UT 03 TMI Signal Recovery
-
Upload
jayant-kirpekar -
Category
Documents
-
view
214 -
download
0
Transcript of UT 03 TMI Signal Recovery
-
8/17/2019 UT 03 TMI Signal Recovery
1/238
Signal recovery from noise in
electronic inst rum entation
T H
Wilmshurst
Department o Electronics and Computer Science
University o Southampton
Second Edition
Institute
of
Physics Pub lishing
Bristol and Philadelphia
Copyright © 1990 IOP Publishing Ltd.
-
8/17/2019 UT 03 TMI Signal Recovery
2/238
OP Publishing Ltd 1990
All rights reserved. N o part of this publication may be reproduced stored
in a retrieval system or transmitted in any fo rm o r by any means elec-
tronic mechanical photocopying recording or otherwise without the
prior permission of the publisher. Multiple copying is only permitted under
the terms of the agreement between the Committee of Vice-Chancellors and
Principals and the Copyright Licensing Agency.
ritish Library Cataloguing in Publication Data
Wilmshurst T. H .
Signal recovery from noise in electronic instrumentation.-
2nd. ed.
1.
Instrumentation systems. Electrical noise
I .
Title
620.0044
ISBN 0-7503-0059-0
ISBN 0-7503-0058-2 pbk
Library o Congress
Cataloging-in-Publication
Data are available
First edition published 1985 by A da m Hilger L td
Second edition published 1990 by I O P Publishing L td
Reprinted 995
Published by Institu te of Physics Publishing wholly ow ne d by T h e Ins titute
of Physics London
Institute of Physics Publishing Tec hno H ou se Redcliffe W ay
Bristol BS1 6NX U K
U S E ditorial Office: Inst i tute of Physics Publishing T h e Public Le dg er
Building Suite 1035 Ind epende nce Squ are Philadelphia P A 19106 U SA
Printed in Great Britain by
J
W Arrow smith Ltd Bristol
Copyright © 1990 IOP Publishing Ltd.
-
8/17/2019 UT 03 TMI Signal Recovery
3/238
-
8/17/2019 UT 03 TMI Signal Recovery
4/238
vi
CONTENTS
4.7 Shot noise amp litude
4.8 Therm al noise amplitude
5.1 General properties
5.2 Measuremen t time independence
5.3 Response to combined ‘noise’
5 4 Experimental result
6.1 Low-pass filter
6.2 Running average
6.3 Visual averaging
6.4 Baseline offset correc tion
6.5 Sloping baseline correction
6.6 Step measu rement
6.7 Spa tial frequency com ponents of
l f
noise
6.8 Frequency response
of
multiple time averaging
Frequency domain view of the phase sensitive detector
7.1 Analogue multiplier
PSD
7.2 White noise erro r
7.3 Reversing
PSD
7.4 Chopping
PSD
7.5 Comp romise arrangem ents
7.6 PSD mperfections
7.7 Non-phase-sensitive detec tors
7.8 Recovery
of
a signal submerged in noise
8.1 Add itional noise du e t o digitisation
8.2 Analogue-to-digital converters
8.3 Suitability of real ADC circuits for taking average
samples
8.4 Point sampling gate
8.5
Aliasing
Magnitude determination
for
transient signals of known
shape and timing
9.1 Flash spectrome ter
9.2 Limit setting
9.3 Weighted pulse sampling
9.4 Matched filtering
9.5 Boxcar sampling gate
9.6 Offset drift an d
l f
noise
9.7 Pe ak an d valley detectors
9.8 Software realisation of weighted pulse sampling
5
l f noise
Frequency response calculations
7
8
Digitisation and noise
56
58
60
60
62
65
66
69
70
72
73
74
80
81
83
85
91
92
96
98
100
103
1 4
105
109
111
111
113
117
118
121
126
126
128
131
136
138
140
152
156
Copyright © 1990 IOP Publishing Ltd.
-
8/17/2019 UT 03 TMI Signal Recovery
5/238
CONTENTS
vii
10
Measurement of the time of Occurrence of a signal
transient
10.1
Method for timing a transient with minimum white
noise erro r
10.2 Signal pulse timing
10.3 Signal ste p timing
10.4 Co m pa riso n of the step an d pulse method s
10.5 Offset drift and l f noise
10.6 Lock-in amplifier
10.7 Autocorrelation methods
159
160
164
176
178
180
184
185
11 Frequency measurement
188
11.1 Period ic transien ts of fixed width 189
11.2 Ratem eter 193
11.3 Phase-locked loo p 196
11.4 Periodic transients of width pro po rtio na l to period 199
1
I
.6 Single transien ts 209
11.8 Noise-optimised averaging of frequency measurem ents 217
11.9 Frequency -domain methods 219
Appe ndix: Fourier analysis 223
Tu tori al questions 225
Answers to tutorial questions 230
Index 23 1
11.5 Frequenc y tracker 202
11.7 Correlation matching 212
Copyright © 1990 IOP Publishing Ltd.
-
8/17/2019 UT 03 TMI Signal Recovery
6/238
Preface to the second edition
Th e second edition add s a furthe r chapter t o cover the subject of fre-
quency measurem ent. Also included ar e a set
of
tutorial q uestions with the
numeric answers given. These cover the material in all chapters and will
consolidate understanding of the material in the book.
T
H Wilmshurst
July
1990
Copyright © 1990 IOP Publishing Ltd.
-
8/17/2019 UT 03 TMI Signal Recovery
7/238
Preface
to
the
first edition
The subject of this book is the recovery of signals from noise in electronic
instrumentation. Th e term ‘electronic instrumentation’ is sometimes reserved
for instruments such as the oscilloscope and testmeter which are used to
measure specifically electrical variables. It should therefore be m ade clear that
the scope of the book goes much beyond this an d covers instrume ntation for
the measurement of any variable whether electrical o r not. The term
‘electronic’simply implies that electronic techniques a re used in the process of
measurement
The term ‘noise’ also requires clarification. Until recently it was used mainly
to refer to rand om fluctuations such as white and
l f
noise. Now however the
term is used to refer to alm ost any kind of unw anted signal in a n electronic
system. It is in the b road er sense th at the term is used in the title since the
recovery
of
the required signal from many kind s of unw anted signal is covered.
These comprise offset drift random noise comprising white an d l f noise and
interference. However to avoid confusion the term ‘noise’ will be reserved for
rand om noise. Th e other types of unw anted signal will be referred t o by their
specific nam es.
There are broadly two ways of reducing the errors du e to unw anted signals
in an electronic system. The first is to prevent the u nwanted signal from being
introduced . This is a m atter of low-noise amplifier design screening
decoupling etc and is reasonably well covered by existing texts. Th e second
app roac h which is taken up when the first has been exploited as far as possible
is to devise ‘signal recovery’ techniques which distinguish the required signal
as well as possible from whatever unw anted signals may remain. This second
app roac h is less well covered a t present a nd therefore represents the subject
matter of the book.
Th e text is intended for anyone w ho is to be involved in the development
of
Copyright © 1990 IOP Publishing Ltd.
-
8/17/2019 UT 03 TMI Signal Recovery
8/238
xii
PREFACE
electronic instrumentation a t grad ua te level o r beyond. In or de r to indicate
better the intended readership it will be helpful to consider briefly the way in
which electronic instruments are o r should be developed.
A
number of
different disciplines is nearly always involved. Consider for exam ple such a
system a s the laser anem om eter. This is a device which uses bo th electronic and
optical techniques for the measurement of fluid velocity. Thus the
development of this instrument would require experts in optics electronics
an d fluid dynamics. When such a team is formed it is impo rtan t that each
mem ber sho uld atte m pt to maste r all aspects of the system no t just those of his
own discipline. This is as true for the electronic aspects of the system as for any
othe r. Thus the book ha s to be acceptable to two rather different classes of
reader. The first is the undergraduate electronics engineer who intends to
specialise in instrumentation. He will meet the subject probably in the third
year of his undergra duate studies an d will need to obta in a good grasp of the
entire contents. T he o the r class of reader will be the non-electronics specialist
who will probably come to the subject after graduation and at the time of
joining t he team . He will usually need t o stud y in de pth a limited selection of
topics from the book. In orde r to help this class of reader attem pts are made to
cover the material in descriptive a s well as analytical m anne r. Also rather
more extensive subsection titling than usual has been adopted in order to
facilitate ‘random access’.
The book stems from two courses that are given in the Department of
Electronics of the University of S ou tham pton . The first has run for over a
decade and is intended fo r graduates in disciplines other th an e lectronics who
ar e engaged in the development of electronic instrum entation o r sometimes
just in its use. These may be studen ts working for higher degrees postdo ctoral
research fellows technicians and sometimes academic staff.
The second course is mo re recent an d is intended primarily for third-year
undergraduates in electronic engineering. It covers that material outside the
‘core’ electronics sub jects tha t is needed for a gra du ate w ho is to specialise in
electronic instrument development. The course is also followed by students
tak ing ‘with-electronics’ courses such a s ‘physics-with-electronics’ ‘chemistry-
with-electronics’ etc.
T H Wilmshurst
Copyright © 1990 IOP Publishing Ltd.
