Novel Sensors and Sensing

Chapter 1
The instrumentation system

1.1 Introduction
Whether mankind is simply another part of the animal world or is somehow
unique is a question that has occupied the minds of philosophers for genera-
tions. The human attributes of self-awareness, empathy, morality and what
constitutes consciousness and creativity have become part of the current
debate surrounding the development of machine intelligence. What is
beyond question is the insatiable thirst mankind has to discover more
about the individual self and the universe around. Observations for scientic
purpose have been made since the beginnings of history and instruments to
facilitate this have existed for millennia. However, during the latter half of
the 20th century the growth in instrumentation technology, and in con-
junction control theory, has been phenomenal.
For a nations instrumentation industry to remain at the fore in world
markets sensor development is essential. In addition to the general technolo-
gical development of traditional sensors, during the past twenty years a
number of specic elds have emerged centred on novel sensing techniques
and materials. The global market for all sensor types is estimated at 15
30 billion per annum. The current growth in sales of traditional sensors is
around 4% per annum compared with that of 10% per annum for novel
sensors. (Source: UK Foresight Programme.)
Initially, the Japanese were the leaders at commercially exploiting
academic innovations. However, the established industrial countries of the
west are now rmly committed, both at governmental as well as company
level, to strategic programmes for sensor development. Key elds are
regularly identied by funding agencies in order to stimulate research and
to assist in commercial exploitation.
The benets of improvements made in sensors will enhance the quality
and performance of industry and other elds. For example, it has been said
that advances in process automation are presently limited to the quality of
Copyright 2004 IOP Publishing Ltd
the instrumentation and in particular the plantinstrument interface. In
medicine, novel developments occurring in diagnostic instrumentation will
reduce response time and costs, and increase throughput. Non-invasive
sensors and invasive microsensors can reduce patient trauma. In the eld
of transportation, modern vehicles contain large numbers of sensors for
safety, engine eciency and global location systems. Medical, environmental
and automotive applications have the largest potential market growth.
In any new device it is important that the specication or transfer char-
acteristic should be well dened and stable. This is often the most dicult
problem to resolve when going from laboratory prototype to commercial
product. Many novel microsensor-prototypes have been reported measuring,
for example, position, ow, humidity, acceleration, liquid level, ionic
concentration, temperature, pressure and dissolved oxygen. However, the
percentage of devices that has become commercially available is relatively
small. Nevertheless, many of those that have enjoy large volume production
which is encouraging the general growth of the novel sensor industry.
It is important to realize that to be commercially successful a device
must be able to meet the cost constraints of the market for which it is
Table 1.1. Acceptable sensor parameters in some common elds.
Domestic Industrial Medical Automotive Environmental
Acceptable
cost ()
1 500 Disposable 10,
non-disposable
250
5 Disposable 10,
non-disposable
250
Acceptable
error (%)
10.0 0.15.0 0.12.0 1.05.0 1.010.0
Mean time
between
failures
(hours)
10
3
during
10 years of
intermittent
use.
10
5
during
10 years of
continuous
use.
Disposable 150
during 1 week
of continuous
use, non-
disposable
5 10
3
during 5
years of
intermittent
use.
1.5 10
4
during 10
years of
intermittent
use.
Disposable 150
during 1 week
of continuous
use, non-
disposable 10
4
during 10 years
of intermittent
use.
Temperature
range (8C)
20 to 400 200 to
1500
Disposable 10
to 50, non-
disposable 0 to
250
30 to 400 30 to 100
Response
time (secs)
1 0.110 1 10
3
1 100
2 The instrumentation system
intended. Also, some elds of application require a higher specication than
