MSC Monkhorst 31mar04

Systems & Control
Dynamic Error Budgeting

a design approach
Wouter Monkhorst
Supervisors ir. L. Jabben

prof. dr. ir. D.L. Trumper
dr. ir. R. Tousain
Professor prof. dr. ir. O.H. Bosgra
Delft, March 31, 2004
Faculty of Design, Engineering and Production

Preface
This thesis is the result of a very exciting and adventurous year of doing research at different
locations. It has always been a dream for me to do research at the Massachusetts Institute
of Technology (MIT) in Boston, U.S.A. and thanks to professor Jan van Eijk (Advanced
Mechatronics, Delft University of Technology (DUT)), this dream has become a reality. The
resulting research project was done in co-operation with Philips Center for Industrial Tech-
nology (Philips CFT) and MIT.
After I finished a feasibility study of a design approach applicable to high end mechatronic
systems, I have done a five months research project at MIT, Boston. Back in the Nether-
lands, I finished the research at Philips CFT, Eindhoven. In this research I was able to
bring together the theory I have learned at the Systems and Control group DUT, with the
practical knowledge of mechatronic system design at MIT, a very challenging job!
First of all, I would like to acknowledge my supervisor ir. Leon Jabben for all his time and
support supervising my graduation project and help with the theoretical background and
interpretation of results. I’m very thankful to Professor Trumper for the guidance he gave
me during my research and for making it possible for me to do this at MIT. I would also
like to thank dr. ir. Rob Tousain for his guidance and support during my research at Philips
CFT.
Last, but not least, I would like to thank other colleague researchers who helped me with
various difficulties and for making this year a pleasant time. A lot of thanks to Katie
Lilienkamp, Rick Montesani and David Otten for their help with the design of the exper-
imental setup at MIT, thanks to Michiel Vervoordeldonk, Toon Blom, Dick Goossens and
Ger Jansen for helping me out with the noise modelling at Philips CFT.
iii
iv
Contents
Preface iii
Summary ix
Nomenclature xi
1 Introduction 1
1.1 Context of this thesis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.2 Dynamic error budgeting . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.3 Vibration isolation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2
1.4 Challenge definition . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
1.5 Outline of thesis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
2 Dynamic Error Budgeting 7

2.1 Motivations of the DEB design approach . . . . . . . . . . . . . . . . . . . . . 7
2.2 DEB design process . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8
2.3 Assumptions and research boundaries . . . . . . . . . . . . . . . . . . . . . . 8
2.4 Theoretical background . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
2.4.1 Variance as a performance measure . . . . . . . . . . . . . . . . . . . . 10
2.4.2 Parseval’s Theorem: linking the frequency to the time domain . . . . 10
2.4.3 Power spectral density function . . . . . . . . . . . . . . . . . . . . . . 10
2.4.4 Propagation of disturbances . . . . . . . . . . . . . . . . . . . . . . . . 12
2.5 H2 control, maximizing performance . . . . . . . . . . . . . . . . . . . . . . . 14
2.5.1 Use of H2 control in DEB . . . . . . . . . . . . . . . . . . . . . . . . . 14
2.5.2 The H2 norm and variance of the output . . . . . . . . . . . . . . . . 15
2.5.3 The H2 control problem . . . . . . . . . . . . . . . . . . . . . . . . . . 15
2.5.4 Using weighting filters to model disturbances . . . . . . . . . . . . . . 15
2.5.5 Output scaling and the Pareto curve . . . . . . . . . . . . . . . . . . . 17
2.6 Conclusions on the theory . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17
3 Case study, active vibration isolation 19

3.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19
3.2 An active vibration isolation platform . . . . . . . . . . . . . . . . . . . . . . 20
3.2.1 Description of the experimental setup . . . . . . . . . . . . . . . . . . 20
v
3.2.2 Performance definition . . . . . . . . . . . . . . . . . . . . . . . . . . . 21
3.2.3 System modelling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21
3.2.4 Mutual inductance of the speaker and geophone . . . . . . . . . . . . 25
3.2.5 Closing the loop . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25
3.3 Modelling of disturbance sources . . . . . . . . . . . . . . . . . . . . . . . . . 27
3.3.1 Measurements and model of ground vibrations . . . . . . . . . . . . . 28
3.3.2 Sensor noise sources . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29
3.3.3 A/D and D/A electrical noise . . . . . . . . . . . . . . . . . . . . . . . 30
3.3.4 A/D and D/A quantization noise . . . . . . . . . . . . . . . . . . . . . 31
3.4 Performance analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 32
3.5 Using a preamplifier . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 34
3.5.1 Introduction preamplifier . . . . . . . . . . . . . . . . . . . . . . . . . 34
3.5.2 Preamplifier transfer function . . . . . . . . . . . . . . . . . . . . . . . 35
3.5.3 Preamplifier noise model . . . . . . . . . . . . . . . . . . . . . . . . . . 36
3.5.4 Performance analysis using a preamplifier . . . . . . . . . . . . . . . . 37
3.6 Experiments and comparison with theory . . . . . . . . . . . . . . . . . . . . 40
3.7 Conclusions and recommendations . . . . . . . . . . . . . . . . . . . . . . . . 42
3.7.1 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 42
3.7.2 Recommendations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 43
4 H2 control in active vibration isolation 45

4.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 45
4.2 Design of an H2 controller . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 46
4.2.1 Weighting filters for the 10-1000 preamp case . . . . . . . . . . . . . . 46
4.2.2 Meeting the standard assumptions . . . . . . . . . . . . . . . . . . . . 47
4.2.3 Closed transfer functions using an H2 controller . . . . . . . . . . . . 49
4.2.4 The Pareto curve . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 50
4.3 System performance using the H2 controller . . . . . . . . . . . . . . . . . . . 51
4.3.1 Performance analysis using an H2 controller and a preamplifier . . . . 51
4.3.2 Experimental results using an H2 controller . . . . . . . . . . . . . . . 51
4.4 Making a design decision using H2 control in DEB . . . . . . . . . . . . . . . 56
4.4.1 The best preamplifier . . . . . . . . . . . . . . . . . . . . . . . . . . . 56
4.4.2 The constant gain preamplifier . . . . . . . . . . . . . . . . . . . . . . 57
4.5 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 58
5 Case study, Advanced Isolation ModuleS 61

5.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 61
5.2 Advanced vibration isolation . . . . . . . . . . . . . . . . . . . . . . . . . . . 62
5.2.1 Performance specification . . . . . . . . . . . . . . . . . . . . . . . . . 63
5.3 Transfer function of the AIMS . . . . . . . . . . . . . . . . . . . . . . . . . . 63
5.3.1 Reference loop transfer function . . . . . . . . . . . . . . . . . . . . . 64
5.3.2 Payload loop transfer function . . . . . . . . . . . . . . . . . . . . . . 66
5.3.3 Open and closed loop transfer function of total system . . . . . . . . . 68
5.4 Modelling of disturbances and noise sources . . . . . . . . . . . . . . . . . . . 69
vi
5.4.1 Ground vibrations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 70
5.4.2 Electrical noise sources . . . . . . . . . . . . . . . . . . . . . . . . . . . 71
5.4.3 Other disturbances and noise sources . . . . . . . . . . . . . . . . . . . 77
5.4.4 Summary of disturbances and noise sources . . . . . . . . . . . . . . . 77
5.5 System performance . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 78
5.5.1 Modelling the closed loop system . . . . . . . . . . . . . . . . . . . . . 79
5.5.2 Noise propagation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 79
5.5.3 Vibration isolation performance of the AIMS . . . . . . . . . . . . . . 87
5.6 Conclusions and recommendations . . . . . . . . . . . . . . . . . . . . . . . . 89
5.6.1 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 89
5.6.2 Recommendations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 90
6 Conclusions and recommendations 91

6.1 Conclusions on the DEB design approach . . . . . . . . . . . . . . . . . . . . 91
6.2 Recommendations and further work . . . . . . . . . . . . . . . . . . . . . . . 92
A Mathematics 95
A.1 MATLAB code for computing the PSD and CPS . . . . . . . . . . . . . . . . 95
A.2 Standard assumptions in the H2 framework . . . . . . . . . . . . . . . . . . . 96
B Noise and system modelling 99

B.1 State space model of the vibration isolation setup . . . . . . . . . . . . . . . . 99
B.2 Peterson New High/Low Noise Model . . . . . . . . . . . . . . . . . . . . . . 102
B.3 Building the preamplifier . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 103
B.3.1 Preamplifier transfer function . . . . . . . . . . . . . . . . . . . . . . . 103
B.3.2 Preamplifier circuit design . . . . . . . . . . . . . . . . . . . . . . . . . 103
B.3.3 Preamplifier transfer function in physical component values . . . . . . 104
B.3.4 Physical parameter values for the 1000-10 preamplifier . . . . . . . . . 105
B.4 A noise model for the preamplifier . . . . . . . . . . . . . . . . . . . . . . . . 105
B.4.1 Introduction preamplifier noise modelling . . . . . . . . . . . . . . . . 105
B.4.2 Voltage and current noise of an operational amplifier . . . . . . . . . . 106
B.4.3 Noise modelling of the preamplifier . . . . . . . . . . . . . . . . . . . . 107
B.4.4 A noise model for the 1000-10 preamplifier . . . . . . . . . . . . . . . 111
B.4.5 Op07 and op27 performance comparison . . . . . . . . . . . . . . . . . 112
C Experimental setup MIT 115

C.1 List of used equipment . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 115
C.2 List of used software . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 115
C.3 Hardware connection . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 115
C.4 Practical notes on using the HP35665A DSA . . . . . . . . . . . . . . . . . . 116
vii
D CD-ROM 119
D.1 MATLAB° r
m-files . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 119
D.1.1 tools . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 120
D.2 LATEXfiles . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 121
D.3 Files . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 121
viii
Summary
In today’s industry the performance demands on high end mechatronic systems are increased
continuously, systems are to be faster, more accurate and less expensive. System designers
face the challenge to design mechatronic solutions that are to fulfill these demands. The
high design costs make it that one cannot afford to make a design failure, resulting in an
important design demand: the system must be ’first time right’. High end mechatronic sys-
tems have become more and more complex. The system is often designed per sub-system by
different designers, sometimes even at different locations, making the design challenge even
bigger. In this environment the designer needs a structural approach which enables him to
face this challenge. This thesis is about the development of such a design approach, which
leads to the problem statement of this thesis:
problem statement: develop a tool that enables a system designer to predict the final
performance of a system subject to disturbances and gives insight in what system property
limits the performance. Show the feasibility of this approach.
Using this approach the designer must be able to predict the final performance a priori,
accounting for all the disturbances that act on the system, enabling him to design a system
such that it will be first time right.
In this thesis such an approach is developed using theory of spectral analysis and is referred
to as the Dynamic Error Budgeting (DEB) design approach. The theory H2 -control is
then added, reaching the designer a tool to improve his design strategy even further. The
approach is then put to the test in two case studies, both concerning the vibration isolation
problem. A vibration isolation system is to isolate a payload, e.g., very sensitive equipment,
from ground vibrations. The vibration isolation problem can be solved passively (e.g., air-
mount) or actively. In this thesis the focus is on the active systems, which use a control
loop to achieve the isolating performance. The performance of an active vibration isolation
system is determined by the disturbances that act on the system, e.g., the ground vibrations
themselves, but also the noise generated by its electrical components.
In the first case study the final performance of an active vibration isolation system is ana-
lyzed by modelling all the disturbances and calculating the performance level at the output.
The system is then redesigned to improve the final performance. Using the DEB approach
it is shown that pre-amplifying the sensor signal gives a huge performance improvement.
ix
Next, an H2 - controller is synthesized improving the performance even more. Finally, the
performance when using two different preamplifiers is compared using H2 controllers in the
loop, enabling the designer to make an objective performance comparison of the two different
preamplifiers. A number of experiments is done to validate the theory.
The second case study is about an active vibration isolation module that can be used parallel
to an existing passive isolator. The drawback of introducing an additional active isolator, is
that its active components also introduce additional disturbances, like electronic noise. The
improvement in isolating performance must therefore at least counter balance introduction
of the additional disturbances due to its active components, which is analyzed using the
DEB approach. The approach shows that the sensor used in this active isolator is the main
performance limiter of the system and that the total performance is indeed improved.
The two case studies show that the DEB design approach is a powerful tool for the designer in
today’s challenging industrial environment. The approach predicts the final performance of
the system during the design process, and gives valuable insight in the limiting components
and/or system properties. It also helps the designer to keep track of the total system in
a modular design approach. Using H2 -control as an add-on to the DEB approach, the
designer can objectively compare different system concepts using H2 control, helping him
to justify important design decisions. Further research can be done to the choice of the
performance variable, the use of higher order disturbance models, to the applicability of the
information given by the H2 controller in practice and to the development a strategy that
can be followed to improve the system performance.
x
Nomenclature
Symbols
u(t) controller output signal [V ]
w(t) disturbance signal [SI]
z(t) performance signal [SI]
Eg (t) geophone output voltage [V ]

Ein (t) power amplifier input voltage [V ]
Eout (t) preamplifier output voltage [V ]
Esp (t) voice coil input voltage [V ]
Xb (t) base plate position [m]
Xl (t) cone plate or payload position [m]
C(s) controller
CH2 (s) H2 controller
Cz (f ) cumulative power spectrum of signal z(t) [SI2 ]
f frequency [Hz]
G(s) open loop weighted transfer function
Ḡ(s) open loop transfer function
Gg geophone generator constant [V /(m/s)]
H(s) closed loop transfer function
Hg (s) sensor dynamics
L(s) loop transfer function
Nf speaker force constant [N/A]
P (s) general open loop plant
Pgeo (s) geophone velocity sensitivity Eg /Ẋl
Pspeaker (s) speaker transfer function Ẋl /Ein
Psp geo (s) least square fit on Eg /Ein
Psp vib (s) open loop speaker model Ẋl /Ẋb
Rc geophone coil impedance [Ω]
Rs speaker voice coil impedance [Ω]
S(s) loop sensitivity function (1 + L(s))−1
Sz (f )/Sz (ω) power spectral density function of signal z(t) [SI2 /Hz ]
Vw (f ) input weighting filter to model disturbances [SI]
xi
Wz (f ) output weighting filter for relative output scaling [−]
Greek Symbols
σx standard deviation of the signal x(t) [SI]
σx2 variance of the signal x(t) [SI2 ]
µx mean of the signal x(t) [SI]
Abbreviations and Acronyms

ade ADC electrical (noise)
ade ADC quantization (noise)
dae DAC electrical (noise)
dae DAC quantization (noise)
emf electro motive force
pa Preamplifier (noise)
sus Suspension (noise)
coil Geophone coil (noise)
lsb Least significant bit
opamp Operational Amplifier
AC Alternating Current
ADC Analogue to Digital Converter
AIMS Advanced Isolation ModuleS
√
ASD Amplitude Spectral Density [SI/ Hz ]
AVI Advanced Vibration Isolation
CPS Cumulative Power Spectrum [SI2 ]
CAS Cumulative Amplitude Spectrum [SI]
DAC Digital to Analogue Converter
DC Direct Current
DEB Dynamic Error Budgeting
DOF Degree Of Freedom
DSA Dynamic Signal Analyzer
FFT Fast Fourier Transform
MIMO Multi Input Multi Output
NHNM New High Noise Model
NLNM New Low Noise Model
P(I)D Proportional, (Integral,) Derivative
PSD Power Spectral Density [SI2 /Hz ]
SISO Single Input, Single Output
SNR Signal to Noise Ratio [dB]
xii
Chapter 1
Introduction
1.1 Context of this thesis

In modern society the use of high performance mechatronic systems is rapidly increasing.
Systems like DVD players and computers are now part of every day life, and the desire for
even better performing systems is clearly present. Manufacturers are forced to push the
performance of their existing system concepts to the limit and develop new system concepts
that can achieve even higher performance levels. The design of these high end mechatronic
systems forms the context of this thesis.
High performance mechatronic systems have become more and more complex. The system is
often designed per sub-system by different designers, sometimes even at different locations.
The design is often time consuming and very expensive. The modular character and the high
costs of the design process result in an important demand on the design of the system; the
designed system must be first time right. This means that, when the system is finally build,
its performance level must satisfy the specifications. No significant changes are allowed in the
post-design phase! Because of this, the designer wants to be able to predict the performance
of the system a-priori and gain insight in the performance limiting factors of the system. A
design approach that enables designer to do so is the subject of this research.
1.2 Dynamic error budgeting

Mechatronic systems often comprise a certain number of sub-systems and a controller C,
together forming a control loop which enforces a certain desired closed loop system behavior.
The performance of such a system is generally defined by the error made, which is caused by
the disturbances d that act on the system. This idea is illustrated in Figure 1.1 (left hand
side). The error e can be seen as a measure of performance of the system.
In the design of precision machines an approach called error budgeting is often used to
attribute an allowable amount of error to the machine’s different components, see e.g.,
Slocum [21, p.61]. Every disturbance acting on the closed loop system is then allowed
to contribute a certain part of the total error budget. The design is satisfying if the total
1
2 CHAPTER 1. INTRODUCTION
Figure 1.1 left: The error e made by a mechatronic system determines its performance and results
from the disturbances d that act on the closed loop system. right: The error is a measure of
performance of the closed loop system and can be seen as the result of the separate contributions
caused by the different input disturbances.
budget does not exceed the performance specification, if not, the system has to be redesigned.
The basic idea behind this approach is that the error can be seen as the result of the separate
contributions caused by the different input disturbances, as is illustrated on the right hand
side of Figure 1.1. For the designer to be able to predict the performance (indicated by
the right-arrow) and to attribute a part of the error to the system disturbances (indicated
by the left-arrow), he needs to know how each disturbance propagates to the error. In the
error budgeting approach this relation is approximated using rules of thumb, which are (very
often) based on simulations with sinusoidal signals. In practise most disturbances are of a
stochastic nature, to which the rules of thumb do not apply.
However, the propagation of stochastic disturbances to the output can be described exactly
(in theory) when one uses a frequency dependent relation to describe the propagation. This
relation is described in the theory of spectral analysis and uses frequency dependent models
of the disturbances and sub-systems. Because (almost all) the disturbances in high perfor-
mance mechatronic system are stochastic of nature, they can be modelled with their power
spectral densities (PSDs). This results in a new design approach which will be referred to
as Dynamic Error Budgeting (DEB), where ’dynamic’ refers to the use of the frequency
dependent models. This approach enables the designer to accurately budget the error, to
predict the final output error and to point out critical sub-systems and/or disturbances.
1.3 Vibration isolation

To put the DEB design approach to the test, it will be applied to two design cases, both
concerning the vibration isolation problem. Vibration isolation is much used in modern high
precision machines, where the obtained precision would not be possible without the use of
vibration isolation, e.g., diamond turning machines, optical measurement equipment and
gravity wave detection antennas. Another good example where extreme high precision is
needed is in the production of Integrated Circuits (IC). So-called wafer stages move wafers
1.3. VIBRATION ISOLATION 3
of silicon with a repeated accuracy of ±50 nm in order to project the image of the circuit
onto a photoresist layer on the wafer (from [20]). Production of ICs would not be possible
without the use of vibration isolation. As such, vibration isolation receives a lot of research
attention.
Figure 1.2 Vibrations of the ground, will cause systems mounted on that ground to vibrate as
well. Internal flexibilities of the system will result in internal deformations (indicated with the
arrow), which, in general, have a negative effect on the accuracy (performance) of the system.
In order to improve the performance of the system a vibration isolator can be used to isolate the
system (now called the payload) from vibrations.
The goal of vibration isolation is to isolate a payload system from ground vibrations, see
Figure 1.2. Isolation from ground vibrations requires a soft (low stiffness) mounting of the
payload to that ground. However, this results in a high sensitivity to disturbance forces
originating from the payload itself; suppression of payload disturbances forces requires a
high stiffness mounting to the ground, see Harris [9] and Subrahmanyan [22].
When using a passive isolation device this results in a fundamental trade-off between the
rejection of ground vibrations and payload disturbances. Adding active control to the mount-
ing, gives means to improve the performance by dampening of the suspension resonances
in the system. However, when a relative sensor between ground and payload is used in the
control system, the fundamental trade-off between payload and ground disturbance rejection
still exists, limiting the performance. This fundamental limitation in performance can be
circumvented by the use of an absolute measurement of the payload movement. This can be
done by using an accelerometer measuring absolute accelerations, or a geophone measuring
absolute velocity.
With an active control system using an absolute sensor, the fundamental trade-off no longer
exists. The performance is now constrained because of the limited performance of the
components used in the control loop, e.g., acceleration and velocity can only be measured over
a limited frequency range and the active components in the control loop introduce additional
disturbances like e.g., sensor noise and amplifier noise. These disturbances are stochastic of
nature and can be modelled by their PSDs. This gives the designer the opportunity to use
the DEB design approach for the design of the system!
1.4 Challenge definition

In the previous paragraphs it was explained that there is a need for a more accurate and
structured design approach that enables the designer to account for stochastic disturbances
during the design of the system. The active vibration isolation problem is introduced as a
possible field of application for this approach. This leads to the challenge definition of this
thesis:
Develop a tool which enables the designer to account

for stochastic disturbances during the design of a
mechatronic system.
From this challenge definition the following research goals can be formulated:
Develop tools that enable the designer to:
• Predict the final performance level of a system which is subject to stochastic distur-
bances.
• Gain insight in performance limiting factors of the system. This insight should enable
the designer to point out critical system components/properties and to improve the
performance of the system.
• Objectively compare the performance of different system designs. When a system

designer starts off with several different system concepts, he needs to find out which
concept is the most promising.
Show the feasibility of these tools by applying them in the design of two (one DOF) active
vibration isolation systems. Steps to be taken in this research are:
• Investigate if there is a gap between the theoretically predicted results and the exper-
imentally measured results. If there is a gap, try to explain the difference.
• Show that is possible to point out critical components and to improve the system
performance, using the obtained insight in performance limiting factors.
• Show that it is possible to objectively compare the performance of two different sys-
tems.
• Apply the approach to a system which is in the design phase and show the feasibility
of the approach in practice.
1.5. OUTLINE OF THESIS 5
1.5 Outline of thesis

In Chapter 2 the DEB approach is covered in more detail. First, motivations are given that
support the development of this new approach and the design process itself is described.
After that the limitations of the DEB approach are given. What assumptions should be
met and what things should be considered to make sure the theory is applicable? Next,
the theoretical background is given which forms the basis of the approach. The theoretical
definition of performance and power spectral densities is given, and it is described how the
performance can be calculated and analyzed using the theory of propagation. In the last
part of this chapter the theory of H2 control is given, which can be can be seen as an add-on
to the DEB design approach. The theory of H2 control describes how to control the system
in order to maximize its performance given the system and its disturbance models.
To put the approach to the test, the approach is applied in two case studies, both concerning
a design of an active vibration isolator. The first case study covers both chapters 3 and 4.
The second case study is described in Chapter 5.
The focus of the first study is to validate the approach in practice. Is the theory applicable
in practice and does the approach give the designer more design insight? In Chapter 3 an
active vibration isolator is described, modelled and its disturbances identified and modelled
as well. The performance of the system is calculated in theory and compared with experi-
mental results. In Chapter 4 the H2 control strategy is applied to the same system and the
theoretically predicted performance it achieves is validated using experimental results.
The second case study is about an active vibration isolator with a more advanced working
principle. This principle is explained and the system and disturbances are modelled. This
system is still in the design phase and the main goal of this case study is to apply the design
approach in practice as it is developed for.
Finally, in Chapter 6 conclusions and suggestions for further work are given regarding the
DEB design approach and the use of H2 control in this approach.
Chapter 2
Dynamic Error Budgeting
In this chapter the DEB design approach is covered in more detail. First, the motivations
behind the DEB approach are explained in more detail. Next, the DEB process itself is
described and the limitations of the approach are discussed. In the remainder of the chapter
the theoretical background of the approach is given.
2.1 Motivations of the DEB design approach

The main motivation for the development of the DEB approach is the need for the designer
to account for disturbances while designing a high performance mechatronic system, as is
already mentioned in the introduction of this thesis. The following advantages of this ap-
proach can be mentioned.
Cutting costs in the design phase. If the error is not simulated during the design phase,
the final performance level can only be found when a costly prototype is build and the per-
formance can be measured physically. If the performance level is not met, the designer has
to find out what component or disturbance causes the output to exceed the error budget
and then redesign the system, a rather time consuming and costly job. If the error could be
simulated beforehand however, changes can be made when the system is still in the design
phase, cutting down the costs of the system.
Speeding up the design process. Since the simulations can be done on a computer, it
can give a quick indication if a concept is feasible or not. Several concepts can be analyzed
in a short period of time and the most promising concept can be chosen, speeding up the
design process.
Enhancing design insight. If the performance specification is not met, the designer wants
to know which component or what system property is limiting the performance most. The
designer also wants to know how much each disturbance contributes to the output error. For
example, a very expensive low-noise sensor can possibly be replaced with a cheaper sensor
if the simulations show that the cheaper sensor performs well within the specifications. The
7
8 CHAPTER 2. DYNAMIC ERROR BUDGETING
analysis of the simulated error shows all the information needed for the designers to make
these kind of trade-offs, helping the designer to make design decisions.
2.2 DEB design process

The DEB design process can be summarized as follows: choose a system concept and simulate
the output error. If the total error is meets the performance specification, the design is
satisfying. If the error exceeds the specified budget, the designer has to change the system
such that the specification is met. Step by step, the process is as follows:
• Design a concept system.
• Model the concept system, such that the closed loop transfer functions can be deter-
mined.
• Identify all significant disturbances. Model them with their Power Spectral Den-
sity(PSD).
• Define the performance (error) outputs of the system and simulate the output error.
Using the theory of propagation, the contribution of each disturbance to the output
error can be analyzed and the critical disturbance(s) can be pointed out.
• Make changes to the system that are expected to improve the performance level, and
simulate the output error again. Iterate until the error budget is met.
The theoretical background of this approach (e.g., the use of PSDs and theory of propaga-
tion) is covered in § 2.4.
2.3 Assumptions and research boundaries

The DEB approach can only be applied when the system and its disturbances satisfy the
following assumptions:
• It is assumed that the (sub-)system can be accurately described with a linear time
invariant model. This assumption is not expected to cause much problems, for the
following reasons: First of all, in today’s mechatronic system design a lot of effort is
put in to make systems have a linear behavior. Second, high precision mechatronic
machines often comprise a feedback loop, which has a ’linearizing’ effect on the closed
loop behavior. Many high precision machines have a small working range, which makes
it more likely for the system to behave linearly.
Anyhow, if a system does shows non-linear behavior, the deviation from the linear
behavior could be modelled as a disturbance and added to the analysis. This approach
might not be so straightforward as it sounds and is considered outside the scope of
this research.
2.4. THEORETICAL BACKGROUND 9
• The disturbances acting on the system must be stationary; their statistical properties
are not allowed to change over time. In practise this can dealt with by applying some
sort of averaging of the measured PSDs over longer periods of time. For example,
when ground vibrations are modelled, it is customary to measure the worst case PSD
over a few hours, sometimes even a whole day.
• The third assumption also applies to the disturbances and says that the disturbances
are assumed to be uncorrelated with each other. This condition is more difficult
to satisfy under certain circumstances, especially for a MIMO (Multi Input, Multi
Output) systems. For example, it is unlikely that ground vibrations in one direction
are uncorrelated with another direction. The designer must therefore make sure that
the separate disturbances all originate from separate independent sources. The theory
in this chapter can be easily extended to be also valid for correlated signals, however
this is outside the scope of this research.
• The disturbance signals will modelled by their PSD. This implies another assumption
on the disturbances: they must have a defined PSD, so only stochastic disturbances are
allowed. Deterministic components, like sinusoidal and DC signals, in a disturbance
signal give infinite peaks in their PSD, making it meaningless in the DEB approach.
In practise however, there will be periodic disturbance sources like e.g., pickup noise in
electrical circuitry due to electromagnetic waves in the air caused by the periodically
(60 Hz) alternating current of the electricity grid and disturbances originating from
rotating machines, like ground vibrations and acoustic noise. The DEB analysis can
then still be used, but should then only be applied to the stochastic part of the dis-
turbances. For the deterministic part other (straightforward) techniques can be used
to determine their influence to the error. In this research it will be assumed that all
disturbances are stochastic of nature.
• It should be noted that the calculation method makes no assumption on the distri-
bution of the distribution functions of the disturbances. In practise, many (stochas-
tic) disturbances will have a normal (or Gaussian) like distribution. Although not
all disturbances have a normal distribution, the performance output channel is most
likely characterized accurately with a normal distribution; the performance channel is
the sum of contributions by many disturbances, the Central Limit Theorem, see e.g.,
Priestley [16, p95] and Papoulis [14, p266] then states that the output will approach a
normal distribution for an increasing number of disturbances.
2.4 Theoretical background
The theory covered in this section is discussed in more detail in Monkhorst [12]. Here, a
short summary of the results is given.
2.4.1 Variance as a performance measure