-
8/17/2019 UT 03 TMI Signal Recovery
9/238
1 Low-pass filtering and visual
averaging
1.1 OVERVIEW
Before discussing low-pass filtering and visual averaging, the two signal
recovery m ethods t o be discussed as the main topic of the present ch ap ter, we
devote the first section to a n overview of the entire book. Here th e classes of
unw anted signal, an d also all of the signal recovery methods to be described,
will be briefly introduced.
Resistor bridge strain gauge
The first step will be to describe the operation of a simple example of an
electronic instru m en t: the resistor bridge strain gauge
of
figure 1.1. Here the
variable t o be measured is the strain (extension)
of
the mechanical com pon ent
show n, such strain usually resulting from a m echanical force (stress) applied t o
the component.
Th e principal compo nent of the strain gauge is the resistor
R,.
This is a mesh
offine wires which is atta ch ed to the mechanical com po ne nt by an adhesive.
As
the mechanical compon ent is strained,
R ,
increases; this causes the bridge to
become unbalanced, producing a voltage U , which is approximately
prop ortion al to strain.
U,
s then amplified, low-pass filtered and d isplayed o n a
chart recorder.
Th e variety of instruments of this kind is vast, the single common feature
being that some variable, which is usually non-electrical, is converted to an
electrical signal which is processed an d displayed. The device which converts
the variable of interest t o the electrical signal is called th e inp ut ‘transducer’
an d for the present system the transduce r is the resistor bridge. This conver ts
the variable of interest, the strain, into the electrical signal U , . Although the
range of input transduce r types in use may con stitute a n im po rtan t topic for
Copyright © 1990 IOP Publishing Ltd.
http://0.0.0.0/http://0.0.0.0/http://0.0.0.0/
-
8/17/2019 UT 03 TMI Signal Recovery
10/238
2 LOW-PASS
FILTERING/VISUAL AVERAGING
Resistor
bridae
Balance
Time
Chart
Amplifier
Low-pass
filter recorder
‘ R -
I ‘ Y
St re ke d mechanical
component
Figure 1.1
Resistor bridge strain gauge.
the instrument engineer, this is not the subject of the present book, and is
adequately covered elsewhere. Instead, our objective is to establish the
principles whereby the electrical signal developed by such a transducer is
recovered from the various kinds of unwanted signal.
This objective will be best served if, wherever possible in the course of
introducing the signal recovery methods, the same simple example of the strain
gauge is retained. This is in con tras t to the alternative a pp roa ch of constantly
altering the example in ord er at the sam e time to try to give some idea of the
differing transducer types. With the principles of signal recovery thus firmly
established, using one type
of
transdu cer, it will be a simple m att er to a pply the
principles to any other type of transducer that might be encountered.
Low-pass filtering
of
white noise
Th e purpose of th e low-pass filter in the stra in gauge of figure 1.1 is to reduce
the amplitude of the unwanted signal components, of which the first to be
considered is white noise. Figure
1.2
illustrates the way in which the low-pass
filter reduces such noise. Here a fixed stress is applied to the mechanical
component at time
t =0
and the requirement is to measure the difference in
recorded strain before and after application of the stress. Figure 1.2 a)shows
what might be observed a t the ou tpu t of the amplifier in this situation. Here
there is a moderately high level of white noise superimposed upon the signal
component.
In du e course it is shown th at th e effect of the filter is to produce the o utp ut
voltage U , which is the ‘runn ing average’ of the inp ut voltage U,. Th e function is
illustrated in figure 1.2(b). Here, at time t , U, is th e average of U , over the period
shown from t- T, to t . Thus T, is termed the ‘averaging time’ for the filter.
In order to construct the output function
v o ,
the averaging box in figure
1.2(b)
must be made to ‘run’ along the input function U , in a), thereby
describing the o ut pu t function
U ,
in c) .Clearly the effect is to reduce th e noise.
It is shown subsequently th at the am plitude
V
of the noise compone nt of U , is
proportional to
T,-’/’,
i.e. V;
T - ’ ’ ~ .
Th e oth er feature t ha t is clear from figure
1.2
is that it takes th e finite time
T,
Copyright © 1990 IOP Publishing Ltd.
http://0.0.0.0/http://0.0.0.0/http://0.0.0.0/http://0.0.0.0/
-
8/17/2019 UT 03 TMI Signal Recovery
11/238
OVERVIEW 3
T-
I
I
Noise
I
t
I
--I
a-6
6)
kl
I
I t
I I
I f: 0
Figure 1.2
Diagram showing how the low-pass filter of figure 1.1 reduces
the
white noise
amplitude
by
forming a running average
of
the filter input.
( a )
Filter
input
U,
in
figure
l . l ) ,
( b )
running average process,
( c )
filter
output U,.
for U , to respond t o the step in
uin.
Thus
time'.
is also know n as th e filter 'response
Visual averaging
The second method of signal recovery to be considered is visual averaging.
Looking a t the filtered noisy signal step in figure 1.2, he basic requirement is to
determine the difference between the signal before an d after the step. Here the
eye forms two averages, one before and one after the step, and subtracts one
from the other. Now, because there are several noise fluctuations over each
averaging period, the eye is able to determine the signal voltage t o a n accuracy
greater than tha t given by the noise amplitude, even a t the filter ou tpu t.
The reason for this improvement is that basically it is the process of
averaging the noise which reduces its effect and filter and eye are just two
different me thods of averaging. The reason that the noise am plitude
V;
at the
filter ou tpu t is proportional to
T-'l2
is that when w hite noise is averaged by
any means over a period
T,,
the value of the typical noise e rro r in the average is
proportional to T,; l i 2 . Then, since for the filter T,, = T,, the noise error is
proportional to T - l i 2 .When, however, the eye extends the averaging period
to the period TJ2 of the half step the noise err or is reduced from the value
of
say k T , - ' i Z given by the amplitude to
k(TJ2)- 12.
Copyright © 1990 IOP Publishing Ltd.
http://0.0.0.0/http://0.0.0.0/http://0.0.0.0/http://0.0.0.0/
-
8/17/2019 UT 03 TMI Signal Recovery
12/238
4
LOW-PASS FILTERING/VISUAL AVERAGING
It is the recovery of signals from white noise by low-pass filtering and by
visual averaging that constitutes the main topic of the present chapter.
How ever, before proceeding t o this detailed study we com plete the overview
by briefly introduc ing the other types of unwanted signal an d signal recovery
methods.
Drift and offset
The next two types of unwanted signal
to
be covered are offset and drift. Of
these, offset is defined as that component of the total of all unwanted signal
components which does not vary with time. For the resistor bridge strain
gauge of figure 1.1 the two main sources of offset are the transducer (the
resistor bridge) and the signal amplifier. The transducer offset arises from
initial imbalance in the bridge and is easily removed by adjusting the balance
resistor sh ow n. Similarly the DC coup led signal amplifier will also be provided
with
a
suitable balancing potentiometer. Thu s, for the present example, offset
is not a major problem. However, for some systems the offset cannot be
so
simply removed. Then an acceptable altemative is the baseline correction
method to be discussed in 1.4.
Unfortuna tely the offset, even if initially adjusted to z ero , tends to ‘drift’ as
time proceeds. This is usually because of slow changes in the temperature of
the circuit.
Now, drift error tends to increase as the time
T,
spent making a
measuremen t increases; this is directly opposed to the trend for white noise, for
which the noise error varies as T,-li2.hus, drift tends t o frustrate a ttem pts to
obtain a low white noise error by using a long measurement period, and we
shall find tha t much of the strategy in signal recovery lies in devising methods
of overcoming this problem.
Multiple time averaging
Two main m ethods are used t o reduce the drift error and so allow the low
white noise e rro r associated with a long period of measurement to be realised.
These ar e multiple time averaging (MTA) an d the phase-sensitive d etector
(PSD)
me thod. Of these
MTA
is a matter of carrying o ut the experimen t rapidly rather
than slowly, thus reducing the drift error, and then extending the averaging
time t o o bta in the low white noise err or by repeating the experiment m any
times and averaging the results. The use of
MTA
to avoid effects of drift in this
way constitutes the main topic of chapter 2.
Phase-sensitive detection
Sometimes the experimental constraints do not allow the time scale of the
experiment to be reduced. Here a suitable altemative is to modify the
experiment in such a way that the transducer produces an
A C
rather than a DC
signal. This allows a n
AC
coupled am plifier to be used which does not respond
to offset or drift.
Copyright © 1990 IOP Publishing Ltd.
http://0.0.0.0/http://0.0.0.0/
-
8/17/2019 UT 03 TMI Signal Recovery
13/238
OVERVIEW 5
Here some kind of rectifier is usually required in order to convert the A C
amplifier ou tpu t into a DC signal suitable for display. This, then, is the function
of the phase-sensitive detector. It is the use of the P m m eth od to overcom e drift
that constitutes the main topic of chapter 3.
l f noise
Th e majo r remaining class of unwanted signal to be considered is l/f noise.
Fo r both white and l/f noise the names refers to the spectral distribution of the
noise. For white noise the spectral density
G
is independent of frequency f
while for l/f noise G f ’. hese terms belong to the frequency dom ain view
an d will be elaborated later. F o r the present, the waveforms of figure 1.3 (a) n d
( b )will suffice to identify the difference between the two. T hese ar e specimens
of white an d l/f noise as might be observed a t the o ut pu t of the signal amplifier
in the s train gauge of figure 1.1. Clearly the
l/f
noise ha s
a
greater tendency to
‘walk away’ from th e mean. This mean s th at the ad va nta ge s of using a larger
measurement period T, ar e less th an for white noise. In fact, it will be sho wn
that the
l/f
noise error is independent of T,. This con tras ts with w hite noise,
for which the error is proportional to T , - l i 2 , nd drift, for which the error
increases with T,.
