Guidance, and
Control Processing
b y ChingFang L i n
ChingFang Lin
American Gh'C Corporation
Lln. ChlngFang.
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 0135962307
1. F l l g h t c o n t r o l . 2. G u i d e d ntssllesControl s y s t e m s .
I. T l t l e . 11. S e r l e s .
TL589.4.L55 1991
629.10C20 9043009
CIP
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
ISBN 01359b2307
2 ModelingDesignAnalysisSimulation
Evaluation (MDASE) of NGC
Processing
2.1 LinearlNonlinear Intercept NGC
System, 13
2.1.1 LinearlNonlinear Intercept N G C
Processing, 14
Contents
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 SingleMode, DualMode, and Multimode
Guidance, 379
6.6.1 SingleMode Guidance, 379
6.6.2 DualMode Guidance, 380
6.6.3 Multitnode Guidance Applications, 384
Contents
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 alltime high. It is against such a background
that the present series was conceived to assess stateoftheart 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 selfcontained, 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
situation.
Preface
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 aerospacerelated 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 mathematicallyoriented 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 selfsufficient 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 avantgarde 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 selfteaching, bridging the gap betwccn thc theoretical
and the practical.
Preface
Introduction
ModelingDesignAnalysisSimulationEvaluation
(MDASE) of NGC Processing
I
 1
Hodern Multivariable Performance
Control Analysis

rn
Modern Filtering
Advanced Adaptive NGC
1
KnowledgeBased
Design and Estimation Multivariable I Systems: Detection and NeuralNetwork
for Advanced NGC Techniques Control Design I and Identification Approaches to
Systems Design (Chap. 4) Techniques (Book 4) Systems Control
(Book 4)
  I    1

1
Fundamentals of Guidance Processing Flight Control
NGC Processing Processing
(Bogk 3)

7
Design +
Advanced Flight Control System Design
(Book 3)
I
r 7   C    1
Integrated, Adaptive, and Digital NGC Systems Design
Intelligent NGC Systems Design (Book 5)
(Book 4)
1.1 OVERVIEW
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 microprocessorbased 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
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 rcalworld 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
systems.
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, 19831985;
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.
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 longterm 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 fireextinguishers, 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 fireprotection 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 1la, 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
1lb. 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
I
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 747400 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
design
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
concepts.
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 stateoftheart 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 35 of the series is to meet this challenge.
* Ibid.. p 3435.
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.
1.2 OUTLINWSCOPE
+
Advanced Multivariable
(Chap. 4)
+
Adaptive NGC System
Control Design Techniques Design Twhniques
(Book3) CBook 4)
v v
KnowledgeEased and NeuralNetwork
Awroacha to Svstems ConUol
Figure 12 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 knowledgebased and neural network intelligent system.
The interactive relationships between components and techniques are depicted in
Figure 12b. This figure illustrates how the design algorithms are made more in
telligent by incorporating a knowledgebased 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).
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
I I I
Tracking Sensing Information
+
I
Guidance
+
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)
night
(Chap. 5) Control
(Chap. 8)
System Control Surface
(Books 1, 3,4, 5)
Guidance Commands Deflection Commands
b
I I
Navigation
Steering Signal and Attitude Information
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.
ModelingDesign
AnalysisSimulation
Evaluation (MDASE)
of NGC Processing
It is brought to the reader's attention at this point that a very broad picture of the
modelingdesignanalysissimulationevaluation (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.
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 lowaccuracy inertial components (gyros and
accelerometers). Small tactical missiles and most large transport vehicles often have
openloop control instead of the more complex FCS control.
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 pursuertarget 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 21 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 highfrequency 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 72 depicts a
missdistance 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: rangcdepcndcnt (thermal) noisc o,,,, rangcindc
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 22, u, is
normally taken to be the lateral acceleration command. As long as saturation effects
are ignored, the model shown in Figure 22 is a linear timevarying system driven
by stochastic inputs. The inclusion of acceleration saturation effects causes the model
to become nonlinear.
The digital simulation technique in which the equations of motion are represented
by' a set of firstorder nonlinear differential equations is widely recognized. The
modeling vector form is given as (see Figure 23):
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 (2la) corresponds to Figure 23, 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 24, defines all of the state equations. Referring again to
Nonlinear Function 
t
f(x(t).uo),t)
Figure 23, 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) (21 b)
where w(t) usually, although not necessarily, .denotes a whitenoise input vector.
A more general case is considered in Figure 24, in which the model includes certain
error sources resulting from imperfect measurements or unmodeled dynamics. In
corporating these error sources in Equation (2la) 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 statedependent random vector; and
nz(t) is a stateindependent random vector. Equation (2lc) could also apply to
Figure 23, in which case w ( t ) is simply the sum of 6, nl(x, t), and n2(t). If M I or
n2 represents a colorednoise process, a shapingfilter technique can be used to model
either or both so that the new state vector can be augmented to Equation (2lc).
The resulting augmented state equation is driven by the whitenoise 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 closedloop 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
highfrequency 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 lowfrequency 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
21. 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 24 can be devclopcd for a particular pursuertargct
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 selftest. Their relationships
are dcscribcd in Figure 25.
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 wnlinuouslime nonlinear
system is dcscribcd by
T
i t =x I I t I , X I =x I ~ 4q1)*TI
Q ~ I= 0. I = at)b(tr)
T
.,
(la)
Z ( I ) s h(x(1). I) t "(I), Qv(t)] = 0. Qv(l) "(1) ] = R(t) b(1z)
(lb)
The continuoustime nonlinear system with d*crete.time nonlinear mcasuremcnls is repnscnled by:
i(1) = f(X(1). Nt), 11 + *I)
(2=4
T
q k ) = h[x(tkh kl + v(k); qv(kl1 = 0. Qv(k) v(k) I = Q
Gb)
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
(2c)
h applying many MDASE uchniqucs. the system quations (1)need to be linearized as:
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 missiletarget line of sight (LOS). In one of these, the missile is constantly
flying
.  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 constantbearing
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 missiletarget velocity aligned
with the LOS. As shown in Figure 22, 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 constantbearing 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
StrapdownAFtU

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
o
o
Vertical Launch
Separation Mode
Stability
t
I
Seeker
Interface
o Aim Point
Self Test
o Test
Right Path Planning
o Stored Mission
Actuation1
o Fin Rate Limitation
4 Sequence 4W
t
o Initialization Profile o Fin Angle Limitation
I 0 Status
 o Change Profile
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
vmf
Missile
Mi&
vector
velocity MiSc 
without ,
, R ~ V target
~ velocity vector
maneuver
N [NORMAL FORCE1
X AXIS
(PITCH TOTAL MOMENT)
REFERENCE

CN
NORMAL
FORCE
COEFFICIENT
0
6
MAX (TRIM1
 .30
c~
TOTAL
MOMENT
COEFFICIENT

0 k@ya MACH CONSTANT
ALTITUDE =CONSTANT
OMAX ITRIM)
& = . 3 0 = 6MAX
6 = .20
6 = 00 b = 100
MACH =CONSTANT
ALTITUDE = CONSTANT TOTAL MOMENT COEFFICIENT CM
VERSUS ANGLE OF ATTACK n
(D)
Figure 27 (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 TailControlled 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)
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 (0Y); 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 TRADEOFFS
E = M* = a,CM6 TO REDUCE 1/A &INCREASE MANEUVERABILITY
n = i.,= a, cNa WHILE MEhTlNG PHYSICALCONSTRAINTS
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
Wnpw5253. 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 =
d
.
B., = 11o:" = lllC+r\o,
&
. .
Ma > 0 and 2Sdwd i s small
rA
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 23  s [Neslines. 19851 with permission from AACC
HIOH ~ R E Q ~ E~ N ~~ ~1 l From
a(~)
Aaua10r: 
4(5)

I+st
25,
1
s2
(1) 
Rate Gyro: ~ G ( S )
0 )

2h
1+s+
1
s2
(2)
 MODE MODE
.
5
5
.
FREOUENCY
GENERALtYD
MASS t l b / ~/in) 0.0517 0.0373
SECOND MODE
0.6.
2
0.4 i Gym Location
0.2 )X
4.6 LOCATION
a) Fint and E m n d Mode Shapes

bl Normalized Bodv Bendinn Sham
 ~ ~ I Y Y &
.
&B 
Xffi XBL
6
,1 2Ei
l+s+l)
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 
.ode
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

*
(lbftsec2)
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
 
Xm
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 )
10.02 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 . , lbfts.c2)
0.03131 lbsec / i n . )
Defining