do others. Table 1.1 indicates acceptable costs and other desired parameters
in some common elds.
Summarizing, it is important that for any newly developed sensor to
become commercially successful it must not only be eectively ruggedized,
when going from the laboratory prototype to marketed product, but also t
within the constraints of the application eld.
1.2 The philosophy of measurement
Lord Kelvin is reputed to have said,
When you can measure what you are speaking about and express it
in numbers, you know something about it; but when you cannot
measure it, when you cannot express it in numbers, your knowledge
is of a meagre and unsatisfactory kind.
Experimental observations are the basis on which our understanding of the
universe around us develops. In science and engineering all hypotheses
and theories before they develop into rules and laws have to be rigorously
examined to discover their validity and its extent. Universal constants and
parameters relating to materials and structures and other artefacts all have
to be painstakingly examined to produce the most reliable data possible. A
good example is the measurement of the speed of light, which has taxed
some of the cleverest experimenters for centuries and for one man in particu-
lar, A A Michelson (18521931), took up a large proportion of his life
earning him the Nobel Prize in 1907.
The philosophy of measurement is that of striving to reach some
notional real or true value. The more eort made the greater the precision
of the measurement and the smaller the uncertainty.
1.2.1 Random errors
Whenever the measurement of a parameter or variable is made an error will
occur. The measured value or measurand will have an associated uncertainty
due to the natural uctuation of real physical quantities. Gauss postulated
that such variations are as equally likely to be above the mean as below
it. This random error may not be noticed if the measurement system is
insensitive, and may not be of any concern in some applications.
However, at the microscopic scale nature is never still and highly sensitive
instruments will reveal such uctuations. The instrumentation system is
part of the physical world and will itself introduce an observable randomness
to any measurements taken. All such random errors have to be treated
statistically.
The philosophy of measurement 3
The best estimate of the true value X of some variable or parameter x is
the mean " xx. In the usual way,
" xx
P
N
i 1
x
i
N
1.1
from N observations. It is expected that for large N, " xx !X. Incidentally,
when calculating the mean it is argued that the numerical value obtained
from a large number of observations can be rounded to an extra signicant
gure more than that of the precision of the equipment used.
A measure of the spread or dispersion of the results around the mean is
very useful as an indication of the repeatability of the measurement. The
deviation of a single result from the mean is x
i
" xx. Because the deviations
can be positive or negative the sum of the deviations is likely to be zero for
a large set of values. Consequently, the square of the deviations is used.
The mean of the squared deviations is called the variance. To obtain a
measure of the dispersion in the same units as the original measurements
the square root of the variance must be taken. This is the standard deviation.
It is simply determined using
s
P
N
i 1
x
i
" xx
2
N
s
. 1.2
The standard deviation is also known as the standard error. It is a measure of
the randomness or uctuations in the results obtained. It is independent of N,
the number of observations made. As can be seen in equation (1.2) making
many, repeated observations does not reduce the value of s but simply
gives a greater condence in its estimate.
Intuitively it is felt that taking more measurements ought to improve the
quality of the mean in some quantiable way, more must be better. To
provide this extra quantier imagine that instead of one set of N results
the measurement set was repeated p times under identical conditions, each
set having N results. Each of the p sets of results could produce its own
mean " xx and standard deviation as before. But now these p means could be
averaged to give the mean of the means
"
" xx" xx:
"
" xx" xx
P
p
k1
" xx
k
p

