@Seismicisolation

@Seismicisolation
Mechanical Engineering Series
Frederick F. Ling
Editor-in-Chief

@Seismicisolation
@Seismicisolation
Mechanical Engineering Series

Wodek Gawronski, Modeling and Control of Antennas and Telescopes

Makoto Ohsaki and Kiyohiro Ikeda, Stability and Optimization of Structures: Generalized
Sensitivity Analysis
A.C. Fischer-Cripps, Introduction to Contact Mechanics, 2nd ed.
W. Cheng and I. Finnie, Residual Stress Measurement and the Slitting Method
J. Angeles, Fundamentals of Robotic Mechanical Systems: Theory Methods and Algorithms,
3rd ed.
J. Angeles, Fundamentals of Robotic Mechanical Systems:
Theory, Methods, and Algorithms, 2nd ed.
P. Basu, C. Kefa, and L. Jestin, Boilers and Burners: Design and Theory
J.M. Berthelot, Composite Materials:
Mechanical Behavior and Structural Analysis
I.J. Busch-Vishniac, Electromechanical Sensors and Actuators
J. Chakrabarty, Applied Plasticity
K.K. Choi and N.H. Kim, Structural Sensitivity Analysis and Optimization 1: Linear Systems
K.K. Choi and N.H. Kim, Structural Sensitivity Analysis and Optimization 2: Nonlinear
Systems and Applications
G. Chryssolouris, Laser Machining: Theory and Practice
V.N. Constantinescu, Laminar Viscous Flow
G.A. Costello, Theory of Wire Rope, 2nd ed.
K. Czolczynski, Rotordynamics of Gas-Lubricated Journal Bearing Systems
M.S. Darlow, Balancing of High-Speed Machinery
W. R. DeVries, Analysis of Material Removal Processes
J.F. Doyle, Nonlinear Analysis of Thin-Walled Structures: Statics, Dynamics, and Stability
J.F. Doyle, Wave Propagation in Structures:
Spectral Analysis Using Fast Discrete Fourier Transforms, 2nd ed.
P.A. Engel, Structural Analysis of Printed Circuit Board Systems
A.C. Fischer-Cripps, Introduction to Contact Mechanics
A.C. Fischer-Cripps, Nanoindentation, 2nd ed.
J. Garcı́a de Jalón and E. Bayo, Kinematic and Dynamic Simulation of Multibody Systems:
The Real-Time Challenge
W.K. Gawronski, Advanced Structural Dynamics and Active Control of Structures
W.K. Gawronski, Dynamics and Control of Structures: A Modal Approach
G. Genta, Dynamics of Rotating Systems
D. Gross and T. Seelig, Fracture Mechanics with Introduction to Micro-mechanics
K.C. Gupta, Mechanics and Control of Robots
R. A. Howland, Intermediate Dynamics: A Linear Algebraic Approach
D. G. Hull, Optimal Control Theory for Applications
J. Ida and J.P.A. Bastos, Electromagnetics and Calculations of Fields
M. Kaviany, Principles of Convective Heat Transfer, 2nd ed.
M. Kaviany, Principles of Heat Transfer in Porous Media, 2nd ed.
(continued after index)

@Seismicisolation
@Seismicisolation
Wodek Gawronski

Modeling and Control

of Antennas and Telescopes

123
@Seismicisolation
@Seismicisolation
Wodek Gawronski
California Institute of Technology
Jet Propulsion Laboratory
Pasedena, CA, USA
[email protected]

ISSN: 0941-5122
ISBN: 978-0-387-78792-3 e-ISBN: 978-0-387-78793-0
DOI: 10.1007/978-0-387-78793-0

Library of Congress Control Number: 2008922734

c 2008 Springer Science+Business Media, LLC

All rights reserved. This work may not be translated or copied in whole or in part without the written
permission of the publisher (Springer Science+Business Media, LLC, 233 Spring Street, New York,
NY 10013, USA), except for brief excerpts in connection with reviews or scholarly analysis. Use in
connection with any form of information storage and retrieval, electronic adaptation, computer
software, or by similar or dissimilar methodology now known or hereafter developed is forbidden.
The use in this publication of trade names, trademarks, service marks, and similar terms, even if
they are not identified as such, is not to be taken as an expression of opinion as to whether or not
they are subject to proprietary rights.

Printed on acid-free paper

springer.com

@Seismicisolation
@Seismicisolation
Mechanical Engineering Series

Frederick F. Ling
Editor-in-Chief

The Mechanical Engineering Series features graduate texts and research monographs
to address the need for information in contemporary mechanical engineering, in-
cluding areas of concentration of applied mechanics, biomechanics, computational
mechanics, dynamical systems and control, energetics, mechanics of materials, pro-
cessing, production systems, thermal science, and tribology.

Advisory Board/Series Editors

Applied Mechanics F.A. Leckie
University of California,
Santa Barbara
D. Gross
Technical University of Darmstadt
Biomechanics V.C. Mow
Columbia University
Computational Mechanics H.T. Yang
University of California,
Santa Barbara
Dynamic Systems and Control/ D. Bryant
Mechatronics University of Texas at Austin
Energetics J.R. Welty
University of Oregon, Eugene
Mechanics of Materials I. Finnie
University of California, Berkeley
Processing K.K. Wang
Cornell University
Production Systems G.-A. Klutke
Texas A&M University
Thermal Science A.E. Bergles
Rensselaer Polytechnic Institute
Tribology W.O. Winer
Georgia Institute of Technology

@Seismicisolation
@Seismicisolation
Series Preface

Mechanical engineering, and engineering discipline born of the needs of the

industrial revolution, is once again asked to do its substantial share in the call
for industrial renewal. The general call is urgent as we face profound issues of
productivity and competitiveness that require engineering solutions, among others.
The Mechanical Engineering Series is a series featuring graduate texts and research
monographs intended to address the need for information in contemporary areas of
mechanical engineering.
The series is conceived as a comprehensive one that covers a broad range of con-
centrations important to mechanical engineering graduate education and research.
We are fortunate to have a distinguished roster of series editors, each an expert in
one of the areas of concentration. The names of the series editors are listed on page
vi of this volume. The areas of concentration are applied mechanics, biomechanics,
computational mechanics, dynamic systems and control, energetics, mechanics of
materials, processing, thermal science, and tribology.

@Seismicisolation
@Seismicisolation
Preface

This book is based on my experience with the control systems of antennas and
radiotelescopes. Overwhelmingly, it is based on experience with the NASA Deep
Space Network (DSN) antennas. It includes modeling the antennas, developing
control algorithms, field testing, system identification, performance evaluation, and
troubleshooting. My previous book1 emphasized the theoretical aspects of antenna
control engineering, while this one describes the application part of the antenna
control engineering.
Recently, the increased requirements for antenna/telescope pointing accuracy
have been imposed. In the case of DSN antennas, it was the shift from the S-band
(4 GHz) and X-band (8 GHz) communication to Ka band (32 GHz). On the other
hand, the Large Millimeter Telescope will operate up to 200 GHz, which requires
extremely accurate pointing. These requirements bring new challenges to the an-
tenna control engineers. Classical PI controllers cannot assure the required accuracy,
while model-based controllers (LQG, and H∞ ) increase the antenna accuracy by a
factor of 10. These controllers are new for antenna/telescope engineering. This book
describes their development and application.
The book is addressed primarily to the antenna, telescope, and radiotelescope
engineers, engineers involved in motion control, as well as students and researches
in motion control and mechatronics (antenna is a combination of mechanical, power,
and electronics subsystems). The book consists of two parts: Modeling (Chapters
2–5) and Control (Chapters 6–13).
In the modeling part, Chapter 2 describes the development of the analyti-
cal model of an antenna, including its structure (finite-element model) and drives
(motors, reducers, amplifiers). This modeling is useful in the design stage of an an-
tenna. Chapter 3 describes the determination of the antenna model using field tests
and the system identification approach. These models are quite accurate, and are
used in the development of the model-based controllers. Because the order of mod-
els (analytical as well as from the identification) is often too high for the analysis and

1Gawronski W. (2004). Advanced Structural Dynamics, and Active Control of Structures.

Springer, New York.

vii
@Seismicisolation
@Seismicisolation
viii Preface

for controller implementation, it is reduced to a reasonable size while preserving the

critical properties of the full model. Chapter 4 briefly describes the reduction tech-
niques as applied to antennas and telescopes. Finally, the wind disturbance models
are developed in Chapter 5. Wind is the major disturbance source for antennas and
telescopes. The steady (or static) wind model is presented, based on wind tunnel
data and confirmed by the field data. Also, three wind gusts models are presented.
The control part presents the performance criteria and shows how to transform
the antenna model to be the most suitable for controller tuning. In Chapter 7, the
book presents the development, properties, and limitations of the PI controller. It
shows the impact of the proportional and the integral gains on the antenna closed-
loop performance. It also analyzes the limits of the PI controller performance.
Chapter 8 describes the tuning process of the LQG controller. It analyzes the per-
formance of the LQG controller, including its limits. It presents the graphical user
interface (GUI) that allows us to tune the antenna LQG controller without analysis,
but by playing with the GUI sliders and buttons (“LQG for dummies”). It is shown
in this chapter that the performance of the LQG controller depends on its location:
either at the position loop or at the velocity loop, or at both. It shows also that in the
case of the 34-m DSN antennas the servo error with LQG controller decreased by a
factor of 6.5 when compared to the PI controller. In Chapter 9, H∞ controller tun-
ing is presented: gain determination, closed-loop equations, limits of performance.
In the following chapter, the non-linear control issues are addressed. The velocity
and acceleration limits often interfere with antenna dynamics. Two solutions are
proposed: a command preprocessor, and the anti-windup technique. Friction is the
source of a deteriorated pointing. Dither can be a solution, as presented in this book.
Finally, gearbox backlash is described and the counter-torque solution to minimize
it is presented.
A typical antenna control system consists of two loops: velocity and position
loops. A single-loop solution is studied in Chapter 11. The antenna control system
uses the encoder measurements to estimate RF beam position, but it is only a rough
estimate. The encoder measures the antenna angular position at its location, which
is different from the RF beam location; thus, antenna structural compliance is the
source of the error. Chapter 12 describes two techniques to control the RF beam
position: monopulse and conscan. Finally, in Chapter 13 we describe an open-loop
RF beam control: the look-up table that corrects for the uneven azimuth track that
impacts the RF beam position.
Analysis is a skill. However, even complex theories and the supporting analysis
are, in a sense, simple because they assume certain properties that simplify the anal-
ysis, or to make it possible. Still, although the theoretical path delivers an answer,
engineers have to ask if the answer is applicable to the real environment.
Engineering is an art. Some aspects of engineering can be described rigorously
by theoretical analysis, but not all. There are cases where engineering reality does
not satisfy analytical assumptions. Hence, engineering solutions are often ad hoc
solutions. This engineering dilemma is summarized by Scholnik: “Who cares how
it works, just as long as it gives the right answer.”

@Seismicisolation
@Seismicisolation
Preface ix

The art of engineering often includes features that theoretically cannot work. For
example:
r LQG controllers cannot be applied to antennas, because the antenna model have
poles at zero (rigid-body mode). But the LQG controllers do control the antennas.
r The noise in the tuning of the LQG controller should be the Gaussian noise,
which is not the case of antennas or telescopes.
r Rigidly applied model reduction algorithms do not produce the best reduced
model (best in terms of the antenna performance).
r Every LQG controller is an optimal controller, but not every one is acceptable.
r In the H∞ controller tuning the plant uncertainity should be either additive or
multiplicative. The antenna uncertainity, which depends on its elevation position,
is neither additive nor multiplicative.
Despite these difficulties, engineers succeed in developing acceptable control
systems, including antennas and telescopes.
Readers who would like to contact me with comments and questions are invited
to do so. My e-mail address is [email protected] or w.gawronski@sbc
global.net.

W. Gawronski
Pasedena, California

Acknowledgements

Part of the work described in this book was carried out at the Jet Propulsion Labora-
tory, California Institute of Technology, under contract with the National Aeronau-
tics and Space Administration. I thank Alaudin Bhanji, Mark Gatti, Pete Hames,
Wendy Hodgin, Andre Jongeling, Scott Morgan, Jean Patterson, Daniel Rascoe,
Christopher Yung, and Susan Zia, the managers at the Communications Ground
Systems Section, Jet Propulsion Laboratory, for their support of the Deep Space
Network antenna study.
I would like to acknowledge the contributions of my colleagues who have had
an influence on this work:
r from the Jet Propulsion Laboratory: Harlow Ahlstrom, Mimi Aung, Farrokh
Baher, Abner Bernardo, John Cucchissi, Jeffery Mellstrom, Martha Strain;
r from the NASA Goldstone Deep Space Communication Complex (California):
Michael Winders and Jeffrey Frazier;
r from the NASA Madrid Deep Space Communication Complex (Spain): Angel
Martin, David Munoz Mochon, and Pablo Perez-Zapardiel;
r from the NASA Canberra Deep Space Communication Complex (Australia):
Paul Richter and John Howell;
r visiting students: Jason Brand, Emily Craparo, Brandon Kuczenski, and Erin
Maneri whose work is also included in this book.

@Seismicisolation
@Seismicisolation
x Preface

I also acknowledge the help of Kamal Souccar in the study of the Large Milli-
meter Telescope; Bogusz Bienkiewicz of Colorado State University in the wind dis-
turbance study; and Toomas Erm of the European Southern Observatory, Munich,
Germany, for many interesting discussions on control systems of antennas and
telescopes.

@Seismicisolation
@Seismicisolation
Contents

1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.1 Examples of Antennas and Telescopes . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.1.1 NASA Deep Space Network . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.1.2 Large Millimeter Telescope . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.1.3 ESA Deep Space Antennas . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2
1.1.4 Atacama Large Millimeter Array . . . . . . . . . . . . . . . . . . . . . . . 2
1.1.5 Thirty Meter Telescope . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
1.1.6 Green Bank Telescope . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
1.1.7 Effelsberg Telescope . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
1.2 Short Description of the Antenna Control System . . . . . . . . . . . . . . . . 5
1.2.1 Velocity Loop . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
1.2.2 Position-Loop . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
1.3 Antenna and Telescope Literature . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8

Part I Modeling

2 Analytical Models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
2.1 Rigid Antenna Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
2.2 Structural Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12
2.2.1 Finite-Element Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12
2.2.2 Modal Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14
2.2.3 State-Space Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17
2.2.4 Models with Rigid Body Modes . . . . . . . . . . . . . . . . . . . . . . . . 19
2.2.5 Discrete-Time Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21
2.3 Drive Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24
2.3.1 Motor Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24
2.3.2 Reducer Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25
2.3.3 Drive Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25
2.4 Velocity Loop Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26
2.5 Drive Parameter Study . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26
2.5.1 Drive Stiffness Factor . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27

xi
@Seismicisolation
@Seismicisolation
xii Contents

2.5.2 Drive Inertia Factor . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28

References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30

3 Models from Identification . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31

3.1 White Noise Testing of the Antenna . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31
3.1.1 Purpose and Conditions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31
3.1.2 Test Input and Output . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 32
3.1.3 Test Configuration . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 32
3.1.4 Data Processing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 34
3.1.5 Basic Relationships for the Discrete-Time Data . . . . . . . . . . . 37
3.2 Identification of the Velocity Loop Model . . . . . . . . . . . . . . . . . . . . . . . 38
3.2.1 Description of the Velocity Loop Model . . . . . . . . . . . . . . . . . 39
3.2.2 Identification of the Velocity Loop Model . . . . . . . . . . . . . . . . 39
3.2.3 A Comparison of the Analytical and Identified Models . . . . . 41
3.2.4 Azimuth Model Depends on the Antenna
Elevation Position . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 41
3.2.5 Fundamental Frequency Depends on Antenna Diameter . . . . 43
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 44

4 Model Reduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 45
4.1 Why Reduction? . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 45
4.2 Balanced Model Reduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 45
4.3 Modal Model Reduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 47
4.3.1 Norms of a Single Mode . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 47
4.3.2 Norms of a Structure . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 48
4.4 Antenna Model Reduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 49
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 50

5 Wind Disturbance Models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 51

5.1 Steady-State Wind Disturbance Model . . . . . . . . . . . . . . . . . . . . . . . . . . 51
5.1.1 Dimensionless Wind Torques . . . . . . . . . . . . . . . . . . . . . . . . . . 52
5.1.2 Obtaining Wind Torques from Field Data . . . . . . . . . . . . . . . . 54
5.1.3 Comparing Wind Tunnel Results and the Field Data . . . . . . . 56
5.2 Wind Gusts Disturbance Models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 59
5.2.1 Model of Wind Forces Acting on the Dish . . . . . . . . . . . . . . . 60
5.2.2 Model of Wind Torque Acting at the Drives . . . . . . . . . . . . . . 64
5.2.3 Algorithm to Generate a Time Profile of Wind
Gusts Torque . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 65
5.2.4 Model of Wind at the Velocity Input . . . . . . . . . . . . . . . . . . . . 66
5.2.5 Algorithm to Generate Time Profile of Wind
at the Velocity Input . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 67
5.2.6 The Equivalence of Wind Torque and Wind
Velocity Models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 67
5.2.7 Closed Loop Pointing Accuracy with Wind Gusts
Disturbances . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 68
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 70

@Seismicisolation
@Seismicisolation
Contents xiii

Part II Control

6 Preliminaries to Control . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 73
6.1 Performance Criteria . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 73
6.2 Transformations of the Velocity Loop Model . . . . . . . . . . . . . . . . . . . . 77
6.2.1 Transformation into Modal Coordinates . . . . . . . . . . . . . . . . . 78
6.2.2 Antenna Position as the First State . . . . . . . . . . . . . . . . . . . . . . 78
6.2.3 Augmentation with the Integral of the Position . . . . . . . . . . . . 78
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 79

7 PI and Feedforward Controllers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 81

7.1 Properties of the PI Controller . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 81
7.1.1 Closed Loop Transfer Functions . . . . . . . . . . . . . . . . . . . . . . . . 83
7.1.2 The Proportional Gain Analysis . . . . . . . . . . . . . . . . . . . . . . . . 83
7.1.3 The Integral Gain Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . 85
7.2 PI Controller Tuning Steps . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 86
7.3 Closed Loop Equations of a Flexible Antenna
with a PI Controller . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 87
7.4 Performance of the PI Controller . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 87
7.4.1 Performance Characteristics . . . . . . . . . . . . . . . . . . . . . . . . . . . 87
7.4.2 Limits of Performance . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 90
7.5 Feedforward Controller . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 91
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 93

8 LQG Controller . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 95
8.1 Properties of the LQG Controller . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 95
8.1.1 LQG Controller Description . . . . . . . . . . . . . . . . . . . . . . . . . . . 95
8.1.2 Tracking LQG Controller . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 98
8.1.3 Closed Loop Equations of the Tracking LQG
Control System . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 101
8.1.4 LQG Weights . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 101
8.1.5 Resemblance of the LQG and PI Controllers . . . . . . . . . . . . . . 104
8.1.6 Properties of the LQG Weights . . . . . . . . . . . . . . . . . . . . . . . . . 105
8.1.7 Limits of the LQG Gains . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 107
8.2 LQG Controller Tuning Steps . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 108
8.3 Performance of the LQG Controller . . . . . . . . . . . . . . . . . . . . . . . . . . . . 110
8.3.1 Summary of the Antenna Servo Performance
Characteristics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 110
8.3.2 Performance of the DSN Antennas
with LQG Controllers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 111
8.3.3 Disturbance Rejection Properties and the Position-Loop
Bandwidth . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 113
8.3.4 Performance Comparison of the PI and LQG Controllers . . . 115
8.3.5 Limits of Performance . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 117

@Seismicisolation
@Seismicisolation
xiv Contents

8.4 Tuning a LQG Controller Using GUI . . . . . . . . . . . . . . . . . . . . . . . . . . . 117

8.4.1 Selecting LQG Weights . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 117
8.4.2 GUI for the LQG Controller Tuning . . . . . . . . . . . . . . . . . . . . . 119
8.4.3 Fine Tuning of the LQG Controller . . . . . . . . . . . . . . . . . . . . . 121
8.5 LQG Controller in the Velocity Loop . . . . . . . . . . . . . . . . . . . . . . . . . . . 124
8.5.1 Position Loop Bandwidth Depends on the Velocity Loop . . . 124
8.5.2 Four Control Configurations . . . . . . . . . . . . . . . . . . . . . . . . . . . 127
8.5.3 PP Control System . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 127
8.5.4 PL Control System . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 129
8.5.5 LP Control System . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 132
8.5.6 LL Control System . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 132
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 133

9 H∞ Controller . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 135
9.1 Definition and Gains . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 135
9.2 Tracking H∞ Controller . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 138
9.3 Closed-Loop Equations of the Tracking H∞ Controller . . . . . . . . . . . . 138
9.4 34-M Antenna Example . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 139
9.5 Limits of Performance . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 141
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 142

10 Single Loop Control . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 145

10.1 Rigid Antenna . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 145
10.1.1 Rigid Antenna with Velocity and Position Loops . . . . . . . . . . 145
10.1.2 Rigid Antenna with Position Loop Only . . . . . . . . . . . . . . . . . 147
10.1.3 Simulation Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 148
10.2 34-M Antenna . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 149
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 155

11 Non-Linear Control . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 157

11.1 Velocity and Acceleration Limits . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 157
11.1.1 Command Preprocessor . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 157
11.1.2 Anti-Windup Technique . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 162
11.2 Friction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 166
11.2.1 Dry Friction Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 168
11.2.2 Low-Velocity Tracking Using Dither . . . . . . . . . . . . . . . . . . . . 170
11.2.3 Non-linear Simulation Results . . . . . . . . . . . . . . . . . . . . . . . . . 173
11.3 Backlash . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 174
11.3.1 Backlash and Its Prevention . . . . . . . . . . . . . . . . . . . . . . . . . . . . 175
11.3.2 The Velocity Loop Model with Friction and Backlash . . . . . . 177
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 181

12 RF Beam Control . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 183

12.1 Selecting the RF Beam Controller . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 183
12.2 Monopulse . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 187

@Seismicisolation
@Seismicisolation
Contents xv

12.2.1 Command following . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 187

12.2.2 Disturbance Rejection Properties . . . . . . . . . . . . . . . . . . . . . . . 188
12.2.3 Stability Due to the Gain Variation . . . . . . . . . . . . . . . . . . . . . . 190
12.2.4 Performance Simulations: Linear Model . . . . . . . . . . . . . . . . . 190
12.2.5 Performance Simulations: Nonlinear Model . . . . . . . . . . . . . . 191
12.3 Scanning . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 194
12.3.1 Conical Scan . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 194
12.3.2 Sliding Window Conscan . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 200
12.3.3 Lissajous Scan . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 201
12.3.4 Rosette Scan . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 204
12.3.5 Performance Evaluation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 206
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 209

13 Track-Level Compensation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 211

13.1 Description of the Track-Level Problem . . . . . . . . . . . . . . . . . . . . . . . . 211
13.2 Collection and Processing of the Inclinometer Data . . . . . . . . . . . . . . . 213
13.3 Estimating Azimuth Axis Tilt . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 215
13.4 Creating the TLC Table . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 217
13.5 Determining Pointing Errors from the TLC Table . . . . . . . . . . . . . . . . . 219
13.6 Antenna Pointing Improvement Using the TLC Table . . . . . . . . . . . . . 221
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 222

Index . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 223

@Seismicisolation
@Seismicisolation
Chapter 1
Introduction

This chapter presents examples of antennas and telescopes, shortly describes the
antenna control system, and presents references on the antenna mechanical and
control engineering.

1.1 Examples of Antennas and Telescopes

1.1.1 NASA Deep Space Network

The NASA Deep Space Network (DSN) antennas communicate with spacecraft by
sending commands (uplink) and by receiving information from spacecraft (down-
link). To assure continuous tracking during Earth’s rotation, the antennas are located
at three sites: Goldstone (California), Madrid (Spain), and Canberra (Australia). The
signal frequencies are 8.5 GHz (X-band) and 32 GHz (Ka-band). The dish size of
the antennas is either 34 m or 70 m. An example of the 70-m antenna is shown in
Fig. 1.1. The antenna dish rotates with respect to the horizontal (or elevation) axis.
The whole antenna structure rotates on a circular track (azimuth track) with respect
to the vertical (or azimuth) axis. For the Ka-band frequency the required tracking
accuracy is on the order of 1 mdeg. This requirement is a driver for the control
system upgrade of the antennas. In [2], [3], and [4] you can find the description
of the DSN antenna control systems, and at the webpage http://ipnpr.jpl.nasa.gov/
index.cfm the DSN antennas research reports, including control systems. The Deep
Space Network webpage is at the webpage deepspace.jpl.nasa.gov/dsn/.

1.1.2 Large Millimeter Telescope

The Large Millimeter Telescope (LMT) is the joint effort of the University of
Massachusetts at Amherst and the Instituto Nacional de Astrofı́sica, Óptica, y
Electrónica (INAOE) in Mexico (Fig. 1.2). The LMT is a 50-m diameter telescope,
designed for operation at wavelengths between 1 mm and 4 mm. The telescope
rotates with respect to elevation and azimuth axes. It is built atop of Sierra Negra

W. Gawronski, Modeling and Control of Antennas and Telescopes, 1

@Seismicisolation
@Seismicisolation
DOI: 10.1007/978-0-387-78793-0 1, 
C Springer Science+Business Media, LLC 2008
2 1 Introduction

Fig. 1.1 NASA/JPL 70-m

antenna at Goldstone, CA
(courtesy of
NASA/JPL/Caltech)

(4640 m), a volcanic peak in the state of Puebla, Mexico. The LMT is a significant
step forward in antenna design: in order to reach its pointing accuracy specifica-
tions, it must outperform every other telescope in its frequency range. The antenna
designers expect that the telescope will point to its specified accuracy of 0.3 mdeg
under conditions of low winds and stable temperatures, with radiofrequency up to
110 GHz. More about LMT, see www.lmtgtm.org/.

1.1.3 ESA Deep Space Antennas

For use in deep space, high elliptical orbit missions, and future missions to Mars,
the European Space Agency (ESA) erected 35-m deep space ground stations; see
http://www.esa.int/SPECIALS/ESOC/SEMZEEW4QWD 0.html. The antennas are
designed for frequencies up to 35 GHz and a pointing accuracy of 6 mdeg. The first
antenna (that moves in azimuth and elevation axes) has been installed in Australia
and has proven its compliance to the specifications. The second antenna is under
construction in Spain. The 35 m antenna incorporates a full motion pedestal with a
beam waveguide system.

1.1.4 Atacama Large Millimeter Array

The Atacama Large Millimeter Array (ALMA) is an international astronomy facility.

A brief description can be found at www.alma.nrao.edu/info. The pointing accuracy
of the 12 m ALMA telescope is 0.16 mdeg. The control system must handle very

@Seismicisolation
@Seismicisolation
1.1 Examples of Antennas and Telescopes 3

Fig. 1.2 Large Millimeter

Telescope (courtesy of the
Large Millimeter Telescope
Project)

accurate movement at sidereal tracking velocities as well as several extremely fast

switching functions. To do this, the drives are designed to accelerate up to 24 deg/s2 ,
which is very unusual for a telescope of this size.

1.1.5 Thirty Meter Telescope

The Thirty Meter Telescope (TMT) will be the first of the giant optical/infrared
ground-based telescopes (of 30-m diameter of primary mirror) addressing one of
the most compelling areas in astrophysics: the nature of dark matter, the assembly
of galaxies, the growth of structure in universe, and the physical processes involved
in star and planet formation. TMT will operate over 0.3–30 ␮m wavelength range,
providing nine times the collecting area of the current largest optical telescope, the
10-m Keck telescope. It will use adaptive optics system to allow diffraction-limited
performance, resulting in spatial resolution 12.5 times sharper than is achieved
by the Hubble Space Telescope. For more about TMT see http://www.tmt.org/ or
http://www.astro.caltech.edu/observatories/tmt/.

@Seismicisolation
@Seismicisolation
4 1 Introduction

1.1.6 Green Bank Telescope

The National Radio Astronomy Observatory operates the Green Bank Telescope
(GBT), the world’s largest (along with the Effelsberg telescope) single aperture
telescope (Fig. 1.3). It is located in Green Bank, West Virginia. The GBT has an
unusual design. Unlike conventional telescopes, the GBT’s aperture is unblocked
(using an off-axis feed arm) so that incoming radiation meets the surface directly.
This increases the useful area of the telescope and eliminates reflection and diffrac-
tion that ordinarily complicate a telescope’s pattern of response. More information
can be found at http://www.gb.nrao.edu/.

1.1.7 Effelsberg Telescope

The Effelsberg 100-m telescope is one of the world’s largest fully steerable tele-
scopes (Fig. 1.4). It is operated by the Max Planck Institute for Radio Astronomy
in Bonn, Germany, at wavelengths from about 3.5 mm to 90 cm. The telescope is
used to observe pulsars, cold gas and dust clusters, the sites of star formation, jets
of matter emitted by black holes, and the nuclei of distant far-off galaxies. Its steel
construction weighs 3,200 tons. Its azimuth velocity is 0.5 deg/s, and 0.25 deg/s in

Fig. 1.3 The Green Bank Telescope (courtesy of National Radio Astronomy Observatory)

@Seismicisolation
@Seismicisolation
1.2 Short Description of the Antenna Control System 5

Fig. 1.4 The Effelsberg

telescope (courtesy of Max
Planck Institute for Radio
Astronomy, Bonn, Germany)

elevation. More information can be found at http://www.mpifr-bonn.mpg.de/public/

index e.html.

1.2 Short Description of the Antenna Control System

A Deep Space Network antenna with a 34-m dish is shown in Fig. 1.5. This antenna
can rotate around azimuth (vertical) and elevation (horizontal) axes. The rotation
is controlled by azimuth and elevation controllers. The combination of the antenna
structure and its azimuth and elevation drives makes the velocity loop of the antenna.
The velocity loop plant has two inputs (azimuth and elevation velocities) and two
outputs (azimuth and elevation position), and the position loop is closed between
the encoder outputs and the velocity inputs. The drives consist of gearboxes, electric
motors, amplifiers, and tachometers.
The antenna controller consists of two independent subsystems: azimuth and
elevation controllers. Because both subsystems are independent, a single system is
considered further, and azimuth or elevation system is specified, if necessary. Also,

@Seismicisolation
@Seismicisolation
6 1 Introduction

Fig. 1.5 The Deep Space

Network antenna at
Goldstone, California
(courtesy of
NASA/JPL/Caltech). It can
rotate with respect to the
azimuth (vertical axis) and
with respect to elevation
(horizontal axis)

the antenna control system consists of the open loop (or velocity loop) system and
the closed loop (or position loop) system.

1.2.1 Velocity Loop

The velocity loop includes the structure and the drives (motors, gearboxes and
amplifiers). It is driven by the velocity input signal uc (deg/s), see Fig. 1.6, while the
encoder reading y (deg) is the output. The elevation and azimuth loops are similar.
The velocity-loop allows for the manual control of antenna.

encoder, y
uc + velocity V
controller
–
ANTENNA

tach velocity

Fig. 1.6 The velocity loop model of the Deep Space Network antenna

@Seismicisolation
@Seismicisolation
1.3 Antenna and Telescope Literature 7

VELOCITY LOOP
command, r encoder, y
V
position uc velocity
controller controller

tach velocity

encoder

Fig. 1.7 The velocity- and position-loops of the Deep Space Network antenna

1.2.2 Position-Loop
The block diagram of the closed loop (or position loop) system is shown in Fig. 1.7.
It consists of the antenna velocity loop and the position controller. The controller
is run by computer software that drives the antenna depending on the actual
antenna position and the commanded position. The controller has two inputs: the
encoder position y, and the commanded position (or shorter: command) r (deg). The
controller output is the velocity uc (deg/s) that drives the antenna.

1.3 Antenna and Telescope Literature

This book is based predominantly on the author’s experience with the NASA Deep
Space Network, although the experience of others is included in the references.
Obviously, there are too many references to be included here, but readers inter-
ested in further study of antenna and telescope control and pointing issues are
referred to:
• the IEEE Antennas and Propagation Magazine,
• the Proceedings of SPIE (the International Society for Optical Engineering),
Optical Engineering (SPIE Journal),
• the Interplanetary Network Progress Report (JPL), http://ipnpr.jpl.nasa.gov/
index.cfm

There are three books that address antenna pointing and control issues, see Refs.
[1], [7], and [8], as well as this author’s book [3], which addresses general issues of
structural dynamics and control, but includes many examples of antenna dynamics
and antenna control systems. Karcher [6] gives an interesting overview of the tele-
scope structure, mechanism, and control system design. The handbook [10] presents
the control design of a simple (rigid) antenna. General antenna theory is presented
in a comprehensible form in [9] and [11].

@Seismicisolation
@Seismicisolation
8 1 Introduction

References
1. Biernson G. (1990). Optimal Radar Tracking Systems. Wiley, New York.
2. Gawronski W. (2001). Antenna control systems: from PI to H∞ . IEEE Antennas and Propa-
gation Magazine, 43 (1).
3. Gawronski W. (2004). Advanced Structural Dynamics and Active Control of Structures.
Springer, New York.
4. Gawronski W. (2007). Control and pointing challenges of large antennas and telescopes. IEEE
Trans. Control Systems Technology, 15, (2).
5. Gawronski W, Mellstrom JA. (1994). Control and dynamics of the Deep Space Network
antennas. In: Control and Dynamic Systems, vol. 63, C.T. Leondes, ed. Academic Press, San
Diego.
6. Karcher HJ. (2006). Telescopes as mechatronic systems. IEEE Antennas and Propagation
Magazine 48 (2):17–37.
7. Kitsuregawa T. (1990). Advanced Technology in Satellite Communication Antennas: Elec-
trical and Mechanical Design. Artech House, Boston.
8. Levy R (1996) Structural Engineering of Microwave Antennas. IEEE Press, New York.
9. Macnamara T. (1995). Handbook of Antennas for EMC, Artech House, Boston.
10. Nise NS (1995) Control System Engineering. The Benjamin/Cummings Publishing Company,
Redwood City, CA.
11. Toomay JC. (1989). Radar Principles for the Non-Specialist. Van Nostrand Reinhold,
New York.

@Seismicisolation
@Seismicisolation
 Part I
Modeling

@Seismicisolation
@Seismicisolation
Chapter 2
Analytical Models

This chapter presents the development of the analytical model of the antenna
velocity loop. The analytical model includes the antenna structure, its drives, and
the velocity loop itself. First, a rigid antenna model is discussed. Next, the modal
model of a flexible antenna structure is analyzed based on the finite-element data.
The modal model is transferred into the state-space model, in both continuous- and
discrete-time. Next, the drive model is derived, and combined into the velocity loop
model. Finally, the impact of the drive parameters (gearbox stiffness and motor
inertia) on the velocity loop properties is analyzed. It is worth to remind that the
analytical model is mainly used in the design stage of the antenna. Due to limited
accuracy of the analytical modeling, these models cannot be used in implementa-
tion, such as the model-based controllers.

2.1 Rigid Antenna Model

In this section the velocity loop of a rigid-body antenna is analyzed. Such models
represents an antenna without flexible deformations in the disturbance frequency
band, and are well beyond antenna bandwidth. This is a model of an idealized
antenna, with rigid gearboxes and a rigid structure. It might be applicable to small
antennas, but in our case it serves as a tool for explaining and deriving basic velocity
loop properties in a closed form (while for larger, flexible antennas the analysis is
based on Matlab and Simulink simulations). This simple analysis is extended later
to illustrate properties of a flexible antenna control system, and for the better insight
into more complicated models and their properties.
A rigid antenna in the open loop configuration has torque input and velocity
output. The relationship between the torque τ and angular velocity ω follows from
the Newton inertia law

J ω̇ = τ (2.1)

where J is the antenna inertia. The Laplace transform of the above equation is
J sω(s) = τ (s); hence, the antenna transfer function is represented as an integrator

W. Gawronski, Modeling and Control of Antennas and Telescopes, 11

@Seismicisolation
@Seismicisolation
DOI: 10.1007/978-0-387-78793-0 2, 
C Springer Science+Business Media, LLC 2008
12 2 Analytical Models

d
rigid antenna

u τ + 1 ϕ 1 ϕ
ko s
Js

Fig. 2.1 Velocity loop model of a rigid antenna

ω(s) 1
G(s) = = (2.2)
τ (s) Js
The velocity loop model of a rigid antenna with a proportional controller is
shown in Fig. 2.1. The transfer function of the closed velocity loop system, from
the velocity command u to the antenna velocity ϕ̇, is as follows:

ϕ̇(s) ko G(s) 1
G rl (s) = = = (2.3)
u(s) 1 + ko G(s) 1 + Ts

where
 T =  J ko . The bandwidth of the velocity loop system is equal to B =
1 T = ko J rad/s. The gain, ko , is tuned to obtain the required bandwidth of the
system. For example, if the required bandwidth is B = 20 rad/s, and the inertia is J =
1 Nms2 /rad, then ko = 20 Nms.

2.2 Structural Model

The structural model is typically derived from the finite-element model. It is not
the finite-element model per se, but its by-product, the modal model. Note that
the modal model includes the rigid-body mode (antenna-free rotation). For the
controller tuning purposes the modal model is represented in state-space form,
which it is finally given in the discrete-time representation.

2.2.1 Finite-Element Model

The analytical model of an antenna is obtained from its finite-element model.

The finite element model of a 34-m antenna is shown in Fig. 2.2. It consists of
hundreds of nodes and elements, and thousands of degrees of freedom. The model
is characterized by the mass, stiffness, and damping matrices, and by the sensors
and actuator locations.

@Seismicisolation
@Seismicisolation
2.2 Structural Model 13

1 2 3 4 5 6
1 2 3 4 5 6
1 2 3 4 5 6

1 .

Fig. 2.2 The finite-element model of the 34-m antenna

Let N be a number of degrees of freedom of the finite-element model, with

antenna displacement and velocity at the encoder location as the output and force
at the motor pinion attachment as the input. Antenna structure is represented by the
following second-order matrix differential equation:

M q̈ + D q̇ + K q = Bo u,
(2.4)
y = Coq q + Cov q̇.

In this equation q is the N × 1 nodal displacement vector; q̇ is the N × 1

nodal velocity vector; q̈ is the N × 1 nodal acceleration vector; u is the input; y
is the output; M is the mass matrix, N × N ; D is the damping matrix, N × N ;
and K is the stiffness matrix, N × N . The input matrix Bo is N × 1, the output
displacement matrix Coq is 1 × N , and the output velocity matrix Cov is 1 × N .
The mass matrix is positive definite (all its eigenvalues are positive), and the

@Seismicisolation
@Seismicisolation
14 2 Analytical Models

stiffness and damping matrices are positive semi-definite (all their eigenvalues are
non-negative).

2.2.2 Modal Model

Modal models of antennas are expressed in modal coordinates. Because these coor-
dinates are independent, they simplify the analysis. The modal model is obtained by
transforming equation (2.4) using a modal matrix; see [2].
Consider free vibrations of an antenna structure without damping, and without
external excitation, that is, u = 0 and D = 0. The equation of motion (2.4) in this case
turns into the following equation:

M q̈ + K q = 0. (2.5)

The solution of the above is q = φe jωt ; hence, the second derivative of the
solution is q̈ = −ω2 φe jωt . Introducing q and q̈ into (2.5) gives

(K − ω2 M)φe jωt = 0. (2.6)

This set of homogeneous equations has a nontrivial solution if the determinant of

K − ω2 M is zero, that is,

det(K − ω2 M) = 0. (2.7)

The above determinant equation is satisfied for a set of n values of frequency ω.

These frequencies are denoted ω1 , ω2 , ..., ωn , and their number n does not exceed
the number of degrees of freedom, i.e., n  N . The frequency ωi is called the ith
natural frequency.
Substituting ωi into (2.6) yields the corresponding set of vectors {φ1 , φ2 , ..., φn }
that satisfy this equation. The ith vector φi corresponding to the ith natural frequency
is called the ith mode shape. The natural modes are not unique, since they can be
arbitrarily scaled. Indeed, if φi satisfies (2.6), so does αφi , where α is an arbitrary
scalar.
For a notational convenience define the matrix of natural frequencies

⎡ ⎤
ω1 0 ··· 0
⎢ 0 ω2 ··· 0 ⎥
⍀=⎢
⎣ ··· ···
⎥ (2.8)
··· ··· ⎦
0 0 · · · ωn

@Seismicisolation
@Seismicisolation
2.2 Structural Model 15

and the matrix of mode shapes, or modal matrix ⌽, of dimensions N × n, which

consists of n natural modes of a structure
⎡ ⎤
φ11 φ21 ... φn1
⎢ φ12 φ22 ... φn2 ⎥
⎢ ⎥

⎢ ... ... ... ... ⎥
⌽= φ1 φ2 ... φn =⎢
⎢
⎥,
⎥ (2.9)
⎢ φ1k φ2k ... φnk ⎥
⎣ ... ... ... ... ⎦
φ1n d φ2n d ... φnn d

where φi j is the jth displacement of the ith mode, that is,

⎧ ⎫
⎪ φi1 ⎪
⎪
⎪ ⎪
⎨ φi2 ⎪
⎬
φi = .. . (2.10)
⎪
⎪ . ⎪
⎪
⎪
⎩ ⎪
⎭
φin

The modal matrix ⌽ has an interesting property: it diagonalizes mass and stiff-
ness matrices M and K,

Mm = ⌽T M⌽, K m = ⌽T K ⌽. (2.11)

The obtained diagonal matrices are called modal mass matrix (Mm ) and modal
stiffness matrix (K m ), respectively. The same transformation, applied to the damping
matrix

Dm = ⌽T D⌽, (2.12)

gives the modal damping matrix Dm , which is a diagonal if it is, for example, propor-
tional to the stiffness matrix.
In order to obtain the modal model a new variable, qm , is introduced. It is called
modal displacement, and satisfies the following equation:

q = ⌽qm . (2.13)

Equation (2.13) is introduced to (2.4), and the latter additionally left-multiplied

by ⌽T , obtaining

Mm q̈m + Dm q̇m + K m qm = ⌽T Bo u,
y = Coq ⌽qm + Cov ⌽q̇m .

@Seismicisolation
@Seismicisolation
16 2 Analytical Models

using (2.11), and (2.12) notations. Next, the left multiplication of the latter equation
by Mm−1 , which gives

q̈m + 2Z⍀q̇m + ⍀2 qm = Bm u,
(2.14)
y = Cmq qm + Cmv q̇m .
In (2.14) ⍀ is a diagonal matrix of natural frequencies, defined before, and Z is
the modal damping matrix. It is a diagonal matrix of modal damping,
⎡ ⎤
ζ1 0 ··· 0
⎢ 0 ζ2 ··· 0 ⎥
Z=⎢
⎣ ··· ···
⎥. (2.15)
··· ··· ⎦
0 0 · · · ζn

where ζi is the damping of the ith mode. This matrix is obtained using the following
relationship Mm−1 Dm = 2ZΩ, thus,

−1 −1
Z = 0.5Mm−1 Dm Ω −1 = 0.5Mm 2 K m 2 Dm . (2.16)

The modal input matrix Bm in (2.14) is as follows

Bm = Mm−1 ⌽T Bo . (2.17)

Finally, in (2.14) the following notations is used for the modal displacement and
velocity matrices:

Cmq = Coq ⌽, (2.18)

Cmv = Cov ⌽. (2.19)

Note that (2.14) (a modal representation of a structure) is a set of uncoupled

equations. Indeed, due to the diagonality of ⍀ and Z, this set of equations can be
written, equivalently, as

q̈mi + 2ζi ωi q̇mi + ωi2 qmi = bmi u

yi = cmqi qmi + cmvi q̇mi , i = 1, . . . , n,
n (2.20)
y= yi ,
i=1

where bmi is the ith row of Bm and cmqi , cmvi are the ith columns of Cmq and Cmv ,
respectively. In the above equations yi is the system output due to the ith mode
dynamics. Note that the structural response y is a sum of modal responses yi , which
is a key property used to derive structural properties in modal coordinates.
Theoretically the determination of the input and output matrices as in (2.17),
(2.18), and (2.19) requires large matrix ⌽ (of size: number of degrees of freedom

@Seismicisolation
@Seismicisolation
2.2 Structural Model 17

by number of modes). In fact, one needs only one row of it for each input/output
matrix. Indeed, the input matrix Bo is all zero, except the location of the actuator
(pinion) to the structure. The same, the output matrix Cmq is all zero, except the
location of the sensor (encoder) to the structure. Let the location of the sensor is at
the kth degree of freedom of the finite-element model. The modal matrix is as in
(2.9). Because Cmq is all zero but the kth entry, therefore

⎡ ⎤
φ11 φ21 ... φn1
⎢ φ12 φ22 ... φn2 ⎥
⎢ ⎥

⎢ ... ... ... ... ⎥

Cmq = Coq ⌽ = 0 0 ... 1 ... 0 ⎢

⎢
⎥ = φ1k φ2k ... φnk
⎥
⎢ φ1k φ2k ... φnk ⎥
⎣ ... ... ... ... ⎦
φ1n d φ2n d ... φnn d
(2.21)
that is, Cmq consists of modal displacement at the encoder location. Similarly, if the
actuator (pinion) is located at the kth degree of freedom, the corresponding input
matrix is

⎡ ⎤
φ1k /m m1
⎢ φ2k /m m2 ⎥
Bmq =⎢
⎣ ··· ⎦
⎥ (2.22)
φnk /m mn

where m mi is the ith modal mass. Thus, Bmq consists of modal displacement at the
actuator (pinion) location scaled by the modal masses.
In summary, from comparatively few parameters of the finite-element model,
such as natural frequencies, modal masses, modal damping, and modal displace-
ments at the actuator and sensor locations one creates the antenna structural model
in equation (2.14). Note that this structural dynamics model of an antenna is much
simpler than its static finite-element model.
As an example, determine the first four natural modes and frequencies of the
antenna presented in Fig. 2.2. The modes are shown in Fig. 2.3. For the first mode
the natural frequency is ω1 = 13.2rad/s (2.10 Hz), for the second mode the natural
frequency is ω2 = 18.1 rad/s (2.88 Hz), for the third mode the natural frequency is
ω3 = 18.8 rad/s (2.99 Hz), and for the fourth mode the natural frequency is ω4 =
24.3 rad/s (3.87 Hz).