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
1 HITMU
ECM
Target Dynamicl&
+ Relative ~ c o r n c n y t
Sewor
Target Stale
Compurarion
Target Starc
bra 4 Sellsor Models el Filtering &A Estimation
T
Esbres
Pmdicnon
airtoground 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
ALTITUDE
(BAR0 OR RADAR ALTIMETER)
ANGLE RATE
DIRECT ( E  J TPACKER)
Figure 29 (a) Visual LineofSight (LOS) Angle Measurements Determine T w o Coordi
nates of AircrafttoTarget Position (b) Determination of Slant Range (c) External Grid Ref
erence Targeting (Courfesy 0jJ.S.Ausman, 1986)
permitting target identification day or night. Lowaltitude 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
GROUNDBASED RADAR
AIRCRAFT TO
RGET
/ \
X
=TARGET
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
,
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 closingvelocity radar information are available
through measurements, but in the semiactive mode, these measurements are bistatic.
2.2.1 Targeting
to locate the pursuer with respect to a pseudotarget 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 24. A particularly com
mon typc of targeting sensor is one that measures the azimuth angle and depression
angle of the pursuertotarget LOS. Examples of this typc of sensor system include
gunsights and headup displays (HUD) in combination with an attitude reference.
As shown in Figure 29a, these sensor systems measure just two angular position
coordinatcs. A slantrange sensor such as an airtoground ranging radar or a laser
range tinder can be used to directly measure the third coordinate of pursuerto
target position. This is illustrated in Figure 29b, which also shows two additional
more indirect techniques for determining the third (slantrange) 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 24 corresponds to the external reference type of targeting
sensor, which is illustrated in Figure 29c. 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
groundbased radar, LORAN, TACAN, DME, and GPS.
bombing run. The LORAN system is capable of obtaining coverage beyond the
LOS and at low altitudes because it uses groundwave transmission oflowfrequency
radio waves. However, since it is only a twodimensional positioning system,
LORAN must rely on another method such as a baroaltimeter 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.
Because only kinematic equations are considered in the kinematic method, the pur
suer may be treated as a geometric point as illustrated in Figure 210. 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
210 and the measurable variables are
2 I12
R = (x? + y,) , o = tan' y , / x , (23)
The closing velocity V , =  R of the pursuer and target is determined by Equation
(22a), while the LOS rate u is related to the parameters of motion of the pursuer
and target by Equation (22b). A set of analogous equations is used for the vertical
plane. Differentiation of Equation (22b) and substitution of Equation (22a) yield
R6 + 2db =  VT sin(u  yT) + V T cos(u  ~T)+T
(24)
+ i',,,  y)  V,,, COS(U  y)+
sin(u
Because kinematic Equations (22) and (24) 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
where
t, = tf  t = time to go and t f = flight time
For analysis purposes, assuming the intercept geometry is simplified to a headon
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 (2SC)
y, = yT y,,, = AT"  A m y = V T j T  V,,,? (2Sd)
where
AT^ = y ~ =. actual target lateral acceleration = V T jT cos y ~ , =
! VT jT
(25el
A,, = , = actual pursuer acceleration = V,,,j cos yo  V,, j
Equations (25) 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 (2Sa) 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 (2Sa) leads to the following formula for the
L O S rate u:
tegrator contained in the positive feedback loop, as shown in Figure 211b. 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 (28),
AT, is given by
AT, = (Rs  2V,)u + A ,,,, =  2)u
V<(t,?s + A,, (29)
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
210 shows the essential seeker angles and geometrical quantities. As shown in
Figure 212, 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 213a 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 closedloop system encompassing the dish rate feedback loop.
MDASE of NQC Processing Chap. 2
 RADOh
ERROR
rsuer
laation
Gyro Acceleration
(B) Sensitivity
Figure 213 Seeker Block Diagram (a) Without Kadomc Effect (b) With Radomc
Effect
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 angleoff 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 singleplane 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 inplane and crossplane radomc errors to become lookangle dependent
with the slopes also depending on the roll attitude of the in~erceptor.Specification
~
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 210. 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 (2lo), the tracking error
measurement (see Figure 210) then becomes

 ~ r , e  e  ~ g + ~ r + V , = ~ ~  u , + ~ , + ~ ,
The angular relationship of Equation (213) and the dish dynamics from Figure 2
13a are employed to develop a seeker block diagram with radome effects shown in
Figure 213b. Since the radomeinduced 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 crossplane 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 22 to compute the LOS angle measurement.
Combining Equations (210) and (2ll), another simplified equation that can be
used to calculate the LOS angle measurement 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 213b) to isolate
body rotations and track loop to maintain the seeker axis along the missiletarget
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 slewrate 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 90dcg 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
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
,~
Sensitivity
(B)
Figure 214 (a) Typical Tracking Error Guidance Seeker Loop (b) LOS Rate Guid
ance System Seeker Model
Sec. 2.2 Target Slgnal Rocesslng 43
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 (213). 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 missdistance impact [Fossier, 19841. The preceding two parasitic
feedbacks become important at high closing velocities occurring in shortrange or
highaltitude 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 33 and 24, 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 lowfrequency 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 213b). 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 closedout 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 previouslymentioned 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 highfrequency 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
a) Simplified P u r s ~ Dynamics
r
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 I
I *kc, Dirh I
qne I Ratc Command
I b;la I
Pursuer mnamicr
Simplified
Punuer D y o a k
(See F i e a)
Figure 215 (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 bandlimited 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 22. The PSD
of y , is
+,\. = 27CLTrCLT, ~ C L T= glint noise variance (22Ob)
The filter was excluded in Figure 215b 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
u,  N(0, r,) with equivalent white clutter noise PSD +, = 2~,r, (220g)
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
In addition, the tracker noise, or.., can be added to target position noise, y,, so
that the measurement equation is modified from Equation (219) as:
z(t) = y T +n = target position measurement (224)
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: '
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.
Model helupmnent
' Wind tunnel test
L CFD. CSM
* Physical laws
Design Specifieatim .
Military speciliatiow
* Flyingqualities
Stability requirements 4
SyPtem identification M i u distance 4
'
Model validillim Performance
I
Open L m p Analysh
Design model
c Model reduction
* System limitation
C
Conlml Syslem D e i g n
,
CACSD
Advanced guidanceand control laws
'Design knowledge h w
Implcmentatia,
Fast prototype Figure 216 Robust Integrated
Envirol~mmlled
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 openloop 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.
Xl'W
c
LINEONIOW1 U T E
C Low C A N
FILTER
FILTERED LINEOF.
LlOn M T Z
C CRWAV
LAW
KCELERATIW
COMMAND

FILTERED LlNEOFalOHT
nus rson~orr~~

OTHER MTE
SENSORS EIIIMATU
LINEOFJIOHT RAT
XltW
L1 u
/L=. bDVU(CED
OUlOANCE
U X E6ETlER
LERATIW
COYUAND
EITIMATOR U W
TARGET
C L W K A L LIIROhCH
NOISE OUIDANCE
SEEKER LAW AUTOPILOT AIRFRAME
~mowAv)
MISSILE
KINEMATICS
TARGET
KINEMATICS
ESTIMATOR
MISSILE
KINEMATICS
Figure 217 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
The guidance analysis model shown in all parts of Figure 215 is composed of
elements representing the homing kinematics plus other systems that affect the hom
ing guidance dynamics. The rootlocus 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 215a. The stability analysis pertains only to the guidance
system transfer function, whereas the missdistance analysis will apply to the entire
homing loop of Figure 215b. The homing loop shown in Figure 215b 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 fifthorder 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 23 through
25 later will discuss radome error and its severity on the guidance and control
responsiveness of homing missiles. As will be shown in Example 24, 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 missdistancc 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 rootlocus 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
rootlocus study, concentrating on the optimum choice of N G C gains.
Substituting jo for s in Equation (226) 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$A(o) is the autopilot phase response.
Although the linear mode of the autopilot can easily be solvcd, it is difficult to
analytically obtain the contribution to this response by the nonlinear autopilot. Thus,
the nonlinear autopilot phase and gain characteristics are obtained through computer
simulation.
analytical expressions for effective slope and effective noise are derived for a sinu
soidal radome error. The guidance system transfer function with radome effect is:
7; = effecti4e guidance system constant
'2
4 rn =
A' Vcs =1 + ~ A V / V , ) ' ( T T , ) I I (+~ A V V (228)
U l + Tis ' A! = A
= effective navigation ratio
1 + rAV,/V,
where T~ = T, + +
T, T~ Figure 215a(iii) illustrates how the radome slope changes
the missile system dynamics. The radome slope r appears twice in A,,,/u. A' is
altered in terms of r effect by a factor of 1 over the quantity 1 + rAV,/V,,. For
positive radome slope, A' is reduced and the response time of the system is increased,
the effect of which is to increase the miss distance due to target maneuver. T h e
opposite occurs for negative radome slope, the result of which is to increase the
effect of A' and thereby increase the miss distance due to RIN [Fossier and Hall,
19671. System damping is reduced in the case of a negative radome error when
higherorder terms are concerned. Figure 218a shows a typical missdistance sen
sitivity to r. The figure shows that there exists a negative threshold value of r, below
which miss rapidly escalates. This value depends on the flight condition through
a/$ 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 2lja(iii) and 215b gocs unstable for r =
0.045. This secondorder 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 thirdorder autopilot as will be given later in Examples 24 and 25,
together with a firstorder 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 218b shows, the miss distance increases with ra
Figure 218 (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 )
. ~
89 CONSTANT SLWE I?
8.
4.
8 2.
5
0.
.2.
4.
'
1 2 3 4 d
TARGET MANEUVERTIME TOGO i rl
3 9 TARGET MANEUVER
7. mred RANGE INDEPENDENT NOISE
(D)
TOTAL

.O.W .0.02 0 0.02 0.04 0.06 0.08
RADOME SLOPE
 

TARGET
MANEUVER:
 ...  . .
RAWME
RANGE
0, I INDEPENDENT
!0.08.o.m  0 . ' ~ .o.b2 o 0.b2 0.04 0.05 0.08 NOISE:
MAXIMUM RADOME SLOPE
(C) t
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 3g step target maneuver is uniformly distributed with 25 samples over
the 5sec 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 218b, the RMS
value for zero radome slope is 1.0 because all RMS miss distances in this example
are normalized by the same factor.
Evaluating Equations (229) for the sinusoidal radome given in Figure 218f results
in thc following normalized effective slopcs and effective noisc variance:
rT
~.WAX 0s( ) [[sin (y)
 rf  COS
~.\IAX
where r ~ f . 4= ~2 ~ r , ~ and
/P

(J,
  sin
2 ~ 4 ~
(Y) sin
ro P
Because it is not significant for long flight times, the bias is neglected. Equations
(230b and 230c) are plotted in Figure 219a 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.
SlNUSOlDAL RADOME
DENOTES EFFECTIVE SLOPE V A L U E
GIMBAL ANGLE
STANDARD
I
D E V I A T I O N ldegl