6)
a )
Figure 1.3 Types of noise seen a t the outp ut of th e filter in figure
1.1
:
a )
white noise,
( b )
l/f
noise.
It is thu s clear tha t l/f noise too frustrates any attem pt to obta in a low white
noise error by increasing T,. For low values of T, the white noise may
dominate, a nd then increasing T, will cause a reduction. H oweve r, a point will
be reached a t which t he white noise erro r falls below the l/f noise e rro r, which
does not vary with
T,.
Then any fu rther increase in T, will have
no
effect.
Fortunately, both MTA and PSD methods ar e effective in rem oving l/f noise
err or b ut, in o rde r to show how this is do ne , it is necessary to use the spectral
view of both signal and noise. This is in contrast with the rather more direct
‘time-domain’ view used in the chapters listed so far.
. Frequency-dom ain view
T he frequency-domain view is a comm only a do pte d one an d th e present text is
perhaps a little unusual in not adopting it from the start. However,
I
have
preferred the directness of the time-dom ain ap pro ach for the initial treatmen t.
Thus, in chap ter 4 the frequency-domain o r ‘spectral’ view will be introduc ed
and the ground previously covered from the time-domain viewpoint re-
covered from the frequency-domain viewpoint. Cha pte rs 5 to
7
then use the
Copyright © 1990 IOP Publishing Ltd.
http://0.0.0.0/http://0.0.0.0/
-
8/17/2019 UT 03 TMI Signal Recovery
14/238
6
LO W-PASS FILTERING/VISUAL AVERAGING
frequency-dom ain view to show how the MTA and PSD methods remove the l/f
noise.
Digitisation
These seven chap ters com prise the first pa rt of the book , which deals with the
recovery of a con tinuously varying analogue signal from the various kinds of
unw anted signal. Th e remainder deals with one o r two closely related topics.
Frequently tod ay a n analogue signal must be digitised for com pu ter storage
an d p erhaps for further processing. In ch apter 8 the problems of digitisation
are covered. It transpires tha t the main factor of importance is tha t n o da ta
should be wasted.
Pulsed signals
Chapter
9
deals with pulsed signals o r more general types of signal transien t.
Of particular interest is the case where the shape of a transien t is know n but the
amplitude is not, being the factor to be measured. It is shown that the
processing required t o give minimum noise er ro r is what is know n as 'matched
filtering'.
Signal timing
In chapter
10
the problem is a little different. Here again the shape of th e signal
transient is known but the factor to be measured is the time of occurrence,
rather than the amplitude, of the signal. The necessary modifications to the
matched filtering in this case are discussed.
The need to use a matched filter when determining the amplitude of a
signal transient of known shape is well established. Also the modifications
needed when signal timing is to be determined are reasonably well
documented. However, such coverage is usually for white noise only. In
instrumentation, as distinct from communications applications, it is highly
likely tha t drift and l/f noise will be the predom inant unw anted signals. It is
the distinctive contribution of the present text that the case of l/f noise is
covered also.
1.2 LOW-PASS FILTERING O F SHOT NO ISE
Th e overview is now com plete an d we proceed to the main topic of the present
chapter. This is the way in which low-pass filtering and visual averaging
recover a required signal from white noise. There a re two m ain types of white
noise: thermal and shot noise. For all purposes of signal recovery these are
indistinguishable. Thus in this section the effect of low-pass filtering on shot
noise will be examined. It will be shown in particular tha t the am plitude V; of
the noise at the filter output is proportional to T,- 2 , where T, is the filter
averaging time. This relation also holds for thermal noise.
Copyright © 1990 IOP Publishing Ltd.
-
8/17/2019 UT 03 TMI Signal Recovery
15/238
LOW-PASS FILTERING O F
SHOT
NOISE
7
Thermal noise
It will, nevertheless, be useful to indicate briefly the origins of thermal noise.
This originates mainly from the resistors in
a
circuit a nd arises from th e fact
that the mobile charge carriers in a resistor are in constant thermal motion.
Thus, at any one time the charge distribution over the resistor is not quite
uniform. Th is causes a potential difference across the ends of the device. As the
therm al motion p roceeds, the potential fluctuates in
a
random manner abou t a
mean of zero. Th is class of noise is discussed further in 4.8, with qu antitative
precision.
Shot noise
Shot noise tends to predominate in the semiconductor com pon ents in a circuit,
i.e. in the d iodes a nd the transistors. It is actually
a
manifestation of the fact
that an electric current is not a continuum but consists of discrete current
carriers, i.e. electrons and ‘holes’. Thu s the commo nly used analogy of a stream
of flowing water to represent an electric current might more properly be
replaced by a stream of grains of sand . It is then the ‘grainy’ na tu re of the flow
that constitutes the shot noise.
Shot noise model
Figure 1.4 gives a simple model for the shot noise in the amplifier of figure 1.1.
Here it is reasonable to assum e tha t all of the shot noise originates from the
first stage, because t ha t from later stages is subject t o less gain. Th e transistor
+----*-- yu p p l y
LC
t
a b )
C)
Figure 1.4 Circuit used t o sh ow the effect of th e low-pa ss filter in figure 1.1
upon the amplifier sh ot noise.
a)
First stage, ( b ) urther stages,
( c )
ow-
pass filter.
current
i
actually consists of a random train of pulses, with each pulse
corresponding to the passage of an electron from the em itter to th e collector of
the transistor. T he ar ea of each pulse will be qe, he electron charge. Th us the
area
q
of the corresponding pulse at th e ou tpu t of the amplifier A will be given
Copyright © 1990 IOP Publishing Ltd.
http://0.0.0.0/http://0.0.0.0/http://0.0.0.0/
-
8/17/2019 UT 03 TMI Signal Recovery
16/238
8
LOW-PASS FILTERING/VISUAL AVERAGING
by
4
= @ , A .
(1.1)
Figure 1.5(a) shows the sequence of these random pulses that make up the
waveform for
U,
=
U;,,
in figure 1.4. This is somewha t idealised, assuming that
the amplifier co ntains no low-pass filter, but will d o for o ur present purposes.
p T r 4
- - - I
Figure
1.5
Memory box running average approximation to filtering of
sho t noise pulse train for transistor in figure 1.4. a)Sho t noise pulse train
c,,,.
( b )
running average
c
Running average
It will next be show n t ha t the essential feature of the filter,in t he present view, is
that it forms a ‘ru nning average’ of the inpu t. Figu re
1.5
illustrates this function
once more. Here the outpu t
U ,
a t time
t
is the average of
U,,
over the period
r,
of
the mem ory box. Then the outp ut function U, is constructed as the box ‘runs’
along the input function. In mathematical terms
t.,(t)= T- ’
U, ‘)
dt‘.
1
- T,
Th e fluctuation component
U ,
of U , is the sho t noise an d arises because
U, is
pro portio nal to the num ber of pulses p in the box and
p
fluctuates because of
th e rand om placing of the pulses. T he fluctuation time is T,because this is the
time taken for all of the pulses in the box to be replaced.
Filter response
The next step is to determine the exact relation between the input U,, and
output U , of the low-pass filter and thus to show that 0, is essentially the
running average of uin .
Consider first th e ‘impulse response’
gi
of the filter, shown in figure
1.6(f).
Copyright © 1990 IOP Publishing Ltd.
http://0.0.0.0/http://0.0.0.0/http://0.0.0.0/http://0.0.0.0/http://0.0.0.0/http://0.0.0.0/http://0.0.0.0/
-
8/17/2019 UT 03 TMI Signal Recovery
17/238
LOW-PASS F IL T E R IN G O F SHOT NOISE
9
I
M e m o r y
function I + rCR
d ) f )
0 I \ \
I t
I
Figure
1.6 Relation between ou tpu t U, and input U,, for a low- pass filter. a) n p u t
U,,, ( b ) ilter,
(c)
components of outp ut U, originating from elements in a), d )
memory function, ( e ) nput impulse,
(f)
mpulse response.
This is the response of the filter to the unit impulse shown in figure 1.6(e),
which is applied a t time t
= 0.
Here th e unit impulse is a pulse of unit are a an d
of width approaching zero.
The input impulse leaves the capacitor with a charge. This decays
exponentially through the resistor
R
as shown, with the decay time C R.
Next consider the shaded element a t time t of the general inp ut waveform uin
shown in a).For 6 t small, this approxim ates to a n impulse of are a ui,(t ) 6 t .
This will produce a t the filter outp ut th e shaded transient shown in
c ) .
At the
time
t ,
he output component
6 u ,
will then equa l the impulse are a multiplied by
gi(t ), .e. ui,(t )gi(t
)
6t. Then summing 6 u , for all values of uin,
Looking a t the transients in
c)
following th e various elements of uin shown
in a) ,and numbered 1 to
5 ,
it is clear th at U , consists of contr ibu tion s from U , ,
extending back from time
t
at w hich th e filter ou tp ut is observed for a period
roughly equal to
CR
This is highly reminiscent of the ru nning average, with
the averaging period T = C R . The main difference is that the 'memory
function'
g i ( t - t ' )
from equation (1.3) and shown in ( d ) has an exponential
'taper'. This co ntra sts with the a br up t cut-off of th e 'memory-box'
approximation corresponding to the running average. This property is
sometimes referred to as a 'fading' memory of the input.