The term ’performance of systems’ is a very general term and needs interpretation. In this
thesis the performance of a system is measured by the variance of the error signal. The
variance gives an interpretation of how much power is spread around the mean of the signal,
cancelling out the power of the DC component. A (deterministic) DC component has an
infinite power density (at 0 Hz), making it impossible to represent it with a PSD. Therefore
it can not be accounted for in the DEB approach. In the remainder of this thesis, all signals
are assumed to have a zero mean. The variance σx2 of a signal x(t) is defined as the signal
power or mean square x¯2 minus the squared mean µ2x :
variance σx2 = x¯2 − µ2x (2.1)

Z T
1
power x¯2 = lim x(t)2 dt (2.2)
T →∞ 2T −T
Z T
1
mean µx = lim x(t)dt (2.3)
T →∞ 2T −T
In this thesis the term power refers to this mathematical definition (2.2) of signal power [SI2 ]
and not to physical power, which is expressed as energy per unit of time [J/s].
2.4.2 Parseval’s Theorem: linking the frequency to the time domain

In this thesis all disturbances and noise sources will be modelled using a Power Spectral
Density function (PSD). Before the definition of a PSD is given in § 2.4.3, the relation be-
tween energy in the time domain and energy in the frequency domain will be described.
For non-periodic finite duration signals the energy in the time domain is described by:
Z ∞
Energy: = x(t)2 dt (2.4)
−∞
Parseval’s Theorem now states that energy in the time domain equals the energy in the
frequency domain as follows:
Z ∞ Z ∞
2
Energy: = x(t) dt = |X(f )|2 df (2.5)
−∞ −∞
where X(f ) is the Fourier Transform of the time signal x(t) and is given by:
Z ∞
X(f ) = x(t)e−2πjf t dt (2.6)
−∞
2.4.3 Power spectral density function

Every disturbance signal that enters the system has a certain variance (or power) level,
defined by (2.1). However, since the propagation transfer functions of a system are likely
to be frequency dependent, one needs the power distribution over frequency to describe the
propagation (see § 2.4.4). The power distribution over frequency of a time signal x(t) is
described by the PSD denoted with Sx (f ). A PSD is a power density function with units
[SI 2 /Hz], meaning that the area underneath the PSD curve equals the power (units [SI 2 ])
of the signal (SI is the unit of the signal, e.g., m/s). When the signal has a zero mean, the
area of the PSD equals the variance of the signal. A DC component, or other deterministic
signal content, would cause the PSD to have infinite peaks, so a PSD can only reliably
represent the power density of a stochastic signal.
One can also use a Power Spectrum (PS) to represent the signal power distribution. A PS
has units [SI 2 ] and can only be defined on a discrete number of frequency points, since
it would otherwise define an infinite power contents of the signal. A PS can only reliably
represent deterministic signals like a sinusoid or a DC component, because the magnitude of
a PS representing a stochastic signal would be dependent of the resolution of the frequency
grid. In this research only stochastic signals are considered, so the PS is not used in this
thesis.
Using the definition of signal power (x¯2 ) and Parseval’s theorem we can link power in the
time domain with power in the frequency domain:
Z T Z ∞
1 2 1
power = lim xT (t) dt = lim |XT (f )|2 df
T →∞ 2T −T T →∞ 2T −∞
Z ∞ µ ¶
|XT (f )|2
= lim df (2.7)
−∞ T →∞ 2T
where XT (f ) denotes the Fourier transform of xT (t), which equals x(t) on the interval
T ≤ t ≤ T , and is zero outside this interval. We will refer to the term within the integral
as the two-sided power spectral density Sx (f ), since its integral over frequency equals the
signal power.
|XT (f )|2
Sx two (f ) = lim , −∞ ≤ f ≤ ∞ (2.8)
T →∞ 2T
In practice the one-sided PSD is used, which is only defined on the positive frequency axis,
0 ≤ f ≤ ∞, but also contains all the power. The one-sided PSD is therefore simply defined
as the double of the two-sided PSD:
|XT (f )|2
Sx one (f ) = lim , 0≤f ≤∞ (2.9)
T →∞ T
In the remainder of this report, Sx (f ) refers to a one-sided PSD.
Discrete formulation
For discrete time signals the one-sided PSD estimate is defined as:
|XL (fk )|2

Ŝ(fk ) = (2.10)
LTs
where L equals the number of time samples and Ts the sample time, XL (fk ) is the N -point
discrete Fourier Transform of the discrete time signal xL [n] containing L samples:
N
X −1
XL (fk ) = xL [n]e−j 2πkn/N (2.11)
n=0
It is wise to choose the number of frequency points N larger than the time vector length
L such that N is the next power of two larger than L. To evaluate XL [fk ], we simply pad
xL [n] with zeros to length N .
The mean power of the signal can be estimated with the following sum:
N
X −1
|XL (fk )|2
power ≈ (2.12)
L2
k=0
The term after the summation symbol is called the two-sided discrete Power Spectrum and
can be easily computed using the MATLAB° r
command fft, which computes the two-sided
Fast Fourier Transform. Appendix A.1 shows the code to compute the two sided PSD from
a sampled time signal.
2.4.4 Propagation of disturbances

Key to the DEB design approach is to realize that the final performance of the system is
determined by all the disturbances that act on the closed loop system. These disturbances
apply at different locations in the closed loop system, as illustrated in Figure 2.1. The
disturbances w(t) are assumed to be stochastic, and can therefore be easily modelled with
their PSD (In this thesis, the PSD of a signal is denoted by a dashed box, this is not a dynamic
filter!). Examples of disturbances are amplifier noise, ground vibrations, sensor noise, etc..
This section gives the mathematical description of how each disturbance propagates through
the closed loop system and contributes to the output error.
Figure 2.1 Schematic of a closed loop system with various disturbances. The disturbances w(t)
can be modelled with their power spectral density function Sw (f ), denoted with a dashed box.
The theory presented here is described in Papoulis [14], Balmer [1], Priestley [16] and Har-
ris [9]. As said, the disturbance signals zi can be modelled using PSDs, which will be denoted
by Swi (f ). To calculate the performance of the system, we consider the closed loop system
as a black box system, with transfer function H(jω) = H(j 2πf ):
    
z1 H11 · · · H1n w1
 ..   .. .. ..   ..  =
 . = . . .  .  (2.13)
zk Hk1 · · · Hkn wn
where the input w stacks n disturbances, and output z stacks k performance outputs. For
multi-variable systems, see Ljung [11], the PSD of the output is then given by:
Sz (f ) = H(j 2πf ) Sw (f ) H t (−j 2πf ) (2.14)

where Sz is a matrix with the PSDs of the performance signals z1 , . . . , zk on the diagonal.
When that the disturbance inputs w(t) are uncorrelated, the input PSD Sw will be diagonal,
which simplifies the calculations. Since we are only interested in the PSDs of the performance
signals, we only look at the diagonal terms of Sz , which follow from (2.14). The PSD of the
jth performance output Szj (f ) due to all the noise PSDs Swi (f ), is calculated by propagating
all the corresponding input PSDs to that output. Since Hji (j 2πf ) Swi (f ) Hjit (−j 2πf ) equals
|Hji (j 2πf )|2 Swi (f ), the PSD of the jth output can be written as:
k
X
Szj (f ) = |Hji (j 2πf )|2 Swi (f ) (2.15)
i=1
In words; the PSD of a performance channel is the linear sum of each disturbance PSD
multiplied with the squared magnitude of the corresponding SISO transfer function. Equa-
tion 2.15 offers the designer an important opportunity to analyze the error and to point
out critical components or system properties, since it describes the PSD of the performance
signal in terms of input disturbance PSDs and closed loop system transfer functions. This
calculation method was implemented in the MATLAB° r
environment.
Now it follows from (2.1), (2.7) and 2.9 that

Z ∞
¯2
z = Sz (f )df (2.16)
0
In words: since a PSD describes the power density over frequency [SI2 /Hz ], the integral of
the output PSD Sz (f ) over frequency equals the power z¯2 of the performance signal z(t).
Because all signals are assumed to have zero mean, this integral equals the variance (2.1) of
the performance signal:
Z ∞
2
σz = Sz (f )df (2.17)
0
where Sz (ω) is the one-sided PSD of the signal. Another very useful function in performance
analysis is the Cumulative Power Spectrum (CPS), which is defined as:
Z fo
Cz (fo ) = Sz (f )df (2.18)
0
where Sz (f ) is the corresponding one-sided output PSD. The CPS is defined from fo = 0 to
fo = ∞ and is a monotone increasing function with its magnitude increasing from zero to
Cw (∞) = σz2 . Because the final value of the CPS equals the variance of the signal, it clearly
indicates how much power is located at which frequency.
The ASD and CAS
In this thesis we will not analyze the PSD and CPS of the performance signal, but at their
√
square rooted versions (element-wise): the Amplitude Spectral Density (ASD) [SI/ Hz ] and
the Cumulative Amplitude Spectrum (CAS) [SI], which actually contain the same information.
In formula:
p
ASD(f ) = |P SD(f )| ∀ f (2.19)
p
CAS(f ) = |CP S(f )| ∀ f (2.20)
A motivation to use the CAS is that its final value, at (f = ∞), equals the standard deviation
of the signal, so e.g., σz . Since performance specifications are often given in terms of a σ-
value, using the CAS makes it easier to compare the results with the specifications. Besides
this, the standard deviation is easier interpretable in the time domain than the variance. For
example, the 3-σ value gives a good approximation of the maximum amplitude of a stochastic
signal. A motivation to use the ASD instead of the PSD is that its units correspond better
to the CAS, which is more a matter of taste.
2.5 H2 control, maximizing performance

2.5.1 Use of H2 control in DEB
Precision machines often use a feedback control loop to achieve a certain system behavior.
When the used controller C is not fixed, it acts as a degree of freedom in the closed loop
transfer function defining the disturbance propagation as described in § 2.4.4. This idea
is illustrated in Figure 2.2. One can use this degree of freedom in the system in order to
minimize the output error, giving the designer an upper bound on the achievable system
performance. The controller that achieves this goal, is called the H2 controller CH2 and is
unique given the system and its disturbances.
The resulting H2 controller is also very useful to the control engineer, since it shows how to
control the system optimally, given the disturbances. In the following sections the theory of
H2 control is explained.
2.5. H2 CONTROL, MAXIMIZING PERFORMANCE 15
Figure 2.2 The controller C acts as a degree of freedom in the closed loop transfer function H̄,
and therefore in the determination of the output error.
2.5.2 The H2 norm and variance of the output

The H2 norm is a norm defined on a system and can be stated as follows:
Z ∞
2
kHk2 = |H(j 2πf )|2 df (2.21)
−∞
In order to relate the H2 system norm to the performance definition used in this thesis,
we will look at the stochastic interpretation of the H2 norm; the squared H2 norm can be
interpreted as the output variance of a system with zero mean white noise input. So, the
value obtained in the (2.21) equals the output variance of the system H(j 2πf ) having zero
mean white noise input.
When the total system with its disturbance models is defined as a system with white noise
inputs, the H2 controller will minimize the output variance of that system. In § 2.5.4 it is
explained how the disturbance models Sw can be included in the system using weighting
filters with zero mean white noise input. First, the H2 control problem is formulated.
2.5.3 The H2 control problem

The H2 control problem can be stated as follows: ’Find a controller CH2 which minimizes
the H2 norm of the closed loop system H’, or expressed in a formula:
CH2 ∈ arg min kH(j 2πf )k2 (2.22)

C
H2 controller synthesis can be done using various methods. One of them is solving two
Ricatti equations as described in Doyle (et.al.) [6], using the MATLAB°
r
command hinfsyn.
In order to synthesize an H2 controller, one needs to model the open loop system as a
generalized plant and the generalized plant needs to meet several standard assumptions as
described in Appendix A.2. But, before the generalized plant can be set up, the disturbance
models need to be included in the open loop model, which is done in the next section.
2.5.4 Using weighting filters to model disturbances

In order to synthesize an H2 controller that will minimize the output error, the total system
including disturbances needs to be modelled as a system with zero mean white noise inputs.
This is done by using weighting filter Vw (j 2πf ), of which the output signal has the previ-
ously discussed PSD Sw (f ) when the input is zero mean white noise. This is illustrated in
Figure 2.3.
Figure 2.3 The use of a weightingfilter Vw (f ) [SI] to give the weighted signal w̄(t) a certain
PSD Sw (f ).
The white noise input w(t) is dimensionless [-], and when the weightingfilter has units [SI],
the resulting weighted signal w̄(t) has units [SI]. The PSD Sw (f ) of the weighted signal is
now given by (also see Equation 2.14):
|Sw (f )| = Vw (j 2πf ) Sin (f ) Vwt (−j 2πf ) (2.23)
Since the PSD of a white noise signal Sin (f ) equals unity (or the identity matrix in a
multivariate case) and has units [−/Hz], and Vw has units [SI], Sw (f ) now indeed has the
units of a PSD [SI2 /Hz ].
Since the disturbances are assumed to be uncorrelated Sw (f ) is diagonal, so the resulting
input weighting filter Vw (f ) will also be diagonal. Given Sw (f ), Vw (f ) can be obtained using
a technique called spectral factorization. However, the step of spectral factorization can be
avoided by modelling the disturbances directly in terms of a weighting filter, which is the
strategy followed in this research.
H2 controller synthesis also allows to use output weighting filters. We will use this possibility
to scale different defined outputs relative to each other as will be explained in more detail
in § 2.5.5. The output weighting filter is denoted by Wz (f ).
In Figure 2.4 it can be seen that the generalized plant G now consists of the open loop plant
Ḡ in series with the weighting filter Vw (f ) and output scaling filter Wz (f ).
Figure 2.4 The open loop system Ḡ in series with the diagonal input weighting filter Vw and
diagonal output scaling filter Wz defining the generalized plant G.
Now that a generalized plant is defined, is has to meet the assumptions in Appendix A.2 in
order for the H2 controller synthesis algorithm to work. If the assumptions are not met, the
weighting and/or scaling filters have to be adjusted appropriately.
2.6. CONCLUSIONS ON THE THEORY 17
2.5.5 Output scaling and the Pareto curve

The closed loop system can have more than one output. In this research the performance
(error) signal e and the controller output u are defined as the outputs of the closed loop
system, see Figure 2.5. In this way the system designer can analyze how much control effort
(defined here as the variance of the controller signal u) is used to achieve the performance
level at the performance output. When the system designer designs several different system
concepts, he wants to compare them by their maximum achievable performance, to decide
what concept is the most promising. When he wants to make an objective performance
comparison, he can choose the amount of control effort to be equal for each system concept.
The resulting problem is actually a multi-objective control problem; while constraining the
variance of the controller output signal u the variance of the performance channel should
be minimized. This problem can be solved by scaling e.g., the controller output u with a
factor α during H2 controller synthesis, such that the control power used by the different
H2 controllers is equal (see Boyd [3]). The output scaling factor α is then contained in the
output scaling filter Wz (f ). The designer now only has to find the right scaling factor α by
iteration.
Figure 2.5 The closed loop system, with weighting filters included. The system has n disturbance
inputs and two outputs; the error e and the control signal u. The H2 controller minimizes the
H2 -norm of this system.
When varying the scaling factor α, one can plot the amount of control effort at one axis
and the achieved performance on the other axis. The resulting points lie on the so-called
Pareto curve (see Boyd [3]) . Every point on the curve corresponds to a certain scaling
α and describes the control effort needed to achieve the corresponding performance level.
Examples of Pareto curves can be found in Sections 4.2 and 4.4.
2.6 Conclusions on the theory

The theory covered in this chapter gives the designer a tool with which he can model stochas-
tic disturbance sources, predict the performance and analyze the resulting error. He is also
able to point out critical transfer functions and/or disturbance sources. All the tools to
make the DEB design approach work in theory are now available.
H2 controller synthesis offers the designer the opportunity to optimize over the degree of
freedom given by the controller. This enables the designer to calculate the maximum per-
formance level that is achievable with a certain system concept (given its configuration and
disturbance models). With the use of controller output scaling during H2 controller synthe-
sis, the designer is capable of objectively comparing different system concepts by synthesizing
H2 controllers that use the same amount of control effort in each case.
Chapter 3
Case study, active vibration

isolation
3.1 Introduction
Chapters 3 and 4 describe the research done during a five months stay at the Massachusetts
Institute of Technology in Boston, U.S.A. in the Precision Motion Control laboratory of
Professor D.L. Trumper. The main challenge taken in the research at MIT is to validate
the dynamic error budgeting design approach (as theoretically described in Chapter 2) in
practice. The main question that is tried to answer is: Does the Dynamic Error Budgeting
(DEB) design approach also work in practice?
In order to apply the approach in practice, a complete vibration isolation setup is to be

designed and build. The system that was build (see 3.1) is an improved version of the setup
described in Trumper [24]. The practical implications of building the setup is not further
discussed here. The goal of this chapter is to apply the DEB approach to the design of this
system and identify the performance limiting factors and use this information to improve
the performance of the system. Finally, the theoretical results must be compared with ex-
perimental results to validate the theoretically predicted performance of DEB approach and
find out what the cause is of the possible gap between theory and practice.
The experimental setup is described and modelled in Section 3.2. In Section 3.3 all distur-
bances that act on the system are modelled as stochastic signals. With all the necessary
system and disturbance models obtained, the performance of the experimental setup can be
simulated and analyzed using the DEB approach, which is done in Section 3.4. The insight
that is obtained with the DEB analysis is used to redesign the system; in Section 3.5 a
preamplifier is added to the configuration and the performance improvement is analyzed.
The theoretical results are compared with experimental results in Section 3.6. The chapter
ends with conclusions on applying the DEB design approach in this case study.
19
20 CHAPTER 3. CASE STUDY, ACTIVE VIBRATION ISOLATION
3.2 An active vibration isolation platform

3.2.1 Description of the experimental setup
The experimental setup used in this case study can be seen in Figure 3.1, and consists of an
8-inch speaker as a one-axis vertical actuator and a geophone as an absolute velocity sensor.
In the center of the speaker a cone plate is glued on which the geophone is rigidly mounted.
The cone plate represents the payload location, and is to be isolated from vibrations.
Figure 3.1 The experimental speaker-geophone setup. An 8 inch speaker is rigidly connected
to a base plate with aluminum struts. A velocity sensor (geophone) is mounted in the speaker
cone. (The sensor on the base plate is not part of the actual setup, but was used to measure
the input vibration level).
A schematic configuration of the speaker and geophone can be seen in Figure 3.2. Big
permanent magnets provide a permanent magnetic field. When a current is sent through
the voice coil a Lorentz force Fs is created which can move the cone up or down. Ground
vibrations enter the system via the base plate (Xb ) and propagate through the speaker via
the suspension, and cause the payload (Xl ) to vibrate.
In Figure 3.3 a block scheme of the complete active vibration isolation system is given. The
system was controlled using a dSpace1104 board. The voice coil is voltage driven by a Crown
DC300R power amplifier. Psp geo (s) denotes the transfer function from the power amplifier
input Ein to geophone output voltage Eg . Psp vib (s) denotes the open loop speaker model
from base velocity Ẋb to payload velocity Ẋl , and describes the open loop propagation of
3.2. AN ACTIVE VIBRATION ISOLATION PLATFORM 21
Eg
suspension
geophone
Xl
aluminum cone plate
strut cylinder strut
magnet voice coil magnet

Xb steel Esp
Fs
base plate
Figure 3.2 Schematic of speaker-geophone setup. Big permanent magnets provide a permanent
magnetic field. When a current is sent through the voice coil a force Fs is created which can
move the cone up or down. In the center of the speaker cone a cone-plate is glued in. A velocity
sensor (geophone) is mounted rigidly on the cone-plate. The cone plate is to be isolated from
vibrations.
ground vibrations into the cone plate. The dynamics of the sensor are encompassed by
Pgeo (s).
3.2.2 Performance definition

To analyze the performance of the vibration isolation setup, one needs a performance variable
as a model output. In this research the variance of the payload velocity Ẋl signal is chosen
as the performance variable. To extract the performance variable Ẋl from the model, the
transfer function Pspeaker (s) is needed, as can be seen in Figure 3.3. To obtain Pspeaker (s),
Psp geo (s) can be multiplied with the inverse of the geophone dynamics Pgeo (s), such that
−1 (s) = Ẋ /E . P
Pspeaker (s) = Psp geo (s) Pgeo l in sp geo (s), Psp vib (s) and Pgeo (s) are all modelled
in the next section.
3.2.3 System modelling

Vibration isolation platform
A vibration isolation platform very similar to this setup is also modelled in Trumper [24],
where a 4th order state space model is developed to describe the platform dynamics. This
state space model was initially used to model the setup, but it turned out that the state
space model differed from empirical results significantly in a crucial frequency region as will
be explained at the end of this paragraph. (also see Figure 3.4). Since the model of the
setup is part of the loop transfer function L(s), and its accuracy will therefore influence the
stability of the loop, it is chosen to base the model for the setup on the empirical data. The
derivation of the state space model from Trumper [24] for this setup is done for modelling
insight anyway and can be found in Appendix B.1.
Figure 3.3 A block scheme of the experimental setup. Ground vibrations enter the loop via the
transfer function Psp vib (s), which defines the propagation of vibrations from the base plate to
the cone plate. The cone plate velocity is defined as the performance variable. The system is
controlled using a dSpace1104 controller board. Psp geo (s), Pspeaker (s) and Pgeo (s) define the
transfer function of the speaker-geophone setup in series with the power amplifier.
To model the transfer function Psp geo (s) (see Figure 3.3), a swept sine experiment was done
using an HP35665 dynamic signal analyzer and the Crown power amplifier. The gain of the
power amplifier (Esp /Ein ) was set to unity for modelling simplicity. The transfer function
between the power amplifier input Ein and geophone output Eg was measured, on which
a least square fit was made. The measured data and the resulting fitted transfer function
Psp geo (s) can be seen in Figure 3.4 on the left hand side.
Speaker−Geophone transfer function Pspeaker from ls fit and from ss model

0
Measured data
Magnitude [dB]
Magnitude [dB]
0 Fitted model
−50
−50
−100
−2 −1 0 1 2 3
−100 −1 0 1 2 3
10 10 10 10 10 10 10 10 10 10 10
180 180
Phase [degrees]
Phase [degrees]
90 90
0 0
−90 −90
−180 −180 Pspeaker from ls fit
−270 −270 Pspeaker from ss model
−360 −2 −360
−1 0 1 2 3 −1 0 1 2 3
10 10 10 10 10 10 10 10 10 10 10
Frequency [Hz] Frequency [Hz]
Figure 3.4 left: The fitted transfer function Psp geo (s) together with the measured frequency
response of the vibration isolation setup Eg /Ein . right: The fitted model Pspeaker (s) and the
transfer function Esp = Ein to Ẋl extracted from state space model.
The low frequency region, where the magnitude of the measured transfer function seems to
level off, needs some special attention. Since the graph represents the transfer function from
power amplifier input voltage Ein towards geophone output voltage Eg , the graph suggests a
DC gain, which implies a certain DC geophone output voltage when a DC voltage is applied
to the power amplifier. A constant geophone output voltage represents a constant velocity
of the geophone coil relative to its casing, which is not possible in reality. It is not exactly
known what causes the transfer function to level off, but the flat part in the low frequency
region is not believed to represent the true system dynamics. The true dynamics are believed
to follow the +3 slope down towards the lower frequencies. In order to match the fit with
believed reality, all the poles and zeros with an absolute value smaller than 1 are placed in
the origin. The resulting pole/zero pairs in the origin cancel out when obtaining a minimal
realization, using the MATLAB°r command minreal. The fitted model now has a +3 slope,
starting the origin, and is given by:
(s − 1.20 · 104 )(s + 3456)(s − 3.33 · 10−3 )

Psp geo (s) = 5.20 · 10−3 ·
(s + 1058)(s2 + 19.93s + 936.9)
(s2 + 3.33 · 10−3 s + 1.11 · 10−5 )(s2 − 8757s + 4.94 · 108 )

(3.1)
(s2 + 57.42s + 8423)(s2 + 2012s + 1.82 · 108 )
In the 5 Hz region the geophone dynamics can be seen (see § 3.2.3) and the speaker reso-
nance frequency around 15 Hz. The peak around 2 kHz is thought to be due to a structural
resonance of the speaker system (there is some flexibility between the voice coil cylinder and
the aluminum cone plate, see Figure 3.2).
On the right hand side of the Figure 3.4, the model Pspeaker (s) can be seen together with
the transfer function Ẋl /Esp extracted from state space model derived in Appendix B.1.
The mismatch between the two models is clearly visible. The difference in phase shift at
frequencies above 100 Hz is crucial for loop stability, since the loop transfer function L(s)
cross-over frequency will be located in this region (see § 3.2.5). Therefore it is chosen to use
the fitted model instead of the state space model.
Ground vibration propagation model Psp vib (s)
Because the vibration propagation transfer function Psp vib (s) is not part of the loop transfer
function L(s) and thus will not influence the stability of the loop, a less accurate model can
be used to model Psp vib (s). Next to that, Psp vib (s) describes the propagation of the ground
vibration model towards the loop, and this ground vibration model itself is a rather rough
approximation of reality as well (see § 3.3.1), so there is no issue of using a high order
propagation model. Using the speaker parameters found in Appendix B.1 a second order
model can be derived in terms of the speaker parameters:
Nf2 /Rs s + kl
Psp vib (s) = (3.2)
ml s2 + Nf2 /Rs s + kl
This model assumes that the geophone dynamics have no significant contribution to the
speaker dynamics. This conclusion can be drawn from the two models shown in Figure 3.4
on the right hand side. The geophone dynamics (resonance frequency at 4.8 Hz) are not
visible in the extracted transfer function Ẋl /Ein from the state space model. Psp vib (s) is
shown in Figure 3.8 on the right hand side.
Geophone
A geophone is a mechanism containing a proof mass which is suspended by leaf springs.