1
p
X
p
k1
1
N
X
N
i 1
x
i
. 1.3
This does not look anything of signicance. After all, it is the same result as
would have been obtained having made pN observations initially:
"
" xx" xx " xx
P
pN
i 1
x
i
pN
. 1.4
The most signicant and useful result is obtained from the standard devia-
tion of the p means. This standard deviation s
m
is a measure of the spread
of the means and indicates the precision of the value of the mean of the
means. In principle this could be determined using
s
m

P
p
k1
" xx
k
"
" xx" xx
2
p
s
. 1.5
In practice equation (1.5) is not used. It can be shown (see Squires 1968) that
s
m

s
N 1
p
P
N
i 1
x
i
" xx
2
NN 1
s
. 1.6
So for large N,
s
m
!
P
N
i 1
x
i
" xx
2
N
2
s
. 1.7
As equation (1.7) indicates, the estimate of s
m
can be seen to improve by
increasing the number of observations made. More is better is correct
here. However, this rule suers from diminishing returns. Large values of
N may not be worth the eort. To appreciate the dierent meanings of s
and s
m
the following examples may be helpful.
Standard error of the measurements (s)
Consider the manufacture of 10 k resistors on a production line. Each
day 100 samples could be taken from the production run and the value of
s determined. This denotes the spread of the results about the mean and
indicates to the production manager whether the batch is within the tolerance
limits specied on the packets, e.g. 10%. The probability of a single resistor
being outside of the 10% is related to s. In section 6.3.2 this is described
further within the topic on probability distributions. For a value to be
more than three standard errors from the mean is statistically unlikely
at about 2.7 in 10
3
. So, if s is arranged to be about 3%, by using good
production technology, then most resistors produced will be within 10% of
the marketed value of 10 k.
This 3s concept can also be applied to identify an erroneous result. If a
reading from a set of identical measurements is more than three standard
errors from the mean of the set it can be considered dubious and be dis-
carded.
Standard error in the mean (s
m
)
In experiments where some parameter is to be determined, such as the speed
of light, the value of s
m
is quoted. This provides an estimate of the precision
of the published value and incorporates the eort made in producing the
repeated sets of measurements. There have been occasions when researchers
have repeated experiments and obtained results outside of the standard error
quoted originally. This has led to the discovery of systematic errors.
Combining standard deviations
If a parameter or variable Q is to be determined from measurements of two
or more parameters or variables X, Y, Z each having their own standard
deviations s
X
, s
Y
, s
Z
, it will be necessary to determine the standard deviation
of Q, s
Q
, by combining these contributing standard errors. It can be shown
that if Q is a function f X. Y. Z then
s
2
Q

0f
0x
2
s
2
X

0f
0y
2
s
2
Y

0f
0z
2
s
2
Z
. 1.8
Table 1.2 shows how to combine the standard errors for some common
functions.
1.2.2 Systematic errors
A further error may occur due to poor calibration or poor performance in the
instrumentation system. Because such errors are inherent in the system and
are transmitted through to the measured value they are termed systematic
errors. They cannot be overcome by the statistical analysis of multiple
readings but are minimized by diligence in calibration and system design.
1.2.3 Environmental disturbances
The engineering environment within which the instrumentation system is
integrated can aect the inputoutput relationship of the sensor or of the
instrumentation system as a whole. The engineering environment includes
Table 1.2. Combination of standard errors.
The function f X. Y. Z Combination of standard errors
Q X Y Z or Q X Y Z s
2
Q
s
2
X
s
2
Y
s
2
Z
Q XYZ or Q
XZ
Y
s
Q
Q
s
X
X
s
Y
Y
s
Z
Z
2
Q X
n
s
Q
Q
n
s
X
X
Q lnX s
Q