Copyright © 1990 IOP Publishing Ltd.
-
8/17/2019 UT 03 TMI Signal Recovery
18/238
10
LOW-PASS FILTERING/VISC:AL AVERAGING
Weighted running average
Clearly U, is no t a simple running average of vi over T,but a 'weighted' runn ing
average, the weighting given by the fading memory function gi(t- ) . Actually
it is only clear from eq uation (1.3) that U, is a w eighted integral of
uinr
over the
approximate period
T,.
An average, whether weighted o r otherwise, is said to
be prop erly normalised if, when th e series of values that a re being averaged a re
all the sam e, the average is equal t o an y on e of the values. Bu t, for the filter, if uin
is cons tant, then U, =
uin.
Thus U, is
a
properly normalised weighted running
average of
uin,
with the weighting given by the memory function which is the
impulse response gi reversed in time.
We now return to the main theme of the present section, which is to
calculate the amplitude
6
of the noise com ponen t U, of the filter outpu t U, when
the random train
of
impulses shown in figure 1.5 a )constitutes the input
uin.
The approximation that U, is the running average of
uin
will be assum ed, with T,
the averaging period. T he n, with
q
the pulse area, U, at any t ime
t
is given by
where
p
is the number of pulses in the box at time
t .
Probability density function and standard deviation
T he fluctuation in
U,
arises from the fluctuation in
p .
T o determine this it is first
necessary t o introduce the concepts of stan da rd deviation a nd the probability
density function.
If
the memory box is placed at a very large number of
different an d mutually exclusive locations on the inpu t pulse train,
a
histogram
of the results can be p lotted as in figure 1.7, together with th e overall mean
e.
He re the horizontal coord inate is the num ber of times, n, th at each value of
p
was obtained.
t
n,vp
Figure
1.7
Prob ability density function for the num ber of pulses
p
spanned
by the memory box of
figure 1.5.
For present purposes it is sufficient to say that the half-width
sp
of the
histogram is the 'standard deviation' and that this is given by the relation
g p=
P ) Z 1.5)
Copyright © 1990 IOP Publishing Ltd.
http://0.0.0.0/http://0.0.0.0/http://0.0.0.0/http://0.0.0.0/
-
8/17/2019 UT 03 TMI Signal Recovery
19/238
LOW-PASS
FILTERING OF SHOT NOISE
11
provided 1. The n op s a good indication of the am plitude of the fluctuation
in p . However, a little more precision may be preferred. Thus, if the count
values n in the histogram are all divided by a co mm on factor k to give v p n / k
and k is such tha t
J
v p dp = 1, the scaled histogram v p s termed th e 'probability
density function'
(PDF)
for
p .
Strictly,
v p
has the 'gaussian' form
and op s the half-width when v p alls from the central maximum by a factor
exp(- /2).
It will be noted that , unlike the rest of the text, the obse rvations m ade in this
subsection are stated rather than reasoned. They are, in fact, stand ard results of
statistical theory a nd equation (1.5), in particula r, is a 'foundation' statem ent
upon which much of wha t follows must stand . It will therefore be of value t o
reflect upon its plausibility for varying values of
P.
Notice for instance how
when e 1 the ratio ode becomes small compared with 1.
Fr om equation (1.4), he stan dard deviation U for the noise com ponent U, of
U, is given by
Then, from equation (1.5),
6
= C p q / T , * (1.7)
on
p 2q/
T,.
(1.8)
Here cc
T,.
Let be the long-term mean pulse rate. The n p YT and
U =
q(Y /T , ) ?
This can be written as
on
k T , - ' I 2
since q and r are both independent of T,.
(1.9)
(1.10)
RMS
value
The usual way to represent the amplitude of a fluctuating variable is by its
'root-mean-square' (RMS) value. F o r the general fluctuating variable
x ,
the RMS
value of x is
( ? ) I z ,
where the average is over a period very large compared
with th e fluctuation time for
x .
Th e more concise symbol x'is usually a do pte d,
i.e.
X E
(x2) ' I2 .
It is clear that the
RMS
value ?a nd the standa rd deviation nX re com parable.
In fact they a re equal an d, in this sense, the two terms a re interchangeable.
However,
x'
expresses the amplitude of the fluctuating variable, while
ox
indicates the typical deviation from the mean for just one measurement.
Since for U,, 6= U equa tion (1.10) becomes
fin =
kT,-'I2. (1.11)
Th us o ur initial objective of showing tha t 6
T - l 2
has been achieved.
Copyright © 1990 IOP Publishing Ltd.
-
8/17/2019 UT 03 TMI Signal Recovery
20/238
12 LOW-PASS FILTERING/VISUAL AVERAGING
Experimental result
Figure
1.8(a)
to (c)shows the results of a series of measurements made upon a
signal step. These confirm tha t each time the averaging period T,, i.e. the filter
time constant
C R ,
s increased by a fac tor of
10
the noise amplitude 6 decreases
by a factor
of
It
can also be seen how the filter response time an d the
noise fluctuation time increase in proportion to T,. Note also how it is
necessary to increase the time taken to make the measurement in order to
accommodate the increase in the response time T,.
T r
20s
(C 1 id1
T 200ms
T
:IO00 s
Signal
i e )
s t e p i f
Figure 1.8 Step+white noise recorded at the output of a low-pass filter. @)-(e)
Recorded traces, (f) xperim ental system. In these traces, T, s the filter response time
and
T ,
the time taken to record the step.
Noise error
in
integral
Suppose th at, instead of the average, the integral of the shot noise is formed
over the period T,. Since the integral is the average m ultiplied by T,, an d since
the stan da rd deviation
on
f the average is prop ortiona l to
T,-
' I z , the standard
deviation ol f the integral will be proportional to T+1 2. ore generally, for
Copyright © 1990 IOP Publishing Ltd.
-
8/17/2019 UT 03 TMI Signal Recovery
21/238
VISUAL AVERAG ING
13
an integration period
ai
=
kT,’”.
(1.12)
Amplifier filter
It should, perhaps, be noted that the noise at the output of
a
real signal
amplifier would never appear as in the idealised model of figure 1.5 a).
In
reality there would always be some sort of low-pass filter in the amplifier,
whether by accident o r design, an d this would sm oo th the train of random ly
placed impulses to prod uce
a
waveform m ore like th at of the noise in figure
1.2(a).Here the small fluctuation time will be the response time of the filter
inte m al to the amplifier. The ex ternal filter, with the m uch longer response
time T,, then extends the fluctuation time to
T,.
1 3
VISUAL AVERA GING
Visual averaging is the process whereby the averaging of a displayed noisy
signal by the eye of the observer makes the overall noise er ro r less th an th at
given by th e noise amplitude Cn. igure
1.9
shows noise a t the o utp ut of
a
filter
of averaging time T,superimposed up on a required signal U, an d displayed for
the period T,.
If
then V; =
T , - ’ ” ,
it follows that the sta nda rd deviation for the
noise averaged over the period T, will be given by
a,
where
a, = k T,- (1.13)
Then, since T ,g T,, ov< O; and a reduction is obtained.
I I
I
I T”
I
I
Figure
1.9 Noise at low-pass filter output superimposed on required DC signal t’,
and averaged by eye over the period
T,.
Another way of looking a t the process is as follows. Th e fluctuation time for
the noise is T,. Th us th e eye averages
n
= T,/T, ndependent fluctuations. The
reduction in noise erro r is by cT /on=n -I/’. This is an exam ple of the general
tru th th at, if
n
independent ran do m samples are averaged, the resulting noise
error is reduced by n-’/’.
N ote tha t the final noise erro r av s independent of the filter time con stan t T,,
althou gh the noise amplitude
Cn T - l i Z .
his is because
if
T, s reduced, say,
causing V; to rise, the num ber n=
T,/T
is increased, so that the eye, having
more fluctuations to average, can obtain a better reduction. In fact the two
Copyright © 1990 IOP Publishing Ltd.
http://0.0.0.0/http://0.0.0.0/http://0.0.0.0/http://0.0.0.0/http://0.0.0.0/http://0.0.0.0/http://0.0.0.0/http://0.0.0.0/http://0.0.0.0/http://0.0.0.0/
-
8/17/2019 UT 03 TMI Signal Recovery
22/238
14 LOW-PASS FILTERING/VISUAL AVERAGING
effects exactly cance l, an d this is obvious when it is understood tha t the only
thing tha t varying T, does is to alter the division of la bo ur between th e filter
and the eye in forming the final average over the period q.
The relative inconsequence of the se tting of T,, regardless of its effect upon
the noise am plitude, is an effect frequently not realised. In fact,
T,
need only be
large enough to reduce the num ber of fluctuations t o a value small enough to
allow the eye to average effectively.
Experimental result
Th e sequence of measurem ents in figure 1.8(a), (d) and e)confirms this point.
Recall that the sequence from a ) to (c) corresponds to increasing the
measurement time
T,
an d the filter response time T, ogether in decade steps.