The proof mass consists of a cylinder with a copper wire coil that moves in a magnetic field
produced by permanent magnets fixed within the case of the geophone, see Figure 3.5. A
movement of the coil generates a voltage Eg across coil output terminals, which is propor-
tional to the velocity of the proof mass relative to the geophone’s casing.
Geophone
Eg
leaf spring steel casing
cilinder
coil
magnet Xg
coil
Xl
steal cap
Figure 3.5 left: Photograph of a geophone cut open. The copper coil is clearly visible. right:
Schematic of a geophone, which generates a voltage Eg proportional to the relative velocity
between the outer casing (Ẋl ) and the coil (Ẋb ).
The geophone used in the setup is the GS11D 4.5Hz with a 4000Ω coil. A one degree of
freedom model, describing the velocity sensitivity of the geophone is developed in Barzilai [2]
and can be written as:
Eg Gg s2
Pgeo (s) = (s) = 2 (3.3)
Ẋl s + 2ζωo s + ωo2
and for the acceleration sensitivity we have:
Eg Gg s
(s) = 2 (3.4)
Ẍl s + 2ζωo s + ωo2
where Eg is the geophone output voltage, and Ẋl and Ẍl are the velocity and acceleration
of the payload respectively. Gg = 100 V /(m/s) is the generator constant, ωo = 2πfo =
2π(4.5 ± 0.75) is the natural frequency in rad/sec and ζ = 0.35 the damping ratio of the
geophone. The natural frequency of the geophone used in this setup turned out to be 4.8
Hz. The velocity and acceleration sensitivity can both be seen in Figure 3.6.
Geophone Velocity Sensitivity Geophone Accelaration Sensitivity
1
10
2
10
[V/(m/s)]
[V/(m/s2)]
0
10
1
10
−1
0 1 2
10 0 1 2
10 10 10 10 10 10
Figure 3.6 left: The velocity sensitivity Pgeo (s) of the GS11D 4.5Hz geophone. right: The
acceleration sensitivity of the GS11D 4.5Hz geophone.
3.2.4 Mutual inductance of the speaker and geophone
Since the speaker and the geophone both contain coils, a phenomena called mutual inductance
can occur. Mutual inductance is based on the idea that alternating magnetic field lines
originating from one coil, generates voltages in another coil. This effect can cause the speaker
and the geophone to be magnetically AC coupled. In this case, the magnetic field lines
produced by the voice coil then dynamically couple in the geophone dynamics. During the
swept sine experiments discussed in § 3.2.3 a magnetically shielded geophone was used. The
swept sine experiment was repeated with an unshielded geophone. The shielded geophone
is covered with an extra layer of shielding material as can be seen on the left hand side of
Figure 3.7.
In the low frequency region the magnetic coupling has a clear effect as can be seen on the
right hand side of Figure 3.7. The unshielded version shows a zero at 0.3 Hz and a much
higher gain at the lower frequencies. Also, the phase shifts 180◦ at the frequency of the new
zero. The alternating voltages in the speaker voice coil apparently couple in the geophone
coil at frequencies below 1 Hz, causing the geophone output voltage not to represent the
relative velocity between the casing and proof mass. Since this transfer function is part of
the loop gain, this phenomena can have a dramatic effect on the stability of the control loop
and must be avoided.
3.2.5 Closing the loop
The controller design for a vibration isolation platform like the one used in this setup, is
covered extensively in Trumper [24]. This will not be repeated here, but a short summary
of the results is given instead.
The controller used to control this setup is denoted as C(s) and can be written as:
Speaker−Geophone Magnetic Coupling
Magnitude [dB]
0
−50
Swept sine data
−100 Unshielded swept sine data
−1 0 1 2 3
10 10 10 10 10
Frequency [Hz]
180
Phase [degrees]
90
0
−90
−180
−270
−360
−1 0 1 2 3
10 10 10 10 10
Frequency [Hz]
Figure 3.7 left: Photograph of a shielded geophone cut open (left) next to an unshielded
geophone (right). The shielded geophone is covered with an extra layer of magnetic shielding
material. right: Magnetic coupling of the speaker voice coil and the geophone coil. The magnet-
ically unshielded geophone dynamically couples with the speaker voice coil, causing a zero and a
180◦ phase shift at 0.3 Hz.
µ ¶µ ¶2 µ ¶2 µ ¶3 µ ¶
τ0 s τ1 s + 1 τp s + 1 τ2 s + 1 1
C(s) = Kp Gpa (3.5)
τ0 s + 1 α 1 τ1 + 1 α p τp + 1 α 2 τ2 + 1 τ3 s + 1
with the parameters: Kp = 64 dB, τ0 = 100, t1 = 0.5, a1 = 15, Gpa = 1000, αp = 10,
τp = 1/(2π4.5), τ2 = 0.01, α2 = 3 and τ3 = 1.355 · 10−4 .
An important transfer function for control design is the loop transfer function L(s). The
loop transfer is defined as the transfer function when going around the loop. In this case
L(s) can be written as:
L(s) = C(s) Pspeaker (s) Pgeo (s) = C(s) Psp geo (s) (3.6)
The loop transfer functions then defines the loop sensitivity function S(s) = (1 + L(s))−1 ,
which defines the vibration isolation performance of the closed loop system as explained
below. L(s) and S(s) are both shown in Figure 3.8 on the left hand side.
The open loop speaker model Psp vib (s) and closed loop speaker model Psp vib (s) S(s) define
the propagation of ground vibrations towards the payload in the open loop respectively,
closed loop case. Both transfer functions are shown in Figure 3.8 on the right hand side.
Since loop sensitivity function S(s) makes the difference between the open loop and closed
loop vibration propagation (Psp vib (s) vs. Psp vib (s) S(s)), it defines the increase in vibration
isolation performance of the closed loop setup. Because S(s) = (1 + L(s))−1 , S(s) will cause
3.3. MODELLING OF DISTURBANCE SOURCES 27
an isolating effect when 1 + L(s) has a gain greater than unity, as can be seen from in
Figure 3.8 on the left.
It is common to refer to Psp vib (s) and Psp vib (s) S(s) as the open resp. closed loop trans-
missibility of the system. Comparing the open and closed loop transmissibility in Figure 3.8
(right) we can conclude that the closed loop system now isolates better than the open loop
system in the 100 mHz - 100 Hz region.
Loop Sensitivity and Loop Gain, 100 mHz bandwidth controller Vibration Isolation, 100 mHz bandwidth controller
3 1
10 10
Loop Sensitivity Open loop
Magnitude [dB]
Magnitude [dB]
1
Loop gain Closed loop
10
−1
10
−1
10
−3 −3
10 −2 −1 0 1 2 3
10 −2 −1 0 1 2 3
10 10 10 10 10 10 10 10 10 10 10 10
180 90
Phase [deg]
Phase [deg]
0
0
−180
−90
−360
−540 −2 −1 0 1 2 3
−180 −2 −1 0 1 2 3
10 10 10 10 10 10 10 10 10 10 10 10
Figure 3.8 left: Loop transfer function L(s) and loop sensitivity function S(s). right: Open
loop speaker model Psp vib (s) and closed loop speaker model Psp vib (s) S(s), which are referred
to as the open, respectively closed loop transmissibility of the system.
3.3 Modelling of disturbance sources

In the real world nothing works perfectly. Most components of the setup contribute a cer-
tain amount of noise to the loop, which will set a limit to the achievable performance of
the system. Examples of noise sources are electrical noise sources, such as random voltage
sources and current sources originating in electrical components, e.g., resistors and opera-
tional amplifiers. In Figure 3.9 the noise sources for this setup can be seen in a schematic.
The ground vibrations ground vib represent the ground vibrations entering the loop via the
speaker base plate. The thermal noise from the coil of the geophone is denoted by geo-
coil noise. Brownian noise is a voltage noise caused by the Brownian motion of the proof
mass. ADE and ADQ represent the A/D Converter (ADC) electrical and quantization noise
respectively. (DAQ and DAE similar, only for a D/A converter). All these noise sources are
modelled in the paragraphs below.
The Crown power amplifier has very low-noise characteristics. Its input equivalent standard
deviation equals 1.3 µVrms over a 20 Hz - 20 kHz bandwidth, according to the manufacturer’s
specifications. This will turn out to be so low compared to e.g., the DAE electrical noise
(applying at the same location), that is is not accounted for in this analysis.
!!"#
Figure 3.9 A schematic of all the noise sources in the experimental setup.
3.3.1 Measurements and model of ground vibrations

Using the HP35665 DSA the PSD [Vrms 2 /Hz] was measured of the output voltage of a
geophone that was glued to the base plate of the setup (see Figure 3.1). To convert the
obtained data to equivalent ground velocity data [(m/s)2 /Hz], the PSD was divided by
the squared magnitude of the geophone velocity sensitivity (3.3). The square root of the
resulting equivalent
√ ground velocity PSD data is called the Amplitude Spectral Density
(ASD) [m/s/ Hz] and can be seen in Figure 3.10 on the left hand side.
Remarkable is the steep rise around 2 Hz. Apparently, most of the power of the ground
vibrations is located above this frequency. In the same figure the velocity equivalent channel
noise of the HP35665 DSA is plotted. Comparing the ground vibration data with the channel
noise, it can be concluded that the measured vibration data is not reliable below 1 Hz,
because the magnitude of the channel noise exceeds the vibration data. (It is possible to
overcome this channel noise problem, by pre-amplifying the geophone signal before it enters
the DSA, as was done in Barzilai [2]).
The New High Noise Model (NHNM) and New Low Noise Model (NLNM) of Peterson [15]
are also included for reference. It represents the largest and lowest seismic noise measured
at seismometer stations across the globe. The seismic noise at the Precision Motion Control
laboratory which in the midst of Boston city is expected to be in the order of the NHNM. The
line parameters to construct the NHNM and NLNM curves can be found in Appendix B.2.
The ground velocity ASD is approximated with a 4rd order transfer function, also shown in
Figure 3.10. The model overestimates the acceleration level at most frequencies, especially
in the 1 Hz and 30 Hz regions. The measurements are not reliable below 1 Hz, and √ the
model is assumed to levels off in between the NHNM and NLNM, at 2 · 10−7 m/s/ Hz.
The ASD model can be written as:
Ã !3 Ã !
3 100 √
2π s + 1
Vvib (s) = 2 · 10−7 3
2π
[m/s/ Hz] (3.7)
2π7 s + 1 s + 100
2π
The model that approximates the ground velocity PSD can now be written as:
1/2
Ground Velocity Model and other levels [m/s/Hz ] ASD of Brownian/Suspension noise, σ = 43.9nVrms
−7
10
−4
10
−8
10
−6
10
[m/s/Hz1/2]
[Vrms/Hz1/2]
−9
−8 10
10
−10
−10
10 10
Measured Velocity Level
Ground Velocity Model
Channel Noise HP35665
Peterson New High Noise Model vel
−12 Peterson New Low Noise Model vel −11
10 −2 −1 0 1 2 3
10 −2 −1 0 1 2 3
10 10 10 10 10 10 10 10 10 10 10 10
√
Figure 3.10 left: ASD of the measured ground velocity level [m/s/ Hz] and the obtained model
Vvib (jω), together with the channel noise of the HP35665. The New High/Low Noise Models
of Peterson are shown for reference. right: ASD of the Brownian/Suspension noise expressed in
geophone output voltage.
Svib (ω) = Vvib (jω)Vvib (−jω) [(m/s)2 /Hz] (3.8)
3.3.2 Sensor noise sources

Geophone coil thermal noise
Because the geophone coil has a certain resistance Rc [Ω], there will be thermal (also called
Johnson) noise present at the output terminals of the geophone. The thermal noise of a
conductor is described in Fish [7], and is the random voltage produced across the resistance
Rc caused by the thermal agitation of electrons. Thermal or Johnson noise is a white noise
source, having a constant PSD over frequency. The voltage PSD of thermal noise of the
geophone coil is given by
Scoil = 4kT Rc = 6.47 · 10−17 2

[Vrms /Hz] (3.9)
where k = 1.38 · 10−23 J/K is the Boltzmann constant, and T is the room temperature in
Kelvin, which is assumed to be 293 K.
Brownian noise
Because the proof mass is suspended in free air, air molecules constantly collide with it,
causing the proof mass to follow a certain random path, called a Brownian motion. The
Brownian motion of suspended objects is also referred to as suspension noise. The resulting
motion of the proof mass can be modelled as an equivalent input acceleration spectrum as
described in Usher [25, p505], Riedesel [17, p1728] and Rodgers [18, p1074]. The resulting
acceleration PSD is constant over frequency and can be written as:
πkT ζfo
Sacc = 16 = 1.45 · 10−17 [(m/s2 )2 /Hz] (3.10)
mg
where k is Boltzmann’s constant, T is the room temperature in Kelvin, ζ is the damping

ratio of the geophone system, mg = 0.0236 kg is the proof mass and fo = 4.8 Hz is the
geophone resonant frequency. In order to get a PSD in terms of geophone output voltage,
this model is multiplied with the squared magnitude of the geophone acceleration sensitivity
(3.4):
¯ ¯2
¯ Gg jω ¯
¯
Ssus (ω) = Sacc ¯ ¯ 2
[Vrms /Hz] (3.11)
(jω) + 2ζωo jω + ω ¯
2 2
o
√
The ASD [Vrms / Hz] of the suspension noise can be seen in Figure 3.10.
3.3.3 A/D and D/A electrical noise

Analog to Digital Converter electrical noise
When short circuiting the ADC input, the ideal output equals zero. In practice the output
is not zero at all as can be seen in Figure 3.11 on the left hand side, where a time trace of the
short circuited ADC is shown. The ADC of the dS1104 board uses a 16 bit converter, and
was running at a sampling rate of 20 kHz. The range of this converter is fixed to ±10 Vpeak ,
giving it a total range of 20 Vpp . The least significant bit (lsb) equals to 20/216 = 0.305 mV .
The output of the short circuited ADC standard deviation was 1.35 bits, or, expressed in
volts, σA/D elec = 0.412 mVrms . The PSD representing the A/D electrical noise is flat and
defined up to the Nyquist frequency Fn = 10 kHz:
2
σA/D
= 1.7 · 10−11 2
elec
Sade = [Vrms /Hz] (3.12)
Fn
Digital to analog converter electrical noise
A Digital to Analog Converter (DAC) with a zero input, ideally has a zero output. As is to
be expected, in practice the output is not zero, but is contaminated with noise. The DAC
output was measured an had a standard deviation of σD/A elec = 0.12 mVrms and had white
noise characteristics. The PSD representing the D/A electrical noise is flat and defined up
to the Nyquist frequency Fn :
2
σD/A
= 1.4 · 10−12 2
elec
Sdae = [Vrms /Hz] (3.13)
Fn
Time trace of short circuited A/D channel, σ = 1.35 counts

Sensor signal and quantized signal
8 2
6 1.8
1.6
4
A/D converter counts
1.4
Sensor signal [mV]

2
1.2
0 1
−2 0.8
0.6
−4
0.4
−6
0.2 Sensor signal
Quantized signal
−8 0
0 20 40 60 80 100 0.164 0.1645 0.165 0.1655 0.166 0.1665 0.167 0.1675 0.168
time [msec] Time [sec]
Figure 3.11 left: Time trace of the A/D electrical noise of a dS1104 controller board. The
signal value was multiplied with a factor 216 /20, making one unit in the graph correspond to
one count of the converter. right: Time trace of a typical sensor signal, with a sampled and
quantized version of the same signal, simulating an A/D conversion. At every sample instant
a quantization error is made. defined as the mismatch between the true signal value and the
quantized signal value. To quantize the signal value, the true value is rounded off towards toe
nearest integer multiple of the lsb value of 0.305 mV .
3.3.4 A/D and D/A quantization noise

A/D quantization noise
When a sensor signal enters an Analog to Digital Converter (ADC) a quantization error is
introduced. At every sample instant, every 0.05 msec, a sample is taken which is then to
be quantized by the ADC. Since only a limited set of bits are available, the sampled value
must be rounded off towards the nearest multiple of the least significant bit (lsb) value of
0.305 [mV], introducing a quantization error as is illustrated in Figure 3.11 on the right
hand side. Oppenheim [13] treats the quantization error as uniformly distributed zero mean
white noise. This simple model is accurate enough in this case study, since the quantization
error is small compared to the A/D electrical noise and they apply at the same location in
the loop. The model describes the theoretical variance of the error and can be written as:
2 lsb2 2
σA/D quant = [Vrms ] (3.14)
12
Using the lsb value of 0.305 mV and taking the Nyquist frequency into account, this leads
to the following model:
2
σA/D
= 7.8 · 10−13 2
quant
Sadq = [Vrms /Hz] (3.15)
Fn
D/A quantization noise

Every output signal send to the DAC, is quantized in the same way as the A/D quantization,
introducing a quantization error similar as with the ADC quantization. Since the DAC has
the same lsb value as the ADC, the same model as for the A/D quantization noise can be
used:
Sdaq = Sadq = 7.8 · 10−13 2

[Vrms /Hz] (3.16)
3.4 Performance analysis

In order to analyze the system performance, the PSD of the performance output Ẋl will
be simulated. Besides the PSD of the performance output, one can look at the PSD of
the control output u, which reveals the amount of control effort needed to achieve the
performance level.
Figure 3.12 Closed loop input-output configuration.
In Figure 3.12 the inputs w1 -w3 and outputs u and Ẋl can be seen. The closed loop transfer
function from these inputs towards the outputs is denoted with H(jω) and can be written
as:
 
· ¸ · ¸ w1
u CS CHg P S CHg S 
= w2 
Ẋl P CS PS S (3.17)
| {z } w3
H(jω)
where C is the controller, S the closed loop sensitivity function, P equals Pspeaker and Hg
the geophone velocity sensitivity Pgeo , see Figure 3.12. The PSDs at the ith disturbance
input, Swi (ω) simply equals the linear sum of the noise PSDs applying at that input, so:
Sw1 = Scoil + Ssus + Sade + Sadq

Sw2 = Sdae + Sdaq (3.18)
Sw3 = Svib (ω) |Psp vib (jω)|2
All the output PSDs and corresponding CPSs can now be calculated using Equation (2.15)
and their corresponding ASDs and CASs can be seen in Figure 3.13. Not only the total
output ASDs/CASs are shown, but also the contribution of every separate source by itself,
3.4. PERFORMANCE ANALYSIS 33
such that one can analyze how much power each source is contributing to the total power
over frequency.
−4 CAS towards performance output, σ = 86 µm/s
ASD towards performance output, σ = 86 µm/s x 10
1
ground
0.9 coil
−4 sus
10
adq
0.8 ade
dae
0.7 daq
−6
10 tot
0.6
[m/s/Hz1/2]
[m/s]
0.5
−8
10
0.4
ground
coil 0.3
sus
−10
10 adq 0.2
ade
dae
daq 0.1
−12
tot
10 −3 −2 −1 0 1 2 3
0 −4 −3 −2 −1 0 1 2 3
10 10 10 10 10 10 10 10 10 10 10 10 10 10 10
CAS towards controller output, σ = 0.9V

ASD towards controller output, σ = 0.9V 1
2
10 ground
0.9 coil
0
sus
10 adq
0.8 ade
dae
−2 0.7
10 daq
tot
0.6
−4
[V / Hz1/2]
10
[V]
0.5
−6
10 0.4
ground
−8 coil 0.3
10 sus
adq 0.2
−10 ade
10 dae
daq 0.1
−12
tot
10 −3 −2 −1 0 1 2 3
0 −4 −3 −2 −1 0 1 2 3
10 10 10 10 10 10 10 10 10 10 10 10 10 10 10
Figure 3.13 Simulation of performance output spectra (Ẋl ) and the controller output spectra
(u) using the 100 mHz controller (3.5). above: Simulated ASD ( left) and CAS ( right) of the
performance signal Ẋl . below: Simulated ASD ( left) and CAS ( right) of the control signal u.
In the simulated performance ASD, one can see that there is a large power contribution in
the lower frequency region, mainly due to the two A/D noise sources Sade and Sadq . The
simulated performance CAS confirms this, since it shows a steep rise in power between 0.01
Hz and 0.1 Hz and levels of at 86 µm/s, which is the 1-σ value of the closed loop load
velocity. In the time domain this can be interpreted as follows: the cone plate moves up and
down in a 10 sec - 100 sec period, causing the payload to drift.
Furthermore, one can analyze the vibration isolation performance. In the performance ASD
it can be seen that the ground vibrations are the dominant noise sources in a very small
frequency span from 5 Hz up to 200. The fact that other noise sources dominate the perfor-
mance CAS, tells us that the isolation system itself contributes more power to the perfor-
mance output at these frequencies, that the ground vibrations themselves. The open loop
performance value (1-σ) equals 98.8 µm/s, so closing the loop of the isolator only increases
the performance to 86 µm/s.
From the simulated ASD and CAS of the control signal, one can conclude almost all the
control effort is put in ’controlling’ the two dominant noise sources. In the ideal situation (no
other noise sources than the ground vibrations), all the control effort is due to the ground
vibrations. The standard deviation of the control signal equals 0.9 Vrms , and almost all its
power is located in the 0.001 Hz - 0.1 region, meaning that the control output is slowly
(about 100 sec) drifting up and down several volts. For comparison, the part of the control
signal that is used to isolate for ground vibrations has a standard deviation of 3 mVrms (too
small to be visible in this CAS).
The cause of the damage done by these two sources, is the low Signal to Noise Ratio (SNR)
at the location where these noise sources apply (The SNR is defined as 10·log(σs2 /σn2 ) [dB] or
equally 20·log(σs /σn ) [dB]). The noise sources are both located in between the geophone and
the A/D converter as can be seen in Figure 3.9. To improve the performance of the isolator,
the SNR at this location should be improved. This can be done by using a preamplifier, as
is discussed in the next section.
3.5 Using a preamplifier
3.5.1 Introduction preamplifier
From the performance analysis in the previous section it was concluded that the Signal to
Noise Ratio (SNR) of the signal entering the ADC is far too low. There are two ways to
overcome this problem.
First, one can try to decrease the magnitude of the noise sources. There is a possibility to
decrease the A/D quantization noise by adjusting the range of the converter, such that it
better matches the magnitude of the closed loop sensor signal. At the start-up of the system
though, the geophone output is in the order of a few millivolts, whereas in the closed loop
case it is in the order of a few tenths of millivolts. A dual range ADC would overcome this
problem, but unfortunately, the magnitude of the A/D electrical noise does not decrease
when the range of the converter is decreased, since it originates in the electronics within the
converter itself.
A second option to solve this problem is to pre-amplify the sensor signal before it enters
the ADC, increasing the signal level. Although a preamplifier can be used to improve the
SNR, the preamplifier itself introduces noise sources as well! One has to find out if the
pre-amplification of the sensor signal indeed counter balances the newly introduced noise
sources and thus improves the system performance. This will be the topic of this section.
The system configuration with a preamplifier included can be seen in Figure 3.14.
3.5. USING A PREAMPLIFIER 35
!!"#
Figure 3.14 System configuration with a preamplifier included and noise sources.
3.5.2 Preamplifier transfer function
As can be seen in Figure 3.15 on the left hand side, the amplitude of the geophone velocity
sensitivity drops off towards the lower frequencies with a +2 rate. As a consequence, input
signals in this low frequency region will generate a very small output voltage and will there-
fore be more sensitive to noise sources. In order to improve the SNR in that frequency region
it seems to make sense to amplify this low frequency region more than the high frequency
part. Therefore, the preamplifier transfer function is chosen to be a second order lag filter,
as can be seen in Figure 3.19 on the right hand side. The improved ’velocity sensitivity’ of
the geophone in series with the preamplifier can be seen in Figure 3.15 on the left hand side.
Velocity sensitivities Preamplifier Transfer Function

4
10
3
10
3
10
Gain
2
10
2
10
1
10
[V/(m/s)]
−1 0 1 2
1 10 10 10 10
10
0
0
10
Phase [deg]
−30
−1 −60
10
Geophone with preamp −90
−2
Geophone
10 −1 0 1 2
−120 −1 0 1 2
10 10 10 10 10 10 10 10
Figure 3.15 left: The velocity sensitivities of the geophone and of the geophone in series with
the preamplifier. right: Transfer function of the preamplifier.
The output voltage signal of the preamplifier is denoted by Eout . Its input signal is the
geophone output Eg . The transfer function of the used preamplifier can be written as a
series connection of two first order lag filters:
µ ¶2
Eout τp s + 1
= Gpa (3.19)
Eg αp τp s + 1
with Gpa = 1000, τp = 1/ω1 = 1/(2π 0.45) and αp = 10, which puts the corner frequencies at
0.45 hz and 4.5 Hz. The preamp now has a low frequency gain of 1000 and a high frequency
gain of 10 and will be referred to as the ’1000-10 preamp’.
This preamplifier was physically build using an operational amplifiers in both preamplifier
stages. The physical configuration of the circuit and the derivation of the transfer function
in terms of the physical components can be found in Appendix B.3.
3.5.3 Preamplifier noise model

The noise model of the preamplifier developed here, is based on the models developed in
Rodgers [18], [19], Riedesel (et al.) [17] and Horowitz and Hill [10]. The physical preamplifier
is build as series connection of two separate, first order stages. The output of the first and
second stage are denoted with Eout,1 and Eout respectively. The two stages have transfer
functions:
µ ¶ µ ¶
Eout,1 τp s + 1 Eout τp s + 1
(s) = 100 , (s) = 10 (3.20)
Eg αp τp s + 1 Eout,1 α p τp s + 1
In Figure 3.16 a schematic of the first stage can be seen with all its noise sources included.
(The geophone-coil noise has been left out, since it is already accounted for in Scoil ).
Figure 3.16 First stage of the preamplifier and all its noise sources
The current noise sources, in− and in+ of the operational amplifier, are shown as Norton
generators from the inverting and non inverting inputs to ground. The voltage noise source
of the operational amplifier en is shown as a Thevenin generator in series with the summing
junction at the intersection of R11 and Zf 1 . The preamplifier is build using the op07 low-
noise operational amplifier and its noise characteristics can be found in the manufacturer
specifications.
The noise sources eR11 , eR21 , and eR31 are due to thermal noise of the resistors and are
included in series with the resistors. The total noise model at the output of the first stage
will be denoted by Spa,1 (ω) and will not be written out here, and its derivation can be found
in Appendix B.4. For the second stage a similar output noise model can be developed, and
will be denoted by Spa,2 (ω). The total noise model at the output of the preamplifier Spa (ω)
can now be written as:
¯ ¯2
¯ Eout,1 ¯
¯
Spa (ω) = Spa,1 (ω) ¯ (jω)¯¯ + Spa,2 (ω) 2
[Vrms /Hz] (3.21)
Eg
The preamplifier noise model Spa (ω) and the separate terms on the right hand side of (3.21)
can be seen in Figure 3.17. To validate this noise model, the output of the preamplifier with
short circuited input was measured using the HP35665 analyzer. The result can also be seen
in Figure 3.17. The measurements match the theory very well, except in the low frequency
region (<1 Hz) and at 60 Hz, where a clear ’60 Hz pick-up’ noise peak can be seen. The
overestimation of the low frequency noise and the un-modelled 60 Hz pick up peak will be
subject of discussion later in this chapter.
Preamplifier output noise and measured noise

−6
10
total: both stages
due stage 1
due stage 2
−8 measurement
10
V2rms / Hz
−10
10
−12
10
−14
10
−2 0 2
10 10 10
Frequency [Hz]
Figure 3.17 Total preamplifier noise model PSD Spa (ω) and the two separate contributions of
the preamp stages at the preamplifier output, together with the measured noise. The noise model
matches the measured noise very well in the 1 Hz - 100 Hz region except for the pick up peak
at 60 Hz. The noise model overestimates the noise model in the low frequency region (<1 Hz).
3.5.4 Performance analysis using a preamplifier

To make the performance comparable with the previous non-preamp case, the controller is
chosen such that the loop gain transfer function L(s) remains the same. This was done
by multiplying the controller (3.5) used in the non-preamp case (see Section 3.4) with the
inverse of the preamplifier transfer function (3.19). The output ASD’s and CAS’s are calcu-
lated as was done in the non-preamp case, and can be seen in Figure 3.18.
−6 CAS towards performance output, σ = 2.9 µm/s

