Stab.

Platform Pitch, Roll & Heading

Vehicle Altitude
Display Indicator (ADI)
Figure 5-17 Interface Information for Gimballed INS
See. 5.5 Strapdown Inertial Navlgatlon System (INS)

J Figure 5-19 Doppler Aiding

Error models. In the Doppler-aided INS error model shown in Figure 5-

21, there arc nine states in the loop, five statcs in the A.y loop and four statcs in the
A loop (not shown). For altitude aiding, the error model is depicted in Figure 5-
20. A self-contained navigation system without aiding has the error model shown
in Figure 5-22a. Figure 5-22b illustrates a block diagram of a generalized mecha-
nization for a local-level free-azimuth INS.

Performance trade-off analysis. The errors 6x in any inertial system

mechanization (for example, nine states as in Figure 5-21) can be cast in the usual
state-variable form 6x = F6x + Gq where the coefficient matrices are functions of
the system mechanization and vehicle dynamics [Schmidt, 19781. This allows the
standard covariance matrix propagation equations to be used in predicting inertial
system performance. If external aids (see Section 5.7) are to be considered, the
Kalman filter update equations can be applied to study requirements on aiding ac-
curacy, frequency of updates and so on, to predict accuracies in the estimates 6 i
that could be achieved in the actual implementation [Schmidt, 19871.


A generalized strapdown INS consists of IMU, electronic unit (EU), and computer
unit (CU). The IMU is a sensor package with a gyro triad and an accelerometer
triad without a gimbal device. This is the main difference between strapdown nav-
igation systems and gimbal navigation systems. Since there are no gimbals to keep
the sensor package in a prescribed orientation, there is an extraordinary computa-
tional load imposed on the CU. The basic concepts of the strapdown inertial nav-
igation are that the body-to-NAV direction cosine matrix is computed using strap-
down body-mounted gyro outputs. Then the horizontal and vertical accelerations
are computed analytically using this direction cosine matrix to transform acceler-
ometer outputs into local-level navigation coordinates.

5.5.1 Generalized Strapdown Mechanization

A simplified block diagram showing the IMU and navigation computer is given in
Figure 5-23. An overview of the strapdown mechanization is given in Figure 5-
24. From this figure, it is seen that the acceleration Ab (in body coordinates) is
8 Vertical Definition of Variables
(Error Model)
9 .



R ~ Ry
% '
CONSTANT 32.2 ft/sec2
X'YZ ,
I T 7
i x, y, z
1 I
Sec. 5.5 Strapdown lnertlal Naulgatlon System (INS)

Figure 5-21 Srarc Error 6.4, Model

transformed onto local-level axes to compute the NAV axes acceleration compo-
nents A,, A,, and A,. The internal structure of the strapdown navigation system
is given in Figure 5-23. Note that in this figure, the angle and velocity output of
the inertial sensor components are small increments to be integrated in the body
attitude and velocity computations. Thc small increments in A 0 and A V are pro-
cessed as shown in Figure 5-26. Figure 5-27 Uoos and Krogmann, 19811 is another
view of the same system depicting the gyros and accelerometers. A detailed block
diagram of a rate gyro strapdown inertial navigation system is presented in Boggess

5.5.2 Strapdown Navigation Computer

The strapdown navigation computer may transform the body-axis sensor data into
either inertial coordinates or into geographic local-vertical coordinates. The strap-
down system computing in inertial coordinates is given in Figure 5-28a. The strap-
down system computing in geographic local-vertical coordinates is given in Figure
3-28b. Figure 5-29 shows a generalized block diagram for strapdown inertial com-
putations. This diagram shows that the computations are partitioned into two parts:
a high-speed section that transforms the body-axis measurements of acceleration
and attitude rates into navigation axes, and a slower-speed section that performs
the navigation routines just as they would be computed in a gimballed system.

5.5.3 Strapdown Computer Requirements

Typical computer requirements for a strapdown NAV computer call for a 16-bit com-
puter whose properties include a speed of 250-350 KOPS (the computer throughput
MEClL4NlZATlffl b P ~
tan* &P
-.pX sccZ+b $
FRE-AZIMUIH 5 4 * "a '%
box *I bo
g Gikw
- I - IC, b y + 4. bpi]


1. I. 2 m ? i p t i m ~ l t d li 0 8 t e .XU

A,. AY. A, h l e m r t e r artpts

Uechmuet ton i "*
v,, v,. vz v.locltY -ant8 relatire to e u t h
mrtb-0i.tiu ( a = O)
(no 0181 a p h i l i t l ) 0
$#, + =x.J?, %,
I-% P,, P,. p, ?l.tlan mrulu n t e s-at. relatlra to b r t h

wee-uiath An o n., ~9l,.u n h rat. C-~S

P ~ u i -(i+qclZ
~ ~ - cclZ -
b tic altitude
~~polar rn UUd -to1181 ndiuS Of the
Vppr s i w
(DlU f n ~ t t . s i y 01 tho u r t h
Umbilib) I*.isobe~cn iA: $49 ' %+1
4 -tic latitld*
-1 siw I$, t 1)
htbern h laeitpd*
a h h t h d e r -la

C~ Directim c m i u s d * f i n i ~latit*. 1-1-

and u i u t b -10. 8prifialb
C,, = uuacmk-sinasin48iah
c, = -.hawk-caasi.Csi.A
C, = w4si.k
$, = s i n a co 6
c , = solac04
c,, = .lo 4

Figure 5-22 (a) Level Channel Error Model (b) Generalized Mechanization for Local-Level
Free-Inertial Navigation Systems (From [Martin, 19861 with permissionf*om Litton)
204 Inertial Navigation Chap. 5



I I -
( I





1 IMU (Electronic Gimbal)

Accelerometers Navigation
.. Attitude Rotation Rates
Strapdown Attitude Pseudo-Platform Heading
Gyros * Computation ,Attitude

Figure 5-23 (a) T\pical Strapdo\vn System Simplified Block Diagram (b) Strap-
down IMU and Na\.igation Computer ((aifitn marf fit^, 19861 u~irltpcrmissiotl.~otn

may rcach 3.0 MIPS to support a wide bandwidth autopilot for high dynamic ap-
plications), a memory sizc of 12-15 K, 1 p S for addition, 8 p S for multiplication,
and 10 FS for division. Thesc con~putationrequircmcnts are for 1 NMIIhr navi-
gation systclns. Thc eri~~irotirtictitalreqrrirctr~etitsinclude an accclcration of 13-30 g,
500-10000/scc maximum angular ratcs, a 100 g (20 MS) shock, a MIL-E-5400 (curve
IV) vibration, and a MIL-E-5400 (Class 2) thcrmal.

5.5.4 Strapdown Error Model and Analysis

Error sources. l3asically, thc strapdown crror sourccs can bc grouped illto
thc two following catcgorics.
Instrument Errors. Thcrc arc significant diffcrcnccs in the navigation per-
formances bctwccn gimbal and strapdown systclns bccausc o f calibration limita-
tions, sensor inaccuracics, colnputational errors, and sensor crror propagatioll cf-




t C
OCI 1 = [OX' Oy' \]
: Oi 'Pi


+ni : i = X. y, z

Figure 5-24 Generalized Strapdowll M r c l ~ ~ n i z a t i o (1:rorrl

nl /Alclrrirl, IY#6/ wirh prrrr~issi~~r,li~~,,~
206 Inertial Navigation Chap. 5

Reference h r d i i a t e System
Vehicle Body CkxdiiteSystem (Locally Level,A Space E m )

0 -A0
P' &
R' Attitude -)Attitude
MY (0, $, q)
4 Inertial and
AVx (Y ,V y ,V,)
Bod' sensor

Campubtiom +Present Pasition (k,h a)


Conuol System

Figure 5-25 Internal Structure of the Strapdown Navigation System

fects. O f these, the instrument errors can be caused by the following error sources
[Schmidt, 19871. (1) Calibration limitations on strapdown systems arise because the
inertial sensors, which are rigidly attached to the vehicle, cannot be arbitrarily ori-
ented at different angles to provide known inputs from the earth's gravity and
angular velocity vectors for a stationary vehicle. (2a) In single-degrce-of-freedom
gyros, a gyro drift, called anisoinertia-induced drift, will be prescllt whenever the
gyro experiences accelerations such as vibration concurrently about its spin and input
axes. To eliminate such an error source requires complicated compensation, so the
system designer must understand the interplay between vehicle environment, sensor

-(We @,*I





Figure 5-26 Typical Strapdown Software Requircmcnts (F~OIII
/hlorritr, 191161 urirh
perrnisriorl j o ~ nLirrort)
Sec 5.5 Strapdown Inertial Navigation System (INS)

Figure 5-27 Gyro and Accelerometer Components of Strapdown Navigation Sys-

tcrn (Frorrr Uooi nrrd Kroyrnarm, 19811 wirl~purnrrisio~rjorrr AGARD)

design, and the resulting navigation system performance when selecting a particular
type of gyroscope for a strapdown system. (2b) Gyro-torquing scale factor error,
particularly asymmetrical scale factor, causes errors in a vibratory environment. (2c)
Output axis rotation error causes the strapdown gyro to have a drift error propor-
tional to the angular acceleration along its output axis. This error source usually
dictates how the gyros are mounted in the vehicle; the output axes of the gyros can,
for example, be placed along the yaw andlor pitch axes in an aircraft application.
(3) Alignment errors incurred by initializing the transformation matrix using the
accelerometers for leveling information result in tilt errors which produce navigation
errors. (4) In strapdown navigation systems that have external position or velocity-
aiding sensors one can calibrate accelerometer biases in flight, since they produce
observable errors in a strapdown system, but not in a gimballed system.