EFFECTIVE
SLOPE, rf TRAJECTORY
DYNAMIC
MODEL WITH
2 mrad
RANGE
INDEPENDENT
TARGET MANEUVER NOISE
EFFECTIVE
3 0 d q PERIODSINUSOIDAL R A D W E
0 180d.p
INDEPENDENT
EFFECTIVE EFFECTIVE NOISE
SLOPE. rf
:;;,,
STANDARD
DEVIATION 1rnr.d)
.0.0(5
. o . m h ,
0
0 5 10 15 0 5 10 15
GIMBAL ANGLE GIMBAL ANGLE
STANDARD DEVIATION Id41 STANDARD DEVIATION 11
(C)
(A)
SlNUSOlDAL
RADDME
1 0
.o.a . 0 . b .o.b2 o 0.b2 0.k o.'a 0.h
MAXIMUM RADOME SLOPE
Figure 219 (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
3 9 TARGET MANEUVER
SIMULATION SIMULATION
ANALYTICAL
1 1
Y
%
40
TARGETMANEUVERAT
5wTOW 1 TARGET MANEUVER AT
3ucTOGO
0 5
TIME TO GO lul O TIME TOGO (ucl
(E)
SlNUSOlDAL
TRAJECTORY RADOME
INDEPENDENT
SLOPE
A N 0 NOISE
RADOME
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
rf = .0.004 RAOOME SLOPE MAXIMUM RADOME SLOPE
(H)
30 dog PERIOD SlNUSOlDAL RADOME
30 deg PERIODSlNUSOlDAL RADOME 4 (0.045( MAXIMUM SLOPE
0.014
EFFECTIVE SLOPE
AN0 NOISE
6 0 3 6 RADOME
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
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
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(235)
n anh is the same as that in Equation
(228). Here, as in the case of the firstorder system, the effective navigation ratio,
A', is modified by radome slope according to Equation (228). 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 firstorder 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,
19611.
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 lefthand side of the preceding inequality gives
the lower limit of the allowable negative radome slope, while the righthand 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
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 (235) and Equation (4) of Table 22, the fin rate response
due to LOS rate for the fifthorder binominal guidance system becomes
As was true in the case of fin angle, zerolag guidance system results are very useful
for determining fin rate response due to deterministic inputs. When guidance system
dynamics are neglected, Equation (237) becomes [Nesline and Zarchan, 1984(a)]
Following Examples 22 and 23, 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 thirdorder
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 221d. These are two important factors that contribute to
high miss distance in a radarguided PNG system. An adjoint model similar to that
to be presented in Chapter 3 for Example 22 is constructed based on the homing
loop shown in Figures 215a(iii) and 2ljb, 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 220. Neglecting parasitic effects, the guidance system
transfer function is given by
TOTAL rmr
TARGET
MANEUVER
CONTRIBUTION
NOISE
CONTRIBUTION
o 2 i 6
E F F E C T I V E NAVIGATION RATIO, A
(A)

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 firstorder term in the denominator of Equation (240a) 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.
where 7; and A' are the same as those in Equation (228) except that T, is given by
Equation (240b). 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 220c [Nesline and Zarchan, 19821. Therefore, the optimal value
66 MDASE of NGC Processing Chap. 2
A,, 
 A V<u;
1 + h A s + 275s2 + 7is3
where 7; and A' are the same as those in Equation (228) 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
highaltitude 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 22. 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

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 thirdorder transfcr
function as described in Equation (23'9) or (342a). If thc nonlinear elcmcnts are
not included in the analysis, Equation (243a) 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 frequencyresponse
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 (243a).
T,;
(1 + G,A V,i C',,,)(T, + T,) + 7, + (G, + r) T,A V,I V,,
(243b)
1 + (G, + r)AV,/V,,,
'=
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 (228) and
(243), 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 rootlocus technique is applicable
only to a linear, timeinvariant system. The root locus for a linear, timevarying
system of Equation (243a) can be obtained with the roots at different t,. The
evaluation can be performed on the transfer function given in Equation (243a) for
T,, T,, and G,, which vary with respect to t,.
on the rigidbody angular rates is valid. The transfer function from the generalized
acceleration f to the normalized bodybending curve can be written as
Targct noise has been presented in Section 2.3. I , and is applied hcrc to gencratc the
target mancuver signaltonoise ratio. Figure 221 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
Target
. 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)
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 21% and 221 is a firstorder GaussMarkov 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 (225). Figure 221a 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 221b through
221d).
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 221d). The uncertainty in target position y + is described by a
thirdorder integrator with a constant PSD 4, whitenoise 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 221d.
70 MDASE of NGC Processing Chap. 2
With regard to the uniformly distributed target maneuver, it can be seen by ex
amining Figure 221d 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 Figur6222 [Cloutier et al., 19881. Initially, as shown in
Figures 222a and 221c, 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.
and Stcar, 1974; Singer, 1970; Vcrgez and Liefer. 19841. The wellknown Singer
model [Singcr, 1970) has been used with various Kalmantype 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 modelbased 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 singlepolc 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 221e
AT, = 2fAr, + 2B d? w , = A,AT, + h,u, (250)
where u, is a zeromean Gaussian noise, that is,

~v, with PSD =  (251)
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
semiMarkov process. The mean of their process, Figure 222c, 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 222d. Larimore and Lebow [I9871 de
veloped a similar model based on a parameterized GaussMarkov 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 continuoustime Gaussian process and a finite
basisljump process. The continuoustime 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.
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 firstorder
GaussMarkov. It was assumed that the orientation of the threedimensional target
could be obtained by means of an EO seeker in conjunction with pattern recognition.
Bullock and SangsukIam [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 statedependent noise. Finally, the
parameterized GaussMarkov 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 continuoustime 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 222d.
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 (23) and is given 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
where
74 MDASE of ROC Processing Chap. 2
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 221d 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 (251). The states A,,, A,,, A,,, represent the pursuer accelerations,
and the firstorder model defined in Example 22 is assumed here with its bandwidth
o defined as
o = (262)
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
as
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 29). 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 (263b)
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,,,
(263~)
ut, + u , + a,, + 0,,,,
urn + Of + u, + ~gl,,,#
where the measurement equations are the same as Equations (263), 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 (261). 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 (264) is
x ( k ) = O ( T ,A)x(k  1 ) + W u ( k  1) + w ( k  1 ) (265a)
where the state matrix @ is
D(A, A)
A'(13  e K A A )
e AA I (265b)
If U(T)is assumed to be u(k  1) between the time intervals (k  1)T and kT, then
V( k1 = ( I ,
T2
T I 0
T
u(k  1) (265c)
and the discrete form of the process noise w(k  1) is a whitenoise sequence
where
D ( k T  T, A)
 e(kT~)A]
(kT,)A
(265e)
E[w(k  I)] = 0, E[w(k  l)w(k  = a
and the discrete process noise coGariance matrix a is
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 singleinput
singleoutput (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.
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
31. Thc transfcr function C(s) incorporatcs both opcnloop plant dynamics and
Sec. 3.1 Singular Value Analysis
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 halfplane 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 openloop 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 joaxis 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 (31)
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
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
32 satisfying the inequalities, ci < C , < c 2 (i= 1, 2, . . . , m), the closedloop
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 32
satisfying the inequalities  c < &(o)< c ( i = 1 , . . . , I N ) , the closedloop 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,
suitable robustness measures are identified and more general robustness analyses are
discussed.
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
bation
produces a singular matrix which cannot be predicted by the eigenvalues. The sin
s
gular values of M 6 defined by Equation (32) 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
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
Consider the system depicted in Figure 33. In control system design, typical re
quirements are (1) closedloop 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 lowfre
quency parameter errors or the changing number of right halfplane 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
Controller
K(s)
glected highfrequency 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 highfrequency region is guaranteed if the nominal system is asymp
totically stable and the following condition is satisfied [Doyle and Stein, 19811:
The system remains robustly stable if Equation (37) 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
functions.
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 tradeoffs 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 highfrequency range (to be robust against highfrequency un
certainties). Since at low frequencies the loop gain is high, the sensitivity function
in the lowfrequency 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 highfrequency 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, blockdiagonal, 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 lowfrequency requirement; (2) fast response characteristic,
an intermediatefrequency requirement; and (3) large stability margins, a highfre
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.
where, as defined in Equation (3I), 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 wellknown inequality
 + A] 2 g [ A ]  1 in Equation (38), a sufficient condition of Equation (38)
o[l
can be given as
1
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 openloop poles, the upperbound frequency in Equation (3
9) has to be lower than the maximum openloop 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
steadystate error will be decreased by a factor of p(0) as follows. When r(s) in
Figure 33 is a step function given by r(s) = [I, . . . , 1IT/s = Hls, following
Figure 33, the steadystatc crror e, can be written as
1
P=
5
e , = lim SSO H = So(0)H (3 10)
T , is inversely proportional to the real part of the closedloop 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 (311). A
typical requirement is shown in Figure 34. Design engineers then shape the loop
transfer function to meet the desired requirements. There exists a fundamental lim
itation here due to the right halfplane poles and zeros. For the system to be stable,
when the plant (or controller) has right halfplane zeros, the sensitivity function
dB
RollOff C.baracIcristiu
Robust Stability
Requirement
Figure 34 Singular Value Plots and Constraints for the Loop Transfer Function
GK(jw)
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 halfplane 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 singleinput 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 normbounded 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 35. Hcre, K(s) is a stabilizing controller for the nominal plant G(s). In
a straightforward way Figure 35 can be equivalently redrawn as shown in Figure
36. In Figure 36, the combination of G(s) and K(s) inside the square box can be
denoted by a nominally stable closedloop 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 36 is just a special casc of the gcncral
structured block diagonal perturbation problem as shown in Figurc 37 with 11 =
2, A = block diag { A , , A,), 2 = [ d l , d2IT. Z = [n,, n2IT, 1,. = r, and = E . 11,
Figure 37, M(s) has the following form:
Sec. 3.4 Structured Singular Value 87
The issue of robust stability becomes one of characterizing the set of allowable
perturbations such that the closedloop system remains stable. Note also the ap
pearances of the sensitivity and the complementary sensitivity functions in the matrix
M(s).
It is well known that the systcm in Figure 37 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
e +
Figure 36 Equivalent System Diagram of the System Shown in Figure
88 Modem Multluarlable Control Analusis Chap. 3
.
a = [dl, 4 .... %? n = [nl,
T
nz, .... bl Figure 37 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
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 (317) 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 (315)
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 socalled 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 blocklimiting
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 blocklimiting 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 blocklimiting norm is cqual to thc structurcd singular value. Unfor
tunatcly, a computational mcthod for the blocklimiting norm is also unavailable
to date.
See. 3.5 General Robustness Analysis 8Q
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 31, 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 (314), 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 31 are
stated. Let