2.2.3 State-Space Model

State-space representation is the standard representation of the control system

models. Thus, the antenna structure needs to be represented in this form, to allow the
use of control system software such as Matlab or Simulink. The modal state-space

@Seismicisolation
@Seismicisolation
18 2 Analytical Models

(a) (b)

(c) (d)

Fig. 2.3 Antenna modes: (a) First mode (of natural frequency 2.10 Hz); (b) second mode (of
natural frequency 2.87 Hz); (c) third mode (of natural frequency 2.99 Hz); and (d) fourth mode
(of natural frequency 3.87 Hz). For each mode the nodal displacements are sinusoidal, have the
same frequency, and the displacements are shown at their extreme values. Gray color denotes
undeformed state

representation of the antenna structure is a triple ( Am , Bm , Cm ) characterized by the

block-diagonal state matrix, Am ; see [2]
⎡ ⎤
× ×  0 0  · · · · · ·  0 0
⎢ × ×  0 0  · · · · · ·  0 0 ⎥
⎢----------------------------- ⎥
⎢ ⎥
⎢ 0 0  × ×  · · · · · ·  0 0 ⎥
⎢ ⎥
⎢ 0 0  × ×  · · · · · ·  0 0 ⎥
⎢ ⎥
Am = diag(Ami ) = ⎢ - - - - - - - - - - - - - - - - - - - - - - - - - - - - - ⎥, i = 1, 2, . . . , n,
⎢··· ··· ··· ··· ··· ··· ··· ··· ⎥
⎢    ⎥
⎢··· ···  ··· ···  ··· ···  ··· ··· ⎥
⎢    ⎥
⎢----------------------------- ⎥
⎣ 0 0  0 0  ··· ···  × × ⎦
  
0 0  0 0  · · · · · ·  × ×
(2.23)

where Ami are 2 × 2 blocks (their nonzero elements are marked with ×), and the
modal input and output matrices are divided, correspondingly,

@Seismicisolation
@Seismicisolation
2.2 Structural Model 19
⎡ ⎤
Bm1
⎢ Bm2 ⎥

⎢ ⎥
Bm = ⎢ . ⎥ , Cm = Cm1 Cm2 · · · Cmn , (2.24)
⎣ .. ⎦
Bmn

The blocks Ami , Bmi , and Cmi are as follows:

     
0 ωi 0 cmqi
Ami = , Bmi = , Cmi = cmvi ; (2.25)
−ωi −2ζi ωi bmi ωi

The state x of the modal representation consists of n independent components,

xi , that represent a state of each mode
⎧ ⎫
⎪
⎪ x1 ⎪
⎪
⎪
⎨ x2 ⎪
⎬
x= .. , (2.26)
⎪
⎪ ⎪
⎪ . ⎪
⎩ ⎪
⎭
xn
The ith state component is as follows:
 
ωi qmi
xi = , (2.27)
q̇mi
The ith component, or mode, has the state-space representation ( Ami , Bmi , Cmi )
independently obtained from the state equations

ẋi = Ami xi + Bmi u,

yi = Cmi xi ,

n (2.28)
y= yi .
i=1

This decomposition is justified by the block-diagonal form of the matrix Am ,

The poles of a structure are the zeros of the characteristic equations . The equa-
tion s 2 + 2ζi ωi s + ωi2 = 0 is the characteristic equation of the ith mode. For small
damping the poles are complex conjugate, and in the following form:

s1,2 = −ζi ωi ± jωi 1 − ζi2 . (2.29)

The plot of the poles is shown in Fig. 2.4, which shows how the location of a
pole relates to the natural frequency and modal damping.

2.2.4 Models with Rigid Body Modes

Antenna structure is unrestrained—it can free rotate with respect to azimuth and
elevation axes. Modal analysis for such structures shows that they have zero natural
frequency, and that the corresponding natural mode shows structural displacements

@Seismicisolation
@Seismicisolation
20 2 Analytical Models

Im

s1 2
ωi sqrt(1– ζi )

ωi

–arcsin(ζi)

–ζiωi 0 Re

ωi
s2
– ωi sqrt(1– ζi )
2

Fig. 2.4 Pole location of the ith mode of a lightly damped structure: It is a complex pair with the
real part proportional to the ith modal damping; the imaginary part approximately equal to the ith
natural frequency; and the radius is the exact natural frequency

but no flexible deformations. A mode without flexible deformations is called a rigid-

body mode. Corresponding zero frequency implies that the zero frequency harmonic
excitation (which is a constant force or torque) causes rigid-body movement of the
structure. Structural analysts sometimes ignore this mode, as there is no deformation
involved. However, it is of crucial importance for a control engineer, because this
mode is the one that allows a controller to move a structure and track a command.
The rigid-body modes are obtained by solving the same eigenvalue problem as
presented for the standard models. Because the natural frequency is zero, the modal
equation becomes det(K ) = 0, that is, the stiffness matrix becomes singular. The
corresponding rigid-body mode φr b is the one that satisfies the equation

K φr b = 0. (2.30)

The modal equations for the rigid-body modes follow from (2.20) by assuming
ωi = 0, that is,

q̈mi = bmi u,

yi = cmqi qmi + cmvi q̇mi , (2.31)


n
y= yi .
i=1

@Seismicisolation
@Seismicisolation
2.2 Structural Model 21

The state-space modal model for a rigid-body mode is as follows,

   
01 0

Ami = , Bmi = , Cmi = cmqi cmvi , (2.32)

00 bmi

For the experienced engineer the rigid-body frequency and mode are not diffi-
cult to determine: rigid-body frequency is always zero, and rigid-body mode can be
predicted as a structural movement without deformation. The importance of distin-
guishing it from “regular” modes is the fact that they make a system unstable, and
thus a system that requires special attention.
The Deep Space Network antenna has two rigid-body modes: rigid-body rotation
with respect to the azimuth (vertical) axis, and rigid-body rotation with respect to the
elevation (horizontal) axis. Figure 2.5 shows the azimuth rigid-body mode. Figure
2.5(a) presents the initial position from the top view; Fig. 2.5(b) presents the modal
displacement (rigid-body rotation with respect to the azimuth axis) from the top
view,

2.2.5 Discrete-Time Model

For the antenna performance simulations and for the controller tuning purposes
the discrete-time models are required. The discrete time model is obtained from
a continuous-time state-space representation (A,B,C). Because the discrete-time
sequences of this model are sampled continuous-time signals, that is,

xk = x(k⌬t), u k = u(k⌬t), and yk = y(k⌬t), (2.33)

(a) (b)

Fig. 2.5 Antenna in neutral position (a), and the azimuth rigid-body mode (b), where no flexible
deformations are observed

@Seismicisolation
@Seismicisolation
22 2 Analytical Models

for k = 1, 2, 3 . . ., thus the corresponding discrete-time representation for the

sampling time ⌬t is (Ad , Bd , Cd , Dd ), where
 ⌬t
Ad = e A⌬t
, Bd = e Aτ Bdτ , Cd = C, (2.34)
0

and the corresponding state-space equations are

xk+1 = Ad xk + Bd u k ,
(2.35)
yk = C xk .

The discretization can be carried out numerically using the c2d command of
Matlab.

Next, the discrete-time models is presented in modal coordinates. Assume small

damping and the sampling rate sufficiently fast, such that the Nyquist sampling
theorem is satisfied (i.e., ωi ⌬t  π for all i), see, for example, [1, p. 111], then
the state matrix in modal coordinates Adm is block-diagonal,

Adm = diag(Admi ), i = 1, . . . , n. (2.36)

The 2 × 2 blocks Admi are in the form (see [4]),

 
cos(ωi ⌬t) − sin(ωi ⌬t)
Admi = e−ζi ωi ⌬t , (2.37)
sin(ωi ⌬t) cos(ωi ⌬t)

where ωi and ζi are the ith natural frequency and the ith modal damping, respec-
tively. The modal input matrix Bdm consists of 2 × 1 blocks Bdmi ,
⎡ ⎤
Bdm1
⎢ Bdm2 ⎥
⎢ ⎥
Bdm = ⎢ . ⎥, (2.38)
⎣ .. ⎦
Bdmn

where
 
1 sin(ωi ⌬t) −1 + cos(ωi ⌬t)
Bdmi = Si Bmi , Si = , (2.39)
ωi 1 − cos(ωi ⌬t) sin(ωi ⌬t)

@Seismicisolation
@Seismicisolation
2.2 Structural Model 23

and Bmi is the continuous-time modal representation. The discrete-time modal

matrix Cdm is the same as the continuous-time modal matrix Cm .
The poles of the matrix Adm are composed of the poles of matrices Admi , I =
1,. . ., n. For the ith mode the poles of Admi are

s1,2 = e−ζi ωi ⌬t (cos(ωi ⌬t) ± sin(ωi ⌬t)). (2.40)

The location of the poles is shown in Fig. 2.6, which is quite different from the
continuous-time system, cf., Fig. 2.5. For a stable system they should be inside the
unit circle.
The question arises how to choose the sampling time ⌬t. Note that from the
Nyquist criterion the ith natural frequency is recovered
 if the sampling rate is at
least twice the natural frequency in Hz ( f i = ωi 2π ), that is, if

1
 2 fi
⌬t

or, if
π
ωi ⌬t  π or ⌬t  . (2.41)
ωi

Considering all modes, the sampling time will be smaller than the smallest π ωi ,
π
⌬t  . (2.42)
max(ωi )
i

Im

–ri = exp(–ζiωiΔ t)
s1
ri sin(ωiΔ t)

ωiΔt Re
ri cos(ωiΔ t)

r=1 –ri sin(ωiΔ t) s2

Fig. 2.6 Pole location of the ith mode of a lightly damped structure in discrete time: It is a complex
pair with angle proportional to natural frequency and magnitude close to 1

@Seismicisolation
@Seismicisolation
24 2 Analytical Models

2.3 Drive Model

The drive model consists of motor, reducer, amplifiers, and tachometers.

2.3.1 Motor Model

The motor model is shown in Fig. 2.7. The motor position (θm ) is controlled by the
armature voltage (v a )
di o
va = L a + Ra i o + kb ωm (2.43)
dt

where Ra is motor resistance, L a is motor inductance, and kb is the armature

constant. The motor torque (To ) is proportional to the motor current (i o )

To = km i o (2.44)

where km is the motor torque constant. The motor torque, To is in equilibrium with
the remaining torque acting on the rotor; therefore,

To = Jm θ̈m + T (2.45)

where T is the drive output torque, and Jm is a total inertia of the motor and the
brake.
The above equations give the following (after Laplace transform, where s is the
Laplace variable)
v a − kb ωm
io =
L a s + Ra
1
ωm = (To − T )
Jm s

motor reducer
Ra io La Tm
θm
N
va Jm Tg T
kg

θg

Fig. 2.7 Motor and reducer model

@Seismicisolation
@Seismicisolation
2.3 Drive Model 25

2.3.2 Reducer Model

The reducer model is shown in Fig. 2.7 as well. The torque T at the output of the
gearbox can be expressed with the torsional deformation of the gearbox

T = k g (θg − θ p ) (2.46)

θ p is the angle of rotation of the pinion, k g is the stiffness of the reducer and drive
shaft, N is the reducer gear ratio, θg is the angle of rotation of the reducer output
shaft

θm
θg = (2.47)
N

2.3.3 Drive Model

Based on the above equations the drive model (e.g., Simulink) is created. Besides
motor and reducer, amplifiers are added. A block-diagram of the drive is created;
see Fig. 2.8. The parameters of this model are given in Table 2.1.

AMPLIFIERS ARMATURE MOTOR AND REDUCER

v2 io TN
kc 1/N

_ _
vo k1k5 v1 kr(1 + τ2 s) + kf ki(1 + τ4 ) 1 1
ωm kg T
km +
1 + τ1 s + _ s(1 + τ3 s) s(1 + τ5 s) + _ Ra + La s Jm s + _ Ns

ωm ω
vs N
kb
ωm
kt

Fig. 2.8 Block diagram of the drive model

Table 2.1 Drive parameters

k1 = 716.2 [V s/rad] kf = 54 [-] [-] τ1 = 0.00637 [s]
km = 1.787 [Nm/A] kg = 1.7× 106 [Nm/rad] τ2 = 0.094 [s]
kb = 1.79 [V s/rad] Jm = 0.14 [Nm/s2 ] τ3 = 0.002 [s]
ks = 0.8 [-] Ra = 0.456 [⍀] τ4 = 0.00484 [s]
kt = 0.0384 [V s/rad] La = 0.011 [H] τ5 = 0.0021 [s]
kr = 80 [V/s/V] N = 354 [-] τ6 = 0.7304 [s]
kc = 0.1266 [V/A]

@Seismicisolation
@Seismicisolation
26 2 Analytical Models

1 Velocity Command
y(n) = Cx(n)+Du(n) 1
Velocity Torque
Pinion velocity x(n+1) = Ax(n)+Bu(n) Encoder
Command
Drive Antenna Structure

pinion velocity

Fig. 2.9 Velocity loop model

100

10–1
magntiude

10–2

from analysis
10–3
from identification

10–1 100 101

frequency, Hz

Fig. 2.10 The magnitude of the transfer function of the velocity loop model

2.4 Velocity Loop Model

The Simulink velocity loop model is a combination of the structural and drive
models; see Fig. 2.9. The drive torque moves the antenna. The pinion velocity is
fed back to the drive to calculate the reducer flexible deformation. The velocity
command controls the antenna velocity, and the encoder measures its position.
The magnitude of the velocity loop transfer function is shown in Fig. 2.10,
solid line.

2.5 Drive Parameter Study

In this section the effect of the drive inertia and stiffness on the antenna velocity
loop properties is investigated. The inertia of the drive includes motor, brake,
and gearbox, and the drive stiffness includes gearbox stiffness and shaft stiffness.
Because the motor dominates the drive inertia, the motor inertia is investigated,
and because gearbox stiffness dominates the drive stiffness, the drive stiffness is
investigated.

@Seismicisolation
@Seismicisolation
2.5 Drive Parameter Study 27

102

101

100
elevation rate

10–1

kg
10–2 10*kg
100*kg

10–3 –2
10 10–1 100 101 102
frequency, Hz

Fig. 2.11 The magnitudes of the transfer function of the velocity loop model for the nominal and
increased gearbox stiffness

It is intuitively obvious that a rigid gearbox and a powerful but weightless motor
are the best choice for the antenna drive (if they exist). So, what is the smallest allow-
able gearbox stiffness, or the largest allowable motor inertia? The 34-m antenna
velocity loop bandwidth is simulated as a function of motor inertia, and as a func-
tion of the gearbox stiffness [3], using the model as in Fig. 2.8 and Fig. 2.9. The
nominal motor inertia is 0.14 [Nms2 ], and the nominal gearbox stiffness is 1.7 ×
106 [N m/rad]; see Table 2.1. The transfer function, from the velocity command to
the output velocity for the nominal values, is shown in Fig. 2.11, solid line.

2.5.1 Drive Stiffness Factor

The impact of the gearbox stiffness on the velocity loop properties is simulated, for
the nominal gearbox stiffness and for increased stiffness by factors of 10 and 100,
respectively. The corresponding transfer functions are shown in Fig. 2.11, dotted (for
factor 10) and dot-dashed lines (for factor 100), respectively. One can see that the
transfer function bandwidth changed when k g incresed to 10k g . The resonant peak,
related to the drive vibration mode shifted to higher frequencies. At the nominal
value of the stiffness, the resonant frequency of the drive mode coincided with
the structural fundamental frequency. The plots show that the nominal stiffness is

@Seismicisolation
@Seismicisolation
28 2 Analytical Models

102

101

100
elevation rate

10–1

kg
10–2
0.1*kg
0.01*kg

10–3
10–2 10–1 100 101 102
frequency, Hz

Fig. 2.12 The magnitudes of the transfer function of the velocity loop model for the nominal and
decreased gearbox stiffness

slightly too low, because the inreased gearbox stiffness increased the velocity loop
bandwidth.
Next, the gearbox stiffness is decreased by factors of 0.1 and 0.01, and the corre-
sponding transfer functions are shown in Fig. 2.12 as dotted and dot-dashed lines,
respectively. One can see that the transfer function bandwidth changed significantly.
The resonant peak, related to the drive vibration mode is shifted to lower frequen-
cies, impacting the bandwidth. The plots show that the nominal stiffness should
not be decreased, because the lower gearbox stiffness impacts the velocity loop
bandwidth.

2.5.2 Drive Inertia Factor

Here, the impact of the motor inertia on the velocity loop properties is investigated.
The transfer function from the velocity command to the output velocity for the
nominal inertia is shown in Fig. 2.13, solid line. First, the motor inertia is decreased
by factors 0.1 and 0.01, and the corresponding transfer functions are shown in
Fig. 2.13, dotted and dot-dashed lines, respectively. One can see that the transfer
function bandwidth changed insignificantly. The resonant peak, related to the drive
vibration mode shifted to higher frequencies. The plots show that the nominal inertia

@Seismicisolation
@Seismicisolation
2.5 Drive Parameter Study 29

102

101

100
elevation rate

10–1

Jm
10–2 0.1*Jm
0.01*Jm

10–3
10–2 10–1 100 101 102
frequency, Hz

Fig. 2.13 The magnitudes of the transfer function of the velocity loop model for the nominal and
decreased motor inertia

102

101

100
elevation rate

10–1

Jm
10–2
10*Jm
100*Jm

10–3
10–2 10–1 100 101 102
frequency, Hz

Fig. 2.14 The magnitudes of the transfer function of the velocity loop model for the nominal and
increased motor inertia

@Seismicisolation
@Seismicisolation
30 2 Analytical Models

is sufficient, because the lower motor inertias does not impact the velocity loop
bandwidth.
Next, the motor inertia is increased by factors 10 and 100, and the corresponding
transfer functions are shown in Fig. 2.14, dotted and dot-dashed lines, respectively.
In this case the transfer function bandwidth changed significantly. The resonant peak
related to the drive vibration mode is shifted to the lower frequencies, impacting the
bandwidth. Note that the antenna fundamental frequency has not been changed.
The plots show that the nominal inertia cannot be increased, because it impacts the
velocity loop bandwidth.
This simple investigation shows that the gearbox stiffness has its lower limit
that should not be violated, because it impacts the velocity loop bandwidth and the
closed loop performance. Similarly, the motor inertia has its upper limit. The limit
is defined by the antenna fundamental frequency: drive natural frequency should be
at least the fundamental frequency of the structure.

References
1. Franklin GF, Powell JD, Workman ML. (1992). Digital Control of Dynamic Systems. Addison-
Wesley, Reading, MA.
2. Gawronski W. (2004). Advanced Structural Dynamics and Active Control of Structures.
Springer, New York.
3. Gawronski W, Mellstrom JA. (1992). A Parameter and Configuration Study of the DSS-13
Antenna Drives. TDA Progress Report, No.42-110, available at http://ipnpr.jpl.nasa.gov/
progress report/42-110/110S.PDF.
4. Lim KB, Gawronski W. (1996). Hankel Singular Values of Flexible Structures in Discrete
Time,” AIAA J. Guidance, Control, and Dynamics, 19(6):1370–1377.

@Seismicisolation
@Seismicisolation
Chapter 3
Models from Identification

This chapter describes the process of deriving the velocity loop model from field
test data. It includes the description of the white noise testing of an antenna, test
conditions, input and output signals, open- and closed-loop testing, the instrumenta-
tion set-up, and the system identification test. Analytical background for the data
description and data processing are also included. Next, the dependency of the
antenna model on its elevation position, and on the dish diameter is discussed.
Finally, analytical and identified models are compared.

3.1 White Noise Testing of the Antenna

The white noise signal is applied at the input of the velocity loop to excite the
antenna dynamic, and the antenna reaction is measured at the encoder. The white
noise is a random signal of even spectrum within the antenna bandwidth; hence
each frequency component of the antenna dynamics is excited evenly. It should guar-
antee to obtain a meaningful transfer function and trustworthy model of the velocity
loop.

3.1.1 Purpose and Conditions

The purpose of the white noise testing of the antenna is the determination of the
antenna velocity loop model. The accurate model is necessary to obtain a stable
and precise model-based controller. Other signals, beside the white noise, can be
used for the testing and model identification, for example, sweep-sine. It was found,
however, that the white noise testing is the most effective and least damaging
when applied to antennas. The sweep-sine excites antenna resonances, causing
wear, and excites antenna backlash since resonance forces exceed anti-backlash
torques.
During the tests the antenna is positioned at 45 deg elevation, azimuth position is
arbitrary, and wind shall be mild (below 12 km/h).

W. Gawronski, Modeling and Control of Antennas and Telescopes, 31

@Seismicisolation
@Seismicisolation
DOI: 10.1007/978-0-387-78793-0 3, 
C Springer Science+Business Media, LLC 2008
32 3 Models from Identification

3.1.2 Test Input and Output

The following signals are generated and/or measured during the test:
• The test input signal, denoted u o It is a discrete white noise, generated by the data
collection computer. The noise is converted into analog signal (voltage).
• The antenna input signal, denoted u. It is the total voltage at the velocity input.
• The antenna output signal, denoted y. It is the encoder reading (digital signal).

The common characteristics of the input and output signals are their sampling
rate, the test duration, signal amplitude, and offset.

• Sampling rate is obtained from the Nyquist criterion, described later. Typical
sampling rate for 34-m antennas is 30 Hz. It can be chosen, however, between 25
and 50 Hz.
• Duration. The test duration determines the system identification accuracy, and
the lowest frequency of the obtained transfer function. Because the input is a
random signal, the accuracy of the transfer function depends on the number of
averaging of the collected data. The longer is the collected data file, the smoother
is the average. Also, the longer is the data segment, the lower is the frequency of
the obtained transfer function. Thus, a large data file is needed. But the limiting
factor for data size is the computer memory (less and less important nowadays),
and patience of a test engineer (more and more important nowadays). Our expe-
rience shows that 40,000 samples are sufficient, which for 30-Hz sampling rate
translates into 20–30 min of test time.
• Amplitude. Amplitude of the test noise shall be large enough to excite antenna
dynamics, and be larger than the disturbances (mostly wind gusts). However, it
shall be low enough in order to prevent antenna wear and damage; thus the white
noise amplitude shall be at least 5% of maximal allowable amplitude.
• Velocity offset. A non-zero input signal offset is applied in order to overcome
antenna static friction. The nonzero velocity offset forces an antenna to move
with a constant velocity. The offset is about 1% of the maximum velocity.

A typical white noise input is shown in Fig. 3.1a. Note that the amplitude is 0.5 V,
and the offset is –0.1 V.
The output signal (encoder reading) is a digital signal of the same sampling rate
as the input signal. Due to the input offset, the output signal shows a slope propor-
tional to the input offset; see Fig. 3.1b.

3.1.3 Test Configuration

The test can be conducted with the antenna position loop open or closed; see
Fig. 3.2 and Fig. 3.3.
Open loop test. The open loop test configuration is shown Fig. 3.2. By setting
the antenna in the manual mode the position loop is broken, the position controller

@Seismicisolation
@Seismicisolation
3.1 White Noise Testing of the Antenna 33

velocity input, V
0.5 offset

–0.5

0 20 40 60 80 100 120 140 160 180 200

sample number

35.55
EL encoder, deg

35.54

35.53

35.52

0 20 40 60 80 100 120 140 160 180 200

sample number

Fig. 3.1 White noise input signal of 0.14 V offset, the first 200 samples, and antenna response to
the white noise signal the first 200 samples

is disconnected, and the white noise input (u) is injected into the velocity loop. The
white noise is recorded by the data computer. The encoder digital signal (y) and
tachometer voltage (v t ) are recorded synchronously by the computer.

encoder (y)

test signal
(uo) u = uo

command Antenna
Position
velocity
+_ controller encoder
loop
velocity input

Fig. 3.2 Open loop test configuration (u = u o )

@Seismicisolation
@Seismicisolation
34 3 Models from Identification

encoder (y )

test signal
(uo) u = uo + uf

command Antenna
Position uf
velocity
+_ controller encoder
loop
velocity input

Fig. 3.3 Closed loop test configuration (u = u o )

Closed loop test. The closed loop tests are useful when the antenna is unstable
or drifting when the position loop is open. An interesting review of the closed loop
identification is given in [5]. The closed loop test configuration is shown Fig. 3.3.
The white noise input (u o ) is injected into the velocity loop, with the position loop
closed. The total signal (u) after the summing junction is recorded by the data
computer. Note that the signal after the summing junction is not identical with
the test input signal u o , because it is a sum of u o and the feedback signal u f , that
is, u = u o + u f . The encoder digital signal (y) is recorded synchronously by the
computer.

3.1.4 Data Processing

The collected input (u) and output (y) data are used for the determination of the
velocity loop transfer function and for the system identification (i.e., the velocity
loop state-space model).
The open loop transfer function is determined using the following steps; see [1].
Let the u and y records of length of 20,000 samples be divided into n =10 segments
of length T = 2000 Δt s. The segments are denoted u i and yi , where i =1,. . .,n.
The FFT algorithm applied to the segments gives u i and yi spectra, denoted Ui
and Yi , respectively. The mean value of their product represents the estimate of the
cross-spectral density function Puy (ω):

2  ∗
n
Puy (ω) = U (ω)Yi (ω) (3.1)
nT i=1 i

where Ui∗ denotes the complex conjugate of Ui . The estimate is smoothed through
averaging (this was the reason for cutting the records into segments).

@Seismicisolation
@Seismicisolation
3.1 White Noise Testing of the Antenna 35

Similarly, one obtains a smooth estimate of the input power spectral density func-
tion Pu (ω)

2 
n
Pu (ω) = |Ui (ω)|2 (3.2)
nT i=1

and the output power spectral density function Py (ω)

2 
n
Py (ω) = |Yi (ω)|2 (3.3)
nT i=1

From these functions one obtains an estimate Ĝ(ω) of the open loop transfer
function G(ω) as follows

Puy (ω)
Ĝ(ω) = (3.4)
Pu (ω)

The above transfer function is an approximate one. The coherence function is an

indicator of the quality of the approximation. It is defined as
 
 Puy (ω)
γ (ω) =  (3.5)
Pu (ω)Py (ω)

Its value varies between 0 and 1. The unity value indicates that a perfect linear
relationship exists between the input and the output, and that no disturbances were
recorded. A non-linear behavior of the antenna, and/or a presence of excessive
disturbances (wind) cause low value of coherence.

The transfer function and coherence are calculated using the Matlab command
p=spectrum(u,y,n) where n is the length of the segment, and P consists of eight
columns, such that
Pu ( f ) is the first column of p,
Py ( f ) is the second column of p,
Ĝ( f ) is the fourth column of p,
γ ( f ) is the fifth column of p,
For the sampling frequency of the data is f s , the frequency range f is
fs  n 
f = 1: (HZ )
n 2

The magnitude and phase of the velocity loop transfer function obtained from the
data presented in Fig. 3.1 is shown in Fig. 3.4. The coherence is shown in Fig. 3.5.
The mean value of the coherence is larger than 0.8 for frequencies between 0.06 Hz

@Seismicisolation
@Seismicisolation
36 3 Models from Identification

magnitude 0

0.01

0.0001

10–2 10–1 100 101

frequency, Hz

–100
phas e, deg

–200

–300

–400
10–2 10–1 100 101
frequency, Hz

Fig. 3.4 The magnitude and phase of the 34-m antenna transfer function obtained from the test
data

0.8

0.6
coherence

0.4

0.2

0
10–2 10–1 100 101
frequency, Hz

Fig. 3.5 The coherence function for the data in Fig. 3.1

@Seismicisolation
@Seismicisolation
3.1 White Noise Testing of the Antenna 37

and 6 Hz. That is, the obtained transfer function is considered accurate within this
frequency interval. Note that for selected frequencies (e.g., 2.8 Hz and 4.6 Hz) the
value of coherence is low. These two frequencies correspond to the anti-resonances
of the transfer function (cf. Fig. 3.4), where antenna does not respond to the
excitation.

3.1.5 Basic Relationships for the Discrete-Time Data

Let u i = u(iΔt) be sampled data of N samples, i = 1, . . . , N ; see Fig. 3.6a. The

sampling time of a discrete-time data is denoted Δt. The corresponding sampling
rate f s is the inverse of the sampling time:

1
fs = (3.6)
Δt

The record length is T. The number of samples in the record is N, thus

N
T = N Δt = (3.7)
fs

(a)

Δt
data, ui

time, t[s]
(b)

fn
data spectrum, Ui

Δf

frequency, f [Hz]

Fig. 3.6 (a) Discrete time signal and (b) its spectrum: Δt, sampling time; T, record length; Δf ,
frequency resolution; and f n , Nyquist frequency

@Seismicisolation
@Seismicisolation
38 3 Models from Identification

Denote Ui the spectrum of the data, which is shown in Fig. 3.6b. The resolution
of lines of the frequency spectrum is Δf . It is obtained either from the record length
or from the sampling time or from the sampling rate

1 fs 1
Δf = = = (3.8)
T N N ⌬t

It is also the lowest frequency of the spectrum. For N = 2048 samples, and for
sampling frequency f s = 30 Hz one obtains the resolution Δf = 30/2048 = 0.015
Hz. The range of the spectrum f n is called the Nyquist frequency, obtained as

N 1 1
fn = = = fs (3.9)
2T 2⌬t 2

It is the maximum frequency of a digital signal. It is equal to the half of the

sampling frequency of this signal. It is also called the cut-off frequency, or the
bandwidth of the signal.
Because the desired range is usually known, the sampling frequency is obtained
from the above equation as

fs = 2 fn (3.10)

The maximal frequency must be larger then the velocity loop bandwidth, f b ,

fn > fb (3.11)

thus

fs > 2 fb (3.12)

The sampling rate, fs , must be at least twice the bandwidth of the velocity loop.
For the 34-m antenna the velocity loop bandwidth is below 10 Hz; hence 30 Hz
satisfies the Nyquist criterion (3.9). High sampling rates (above 50 Hz) add no new
information in the test; however, they significantly increase the size of data files.

3.2 Identification of the Velocity Loop Model

In this section the process of obtaining an antenna velocity loop model from the
collected input and output data is described.

@Seismicisolation
@Seismicisolation
3.2 Identification of the Velocity Loop Model 39

3.2.1 Description of the Velocity Loop Model

The antenna control system is shown in Fig. 3.7. It consists of the antenna open loop
system (velocity loop), position controller, and velocity and acceleration limiters.
The controller output u represents the commanded velocity. The antenna angular
position (measured at the encoder) is the output of the velocity loop system. The
antenna velocity loop model is obtained from field tests and the system identification
in the state-space form

x(i + 1) = Ax(i) + Bu(i) + w(i), (3.13)

y(i) = C x(i),

where x is the state, w is the disturbance, y is the antenna position, and (A,B,C) is
the state-space triple.
The size of the matrices and vectors in Eq. (3.13) is as follows: x(i) is n × 1,
w(i) is n × 1, A is n × n, B is n × 1, C is 1 × n, u(i) is a scalar, y(i) is a scalar,
n is the order of the system, and i is the sample number. The order n is the input
parameter to the system identification program, selected by a user. Typically, it is
selected n = 40, which is oversized, but necessary in order to assure the inclusion of
all significant system dynamics in the model. Later, however, the order is reduced
to n ≤ 14, suitable for the DSN antenna controller tuning purposes.

3.2.2 Identification of the Velocity Loop Model

The determination of the antenna velocity loop model is part of the LQG controller
tuning process (as the LQG estimator). Thus, the accuracy of the model and its
coordinate selection are the factors that influence the controller performance. The
accuracy is solved by determining the model through field testing of the velocity
loop antenna, applying system identification, and comparing the identified model
transfer function with the transfer function obtained from the data. Although the

r y
antenna
controller uo u (G)
(K)
rate acceleration
y limit limit

Fig. 3.7 Antenna control system: G, velocity loop system; K, position controller; r, position
command,; y, antenna angular position; u, limited velocity command; uo , unlimited velocity
command; w, disturbance

@Seismicisolation
@Seismicisolation
40 3 Models from Identification

antenna transfer function has been already determined, there is no equation that
describes the velocity loop dynamics. These equations, or a mathematical model
of the velocity loop, are obtained using the system identification software, such as
SOCIT program of the NASA Langley Research Center, see [4], or Matlab program
n4sid. The complete description of the SOCIT algorithm can be found in [3], or
in [2], Ch. 9. In our case the identified model is determined in the form of the
state-space representation (A,B,C).

One also can use the Matlab system identification program (n4sid). In this
case, having the input (u) and output (y) signals one obtains the state-space
representation (A,B,C,D) as follows
DAT = iddata (y,u,dt);
MODEL = n4sid(DAT,40, CovarianceMatrix , None );
A=MODEL.a;
B=MODEL.b;
C=MODEL.c;

The transfer function plot, using the identified models (A,B,C) is shown in
Fig. 3.8, solid line, and compared with the transfer function obtained from the
Fourier transform of the data (dashed line). The estimated and identified transfer
functions show good coincidence.

100

10–1
magnitude

10–2

10–3

from id model
from data

10–1 100 101

frequency, Hz

Fig. 3.8 Magnitude of the transfer function, from the test data (dashed line) and from the identified
model (solid line)

@Seismicisolation
@Seismicisolation
3.2 Identification of the Velocity Loop Model 41

3.2.3 A Comparison of the Analytical and Identified Models

Figure 2.10 (dashed line) shows the magnitude of the azimuth transfer function of
the 34-m antenna obtained from the identification. The transfer function plotted in
Fig. 2.10, solid line, was obtained from the analytical model of the same antenna.
One can see that the character of the transfer function is the same; however, there
are also differences. The first resonant frequency of the identified model is lower
than the analytical model. Amplitudes of the resonances at higher frequencies in the
identified model are significantly lower than in the analytical model.

3.2.4 Azimuth Model Depends on the Antenna Elevation Position

In the system identification process the antenna azimuth and elevation models are
determined separately. The models are independent of each other because there is a
very weak coupling between the axes. However, the azimuth model depends on the
elevation position of the antenna.
The antenna structural properties, as measured at the elevation axis, do not
depend on the elevation position. Thus, the elevation structural modes and frequen-
cies are also independent on the antenna elevation position. This is shown in
Fig. 3.9, with the measurements of the elevation transfer function at the antenna
elevation position 13 deg, 45 deg, and 84 deg. The magnitudes of the transfer func-
tion are identical at all three positions.
A different situation is observed for azimuth models. Azimuth model properties
depend on the antenna elevation position because the antenna inertia and stiffness
with respect to azimuth axis depend on the elevation position. It is visible with the
measurements of the azimuth transfer function at the antenna elevation position 10,
30, 45, 60, and 90 deg; see Fig. 3.10.

100
elevation magnitude

10–1

10–2

EL = 13 deg
EL = 45 deg
10–3
EL = 84 deg

10–1 100 101

frequency, Hz

Fig. 3.9 Elevation transfer function does not depend on antenna elevation position

@Seismicisolation
@Seismicisolation
42 3 Models from Identification

(a)

natural frequencies: first second third fourth

100
azimuth magnitude

10–1

10–2
EL = 10 deg
EL = 30 deg
EL = 45 deg
10–3
EL = 60 deg
EL = 90 deg

100 101

(b)

natural frequencies: first second third fourth

100
azimuth magnitude

10–1

10–2
EL = 10 deg
EL = 30 deg
EL = 45 deg
10–3 EL = 60 deg
EL = 90 deg

100 101
frequency, Hz

Fig. 3.10 Azimuth transfer function depends on antenna elevation position: (a) 34-m HEF antenna,
and (b) 34-meter BWG antenna

@Seismicisolation
@Seismicisolation
3.2 Identification of the Velocity Loop Model 43

In Fig. 3.10a the damping of the first mode decreases with an increase of the
elevation position, while the natural frequency of the third mode increases with
an increase of the elevation position. In Fig. 3.10b the damping of the first mode
does not change with the elevation position, while the natural frequency of the third
mode decrease with the increase of the elevation position. Both antennas are 34-m
antennas, but different types. Thus, there is no obvious pattern in the dynamical
properties of the antennas. For the controller tuning purposes the mean model is
chosen, the one at elevation position of 45 deg.

3.2.5 Fundamental Frequency Depends on Antenna Diameter

Antenna natural frequencies depend also on the antenna size. It is a general tendency
that the natural frequencies decrease with the increase of the antenna size (the
structure becomes “softer”). The lowest natural frequency (called the fundamental
frequency) is considered a measure of the compliance of a structure.
The data of fundamental frequencies of many antenna structures have been
collected by D.D. Pidhayny, A.R. Lewis, and S.R. Bandel of Aerospace Corpo-
ration. In these data the decreasing tendency of frequencies with the increase of
antenna dish size is observed. Based on these data, the best-fit line is determined (in
logarithmic scale). It relates the antenna dish diameter (in meters) with the antenna
fundamental frequency (in Hertz). It is given by the following equation

f = 20 d −0.7 (3.14)
antenna fundamental frequency, Hz

10

1 10 100
antenna diameter, m

Fig. 3.11 The best-fit line that fits the Aerospace Corp. chart of the antenna fundamental frequen-
cies

@Seismicisolation
@Seismicisolation
44 3 Models from Identification

The line is presented in Fig. 3.11. This equation represents the average natural
frequency for a given antenna diameter. Equation (3.14) allows evaluating the struc-
tural soundness of a particular antenna: if the fundamental frequency of the consid-
ered antenna structure is higher than the frequency obtained from Eq. (3.14), the
structure is stiffer than the average (thus, better pointing performance is expected);
if it is lower, the structure is softer than average (thus causing inferior pointing
performance).

References
1. Bendat, JS, Piersol AG. (1986). Random Data. John Wiley, New York.
2. Gawronski W. (2004). Advanced Structural Dynamics and Active Control of Structures.
Springer, New York.
3. Juang JN. (1994). Applied System Identification, Prentice-Hall, Englewood Cliffs, NJ.
4. Juang JN, Horta LG, Phan M. (1992). System/observer/controller identification toolbox. NASA
Tech Memo 107566.
5. Landau ID. (2001). Identification in closed loop: a powerful design tool (better design models,
simpler controllers). Control Engineering Practice 9: 51–65.

@Seismicisolation
@Seismicisolation
Chapter 4
Model Reduction

The velocity loop model usually consists of too many state variables (i.e., a smaller
number would reflect the system dynamics with similar precision). There are
different methods of reducing the number of variables in the velocity loop models;
see, for example, [3]. The balanced and modal model reduction methods have been
selected, which are well tested in the antenna environment. Also, the reduction of
a system with rigid-body modes (such as antennas) is discussed; it requires special
attention because the standard approaches fail.

4.1 Why Reduction?

Analytical models have an excess of state variables that describe the antenna
dynamics, due to the large number of degrees of freedom of the structural model
and variables of the drives. For example, some drive variables may be important
in the drive dynamics, but less important in the overall velocity loop dynamics.
Thus, the model order reduction is recommended. The identified model also has
excessive order, because the unimportant structural modes (weakly observed at the
output and/or weakly excited at the input) are part of the model. Additionally, the
measurement noise is included in the model.
Besides the reduction of the model complexity, there is another important factor
that requires model reduction: the model-based controller property. The order (and
complexity) of the controller is equal to the order of the antenna model. In the imple-
mentation one requires an accurate model, which is as simple as possible (but not
simpler!) to ease the implementation. This is the task of model reduction.

4.2 Balanced Model Reduction

The order of the antenna model obtained from the system identification is too
high for the implementation, and it includes dynamics that do not represent the
antenna, but possible disturbances and measurement noise. The mode order has to
be reduced. The reduced model shall include all relevant antenna dynamics in order
to assure a stable LQG controller.

W. Gawronski, Modeling and Control of Antennas and Telescopes, 45

@Seismicisolation
@Seismicisolation
DOI: 10.1007/978-0-387-78793-0 4, 
C Springer Science+Business Media, LLC 2008
46 4 Model Reduction

In order to reduce the model the model (e.g., obtained from the system identifi-
cation) is transformed into the balanced representation. The balanced representation
is described in [3].

It can be executed in Matlab as follows:

[sysb, g] = balreal(sys);
or, equivalently
[Ab, Bb, Cb, g] = balreal(A, B, C);
where sys is the original system (or with the state space representation (A, B,
C)), while sysb is the balanced system (or with the state space representation
(Ab, Bb, Cb)), and g is the vector of the Hankel singular values (HSV).

As part of the transformation the Hankel singular values (HSV, also called Hankel
norms) of the system are obtained. The model is presented in descending order of
the HSV, as follows:
⎧ ⎫
⎪
⎪ x1 ⎪⎪
⎪
⎨ x2 ⎪⎬
x= .. (4.1)
⎪
⎪ . ⎪
⎪
⎪
⎩ ⎪ ⎭
xn

where xi is the ith state and x1 is the state with the highest HSV, while xn is the state
with the lowest HSV.
The reduced-order model is obtained by truncating the least important states.
Because the states with the smallest HSVs are the last ones in the state vector, a
reduced-order model is obtained here by truncating the last states in the state vector.
How many such states? Enough that the remaining states represent the significant
portion of the antenna dynamics (e.g., the part represented in the magnitude of
the transfer function above a selected significance or noise level). “The significant
portion of the antenna dynamics” is an ambiguous statement, but it is a part of the
art of engineering, which includes experience and intuition.
Let (Ab , Bb , Cb ) be the balanced representation corresponding to the state vector
x as in (4.1). Let x be partitioned as follows:
 
xr
x= , (4.2)
xt

where xr is the vector of the retained states and xt is a vector of truncated states. If
there are k < n retained states, xr is a vector of k states, and xt is a vector of n − k

@Seismicisolation
@Seismicisolation
4.3 Modal Model Reduction 47

states. Let the state triple (Ab , Bb , Cb ) be partitioned accordingly,

   
Ar Ar t Br

Ab = , Bb = , Cb = Cr Ct . (4.3)
Atr At Bt
The reduced model is obtained by deleting the last n − k rows of Ab , Bb , and
the last n − k columns of Ab , Cb , obtaining the reduced triple ( Ar , Br , Cr ), marked
boldface in equation (4.3).

4.3 Modal Model Reduction

Similar results can be obtained using the modal model described in Chapter 2. For
this, a value of each mode should be obtained to decide which mode should be
retained or removed in the reduction process. Each mode can be “valued” in three
different ways: using H2 , H∞ , or Hankel norms.

4.3.1 Norms of a Single Mode

For structures in the modal representation, each mode is independent; thus the norms
of a single mode are independent as well (they depend on the mode properties, but
not on other modes).
The H2 norm of the ith mode is

Bmi 2 Cmi 2
G i 2 ∼
= √ (4.4)
2 ζi ωi
The H∞ norm is estimated as

Bmi 2 Cmi 2
G i ∞ ∼
= (4.5)
2ζi ωi
The Hankel norm is determined from

Bmi 2 Cmi 2
G i h ∼
= (4.6)
4ζi ωi

where Bmi and Cmi are defined in Chapter 2; equations (2.32).

For single-input single-output systems—which is the case for antennas or
telescopes—the above norms are easily determined from the magnitude of the
transfer function. Namely, the H∞ norm is of the ith mode and is the height of its
resonant peak; see Fig. 4.1. The Hankel norm is half of this peak. The determination
of the H2 norm of a mode is a little more complicated. Let Δωi be a half-power
frequency at the ith resonance, Δωi = 2ζi ωi ; see [1], [2]. It is the width of the

@Seismicisolation
@Seismicisolation
48 4 Model Reduction

4
G1 ∞
= 2 G1 h
mode 1
3
magnitude

G1 ∞
2

2 mode 2

1
Δω1 = 2ζ 1ω1
0
0 1 2 3 4 5 6 7 8
frequency, Hz

Fig. 4.1 The determination of the half-power frequency, H2 , Hankel and H∞ norm for the first

shaded area in √ this figure, obtained as a cross section of the resonance peak at the
height of h i / 2, where h i is the height of the resonance peak.
In practice, the norm can be evaluated visually through the inspection of the
magnitude of the transfer function. When Hankel or H∞ norm is used, the reduction
is carried out by eliminating modes with lowest resonances; when H2 norm is used,
the height and the width of the resonances is used.
Comparing (4.4)–(4.6) the approximate relationships between H2 , H∞ , and
Hankel norms is obtained


G i ∞ ∼
= 2 G i h ∼
= ζi ωi G i 2 . (4.7)

The above relationship is illustrated in Fig. 4.2

4.3.2 Norms of a Structure

Next, the norms of the whole structure are expressed in terms of the norms of its
modes.
The structure H2 norm is approximately the rms sum of the modal norms

 n

G2 ∼
= G i 22 , (4.8)
i=1

where n is the number of modes, and G i 2 is given by (5.21).

@Seismicisolation
@Seismicisolation
4.4 Antenna Model Reduction 49

H2
102 Hinf
Hankel
modal norm

100

10–2

10–3 10–2 10–1 100 101 102

ζiωi, rad/s

Fig. 4.2 Modal norms versus ζi ωi

The structure H∞ norm is the largest of the mode norms, i.e.,

G∞ ∼
= max G i ∞ , i = 1, . . . , n. (4.9)
i

where n is the number of modes, and G i ∞ is given by (5.22).