Computational Errors. Computation-induced errors in strapdown systems

refer to those navigation errors introduced by the incorrect transformation of the
body-measured specific forces to the computational frame. The factors causing the
incorrect transformation matrix include: the algorithm used to generate the trans-
formation matrix, the speed at which the transformation occurs, the roundoff due
to computer word length used, quantization of gyro outputs, truncation due to
integration scheme, and renormalization of the quaternion vector. They all play a
part in contributing to computational errors. Since this type of error also depends
on the vehicle's dynamic environment, most designers evaluate it by computer
simulation [Savage, 1984(a)].
Inertial Naulgation Chap. 3

- s p c ~ f ~cone
c m

anoulor Hlocnv of
thl body wflh RsWct
to ncrtal w e

Altitude ~ n f m t r n

Figure 5-28 (a) Strapdown System Computing in Inertial Coordinates (b) Strap-
down System Computing in Geographic Local-Vertical Coordinates (Rrpritlted with
permisriot~f*om(Schmidt, Sysrmis G- Control E~lryclopedia),C o p y r ~ k (1987),
t Pergamon
Prrss P L C )

Error model. A strapdown error ~ n o d c lis depicted in Figure 5-30.

5.5.5 Operation Flow Diagram

The functional flow of strapdown navigation systems is dividcd into five blocks as
shown in Figure 5-31.
Sec. 5.6 Cornparkon and Analysls of QLmbal versus Strapdown

NavIpcrtlm Conpltath
II lbmmon to strquklvm and Ohlbalkd Syrtrml

Figure 5-29 Block Diagram of Strapdown Inertial Computations (Frorn U o o s a d K r o , q ~ t l ~ ~ t ~ n ,
1981/ wirh p r n n i s s i o t t j o t n ACdRD)



5.6.1 Feature Comparison

Not only is a gimbal system inherently more accurate than a strapdown system,
but calibration is also easier with a gimbal system. In addition, there is less com-
putation involved and less demand placed on the gyro. O n the other hand, a strap-
down system is smaller in size and more reliable than a gimbal system, eliminating
AID conversion of synchro signals as well as electromechanical components. At-
titude and heading data in a strapdown system are in particular most accurate. Fur-
thermore, angular rates are inherently available in body-referenced axes, making
strapdown INS suitable as a flight control sensor. Multiple redundant systems of
strapdown systems are required when used for flight control because in that role
they become critical to flight safety.

5.6.2 Navigation Errors Comparison

This section presents a comparison of navigation errors based on a detailed analysis

appearing in Schmidt [1987]. While the strapdown system errors listed in the dis-
cussion that follows appear in Section 5.5.4 (Instrument Errors), they are again
mentioned here for the purpose. of a direct comparison between strapdown and


C $
Sec. 5.6 Cornparkon and Analysis of Glmbal versus Strapdown

Main Lie
Senror Ibt
Stan & Self-%st Block 1 Ucctronicr %t

Pmcessor %st

Read-In Poaition & Velocity

Level (Self-Alignment)
Initialize Herding Alignment

'Iiansfomtion Matrix
Attitude Update
1 Block3{ {Sensor Uibntion

Gym OP Compensated
Add Lnb Rate & ?hnrpon Rate


Guidance & Control

Figure 5-31 Operation Flow in Strapdown Navigation Systems

gimbal systems. There are significant differences in the navigation performance

between gimbal and strapdown systems because of calibration limitations, sensor
inaccuracies, computational errors, and sensor error propagation effects. The ina-
bility to arbitarily orient the rigidly attached inertial sensors of strapdown systems
to provide known inputs from the earth's gravity and angular velocity vectors for
a stationary vehicle results in calibration limitations. Conversely, a gimbal system
is mechanized so that the inertial component outputs at different angular orientations
can be compared with the known input components of gravity or earth rate and
the sensor errors calibrated in a series of test positions that expose the error sources.
Since only a few of the strapdown sensor errors (compared with a gimbal system)
can be calibrated, the system designer must depend on stability of the instruments
between removals of the system from the vehicle for calibration. The errors that
cannot be calibrated in a strapdown system propagate into navigation errors when
the system begins to navigate. Sensor inaccuracies that arise from motion of the
vehicle are even more difficult to calibrate in a strapdown system. The dynamic
response characteristics of the inertial sensor under consideration for use in the
navigator must be well understood and the limitations on navigation performance
must be evaluated early in the system design.
Gyro-torquing scale factor errors such as asymmetrical scale factor errors in
vibratory environments represent another significant source of strapdown system
212 Inertial Navigation Chap. 5

sensor error. In a local-vertical gimbal system, gyro scale factor errors usually have
a small effect on navigation system performance, since the torquing angular rates
are small, being nearly equivalent to earth rate plus the angular velocity of the vehicle
over the earth's surface. In a strapdown system, the gyros must, in addition, be
torqued for the attitude rate of the vehicle; consequently, gyro scale-factor-induced
navigation errors can be quite large. Misalignments of gyro input axes in strapdown
systems are also important, for similar reasons. Another major sensor error source
in strapdown single-degree-of-freedom gyros is output axis inertia. In this case, the
strapdown gyro has a drift error proportional to the angular acceleration about its
output axis.
Computation-induced errors in strapdown systems are those navigation errors
introduced by the incorrect transformation of the body-measured specific forces to
the computational frame. The transformation algorithm, its iteration rate, the com-
puter word length and the inertial sensor quantization all play a part in determining-
computational errors. In addition, this type of error depends on the vehicle's dy-
namic environment [Savage, 1984(b)].
Sensor-induced navigation errors propagate differently in a strapdown system.
As noted in Schmidt [1987], in a gimbal system during self-alignment, the acceler-
ometer outputs are used to level the stable member and the misleveling of the stable
member is due primarily to accelerometer bias. When the system starts to navigate,
no significant position or velocity errors occur due to this error source, since the
platform tilt-error contribution exactly cancels the accelerometer bias-error contri-
bution. In a strapdown system, where self-alignment occurs by initializing the trans-
formation matrix using the accelerometers for level information, the identical type
of tilt error results. However, when the system starts to navigate and the vehicle
heading changes, the orientation of the inertial sensors changes with respect to the
alignment (or navigation) axcs and the net rcsult is to introduce a step of acceleration
error, equal to the value of the accelerometer bias if the heading changes by 90 dcg.
Subsequently, velocity and position errors arise both from accelerometer bias and
from the initial tilt error that bias introduced during alignment.
These error effects in strapdown navigation systems that have external position
or velocity-aiding sensors means that one can calibrate accelerometer bias in flight,
since it produces an observable error in a strapdown system. Detailed computer-
based evaluation of performance, including algorithm errors, in a strapdown systcm
design requires the use of a whole-number type of simulator rather than a statistical
(or covariance) approach. This simulator can also evaluate the effects of nonlinear
sensor errors and quantization errors. Such computer programs are used widcly in
the inertial systcnl manufacturing industry. Generally, designers will use the rule
of thumb that for gimbal systems 0.0l0h-' of gyro drift will produce a position
error growth rate of 1.85 km h-' and a velocity error of approximately 1 ms-' for
a terrestrial navigation system; for a strapdown system, the errors would be 20
percent to 50 percent worse. However, as mentioned in Schmidt [1987], thc dif-
ference in errors can be small, depending on the vchiclc environ~nentand thc use
of other aiding sensors (Doppler velocity, position-fixing radar, satcllitc navigatiol1
Sec. 5.7 External Navlgatfon AIds 213

receivers, and so on). Consequently, the process of preliminary system selection

involves the choice not only of system conlponcnts and mechanization but also of
aiding devices and frequency of updates. These trade-offs are usually conducted
with a general-purpose covariance analysis program [Schmidt, 19781.


An unaided INS has the advantage of possessing all-wcather, all-attitude, and non-
radiating characteristics; high-dynamic response; and self-contained navigation. Dis-
advantages include unbounded position errors and oscillatory velocity errors, ab-
sence of in-flight alignment, long reaction from cold start, and poor recovery from
power interruption. Among the disadvantages, unbounded position error and os-
cillatory velocity error are of most concern to the designer. Thus, it is often necessary
for navigation systems to be aided by some other external navigation sensors.