BA = {A E Q 1 ll A 112 5 1 )
This leads to the following theorems.
w Perturbatiqn P e r f o m Robustness
Measure
1
YIc(t) I ~ ) =I I b(tII) 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
llAll.c < 1
~ ~ I 5e 1~ ~ 2 A = Diag (A, ... A$ JlelJ2c oc l l ~ ~ l <l 1
p
llAll= < 1
I I I c 5~ ~1 ~ A = a g ( A A Ilel12<1 IIMllr < 1
back, where the variable feedback matrix is restricted to the block diagonal elements
represented in Figure 37. 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 wellknown 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
inequalities.
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 halfplane. The p x 4 loop transfer matrix Gll(s) in
corporates both openloop plant dynamics and any compensation employed. The
advantage of the statespace 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 realvalued
92 Modem Multiuariable Control Analusis Chap. 3
A E Q and det(jw1  A
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 thirdorder polynomials and those predicted by SSV
For secondorder polynomials they are identical, but for thirdorder systems they
are distinct. Let g ( s ) denote a strictly proper nthdegree transfer function, and let
( A , b2, c2) be the minimal controlcanonical form realization such that g ( s ) =
c2 ($1  A )  ' b 2 . Using the basic matrix identity Equation (321) results in the
following equation which relates root locus and the spectral set e(6), for some pa
rameter A:
+
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 thirdorder 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 wellknown 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
(
A = diag 0, l i m )
As an application of this theory, the Monte Carlo search algorithm is used t o com
pute a pe(A, Bz, CZ) for a thirdorder system, following Hewer et al. [1988(b)].
The system matrices are structured as in Equation (322) with a" = 1, a, = 2, a 2
= 2. T o estimate the 1.1. function, Equation (317) 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 Fnorm, which represents a 2sphere for the
Qa set that describes the robustness region. The Monte Carlo algorithm is described
as follows:
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.
Frequency At A2 A3
Monte
Carlo 7.753 .I473 e.2645 .1933
Penalty
Function 7.697 ,1229 .3 164 .?391
S 
L
...
1
1
S
V ,,,,,..,.. .""
., ......
,,...'
. .' ..I.
'1
(,,,..,.,.......l,.#.....'~~
E
I' 10' q
9
T
M
A
..
T
E .
S *. .
10"
1oO
 *C*r
'0'
FREQUENCY
('3
 RAD/SEC
Figure 38 Robustness of Real Perturbations (From [Hewer et at., 1988(a)] with
permission from S C S )
98 Modern Multiuarlable Control Analysis Chap. 3
For the example problem considered here, the original block diagram is shown in
Figure 38a. 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
statespace modeling and to illustrate Doyle's claim that a linear system can be
rearranged to match Figure 37 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 38a is transformed into one with scaled
perturbations, and results in the diagram shown in Figure 38b. 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 38a 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
38d. The original system inputs and outputs are no longer represented in Figure
38c. The exact entries for M are found in Figure 38e. 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 " 38a can be derived
using the following state variables x l , x2, x,, and x4 that are identified in the same
figure.
B2 = [ ; !]
0
0 1
2 =
c [. : :]
0 0 0 1
A = diag(Al, A2, A;)
Sec. 3.6 Robustness of Real Perturbatlons B9
factors into the product 6A = BzSACz. Before the statespace equivalence of M(s)
in Figure 38e 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
It is not difficult to verify that the matrix M(s) = SC2(sl  A )  ' B 2 is identical to
the transfer matrix in Figure 38e. 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 inputoutput 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 rootlocus techniques on the nominal system in Figure 3Sa, 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 38g 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 38f lists two possible near
minimum values for the joint variation in A. Figure 38g 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 noisedriven timevarying 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 whitenoise
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 (323) 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 (323) 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
NORMALIZED
CONFIDENCE
INTERVAL
I I I I I I
0 50 100 150 200 250
NUMBER OF SAMPLES
Figure 39 (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 39a. 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.
where jj(tf) is the miss distance from the ith run. Following Example 22, a 200
run Monte Carlo simulation was performed of the homing loop of Figure 2l5a(ii)
and 215b, 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 39b, 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.
...
u u:.> ... 1
u;,,,
u?.~~,~
u;, is used for the variances of the
individual states .u,.
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 offdiagonal 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 21 without the input term Bu. The covariance
matrix is defined as in Equation (326a) 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 (327) 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
(327) 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 (326a),
the covariance matrix Equation (326b) is given by P, = CPCr. In the discrete
state model given by Equation (5) of Table 21, assuming 4' ' = 0, the equation for
projecting the error covariance is given by
Fi ure 310a plots the RMS relative separation between the missile and target,
h, as a function of time, which war found by intcgrating Equation (327).
This value becomes the RMS miss distance at if, or the end of flight. That is, RMS
TlME lrecl
Figure 310 Covariance Analysis Provides RMS Trajectory Profile (From [Zar
than, 1979(a)], 0 1979 AIAA)
106 Modem MultIuarlable Control Analysis Chap. 3
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,
timevarying noisedriven 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
problems.
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) (332)
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 3lla. For the adjoint
system, however, if the observation time is the final time (to = t,), then only one
adjoint response must be generatcd. Equation (332) 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 31 la
and 3llb illustrate the way in which the two responscs are relatcd [Zarchan,
1979(a)].
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
(332) is substituted into Equation (333) to obtain [Pctcrson, 19611
~ c 3c.9 Aaolnt Method
I
Uto, t , )
crvation Time
0

Figure 311 (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
AIAA)
using
y(t) = ul 0
t
I*(tf  T . tf  t) U(T) di
Z = t  T
3 y(t)
=lI*(tft+z,yt)u(rz)dz (334)
108 Modem Multiuariable Control Analysis Chap. 3
If the final time t k is the time at which the response is desired, then Equation
(334) becomes
The concept of adjoint system can be extended to account for stochastic inputs. A
linear timevarying system that is driven by white noise has a mean square response
[Zarchan, 1979(a)]
where 4, is the spectral density of the whitenoise input in units of Hertz, and is
assumed doublesided and stationary. Unfortunately, the many computer runs re
quired to generate I(tf, t i ) render the direct simulation of Equation (336) imprac
tical, as mentioned previously. Substitution of Equation (332) into Equation
(336) yields
For the case in which the final time is of interest ( t = tf), Equation (337) 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 meansquare response at time t, of
a linear timevarying system driven by white noise. The RMS value of the output
of the terminal time is the square root of Equation (338). The RMS value of the
output of the original system in the presence of a whitenoise 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)]:
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 timevarying 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 33. 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 215a(iii) and 215b
of Examplc 22. 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 215b, 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
4
GLINT NOISE
4r ('t)
FAD. NOISE
=m
M = 11s 1 I
(1)
~2
AT,
2 q1( I f )
+ FAD. NOISE 1' I
+
9,( 1 0
GLINT NOISE I'
1
= ,/+ ~ E) IIY($)I
=~ ~ Y L ? + ~ + YI: (141
ADJOINT TlME l u )
(C)
Figure 312 (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 312b 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
noise.
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 twolevel 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. Quasilinear 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 313, using statistical lin
earization. T h e input x in the nonlinear system can be decomposed into a mean
component m plus a zeromean illdependent random process I., giving x = PI + r.
For this systcm, it is desired to replace the nonlinear element by an equivalent gain
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 zeromean 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
1
dr, K, =
1 ,.f(,,, + r)er21(zo2) dr
7 utn d 2 7 u~
Now, when it is known that the signal x itself is a zeromean random Gaussian
process, m is zero, that is, x = r and the equivalent gain reduces to
In Gelb and Vander Velde [I9681 inputsensitive 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
earities.
Example 34. This example, following the work of Zarchan [1979(a)] and
Example 22, centers around the limiter shown in Figure 314a. The describing
function given by Equation (343) becomes
xe  ~ 2 / ( 2 ~ 3
dx + * 1 lim
Jlim x2ex21(2u:) dx
(344)
+
lim
Lim
m
xe s1(2u:) dx
Modem Multlvariable Control Analysls Chap. 3
0 lb h ;O do 50 60
MISSILE ACCELERATION LIMIT (g'sl
(C)
Figure 314 (a) InputOutput 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)
REVERSED GAIN FROM CADET
PORTION OF PROORAM
('3
116 Modem Multluarhble Control Analysis Chap. 3
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
places
K,, = 1  (21fi)e""'~'(~":) (0.43618360  0. 1201676w2 + 0.937298w3)
(347)
where
Figure 314b 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.
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 meansquare
output at t,, for example, meansquare miss distance at the intercept time.
Example 35. This example follows the work of Zarchan [1979(a)] and
Example 22. It illustrates how CADET can be applied to the system of Figure 2
lja(ii) and 215b 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
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:
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 steptarget maneuver, and provide
insight into the relative stability of the system. Finally, another meritorious feature
of the SLAM concept lies in its selfchecking 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.
is entered into a linearized adjoint model of the original nonlinear system. This is
shown in Figure 314d. 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 314e 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 314f plots the total RMS missdistance 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
35 (see Figure 314c). 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 312b. 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 steptarget maneuver is also plotted in Figure
3149 as a function of adjoint time for various values of titi,,,. These functions
represent mathematically the impulse response of the quasilinearized system, but
cannot, as mentioned before, be used to compute the miss distance due to step
target maneuver. The sensitivity curve of Figure 314g 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 312c, would be found to
occur in a welldesigned 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 314g demonstrates that
the missdistance 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
8
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
.F
b the Missile in Order to Intercept the arget Using APr$or Advanced
duidance Law.
Linear Analysis
I I
10
I
20 30
I
40
I
50
I I
60
I
70
Missile Acceleration Limit (g's)
Figure 315 Missile Acceleration for Target Intercept: Nonlinear System Analysis
celeration limit greatly influences the RMS miss distance. Figure 315 shows that,
for 7g 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.
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 11state 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 11stable 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
Figurcs 316a and 316b 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.
COS
NUMBER OF RUNS
COVARIANCE
ANALYSIS
100
10 100 1000
NUMBER OF STATES
a) coat Cmparlmn lor Linear Syatema
COST
NUMBER OF STATES
Figure 316 (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 30ftlsec horizontal crosswind hitting a flight vehicle with a 300ftlsec
forward speed will create a 5.7deg 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 (centerofgravity 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 steadystate 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 closedloop 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
computed.
Modem Filtering
and Estimation
Techniques
As the caption of Figure 41 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
Deterministic Stochastic
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 online applications
involving digital computers. Modeling of a timevarying 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
termination of spacecraft orbits, satellite tracking and navigation, digital image pro
cessing, economic forecasting, industrial processes, and nuclear powerplant 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, spacecraftbased optical observations were
combined with earthbased radiometric observations to accurately determine the
orbit during the Jupiter encounter approach phase. In the case of nuclear powcr
plant instrumentation, online instrument failures are detected and faulttolerant
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
42.
Sec. 4.1 Linear Minimum Variance Estimation: Kalman Filter
i(k,+
Figure 42 (a) System Model and Discrete Kalman Filter (b) Linear Continuous Kalman
Filter
Table 41 defines the wellknown 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 42a. Table 42 defines the multiple
rate Kalman filter.
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
21 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 21 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
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.
  