ASD towards performance output, σ = 2.9 µm/s x 10
−5
10 ground
3 coil
−6
sus
10 pa
2.5 adq
ade
−7
10 dae
2 daq
tot
[m/s/Hz1/2]
−8
10
[m/s]
1.5
−9
10 ground
coil
−10 sus 1
10 pa
adq
−11 ade 0.5
10 dae
daq
−12
tot
10 −4 −3 −2 −1 0 1 2 3 4
0 −4 −3 −2 −1 0 1 2 3
10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10
CAS towards controller output, σ = 0.031V

ASD towards controller output, σ = 0.031V
0 0.035
10 ground
coil
0.03 sus
−2 pa
10
adq
0.025 ade
−4 dae
10 daq
tot
0.02
[V / Hz1/2]
[V]
−6
10
0.015
ground
−8 coil
10 sus
0.01
pa
adq
−10 ade
10 0.005
dae
daq
−12
tot
10 −4 −3 −2 −1 0 1 2 3 4
0 −4 −3 −2 −1 0 1 2 3
10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10
Figure 3.18 Simulation of performance output spectra (Ẋl ) and the controller output spectra
(u) using the 1000-10 preamplifier and a controller such that the loop transfer function L(s)
is the same as the in the non-preamp case (see Figure 3.8).above: Simulated ASD ( left) and
CAS ( right) of the performance signal Ẋl . below: Simulated ASD ( left) and CAS ( right) of the
control signal u.
The performance ASD in Figure 3.18, shows that the magnitude of the contribution of the
ADC noise sources is now much smaller compared to the non-preamp case (see Figure 3.13),
due to the use of the preamplifier. As a price though, preamplifier noise is introduced. The
preamp noise dominates the ASDs in the low frequency region, but at a much lower level
than was the case with the ADC noise sources in the non-preamp case. Apparently, the
introduction of the preamplifier indeed counter balances the introduction of its own noise.
The standard deviation of the performance output dropped from 86 µm/s to 2.9 µm/s, a
performance increase with a factor 30!
The control effort dropped from 0.89 Vrms in the non-preamp case to 31 mVrms in this case.
The preamplifier noise is now the major contributor to the control effort, accounting for
90 % of the control effort, meaning that the system is still mainly compensating for its own
noise, instead of for ground vibrations. Because the power of the controller output is mainly
located in the 0.001 Hz - 0.1 region, the controller output will slowly drift up and down circa
50 mV causing the platform to make a slow drifting motion. This causes undesired system
behavior, since there is the risk that the platform will reach its physical travel limits.
Using the analysis tools available, one can try to find possible improvements that could be
made to avoid the controller output to have its power located mainly in the low frequency
region. The most straight forward approach is to decrease the amount of low frequency noise
generated by the preamplifier. Since the low frequency noise is caused by the inherent 1/f
noise of the opamps, this is hard to achieve with the current circuit configuration. On the
other had, one can look at the closed loop transfer function from preamp input to controller
output; C · S (not shown here). One can see that this transfer has a huge gain in the 0.001
Hz - 0.1 region, due to the controller used. This huge controller gain is needed to achieve
the 0.1 Hz bandwidth. This analysis raises the question if the desire to have this 0.1 Hz
bandwidth is a good idea concerning the effect is has on the propagation of the low frequency
preamp noise.
This points out a degree of freedom in the DEB design approach: the control strategy used.
The controller defines the closed loop transfer function H(jω) and thus the propagation of
the disturbances towards the outputs. The designer can use this degree of freedom, in order
to optimize for performance. This topic is discussed in Chapter 4.
As mentioned earlier in § 3.5.3, the preamplifier noise model (see Figure 3.17) overestimates
the measured noise in the low frequency region. Since it is this part of the preamp noise
spectrum that causes the huge power contribution to the controller output, one can say that
the results of the performance simulation are therefore not reliable in the low frequency
region. In order to get a more reliable picture of the effect of the low frequency preamplifier
noise, the propagation should be done with the actually measured noise. This will be done
in the next section.
ASD of simulated controller input, σ = 2.7 mV ASD of measured controller input, σ = 1.2 mV
0 −2
10 10
−3
10
−5
10
−4
10
[V/Hz1/2]
[V/Hz1/2]
−10
10
−5
ground 10
coil
sus
−15 pa
10
adq −6
10
ade
dae
daq
−20
tot −7
10 −4 −2 0 2 4
10 −1 0 1 2 3
10 10 10 10 10 10 10 10 10 10
−3 CAS of simulated controller input, σ = 2.7 mV −3 CAS of measured controller input, σ = 1.2 mV
x 10 x 10
3 1.4
ground
coil
sus 1.2
2.5 pa
adq
ade 1
dae
2
daq
tot
0.8
[V]
[V]
1.5
0.6
1
0.4
0.5
0.2
0 −4 −2 0 2 4
0 −1 0 1 2 3
10 10 10 10 10 10 10 10 10 10
Figure 3.19 Validation of the controller input y spectra using the 1000-10 preamplifier. above:
Simulated ASD ( left) and measured ASD ( right) of the controller input signal. below: Simulated
CAS ( left) and measured CAS ( right) of the controller input signal.
3.6 Experiments and comparison with theory
In this section it is tried to close the gap between the theoretical and experimental results.
It is chosen to validate the noise levels of the closed setup at the controller input, right
after the ADC. This signal can be comfortably ’measured’ inside the computer, avoiding the
introduction of new measurement noise sources. The simulated and measured results can be
seen in Figure 3.19. From the simulated and measured CAS in this figure one can see that
the simulated controller input has a standard deviation σ = 2.7 mVrms and the measured
input has a deviation of σ = 1.2 mVrms .
In the simulated CAS of the controller input (Figure 3.19 below left) one can see that the
ground vibrations account for the majority of the controller input power. Because the ground
vibration model Vvib (ω) (see Figure 3.10 on page 29) overestimates the measured vibration
3.6. EXPERIMENTS AND COMPARISON WITH THEORY 41
−3 CAS towards controller output, σ = 0.0032V CAS of simulated controller input, σ = 2.04 mV
x 10 −3
x 10
3.5 2.5
ground ground
coil coil
3 sus sus
pa pa
adq 2 adq
2.5 ade ade
dae dae
daq daq
tot 1.5 tot
2
[V]
[V]
1.5
1
0.5
0.5
0 −4 −3 −2 −1 0 1 2 3
0 −4 −2 0 2 4
10 10 10 10 10 10 10 10 10 10 10 10 10
Figure 3.20 Simulation results using the actually measured ground vibration spectrum and pream-
plifier noise spectrum. left: Simulated CAS of the controller output (compare with Figure 3.18,
below right). right: Simulated CAS of the controller input (compare with Figure 3.19, below
left).
level at certain frequency regions, the simulated result is likely to overestimate the power in
the controller input signal.
In the measured CAS (Figure 3.19 below right) electrical pick up noise can be identified as
stepwise increases in power at 60 Hz, 120 Hz, 180 Hz and 240 Hz. This pick up can be the
result of electromagnetic/static noise sources which are not accounted for in this analysis.
From the measurements, shown in Figure 3.17 on page 37, one can see that the preamplifier
noise spectrum shows a clear 60 Hz pick up peak. One might expect that part of the pick up
noise is due to the preamplifier. Looking at the simulated ASD in Figure 3.19 though, it can
seen the preamp noise lies far below the DAC and ADC sources. This means that the DAC
and ADC are the most sensitive to pick up noise (concerning the effect it can have on the
controller input), and that it is less likely that the pick up noise measured at the controller
input, including the 120, 180 and 240 Hz pick up peaks, can be attributed to the preamplifier
electrical circuitry. The magnitude of the pick up noise can be reduced by applying electri-
cal/magnetic shielding to the electrical circuitry and careful cabling. If the pick up remains
to have a significant impact, one should account for them in order to increase the predic-
tion accuracy. Discarding the power due to electrical pick up, the measured level would be
around 0.8 mVrms . This value will be used to compare with the theoretically obtained results.
In order to increase the prediction accuracy of the simulation, the simulation is done again
with the actually measured preamplifier noise and measured ground vibration spectra. Not
all results are shown, only the CAS of the controller output and the the CAS of the controller
input, which can be seen in Figure 3.20.
The simulated CAS of the controller output using the measured preamplifier noise and
ground vibration level shows a drastic decrease in predicted controller output power when
compared with the simulation results using the actual noise models (see Figure 3.18 below
right); instead of 31 mVrms the predicted controller output decreased to 3.2 mVrms , indi-
cating that the prediction done in the previous section was indeed not reliable. This result
shows the importance of accurate disturbance modelling, especially when the disturbance is
the dominant source at the output.
The simulated CAS of the controller input in Figure 3.20 shows that ground vibration still
contribute the majority of the power. In this figure one can see that more than half of the
power (about 1.3 mVrms ) is now added at 120 Hz and 240 Hz. Looking at the measured
ground vibration spectrum in Figure 3.10 on page 29 one can see that it indeed contains
two power peaks at these frequencies. However, these rather large increases in power are
not visible in the measured CAS of the controller input! A reason for this discrepancy can
be the following: the peaks in the measured ground vibration level are due pick up noise
in the measurement circuit and do not represent ground vibrations. When leaving out the
contribution due to this measurement pick up noise, the simulated controller input level
would be around 0.7 mVrms which matches up well with the measured controller input level
of 0.8 mVrms where the pick up noise also has been left out.
Summarizing, the theoretical results match the measured results if the simulation is done
with the actual measured disturbance spectra and the power contribution thought to be due
to electrical pick up noise are left out of the comparison.
The predicted performance in terms of Ẋl drops from 2.9 µm/s when using the disturbance
models (see Figure 3.18) to about 0.8 µm/s when using the measured spectra and the power
due to the measurement pick up noise in the ground vibration spectrum is discarded.
3.7 Conclusions and recommendations

3.7.1 Conclusions
The improved vibration isolation setup is successfully build and made operational. The
system was modelled using theoretical modelling and standard identification techniques.
By means of experiments, the existence of magnetic coupling between the voice coil of the
speaker and the coil of the geophone is shown. With the use of a magnetically shielded
geophone the coupling could be eliminated.
All the (stochastic) disturbance thought to act on the setup are modelled by either measure-
ment of their PSD or by theoretical modelling. Using these models in the DEB simulation
tool, the error (or performance) of the open and closed loop system is predicted, resulting
in σ = 98.8 µm/s resp. σ = 86 µm/s (1-σ). By analyzing the DEB simulation results
the performance limiting factors could be identified; the disturbances introduced by the AD
converter turned out to be the absolute dominant contributors to the error. The system
could be improved by introducing a preamplifier before the AD converter in the loop. The
DEB simulations show a huge performance improvement; a closed loop performance level of
(theoretically) 2.9 µm/sec! These results show that, using the DEB approach, the designer
is indeed able to predict the performance of a system concept and to gain insight in the
3.7. CONCLUSIONS AND RECOMMENDATIONS 43
performance limiting factors of the system.
To validate the prediction accuracy of the DEB approach, simulated results are compared
with experimental results; the closed loop controller input level was simulated and measured.
The simulations showed that the almost all the power of the controller input signal was due
to ground vibrations. The measured results showed a significant amount of electrical pick
up noise which was not accounted for in the DEB analysis. The simulated results matched
the experimentally measured level when the simulations were done with the actual measured
ground vibration spectra (the model representing the measured ground vibrations was too
inaccurate) and the power due to electrical pick up noise in the measured controller input
(which accounted for 30 % of the measured power) were discarded. This result shows that
the predicted results by the DEB approach can be accurate if realistic disturbance models
are used and electrical pick up noise is accounted for.
3.7.2 Recommendations
To increase the predicting accuracy of the DEB analysis the dominant disturbance sources
should be measured very accurately (if a prototype is available) and it is then strongly
recommended to use the actual measured spectra in the simulations. For this case study,
better measurements of the ground vibrations should be made e.g., by averaging several
measurements over longer periods of time and taking special care not to contaminate the
measurement with electrical pick up noise.
To further increase the prediction accuracy, effort should be put in to identify and model the
electrical pick up noise sources that act on the closed loop system and account for them to
the DEB analysis. If possible, electrical pick up noise should be avoided by careful electrical
circuit design, grounding and cabling.
If there remains to be a discrepancy between the theoretical and measured output PSD, a
way to deal with it is to accredit the difference to an additional (artificial) disturbance, such
that the difference in output power is contributed by this input disturbance. In this way the
designer has a possibility to account for this difference in power in the simulations.
Chapter 4
H2 control in active vibration

isolation
4.1 Introduction
When a system designer has designed different system concepts, he needs to find out which
concept is the most promising. In order to compare the performance of these different
concepts, the system designer needs an objective measure of performance. In this thesis
an objective measure is interpreted as follows: the same amount of control effort should be
used for each different system concept and the used controller should push the system to its
maximum performance level given the system and disturbance models.
As is explained in Section 2.5, the H2 control strategy (theoretically) offers the system
designer the opportunity to put a constraint on the amount of control effort used, while
maximizing the performance of the system. So, H2 control offers the ability to objectively
compare different system concepts and can be seen as an extension of the DEB design ap-
proach. The goal of this chapter is to find out if the use of H2 control indeed adds value to
the DEB approach in practice, by applying it to the case study of Chapter 3.
In this chapter an H2 controller will be synthesized for the system concept using the pream-
plifier developed in the previous chapter. The disturbance models developed in Chapter 3 are
used to obtain the weighting filters needed for H2 controller synthesis in § 4.2.1. In § 4.2.2
it is taken care of that the weighting filters meet the standard assumptions as discussed
in Appendix A.2. Then the H2 controller is synthesized and compared with old controller
(3.5) and the Pareto curve is calculated. In Section 4.3 the performance is simulated and
compared with experimental results. Next, in Section 4.4 it is tried to find out what the
best preamplifier design is by comparing the Pareto curves using different preamplifiers.
45
46 CHAPTER 4. H2 CONTROL IN ACTIVE VIBRATION ISOLATION
4.2 Design of an H2 controller

4.2.1 Weighting filters for the 10-1000 preamp case
When designing an H2 controller for the 10-1000 preamp configuration all the noise sources
and disturbances models obtained in the previous chapter can be used to obtain the weighting
filters. The chosen configuration of the weighting filters can be seen in Figure 4.1.
Figure 4.1 Configuration of the weighting filters Vw1 , Vw2 , and Vw3 for H2 controller design.
Two outputs are defined, a control effort output u with scaling filter Wz1 and a performance
output (Ẋl ) with scaling filter Wz2 .
The weighting filter Vw1 accounts for five noise sources; the two sensor noise sources Scoil and
Ssus , the preamp noise source Spa and the ADC noise sources Sade and Sadq . The two sensor
noise sources must be propagated to the location where the disturbance signal w1 applies,
by multiplying them with the squared magnitude of the preamplifier transfer function. The
PSD of the fictional disturbance signal w1 can then be written as:
¯ ¯2
¯ Eout ¯
¯
Sw1 (ω) = (Scoil + Ssus ) ¯ (jω)¯¯ + Sade + Sadq + Spa (4.1)
Eg
and the weighting filter Vw1 (jω) equals the spectral factorization of Sw1 (ω). A way to avoid
the problem of obtaining a spectral factorization is to make a least square fit on the square
root of a frequency response data of Sw1 (ω) (so directly on ASD data) on a certain grid.
This approach is followed using the MATLAB° r
command fitmag and the resulting 6th order
weighting filter Vw1 (s) can be seen in Figure 4.2.
(s + 0.57)(s + 1.43 · 10−3 )

Vw1 (s) = 4.22 · 10−6 ·
(s + 0.187)(s + 4.56 · 10−2 )
(s2 + 9.95 · 10−2 s + 2.93 · 10−3 )(s2 + 8.63s + 31.0)

[Vrms ] (4.2)
(s + 1.75 · 10−2 )(s + 1.45 · 10−3 )(s2 + 3.95s + 3.98)
4.2. DESIGN OF AN H2 CONTROLLER 47
Weighting filter Vw1

−3
10
−4
10
−5
10
−6
10
Vrms
−7
10 filter Vw1
adq
−8 ade
10 pa
gcp
−9
gsp
10 −2 0 2
10 10 10
Frequency [Hz]
Figure 4.2 The weighting filter Vw1 (s). The filter is a least square fit to square rooted values of
Sw1 (ω) on a certain frequency grid.
To obtain Vw2 (s) a similar approach can be followed. Vw2 (s) accounts for the two DAC noise
sources Sdae and Sdae , which are both constant over frequency (see Section 3.3), and thus
Vw2 is easily calculated by hand as:
Vw2 = 1.48 · 10−6 [Vrms ] (4.3)
The third filter accounts only for the ground vibrations and is given by the floor noise model
Vvib (jω) (3.7) multiplied with the open loop speaker model Psp vib (s) (3.2):
Vw3 (s) = Vvib (s) Psp vib (s) [Vrms ] (4.4)
The two output scaling filter Wz1 scales the controller output u and Wz2 scales the perfor-
mance output Ẋl . Actually, only their ratio is of importance. They are given the values:
Wz1 = 1, Wz2 = 1 · 105 (4.5)
In § 4.2.4 it is explained why these particular values are chosen.
4.2.2 Meeting the standard assumptions
As discussed in Appendix A.2, the open loop system, including the weighting filters, has to
meet the standard assumptions. From Figure 4.1 one can see that the open loop system can
be written as:
  
  .. w1
z1  0 0 0 . Wz1  
 z2   ..  w2 
  = 
 0 Wz2 P Vw2 Wz2 Vw3 .Wz2 P  w3 
 ···   ···  
 ··· ··· ···   ··· 
y ..
Vw1 Hgpa P Vw2 Hgpa Vw3 . Hgpa P u
  (4.6)
w1
· ¸  w2 
G11 G12   w3


=
G21 G22   ···


u
corresponding to Figure 2.4 on page 16 (repeated in Appendix A.2 as Figure A.1). For the
open loop plant to meet the standard assumptions, implies the following:
Assumption 1: D11 must be 0

The direct feedthrough from w to z, D11 must be zero. Since four of the six are already zero,
we only have to pay attention to the remaining two. The two transfer functions Wz2 P Vw2
and Wz2 Vw3 must have high frequency roll-off such that D11 equals zero. To meet this
assumption, the filter Vw2 and Vw3 are adjusted to have high frequency roll-off. When the
roll-off is given well beyond the bandwidth (which is ca. 200 Hz) of the system, it will not
influence the performance analysis. Both the filters are multiplied with a first order transfer
function which has its cut-off frequency at 1 kHz:
Ã !
−6 1
Vw2 (s) = 1.48 · 10 1 [Vrms ] (4.7)
2π1000 s + 1
Ã !
1
Vw3 (s) = Vvib (s) Psp vib (s) 1 [Vrms ] (4.8)
2π1000 s + 1
Assumption 2 and 3: (A, B2 ) is stabilizable and (A, C2 ) is detectable

In a well defined plant (stabilizable and detectable), the only uncontrollable or unobservable
unstable modes can come from the weighting filters. Since none of the weighting filters have
unstable poles, this assumption is met.
Assumption 4: D12 has full column rank

The matrix D12 denotes the direct feedthrough matrix of G12 , the transfer function from
the control signal u to the controlled output z:
· ¸
Wz1
G12 = (4.9)
Wz2 P
4.2. DESIGN OF AN H2 CONTROLLER 49
When one of the matrices has full column rank the assumption is met. Since Wz1 is a nonzero
scalar, this assumption is met.
Assumption 5: D21 has full row rank
The matrix D21 denotes the direct feedthrough matrix of G21 , the transfer function from
noise and disturbance signals w to controller input y:
£ ¤
G21 = Vw1 Hgpa P Vw2 Hgpa Vw3 (4.10)
When one of the matrices has full row rank the assumption is met. Since Vw1 (jω) is nonzero
for ω ∈ R ∪ ∞ this assumption is met.
Assumption 6 and 7
When these assumptions are not met the plant has poles on the imaginary axis or one of the
transfer functions G12 or G21 has transmission zeros. This is not the case so the assumption
is met.
4.2.3 Closed transfer functions using an H2 controller

The weighting filters are build and all the assumptions are met. An H2 controller can now
be synthesized. The resulting transfer functions can be seen in Figures 4.3 and 4.4. The
transfer functions when using the 100 mHz controller (3.5) developed in § 3.2.5 are also
shown for comparison.
100 mHz and H2 Transmissibility 100 mHz and H2 Sensitivity
20 20
0 0
Magnitude [dB]
Magnitude [dB]
−20 −20
−40 −40
−60 −60
−80 100 mHz transmissibility −80 100 mHz sensitivity

H2 transmissibility H2 sensitivity
−2 0 2 4 −2 0 2 4
10 10 10 10 10 10 10 10
Figure 4.3 left: Closed loop transmissibilities for the 100 mHz and the H2 controller case. The
H2 controller tends to isolate the system more at the higher frequencies. right: Loop sensitivity
functions S(s) and SH2 (s) when using the 100mHz and the H2 controller respectively. The loop
sensitivity defines the amount of isolation. It is clearly visible in this graph, that the H2 controller
tends to isolate more in the high frequency region.
In Figure 4.3 the closed loop transmissibilities (see § 3.2.5) when using an H2 controller
and a 100 mHz controller can be seen. Analyzing the transmissibilities, a clear difference in
control strategy can be seen. The 100 mHz controller puts the first 0 dB cross-over at 0.1 Hz,
as it was designed for. The H2 controller puts its first cross-over at about 0.3 Hz, decreasing
the vibration isolation performance in this region, but apparently, it is better when taking
the disturbances in account. This can also be concluded from the sensitivity functions in
Figure 4.3; the H2 control strategy tends to shift the whole sensitivity function SH2 (s)
towards a higher frequency region. Since the sensitivity defines the increase in vibration
isolation performance (see § 3.2.5), the isolation performance of the system is shifted to a
higher frequency region.
100 mHz and H2 Loop−gain 100 mHz controller and H2 controller

100 80
50 60
Magnitude [dB]
Magnitude [dB]
0 40
−50 20
−100 0
100 mHz loop gain 100 mHz controller
−150 −20
H2 loop gain H2 controller
−200 −2 0 2 4
−40 −4 −2 0 2 4
10 10 10 10 10 10 10 10 10
−90 135
−180 90
Phase [deg]
Phase [deg]
−270 45
−360 0
−450 −45
−540 −90
−630 −2 0 2 4
−135 −4 −2 0 2 4
10 10 10 10 10 10 10 10 10
Figure 4.4 left: Loop transfer functions L(s) for the 100 mHz and H2 controller case. In the
H2 case the 0 dB cross-overs lie at higher frequencies, suggesting a different control strategy.
right: The 100 mHz controller together with the H2 controller. The H2 controller uses less gain
in the low frequency region, and more gain in the high frequency region.
From the loop gains in Figure 4.4 one can see that the H2 controller gives good phase
and gain margins at both the cross-over frequencies, making it a robust and implementable
controller. The H2 controller puts both the loop transfer function cross-over frequencies at
a higher frequency than the 100 mHz controller, suggesting a different control strategy when
one wants to optimize performance. Looking at the controllers in Figure 4.4, it can be seen
that this difference in closed loop behavior is due to less gain at the low frequency region.
At the higher frequencies, the H2 controller has more gain, achieving more isolation there.
4.2.4 The Pareto curve

As was discussed in the previous section, every ratio between the values of the scaling filters
returns a different controller. Each controller uses a certain control effort level to achieve a
certain performance level (given the disturbances filters). These levels can be plotted against
each other, defining a Pareto curve. In Table 4.1 the different ratio’s and obtained levels can
be seen, which are plotted in Figure 4.5. One can see that there is a part on the curve where
4.3. SYSTEM PERFORMANCE USING THE H2 CONTROLLER 51
an increase of control effort stops causing a significant increase in performance (a decrease in

the actual value). The sharp bend in the curve corresponds to an output filter ratio of 104 .
The H2 controller above was synthesized with ratio of 105 , just to be sure the maximum
achievable performance is squeezed out of the system.
Ratio W2 /W1 effort u [mVrms ] Performance Ẋl [µm/sec]

100 : 0.00469 - 98.6
101 : 0.349 - 83.2
5·101 : 1.97 - 29
102 : 2.66 - 11.6
5·102 : 3.18 - 1.25
103 : 3.24 - 0.91
104 : 3.65 - 0.696
105 : 5.73 - 0.667
106 : 12 - 0.666
107 : 16.7 - 0.666
Table 4.1 Ratios and obtained levels defining the Pareto curve
4.3 System performance using the H2 controller

4.3.1 Performance analysis using an H2 controller and a preamplifier
Now an H2 controller is synthesized with a ratio of 105 , its performance can be analyzed
as in the previous cases. The simulation results can be seen in Figure 4.6. Comparing
these results with the fixed loop-gain case using a preamp (see Figure 3.18 on page 38), two
things must be said. To objectively compare the two performance levels, the effort to achieve
this should be the same. The control effort level in the that case was 31 mVrms achieving
a performance level of 2.9 µm/s. Comparing this control effort level with Table 4.1, the
performance achieved using an H2 controller at that effort level would be near 0.666µm/sec.
So, with the use of an H2 controller the performance value decreases with a factor of 4.35!
From the Pareto curve it can be observed that this effort level is a little bit overdone, and a
lower level was chosen: 5.73 mVrms resulting in a performance level of 0.667 µm/sec. The
use of an H2 controller thus gives, even on this lower effort level, a serious performance
improvement compared to the 100 mHz controller.
4.3.2 Experimental results using an H2 controller