s
X
X
Q expX
s
Q
Q
s
X
variables such as the ambient temperature and humidity and power supply
behaviour. There are two common ways in which the environment can
aect the response and are described below for an idealized, linear sensor
or system.
Zero-shift or the interfering eect
This disturbance has the eect of changing the intercept value of the straight-
line graph, shown in gure 1.1. The interfering input is the change in or
deviation of some environmental parameter. This input couples into the
system via a sensitivity constant.
For example, if a drift in the ambient temperature of 5 8 C caused a
pressure measurement reading to change by 150 Pa then the system would
have an environmental sensitivity of 30 Pa K
1
.
Variation in sensitivity or the modifying eect
This environmental disturbance modies the value of the system sensitivity,
the gradient of the graph (gure 1.2). An example of this eect could be of the
oat-type liquid level gauge as commonly used in automobiles. A potenti-
ometer fastened to the oat provides a variable resistance related to the
level of the liquid, which in turn alters the current through the milliammeter
petrol gauge. If the supply voltage alters then so does the reading. Typically,
when the alternator and regulator are operating, an automobile power
supply is held steady at 14 V. However, when the engine is stopped the
power supply system is reduced to that of the 12 V battery. An environmental
sensitivity constant could be determined in gallons mA
1
/V.
A simple way to monitor whether there has been some environmental
disturbance is to repeat the rst measurement taken after the completion
Figure 1.1. The interfering eect.
of a set of readings. More sophisticated methods can be devised. For
example, for a linear system denoted by y mx c the plotting of y mx
versus x should yield a xed value if c is constant (no interfering eect).
Similarly, plotting y c,x versus x should give a constant if m is xed
(no modifying eect).
1.2.4 System design and the contribution of errors
It was shown in section 1.2.1 how the standard errors of a number of random
variables can be combined to give the standard error in some derived quan-
tity Q. An appreciation of the way errors combine is also useful when an
experiment or system is being designed. It can be instructive to see how
errors in the component measurands of the quantity to be determined and
any additional system element errors contribute to the nal result. In this
way more care can be taken over those components that contribute most
to the global error.
As previously, suppose a quantity Q is to be determined from observa-
tions of one or more variables. In addition, other elements in the instrumen-
tation system such as a lter or an amplier may also introduce errors.
Again, Q is some function f X. Y. Z where X and Y might be the measur-
ands and Z the gain of an amplier. The values of X, Y and Z have associated
uncertainties cX, cY, cZ which combine to produce an uncertainty cQ in Q.
The relationship between cQ and cX, cY, cZ can be found using the partial
dierential equation
cQ
0f
0X
cX
0f
0Y
cY
0f
0Z
cZ. 1.9
Table 1.3 gives examples of some common functions.
Figure 1.2. The modifying eect.
Addition/subtraction
Unless there is some a priori knowledge about the system behaviour, the
worst case is assumed and all the absolute errors are added. The error
estimates must all be in the same units and cQ will also have these units.
Multiplication/division
Here, again assuming the worst case, all the relative errors are added. They
have no units and can be described as fractional errors or multiplied by
100% to give percentage errors. The relative error is also used with the loga-
rithmic function.
Simple determination of the global error
The relationships in table 1.3 or others derived from equation (1.9) can also
be used to make a simple estimate of the error in a derived quantity, whether
determined from a number of measurands, or as the output from an instru-
mentation system. These contributory errors might arise from
.
a rough guess at the random errors,
.
estimated systematic errors,
.
calibration limits,
.
component tolerances,
.
estimated likely environmental disturbance,
.
the resolution limits,
.
hysteresis or dead band eects,
.
estimated nonlinearities,
.
digitization errors.
In the paragraph preceding equation (1.9) the word uncertainty was used
when describing cX, cY, cZ. When dealing with the quantities listed above
Table 1.3. Simple combination of errors.
The function f X. Y. Z Combination of standard errors
Q X Y Z or Q X Y Z cQ cX cY cZ
Q XYZ or Q
XZ
Y
cQ
Q

cX
X

cY
Y

cZ
Z
Q X
n
c
Q
Q
n
c
X
X
Q lnX c
Q

c
X
X
Q expX
c
Q
Q
c
X
this is a preferred expression to error since it is all encompassing. Items such
as calibration limits and tolerances are then included which, strictly speaking,
are not errors.
WORKED EXAMPLE
The density of a material is to be determined from the mass and volume of a
wire sample and a simple method used to estimate the uncertainty in the
answer obtained. The density is derived from
,
m
d,2
2
h
.
From equation (1.9) it can be seen that
c,
,

cm
m
2
cd
d

ch
h
revealing that the fractional uncertainty in the diameter d is the most sensitive
term in the determination of ,. Greater care must therefore be taken with this
measurement.
The mass of the specimen is found to be 1.25 10
3
kg and the balance
used is calibrated to 1 mg. The length is measured as 0.1845 m using a rule
readable to the nearest 0.5 mm. The thickness is measured ten times at
various points and orientations along the length using a micrometer
capable of measuring to 10
5
m and is found to be 0.984 mm with a
standard error in the mean of 0.024 mm. The density is determined as
,
1.25 10
3
9.84 10
4
2
2
0.1845
8.909 10
3
kg m
3
.
The fractional uncertainty is determined,
c,
,