5.7.1 Aided Inertial Navigation Mechanization

Referring to the nine-state error propagation equation shown in Figure 5-10, for
example, there are generally three types of aiding available to inertial navigation
systems. These are position aiding, velocity aiding, and attitude aiding. Each of
these augmentations
. directly observe one or more of the error states in the INS. In
addition, various combinations of these three aiding techniques may be used.
GPS, T E R C O M , LORAN, and T A C A N are position-aiding devices. Doppler
radar and GPS can both provide velocity updates. Another very important velocity
update is available on stationary vehicles, namely a zero-velocity update. Stellar
observations made by a telescope mounted directly on the stable element of an INS
provide attitude updates, actually psi-angle updates. Other examples of attitude
updating are stored heading alignment and attitude matching transfer alignment
from a master INS to a slave INS.

Effect of augmented inertial mechanization

Error Improvement. From Figure 5-32, it is seen that the augmented inertial
mechanization may keep acceleration, velocity, and position errors in a fixed region.
Figures 5-32a through 5-32c represent self-contained navigation systems. Among
them, the first figure represents the worst case involving free-inertial navigation,
and the last figure represents the best case involving position navigation. However,
augmented inertial mechanization as shown in Figure 5-32d, provides acceleration,
velocity, and position errors which are much better than those of the three self-
contained systems.
Cost and Performance Improvement. From Figure 5-33, it is seen that, in
the region of drift rate O.OOlO/hr,and 0.02"/hr, the position- or velocity-aided inertial
214 Inertial Navigation Chap. 5

TIME --+
Figure 5-32 Augmented Inertial Rationale (From [Martin, 19861 with permission from Litton)


COST 100

0.001~1~~ 0.01~1~~ O.l0lH~
L0 ' 1 ~ ~
0. lMlHR LO MllHR 10 MllHR
Figure 5-33 Typical Cost vs. Performance: Unaided lnertial 8: Augn~entedInertial
Navigation (From [Martin, 19861 with prrmissiot~,kom Littoti)
Sec. 5.7 btemal Naulgation Aids 215

mechanization has a lower cost than the unaided one. This is not true for stellar-
aided systems which cost considerably more than an unaided INS.
Advantage of Augmented Position-Velocity Inertial. The advantages of aug-
mented position-velocity inertial mechanization include: multibackup modes; pro-
tection against spoofing o r position; rate-aid for position sensor; good dynamic
response; in-flight startup and alignment; and bounded position, velocity, and head-
ing errors. When an augmented inertial mechanization is no longer self-contained.
that is, it requires external cooperation, its dominant disadvantage is its susceptibility
to RF noise and propagation errors.

5.7.2 Global Positioning System (GPS)

Among the current generation of positioning and navigation aids, it is perhaps GPS
coupled with newer and better-designed field-qualified receivers that provide un-
precedented positioning and navigation capability.

Requirements. The objective of a GPS is to develop a worldwide precision

navigation system based on satellites so that it can be applied to all classes of military
and possibly commercial operations with an accuracy of 25 ft in three dimensions.
In conjunction with other systems, a GPS, with its precise position, velocity, and
time is very versatile and is applicable to a variety of air, land, and sea services and
operations. GPS will be an integral part of future navigation systems. The appli-
cations of a GPS include: time transfer; range instrumentation; geodesy and survey;
antisubmarine warfare; launch vehicle guidance improvement; mine and sensor de-
livery; missile inertial system updating; field artillery and shore bombardment; close
support and common grid coordination; riverine and small-craft operations; land
vehicle navigation; enroute nautical and aeronautical navigation; VTOL, STOL, and
helicopter takeoff, landing and cruise operations; and photomapping, phototargeting
coordinate bombing.

GPS overview. GPS is a space-based radionavigation system which is func-

tionally divided into the space, control, and user segments. The space segment
consists of a constellation of navigation signal time and range (NAVSTAR) satellites.
A full constellation will have 21 satellites plus 3 active spares. They will be uniformly
distributed in 6 orbit planes, providing 4 to 7 visible satellites at any time anywhere
on earth. The planes are inclined 55 deg with respect to the equatorial plane. The
orbital altitude is 10,890 nautical mi. NAVSTAR measures range to a set of four
satellites in 12-hour orbits, by timing the arrival of radio signals transmitted from
the satellites at precisely known times. Theoretically, a minimum of only three
satellites would allow a fix to be obtained but since three satellites may not always
be in suitable positions, and because timing errors in the receiving system have to
be eliminated, a fourth satellite is necessary [Barnes, 19871. Each satellite trarismits
specially coded signals that allow individual satellites to be distinguished, and the
216 Inertial Navigation Chap. 5

range and rate of range change to the user to be measured. The signals are pseudo-
random binary noise codes (PRN). T w o different codes are transmitted in phase
cluadrature, providing a standard positioning service (SPS) and a precise positioning
service (PPS). A low-rate data message is also transmitted. T h e control segment
has five monitor stations that track all satellites in view of their antennas. Data are
transmitted to a master control station where processing takes place to determine
orbital and clock modeling parameters for each satellite. The information is then
uploaded to the satellites by one of three upload stations. The satellites incorporate
this information into the data message.
The user segment consists of equipment designed to receive and process the
satellite signals. The unique codes transmitted by each satellite allow the use of
common RF carrier frequencies throughout the constellation, a process known as
code division multiplexing. Measurements from four satellites are required in the
general case. The United States Department of Defense has developed two classes
of receiver, continuous tracking and sequential tracking. The continuous tracking
receiver provides a dedicated channel for each satellite being tracked and a fifth
channel that performs ancillary functions (for example, ionosphere correction, in-
terchannel bias measurement). The sequential receiver has one or two channels, for
low- and medium-dynamic applications, respectively. In these receivers the channels
are time shared among satellites and housekeeping chores [McGowan, 19871.

Measurement principles. The following discussions on measurement

principles are based on McGowan [I9871 as follows.
Range Measurement. The PRN code in each satellite is reset to its initial
state precisely at midnight, Saturday (Universal Coordinated Time). The codes have
sufficient period so that no state will be repeated bctween the weekly resets. The
receivers generate a replica of the code and phaselock on it using correlation tcch-
niques. The receivcd code phase provides a dircct indication of the time of trans-
mission of the code and allows the transit time of the signal to be calculated against
a local clock. Scaling this measurement by the speed of light (c) determines the range
betwecn the user and satellite antennas. This quantity is known as pseudo-range
(PR) because it contains errors. T w o dominant error sources are the error in the
local clock (its quality is at least an order of magnitude worse than that of the satellite
clock) and atmospheric dclays suffcred by the signal.
T i m e Measurement. The mcasurcn~cntof transit time by the rcceivcr con-
tains thc bias bctwecn the satellite and uscr clocks. Since it affccts all mcasurcmcnts
cqually, it can bc cstimatcd as part of thc solution. Whcrc three 111easurcmcnts arc
nccdcd to dctcrminc thrcc position coordinatcs, a fourth allows clock bias to bc
Range Rate Measurement. During pcriods of sufficient rcccived signal qual-
ity (CINo > 29 dB), thc rcccivcr can phaselock onto the carricr. Thc Doppler shift
observed 1s uscd to dctcrm~ncthe linc-of-sight (LOS) velocity betwccn the satcllltc
and uscr. This mcasurcn~cntis implcmcnted as an integrated Doppler known as the
Sec. 5.7 External Naulgatlon AIds 217

pseudo-delta-range (I'DR), which is a measure o f the range change during the in-
tegration interval.

Performance. Table 5-1 lists the error budget for the PR measurement,
assuming that the dual-frequency ionospheric co~npensationtechnique is used. As
shown, a measurement error of 5 m is predicted. The test results indicate that the
high-frequency fluctuation of the error is less than 1 m. The error was dominated
by a bias component that ranged from 7-10 m [McGowan, 19871.

INS updating. Figure 5-34 illustrates a block diagram of a typical imple-

mentation using the GPS system as an update aid for an inertial system. Note that
the filter estimates receiver errors as well as inertial navigation errors. However,
realistic modeling of all error sources usually results in a state vector dimension that
prohibits full implementation in the vehicle computer. Trade-offs to determine the
appropriate suboptimal filter to implement are then required [Schmidt, 1976; Set-
terlund, 19861. From these performance trade-off analyses, a set of system perfor-
mance specifications can be developed prior to implementation and integration
[Schmidt, 19871.