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 multirate measurement equation.
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.
This last equation is nothing more than Equation (4la) with P = 0. The resultant
Kalman filter gain is the constant matrix
Given that these arc the statc equations o i thc stcadystate 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 (44) can be
obtaincd by analyzing the roots of the characteristic cquation corresponding to det[sl
 (A  K,H)] = 0. A sufficient condition for the steadystate 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.
;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 problemoriented 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 linearoriented Kalman filter known
as the extended Kalman filter (EKF).
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, timevarying 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 21. 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] = (45a)
where
K[2(t), t] = p(t)HT[2(t), t ] ~  ' ( t ) (45d)
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 (45b and 45c). 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
43. 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 (45a to 45d) 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.

E~s.(4k&d)
Calculation of
Eqs. (45a)
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.(4Sb) Estimates
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 secondorder 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 continuoustime nonlinear system with discretetime nonlinear measure
ments (see Equations (2a & b) of Table 21) is presented in Table 43, from which
the linear case of continuous dynamics, discrete measurements can be easily deduced.
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 21, 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 semidefinite 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} PI
(4)
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} pl(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 $K,,(k)] PC
where Kk= Pi K L ( ~() q ( k ) Pi ~ l ( k+) RI,)'
,=I
zk
= { z ( l )... z ( k ) }and, under the Gaussian assumption, the probability density func
tion of the innovation from model e is
Using Bayesian rule, the a posteriori probability that model E is correct at time k
is
PIAe I z k } = P { A ~ } P { ( zAe}
~ P{A,}P{zk I A,} = L e ( k ) p e
j=l j= 1
(46)
It is assumed in Equation (46) that the same model has been in effect from the
initial time. However, this is obviously not true if the model changes at some time
during the interval [ I , k ] . In view of this, a "slidingwindow" likelihood function
has to be used. It is now assumed that the model changes at time k  N, so that
one considers only the likelihood functions of the measurements from k  N + 1
to k. Each filter starts from the same initial condition at k  N . The likelihood
function corresponding to model e is then
where v j ( k ) is the innovation from the filter based on model j. The "effective"
memory length is (1  a j )  I . Let S j ( k ) = S, be the steadystate covariance of the
innovations from filter j. Then, it is possible to replace Equation (47) with
 I
1 1
P_I ( k ) = ~ , e  ' ~ " ~P;(O)
' Cce1 ,P<(k)
Pe(O), C , = I S j I ?(I  a,) (4 10)
/:I
where Pc ( 0 )are the prior model probabilities. If these prior probabilities are equated
to the posterior probabilties, then Equation (410) simplifies to Pi =
I
/
c,~$"~) cce 7,:')
1
 A typical multiple model estimation is drawn in Figure
44 in which the combination of the estimates is done according to Equation
(48).
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~
(411)
Ssc. 4.3 Predlctlon and Smoothing
System
+
Dynamics
Measurement
System
Combination
of Estimates 
$ p
+%
ICalculation of
Probabilities I
Figure 44 Multiple Model Estimation
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
(411) 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:
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 EKFEquations (45b and 45c) with R' = O and
K=O
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)
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 4Sa, while the linear continuous smoother can be derived
and is summarized in Table 45b. Finally, the stnoother of this section is called a
fixedinterval 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].
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 reducedorder filters. At this stage, possible
approximations to the optimal gain are also evaluated. Following this, a covariance
sensitivity analysis and Monte Carlo simulationsare conducted, after which it is
then possible to select a final design based o n a tradeoff 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.
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 (4la), 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 openloop
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 lowbandwidth 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 halfplane), the
filter is a highbandwidth 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] (41 8)
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.
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 45a and is discussed here. Using Equation (417). the transfer function
from x(t) to 2(t) is
k(3)  kh k(0) kh
 where the DC gain is   (420)
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 45b. For a given value of r, increasing the value of q in the
design results in increased values of the gain k. From Equation (420), 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 steadystate conditions the model dynamics equations imply that x = w l a
= constant, while, from the state estimate equation, i= 0 and
Imaginary
Real
b\ a
Imaginary
qh qir = 0
k = k =0
1
a a
Real
OpenLoop
Model ~ ~ n a m iPole
cs
Figure 45 Selection of Noise Intensity Parameters q and r
=)[a + g a 2
i ( t ) , given by Equation (418) 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 45c, x(s)/z(s) = Ilh.
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 46a. As shown in Figure 46b, 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 halfplane, 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 (4lb), the error dy
namics Equation (422) is a stable linear system driven by white noise. Substituting
in Equation (423) the Kalman filter gain gives the same equation for Pas in Equation
(4la). As t becomes large, the error covariance P(t) in Equation (423) satisfies
P(;) = 0 and the steadystate covariance matrix is obtained by solving the linear
Lyapunov matrix equation,
The Kalman filter gain minimizes P(t). That is, P ( t ) from Equation (424) using a
+
AKH
b I 'f+@ I
general filter gain matrix is always greater than P(t) from Equation (42) using
the Kalman filter gain Equation (43).
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'= (425)
2 2
+ ... + 
u*;pi
((t)  2,))
u*;,,
+ ... +  @,,(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 ) .
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
ducedstate 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.
Trace P(t)
r
I
For Minimum Settling Time Filter
Trace P
Figure 47 (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 tradeoff)
I I
I I
I I n>m>P
I I
I I
I I
I I
i'
al
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 squareroot filter.
filter dcsign proccss in the future and may at the same time substantially improve
performance (Bahcti, 19871.
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 leastsquares 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 longterm op
eration of the filter. Note that the steadystate 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 online 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.
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
roundoff 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 boundedstate 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 unboundedstate 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 roundoff 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 squareroot propagation, which are capable of producing an ef
fective precision that is double that of the conventional Kalman filter. A procedure
similar to the squareroot formulation and called UD 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 UD 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 squareroot 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 (424) arc in the right halfplane, 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 timevarying 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 (423). ,\.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 halfplane 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 (423) and C ( K ) > 0 is
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 firstorder or secondorder GaussMarkov 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 discretetime 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 "biasfree" 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 biasfree 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 reducedorder model may,
however, be degraded filter performance. Algorithms for analyzing the meansquare
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.
Timccorrclatcd 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 timccorrelated 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 discrctctime niodcl
cases,
where wl ( t ) represents zeromean 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 (426)) is a GaussMarkov process, then it can always be expressed by
a linear dynamic system with white noise forcing function only
where li, is zeromean white noise with variance Q. Equation (427) 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 (427). The coloredmeasurement 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 (428)
where w and v l are uncorrelated zeromean white noise with variance Q and R ,
respectively, and Equation (428) defines the correlation properties of the mea
surement noise v, in which E[v(to)]= 0 and ~ [ v ( t ~ ) v ( t =o )Vo.
~ ] Define
However, this case is easily handled by employing Equation (4ld) instead of Equa
tion (4lb) with S(t) = Q ( t ) N ( t ) T ~ ( t )Thus
T.
x* 9 i ( t )  K ( t ) z ( t )J i* . (432)
= (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 (431) in the filter calculations. Finally, to
determine 2(t) and P(t) from Equations (431), 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 (429). 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
(431) 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 (431) with P = Po.
This section presents, derives, and verifies the continuous form of first, second,
and thirdorder 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 loworder 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
filters.
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 48a ~vhicll
Sec. 4.6 Fllter Approach to Estimator Deslgn
Complementa~~lKalman 159
Kalman Filter
+Rate Limit
I
Complementary Filter
LowPass
Comparison
k, t bu, 4
Z= x t v = xl+bu1+v z = xtv. v = high frequency noise
1
U X
includes the optimal state estimator. With regard to Kalman filter formulation,
Z(s) = V c =
s+k s+k s+k s+k s+k
1
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.
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 (436) yiclds
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.
+
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 (436),
u = V I = inertial acceleration, ul = 0
) c l ( s )  K V C ( S ) I ( Sf k ) . Thus, V r ~ s
Applying Equation (439) 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
Roll Rate
Pitch Rate
Yaw Rate
 Body to
Euler Axis
Transformation
Kalman a l t e r
L.
Complementary Filter
High Pass
z
4 =Fl
nT nv
s*
Low Pass
+ 2 j wn s +a);;
G(s) Low Pass
Comparison
el + bu2 L
z = r + v = n l + b%+v z = s+v. v = high frequency noise
eration measurements. Consider the model for the secondorder dynamics system
in which x l = x = position, xz = i = rate, and 14 = x = acceleration. T h e equation
o f motion in statespace 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 zeromean 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 zeromean 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 secondorder complementary filter and Kalman filter is also shown in the figure.
As in the firstorder 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 zeromean 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) (440)
i2(t) = ~ ( t+
) k I [ z ( t )  . t l ( t )  6112(t)]
where gains k and k l have been expressed in terms of the secondorder parameters.
Therefore,
i(s) = 2(s) = kl(s) + bU2(s) (442a)
= [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) (442b)
= [(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 lowpass filters. The estimator gains k and kr
166 Modem Filtering and Estimation Techniques Chap. q
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 (441a),
L
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
U.
the integral of the accelerometer input. The position estimate is the firstorder com
pletnentary/Kaltria~~
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 (423) are employed here with
K = 0 since they hold for an arbitrary equation Kmatrix.
The error covariance will grow according to Equation (447) 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.
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
A