10
6
1.25 10
3
2
0.024
0.984
0.5
184.5
8 10
4
0.04878 2.71 10
3
0.05229.
The uncertainty in , is
c, , 0.05229 465.9 kg m
3
.
It is usual to express the result in the form
, 8.909 0.4659 10
3
kg m
3
.
The question of the number of signicant gures quoted needs to be
considered. A common sense approach must be adopted. In the example
above, considering that the uncertainty is about 5% (mostly from the
diameter error), it would seem meaningless to quote the nal value for , to
more than 2 signicant gures
, 8.9 0.5 10
3
kg m
3
.
The uncertainty or standard error is normally quoted to a single digit or to
two digits if the digit is 1 or 2.
Finally, it is important to realize that this simple method of combining errors
does not have the statistical rigour of that used for combining standard errors
achieved by equation (1.8). This is because the uncertainties here do not belong
to probability distributions but are bands of vagueness. However, on
occasion it may be possible to consider the uncertainty as inherently part of
a gaussian distribution. The uncertainty could then be interpreted as a stan-
dard error and equation (1.8) invoked to combine it with other standard
errors. The advantage of this is that the standard error produced gives a
precise, statistical weighting as to the reliability of the quantity. In section
6.3.2 the method of calculating such probabilities is illustrated.
1.3 The general instrumentation system
The instrumentation system can be divided conveniently into the following
general components. Not all systems contain all components.
Sensor: This is in good contact with the variable to be measured and
some physical property of the sensor is modied by the variable. On
occasions the sensor may be composed of two elements termed
primary and secondary.
Signal conditioning: The output of the sensor may not be in a
convenient format. The signal conditioning circuit will most
commonly be used to provide an output of a few volts.
Signal processing: The few volts of analogue data may require
ltering, linearizing, converting to digital format or preparing for
transmission if a telemetry link is required.
Telemetry link: May be wire, pneumatic, microwave or optical bre
and use AM, FM or pulse code.
Data display: This may be analogue or digital as appropriate for
the application. The ergonomics of data displays canbe of signicance.
Table 1.4 lists some of the relative characteristics of each type.
The general instrumentation system 11
To conclude this section it was felt that a comment on the terminology in
current use concerning sensors would be appropriate. The word transducer,
strictly dened, is a device that converts energy from one form to another. A
microphone could be described as a transducer but so could a loudspeaker.
A sensor might be a transducer but a transducer may not be a sensor.
However, a recent advertisement for a pressure transducer, on inspection,
was found to be a diaphragm with strain gauge and associated signal
conditioning. A power supply was input on two wires and the output on a
further two. This, clearly, is not a transducer in the strict sense.
A further example of misleading terminology is the use in the process
industries of the word transmitter. A pressure transmitter, for example, is a
device that measures the required pressure and outputs the information on
a wire pair usually in the standard industrial range of 420 mA. Such
transmitters are sensor, signal conditioning and signal processing.
It is of paramount importance in science and engineering that terms are
clearly dened and upheld. History is littered with accidents caused by
misunderstandings between engineers. The modern trend is to use the
word sensor for an input device and actuator for the output device attempting
to cause some change to a system or plant. The word transducer is best
avoided. Sensors can be subdivided into active, when energy conversion
occurs producing an electrical signal, and passive when a power supply is
required to obtain an electrical output. For example, a thermocouple is an
active sensor whereas a resistance thermometer a passive sensor. This
distinction can be signicant when dealing with explosive environments
where the energy available within the hazardous area must be strictly
controlled.
1.4 The overall transfer function
The output of the instrumentation system is a product of the input variable
and the various elements within the system. In analysing the overall system
performance it can be useful to mathematically model each element. Finkel-
stein and Watts (1971) proposed this systems approach to instrumentation.
Table 1.4. Analogue versus digital displays.
Advantages Disadvantages
Analogue Continuous scale. Faster
recognition by user
Less easy to read small divisions.
Easier to misread
Digital Cheaper. Faster response More stressful in continuous use.
Wrong value shown if a segment fails
The transfer function for each element is
output
input
.
For example, if the sensor were a resistance temperature element then
G
1