This shows the variation of
V
with T , - 1 / 2 .n con trast, the sequence a), d ) , e)
increases T, in the same way but leaves T,unaltered. Here 1.7 remains the sam e
as for (a)bu t the accuracy with which the step height ca n be determined plainly
increases, because of the increasing number of fluctuations for the eye to
average. Following the abov e contention tha t the noise erro r is independent of
T,,we assert th at , comp aring say ( b ) nd (d) (although V for
( b )
s less th an for
d)) , he increased potential fo r visual averaging in (d) reduces the final noise
error t o the same a s for (b ) .Similarly
(e)
should give the same final err or as c) .
Resolution
time
Fo r most measurements it is possible t o define a required 'resolution time' a nd
this is the period over which visual averaging is carried out. Consider, for
exam ple, the result of figure 1.10, He re the strain gauge of figure 1.1 has been
used t o plot the stress-strain curve for the mechanical comp onent. T he sample
com pone nt is placed in a testing machine which applies
a
controlled stress to
the sample. The stress is increased a t a co nsta nt rate (scanned). This is the
independent variable. Then the resulting strain (the dependent variable) is
plotted against the stress.
Here it is necessary to decide upon the number
of
independent points
needed t o define the stress-strain curve. This may be expressed as the nu m be r
npof poin ts o r the 'fractional resolution'
i
=
np- .
F o r the present rather simple
function,
np
will be fairly small. Fo r som ething more complex, such as a m ulti-
line optical spectrum,
np
would be a good deal larger.
F or the stress-strain curve example, T,,,
=
T,/n,. Th e eye averages over each
period of length T,, giving, for each time-resolved point, the noise error
IS =
T ; l l2 .
(1.14)
But
T,,= T,/np=XT,,
so
IS^
= ~(xT,) (1.15)
Since i is independent of
T,,
this means 6,
T,- 12,
and
so
the larger
T,,
the
better.
Copyright © 1990 IOP Publishing Ltd.
http://0.0.0.0/http://0.0.0.0/http://0.0.0.0/http://0.0.0.0/http://0.0.0.0/http://0.0.0.0/http://0.0.0.0/http://0.0.0.0/http://0.0.0.0/
-
8/17/2019 UT 03 TMI Signal Recovery
23/238
BASELINE SUBTRACTION
t
15
Figure
1.10 Output
U, of
the strain gauge in figure1.1 as the stress applied
to the sample is scanned at a constant rate.
1.4 BASELINE SUBTRACTION
Figure
1.11
shows an othe r method of dealing with offset. He re the example is
the stress-strain curve measurement of figure 1.10 but, instead of applying the
scanned stress immediately, zero stress is applied for the initial period G,
Figure 1.11 Baseline subtractio n: a) pplied stress,
( b )
train gauge out pu t
show ing offset
U,,,.
Copyright © 1990 IOP Publishing Ltd.
http://0.0.0.0/http://0.0.0.0/http://0.0.0.0/http://0.0.0.0/
-
8/17/2019 UT 03 TMI Signal Recovery
24/238
16 LOW-PASS FILTERING/VISUAL AVERAGING
during which the baseline (offset)
uOb
s measured.
uob
is then subtracted from
each value of measured strain when the stress is applied.
T i
needs to be of non-zero length in ord er t o allow for noise. Th e noise err or
in the baseline measurement is equal to
kTi,- '.
Since uob is subtracted from
each time-resolved value
of
v o
for which the error is
kTe; l i ,
we require that
B T,,, if
the tot al noise err or is not to be significantly increased. No tice t ha t
the requirement is that
Ti
B T,,, not
B
T,.
Copyright © 1990 IOP Publishing Ltd.
-
8/17/2019 UT 03 TMI Signal Recovery
25/238
2 Multiple time averaging and drift
In the overview ( 1.1)multiple time averaging MTA) was introduced as a means
of avoiding drift and so allowing the low white noise err or associated with a
long period of measurem ent t o be realised. This po int is covered in detail in the
present chapter, but first we consider a use of MTA where the object is merely to
reduce the white noise error.
Figure 2.1 shows the arrangem ent. Here the resistor bridge strain gauge of
figure 1 1is used to m on itor the oscillatory strain in a vibrating beam after the
beam has been struck by an im pacting device. If the experimen t is repeated a
num ber of times an d the results averaged using MTA, the white noise er ro r is
reduced a s the averaging time is increased.
This kind of measurement should be con trasted with one where there is an
independent variable and it is possible to vary the length of time T, used to
scan through the required range of the variable. An example of this kind of
measuremen t is the use oft he strain gauge to measure the stress-strain curve of
a samp le, as discussed in 1.3. Here the scanned independent variable is the
stress.
For such measurements, the easiest way to increase the noise averaging
time, an d thus reduce the white noise e rro r, is by increasing T,. Leaving
T
at
the original value a nd repeating the scan u p t o the full time available has
n o
advantag e over this. How ever,a s will be shown sho rtly,it is when drift (and
l/f
noise) are present that MTA can be used to advantage with the stress-strain
curve type of measu rement. Here, if a sh or t
T
is used an d MTA employed with a
repeated scan, the low drift and l/f noise associated with the short
T
are
combined with the low white noise error associated with a large averaging
period.
Copyright © 1990 IOP Publishing Ltd.
http://0.0.0.0/http://0.0.0.0/http://0.0.0.0/
-
8/17/2019 UT 03 TMI Signal Recovery
26/238
18
MULTIPLE TIME AVERAGING AND DRIFT
I 1
reco rde r
Fil ter
Figure
2.1
Use
of
a resistor bridge strain gauge to record the oscillatory
response of a vibrating beam to an impact.
2.1 MULTIPLE TIME AVERAGING METHODS
Overlay
averaging
There are two commonly occurring ways in which the results of a repeated
measuremen t ar e averaged. Th e first is shown in figure
2.2 a).
Here , each time
the beam is struck
a
new noisy transient is plotted
on
the paper
on
top
of
the
last, giving a series of ‘overlaid’ traces. These a re then averaged by eye to give
the reduced white noise error.
Figure 2.2 a) Overlay
of
three traces taken from a section
of
the output
transient from the system
of
figure
2.1.
( b ) Computer averaging of the
traces in
a).
Copyright © 1990 IOP Publishing Ltd.
-
8/17/2019 UT 03 TMI Signal Recovery
27/238
DRIFT 19
Computer averaging
Th e second, an d m ore sophisticated, method of averaging is to use a com puter.
Here all of the values of u are input t o the compu ter, the traces are averaged by
the com puter an d the result is displayed. F o r the three comp onent traces of
figure 2.2(a), this would give the result of
(b ) .
Whichever method is used,if the num ber of traces averaged is n,, the am ount
of time available for noise ave raging is increased by the fac tor
n,.
F o r one of the
time-resolved values of strain
on
a single transient the white noise error
IT
s
given by
u n = k T e ; l i z ,
here Te, s the resolution time. Then for
n,
averaged
traces, the final erro r for one of the time-resolved po ints becomes
Thus, the averaging of n, traces gives an improvement of n;’/’.
2.2 DRIFT
Before considering how multiple time averaging can be used to reduce drift, we
first consider the subject more generally. Drift usually arises mainly from
changes in the am bien t temperature , draughts, etc. These influence the offset a t
the signal amplifier input. Th e system enclosure tends t o act as a low-pass filter
with a time cons tant of a minute o r
so
Th us the fluctuation time has this
value also.
Sloping baseline correction
Figure 2.3 shows an exam ple of such drift, a s seen at the ou tp ut of the strain
gauge. F o r an experiment such as the stress-strain curve measurement, if the
stress scan time T is somewha t less th an the drift fluctuation time a s shown ,
then the drift approximates reasonably well to a cons tant rate ram p over the
scan period.
When the drift rate is truly constant, it is possible to correct for the drift
exactly, using the method of figure 2.4. This
is
similar to the simple baseline
correc tion system of figure 1.11, but now the baseline is measured bo th before
Figure
2 3
Sample of drift of fluctu ation time T, scan time T, for
stress-strain curv e m easu rem ent.
Copyright © 1990 IOP Publishing Ltd.
http://0.0.0.0/http://0.0.0.0/http://0.0.0.0/http://0.0.0.0/
-
8/17/2019 UT 03 TMI Signal Recovery
28/238
20 M L L T I P L E T I M E AVERAGING AND DRIFT
Figure 2.4
Sloping baseline correction m etho d for drift
of
constant rate, as
applied to stress-strain curve measu reme nt. a ) Applied stress, ( b )
measured strain +drift,
(c)
reconstructe d baseline.
an d after the scan, in each case with t he applied stress zero. Fro m the m easured
values, the entire sloping baseline can be reconstructed. Then th e app ropr iat e
baseline value can be subtracted from each recorded value of measured strain
during the scan.
As before, the finite period G s needed for the baseline measurem ent. T his is
to reduce white noise error. In order for the noise err or in the baseline value to
be small compared with that in the measured strain, we again require that
q
,,, rather than & T .
In reality th e drift rate is no t constan t. Figure 2.5 shows the err or tha t results
in the stress-strain curve measurement for bo th simple and sloping baseline
Figure
2.5 Drift err ors rem aining after simple baseline correctio n
of
figure
1.1
1
an d sloping baseline correction
of
figure 2.4. a ) Simple baseline,
( b )
sloping baseline, c ) actual drift, d ) error for
a ) ,
e ) error for ( b ) .
Copyright © 1990 IOP Publishing Ltd.