The Hankel norm of the structure is the largest norm of its modes, and it is half
of the H∞ norm, that is,

Gh ∼
= max G i h = 0.5 G∞ , (4.10)
i

where n is the number of modes, and G i h is given by (5.23).

4.4 Antenna Model Reduction

Antennas and telescopes have rigid-body modes, that is, they have poles at zero.
But, the H2 , H∞ , and Hankel norms for systems with poles at zero do not exist
as their values tend to infinity; see equations (4.4)–(4.6) for ωi = 0. However, the
infinite values of the norms of modes are not an obstacle in the reduction process.
These values indicate that the corresponding states should be retained in the reduced
model, regardless of the norms of other modes. Here a simple approach is used for
the reduction of systems with poles at zero.
In this approach, the system with a pole at zero is represented in modal coordi-
nates by the following system triple:
   
0 0 Br

A= , B= , C = Cr C o , (4.11)
0 Ao Bo

@Seismicisolation
@Seismicisolation
50 4 Model Reduction

100

10–1
magnitude

10–2

10–3
full model
reduced model

10–1 100 101

frequency, Hz

Fig. 4.3 Magnitudes of the AZ transfer functions of the 40-state full model (solid line) and 10-state
balanced reduced model (dashed line)

where the triple ( Ao , Bo , Co ) has no poles at zero, but is itself in modal coordinates.
The vector of the corresponding modal norms is denoted h o . This vector is arranged
in descending order, and the remaining infinite norms are added

h = {∞, h o } (4.12)

to obtain the vector of norms of the (A, B, C) representation. The system is reduced
by truncation, as described at the beginning of this chapter.
Figure 4.3 shows the transfer functions of the 40-state antenna AZ model, and
the transfer functions of the reduced (14-state) model using balanced representation.
One can see that the antenna rigid-body motion (represented by the straight line at
frequencies up to 1 Hz) and the antenna major resonances have been included in the
reduced-order model, while the high-frequency dynamics, above 9 Hz, have been
eliminated.

References
1. Clough RW, Penzien J. (1975). Dynamics of Structures. McGraw-Hill, New York.
2. Ewins DJ. (2000). Modal Testing. Research Studies Press, Baldock, England.
3. Gawronski W. (2004). Advanced Structural Dynamics and Active Control of Structures.
Springer, New York.

@Seismicisolation
@Seismicisolation
Chapter 5
Wind Disturbance Models

Wind is the main source of antenna/telescope disturbance. It loads the antenna struc-
ture and causes pointing problems. The wind load is divided into mean (or static
load) and variable load (gusts). The load depends on the wind speed, on direction of
the wind with respect to the antenna structure (yaw), and on the antenna elevation
position (pitch).
This chapter discusses models of steady-state wind, and a model of wind gusts.
The steady-state model is obtained from wind tunnel data for different yaw angles
and different antenna elevation positions. The model is verified with the 34-m
antenna wind data obtained in the field.
The wind gust models include forces acting on the dish, wind torques at the
drives, and wind disturbance at the velocity input of the velocity loop.

5.1 Steady-State Wind Disturbance Model

The antenna wind loads are typically separated into two components: steady-state
(or mean, or static) and dynamic (gusts) load. Both the static and dynamic wind
loads are used to estimate antenna pointing accuracy. The dynamic loads are
discussed in [5] and [6].
The mean wind loads are used to size antenna drives and motors. They are
typically obtained by scaling the wind tunnel data. The primary source of such
data is the result of the wind tunnel tests of antenna wind loading. One of the
tests was conducted over 40 years ago at California Institute of Technology, as
reported in [1–3]. The wind tunnel data presented in the CP-6 Memorandum [1]
are regarded as the most reliable because they were collected using model of the
entire antenna rather than the dish-only as in [2] and [3]. In addition, the ground
effects and a 25% porosity of the outer portion (25% of the radius) of the antenna
were modeled during the testing described in [1]. As a result, the modeled antenna
represented geometrical properties of a typical radiotelescope (including the DSN
antennas). The geometrical scale employed in wind tunnel testing described in the
CP-6 Memorandum was 1:75, which resulted in the 0.46 m (diameter) model of the
antenna 34-m reflector.

W. Gawronski, Modeling and Control of Antennas and Telescopes, 51

@Seismicisolation
@Seismicisolation
DOI: 10.1007/978-0-387-78793-0 5, 
C Springer Science+Business Media, LLC 2008
52 5 Wind Disturbance Models

5.1.1 Dimensionless Wind Torques

Wind-induced torque in an antenna drive, Tw , depends on the antenna geometry and
orientation relative to the wind, and also on the wind characteristics. The principal
geometrical parameters affecting this loading are shown in Fig. 5.1, where the eleva-
tion and plan views of the antenna are schematically depicted. These parameters
include the dish diameter, D; the distance, d, of the dish vertex from the elevation
axis; the elevation angle, βe ; the yaw angle βa , and the antenna porosity.
The wind-induced torques in the antenna drives are represented in a dimension-
less form as torque coefficients τw ; see [10], and [11],

Tw
τw = (5.1)
p AD

where Tw is the on-axis

 (drive) torque (Nm), p is the wind dynamic pressure (N/m2 ),
and A = π D 4 is the frontal area of the dish (m2 ). Defined in this way the
2

dimensionless torques are independent of the antenna size and wind speed, and
their dependence on yaw angle (wind direction with respect to the antenna azimuth
position) and on the antenna elevation (or pitch) angle. They have been obtained
from the wind tunnel tests, see [1], and are shown in Figs. 5.2, 5.3, and 5.4, solid
line.
When the torque coefficient, τw , is known (e.g., determined from wind-tunnel
testing), the actual torque for the prototype antenna can be obtained from
equation (5.1)

Tw = p ADτw (5.2)

which takes into account the size of the antenna (D, A), and the wind speed in form
of the dynamic wind pressure (p).

(a) (b)

d wind D

βe

Fl
βa

wind

Fig. 5.1 Antenna configuration with respect to wind: (a) side view and (b) top view

@Seismicisolation
@Seismicisolation
5.1 Steady-State Wind Disturbance Model 53

0.30

0.25

dimensionless yaw torque

0.20

0.15

0.10

0.05

0.00

–0.05

–0.10
0 20 40 60 80 100 120 140 160 180
yaw angle, deg

Fig. 5.2 Field and wind tunnel yaw torque coefficients, elevation angle 10 deg

0.10

0.08
dimensionless yaw torque

0.06

0.04

0.02

0.00

–0.02
0 20 40 60 80 100 120 140 160 180
yaw angle, deg

Fig. 5.3 Field and wind tunnel yaw torque coefficients, elevation angle 60 deg

0.20
dimensionless EL torque

0.15

0.10

0.05

0.00

–0.05
0 10 20 30 40 50 60 70 80 90
EL position, deg

Fig. 5.4 Field and wind tunnel elevation (pitch) torque coefficients

@Seismicisolation
@Seismicisolation
54 5 Wind Disturbance Models

5.1.2 Obtaining Wind Torques from Field Data

It has been recognized that wind tunnel test results need verification, because the
tests were run on small models of antennas (Reynolds number effects), and the
wind tunnel approach flow characteristics (mean wind profile, turbulence) were not
necessarily the same as those in the field wind conditions.
The wind tests were conducted on the 34-m antenna located at Goldstone, CA;
see [8] and [9]. During the test the antenna orientation was changed in azimuth at
a constant velocity of 0.05 deg/s, and the wind speed and direction were recorded
at the rate of 40 samples per second, along with the currents in the antenna motor
drives in the azimuth and elevation axes. The currents were used to calculate the
antenna torques.
The dynamic wind pressure can be expressed in terms of the reference wind
speed, U, typically the mean wind speed, at the reference location (the elevation of
the geometrical center of the antenna)

p = 0.5ρU 2 (5.3)

where ρ = 1.29 kg/m3 is the air mass density. For the 34-m antenna, AD =
30,874 m3 , which (when combined with equation (5.3)) gives equation (5.1) as
follows

Tw
τw = αt (5.4)
U2

where αt = 0.506 × 10−4 (for Tw in Nm and U in m/s).

In the field, the torque was measured while moving the antenna in azimuth direc-
tion with a constant velocity of 0.05 deg/s. Due to antenna movement the measured
torque, Tm , comprised of the torque generated by wind, Tw , and of the torque due to
the internal friction, Tf , thus,

Tw = Tm − T f (5.5)

The friction torque was measured when slewing the antenna with the same
azimuth velocity and in quiet weather, with the wind speed close to zero. Hence,
the dimensionless torque in equation (5.4) is modified to read

Tw − T f
τw = αt (5.6)
U2

The dimensionless torques depend on the porosity of the antenna dish; the
distance, d, of the dish vertex from the elevation axis; the elevation angle of the
antenna, β e ; and the yaw angle, β a . The distance, d, has the same effect on the
torques of the 34-m antenna and of the CP-6 Memorandum model. To prove it, note

@Seismicisolation
@Seismicisolation
5.1 Steady-State Wind Disturbance Model 55

that the total azimuth axial torque is a sum of the dish torque, Td , and the torque, Tl ,
generated by the lateral force, Fl (see Fig. 5.1). From this figure, one obtains

Tl = Fl d cos βe (5.7)

or dividing by pDA one obtains

fl d
τl = cos βe (5.8)
D

where βe is the elevation angle, τl is the dimensionless lateral torque, and fl is the
dimensionless lateral force,

Tl Fl
τl = , fl = (5.9)
p AD pA
The ratios: d/D = 0.14 for the CP-6 Memorandum case, and 0.11 for the 34-m
antenna; are close enough so that from equation (5.8) it follows that the dimension-
less lateral force is about the same in the CP-6 experiment and the 34-m antenna
experiment.
The porosity of the 34-m antenna dish was about 25% for the outer 25% of the
dish radius, which is similar to the CP-6 model. The torque was tested with respect
to the two remaining parameters: elevation angle, βe , and yaw angle, βa . Thus, the
dimensionless torque was a function of βe and βa , that is, τw = τw (βe , βa ). The
dimensionless yaw torques from the CP-6 Memorandum are shown in Figs. 5.2 and
5.3 (for the elevation angle 10 and 60 deg, respectively). Figure. 5.4 shows the pitch
torques.
It will be useful to express the wind tunnel data in an analytical form, to make
them easily available (CP-6 memo is not needed). Thus, the wind-tunnel torques are
approximated with Fourier series, and the analytical model is obtained. For azimuth
torques the polynomial order was 5, and for elevation torque it was 1. These approx-
imations can be written for both the azimuth and elevation cases as follows


5
α α
tw (α) = ao + ai sin(2πi ) + bi cos(2πi ) (5.10)
i=1
αo αo

where the Fourier coefficients are given in Table 5.1. The variable α is the wind yaw
angle if equation (5.10) is used as the yaw torque equation, and α is the antenna
elevation position if equation (5.10) is used as the elevation torque equation. The
wind tunnel data and the approximating functions are shown in Fig. 5.5.
In summary, the torque comparison of field measurements with the wind tunnel
data showed a good agreement for the steady-state winds. The differences between
the wind-induced steady-state dimensionless torques on the 34-m antenna and the
wind tunnel data were less than 10% of the field-measured torques.

@Seismicisolation
@Seismicisolation
56 5 Wind Disturbance Models

Table 5.1 Fourier coefficients for the dimensionless azimuth and elevation torque
Fourier coefficient AZ torque for EL = 10 deg AZ torque for EL = 60 deg EL torque
␣o 180 180 120
ao 0.06197 0.03538 0.06066
a1 –0.05198 –0.02233 –0.03338
a2 0.005744 –0.004266 0
a3 –0.01203 0.0003286 0
a4 –0.001404 –0.001649 0
a5 0.002014 –0.000888 0
b1 –0.07083 –0.009118 0.005447
b2 –0.008286 –0.005024 0
b3 0.01181 0.0007024 0
b4 –0.005739 0.0005609 0
b5 0.003556 –0.0007596 0

5.1.3 Comparing Wind Tunnel Results and the Field Data

In the following, the field and wind tunnel data are compared. The dimensionless
torques were obtained for the range of yaw angles, from 0 to 180 degrees, with the
elevation angle fixed; thus, τw = τw (βa , i), where i is the sample number. They
were also measured for a range of elevation angles, from 10 to 89 degrees, with
the yaw angle fixed, thus τw = τw (βe , i). Denote the dimensionless torque from
the CP-6 Memorandum τw6 (βa ) if it depends on yaw angle or tw6 (βe ) if it depends
on elevation angle. The new dimensionless (rescaled) torques, τwn , are obtained
through the following linear transformation

τwn (β) = a1 τw6 (β) + a2 (5.11)

where β = βa or β = βe . Define the error, ε, between the CP-6 Memo data and the
field data as

ε(a1 , a2 ) = (τw (β, i) − τw6 (β))2 (5.12)
i

The coefficients a1 and a2 are determined such that the error is minimal. The
parameter a1 is the scaling coefficient that adjusts the CP-6 Memorandum curve τw6
to best fit the field data. The parameter a2 , the shifting coefficient, shifts the field
data to compensate for undetermined friction forces.
The wind data were collected when the antenna rotated by 360 deg in azimuth
with a fixed elevation angle (at 10 or 60 deg) and the azimuth torques were measured.
The torques were measured for the average wind speed of 16.7 m/s and an elevation
angle of 10 deg. The wind direction was 251 deg with respect to the azimuth zero
position. The yaw angle is the difference between the antenna azimuth position
and the wind direction. The wind speed ranged from 13 to 19 m/s, as illustrated
in Fig. 5.6.
The plot of the azimuth torque versus yaw angle is shown in Fig. 5.7. It was
obtained by averaging the torques every 5 s, and by subtracting the friction torque,

@Seismicisolation
@Seismicisolation
5.1 Steady-State Wind Disturbance Model 57

0.15
(a)

EL = 10 deg
0.1
yaw torque coefficient

EL = 60 deg
0.05

–0.05
0 20 40 60 80 100 120 140 160 180
yaw angle, deg

0.12
(b)
EL torque coefficient

0.1
0.08
0.06
0.04
0.02
0
10 20 30 40 50 60 70 80 90
elevation angle, deg

Fig. 5.5 Polynomial approximations of yaw and elevation torque coefficients

20
mean wind speed, m/s

18

16

14

12
–150 –100 –50 0 50 100 150
yaw angle, deg

Fig. 5.6 Mean wind speed versus yaw angle

@Seismicisolation
@Seismicisolation
58 5 Wind Disturbance Models

which was measured for a non-windy day at 17.9 Nm. This plot is expected to
be anti-symmetric, and a small departure from the anti-symmetry is caused by the
varying wind speed during this experiment. The mean wind speed (a linear fit to the
wind speed data) is shown in Fig. 5.6. In order to obtain the results for the constant
wind speed, say Uo , the data were scaled by the factor (Uo /U )2 . By further rescaling
the torques according to equation (5.6), the dimensionless torque was obtained
(Fig. 5.2, dots). The dimensionless torque from the CP-6 Memorandum was best
fitted to this data (Fig. 5.2, solid line). It required a scaling factor of a1 = 1.016
and a shifting factor of a2 = 0.0066, which means that the wind tunnel data are
1.6% apart from the field data. The friction torque estimation is about 10% error,
because the shift a2 = 0.0066 corresponds to 1.8 Nm friction torque. Two other
experiments for an elevation angle of 10 deg required scaling of 1.07 and 1.11; thus,
the tunnel data for this case can be considered to be accurate to within 11% or
better.
A similar experiment was conducted for an elevation angle of 60 deg. The aver-
aged field-measured torques (dots) and the rescaled CP-6 Memorandum data (solid
line) are shown in Fig. 5.3. For this case, the scaling was a1 = 1.08 and the shifting
was a2 = 0.0008; thus, the difference between the field and the CP-6 Memorandum
data was within 8%, and the friction torque correction was about 1%.
The elevation (pitch) dimensionless torques were verified for winds blowing from
behind the antenna (at yaw angle of 180 deg). The torques were measured for eleva-
tion angles ranging from 10 to 89 deg. The friction torques were 5.2 Nm, and the
unbalanced torques were 2.0 Nm, with opposite direction to the friction torques (the
antenna dish was driven down). These torques were scaled to obtain the dimension-
less torques, and the dimensionless torques from the CP-6 Memorandum were fitted
to the field data; see Fig. 5.4. The fitting coefficients were: scaling a1 = 0.87, and
shifting a2 = –0.013, showing that the CP-6 and field measurements are within a
13% error margin. A slightly larger a2 was due to poorer friction torque estimation
(friction torques were obtained for different velocities, and the unbalance torque was
not known exactly).

50
azimuth torque, Nm

–50

–150 –100 –50 0 50 100 150

yaw angle, deg

Fig. 5.7 Azimuth torque versus yaw angle, elevation angle 10 deg

@Seismicisolation
@Seismicisolation
5.2 Wind Gusts Disturbance Models 59

5.2 Wind Gusts Disturbance Models

There are three ways to model wind gusts acting on an antenna structure:

• The first model, where the wind gust disturbances are modeled as a force acting
on the antenna dish; see Fig. 5.8, and [5–7]. The white noise of unit standard
deviation is filtered by the Davenport filter (on the spectrum of the Davenport
wind gusts, see [11]), and appropriately scaled, with scale factor kf . This model
is used when the antenna finite-element model is available, and includes forces
at the dish panels. It is mainly used during an antenna design stage.
• The second model is presented as a time-varying torque, Tw , acting at the antenna
drives; see also Fig. 5.8. The wind torque adds-up to the antenna drive torque
Tc , producing the total torque T acting at the structure. The Davenport filter is
again used to shape white noise into wind gusts, but a different scaling factor
(kt ) is now used. This model is often used in antenna analyses, giving adequate

wind 3 wind 2 wind 1

velocity torque force

input

command + + encoder
+
CONTROLLER +
DRIVE
+ _

velocity feedback

position feedback

wind 1 wind 2 wind 3

white noise white noise white noise

Davenport Davenport Davenport

filter filter filter

Δvo Δvo Δvo

kf kt kt

Tw
force Tw
Filter, F

velocity
input

Fig. 5.8 Wind gust models: (wind 1) wind forces acting on the dish, (wind 2) wind torques acting
at the drive motors, and (wind 3) wind velocity acting at the velocity input

@Seismicisolation
@Seismicisolation
60 5 Wind Disturbance Models

approximation of servo errors. It is used, for example, when antenna structure

and antenna drives are modeled as separate units.
• The third model is used when the antenna and drives are inseparable. It is the case
of the antenna velocity loop model obtained from system identification, where an
integrated model of the antenna structure and the drives is obtained. In this model,
as shown in Fig. 5.8, the velocity command is the input to the model, the encoder
reading is its output, and the drive torque is an internal variable. In this situation,
the only possible entry point for the wind gusts disturbances is the velocity input.
This model is obtained by adding a filter F; see Fig. 5.8. The wind torques are now
filtered to obtain an equivalent velocity input signal. The filter transfer function F
is such that the action of white noise (see the model in Fig. 5.8) causes the same
servo error as the action of white noise in the model shown in the same figure.

This chapter presents the development of the Davenport filter, the determination
of force (kf ) and torque (kt ) gains, and the determination of the filter transfer function
(F) for the 34-m antenna. It also compares servo errors in wind gusts produced
by the three wind models and measured at the antenna sites. The Davenport wind
gusts model is used; however, other models (e.g., Kaimal, Antanoui) may be more
appropriate, depending on the terrain conditions.

5.2.1 Model of Wind Forces Acting on the Dish

In this model, as shown in Fig. 5.8, the wind gust is represented as a uniformly
distributed force acting on the antenna dish, either from its front, its back, or its
side. The gust force is obtained from the gust velocity Δv o of the unit standard
deviation, and the velocity Δv o , on the other hand, is obtained from the Davenport
spectrum. Consider first the determination of wind velocity Δv o .
In this step, wind velocity from the Davenport spectrum is derived. The wind
velocity v is a combination of a steady-state, or mean velocity v m , and a turbulence
(gust) Δv, that is,
v = v m + Δv (5.13)

The gust component is a random process with zero mean and with a spectrum,
called the Davenport spectrum. The Davenport spectrum Sv (ω) depends on average
wind speed and terrain roughness, and is given by the following equation:

βω
Sv (ω) = 4800v m κ 4
(5.14)
(1 + β 2 ω2 ) /3

where β = πv 600
m
, and κ is the surface drag coefficient, obtained from the roughness
of the terrain; see [11]:

1
κ= (5.15)
(2.5 ln(z/z o ))2

@Seismicisolation
@Seismicisolation
5.2 Wind Gusts Disturbance Models 61

In these equations, z is the distance from the ground to the antenna dish center,
and zo is the height of the terrain roughness (e.g., z o = 0.1 to 0.3 m at Goldstone,
CA, where the DSN antennas are located).
The wind gust velocity Δv o of the unit standard deviation is obtained by applying
a white noise input of unit standard deviation to a filter that approximates the Daven-
port spectrum. The unit standard deviation is obtained by appropriate scaling of
filter gain. This filter is further called a Davenport filter. The filter transfer function
is of fourth order and was obtained in [6] and [7] by adjusting the filter parameters
such that the magnitude of the filter transfer function best fits the Davenport spec-
trum within the antenna bandwidth of [0.001, 20] Hz. The resulting filter transfer
function is

3.9021s 3 + 230.1426s 2 − 686.3151s + 3.4197

H= (5.16)
0.331s 4 + 38.2997s 3 + 224.7118s 2 + 22.7788s + 0.3538

The corresponding digital filter for sampling time of 0.02 s is as follows

0.1584z3 − 0.3765z2 + 0.2716z − 0.0534

Hd = (5.17)
z4 − 2.9951z3 + 3.0893z2 − 1.1930z + 0.0988

This filter is scaled such that applying white noise of unit standard deviation one
obtains an output Δv o of unit standard deviation as well.
The plot of the square root of the Davenport spectrum and of the magnitude of
the filter transfer function is shown in Fig. 5.9, and a sample of the wind speed
generated by the filter is shown in Fig. 5.10.
In the next step wind force from wind velocity is obtained. In order to do this,
consider first a steady wind for which the quadratic law relates its velocity (v n ) and
force (Fn )

Fn = k F v n2 (5.18)

The constant k F depends on the scaling of the structural model. In our case modes
were scaled such that 44.7 m/s wind corresponds to a force of 4.4482 N. Thus, from
the above equation, for this case one obtains

k F = 0.0022 N/(m/s)2 (5.19)

with velocity in m/s, and force in N. Equation (5.18) represents a steady (or static)
wind force.

@Seismicisolation
@Seismicisolation
62 5 Wind Disturbance Models

102
square root of Davenport spectrum and filter t.f. Davenport spectrum
filter

101

100

10–1
10–4 10–3 10–2 10–1 100 101
frequency, Hz

Fig. 5.9 The Davenport spectrum of the wind velocity and magnitude of the filter transfer function

The next step is to obtain time-varying forces generated by wind gusts using the
velocity time history Δv o (t). Consider a long enough time interval (e.g., of 200 s
or more). The wind speed over this interval can be decomposed into its constant

2
Davenport filter output

–1

–2

–3

–4
0 10 20 30 40 50 60 70 80 90 100
time, s

Fig. 5.10 Generic wind speed of unit standard deviation

@Seismicisolation
@Seismicisolation
5.2 Wind Gusts Disturbance Models 63

(or mean) component v m and variable component Δv of zero mean value, see equa-
tion (5.13). The corresponding wind forces are similarly decomposed:

F = Fm + Fw (5.20)

Variable Fm represent the steady-state (static force) component. The wind gust
variations are 10%–20% of the static force. The wind gust force variations (Fw )
are related to wind velocity variations
!  " (Δv) by expanding! in  Taylor
" series equa-
tion (5.18), obtaining Fw = ⭸F ⭸v v=vm Δv. Because ⭸F ⭸v v=vm = 2v m ,,
therefore

Fw = 2k F v m Δv (5.21)

The wind gust Δv obtained in the previous section is scaled to obtain its unit
standard deviation, that is, the speed Δv o (t) is such that

Δv(t)
Δv o (t) = (5.22)
σv

where σv is the standard deviation of Δv. However, the standard deviation of the
wind gust is proportional to the mean wind speed; see [8]:

σv = αv n (5.23)

Combining (5.22) and (5.23) one obtains

Δv = αv m Δv o (5.24)

In the above equations

√
α= 6κ (5.25)

and κ is the surface drag coefficient, defined in equation (5.15).

Introducing (5.24) to (5.21) one obtains the final relationship between wing gust
velocity and force

Fw = k f Δv o (5.26)

where

k f = 2k F αv n2 (5.27)

For the 34-m k F = 0.00223 Ns2 /m2 , and α = 0.20, thus

k f = 0.000892v m2

@Seismicisolation
@Seismicisolation
64 5 Wind Disturbance Models

with velocity in m/s, and force in N. For 8.94 m/s wind the force gain is k f =
0.0713 Ns/m.

5.2.2 Model of Wind Torque Acting at the Drives

In this model the wind torque disturbance is added to the drive torque. It is a time
function, Tw (t), determined from the velocity gusts Δv o . This model is shown in
Fig. 5.8, where the white noise is applied to the Davenport filter. The filter output
is the velocity gust Δv o of unit standard deviation, which is consequently scaled
to obtain wind torque that is added to the antenna drive torque. In this model the
Davenport filter is identical with the filter presented in previous section. The scaling
factor, kt , from the velocity to torque is obtained from the wind quadratic law for
torques (Tn ) and for steady wind speed v m :

Tn = k T v m2 (5.28)

The constant k T in this equation is particular for an antenna, antenna elevation

position, antenna site, terrain profile, and wind direction. It is determined as follows.
First, the wind torque depends on wind pressure, and the pressure–torque relation-
ship was determined experimentally in wind tunnels [1–3] and in field tests [8]

Tn = ct ADpn (5.29)

where D is the antenna dish diameter (m), and A is the antenna dish frontal area, A =
π D2 /4 (m2 ), and ct is a dimensionless torque coefficient. This coefficient depends
on the wind direction and the antenna elevation position, and varies from –0.05 to
0.25.
Next note that the dynamic pressure of wind, p, depends of wind velocity, simi-
larly to the torque, see [3] and [8]:

pn = α p v m2 (5.30)

where α p is the static air density, α p = 0.6126 Ns2 /m4 (pressure is in N/m2 , and
velocity is in m/s). Introducing (5.30) to (5.29) one obtains

Tn = ct α p ADv m2 (5.31)

Comparing (5.30) and (5.28) the quadratic law coefficient k T is obtained

π D3
k T = ct α p AD = ct α p (5.32)
4
For the 34-m antenna D = 34 m, and for ct = 0.25 (front and side wind) one
obtains k T = 4661 Ns2 /m.

@Seismicisolation
@Seismicisolation
5.2 Wind Gusts Disturbance Models 65

The torque equation (5.28) is valid for steady wind only. However, the wind gust
torque can be derived from it, using a linear expansion (again, the gust part of the
wind is between 10% and 20% of the steady wind that justifies the linearization).
Indeed, it is determined from the velocity gusts Δv o by linearizing equation (5.28),
which gives

ΔT = 2k T v m Δv (5.33)

The above represents the antenna axis torque. The wind torque at the pinion axis,
Tw , is the axis torque ΔT divided by the axis-to-pinion ratio Np ; thus the pinion
torque is also proportional to the velocity
ΔT 2k T v n
Tw = = Δv (5.34)
Np Np
In simulations the wind speed model with unit standard deviation Δv o (t) is used.
It was previously derived that Δv o (t) is related to an arbitrarily scaled wind speed
as in equation (5.24).
Introducing (5.24) to (5.34) one obtains

Tw = kt Δv o (5.35)

where
2k T α 2
kt = v (5.36)
Np m
For the 34-m antenna k T = 4661 Ns2 /m, α = 0.20 and Np = 42, thus

kt = 44.4v m2 (5.37)

where velocity is in m/s, and torque in Nm. For the elevation axis one obtains

kt = 294.7v m2 (5.38)

5.2.3 Algorithm to Generate a Time Profile of Wind Gusts Torque

1. Apply white noise to the filter H, with the following transfer function

0.1584z3 − 0.3765z2 + 0.2716z − 0.0534

Hd =
z4 − 2.9951z3 + 3.0893z2 − 1.1930z + 0.0988

and obtain time series of wind Δv(t). The above filter represents the Davenport
wind gusts model.
2. Calculate ␴v the standard deviation of Δv(t).

@Seismicisolation
@Seismicisolation
66 5 Wind Disturbance Models

3. Divide Δv(t) by ␴v , obtaining wind speed with unit standard deviation

Δv o (t) = Δv(t)/σv

The standard deviation of Δv o (t) is 1.

4. Scale Δv o (t) by kt to obtain Tw (t)

Tw (t) = kt Δv o (t)

where

2kα 2
kt = v
Np m

and vn is the mean wind speed, Np is the gear ratio between antenna √ main
axis and the drive pinion, k is the “wind quadratic law,” α = 6κ, and
κ = (2.5 ln(z/z o ))−2 . In the latter equation, z is the distance from the ground
to the antenna dish center, and zo is the height of the terrain roughness.

5.2.4 Model of Wind at the Velocity Input

In this model the wind gusts is applied at the velocity input. It is shown in Fig. 5.8,
where the white noise is applied to the Davenport filter; its output is the velocity gust
Δv o of unit standard deviation, appropriately scaled with gain kt to obtain wind
torque. Next, the torque is filtered with filter F that produces a velocity, which is
added to the velocity input of the antenna. In this model the Davenport filter and the
scaling factor, kt , are identical with the filter and factor presented in the previous
section. The task of this section is to find the filter transfer function F such that the
encoder response to the wind disturbances is almost the same as the response for the
wind torque acting at the drives.
In order to have the antenna response the same as for the wind torque acting at
the drives the transfer function of the filter should be the inverse of the drive transfer
function. The drive transfer function Fd (s), from the velocity input to the torque
output is shown in Fig. 5.11, solid line. Within the antenna bandwidth it can be
approximated with the integrator, that is, with Fd (s) = ksd , where kd is the drive
gain. The magnitude of the approximate transfer function is shown in Fig. 5.11,
dashed line. Thus, the filter transfer function, as an inverse of the approximate drive
transfer function, is a derivative with the inverse gain

s
F(s) = (5.39)
kd

@Seismicisolation
@Seismicisolation
5.2 Wind Gusts Disturbance Models 67

El drive
109
approx. EL drive

gain = 1.1*108
108
torque, Nm

f = 1/2π
107

106 –2
10 10–1 100 101
frequency, Hz

Fig. 5.11 Transfer function of the EL drive (solid line), and its approximate (dashed line)

The corresponding discrete-time filter of sampling time Δt is as follows

1 z−1
F(z) = (5.40)
kd z Δt
For the 34-m antenna the drive gain is kd = 1.1 × 108 [Nms/deg].

5.2.5 Algorithm to Generate Time Profile of Wind

at the Velocity Input

1. Follow the steps of the algorithm in Section 5.2.3 to obtain Tw (t).

2. Filter Tw (t) to obtain uw (t). Use filter F with the transfer function as follows
1 z−1
F(z) =
kd z Δt

where kd is the drive gain from input u to torque T, and Δt is the sampling time.

5.2.6 The Equivalence of Wind Torque and Wind Velocity Models

The nature of the first model, which consists of wind forces acting on the dish, is
different than the second and third models, which consists of torque and velocity
signals acting on the antenna drives. Thus, the first and the second model and the
first and the third model can be only compared in statistical terms, such as standard
deviation of the resulting antenna servo error. However the second and the third

@Seismicisolation
@Seismicisolation
68 5 Wind Disturbance Models

(a)
3
2
EL encoder, mdeg

1
0

–1
–2
–3
0 10 20 30 40
time, s
(b)
3
EL encoder, mdeg

2
1

0
–1

–2

–3
0 1 2 3 4 5 6 7 8 9 10
time, s

Fig. 5.12 Open loop EL encoder response to the wind: torque input (solid line) and velocity input
(dashed line): (a) 40 s response, and (b) 10 s segment

models are closely related and they can be compared directly. Thus, the antenna
positions excited by the wind gusts generated by the second model (wind torque),
and by the third model (wind input velocity) are compared. In order to do this,
one applies the white noise as marked in Fig. 5.8, Wind 2 and Wind 3, and with the
position loop open. The resulting elevation encoder reading is shown in Fig. 5.12a,b.
The plots show that the encoder outputs for the two wind cases are very close,
indicating close equivalence of the second and the third models.

5.2.7 Closed Loop Pointing Accuracy with Wind Gusts

Disturbances

The wind gusts using all three models were simulated. Then the antenna servo
error in azimuth and elevation for the front and side wind directions were deter-
mined, and compared with the field data. The Simulink model was developed for
the 34-m antennas, and allows for simulation three wind models; shown in Fig.
5.13a. The elevation drive model is shown in Fig. 5.13b. A sample of simulation
results is shown in Fig. 5.14. It represents the azimuth and elevation position errors
obtained from model of wind velocity acting at the velocity input for 32 km/h wind

@Seismicisolation
@Seismicisolation
5.2 Wind Gusts Disturbance Models 69

(a)
EL wind 3 EL wind 2 side wind

5 3 6

EL encoder
1
EL pinion rate EL

4 2
EL point
servo error vel. comm.
1
y(n) = Cx(n) + Du(n) m
EL comm EL drive x(n+1) = Ax(n) + Bu(n)
EL controller
3
Structure
8 XEL point
servo error vel. comm.
AZ comm

AZ controller 8

AZ drive AZ pinion rate

4
AZ encoder AZ

4 2 7
AZ wind3 AZ wind2 front wind

(b)

kcur 1/N

KTsz
1 k1*ks KTsz KTsz
kf*ki
kr 1/Jm KTsz
rate z-1 1/Ra km
z-1 z-1 kg/N 1
rate to V z-1
Torque
MOTOR
AMPLIFIER GEARBOX

kb

ktach N 2
Pinion Rate

Fig. 5.13 (a) The Simulink model of the 34-m antenna with three wind model inputs: front wind1
and side wind1 for model 1, AZ wind2 and EL wind2 for model2, and AZ wind3 and EL wind3
for model3; (b) EL drive Simulink model composed of amplifier, motor and gearbox models

1.5

1
elevation servo error, mdeg

0.5

–0.5

–1

–1.5

–2
–2 –1 0 1 2
azimuth servo error, mdeg

Fig. 5.14 34-m antenna azimuth and elevation position in 32-km/h wind gusts blowing from its side

@Seismicisolation
@Seismicisolation
70 5 Wind Disturbance Models

Table 5.2 Standard deviations of 34-m antenna servo errors due to 32 km/h wind gusts
Front wind Side wind
AZ (mdeg) EL (mdeg) AZ (mdeg) EL (mdeg)
Field data1 0.23–0.48 0.6–1.2 0.23–0.48 0.9–1.5
Model 1 0.33 1.73 0.30 1.02
Model 2 0.32 1.20 0.32 1.20
Model 3 0.44 1.21 0.44 1.07
1
From [4]

gusts blowing from the side of 34-m antenna. The simulation results are shown in
Table 5.2, where they are compared to the measured servo errors from [4]. The servo
errors were measured repeatedly, and the table shows the minimal and maximal
values of the standard deviation of the errors. The table shows that the simulation
results are within the range of measurements; except for the elevation error in front
wind of model 1, which exceeds the measurements.

References
1. Blaylock RB. (1964). Aerodynamic Coefficients for Model of a Paraboloidal Reflector Direc-
tional Antenna Proposed for a JPL Advanced Antenna System. JPL Memorandum CP-6
(internal document), Jet Propulsion Laboratory, Pasadena, CA.
2. Fox NL. (1962). Load Distributions on the Surface of Paraboloidal Reflector Antennas. JPL
Memorandum CP-4 (internal document), Jet Propulsion Laboratory, Pasadena, CA.
3. Fox NL, Layman, Jr. B. (1962). Preliminary Report on Paraboloidal Reflector Antenna
Wind Tunnel Tests. JPL Memorandum CP-3 (internal document), Jet Propulsion Laboratory,
Pasadena, CA.
4. Gawronski W. (1995). Wind Gust Models Derived from Field Data. TDA Progress Report
42-123, pp. 30–36. Available at http://ipnpr.jpl.nasa.gov/progress˙report/42-123/123 G.pdf.
5. Gawronski W. (2004). Modeling Wind Gusts Disturbances for the Analysis of Antenna
Pointing Accuracy. IEEE Antennas and Propagation Magazine, 46(1): 50–58.
6. Gawronski W, Bienkiewicz B, Hill RE. (1994). Wind-Induced Dynamics of a Deep Space
Network Antenna. Journal of Sound and Vibration, 178(1).
7. Gawronski W, Mellstrom JA. (1994). Control and Dynamics of the Deep Space Network
Antennas. In: Control and Dynamics Systems, ed. C.T. Leondes, vol. 63, Academic Press, San
Diego, pp. 289–412.
8. Gawronski W, Mellstrom JA. (1994). Field Verification of the Wind Tunnel Coeffi-
cients. TDA Progress Report 42-119, pp. 210–220. Available at http://ipnpr.jpl.nasa.gov/
progress˙report/42-119/119 G.pdf.
9. Gawronski W, Mellstrom JA, Bienkiewicz B. (2005). Antenna Mean Wind Torques: A
Comparison of Field and Wind Tunnel Data. IEEE Antennas and Propagation Magazine,
47(5).
10. Rae, Jr., WH, Pope A. (1984). Low-Speed Wind Tunnel Testing. John Wiley, New York.
11. Simiu E, Scanlan RH. (1996). Wind Effects on Structures. John Wiley, New York.

@Seismicisolation
@Seismicisolation
 Part II
Control

@Seismicisolation
@Seismicisolation
Chapter 6
Preliminaries to Control

Before starting the controller tuning process, one has to know the criteria that the
closed loop system should satisfy. Also, one has to prepare the plant model (in
our case the velocity loop model) such that it is the convenient configuration for
the controller tuning. This chapter defines the performance criteria used in the
controller tuning process. Also, before starting the tuning process, the plant model
(antenna velocity loop model) is transformed and augmented to make the tuning
more straightforward, and analysis physically interpretable.

6.1 Performance Criteria

The controller performance is evaluated using the following criteria:

• Settling time of a small step response, shown in Fig. 6.1. Settling time is defined
as the time at which the antenna encoder output remains within ±3% threshold of
the nominal value of the step command. For example, a 5.1 s settling time due to
a 0.02 deg step command is illustrated in Fig. 6.1. The settling time indicates how
fast antenna reacts to the command. One does not use the settling time obtained
at large position offsets (larger than 0.1 deg), because at these offsets antenna
reaches acceleration and velocity limits.
• Overshoot of a small step response is illustrated in Fig. 6.1. Overshoot (in
percent) is the relative difference between the maximal encoder output and the
commanded step with respect to the value of the commanded step. In Fig. 6.1 the
overshoot is 20%. It is measured at small position offsets (e.g., 0.02 deg).
• Overshoot of a large step response is measured in absolute values (deg) rather
than in percentage.
• Steady-state error in velocity offsets; see Fig. 6.2. This represents a lagging of
the antenna when commanded at constant velocity.
• Bandwidth of the closed loop transfer function, illustrated in Fig. 6.3. Bandwidth
is the frequency at which the magnitude drops 3 dB, or to 70.7% below zero dB
level. This is illustrated in Fig. 6.3, where the bandwidth is 0.37 Hz. The wider is
the bandwidth, the faster and more precise is the antenna.

W. Gawronski, Modeling and Control of Antennas and Telescopes, 73

@Seismicisolation
@Seismicisolation
DOI: 10.1007/978-0-387-78793-0 6, 
C Springer Science+Business Media, LLC 2008
74 6 Preliminaries to Control

0.025
X: 5.1
Y: 0.02055 overshoot
0.02
settling time
antenna pos ition, deg

+/– 3% threshold
0.015

0.01

0.005

0
0 1 2 3 4 5 6 7 8 9 10
time, s

Fig. 6.1 Settling time (5.1 s) and overshoot (20%) in a step response

12
antenna c ommand and pos ition, deg

10
servo error

2 command
antenna position
0
0 2 4 6 8 10 12 14 16 18
time, s

Fig. 6.2 Steady-state (mean) servo error in velocity offsets

• Amplitude and settling time of a disturbance step; see Fig. 6.4. Wind disturbance
in a form of rapid (stepwise) action is suppressed by the controller counteraction.
The time required to suppress it, and amplitude of the antenna movement, are
measures of the controller performance (the smaller is the amplitude and reaction
time, the better is the controller).
• Magnitude of the disturbance transfer function; see Fig. 6.5. The lower is
the magnitude, the better are the wind disturbance rejection properties of the
controller.

@Seismicisolation
@Seismicisolation
6.1 Performance Criteria 75

100

X: 0.37
Y: 0.7
magnitude

10–1
bandwidth

10–2 –3
10 10–2 10–1 100 101
frequency, Hz

Fig. 6.3 Determining bandwidth (0.3 Hz) of the closed loop system from the magnitude of its
transfer function

0.8

0.7

0.6
antenna pos ition, deg

0.5

0.4 dmax

0.3

0.2

0.1

0
0 1 2 3 4 5 6 7 8 9 10
time, s

Fig. 6.4 Max servo error due to step disturbance

• Root-mean-square servo error in 10 m/s wind gusts; see Fig. 6.6. In this figure
the wind gust simulations are used to evaluate the servo performance in wind.
• Phase and gain stability margins; see Fig. 6.7. Note that gain crossover is the
frequency at which the open loop magnitude first reaches the value of unity, and
phase crossover is the frequency at which the open loop phase angle first reaches
the value of –180 deg. Thus, gain margin is the factor by which the open loop
magnitude must be multiplied to destabilize the system, and phase margin is the

@Seismicisolation
@Seismicisolation
76 6 Preliminaries to Control

100
magnitude

10–1

10–2
10–3 10–2 10–1 100 101
frequency, Hz

Fig. 6.5 Disturbance transfer function

2
elev at ion s erv o error, mdeg

–1

–2

–3
–3 –2 –1 0 1 2 3
azimuth servo error, mdeg

Fig. 6.6 AZ and EL servo errors in wind gusts

number of degrees of delay to destabilize the system. The gain and phase margins
are measures of stability robustness. They show how much the system gain can
be changed to destabilize the closed loop system, and how much signal delay the
closed loop system can tolerate.

@Seismicisolation
@Seismicisolation
6.2 Transformations of the Velocity Loop Model 77

101

gain = 1
100
gain
margin
magnitude

10–1
gain
crossover
10–2

10–3
–2 –1 0 1
10 10 10 10
frequency, Hz

phase
–100 crossover
phase
margin
phas e, deg

phase = –180 deg

–200

–300

–400
10–2 10–1 100 101
frequency, Hz

Fig. 6.7 Gain and phase margins

6.2 Transformations of the Velocity Loop Model

The selection of the suitable coordinates for the antenna velocity loop model is
crucial in the tracking controller tuning process. The state-space model (as obtained
from the system identification) is transformed into coordinates that are convenient
for the tuning of the antenna controller.
The proposed velocity loop model transformation consists of three stages:

1. transformation into modal coordinates (makes the structural modes indepen-

dent);
2. transformation to coordinates with the antenna position as the first state (allows
for addressing the tracking requirements);
3. augmentation with the integral of the position (eliminates lagging in constant
velocity tracking).

@Seismicisolation
@Seismicisolation
78 6 Preliminaries to Control

6.2.1 Transformation into Modal Coordinates

In the first step, the velocity loop model (A, B, C) is transformed into the model
in modal coordinates (Am , Bm , Cm ), with the modal state vector xm ; for the details
see Chapter 2. The modal state-space representation is characterized by the block-
diagonal state matrix Am , and by the weakly coupled modal states, xmT =
# 1m , x2m , ·$· ·], where the ith component xim is the state of the ith mode, xim =
T T T
[x
xi1m xi2m . The weak coupling allows adjusting gains for each modal state inde-
pendently, which vastly simplifies the controller tuning process.

6.2.2 Antenna Position as the First State

Antenna position is a variable that defines antenna performance; therefore it shall

be directly controlled. It could be achieved if the position is a part of the antenna
state-space representation. Thus, the modal model is transformed further, to obtain
the new state x p , where the first state is the antenna position, while the remaining
states are unchanged. The new state, after transformation, is as follows
 
y
xp = (6.1)
xf

where y is the antenna angular position, and x f are the remaining (unchanged) states.
The new state-space representation (A p , B p , C p ) is obtained using the following
transformation

x p = P xm (6.2)

where, for n states, one has


Cm
P= , (6.3)
Po

where Po = 0n−1,1 In−1 , and 0n−1,1 is zero vector (n − 1) × 1, and In−1 is

the identity matrix of order n − 1. The transformation was derived by noting that
y = Cm xm , and partitioning Cm as follows: Cm = [Cm1 Cm2 ], where Cm1 is the
first component of Cm , and Cm2 are the remaining components.

6.2.3 Augmentation with the Integral of the Position

The model is also augmented with the integral of the position, as recommended
in [1, 2, 3], to eliminate the steady-state errors in constant-velocity tracking and
to improve the antenna disturbance rejection properties. Thus the integral of the

@Seismicisolation
@Seismicisolation
References 79

position (denoted yi ) is added to the state-space model as a new state, obtaining the
state xo
⎧ ⎫
  ⎨ yi ⎬
yi
xo = = y (6.4)
xp ⎩ ⎭
xf

The new state yi satisfies the following equation ẏi = y, that is,

ẏi = C p x p (6.5)

which combined with the antenna state space representation ( A p , B p , C p )

ẋ p = A p x p + B p u
y = Cpxp (6.6)

gives the new representation ( Ao , Bo , Co )

   
0 Cp 0

Ao = ; Bo = ; Co = 0 C p (6.7)
0 Ap Bp

The state xo as in (6.4) is the desired state vector of the antenna velocity loop
system, and (6.7) is the preferred velocity loop model of the antenna. The velocity
loop state consists of the integral of the position, the position, and the flexible defor-
mations in modal coordinates.