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 lowpass filter. In the Laplace domain,
s
from Equation (442) 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 419: 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
of a pure a 1and a filtered tic; and (2) an & that consists of a pure &I,, and a filtered
kc.
Example 421: Secondorder sideslip estimator. This example fol
lows the firstorder sideslip estimator problem formulation in Example 411.
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 firstorder esti
mator, while 6 is not provided by the firstorder estimator but is provided here by
this secondorder 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
U 4
a) Formulation 1
U

,A
Wi3
b) Formulation 2
Figure 411 ThirdOrder Complementary/Kalrnan Filter
Sec. 4.6 Comp(ementarylKa1manNlter Approach to Estimator Design 171
rate, and x = acceleration, and the equations of motion in statespace form are as
follows:
XI = X? + 111 + LVI, E \ W I ( ~ )=
] 0. )] 
E [ I v ~ ( ~ ) w ~ =( Tqlb(t T)
The preceding model dynamics are illustrated in Figure 4lla 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
r
+ b112, x = i 2 + U I , and x = 23+,,.
A
As shown in Equation (452), 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 lowpass filters. The algebraic Riccati equation is
43 = 2317 + k l ( z  21  bu2)
Applying Equation (4la),
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 (449b) to be a colored noite,
then following Equations (428) to (431) 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 (431) 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:
5.1 INTRODUCTION
'This chapter has becn technically cditcd and improvcd by I l r . J. Stanley Ausrnan of Litton
Guidance and Control Systcms.
176
Sec 5.1 lntroductlon
Common Requiremenu
Comparison of
Gimballed INS &
Strapdown INS
I I (Sec. 5.6)
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
INERTIAUDOPPLER
1 MILE ABSOLUTE
Sec. 5.2 Common Requirements for lnertlal Navigators 181
CHARACTERISTIC
these navigation devices are given in Figure 52c. 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.
system for tactical aircraft, whereas landfall position fix, velocity, and nuclear hard
ness are all required of the navigation system for longrange 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 shortrange missiles and tanks, size, weight, and low cost figure
predominantly in the requirements, whereas mediumrange 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 fixedwing military aircraft to be equipped with an inertial
navigator [Barnes, 19871.
TARGET
/SEARCH
  ~
TARGET
LAUNCHING
AIRCRAFT
FLIGHT ORIGINATION
POINT
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 crosstrack error budget should be limited to
one half of this, or 0.5 km. The alongtrack error is dependent on the allowable
false alarm and desired detection probability criteria. Obviously, as the alongtrack
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 alongtrack error ofless than 1.5 percent of distance traveled.
The future will see an increased demand for more agile missiles for both air
and groundlaunched applications. Gunlaunched 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
fiberoptic 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.
5.3.1 Coordinate S y s t e m s
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 earthfixed set of coordinates (Figure 54), we can, by a
series of two rotations first about the zaxis through the longitude angle and then
about the negative yaxis through the latitude angle plus 90 deg, transform from
the earthfixed coordinates into a navigation coordinate system. In this case the
navigation coordinate system is defined as one in which the xaxis is north, the y
axis is east, and the zaxis is down. The transformation is depicted graphically in
Figure 55. Figure 56 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
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
NAV FAR'IH FARlH
angle +. The second rotation is about the new y (rightwing) 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
7
1.
58. In the figure and xSare inertial position and velocity vectors, respectively,
in a spacefixed 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
MEASURED
INERTIAL

A ~ I

I:

;I I    ...
is
* (X*,Y.,Zs)
VELDCITY
  INCREMENTAL
&AI

INERTIAL VELOCITY VECTOR
I SPACEFIXED
VELOCITY VECTOR

XI. TOTAL INERTIAL VELOCITY VECTOR

x l TOTAL INERTIAL ACCELERATION VECTOR

Ap TOTAL GRAVITATDNAL VELOCITY VECTOR

i p TOTAL GRAVITATIONAL ACCELERATION VECTOR  1
NOTE: SUBSCRIPT i DENOTES INITIAL VALUE. I GRAVITY
CotAFWfATuN
SPACEFIXED
POSITION VECTOR
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 58 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 spacefixed coordinate system which has its origin at the center
of the earth. The diagram further shows the calculation of the acceleration in the
spacefixed coordinate system as thrusttomass 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. Forovertheearth flight vehicles,
the required position vector and velocity vector x are not expressed in space
fixed coordinates but in earthfixed coordinates, as seen in Section 5.3.1 [Haeus
sermann, 19811.
Inertial Navigation Chap. 5
Following BarItzhack 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)
   +
Coordinate
Thnsforma~on Earth Caord.
lian~form~ti~n
sq     
We, Ve
Navigation & 6 6
Procasing +e"E
h
Be
2) AV Ptocssing: (Vekxity Inaancnt)
+ xD YES%
&Ordinatc +
V~
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 59 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 [BarItzhack and Bcrman,
19881.
where v , r, and I) are, respectively, the velocity, position, and attitude error vectors;
0 is the earthrate vector; w is the ang~~larrate vector of the true coordinate system
with respect to an inertial frame; V is the accelerometererror 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 gyrodrift vector.
190 Inertial Naulgatlon Chap. 5
It can be shown [Leondes, 19631 that in the local north, east, and down coordinate
system,
a=
R cos L
[ 0
 R sin L
] (52)
.=[ icos L
.i
 A sin L
1
and ); is the longitude rate. When Equations (51) are resolved in the true frame,
(54)
nine scalar differential equations are obtained that can be put in a statespace model.
If the expressions given in Equations (52) to (54) are used, the resulting state
space model is shown in Figure 510. 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 (51) 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
dt
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 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 (55) is not suitable for use in a Kalman
filter since the accelerometer error and gyrodrift rates are not whitenoise 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 whitenoise 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
is,
When r . ~ ,YE, r D , and v ~ are removed from the state of the INS error model of
Equations (55) and the resulting model is augmented with Equations (56). the
following is obtained:
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 510 State Equation of Terrestrial INS Error Model (From [Barltzhatk 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
(55) and (57) from the point of view of a control analyst is presented in Bar
Itzhack and Berman [1988].
Outer
Accelerometers)
the gyros sense the direction of zero earth rate (eastwest). 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 eastaxis. The north accelerometer then generates a pseudo velocity due to the
platform tilting about the eastaxis. The pseudo velocity is then used to drive a r
gyro torquer and its associated gimbal servo to achieve the proper alignment.
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 errorfree 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 514. The two mechanizations for the platform attitude reference are the
space stable and local vertical. The platform spacestable inertial mechanization is
given in Figure 5l5a. Figure 5l5b illustrates the case in which the geographical
northpointing locallevel navigation frame is the instrumented stable member frame
for a gimbal system. The local~levelframe is the most common for gimbal systems.
The localvertical frame is usuallv a form of a .,fiecarimirth reference frame in which
no attempt is made in a gimbal system to point one of the stablemember axes north,
as illustrated in Figure 515c. 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,
19871.
Figure 516 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
&Attitude
L     J
Figure 514 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 xaxis 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.
t
Alt~tudeInformation
Init~olc o n d ~ t m s
I
Nav~got~on
Posttlon ~n computer
* Lotitude ( L)
Lonpitude ( I I
Alt~tude( h )
lnert~olcwrdtnotes
t
itotiliied
platform
(B)
Indial Altitude
Wndtt~otn mfamoton
Velocity lv7
ws~tlon(L,I,&
Figure 515 (a) I'latforn~ SpaceStable Inertial Mcchanization (b) I'latform Geo
graphic LocalVertical Mechanization (c) Platform FreeAzimuth 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)
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 socalled 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 320), the altitude reference h~ is given by an
altimeter and thus generates the flight height error E,, to damp the loop. In Figure
520, thc feedback gains K I and K 2 can be selected so as to nicely damp the vertical
100~.