5.7.3 Tactical Air Navigation (TACAN)

T A C A N is a two-dimensional navigation system used for piloted aircraft and is

based on generating azimuth and distance to a fixed ground station for the aircraft.
The mechanization of TACAN centers around an airborne-interrogator and ground
transponder. The flight vehicle is provided with range and bearing lines of position


System budget (1)
Segment Error mechanism meters

Space Segment Clock Stability

Track Perturbations
Segment RSS
Control Segment Ephermeris Prediction and Modeling
Segment RSS
User Segment Ionospheric Delay
Tropospheric Delay
Receiver Noise/Resolution
Multipath Interference
Segment RSS
System RSS

From [McGowan, 19871 with permission from AGARD

Inertial Navlgatlon Chap. 3

vorlo ICS ~ovigationcorrections

and pseudo
delta range
Figure 5-34 Aided Inertial-GPS Mechanization (Reprinted with permission jot11
(Schmidt, Systems fi Control Enrydopedio), Copyrishr (1987), Pergemort Press P L C )

by a ground transponder which uses a nearly pulse type of transmission for range1
bearing loci measurements. It operates on a frequency in the L-band (1.0-GHz)
range, and has an operational range of 150 N M (LOS) at an altitude of 15,000 ft.
It has an accuracy that is on the order of 200-1,500 ft in range/(bounded) radial
position error and 0.5 deg-1.0 deg (100 ft/NM) in bearing/(bounded) angular po-
sition error. The resolution is approximately 100 ft in range and 0.5 deg in heading.
It is the bandwidth that limits the range resolution to 100 ft, while the L-band
frequency and the beamwidth of the transponder antenna limits the bearing reso-
lution. In Figure 5-35a, it is seen that the angular position error is 10,000 ft at 100
NM, corresponding to the bearing accuracy mentioned earlier. O n e way to over-
come the large angular position error is to go to a dual-TACAN mechanization.
By implementing a ~ U ~ ~ - T A CsystemA N or range-range mechanization, the po-
sition error is reduced to 500 ft (CEP), as illustrated in Figure 5-35b [Martin, 19861.

TACAN advantages and limitations. Advantages of T A C A N include

excellent enroutelhoming navigation aid, and accurate range and bearing over short
distances. Furthermore, it is amenable to fixed and mobile (that is, ships) ground
stations. One of the limitations of TACAN is that it is not a navigation system and
thus requires a heading device. In addition, it is not sclf-contained because it requires
a radiating ground station. At long ranges, its cross-track errors are large. There
are also limitations on its LOS, which require aircraft to fly high in mountainous
regions and for long-range measurements. There is a disadvantage in the fact that
radiation from the aircraft can be detected during military OPS. T A C A N is also
subject to natural and enemy environments as well as to ground maintenance ell-
Sez. 5.7 External Navigation Aids 218





Figure 5-35 (a) TACAN Characteristics (b) Dual TACAN Capability Reduces
Large TACAN Cross Track Position Errors (From [Martin, 19861 with permission
from Litton)
220 Inertial Navigation Chap. 5

vironment, Due to the earth's curvature, as the range to the T A C A N station in,
creases, so must the altitude of the flight vehicle increase. A relationship between
the TACAN range and the altitude of the flight vehicle is given in Figure 5-36.

5.7.4 Long-Range Navigation (LORAN)

LORAN is a time-difference of arrival system that requires one master and at least
two slave ground stations that transmit radiation to the flight vehicle. Like TACAN
it is a two-dimensional system. The transmission of the ground station is approx-
imately in the form of 8 or 16 pulsed (groups) at 10 KHz PRF, entailing measure-
ments of hyperbolic loci of position provided by pulse envelope and phase difference.
It has an operating frequency of nearly 100 KHz and an accuracy that is on the order
of 50-500 ft or more, contingent upon the geometric dilution of precision (GDOP).
The resolution of the transmission is approximately 0.01 psec X GDOP.

LOBAR characteristics. LORAN characteristics are depicted in Figure

5-37a, which include an operational range of 600-1,000 N M on land and 1,000-
1,500 N M over water and a radial position error of approximately 300-500 ft in
the primary zone. The radial position error is the region of interception of the
M-S#l line of position (LOP) (see Figure 5-37b), and the M-S#2 LOP as given
in Figure 5-37a. M-S#1 and M-S#2 LOP are hyperbolic curves between M-S#l
and M-S#2, respectively. The hyperbolic LOP is constructed based on the receiving
time (that is, range) difference measurements between the master station and the
slave station.

LORAN advantages and limitations. Advantages of LORAN include

simplicity and reliability of airborne units, accurate medium-range navigation aid,
and nonradiation of passive receiver. There are some limitations and disadvantages
associated with LORAN, one of which is that radiating ground stations are required
for LORAN. In addition. LORAN, like TACAN, is subiect " to natural and enemy
environments as well as to ground maintenance environments. Also, like TACAN,
LORAN is not a full navigation system and needs to be augmented with heading,
velocity, and altitude. Its performance is reduced by the GDOP. The long smoothing
circuits of LORAN require low dynamics during measurement or rate aid to main-
tain accuracy/track.

5.7.5 Terrain Contour Matching (TERCOM)

TERCOM is a form of correlation guidance based on a comparison between the

measured and the prestored features ofthe profile of the ground (terrain) over which
a missile or aircraft is flying. Generally, terrain height forms the basis of this com-
parison. Obtaining the reference data requires prior measurement of the ground
contours of interest. This type of guidance is used for updating a midcourse inertial
guidance system on a periodic basis, and has been applicd to the guidance of cruise
222 Inertial Naulgatlon Chap. 5

missiles, which usually fly at subsonic speeds and fairly constant altitude [Heaston
and Smoots, 19831.
A TERCOM system uses an airborne altimeter and a data processor to correlate
the measured terrain contours with the prestored contours to obtain the best estimate
of position. The transmission characteristics of the airborne altimeter include an
operating frequency of approximately 4.4 GHz (incidental to operation) and a trans-
mission type that is pulsed or CW. As the missile flies, the radar altimeter first
measures the variations in the ground's
- ~rofile.These measured variations are then
digitized and processed for input to a correlator for comparison with stored data.
High-density digital storage permits large quantities of data to be stored, and modern
microprocessors allow the comparison to be done sufficiently and quickly [Barnes,
19871. Through this process, the missile can determine its position and correct any
errors that have developed since the previous update. Accuracy is 100-1,000 ft and
resolution is 100-200 ft. Figure 5-38 illustrates the basic concept employed by
TERCOM. The terminal guidance stage may bebased on the final TERCOM update
and a preprogrammed course relying on the inertial system, or a separate terminal
homing seeker may be employed that can recognize the target and provide the final
guidance commands [Heaston and Smoots, 19831. The block diagram shown in
Figure 5-39 shows the mechanization of the TERCOM system.


-300 - 500 FT (PRIMARY ZONE)


Figure 5-37 (a) LORAN Characteristics (b) LORAN Lines o f Position ( a ) j o m

[Martin, 19861 with permission j o m Liftot~( b ) courfesy 0jF.U'. Hardy, 1986)
Sec. 5.7 external Navlgatlon Alds

Figure 5-37 (Continued)

TERCOM advantages and limitations. The advantages of TERCOM

include total automation of its operation, accurate position location with a good
heading device, and self-contained operation requiring no external support. The
224 Inertial Navigation Chap. 5

Cruise Cruise
Missile Elevation Map
Stored In Onboard

Terrain Terrain Mapped







Figure 5-38 (a) Terrain Corltour

Matching (b) T E R C O M ~haracteristics
((aJjom /Heaston and Smoors, 198.71 lr'irh
prrmissiot~j o m GACIAC (b)jottr /hlnr.ritlt
19861 with prrmission j o m Lirrorr)
Sez. 5.7 External Navigation Aids

Data Recommended
Cornlation Flight Path AutopiSol
h I
Figure 5-39 TERCOM Mechanization

limitations of TERCOM are as follows. TERCOM requires varying-altitudc tcrrain

and a radiating altimeter. Moreover, thc prestored maps must be obtained, gen-
erated, and programmed in advance. In order for the updating tcchnique described
earlier to perform properly, there must be sufficient variation in the terrain to gen-
erate a usable signal. Obviously, such a system will not work over water. It may
have difficulty in the midwestern region of the United States where the ground is
very flat. Therefore, consideration must be given to the terrain surrounding the
target when employing this type of guidance. The data collected for the reference
map will provide the information necessary to predetermine if the system will work
in the areas of interest [Heaston and Smoots, 19831.

5.7.6 Doppler Radar

Doppler radar on an aircraft provides a measure of the aircraft's velocity relative to

the earth. It usually consists of four beams, one forward, one aft, one to port, and
one to starboard. They are all angled downward at a 45-deg to 60-deg angle. The
fore and aft beams measure velocity along the track heading of the aircraft, while
the left and right beams measure cross-heading velocity. All four beams measure
vertical velocity. The Doppler set measures velocity in aircraft axes. The INS, or
sometimes an AHRS, provides the reference coordinate system for resolving the
Doppler velocity components onto along-track and cross-track axes or onto earth-
fixed axes.

Doppler advantages and limitations. Doppler radar has the advantage

of being autonomous; it does not require any ground-based or space-based facilities.
Because it does not have its own attitude reference, it must always be used as an
adjunct to an INS or an AHRS. The Doppler return signal over water is poor due
to more specular and less diffuse reflection off the surface. In addition, the scattering
cells on the water surface may be moving due to currents or wind-blown spray.
For this reason Doppler radars usually have two modes, one for land and one for
over water. This landlsea switch can be used to reduce the amount of feedback gains
to the INS during over-water operations.