References
1. Athans M. (1971). On the Design of PID Controllers Using Optimal Linear Regulator Theory.
Automatica, vol. 7: 643–647.
2. Johnson CD. (1968). Optimal Control of the Linear Regulator with Constant Disturbances.
IEEE Trans. Automatic Control, vol. 13: 416–421.
3. Porter B. (1971). Optimal Control of Multivariable Linear Systems Incorporating Linear Feed-
back. Electronic Lett., vol. 7: 170–174.

@Seismicisolation
@Seismicisolation
Chapter 7
PI and Feedforward Controllers

Tuning process of the antenna controller is discussed in two steps:

1. analysis of a simple (i.e., rigid) antenna and simple (PI) controller (to derive
basic properties of the closed loop system),
2. extension of the properties of a simple system to the real (flexible) antenna and
a complex (LQG or H∞ ) controller.
The first step is discussed in this chapter, the second step in the next two chapters.
The combination of PI controller and rigid antenna allows for the closed-form
analysis. The analysis includes the relationship between the proportional and inte-
gral gains and the closed loop properties of the antenna. This leads to the tuning
method for the PI controller and to the closed loop equations of a flexible antenna
with the PI controller. Next, the performance of the PI controller and its limits are
discussed. Finally, the feedforward controller is analyzed.

7.1 Properties of the PI Controller

The block diagram of the antenna control system is shown in Fig. 7.1a. For a rigid
antenna the position proportional-and integral (PI) controller is selected, as shown
in Fig. 7.1b.
The velocity loop of a rigid antenna is a pure integrator, while the flexible antenna
velocity loop is designed so that it is approximately an integrator, as illustrated in
Fig. 7.2. The magnitude of the transfer function of a perfect integrator is shown
in Fig. 7.2 as a straight dashed line sloping at –20 dB/dec. The magnitude of the
transfer function of the 34-m flexible antenna obtained from the field test is shown
in Fig. 7.2, solid line, as a straight line sloping at –20 dB/dec for low frequencies
(up to 1 Hz), and showing flexible deformations (resonances) at higher frequen-
cies. (Structural and drive flexibility that cause deformations are not a part of the
integration.)
The closed loop system is shown in Fig. 7.1.a, where K denotes the controller
transfer function, and G is the antenna transfer function. A rigid antenna as a pure

W. Gawronski, Modeling and Control of Antennas and Telescopes, 81

@Seismicisolation
@Seismicisolation
DOI: 10.1007/978-0-387-78793-0 7, 
C Springer Science+Business Media, LLC 2008
82 7 PI and Feedforward Controllers

(a)
w

r antenna y
controller uo (G)
(K)

(b)

PI controller

kp
+
r + e uo
_
y ei +
∫ ki

Fig. 7.1 Antenna position loop (a) and its PI controller (b): k p , proportional gain; ki , integral
gain; r, command; y, antenna position; e. servo error; ei , integral of the servo error; uo . velocity
command

102

101
flexible antenna

100
magnitude

10–1
rigid antenna

10–2

10–3
10–2 10–1 100 101
frequency, Hz

Fig. 7.2 Magnitudes of the transfer function of the velocity loop model of the 34-m antenna (solid
line), and a rigid antenna (dashed line)

@Seismicisolation
@Seismicisolation
7.1 Properties of the PI Controller 83

integrator and the PI position controller have the following transfer functions

1 ki
G= and K = k p + (7.1)
s s

where k p is the proportional gain and ki is the integral gain of the PI controller.
Also in Fig. 7.1, r is a position command, uo is the controller output (called velocity
command), w is wind disturbance, y is the antenna angular position, and e = r – y is
a servo error.

7.1.1 Closed Loop Transfer Functions

Besides the variables listed in Fig. 7.1, also the commanded acceleration a (a deriva-
tive of the controller input uo ) is considered, in order to explain the impact of the
command r and disturbances w on the antenna position y and the commanded accel-
eration a. Thus, the following transfer functions are analyzed:
• Try – from the command to the angular position encoder;
• Twy – from the disturbance to the encoder;
• Tra – from the command to the acceleration; and
• Twa – from the disturbance to the acceleration.

From the block diagram in Fig. 7.1 one obtains

GK G
Tr y = and Twy = (7.2)
1 + GK 1 + GK
sK −s K G
Tra = and Twa = (7.3)
1 + KG 1 + KG

and introducing (7.1) – (7.3) gives

k p s + ki s
Tr y = and Twy = 2 (7.4)
s 2 + k p s + ki s + k p s + ki

(k p s + ki )s 2 −(k p s + ki )s
Tra = and Twa = 2 (7.5)
s + k p s + ki
2 s + k p s + ki

7.1.2 The Proportional Gain Analysis

The controller tuning starts with the selection of the proportional gain, thus assuming
ki = 0 in the above transfer functions, one obtains
1
1 kp
Tr y = 1 and Twy = (7.6)
kp
s +1 1
kp
s +1

@Seismicisolation
@Seismicisolation
84 7 PI and Feedforward Controllers

s2 −s
Tra = and Twa = (7.7)
1
kp
s+1 1
kp
s +1

The magnitudes of the above transfer functions, for different values of the propor-
tional gain, are shown in Fig. 7.3. One can see from this figure that the increase of
the proportional gain:
1. increases the bandwidth of the transfer function Tr y (from the command to the
antenna position); see Fig. 7.3a,
2. improves the disturbance rejection properties of the antenna by lowering the
magnitude of the disturbance rejection transfer function Twy ; see Fig. 7.3b,
3. increases the impact of the command on the antenna acceleration (increases
the magnitude and the bandwidth of the acceleration transfer function Tra ); see
Fig. 7.3c,
4. increases the impact of disturbances on the antenna acceleration (increases the
magnitude and the bandwidth of the acceleration transfer function Twa ); see
Fig. 7.3d.

1 1
10 10
(a) (b)
0
10
magni tude, Twy
magnitude, Try

0
10
–1
10
–1
10
kp 10
–2
kp
–2 –3
10 –2 0 2
10 –2 0 2
10 10 10 10 10 10
frequency, Hz frequency, Hz
4 2
10 10
(c) kp (d)
kp
magni tude, Twa
magni tude, Tra

2 1
10 10

0 0
10 10

–2 –1
10 –2 0 2
10 –2 0 2
10 10 10 10 10 10
frequency, Hz frequency, Hz

Fig. 7.3 Magnitude of transfer functions of the proportional controller for kp = 1 (solid line), for
kp = 4 (dashed line), and for kp = 16 (dotted line): (a) Try , from the command to the encoder,
(b) Twy , from the disturbance to the encoder, (c) Tra , from the command to the acceleration, and
(d) Twa , from the disturbance to the acceleration

@Seismicisolation
@Seismicisolation
7.1 Properties of the PI Controller 85

The first two transfer functions show the improvement of the antenna perfor-
mance with the increase of the proportional gain. However, the last two functions
show a potential problem: antenna acceleration increases at high frequencies, both
due to command and due to disturbances. The increased acceleration indicates that
the antenna can hit the acceleration limit and enter a nonlinear regime; consequently
its performance will deteriorate, leading even to instability. Thus, the proportional
gain increase is limited by the acceleration limits imposed at the antenna drives.

7.1.3 The Integral Gain Analysis

Before the impact of the integral gain on antenna dynamics is analyzed, two critical
values are introduced: the critical integral gain and the critical frequency.
Critical integral gain. It is well known that large integral gain causes low-
frequency oscillations of the closed loop system. Consider the poles s1 and s2 of the
closed loop system (7.4) and (7.5), that is, the roots of the polynomial s 2 + k p s + ki :
  
s1,2 = 0.5 −k p ± k 2p − 4ki (7.8)

√
The system is non-oscillatory if poles are real, which happens when k p > 2 ki ,
or when

ki ≤ 0.25k 2p (7.9)

Thus
kic = 0.25k 2p (7.10)

is the upper limit of the integral gain, called the critical integral gain.
Critical frequency. For the critical integral gain the denominator of the transfer
functions shown in (7.4) and (7.5) is as follows: s 2 + k p s + 0.25k 2p = (s + 0.5k p )2 .
At frequency

ωo = 0.5k p = ki (7.11)

the slope of the transfer function Tr y drops by –40 dB/dec. This is the critical
frequency of the closed loop system that determines the antenna bandwidth. In the
following, the frequencies significantly smaller than ωo are called low frequencies;
frequencies significantly larger than ωo are called high frequencies; and frequencies
in the neighborhood of ωo are called medium frequencies.
By considering low, medium, and high frequencies in (7.4) and (7.5) one can
show how the transfer functions depend on the integral gain. Note first that for
medium frequencies the variations of all four transfer functions are minimal (see
Fig. 7.4) because the integral gain is smaller than the critical integral gain. For low
and high frequencies the transfer functions behave as follows:

@Seismicisolation
@Seismicisolation
86 7 PI and Feedforward Controllers

1 0
10 10
(a) (b)

magnitude, Twy
magnitude, Try

0
10 10–1

–1
kp
10 10–2 kp

–2
10 –2 0 2
10–3 –2 0 2
10 10 10 10 10 10
frequency, Hz frequency, Hz
2
10
10
4 (c) (d)

magnitude, Twa
magnitude, Tra

1
10
10
2 kp

0
10
0 10 kp

–2 –1
10 –2 0 2
10 –2 0 2
10 10 10 10 10 10
frequency, Hz frequency, Hz

Fig. 7.4 Magnitude of transfer functions of the PI controller for kp = 16, and ki = 1 (solid line), for
kp = 16, and ki = 4 (dashed line), for kp = 16, and ki = 16 (dotted line), and for kp = 16,: (a) Try ,
(b) Twy , (c) Tra , and (d) Twa

1. The transfer function Tr y does not depend on ki , because for low frequencies
Tr y ∼
= 1 and for high frequencies Tr y ∼= k p /s, as shown in Fig. 7.4a.
2. The transfer function Twy is inversely proportional to ki for low frequencies,
because Twy ∼ = s/ki ; for high frequencies it does not depend on ki because
Twy ∼= 1/s, as shown in Fig. 7.4b.
3. The transfer function Tra does not depend on ki , because for low frequencies
Tra ∼= s 2 , for high frequencies Tra ∼
= k p s, as shown in Fig. 7.4c.
4. The transfer function Twa does not depend on ki , because for low frequencies
Twa ∼= −s; for high frequencies Twa ∼ = −k p as shown in Fig. 7.4d.
The above analysis showed that the integral gain impacts the disturbance rejec-
tion transfer function Twy only, and that the impact is at low frequencies.

7.2 PI Controller Tuning Steps

The PI controller tuning involves:
1. Tuning the proportional gain. Increase the gain until the antenna reaches the
acceleration limits at typical commands and at expected disturbances, and set it
at 75% of this value.

@Seismicisolation
@Seismicisolation
7.4 Performance of the PI Controller 87

2. Tuning the integral gain. Increase the gain until oscillations or undershoot
appear, and set it at 75% of this value. It should be smaller than the critical
integral gain.
In summary, the proportional gain shapes the bandwidth of the transfer func-
tion Tr y . The larger is the gain, the wider is the bandwidth. For a rigid antenna the
limit for the proportional gain is set by the antenna acceleration limits, because the
increase of proportional gain increases antenna acceleration caused by commands
and disturbances; see Fig. 7.3c,d. The integral gain shapes the disturbance transfer
function at low frequencies. It is increased to obtain better disturbance rejection
properties. But there is a limit: the integral gain should not exceed the critical inte-
gral gain (i.e., should satisfy condition (7.9), to prevent antenna oscillations.

7.3 Closed Loop Equations of a Flexible Antenna

with a PI Controller

Consider a flexible antenna with the state-space representation of the velocity loop
(A, B, C). For the system as in Fig. 7.1, the closed loop equations are as follows

ẋcl = Acl xcl + Bcl r + Bw w

y = Ccl xcl (7.12)

where
   
ei 0 −C
xcl = , Acl = , (7.13)
x Bki A − Bk p C
   
1 0

Bcl = , Bw = , Ccl = 0 C (7.14)

Bk p I

7.4 Performance of the PI Controller

The performance of the closed loop system depends on the PI controller gains.
However, the performance characteristics of the controller, such as settling time,
bandwidth or the rms servo error in wind gusts, are inter-related.

7.4.1 Performance Characteristics

Our analysis includes a rigid antenna. The rigid antenna is considered to be a perfect
antenna, with no flexible deformations, for which the controller needs to overcome
inertia forces only. The closed loop transfer function, from the command to the

@Seismicisolation
@Seismicisolation
88 7 PI and Feedforward Controllers

14
ki = 0.2 kp
ki = 0.4 kp
12
ki = 0.6 kp
ki = 0.8 kp
10
setting time, s

(a)
8

2
0.5 1 1.5 2 2.5 3
kp

ki = 0.2 kp
0.6 ki = 0.4 kp
ki = 0.6 kp
0.5 ki = 0.8 kp
bandwidth, Hz

(b)
0.4

0.3

0.2

0.1
0.5 1 1.5 2 2.5 3
kp
5
ki = 0.2 kp
4.5
rms servo error in wind gusts, mdeg

ki = 0.4 kp
4 ki = 0.6 kp
ki = 0.8 kp
3.5

3
(c)
2.5

1.5

0.5
0.5 1 1.5 2 2.5 3
kp

Fig. 7.5 The performance parameters of the rigid antenna with PI controller as a function of the
controller gains: (a) settling time, (b) bandwidth, and (c) rms servo error in 20-mph wind gusts

@Seismicisolation
@Seismicisolation
7.4 Performance of the PI Controller 89

antenna position, and from the wind disturbance to the antenna position is given
by (7.12). Using these equations, the step responses, frequency responses and wind
errors simulated (see [1]) for the proportional gain from the range of k p = [0.5–3.0],
and the integral gain, from the range ki = [0.2kp – 0.9 kp ]. As a result, the settling
time, bandwidth and servo error in wind gusts (wind gusts are modeled as a random
process with zero mean and a Davenport spectrum) to obtained. The results are
plotted in Fig. 7.5a–c.
The plots show that

0.6
ki = 0.2kp
ki = 0.4kp
0.5 ki = 0.6kp
ki = 0.8kp
bandwith, Hz

0.4
(a)

0.3

0.2

3 4 5 6 7 8 9 10 11 12
setting time, s
4
ki = 0.2kp
ki = 0.4kp
rms servo error in wind gusts, mdeg

3.5
ki = 0.6kp
ki = 0.8kp
3

(b)
2.5

1.5

1
3 4 5 6 7 8 9 10 11 12
setting time, s

Fig. 7.6 The performance parameters of the rigid antenna with PI controller as a function of the
settling time: (a) bandwidth, and (b) rms servo error in 20-mph wind gusts

@Seismicisolation
@Seismicisolation
90 7 PI and Feedforward Controllers

• the settling time significantly decreases with the increase of the proportional gain,
and also decreases with the increase of the integral gain.
• The bandwidth increases linearly with the proportional and integral gains.
• The servo error in wind decreases with the increase of the proportional gain and
with the increase of the integral gain.

Figures 7.5a–c can be recombined to show how the bandwidth and the servo error
in wind gusts depend on the settling time; see Fig. 7.6a,b. The bandwidth decreases
and the servo error increases with the increase of the settling time.

7.4.2 Limits of Performance

It was shown in the above sections that the performance of the PI controller improves
with the increase of its proportional and integral gains. Unfortunately, when applied
to a flexible antenna, the increase of the gains has its limits: high proportional
gain causes instability. Indeed, assume a zero integral gain, and the proportional
gain k p = 0.6, then the response of the closed loop system to a 20-mdeg step
command is shown in Fig. 7.7, dash line. It has no overshoot and a settling time
is 6 s. However, increasing the gain to 1.6 produces 2 s settling time, but the system
becomes unstable, with undying vibrations (see Fig. 7.7). Thus, antenna compliance
is the source of the controller instability. A controller that suppresses the antenna
vibrations when the gains are increased would significantly improve antenna preci-
sion. Model-based controllers, considered in subsequent chapters, act in the above-
described manner.

25

20
antenna position, mdeg

15 kp = 1.6

kp = 0.6

10

0
0 1 2 3 4 5 6 7 8 9 10
time, s

Fig. 7.7 Step responses of the PI controller with flexible antenna: low gain controller (dashed line),
and high gain controller (solid line)

@Seismicisolation
@Seismicisolation
7.5 Feedforward Controller 91

7.5 Feedforward Controller

The feedforward loop is added to the PI controller to improve its tracking accuracy
at high velocities [2]. In the feedforward loop the command is differentiated and
forwarded to the velocity loop input, as a velocity command, see Fig. 7.8.
Consider the open loop system with feedforward and velocity loop in series; see
Fig. 7.8b. The transfer function of this system is approximately equal to 1. Indeed,
the derivative is an approximate inversion the velocity loop transfer function, so the
series of the two gives transfer function equal to 1. It is illustrated with the magni-
tude of the velocity loop transfer function G r is shown in Fig. 7.9, approximated
(up to 1 Hz) with an integrator (G ra = 1/s), shown in the same figure, dashed line.
The feedforward transfer function is a derivative (G f f = s) shown in Fig. 7.9, dash-
dotted line. Thus the overall transfer function is a series of the feedforward and the
velocity loop G o = G r G f f , which is approximately equal to 1 up to the frequency
1 Hz (dashed line). The position feedback is added to compensate disturbances and
system imperfections.
Note, however, that the disturbance transfer function is not affected by the feed-
forward loop. Thus, feedforward improves command following, but not disturbance
rejection properties.
The impact of the feedforward loop on the 34-m antenna performance is illus-
trated as follows. The step response of the 34-m antenna with a PI controller is
shown in Fig. 7.10, solid line. The settling time is 15 s. Adding the feedfordward
controller makes the antenna response to the command significantly faster (4 s
settling time), but response is more oscillatory. The latter is eliminated by using
command pre-processor, discussed in Chapter 10. Note also that the feedforward

(a)
velocity
command
d/dt

+ position + + position
velocity
command controller loop
–

(b)
velocity
command command position
d/dt velocity
loop

Fig. 7.8 Antenna control system with the feedforward controller: (a) closed loop, and (b) open
loop

@Seismicisolation
@Seismicisolation
92 7 PI and Feedforward Controllers

2
10

1
10

0
10

–1
10

velocity-loop
–2
10 velocity-loop approx.
feedforward
velocity-loop+feedforward
–3
10
–2 –1 0 1
10 10 10 10
frequency, hz

Fig. 7.9 The feedforward action is illustrated with the magnitudes of the transfer functions: the
velocity loop transfer function G r , the velocity loop transfer function approximation, G ra , the
feedforward loop transfer function, G f f , and G o = G f f G r is the transfer function of the series
connection of feedforward and velocity loop

loop increases the bandwidth of the antenna transfer function from the command
to the encoder from 0.1 Hz to 1.4 Hz; see Fig. 7.11. However, the transfer function
from the disturbance to the encoder remains unchanged.

0.03

0.025
azimuth position, deg

0.02

0.015

0.01

0.005 with feedforward

without feedforward
0
0 5 10 15 20

Fig. 7.10 Step responses of the antenna with the PI controller, and the antenna with PI and feed-
forward controller

@Seismicisolation
@Seismicisolation
References 93

1
10

0
10

–1
magnitude

10

–2
10

–3
10
with feedforward
without feedforward
–4
10
–2 –1 0 1
10 10 10 10

Fig. 7.11 Magnitudes of the transfer functions from the command to the encoder of the antenna
with PI controller, and the antenna with PI and feedforward controller

References
1. Gawronski W. (2007). Servo Performance Parameters of the NASA Deep Space Network
Antennas. IEEE Antennas and Propagation Magazine, 49(6).
2. Gawronski W, Mellstrom JA. (1994). Antenna Servo Design for Tracking Low-Earth-Orbit
Satellites. AIAA J. Guidance, Control Dynamics, 17(6):1179–1184.

@Seismicisolation
@Seismicisolation
Chapter 8
LQG Controller

This chapter presents the tuning process of the LQG controller. Its applications in
the antenna industry are discussed in many papers, but scarcely implemented. LQG
controllers are implemented at the 34- and 70-m Deep Space Network antennas.
The discussion begins with the short description of the LQG controller, and its
tracking version, gives the closed loop equations of a flexible antenna with the LQG
controller, and shows the similarity of the PI and LQG controllers. It discusses how
to select the LQG weights to obtain the required performance, present the controller
tuning steps, and show the limits of the performance. Next, it presents two graphical
user interfaces (GUI) to tune the LQG controllers. The GUIs do not require detailed
knowledge of the tuning process to obtain excellent closed loop performance. The
chapter also discusses the LQG controller in the velocity loop.

8.1 Properties of the LQG Controller

The LQG controllers are well established; see, for example, [4–6], [8], [10],
[12–17]. However, their tuning (called also design) for antenna tracking purposes
is a tricky process. The controller must address the antenna tracking requirements
(such as minimization of the antenna servo error in wind gusts, and fast response
to commands) while complying with antenna limitations (such as acceleration and
velocity limits). Thus, blindly applying the standard LQG tuning steps can lead to an
optimal controller that minimizes the LQG performance index but might not satisfy
the performance requirements, because “optimal” does not mean “satisfactory.”
One has to remember that every LQG controller is an optimal controller because
it minimizes the LQG performance index, defined later. But, the LQG index seldom
includes engineering requirements, thus the LQG controller with poorly defined
index would have non-satisfactory performance requirements.

8.1.1 LQG Controller Description

A block diagram of the antenna LQG control system is shown in Fig. 3.7. Assuming
the command zero, and no limits imposed on the velocity and acceleration, the

W. Gawronski, Modeling and Control of Antennas and Telescopes, 95

@Seismicisolation
@Seismicisolation
DOI: 10.1007/978-0-387-78793-0 8, 
C Springer Science+Business Media, LLC 2008
96 8 LQG Controller

Fig. 8.1 The LQG closed u y

loop system: G, antenna G
velocity loop; K, controller;
u, velocity input; and y,
encoder
K

control system is reduced to the one in Fig. 8.1. It consists of the antenna velocity-
loop (G) and controller (K). The antenna output y is the encoder measurement
and is supplied to the controller. Using the output y the controller determines the
control signal u that drives the antenna. The inside structure of the antenna and the
controller is shown in Fig. 8.2. The plant is described by the following state-space
equations:

ẋ = Ax + Bu + v,

y = C x + w, (8.1)

as shown in Fig. 8.2. In the above description the plant (antenna velocity loop)
state vector is denoted x. The plant is perturbed by random disturbances, denoted
v, and its output is corrupted by noise w. The noise v, called process noise, has
covariance V = E(vv T ); the noise w is called measurement noise, and its covariance
is W = E(ww T ). Both noises are assumed uncorrelated, that is, E(vw T ) = 0,

Antenna (G)

v A
w
+ +
u + x x + y
B ∫ C
–

y
u
Controller (K ) ŷ − y
C +

A Ke

u x̂ x̂ + +
Kc ∫
–

u
B

Fig. 8.2 The inner structure of the LQG closed loop system

@Seismicisolation
@Seismicisolation
8.1 Properties of the LQG Controller 97

where E(.) is the expectation operator. Without loss of generality, it is assumed that
the covariance of the measurement noise is unity, that is, W = I
The controller is driven by the plant output y. The controller produces the control
signal u that drives the plant. This signal is proportional to the plant estimated state
denoted x̂, and the gain between the state and the controlled signal u is the controller
gain (K c ). The estimated state x̂ rather than the actual state x is used, because
the latter is not available from measurements (except for the antenna position).
The estimated state is obtained from the estimator, which is part of the controller,
as shown in Fig. 8.2. The estimator equations follow from the block-diagram
in Fig. 8.2:

x̂˙ = A x̂ + Bu + K e (y − C x̂). (8.2)

Assuming that the plant model (A,B,C) is known exactly (e.g., from system iden-
tification), one can see that the estimated state is an exact copy of the actual state,
except for the initial (transient) dynamics. From the above equation it follows that
in order to determine the estimator one has to determine the estimator gain, K e .
The controller gains are obtained by minimizing the performance index J,
⎛ ⎞
∞
J 2 = E ⎝ (x T Qx + u T Ru) dt ⎠ (8.3)
0

In the above definition Q is a positive semi-definite weight matrix and R is a

positive scalar. Assume R = 1 because the case R = 1 is equivalent to R = 1 with
the scaled weight matrix Q/R. The minimum of J is obtained for the feedback

u = −K c x̂ (8.4)

with the gain K c obtained as

K c = BoT Sc , (8.5)

and Sc is the solution of the controller algebraic Riccati equation (called CARE)

A T Sc + Sc A − Sc B B T Sc + Q = 0. (8.6)

One can see that the controller gain K c depends solely on the weight matrix Q
(A and B are fixed).
The optimal estimator gain is given by

K e = Se C T , (8.7)

@Seismicisolation
@Seismicisolation
98 8 LQG Controller

where Se is the solution of the estimator (or filter) algebraic Riccati equation (called
FARE)

ASe + Se A T − Se C T C Se + V = 0. (8.8)

Using (8.2) and the block-diagram in Fig. 8.2 the controller state-space equations
from input y to output u are derived:

x̂˙ = (A − B K c − K e C)x̂ + K e y,
u = −K c x̂. (8.9)

From these equations the controller state-space representation (Alqg ,Blqg ,Clqg ),
are obtained

Alqg = A − B K c − K e C,
Blqg = K e ,
Clqg = −K c . (8.10)

The controller gain K c is obtained in Matlab as follows:

Kc = lqr(A,B,Q,R)

and in the discrete-time case

Kc = dlqr(A,B,Q,R)

Similarly, the estimator gain K e is obtained in Matlab as follows:

Ke = lqe(A,I,C,Q,R)

or, in the discrete time case,

Ke = dlqe(A,I,C,Q,R)

The matrices A, B, C, Q, and R are described earlier, and I is the identity

matrix.

8.1.2 Tracking LQG Controller

Above, the LQG controller was described for the zero (or constant) command.
However, the antenna must follow a variable command, and the controller should

@Seismicisolation
@Seismicisolation
8.1 Properties of the LQG Controller 99

Structure (G)

v A
w
+ + y
u + x x +
B ∫ C
+

u y

Controller (K )
+ – ŷ − +y
kf Cf C
+
ki kp A Ke
e +
x̂ x̂ +
e ∫
∫

+
r
Cp
+ −
u
B

Fig. 8.3 The tracking LQG controller with an integral upgrade

assure zero steady-state tracking error. It is achieved by adding an integral of the

plant position to the plant state-space representation; see Section 6.2.3. The closed
loop system configuration of the tracking LQG controller is shown in Fig. 8.3. In
this figure (A,B,C) is the plant state-space triple, x is the state, x̂ is the estimated
state, x̂ f is the estimated state of a flexible part, r is the command, u is the control
input, y is the output, ŷ is the estimated output, e = r − ŷ is the servo error, ei is the
integral of servo error, v is the process noise of intensity V, and the measurement
noise w is of intensity W. Both v and w are uncorrelated: E(vw T ) = 0, V = E(vv T ),
W = E(ww T ) = I , E(v) = 0, and E(w) = 0.
Similarly to the actual state x, see (8.4), the estimated state consists of three
components: estimated integral of the position ŷi , estimated position ŷ, and esti-
mated flexible mode state x̂ f :
⎧ ⎫
⎨ ŷi ⎬
x̂ = ŷ (8.11)
⎩ ⎭
x̂ f
In order to make a stabilizing controller into a tracking controller, the antenna
estimated position ŷ is replaced with the antenna servo error e = r − )y, and the
integral of position is replaced with the integral of the servo error ei = e dt; see
[2, pp. 603–604], thus
⎧ ⎫
⎨ êi ⎬
x̂ = ê (8.12)
⎩ ⎭
x̂ f

@Seismicisolation
@Seismicisolation
100 8 LQG Controller

(a) w

r y
uo antenna
controller (G)
( K)
y

(b)
LQG controller

PI controller

kp
+
r + e u

y _ + +
ei
∫ ki

Flexible Mode controller

Kf
xˆ f
u
Estimator
ŷ _
+ ê
Ke

Fig. 8.4 Antenna position-loop (a) and its LQG controller (b): k p , proportional gain; ki , integral
gain; K f , flexible mode gain; K e , estimator gain; r, command; y, antenna position; e, servo error, ei ,
integral of the servo error; u, velocity command; y, antenna angular position; ŷ, antenna estimated
position; ê, estimation error; and x̂ f , estimated flexible mode states

The antenna position loop with the LQG controller is shown in Fig. 8.4a (the
same structure as with the PI controller), and the LQG controller structure is shown
in Fig. 8.4b. The LQG controller structure corresponds to the above representation.
In this representation the controller gain is divided into the proportional gain k p ,
integral gain ki , and flexible mode gain K f , that is, K c = [K i K p K f ], which gives
(8.9) in the form u = −K i ei − K p e − K f x̂ f .
The missing part of the controller is the estimated state x̂ in (8.4). It is obtained
from the estimator, which is a computer model of the antenna. Its equation is as
follows:

x̂˙ = A x̂ + Bu + K e (y − C x̂). (8.13)

with the estimator gain given by (8.7).

The controller gain is divided according to the state (8.12), that is,

@Seismicisolation
@Seismicisolation
8.1 Properties of the LQG Controller 101

K c = ki kp Kf (8.14)

Introducing (8.14) and (8.12) to (8.4) one obtains

u = −ki ei − k p e − K f x̂ f (8.15)

8.1.3 Closed Loop Equations of the Tracking LQG Control System

For the system as in Fig. 8.3 the closed loop equations are as follows

ẋcl = Acl xcl + Bcl r + Bv v

y = Ccl xcl + w (8.16)

where
⎧ ⎫ ⎡ ⎤
⎨ei ⎬ 0 0 −C
xcl = x , Acl = ⎣ Bki A −Bk p C − Bk f C f ⎦ , (8.17)
⎩ ⎭
x̂ Bki ke C A − ke C − Bk p C − Bk f C f

⎡ ⎤ ⎡ ⎤
1 0

Bcl = ⎣ Bk p ⎦ , Bv = ⎣ I ⎦ , Ccl = 0 C 0 (8.18)

Bk p 0

8.1.4 LQG Weights

For the open loop state-space representation (A,B,C ) of a flexible structure the state
vector x is divided into the tracking, xt , and flexible, x f , parts, that is,
 
xt
x= . (8.19)
xf

The tracking part includes the position error (e), and its integral (ei ),
 
ei
xt = (8.20)
e

@Seismicisolation
@Seismicisolation
102 8 LQG Controller

while the flexible mode part includes modes of deformation. For this division the
system triple can be presented as follows:
  

At At f Bt

A= , B= , C = Ct 0 . (8.21)
0 Af Bf

The gain, K c , the weight, Q, and solution of CARE, Sc , are divided similarly
to x,

K c = K ct Kc f ,

 
Qt 0
Q= ,
0 Qf

 
Sct Sct f
Sc = T . (8.22)
Sct f Sc f

The tuning process is carried out such that the tracking system is of low authority,
that is, such
* that* the flexible weights are much smaller than the tracking ones:
Q t   * Q f *. It was shown in [1] that for Q f = 0 one obtains Sc f = 0 and
Sct f = 0. This means that the gain of the tracking part, K ct , does not depend on
the flexible part. And, for the low-authority tracking system (with small Q f ), one
obtains weak dependence of the tracking gains on the flexible weights, due to the
continuity of the solution. Similar conclusions apply to the FARE equation (8.8).
This property can be validated by observation of the closed loop transfer func-
tions for different weights. Consider the transfer function of the Deep Space Network
antenna, as in Fig. 8.5. Denote by In and 0n the identity and zero matrices of order
n, then the magnitude of the closed loop transfer function (azimuth angle to azimuth
command) for Q t = I2 and Q f = 010 is shown as a dashed line, for Q t = 8× I2 and
Q f = 010 as a dot–dashed line, and for Q t = 8I2 and Q f = 5 × I10 as a solid line
in Fig. 8.5. It follows from the plots that variations in Q f changed the properties of
the flexible subsystem only, while variations in Q t changed the properties of both
subsystems.
Note, however, that large Q f increases dependency of the gains on the flexible
system; only quasi-independence in the final stage of controller tuning is observed,
while separation in the initial stages of controller tuning is still strong. The tuning
consists therefore of the initial choice of weights for the tracking subsystem, and
determination of the controller gains of the flexible subsystem. It is followed by the
adjustment of weights of the tracking subsystem, and a final tuning of the flexible
weights, if necessary.
In conclusion, in modal coordinates the LQG weight matrix is selected as a diag-
onal matrix (due to independence of states in modal coordinates), that is,

Q = diag(qi , q p , q f ) (8.23)

@Seismicisolation
@Seismicisolation
8.1 Properties of the LQG Controller 103

100

10–1
magnitude

10–2
Qt = I2, Qf = 010
Qt = 8*I2, Qf = 010
Qt = 8*I2, Qf = 5*I10
10–3
10–3 10–2 10–1 100 101
frequency, Hz

Fig. 8.5 Magnitudes of the transfer function of a closed loop system for different LQG weights:
Q f impacts the flexible modes (higher frequencies), while Q t impacts the low and high frequencies

where qi is the integral weight, q p is the proportional weight, and q f is a vector

of flexible mode weights. It is convenient to present the LQG weights in the vector
form as the LQG weight vector q

⎧ ⎫
⎨ qi ⎬
q = qp (8.24)
⎩ ⎭
qf

The weight vector q corresponds to the state vector xo , and the flexible mode
weights are divided into weights of each mode

⎧ ⎫
⎨q f 1m ⎪
⎪ ⎬
q f = q f 2m (8.25)
⎩ .. ⎪
⎪ ⎭
.

# $
with each mode weight q Tfim = q f i q̇ f i corresponding to the two-state mode.
For the antenna model in modal coordinates the modal states are weakly coupled.
They are also almost independent from the antenna position and integral of the posi-
tion. Thus, the corresponding weights act independently on each flexible mode, and
almost independently on the position and on the integral of the position states. This
adds to the flexibility to the controller tuning.

@Seismicisolation
@Seismicisolation
104 8 LQG Controller

8.1.5 Resemblance of the LQG and PI Controllers

It was shown in Chapter 7 that for a rigid antenna the increase of the proportional
gain improves antenna bandwidth and the disturbance rejection properties. However,
an increase of proportional gain, when applied to a flexible antenna, is drastically
limited: even a moderate gain can excite structural vibrations and cause instability.
However, because the LQG controller consists of the PI part and the flexible mode
part, see Fig. 8.4, the flexible mode controller is able to suppress antenna vibra-
tions. In this way, the increased proportional gain does not cause instability: a flex-
ible antenna behaves approximately as a rigid one. Therefore the controller tuning
approach used for a rigid antenna with a PI controller can be extended (with certain
limitations) for tuning the LQG controller of a flexible antenna. The limitations are
formulated as follows: the action of the flexible mode gains should not be excessive;
the gains should be large enough to assure vibration suppression. Such a controller
is a low-authority LQG controller.
The following example shows the similarity between the rigid antenna with a PI
controller and a flexible antenna with a LQG controller. Consider the 34-m antenna
velocity loop model with a transfer function from the velocity input to the position
output shown in Fig. 7.2, solid line. At lower frequencies the transfer function is
identical to the transfer function of a rigid antenna, and at higher frequencies it
shows flexible mode resonances.
The LQG controller to this antenna is applied as follows. First, the weights
of three LQG controllers are selected, with their integral gain zero, and propor-
tional gains 1, 4, and 16, respectively. For these cases the plots of magnitudes
of the transfer functions Try , Twy , Tra , and Twa are shown in Fig. 8.6. Comparing
Fig. 8.6 and Fig. 7.3 one can see similarities between the rigid antenna with a PI
controller and a flexible antenna with a LQG controller. Namely, the plots of Try
show the expanding bandwidth with the increase of the proportional gain. The plots
of Twy show the decreasing antenna response to disturbances with the increase of the
proportional gain. The plots of Tra and Twa show increased acceleration response at
high frequencies.
Next, the weights of the LQG controller are selected to obtain a fixed propor-
tional gain, k p = 16 and the integral gains 1, 4, and 16, respectively. Note from
(8.16) that the critical integral gain is 64 in this case. The plots of magnitudes of
the transfer functions Try , Twy , Tra , and Twa for the above three cases are shown in
Fig. 8.7. Comparing Fig. 8.7 and Fig. 7.4 one can observe similarities between the
rigid antenna with a PI controller and a flexible antenna with a LQG controller. The
integral gain impacts significantly the disturbance rejection properties (Twy ) only,
and there is no significant impact on the closed loop bandwidth (see Try plot) or on
the system acceleration; see the plots of Tra and Twa .
Finally, the impact of flexible mode weights on antenna dynamics is analyzed.
Figure 8.8 presents the magnitudes of the transfer functions Try , Twy , Tra , and
Twa for fixed proportional and integral gains (k p = 9.5 and ki = 6.3) and
for small flexible mode weights (solid lines) and for large flexible mode weights
(dashed lines). The plots show that the excessive flexible mode weights reduce the

@Seismicisolation
@Seismicisolation
8.1 Properties of the LQG Controller 105

101
kp
0
10

magnitude, Twy
magnitude, Try

100

kp
10–1
10–2

10–2
10–2 100 102 10–2 100 102
frequency, Hz frequency, Hz

104 102

magnitude, Twa
magnitude, Tra

102 kp 101

100 100
kp
–2 –1
10 10
10–2 100 102 10–2 100 102
frequency, Hz frequency, Hz

Fig. 8.6 Magnitude of transfer functions of the LQG controller for ki = 0: kp = 1 (dot-dash line),
kp = 4 (dotted line), and kp = 16 (solid line): (a) Try , (b) Twy , (c) Tra , and (d) Twa .

closed-loop bandwidth (Fig. 8.8a), and deteriorate the disturbance rejection proper-
ties (Fig. 8.8b).

8.1.6 Properties of the LQG Weights

The above comparison indicates that the LQG weights have similar impact on a
flexible antenna performance as PI gains on a rigid antenna performance, assuming
the proper choice of coordinate system. The following list summarizes the properties
of the LQG weights:

• The increase of the flexible mode weights causes antenna vibration suppression.
A single mode weight impacts only states corresponding to this particular mode
(the flexible mode coordinates are weakly coupled).
• The increase of the proportional weight increases the closed loop bandwidth and
improves the disturbance rejection properties.

@Seismicisolation
@Seismicisolation
106 8 LQG Controller

101 100

magnitude, Twy
magnitude, Try

100 10–1
ki

10–1 10–2

10–2 10–3
10–2 100 102 10–2 100 102
frequency, Hz frequency, Hz

105 102

magnitude, Twa
magnitude, Tra

101

100
100

10–1
10–2 100 102 10–2 100 102
frequency, Hz frequency, Hz

Fig. 8.7 Magnitude of transfer functions of the LQG controller for kp = 16: ki = 1 (dash-dot line),
ki = 4 (dot line), and ki = 16 (solid line): (a) Try , (b) Twy , (c) Tra , and (d) Twa

• The increase of the integral weight improves the disturbance rejection properties,
but does not impact the bandwidth.

The position and integral of the position weights are coupled, but they are easily
manageable because the coupling involves two variables only.
Finally, Figs. 8.9 and 8.10 show the dependency of the proportional and inte-
gral gains on the proportional and integral weights of the 34-m antenna. First, one
can see that the increase of the proportional weight increases both proportional and
integral gains. Similarly, the increase of the integral weight increases simultaneously
proportional and integral gains. Next, one can see that the weight-gain dependency
relies on the ratio of the integral-to-proportional gain (except for the relationship
between the proportional weight and integral gain). The plots may serve as a guide
in selection of the proportional and integral weights.

@Seismicisolation
@Seismicisolation
8.1 Properties of the LQG Controller 107

(a) (b)
101

100

magnitude, Twy
magnitude, Try

100

10–1
10–2

10–2
10–2 100 102 10–2 100 102
frequency, Hz frequency, Hz

(c) (d)
104 magnitude, Twa 102
magnitude, Tra

102 101

100
100

10–2 –2 10–1
10 100 102 10–2 100 102
frequency, Hz frequency, Hz

Fig. 8.8 Magnitude of transfer functions of the LQG controller for kp = 9.5 and ki = 6.3: flexible
weights small (solid line), and flexible weights large; overdamped modes (dotted line): (a)Try ,
(b) Twy , (c) Tra , and (d) Twa

8.1.7 Limits of the LQG Gains

The increases of the proportional, integral, and flexible mode gains have their limits,
namely:

• Increase of flexible mode weights should not be excessive. Large weights lead to
the overdamped dynamics and deteriorate overall antenna performance (reduced
bandwidth, depreciated disturbance rejection properties).
• The position weight is restricted by the antenna acceleration limit: Large posi-
tion weight causes excessive acceleration of the antenna that hits the acceleration
limit, and leads to non-linear dynamics and deterioration of the performance.
• The integral weight should not exceed the critical integral gain in order to prevent
low-frequency oscillations (the frequency of these oscillations is lower than the
antenna fundamental frequency).

@Seismicisolation
@Seismicisolation
108 8 LQG Controller

2.5

2
kp

1.5

ki = 0.2kp
1 ki = 0.4kp
ki = 0.6kp
ki = 0.8kp
0.5
0 1 2 3 4 5 6
proportional weight

2.5

2
kp

1.5

ki = 0.2kp
1 ki = 0.4kp
ki = 0.6kp
ki = 0.8kp
0.5
0 1 2 3 4 5 6 7 8
integral weight

Fig. 8.9 The dependency of the proportional gain on the proportional and integral weights

8.2 LQG Controller Tuning Steps

Based on the above analysis, the LQG controller tuning involves:

1. Tuning the flexible mode weights. Apply small weights of the integral of the posi-
tion and position (which result in small PI gains), and also apply small flexible

@Seismicisolation
@Seismicisolation
8.2 LQG Controller Tuning Steps 109

2.5

1.5
ki

ki = 0.2 kp
0.5 ki = 0.4 kp
ki = 0.6 kp
ki = 0.8 kp
0
0 1 2 3 4 5 6 7
proportional weight
2.5
ki = 0.2 kp
ki = 0.4 kp
2 ki = 0.6 kp
ki = 0.8 kp

1.5
ki

0.5

0
0 1 2 3 4 5 6 7
integral weight

Fig. 8.10 The dependency of the integral gain on the proportional and integral weights

mode weights. Check the closed loop transfer function for the appearance of flex-
ible mode resonances. If they are excessive, increase the corresponding flexible
mode weights. Do not apply unnecessarily large weights, because overdamped
modes impact the antenna tracking performance.
2. Tuning the proportional weight. Increase the position weight: the proportional
gain increases accordingly. The increase of the position gain causes the expansion

@Seismicisolation
@Seismicisolation
110 8 LQG Controller

of the closed loop bandwidth. Increase the weight until bandwidth reaches the
antenna fundamental frequency.
3. Tuning the integral weight. Increase the weight the position integral, causing the
increase of the integral gain. The weight should increase until oscillations appear.
The integral gain should satisfy the condition (8.15).
4. Correction of the flexible mode weights. Check the flexile mode dynamics. If
resonances resurface after tuning the proportional and integral parts, increase the
corresponding flexible mode weights.

8.3 Performance of the LQG Controller

The primary measure of antenna control system performance is the antenna servo
error while tracking in wind gusts. However, the error is not easily measured, and
antenna performance is frequently poorly estimated. The difficulty of evaluating the
servo error in wind gusts led to the search for alternative (or indirect) measures of
the antenna pointing performance in the wind.
The control systems of the 34-m and 70-m Deep Space Network antennas [7] is
analyzed. It was observed that the servo performance parameters of the antennas,
such as settling time and bandwidth, are related to the servo error in wind gusts.
Qualitatively, the shorter the settling time, or the wider the bandwidth, the smaller is
the servo error in wind gusts. In this section, the qualitative knowledge is trans-
formed into a quantitative one. The relationships between controller gains and
performance parameters have been established, as well as between the parameters
themselves. They allow one to estimate the antenna servo error in wind (which is not
easy to measure) using the settling time (which is simple to measure). Because the
results of the analysis are similar for the 34-me and 70-m antennas, it is presumed
that they are valid for many other antennas and radio-telescopes.

8.3.1 Summary of the Antenna Servo Performance Characteristics

In this section, the antenna servo performance is characterized by the following

parameters:
• Settling time (s)
• Bandwidth (Hz)
• Root-means-square (rms) servo error in wind gusts (mdeg)

The ultimate measure of antenna servo performance is the rms servo error
(called further servo error, for simplicity) in wind gusts during spacecraft tracking.
It consists of the constant (mean) component and the variable component. The
constant component is required to be zero. The variable component is characterized
by its standard deviation (or rms error).

@Seismicisolation
@Seismicisolation
8.3 Performance of the LQG Controller 111

The servo error measurements are straightforward, but require long logging time
to obtain reliable statistical estimates. They depend on the wind velocity, wind
direction with respect to the antenna position (yaw angle), and antenna elevation
position. The important factor, not to be ignored, is the availability of wind gusts of
significant power during measurements. For this reason, measurements of the servo
error in wind gusts require time, patience, and significant data processing, which
explain why they are rather scarce.
Occasional servo performance measurements and analysis results indicated a
connection between the servo error and other antenna parameters, such as settling
time or bandwidth, and also a connection between the servo error and antenna
servo proportional and integral gains. The importance of these facts lays in the fact
that measurement of the settling time or bandwidth is much simpler and less time
consuming than measurement of the error in wind gusts, while the servo gains do
not need to be measured—they are given.