Gyro Master
28V DC
I
26AC 400 Hz
Vehicle Altitude
Display Indicator (ADI)
Figure 517 Interface Information for Gimballed INS
See. 5.5 Strapdown Inertial Navlgatlon System (INS)
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 bodytoNAV direction cosine matrix is computed using strap
down bodymounted gyro outputs. Then the horizontal and vertical accelerations
are computed analytically using this direction cosine matrix to transform acceler
ometer outputs into locallevel navigation coordinates.
A simplified block diagram showing the IMU and navigation computer is given in
Figure 523. 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 .
bVX, bVyv 6VZ VELOC I l Y ERROR
transformed onto locallevel 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 523. 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 526. Figure 527 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
[1986].
The strapdown navigation computer may transform the bodyaxis sensor data into
either inertial coordinates or into geographic localvertical coordinates. The strap
down system computing in inertial coordinates is given in Figure 528a. The strap
down system computing in geographic localvertical coordinates is given in Figure
328b. Figure 529 shows a generalized block diagram for strapdown inertial com
putations. This diagram shows that the computations are partitioned into two parts:
a highspeed section that transforms the bodyaxis measurements of acceleration
and attitude rates into navigation axes, and a slowerspeed section that performs
the navigation routines just as they would be computed in a gimballed system.
Typical computer requirements for a strapdown NAV computer call for a 16bit com
puter whose properties include a speed of 250350 KOPS (the computer throughput
MEClL4NlZATlffl b P ~
NORTHPOIM ING
{NO POlAR CAPABILITI)
tan* &P
Y
.pX sccZ+b $
FREAZIMUIH 5 4 * "a '%
WANDlR FWUKl u
Y
box *I bo
Y
AZIMUIH
IWUR VNlPMR
CAPABILITYI
g Gikw
lMISPHRL
 I  IC, b y + 4. bpi]
L M P SIGN
SOUIHRN
lMISPHiE
.

1. I. 2 m ? i p t i m ~ l t d li 0 8 t e .XU
P ~ u i (i+qclZ
~ ~  cclZ 
b tic altitude
b&r
~~polar rn UUd to1181 ndiuS Of the
hhth
Vppr s i w
(DlU f n ~ t t . s i y 01 tho u r t h
I(0rtb.m
Umbilib) I*.isobe~cn iA: $49 ' %+1
'
4 tic latitld*
1 siw I$, t 1)
htbern h laeitpd*
Hais~b*r*
a h h t h d e r la
Figure 522 (a) Level Channel Error Model (b) Generalized Mechanization for LocalLevel
FreeInertial Navigation Systems (From [Martin, 19861 with permissionf*om Litton)
204 Inertial Navigation Chap. 5
I
I I 
r7
i
I
I
I
I
SENSOR
BLOCK
 IMU ELECTRONICS
REBALANCE
COMPUTER
110
( I
I
I
I I
I ' CONTROU
I
I
I
I
GYROS
ACCELEROMETERS  ELECTRONICS
POWER SUPPLIES.
TIMING
CPU
MEMORY 4 n
I I
I
I
DISPLAY
UNIT I
I
I
I J
I I A
I
LI.II.IIII.IIIIIIIJ Ld
(A)
Figure 523 (a) T\pical Strapdo\vn System Simplified Block Diagram (b) Strap
down IMU and Na\.igation Computer ((aifitn marf fit^, 19861 u~irltpcrmissiotl.~otn
Lirrot!)
may rcach 3.0 MIPS to support a wide bandwidth autopilot for high dynamic ap
plications), a memory sizc of 1215 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 1330 g,
50010000/scc maximum angular ratcs, a 100 g (20 MS) shock, a MILE5400 (curve
IV) vibration, and a MILE5400 (Class 2) thcrmal.
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
SAME AS LOCAL LEVEL AND
ALTITUDE
f REE INERTIAL
NAVIGATION EQUATIONS
TO BODY AXES
LOCAL LEVEL AXES
RESOLUTION
BODY AXES
t C
OCI 1 = [OX' Oy' \]
T
: Oi 'Pi
Reference h r d i i a t e System
Vehicle Body CkxdiiteSystem (Locally Level,A Space E m )
A
0 A0
P' &
R' Attitude )Attitude
MY (0, $, q)
Vchick
4 Inertial and
AVx (Y ,V y ,V,)
Bod' sensor
mponents
a;
A
velocity
+Velocity
Conuol 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 singledegrceoffreedom
gyros, a gyro drift, called anisoinertiainduced 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
BODY COMPUTATIONAL
FR M FRAME
(We @,*I
 SELFCONTAINED
ALIGNMENT 
EQUATIONS
.
Figure 526 Typical Strapdown Software Requircmcnts (F~OIII
/hlorritr, 191161 urirh
perrnisriorl j o ~ nLirrort)
Sec 5.5 Strapdown Inertial Navigation System (INS)
design, and the resulting navigation system performance when selecting a particular
type of gyroscope for a strapdown system. (2b) Gyrotorquing 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.
 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 528 (a) Strapdown System Computing in Inertial Coordinates (b) Strap
down System Computing in Geographic LocalVertical 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 )
The functional flow of strapdown navigation systems is dividcd into five blocks as
shown in Figure 531.
Sec. 5.6 Cornparkon and Analysls of QLmbal versus Strapdown