5.7.7 Star Trackers

Astro-inertial systems consist of a telescope mounted in two gimbals, azimuth1

elevation, which in turn are mounted directly onto an inertial system. The systems
in operational use today are integrated into a !gimballed inertial system. This means
226 Inertial Navigation Chap. 5

that the whole system must include at least five gimbals: the pitch, roll, and azimuth
gimbals of the inertial system plus the azimuth and elevation gimbals of the stellar
telescope. Consequently, these systems tend to be quite large and expensive. Con-
ceptually, the gimballed inertial system could be replaced by a strapdown inertial.
thereby reducing the size of the overall system. T o date, however, this development
is still in the early stages of design.
The basic idea of the astro-inertial system is as follows. Given the position of
the aircraft from the inertial system, the computer looks up the ephemeris data for
a particular star from the star catalog contained in its memory, and calculates the
azimuth and elevation angles for that star at that time. It then drives the telescope
to those angles and looks for the star. If the star is off the axis of the telescope, the
star tracker provides correction signals to center the star on the telescope axis. These
same azimuth and elevation correction signals go the inertial system as psi-angle
corrections (see Figure 5-10).
One such star fix can update two attitude error components. For a full three-
axis update, the telescope is driven to another star, widely separated in angle from
the first, and the process is repeated.

Star tracker advantages and limitations. Like inertial and Doppler

systems, the star tracker is autonomous, requiring no outside facilities, except of
course for the stars whlch are always there. Cloud cover is an obvious limitation
for star trackers, and this relegates their usefulness to high-altitude flight. Because
it is a psi-angle measuring device, it basically updates the gyros and not the accelero-
meters. (Recall that the psi equations are independent of the acceleration equations.)
This means that acceleration errors can grow without bound. The astro-inertial
system will simply apply a tilt error to compensate for the position error; it cannot
distinguish between the two. However, with an occasional position fix or with
Doppler radar updating, the integrated system can produce a navigation system
with bounded errors over very long periods of time.

5.7.8 Kalman Filtering

In order to combine the INS data with one or more of the foregoing augmentations,
most modern avionics systems use a Kalman filter (KF). The KF is a linear feedback
system with time-varying gains. The gains are varied in an optimal or nearly optilnal
fashion so as to take account of the relative accuracy of the INS and the updating
measurements as well as the geometry of the measurements. Because the accuracy
of the INS varies with time, the KF must include an error model of the INS. This
error model includes a number of error states, the dynamic coupling betwccn these
states, and the noise which drives these states. A similar but simpler error model
for the measurement noise allows the KF to properly weight the feedback gains to
the INS.
The feedback corrections themselves may be applied in two different ways,
One is called open-loop correction, and the other is called closed-loop correctioll,
Sec. 5.7 External Naulgatlon Aids 227

In the first, the corrcctions are added to the INS outputs and do not affect the
operation of the INS itself. Figure 5-40 illustrates such a mechanization. The ad-
vantage of this type of mechanization is that spurious measurements affect only the
corrected INS outputs and not the INS itself. If the KF diverges as a result o f these
spurious updates, the INS is still operational, and the system can recover by rein-
itializing the KF. The disadvantage of the open-loop KF is that the state errors in
the error model arc not driven to zcro and can become large enough for nonlinearity
effects to disrupt the system operation.
In thc closed-loop feedback system, the feedback corrcctions actually correct
the INS itself so as to maintain the INS state crrors near zcro. This eliminates the
nonlinearity concerns. O n the other hand, spurious updates can drive the INS so
far off that it may not be able to recover.

Kalman filtering approach. A standard Kalman filter minimizes the

state-vector variance and estimates the state errors. The Kalrnan filter is constructed
based on an error model, as mentioned previously. First, an INS model and a mea-
surement model arc given as
jc = .4.u + lu (INS) E[ru(t)] = 0, E[~(~)Iu(T = ) Q6(t
] - 7)
Z = Hs + v (measurement) E[v(t)] = 0, E[v(I)v(T)] = R6(t - T)





Figure 5-40 Generalized Mechanization of Filtering Open-Loop Error Control

System (From ['Martin, 19861 with permissionfrom Lirton)
228 Inertial Navigation Chap. 5

where x is the error state with X E R ~ ,and w and v are white noises. The measurement
z can be a combination of attitude, velocity, o r position measurements. If position
measurement is available, position error 6, must be included in the state vector x,
An example application of error-covariance analysis to inertial platform errors
is constructed as illustrated in Figure 5-41. As presented in Chapter 4 (Sections 4.1
and 4.2), the fiow diagram in Figure 5-41 [Tucker and Stern, 19861 can be con..
structed, with v = a as the input. From this figure, it is seen that there are nine
error state variables: three position errors, one along each of the coordinate axes;
three corresponding velocity errors; and three misalignment errors, one about each
of the coordinate axes. The velocity errors are obtained by integrating the accel-
eration error inputs, and the position errors are obtained by a second integration.
In addition, the misalignment error about any axis leads to an additional acceleration
input along each of the other two axes. The resulting dynamics matrix (A) from
Figure 5-41 is as follows: -

0 1 0 0 0 0 0 0 0
-k? 0 0 0 0 0 0 a28 a29
0 0 0 1 0 0 0 0 0
0 0 k 0 0 0 a47 0 049
0 0 0 0 0 1 0 0 0
0 0 0 0 kz 0 aj7 ass 0
0 0 0 0 0 0 0 0 0
0 0 0 0 0 0 0 0 0
0 0 0 0 0 0 0 0 0
Examination of the A matrix and the flov,. diagram in Figurc 5-41, howcver, in-
dicates that the systcm can bc decoupled into three indcpcndcnt subsystems, cach
of dimension four. For example, states 1, 2, 8, and 9 form a subsystem. It IS noted
that cach of thc misalignment error states 7, 8, and 9 feed into two of the three
subsystems, but this does not cause any direct coupling among the subsystcms, since
there are no feedback inputs into the misalignrncnt states. T h e four-dimensional
problem for states 1, 2, 8, and 9 is of the form

5.7.9 Kalman Filtering Performance

For conventional filtering, the position error (in ft) ncvcr converges to zcro, as shown
in Figure 5-42a. For Kalman filtcring (optimal filtcring), on the other hand, the
position error (in ft) convcrgcs to, but ncvcr rcachcs, zcro, as shown in Figure
5-4%. Figure 5-42c displays a table showing thc relationship of mechanization and
Sec 5.8 Integrated Inertial Navigation System (IINS)

9 x Position n2 - x velocity n7 ' misalignment about x-axis

n3 = y Position n4 = y velocity nB = misalignment about y-axis
n5 = z position n6 = z velocity ng - misalignment about z-axis

Figure 5-41 Flow Diagram (From [Tucker and Stern, 19861 with permission from


When inertial systems were first included as part of the avionics package on aircraft
circa 1960, they performed only the navigation function. Typically, there would
also be an AHRS for driving the flight instruments, and there would be rate gyros
230 Inertial Naulgatlon Chap. 3


(A) (B)


Figure 5-42 (a) & (b) Integrated Performance (c) Augme~~red Navigation Per-
formance Summary (From /Manin, 19861 ulith pcrmir~ionJiom Lirrotr)
Sec. 5.8 Integrated Inertlal Naulgatlon System (IINS) 231

for flight control sensors. Gradually, as the INS became more reliable they began
to take over the AHRS functions. Then as the strapdown INS came into being,
their use for flight control became a real possibility.
In order for the INS to be part of the flight control system (FCS), it must be
made ultra reliable. This leads to the design of multiplcly redundant INS. Such
systems, properly updatcd, can providc reliable and accurate position, velocity, and
attitudc for the navigation function, as well as body-axis accelerations and attitudc
rates for the FCS. Such systems are called intcgratcd inertial navigation systems
(IINS). Only strapdown INS arc candidates for IINS, because the gimballcd INS
cannot providc the attitude rates for the FCS with a wide enough bandwith.