-
8/17/2019 UT 03 TMI Signal Recovery
29/238
DRIFT
21
correc tion. Clearly, provided
T, <
z, he second method still gives a reduced
drift error, although not now zero.
Drift reduction by multiple time averaging
It is clear that for either
of
the baseline correction m ethod s offigure
2.5
the drift
error can be reduced by reducing T,. With T x 4 as shown, the simple
correction gives a drift error approxim ately propo rtional to T, an d the sloping
baseline correc tion method gives a decrease with T, which is even mo re rapid.
Unfortunately, white noise er ro r is prop ortional to T,-
’ I 2 , so
reducing T, in
this way gives an increased noise error. This is avoided by repeating the scan
over and over in the time available and averaging the results. In this way a
large measurement period can be used to give the same low w hite noise err or as
for one long scan taking the entire period, but with the actual scan time
reduced to the degree necessary to mak e the drift err or negligible.
Here it is assumed that, when a short scan is repeated and the results
averaged, the final drift error is the sam e as for th e original short scan. It is
conceivable th at the effect of repetition an d averaging would be an increase in
drift erro r with the numb er
n
of scans averaged. Figure
2.6
shows tha t this is
not so. This shows the drift error fo r a series of scans, using simple baseline
correction. Here the overall period
n,T,
is somewhat less than the drift
fluctuation time
z.
When n,T,<
the drift rate is nearly con stant over the
overall period
n,T,,
an d then the averaged drift er ro r is the sam e as for on e of
the short component scans. When, on the other ha nd ,
n,T,
> but still with
T
-
8/17/2019 UT 03 TMI Signal Recovery
30/238
22
MU LTIPLE TIME AVERAGING AND DRIFT
does increase with the num ber
n
of scans averaged a nd , for con stan t rate drift,
is proportional to n .
For some measurements, offset is of no importance. For example, for the
measurem ent of a signal step the offset ca n simply be ignored. F or this class of
measurement therefore the baseline correction is not strictly required. This
con trasts with the above stress-strain curve measurem ent where offset er ro r is
important. Here, for multiple time averaging to be effective, some type of
baseline correction is required.
2.3 MULTIPLE TIME AVERAGING BY OSCILLOSCOPE
It is im po rta nt t o realise, when predicting the noise er ro r of a measurement,
tha t the norm al oscilloscope display may incorp orate a high degree of overlay
multiple time averaging of the type shown in figure 2.2(a). Here a number
n,
of
traces a re overlaid, where
n
T JIT; an d where
I T
is the oscilloscope time-base
period and
T
is a ‘memory time’ compounded from the
CRT
screen
persistence, the persistence of the observer’s eye and the observer’s visual
memory. The noise error is thus reduced from the value kT,;”’ given by
equation (1.14) by the factor (T,JIT;)-”’ to give
n k TeST,,/IT;)’”.
(2.2)
Clearly , for very large values of n
T JIT;
this is somewhat d egraded, as the
eye becomes ‘saturated‘.
2.4 MULTIPLE TIME AVERAGING BY COMPUTER
Hardware and program
Figure 2.7 indicates how a com puter is used for multiple time averaging. Here
the signal averaged is tha t from the vibrating beam of figure 2.1. T he beam is
struck periodically using the impact transducer which is drawn by the pulse
generator output A ) .Here the pulse period 6 is made somewhat larger than
the decay time for the transient oscillators of the beam, which is shown in B ) .
Th e overall program MTAV for averaging an d display is summ arised in figure
2.8 and includes the procedures ADNT,
DOT
and
DISP
(figures 2.9-2.1
1).
Th e signal transient is sampled every T, seconds, as shown in B ) ,giving a
total of n s = 500 samples between each impact. Within th e com pu ter is a d ata
buffer of
n
locations which is initially cleared. This is the first step of
MTAV.
Then, following each impact, the n samples are added to the value stored a t
each successive location in the d at a buffer. F o r o ne transient this is detailed in
the procedure ADOT This is repeated for the required nu m be r n of transien ts as
indicated in ADNT, hich is the second step of MTAV.
Then the data buffer contents represent the sum of n ransients. Thus, to
Copyright © 1990 IOP Publishing Ltd.
http://0.0.0.0/http://0.0.0.0/http://0.0.0.0/http://0.0.0.0/http://0.0.0.0/http://0.0.0.0/http://0.0.0.0/http://0.0.0.0/http://0.0.0.0/http://0.0.0.0/http://0.0.0.0/http://0.0.0.0/http://0.0.0.0/http://0.0.0.0/http://0.0.0.0/http://0.0.0.0/http://0.0.0.0/http://0.0.0.0/http://0.0.0.0/http://0.0.0.0/http://0.0.0.0/http://0.0.0.0/
-
8/17/2019 UT 03 TMI Signal Recovery
31/238
W
Copyright © 1990 IOP Publishing Ltd.
-
8/17/2019 UT 03 TMI Signal Recovery
32/238
24
MUL T IPL E
TIME AVERAGING
AND
DRIFT
Output contents
of
d n t a b u f f e ra t p oint er l o ca t i o n t o
digital-to-analogue converter
MTAV
Cle ar d a t a b u f f e r
ADNT Sum
nt
reco rded t rans i en ts in to the d a ta bu f fe r
Div ide each sample in the b uf fe r by n t
D l S P Display the b uf fe r conte nts on the oscil loscope
Figure
2.8
Structure
of
wAv-main program for multiple time averaging,
using the system of figure
2.7.
ADNT
Transient counter t
A DO T: Add o ne recorde d t ra n s ie n t i n t o t h e d a t a b u f f e r
Decrement t rans ien t counter; O ?
YES
Figure
2.9
Procedure ADNT sums
n
ecorded transients into the d ata buffer.
DlSP
I
ai t for short per iod to a l low osci lloscope scan t o
complete and th en for f lybac k
igure
2.10
Procedure DISP displays the contents
of
the data buffer
continuously on the oscilloscope screen.
Copyright © 1990 IOP Publishing Ltd.
http://0.0.0.0/http://0.0.0.0/http://0.0.0.0/http://0.0.0.0/
-
8/17/2019 UT 03 TMI Signal Recovery
33/238
-
8/17/2019 UT 03 TMI Signal Recovery
34/238
26
M U L T I P L E T I M E A V ER A GI NG A N D D R I F T
adjustment an d a restart, the o per ato r is not m ade aw are of this until the full
sequence of n, transients has been accumulated.
One way to obtain continuous monitoring is to run the display routine
interleaved w ith the logging rou tine, on some so rt of interrupt basis. The n for
the present arrangement the eight most significant bits from the 16-bit data
buffer word are ou tput to the
DAC
for display. Th is has th e effect of dividing the
displayed value by 2', the presen t value of n,. Then, as the sequence of 2'
transien ts builds u p in the buffer, the displayed trans ient is seen to grow from
zero to the full range of the DAC. Because the signal amplitude increases directly
with the number of transients
n,
summed to da te, while the noise am plitude
increases with n:I2, the signal appears t o grow o ut of the less rapidly increasing
noise.
The above statement concerning the increase in noise should perhaps be
justified.
In
42.1
it is shown th at , when
n,
transients are averaged, the stan da rd
deviation for th e averaged white noise is reduced by
n [ l 2 .
Thus, for the sum
formed prior to averaging, the noise amplitude will be proportional to n: .
A further advantage of continuous m onitoring is tha t if, during
a
successful
run, it is observed th at an adequately noise-free signal has been ac cumulated
before the end of the sequence of n, transients, it is possible to save time by
stopping the logging sequence early. At this stage the usual display-only mod e
will be en tered, and it may then be necessary to increase the effective gain of the
display, because fewer than the usual number of transients will have been
accumulated. The simplest method is to output to the DAC, not the most
significant eight bits of the 16-bit da ta buffer word , but a suitable sequence of
bits of lower significance.
Continuous time averaging
With o r without continuo us display, the above method of time averaging still
has the property that averaging is carried out over a given, usually
predetermined, period and then has t o sto p, for fear of overflowing the da ta
buffer. In some instances a mode of averaging more akin to the running
average formed by a low-pass filter is more suitable. Here the displayed
transient would represent the average of the last n, transients t o be recorded.
This would have the dua l advantage of giving a continuously updated account
of the signal transient, and also of not overflowing the data buffer.
T he low-pass filter is no t suitable for multiple time averaging in its norm al
form, because the values tha t it averages are consecutive in time. F or multiple
time averaging, in con tras t, the averaged values are taken fram points widely
separated in time, with the values for any one data buffer location all taken
from different input transients.
If, however, a method can be found for making the computer perform the
basic function of the low-pass filter, then the inherent flexibility of the
computer should make it an easy matter to adapt the process to the
requirements of multiple time averaging.
Copyright © 1990 IOP Publishing Ltd.
-
8/17/2019 UT 03 TMI Signal Recovery
35/238
MTA BY
COMPUTER
27
Figure 2.12 Simple low-pass filter to be simulated
by
computer.
We therefore sta rt by determining how the com puter can be made to carry
ou t the fu nction of the low-pass filter shown in figure 2.12. F o r the filter show n,
dudd t C and
i uin- ,) /R, so
that
(2.3)
where T,
CR.
Th us, if the in itial value of
U ,
is know n, the subsequen t values are determined
by the input function uinand equation (2.3).