NavIpcrtlm Conpltath
II lbmmon to strquklvm and Ohlbalkd Syrtrml
I
Figure 529 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)
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 bodyreferenced 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.
+"I
P
a"'
+
uX
C $
OX
C
Sec. 5.6 Cornparkon and Analysis of Glmbal versus Strapdown
Main Lie
4
{
Senror Ibt
Stan & Self%st Block 1 Ucctronicr %t
4
Pmcessor %st
'Iiansfomtion Matrix
Attitude Update
1 Block3{ {Sensor Uibntion
Gym OP Compensated
Add Lnb Rate & ?hnrpon Rate
No
<
Navigation
sensor error. In a localvertical 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 scalefactorinduced
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 singledegreeoffreedom gyros is output axis inertia. In this case, the
strapdown gyro has a drift error proportional to the angular acceleration about its
output axis.
Computationinduced errors in strapdown systems are those navigation errors
introduced by the incorrect transformation of the bodymeasured 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)].
Sensorinduced navigation errors propagate differently in a strapdown system.
As noted in Schmidt [1987], in a gimbal system during selfalignment, 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 tilterror contribution exactly cancels the accelerometer biaserror contri
bution. In a strapdown system, where selfalignment 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 velocityaiding 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 wholenumber 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, positionfixing radar, satcllitc navigatiol1
Sec. 5.7 External Navlgatfon AIds 213
An unaided INS has the advantage of possessing allwcather, allattitude, and non
radiating characteristics; highdynamic response; and selfcontained navigation. Dis
advantages include unbounded position errors and oscillatory velocity errors, ab
sence of inflight 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.
Referring to the ninestate error propagation equation shown in Figure 510, 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 positionaiding devices. Doppler
radar and GPS can both provide velocity updates. Another very important velocity
update is available on stationary vehicles, namely a zerovelocity update. Stellar
observations made by a telescope mounted directly on the stable element of an INS
provide attitude updates, actually psiangle updates. Other examples of attitude
updating are stored heading alignment and attitude matching transfer alignment
from a master INS to a slave INS.
TIME +
AUGMENTED INERTIAL NAVIGATION
Figure 532 Augmented Inertial Rationale (From [Martin, 19861 with permission from Litton)
REQUl RES
OCl TY
COST 100
10
0.001~1~~ 0.01~1~~ O.l0lH~
I
L0 ' 1 ~ ~
0. lMlHR LO MllHR 10 MllHR
PERFORMANCE
Figure 533 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 PositionVelocity Inertial. The advantages of aug
mented positionvelocity inertial mechanization include: multibackup modes; pro
tection against spoofing o r position; rateaid for position sensor; good dynamic
response; inflight startup and alignment; and bounded position, velocity, and head
ing errors. When an augmented inertial mechanization is no longer selfcontained.
that is, it requires external cooperation, its dominant disadvantage is its susceptibility
to RF noise and propagation errors.
Among the current generation of positioning and navigation aids, it is perhaps GPS
coupled with newer and betterdesigned fieldqualified receivers that provide un
precedented positioning and navigation capability.
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 lowrate 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 mediumdynamic applications, respectively. In these receivers the channels
are time shared among satellites and housekeeping chores [McGowan, 19871.
pseudodeltarange (I'DR), which is a measure o f the range change during the in
tegration interval.
Performance. Table 51 lists the error budget for the PR measurement,
assuming that the dualfrequency ionospheric co~npensationtechnique is used. As
shown, a measurement error of 5 m is predicted. The test results indicate that the
highfrequency fluctuation of the error is less than 1 m. The error was dominated
by a bias component that ranged from 710 m [McGowan, 19871.
Uncorrected
nOviVion
vorlo ICS ~ovigationcorrections
pseudoronge
and pseudo
delta range
Figure 534 Aided InertialGPS 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 Lband (1.0GHz)
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 2001,500 ft in range/(bounded) radial
position error and 0.5 deg1.0 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 Lband
frequency and the beamwidth of the transponder antenna limits the bearing reso
lution. In Figure 535a, 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 dualTACAN mechanization.
By implementing a ~ U ~ ~  T A CsystemA N or rangerange mechanization, the po
sition error is reduced to 500 ft (CEP), as illustrated in Figure 535b [Martin, 19861.
I TACAN STATION
ACAN STATION R
Figure 535 (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 536.
LORAN is a timedifference 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 twodimensional 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 50500 ft or more, contingent upon the geometric dilution of precision (GDOP).
The resolution of the transmission is approximately 0.01 psec X GDOP.
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.
Highdensity 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 1001,000 ft and
resolution is 100200 ft. Figure 538 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 539 shows the mechanization of the TERCOM system.
MASTER
STAT l ON
(A)
(B)
Figure 537 (Continued)
Cruise Cruise
Missile Elevation Map
Stored In Onboard
Computer
Guidance
Update
T A R G E T AREA
Data Recommended
Cornlation Flight Path AutopiSol
4
h I
Figure 539 TERCOM Mechanization
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 astroinertial 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 psiangle
corrections (see Figure 510).
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.
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 timevarying 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 openloop correction, and the other is called closedloop 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 540 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 openloop 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 closedloop 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.
SUBSYSTEM
SENSOR ERROR SUBSYSTEM
SUBSYSTEM ESTIMATES
c tl  CORRECTED OUTPUT
ERROR 1
MEASUREMENT
ERROR
VARIABLE
DIFFERENCES MATRIX
SUBSYSTEM SUBSYSTEM
ERROR CORRECTED
OUTPUT
ERROR
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 errorcovariance analysis to inertial platform errors
is constructed as illustrated in Figure 541. As presented in Chapter 4 (Sections 4.1
and 4.2), the fiow diagram in Figure 541 [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 541 is as follows: 
I
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 541, 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 fourdimensional
problem for states 1, 2, 8, and 9 is of the form
For conventional filtering, the position error (in ft) ncvcr converges to zcro, as shown
in Figure 542a. For Kalman filtcring (optimal filtcring), on the other hand, the
position error (in ft) convcrgcs to, but ncvcr rcachcs, zcro, as shown in Figure
54%. Figure 542c displays a table showing thc relationship of mechanization and
performance.
Sec 5.8 Integrated Inertial Navigation System (IINS)
Figure 541 Flow Diagram (From [Tucker and Stern, 19861 with permission from
AACC)
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
Figure 542 (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 bodyaxis 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.
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 543.
The sensors are numbered from 1 to 6 with 1, 2, and 3 corresponding to the
x, y, and zaxes of the body aligned unit, and 4, 5, and 6 corresponding to the
x, y, and zaxes of the skewed unit. A functional block diagram of an ISFCRS is
presented in Figure 544. 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 545. Figure 546 shows
a conventional approach of comparison with MIRA [Boggess, 1986; Young et al.,
19831.
ISS DHS is shown in Figure 347a. 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 2fail 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 FCSIbendingmode 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
547b.
Another type o f IINS is the IISA. This is an integrated inertial sensing assct1lbJ~
(IISA) that uses six ringlaser gyros and six inertialgrade accelcromcters in two,
separated clustcrs. It is a system that is well suited to the combined navigatioll and
flight control sensing needs of modern highperformance flybywire 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
I
I
iI
I
?={
I
I
MILSTD.1553A
USERS
Right Control
Navigation
Fire Control
Weapon Delivery & Trajectory
Sensorhcker Stabilization
NAV Axis
I Cockpit Display
I
L  I
l
Figure 545 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 5482,
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 548b. Figure
548a 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 180deg rotation about yaw relative to the left INA, the six
sensor axes are then distributed uniformly about a 54.7deg halfangle cone. No
two axes are coincident, nor are three in the same plane. Thus, any three sensors
may be used to derive threeaxis outputs in vehicle axes after suitable computer
transformation. The IISA development uses highreliability 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 548c) consists of five assemblies: two identical INAs containing the
inertial sensors and navigation computers (shown in Figure 548d); 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 546 Conventional ~ p p r o a c ~
to Sensing and Flight Control Referellce
System
ORTHWWAL OW111
RAT1b UtR
ACCIL DATA FIST#MS
PITCH L ROLL
IU OAT A
HANDLINO
SYSTEM
DELIVERY
INTEORATID
N A V SVSTEM
CONTROL
SYSTEM
SKEWLD
OVROS   AIRCRAFT
DYNAMICS
I
ACCEL8
I I
SKEWED
ACCELERWETER
USE EXTRAPOLATED ESTIMATES
TO REMOVE BENDINO
ACCELERATIONS
USE GYRO RDMS OUTPUT 6
EXTRAPOLATED ROTATIONAL
ACCELERATION ESTIMATES
TO REMOVE KINEMATIC
pr4 ACCELEROMETER
ACCELERATIONS
7 ESTIMATOR
TE t+ TO SAHRS 6
CONTROL L A M
PITCH 6 ROLL
ATTITUDE
FROM SAHRS
ELBP
USE EXTRAPOLATED
ESTIMATES TO REMOVE
BENDINO RATES
I 1
(B)
Figure 547 (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.
INA
NAY
ILICIRONIC~
MILStDIS538
 N A V TO * I C
 
3 RLOS 1014 Hx
rcro AIC(MH.,

3 ACCtU d,
3 HVPS
3 fC ILICTRONICS
PRoClssoR = +(I
PC T O A I C (80 Ha)
DCA
INA

A
3 RLOS
3 ACCIU
3 HVPS

a DCA
PROCESSOR 
  ,I
I,
 rc TO AIC (80 HZ)

3 PC I L I C I R O N I C S
NAV
c rcc TO AIC POHZ)
tLtCIRONIU     N A V TO AIC
(C)
INSTALL
Figure 548 (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
TABLE 52
Config
CONP'IGURATION VS. ATTITUDEWELOCITY REDUNDANCY
Attitude
redundancy
Velocity
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
Effect of gyro dither. Basically, IISA utilizes ring laser gyros that dcpcnd
on a small amount of angular mechanical dither motion to avoid lockin 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 lowfrequency 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.
~ 
Figure 549 (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 )
SlWSql DATA
o i s taurrrons
o ISOLATE 2 S x n a t A N E o u s FAILORES
o RAPID DETECTION mon UNFILTERED omms
0 FILTERED OUTPUTS MR P E R W N I FAIL
O SELFADJOSTING TRIP E E L S
( SENSOR SELECTION
+I 1
o SENSOR PERCORCUNCL INDEX
o RESULTS
0 USE BEST 3 OR 4 SENSORS
o 29 COI4BINATIOll WJATIONS
0 LEAST SQULRE CSTItiATE FOR I) SEXSORS I
1
h I G 8 T CONTROL OUTPUTS
(A)
1 INA1
TRAJECTORY I
I
u I
SYSTEM
DYNAMICS MODEL
I
G EFFECT
I
a
GYRO DESIGN EQUATION
LEVER ARM
COMPENSATION
? I l
ERRORMODEL
I I ACCELEROMETER
PARITY
I QUANTIZATION
FILTER ANDSAMPLER
I
ACCELEROMETER DESIGN
EQUATION I
240 Inertial Navigation Chap.
cessing to derive bodyaxis 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 549b. 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 549c.
Scale factor
Bias
Alignn~etit
llcsolutioi~
Range 4(X) drciscc 20 g
Error sources. The primary sources of error associated with llSA follow.
(1) Axis alignt~~cr~remor: Achieving the 12 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
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
550. 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.
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.
Inertial Navigation
+ Data Bus
MILSTD1553B
(h4ILSTD1750A) Reference
curacy requirements given by this table have the same attitude and heading accuracy
requirements and slightly different position accuracy requirements.
r 
1 HUNS 7I
I
1I DVS
RDN 808
I
I
I
I
t  1I 

Q
II I
I
I I
I I
I I
I
I
I mu  INS
I
,
I , ucs
II I
I
RADM
ALTII(E1ER
 *
I
I
I
I L II I
 
L
fl _I
DUAL MILBUS lSS3B
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 353a. 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 533b.
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 553b, 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 553c. The block diagram is intended to show that the output of the filter
u, as well as the state estimates 8 and 4 are continuoustime variables, whereas the
inputs to the filter are discretetime variables. (In actual implementation, the con
tinuoustime variables are also discretetime 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 seekersampling 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 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 .
Sight
Angler

.
PI C001101NATE
YziT
.la



*rr.L TIIANSFOR
UTIO* A"TOPILOT
".A
C
CO*I"OL
b SURFACE
DEFLECTIONS
.
8 0 V
DIIECTIOY
UKI*E
UIlllX
l*TEt"AT,O*

 UlC"
Q tf'
P
 I O U
 "A.
R
Figure 553 (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)
MEASURED
INERTIAL
LINEOFSIGHT
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 553c remain open and the missile
continues by dead reckoning to the interception.
(B)
Figure 554 (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) LowCost Version with
an Advanced Processor (c) Structure of the Navigation Software Development Sys
tem (NSD) (d) Block Diagram of LineofSight 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
nlmAm MIUL SDGDR MIA RUS REAL TIME M R ~ D SIW
m 1 m MTA ElWl
m 1 m
AafWT. RN
PATE RDCESSIrn
CONVERSION
m 80285187
PISUL
(C)
1 LOS CLOS
 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 61 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
conditions.
GuidanaRarssor
( k c . 6.1)
I I
Guidasc Mission & Pctfonnance Guidana Algorithms
I I
LOS An& Guidance LOS Rate Guidance
( k c . 65.1) ( k c . 6.5.2)

FixedLea Navigation
Guidance (PNG)
and Its Variations
I
Sensitivity and Comparison of Guidance Laws
(Sec. 6.5.3 & Chap. 8)
I
Guidance Mode Options
(Sec.6.6)