5.8.1 Integrated SensinglFIight Control Reference

System (ISFCRS)

One way to improve the reliability of the INS is to employ redundant sensors, as
in the case of an integrated sensinglflight control reference system (ISFCRS). The
ISFCRS uses six ring laser gyros and accelerometers in an arrangement of two
orthogonal triads, onc of which has its axes aligned with the flight vehicle's body
axis. The geometry is depicted in Figure 5-43.
The sensors are numbered from 1 to 6 with 1, 2, and 3 corresponding to the
x-, y-, and z-axes of the body aligned unit, and 4, 5, and 6 corresponding to the
x-, y-, and z-axes of the skewed unit. A functional block diagram of an ISFCRS is
presented in Figure 5-44. The ISFCRS contains signal selection, lever arm com-
pensation, and parity equations. The ISFCRS was proposed for the MFCRS and
MIRA systems that provide all required incrtial reference and air data in advanced
transport and fighter aircraft for flight control, navigation, weapon delivery and
targeting, terminal area functions, automatic TFITA, sensorltracker stabilization,
flight instruments, and displays, as illustrated in Figure 5-45. Figure 5-46 shows
a conventional approach of comparison with MIRA [Boggess, 1986; Young et al.,

5.8.2 Integrated Sensory Subsystem (ISS)

Detailed discussions of the integrated sensory subsystem (ISS) presented in this

example can be found in Toolan and Grobert [1983]. The ISS is composed of three
major elements. The first element consists of the various sensors/sensory data and
preprocessing electronics. The second element is a redundant I 1 0 system which
provides data transfer between the sensors/transducers and redundant computers.
The third ISS element is the data handling system (DHS) contained within the
redundant digital computers in which redundancy data management and output
parameter calculations are performed. The ISS sensor set consists of redundant
skewed arrays of strapdown integrating rate gyros and accelerometers, redundant
air data probes and transducers, magnetic heading reference sensors, radio aids, and
command inputlsurface position transducers. A system-level block diagram of the
232 Ineriial Navigation Chap. 5

Figure 5-43 ISFCRS Geometry

ISS DHS is shown in Figure 3-47a. The skewed and dispersed gyros and acceler-
ometers supply raw input data to the DHS. The DHS removes the vehicle bending
and kinematic effects and provides orthogonal rate and acceleration data to the
SAHRS (strapdown AHRS) algorithms, the FCS, and the user systems. The six
gyros and six accelerometers are configured in a symmetrical conical array and
provide a 2-fail operational sensor capability. In addition, the sensors are dispersed
to assure a survivable system. The ISS design philosophy for inertial sensor pack-
aging and dispersion for survivability permits installation o f gyros and accclero-
meters in less than ideal locations with respect to bending nodeslantinodes and the
aircraft center of gravity. Without appropriate compensation generated by the use
o f a state estimator, the sensor locations can result in FCSIbending-mode coupling
andlor nuisance trips of the RDM failure detection routines. The configuration of
the gyro and accclerometer RDM routines and the state estimator is shown in Figure

5.8.3 Integrated Inertial Sensing Assembly (IISA)

Another type o f IINS is the IISA. This is an integrated inertial sensing assct1lbJ~
(IISA) that uses six ring-laser gyros and six inertial-grade accelcromcters in two,
separated clustcrs. It is a system that is well suited to the combined navigatioll and
flight control sensing needs of modern high-performance fly-by-wire vchiclcs. The
background of integrated navigationlflight control sensors is given by Ebncr and
Wei [1984]. Krasnjanski and Ebncr [1983], and Jar~kovitzct al. 119871 includillg a
spectrum of possible system configurations as a function of avionics redundancy
234 Inertial Navigation Chap. 5



Right Control
Fire Control
Weapon Delivery & Trajectory
Sensorhcker Stabilization
NAV Axis
I Cockpit Display
L --- I
Figure 5-45 MlRA System

requirements. Within each inertial navigation assembly (INA), sensor axes are or-
thogonal but skewed relative to the vehicle yaw axis, as shown in Figure 5-482,
One accelerometer and one gyro in an INA are oriented along each skewed axis.
Each of the three channels is largely independent, as shown in Figure 5-48b. Figure
5-48a also depicts the orientation of axes when the INA is installed into the left
equipment bay of the vehicle. When an identical INA is installed into the right
equipment bay, with 180-deg rotation about yaw relative to the left INA, the six
sensor axes are then distributed uniformly about a 54.7-deg half-angle cone. No
two axes are coincident, nor are three in the same plane. Thus, any three sensors
may be used to derive three-axis outputs in vehicle axes after suitable computer
transformation. The IISA development uses high-reliability laser gyro and acceler-
ometer sensors packaged in a strapdown system configuration to provide a common,
efficient source of aircraft body rates, attitude, and accelerations. These measure-
ments provide the essential inertial data inputs for all core and mission avionic
functions including stability control and stability augmentation system (CSAS). The
IISA (Figure 5-48c) consists of five assemblies: two identical INAs containing the
inertial sensors and navigation computers (shown in Figure 5-48d); two identical
digital computer assemblies (DCAs) each containing dual redundant flight control
redundancy management and sensor selection logic computers; and a multifunction

L User

k Air Data
Figure 5-46 Conventional ~ p p r o a c ~
to Sensing and Flight Control Referellce















I 1



Figure 5-47 (a) Inertial System Block Diagram (b) Inertial Sensor RDM (From [Toolan and
Groberf, 19831, 0 1983 IEEE)
Inertial Navigation Chap. 5

control display unit (CDU) for displaying IISA data and providing the operator
interface for initializatior~and mode selection.

Configuration of sensor locations. Table 5-2 is a comparison of con-

figuration and attitude/velocity redundancy for the IISA. T h e four different system
configurations with variable levels of attitude/velocity redundancy are described as
follows [Ebncr and Wci, 19841. (1) Configuration 1 assumes that neither attitude
nor velocity is required for flight safety. The INS can be nonskewed relative to
aircraft pitchlrolllyaw axes for minimum size. A companion unit would be required
Sec 5.8 Integrated Inertial Naulgatfon System (IINS)


-- N A V TO * I C

- --
3 RLOS 1014 Hx
rcro AIC(MH.,

3 ACCtU d,
PRoClssoR =- +-(I

PC T O A I C (80 Ha)



-- - -,I

- rc TO AIC (80 HZ)

3 PC I L I C I R O N I C S

c rcc TO AIC POHZ)




Figure 5-48 (a) IISA Sensor Orientation (b) INA Functional Block Diagram (c) IISA Func-
tional Block Diagram (d) INA Installation Configuration (a) G. ( c ) j o m [Krasnjanski and Ebner,
19831, O 1983 IEEE (b) G. ( d ) j o t n uankovitz yr al., 19871 with permissionjotn AGARD)
238 Inertial Navigation Chap, 5


redundancy Unit 1 Unit 2
1 None None INS (nonskew) Sensors only (skew)
2 Dual None INS (nonskew) SD AHRS (skew)
3 Dual Dual INS (skew)
4 Quad

From [Ebner and Wei, 19841, O 1984 AlAA

Quad INS (3NS+lS) INS (3 N S + IS)
containing three gyros and three accelerometers, skewed relative to the INS axes
s o that n o t w o are coincident and three are not in the same plane. (2) Configuration
2 would be virtually identical to configuration 1 except that a computer is added to
unit 2 to solve the strapdown equations in order t o derive the aircraft attitude and
heading. (3) Configuration 3 consists basically o f t w o identical INS (skew) to provide
the required fail-operationlfail-operationifail-safe (FO-FO-FS) angular rate and ac-
celeration outputs, that is, dual attitude and navigation outputs. (4) FO-FO-FS nav-
igation can be achieved by adding one gyro and one accelerometer to each INS.
Since the four sensors within an INS are bolted together rigidly, with proper ge-
ometry good navigation can be achieved using any three of thc four.
Among the four configurations, configuration 3 is commonly used for re-
dundancy. For this configuration, thc sensors are contained in t w o INA, each of
which provides full, independent incrtial navigation outputs.

Effect of gyro dither. Basically, IISA utilizes ring laser gyros that dcpcnd
on a small amount of angular mechanical dither motion to avoid lock-in between
C W and C C W lascrs to achieve full navigation accuracy. This mcchanical dither
produces accelerations sensed by the acccleronlcters. If nonlinearitics occur some-
where in the proccss, diffcrent frcqucncics bctwectl gyro dithers o r aliasing with
the sampling frequcncy can causc low-frequency beats to occur on acccleration out-
puts that could enter the FCS and causc wander or actuator flutter. The acccleration
output from thc actual operational IlSA hardware with gyro dithcring is random
with a noise arnplitudc of 0.05 ftlsec' (less than 2 milli - g) [Ebner and Wei, 19841.

Redundancy management. Redundancy managcmcnt providcs FO-

FO-FS opcration through the usc of parity cquations which lincarly conlbinc the
sensor outputs to dctcrminc thc residual crrors. Thc opcration scqucncc of rcdun-
dancy managcmcnt is givcn in Figure 5-49a. Thc six axcs of skcwcd angular rate
and accclcration arc sent to an cxtcrnal computcr for redundancy managcmcnt pro-

~ -
Figure 5-49 (a) Itcdundancy Management Operation (b) Trajrrtory Flow Chart
and System Response Sinlulator (c) I'arity Flow Chart and llesign Equation Sinl-
. f i ~ n A C A R D ( h ) G (0fro111
ulator (a) .font Oankovirz rr a/.. 1Y87/ luirh prn~~issiotl
/Ebttrr and Wri, 19841, 0 I984 A I A A )


o I/O A V A I L U I L R I
,o D Y X L X C E I 0 N A B L I : I E S S

o i s taurrrons
o ISOLATE 2 S x n a t A N E o u s FAILORES

+I 1


1 INA-1

u I



? I l



240 Inertial Navigation Chap.

cessing to derive body-axis rates and acceleration, free from the effects of any tx~,
hard or soft sensor failures. The redundancy management consists of parity equa-
tions, sensor selection logic, and design equations. The parity equations ensure
detection of up to three sensor failures by comparison of redundant sensor data.
Sensor errors above a predetermined level (threshold) are defined as failures. Because
of information limitations, a third sensor failure of the same type can be detected
but not compensated. Sensor selection logic considers the state of all the parity
equations and determines which sensors are to be used to derive angular rate and
acceleration outputs. Based on selected sensors, design equations are used in Ebner
and Wei [I9841 to derive the required outputs, removing skew angles and accounting
for redundant data where applicable.
Citing Jankovitz et al. [1987], realistic simulations have been performed, to
evaluate the effects of factors such as vibration isolators, antialiasing filters, and
misalignments on the redundancy management process. The redundancy manage-
ment simulator consists of the following three programs: trajectory simulator, sys-
tem response simulation, and parity and design equation simulator. The trajectory
and system response simulator for trajectory and system response simulation are
illustrated in Figure 5-49b. The system dynamic model in this figure is based on
the sensor assembly model that results from rigid body dynamics, including effects
of vibration isolators and gyro dither mechanisms. The input excitation is obtained
from the trajectory sin~ulator[Ebner and Wei, 19841. The parity and design equation
simulator is flowcharted in Figure 5-49c.