8.3.2 Performance of the DSN Antennas with LQG Controllers

In this section only the elevation axis controller of the 34-m DSN antennas is
analyzed. For the azimuth axis the relationships are similar; see [7]. The perfor-
mance parameters have been determined as follows: For a selected proportional and
integral gain, the LQG controller was tuned and the performance parameters for this
controller ware simulated (using Matlab). The process was repeated for another set
of proportional and integral gain values.
Figure 8.11a–d show the dependency of the settling time, overshoot, bandwidth,
and servo error in wind gusts on the proportional and integral gains of the LQG
controller. Figure 8.12a,b show the dependency of the bandwidth and the servo error
in wind gusts on the antenna settling time.
In Figure 8.13a–c, the performance parameters of a rigid antenna, a 34-m
antenna, and a 70-m antenna is compared. The comparison is presented for the
integral gain being half of the proportional gain, ki = 0.5 k p , which is typical for
the DSN antennas. The figures show that for the same gains

• The settling time is smallest for the rigid antennas. However, for other antennas,
it is similar; the difference does not exceed 0.5 s.
• The 34-m and 70-m antenna bandwidths slightly exceed the rigid antenna band-
width.
• The servo error in wind gusts is definitely smallest for the rigid antenna.
• All the DSN antennas have very similar servo error in wind gusts.

The unexpected result of the lowest bandwidth of the rigid antenna (compared
to 34-m and 70-m antennas) is due to the integral action of the controllers. For the
rigid antenna, the magnitude of the transfer function is less “inflated” by the integral
gain, giving slightly lower bandwidth.

@Seismicisolation
@Seismicisolation
112 8 LQG Controller

14
ki = 0.2 kp
ki = 0.4 kp
12 ki = 0.6 kp
ki = 0.8 kp
settling time, s 10

8
(a)

2
0.5 1 1.5 2 2.5 3
kp

ki = 0.2 kp
ki = 0.4 kp
0.6 ki = 0.6 kp
ki = 0.8 kp
0.5
bandwidth, Hz

(b)
0.4

0.3

0.2

0.1
0.5 1 1.5 2 2.5 3
kp

5
rms servo error in wind gusts, mdeg

ki = 0.2 kp
4.5 ki = 0.4 kp
ki = 0.6 kp
4
ki = 0.8 kp
3.5
(c)
3

2.5

1.5

1
0.5 1 1.5 2 2.5 3
kp

Fig. 8.11 The performance parameters of the 34-m antenna (elevation axis) ith LQG controller
as a function of the controller gains: (a) settling time, (b) bandwidth, and (c) rms servo error in
20 mph wind gusts

@Seismicisolation
@Seismicisolation
8.3 Performance of the LQG Controller 113

ki = 0.2 kp
0.55 ki = 0.4 kp
ki = 0.6 kp
0.5 ki = 0.8 kp

0.45
bandwidth, Hz

0.4
(a)
0.35

0.3

0.25

0.2

0.15
3 4 5 6 7 8 9 10 11 12
settling time, s

4.5
ki = 0.2 kp
4 ki = 0.4 kp
rms servo error in wind gusts, mdeg

ki = 0.6 kp
3.5 ki = 0.8 kp

3
(b)
2.5

1.5

1
3 4 5 6 7 8 9 10 11 12
settling time, s

Fig. 8.12 The performance parameters of the 34-m antenna (elevation axis) with LQG controller
as a function of the settling time: (a) bandwidth, and (b) rms servo error in 20-mph wind gusts

8.3.3 Disturbance Rejection Properties and the Position-Loop

Bandwidth

It has been observed that the expansion of the bandwidth of a telescope improves
its disturbance rejection properties (resulting in smaller servo errors in wind distur-
bances). It can be explained using Fig. 8.4a. The position loop transfer function

@Seismicisolation
@Seismicisolation
114 8 LQG Controller

(a) 8
70 m antenna
7.5 34 m antenna
rigid antenna
7

settilng time, s
6.5

5.5

4.5

4
0.5 1 1.5 2 2.5 3
proportional gain,kp

(b) 0.6
70 m antenna
0.55 34 m antenna
0.5 rigid antenna

0.45
bandwidth, Hz

0.4
0.35
0.3
0.25
0.2
0.15
0.1
0.5 1 1.5 2 2.5 3
proportional gain,kp

(c) 2.5
70-meter antenna
rms servo error in wind gusts, mdeg

34-meter antenna
rigid antenna

1.5

1
1 1.5 2 2.5 3
proportional gain,kp

Fig. 8.13 The comparison of elevation axis performance of the rigid antenna, 34-m antenna, and
70-m antenna: (a) settling time, (b) bandwidth, and (c) rms servo error in 20-mph wind gusts

@Seismicisolation
@Seismicisolation
8.3 Performance of the LQG Controller 115

Tr y (from r to y) and the disturbance transfer function Twy (from w to y) are as

follows:
GK G
Tr y = , Twy = ,
1 + GK 1 + GK

where K is the controller transfer function and G is the velocity loop transfer func-
tion. It follows from the above equations that
Twy
Tr y + = 1,
G

which, consequently, gives

Twy = G(1 − Tr y ).

Note that within the bandwidth, by definition, Tr y ∼

= 1, which implies – from the
above equation – that Twy ∼ = 0, i.e., the disturbance error is small within the band-
width. Thus, if the bandwidth covers the wind disturbance spectrum, the telescope
servo error due to wind is small. Note, however, that G is a part the above equation,
and that for telescopes G is large for low frequencies and small for high frequencies.
Thus, Tr y ∼
= 1 should be tight at low frequencies, and not necessarily tight at high
frequencies.
The bandwidth of Tr y is limited by the fundamental frequency of the telescope.
Thus, if it reaches the fundamental frequency, there is little room left for the
improvement of the disturbance rejection properties.
In conclusion, the following rules of thumb can be introduced:
1. The bulk improvement of the disturbance rejection properties is due to the band-
width expansion.
2. The fundamental frequency of the telescope limits its disturbance rejection prop-
erties.

8.3.4 Performance Comparison of the PI and LQG Controllers

The servo errors were measured at the 34-m antenna at the Madrid Deep Space
Communication Complex, during 15–16 km/h wind; see [9]. The measurements
were taken with the PI coefficients and with the LQG coefficients, one after another,
so that similar wind conditions were observed. The wind data were obtained one
sample per minute, while the servo data were obtained 10 samples per second.
The mean wind velocity during the PI test was 15.0 km/h, and during the LQG
test was 16.2 km/h. The measured servo errors were scaled to wind velocity 16 km/h.
! "2
The scaling factor for the PI measurements was 16 = 1.15, and for the LQG
! 16.0 "2 15
measurements it was 16.2 = 0.99.
The scaled measured servo errors are plotted in Fig. 8.14. The plots show that the
elevation servo errors are significantly larger than the azimuth servo errors, which is

@Seismicisolation
@Seismicisolation
116 8 LQG Controller

20

15 PI controller

10
EL servo error, mdeg

–5

–10

–15

–20
–20 –15 –10 –5 0 5 10 15 20
AZ servo error, mdeg

20

15 LQG controller

10
EL servo error, mdeg

–5

–10

–15

–20
–20 –15 –10 –5 0 5 10 15 20
AZ servo error, mdeg

Fig. 8.14 The servo errors in 16.1 km/h (10 mph) wind, for the DSS55 antenna with the PI coeffi-
cients (upper plot), and with the LQG coefficients (lower plot)

@Seismicisolation
@Seismicisolation
8.4 Tuning a LQG Controller Using GUI 117

Table 8.1 Servo errors in 16.1 km/h wind for 34-m antenna with PI and LQG controllers
AZ rms error EL rms error Mean radial
Controller mdeg mdeg error mdeg
PI 0.32 2.69 2.71
LQG 0.066 0.42 0.42
Error ratio (PI/LQG) 4.9 6.5 6.5

typical for the 34-m antennas. They also show that the servo errors for the antenna
with LQG coefficients are significantly smaller than the errors with the PI coeffi-
cients (4.9 times smaller in azimuth and 6.5 times smaller in elevation); see Table
8.1. The mean radial error (rms of azimuth and elevation errors) dropped by factor
6.5.

8.3.5 Limits of Performance

It is not difficult to obtain a high-gain LQG controller with outstanding wind distur-
bance rejection properties. However, simulations and measurements results indicate
that such a controller can be implemented only if the drives are powerful enough.
The power of a drive is expressed in the antenna acceleration limits. Because the
bulk of the drive power is used to accelerate the antenna, the acceleration limit
thus expresses the drive capability to react to the antenna dynamics. The limit (on
the input signal u that drives the antenna) is applied to prevent overheating of the
motors. For the 34-m antenna, the acceleration limits are ±0.4 deg/s2 .
But the acceleration limit effectively cancels the benefits of a high-gain controller.
Even small motion may cause the limit violation. Consider the high-gain LQG
controller in response to a 20-mdeg step offset. It accelerates up to 10 deg/s2 .
Because the acceleration is limited to 0.4 deg/s2 (Fig. 8.15b), the antenna control
system is in a nonlinear regime that causes the oscillations, see Fig. 8.15a.

8.4 Tuning a LQG Controller Using GUI

A user-friendly graphical user interface (GUI) was developed to simplify the tuning
process and create an environment in which one with limited control engineering
background may “play” to obtain a tracking algorithm that ensures the desired
performance [15]. In this section two Matlab based–software packages for the
controller tuning—basic LQG tuning and fine-tuning of the algorithm—are
described.

8.4.1 Selecting LQG Weights

The process of determining the LQG gains is explained in detail in Section 8.1.
Here, the approach is explained in a heuristic way. The controller gains depend

@Seismicisolation
@Seismicisolation
118 8 LQG Controller

40

azimuth position, mdeg

30

20

10

0
0 1 2 3 4 5
time, s

0.5
azimuth acceleration, deg/s2

accel. limit 0.4 deg/s2

accel. limit 0.4 deg/s2

–0.5
0 1 2 3 4 5
time, s

Fig. 8.15 Step response and acceleration of a high-gain LQG controller

on the weight matrix Q and covariance matrix V. The weight matrix shapes the
optimization index, a positive variable to be minimized. The covariance matrix
specifies the noise in the system, and it impacts the estimator gains. The diffi-
culty arises when one tries to express the performance of the antenna (such as
the rms servo error in wind, closed loop bandwidth, etc.) through the values of
Q and V.
A closed-form relationship does not exist between those matrices and the required
behavior of an antenna. Thus, an immediate solution is a “trial-and-error” approach.
However, there are too many parameters in Q and V to make this approach effec-
tive. Moreover, Q and V depend on the choice of the state-space coordinates, with
some coordinates more useful than others. Physical coordinates, such as structural
displacements, motor torques or currents, although easy to interpret, are highly
coupled; they therefore create undesirable difficulties in the tuning process. The
modal coordinates simplify the tuning because they are weakly coupled (i.e., modi-
fications of one of them weakly impact the remaining ones). In consequence, Q and
V matrices are diagonal (therefore there is a smaller number of parameters to control
the closed loop dynamics). Additionally, Q = V is assumed. This simplification eases
the search for the “best” controller and is rather opportunistic, with the goal to
simplify the GUI approach. Experience shows, however, that the results obtained

@Seismicisolation
@Seismicisolation
8.4 Tuning a LQG Controller Using GUI 119

using this assumption are satisfactory or even exceed the expectations. The reader
shall note that in individual cases, the tuning of Q and V separately may improve the
performance.
Taking into account the above considerations, the matrices Q and V are equal
and diagonal, that is, Q = V = diag(q), where q is a weighting vector. Note that
components of q correspond to the components of the state vector x. Thus, for the
state vector as in (8.24), the weighting vector has the form similar to x
# $
q = qei qe qf1 qf1 ··· qfn qfn (8.26)

where qei is the weight of the integral of the servo error, qe is the weight of the servo
error, q f 1 is the weight of the first natural mode, and q f n is the weight of the nth
natural mode.
Now, one has to specify 2n + 2 weights, which, for a typical number n = 4
it makes 10 weights. The critical fact is that the weights have weak dependence
amongst each other: the ith modal weight, q f i , impacts mostly the ith mode, that
is, state variables, x f i1 , and x f i2 . Also, the “tracking” weights, qe , and qei , impact
mostly the “tracking” states, e, and ei . The tuning process is illustrated in the next
section.

8.4.2 GUI for the LQG Controller Tuning

The GUI display is shown in Fig. 8.16. The GUI allows for simple manipulations of
the tuning parameters and observations of the antenna performance to make tuning
decisions.

1. On the left-hand side of the interface, there is a short description of the tool.
2. The frame below the description contains a Matlab’s data file that holds the A, B,
and C matrices.
3. The eight sliders and their three displays are the user tools to modify the
controller’s performance.
4. Each slider (except the wind-speed slider) ranges from 0 to 100 (these are dimen-
sionless numbers).
5. In the bottom right corner of the GUI screen are the simulation results and the
values of the proportional and integral gains.
6. The upper-left plot is the step response to a one-degree offset. The upper-right
plot is the magnitude of the transfer function of the closed loop system.
7. The “simulate” button is used to execute the simulation with the parameters on
the screen.
On the GUI, the antenna performance is displayed, such as settling time, band-
width, overshoot, rms error in wind gusts for the selected wind speed, steady-state
error when the antenna is moving with a constant velocity. Also, the proportional
and integral gains are displayed.

@Seismicisolation
@Seismicisolation
120 8 LQG Controller

Step response Transfer function

0
10
1
Step Response

0.8 –1

Magnitude
10
0.6

0.4 –2
10
0.2

–3
0 10 –2 –1 0 1
0 2 4 6 8 10 12 14 16 18 20 10 10 10 10
Time (s) Frequency (Hz)

PI Controls Weights
100

90

0 20 40 60 80 100
80

70

60

50

Wind Speed 40
1

0.5 30

0 20
0 10 20 30 40 50

10

0
0 0.5 1

Weights Results

Fig. 8.16 Interface of the LQG tuning

The performance of the antenna depends on all controller weights. Generally,

changing a single weight impacts more than one performance feature. However, in
the modal coordinates of the antenna model each weight influences predominantly
one or two performance features, and is weakly coupled with the remaining ones.
This helps to follow the improvement of the antenna performance, because each
slider predominantly addresses a single performance feature. Table 8.2 summarizes
these relationships.
In the 34-m antenna example the following steps correspond to the interfaces
pictured in Fig. 8.17a–b. These steps reflect antennas’ common features, and can be
followed as a guideline to the controller tuning process:

1. The tuning starts with the input of the state space representation of the open
loop antenna model (A,B,C,D), a file abc.mat at the directory c:/lqg/gui/az. In
the lower left window enter: load c:/lqg/gui/az/abc.mat.
2. Next, select arbitrary, but small, values of proportional and integral weights
(say qei = 0.1 and qe = 0.3) and zero for frequency weights see Fig. 8.17a.
After simulation the step response shows excessive settling time and overshoot
(8.52 s and 18%, respectively), visible flexible oscillations and an unaccept-
able maximal disturbance step value (0.82 deg). The magnitude of the transfer

@Seismicisolation
@Seismicisolation
8.4 Tuning a LQG Controller Using GUI 121

Table 8.2 Relationship between LQG weights and performance

Slider movement
(LQG weight) Result (performance)
Proportional weight Settling time, bandwidth, disturbance rejection
Integral weight Overshoot, velocity offset error, disturbance rejection
Frequency weight 1 Vibration amplitude of the lowest (fundamental) frequency
Frequency weight 2 Vibration amplitude of the second frequency
Frequency weight 3 Vibration amplitude of the third frequency
Frequency weight 4 Vibration amplitude of the fourth frequency

function shows low bandwidth (0.21 Hz), sharp resonance peaks, and a larger-
than-desired magnitude of the disturbance transfer function.
3. The most excessive oscillations come from the first (or fundamental) mode.
Therefore the weight of the first flexible mode is increased to 5, and other
weights are unchanged, as shown in Fig. 8.17a. The results are visible in the step
responses where the oscillations are now invisible and in the transfer function
where the first resonant peak disappeared.
4. The proportional weight is increased to 10, and the integral gain was increased to
3, c.f., Fig. 8.17b. This move reduced the settling time to 4 s, overshoot to 13%,
and expanded bandwidth to 0.8 Hz.

8.4.3 Fine Tuning of the LQG Controller

Certain features of the controller obtained through the approach described above can
be improved by further modifying the controller weights. For example, the oscilla-
tions of the closed loop response can be further reduced or the closed loop response
can be modified such that the wind disturbance does not exceed its specification.
However, at this stage, the closed loop states are not as easily decoupled as the
open loop states; therefore the relationship between the weights and the closed loop
dynamics is less clear, and often difficult to track. For this reason a constrained
optimization approach is used to tune the already tuned controller.
In this approach the weights are tuning parameters, and the following variables
are either optimized or constrained:
1. steady state servo error in the 0.1 deg/s velocity offset,
2. max value of the servo error due to the unit step disturbance (deg), edmaz =
max(|ed |),
3. rms servo error in 24 km/h wind gusts, ew (mdeg), for 0 ≤ t ≤ 100 s,
4. overshoot of the unit step response, eo (%),
5. settling time of the unit step response, ts (s),
6. bandwidth, f o (Hz),
7. magnitude of the closed loop transfer function (from the command to the
encoder) for the high-frequency range (above 3.5 Hz), that is, m h = max m( f ),
for f ≥ 3.5Hz

@Seismicisolation
@Seismicisolation
122 8 LQG Controller

Fig. 8.17 Interfaces for the 34-m antenna example

8. magnitude of the closed loop transfer function (from the disturbance to the
encoder), i.e., m d max = max m d , for f ≥ 0Hz.
The optimization index (a positive function to be minimized) is defined as
follows:

@Seismicisolation
@Seismicisolation
8.4 Tuning a LQG Controller Using GUI 123

f = w1 er +w2 ed max +w3 ew +w4 eo +w5 ts +w6 (3− f o )+w7 m h +w8 m d max (8.27)

Each variable in the function is weighted. The weight, wi , indicates the relative
importance of each variable. If wi = 0, the ith variable is not optimized. In the
above-defined function, the bandwidth is maximized by using 3 − fo variable rather
than the bandwidth fo itself. For the 34-m antennas the bandwidth will not exceed
3 Hz; therefore the minimization of 3 − fo leads to expansion of fo . The variables
above are functions of the LQG weights; therefore, f is function of weights as well.
The initial values of the LQG weights are taken from the tuning part of GUI.
This tool is quite similar to the previously described tool, see Fig. 8.18. The two
plots of the fine-tuning GUI are identical to those of the tuning GUI. The user is
given a line on which she/he enters the file containing the discrete time state space
representation of the velocity loop model. In this GUI, there are eight variables
explained above that can be either minimized or constrained. Different variables
and their changes are displayed and can be tracked throughout.
The Matlab optimization function, constr, is used here to find a local minimum
of a constrained, nonlinear, multivariable function. Before attempting to minimize a
function, it will ensure that the constraints have been met. Two main decisions must
be made when using the constr function. One is whether to optimize or constrain a

Step Response Transfer Function

0
1.2 10

1
Step Response

–1
0.8 10
Magnitude

0.6

0.4 10–2

0.2

0 10–3 –2 –1 0 1
0 2 4 6 8 10 10 10 10 10
Time(S) Frequency (Hz)

Weights Constrain or optimize

Fig. 8.18 The fine-tuning interface

@Seismicisolation
@Seismicisolation
124 8 LQG Controller

certain variable. The other is how many iterations are necessary to run in a block.
Too many iterations may require long simulation time and give the user the feeling
of being out of control; too few requires frequent re-starting.
The following steps are recommended to tune the controller using the GUI:
1. For each variable verify the starting values by clicking on the “optimization
weights” (rather than “constrain”) radiobutton, and set the weights to zero.
2. Construct the weight vector:
⎡ ⎤
I ntegral
⎢ Pr opor tional ⎥
⎢ ⎥
⎢ Special ⎥
⎢ ⎥
⎢ Fr equency1 ⎥
⎢ ⎥
⎢ Fr equency2 ⎥
⎢ ⎥
⎣ Fr equency3 ⎦
Fr equency4

3. Using the last values from the tuning GUI, run the optimizer by pressing the
“Optimize” button.
4. Keep all of the “optimize” radiobuttons on, but scale the different variables
according to your priorities, that is, change the weights to nonzero prioritized
values. Run this optimization in increments of about 300 iterations, or until the
system is optimized.
5. After the system is optimized (“Converged Successfully” is printed in the Matlab
command window), look through the values of the variables and decide which
performance values are acceptable and which are not. Of the ones that are not,
choose their “constrain” radiobutton and select constraint values tighter than the
current values. Run the optimization until it converges successfully.
The display in Fig, 8.18 show the optimization results for the 34-m antenna.

8.5 LQG Controller in the Velocity Loop

The LQG controllers presented above were located in the position loop, while the
velocity loop was controlled by a simple proportional controller. The question arises,
What happens to the antenna performance when the LQG controller is placed in the
velocity loop? The advantage of placing the LQG controller in the velocity loop is
first explained with a rigid antenna and proportional controller, and later analyzed
with the LMT radiotelescope.

8.5.1 Position Loop Bandwidth Depends on the Velocity Loop

In this section a simple (rigid) antenna and a simple (proportional) controller is
analyzed, to obtain the impact of the velocity loop bandwidth on the position loop
bandwidth. Consider a rigid antenna, the transfer function of which is an integrator,

@Seismicisolation
@Seismicisolation
8.5 LQG Controller in the Velocity Loop 125

1
s
.When the velocity loop of is closed with a proportional controller, its transfer
function is as follows, see Chapter 2, equation (2.3):

1
G(s) = (8.28)
Ts + 1

where T is inverse proportional to the gain of the velocity loop. The time constant
T obviously defines the velocity loop bandwidth. Figure 8.19 shows the magnitudes
of the transfer functions for T = 0.5/␲ and for T = 5/␲, with bandwidth 0.1 Hz and
1 Hz, respectively.
Now, if the position loop is closed with the proportional controller of gain k, its
transfer function H is

k/T
H (s) = (8.29)
s 2 + s/T + k/T
The plot of this transfer functions and step responses for k = 5, T = 0.5/␲, for
k = 0.5, T = 5/␲ and for k = 5, T = 5/␲ is shown in Fig. 8.20a. The figures show
that the antenna with low velocity loop bandwidth (0.1 Hz, for T = 5/␲) has low
position loop bandwidth, when compared to the antenna with high velocity band-
width (1 Hz, for T = 0.5/␲). Increasing the low bandwidth by increasing the position
controller gain (from k = 0.5 to k = 5) causes antenna oscillations, and a resonance
peak appears; see Fig. 8.20b. This concludes that the bandwidth of the position loop
depends on the bandwidth of the velocity loop. Thus, wide bandwidth of the velocity
loop is essential to the position loop performance.

101
T = 5/π
T = 0.5/π

100
magnitude

10–1

10–2 –2
10 10–1 100 101
frequency, Hz

Fig. 8.19 Magnitudes of the velocity loop transfer functions, for T = 5/␲ (bandwidth 0.1 Hz) and
for T = 0.5/␲ (bandwidth 1 Hz)

@Seismicisolation
@Seismicisolation
126 8 LQG Controller

(a)
101
T = 5/π, k = 0.5
T = 5/π, k = 5
T = 0.5/π, k = 5

100
magnitude

10–1

10–2 –2
10 10–1 100 101
frequency, Hz
(b)
1.6

1.4

1.2

1
position, deg

0.8

0.6

0.4
T = 5/π, k = 0.5
T = 5/π, k = 5
0.2
T = 0.5/π, k = 5

0
0 2 4 6 8 10 12 14 16 18 20
time, s

Fig. 8.20 Position loop for T = 5/␲, k = 0.5 (bandwidth 0.1 Hz), for T = 5/␲, k = 5 (bandwidth
0.4 Hz but oscillatory response) and for T = 0.5/␲, k = 5 (bandwidth 1 Hz): (a) magnitudes of the
transfer functions and (b) step responses

@Seismicisolation
@Seismicisolation
8.5 LQG Controller in the Velocity Loop 127

8.5.2 Four Control Configurations

The LQG controller placed at different loops is analyzed. The problem is investi-
gated at the Large Millimeter Telescope (LMT); see Chapter 1. The control system
of the LMT consists of velocity- and position loops, as shown in Fig. 8.21. Four
control systems are analyzed [11]. They have the following structure:
• PP control system, with PI controller in the position loop, and PI controller in
the velocity loop,
• PL control system, where the PI controller is in the velocity loop, and LQG
controller is in the position loop.
• LP control system, where the LQG controller is in the velocity loop, and PID
controller is in the position loop
• LL control system, where the LQG controller is in the velocity loop, and LQG
controller is in the position loop.

The PP control system is a typical telescope control system configuration. The

PL case is the configuration of the Deep Space Network antenna control system,
which has been considered by MAN Technologie [3], [12]. The LP and LL control
systems have not yet been implemented.
In this section the performance analysis (in terms of bandwidth, step responses,
and wind disturbance rejection properties) of the four control systems as applied
to the LMT is analyzed. This analysis shall help to evaluate and select the control
system, not only for the LMT but also for other antennas and radiotelescopes of a
similar design.

8.5.3 PP Control System

The PP control system consists of the PI controller in the position loop, and the
PI controller in the velocity loop. Its Simulink model is shown in Fig. 8.22a, and
the velocity loop subsystem in Fig. 8.22b. The controller is shown in Fig. 8.22c,

position velocity
command command position
position velocity torque structure
controller controller and drives

velocity
position

Fig. 8.21 Four control systems of the LMT: (1) PP control system, where velocity controller is PI,
and position controller is PI, (2) PL control system, where velocity controller is PI, and position
controller is LQG, (3) LP control system, where velocity controller is LQG, and position controller
is PID, and (4) LL control system, where velocity controller is LQG, and position controller is
LQG.

@Seismicisolation
@Seismicisolation
128 8 LQG Controller

(a)
AZ position feedback

AZ controller
y
AZ command u
r rp rp AZ velocity
AZ FF AZ acc.
AZ CPP AZ vel.
z1 limit AZ vel.
limit
dt*[10](z) AZ
AZ encoder
EL vel.
EL FF
z1
dt*[10](z) wind_AZ AZ wind
EL command
EL vel. EL acc. EL encoder
r rp rp limit EL
u limit
y EL velocity wind_EL EL wind
EL CPP
EL controller
velocity loop
EL position feedback

(b)
AZ drive velocity
AZ vel.
controller

AZ vel.
AZ
1 1

2
EL vel. 2
EL
AZ wind EL vel.
controller
3
Antenna structure
4 FEM model
EL wind
EL drive velocity

(c)
Integrator ki
K Ts
K*u
z–1
kp

K*u 1
+ kf u
–
rp
1 K*u

2 +–
y

y(n) = Cx(n)+Du(n)
u
x(n+1) = Ax(n)+Bu(n)
Estimator

Fig. 8.22 The Simulink model of (a) the position loop, and (b) the velocity loop system, and
(c) the controller (for kf = 0 it is a PI controller, for k f = 0 it is an LQG controller)

@Seismicisolation
@Seismicisolation
8.5 LQG Controller in the Velocity Loop 129

assuming that the flexible mode gain is zero, kf = 0. Here, the position loop
PI controller is complemented with the feedforward (FF) loop to improve the
tracking properties, especially at high velocities, and with a command preprocessor
(CPP) to avoid large overshoots during target acquisition and limit cycling during
slewing.
The velocity loop model is shown in Fig. 8.22b. It consists of the finite-element
model of the telescope structure, which includes the drives and azimuth and eleva-
tion velocity loop controllers. It is a discrete-time (digital) control system, with
0.001-s sampling time. The proportional and integral gains of the azimuth controller
are 300, and for the elevation controller proportional gain is also 300, and the inte-
gral gain is 400. The bandwidth of the velocity loop transfer function is 1.0 Hz, both
in azimuth and in elevation.
The position loop model is shown in Fig. 8.22a. It consists of the velocity loop
model, PI, and feedforward controllers in azimuth and elevation, command prepro-
cessors in azimuth and elevation, and velocity and acceleration limiters in azimuth
and elevation. The telescope velocity limit is 1.0 deg/s, and the acceleration limit
is 0.5 deg/s2 , both in azimuth and elevation. The PI controller gains were selected
to minimize settling time and servo error in wind gusts. They also guarantee zero
steady-state error for constant velocity tracking. The proportional gain is 3.0, and
the integral gain is 1.0.
The telescope performance is illustrated with the elevation axis performance
(azimuth axis performance is similar). The position loop transfer function is shown
in Fig. 8.23, dotted line. It follows from this figure that the bandwidth is 1.8 Hz.
In order to evaluate settling time a small (0.01 deg) step response was simulated,
shown in Fig. 8.24. From the plot one can see that there is no overshoot and that the
settling time is 3.0 s.
The wind gusts time history was obtained from the wind spectrum; see Chapter 5.
The plots of the servo error in azimuth and elevation are shown in Fig. 8.25a.
The PP control system analysis showed also that LMT is a sturdy structure. Its
fundamental frequency is 1.7 Hz (it is expected 1.3 Hz for a 50-m radiotelescope, as
shown in Fig. 3.12). The more than usual stiffness of the LMT structure (expressed
as a higher-than-expected fundamental frequency) allows for higher gains in the PP
controller, thus enhancing pointing performance.

8.5.4 PL Control System

The PL control system consists of the LQG controller in the position loop, and PI
controller in the velocity loop. In this case the velocity loop model of telescope is as
in Fig. 8.22b. It consists of the structure and drive model, and azimuth and elevation
velocity loop controllers (PI type).
The position loop is presented in Fig. 8.22, where the PI controllers are replaced
with the LQG controllers, as in Fig. 8.22c. It consists of the same velocity loop
model as PP control system, the LQG controller with feedforward loop, command

@Seismicisolation
@Seismicisolation
130 8 LQG Controller

1
10

0
10
EL magnitude

–1
10
PP
–2 PL
10
LP
LL
–3
10
–2 –1 0 1 2
10 10 10 10 10
frequency, Hz

-50
EL phas e, deg

–100

–150 PP
PL
–200 LP
LL
–250
–2 –1 0 1 2
10 10 10 10 10
frequency, Hz

Fig. 8.23 Telescope position loop transfer functions: (a) magnitude and (b) phase

0.014

0.012
EL telescope position, deg

0.01

0.008

0.006
LL
0.004
LP
0.002 PL
PP
0
0 0.5 1 1.5 2 2.5 3 3.5 4
time, s

Fig. 8.24 Telescope responses to 0.01 deg step

@Seismicisolation
@Seismicisolation
8.5 LQG Controller in the Velocity Loop 131

4 4
EL servo error, mdeg PP PL

EL servo error, mdeg

2 2

0 0

–2 –2

–4 –4
–4 –2 0 2 4 –4 –2 0 2 4
AZ servo error, mdeg AZ servo error, mdeg

0.5 0.5
LP LL
EL servo error, mdeg

EL servo error, mdeg

0 0

–0.5 –0.5
–0.5 0 0.5 –0.5 0 0.5
AZ servo error, mdeg AZ servo error, mdeg

Fig. 8.25 Telescope servo error in 12 m/s wind gusts: with PP, PL, LP, and LL control systems

preprocessor, and velocity and acceleration limits. The CPP parameters are as
follows: kv = 6, ko = 0.6, and β = 20, for azimuth and elevation.
The performance of the PL control system was evaluated, using settling time,
bandwidth, and servo error in wind gusts. The step response for small step (0.01 deg)
is shown in Fig. 8.24, dash-dot line. The figure shows 1.2 s settling time and 35%
overshoot. The position loop transfer function is shown in Fig. 8.23, dash-dot line,
showing wide bandwidth of 1.5 Hz.
The wind gusts simulations show 0.15 mdeg rms servo error in azimuth and
0.74 mdeg rms servo error in elevation, as shown the servo error plot in Fig. 8.25b.
These numbers compared with the PP control system (0.35 mdeg in azimuth and
1.4 mdeg in elevation) show that the LQG controller improves the servo error in
wind over the PI controller by factor 2.3 in azimuth, and by factor 1.9 in eleva-
tion. The bandwidth of the PL system is smaller than the PP system, due to the PP
system transfer function hump seen in Fig. 8.23, which extends the bandwidth but
deteriorates the system performance.

@Seismicisolation
@Seismicisolation
132 8 LQG Controller

8.5.5 LP Control System

The LP control system consists of the PID (proportional, integral, and deriva-
tive) controller in the position loop, and the LQG controller in the velocity loop.
Its Simulink model is shown in Fig. 8.22a, and the velocity loop subsystem in
Fig. 8.22b.
The Simulink model of the open loop telescope is shown in Fig. 8.22b, with
the velocity feedback removed. The open loop model is scaled to obtain maximal
velocity of 1 deg/s for 10 V command (a standard input to motor drives). For this
open loop model an LQG controller was tuned and then evaluated the performance
of the velocity loop LQG controller by using step responses and transfer functions of
the elevation axis. For this velocity loop the settling time is 0.2 s and the bandwidth
is 1.8 Hz.
The position loop is as in Fig. 8.22, where the velocity controllers are now of
the LQG type. Besides the velocity loop, it consists of the PID controller with feed-
forward loop, the command preprocessor, and velocity and acceleration limiters.
The feedforward loop forwards the command velocity to the velocity loop input.
The following PID gains were selected: proportional gain 10, integral gain 6, and
derivative gain 5. The CPP parameters are as follows: kv = 6, ko = 0.93, and
β = 30.
The position loop step responses for small step (0.01 deg) is shown in Fig. 8.24,
dashed line, showing 0.6 s settling time. The position loop transfer function is shown
in Fig. 8.23, dashed line, showing wide bandwidth of 20 Hz.
The wind gust simulations to 12 m/s wind are plotted in Fig. 8.25c. The figure
shows 0.012 mdeg rms servo error in azimuth and 0.150 mdeg rms servo error in
elevation. These small numbers show that the LQG controller in the velocity loop
improves the servo error in wind over the PP control system by factor 30 in azimuth,
and by factor 10 in elevation.

8.5.6 LL Control System

Finally, the telescope control system with the LQG controller in the velocity- and
position loops was tuned. The velocity loop of the LL system is identical to that of
the LP control system. The position loop controller was tuned to minimize the servo
error in the wind gusts. The position loop characteristics are plotted in Figs. 8.23
and 8.24. From Fig. 8.24 it follows that the system settling time is 0.5 s, and there is
no overshoot. From Fig. 8.23 one can find that the bandwidth is 40 Hz. Finally, the
wind gusts simulations to 12 m/s wind are plotted in Fig. 8.25d. The figure shows
0.0012 mdeg rms servo error in azimuth and 0.0057 mdeg rms servo error in eleva-
tion, which gives the total rms error of 0.0058 mdeg. It is 250 times smaller than the
error of the PP control system. Thus, the LL control system performance is the best
of all presented system, although the system is the most complex and will require
careful tuning of both velocity and position loop LQG controllers in order to obtain
the predicted performance.

@Seismicisolation
@Seismicisolation
References 133

Table 8.3 Performance of the PP, LP, PL and LL control systems

Control Settling Overshoot Bandwidth Servo error in
system time s % Hz wind mdeg
PP 3.0 20 1.2 1.48
LP 0.6 0 20 0.15
PL 1.4 20 1.4 0.76
LL 0.5 0 20 0.004

The LMT control systems performance is summarized in Table 8.3. Based on the
performed analysis one concludes that:
• The PP control system shows improved pointing accuracy when compared to
similar control systems applied to typical antennas or telescopes. It was achieved
because the analysis showed the exceptionally rigid LMT structure.
• The PL control system has twice better pointing precision in wind than the PP
system.
• The analysis shows that pointing accuracy in the wind of the LP control system
is ten times better than the PP system. This significant reduction was achieved
because of the expanded bandwidth of the velocity loop.
• The analysis shows that its pointing accuracy in the wind of the LL control system
is 250 times better than PP system.

Finally, some comments on the obtained performance estimates of the telescope

are necessary. The estimates include unknown factors. First, the presented telescope
performance is based on the analytical models the structure and the drives, which
do not represent an accurate dynamics of the telescope. To improve the accuracy
a model shall be derived from the system identification and data collected at the
real telescope. Next, the wind disturbance torques are applied to the drives, while
in reality the wind acts on the entire structure, including the dish surface. Finally,
the RF beam movement, the ultimate goal of the control, was not simulated. Instead
azimuth and elevation encoders were used, which only partially reflected the beam
position. The encoders—although relatively precise—cannot exactly measure the
actual beam position due to their distant location from the beam focal point, which
is the RF beam location.

References
1. Collins, Jr., EG, Haddad WM, Ying SS. (1994). Construction of Low-Authority, Nearly
Non-Minimal LQG Compensators for Reduced-Order Control Design. 1994 IEEE American
Control Conference, Baltimore, MD.
2. Dutton K, Thompson S, Barraclough B. (1997). The Art of Control Engineering. Addison-
Wesley, Harlow.
3. Eisentraeger P, Suess M. (2000). Verification of the Active Deformation Compensation
System of the LMT/GMT by End-to-End Simulations. In: Radio Telescopes, Proceedings of
SPIE, Vol. 4015.
4. Gawronski W. (1994). Design of a Linear Quadratic Controller for the Deep Space Network
Antennas. AIAA J. Guidance, Control, and Dynamics, vol. 17.

@Seismicisolation
@Seismicisolation
134 8 LQG Controller

5. Gawronski W. (2001). Antenna Control Systems: From PI to H∞ . IEEE Antennas and Prop-
agation Magazine, 43(1).
6. Gawronski W. (2004). Advanced Structural Dynamics and Active Control of Structures,
Springer, New York.
7. Gawronski W. (2007). Servo Performance Parameters of the NASA Deep Space Network
Antennas. IEEE Antennas and Propagation Magazine, 49(6).
8. Gawronski W, Mellstrom JA. (1994). Control and Dynamics of the Deep Space Network
Antennas. In: Control and Dynamics Systems, ed. Leondes, CT vol. 63, Academic Press, San
Diego, CA, pp. 289–412.
9. Gawronski W, Perez-Zapardiel P. (2007). Performance Comparison of the LQG
and PI Controllers in Wind Gusts.IPN Progress Report, 42–171. Available at
http://ipnpr.jpl.nasa.gov/progress report/42-171/171D.pdf.
10. Gawronski W, Racho C, Mellstrom JA. (1995). Application of the LQG and Feedforward
Controllers for the DSN Antennas. IEEE Trans. Control Systems Technology, vol. 3.
11. Gawronski W, Souccar K. (2005). Control Systems of the Large Millimeter Telescope. IEEE
Antennas and Propagation Magazine, 47(4).
12. Kaercher HJ, Baars JWM. (2000). The Design of the Large Millimeter Telescope/Gran Tele-
scopio Milimetrico (LMT/GTM). In: Radio Telescopes, Proceedings of SPIE, vol. 4015.
13. Li K, Kosmatopoulos EB, Ioannou PA et al. (1998). Control Techniques for a Large
Segmented Reflector. Proc. 37 th IEEE Conf. Decision and Control, Tampa, FL.
14. Li K, Kosmatopoulos EB, Ioannou PA et al. (2000). Large Segmented Telescopes: Central-
ized, Decentralized and Overlapping Control Designs. IEEE Control Systems Mag., 20(5):
59–72.
15. Maneri E, Gawronski W. (2000). LQG Controller Design Using GUI: Application to Antennas
and Radio-Telescopes. ISA Transactions, 39(2): 243–264.
16. Olberg M, Lindeborg C, Seyf A et al. (1995). A Simple Robust Digital Controller for the
Onsala 20-m Radio Telescope. Proc. SPIE, 2479: 257–265.
17. Whorton M, Angeli G. (2003). Modern Control for the Secondary Mirror of a Giant
Segmented Mirror Telescope. Proc. SPIE, Future Giant Telescopes, vol. 4840.

@Seismicisolation
@Seismicisolation
Chapter 9
H∞ Controller

This chapter presents the tuning process of the H∞ controller. It describes the
H∞ controller modified for tracking purposes, gives the closed loop equations, and
presents the 34-m antenna example, with the limits of the performance.

9.1 Definition and Gains

Application of the H∞ controllers to antennas and telescopes is discussed in [1, 2,

3, 4, 5, 6, 7, 8]. The structure of an H∞ controller is similar to that of the LQG
controller, although its parameters are obtained from a different algorithm. The
algorithm minimizes the system H∞ norm, which is, in the case of a SISO system,
the maximal magnitude of its transfer function. Also, in the algorithm the control (u)
and the disturbance (w) inputs of a system are separated (see Fig. 9.1). The control
(y) and performance (z) outputs are also parted. In our case the plant-controlled
input (u) is the velocity input to the drives, the disturbance input (w) is the wind
pressure at the antenna structure, the controlled output (y) is the encoder reading,
and the performance output is the RF beam position.
A significant portion of the antenna tracking error is generated by the antenna
vibrations excited by wind gusts. The LQG controller improves its tracking, but
the disturbance rejection properties have not been addressed directly in the tuning
process. The H∞ controller allows for addressing simultaneously its tracking and
disturbance rejection properties, as it will be shown in the following.
The closed loop system architecture is shown in Fig. 9.1. In this figure G is the
transfer function of an antenna, while K is the transfer function of a controller. For a
closed loop system as in Fig. 9.1 the plant transfer function G(s) and the controller
transfer function K(s) are such that
+ , + ,
z(s) w(s)
= G(s) ,
y(s) u(s) (9.1)
u(s) = K (s)y(s),

where u, w are control and exogenous inputs and y and z are measured and controlled
outputs, respectively. The related state-space equations of a structure are as follows:

W. Gawronski, Modeling and Control of Antennas and Telescopes, 135

@Seismicisolation
@Seismicisolation
DOI: 10.1007/978-0-387-78793-0 9, 
C Springer Science+Business Media, LLC 2008
136 9 H∞ Controller

Fig. 9.1 The H∞ closed loop w z

system configuration: G, G
plant; K, controller; u,
velocity input; w, u y
disturbances; y, encoder;
and z, RF error
K

ẋ = Ax + B1 w + B2 u,
z = C1 x + D12 u, (9.2)
y = C2 x + D21 w.

Hence, the state-space representation in the H∞ controller description consists

of the quintuple (A, B1 , B2 , C1 , C2 ). For this representation, (A, B2 ) is stabilizable
and (A, C2 ) is able to be detectable, and the conditions

D12 T
C1 D12 = 0 I ,


(9.3)
D21 B1T D21 T
= 0I

are satisfied. When the latter conditions are satisfied the H∞ controller is called the
central H∞ controller. These are quite common assumptions, and in the H2 control
they are interpreted as the absence of cross terms in the cost function (D12 T
C1 = 0),
and the process noise and measurement noise are uncorrelated (B1 D21 = 0).
T

The H∞ control problem consists of determining controller K such that the H∞

norm of the closed loop transfer function G wz from w to z is minimized over all
realizable controllers K, that is, one needs to find a realizable K such that

G wz (K )∞ (9.4)

is minimal. Note that the LQG control system depends on y and u rather than on w
and z, as above.
The solution says that there exists an admissible controller such that G wz ∞ <
ρ, where ρ is the smallest number such that the following four conditions hold:
1. S∞c ≥ 0 solves the following central H∞ controller algebraic Riccati equation
(HCARE),

S∞c A + A T S∞c + C1T C1 − S∞c (B2 B2T − ρ −2 B1 B1T )S∞c = 0. (9.5)

2. S∞e ≥ 0 solves the following central H∞ estimator (or filter) algebraic Riccati
equation (HFARE),

S∞e A T + AS∞e + B1 B1T − S∞e (C2T C2 − ρ −2 C1T C1 S∞e = 0. (9.6)

@Seismicisolation
@Seismicisolation
9.1 Definition and Gains 137

3.

λmax (S∞c S∞e ) < ρ 2 , (9.7)

where λmax (X ) is the largest eigenvalue of X.

4. The Hamiltonian matrices
 
A ρ −2 B1 B1T − B2 B2T
,
−C T C1 −A T
 1T  (9.8)
A ρ −2 C1T C1 − C2T C2
,
−B1 B1 T
−A

do not have eigenvalues on the jω-axis.

With the above conditions satisfied the optimal closed loop system is presented
in Fig. 9.2, and the controller state-space equations, from the input y to the output
u, are obtained from the block-diagram in Fig. 9.2,

x̂˙ = (A + ρ −2 B1 B1T S∞c − B2 K c − K e C2 )x̂ + K e y,

(9.9)
u = −K c x̂.

w
Structure (G) z
D21

B1
d
A

+ +
u + x x + y
B2 ∫ C2
–

C1
+
+
D12 y
u

Controller (K)
ŷ – + y
C2

A Ke

u x̂ x̂ + +
Kc ∫
+ –
ρ–2B1B1TSα c d̂

u
B2

Fig. 9.2 An H∞ closed loop system

@Seismicisolation
@Seismicisolation
138 9 H∞ Controller

According to the above equations the H∞ controller state-space representation

(A∞ , B∞ , C∞ ) is as follows:

A∞ = A + ρ −2 B1 B1T S∞c − B2 K c − K e C2 ,
B∞ = K e , (9.10)
C∞ = −K c ,

where

K c = B2T S∞c (9.11)

and

K e = So S∞e C2T ,
(9.12)
So = (I − ρ −2 S∞e S∞c )−1 .