An interesting question is if this controller strategy also works in practice. Is the perfor-
mance improvement a pure theoretical result or does it also work when an H2 controller is
really implemented? To validate the theoretical result an H2 controller was implemented
on the dSpace 1104 board and the results were measured. Before the H2 controller could
be implemented, the order was reduced from 25 to 15 and it was put in a discrete state
Pareto Curve for the 1000−10 preamp cases
4.5
3.5
Performance [µ m/sec]
2.5
1.5
0.5
0
0 2 4 6 8 10 12 14 16 18
Control effort [mV]
Figure 4.5 Pareto curve of the 10-1000 preamp case. The curve shows how much effort is
needed to achieve a certain performance level (More performance implies a lower performance
value). The dots represent calculated points. When travelling the curve from left to right, the
performance increases rapidly (value decreases) when the control effort increases. Somewhere
near a performance value of 0.66 µm/sec the curve levels off, implying that a large increase in
control effort would only gives a small increase in performance. The curve bends at a ratio of
approximately 104 , as can be seen in the table. The H2 controller used above was computed
with the ratio 105 , which position on the curve is marked with an asterisk.
space representation using MATLAB command d2c. The order reduction was done with the
MATLAB command modred, deleting 10 weakly coupled states. After the model reduction,
the gain and phase margins of the loop transfer function L(s) with the reduced controller
were carefully checked to make sure closed loop stability was not lost by the reduction step.
To validate the performance of the H2 controller the controller input level (as in Section 3.6,
the signal right after the ADC) is simulated and compared with the theoretical results. Next
to the controller input level, the controller output level is also simulated and compared with
the theoretical results. The controller output signal is defined as the signal value just before
entering the DAC. These signals are conveniently available inside the computer, avoiding
the introduction of new noise sources introduced by new measurements. The theoretical
and experimental results can be seen in figures 4.7 and 4.8.
Analyzing the controller input results in Figure 4.7, one can see that the ground vibrations
account for the majority of the (theoretical) control input power. The simulated level of 0.66
mVrms matches almost perfectly with the measured level of 0.69 mVrms . Since the ground
vibration model overestimates the true ground vibration level, the simulated level is likely to
−7 CPS towards performance output, σ = 0.67 µm/s

x 10
7
ASD towards performance output, σ = 0.67 µm/s ground
−5
10 coil
6 sus
pa
−6
10 adq
5 ade
−7
dae
10 daq
tot
4
[m/s]
[m/s/Hz1/2]
−8
10
−9
3
10 ground
coil
−10 sus 2
10 pa
adq
−11 ade 1
10 dae
daq
−12
tot
10 0 −4 −3 −2 −1 0 1 2 3
−4 −3 −2 −1 0 1 2 3 4
10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10
−3 CAS towards controller output, σ = 0.0057V

x 10
6
ASD towards controller output, σ = 0.0057V ground
0
10 coil
sus
5 pa
−2 adq
10
ade
dae
4
−4 daq
10 tot
[V / Hz1/2]
[V]
3
−6
10
ground
−8 coil 2
10 sus
pa
adq
−10
10 ade 1
dae
daq
−12
tot
10 0 −4 −3 −2 −1 0 1 2 3
−4 −3 −2 −1 0 1 2 3 4
10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10
Figure 4.6 Simulation of performance output spectra (Ẋl ) and the controller output spectra (u)
using the 1000-10 preamplifier and an H2 controller. above: Simulated ASD ( left) and CAS
( right) of the performance signal Ẋl using an H2 controller and a preamplifier. below: Simulated
ASD ( left) and CAS ( right) of the control signal u using an H2 controller and a preamplifier.
ASD of simulated controller input, σ = 0.662 mV ASD of measured controller input, σ = 0.69 mV
0 −2
10 10
−3
10
−5
10
−4
10
[V/Hz1/2]
[V/Hz1/2]
−10
10
−5
ground 10
coil
sus
−15 pa
10
adq −6
10
ade
dae
daq
−20
tot −7
10 −4 −2 0 2 4
10 −1 0 1 2 3
10 10 10 10 10 10 10 10 10 10
−4 CAS of simulated controller input, σ = 0.662 mV −4 CAS of measured controller input, σ = 0.69 mV
x 10 x 10
8 8
ground
coil
sus
pa
adq
6 ade 6
dae
daq
tot
[V]
[V]
4 4
2 2
0 −4 −2 0 2 4
0 −1 0 1 2 3
10 10 10 10 10 10 10 10 10 10
Figure 4.7 Validation of the controller input y spectra using the 1000-10 preamplifier and an
H2 controller. above: Simulated ASD ( left) and measured ASD ( right) of the controller input
signal using an H2 controller and a preamplifier.below: Simulated CAS ( left) and measured CAS
( right) of the controller input signal using an H2 controller and a preamplifier.
ASD towards controller output, σ = 0.0057V ASD of measured controller output, σ = 3.99 mV
0 −2
10 10
−2
10 −3
10
−4
10
−4
10
[V / Hz1/2]
[V/Hz1/2]
−6
10
−5
ground 10
−8 coil
10 sus
pa
adq −6
−10 10
10 ade
dae
daq
−12
tot −7
10 −4 −3 −2 −1 0 1 2 3 4
10 −1 0 1 2 3
10 10 10 10 10 10 10 10 10 10 10 10 10 10
−3 CAS towards controller output, σ = 0.0057V

x 10
6 −3 CAS of measured controller output, σ = 3.99 mV
ground x 10
4
coil
sus
5 pa
adq 3.8
ade
dae
4
daq 3.6
tot
[V]
3
[V]
3.4
2 3.2
1 3
0 −4 −3 −2 −1 0 1 2 3 2.8 −1 0 1 2 3
10 10 10 10 10 10 10 10 10 10 10 10 10
Figure 4.8 Validation of the controller output u spectra using the 1000-10 preamplifier and an
H2 controller. above: Simulated ASD ( left) and measured ASD ( right) of the controller output
signal using an H2 controller and a preamplifier.below: Simulated CAS ( left) and measured CAS
( right) of the controller output signal using an H2 controller and a preamplifier.
overestimate the measured level, therefore, the perfect match is more of a coincidence than
due to prediction accuracy. The H2 controller achieves a lower controller input level than
the 100 mHz controller. The measured controller input level decreased from 1.2 mVrms to
0.67 mVrms , an improvement of 50 %! On the right hand side of Figure 4.9 one can see the
10 sec time trace on which the measured ASD and CAS in Figure 4.7 is based.
Looking at Figure 4.8, one can see that the simulated controller output (which is the same
as the control effort) of 5.7 mVrms is still mainly due to the low frequency preamp noise,
as can be seen in the simulated CAS. The measured controller output level also shows this
low frequency power, but because the measured output was only measured over a period
of 10 seconds, the lowest measured frequency is only 0.1 Hz. The low frequency noise of
the preamp can therefore not be validated from this experiment. It is very apparent though
when one looks at a controller output time trace as shown on the left hand side in Figure 4.9;
Time trace, σ = 3.99 mV Time trace, σ = 0.69 mV
2.5
−20
2
1.5
−25
1
Volts [mV]
Volts [mV]
0.5
−30
0
−0.5
−35
−1
−40 −1.5
−2
−2.5
0 2 4 6 8 10 0 2 4 6 8 10
Time [sec] Time [sec]
Figure 4.9 left: Time trace of a controller output signal using an H2 controller and the 10-1000
preamp. The low frequency noise of the preamp is clearly visible as the controller output signal
slowly moves up and down. right: Time trace of the controller input signal using an H2 controller
and the 10-1000 preamp.
the control signal is slowly drifting up and down.
4.4 Making a design decision using H2 control in DEB

4.4.1 The best preamplifier
With the DEB strategy in mind, one can ask what the best preamplifier transfer function is
considering system performance? The preamplifier was introduced to improve the SNR of
the geophone signal and its transfer function was designed with the geophone velocity sensi-
tivity in mind, see § 3.5.2. But it is not guaranteed at all that the used preamplifier transfer
function is the best transfer function to use when considering performance. In this para-
graph the performance of a different preamplifier transfer function will be investigated using
the DEB analysis approach. An H2 controller will be used to cancel the degree of freedom
introduced by the controller, optimizing the system configuration for maximum performance.
To improve the SNR of the geophone signal the most simple strategy that comes in mind is to
pre-amplify the signal as much as possible. There are two things that limit the preamplifier
gain though; first, the preamplifier introduces noise. In general can be said, the more gain
the preamplifier provides, the more noise it will introduce (see § 3.5.3). So, introducing gain
comes with a price: an increase of noise in the circuit. Second, the gain of the preamplifier
is limited by clipping of the A/D converter, which can only cope with signals in between
±10 V . A practical limit in the open loop setting turned out to be a gain of 100 (the open
loop setting is needed during start up of the system and is the worst case; it has the largest
preamplifier output values, since the vibrations are not yet compensated for).
4.4. MAKING A DESIGN DECISION USING H2 CONTROL IN DEB 57
balancing of disturbances
A good design consideration to make is: how much preamplifier gain is actually needed?
How high must the SNR be? Actually, the SNR must be considered frequency dependent;
in one frequency region the SNR can be high enough, while in an other region the noise
dominates the signal. The DEB performance analysis can give some clarity in this matter.
One can say: the preamplifier has to amplify a certain frequency region just enough, such
that on the one hand, the ADC noise are suppressed and on the other hand, the introduced
preamplifier noise does not dominate the region. There is a balance that has to be found.
In short, balancing of disturbances implies that the disturbance introduced by e.g., an extra
gain and the disturbance itself (which the gain is supposed to suppress by improving the
SNR) are balanced against each other such that they contribute an equal amount of power
(over frequency) at the performance output.
Analyzing the performance CAS in Figure 4.6, one can see that the preamplifier noise now
contributes a significant amount of power in the low frequency region up to 0.1 Hz. The
A/D electrical noise then adds a significant part in the 1 kHz region. From this analysis
the following questions arise: Does the preamplifier have to much gain in the low frequency
region, introducing more preamplifier noise than necessary? Does the preamplifier have
enough gain in the high frequency region, since the A/D electrical noise still pops up? In
the following paragraph a different preamplifier will be analyzed in order to try to answer
the questions above.
4.4.2 The constant gain preamplifier

To improve the performance as discussed above, less gain was needed in the low frequency
region and more in the high frequency region. The old preamplifier transfer function (see
Figure 3.15) then suggest that a constant gain preamplifier with the gain somewhere in
between its low frequency gain of a 1000 and its high frequency gain of 10 would be a better
choice. Therefore a preamp with a gain of 100 is chosen. Since the transfer function of the
preamplifier is changed, its noise model is changed accordingly, as well and the weighting
filter Vw1 used to synthesize the H2 controller.
Judging the preamp performance by its Pareto curve
Every different system configuration will obtain a certain performance level given a certain
control effort level, defined by the Pareto curves. To compare the configurations, the two
Pareto curves are computed and shown in Figure 4.10 on the left hand side.
Both the curves start at the open loop performance level of 98.8 µm/sec. As the control effort
increases, the performance increases as well. Up to 3 mVrms control effort the curves almost
coincide, meaning that it does not really matter if the 1000-10 preamp is used or the 100
preamp. But, above 3 mVrms control effort the 100 preamp curve drops below the 1000-10
preamp curve, meaning the the 100 preamp configuration can achieve a higher performance
level than the 1000-10 preamp case. The absolute performance of the 1000-10 preamp case
is 0.66 µm/s, compared to 0.55 µm/s in the 100 preamp case, a 20 % improvement!
−7 CAS towards performance output, σ = 0.55 µm/s

x 10
Pareto curves for the 1000−10 and 100 preamp 6
ground
coil
2 sus
10 5 pa
adq
ade
dae
Performance [µ m/sec]
4
daq
tot
[m/s]
1
10 3
0 curve using 1000−10 preamp

10 1
curve using 100 preamp
used level with 1000−10 preamp
used level with 100 preamp
0 −4 −3 −2 −1 0 1 2 3
−2 −1 0 1
10 10 10 10 10 10 10 10 10 10 10 10
Control effort [mV] Frequency [Hz]
Figure 4.10 left: Pareto curves using the 1000-10 preamp and the 100 preamp. right: Simulated
CAS of the performance signal Ẋl using the 100 preamp and an H2 controller.
On the right hand side Figure 4.10 the simulated CPS of the performance output can be seen.
When analyzing the power contribution of the ADC electrical noise and the preamplifier noise
one can see that their contributions are now well balanced (both contribute about the same
amount of power at each frequency). So, designing the preamplifier by analyzing the noise
levels of the preamplifier and the ADC and try to balance them, seems a good strategy to
improve the system performance.
What can be concluded from these results is that the frequency dependent second order (lag
shaped) preamplifier based on the geophone velocity sensitivity, does not result in a system
which handles noise best. With the much simpler first order constant gain preamplifier a
higher performance level can be achieved. There is one disadvantage to the constant gain
preamplifier though, the constant gain 100 preamp makes the risk of clipping the A/D
converter about 10 times higher. If the system clips when the loop is not yet closed, this
can result in start-up problems. If clipping is not a problem e.g., when using a dual gain
preamplifier, which can switch from low gain to high gain when the system is at its normal
working point, then the constant gain 100 preamp is the preferred preamplifier. It is not
only much easier to build (only one stage using one opamp and two resistors) but it can also
achieve a higher performance level than the 1000-10 preamp.
4.5 Conclusions
The H2 control strategy is successfully applied to the vibration isolation system. An
H2 controller is synthesized and the (theoretical) performance of the system is maximized,
showing the system designer the performance potential of the system concept. The simulated
results show a significant increase in performance when using an H2 controller compared to
the (standard) controller used in the previous chapter. The performance level dropped from
2.9 µm/s, when using a preamplifier and standard controller, to 0.66 µm/s using the same
4.5. CONCLUSIONS 59
preamplifier and an H2 controller.
The performance level achieved by the H2 controller is validated with experiments, show-
ing that the theoretical increase in performance displays a realistic image. The synthesized
H2 controller is therefore a good indicator for the control designer how to control the system
in practice to optimize the system performance. The 0.1 Hz bandwidth used in the previous
chapter turned out not to be such a good control strategy, since it needed in a very high low
frequency controller gain, resulting in a high level of low frequency noise at the performance
output. The H2 controller suggested a 0.3 Hz bandwidth, resulting in significantly less low
frequency noise at the performance outputs.
The use of Pareto curves to judge different system also turned out to be feasible in practice.
Two different preamplifier designs were compared and their performance analyzed. Their
Pareto curves showed that the much simpler constant gain 100 preamp can achieve a higher
performance level than the second order 1000-10 preamp designed in the previous chapter.
This design exercise shows that the use of H2 control in the DEB design approach can be
of great value for the designer.
Chapter 5
Case study, Advanced Isolation

ModuleS
5.1 Introduction
The research presented in this chapter is done during a 3 months stay at Philips Center
for Industrial Technology (Philips CFT) in Eindhoven, the Netherlands. At Philips CFT a
new vibration isolation system is being developed which carries the name ’Advance Isola-
tion ModuleS’ (AIMS). The main challenge of this case study is to apply the DEB design
approach in the way it was developed for; to predict the performance of a vibration isolation
system which is still on the drawing board, check if the desired performance specification is
met and point out critical performance limiting components.
The AIMS system is designed to be used parallel with another passive isolator system.
In Figure 5.1 this configuration is illustrated. The performance of the passive isolator,
here characterized by its main resonance frequency of 2 Hz, is improved by the parallel
introduction of the AIMS. A crucial question to the system designer now is: is the total
isolating performance indeed improved by the introduction of the AIMS? Another interesting
question is: Is the performance of the AIMS parallel with the passive isolator, comparable
with the performance of the better and more expensive 0.5 Hz passive isolator? Since it is
a design aim that the AIMS and 2 Hz isolator together is less expensive than the 0.5 Hz
isolator, this would imply a reduction in costs of the total isolation system.
The goals of this chapter can be formulated as follows: Show that the DEB approach can
be applied during the design phase of the AIMS system. By doing this, find out if the
AIMS concept is feasible and point out the performance limiting components. Investigate
the influence of different control strategies to the final performance of the system. Compare
the performance of the AIMS (parallel with the 2 Hz passive isolator) with the performance
of a 0.5 Hz passive isolator.
61
62 CHAPTER 5. CASE STUDY, ADVANCED ISOLATION MODULES
Figure 5.1 The AIMS vibration isolator is designed to be used parallel with a passive isolator.
5.2 Advanced vibration isolation

The AIMS works with a different isolation concept than the isolator in the previous case
study. This new concept is referred to as Advanced Vibration Isolation (AVI) and is ex-
plained in this section. The AVI concept is based on a ’payload system’ mimicking an other,
’ideal system’. The payload system and the ideal system are both subject to the same in-
put vibrations, and since the payload system ’follows’ the response of the ideal system, the
payload system now also has ideal system behavior. The output of the ideal system actually
acts as a reference for the payload system.
An ideal vibration isolation system, is softly suspended to the ground and very strongly
attached to a virtual point in the sky, a configuration referred to as a ’sky hook configura-
tion’. This way the ground vibrations do almost not propagate into the system and payload
disturbance forces almost have no effect on the payload since the system is hooked to the
sky. In the AVI concept this behavior is approached as follows. The ideal system consists
of a reference mass which is suspended in a control loop at 0.5 Hz, which we will call the
reference loop. The main system follows the position of the reference mass via another con-
trol loop, the payload loop. The payload reacts on vibrations as an (almost ideal) 0.5 Hz
suspended system. Payload disturbance forces Fd act on the payload part of the AVI system
and not on the reference mass and thus do not cause any disturbance forces on the refer-
ence mass. Because the payload follows the reference mass, the payload loop will suppress
displacements due to payload disturbance forces, achieving the sky hook effect. A model of
the system configuration can be seen in Figure 5.2.
Besides the AIMS, the payload mass is also suspended with a simple vibration isolation
system, which is modelled with a spring kp . This simple isolator can be seen as a basic
passive isolation system and the AIMS as an add-on system, to further improve the isolation
performance.
The relative position xs − xh between the reference mass and the ground is measured with
a capacitive position sensor, and fed back via a reference controller into a Lorentz actuator,
5.3. TRANSFER FUNCTION OF THE AIMS 63
Figure 5.2 AVI model
which generates a force Fs . The closed loop would typically achieve a 0.5 Hz suspension
behavior.
The relative position xp − xs between the reference mass and payload mass is also measured
with a capacitive position sensor, which is fed back via a payload controller into a second
lorentz actuator which generates a force Fp between the ground and the payload mass. The
force Fp is the actual compensating force which actively isolates the payload mass from
vibrations.
5.2.1 Performance specification

The AIMS is designed to isolate high performance machinery from vibrations. A typical
maximum acceleration level for this kind of machinery has a 1-σ value of 1 mm/s2 . In
order to achieve this acceleration level it is believed that the isolator may only contribute
about 10% of that value (Vervoordeldonk [26]). So in this case, the noise level of the AIMS
(excluded the contribution by ground vibrations) may not be larger than 0.1 mm/s2 . This
performance value will be the aim of this case study. The frequency distribution of the noise
is also of importance. In general the power of the noise should be located well below the
first internal resonance frequency of the payload.
5.3 Transfer function of the AIMS

The AIMS system contains two separate control loops, the reference loop and the payload
loop, as can be seen in the block diagram in Figure 5.3.
Figure 5.3 AVI control configuration
The reference loop is totally independent from the payload loop. It acts as a reference gen-
erator for the payload loop, giving the payload loop a position reference: xs . This reference
acts as a disturbance on the payload loop. The better the payload loop can suppress this
disturbance, the more the payload will mimic the 0.5 Hz behavior of the reference mass,
achieving vibration isolation within the bandwidth of the payload loop.
The ground vibrations xh enter the system at two different locations; as a position distur-
bance in the reference loop and, via the basic isolation system, as a disturbance force in the
payload loop. Payload disturbance forces Fd enter the system in the payload loop and can
not influence the reference loop, giving the system the ’sky hook’ effect within the payload
loop bandwidth.
5.3.1 Reference loop transfer function

In this section all the components of the reference loop will be discussed briefly and the loop
transfer function will be determined.
The first component in the reference loop is the capacitive reference sensor, which converts
a position into a voltage signal using a capacitive measuring technique. It has a sensitivity
of 3300 V /m and contains an internal second order low pass filter with cut-off frequencies
at 100 and 160 Hz (transfer function not shown):
µ ¶µ ¶
2π100 2π160
Psen,r (s) = 3300 [V/m] (5.1)
s + 2π100 s + 2π160
The set-point, where the height between the ground and the reference mass can be set, has
a gain on its own. This can be modelled as a addition point as shown in Figure 5.3 in series
with a static gain Pset,r . The gain of the reference set-point is 0.2. The set-point is discussed
in more detail in Section 5.4.2 on page 72. Looking at Figure 5.11 on page 73, the (static)
transfer function can written as −Rspf /Rspi .
Pset,r = 0.2 [V/V] (5.2)
The reference controller is designed to achieve a closed loop 0.5 Hz suspension behavior of
the reference mass. To achieve this a PID controller is used with a transfer function which
can be seen in Figure 5.4. The transfer function of the PID controller is discussed in more
detail in Section 5.4.2 on page 72.
µ ¶µ ¶
s + 0.45 1.06s + 1
Ppid,r (s) = 0.051 [V/V] (5.3)
s + 0.041 0.21s + 1
Reference PID controller

−5
Magnitude [dB]
−10
−15
−20
−2 −1 0 1 2
10 10 10 10 10
210
Degree [deg]
180
150
120 −2 −1 0 1 2
10 10 10 10 10
Frequency [Hz]
Figure 5.4 Reference PID controller
The controller voltage signal is transformed into a current signal using a servo amplifier.
The reference servo amplifier is referred to as the ’umi’ and has a static gain of 0.0096 and
an internal first order low pass filter with a cut-off frequency of 2 kHz (transfer not shown).
µ ¶
2π2000
Pumi (s) = 0.0096 [A/V] (5.4)
s + 2π2000
The Lorentz actuator converts the current signal into the force Fs . The gain of the actuator
is static and has the value 1.2 N/A.
Plor,r = 1.2 [N/A] (5.5)
Finally, the open loop model of the reference mass dynamics is given by
xs 1
= 2
(5.6)
Fs ms s + bs s
where ms = 0.12 kg is the mass of the reference sensor, and bs = 0.01 N s/m models the
parasitic damping of the air bearing of the reference mass.
The series connection of all these components returns the loop gain function Lr (s).
xs
Lr (s) = Psen,r (s) Pset,r Ppid,r (s) Pumi (s) Plor,r [m/m] (5.7)
Fs
A Bode diagram of the reference loop gain function Lr (s), together with its complementary
sensitivity function Lr (s)(Lr (s) + 1)−1 , can be seen in Figure 5.5. From the complementary
sensitivity one can conclude that the closed loop reference system indeed has as a 0.5 Hz
suspension behavior.
Reference loop transfer functions
100
reference loop transfer function
complementary sensitivity function
Magnitude [dB]
50
−50
−100
−2 −1 0 1 2
10 10 10 10 10
0
−60
Degree [deg]
−120
−180
−240
−300
−2 −1 0 1 2
10 10 10 10 10
Frequency [Hz]
Figure 5.5 Reference loop gain function Lr (s), together with the complementary sensitivity
function. From the latter one can conclude that the closed loop reference system has as a 0.5
Hz suspension behavior.
5.3.2 Payload loop transfer function

In this section all the components of the payload loop will be discussed briefly and the loop
transfer function will be determined.
The capacitive payload sensor has a sensitivity of 9000 V /m and an internal second order
low pass filter with cut-off frequencies at 1 kHz and 2 kHz (transfer function not shown).
µ ¶µ ¶
2π1000 2π2000
Psen,p (s) = 9000 [V/m] (5.8)
s + 2π1000 s + 2π2000
The gain of the payload set-point is 10.
Pset,p = 10 [V/V] (5.9)
The payload controller is designed such that the closed payload loop can suppress reference
disturbances up to 30 Hz. To achieve this, a PID controller Ppid,p (s) is used, of which the
transfer function can be seen in Figure 5.6.
µ ¶µ ¶
s + 27 0.013s + 1
Ppid,p (s) = 0.58 [V/V] (5.10)
s + 2.4 0.0027s + 1
Payload PID controller

20
15
Magnitude [dB]
10
5
0
−5
−2 −1 0 1 2
10 10 10 10 10
210
Degree [deg]
180
150
120
−2 −1 0 1 2
10 10 10 10 10
Frequency [Hz]
Figure 5.6 Payload PID controller
The payload servo amplifier, referred to as ’uma’, has a gain of 15 and also an internal first
order low pass filter with a cut-off frequency of 2 kHz (transfer not shown).
µ ¶
2π2000
Puma (s) = 15 [A/V] (5.11)
s + 2π2000
The gain of the payload Lorentz actuator is 9.5 N/A.
Plor,p = 9.5 [N/A] (5.12)
The open loop model of the payload mass xp /F also contains the basic isolator dynamics
and is given by
xp 1/mp
= 2 (5.13)
F s + 2ζp ωn s + ωn2
where mp = 500 kg is the mass of the payload, ζp = 0.05 and ωn = 2π(2Hz) rad/sec (The
damping of the airmount is not shown in Figure 5.2).
The series connection of all these components returns the payload loop gain function Lp (s).
xp
Lp (s) = Psen,p (s) Pset,p Ppid,p (s) Puma (s) Plor,p [m/m] (5.14)
F
The loop gain function Lp (s) can be seen in Figure 5.7, together with the sensitivity function
(Lp (s) + 1)−1 of the payload loop. From the sensitivity function one can conclude that the
closed loop payload system can suppress reference disturbances up to 30 Hz, enabling the
AIMS to improve the isolation performance compared to the basic isolation system up to
this frequency.
Payload loop transfer functions

80
Magnitude [dB]
40
−40
−80 payload loop transfer function

sensitivity function
−2 −1 0 1 2
10 10 10 10 10
240
Degree [deg]
120
0
−120
−240
−360
−2 −1 0 1 2
10 10 10 10 10
Frequency [Hz]
Figure 5.7 Payload loop gain function Lp (s) together with its sensitivity function. From the
sensitivity function one can see that payload loop reference disturbances are suppressed up to 30
Hz.
5.3.3 Open and closed loop transfer function of total system

The payload has to be isolated from vibrations. An important measure of performance is the
acceleration of the payload x¨p , since below the first resonance frequency of the payload (e.g.,
100 Hz) accelerations of the payload are proportional with physical deformations within the
payload, causing measurements errors etc. When the ground vibrations are described by an
acceleration spectrum [x¨h 2 /Hz], the transfer function to analyze is x¨p /x¨h , which is equal to
the transfer function xp /xh . One can see that this transfer function is given by:
xp
= Tr Tp + kp [xp /F ]Sp (5.15)
xh
5.4. MODELLING OF DISTURBANCES AND NOISE SOURCES 69
where Tr and Tp are the complementary sensitivity function of the reference loop and payload
loop respectively, and Sp is the sensitivity function of the payload loop. This closed loop
transfer function is called the transmissibility, and defines the performance of the AIMS in
terms of vibration isolation. The open loop transfer function is also given for comparison:
open loop = kp [xp /F ] (5.16)
Both these transfer functions are plotted in Figure 5.8. Comparing the two transfer functions,
one can see that the closed loop AIMS improves the vibration isolation performance in the
0.8 Hz - 20 Hz region.
Open and closed loop 30 Hz transmissibility
20
Magnitude [dB]
0
−20
−40
−60
−80 open loop
−100 30 Hz transmissibility
−1 0 1 2
10 10 10 10
60
−60
Degree [deg]
−180
−300
−420
−540
−1 0 1 2
10 10 10 10
Frequency [Hz]
Figure 5.8 The open and closed loop transfer function (30 Hz transmissibility) of the AIMS.
5.4 Modelling of disturbances and noise sources

In this section the disturbance and noise sources acting on the AIMS are discussed. Several
physical disturbances act on the AIMS, like ground vibrations, payload disturbance forces,
and air pressure variations. Next to that, there are several electrical noise sources originating
from the electrical circuitry, e.g., the capacitive sensor and controller electronics.
Figure 5.9 presents the identified noise sources in a schematic. Transfer functions are denoted
with a P , and the power spectral densities [SI2 /Hz ], used to model the disturbance sources,
are denoted with a S. In total, 11 noise sources are identified, giving the closed loop system
11 inputs, w1 ...w11 . In the following sections the disturbance sources will be modelled.
The transfer functions P.. are analogue to Figure 5.3 on page 64, except that the transfer
function Psys,p now has two outputs: xp and x¨p , the position and acceleration of the payload
respectively. The payload acceleration is defined as the performance output of the closed
loop system.
Figure 5.9 Closed loop AIMS with disturbance and noise sources modelled with Power Spectral
Densities S.. . The acceleration of the payload is defined as the performance output.
5.4.1 Ground vibrations

The most important disturbance source in this vibration isolation case study are the ground
vibrations. To model the ground vibrations a filter is used which is based on the ground
vibration measurements shown in Figure 3.10 on page 29. The filter is developed in a similar
manner as the filter (3.7), but now it is expressed in terms of acceleration. The filter is given
by:
Ã !3 Ã !3
3 5 √
−6 2π jω + 1 2π jω
V (jω)f loor = 10 3 5 [(m/s2 )/ Hz] (5.17)
2π10 jω + 1 2π jω +1
The 3rd order high pass filter (last factor on the right hand side) was added because one
does not want the filter to define a DC position disturbance on the AIMS (this would imply
that the ground itself is off-set with some constant value, which is unrealistic). The relation
with the PSD Sf loor shown in Figure 5.9 is as follows:
Sf loor (ω) = Vf loor (jω)Vf∗loor (jω) (5.18)

where V ∗ (jω) denotes the complex conjugate transpose of V (jω).
As stated earlier, the ground vibrations enter the system at two different locations, via the
airmount and as a position disturbance in the reference loop.
5.4.2 Electrical noise sources

The payload loop and reference loop both consist of the same electrical components: a ca-
pacitive sensor, a set-point, a PID controller and a servo amplifier. Each of these components
introduces an amount of noise. The noise models will be described below.
Capacitive sensor noise
At the time of this research, the two capacitive sensors are still in development. Therefore
the modelled noise level here can deviate from the final noise model and the model must thus
be seen as a preliminary model. The noise of the capacitive payload sensor was measured
using an HP Dynamic Signal Analyzer. The measured power spectral density can be seen
in Figure 5.10. The PSD of the sensor noise can be modelled with a fourth order model
with corner frequencies at 70 Hz and 700 Hz and an asymptotic high frequency level of
(10n2 ) Vrms
2 /Hz. The sensor noise model is modelled such that its power matches the
measured power level over the 4Hz - 800Hz frequency grid. This can be seen in the right
part of Figure 5.10. The payload sensor has an internal low pass filter as was discussed in
the previous section. The noise model must therefore be multiplied with the square of the
low pass filter transfer function to make sure the high frequency energy content of the model
corresponds with reality. The model is denoted by Ssen,p (ω), where the subscript refers to
payload sensor, and can be written as:
µ ¶4 µ ¶2
jω + 2π700 2π1000 2π2000
Ssen,p (ω) = (10n)2 [V 2 /Hz] (5.19)
jω + 2π70 jω + 2π1000 jω + 2π2000
This model can also be seen in Figure 5.10.