F o r the compu ter to carry o ut this function, the simple C R circuit must be
replaced by the ADC, the computer, and a DAC. The ADC converts the analogue
inpu t voltage uin o digital form, the com puter executes the process of equation
(2.3) an d the DAC converts the result int o the ou tpu t voltage
U .
Suppose th at u n is digitised at regular brief intervals of length
A t .
Then, to
reflect th e process of equa tion (2.3), the dig ital num ber representing
U ,
must
increase at each interval by the amount
A V ,
given by
2.4)
(2.5)
du,/dt vi,,- ,)/T,
AUJAt vi,,-
, /T,.
U ,
+ U , + vin- , ) AtlT,.
This implies the algorithm
Here T, is the period over which both
C R
filter and comp uter average uin .
F o r the co mpu ter, this will equal
n, At,
where
n,
is the num ber of digitised input
values averaged. Thus the algorithm can be expressed in the form
uo o
+
uin- o)/nt*
(2.6)
Ad aptation t o multiple time averaging is simple; all that is necessary is to
replace the simple summing algorithm in the present program by that of
equa tion (2.6). Here, uin represents the value output by the ADC, and
U ,
the
current value of the corresponding word in the data buffer.
For the present arrangement the most suitable way of executing the
continuously averaging algorithm would be as follows. First the 8-bit input
representing
u n
is regarded as the
8
most significant bits of a 16-bit word , fo r
which the remaining
8
bits are zero. Then the stored 16-bit value representing
U ,
is sub tracted from this value. Next the division by n,,here 2*, is performed by
shifting the 16-bit differencedow n by eight bits. The n the divided difference is
added to the stored value for U , .
If this metho d is used, an d the eight most significant bits of the d at a bu ffer
Copyright © 1990 IOP Publishing Ltd.
-
8/17/2019 UT 03 TMI Signal Recovery
36/238
28 MULTIPLE TIME AVERAGING AND DRIFT
are output to the DAC as before, then the initial growth of the displayed signal
will be the sam e as for the simple summ ing algorithm. H owever, when the 16-
bit numbers representing the values
U
of the stored transient begin to
ap pro ach the effective 16-bit num bers representing t he inp ut transient
u in ,
he
rate
of
growth decreases. The rise then becomes exponential and eventually
terminates when the magnitude of the store d transient becomes equal to that
of
the input transient. Thereafter the signal transient becomes largely stable,
reflecting only changes th at are sustained for
a
period comparable with the
time taken to record n, transients, as for example m ight occur if the am plitude
of
the signal transient was gradually increasing due to some deterioration in
the structure
of
the vibrating beam.
Throughout, the noise amplitude will remain at the reduced level
corresponding to the averaging of
n,
samples.
Copyright © 1990 IOP Publishing Ltd.
-
8/17/2019 UT 03 TMI Signal Recovery
37/238
3
Phase sensitive detector methods
It has been stated previously th at the phase-sensitive de tecto r
PSD)
method is a
useful altemative to multiple time averaging
MTA)
for avoiding errors due t o
drift and
l/f
noise, and thus th at either can be used to prevent these effects from
frustrating the low white noise error obtainable by using a long period of
measurement. Th e principal object of the last cha pt er was t o show th at MTA
can be used to reject drift in this way. In this chapter the drift rejecting
properties of the
PSD
system are explained. The discussion of the ability of
either system to reject
l/f
noise is considered later, for the PSD in chap ter 7.
Th e present chapter en ds with a nu mber of miscellaneous app lications of the
PSD method.
3.1
ADAPTATION
OF
RESISTOR BRIDGE STRAIN GAUGE
The example used t o introduce the PSD method is the usu al on e of the resistor
bridge stra in gauge of figure 1.1, an d figure
3.1
shows this ad apted to the
PSD
method. H ere the DC supp ly to the bridge is replaced by th e AC supply eg and
trans form er shown in figure 3.1. Figures 3.1 ( b ) , c) and e ) hen show how the
strain
s
is converted to the AC signal U at the amplifier output.
The phase-sensitive detector then dem odu lates the signal to give the voltage
up
in figure 3. l( f) . Th e low-pass filter then extracts the slowly varying m ean,
which is displayed as
U .
Th e filter operates by forming the runnin g average of
up over the period T,of the filter and this requires th at T ,+ f ; , where
f,,
s the
signal frequency.
Notice tha t
U
K
s
the strain . W ere the simple full-wave rectifier of figure
3.2
to be used in place of the
PSD,
the o utp ut would be as in figure
3.1 9)
and
U
would become a distorted version of
s.
Copyright © 1990 IOP Publishing Ltd.
http://0.0.0.0/http://0.0.0.0/http://0.0.0.0/http://0.0.0.0/http://0.0.0.0/http://0.0.0.0/
-
8/17/2019 UT 03 TMI Signal Recovery
38/238
30
PHASE-SENSITIVE
DETECTOR METHODS
U )
Isolat ing
transformer Phase-sensitive
detector
I
-- - - - - - - -
Adjustable Switch
phase driver
--+=-
ine wave sh if te r
signal generator
(frequency fol
Figure 3 1 a )
The phase-sensitive detector method as applied to the
resistor bridge strain ga uge of figure
1.1:
system diagram.
The p hase shifter in figure 3.1 is needed t o ensu re that the switching of the
PSD is correctly phased , as in figure 3.1 d ) . This is to correct for phase erro rs in
the transformer, etc, which, in som e cases, are so small as to make the phase
shifter unnecessary.
It is actually unlikely that the
PSD
switch will be mechanical as shown,
altho ugh a relay is sometimes used. The use of som e sort of tran sistor switch is
much more likely.
3.2 AC COUPLED AMPLIFIER
There are two ways in which the
PSD
system rejects drift: one is tha t a n
AC
coup led amplifier can be used in place of the usua l
DC
coup led signal amplifier
and the other
is
that the PSD itself tends t o reject drift. In this section th e drift-
rejecting p roperties of the AC coupled amplifier are considered.
Figure 3.1 shows the AC coup led amplifier replacing the
DC
coup led amplifier
of figure 1.1. Figure 3.3 show s the essential features of the
A C
coup led am plifier.
Here th e single-stage transis tor amplifiers are sepa rated by simple high-pass
C R filter sections, and the arrangement is terminated with a unity-gain
isolator to drive the
PSD
The exact arrangement would now be a little ou t of
da te a nd a feedback amplifier section, such as th at in figure 3.3(b), would
probab ly replace each single stage
of
figure 3.3(a). However, the response is the
same, and for ou r present p urpose it is convenient t o analyse system a).
offs t
The first step is to show that the AC coupled amplifier rejects offset. One way of
showing this is to consider the response of one of the C R networks to a step
Copyright © 1990 IOP Publishing Ltd.
http://0.0.0.0/http://0.0.0.0/
-
8/17/2019 UT 03 TMI Signal Recovery
39/238
-
8/17/2019 UT 03 TMI Signal Recovery
40/238
32 P HAS E- SENS ITIVE DETEC TOR M ET HOD S
:i l te r
ow-pass
Figure
3.2
Full-wave rectifier
Single- stage
t rans istor ampl i f ie r
\
A C coupling
O P
amp
Figure 3.3 AC coupled amplifier circuits. a) Cascaded single-stage (time
constant =
C R )
ransisto r amplifiers, ( b ) eedback amplifier (gain =
R J R , ,
t ime constant =
C R , ) .
applied t o the input. Figure 3.4 shows the s tep
( b )
applied to th e network a )
giving the response c). Here the o utp ut responds initially in full to th e inpu t
change, but then decays to zero a s the capacitor C charges thro ug h the resistor
R . Thu s, after the initial transient, the response t o DC applied to the input is
zero.
Square wave signal response
It is important to show that the AC coupling networks d o not attenuate the
required signal. De pend ing up on the way in which the system is ad ap ted , this
can be a square wave or a sine wave, simply depending upon whether the
generator
eg
in figure 3.1 is a sq uare o r a sine wave generator. Figure 3 5 shows
the effect when a square-wave signal is applied to one of the C R sections of
figure 3.3.It is clear that, to avoid atten uati on , we require tha t T, To,where To
is the signal period and T, is the filter time constant C R .
Copyright © 1990 IOP Publishing Ltd.
http://0.0.0.0/http://0.0.0.0/http://0.0.0.0/http://0.0.0.0/http://0.0.0.0/http://0.0.0.0/http://0.0.0.0/http://0.0.0.0/http://0.0.0.0/http://0.0.0.0/http://0.0.0.0/
-
8/17/2019 UT 03 TMI Signal Recovery
41/238
AC COUPLED AMPLIFIER
33
t ”In
bl
Figure 3.4 Step response of A C coupling element in
figure 3.3 a). a )
A C
coupling network,
( b )
nput step,
(c)
output response
U,=
Vexp(
-t/q .
Figure
3.5
Response of A C coupling element in
figure 3.3 a)
to a square
wave of period
To
omewhat less than the filter response time T,. a) nput,
( b )
output.
Drift response
In this chap ter we shall consider mainly th e ability of the AC coupled amplifier
to reject drift of cons tant rate. Drift of varying rate requires th e spectra l view of
the next chapter and is considered later, in chapter
5.
To demonstrate the constant-rate drift response, we first consider the
response of one
C R
section to the input signal of figure
3.6 a).