The gain K c is called the controller gain, while K e is the estimator gain. The
order of the controller state-space representation is equal to the order of the plant.
Note that the form of the H∞ solution is similar to the LQG solution. However, the
LQG gains are determined independently, while the H∞ gains are coupled through
the inequality, and through the component So in as in the above equation.

9.2 Tracking H∞ Controller

For tracking purposes the plant is upgraded with the integrator to maintain zero
steady-state error for constant-rate command, as was already discussed in the LQG
controller tuning. An H∞ tracking controller with an integral upgrade is presented
in Fig. 9.3.

9.3 Closed-Loop Equations of the Tracking H∞ Controller

The closed loop equations for the controller as presented in Fig. 9.3 are as follows,

ẋcl = Acl xcl + Bcl r + Bw w

y = Ccl xcl + D21 w (9.13)
z = C1 x + D12 u

where
⎧ ⎫
⎨ei ⎬
xcl = x (9.14)
⎩ ⎭
x̂

@Seismicisolation
@Seismicisolation
9.4 34-M Antenna Example 139

w Structure (G) z
D21

B1
d
A

+ +
u + x x + y
B2 ∫ C2
+
C1
+ + z
D12 y
u

Controller (K)
u – ŷ – + y
Kf C1 C2
+ +
A Ke
ki kp
ei x̂ x̂ + +
e ∫
∫

+
+
r ŷ ρ–2B1B1TSαc dˆ
C2
+ –
u
B2

Fig. 9.3 An H∞ tracking controller with an integral upgrade

⎡ ⎤
0 0 −C2
Acl = ⎣ B2 ki A −B2 k p C2 − B2 k f C1 ⎦,
B2 ki ke C 2A − ke C2 − B2 k p C2 − B2 k f C1 + ρ −2 B1 B1T S∞c
(9.15)
⎡ ⎤ ⎡ ⎤
1 0

Bcl = ⎣ B2 k p ⎦ , Bw = ⎣ B1 ⎦ , Ccl 0 C p 0 (9.16)

B2 k p 0

9.4 34-M Antenna Example

In Chapter 5, the wind spectra were determined from the wind field data. Based on
these spectra, and using the antenna model in the modal representation, the wind
filter is added by an appropriate scaling of the input matrix B1 of the antenna. The
scaling factors are the filter gains at the natural frequencies of the antenna.
The H∞ controller for the elevation axis is obtained and simulated. Its tracking
performances are and compared with the LQG and PI controller performances. First,
the responses to small (10 mdeg) step commands are compared; see Fig. 9.4. The
settling time of the PI controller is 15 s, LQG controller is 2 s, and the H∞ controller is
1 s. The responses to the step disturbance are shown in Fig. 9.5. The integral over the

@Seismicisolation
@Seismicisolation
140 9 H∞ Controller

14

12

10
s tep res pons e, mdeg

4
PI
Hinf
2
LQG
0
0 2 4 6 8 10 12 14 16 18 20
time, s

Fig. 9.4 Comparison of the step responses of the PI, LQG and H∞ controllers

20
PI
dis turbanc e s tep res pons e, mdeg

Hinf
15
LQG

10

–5
0 2 4 6 8 10 12 14 16 18 20
time, s

Fig. 9.5 Comparison of the disturbance step responses of the PI, LQG and H∞ controllers

absolute value of the responses is 98, 3.2, and 0.9, for the PI, LQG, and H∞ controller,
respectively. The transfer functions from the command input to the encoder output
are shown in Fig. 9.6. The plot shows improved tracking performance of the H∞
controller (the bandwidth is 6 Hz for the H∞ controller, 1.2 Hz for the LQG controller,
and 0.09 Hz for the PI controller). The wind disturbance rejection properties are repre-
sented by the transfer functions from the wind disturbance input to the encoder output
(Fig. 9.7), where the H∞ controller disturbance transfer function is about a decade
lower than (factor 0.1) the LQG controller, and three decades lower (factor 0.001)

@Seismicisolation
@Seismicisolation
9.5 Limits of Performance 141

1
10

0
10
magnitude

–1
10

–2
10 PI
Hinf
LQG
–3
10
–2 –1 0 1
10 10 10 10
frequency, Hz

Fig. 9.6 Comparison of the transfer functions of the PI, LQG and H∞ controllers

1
10
PI
Hinf
0
10 LQG
magnitude, dis turbanc e

–1
10

–2
10

–3
10
–2 –1 0 1
10 10 10 10
frequency, Hz

Fig. 9.7 Comparison of the disturbance transfer functions of the PI, LQG and H∞ controllers

than the PI controller, showing improved disturbance rejection properties of the H∞

controller.

9.5 Limits of Performance

The simulations and measurements show that the high-gain H∞ controller is hard to
implement due to acceleration limits imposed on the velocity input signal u that
drives the antenna. The limit is imposed to prevent overloading of the motors.
For the 34-m antennas the acceleration limits are ±0.4 deg/s2 . The acceleration
@Seismicisolation
@Seismicisolation
142 9 H∞ Controller

(a)
1.5

v eloc ity , deg/s

1
0.5
0

–0.5
0 0.2 0.4 0.6 0.8 1
(b) time, s

50
ac c eleration, deg/s 2

–50
0 0.2 0.4 0.6 0.8 1
time, s

Fig. 9.8 H∞ controller response to 10-mdeg step offset: (a) velocity at the antenna input, (b)
acceleration at the antenna input

of the input u during 10 mdeg step offset are shown in Fig. 9.8. The acceleration
reaches the large value of 40 deg/s2 . The acceleration limiter reduces these values to
±0.4 deg/s2 , but this non-linear operation destabilizes the system. In order to avoid
the excessive accelerations a command pre-processor is implemented. It commands
the antenna with max velocity and/or acceleration if the velocity and/or accelera-
tions are exceeded, and commands with the original command if the limits are not
exceeded. This helps in no-wind weather. However, in windy weather the antenna
is moved not only by the command, but also by the wind gusts. In this scenario the
input u is a combination of the command and the feedback of the wind disturbance.
Although the command, due to the preprocessor, is below the acceleration limits, the
encoder signal from the feedback—due to the strong reaction of the H∞ controller—
exceeds the acceleration limits and destabilizes the system. This behavior shows that
the improvement of the H∞ controller performance is tied to the acceleration limits
imposed on the antenna drives. The higher are the limits the smaller is the servo
error.

References
1. Erm T, Bauvir B, Hurak Z. (2004). Time to Go H-infinity? Proc. SPIE, Advanced Software,
Control, and Communication Systems for Astronomy, Glasgow, UK, vol. 5496.
2. Gawronski W. (1996). H∞ Controller for the DSS13 Antenna with Wind Disturbance Rejec-
tion Properties. TDA Progress Report, vol. 42–127, 1996. Available at http://ipnpr.jpl.nasa.gov/
progress report/42-127/127 G.pdf
3. Gawronski W. (2001). Antenna Control Systems: From PI to H∞ . IEEE Antennas and Propa-
gation Magazine, 43(1).
4. Gawronski W. (2004). Advanced Structural Dynamics and Active Control of Structures,
Springer, New York.

@Seismicisolation
@Seismicisolation
References 143

5. Li K, Kosmatopoulos EB, Ioannou PA et al. (2000). Large Segmented Telescopes: Central-

ized, Decentralized and Overlapping Control Designs. IEEE Control Systems Magazine, 20(5):
59–72.
6. Li K, Kosmatopoulos EB, Ioannou PA et al. (1998). Control Techniques for a Large Segmented
Reflector. Proc. 37th IEEE Conf. on Decision and Control, Tampa, FL.
7. Schoenhoff U, Klein A, Nordmann R. (2000). Attitude Control of the Airborne Telescope
SOFIA: ␮-Synthesis for a Large-Scaled Flexible Structure. Proc. 39th IEEE Conf. Decision
and Control, Sydney, Australia.
8. Whorton M, Angeli G. (2003). Modern Control for the Secondary Mirror of a Giant Segmented
Mirror Telescope. Proc. SPIE, Future Giant Telescopes, vol. 4840.

@Seismicisolation
@Seismicisolation
Chapter 10
Single Loop Control

Control systems of most antennas and radiotelescopes consist of two feedback

loops: the velocity and position loops. The velocity feedback is an inner loop that
controls motor velocity. The position feedback is an outer loop that controls the
antenna position. The questions arise: Is the velocity loop necessary, and for what
reason? Is the antenna pointing performance degraded when the velocity feedback
is removed? This chapter tries to answer these questions. Simplicity is the reason
for the proposed removal of the velocity feedback, because removing the velocity
feedback simplifies the hardware, lowers the cost of the antennas, and lowers
the cost of their maintenance. A trend toward eliminating the velocity feedback
loop is observed in spacecraft and robot control. Recent papers (e.g., [1] and [2])
show that spacecraft can be stabilized and controlled without velocity feedback,
and that robots can be controlled without velocity measurements; see [3], [4], [6],
and [7].
The analysis starts with an ideal (or rigid) antenna to obtain initial assessment of
the performance in a closed-form and then proceeds to a 34-m antenna that displays
drive and structural flexibility [5]. The performance is compared for antennas with
and without a velocity feedback loop, including their responses to commands and
disturbances.

10.1 Rigid Antenna

The rigid-body antenna model is a model of an idealized antenna, which allows for
analysis of basic properties in a closed form.

10.1.1 Rigid Antenna with Velocity and Position Loops

A typical control system for a rigid-body antenna consists of the velocity and posi-
tion loops, as shown in Fig. 10.1a. A proportional controller is used to control the
velocity loop. The transfer function of the closed velocity loop system, from velocity
command u to antenna velocity ϕ̇ is as follows

W. Gawronski, Modeling and Control of Antennas and Telescopes, 145

@Seismicisolation
@Seismicisolation
DOI: 10.1007/978-0-387-78793-0 10, C Springer Science+Business Media, LLC 2008
146 10 Single Loop Control

(a)
d rigid antenna

r e u τ + 1 ϕ 1 ϕ
K(s) ko s
Js

(b) d

r e τ + 1 ϕ
K(s)
Js2
rigid antenna

Fig. 10.1 Rigid antenna control system: (a) with the velocity and position feedback, and (b)
with the position feedback only: r, position command; e, servo error; τ , torque; d, disturbance; φ,
antenna position; ϕ̇, antenna velocity; ko , gain of the velocity controller; K(s), position controller
transfer function

ϕ̇(s) ko G(s) 1
G rl (s) = = = (10.1)
u(s) 1 + ko G(s) 1 + Ts

where T = J / ko . The bandwidth of the velocity loop system is equal to B =

1 / T = ko / J rad/s. The gain, ko , is tuned to obtain the required bandwidth of the
system. The required bandwidth is B = 10 rad/s, the inertia is J = 1 Nms2 /rad, thus
ko = 10 Nms.
The position loop is closed using a proportional-and-integral (PID) controller.
The PID controller transfer function is
ki
K (s) = kd s + k p + (10.2)
s

where kd , k p , and ki are derivative, proportional, and integral gains, respectively.

The transfer function of the closed position loop (from the command r to the encoder
position φ) is as follows

ϕ(s) K (s)G rl (s)/s ko (kd s 2 + k p s + ki )

G cl (s) = = =
r (s) 1 + K (s)G rl (s)/s J s 3 + ko (1 + kd )s 2 + ko k p s + ko ki
(10.3)

This system is stable for k p > T ki . The above equation shows that for low
frequencies (s → 0) the magnitude of the transfer function is unity; thus, for
these frequencies ϕ = r , i.e., the antenna follows exactly the command. For high

@Seismicisolation
@Seismicisolation
10.1 Rigid Antenna 147

frequencies (s → ∞) the magnitude of the transfer function tends to zero, that is,
the antenna does not respond to the command or to the high-frequency noise.
The transfer function from the disturbance d to encoder position φ for the PI
controller is as follows (see Fig. 10.1a):

ϕ(s) s
G d (s) = = (10.4)
d(s) J s 3 + ko s 2 (1 + kd ) + ko k p s + ko ki

It shows that for low frequencies (s → 0) the magnitude of the transfer function
tends to zero and for high frequencies (s → ∞) the magnitude of the disturbance
transfer function tends to zero as well, that is, for these frequencies the disturbances
are completely rejected.

10.1.2 Rigid Antenna with Position Loop Only

When the velocity loop feedback is removed, the antenna transfer function from the
torque input to the position output is

ϕ(s) 1
G(s) = = (10.5)
τ (s) J s2

The closed loop control without the velocity feedback is shown in Fig. 10.1b.
By closing the position feedback with a PI controller the system becomes unstable.
Indeed, the closed loop transfer function is

ϕ(s) K (s)G(s) k p s + ki
G cl (s) = = = (10.6)
r (s) 1 + K (s)G(s) J s + k p s + ki
3

From Routh–Hurwitz criterion the above system is unstable for arbitrary positive
gains, that is, for k p > 0 and ki > 0.
However, adding a derivative gain to the PI controller one obtains the PID
controller. For this controller the position-only loop is stable. The transfer function
of the PID controller is K (s) = ksi +k p +kd s; thus the closed loop transfer function is

ϕ(s) K (s)G(s) kd s 2 + k p s + ki
G cl (s) = = = (10.7)
r (s) 1 + K (s)G(s) J s 3 + kd s 2 + k p s + ki

From the Routh–Hurwitz criterion, this system is stable for kd > 0. The above
transfer function shows that for low frequencies (s → 0) the magnitude of the
transfer function is unity; thus, for these frequencies one has ϕ = r , that is, the
antenna follows the command exactly. For high frequencies (s → ∞) the magnitude
of the transfer function tends to zero, that is, the antenna does not respond to high-
frequency noise.

@Seismicisolation
@Seismicisolation
148 10 Single Loop Control

The disturbance transfer function is derived from Fig. 10.1b, obtaining

ϕ(s) s
G d (s) = = (10.8)
d(s) J s + kd s + k p s + ki
3 2

It shows that for low frequencies (s → 0) as well as for high frequencies

(s → ∞) the magnitude of the disturbance transfer function tends to zero |G d | ≈ 0,
that is, for these frequencies the disturbances are rejected.

10.1.3 Simulation Results

The rigid-body antenna with and without velocity loop feedback is simulated. The
gains of the PID controller for the rigid antenna with and without velocity loop are
given in Table 10.1. For these controllers the closed loop system responses to the
1 deg step command and to the 1 deg/s disturbance step were simulated. They are
shown in Fig. 10.2a,b, respectively. The figures show that the settling time of the
antenna with the velocity loop and the antenna without a velocity loop are approxi-
mately the same (3.5 s). However, the disturbance rejection properties of the antenna
with the velocity loop are much worse than of the antenna without the velocity loop
(the amplitude and the duration of the response of the antenna with a velocity loop
are much larger than the antenna without the velocity loop).
Magnitudes of the transfer functions of the closed loop systems, from the
command to the encoder, are shown in Fig. 10.3a and from the disturbance to the
encoder in Fig. 10.3b. The first figure shows that the bandwidth of the antenna with
and without the velocity loop are the same (0.6 Hz). However, the magnitude of the
disturbance transfer function of the antenna without the velocity loop in Fig. 10.3b
is smaller that the antenna with the velocity loop (for all frequencies). It shows again
that the antenna without a velocity loop has better disturbance rejection properties
than the antenna with the velocity loop.
Finally, note that the controller gains of the antenna without a velocity loop
are higher than the ones of the antenna with the velocity loop, indicating that the
control effort (torques) may be higher in the case of antennas without a velocity
loop.
In summary the tracking performance of the rigid antenna with and without
velocity feedback are similar, but the antenna without the velocity loop has better
disturbance rejection properties, and possibly requires more control effort.

Table 10.1 Controller gains of the rigid antenna

Control system ko kd kp ki
With velocity feedback 10 0 2.2 1.4
Without velocity feedback 0 130 300 240

@Seismicisolation
@Seismicisolation
10.2 34-M Antenna 149

(a)
1.2

1
c ommand s tep res pons e

0.8

0.6

0.4

0.2 with velocity loop

without velocity loop
0
0 2 4 6 8 10
time, s
(b)
0.04
with velocity loop
dis turbanc e s tep res pons e

without velocity loop

0.03

0.02

0.01

0
0 2 4 6 8 10
time, s

Fig. 10.2 Step responses of the rigid antenna with a velocity loop (solid line) and without a
velocity-loop (dashed line): (a) responses to the unit command step, and (b) responses to the unit
disturbance step

10.2 34-M Antenna

The 34-m antenna model, as described in [5] and Chapter 2, is used for this analysis.
Our task is to tune and compare the LQG controllers for the antenna with velocity
feedback and without velocity feedback. The description of the LQG controller for
the antenna with velocity feedback was given in Chapter 8. The LQG control loop
is shown in Fig. 10.4, which includes the antenna drive.
Consider two kinds of drives: one with the velocity feedback, and another without
the velocity feedback. An antenna drive with a velocity loop feedback is shown in
Fig. 10.5. In this figure the ktach gain represents the velocity (tachometer) feedback

@Seismicisolation
@Seismicisolation
150 10 Single Loop Control

(a)

0
10

–1
magnitude

10

–2
10
with velocity loop
without velocity loop
–3
10
–3 –2 –1 0 1 2
10 10 10 10 10 10
frequency, Hz
(b)

–2
10
dis turbanc e magnitude

–4
10

with velocity loop

without velocity loop
–6
10
–3 –2 –1 0 1 2
10 10 10 10 10 10
frequency, Hz

Fig. 10.3 Magnitudes of the transfer functions of the rigid antenna with a velocity loop (solid
line) and without a velocity loop (dashed line): (a) from the command input, and (b) from the
disturbance input

gain. The same drive without velocity feedback is shown in Fig. 10.6. It was
obtained from the original drive by breaking the velocity feedback (by setting the
feedback gain to zero, ktach = 0) and by replacing the Amplifier1 with the gain kr .

wind disturbance

CONTROLLER
DRIVE STRUCTURE
command position
servo error velocity velocity encoder

Pinion torque torque

Pinion
velocity velocity

Fig. 10.4 Control system of the 34-m antenna

@Seismicisolation
@Seismicisolation
 10.2 34-M Antenna

kc
1/N

2 tau2*kr.s+kr
ki*tau4.s+ki 1
kf 1 +
tau3.s2+s tau5.s2+s km kg
Voltage Jm.s – 1/N 1
La.s+Ra s
Amplifier 1 Amplifier 2 Motor Torque
Motor Gearbox
Armature

kb N 1
Pinion
ktach Velocity

Fig. 10.5 The 34-m antenna drive with velocity loop feedback

@Seismicisolation
@Seismicisolation
151
 152

kc
1/N

ki*tau4.s+ki 10
kf 1
2 kr km kg*100(s)
tau5.s2+s Jm.s 1/N 1
velocity La.s+Ra
Amplifier 2 s
input Motor torque
Motor Gearbox
Armature
pinion
velocity
kb N 1

Fig. 10.6 The 34-m antenna drive without velocity loop feedback

@Seismicisolation
@Seismicisolation
10 Single Loop Control
10.2 34-M Antenna 153

The value of the latter gain is determined such that the open loop gain (from velocity
input to the antenna velocity output) is equal to 1.
The LQG controller gains for the antenna with the velocity feedback are deter-
mined in [5]. A similar controller was tuned for the antenna without a velocity
loop, and the performance of the two systems was compared. The command step
responses, disturbance step responses, command transfer functions, and the distur-
bance transfer functions have been compared.
The command step responses are shown in Fig. 10.7a. The figure shows very
similar responses of the antenna with and without a velocity loop (i.e., similar
overshoot and settling time). The disturbance step responses (representing a rapid
increase of wind torque) are shown in Fig. 10.7b. The plot shows an improved

(a)
1.2

1
antenna pos ition, deg

0.8

0.6

0.4

0.2 with velocity loop

without velocity loop
0
0 2 4 6 8 10
(b) time, s
–4
x 10
2
with velocity loop
without velocity loop
antenna pos ition, deg

–1

–2
0 2 4 6 8 10
time, s

Fig. 10.7 Step responses of the 34-m antenna with a velocity loop (solid line) and without a
velocity loop (dashed line): (a) responses to the unit command step, and (b) responses to the unit
disturbance step

@Seismicisolation
@Seismicisolation
154 10 Single Loop Control

(a)

0
10

–1
magnitude

10

–2
10
with velocity loop
without velocity loop
–3
10
–2 –1 0 1
10 10 10 10
frequency, Hz

(b)
–2
10
with velocity loop
without velocity loop
magnitude, dis turbanc e

–3
10

–4
10

–5
10

–6
10
–2 –1 0 1
10 10 10 10
frequency, Hz

Fig. 10.8 Magnitudes of the transfer functions of the 34-m antenna with a velocity loop (solid
line) and without a velocity loop (dashed line): (a) from the command input, and (b) from the
disturbance input

response of the antenna without velocity feedback (although not as dramatic as at

the rigid antenna).
Figure 10.8a shows the antenna transfer function from the position command
to the antenna encoder position. The transfer functions (of the antenna with and
without velocity feedback) are similar. Figure 10.8b shows the antenna disturbance
transfer function, where the antenna without velocity feedback has slightly better
performance (the magnitude of the transfer function is smaller than the one for the
antenna with velocity feedback).
In summary, the tracking performance of the 34-m antenna with and without
velocity feedback are similar, but the antenna without velocity feedback has better
wind disturbance rejection properties.

@Seismicisolation
@Seismicisolation
References 155

Safety is one of the reasons that the velocity feedback is implemented at the
antennas. In case of failure of the position loop the velocity feedback controls
antenna velocity, maintaining it at zero value. Note, however, that for the safety
reason the velocity loop is not necessary: in the case of failure of the control system
the antenna remains motionless because the gearboxes oppose antenna motion and
brakes are applied automatically. Another reason for implementing velocity feed-
back is its ability to increase damping in the system. This is justified in a case when
a simple (PI type) controller is in the position loop. However, if LQG controller
implemented in the position loop significantly increases the damping of the closed
loop system, and the velocity loop is not necessary.

References
1. Akella MR. (2000). Rigid-Body Attitude Tracking without Angular Velocity Feedback. In:
Spaceflight Mechanics 2000, American Astronautical Society, San Diego.
2. Caccavale F, Villani L. (1999). Output Feedback Control for Attitude Tracking. System Control
Lett., vol. 38.
3. Chang YC, Chen BS, and Lee TC. (1996). Tracking Control of Flexible Joint Manipulators
Using Only Position. Int. J. Control, 64(4).
4. Chang YC, Lee CH. (1999). Robust Tracking Control for Constrained Robots Actuated by DC
Motors without Velocity Measurements. IEE Proc. Control Theory and Applications,146(2).
5. Gawronski W. (2002). Single loop Antenna Control. IPN Progress Report 42-151. Available
at: http://ipnpr.jpl.nasa.gov/progress report/42-151/151D.pdf
6. Lim SL, Dawson DM, Hu J et al. (1997). An Adaptive Link Position Tracking Controller for
Rigid-Link Flexible-Joint Robots without Velocity Measurements. IEEE Trans. System, Man,
Cybernetics, Part B: Cybernetics, 27(3).
7. Loria A, Panteley E. (1999). Force/Motion Control of Constrained Manipulators without
Velocity Measurements. IEEE Trans. Aut. Control, 44(7).

@Seismicisolation
@Seismicisolation
Chapter 11
Non-Linear Control

This chapter discusses three main sources of the antenna non-linear behavior:
velocity and acceleration limits, dry friction, and backlash. The velocity and accel-
eration limits are neutralized by using the command preprocessor, or anti-windup
technique. The friction model is presented and the dither is analyzed as a cure for the
antenna sticking at low velocities. The backlash model is given and the performance
of an antenna with applied counter-torque is analyzed.

11.1 Velocity and Acceleration Limits

The antenna control system is shown in Fig. 3.7. Besides the controller, it includes
the acceleration and velocity limiters. The acceleration limiters are implemented
because the motor currents must be limited to avoid overheating (the currents are
proportional to the antenna acceleration). Antenna velocity is limited for safety
reasons. Those limits are not affected during normal tracking, but are reached during
slew operations. When the limits are reached, the antenna is in a non-linear regime
and starts limit cycling.
The limit cycling can be prevented by the following means:
• applying two different controllers, one for tracking and another for slewing;
• implementing an anti-windup technique, see [18, 21];
• applying a controller with variable gains [24, 25];
• using a trajectory calculated in advance, such that it never exceeds the velocity
and acceleration limits [32];
• using a command preprocessor, see [14, 15].

11.1.1 Command Preprocessor

The limit cycling during antenna slewing operations and target acquisitions is
caused by the violation of the antenna velocity and acceleration limits that cause,
in turn, the increase of the integral error. This problem can be avoided by introduc-
tion of a command that does not exceed the limits; see [14] and [15]. A command

W. Gawronski, Modeling and Control of Antennas and Telescopes, 157

@Seismicisolation
@Seismicisolation
DOI: 10.1007/978-0-387-78793-0 11, C Springer Science+Business Media, LLC 2008
158 11 Non-Linear Control

r rp y
CPP uo u antenna
controller
velocity acceleration
y limit limit

Fig. 11.1 Antenna control system with CPP

preprocessor (CPP) is computer software that generates a modified command, iden-

tical to the original one, if the velocity and accelerations of the command are within
the limits; and a command of maximal (or minimal) velocity and acceleration when
the limits are met or violated. The location of the preprocessor in the antenna control
system is shown in Fig. 11.1.
CPP Description. The block diagram of the CPP is shown in Fig. 11.2. It consists
of a derivative, an integrator, and velocity and acceleration limiters. The proportional
feedback loop has a variable gain ki that depends on the preprocessor error ei . The
sampling time is denoted Δt, where Δt = 0.02s, the command at the ith time instant
t = iΔt is denoted ri , the command velocity is νi , the preprocessed command is
denoted r pi , the input to the integrator is u i , and the preprocessor error is ei =
ri − r pi .
Consider a case where the command does not exceed the velocity and accel-
eration limits. In this case the system is linear, and the velocity and acceleration
limiters in Fig. 11.2 are replaced with a unit gain. For the linear case the equa-
tions are:
• for the integrator:

r pi = r pi−1 + Δtu i (11.1)

• for the derivative

ri − ri−1
νi = (11.2)
Δt

d/dt

ri + ei + ui rpi
+
kpi ∫
–
variable accel. velocity rigid
gain limit limit antenna
controller

Fig. 11.2 Command preprocessor: derivative, integrator, velocity limiter, and acceleration limiter

@Seismicisolation
@Seismicisolation
11.1 Velocity and Acceleration Limits 159

• for the error

ei = ri − r pi (11.3)

• and for the integrator input

u i = ki ei−1 + v i (11.4)

Combining the above equations one obtains

r pi − r pi−1 + Δtki r pi−1 = ri − ri−1 + Δtki ri−1 (11.5)

The above equation shows that ri = r pi for zero error initial conditions. As a
consequence, if the preprocessed command reaches the original command it follows
the latter one exactly.
In the following the transient motion of CPP is investigated. In order to do this,
equation (11.5) can be rewritten as

ei = αi ei−1 (11.6)

where αi = 1 − Δtki . This is an equation of the transient dynamics of the CPP error.
From (11.6) it follows that the system is stable if |α1 | < 1. For 0 < αi < 1 there
is no overshoot, and for −1 < αi < 0 the transient is oscillatory, i = 1, 2, . . . ,.
Consider further positive αi only, the case with no overshoot of the preprocessed
command over the original command. Note that in this case the smaller is the gain
αi , the quicker the error dies down. The gain ki controls the value of αi ; therefore,
for large-gain ki the transient between the original and the preprocessed command
is strongly damped.
Non-tracking activities consist of slewing and of target acquisition. Antenna
slews and targets are acquired by inserting step commands. For this reason step
inputs are used to determine the CPP gain. Two kinds of steps are distinguished:
small steps (that do not violate the velocity and acceleration limits) and large steps
(that evidently violate the limits). For the antenna sampling time Δt = 0.02s
and the velocity limit vmax =0.8 deg/s, the maximal step that does not violate the
velocity limit is v max Δt = 0.016deg. The field test results show that steps larger
than 0.150 deg exhibit significant non-linear behavior. Thus, a step of 0.01 deg is
considered small, and one of 10 deg is considered large.
First, assume small gain, ki = 1. For this gain the response to a large step is
almost optimal. The steady value is reached with the maximal velocity and acceler-
ation, and without overshoot. For small steps, however, the steady value is reached
after 4 s, which is too slow for the antenna. Assume now a CPP with a large gain,
ki = 6. It performs outstandingly for small steps (settling time is 1 s), but for large
steps the CPP exhibits limit cycling. This is illustrated in the simulations presented

@Seismicisolation
@Seismicisolation
160 11 Non-Linear Control

in the next section. These two mental experiments indicated that small gain is useful
for large steps, and large gain is needed for fast response to small steps.
In order to help the CPP to tell the difference between small and large steps, note
that steps are tied to the CPP tracking error: large steps produce large error, and,
not surprisingly, small steps produce small error. Thus, in order to produce different
gains for different steps, the gain value shall depend on the error value. The gain
shall be larger for small error, and smaller for large error. It is achieved by using the
following relationship between the gain and the error:

ki = ko + kν e−β|ei | (11.7)

where ko is the constant part of the gain, kν is the variable part of the gain, and β is
the error exponential. The plot of ki (ei ) for ko = 1, kv = 5, and for β = 10, 20, 40,
and 100 is shown in Fig. 11.3.
It was already determined that ki = 1 is recommended for small errors, and
ki = 6 for large errors. In equation (11.7) the gain ko = 1 sets the lower value of
ki , and kv = 5 sets the upper value of ki . The parameter α shapes the transition of
ki from its minimal to its maximal value, within the error interval of [0.016, 0.150]
deg. β = 20 was chosen because for this value the gain covers the error segment
of [0.016, 0.150] deg (marked gray in Fig. 11.3) where the antenna transfers from
linear to non-linear behavior. Additionally, this choice of β was verified with the
simulations of the CPP performance, as illustrated in the following sections.
Finally the small gain ko needs additional correction, because for the antenna
operational purposes it is preferable that the CPP does not have overshoots for
large steps. The overshoots may appear during the deceleration phase due to the

7
linear non-
linear
6

4
gain ki

2 β = 10
β = 20
1 β = 40
β = 100
0
10–4 10–3 10–2 10–1 100
error ei, deg

Fig. 11.3 CPP gain vs. CPP error

@Seismicisolation
@Seismicisolation
11.1 Velocity and Acceleration Limits 161

0.16
ko = 0.95
0.14
ko = 0.90
0.12
ko = 0.85
overshoot, deg

0.1

0.08

0.06

0.04

0.02

–0.02
0.8 0.82 0.84 0.86 0.88 0.9 0.92 0.94 0.96 0.98 1
acceleration reduction, α

Fig. 11.4 Overshoots for large steps vs. acceleration limit reduction and CPP gain ko

finite sampling time. The magnitude of overshoot with respect to the value of the
acceleration limit, and the CPP gain ko , as in equation (11.7), was investigated. The
maximum value of the acceleration was reduced by factor α, where α varied from
0.8 to 1.0, while the gain ko varied from 0.85 to 0.95. The plot of overshoots for
various large steps with respect to ko and α is shown in Fig. 11.4. The figure shows
that for small acceleration and small gain the overshoot is small too, and eventually it
disappears. Because a 10% margin of acceleration should be preserved to guarantee
a stable closed-loop of the antenna, α = 0.9 was chosen. For this acceleration the
gain ko = 0.88 produces no overshoot, as seen in Fig. 11.5.
CPP Dynamics. The CPP dynamics are checked in the two scenarios, typical for
antennas and radiotelescopes:

(1) Step responses, small and large. Small steps do not violate the limits; large
steps do.
(2) Acquiring and tracking a typical trajectory.
The velocity and acceleration limits of the 34-m antenna are 0.8 deg/s and
0.4 deg/s2 , respectively, and the acceleration limits of the CPP were 90% of the
antenna limits, while the velocity limit of the CPP was equal to the antenna velocity
limit.
For the large step of 10 deg, the preprocessed command, its velocity and acceler-
ation are shown in Fig. 11.6a–c the preprocessed command begins with the maximal
acceleration until it reaches the maximum velocity, then continues with the maximal
(and constant) velocity, and finally slows down with the minimal deceleration. After
reaching the steady-state value of 10 deg the error between the original and the
preprocessed command is zero.

@Seismicisolation
@Seismicisolation
162 11 Non-Linear Control

0.1

0.08
overshoot, deg

0.06

0.04

0.02

–0.02
0.8 0.82 0.84 0.86 0.88 0.9 0.92 0.94 0.96 0.98 1
acceleration reduction, α

Fig. 11.5 Overshoots for large steps vs. acceleration limit reduction, and CPP gain ko = 0.88

Finally, a typical azimuth trajectory acquisition and tracking by the CPP is shown
in Fig. 11.7a. The antenna position at the initial time is 10 deg, while the target
position is at 24 deg. The target is acquired in 20 s with the maximal speed (see
Fig. 11.7b) and maximal acceleration (see Fig. 11.7c), and also with very small
overshoot (c.f. Fig. 11.7a). The CPP error (the difference between the original and
preprocessed trajectory) after the acquiring is virtually zero.
Antenna Dynamics. The 34-me antenna step response is shown in Fig. 11.8a. The
velocity and acceleration of the antenna are shown in Fig. 11.8b and c, respectively.
Velocity is within the imposed limits (0.8 deg/s) and acceleration initially hits the
limit (0.4 deg/s2 ), but does not destabilize the system. For large steps, however, the
antenna hits velocity and acceleration limits (see Fig. 11.9a,b,c), which causes limit
cycling. This happens during antenna slewing.
The response of the CPP to the small step input of 0.01 deg is shown in Fig. 11.8,
dashed line. It is a rapid response of less than 1 s settling time. The antenna follows
the preprocessed command with very little overshoot.
The response of the antenna to a large-step input of 10 deg is shown in Fig. 11.10.
For comparison, the response of the same antenna to a non-processed step is shown
in Fig.11.9. Clearly, an unstable limit cycling disappeared for the antenna with CPP.

11.1.2 Anti-Windup Technique

The anti-windup technique is another option to avoid limit cycling of antennas. The
description of this approach one can be found in [12, 18, 19, 21, 27]. Here a simple,

@Seismicisolation
@Seismicisolation
11.1 Velocity and Acceleration Limits 163

(a)
12
10
position, deg

8
6
4
command
2
pre-processed command
0
0 2 4 6 8 10 12 14 16 18 20
time, s
(b)
1

0.8
rate, deg/s

0.6

0.4

0.2

0
0 2 4 6 8 10 12 14 16 18 20
time, s
(c)
0.5
accel., deg/s2

–0.5
0 2 4 6 8 10 12 14 16 18 20
time, s

Fig. 11.6 Preprocessed large-step commands: (a) command, (b) velocity, (c) acceleration

but effective application of the anti-windup technique to the antenna control system
is described.
The antenna control system is shown in Fig. 11.11, solid line. Besides the
controller it includes the acceleration and velocity limiters. When the limits are
reached the antenna is in a non-linear regime and starts limit cycling. Figure 11.12,
dashed line, shows the antenna oscillatory response to 2 deg position offset.
During slewing and re-targeting, when a large position offset is commanded,
it causes saturation of the velocity and/or acceleration limiters. Thus, the antenna
input u is smaller than the controller output uo. Consequently, the antenna response
y is slower, and the servo error decreases slowly, as well. The slow process impacts

@Seismicisolation
@Seismicisolation
164 11 Non-Linear Control

(a)
25
command
antenna position
position, deg

20

15

10
0 20 40 60 80 100 120 140 160
time, s
(b)
1
velocity, deg/s

0.5

–0.5
0 20 40 60 80 100 120 140 160
time, s
(c)
0.5
accel., deg/s2

–0.5
0 20 40 60 80 100 120 140 160
time, s

Fig. 11.7 Azimuth trajectory: original (dashed line), and preprocessed (solid line): (a) command,
(b) velocity, (c) acceleration

mostly the integral term of the controller and the integral error becomes large. When
eventually y approaches r the controller output uo is still saturated due to the large
integral error, pushing the antenna position to pass the command. This is the begin-
ning of the limit cycling. One can limit the integral term, but it causes unacceptably
large overshoot.
The increase of the integral term can be slowed down by applying the additional
feedback loop that feeds the difference between controller output and antenna input,

@Seismicisolation
@Seismicisolation
11.1 Velocity and Acceleration Limits 165

(a)

position, deg 0.01

0.005
antenna
CPP
0
0 0.5 1 1.5 2 2.5 3
time, s
(b)

antenna
velocity, deg/s

0.04
CPP

0.02

0
0 0.5 1 1.5 2 2.5 3
time, s
(c)
0.4
antenna
accel., deg/s2

CPP
0.2

–0.2
0 0.5 1 1.5 2 2.5 3
time, s

Fig. 11.8 Closed-loop step response to 0.01 deg step: (a) antenna position, (b) antenna velocity,
and (c) antenna acceleration

uo – u, through a compensator. G aw Because the integral term has to be reduced, the

compensator is an integral itself:

kaw
G aw = (11.8)
s

This is the “classical” anti-windup approach. The only parameter to be deter-

mined is the compensator gain kaw , which should guarantee required and stable
performance.
Figure 11.12 presents the antenna responses to the 10 deg position offset, one
without the anti-windup feedback (broken line) and one with the feedback (solid
line), for the feedback gain kaw = 2. It is clear that for large position offsets, without
the anti-windup feedback, the antenna is unstable, while the feedback produces a
stable response with minimal overshoot.

@Seismicisolation
@Seismicisolation
166 11 Non-Linear Control

(a)

position, deg 10

0
0 5 10 15 20 25 30 35 40
time, s
(b)
1
velocity, deg/s

0.5

–0.5

–1
0 5 10 15 20 25 30 35 40
time, s
(c)
0.5
accel., deg/s2

–0.5
0 5 10 15 20 25 30 35 40
time, s

Fig. 11.9 Closed-loop step response to 10 deg step: (a) antenna position, (b) antenna velocity, and
(c) antenna acceleration

11.2 Friction

Many telescopes cannot precisely point at a certain part of the sky that requires
tracking with very low azimuth angular velocities (approximately 0.0004 deg/s or
lower). For such low velocities, dry rolling friction is observed at the telescope
drives that cause an unwanted increase of pointing error. (Pointing error is defined
as a difference between the target location and the RF beam position.)
In this section the National Radio Astronomy Observatory’s Green Bank Tele-
scope located in West Virginia (see Fig. 1.1) is analyzed. The telescope’s size,
weight, and especially configuration create difficulties in precision tracking. One
of them is observed during tracking at low velocities. Namely, for velocities lower
than 0.0003 deg/s a non-smooth telescope motion with breakaways may occur. The
peak-to-peak pointing error due to friction is 1.4 mdeg.

@Seismicisolation
@Seismicisolation
11.2 Friction 167

12

10
antenna position, deg

0
0 5 10 15 20 25 30
time, s

Fig. 11.10 Antenna responses to preprocessed 10 deg step command

Gaw

– +

r +
antenna uo u y
controller
antenna
+
velocity acceleration
y limit limit

Fig. 11.11 Block diagram of the antenna control system (solid line) and anti-windup controller
(broken line): uo , controller output; u, antenna input

A high-frequency external signal injected into a system (its frequency much

higher than the system dynamics) is called a dither [2, 11]. It has long been known
that injection of high-frequency signals into a non-linear system results in elimi-
nating the system limit cycles [2, 11, 23]. This phenomenon was detected in elec-
trical circuits by Appleton in 1922. Although dither is a well-known remedy for
the non-linear behavior of control systems, little is known of its impact on the
dynamics of large flexible structures. It is shown here that dither improves the
telescope pointing at low tracking velocities and that high-frequency dither signal
excites local vibrations only (at the wheels), causing the breaking of the friction-
stiction phenomena. The vibrations are not transmitted through the telescope struc-
ture; thus they do not impact its pointing performance. Also, in this section a friction

@Seismicisolation
@Seismicisolation
168 11 Non-Linear Control

12

10

8
position, deg

2 with AW
without AW
0
0 5 10 15 20 25 30 35 40
time, s

Fig. 11.12 The 34-m antenna responses to 10 deg position offset: with and without anti-windup
(AW) feedback

model frequently used in the antenna industry is presented. It includes the velocity
threshold and the determination of the applied torque within the threshold.

11.2.1 Dry Friction Model

The telescope’s non-linear dynamics, at low velocities, is caused by the dry friction
phenomenon. Friction is a torque, or a force, that depends on the relative velocity of
the moving surfaces. In the Coulomb friction model it is constant after the motion
begins, and this constant value is called the Coulomb friction torque. At zero speed,
the friction torque is equal and opposite of the applied torque, unless the latter one is
larger than the stiction torque. In this case, the friction torque is equal to the stiction
torque. The stiction torque is a torque at the moment of breakaway and is larger
than the Coulomb torque. A diagram of the friction torque versus relative velocity
is shown in Fig. 11.13.
Many friction models have been developed; see, for example, [1–3], [7], and [8].
They reflect different aspects of the friction phenomena; their usefulness depends
on application purposes. The model presented below combines basic physical prop-
erties of the dry friction with the numerical features that improve digital simula-
tions. Its accuracy for antenna tracking purposes has been tested on many existing
telescopes and antennas. In this model, denote v the telescope wheel velocity, and
νt > 0 a wheel velocity threshold that is a small positive number. Denote Tc the
Coulomb friction torque, and Ts the stiction torque; then the friction torque model,
T, is defined as follows:

@Seismicisolation
@Seismicisolation
11.2 Friction 169

Tf
velocity threshold
Ts
Tc

–νt νt ν

–Tc
–Ts

Fig. 11.13 Friction torque versus velocity

-
−Tc sign(ν) for |ν| > νt
T = (11.9)
− min(|Td |, Ts )sign(Td ) for |ν| ≤ νt

where sign(ν) = 1, 0, −1 for ν > 0, ν = 0, and ν < 0, respectively, and Td denotes

the total applied torque. In this model, if the surfaces in a contact develop a measurable
relative velocity, such that |ν| > νt , the friction torque is constant, directed opposite
to the relative speed. If the relative velocity is small, namely, within the threshold,
(|ν| ≤ νt ) the torque does not exceed the stiction torque and the applied torque, and
is directed opposite to the applied torque. The velocity threshold νt is implemented
for numerical purposes, because numerically the zero state does not exist.
It follows from equation (11.9) that in order to determine the friction torque T
one has to know the following:
• the Coulomb friction torque Tc ,
• the stiction (breakaway) torque Ts ,
• the applied torque Td ,
• the wheel velocity ν, and
• the wheel velocity threshold νt .

Each variable is determined as follows:

The Coulomb friction torque is proportional to force F, which is normal to the
surface

Tc = μr F (11.10)

where r is the wheel radius and μ is the friction coefficient. For hard steel μ =
0.0012–0.002.
The stiction (breakaway) torque is most often assumed to be 20%–30% higher
than the Coulomb friction that is

@Seismicisolation
@Seismicisolation
170 11 Non-Linear Control

Ts = αTc (11.11)

where α = 1.2 ÷ 1.3.

The total applied torque Td is determined from the plant dynamics as follows. Let
the discrete state-space equation of the plant (which includes the telescope structure
and its drives) be

x(i + 1) = Ad x(i) + Bdr r (i) + Bd f T (i),

(11.12)
ν(i + 1) = Cd x(i + 1)

In this model Δt, denotes sampling time, i denotes the ith sample, v(i) is the
wheel velocity at time instant iΔt, x(i) is the plant state at the instant iΔt, r(i) is
the telescope angular input velocity, and T(i) is the friction azimuth torque. Addi-
tionally, Ad is the telescope discrete-time state matrix, Bdr nd Bd f are telescope
velocity and friction torque input matrices, and Cd is the wheel velocity output
matrix. Left-multiplying the first equation (11.12) by Cd gives

ν(i + 1) = Cd x(i + 1) = Cd Ad x(i) + Cd Bdr r (i) + Cd Bd f T (i) (11.13)

According to the friction model, for the wheel velocity being within the threshold,
that is, such that |ν(i + 1)| ≤ νt one obtains ν(i + 1) = 0. Thus, from equation
(11.13) it follows that

Cd
T (i) = − (Ad x(i) + Bdr r (i)) (11.14)
Cd Bd f

and the applied torque Td has opposite sign to the above friction torque T.
The wheel velocity threshold νt was assumed to be 0.67 mdeg/s.