p
input value
8C
1
. 1.10
Individual elements may contain feedback loops but this is incorporated into
the value of G. The overall transfer function is the cascade of the individual
transfer functions
measured value
input value
G
1
G
2
G
3
G
4
G
5
. 1.11
The overall transfer function is the sensitivity of the system. Clearly then
measured value input value G
1
G
2
G
3
G
4
G
5
. 1.12
However, this will not be the case if an element has a zero or an o-set value.
For example, the standard platinum resistance sensor has a resistance of
100.0 at 0 8C. To overcome this the signal-conditioning element is made
to output zero for an input of 100.0 . It is good practice generally to have
o-set or zero adjust facilities early in the system. Failure to do this could
result in large signals later in the system causing saturation of an amplier
or inecient use of the input span of an analogue to digital converter.
Similarly, the expression (1.12) above takes no account of inter-element
loading. Again it is good practice to construct elements with low output
impedance and high input impedance to minimize loading eects. The
availability of high performance operational ampliers for use as buers
makes this requirement easy to achieve. Inter-element loading is described
further in section 1.6.
1.5 Dynamic response of the sensor
The modelling of the sensor itself can be more involved than equation (1.10)
suggests. For example the warming or cooling of a temperature sensor can
be modelled by a rst-order, linear, dierential equation. This means
that because it is not an algebraic expression the transfer function cannot
be simply set down. Generally, elements that involve energy storage
and transfer are described by models involving time-dependent equations.
Consequently, if a variable is changing, the output measured value will
have a dynamic error. To be able to use the cascade concept above for the
whole system it is convenient to Laplace transform the element equations.
This has the eect of turning an expression involving calculus into one
Dynamic response of the sensor 13
involving algebra. The example below illustrates the dynamic response of a
temperature sensor.
Transfer function of a temperature sensor
At the start of the measurement the sensor will have a dierent temperature
from that of the uid in which it is immersed. There must be an energy
transfer between uid and sensor until they are both at the same temperature.
At the start
time t 0. sensor temperature 0
0
. fluid temperature 0
f
.
During the heat transfer process the sensor temperature will be changing
continuously and will have some value 0
s
t. To cause the temperature of
some material to rise a quantity of heat Q must be supplied,
Q mc0
s
0
0
1.13
where m is the mass of the material and c its specic heat capacity. Assuming
that the uid is hotter than the sensor, the rate of gain of heat by the sensor is
dQ
dt
mc
d0
s
0
0
dt
. 1.14
The heat transfer can be by either convection, conduction or radiation. In
this situation the dominant process will be convection. The rate of heat
loss by the liquid is determined using the convection equation
dQ
dt
UA0
f
0
s
1.15
where A is the surface area and U is a geometry factor.
When using Laplace transforms it is more convenient to have zero
starting conditions. This means that all variables in the equation should be
zero at the start. To arrange for this a substitution is made,
0 0
s
0
0
. 1.16
Equation (1.14) now becomes
dQ
dt
mc
d0
dt
1.17
and equation (1.15) becomes
dQ
dt
UA0
f
0 0
0
. 1.18
Equation (1.18) can be written more conveniently using the substitution
0
f
0
f
0
0
1.19
giving
dQ
dt
UA0
f
0. 1.20
Now,
the heat gained by the sensor the heat lost by the fluid.
So, combining equations (1.17) and (1.20) gives
mc
d0
dt
UA0
f
0 1.21
and re-arranging gives
mc
UA
d0
dt
0 0
f
. 1.22
Science and engineering equations must be homogenous in their units. The
coecient in front of the dierential term must therefore have the dimensions
of time. A time constant term t can be introduced where
t
mc
UA
. 1.23
Substituting t into equation (1.22) gives
t
d0
dt
0 0
f
. 1.24
This can now be conveniently transformed using Laplace tables to give
ts
"
00s
"
00s
"
00
f
s 1.25
"
00sts 1
"
00
f
s 1.26
"
00s
"
00
f
s
1 st
. 1.27
So the Laplace transformed transfer function of the sensor is
"
00s
"
00
f
s
1
1 st
. 1.28
Control engineers commonly use the systems approach in design and, con-
ventionally, express transfer functions in Laplace transformed format.
Upper case symbols are used to signify Laplace format so equation (1.28)
would be given the symbol
G
1
s