The same model is used for the reference sensor, which is equal to the payload sensor (it
was not measured separately), except that its low pass filter is different. To be safe, its
asymptotic noise level is taken 1.5 times higher. The noise PSD model used for the reference
sensor is denoted by Ssen,r (ω):
µ ¶4 µ ¶2
2 jω + 2π700 2π100 2π160
Ssen,r (ω) = (15n) [V 2 /Hz] (5.20)
jω + 2π70 jω + 2π100 jω + 2π160
The surface of the sensor noise PSD is equal to the variance, or squared standard deviation,
of the noise signal. The standard deviation found for the payload sensor equals 7.3 µVrms
(over a frequency grid from 4Hz to 800Hz). When the sensor noise PSD is converted to the
input of the sensor, an equivalent position noise level can be computed, giving a value of
0.81 nm on the same frequency grid.
−11payload sensor noise [V], σv = 7.1 µVrms, σm = 0.81 nm

CPS
PSD payload sensor noise [V/Hz1/2], σv = 7.1 µVrms, σm = 0.81 nm x 10
6
−10
10
Measured noise
Noise model 5
−12
10
4
−14
10
[V2]
[V2/Hz]
−16
10
2
−18
10 1
Measured
Model
−20
10 0
0 100 200 300 400 500 600 700 800 0 100 200 300 400 500 600 700 800
Figure 5.10 left: PSD of the payload sensor noise in the 4-800 Hz range. The noise model
Ssen,p (ω) is also plotted right: The CPS of the measured sensor noise in the 4-800Hz range
and the CPS of the noise model. The noise model is modelled such that the measured power
matches the modelled power.
Set-point noise
The set-points electrical circuits comprises an operational amplifier, the op27, and four re-
sistors, as can be seen in Figure 5.11. The resistors generate thermal noise and the opamp
generates voltage noise and current noise. set-points. The thermal noise generated by the
resistors was accounted for in this analysis as described by Rodgers [18], but is not worked
out here. Since the contribution to the performance output by the set-point noise sources
will turn out to be of minor importance (see § 5.5.2), not all noise sources are modelled here,
only the contribution by the opamp noise is discussed.
The voltage and current noise model of the op27 opamp have already been discussed in
Section B.4.2, but are repeated here for convenience:
µ ¶
2πfv
Sen (ω) = e2o 1+ 2
[Vrms /Hz] (5.21)
ω
µ ¶
2πfc
Sin ± (ω) = i2o 1+ [A2rms /Hz] (5.22)
ω
with parameter values as listed in Table 5.1.

The propagation of the noise towards the output is very similar to the preamp case covered
in Section B.4.3 and is based on Rodgers [18].
Asymptotic high frequency voltage level eo : 3.0 · 10−9 Vrms

Asymptotic high frequency current level io : 0.4 · 10−12 Arms
Voltage noise corner frequency fv : 2.7 Hz
Current noise corner frequency fc : 140 Hz
Table 5.1 Specifications of the op27
Figure 5.11 Set-point circuit configuration with a voltage noise model en and two current noise
models in− and in+ .
Voltage noise
The noise at the set-point output due to en is denoted by Sspo,en (read subscript as: set-point
output due to en ). From the voltage divider relationship between en and Espo , Sspo,en can
be obtained as:
µ ¶
Rspf 2
Sspo,en (ω) = 1 + Sen (ω) (5.23)
Rspi
Current noise
The noise at the set-point output due to in is denoted by Sspo,in . The currents coming from
in+ can not flow through the resistors Rspi and Rspf (below), since they are grounded at the
bottom. Their top side is at virtual ground, so there will be no potential difference across
the two resistors due to in+ .
The currents from in− will not flow through Rspi (above), because the left side is at ground
and the right side is at virtual ground, so there is no potential difference across Rspi due to
in− . Therefore, all of in− flows through Rspf , and none of it flows through Rspi .
So for the current noise at the set-point output due to in− we can write :
2
Sspo,in− (ω) = Rspf Sin (ω) (5.24)
Adding up both obtained PSD’s gives the total PSD of the set-point noise at the set-point
output Sspo :
µ ¶
Rspf 2 2
Sspo (ω) = 1 + Sen (ω) + Rspf Sin (ω) (5.25)
Rspi
The set-point noise PSD at the set-point output in the payload loop will be denoted by Sset,p
and the PSD in the reference loop by Sset,r . The resistor values for Rspi and Rspf for the
payload loop are 5kΩ and 50kΩ resp. and for the reference loop 10kΩ and 2kΩ resp.
Analogue PID controller noise

A schematic of an analogue PID controller can be seen in Figure 5.12. The transfer function
from Epidi to Epido is given by −Zf (s)/Zi (s), which defines the transfer function of the PID
controller. The transfer functions of Zf (s) and Zi (s) are given by:
µ ¶
τ2 s + 1
Zf (s) = Rf 1 (5.26)
τ1 s + 1
µ ¶
1 τ3 s + 1
Zi (s) = (5.27)
Ri1 τ4 s + 1
where the time constants can be written in resistor and capacitor values as follows
τ1 = Cf (Rf 1 + Rf 2 ) (5.28)
τ2 = Cf (Rf 2 ) (5.29)
τ3 = Ci (Ri1 + Ri2 ) (5.30)
τ4 = Ci (Ri2 ) (5.31)
making the total PID transfer function:
µ ¶µ ¶
Rf 1 Rf 2 Cf s + 1 (Ri1 + Ri2 )Ci s + 1
Ppid (s) = − (5.32)
Ri1 (Rf 1 + Rf 2 )Cf s + 1 Ri2 Ci s + 1
which can be written in classical PID form:
Ã !µ ¶
Rf 1 ||Rf 1 s + Rf 21Cf (Ri1 + Ri2 )Ci s + 1
Ppid (s) = − 1 (5.33)
R Ri2 Ci s + 1
| {zi1 } s + (Rf 1 +Rf 2 )Cf | {z }
P
| {z }
D
I
The resistor and capacitor values for the payload and reference controller are listed in Ta-
ble 5.2.
Thermal noise
The thermal noise generated by the resistors was accounted for in this analysis as described
by Rodgers [18], but is not worked out here. Since the contribution to the performance out-
put by the controller noise sources will turn out to be of minor importance (see Section 5.5.2)
only the contribution by the opamp noise is covered.
Figure 5.12 Schematic of analogue PID controller
Payload Reference
Cf : 0.29 µF Cf : 83 mF
Rf 1 : 1.27 M Ω Rf 1 : 264 kΩ
Rf 2 : 0.127 M Ω Rf 2 : 26.4 kΩ
Ci : 53 nF Ci : 1.8 µF
Ri1 : 200 kΩ Ri1 : 470 kΩ
Ri2 : 51.7 kΩ Ri2 : 117 kΩ
Table 5.2 Parameters of the PID controllers
Voltage noise
The noise at the PID output due to en is denoted by Spido,en . From the voltage divider
relationship between en and Epido , Spido,en can be obtained as (see Rodgers [18]):
¯ ¯
¯
¯ Zf (jω) ¯¯2
Spido,en (ω) = ¯1 + Sen (ω) = |1 − Ppid (jω)|2 Sen (ω) (5.34)
Zi (jω) ¯
since −Zf (jω)/Zi (jω) = Ppid (jω).
Current noise
There is no noise generated by in since it directly connected to ground.
For currents from in− the same reasoning as in the set-point case holds, therefore, all of in−
flows through Zf , and none of it flows through Zi (Rodgers [18]).
So for the current noise at the PID output due to in− we can write:
Spido,in− (ω) = |Zf (jω)|2 Sin (ω) (5.35)
Adding up both obtained PSD’s gives the total PSD at the PID output:
Spido (ω) = |1 − Ppid (jω)|2 Sen (ω) + |Zf (jω)|2 Sin (ω) (5.36)
The PID noise in the payload loop will be denoted by Spid,p and the noise in the reference
loop by Spid,r .
Servo amplifier noise

The task of the servo amplifier is to generate a current Iumao to drive the Lorentz actuator.
A simplified electrical circuit configuration can be seen in Figure 5.13.
Figure 5.13 Schematic of servo amplifier
The input of the amplifier is the voltage signal Eumai . The opamp configuration will drive
the power opamp that will provide for the necessary currents. The feedback loop via resistor
Rf 1 closes the loop. It can be seen that the transfer function from input voltage Eumai to
output current Iumao is equal to Rf 1 /(Ri Rs ).
The resistor and capacitor values for the payload and reference controller are listed in Ta-
ble 5.3.
Payload Reference
Cf : 100 nF Cf : 100 nF
Rf 1 : 6 kΩ Rf 1 : 750 Ω
Rf 2 : 10 kΩ Rf 2 : 6 kΩ
Ri : 2 kΩ Ri : 390 kΩ
Rs : 0.2 kΩ Rs : 0.2 kΩ
Table 5.3 Parameters of the servo amplifiers
The following noise analysis is based on Goossens [8]. The noise of the power opamp is
negligible at the output Iumao because of the suppression of the feedback loop. The noise
sources of the opamp cannot be neglected and will be modelled. The used opamp is an
op27, and its noise can be modelled with three equivalent noise sources, also is shown in
Figure 5.13. The currents of source In+ cannot go anywhere, so this source will not cause
noise at the output of the servo amp. The two remaining sources will cause noise at the
current output as follows:
¯ ¯2
¯ Zf (jω) ¯
Suma,en (ω) = ¯¯1 + ¯ Sen (ω) [A2 /Hz] (5.37)
Ri ¯
Suma,in− (ω) = |Zf (jω)||Rf 1 |2 Sin (ω) [A2 /Hz] (5.38)

where Sen and Sin are the voltage and current PSD of the op27 resp. as given in Section 5.4.2.
Adding up both obtained PSD’s gives the total PSD at the servo amp output:
¯ ¯
¯
¯ Zf (jω) ¯¯2
Sumao (ω) = ¯1 + Sen (ω) + |Zf (jω)||Rf 1 |2 Sin (ω) (5.39)
Ri ¯
The servo amp in the reference loop is similar to the one in the payload loop. The noise in
the payload loop will be denoted by Suma and the noise in the reference loop by Sumi .
5.4.3 Other disturbances and noise sources

Reference and payload mass disturbance forces
Disturbance forces can act on the reference mass and the payload mass. Reference distur-
bance forces can be caused by friction in the bearing or air pressure fluctuations above and
underneath the reference mass. Acoustic noise can also be an important noise source. Un-
fortunately, it is unknown what the magnitudes these forces are and it is thus not possible
yet to model these sources. Therefore, reference mass disturbance forces are not analyzed
any further.
Payload disturbance forces can also be caused by air pressure variations and acoustic noise,
but will mainly be due to forces caused by the payload itself, like internal vibrations or
moving parts. These payload disturbance forces are characteristic to a certain payload
application, and since the application is not fixed, it is not possible to make a model at this
point.
5.4.4 Summary of disturbances and noise sources

Now all the disturbance and noise sources are discussed, lets√ summarize the results:
First of all, the floor noise was modelled as Vf loor [m/s2 / Hz]:
Ã !3
3
2π jω + 1
V (jω)f loor = 10−6 ∗ 3 (5.40)
2π10 jω + 1
Then eight electrical noise sources were modelled as PSDs; the capacitive sensor noise in
the payload and reference loop, Ssen,p , Ssen,r , the set-point noise in both the loops: Sset,p ,
Sset,r , and both the analogue PID controller noise sources: Spid,p , and Spid,r , all with units
2 /Hz]. Finally the two servo amp PSDs S

[Vrms umi and Suma were modelled with units
2
[Arms /Hz].
µ ¶4 µ ¶2
2 jω + 2π700 2π1000 2π2000
Ssen,p (ω) = (10n) (5.41)
jω + 2π70 jω + 2π1000 jω + 2π2000
µ ¶4 µ ¶2
2 jω + 2π700 2π100 2π160
Ssen,r (ω) = (15n) (5.42)
jω + 2π70 jω + 2π100 jω + 2π160
µ ¶2
5 · 103
Sset,p (ω) = 1 + Sen (ω) + (50 · 103 )2 Sin (ω) (5.43)
50 · 103
µ ¶2
10 · 103
Sset,r (ω) = 1 + Sen (ω) + (2 · 103 )2 Sin (ω) (5.44)
2 · 103
Spid,p (ω) = |1 − Ppid,p (jω)|2 Sen (ω) + |Zf p (jω)|2 Sin (ω) (5.45)
Spid,r (ω) = |1 − Ppid,r (jω)|2 Sen (ω) + |Zf r (jω)|2 Sin (ω) (5.46)
¯ ¯
¯
¯ Zumaf (jω) ¯¯2
Suma (ω) = ¯1 + ¯ Sen (ω) + |Zumaf (jω)||Rumaf 1 |2 Sin (ω) (5.47)
Rumai
¯ ¯
¯
¯ Zumif (jω) ¯¯2
Sumi (ω) = ¯1 + Sen (ω) + |Zumif (jω)||Rumif 1 |2 Sin (ω) (5.48)
Rumii ¯
The disturbance forces on the payload mass and reference mass are unknown and thus Sdist,p
and Sdist,r are not modelled.
5.5 System performance

In this section the output noise level is calculated, using the noise models from the previous
section. In order to calculate the noise propagation through the closed loop system, one needs
a closed loop model, with inputs where the noise sources apply and with the performance
variable as output. After that, the noise level at the performance output due to all the
internal AIMS noise sources (so excluded ground vibrations) is calculated; every noise model
is propagated towards the performance output of the system and the total noise level due
to the internal AIMS noise sources is then calculated by adding the separate contributions
quadratically. Finally, the vibration isolation performance of the AIMS is evaluated by
comparing the performance of the passive basic isolator with the performance of the basic
isolator together with the AIMS, which is build up out of two parts, one part due to the
internal AIMS noise sources and one part due to the ground vibrations.
5.5. SYSTEM PERFORMANCE 79
5.5.1 Modelling the closed loop system

Making a generalized plant
To propagate the noise models towards the performance output, a closed loop model with
inputs there where the noise sources apply and with the performance variable as output is
required, the closed loop system will have 11 inputs, w¯1 ...w¯11 and one single output z1 , as is
depicted in Figure 5.14. The variable names correspond to Figure 5.9, over-bar denote that
the variable is weighted by a filter V.. (s) such that the signal contains the PSD of the noise
source. A convenient way to make the closed loop model is by defining a generalized plant
P , using the Matlab command sysic and to close the loop afterwards with the Matlab com-
mand starp. In this case the loop is closed with a two-by-two controller, with just the PID
controllers on the diagonal. The result is the closed loop model H shown in Figure 5.15,
which defines 11 transfer functions, one from every noise input towards the performance
output. All these transfers are plotted in Figures 5.15-5.16. In the previous section only
nine noise sources were actually modelled, corresponding to transfer functions 1-5 and 7-10.
In the propagation analysis in the next section, only these transfers are used.
Figure 5.14 The open loop system, defined in a, so-called, ’generalized plant’ setting. The open
loop system is denoted by P .
5.5.2 Noise propagation

Now the closed loop transfer functions are available, the noise sources can be propagated
towards the performance output. This is done for every noise source separately, using the
equation
vibrations w1_z1 [m] payl sensor w2_z1*cp [V]

60 40
20
40
Magnitude [dB]
Magnitude [dB]
0
20
−20
0
−40
−20 −60
−1 0 1 2 3 −1 0 1 2 3
10 10 10 10 10 10 10 10 10 10
payl set−point w3_z1 [V] payl controller w4_z1 [V]
20 10
0
0
−10
Magnitude [dB]
Magnitude [dB]
−20 −20
−40 −30
−40
−60
−50
−80 −60
−1 0 1 2 3 −1 0 1 2 3
10 10 10 10 10 10 10 10 10 10
Figure 5.15 above: The closed loop system H. below: Closed loop transfer functions from
inputs 1-4 to the performance output [m/s2 ]. (1) Ground vibrations [m], (2) Payload sensor
noise [V ], (3) Payload set-point noise [V ], (4) Payload PID controller noise [V ].
uma w5_z1 [A] payl disturbance w6_z1 [N]

−20 −40
−30 −50
Magnitude [dB]
Magnitude [dB]
−40 −60
−50 −70
−60 −80
−70 −90
−80 −100
−1 0 1 2 3 −1 0 1 2 3
10 10 10 10 10 10 10 10 10 10
ref sensor w7_z1*cs [V] ref set−point w8_z1 [V]
−30 −10
−40 −20
Magnitude [dB]
Magnitude [dB]
−30
−50
−40
−60
−50
−70 −60
−80 −70
−1 0 1 2 3 −1 0 1 2 3
10 10 10 10 10 10 10 10 10 10
sens controller w9_z1 [V] umi w10_z1 [A]
−10 30
−20 20
Magnitude [dB]
Magnitude [dB]
−30 10
−40 0
−50 −10
−60 −20
−1 0 1 2 3 −1 0 1 2 3
10 10 10 10 10 10 10 10 10 10
ref disturbance w11_z1 [N]
30
20
Magnitude [dB]
10
−10
−20
−30
−1 0 1 2 3
10 10 10 10 10
Frequency [Hz]
Figure 5.16 Closed loop transfer functions from inputs 5-11 to the performance output [m/s2 ].
(5) Payload servo amplifier noise [V ], (6) Payload disturbance forces [N ], this transfer function
is generally known as the compliance of the vibration isolation system, (7) Reference sensor noise
[V ], (8) Reference set-point noise [V ], (9) Reference PID controller noise [V ], (10) Reference
servo amplifier noise [V ], (11) Reference disturbance forces [N ].
ASD performance, σ = 0.051 mm/s2, σ = 1.27 µm

−4
10
−6
10
−8
10
−10
10
[m/s2/Hz1/2]
−12
10
−14 sens_p
10 set_p
pid_p
−16
10 uma_p
sens_r
set_r
−18
10 pid_r
umi_r
−20
total
10 −2 −1 0 1 2 3 4
10 10 10 10 10 10 10
Frequency [Hz]
Figure 5.17 Amplitude Spectral Density of the propagated noise sources separately and the ASD
of the total signal.
Sout,i (ω) = |Hi (jω)|2 Sin,i (ω) (5.49)

2 /Hz]
where Sin,i is the input PSD and Sout,i the resulting output PSD, both with units [Vrms
and Hi (jω) is the transfer function towards the performance output, corresponding to the
noise source. In order to get the total output PSD, that is, the PSD of the output signal
with all the noise sources acting on the system, all the separate output PSD’s are simply
added.
Sout,tot (ω) = Sout,1 (ω) + · · · + Sout,4 (ω) + Sout,7 (ω) + · · · + Sout,9 (ω) (5.50)
Performance analysis is commonly done using the resulting Amplitude Spectral Density
(ASD) and the Cumulative Amplitude Spectra (CAS) instead of the PSD and CPS (see
§ 2.4.4). In Figure 5.17 the separate (ASDs) and the total ASD can be seen. The CAS can
be seen in Figure 5.18 and in Figure 5.19 (left), which both represent the same CAS but
with a different y-scale.
Analysis of the propagation results

Looking at the ASD in Figure 5.17, one can see that two noise sources are dominant; the
servo amplifier noise in the reference loop is dominant in the lower frequency region and the
sensor noise in the payload loop in the higher frequency region (>4 Hz). The servo amplifier
noise has a peak around 3-5 Hz, and its power contribution to the error will therefore be
CAS_log performance, σ = 0.051 mm/s2, σ = 1.27 µm

−4
10
−6
10
−8
10
[m/s2]
−10
10
sens_p
set_p
−12
pid_p
10 uma_p
sens_r
set_r
−14
pid_r
10 umi_r
total
−2 −1 0 1 2 3 4
10 10 10 10 10 10 10
Frequency [Hz]
Figure 5.18 Cumulative Amplitude Spectrum with a logarithmic y scale. Thanks to the loga-
rithmic y scale one can still see the magnitude of the contribution of the other noise sources.
2
−5CAS_lin performance, σ = 0.051 mm/s , σ = 1.27 µm
x 10 CAS_lin position − drift payload, sigma = 1.27 µm
6
sens_p 1.4
set_p
pid_p
5 uma_p 1.2
sens_r
set_r
pid_r 1
4
umi_r
total
0.8
[m/s2]
[µm]
3
0.6
2
0.4
1
0.2
Position
0 −2 0 2 4
0 −2 0 2 4
10 10 10 10 10 10 10 10
Figure 5.19 left: Cumulative Amplitude Spectrum with a linear y scale. The separate and total
CAS is plotted. The payload sensor contributes about 90% of the energy of the error signal.
right: Cumulative Amplitude Spectrum with a linear y scale of the position error. The error is
build op in the low frequency region (<0.1 Hz) and is due to the reference servo amplifier noise.
mainly in that region. The payload sensor noise peaks at about the bandwidth of the payload
loop, at 30 Hz.
In the logarithmic CAS in Figure 5.18 one can easily see that the payload sensor noise
and the umi servo amplifier noise are indeed thé main contributors to the power of the
performance signal, and in the linear CAS in Figure 5.19 (left) one can see that the payload
sensor contributes about 90% of the power. The total power has a value of 0.051 mm/s2
(1-σ).
Since the capacitive sensors are still in development, the noise model of the capacitive sensor
is a preliminary noise model. Now that it turns out that the sensor is thé main contributor
to the error, the designer knows that the development of this sensor is crucial to the success
of the system. The knowledge that the sensor contributes most of the power at the perfor-
mance output in the high frequency region, can be of help to the sensor designer.
In the logarithmic CAS it is visible that the second largest error contributor is the set-point
noise of the payload loop, which adds a 1-σ value of about 0.002 mm/s2 (relative to a 0.051
mm/s2 from the payload sensor). So, although this noise source is not dominant in some
frequency region, it adds more power than the dominant servo amplifier noise.
Due to the accelerations observed at the performance output, the payload also has a po-
sition disturbance.
√ The ASD of this position error is found when the ASD expressed in
(mm/s )/ Hz is multiplied with 1/s2 (equalling double integration with respect to time).
2
The final value of the corresponding CAS, shown in Figure 5.19 on the right hand side,
gives the 1-σ value of the position disturbance. In this figure one can see that the error is
build up in the low frequency region (<0.1 Hz). From the acceleration ASD of the error in
Figure 5.17, it is visible that the error in this region is due to the reference servo amplifier
noise, which caused a payload to drift 1.27 µm (1-σ).
Case: payload loop with 100 Hz open loop bandwidth

An interesting question is, how does the noise picture change when the payload loop band-
width is increased, e.g., to 100 Hz. A motivation to increase the payload loop bandwidth is
that the vibration isolation performance is then increased. This can be seen in Figure 5.20,
where the 100 Hz transmissibility is shown (the 30 Hz transmissibility is also shown for
reference). When one compares the 100 Hz transmissibility with the 30 Hz case, shown in
Figure 5.8 on page 69, it is visible that the 100 hz case isolates up to 70 Hz, instead of 20
Hz in the 30 Hz case.
Another motivation to increase the bandwidth, is that payload disturbance forces are then
suppressed up to 100 Hz instead of 30 Hz. One can see this when one looks at the transfer
function from payload disturbance forces w¯6 to performance output z1 , or the compliance
of the AIMS. The compliance for the 30 Hz case can be seen in Figure 5.16 above right,
which starts to level at about 30 Hz. This will become 100 Hz in this case, achieving payload
disturbance suppression up to this frequency. So, by increasing the open loop bandwidth of
the payload loop, both the transmissibility and the compliance of the AIMS are increased!
This improvement comes with a cost though, in the form of an increase in noise. The ques-
tion now is, does the increase in performance counter balance the increase in noise? This
Open and closed loop (30 Hz, 100 Hz) transmissibility

50
Magnitude [dB]
0
−50 open loop

30 Hz transmissibility
100 Hz transmissibility
−100 −1 0 1 2
10 10 10 10
180
0
Degree [deg]
−180
−360
−540
−720 −1 0 1 2
10 10 10 10
Frequency [Hz]
Figure 5.20 The open and closed loop transfer function (transmissibility) of the AIMS with a
100 Hz open-loop payload loop bandwidth.
question will be answered below.
The loop transfer function Lp (s) is altered, by changing the PID controller Ppid,p , such that
the open loop bandwidth of the payload loop is 100 Hz. The new coefficients of the controller
are not given here. Since the components in the PID controller are different from the 30 Hz
case, its noise contribution changes as well. The model was altered accordingly.
With the new loop gain function and noise model in place, the output error spectrum is
calculated again. The propagation results can be seen in Figures 5.21-5.22.
Comparing Figure 5.17 with Figure 5.21, one can see that the overall noise picture is the
same, except for that the total ASD being a bit higher. The reference servo amplifier noise
and the payload PID controller noise are still the main contributors to the error spectrum,
but their contribution is increased. So, by increasing the open loop bandwidth of the pay-
load loop, the payload force disturbance suppression increases, but the total error due to
the AIMS components increases as well.
Comparing the logarithmic CAS, one can see that the high frequency contribution of the
payload set-point gain and the payload PID controller is increased, compared to the refer-
ence servo amplifier noise. This is also visible in the linear CAS, where the total error power
even increases in the high frequency region due to these two noise sources.
The 1-σ value of the total error power now is 0.30 mm/s2 , compared to a value of 0.051
ASD performance, σ = 0.3 mm/s2, σ = 1.28 µm

−4
10
−6
10
−8
10
−10
10
[m/s2/Hz1/2]
−12
10
−14 sens_p
10 set_p
pid_p
10
−16 uma_p
sens_r
set_r
−18
10 pid_r
umi_r
−20
total
10 −2 −1 0 1 2 3 4
10 10 10 10 10 10 10
Frequency [Hz]
Figure 5.21 Amplitude Spectral Density of the propagated noise sources separately and the ASD
of the total signal in the 100 Hz case.
−4 CAS_lin performance, σ = 0.3 mm/s2, σ = 1.28 µm

CAS_log performance, σ = 0.3 mm/s2, σ = 1.28 µm x 10
−2
10
3.5
sens_p
set_p
3 pid_p
−4
10 uma_p
sens_r
2.5 set_r
−6
10 pid_r
umi_r
total
2
[m/s2]
−8
10
[m/s2]
1.5
−10 sens_p
10
set_p
pid_p 1
−12
uma_p
10 sens_r
set_r
0.5
pid_r
−14
10 umi_r
total
0 −2 0 2 4
−2 −1 0 1 2 3 4
10 10 10 10 10 10 10 10 10 10 10
Figure 5.22 left: Cumulative Amplitude Spectrum with a logarithmic y scale in the 100 Hz case.
left: Cumulative Amplitude Spectrum with a linear y scale in the 100 Hz case.
mm/s2 in the 30 Hz case. Changing the system to a 100 Hz configuration, apparently de-
creases the performance with a factor of about 6. So, increasing the bandwidth to 100 Hz
comes with rather large costs, and must be traded off with the gain in transmissibility and
compliance. Since the performance specification of 0.1 mm/s2 is exceeded with a factor
three (0.30 mm/s2 ), it is obvious that the increase in transmissibility alone does not counter
balance the gain in noise. An appropriate trade off can only be made, when the ASD of the
payload disturbance forces is known and brought into the analysis as a disturbance source.
The 1-σ value of the total position disturbance is increased from 1.27 µm to 1.28 µm, which
is a negligible increment. This is because the noise due to the reference servo amplifier
noise (which causes the drift) did not increase when the bandwidth was changed, as can be
concluded from Figure 5.17 and 5.21.
5.5.3 Vibration isolation performance of the AIMS