IlSA performance. Following Jankovitz et al. [1987], the performance

requiremcnts arc:
Radial position error rate 1 nmiihr (1.852 kmlhr) (CEP)
Velocity errors, per axis 3 ftlsec (91.44 cmlsec) (rms)
Reaction time 5 min
The specified accuracy of outputs to the FCS follows. Actual accuracy will be sig-
nificantly better since the outputs are derived from inertial navigation grade sensors:

Angular rate Accclcration

Scale factor
Range 4(X) drciscc 20 g

Error sources. The primary sources of error associated with llSA follow.
(1) Axis alignt~~cr~remor: Achieving the 1-2 arc scc axis alignment stability needed,
if a significant portion of flights is to contain tcrrain avoidance and cvasivc n~ancu-
vcring, rcquircs very careful design. (2) Atcelcratint~scale factor stability: Skewing 0f
Sec 5.8 Integrated Inertlal Naulgatlon System (IINS) 241

accelerometer axes requires that accelerometer scale factor stability be signifcantly

better than for a nonskewcd configuration. (3) It~putaxis bending: Input axis bending
is the major error source for strapdown navigators in a vibration environment. (4)
Data synchronization: 1mportant.considerations for flight control are the time delays
and synchronization of data from the IISA when used as part of a digital FCS
controlling the states of an aircraft in real time. Data sampling and processing time
delays in the sensor clement have a destabilizing effect in an aircraft control system
and must be carefully selected. (3) Noise andfilterittg: Modern FCSs are digital, and
sensor data are sampled at some fixed frequency. for example, 80 Hz for modern
fighter aircraft. Sensor noise or vibration inputs at high frequencies can be aliased
by the sampling process to a frequency within the flight-control bandwidth, causing
control surface flutter or pilot discomfort.
In light of the preceding error sources, extensive laboratory testing of the
system must be undertaken to insure that IISA is suitable for installation and flight
test in a flight vehicle. These tests are described in detail Uankovitz et a]., 19871,
which examines IISA system performance for navigation and flight control.

5.8.4 Helicopter Integrated Inertial Navigation

System (HIINS)

Fingerote et al. [I9871 present the subject of HIINS, including the following dis-
cussions on accuracy requirements as they relate to configuration 1 shown in Figure
5-50. The roles of the military maritime helicopter include search and rescue, an-
tisurface surveillance and targeting (ASST), antisubmarine warfare (ASW), and an-
tiship missile defense (ASMD). Many of the missions must be carried out at ultra-
low altitudes under all weather and visibility conditions. The increased range, speed,
and accuracy of modern weapon systems impose stringent accuracy and reliability
requirements on the aircraft navigation system.

Accuracy requirements. In view of the often dangerous missions that

must be carried out, the small crew of the helicopter must not be burdened with
monitoring the functioning of, or updating, the navigation system. Consideration
of these factors has led to the following accuracy requirements.

1. Radial position error (95 percent), with external aids = 2.0 nautical mi (nm),
and without external aids = 1.5 nmlhr.
2. Radial velocity error (95 percent), with external aids = 3.0 ftlsec, and without
external aids = 4.0 ftlsec.
3. Attitude error (95 percent), with or without external aids = 0.5 deg.
4. Heading error (95 percent), with o r without external aids = 0.5 deg.

Another similar requirement proposed by Hassenpflug and Baumker [I9871 is listed

in Table 5-3. Comparing with the preceding requirements, it is seen that the ac-
Inertial Nauigation Chap. 5

Inertial Navigation

+ Data Bus

(h4ILSTD-1750A) Reference

Figure 5-50 HIINS Configuration 1 (Front [Fingerore et a l . , 19871 with permission

jom ACARD)


Parameter Range Refresh- Accuracy (95 %)
Pitch 8 -30 + 45.
rate [HZ 1
INS kquircmcnts
Roll 6 f 90. 50 .5 * .25O
True Heading
Velocity along
Velocity across v f50kmlh 50 .5X+ .25kt .5%+.2kt
Velocity vertical Y f l5mls 50 .6%+.2 kt .Z%+.lkt
geographic vertical z
v tl5mls 50 .6%+.2 kt TED
Ground speed v -60++400km/h 50 .5%+.25kt .5%+.25kt
Acceleration g
ax t. 5g 50 .Olg .01g
Acceleration a f.5g 50 .Olg .Olg
Acceleration Y -.5g++3.5g 50 .Olg .Olg
Angu- P 100~1s 50 .25*ls .2*ls
lar 9 60n/s 50 .25@/s .2*/s
rates r LOO*ls 50 .25*ls .2*ls
Position(Enr0ute) P.P b.25 2% 1.5%
Position(N0E) P. P 6.25 300m/l/b h 25Om/1/4h
Drift 6 f90 6.25 1. .5
Wind vW O++15Okm/h 6.25 1.2m1s 1.2mIs
Direction *W f 90. 6.25 1 1'
TAS u -25++100mls 12.5 Zmls 2mIs
v f l4mls 12.5 2mIs 2mIs
w f l5mls 12.5 Imls lm/s
T e m p e r a t u r e static -45tr70aC 6.25 2*C*:T 1100: 2 * C + : T 11001
Static pressure 480t1100mb 6.25 3mE 3mE
Height above ground OIZSOOf t 50 .5m 0.5% .5m 0 .5X
Target WPT f90*/f180. 12.5 0.5nm 0.5nm
Desired T r a c k DTK 0 + 360' 6.25 1 1'
XTrack XTK t5Okmlh 6.25 I km I km
T r a c k An6le Error TKE flOO* 6.25 1 1
Roll commanded *e - t30' 6.25 0.1. 0.1'
Turnrate dqldt 10.1s 12.5 0.6'16 0.6'1s

From [Hassenpflug and Baumker. 1987) with pcrniission from AGAHI)

Sec. 5.8 Integrated Inertial Navlgatlon System (IINS) 243

curacy requirements given by this table have the same attitude and heading accuracy
requirements and slightly different position accuracy requirements.

Configuration 1. Figure 5-50 illustrates HllNS configuration 1. The rec-

omnlendcd system shown in this figure comprises an airborne processor to which
four primary subsystem/sensors are interfaced by means of a MIL-STD-l553B serial
multiplex data bus: (1) inertial navigation system (INS); ( 2 ) NAVSTARIglobal po-
sitioning system (GPS) receiver; (3) Doppler radar velocity sensor; and (4) magnetic
heading reference.
Also interfaced to the proccssor arc a radar altimeter, a T A C A N receiver, and
an air data subsystem which supplies barometric altitude and true-air spccd infor-
mation. T h e self-contained INS is a standard form-fit function (F3) inertial navi-
gation employing ring laser gyros (RLGs).