"
00s
"
00
f
s
1
1 st
. 1.29
Dynamic response of the sensor 15
1.6 The measurement system as a series of networks
Finkelstein and Watts (1971) proposed that elements in an instrumentation
system could be described as two-port networks. Two-port network analysis
is a technique whereby an electrical circuit, or part of a circuit, is treated as a
generalized element having an input and an output port. The input/output
variables are generally referred to as across (potential dierence) and
through (current) (Kraus 1991).
Of the four variables two are chosen as independent and the others
become the dependent. Depending on the choices made six canonical forms
can be produced relating the inputoutput variables. A set of equations is
derived, using a particular canon, which provides the generalized description.
These equations may involve integral-dierential relationships. Of the six
canonical forms the impedance parameter equations are the most commonly
used. For gure 1.3 these are
V
1
z
11
I
1
z
12
I
2
1.30
and
V
2
z
21
I
1
z
22
I
2
. 1.31
These can be conveniently expressed in matrix form as
V
1
V
2
I
1
I
2
z
11
z
21
z
12
z
22
I
1
I
2
1.32
where Z is termed the impedance parameter matrix. Having described the
elements in a system as sets of two-port networks with associated linear equa-
tions, matrix algebra or software tools such as Matlab
1
can be conveniently
used to determine the overall system response.
Finkelstein and Watts suggested that the process to which the measured
variable belongs could be treated as a one-port network. The process must
have appropriately dened across y and through _ xx variables.
Figure 1.3. The two-port network.
Figure 1.4. The process as a one-port network.
Bentley (1995) describes this approach in more detail. Some examples of
across and through variables are shown in table 1.5.
The comparison of the charging of a capacitor and the heating of an
object provides a simple example of the analogy that can be drawn
between physical systems. For a simple RC network the output voltage V
s
is developed from the xed supply voltage V
f
(gure 1.5). The time dependent
dierential equation is derived below.
For a capacitor,
CV
s
q. 1.33
Dierentiating equation (1.33),
C
dV
s
dt
i 1.34
Table 1.5. Across and through variables for some common systems.
Variable Electrical Thermal Fluid Translational
motion
Rotational
motion
Across y Potential
dierence
(V)
Temperature
dierence
(8C)
Pressure
dierence
(Pa)
Force (N) Torque (Nm)
Through _ xx Current
(A)
Thermal
power (W)
Volume ow
rate (m
3
s
1
)
Velocity
(ms
1
)
Angular velocity
(rads
1
)
x Charge (C) Heat energy
(J)
Volume (m
3
) Displacement
(m)
Angular
displacement
(rad)
xx Rate of
change of
current
(As
1
)
Acceleration
(ms
2
)
Angular
acceleration
(rads
2
)
Impedance
Z
Impedance
()
Thermal
resistance
(8CW
1
)
Pressure/
volume ow
rate
Force/
velocity
damping
constant
(Ns m
1
)
Torque/angular
velocity damping
constant
(Nms rad
1
)
Figure 1.5. The analogy between a capacitor being charged and an object being heated.
The measurement system as a series of networks 17
C
dV
s
dt

V
f
V
s
R
1.35
RC
dV
s
dt
V
f
V
s
1.36
RC
dV
s
dt
V
s
V
f
. 1.37
By comparing with equation (1.22) above it can be seen that
mc
UA
RC. 1.38
So, intuitively
C mc thermal capacity
R
1
UA
thermal resistance.
Concepts such as these are displayed in table 1.3.
If a process is being measured then a two-port network is added to repre-
sent the sensor (see gure 1.6). If, for example, the process is a hot tank whose
temperature is to be measured by a resistance thermometer then the use of
network analysis enables loading to be accounted for. The loading can
occur in the usual way by the subsequent signal conditioning circuitry aect-
ing the sensor. However, it can also occur by the sensor aecting the process
it is measuring. In this example the former could be caused by the signal con-
ditioning circuit heating the sensor element and the latter by the thermal
capacity of the sensor element altering the temperature of the hot tank.
The principles of impedance matching can be applied to the juxtaposition
of process and sensor in order to get optimum response (gure 1.7). If the
requirement is that maximum power is transferred from process to sensor
Figure 1.6. Process and sensor represented as one- and two-port networks.
Figure 1.7. Impedance matching of the process and sensor.
then the usual condition applies that
Z
p
Z
s
. 1.39
If, as is most commonly the case, the across variable generated by the process
y
p
should not be aected by the presence of the sensor then a dierent
condition applies. Assuming that the ow variable is conserved around the
system,
the flow variable. _ xx
y
p
Z
p
Z
s
1.40
the measured across variable. y
s
_ xxZ
s
. 1.41
Eliminating _ xx gives,
y
s