How well does the AIMS√perform concerning vibration isolation? The ground vibration
model V (jω)f loor [m/s2 / Hz], developed in Section 5.4.1 is used in this analysis to inves-
tigate the vibration isolation performance of the AIMS.
To investigate the vibration isolation performance, one can look at the open loop and closed
loop (30 Hz case) ground vibration propagation
√ transfer function. The ground vibration
model (5.17) is defined in units [m/s2 / Hz], so the corresponding propagation transfer
function is the transfer from acceleration of the ground: x¨h towards payload acceleration:
x¨p . This transfer function is the same as the one from ground position xh to payload position
xp . The open and closed loop transfer function are already derived in Section 5.3.3 and can
be seen in Figure 5.8 on page 69.
Using the open and closed loop transfer function, the floor noise spectrum is propagated
towards the performance output. The performance of the AIMS can now be analyzed by
comparing the open loop ASD (using only the basic isolator to isolate from vibrations) with
the closed loop ASD due to vibrations combined with the closed loop ASD due to its noise
sources, which was already calculated in Section 5.5.2. In the context of performance, the
ASD due to the noise sources can be seen as the costs that have to be paid to use the AIMS
in the closed loop setting.
Besides the performance of the open loop AVI using a basic (cheap) 2 Hz isolator, the (open
loop) performance of a more expensive passive 0.5 Hz isolator (without using the AIMS) is
also calculated. A 0.5 Hz isolator with 20% damping is the type of (passive)isolator which
is used nowadays in industry e.g., an airmount. The performance level of the expensive 0.5
Hz isolator is the level that should be achieved by the cheap 2 Hz isolator in combination
with the AIMS. The results can be seen in Figure 5.23 and 5.24.
In this figure it can be seen that the open loop 2 Hz isolator ASD is higher than the total
closed loop ASD in the mid-frequency range. In the low and high frequency regions, the
noise sources due to the AIMS dominate the closed loop spectrum, making the closed loop
performance worse than the open loop performance. In the mid-frequency range, where
the error due to ground vibrations dominate the total closed loop spectrum, the isolation
2
ASD performance, σAIMS = 0.071 mm/s
−2
10
−4
10
−6
]
10
1/2
[m/s /Hz
2
−8
10
no isolation
−10 open loop 2 Hz isolator
10 closed loop due vibrations
closed loop due AIMS
total closed loop
−12
open loop 0.5 Hz isolator
10 −2 0 2 4
10 10 10 10
Frequency [Hz]
Figure 5.23 ASD of the open loop and closed loop AIMS vibration isolation performance, together
with the ASD of the ground itself (no isolation). The closed loop PSD is the ASD due to ground
vibrations added with the PSD due to the noise sources of the AIMS.
2
CAS_log performance, σAIMS = 0.071 mm/s
0
10
−2
10
−4
10
[m/s ]
2
−6
10
−8
10
−10 no isolation
10
open loop 2 Hz isolator
total closed loop
−12
0.5 Hz isolator
10 −2 0 2 4
10 10 10 10
Frequency [Hz]
Figure 5.24 CAS of the open and closed loop (30 Hz) AIMS, together with the CAS of the
ground vibration itself (no isolation).
5.6. CONCLUSIONS AND RECOMMENDATIONS 89
performance of the AIMS proves its value, since it takes a huge ’bite’ out of the open loop
ASD.
To analyze the performance of the AIMS one has to look at the resulting power (1-σ) of
the acceleration signal, and compare the no-isolation, open loop 2 Hz isolator, total closed
loop and the 0.5 Hz isolator case. In Figure 5.24 four different CAS can be seen. The final
value of the CAS equals the 1-σ value of the signal. The CAS of the ground vibrations (no
isolation) gradually rises to 10 cm/s2 at 10 kHz. The open loop CAS shows a steep rise
in power around 2 Hz, where the 2 Hz isolator has its resonance frequency, and levels at a
value of 0.94 mm/s2 . Because of the mass-spring characteristics of the 2 Hz isolator, the
open loop CAS does not show any increase in power at frequencies higher than 2 Hz. The
total closed loop CAS starts at a higher level in the lower frequency region, because of the
power contributed by the electrical components of the AIMS in this region (see Figure 5.24).
Around 1 Hz the level dives below the open loop level, because of the vibration-isolation
performance of the AIMS. After 10 Hz it still rises a little, due to the payload sensor noise,
but finally levels to 0.071 mm/s2 around 100 Hz. The 0.5 Hz isolator still achieves the
highest performance level: 0.021 mm/s2 . Comparing the CAS’s, one can see that the main
difference is made in the 10 - 30 Hz region, where the closed loop CAS shows an increase in
power rises due to the payload sensor noise of the AIMS.
Summarizing; it is clear that the closed loop case performs significantly better than the open
loop 2 Hz isolator; 0.071 mm/s2 in the closed loop case compared to 0.94 mm/s2 in the open
loop case, which is about 13 times better. On the other hand, the more expensive 0.5 Hz
isolator still achieves the highest performance level of 0.021 mm/s2 , 3.4 times better than
the closed loop case!
5.6 Conclusions and recommendations

5.6.1 Conclusions
In this chapter it is shown that the DEB approach can be applied to a system that is still
in the design phase. The DEB analysis shows that the AIMS concept is feasible in practice
and points out the payload sensor noise as the main contributor of power to error signal.
The simulations show that the 1-σ performance specification of 0.1 mm/s2 is met when the
payload loop-gain function has an open loop bandwidth of 30 Hz. In this case the power of
the error due to the noise sources of the AIMS contribute 0.051 mm/s2 to the total budget.
When the ground vibrations themselves are added to the analysis, the final performance
level increased to 0.071 mm/s2 . The closed loop AIMS system (parallel with the passive
2 Hz isolator) does not achieve a higher performance level than the more expensive 0.5 Hz
passive isolator, which achieves an open loop performance level of 0.021 mm/s2 .
When using a 100 Hz payload-loop-bandwidth, the AIMS will not achieve the performance
specification, since the electrical components of the AIMS themselves already contribute
0.30 mm/s2 , which is three times the performance specification of 0.1 mm/s2 . Almost all
of this error is due to the payload sensor. So, the 30 Hz case is the configuration to choose,
unless a better payload sensor is developed.
It has to be said that it is likely that other un-modelled disturbances, like air pressure
fluctuations and disturbance (friction) forces, will also add a significant part to the error
budget. So, in order to achieve the spec of 0.1 mm/s2 in practice, the simulated value has
to be significantly smaller than the specification, which is fortunately the case in the 30 Hz
configuration (0.051 mm/s2 ).
5.6.2 Recommendations
The capacitive payload sensor turns out to be the main contributor of power to the perfor-
mance output, and most of this power is contributed in the high frequency region (>100 Hz).
Since the sensor is still in development, the designer should design the sensor with this knowl-
edge in mind, such that the final system meets the performance specification.
The H2 control strategy can be used in this case study to find out if the control strategy
used in the case study is close to optimal. The two PID controllers used in the setup now are
designed with a certain closed loop behavior in mind. The reference controller is designed
to achieve a closed loop 0.5 Hz suspension behavior of the reference mass. The payload
controller is designed such that the closed payload loop can suppress reference disturbances
up to 30 Hz. It could well be that this control strategy is far from the optimal strategy, which
is given by the H2 control strategy. The H2 control strategy could be used to synthesize
e.g., only the payload loop controller and consider the reference loop controller as fixed
and vice versa. With this approach two SISO controllers can be found. On can also go
one step further, by considering both controllers as a degree of freedom and synthesize a
MIMO H2 controller. The extra control degree of freedom given by the of diagonal terms of
the MIMO controller might offer and opportunity to increase performance even more. One
disadvantage of this approach is that the MIMO controller might be hard to interpret.
Chapter 6
Conclusions and recommendations
6.1 Conclusions on the DEB design approach

In this research the Dynamic Error Budgeting (DEB) design approach is developed. With
the tools offered by the DEB design approach, the designer is able to account for stochastic
disturbances during the design of a mechatronic system. The DEB approach requires distur-
bance models, a model of the system and a controller as input and then offers the designer
tools to predict the performance of the system a priori and gain insight in the performance
limiting factors of the system. With this insight, the designer is able to point out critical
system components/properties and make design decisions that improve the performance of
the system.
H2 control strategy is presented as an extension to the DEB approach. It offers the designer
the opportunity to optimize over the degree of freedom given by the controller, enabling the
designer to predict the maximum achievable performance level of a system concept. Using
this technique, the designer is able to objectively compare the performance potential of
different system concepts.
When different system components are designed at different locations, the DEB design ap-
proach reaches the system designer a tool to keep the overview of the total system and to
keep track of the final performance of the system, taking in account all disturbances. This
helps the designer to design complex systems ’first time right’, significantly reducing design
costs and time.
By applying the DEB approach in two (one DOF) vibration isolation case studies, its feasibil-
ity in practice is shown. In both studies, the disturbances are modelled and the performance
of the system predicted. In the first case study the gap between theory and practice is an-
alyzed and the discrepancies between theoretical and experimental results explained. Both
case studies show that, using the DEB analysis tool, the designer is able to point out critical
components and to make design decisions that improve the performance of the system.
In the first case study the critical system component turned out to be the AD converter. The
design of the system was improved by introducing a preamplifier just before the converter,
91
92 CHAPTER 6. CONCLUSIONS AND RECOMMENDATIONS
increasing the performance level with a factor of 30. Then another improvement was made
by applying an H2 control strategy to control the system, which increased the performance
even further with a factor of about 4. The theoretical performance improvements were val-
idated with experiments. It is also shown that, with the use of Pareto curves, the designer
is able to objectively compare different system designs in practice.
In the second case study the DEB approach is applied to the design of a system in practice.
Using the tools available it is shown that the performance of the concept design can be
predicted and critical disturbances/components can be pointed out a priori, showing that
the approach is feasible in practice. It is shown that the design concept is feasible and that
the critical system component is the capacitive payload sensor. The effect of changing the
bandwidth of the system on performance was analyzed by comparing the performance using
two different bandwidths, showing a big difference in final performance.
6.2 Recommendations and further work

An interesting and very important question in the DEB design approach is what performance
variable to choose? The analysis would have been different when in the first case study, in-
stead of the load velocity, e.g., the load acceleration level was chosen as the performance
measure. This choice greatly depends on the application of the system. An interesting point
for further research is to investigate what the consequences are for the design of the system
when a different performance variable is chosen and how this is related to the performance
objective of the system in practice.
The accuracy of the predicted performance by DEB with respect to the measured results,
can be improved by using higher order models of the disturbances. Increasing the order of
the disturbance model might even allow (approximate) modelling of harmonic disturbances
by using inverse notches. How accurate this can be done is an interesting item for future
research. To achieve the highest degree of prediction accuracy is it recommended to use the
actual measured disturbance spectra in the simulations.
If there is a discrepancy between the theoretical and measured performance PSD, a way to
deal with it is to accredit the difference to an additional artificial disturbance, such that the
difference in output power is contributed by this input disturbance. In this way the designer
has a possibility to account for this difference in power in the simulations. An interesting
question that remains is where to put the input of the additional disturbance.
An explanation for such a discrepancy could be that the input disturbances are not un-
correlated and that the disturbance propagation relation should also contain cross power
spectra such the correlation is accounted for. Another way to deal with correlated distur-
bances, is to (re)model the disturbances such that they originate from the a new single input.
In this way the need to use cross power spectra in the propagation relation is circumvented.
In this research the DEB approach is applied to a one DOF (SISO) system. An interesting
follow-up of this research is use the approach with the design of a MIMO setup e.g., a 2
6.2. RECOMMENDATIONS AND FURTHER WORK 93
DOF system.
When an H2 controller is synthesized for a particular system, it can give the control designer
useful hints about how to control the system best for optimal performance. Drawbacks how-
ever are, that no robustness guarantees can be given and that the order of the H2 controller
will generally be too high for implementation. Research can be done on how to use the
information given by the H2 controller to the designers advantage in practice.
Further work can also be done on developing a DEB design strategy. In this thesis it is
discussed that the DEB analysis tools enable the designer to make performance improving
design decisions. The decisions must be made on the basis of the analysis and designer
instinct. It is interesting to investigate if a strategy can be developed that can be followed
to improve the system performance. A suggestion for this strategy is the balancing approach
as discussed in Section 4.4. In this approach the disturbances introduced by a noise source
and e.g., an extra gain, which is supposed to suppress the negative effect of the disturbance
on performance, must be balanced at the performance output.
94 CHAPTER 6. CONCLUSIONS AND RECOMMENDATIONS
Appendix A
Mathematics
A.1 MATLAB code for computing the PSD and CPS

Consider the discrete time signal xL [n], containing L samples in variable datavec, using
sample time Ts seconds. If the data set is too big, you might want to re-sample the time
signal using the MATLAB° r
command decimate, before applying the code below. Mind, that
the original continuous time signal is not allowed to have power content above the Nyquist
frequency 1/(2Ts ), to avoid anti-aliasing effects.
T_r = L*T_s; % signal time range
d_f = 1/T_r; % width of frequency grid
F_s = 1/T_s; % sample frequency
F_n = F_s/2; % Nyquist frequency
F = [0:d_f:F_n]; % one sided frequency grid
% Discrete Time Fourier Transform Wxx

Wxx = fft(datavec - mean(datavec))/L;
% Two-sided Power Spectrum Pxx [SI^2]

Pxx = Wxx.*conj(Wxx);
% Two-sided Power Spectral Density Sxx_t [SI^2/Hz]

Sxx_t = Pxx/d_f;
% One-sided Power Spectral Density Sxx_o [SI^2/Hz] defined on F

Sxx_o = 2*Sxx_t;
% Two-sided Cumulative Power Spectrum CPxx [SI^2]

CPxx = cumsum(Pxx);
% Variance of signal from time data and frequency data

var_time = var(datavec); var_psd = 2*CPxx(floor(L/2));
95
96 APPENDIX A. MATHEMATICS
A.2 Standard assumptions in the H2 framework

The material covered in this appendix is discussed in more detail in a feasibility study of
the dynamic error budgeting approach by Monkhorst [12].
For H2 controller synthesis using the MATLAB° r
command hinfsyn, the system needs to
be defined in a generalized plant setting, which has to meet the standard assumptions. The
open loop system G in Figure A.1 is shown in a generalized plant setting, having n stacked
disturbance inputs w and k stacked performance outputs z. The controller input y is defined
as output and the controller output u as input of G. The input weighting filters Vw (j 2πf ),
used to model the disturbances can also be seen.
Figure A.1 The open loop system Ḡ in series with the diagonal input weighting filter Vw and
diagonal output scaling filter Wz defining the generalized plant G.
Assume that G(s) can be written as:

· ¸ · ¸· ¸
z̄ G11 (s) G12 (s) w̄
= (A.1)
y G21 (s) G22 (s) u
| {z }
G(s)
and each subsystem Gji (s) is defined by the state space matrices A, Bi , Cj and Dji , as
depicted in Figure A.1.
w u
A B1 B2
z C1 D11 D12
y C2 D21 D22
Table A.1 State space matrices and signals of the open loop generalized plant
For the controller design techniques described by Doyle (et.al.) [6] to be applicable, the
following technical assumptions on the generalized plant G have to be met, see [5, 4, 27].
1. D11 must be zero.
2. (A, B2 ) is stabilizable.
3. (A, C2 ) is detectable.
4. D12 has full column rank.

A.2. STANDARD ASSUMPTIONS IN THE H2 FRAMEWORK 97
5. D21 has full row rank.

· ¸
A − jωI B2
6. has full column rank at all frequencies ω.
C1 D12
· ¸
A − jωI B1
7. has full row rank at all frequencies ω.
C2 D21
Assumption 1 implies that there is no direct feedthrough from disturbance signals w to

performance outputs z. Assumptions 2 and 3 imply that the generalized plant G must be
detectable and stabilizable by the controller. Assumption 4 means that a part of every
entry in u is present in the performance variable z, such that the control effort u is always
penalized. Assumption 5 means that every measured output y must always be contaminated
with a disturbance.
Finally, assumptions 7 and 6 imply that G12 (s) and G21 (s) may not have transmission zeros.
98 APPENDIX A. MATHEMATICS
Appendix B
Noise and system modelling
B.1 State space model of the vibration isolation setup

In Trumper [24] a state space model is derived to model a two degree of freedom vibration
isolation setup. This model is here derived again for the setup used in the first case study
in this thesis.
The speaker, together with the rigidly mounted geophone can be modelled as a fourth order,
two degree of freedom system, with a force fs acting as the driving actuator force as shown
in Figure B.1.
Figure B.1 Model of the speaker with the mounted geophone. The system has two degrees of
freedom, defined as w(t) and W (t).
The parameters ml , kl , bl and Nf represent the weight of the cone structure (including the
geophone casing, cone plate etc.), the speaker suspension-stiffness, the speaker suspension
damping-coefficient and the speaker force constant resp. Note the extra damping term
Nf2 /Rs to account for the back emf in the voice coil. The geophone proof mass mg , leaf
99
100 APPENDIX B. NOISE AND SYSTEM MODELLING
spring stiffness kg and suspension damping bg are derived in § 3.2.3.

To derive an expression for the force fs , let us look at the electrical relations of the voice
coil. When a current Is is send through the voice coil a force Fs is created. This force is
given by Fs = Nf Is , where Nf [N/A] is called the speaker force constant. This constant is
characteristic for a particular speaker type and also varies with the axial position of the voice
coil relative to the magnets. For the sake of simplicity this value is now assumed constant.
More details on calculating Nf can be found in [24]. The electrical relationship for the voice
coil is:
Esp = Rs Is + Nf (Ẋl − Ẋb ) (B.1)

| {z }
back emf
where Esp is the voltage applied at the input terminals of the voice coil and Rs is the voice
coil impedance. The second term on the right is to account for the back emf generated when
the voice coil moves relatively to the magnets. Rewriting for Is and substituting Is in to
Fs = Nf Is gives us for Fs
Esp Nf Nf2 (Ẋl − Ẋb )

Fs = − (B.2)
R Rs
| {zs } | {z }
fs additional damping
Note that the second term on the right-hand side acts as an additional damping term on the
speaker motion and that the term fs is the actual driving force applied to the voice coil, as
shown in Figure B.1.
Defining the relative positions as the states of the system:
w(t) = Xl (t) − Xb (t) (B.3)

W (t) = Xg (t) − Xl (t) (B.4)
(B.5)
and defining input u = [Esp Ẍb Fd ]t and output y = [Eg Ẍl ]t , the state space model
can be written as
ẋ = Ax + Bu
(B.6)
y = Cx + Du
with
B.1. STATE SPACE MODEL OF THE VIBRATION ISOLATION SETUP 101
 
0 0 1 0
 0 0 0 1 
 
A =  kl
 −m kg bl +N 2 /Rs
− mfl
bg 

 l m
³ l ´ m
³ l ´ 
kl bl +Nf2 /Rs
ml −kg m1l + m1g ml −bg m1l + m1g
 
0 0 0
 0 0 0 
 
B =  Nf 1 
 Rs ml −1 ml 
Nf
− Rs m l
0 − m1l
" # (B.7)
0 0 0 Gg
C = kl kg bl +Nf2 /Rs bg
−m l ml − ml ml
" #
0 0 0
D = Nf 1
Rs ml 0 ml
£ ¤t
x = w W ẇ Ẇ
£ ¤t
u = Ein Ẍb Fd
Half of the parameters in the state space model are still unknown, except for the geophone
parameters mg , bg and kg , which are derived in § 3.2.3, and the voice coil impedance Rs = 4Ω
given by the manufacturer. To identify the remaining parameters ml , bl , kl and Nf , the state
space mode was fitted by hand on the the empirical model Pspeaker obtained in § 3.2.3.
The resulting parameters values are given in Table B.1. The following parameters are ob-
tained from fitting the model to the swept sine data: kl , bl , ml and Nf . A remarkable
result is the zero damping coefficient of the speaker, which can be interpreted as follows; the
damping in the speaker system is mainly due to the back emf generated in the voice coil.
The value of bl is probably not zero in reality, but can be assumed zero for the modelling
purposes in this case study.
Stiffness coefficient kl : 4000 N/m

Damping coefficient bl : 0 N/m/s
Mass coefficient ml : 485 grams
Speaker force constant Nf : 11 N/A
Voice coil impedance Rs : 4Ω
Table B.1 Speaker parameter values

B.2 Peterson New High/Low Noise Model

The material in this appendix is copied out Peterson [15] and is repeated here for convenience
in case of later use. The New High Noise Model is a spectrum of average high seismic
background noise at seismometer stations across the globe and can be viewed as an upper
bound of seismic background noise. The Low High Noise Model represents the average low
seismic background noise at the quietest sites across the globe.
The New High Noise Model (NHNM) can be expressed in acceleration, velocity or position
density as follows:
NHNMacc = A + B log(P ) [10 log((m/sec2 )2 /Hz)] (B.8)

2
NHNMvel = NHNMacc + 20.0 log(P/(2π)) [10 log((m/sec) /Hz)] (B.9)
2 2 2
NHNMpos = NHNMacc + 20.0 log(P /(4π )) [10 log(m /Hz)] (B.10)
For the NLNM identical equations hold. Line parameters are given in Table B.2.
Pnhnm Anhnm Bnhnm - Pnlnm Anlnm Bnlnm

0.10 -108.73 -17.23 0.10 -162.36 5.64
0.10 -108.73 -17.23 0.17 -166.70 0
0.22 -150.34 -80.50 0.40 -170.00 -8.30
0.32 -122.31 -23.87 0.80 -166.40 28.90
0.80 -116.85 32.51 1.24 -168.60 52.48
3.80 -108.48 18.08 2.40 -159.98 29.81
4.60 -74.66 -32.95 4.30 -141.10 0
6.30 0.66 -127.18 5.00 -71.36 -99.77
7.90 -93.37 -22.42 6.00 -97.26 -66.49
15.40 73.54 -162.98 10.00 -132.18 -31.57
20.00 -151.52 10.01 12.00 -205.27 36.16
354.80 -206.66 31.33 15.60 -37.65 -104.33
21.90 -114.37 -47.10
31.60 -160.58 -16.28
45.00 -187.50 0
70.00 -216.47 15.70
101.00 -185.00 0
154.00 -168.34 -7.61
328.00 -217.43 11.90
600.00 -258.28 26.60
10000 -346.88 48.75
Table B.2 Line parameters for constructing the NHNM and NLNM curve given the period (P)
or frequency (1/P).
B.3. BUILDING THE PREAMPLIFIER 103
B.3 Building the preamplifier

B.3.1 Preamplifier transfer function
The output voltage signal of the preamplifier is denoted by Eout . Its input signal is the
geophone output Eg . The transfer function of the preamplifier can be written as a series
connection of two first order lag filters:
µ ¶µ ¶ µ ¶
Eout τp s + 1 τp s + 1 τp s + 1 2
= Gpa1 Gpa2 = Gpa (B.11)
Eg α p τp s + 1 αp τp s + 1 αp τp s + 1
| {z }| {z }
f irst stage second stage
with τp = 1/ω1 = 1/(2π 0.45) and αp = 10, which puts the corner frequencies in the right
place. With the gains chosen as Gpa1 = 100 and Gpa2 = 10 (Gpa = 1000), the total transfer
equals the wanted preamplifier transfer shown in Figure 3.15. The two first order parts of
the preamplifier are referred to as the first and the second stage.
As one might have noticed, the preamplifier transfer function resembles exactly a part of
the controller ( 3.5 on page 26). So, by using this preamplifier, a part of the controller is
actually shifted up the control loop to improve the SNR.
B.3.2 Preamplifier circuit design

The wanted preamplifier must have a second order transfer function (B.11), which can be
written as a series connection of two first order stages, the first stage and second stage. A
preamplifier stage with a second order transfer function can be physically achieved with the
circuit configuration as shown in Figure B.2.
Figure B.2 A two stage preamplifier circuit configuration. The geophone output signal Eg is fed
in to the non inverting input of the first opamp. The output of the first stage Eout1 is then fed
into a the second stage. The output of the second stage Eout is the final preamplifier output.
R11 , R21 , R31 , R12 , R22 , R32 , represent resistors and C1 , C2 capacitors. Their values determine
the transfer function of the preamplifier. The variables Zf 1 (s) and Zf 2 (s) represent impedances
defined by the dashed box enclosures. The triangles represent the signal common.
The geophone output Eg is connected to the non-inverting input of the operational amplifier
(opamp) of the first stage, which has the advantage over the inverting input concerning noise
levels, as is discussed in Rodgers [18] and [19]. Since the transfer function of a single stage
(B.11) has 3 DOF’s (e.g.,, Gpa1 , α1 , and τ1 ) and a preamp stage has 4 DOF’s, there is one
DOF left, which gives a degree of freedom in choosing the physical components.
B.3.3 Preamplifier transfer function in physical component values

The total transfer function of the preamplifier as shown in Figure B.2, will be denoted by
Zpa and can be written as ([10]):
Zpa (s) = Zpa1 (s)Zpa2 (s) = (1 + Zf 1 (s)/R11 ) (1 + Zf 2 (s)/R12 ) (B.12)
where Zpa1 (s) is the transfer function of the first stage and Zpa2 (s) of the second stage.
Transfer function of the first stage, Zpa1

The transfer function of the first stage is now written as:
Zpa1 (s) = 1 + Zf 1 (s)/R11 (B.13)
and must be expressed in terms of the resistors values and the capacitor value. The
impedance Zf 1 (s) is given by the parallel combination of R21 and the series connection
of C1 and R31 , or written out:
³ ´
1 R31 C1 s+1
Zf 1 (s) = C1 s + R31 ||R21 = C1 s ||R21 =
R21 (B.14)
R21 R31 C1 s+R21 R21 ||R31 C1 s+ R
21 +R31
= (R21 +R31 )C1 s+1 = C1 s+ R 1
21 +R31
Interpreting Zf 1 (s): at low frequencies (s ≈ 0) the capacitor is open (no connection) and
Zf 1 (s) equals R21 , at high frequencies (s → ∞) the capacitor is closed (short circuited) and
Zf 1 (s) equals the parallel combination of R21 and R31 : R21 ||R31 .
Notice that, when C1 and R31 are set to zero the impedance Zf 1 (s) just equals the resistor
value R21 .
Zpa1 (s) can now be written as

R11 +R21
(R11 + R21 ||R31 )C1 s + R21 +R31
Zpa1 (s) = 1 + Zf 1 /R11 = R11
(B.15)
R11 C1 s + R21 +R31
The low and high frequency gain of this Zpa1 (s) are given by resp.:
R11 + R21 R11 + R21 ||R31

Zpa1 (0) = , Zpa1 (∞) = (B.16)
R11 R11
and the two corner frequencies, in radians per second, are located at
B.4. A NOISE MODEL FOR THE PREAMPLIFIER 105
1 1
ω1 = , ω2 = (B.17)
(R21 + R31 )C1 (R31 + R21 ||R11 )C1
We can now equal the transfer function of Zpa1 (s) with the corresponding part of Equa-
tion B.11 which can result in the following equations (R11 is chosen as the degree of free-
dom):
R21 = R11 (Gpa1 − 1)