11.2.2 Low-Velocity Tracking Using Dither

At low velocities friction significantly impacts the antenna dynamics, and conse-
quently the pointing precision. There are many ways to reduce the system dynamics
due to friction. Most of these methods are based on closed-loop compensation [2, 6,
10, 11, 28]. Here, an open-loop technique is applied by dithering the driving torque
[17]. The block diagram of the closed-loop system with the dry friction and the
velocity dither is shown in Fig. 11.14. In this diagram, according to equation (11.9),
the dry friction torque T is a non-linear function of the azimuth wheel velocity v and
the drive torque Td .
Consider a torque at the azimuth wheel. For low velocities, the driving torque
Td is smaller than the dry friction torque Tc , causing the telescope to stop. While
resting, the error between the commanded position and the actual telescope position

@Seismicisolation
@Seismicisolation
11.2 Friction 171

dither
command + encoder
+ Td + Antenna
Controller Drive
+ structure
– +
T

Dry pinion
friction velocity

Fig. 11.14 Azimuth control system with dry friction and dither

increases. In the closed-loop configuration the increased error causes an increase of

the driving torque Td , and eventually the movement of the telescope is observed. The
cycle repeats itself and is called limit cycling. The plots of the telescope pointing
error in this limit cycling are shown in Fig. 11.15.
A harmonic dither of amplitude do and period to ,

d(t) = do sin(2π t/to ), (11.15)

is introduced at the telescope velocity input. When dither is implemented the torque
level at the wheel is raised. This increase is not a constant; it varies harmonically. If
the amplitude of the driving torque Td exceeds the friction torque, the telescope is
moving continuously and the limit cycling is overcome. Due to the high frequency
of the dither signal (high, when compared to the telescope dynamics), the harmonic
movement is a local phenomenon at the wheels, which is shown later in this section.
It is not propagated through the telescope structure, and has very low impact on the
structural dynamics, and consequently on the telescope pointing.
The above heuristic explanation of the dither action can be derived more formally.
Consider the continuous-time telescope model with the non-linear friction torque
T (ν), driven by the command velocity, r, and dither,d

Fig. 11.15 Pointing error for the simulated tracking, no dither

@Seismicisolation
@Seismicisolation
172 11 Non-Linear Control

ẋ = Ax + Br (r + d) + B f T (ν) (11.16)

The parameters ((A, Br , B f , C)) are the continuous-time counterparts of the

discrete-time parameters (( Ad , Bdr , Bd f , Cd )) as in equation (11.12), and x(t) is the
state variable of the plant. This equation is averaged over the dither period to . The
average value xa of x is defined as

t+to
1
xa = x(τ )dτ (11.17)
to
t

Note that in equation (11.16), the average value of the velocity command, is
almost the same as the instantaneous value, because the command changes insignif-

icantly over the period to , that is, r (t) ∼
= r (t +to ). The average value T (ν) of the non-
linear torque T (ν) is obtained from the dry friction torque as in equation (11.9). The
velocity threshold νt in this equation is assumed to be zero (the non-zero threshold
was previously introduced to avoid numerical difficulties in simulations). Thus, the
wheel friction torque is given as T = −Tc sign(ν + d). The average torque T̂ is
called the smooth image of the dry friction torque. The smooth image T̂ is defined
)o
t+t
as T̂ = t1o T dt; therefore one obtains
t

t+to t+to
1 1 2Tc ν
T̂ (ν) = T (ν)dt = − Tc sign(ν + d)dτ = − arcsin( ) (11.18)
to to π do
t t


The plot of T̂ with respect to ν do for Tc = 1 is shown in Fig. 11.16. One can
see from this figure that although the dry friction T is a discontinuous function of the
velocity, its smooth image T̂ is, by definition, a smooth function of the velocity. It
also follows from Fig. 11.16 that the smooth image exists only for the dither ampli-
tudes that extend the wheel velocity, that is, for do > ν. This is quite understandable
because for the dither amplitude smaller than the wheel velocity there is no change
in the friction torque.
Because the function T̂ is smooth it can be linearized for small velocity varia-
tions. Consequently, for small wheel velocity ν that is proportional to the velocity
of the command r, ν = kr r , one obtains

2Tc kr
T̂ (do ) ∼
= ko r, ko = − (11.19)
π do

The plot of linearized T̂ in Fig. 11.16, shown as a dashed line, reveals a good
coincidence with T̂ for ν < 0.5do .

@Seismicisolation
@Seismicisolation
11.2 Friction 173

0.8

0.6

0.4

0.2

0
T

–0.2

–0.4

–0.6

–0.8

–1
–1 –0.5 0 0.5 1
v/do

Fig. 11.16 Smooth image of dry friction (solid line) and the lineraized image (broken line)

Introducing equation (11.19) to the averaged equation (11.16) one obtains the
following linear system

ẋa = Axa + Br o r + Br d, ya = C xa (11.20)

with the input matrix Br o in the form Br o = Br + ko Bn . The latter equation proves
that the system dynamics with dither is linear one.
Notice that the dither input has no significant impact at the telescope pointing.
Let us write the pointing ya as a superposition of the pointing yar due to the input r
and the pointing yad due to the dither input d, that is, ya = yar + yad . Notice that the
dither is of high frequency; therefore, the response yad is negligible when compared
to yar ; thus ya ∼
= yar . The latter shows that the dither action makes the telescope
dynamics linear, but it does not show itself at the output; thus the telescope pointing
performance is not affected.

11.2.3 Non-linear Simulation Results

Typically, the dither signal should be injected just ahead of the non-linearity. In
the case of the Green Bank Telescope it is a velocity command at the telescope
azimuth drives. In this case the dither is simply added to the feed-forward command
generated by the controller computer.
The dither amplitude and frequency are determined as follows. The frequency
must be much higher than the telescope dynamics. Because no significant telescope

@Seismicisolation
@Seismicisolation
174 11 Non-Linear Control

Fig. 11.17 Pointing error for the simulated tracking with dither

dynamics is observed above 10 Hz, the dither frequency 30 Hz is chosen. The dither
amplitude depends on the level of friction torque. Dry friction torque for tracking at
a velocity of 0.0003 deg/s is smaller than 5600 Nm. For this friction level the dither
amplitude was selected to be 0.18 deg/s using the telescope pointing simulations
with various dither amplitudes and showing that the pointing error was the smallest
for the dither of amplitude 0.18 deg/s. The plot of the pointing error for the above
dither amplitude is shown in Fig. 11.17. The plot shows that the maximum pointing
error dropped 18-fold, from 1.4 to 0.08 mdeg.

11.3 Backlash
Gearboxes and gears are two critical components of the antenna drives. Measure-
ments at the drives of the DSN antennas indicated that the small gap between gear
teeth was causing backlash at the gearboxes and at elevation via the bull gear. A
backlash phenomenon is observed when one gear rotates through a small angle
without causing a corresponding movement in the gear it is driving. Left uncor-
rected, backlash deteriorates the antenna’s pointing precision. Backlash is observed
not only in antennas and telescopes [29, 30], but also in precision instruments [20–
23], and robots [24, 25]. Often, to reduce backlash impact on the closed-loop preci-
sion, anti-backlash controllers are designed [9, 13, 26].
For antennas and telescopes the backlash is eliminated by implementing two
identical drives that impose two non-identical torques [16]. The torque difference
is called a torque bias, or counter-torque. The difference between these two torques
depends on the antenna load and is shaped by the drive electronic circuits. With a
two-motor configuration (see Fig. 11.18), backlash clearance will occur at one drive
while the other is still coupled. The antenna dynamics will be controlled by the
latter drive. The effectiveness of the two-motor approach depends on the amount
of torque bias applied at the drives, which depends on the antenna load. It should
be large enough to lead the antenna through the gap for the maximal allowable

@Seismicisolation
@Seismicisolation
11.3 Backlash 175

(a)

Drive 1 Drive 2

(b)

Tachometer
Motor
Gearbox

Fig. 11.18 Azimuth drive of the 34-m antenna: (a) two-drive configuration, (b) close look at a
single drive

torque load, but small enough that it will not cause excessive local stress, friction and
wear.
High and steady loads do not need a torque bias because the backlash is observed
for low and reversing loads only. Time-varying loads, such as wind gusts, can
produce high torques that become very low within a short period of time, which
causes a backlash gap when the torque bias dynamics are too slow. Reversing loads
from wind gusting were observed and recorded at the DSN antennas.

11.3.1 Backlash and Its Prevention

Consider a simple gearbox with two gears rotating in opposite directions. Let the
angle of rotation of the first gear be β1 , the second gear be β2 , the gearbox ratio
be N, and the gearbox stiffness (measured at the second gear) be k. In this case the
relationship between the gearbox rotation and the torque T at the second gear is as
follows

@Seismicisolation
@Seismicisolation
176 11 Non-Linear Control
⎧
⎪
⎨0 for |Δβ| ≤ b
T = k(Δβ − b) for Δβ > b (11.21)
⎪
⎩
k(Δβ + b) for Δβ < −b


where Δβ = β2 − β1 N , and b is the size of the backlash gap measured at the
second gear. A plot of torque versus angle difference Δβ is shown in Fig. 11.19 for
b =1, and k = 20. It is clear from equation (11.21) that if the angle difference of two
gears is smaller than the backlash gap b, the gear motion is discontinuous and the
gear teeth impact each other.
Implementing a drive with two gears of torques T1 and T2 (that differ by the
amount of ) instead of a single drive will minimize the impact at the disconti-
nuity. This difference, called torque bias or counter-torque, will continually drive
the antenna even if one gear is in backlash. This is the principle of torque sharing.
The question remains as to how large the torque bias must be to prevent backlash.
If the stiffness of the gearbox is k and the backlash gap is b, the torque bias ΔT
should be greater than 2kb. But ΔT is a function of the load applied to the gears as
well: a bias is not required if torque load is high (T1 ∼
= T2  ΔT ) because the angle
difference is large and backlash is not observed (even when ΔT = 0). Plots of the
existing profile of motor torque vs. axial load (as percentage of the maximal load)
are shown in Fig. 11.20, for 10, 20, and 30 percent of the bias. The bias is shaped
such that it is the largest for the low loads, and phases out for higher loads.
In the event of dynamic loading, such as wind gusts, the appropriate magnitude of
the torque bias is not obvious. During dynamic loading, the torque difference deter-
mined for the steady-state case may not be large enough to prevent the backlash,
and assuming a higher counter-torque may lead to premature wear. Additionally,

200

150

100

50

0
T

–50

–100

–150

–200
–10 –5 0 5 10
Δβ

Fig. 11.19 The backlash function: relationship between gear angle difference and the torque

@Seismicisolation
@Seismicisolation
11.3 Backlash 177

0.8

0.6 30% bias

motor torque/max. motor torque

20% bias
0.4
10% bias
0.2

–0.2

–0.4

–0.6

–0.8

–1
–1 –0.8 –0.6 –0.4 –0.2 0 0.2 0.4 0.6 0.8 1
axis torque/max. axis torque

Fig. 11.20 Motor torque vs. axial torque, for 10% (dashed line), 20% (solid line), and 30% (dot-
dashed line) torque bias. Upper lines represent drive 1, lower lines represent drive 2

quickly varying loads with small steady components may cause backlash in both
gears simultaneously, despite the non-zero torque bias. Thus, the torque bias time
response is also an important design factor.

11.3.2 The Velocity Loop Model with Friction and Backlash

The Simulink model of the velocity loop system is shown in Fig. 11.21. The model
contains the antenna structure with the velocity input. The outputs are the encoder,
antenna velocity, and pinion velocity. The antenna structure model is obtained from
the finite-element model, as described in Chapter 2. The drive model which consists
of the velocity input, pinion velocity inputs, and torque output, is shown in Fig.
11.21b.
The drive consists of two motors (with gearboxes) and the torque share circuit.
The block diagram of the drive subsystem is shown in Fig. 11.21c. It consists of two
amplifiers, a motor armature, and a gearbox. The gearbox model includes non-linear
friction and backlash models; see Fig. 11.21d. The friction torque in this model
depends on the motor torque and motor speed, as described in equation (11.9). In
the backlash model the torque depends on the difference between the motor and
the pinion angle, as in equation (11.21). The torque share circuit is shown in Fig.
11.21b.

@Seismicisolation
@Seismicisolation
178 11 Non-Linear Control

(a)

Pinion velocity
y(n) = Cx(n) + Du(n)
Torque m
x(n + 1) = Ax(n) + Bu(n) 1
1 Velocity Comm.
velocity Antenna antenna
command Structure position
Drive

(b) Motor 1
Pinion Rate Torque

Command Voltage Current

Current 1
num(z)
1 2 Bias Voltage Fade Out –8 1
den(z)
Pinion Velocity Current 2 Fade Out Torque
velocity Comm. Filter1
Torque Bias Circuit

Command Voltage Current

Pinion Rate Torque

Motor 2

(c)
2
Current
Current
2 Command Voltage Amp 2 Voltage Amp 2 Voltage Current
Command Amp1 Voltage Amp 1 Voltage
Motor Speed Motor Torque
Voltage Amplifier 2 Motor Speed Motor Torque
Amplifier 1 Motor velocity
Motor Armature
1 Pinion Rate Torque 1
Pinion Motor and Gearbox Torque
Rate

(d) -K-

Gain 2

K Ts K Ts z
2 - K- - K- 2
z–1 z–1
Motor Torque
1/J Integrator 2 Backlash Gain 4
Torque1

_ friction _ u 1
Motor
velocity

1 - K-
Pinion
Rate Gain 3

Fig. 11.21 Simulink model of the antenna backlash: (a) velocity loop, (b) drive, (c) motor, gearbox,
and amplifiers (d) motor and gearbox with backlash

@Seismicisolation
@Seismicisolation
11.3 Backlash 179

The accuracy of the model was verified experimentally, showing significant accu-
racy of the simulations; see [16]. Open-loop tests were conducted at the 34-m
antenna to compare the measured antenna dynamics with the simulated dynamics of
the model that includes backlash and friction. The test data were used to determine
the amount of friction and the backlash angle. The velocity loop experiments were
conducted with a square-wave input of period 6.3 s and amplitude 0.013 deg/s. Two
tests were conducted: one with zero torque bias, and another with a torque bias of
15% of the maximal motor torque (the maximal torque is 308 kGm; thus torque bias
was set at 46 kGm).
With zero torque bias the simulated motor current (excited by the square wave)
and encoder reading are shown in Figs. 11.22 and 11.23, respectively. These plots
coincide with the field data in [16]. In addition, the motor currents, which are propor-
tional to the motor torques, were used to determine the frictional
torques.
The constant current in Fig. 11.22 measures ±1 A, and corresponds to the
constant antenna velocity because inertial forces are not present and the motor’s
effort is totally dedicated to overcoming the frictional forces. The 1 A current corre-
sponds to a 61 kGm motor torque, or 9.1 × 105 kGm axis torque, which is the
measure of frictional torque.
For 15% torque bias the plots of measured and simulated motor currents and
encoder readings are given in Figs. 11.24 and 11.25, respectively. This situation
differs from that of the zero torque bias case in that the encoder shows less chaotic

Fig. 11.22 Motor currents for zero torque bias

@Seismicisolation
@Seismicisolation
180 11 Non-Linear Control

Fig. 11.23 Encoder readings for zero torque bias

movement of the antenna (c.f. Figs. 11.23 and 11.25) and the motor torque plots
indicate the presence of the torque bias (their mean values are non-zero and have
opposite sign; c.f. Figs. 11.22 and 11.24).

Fig. 11.24 Motor currents for 15% torque bias

@Seismicisolation
@Seismicisolation
References 181

Fig. 11.25 Encoder readings for 15% torque bias

References
1. Armstrong-Helouvry B, Dupont P, Canudas de Wit C. (1994). Friction in Servo Machines:
Analysis and Control Methods. Appl. Mechanics Rev., 47(7): 275–305.
2. Armstrong-Helouvry B, Dupont, P, Canudas de Wit, C. (1994). A Survey of Models, Analysis
Tools and Compensation Methods for the Control of Machines with Friction. Automatica,
30(7): 1083–1138.
3. Bliman PAJ. (1992). Mathematical Study of the Dahl’s Friction Model. European J.
Mechanics, A/Solids, 11(6): 835–848.
4. Boddeke FR, VanVliet LJ, Young IT. (1997). Calibration of the Automated z-Axis of a Micro-
scope Using Focus Function. Journal of Microscopy, 186 (3).
5. Bridges MM, Dawson DM, Hu J. (1996). Adaptive Control for a Class of Direct Drive Robot
Manipulators. Int. J. Adaptive Control and Signal Processing, 10(4).
6. Cai L, Song G. (1994). Joint Stick-Slip Friction Compensation of Robot Manipulators by
Using Smooth Robust Controllers. J. Robotic Systems, 11(6): 451-469.
7. Canudas de Wit C, Olsson H, Astrom KJ. (1995). A New Model for Control of Systems with
Friction. IEEE Trans. on Aut. Control, 40(3): 419–425.
8. Dahl PR. (1976). Solid Friction Damping of Mechanical Vibrations. AIAA J., 14(12):
1675–1682.
9. Dhaouadi R, Kubo K, Tobise MI. (1994). Analysis and Compensation of Speed Drive Systems
with Torsional Loads. IEEE Trans. Industry Applications, 30(3): 760–766.
10. Dupont PE. (1994). Avoiding Stick-Slip Through PD Control. IEEE Trans. Aut. Control,
39(5): 1094–1097.
11. Dupont PE, Dunlap EP. (1995). Friction Modeling and PD Compensation at Very Low
Velocities. J. Dynamic System, Measurement, and Control, 117(1): 8–14.
12. Edwards C, Postlethwaite I. (1998). Anti-Windup and Bumpless-Transfer Schemes.
Automatica, 34(2): 199–210.
13. Friedland B, Davis L. (1997). Feedback Control of Systems with Parasitic Effects. Proc.
American Control Conference, Albuquerque, NM.

@Seismicisolation
@Seismicisolation
182 11 Non-Linear Control

14. Gawronski W. (1999). Command Preprocessor for the Beam-Waveguide Antennas. TMO
Progress Report, vol. 42-136. Available at http: //ipnpr.jpl.nasa.gov/progress˙report/42-136/
136A.pdf .
15. Gawronski W, Almassy W. (2002). Command Pre-Processor for Radiotelescopes and
Microwave Antennas. IEEE Antennas and Propagation Magazine, 44(2).
16. Gawronski W, Brandt JJ, Ahlstrom, Jr., HG et al. (2000). Torque Bias Profile for Improved
Tracking of the Deep Space Network Antennas. IEEE Antennas and Propagation Magazine,
42(6): 35–45.
17. Gawronski W, Parvin B. (1998). Radiotelescope Low Rate Tracking Using Dither. AIAA J.
Guidance, Control, and Dynamics, 21: 349–352.
18. Glattfelder AH, Schaufelberger W. (2003). Control Systems with Input and Output
Constraints, Springer, London.
19. Grimm G, Hatfield J, Postlethwaite I et al. (2001). Experimental Results in Optimal Linear
Anti-Windup Compensation. Proc. 40th IEEE Conf. on Decision and Control, Orlando, FL.
20. Hale LC, Slocum AH. (1994). Design of Anti-Backlash Transmission for Precision Position
Control Systems. Precision Engineering, 16(4).
21. Hippe P. (2006). Windup in Control, Its Effects and Their Prevention, Springer, London.
22. Ku SS, Larsen G, Cetinkunt S. (1998). Fast Tool Servo Control for Ultra-Precision Machining
at Extremely Low Feed Rates. Mechatronics, 8(4).
23. Lee S, Meerkov SM. (1983). Generalized Dither. International Journal of Control, 53(3):
741–747.
24. Mancini D, Brescia M, Cascote E et al. (1997). A Neural Variable Structure Controller for
Telescopes Pointing and Tracking Improvement. Proc. SPIE, vol. 3112.
25. Mancini D, Brescia M, Cascote E et al. (1997). A Variable Structure Control Law for Tele-
scopes Pointing and Tracking. Proc. SPIE, vol. 3086.
26. Mata-Jimenez MT, Brogliato B, Goswami A. (1997). On the Control of Mechanical Systems
with Dynamics Backlash, Proc. 36th Conf. Decision and Control, San Diego, CA.
27. Peng Y, Vrancic D, Hanus R. (1996). Anti-Windup, Bumpless, and Conditioned Transfer
Techniques for PID Controllers. IEEE Control Systems Magazine, August, 48–57.
28. Southward SC, Radcliffe CJ, MacCluer CR. (1991). Robust Non-linear Stick-Slip Friction
Compensation. J. Dynamic Systems, Measurement, and Control, 113: 639–645.
29. Stark AA, Chamberlin RA, Ingalls JG et al. (1997). Optical and Mechanical Design of the
Antarctic Submillimeter Telescope and Remote Observatory. Rev. Sci. Instrum., 68(5).
30. Tickoo AK, Koul R, Kaul SK et al (1999) Drive-Control System for the TACTIC gamma-ray
telescope. Experimental Astronomy, vol.9, no.2.
31. Trautt TA, Bayo E (1999) Inverse Dynamics of Flexible Manipulators with Coulomb Friction
or Backlash and Non-Zero Initial Conditions. Dynamics and Control, vol.9, no.2.
32. Tyler SR. (1994). A Trajectory Preprocessor for Antenna Pointing. TDA Progress Report,
42-118, pp. 139–159. Available at: http: //ipnpr.jpl.nasa.gov/progress˙report/42-118/118E.
pdf.
33. Yeh TJ, Pan YC. (2000). Modeling and Identification of Opto-mechanical Coupling and Back-
lash Non-linearity in Optical Disk Drives. IEEE Trans. Consumer Electronics, 46(1).

@Seismicisolation
@Seismicisolation
Chapter 12
RF Beam Control

This chapter discusses the detection and control of the position of the radio frequency
(RF) beam. The position of the beam is slightly different from the antenna position
as measured by the encoders. This happens due to structural deformations caused
either by loads (e.g., gravity, wind) or by the temperature gradient, atmospheric
distortion, or azimuth track unevenness. We also discuss how the beam position is
detected and controlled to minimize pointing error.
First, the RF controller to be used in the RF control is selected. Next, the
monopulse detection and control technique are analyzed, showing its performance
in its linear and nonlinear models. Finally, the scanning techniques, such as conical
scan, Lissajous, rosette, and sliding window conical scan are discussed.

12.1 Selecting the RF Beam Controller

In this section, based on [9], the RF beam position controller (RF controller, for
short) is selected. The RF feedback loop is closed over the antenna position loop
(see the block diagram in Fig. 12.1). The RF control system consists of the antenna
position loop, the RF controller, and the RF receiver. The receiver is either a conical
scan RF detector or a monopulse receiver, discussed later.
Two RF control systems are tuned and evaluated: the first one with the PI
controller in the position loop, and the second one with the LQG controller in the
position loop.
In the tuning process the following conditions were assumed:
1. Azimuth and elevation control loops are independent.
2. Noises in elevation and cross-elevation channels are independent.
The properties of the RF loop depend on the properties of the position loop.
The magnitudes and phases of the position-loop transfer functions are shown in
Figs. 12.2 and 12.3 (solid lines). The position loop is tuned such that its transfer
function for low frequencies (up to the frequency denoted f o ) is approximately equal
to 1. The frequency f o is a bandwidth, and for the PI controller f o = 0.1Hz, and
for the LQG controller f o = 1.0 Hz. For frequencies higher than f o the position
loop transfer function rolls-off, although it contains resonance peaks that reflect

W. Gawronski, Modeling and Control of Antennas and Telescopes, 183

@Seismicisolation
@Seismicisolation
DOI: 10.1007/978-0-387-78793-0 12, C Springer Science+Business Media, LLC 2008
184 12 RF Beam Control

command POSITION-LOOP

RF eb RF beam cc + + Position Velocity

beam controller loop controller loop
–
encoder
RF beam

eb = beam error
cc = command correction

Fig. 12.1 RF beam controller

(a)

100
magnitude

actual
10–2
simplified

10–2 10–1 100 101

frequency, Hz
(b)
0

–100
phase, deg

–200

actual
–300 simplified

10–2 10–1 100 101

frequency, Hz

Fig. 12.2 The transfer function of the PI position control loop: (a) magnitude and (b) phase

the antenna flexible deformations. For the RF controller tuning purposes the peaks
are ignored, and the high-frequency part of the position-loop transfer function is
approximated with the slope of –20 dB/dec. Therefore, for the RF controller tuning
purposes the position loop is approximated with the first-order transfer function:
1
G(s) = (12.1)
1 + Ts

It has a unit gain and the time constant reciprocal to the bandwidth, T = 2π1fo .
The time constant is different for the PI and LQG controllers, namely, T =1.592 s

@Seismicisolation
@Seismicisolation
12.1 Selecting the RF Beam Controller 185

(a)

100
magnitude

10–2 actual
simplified

10–2 10–1 100 101

frequency, Hz
(b)
0

–100
phase, deg

–200

actual
–300
simplified

10–2 10–1 100 101

frequency, Hz

Fig. 12.3 The transfer function of the LQG position control loop: (a) magnitude and (b) phase

for the PI system, and T = 0.159 s for the LQG system. The magnitudes and phases
of the transfer functions of the simplified models are shown in Figs. 12.2 and 12.3
(dashed lines).
Based on the above model the RF controller is tuned. It includes the deter-
mination of its transfer function F(s) and the tracking and disturbance rejection
properties. The block diagram of the simplified RF control system is presented in
Fig. 12.4. In this diagram, α denotes the target location, y is the beam position, e is
beam error, and r is the command (or predict). Denote the transfer function from α
to y by H, that from r to y by Hr , and the transfer function of the position loop by
G. From Fig. 12.4 one obtains

command

+ y
+ e +
F(s) G(s)
–

Fig. 12.4 A simplified block diagram of the RF control system

@Seismicisolation
@Seismicisolation
186 12 RF Beam Control

y(s) GF y(s) G
H (s) = = and Hr (s) = = (12.2)
α(s) 1 + GF r (s) 1 + GF

Using the plant transfer function G(s) as in equation (12.1), one arrives at

F(s) 1
H (s) = and Hr (s) = (12.3)
1 + T s + F(s) 1 + T s + F(s)

The controller transfer function F(s) is determined by shaping the tracking prop-
erties of the RF system, H (s). For good tracking properties it is required that the
magnitude of the transfer function within the bandwidth 0 ≤ ω ≤ 1/T be equal to
1, that is,

1
|H (ω)| = 1, for 0 ≤ ω ≤ (12.4)
T

From equation (12.3), it follows that the above condition is satisfied for

1
F  1 for 0 ≤ ω ≤ (12.5)
T

On the other hand, outside the bandwidth, for ω > 1/T it is required that

|H (ω)| → 0 for ω > 1/T. (12.6)

From equation (12.3) it follows that the above condition is satisfied for

|F(ω)| → 0, for ω > 1/T (12.7)

Finally, for the reasonable stability margin, the roll-off rate of F at the crossover
frequency should be −20 dB/dec; see [5, p. 172].
It is easy to see that the transfer function of an integrator

k
F(s) = (12.8)
s

satisfies all the above conditions. Thus, an integrator is chosen as a RF controller.

Its gain k was chosen to obtain acceptable tracking properties. Namely, k = 0.75
was determined for the PI position loop system, and k = 1.0 for the LQG position
loop system. The selection was backed with extensive simulations of the tracking
properties of the closed-loop system.

@Seismicisolation
@Seismicisolation
12.2 Monopulse 187

12.2 Monopulse
The principle of the monopulse technique is simple. The feedhorns of the monopulse
tracker are slightly displaced so that each receives the signal from a slightly different
position, that is, at slightly different power. The received power of the opposite
horns is added, to form a sum beam, and subtracted, to form a difference beam.
The difference beam characterizes the pointing error. If the difference beam is zero,
the antenna is at the target. If the difference beam is nonzero, the pointing error is
generated. The difference beam is linear near the origin; thus the monopulse sensor
is linear for small antenna deviations.
The PI and LQG controllers in the position loop are analyzed and compared. The
selection of the monopulse controller is presented, along with the analysis of the
monopulse tracking errors in a noisy environment. The simulations tried to reflect
the real antenna environment. For example, the antenna model was derived from
the field test data using system identification test; encoder error disturbances, servo
noise, and wind gusts were obtained from the field data; and received noise was
obtained by propagating white noise from the antenna input to the output of the
digital receiver.

12.2.1 Command following

Introducing equation (12.8) to equation (12.2), one obtains the closed-loop transfer
function H(s) of the second order:

ωo2
H (s) = (12.9)
s 2 + 2ς ωo s + ωo2
√
where ωo = k/T and ζ = 2√1kT . The monopulse closed-loop parameters are
compared in Table 12.1.
The plots of the simplified (dashed line) and full-order (solid line) transfer func-
tions of the monopulse closed-loop system are shown in Fig. 12.5 (with the PI
position controller) and in Fig. 12.6 (with the LQG position controller), showing
good coincidence. Note that the monopulse system with the PI controller is narrow-
banded ( fo = 0.11 Hz) and underdamped, (the damping ratio is smaller than the
critical one, ζ = 0.46 < ζcritical , where ζcritical = 0.71), while the bandwidth of the
monopulse system with LQG controller is wider (fo = 0.40 Hz) and is overdamped,
(the damping ratio is larger than the critical one, ζ = 1.25 > ζcritical ). Thus, the
monopulse system with the PI position loop exhibits overshoot and longer settling

Table 12.1 Monopulse closed-loop parameters

Controller T [s] k[1/s] ωo [rad/s] fo [Hz] ␨[–]
PI 1.592 0.75 0.69 0.11 0.46
LQG 0.159 1.0 2.51 0.40 1.25

@Seismicisolation
@Seismicisolation
188 12 RF Beam Control

(a)

100
magnitude

actual
10–2
simplified

10–3 10–2 10–1 100 101

frequency, Hz
(b)
0
phase, deg

–100

–200
actual
simplified
–300
10–3 10–2 10–1 100 101
frequency, Hz

Fig. 12.5 The transfer function of the monopulse control system with the PI position control loop:
(a) magnitude and (b) phase.

time in the step response, while the monopulse system with the LQG position loop
has no overshoot, and shorter settling time.
Within the bandwidth the simplified and the full-order systems show good coin-
cidence in terms of the properties used in the controller tuning and stability analysis.
However, only the full-order system can give reliable error estimates of the pointing
errors with the required precision.

12.2.2 Disturbance Rejection Properties

The transfer function Hr (s) from the command to the encoder describes the system
disturbance rejection properties and can be obtained by introducing equation (12.8)
to the second transfer function in equation (12.3):

2ζ ωo s
Hr (s) = (12.10)
s 2 + 2ς ωo s + ωo2

The plots of the transfer function from the wind gusts input to the encoder output
are given in Fig. 12.7: the solid line for the PI position loop and the dashed line for

@Seismicisolation
@Seismicisolation
12.2 Monopulse 189

(a)

100
magnitude

10–2 actual
simplified

10–3 10–2 10–1 100 101

frequency, Hz
(b)
0
phase, deg

–100

–200
actual
simplified
–300
10–3 10–2 10–1 100 101
frequency, Hz

Fig. 12.6 The transfer function of the monopulse control system with the LQG position control
loop: (a) magnitude and (b) phase

101

100
Hw magnitude

10–1

10–2 LQG
PI
10–3 –3
10 10–2 10–1 100 101
frequency, Hz

Fig. 12.7 A comparison of the wind disturbance transfer functions of the monopulse control
systems with the PI position control loop and LQG position control loop

the LQG position loop. It is easy to see that the monopulse system with the LQG
position loop has wind-disturbance rejection properties of an order of magnitude
better than has the monopulse with the PI position loop. This will be illustrated later
with the results of the wind gusts simulations.

@Seismicisolation
@Seismicisolation
190 12 RF Beam Control

12.2.3 Stability Due to the Gain Variation

In practice, the monopulse gain, k, is slowly but significantly varying. Therefore, it
is important to check the maximal gain for which the system is stable. The critical
gains for which the system is still stable were obtained from simulations and are
given in Table 12.2. For the PI position loop, the critical gains are 8.3 in azimuth
and 16.7 in elevation. Because the nominal value of the gain for the PI controller
is k = 0.75, the stability margin is large enough. For the LQG position loop the
critical gains are 5.9 in azimuth and 7.1 in elevation. Because the nominal value of
the gain for the LQG controller is k = 1.0, this is also an acceptable stability margin.
However, one should notice that the LQG controller is the model-based one and
that the velocity loop model varies slightly with the antenna elevation angle. These
variations impact mainly the azimuth loop, and the stability margins for this loop
can be slightly lower than those obtained from the simulations. The performance
under the varied gain is discussed in the next section.

12.2.4 Performance Simulations: Linear Model

The full-order model of the antenna obtained from the system identification (rather
than the simplified one), is used to evaluate the monopulse closed-loop performance.
The following disturbances were applied in the simulations:

1. The 25 km/h wind gusts. The wind disturbance model was taken from the field
measured data, see [7] and Chapter 5.
2. Servo noise, generated by the white noise input. This noise generates encoder
jitters of standard deviation about 0.1 mdeg
3. Azimuth encoder error consists of the radial run-out error, jitters, and rapid
changes due to the gaps between the encoder segments.
4. Monopulse receiver noise. It is a white noise of 1 mdeg standard deviation.
The results of pointing simulations due to all of the above-mentioned distur-
bances are given in Fig. 12.8 (solid line for the LQG and dashed line for the PI).
They are obtained for the nominal monopulse gains (k = knom , where knom = 0.75
for the PI controller and knom = 1.0 for the LQG controller). The simulations also
were performed for the varying gain, k = αknom , i.e. for the reduced gain (α = 0.5),
and for increased gains (α = 2.0 and α = 4.0). The figures show that for the
monopulse system with the PI controller the total error is 1.23 mdeg for the nominal
gain. The bulk of this error is due to wind gusts. However, the LQG controller

Table 12.2 Critical gains

Axis PI LQG
Azimuth 8.3 5.9
Elevation 16.7 7.1

@Seismicisolation
@Seismicisolation
12.2 Monopulse 191

(a) (b)
total rms error, mdeg

total rms error, mdeg

1 1

0.5 0.5

0 0
0 1 2 3 4 0 1 2 3 4
gain, multiple of nominal gain, multiple of nominal

Fig. 12.8 Monopulse system error under total disturbances: (a) azimuth, and (b) elevation

suppresses the wind disturbances significantly, with the rms error below 0.2 mdeg
for the nominal gain (see Fig. 12.8).

12.2.5 Performance Simulations: Nonlinear Model

The monopulse pointing-error model combines an antenna, feed, and low-noise

amplifier. This model is a non-linear one, and the elevation and cross-elevation
errors are not independent. The non-linear monopulse pointing-error detector is
incorporated into the control system model. The combined control system model
that includes both axes and the non-linear monopulse model is shown in Fig. 12.9(a).
It consists of the antenna velocity loop model, the position controller, the monopulse
controller with variable gain k, the monopulse sum g(θ, φ) and difference h(θ, φ)
functions, the monopulse processing block (BVR), and the conversion block of
elevation and cross-elevation errors into elevation and azimuth errors (θ and φ are
magnitude and phase of the error, respectively). This model includes the following
signals: encoder errors, commands, servo noises, wind disturbances, and main and
error signal noises. The finite null depth of the detector function was modeled by
adding a coherent signal to the difference pattern h(θ, φ). The imbalance between
the main and the error channels was modeled by adding phase error ⌬φcalib . BVR
noises (noises in error detection) ηel , ηxel , and the receiver demodulation noise ηφ
were added to the BVR monopulse processing block.
The block diagram of the monopulse pointing error detector is shown in
Fig. 12.9(b). This nonlinear device transforms the azimuth and elevation pointing
errors into the elevation and cross-elevation pointing errors. It consists of the
converter of azimuth and elevation angles into (θ and φ parameters, √ nonlinear func-
tions g(θ, φ) and h(θ, φ), and other nonlinear functions (such as (.), sin(.), cos(.)).
The detector functions g(θ, φ) and h(θ, φ) are shown in Fig. 12.10a,b for the 15
dB signal-to-noise-ratio (SNR). They depend on the size of the null-depth. Addi-
tionally, the detector model consists of disturbances and noises that represent its

@Seismicisolation
@Seismicisolation
192 12 RF Beam Control

(a) Az Encoder Error

enc_err2

y
y(n) = Cx(n) + Du(n)
r AZ correction AZ err
x(n + 1) = Ax(n) + Bu(n) wind_az
AZ Position,
AZ Velocity Loop AZ Wind Controller AZ monopulse
(Antenna) controller
AZ Noise servo_na

azerr
conversion
fromXEL to AZ
AZ (deg) .001 g_off
10 deg_el factor
10 deg_el xel_off

1 R1

25 th_calib Monopulse Pointing

Error Detector
1 bl_bvr

50 pc_n0

1/25 T_update el_off

EL(deg) 50 gamma

elerr

EL Noise servo_ne

EL Velocity Loop EL monopulse

EL Wind EL Position,
(Antenna) controller
Controller
y(n) = Cx(n) + Du(n) wind_el
r EL correction EL err
x(n + 1) = Ax(n) + Bu(n)
y

enc_err2

EL Encoder Error
(b)
nxel 5

-1
1
xel_off
1 az theta theta g sqrt -K-
azerr
R/delta
2
6 el phi phi h cos el_off
elerr
(AZ,EL) to (theta,phi) (g,h)

sin 2 nel

dphi_calib 3 4 nphic

Fig. 12.9 Block diagrams of (a) the nonlinear monopulse control model and (b) the monopulse
pointing model detector

imperfections or environmental interactions: the bore sight shift, ⌬φcalib ; the BVR
noises, ηel and ηxel ; and the receiver demodulation noise, ηφ .
The detector couples azimuth and elevation errors; therefore the errors cannot
be determined separately. As a consequence, the nonlinear control system model

@Seismicisolation
@Seismicisolation
12.2 Monopulse 193

(a) (b)
102 200

100
100

h,deg
0
g

10–2
–100

10–4 –200
–0.02 –0.01 0 0.01 0.02 –0.02 –0.01 0 0.01 0.02
θ, deg θ, deg

Fig. 12.10 The detector functions for an SNR of 15 dB: (a) g and (b) h

combines azimuth and elevation loops together, as in Fig. 12.9(a). This model
includes, besides the detector, converters from the cross-elevation, and elevation
coordinates to azimuth, elevation coordinates, time delays (of 0.1 s), variable loop
gains kaz and kel , the monopulse controllers, position controller (LQG), and azimuth
and elevation velocity loop models. This closed-loop system is subject to wind
disturbances, servo noises, encoder errors, and is driven by the azimuth and eleva-
tion commands.
Extensive simulations of the non-linear closed-loop system that depend on the
above-listed parameters have been performed. The results are summarized in
Fig. 12.11, showing the elevation and cross-elevation errors for the 5 mdeg step
offset in elevation and cross-elevation. The target of 5 mdeg in elevation and 5 mdeg
in cross-elevation is acquired. The error depends on the SNR. For a high SNR of 40
dB the radius of the circling of the destination point of (5, 5) mdeg is small, about
0.5 mdeg. For an SNR of 12 dB it increases about 5 mdeg.

10 10
XEL error, mdeg

XEL error, mdeg

5 5

0 0
0 5 10 0 5 10
EL error, mdeg EL error, mdeg

Fig. 12.11 Elevation and cross-elevation errors for the 5-mdeg step offset in elevation and cross-
elevation, and for an SNR of: (a) 40 dB, (b) 12 dB

@Seismicisolation
@Seismicisolation
194 12 RF Beam Control

12.3 Scanning
The spacecraft trajectory (its position versus time) is typically known with high
accuracy, and this trajectory is programmed into the antenna, forming the antenna
command. However, due to environmental disturbances such as temperature gradient,
wind and gravity forces, and manufacturing imperfections, the antenna does not
point precisely towards the spacecraft. Scanning techniques are used to determine
the spacecraft position.
While tracking, antenna besides following the target performs scanning move-
ments. Scanning movements of an antenna are small harmonic axial movements
added to the antenna trajectory. They are used to estimate the true spacecraft posi-
tion, because the scanning motion produces power variations of the received signal,
which are used to estimate the spacecraft position. Three different scanning patterns
(conical scan, Lissajous scan, and rosette scan) are presented and analyzed in this
chapter. The analysis includes the evaluation of the estimation errors due to random
or harmonic variation of the antenna position and due to random and harmonic
variations of the power level. Typically, the estimation of the spacecraft position is
carried out after completing a full scanning cycle. However, it could be done more
frequently using the sliding window scanning, where the spacecraft position estima-
tion is carried out in an almost-continuous manner. The sliding window scanning is
analyzed, showing that it reduces estimation time by half.

12.3.1 Conical Scan

A technique commonly used for the determination of the true spacecraft position
is the conical scanning method (conscan). During conscan, circular movements are
added to the antenna command as shown in Fig. 12.12. These circular movements
cause sinusoidal variations in the power of the signal received from the space-
craft by the antenna, and these variations are used to estimate the true spacecraft
position.
The radius of the conscan movement is typically chosen such that the loss of the
signal power is of 0.1 dBi (dB power relative to isotropic source). Thus, it depends
on the frequency of the receiving signal. For Ka-band signals (32 GHz), the conscan
radius is 1.55 mdeg. Depending on the radius, sampling rate, antenna tracking capa-
bilities, and desired accuracy, the period of the conscan typically varies from 30 to
120 s (in our case it is 60 s). Finally, the sampling frequency was chosen as 50 Hz
to satisfy the Nyquist criterion, which says that the sampling rate shall be at least
twice the antenna bandwidth (of 10 Hz).
The conscan technique is used for antenna and radar tracking; see [2, 4]. For
spacecraft applications it is described in [1, 6, 11]. For missile tracking is described
in [10]. Papers [1] and [5] present the least square and Kalman filter techniques,
respectively. In [3] the nonlinear estimation techniques were used to estimate space-
craft position using conscan, and in [8] there are analyses of different pattern of
scanning.

@Seismicisolation
@Seismicisolation
12.3 Scanning 195

elevation

trajectory
trajectory with conscan

azimuth

Fig. 12.12 Schematic diagram of spacecraft trajectory and antenna conscan

Power variation during conscan. Define a coordinate system with its origin
located at the antenna command position (i.e., translating with the antenna
command). The coordinate system consists of two components: the elevation rota-
tion of the dish and the cross-elevation rotation of the dish. The first component is
defined as a rotation with respect to a horizontal axis orthogonal to the boresight,
and to the second as a rotation with respect to a vertical axis orthogonal to the
boresight and the elevation axis (see Fig. 12.13). Because the spacecraft position in
this coordinate system is measured with respect to the antenna boresight, and the
spacecraft trajectory is accurately known, its position deviations are predominantly

XEL

boresight

EL

Fig. 12.13 Elevation and cross-elevation coordinate system

@Seismicisolation
@Seismicisolation
196 12 RF Beam Control

caused by either unpredictable disturbances acting on antenna (e.g., wind pressure)

or by antenna deformations (such as thermal deformations, and un-modeled atmo-
spheric phenomena such as refraction). During conical scanning the antenna moves
in a circle of radius r, with its center located at the antenna command position.
The conscan data are sampled with frequency of 50 Hz; thus the sampling time
is ⌬t = 0.02s. The antenna position ai at time ti = i⌬t consists of the elevation
component, aei , and the cross-elevation component, axi , as shown in Fig. 12.14. The
position is described by the following equation
   
aeli r cos ωti
ai = = (12.11)
axeli r sin ωti
Plots of aei and axi for conscan radius r = 1.55 mdeg, and conscan period T =
60 s, ω = 2π /T is conscan frequency (ω = 0.1047 rad/s) are shown in Fig. 12.14.
Also,
T = n⌬t (12.12)

where n is the number of samples per 1 circle, n = 3000.

The target position is denoted si , with the elevation and cross-elevation compo-
nents sei and sxi
 
s
si = ei (12.13)
sxi

where sei is the elevation component of the target position at i⌬t, and sxi is the
cross-elevation component of target position at i⌬t (Fig. 12.15),
The antenna position error is defined as the difference between the target position
and the antenna position, that is,
ei = si − ai (12.14)

1.5
XEL
1
EL
EL and XEL, mdeg

0.5

–0.5

–1

–1.5
0 10 20 30 40 50 60
time, s

Fig. 12.14 Elevation and cross-elevation components of conscan

@Seismicisolation
@Seismicisolation
12.3 Scanning 197

XEL
axi ai

si trajectory
sxi
sˆxi sˆi

0 aei sei sˆei EL

Fig. 12.15 Antenna position, target position, and estimated target position during conscan

Like the antenna position, it has two components, elevation and cross-elevation
position errors:
   
eei sei − aei
ei = = (12.15)
exi sxi − axi

The total error at ti = i⌬t is described as the position rms error, that is,

εi = eiT ei = ei 2 (12.16)

Combining (12.11), (12.14), and (12.16), one obtains

εi2 = siT si − 2aiT si + aiT ai = siT si − 2aiT si + r 2 (12.17)

Next, we describe how the error impacts the beam power. The carrier power pi
is a function of the error εi , and its Gaussian approximation is expressed as
μ 2
pi = poi exp(− ε ) + vi (12.18)
h2 i

where poi is the maximum carrier power, v i is the signal noise, h is the half-power
beamwidth (17 mdeg for Ka-band), and μ = 4 ln(2) = 2.7726. Note that although a
spacecraft can move relatively quickly with respect to fixed coordinates, it typically
moves slowly in the selected coordinate frame (this relative movement is caused
by slowly varying disturbances). For example, thermal deformations have period
of several hours, while conscan period is only one minute. Therefore, it is safe to
assume that the target position is constant during the conscan period, that is, si ∼
= s.
It is also safe to assume that the power is constant during the conscan period, that
is, poi = po .