This is a
constant-rate ramp star ting a bruptly a t t ime t
=O
Consider first the ‘steady-
state’ response at a time well afte r the initial transient. Here the o ut pu t U is
con stant. Thus, for the capacitor voltage u c , duJdt will eq ua l du,Jdt. But the
capacitor current i
=C
duJdt,
so
i =
C
dui dt. Since U
=
iR, one can write
U = C R
dui,Jdt. But du,Jdt has the con sta nt value
V’ s
that
U
has the
constant value CRV’, as shown.
Copyright © 1990 IOP Publishing Ltd.
http://0.0.0.0/http://0.0.0.0/http://0.0.0.0/http://0.0.0.0/http://0.0.0.0/http://0.0.0.0/
-
8/17/2019 UT 03 TMI Signal Recovery
42/238
34 PHASE-SENSITIVE DETECTOR
METHODS
,
/
Figure
3.6 Response of
AC
coupling element
of
figure
3.3 a)
to a ramp
input .
a)
Input , b ) output .
It ap pe ars then th at a single high-pass filter section does not entirely reject
drift. The drift is converted to a steady output. This, of course, is easily
eliminated by add ing an ot he r high-pass filter section. This is why a typical
A C
coupled amplifier will have a t least two high-pass section s; the arrangem ent of
figure
3.3 a)
has three.
3.3
DRIm
REJECTION OF PSD
Although the
AC
coupled amplifier commonly used with the
PSD
system has
drift rejecting properties, as explained above, the use of the AC coupled
amplifier is not actually fundamental to the technique, because the
PSD
itself
tends to reject drift, and offset, an d indeed l/f noise. It is the pu rpose of this
section to show how the PSD rejects offset an d drift.
Figure 3.7 illustrates the ab ove point. H ere U represents drift and offset at
the amplifier ou tp ut , i.e. the PSD input. This slowly varying signal is converted
to a square wave
c)
with a corresponding slow variation in am plitude. The n,
because the frequency of the square wave is the signal frequency
fo,
and
because
fo f,,
where
f ,
is the cut-off frequency of the low-pass filter that
follows the PSD, the filter virtually rejects the square wave, leaving only the
small residual component d ) .
PSD imperfections
Th ere are two kinds of imperfection in a real PSD circuit which imp air its ability
to reject drift a nd offset, an d therefore mak e it desirable t o retain the AC
coupled amplifier. The first is th at the magnitudes of the g;in in the forward
an d reversed states of the switch may n ot be exactly equal. This causes
a
small
DC component at the PSD output which, unlike the output square wave
com ponen t resulting from the offset, will be accepted by the low-pass filter.
Then, naturally, if the input offset component drifts, so also will the small DC
component at the output.
Th e oth er potential imperfection of the real PSD circuit is th at , in general, the
Copyright © 1990 IOP Publishing Ltd.
http://0.0.0.0/http://0.0.0.0/http://0.0.0.0/http://0.0.0.0/http://0.0.0.0/http://0.0.0.0/http://0.0.0.0/http://0.0.0.0/
-
8/17/2019 UT 03 TMI Signal Recovery
43/238
WHITE NOISE ERROR 35
Figure
3.7
Waveforms showing how the PSD circuit of figure
3.1
avoids the
effects of drift and offset. a) Drift an d offset a t amplifier ou tpu t, ( b )
PSD
switch states, c )
PSD
output up d ) ow-pass filter output uo .
periods for the forward a nd reversed states of the switch will no t be exactly the
same. This will have essentially the same effect as when the gains fo r the tw o
states are unequal.
3 4 WHITE NOISE ERROR
It has now been shown that when a system is adapted to phase-sensitive
detection, drift is removed. It remains to be shown t ha t the adap tation gives no
increase in white noise error.
Figure 3.8 allows this point to be dem onstrated.
a )
how s the essentials of
the resistor bridge strain gauge of figure
1.1
and
( b )
hows an adaptat ion to
phase-sensitive detection. Here the sine-wave generator eg of figure 3.1 is
replaced by the DC source E and the reversing switch of figure 3.8 b). This
constitutes a square-wave generator a nd allows a fair comparison of a ) and
( b ) ,
because the same primary power source
E
is used in both cases.
The first point t o note is that the signal outp ut for the two systems, U in a )
and u p in ( b )are the sam e. This is because the two reversing switches in ( b )
cancel each other. The w aveforms in c)- f) confirm the po int. In c)there is an
output com ponent
uaa.
This is the signal com ponen t resulting from the strain
which unbalances the resistor bridge.
( (f)
then show tha t the PSD system
output
u p
is the same, i.e. up
= U
Waveforms 9) to j ) now show the result of adding noise. Here the
fluctuation time is determined by the response time of the signal amplifier. As
before, the signal components U are the same at each output. The noise,
Copyright © 1990 IOP Publishing Ltd.
http://0.0.0.0/http://0.0.0.0/http://0.0.0.0/http://0.0.0.0/http://0.0.0.0/http://0.0.0.0/
-
8/17/2019 UT 03 TMI Signal Recovery
44/238
36 PHASE-SENSITIVE
DETECTOR
METHODS
I
I
I
€ g a y
Amplif ier
I
I
I
I
Figure
3.8
Diagram showing how adapting the resistor bridge strain gauge
to the PSD method leaves the system response to both signal and white
noise unaltered.
a)
Simple resistor bridge strain gauge, ( b )adaptation to
PSD. c )and (g), amplifier output for a); d ) and
( h ) ,
amplifier output for
b), e) , i) PS D switch states, ( f ) , j ) PSD output.
however, differs slightly. T hat at the
PSD
ou tp ut is subject t o periodic reversals,
due to the
PSD
reversing sw itch. Fo r either system th e o ut pu t noise is passed
thr ou gh the final low-pass filter of time constan t T,(not show n). It is fairly clear
th at the noise am plitud e after this final filtering will not be influenced by these
periodic reversals, and so the final noise amplitude is the same for both
systems. This , in fact, is so an d is argued more rigorously from the frequency-
domain viewpoint in chapter
7 .
Copyright © 1990 IOP Publishing Ltd.
-
8/17/2019 UT 03 TMI Signal Recovery
45/238
MISCELLANEOUS APPLICATIONS
37
It has now been established that both multiple time averaging and the
phase-sensitive detector method are effective in removing drift and offset
without increasing white noise error. Th us either can be used, in principle, to
allow the low white noise er ro r associated with
a
large measu rement time t o be
realised. In practice, however, the choice between the two m ay be limited. F o r
example, to use
MTA
to reduce drift when measuring a stress-strain curve
requires making the stress scan time
T,
small. In practice
a
sufficiently low
value of
T,
may not be possible.
3 5 MISCELLANEOUS APPLICATIONS
Th e arrangem ent of figure
3.1
is no t th e total solution to drift an d offset for the
resistor bridge strain gauge. While amplifier drift and offset are eliminated,
transducer drift and offset remain: if the bridge is initially unbalanced this
constitutes transducer offset; an d when the imbalance changes-say du e t o
therm al changes in the relative values of the bridge resistors-this cons titutes
transducer drift.
A possible,
if
far-fetched, solution can be found for the stress-strain curve
measu rement. Here the original
DC
supply is used to drive the resistor bridge,
instead of the
A C
supply of figure 3.1. Also, instead of scanning the applied
stress uniformly as in figure 3.9 a), the stepped scan of
( b )
is used. This
produces an A C signal compone nt of the frequency
o
shown, which is then
processed by the
AC
coupled amplifier and PSD in the usual way. For this
arrangem ent, in con tras t with th at of figure
3.1,
the transducer drift an d offset
is not modulated to produce an
AC
signal component, and thus is rejected.
a ) Normal
scan
\
' b f p 4
Figure 3.9 Stress modulation to avoid transducer offset and drift in
stress-strain curv e me asure men t.
a )
Norm al scan ,
( b )
modulated scan.
F o r this particular application, the idea is impractical, because th e inertia of
the testing machine would severely restrict the value of fo. How ever, the idea is
sound in principle a nd serves t o illustrate just on e of the grea t variety
of
ways
in which the
PSD
method can be refined to allow rejection of various kinds of
unwanted response.
Copyright © 1990 IOP Publishing Ltd.
http://0.0.0.0/http://0.0.0.0/http://0.0.0.0/http://0.0.0.0/
-
8/17/2019 UT 03 TMI Signal Recovery
46/238
38 PHASE-SENSITIVE DETECTOR METHODS
It is not really possible to lay dow n g eneral principles beyond this point.
Each system w ill present its own problems, an d it is a m att er for the ingenuity
of the experimenter t o devise a m ethod which will mo dulate the variable tha t it
is wished to measure, while leaving un mo dulated all unw anted effects. Th us we
finish by giving a few rather more practical examples of the kind of strategy
that is adopted.
Choppers
Figure 3.10 shows on e of the simplest metho ds for converting a DC signal into
an AC signal suitable for AC coupled amplification and phase-sensitive
detection. Here the vibrating switch contacts chop th e signal to produc e a
square wave of amplitude equal to the initial slowly varying
DC
voltage.
0
Chopped
output
S ~ O W I ~
arying
signal source
supply
Figure 3.10 Relay-ope rated cho ppin g circuit.
a )
Circuit, ( b )signals, (c) switch states
0
op en, c =closed).
Disadvantages of the arrangement are the obvious problems of wear
associated with a n electromechanical system, an d also th e thermoelectric EM F
th at is generated across the con tacts, even in the absence of a signal.
Both