R31 = R11 R21 ((Gpa1 /αp − 1)/(R21 − R11 (Gpa1 /αp − 1))) (B.18)
C1 = 1/(ω1 (R21 + R31 ))
Transfer function of the second stage, Zpa2

The derivation of the transfer function Zpa2 (s) is completely similar to that of Zpa1 shown
above, and is not repeated here. The transfer function of Zpa2 can be written as:
R12 +R22
(R12 + R22 ||R32 )C2 s + R22 +R32
Zpa2 (s) = 1 + Zf 2 /R12 = R12
(B.19)
R12 C2 s + R22 +R32
B.3.4 Physical parameter values for the 1000-10 preamplifier

For the first stage the parameters are given by: Gpa1 = 100, αp = 10 and ω1 = 1/(2π 0.45) rad/sec
, as obtained earlier in this section. Setting R11 to 3.6 kΩ and using Equations B.18, we
obtain: R21 = 356400 Ω, R31 = 35640 Ω and C1 = 0.9 µF .
For the second stage the parameters are given by: Gpa2 = 10 and αp and ω1 . Plugging in
R12 = 43 kΩ, returns: R22 = 387 kΩ, R32 = 0Ω and C2 = 0.91 µF .
A preamplifier with these component values was build and used in the isolation system. A
photograph of the build preamp is shown in Figure B.3.
It is not a coincidence that the two capacitor values are about 0.9 µF . The output noise
level of the preamplifier increases with increasing resistor values, as will be shown in the next
Subsection. The bigger the capacitor value is chosen, the smaller the resistor values will be
(this is clear when you plot the dependency described by Equations B.18, not shown here).
Since the capacitor size increases rapidly with its capacitance, there is a practical limit to
the usable capacitance value (the size of the circuit board should be as small as possible in
order to avoid noise pick up). 0.9 µF was chosen in this case, and the values of R11 and R12
were chosen to meet this requirement.
B.4 A noise model for the preamplifier

B.4.1 Introduction preamplifier noise modelling
To develop a noise model for the preamplifier, all the noise models of the noise generating
components will be modelled first. After that, a total noise model of the preamplifier will
be derived.
Figure B.3 Photograph of the build preamplifier circuit.
The noise generated by the preamplifier originates in its resistors and operational amplifiers.
(The noise generated by capacitors is negligible, see Fish [7]). The thermal noise generated
by the resistors was already modelled in section 3.3.2 on page 29. The noise generated by
an operational amplifier is modelled in § B.4.2. In § B.4.3 a noise model for the preamplifier
is derived in terms of the physical component values. Finally, in § B.4.4 the actual physical
values found in § B.3 are filled in and the use of two different low noise opamps is compared.
B.4.2 Voltage and current noise of an operational amplifier

In this case study the use of two different operational amplifiers is analyzed; the op07 and the
op27, both low-noise operational amplifiers. The noise produced by an operational amplifier
can be modelled as a equivalent voltage noise source en in series with its inverting input,
an equivalent current noise source in− in parallel with the inverting input and a current
noise source in+ in parallel with the non inverting input. These noise sources can be seen in
Figure B.4.
The PSD’s of the voltage noise source and current noise sources are given by (see e.g.,
Riedesel [17, p1749])
µ ¶
2πfv
Sen (ω) = e2o 1 + 2
[Vrms /Hz] (B.20)
ω
µ ¶
2 2πfc
Sin ± (ω) = io 1 + [A2rms /Hz] (B.21)
ω
where e2o and i2o are the asymptotic high frequency voltage and current noise PSD levels,
respectively, and fv and fc are the voltage and current noise corner frequencies, respectively.
The noise PSD Sin of the two current sources are assumed to be equal, Sin = Sin − = Sin + .
Figure B.4 The equivalent noise sources of an opamp used to model the voltage noise, en , and
current noise, in− and in+ .
The ASDs of these two noise sources of an op07 can be seen in Figure B.5, where the values of
eo , io , fv and fc are obtained from manufacturers product specification (Texas Instruments).
The characteristic noise parameters of the two opamps are listed in Tables B.3 and B.4.
Asymptotic high frequency voltage level eo : 10.0 · 10−9 Vrms

Table B.3 Noise parameters of the op07
Asymptotic high frequency voltage level eo : 3 · 10−9 Vrms

Table B.4 Noise parameters of the op27
B.4.3 Noise modelling of the preamplifier

The noise model of the preamplifier developed here, is based on the models developed in
Rodgers [18], [19], Riedesel (et al.) [17] and Horowitz and Hill [10], and corresponds to the
configuration shown in Figure B.2 on page 103.
Noise model for the first preamplifier stage

In Figure B.6 the first stage can be seen again with all its noise sources included. (The
geophone-coil noise is left out, because it is already accounted for in the geophone-coil noise
Voltage noise ASD of op07 [nV/Hz1/2]

Current noise ASD of op07 [pA/Hz1/2]
0.4
40
30 0.3
[nV / Hz1/2]
[pA / Hz1/2]
20
0.2
10 0.14
6 −1 0 1 2
0.1 1 2 3
10 10 10 10 10 10 10
Figure B.5 left: Voltage noise model (ASD) for the op07 with corner frequency at 1.3 Hz and
an asymptotic high frequency white noise level of 10 nVrms . right: Current noise model (ASD)
of the op07, with corner frequency @ 50 Hz and an asymptotic high frequency white noise level
of 0.14 pArms .
model.)
Figure B.6 First stage of the preamplifier and all its noise sources
The current noises, in− and in+ , are shown as Norton generators from the inverting and non
inverting inputs to ground. The voltage noise en is shown as a Thevenin generator in series
with the summing junction at the intersection of R11 and Zf 1 . The thermal noise voltages
eR11 , eR21 , and eR31 are shown in series with the resistors (Rodgers [18, p1091]).
Voltage Noise, first stage
The noise PSD at the output of the first stage of the preamp due to the voltage noise of the
opamp en is denoted by So1,en . From the voltage divider relationship between en and Eout1 ,
So1,en can be obtained as (see Rodgers [18, p1092]):
¯ ¯2
¯ Zf 1 (jω) ¯
So1,en (ω) = ¯¯1 + ¯ Sen (ω) (B.22)
R11 ¯
Current Noise, first stage
The noise PSD at the stage output due to the current noise of the opamp in− is denoted
by So1,in− . Because the left side of R11 is at ground and the right side is at virtual ground,
there is no potential difference across R11 due to in− . Therefore, all of in− flows through
Zf 1 , and none of it flows through R11 . Therefore (see Rodgers [18, 92,p1092]):
So1,in− (ω) = |Zf 1 (jω)|2 Sin (ω) (B.23)
The noise at the stage output due to the current noise in+ is denoted by So1,in+ . in+ flows
directly through the coil resistance, Rc . (see Rodgers [18, p1094]) The part of in+ which
flows through the coil resistance Rc produces a force on the seismometer mass causing the
mass to move and generate a back emf. Therefore the current noise does not see a resistance
Rc but sees a source impedance Zs . The relations for the source impedance are derived by
Rodgers [18, p1096].
Ã !
G2g s
Zs = Rc + (B.24)
mg s + 2ζωo s + ωo2
2
Accounting for the fact that the current noise sees the impedance Zf 1 instead of only the
resistance Rc , it can be seen that So1,in+ is given by
¯ ¯
¯
2¯ Zf 1 (jω) ¯¯2
So1,in+ (ω) = |Zs (jω)| ¯1 + Si (ω) (B.25)
R11 ¯ n
Thermal noise, first stage
The noise PSD at the output due to Johnson noise (see Fish [7]) generated by R11 is denoted
by So1,R11 . The thermal noise generated by R11 is in series with any source voltage in
the inverting configuration of the op amp, if one were connected at the left side of R11 .
Because the gain of the inverting amplifier is given by −Zf 1 /R11 , So1,R11 can be written as
(Rodgers [18, p1093]):
¯ ¯
¯ Zf 1 (jω) ¯2
So1,R11 (ω) = 4kT R11 ¯¯ ¯ (B.26)
R11 ¯
The noise P SD at the output due to Johnson noise generated by R21 and R31 are denoted
by So1,R21 and So1,R31 . Since the left hand side of Zf 1 is at virtual ground, R21 and R31
both generate a noise component directly at the output, therefore
So1,R21 = 4kT R21 (B.27)
So1,R31 = 4kT R31 (B.28)
Finally, summing up all the noise PSD’s at the output we obtain the total noise PSD at the
output of the first stage of the preamplifier, denoted by Spa,1 :
Spa,1 = So1,en + So1,in− + So1,in+ + So1,R11 + So1,R21 + So1,R31 (B.29)
where the dependency on ω has been left out for convenience. Written out:
¯ ¯2 ( ¯ ¯2 )
¯ Z f 1 ¯ ¯ Zf 1 ¯
Spa,1 (ω) = ¯¯1 + ¯ Sen + |Zf 1 | + |Zs | ¯1 +
¯
2 2
¯
¯ Sin +
R11 R11 ¯
| {z } | {z }
Voltage noise Current noise
( ¯ ¯ )
¯ Zf 1 ¯ 2
+ 4kT R11 ¯¯ ¯ + (R21 + R31 ) (B.30)
R11 ¯
| {z }
Johnson noise
where the dependency on ω and jω has been left out for convenience.
Noise model for the second preamplifier stage

The noise sources of the second stage are identical to the noise sources of the first stage,
except for the current noise at the non inverting input of the opamp.
Figure B.7 Second stage of the preamplifier and all its noise sources
The total noise PSD at the output of stage 2 due to noise sources of stage 2 solely is denoted
by Spa,2 (s):
¯ ¯2
¯ Zf 2 ¯ © ª
Spa,2 (ω) = ¯¯1 + ¯ Sen + |Zf 2 |2 Sin +
¯
R12 | {z }
| {z } Current noise
Voltage noise
( ¯ ¯ )
¯ Zf 2 ¯ 2
+ 4kT R12 ¯¯ ¯ + (R22 + R32 ) (B.31)
R12 ¯
| {z }
Johnson noise
where the dependency on ω and jω has been left out for convenience.
Finally, the noise PSD at the preamp output can be written as
Spa (ω) = Spa,1 (ω) |Zpa2 (jω)|2 + Spa,2 (ω) 2

[Vrms /Hz] (B.32)
B.4.4 A noise model for the 1000-10 preamplifier

Having obtained a general expression for the noise PSD of the preamplifier, it is now possible
to plug in all the numeric values and investigate the resulting PSD’s. The noise level of the
preamplifier is defined by the found resistor and capacitor values at the end of Appendix B.3.
The voltage and current noise levels are defined by the used operational amplifier. In this
case study two opamps are considered, the op07 and the op27. Both cases are studied below.
Numeric values of the op07 noise levels can be found in Tables B.3 and B.4.
using the op07

In Figure B.8 plots can be seen of the output ASD’s corresponding to Spa,1 (ω) and Spa,2 (ω),
when using the op07 in the circuit. In both plots, the different components comprising the
ASD’s can be seen, such as the total thermal noise, voltage noise and current noise. Looking
at the current noise plot of the first stage, the effect of the geophone impedance can be seen
as the little bump in the 5 Hz region.
On the left hand side of Figure B.10 the two ASD’s can be seen again, together with the
total output ASD of the preamplifier corresponding to Spa (ω).
using the op27

The analysis is also made with another operational amplifier, the op27. This opamp has
different noise parameters than the op07, listed in Table B.4. The transfer function of the
preamplifier itself is exactly the same as the one build with the op07. Using the new values,
the resulting ASD’s are calculated as can be seen in
Figure B.9. On the right hand side of Figure B.10 the two ASD’s can be seen again, together
with the total ASD corresponding to Spa (ω) using the op27.
ASD at output of stage1 due to stage1 ASD at output of stage2 due to stage2
Voltage noise Voltage noise

−5
Current noise −5
Current noise
10 Johnson noise 10 Johnson noise
Total noise Total noise
−6 −6
10 10
/ Hz0.5
VRMS / Hz0.5
RMS
V
−7 −7
10 10
−8 −8
10 10
−2 0 2 −2 0 2
10 10 10 10 10 10
Figure B.8 left: Resulting ASD corresponding to Spa,1 (ω) of the first stage output of the
preamplifier using an op07 operational amplifier. right: Resulting ASD corresponding to Spa,1 (ω)
of the second stage of the preamplifier using an op07 operational amplifier.
B.4.5 Op07 and op27 performance comparison

To compare the performance of the two different opamps, one can compare the total output
noise PSD’s when using the different opamps. Both can be seen in Figure B.10. The total
PSD using the op27 has a higher value than the op07 PSD at the lower frequencies, but is
a little lower at the higher frequencies. To compare the performance one should analyze the
noise propagation in closed loop, which is done in Section 3.5.4. The critical region for the
preamplifier noise in the closed loop setting turns out to be the low frequency region. Since
the noise of the op07 has a lower value in this region, it is chosen to use the op07.
ASD at output of stage1 due to stage1 ASD at output of stage2 due to stage2
Voltage noise Voltage noise

−5
Current noise −5
Current noise
10 Johnson noise 10 Johnson noise
Total noise Total noise
−6 −6
10 10
/ Hz0.5
VRMS / Hz0.5
RMS
V
−7 −7
10 10
−8 −8
10 10
−2 0 2 −2 0 2
10 10 10 10 10 10
Figure B.9 left: Resulting ASD corresponding to Spa,1 (ω) of the first stage of the preamplifier
using an op27 operational amplifier. right: Resulting ASD corresponding to Spa,2 (ω) of the
second stage of the preamplifier using an op27 operational amplifier.
Total ASD at output of stage2 due to both stages Total ASD at output of stage2 due to both stages
−3 −3
10 10
due stage 1 due stage 1 op27
due stage 2 due stage 2 op27
total; both stages total; both stages op27
−4 −4 total; both stages op07
10 10
/Hz1/2
Vrms/Hz1/2
−5 −5
10 10
rms
V
−6 −6
10 10
−7 −7
10 10
−2 0 2 −2 0 2
10 10 10 10 10 10
Figure B.10 left: Total noise ASD corresponding to Spa (ω) at output of the preamplifier using
an op07. The separate contributions from the two stages to the output ASD can also be seen.
right: Total noise ASD at preamplifier output using an op27. Again, the separate contributions
of the two stages to the output ASD can also be seen. For performance comparison, the total
output ASD using an op07 is plotted again. The total ASD using the op27 has a higher value
than the op07 ASD at the lower frequencies, but is a little lower at the higher frequencies.
Appendix C
Experimental setup MIT
In this appendix the experimental vibration isolation setup used at MIT is described. First
a list of used hardware and software is given, then it is described how the hardware is
connected.
C.1 List of used equipment

• Hewlett Packard 35665A Dynamic Signal Analyzer
• Crown DC300R stereo power amplifier
• Cerwin-Vega VEGA 84 subwoofer/speaker
• dSpace 1104 controller board, PCI-card in PC.
• Geo Space vertical geophone, GS-11D @ 4.5 Hz @ 4000Ω, http://www.geospacelp.com.
• Texas Instruments, operational amplifier op07c (in preamplifier circuit).
C.2 List of used software

• MATLAB°r R12 and Simulink.
• ControlDesk 2.2: to control the dSpace board from a PC.
• sdftoml.exe: conversion program for HP analyzer *.dat files.
C.3 Hardware connection

In Figure C.1 a schematic of the hardware configuration of the vibration isolation setup can
be seen using the preamplifier. The setup is shown in closed loop configuration while mea-
suring the geophone output signal for analysis with the HP analyzer. The speaker-geophone
combination corresponds to the schematic shown in Figure 3.2 on page 21. In Figure C.2 a
115
116 APPENDIX C. EXPERIMENTAL SETUP MIT
photograph of the preamplifier-speaker-geophone setup can be seen. Most of the hardware

is electrically connected using bnc-cable’s (see figure), which has a inner-wire carrying the
signal and an outer shielding which is connected to earth ground located in the dSpace
board. The outer shielding of the bnc-cable’s protects the delicate signals from electrical
pick up noise. The power amplifier generates relatively large signals and can therefore be
connected to the speaker voice coil using two simple electrical wires.
In Figure C.3 the hardware configuration of the swept sine experiment, which was done to
identify the empirical poweramp-speaker-geophone model Eg /Ein , can be seen. The output
of the analyzer, generating the sinusoidal source signals is connected with bnc-cable’s to the
power amplifier input and to the second analyzer input. The geophone output is fed into
the first input channel of the analyzer. In the same manner the transfer function of the
preamplifier was validated; the output of the analyzer was connected to the preamplifier
input and its output was connected to the second input of the analyzer.
Figure C.1 Closed loop hardware configuration MIT vibration isolation setup measuring the
geophone output signal with the HP analyzer.
C.4 Practical notes on using the HP35665A DSA

When one wants to measure a PSD with the analyzer a few things have to be taken into
account. The analyzer shows the results on an the board display in the units set by the
user. When the results are written to a ∗.dat file on a floppy disk, the data will be saved
as an peak power spectrum with units [SI2 ]. To convert the data to PSD units [SI2rms /Hz ] the
software sdftoml.exe can be used (comes on a floppy disk with the analyzer). With the
following line the data is converted to the right units:
sdftoml sourcefile.dat /X /Y:PRD
The data is now written to a ∗.mat file which can be loaded in MATLAB°r . For more
conversion options see the software manual. sdftoml.exe can also be found on the CD
ROM that comes with this thesis.
C.4. PRACTICAL NOTES ON USING THE HP35665A DSA 117
Figure C.2 Hardware of the MIT vibration isolation setup. The speaker-geophone setup together
with the preamplifier located at the rim of the speaker. The preamplifier is contained in an
aluminum box, to shield it from electromagnetic disturbances. Two 9 Volt batteries can be seen
that are used to power the preamplifier. On the base plate of the setup a similar geophone as is
mounted on the cone plate of the speaker can be seen (this geophone is not used in the setup,
but is used to measure the ground vibration level).
Figure C.3 Hardware configuration MIT vibration isolation setup for swept sine identification
experiment.
118 APPENDIX C. EXPERIMENTAL SETUP MIT
Appendix D
CD-ROM
The files on the CD-ROM are divided over three folders: Matlab, LaTeX and Files.
D.1 MATLAB°r m-files

The /Matlab folder contains two subfolders AVIMIT and AVIPhilips, in which all the
MATLAB° r
files used in the case study at MIT and Philips can be found respectively. Both
case studies have the same m-file structure. The main m-file is a go.m file. From this m-file
all the other files are imported. The sequence of m-files and what happens in the m-files is
quite self-explanatory by the comments. The go.m file is designed to run from top to bottom.
The structure of the go.m files is briefly described below, the commented lines explain what
happens in the imported m-file:
% pick up parameters, all parameters are defined here!

parameters
% define all system models

system_models
% define controller(s)
controllers
% define all noise models

noise_models
% define systems and filters in generalized plant format

genplant_systems
% define generalized plant with filters

genplant
119
120 APPENDIX D. CD-ROM
% H2 controller synthesis
controller_H2
% close the loop and define all propagation transfer functions

closeloop
% simulate performance by propagating noise sources to output

calc_noise_prop
% plot propagations results

plot_prop_res
In the MIT case study a few different system configurations can be chosen. In the file
MIT readme is explained what to change in the m-files in order to simulate a different case.
Only in the MIT case study an H2 controller can be synthesized.
D.1.1 tools
In the folder Matlab/tools a few tools can be found that are needed to run the go.m files
and are self written. This folder must be added to the MATLAB°r search path.
• cumpsd.m, calculates the PSD, PS, CPS and variance of the output of a system with
a PSD at the input.
• spectra.m, calculates the PSD, PS and CPS of a sampled time signal.
• check genP.m, algorithm to check if the standard assumptions on the weighted gener-
alized plant are met for H2 controller synthesis.
• frqrsp.m, calculates frequency response of system.
• angl.m, calculates phase response of system.
• makesys.m, converts a state space system into system-matrix format in which the
generalized plant is defined, a format used in the µ-toolbox.
• makess.m, converts a system in system-matrix format to state space format.
• parr.m, calculates the parallel combination of two values, e.g., resistors.
More information on how to use the m-files can be found in their help text.
D.2. LATEXFILES 121
D.2 LATEXfiles
In this folder all the LATEXfiles used to write this thesis can be found. The main file of this
thesis is a WinEdt project file: MasterThesis.prj. The structure of the LATEXfiles follows
from the comments in the main file report.txt, which can be found when opening the
project file. In the folder image files most of the source files used to generate the images
in the thesis can be found, e.g., in the /PPT folder all the PowerPoint files are located. A
commented line above each figure environment in the LATEXcode points out were the image
was created.
D.3 Files
The folder Files/doc contains all digital literature used in this thesis, a large number of
*.pdf files. The folder Files/Pics contains pictures of the setup build at MIT.
Bibliography
[1] Balmer L. Signals and Systems. sec. ed. Prentice Hall, Europe, 1998.
[2] Barzilai A., VanZandt T., Kenny T. Technique for measurements of the noise of a sensor
in the presence of large background signals, Review of Scientific Instruments, Vol. 69,
No. 7, pp 2767-72, july 1998.
[3] Boyd S., Barrat C. Linear Controller Design - Limits of Performance. Prentice Hall,
1991.
[4] Scherer C. Theory of Robust Control. Lecture notes, Delft University of Technology,
Netherlands, 2001.
[5] Keith G., Doyle J.C. State-space formulae for all stabilizing controllers that satisfy an
H∞ norm bound and relations to risk sensitivity. System and Control Letters, Vol 11,
pp 167-172, 1988.
[6] Doyle J.C., Glover K., Khargonekar P.P., Francis B.A. State-space solutions to standard
H2 and H∞ control problems. IEEE Transactions on Automated Control, Vol. 34, No
8, august 1989.
[7] Fish P.J. Electric Noise and Low Noise Design. McGraw Hill Inc. New York, 1994.
[8] Goossens H.M.W. By speach, Philips CFT, Eindhoven 2003.
[9] Harris C.M. Harris’ Shock and Vibration Handbook. McGRaw-Hill Professional, 5th
ed., 2001.
[10] Horowitz P., Hill W. The Art of Electronics. Cambridge University Press, Cambridge,
1990.
[11] Ljung L. System Identification - Theory for the User. Prentice Hall, New Jersey, 1987.
[12] Monkhorst W. Feasibility Study of the Dynamic Error Budgeting Design Approach.
Mech. Eng. Systems and Control Group, Delft Univ. of Technology, The Netherlands,
2002.
[13] Oppenheim A.V., Schafer R.W., Buck J.R. Discrete-Time Signal Processing. Prentice
Hall, 2nd ed., 1999.
123
124 BIBLIOGRAPHY
[14] Papoulis A. Probability, Random Variables and Stochastic Processes. McGraw-Hill,

1965.
[15] Peterson J. Observations and modeling of seismic background noise, United States
Department of the Geological Survey, Open-File Report 93-322, Albuquerque, New
Mexico, 1993, http://aslwww.cr.usgs.gov/Publications, last accesed 01-2004.
[16] Priestley M.B. Spectral analysis and time series. Academic Press, London, 1989.
[17] Riedesel M.A., Moore R.D., Orcutt J.A. Limits of Sensitivity of Inertial Seismome-
ters with Velocity Transducers and Electronic Amplifiers. Bulletin of the Seismological
Society of America. Vol. 80, No. 6, pp. 1725-1752, December 1990.
[18] Rodgers P.W. Frequency Limits For Seismometers as Determined from Signal-to-Noise
Ratios. Part 1. The Electromagnetic Seismometer. Bulletin of the Seismological Society
of America, Vol. 82, No. 2, pp. 1071-1098, April 1992.
[19] Rodgers P.W. Maximizing the Signal to Noise Ratio of the Electromagnetic Seismome-
ter: the Optimum Coil Resistance, Amplifier Characteristics, and Circuit. Bulletin of
the Seismological Society of America, Vol. 83, No. 2, pp. 561-582, April 1993.
[20] Roover D. de. Motion Control of a Wafer Stepper - A Design Approach for Speeding
Up IC Production. PhD Thesis, Mech. Eng. Systems and Control Group, Delft Univ.
of Technology, The Netherlands, 1997.
[21] Slocum, A.H., Precision Machine Design. Prentice-Hall, Inc., New Jersey, 1992.
[22] Subrahmanyan, P.K. A Model Approach to Precision Motion Control. PhD Thesis,
Medch. Eng., Massachusetts Institute of Technology, Boston, USA, 1999.
[23] Texas Instruments Application Report. Noise Analysis in Operational Amplifier Cir-
cuits. SLVA043A, 1999. Internet document, last accessed Mch-2003.
[24] Trumper D.L., Sato T. A Vibration Isolation Platform. Mechatronics, No. 12, pp. 281-
294, Elsevier Science Ltd. 2002.
[25] Usher M.J. Developments in Seismometry. Journal of Physics E: Scientific Instruments,

Mechatronics, No. 6, pp. 501-507, Great Britain 1973.
[26] Vervoordeldonk M. By speach, Philips CFT, Eindhoven, 2003.
[27] Zhou K., Doyle J.C., Glover K. Robust and Optimal Control. Prentice Hall, Upper
Saddle River, NJ, 1996.

MSC Monkhorst 31mar04

Uploaded by

Copyright:

Available Formats

MSC Monkhorst 31mar04

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

MSC Monkhorst 31mar04

Uploaded by

Copyright:

Available Formats

Systems & Control

Dynamic Error Budgeting

Supervisors ir. L. Jabben

Delft, March 31, 2004

Faculty of Design, Engineering and Production

2 Dynamic Error Budgeting 7

3 Case study, active vibration isolation 19

4 H2 control in active vibration isolation 45

5 Case study, Advanced Isolation ModuleS 61

6 Conclusions and recommendations 91

B Noise and system modelling 99

C Experimental setup MIT 115

Eg (t) geophone output voltage [V ]

Abbreviations and Acronyms

1.1 Context of this thesis

1.2 Dynamic error budgeting

1.3 Vibration isolation

the DEB design approach for the design of the system!

1.4 Challenge definition

Develop a tool which enables the designer to account

Develop tools that enable the designer to:

• Objectively compare the performance of different system designs. When a system

1.5 Outline of thesis

Dynamic Error Budgeting

2.1 Motivations of the DEB design approach

2.2 DEB design process

• Design a concept system.

2.3 Assumptions and research boundaries

2.4 Theoretical background

2.4.1 Variance as a performance measure

variance σx2 = x¯2 − µ2x (2.1)

2.4.2 Parseval’s Theorem: linking the frequency to the time domain

2.4.3 Power spectral density function

|XL (fk )|2

2.4.4 Propagation of disturbances

Sz (f ) = H(j 2πf ) Sw (f ) H t (−j 2πf ) (2.14)

Now it follows from (2.1), (2.7) and 2.9 that

The ASD and CAS

2.5 H2 control, maximizing performance

2.5.2 The H2 norm and variance of the output

2.5.3 The H2 control problem

CH2 ∈ arg min kH(j 2πf )k2 (2.22)

2.5.4 Using weighting filters to model disturbances

|Sw (f )| = Vw (j 2πf ) Sin (f ) Vwt (−j 2πf ) (2.23)

2.5.5 Output scaling and the Pareto curve

2.6 Conclusions on the theory

Case study, active vibration

In order to apply the approach in practice, a complete vibration isolation setup is to be

3.2 An active vibration isolation platform

aluminum cone plate

strut cylinder strut

magnet voice coil magnet

3.2.2 Performance definition

3.2.3 System modelling

Speaker−Geophone transfer function Pspeaker from ls fit and from ss model

(s − 1.20 · 104 )(s + 3456)(s − 3.33 · 10−3 )

(s2 + 3.33 · 10−3 s + 1.11 · 10−5 )(s2 − 8757s + 4.94 · 108 )

Ground vibration propagation model Psp vib (s)

A geophone is a mechanism containing a proof mass which is suspended by leaf springs.