Configuration 2. Configuration 2 of HIINS (based on Hassenpflug and

Baumker, 1987) and its 110 parameter are shown respectively in Figures 5-51. The
HIINS is a heading- and velocity-augmented INS system, providing three-dimen-
sional navigation information in conjunction with a radar altimeter and calculating
the wind vector by means of a TAS system for the entire speed regime of the
helicopter. The latitude range is ?80 deg ( U T M range). Angular rates and linear
acceleration in the body-frame coordinate system for flight control and weapon
delivery purposes are supplied by the INS. The autopilot functions are supported
by the following signals:
Radar altitude h~ Inertial altitude h,
Doppler vertical velocity v , ~ Inertial vertical velocity v ,
Magnetic heading +,L, True heading +
Attitude +, 0 Body velocities v,, v,, v,
Velocities in the navigation frame VE, VN, bv
Besides calculating the present position coordinates, the following navigation func-
tions are available: (1) bearing and distance to the selected waypoint; (2) time to go
to this waypoint based on the momentary speed; (3) optimal steering information
to the selected waypoint; (4) targets of opportunity; (5) position update by flying
over known landmarks whereby the position coordinates of these landmarks are:
(a) already stored, (b) read from the map and manually inserted after freezing the
position flown over, and (c) gathered and inserted by means of a map display after
freezing the position flown over.
Figure 5-52a is another view of an integrated NAVJFCS for helicopters, a
detailed discussion of which appears in Osder [1986]. It is a double fail-operative
(fail-op2) system in which the navigation function achieves the same level of fail-
op2 performance as the flight control computers. Each flight control computer
(FCC) contains its o w n internal redundancy structures and techniques (hardware
and software) needed to detect its own faults, and achieve the required reconfigura-
tion to support the system's fail-op2 specifications. The emphasis here is on the
Inertial Navigation Chap. 5

r---------- ------------

RDN 808
t - 1I -
I mu - INS
I , ucs

- *
---------- -------
fl -----_I

Figure 5-51 (a) HIINS Configuration 2

(b) Configuration 2 110 Parameter (Fr*m
[HasscnpJug and Baumker, 19871 with per-
mission j o m AGARD)
Sec. 5.8 integrated Inertial Nauigatlon System (IINS) 245

system architecture advantages obtained by making the navigationlflight control

sensor function an autonomous navigation subsystem by virtue of the consolidation
of sensor measurements, state estimation algorithms and self-contained reversion1
recollfiguration stratcgics. Figure 5-52c illustrates the data flow interfaces between
the integrated navigationlsensor assembly and the flight control computers and ex-
ternal sensors. The broadcast bus transmissions of flight critical data to the FCCs
are summarized in Figure 5-52d. The tight coupling of data from various mea-
surements is implied in Figure 5-32b by the position vector output which contains
the four states of latitude, longitude, altitude, and GI's clock bias ( C R ) For
. accurate
navigation, the GI's clock bias as well as its first and second derivatives are included
in the navigation filter.

5.8.5 Integrated Missile Guidance Systems

This section discusses the design of integrated navigation, guidance, and control
(NGC) systems for tactical missiles using optimum control and estimation theory
based on Williams et al. [1983]. Sensors used in this design consist of strapdown
accelerometers and rate gyros and a strapdown homing seeker. Guidance and control
algorithms, in addition to performance studies, is presented in Book 3 of the series.
The overall system is shown in Figure 3-53a. The design uses a passive strapdown
seeker and a set of strapdown inertial sensors (gyros and accelerometers). These
sensors provide the two LOS angles between the missile and target as measured in
the missile pitch and yaw planes, and the missile body rates (P, Q, R) and accel-
erations (A,, A,, A,) along the three principal body axes of the missile. T h e LOS
angles, plus information about the missile body rates and accelerations as provided
by the inertial sensors, are inputs to the navigation and FCS. The integrated N G C
system is shown schematically in Figure 5-33b.
Since the guidance algorithm requires knowledge of the inertial LOS rate,
which cannot be directly obtained from the output of a strapdown seeker, an al-
gorithm by which these rates can be estimated given the angular inputs is essential
in order to achieve good performance. In the design of Figure 5-53b, these rates
are not obtained by a finite difference algorithm which could produce serious errors
in the presence of noisy o r infrequent data, but instead an extended Kalman filter
(EKF) includes these rates as state variables. The EKFs are part of the navigation
filters as shown in the figure. Since both channels have identical dynamics, the
navigation filters for each channel are also identical and have the structure shown
in Figure 5-53c. The block diagram is intended to show that the output of the filter
u, as well as the state estimates 8 and 4 are continuous-time variables, whereas the
inputs to the filter are discrete-time variables. (In actual implementation, the con-
tinuous-time variables are also discrete-time variables, but the sampling interval is
much smaller.) The navigation filter automatically extrapolates the state estimates
from one observation until the next in accordance with the missile dynamics. This
is very important when the seeker-sampling interval T, is relatively long compared
with the dynamic behavior of the missile. It also automatically solves the problem
Inertial Navigation Chap. 3
Sec. 5.8 Integrated Inertlal Naulgatlon System (IINS)

Parameter DesCrIDtlon

body angular rates

body I l n e a r accel.

P l t c h , R o l l . Iltadlng
vn. vc, Yz North. East. Dan Veloclty
L.A. h Latitude, Longitude, A l t l t v d e
l m d other coordlnatc
systea equivalents)

Baro A l t l t u d c
Calibrated Alrspeed

True Airspeed

A t r Dtnslty
A l r Temperature


El .

Figure 5-52 (a) Integrated NavigationlFlight Control Overview Block Diagram

(b) Flight ControllNavigation Interfaces (c) Summary (d) Typical Strapdown Block
Diagram for Low Cost/Performance IMU with Kalman Filter Updating (From
[Osder, 19861, O 1986 IEEE)
Inertial Nauigatlon Chap. 5



-PI C001101NATE

8 0 V


- UlC"
Q tf'
- I O U

- "A.

Figure 5-53 (a) Overall System (b) Integrated Guidance and Control System Sche-
matic (c) Navigation Filter Schematic (From /Mrilliatns rt al., 19831, 0 1983 A I A A )
Sec. 5.8 Integrated inertial Naulgation System (IINS)


of seeker blinding which occurs when the target conlpletely fills the field of view
of the seeker and the latter can no longer provide useful directional signals. When
this happens, the switches shown in Figure 5-53c remain open and the missile
continues by dead reckoning to the interception.

Another example of an integrated seeker-guidance navigation

system. In a classical design, a weapon guidance system includes two distinct
inertial sensor sections. Figure 3-34a showsthe separate flight control using a ref-
erence gyro for yaw and pitch corrections. Also shown in this figure is the additional
seeker equipment, which in this case is an optical device including a tracker, that
relies on a second reference gyro system for LOS stabilization. Figure 5-54b(i),
however. shows one stra~downreference svstem made UD of the reference unit and
the strapdown computer. This single unit is capable of referencing both the flight
control and the seeker system. Figure 5-54b(ii) shows how future weapon system
guidance and seeker-processing tasks will be integrated into one processor. Figure
5-54c illustrates a structure of the navigation software development system (NSD).
Figure 5-54d depicts a block diagram for the LOS control. The gimbal control
including the controlled element (gimbal and torquer) together with a high-band-
width angle transducer are both contained in the inner loop. The nonlinear char-
acteristics shown in the figure represent the effects of friction found with all gimbal
Inertial Naulgation Chap. 5

Figure 5-54 (a) Functional Diagram ofa Classical <;uidance& Seeker System with
Separate Gyro Rcfcrenccs (b) Functional Diagram o l a Guidance & Seeker System
with a Strapdown llcferencc (i) Distributed Electronics (ii) Low-Cost Version with
an Advanced Processor (c) Structure of the Navigation Software Development Sys-
tem (NSD) (d) Block Diagram of Line-of-Sight Control (From [RahlJs rr a / . , 19871
with prmrissio!~jo!n A C A R D )
Sa. 5.8 Integrated Inertfal Naulgatlon System (IINS)


II - 1
m 1 m MTA ElWl
m 1 m

m 80285187



- LOS.. . J ~ n eo f s~gnt
Guidance Processing
Current methods used in guidance processing are examined in this chapter based
on AGNC[1986(i)]. Figure 6-1 illustrates the order in which these methods are
presented, including their subsections. The primary functions of the elements that
make up the guidance system include sensing, information processing, and correc-
tion. The guidance algorithm is used for correction. Several different guidance sys-
tem configurations are currently in use as solutions to numerous types of guidance
problems. The guidance problems are mainly determined by the nature and the
location of the target, the circumstances for weapon delivery, and the environmental


As discussed in Book 1 of the series, precision-guided munitions (PGM) depend

heavily on their guidance processors. Imbedded processors are currently present on
all guided vehicles in order that the entire system operates based on a particular
algorithm. A general guidance system is drawn in Figure 2-2. Theguidance algorithm
as part of the guidance loop represents an essential component in the design of an
integrated guidance system. The information subsystem, which is needed to perform
the guidance task of pursuer-target intercept, determines basically the configuration
of necessary sensors and information processing, including the guidance tracking filter.
Sec a1 Quldance Processors

( k c . 6.1)
Guidasc Mission & Pctfonnance Guidana Algorithms

G u i d m Rrfonaance G u i i e Phase Opcratiin PresetGuidancc Dinct Guidance

(Sec. 6.2.1) (Sec. 6.2.2) (Sec. 6.23) (Sec. 6.4.1) ( k c . 6.4.2)
Multiple Mode Guidance Modeling

Active Semi- Passive

hG&omsc 'Itnninal E m r Analysis Model
(Sce. 63.1) (Scc. 6 3 2 )

Guidance Filter Guiance Law

(Chap. 3 ( k c . 6.5)

LOS An& Guidance LOS Rate Guidance
( k c . 65.1) ( k c . 6.5.2)

Pursuit Deviated Pursuit1 Constant Proportional

Fixed-Lea Navigation
Guidance (PNG)
and Its Variations

Sensitivity and Comparison of Guidance Laws
(Sec. 6.5.3 & Chap. 8)
Guidance Mode Options

Single Mode Dual Mode Multimode

(Sec. 6.6.1) ( k c . 6.6.2) (Sec