@Seismicisolation
@Seismicisolation
198 12 RF Beam Control

Using the approximation exp(x) ∼

= 1 + x one obtains equation (12.18) as follows
μ 2
pi = po (1 − ε ) (12.19)
h2 i
Substituting (12.17) into the above equation, and assuming si = s, poi = po , one
obtains
po μ 2 2 po μ T
pi = po − (r + s T s − 2aiT s) + v i = pm + a s + vi
h2 h2 i
where pm is mean power. Using equation (12.11) one arrives at
2 po μr
pi = pm + (se cos ωti + sx sin ωti ) + v i (12.20)
h2
In the above equations pm is the mean power, defined as
 μ 
pm = po 1 − 2 (r 2 + s T s) (12.21)
h
(see [6]). The plot of pi is shown in Figs. 12.16 and 12.17. Additionally, in these
figures the plot of the power as a function of the antenna position is marked. These

antenna off target

1 antenna on target

0.8

0.6

0.4

0.2

20
10 20
0 10
cross-elevation, mdeg 0
–10
–10 elevation, mdeg
–20 –20

Fig. 12.16 Carrier power and the conscan power for perfectly pointed antenna and for elevation
error

@Seismicisolation
@Seismicisolation
12.3 Scanning 199

(a)
1

0.98
power/powermax

0.96

0.94
rosette conscan Lissajous
0.92

0.9
0 10 20 30 40 50 60
time, s
(b)
1

0.98
power/powermax

0.96

0.94
rosette Lissajous conscan
0.92

0.9
0 10 20 30 40 50 60
time, s

Fig. 12.17 A power variation (with respect to maximum power) for the conscan, Lissajous, and
rosette scans: (a) antenna on target, and (b) 0.7 mdeg elevation error

plots are presented for the case of the antenna perfectly pointed at the target, as
well as for the case of an error in antenna elevation position. It can be seen that for
the perfectly pointed antenna the received power is constant and smaller than the
maximum power. For a mis-pointed antenna the received power varies in sinusoidal
fashion, as it is derived in the following paragraph.
The algorithm presented above is a corrected Alvarez algorithm [1]. Here the
Taylor expansion was taken with respect to the error εi , which produces the
maximum power rather than mean power in the second component in equation
(12.20). Denoting the variation from mean power as dpi = pi − pm one obtains
from equation (12.20)

dpi = gse cos ωti + gsx sin ωti + v i (12.22)

where

2 po μr
g= (12.23)
h2

In this equation, g, and ω are known parameters, the power variation dpi is
measured, and the spacecraft coordinates se , sx are to be determined. If no noise

@Seismicisolation
@Seismicisolation
200 12 RF Beam Control

were present, the spacecraft position could be obtained from the amplitude and
phase of the power variation. Because the received power signal is noisy, the least-
square technique is applied.
Estimating spacecraft position from the power measurements. Denoting

ki = g cos ωti sin ωti (12.24)

equation (12.22) can be written as

dpi = ki s + v i (12.25)
 
se
where s = . For an entire conscan circle/period,
sx
⎧ ⎫ ⎧ ⎫ ⎧ ⎫
⎪
⎪ dp1 ⎪⎪ ⎪
⎪ k1 ⎪
⎪ ⎪
⎪ v1 ⎪⎪
⎪
⎨dp2 ⎪ ⎬ ⎪
⎨k2 ⎪⎬ ⎪
⎨v 2 ⎪⎬
dP = .. , K = . , V = . , (12.26)
⎪
⎪ ⎪ ⎪ .⎪ ⎪ .. ⎪
⎪ . ⎪
⎩ ⎪
⎭
⎪
⎪.⎪
⎩ ⎪
⎭
⎪
⎪
⎩ ⎭
⎪
⎪
dpn kn vn

and equation (12.25) is obtained in the following form

dP = Ks + V (12.27)

The estimated spacecraft position ŝ is the least-square solution of the above equa-
tion:

ŝ = K + d P (12.28)

where K + = (K T K )−1 K T .

12.3.2 Sliding Window Conscan

The spacecraft position estimation technique described up to this point has used data
collected during the whole scanning period. Thus, the spacecraft position estimate
is updated every period T, which typically ranges from 30 to 120 s. This is a rather
slow update, causing a significant lag in antenna tracking if the assumption of slowly
varying target position is incorrect. This lag can be improved using a technique
known as sliding window scanning. In sliding window scanning, spacecraft position
is estimated every time period ⌬T, where ⌬T < T . The update moments are shown
in Fig. 12.18 for ⌬T = 13 T . In this case the data used in equation (12.27) do not start
at the beginning of every circle; rather, they begin at times T, T + ⌬T , T + 2⌬T ,

@Seismicisolation
@Seismicisolation
12.3 Scanning 201

(a)
0 T 2T 3T
… time

(b)
0 T 2T
… time

Fig. 12.18 Scans: (a) regular and (b) sliding window, for ⌬T = 13 T

T + 3⌬T , etc. Note that the first estimation is at T rather than ⌬T because an entire
circle is required to begin the estimation process.
To see the usefulness of the sliding window technique in the estimation process,
assume that the target position changes rapidly by 0.15 mdeg at t = 150 s. This type
of shift may be caused by a sudden disturbance in antenna position, such as a large
gust of wind. This is an extreme situation, because in the closed-loop configuration
the control system would “soften” the impact of the gusts. The shift is illustrated in
Fig. 12.19, along with simulated responses of antennas using the traditional conscan
method and the sliding window method. The simulations show that for the conscan
period T = 60 s, the time of 120 s is required to reach the target, whereas the sliding
window conscan with ⌬T = 5 s reaches the target in half the time, that is, in 60 s.
This is especially important when antenna dynamics is involved, because faster
sensors improve the pointing accuracy.

12.3.3 Lissajous Scan

In the Lissajous scanning pattern, the antenna position ai at time ti = i⌬t consists
of the elevation component, aei , and the cross-elevation component, axi , described
by the following equation
   
aei r sin nωti
ai = = (12.29)
axi r sin mωti

where n and m are natural numbers. Again, the components are harmonic func-
tions, which are most desirable for the antenna motion because they do not result in
jerks or rapid motions. The Lissajous curve for r = 1.55 mdeg, n = 3, and m = 4 is
shown in Fig. 12.20, and the individual components’ plots (aei and axi ) are shown in
Fig. 12.21. The radius was chosen such that the mean power loss was equal to the
mean power loss resulting from a conscan sweep with r = 1.55 mdeg.
For this scanning pattern, the power variation is obtained as follows. With the
antenna position error defined as in equation (12.14), the total error at ti = i⌬t is
described as the position rms error:

@Seismicisolation
@Seismicisolation
202 12 RF Beam Control

0.15

target
elevation mdeg

position sliding
0.1 conscan
conscan

0.05

0
0 20 40 60 80 100 120 140 160
time, s

0.15
cross-elevation mdeg

target
position sliding
0.1
conscan
conscan
0.05

0
0 20 40 60 80 100 120 140 160
time, s

Fig. 12.19 Estimated spacecraft position for regular conscan and sliding window conscan

εi2 = siT si − 2aiT si + aiT ai (12.30)

Using the carrier power pi as in equation (12.19), and assuming that the space-
craft position and the carrier power are constant during conscan period, that is, that
si = s and poi = po , one obtains

po μ T
pi = po − (a ai + s T s − 2aiT s) + v i
h2 i
2μ
= pm exp ( 2 aiT s ) + v i (12.31)
h
2μr
= pm exp ( 2 (se sin nωti + sx sin mωti ) ) + v i
h
Using equation (12.29) one obtains

2 po μr
pi = pm + (se sin nωti + sx sin mωti ) + v i (12.32)
h2

@Seismicisolation
@Seismicisolation
12.3 Scanning 203

1.5

1
cross-elevation, mdeg

0.5

–0.5

–1

–1.5
Lissajous conscan
courve circle
–2
–2 –1.5 –1 –0.5 0 0.5 1 1.5 2
elevation, mdeg

Fig. 12.20 Lissajous curve for R = 1.55 mdeg, n = 3, and m = 4

2
XEL
1.5
EL
1
EL and XEL, mdeg

0.5

–0.5

–1

–1.5

–2
0 10 20 30 40 50 60
time, s

Fig. 12.21 Elevation and cross elevation components of Lissajous scanning pattern for n=3, m=4

where pm is the mean power defined as

μ T
pm = po ( 1 − (a ai + s T s) ) (12.33)
h2 i
Denoting the variation from mean power as dpi = pi − pm one obtains the power
variation as a function of the spacecraft position, se and sx

@Seismicisolation
@Seismicisolation
204 12 RF Beam Control

dpi = gse sin nωti + gsx sin mωti + v i (12.34)

In this equation g and ω are known parameters, dpi is measured, and se and sx
are spacecraft coordinates to be determined. The plot of the power variation dpi
is shown in Fig. 12.17. It is seen from this figure that power variation occurs even
when the antenna is perfectly pointed, unlike the conscan situation. The Lissajous
radius, however, was chosen such that in average the loss of the received power is
the same as during conscan. Also, unlike the conscan situation, the maximum power
is reached during the cycle because the Lissajous curve crosses through the origin.
As in the conscan case, the spacecraft position estimate is obtained from equation
(12.28). In the Lissajous case, the matrix K is changed such that its ith row is

ki = g sin nωti sin mωti (12.35)

12.3.4 Rosette Scan

The rosette scanning is used in missile tracking (see [10]) and in telescope infrared
tracking (see [12]). In the rosette scanning, the antenna movement is described by
the following equation
   
aei r cos nωti + r cos mωti
ai = = (12.36)
axi r sin nωti − r sin mωti

The plots of the elevation component aei , and the cross-elevation component, axi
are shown in Fig. 12.22 for the rosette curve of radius r = 1.10 mdeg, n = 1, and
m = 3. The rosette curve itself is shown in Fig. 12.23 (the radius was chosen such
that the mean power loss is the same as for the conscan with r = 1.55 mdeg). Note
that the rosette curve, unlike the conscan circle, crosses the origin and thus receives
peak power if the target and boresight position coincide.
Following the derivation as above, the antenna power is obtained as follows

2 po μr
pi = pm + (se (cos nωti + cos mωti ) + sx (sin nωti − sin mωti )) + v i
h2
(12.37)
therefore the power variation is

dpi = gse (cos nωti + cos mωti ) + gsx (sin nωti − sin mωti ) + v i (12.38)

where g is given by equation (12.23). The plots of variations of dpi are shown in
Fig. 12.17. The plots show that received power varies for the perfectly pointed

@Seismicisolation
@Seismicisolation
12.3 Scanning 205

2
XEL
EL
1
EL and XEL, mdeg

–1

–2
0 10 20 30 40 50 60
time, s

Fig. 12.22 Rosette elevation and cross-elevation components

Fig. 12.23 Rosette curve for 2

r = 1.1 mdeg, n = 1, and m =
3
1.5

1
cross-elevation, mdeg

0.5

–0.5

–1
conscan
circle
–1.5 rosette
courve
–2
–2 –1.5 –1 –0.5 0 0.5 1 1.5 2
elevation, mdeg

antenna. Also, like the Lissajous case, the maximum power is reached during the
cycle, because the rosette curve crosses through the origin.
The spacecraft position estimate is determined from the above equation, using
equation (12.28), where the ith row of the matrix K in this equation is

ki = g cos nωti + cos mωti sin nωti − sin mωti (12.39)

@Seismicisolation
@Seismicisolation
206 12 RF Beam Control

12.3.5 Performance Evaluation

The scanning performance is evaluated in the presence of disturbances. First,
the antenna position is disturbed by random factors, such as wind gusts, and by
deterministic disturbances, which can be decomposed into harmonic components.
Secondly, the carrier power is also modeled as disturbed by random noise (e.g.,
receiver noise), and deterministic variations, such as spacecraft spinning. The latter
disturbances are decomposed into harmonic components.
Position disturbances are caused by antenna motion. These disturbances may
be purely random, and as such are modeled as a white Gaussian noise of given
standard deviation, or more or less deterministic disturbances, which are modeled
as harmonic components within a bandwidth up to 10 Hz (i.e., antenna dynamics
bandwidth).
The random disturbances were simulated separately in the elevation and cross-
elevation directions. The results are shown in Tables 12.3 and 12.4. They show
that the scanning algorithms have quite effective disturbance rejection properties.
In conscan, for example, a disturbance in XEL direction of 0.3-mdeg standard
deviation will cause approximately 0.005 mdeg of error in the EL estimation and
0.01 mdeg of error in the XEL estimation.
Next, the antenna position xa was disturbed by a harmonic motion in the EL
direction of frequency f and amplitude 0.1 mdeg. The disturbance’s impact on esti-
mation accuracy was analyzed for frequencies ranging from 0.0001 Hz (very slow
motion) to 10 Hz (antenna bandwidth). Note that the scanning frequency is fo = 1/60
= 0.0167 Hz. Consider the plots of the estimation error in EL and XEL as shown
in Fig. 12.24a,b. They show that for frequencies higher that the scanning frequency
the disturbances are quickly suppressed. The slope of the error magnitude drops as
follows: for conscan, −60 dB/dec in EL and −40 dB/dec in XEL; for Lissajous, scan
−20 dB/dec in EL and −60 dB/dec in XEL; and for rosette, −20 dB/dec in EL and
−40 dB/dec in XEL. For low frequencies, the disturbance level is constant in EL for
all scans, and drops down in XEL: −20 dB/dec for conscan and rosette scan, and
–40 dB/dec for Lissajous scan.
Random power variations are also simulated. Variations have standard devia-
tions ranging from 0.1 to 10% of the maximum power. For all three scans, the EL

Table 12.3 Estimation error (mdeg) due to 1-mdeg noise in EL

Direction Conscan Lissajous Rosette
Elevation 0.025 0.025 0.037
Cross-elevation 0.013 0.017 0.010

Table 12.4 Estimation error (mdeg) due to 1-mdeg noise in XEL

Direction Conscan Lissajous Rosette
Elevation 0.015 0.019 0.010
Cross-elevation 0.027 0.029 0.037

@Seismicisolation
@Seismicisolation
12.3 Scanning 207

(a)
10–1

10–2
elevation error, mdeg

10–3

10–4

10–5 conscan
Lissajous
rosette
10–6 –4
10 10–3 10–2 10–1 100 101
frequency, Hz

(b)
10–1
conscan
Lissajous
10–2 rosette
cross-elevation error, mdeg

10–3

10–4

10–5

10–6 –4
10 10–3 10–2 10–1 100 101
frequency, Hz

Fig. 12.24 Estimation errors in response to elevation harmonic disturbance of amplitude 0.1 mdeg:
(a) elevation error, and (b) cross-elevation error

and XEL estimation error was proportional to the variation of power with gain of
0.080–0.089 mdeg per 10% of power variation, as shown in Table 12.5. This is very
effective suppression of power noise, because the10% standard deviation of power
variation causes less than 0.1 mdeg estimation error.

@Seismicisolation
@Seismicisolation
208 12 RF Beam Control

Table 12.5 Estimation error (mdeg) due to power noise

Direction Conscan Lissajous Rosette
Elevation 0.089 0.083 0.080
Cross-elevation 0.089 0.080 0.088
The standard deviation of the power noise is 10% of peak power.

Next, the impact of pulsating power on the estimation of the EL and XEL posi-
tion was analyzed. The results are shown in Fig. 12.25a,b. The harmonic power
variations were of frequencies ranging from 0.0001 Hz to 10 Hz, and of amplitude

(a) 101
conscan
Lissajous
100 rosette
elevation error, mdeg

10–1

10–2

10–3

10–4 –4
10 10–3 10–2 10–1 100 101
frequency, Hz

(b) 101
conscan
Lissajous
100 rosette
cross-elevation error, mdeg

10–1

10–2

10–3

10–4 –4
10 10–3 10–2 10–1 100 101
frequency, Hz

Fig. 12.25 Estimation errors in response to harmonic power variation: (a) elevation error, and (b)
cross-elevation error

@Seismicisolation
@Seismicisolation
References 209

0.1 (10% of maximal power). The plots show that the maximal estimation error of
3 mdeg amplitude was observed for frequencies near the scan frequency, and that
for frequencies lower and higher frequencies the amplitude of the estimation errors
quickly drops. Thus, all three scanning algorithms act as effective filters for this kind
of disturbance.

References
1. Alvarez L. (1993). Analysis of Open-Loop Conical Scan Pointing Error and Variance Esti-
mators. TDA Progress Report 42-115, Jet Propulsion Laboratory, Pasadena, CA. Available at
http://ipnpr.jpl.nasa.gov/progress report/42-115/115 g.pdf
2. Biernson G. (1990). Optimal Radar Tracking Systems, Wiley, New York.
3. Chen L, Fathpour N, Mehra R. (2007). Comparing Antenna Conical Scan Algorithms for
Spacecraft Position Estimation. J Guidance, Control and Dynamics, 30(4): 1186–1188.
4. Damonte JB, Stoddard DJ. (1956). An Analysis of Conical Scan Antennas for Tracking. IRE
National Convention Record, 4(1).
5. Dutton K, Thompson S, Barraclough B. (1997). The Art of Control Engineering. Addison-
Wesley, Harlow.
6. Eldred DB. (1994). An Improved Conscan Algorithm Based on Kalman Filter. TDA Progress
Report 42-116, Jet Propulsion Laboratory, Pasadena, CA. Available at http://ipnpr.jpl.nasa.gov/
progress report/42-116/116r.pdf.
7. Gawronski W. (1995). Wind Gust Models Derived from Field Data. TDA Progress Report
42-123, Jet Propulsion Laboratory, Pasadena, California, pp. 30–36. Available at http://ipnpr.
jpl.nasa.gov/progress report/42-123/123G.pdf.
8. Gawronski W, Craparo E. (2002). Antenna Scanning Techniques for Estimation of Spacecraft
Position. IEEE Antennas and Propagation Magazine, 44(6).
9. Gawronski W, Gudim MA. (1999). Design and Performance of the Monopulse Control System.
IEEE Antennas and Propagation Magazine, 41(6).
10. Lee HP, Hwang HY. (1997). Design of Two-Degree-of Freedom Robust Controllers for a
Seeker Scan Loop System. Int. J. Control, 66(4).
11. Ohlson JE, Reid MS. (1976). Conical-Scan Tracking with 64-m Diameter Antenna at Gold-
stone. JPL Technical Report, 32-1605, Jet Propulsion Laboratory, Pasadena, CA.
12. Wan H, Liang Z, Zhang Q et al. (1996). A Double-Band Infrared Image Processing System
Using Rosette Scanning. Detectors, Focal Plane Arrays, and Applications, SPIE Proceedings,
vol. 2894.

@Seismicisolation
@Seismicisolation
Chapter 13
Track-Level Compensation

The track-level compensation technique is an open-loop control of the RF beam

position. The beam position is calculated with respect to the antenna azimuth and
elevation position from the look-up table, and the antenna position is corrected to
reduce the pointing error.
The whole antenna structure rotates in azimuth on a circular azimuth track. It is
manufactured with the level of precision of ±0.5 mm for the 34-m DSN antennas.
For more discussion of the azimuth track design, see [1]. The pointing accuracy
of the antennas and radiotelescopes is impacted by the unevenness of the antenna
azimuth track. The track unevenness causes repeated antenna tilts, hence repeat-
able pointing errors. In this chapter, track-level errors are described along with their
compensation using the look-up table. Also, the creation of the table is described,
which includes the collection and processing of the inclinometer data, and deter-
mining azimuth axis tilt. Next, the antenna pointing errors are derived from the
table. Finally, pointing improvement is discussed using the table.

13.1 Description of the Track-Level Problem

The track is shown in Fig. 13.1. Manufacturing and installation tolerances, as
well as the foundation compliance and the gaps between the segments of the
track, are the sources of the pointing errors that reach over 20 mdeg peak-to-peak
magnitude.
This chapter discusses the improvement of the pointing accuracy of the antennas
by implementing the track-level-compensation look-up table; see [2, 3]. The table
consists of three axis rotations of the antenna lower structure (alidade) as a function
of the azimuth position. The development of the table is presented, based on the
measurements of the inclinometer tilts, and describe the processing the measure-
ment data. Also the determination of the elevation and cross-elevation errors of the
antenna is presented as a function of the alidade rotations. The pointing accuracy of
the antenna with and without a table was measured using various RF beam pointing
techniques. It was shown that the pointing error decreased when the table was used,
from 7.5 mdeg to 1.2 mdeg in elevation, and from 20.4 mdeg to 2.2 mdeg in cross-
elevation.

W. Gawronski, Modeling and Control of Antennas and Telescopes, 211

@Seismicisolation
@Seismicisolation
DOI: 10.1007/978-0-387-78793-0 13, C Springer Science+Business Media, LLC 2008
212 13 Track-Level Compensation

(a)

alidade

track

(b)

Fig. 13.1 The 34-m antenna: (a) alidade and azimuth track, and (b) azimuth track

Track-level compensation (TLC) look-up tables were created for the DSN
antennas in Goldstone, California, in Canberra, Australia, and in Madrid, Spain.
However, the James Clerk Maxwell Telescope1,2 used inclinometers to perform track
profile measurements to overcome possible systematic errors, but the results have
not been published. The track-level unevenness compensation is planned for the

1 http://www.jach.hawaii.edu/ets/mech/mech recent.html
2 http://www.jach.hawaii.edu/JCMT/telescope/pointing/20011006.html

@Seismicisolation
@Seismicisolation
13.2 Collection and Processing of the Inclinometer Data 213

Sardinia Radio Telescope [5]. The Green Bank Telescope memo3 reports on the
pointing errors due to the azimuth track-level unevenness. GBT says4 that “in the
antenna engineering and operations area work on the Green Bank Telescope azimuth
track was seen as the most important.” Inclinometers were used also for the thermal
deformation of the IRAM telescope [4].

13.2 Collection and Processing of the Inclinometer Data

The TLC system hardware consists of four inclinometers, the interface assembly,
and a PC computer. Four digital inclinometers (model D711 of Applied Geome-
chanics) are mounted on the antenna. The inclinometers are located on the alidade,
as shown in Fig. 13.2. Each inclinometer measures tilt in two axis, denoted x
and y. The manufacturer describes the inclinometer rotation as tilts. Note that the
x-axis tilt is equivalent to the y-axis rotation, and vice versa, as shown in Fig. 13.3.
First, the track profile was measured. A shim of 2.5 mm was placed on the track,
and the antenna wheel positioned on the shim. The inclinometer tilts were measured
to determine the relationship between the inclinometer tilt and the azimuth track
unevenness. The shim caused 14.8 mdeg x-tilt of the inclinometer No. 3. Based on
this scaling and the continuous records of the lower inclinometer x-tilt measure-
ments during the antenna constant velocity slewing the azimuth track profile was
determined, and is shown in Fig. 13.4. It is seen from this plot that the maximum,

Z rotation

2 L
y2 x2

1
Y rotation x1 X rotation
y1
4
x4
y4
3

x3 y3

Fig. 13.2 The location of the inclinometers at the alidade and X, Y, and Z rotations of the alidade

3 http://wwwlocal.gb.nrao.edu/ptcs/ptcspn/ptcspn40/AzTrackSpec.pdf
4 http://www.nrao.edu/news/newsletters/nraonews94.pdf

@Seismicisolation
@Seismicisolation
214 13 Track-Level Compensation

Fig. 13.3 x-axis tilt is a

rotation with respect to y-axis
x-tilt

x
y-rotation

Fig. 13.4 Azimuth track profile of the 34-m antenna

peak-to-peak, track profile variation is 1.2 mm, slightly higher than the specifica-
tion (1 mm).
Next, the inclinometer data were collected to determine the alidade rotations. The
antenna moves at constant azimuth axis velocity of 0.05 deg/s. Due to the environ-
mental disturbances the inclinometer data are extremely noisy. Take for example the
x-axis measurement of the inclinometer 1 shown in Fig. 13.5. The unfiltered data are
represented by the gray line. Using a zero-phase filter to prevent filtering delay, the
data is smoothed, as represented by the black line.

20

15
inclinometer 1, x-tilt, mdeg

10

–5

–10

–15

–20
0 50 100 150 200 250 300 350
azimuth encoder, deg

Fig. 13.5 Raw inclinometer data (gray line) and the filtered data (black line)

@Seismicisolation
@Seismicisolation
13.3 Estimating Azimuth Axis Tilt 215

13.3 Estimating Azimuth Axis Tilt

The additional processing includes the removal of the azimuth axis tilt from the
data. The tilt is present in the inclinometer data as harmonic functions in x- and
y-axis tilt, of period 360 deg; see Fig. 13.6a,b. Its amplitude (a) and phase (ϕ) need
to be determined. Let α1x (i) and α1y (i) be the ith sample of the x- and y-tilts of the
inclinometer 1, and e(i) be the ith sample of the azimuth encoder. The inclinometer
harmonics caused by the azimuth axis tilt are described as

α1x (i) = a cos(e(i) + ϕ), and α1y (i) = a sin(e(i) + ϕ) (13.1)

or, in short, as

α1x (i) = ac ci − as si and α1y (i) = ac si − as ci (13.2)

10
inc linometer x -tilt, mdeg

–5

–10
0 50 100 150 200 250 300 350
azimuth position, deg

10
inc linometer y -tilt, mdeg

–5

–10
0 50 100 150 200 250 300 350
azimuth position, deg

Fig. 13.6 Removing the azimuth axis tilt from the inclinometer data (solid line, inclinometer data;
dash-dot line, inclinometer tilt caused by the azimuth axis tilt; and dashed line, inclinometer data
after azimuth axis tilt removal)

@Seismicisolation
@Seismicisolation
216 13 Track-Level Compensation

where ac = a cos(ϕ), as = a sin(ϕ), ci = cos(e(i)), and si = sin(e(i)). For n

samples define the following vectors and matrices
⎧ ⎫ ⎧ ⎫
⎪
⎪ α1x (1)⎪
⎪ ⎪
⎪ α1y (1)⎪
⎪
  ⎪
⎨α1x (2)⎬⎪ ⎪
⎨α1y (2)⎪⎬
α1x
α1 = , where α1x = .. and α1y = .. (13.3)
α1y ⎪
⎪ . ⎪⎪ ⎪ . ⎪
⎪ ⎪
⎪
⎩ ⎪
⎭ ⎪
⎩ ⎪
⎭
α1x (n) α1y (n)

⎧ ⎫ ⎧ ⎫
⎪
⎪lc1 ⎪⎪ ⎪
⎪ls1 ⎪⎪
  ⎪ ⎬
⎨ ⎪ ⎪
⎨ s2 ⎪⎬
c −s c 2
P= , where c = . and s = . (13.4)
s c ⎪
⎪ .. ⎪⎪ ⎪
⎪ .. ⎪
⎪
⎪
⎩ ⎪ ⎭ ⎪
⎩ ⎪ ⎭
cn sn

and
 
ac
A= (13.5)
as

For the above notations equations (13.2) can be rewritten in a compact form

P A = α1 (13.6)

The least-square solution A of equation (13.5) is as follows

A = (P T P)−1 P T α1 (13.7)

But, from equation (13.5) it follows that

a cos(ϕ) = ac , and a sin(ϕ) = as (13.8)

therefore
 + ,
as
a= ac2 + as2 and ϕ = tan−1 (13.9)
ac

Based on several sets of data one obtains from (13.9) the following amplitude
and phase of the azimuth axis tilt

a = 4.2 mdeg and ϕ = 274.5 deg

@Seismicisolation
@Seismicisolation
13.4 Creating the TLC Table 217

that is, the tilt magnitude 4.2 mdeg and phase 274.5 deg. The x- and y-axis move-
ments of the inclinometer 1 after the tilt removal is shown in Fig. 13.6a,b (dashed
line).

13.4 Creating the TLC Table

The TLC look-up table consists of X, Y, and Z rotations of the alidade, as shown
in Fig. 13.2. They are obtained from the inclinometer tilts. Namely, a rotation with
respect to the antenna x-axis, denoted X, is a rotation with respect to the antenna
elevation axis. It is measured as the y-tilt of the second inclinometer (α2y ):

X = α2y (13.10)

The Y rotation is a tilt of the elevation axis. It is an average of the x-tilts of the
inclinometers 1 and 2, that is,

Y = 0.5(α1x + α2x ) (13.11)

The Z rotation of the alidade is a twist of the alidade, and is not directly measured
by the inclinometers. It is determined from x-tilts of inclinometers 3 and 4, as
follows. From Fig. 13.7, which represents the view from the top of the alidade,
one has

d3 − d4
Z= (13.12)
L

where d3 and d4 are horizontal displacements of the locations of inclinometers 3

and 4, and L = 12.396 m is the distance between the two inclinometers. The displace-
ments d3 and d4 are determined from the tilts of inclinometers 3 and 4, respectively,
by assuming that the horizontal displacement of the alidade side due to azimuth
track unevenness is caused predominantly by the rigid-body motion of each side of
the alidade. This assumption has been checked with the finite-element model of the
alidade, giving a 93% accuracy in the estimation of displacements d3 and d4 . It was

Z
d3
d4
L
Inclinometer 4 Inclinometer 3

Fig. 13.7 Top view on the inclinometers 3 and 4

@Seismicisolation
@Seismicisolation
218 13 Track-Level Compensation

also confirmed by the comparison of the rotations of the inclinometers located at the
bottom, the middle, and the top of the alidade.
The rigid-body angle is measured as the x-tilt of the inclinometers 3 and 4
(denoted as α3x and α4x , respectively), therefore

d3 = H α3x and d4 = H α4x (13.13)

where H is the height at which the inclinometers are located, H = 9.292 m. Intro-
ducing (13.13) to (13.12) one obtains the Z rotation of the alidade as

H
Z= (α3x − α4x ) (13.14)
L

where H is the alidade height and L is the distance between the inclinometers 1
and 2. Because for the 34-m antennas L = 12.39 m, the ratio is H / L = 0.75,
therefore

Z = 0.75(α3x − α4x ) (13.15)

The X, Y, and Z alidade rotations obtained from the inclinometer data, for azimuth
angles varied from 0 to 360 deg, are shown in Fig. 13.8. The plots show that the X
rotation (the elevation correction) is comparatively small, and that the largest is the

5
X , mdeg

–5
0 50 100 150 200 250 300 350
10
Y, mdeg

–10
0 50 100 150 200 250 300 350

10
Z, mdeg

–10
0 50 100 150 200 250 300 350
azimuth position, deg

Fig. 13.8 The TLC look-up table of the 34-m antenna

@Seismicisolation
@Seismicisolation
13.5 Determining Pointing Errors from the TLC Table 219

Z rotation. But, it will be shown later that the Z rotation is compensated for by the
azimuth encoder and hence it is not a part of pointing error.

13.5 Determining Pointing Errors from the TLC Table

The antenna elevation and cross-elevation pointing errors are determined. The eleva-
tion error Δ E L is simply determined as the alidade X rotation

ΔE L = X (13.16)

The cross-elevation error, Δ X E L , depends on the antenna elevation position, EL,

and on the alidade Y and Z rotations as illustrated in Fig. 13.9

Δ X E L = Z cos(E L) − Y sin(E L) (13.17)

However, the Z-rotation contributions are assumed zero in the TLC table because
this error is measured by the azimuth encoder and therefore eliminated by the
azimuth servo. The following experiment at the 34-m antenna was conducted at
the Madrid Deep Space Communication Complex to verify this hypothesis. With
the antenna dish positioned at EL = 30 deg a shim of 1 mm thick was placed on the
azimuth track, as shown in Fig. 13.10. The antenna was then moved slowly with
constant speed in azimuth over the shim. The same antenna movement was repeated
when the shim was removed. The difference between azimuth encoder reading with
and without the shim is plotted in Fig. 13.11. It shows the azimuth position rising
sharply (section A) when the antenna is climbing the shim. But the azimuth servo
compensates for the shim disturbance (section B), and the azimuth position returns

xel
Z
Z cos(EL)
EL Δ XEL

Y
y
Y sin(EL)

Fig. 13.9 The relationship between the cross-elevation error and the X and Y rotations of the
alidade

@Seismicisolation
@Seismicisolation
220 13 Track-Level Compensation

Fig. 13.10 Azimuth wheel crosses 1 mm shim

to the initial position (section C). As result the antenna does not need correction in
z-axis, and the Z component of the TLC table shall be zero.
Based on the above experiment the following equation

ΔXEL = −Y sin(E L) (13.18)

is the resulting formula for the cross-elevation error.

10

4
AZ enc oder, mdeg

–2

–4

–6
shim profile
–8

–10 A B C
0 0.5 1 1.5 2 2.5 3
AZ track angle, deg

Fig. 13.11 The azimuth encoder reading when crossing the shim. Section A shows a sharp rise in
encoder reading at the beginning of the shim; Section B shows an azimuth servo correction to the
shim disturbance; and Section C shows the stabilized azimuth position

@Seismicisolation
@Seismicisolation
13.6 Antenna Pointing Improvement Using the TLC Table 221

13.6 Antenna Pointing Improvement Using the TLC Table

The improvement of pointing accuracy with the look-up table was evaluated using
the RF pointing data. The following RF beam measurement techniques were used:
boresight, monopulse, and conscan. The data were measured with the installed TLC
table (“TLC table on”) and without the TLC table (“TLC table off” ). Both methods
are useful in the validation of the effectiveness of the TLC table. Namely, when
the table is on, the pointing errors should be significantly smaller than the errors
predicted from the TLC table (or the errors obtained for the same track with the
TLC table off ). When the table is off, the RF pointing errors should match the
errors predicted from the TLC table.
The measurements with the TLC table off were taken for the trajectory shown in
Fig. 13.12. Figure 13.13a shows that the measured elevation pointing errors and the
errors predicted by the look-up table coincide. The elevation error predicted from the
TLC table varies by 7 mdeg, from –3 to 4 mdeg (dashed line). Figure 13.13b shows
that the cross-elevation errors (predicted and measured) coincide when the antenna
elevation position is below 72 deg, and that a deterministic residual is uncompen-
sated when the antenna elevation position is above 72 deg The cross-elevation error
predicted from the TLC table vary by 12 mdeg, from –5 to 7 mdeg (dashed line); they
show a “deterministic” error (for AZ < 240 deg, where antenna elevation position
above 72 deg).
The measurements of the radio beam position with a TLC table have a standard
deviation of 0.41 mdeg (or 1.2 mdeg peak-to-peak), while the radio beam data with
the TLC table shows its standard deviation of 0.72 mdeg (or 2.2 mdeg peak-to-peak),
for the antenna at elevation position below 72 deg.
Table 13.1 summarizes the antenna tracking accuracy with the TLC table on and
off. The elevation pointing error decreased 6-fold, and the cross-elevation pointing
error decreased 10-fold.
elev at ion pos it ion, deg

80

70

60

50

40
140 160 180 200 220 240 260 280
azimuth position, deg

Fig. 13.12 The 34-m antenna tracking trajectory

@Seismicisolation
@Seismicisolation
222 13 Track-Level Compensation

(a)

5
EL error, mdeg

–5

100 120 140 160 180 200 220 240 260 280
(b)

5
X E L error, mdeg

–5

100 120 140 160 180 200 220 240 260 280
azimuth position, deg

Fig. 13.13 The 34-m antenna pointing errors, measured (black solid line) and predicted from the
TLC table (dashed line): (a) the elevation pointing error, and (b) the cross-elevation pointing error

Table 13.1 Peak-to-peak pointing errors of the 34-m antenna

Elevation error Cross-elevation error
Without TLC table 7.5 mdeg 20.4 mdeg
With TLC table 1.2 mdeg 2.2 mdeg

References
1. Antebi J, Kan FW. (2003) Precision Continuous High-Strength Azimuth Track gor Large Tele-
scopes. Proc. of SPIE, Future Giant Telescopes, 4840: 612–623.
2. Gawronski W, Baher F, Quintero O. (2000). Azimuth track-level Compensation to Reduce
Blind Pointing Errors of the Deep Space Network Antennas. IEEE Antennas and Propagation
Magazine, vol. 42.
3. Gawronski W, Baher F, Gama E. (2006). Track-Level Compensation Look-Up Table Improves
Antenna Pointing Precision. IPN Progress Report 42–164. Available at http://ipnpr.jpl.
nasa.gov/progress report/42-164/164E.pdf
4. Greve M, Bremer J, Penalver P, et al. (2005). Improvement of the IRAM 30-m Telescope from
Temperature Measurements and Finite-Element Calculations. IEEE Trans. Antennas and Prop-
agation, 53(2).
5. Pisanu T, Morisani M, Pernechele C, et al. (2004). How to Improve the High-Frequency Capa-
bilities of SRT. Proc. 7th European VLBI Network Symposium, Toledo, Spain, 2004.

@Seismicisolation
@Seismicisolation
Index

A sliding window, 200–201

Acceleration limit, 39, 85–87, 107, 117, trajectory, 195, 197
129–132, 141–142, 157–165 Coulomb friction torque, 168–170
Antenna model Counter-torque, 168–169
reduction, 45–50, 133, 161–162 CPP, see Command preprocessor
rigid, 11–12 Critical frequency, 85
Anti-windup technique, 157, 162–163 Critical integral gain, 85, 87, 104, 107
Atacama Large Millimeter Array, 2–3 Cross-elevation coordinate, 195
Augmentation of the model, 73, 77, 78–79 Cross-spectral density function, 34–35
Azimuth axis tilt, 211, 215–217
Azimuth model, 41–43 D
Damping matrix, 13, 15–16
Data processing, 31, 34–37, 111
B
Davenport filter, 59–62, 64, 66
Backlash
Davenport spectrum, 60–62, 89
equation, 177–181
Deep Space Network, 1, 5–7, 102, 127
Balanced model reduction, 45–47
34 meter antenna, 95, 110
Balanced representation, 46, 50
70 meter antenna, 95, 110
Bandwidth, 11, 12, 27–28, 30, 31, 38, 61, 66,
Discrete-time data, 37–38
73, 75, 84–85, 87–90, 92, 104–107,
Discrete-time model, 21–23
110–115, 118–119, 121, 123, 124–127,
Disturbance step, 74, 120, 140, 148–149, 153
129, 131–133, 140, 146, 148, 183–188,
Disturbance transfer function, 74, 76, 87,
194, 206
91, 115, 121, 140, 141, 147–148,
Bias, torque, 174–181
153–154, 189
Dither, 157, 167, 170–174
C Drive inertia, 26, 28–30
Carrier power, 197, 198, 202, 206 Drive model, 11, 24–26, 68, 129, 177
Closed-loop test, 31 Drive stiffness, 26, 27–28
Coherence, 35–37 Dry friction model, 168–170
Command preprocessor Dynamic pressure of wind, 64
with antenna, 157–162
description, 158 E
dynamics, 155, 161–162 Effelsberg telescope, 4–5
gains, 160 ESA deep space antennas, 2
gain selection, 158–165
Conical scan, see Conscan F
Conscan Feedforward controller, 81–92, 129
estimated spacecraft position, 200, 202 Finite element model, 12–13, 17, 59, 129, 177
power variation, 195–200, 201, 203–204, Friction
206–208 dry friction model, 168–170

223
@Seismicisolation
@Seismicisolation
224 Index

Friction torque tracking, 98–101

smooth image, 172–173 tuning steps, 108–110
Fundamental frequency, 27, 30, 43–44, 107, tuning using GUI, 117
108, 110, 115, 121, 129 weight properties, 105–107
weights, 101–103
G LQG controller in the velocity loop,
Gain stability margin, 75–76 124–132
Green Bank Telescope, 4, 166, 173, 213
M
H
Mass matrix, 13–15
H2 norm, 47–49
Matlab, 11, 17, 22, 35, 40, 46, 98, 111, 117,
H∞ controller
119, 123, 124
closed-loop equations, 138–139
Modal displacement, 15–17, 21
definition, 135–138
Modal model, 11, 12, 14–15, 21, 45, 78
design, 136–138
reduction, 47–48
with DSN antenna, 139–141
Model reduction, 45–50
gains, 135–138
Modes, 14, 15, 17, 18, 20, 21, 23, 41, 45, 47,
limits of performance, 141–142
48, 49, 61, 77, 102, 103, 107
tracking, 138
H∞ norm, 47–49, 135 Monopulse
Hankel norm, 46, 47, 48, 49 command following, 187–188
disturbance rejection, 188–189
I performance, linear model, 190–191
Identification, velocity loop model, 26, 39–44 performance, nonlinear model, 191–193
Inclinometer data, 211, 213–218 principle, 187
Integral gain, 81, 83, 85–90, 100, 104, 106, stability, 190
107, 110–111, 119, 121, 129, 132, 146 Motor model, 24

K N
Ka band, 1, 194, 197 Natural frequencies, 14, 16, 17, 42,
43, 139
L Natural modes, 14, 15, 17
Large Millimeter Telescope Norm, single mode, 47–48
control systems, 1–3, 127 Norm, structure, 48–49
Limit Nyquist frequency, 37–38
acceleration, 39, 85–87, 107, 117, 129,
131, 132, 141–142, 157–165
O
velocity, 73, 95, 129, 157–165
Open-loop test, 179
Limit cycling, 129, 157, 159, 162, 163,
Overshoot, 73–74, 90, 111, 119–121, 129,
164, 171
131–133, 153, 159–162, 164, 165,
Lissajous scan, 194, 201–204, 206
187–188
Low velocity tracking, 170–173
LQG controller in the position loop
closed-loop equations, 101 P
description, 95–98 Performance criteria, 73–77
fine tuning, 121–124 Phase stability margin, 75–77
gains from Matlab, 98 PI controller
limits of the gains, 107–108 closed-loop equations, 87
limits of performance, 117 closed-loop transfer function, 83
performance, 110 limit of performance, 90
performance with DSN antennas, 111–113 performance, 87–90
performance vs PI performance, 115–117 tuning steps, 86–87
properties of weights, 105–107 Pointing accuracy, closed loop, 68–70
resemblance of PI controller, 104–105 Poles, 19, 23, 49–50, 85

@Seismicisolation
@Seismicisolation
Index 225

Position loop, 5–7, 32–34, 68, 82, 100, Tracking H∞ controller, 138
113–115, 124–132, 145–147, 155, Tracking LQG controller, 101
183–185, 187, 188–190 Tracking PI controller, 91–92
Power spectral density function, 35 Track level compensation
Proportional gain, 82–90, 100, 104, 106, 108, collecting data, 213–214
114, 115, 129, 132 creating look-up table, 217–219
description, 211–213
R determining pointing errors, 219–220
Reducer model, 24–25 estimating azimuth axis tilt, 215–217
Reduction, 45–50, 133, 161–162 inclinometer locations, 213–214
RF beam controller pointing errors, 219–220
Selection, 183–186 pointing improvement, 221–222
Riccati equations, 97, 98, 136 processing data, 213–214
Rigid antenna model, 11–12 table, 217–219
Rigid body model, 11–12, 19–21, 45, 49–50, Track profile, 212–214
145, 148, 217, 218 Transfer function, 11, 12, 26–30, 31, 32,
Rms servo error, 87–89, 110, 112–115, 118, 34–37, 39–42, 47, 48, 50, 60–62,
131–132 66–67, 73–76, 81–88, 91–93, 102–107,
Rosette scan, 194, 199, 204–205, 206 109, 111, 113, 115, 119–132, 135,
140–141, 143, 146–148, 150, 153–154,
S 183–189
Sampling rate, 22, 23, 32, 37–38, 194 Transfer function from data, 34–40
Sampling time, 22, 23, 37–38, 61, 67, 129, Transformations of the velocity loop model,
159, 161, 170, 196 34–37
Scanning Tuning
performance evaluation, 206–208 LQG controller, 108–110
Settling time, 73–74, 88–91, 110–113, 115, PI controller, 86–87
120, 121, 129, 131, 132, 138, 139, 140,
148, 153, 159, 162, 188 V
Single-loop control Velocity limit, 73, 95, 129, 158–159, 161,
34-meter antenna, 149–154 163, 167
rigid antenna, 145–149 Velocity loop, 5–7, 11–12, 26–30, 31, 33–35,
Stability margin, 75, 186, 190 38–43, 45, 51, 60, 73, 77–79, 81–83,
State-space model, 11, 17–19, 34, 77, 79 87, 91–92, 95–96, 104, 115, 123–132,
Steady-state error, 73, 78, 119, 129, 138 145–155, 177–179, 184, 190–193
Stiction torque, 168–169 Velocity loop model
Stiffness matrix, 13, 15, 20 rigid antenna, 11–12
Structural model, 12–23, 45, 61 Velocity offset, 32, 72–73, 121
Surface drag coefficient, 60, 63
W
T White noise, 59–61, 64–66, 68, 187, 190
Test testingm, 31–34
configuration, 32–34 Wind filter, 139
duration, 32 Wind forces, 59, 60–64, 67–68
input, 32–34 Wind model
input amplitude, 32 gusts, 59–70
output, 32 static, 51
velocity offset, 32 Wind torque, 51, 52, 54, 59–60, 64–68, 153
Thirty Meter Telescope, 3 Wind tunnel data, 51, 55–56, 58
Tilt, 211, 213–218 Wind velocity, 59–64, 67, 111, 115, 128
Time profile of wind gusts, 65–66
TLC, see track level compensation X
Torque bias, 174–181 X-band, 1

@Seismicisolation
@Seismicisolation
Mechanical Engineering Series (continued from page ii)

E.N. Kuznetsov, Underconstrained Structural Systems

P. Ladevèze, Nonlinear Computational Structural Mechanics: New Approaches and Non-
Incremental Methods of Calculation
P. Ladevèze and J.-P. Pelle, Mastering Calculations in Linear and Nonlinear Mechanics
A. Lawrence, Modern Inertial Technology: Navigation, Guidance, and Control, 2nd ed.
R.A. Layton, Principles of Analytical System Dynamics
F.F. Ling, W.M. Lai, D.A. Lucca, Fundamentals of Surface Mechanics: With Applications,
2nd ed.
C.V. Madhusudana, Thermal Contact Conductance
D.P. Miannay, Fracture Mechanics
D.P. Miannay, Time-Dependent Fracture Mechanics
D.K. Miu, Mechatronics: Electromechanics and Contromechanics
D. Post, B. Han, and P. Ifju, High Sensitivity and Moiré: Experimental Analysis for Mechanics
and Materials
R. Rajamani, Vehicle Dynamics and Control
F.P. Rimrott, Introductory Attitude Dynamics
S.S. Sadhal, P.S. Ayyaswamy, and J.N. Chung, Transport Phenomena with Drops and Bubbles
A.A. Shabana, Theory of Vibration: An Introduction, 2nd ed.
A.A. Shabana, Theory of Vibration: Discrete and Continuous Systems, 2nd ed.
Y. Tseytlin, Structural Synthesis in Precision Elasticity

@Seismicisolation
@Seismicisolation