y
p
Z
s
Z
p
Z
s
. 1.42
So, for y
s
y
p
, then Z
s
)Z
p
.
If the ow variable _ xx is the desired measurand then the requirement
would be that the sensor impedance should be as low as possible so that _ xx
is not restricted. For the process not to be aected it can be seen from
equation (1.42) that the condition now is that Z
p
)Z
s
.
By the use of the impedance matrix technique it is possible to model the
whole instrumentation system as an interconnecting group of elements
driven by the process. In this way the dynamic response to a time-varying
measurand can be determined before construction, making the design of
complex systems more ecient.
Bibliography
This chapter has outlined the structure and general operation of an instru-
mentation system. The following chapters describe some of the more
recent developments in sensors. Traditional sensors such as thermocouples,
resistance thermometers and orice-plate owmeters have not been included
since these are well covered in other texts. For readers wishing to know more
of traditional sensors the following might be helpful.
Bentley J P 1995 Principles of Measurement Systems 3rd edition (Longman)
Considine D M (ed) 1999 Process/Industrial Instruments and Controls Handbook 5th
edition (McGraw-Hill)
Johnson C D 1988 Process Control Instrumentation Technology 3rd edition (New York:
Wiley)
Sensors: A Comprehensive Survey 1992 (Cambridge: UCH)
Usher M J 1985 Sensors and Transducers (Macmillan)
Bibliography 19
References
Finkelstein L and Watts R D 1971 Systems analysis of instruments Measurement and
Control 4 September
Kraus A D 1991 Circuit Analysis (West Publishing Company)
Squires G L 1968 Practical Physics (London: McGraw-Hill)
Exercises
1. A miniature camera has been developed with a radio telemetry link. It is
cylindrical, 8 mm in diameter, 10 mm long and is battery powered. The
probe costs 300 and the receiver system 2000 with a range of 10 m.
Suggest some applications.
2. Suggest a suitable VDU system for the control panel in the engine room
of a submarine.
3. The following measurements are made of the resistance R of a coil of
wire,
Resistance
R ()
5.615 5.622 5.624 5.618 5.620 5.633 5.628 5.624 5.613
Determine the best estimate of the resistance. Calculate the standard
deviation of the measurements (s) and the standard deviation in the
mean (s
m
). Which standard deviation value should be quoted with R
and why?
4. A owmeter produces an output current which is proportional to the
square of the ow velocity. Under steady ow conditions the readings
of the output current are found to have a standard deviation of 5%.
What is the standard deviation of the ow velocity?
5. The time period of a simple pendulum is given by
T 2
l,g
p
.
A number of experiments are performed and it is found that the
standard error in the mean of the time period T is 0.7% and of the
length l is 0.4%. Calculate the standard error in the mean of g.
6. If the errors given in question 5 were not standard errors but uncertainty
estimates from calibration limits or similar what would the estimate
of the overall uncertainty be then? Why is this estimate less useful
statistically than the standard error produced in the more formal way
of question 5?
7. A certain force sensor can be modelled as a spring, mass and damper
whose dynamic equation is
F m
d
2
x
dt
2
`
dx
dt
kx
where F is the applied force, m the eective mass, ` the damping con-
stant and k the spring stiness constant. Using Laplace transform
tables and assuming zero initial starting conditions show that the trans-
fer function G(s) is
Gs
xs
ys

1
1,.
2
n
s
2
2,.
n
s 1
where the undamped natural frequency .
n
is
k,m
p
and the so-called
damping ratio is `,2
km
p
.
8. Finkelstein and Watts proposed that an instrumentation system could
be described as a series of two-port networks:
(i) Draw a block diagram of a general instrumentation system showing
the various elements.
(ii) In an electrical two-port network what are the across and through
variables?
(iii) In a thermal two-port network what are the across and through
variables?
(iv) A process to be measured can be described as a one-port network.
What would be the across and through variables for a hot tank of
liquid whose temperature is to be measured?
9. It is possible to use software packages such as the matrix based Matlab
1
to model an instrumentation system. The general two-port network can
be described by the following block diagram.
(i) Write down the two impedance parameter equations.
(ii) Re-write the impedance parameter equations in matrix format.
10. The temperature of a hot liquid of small volume is to be measured by a
thermometer of thermal impedance Z
t
.
(i) What does it mean to say that the sensor might load the process?
(ii) The block diagram for such a system can be represented by the
following diagram
Exercises 21
where I
n
is the rate of supply of heat energy to the process of
thermal impedance Z
p
, R is the thermal resistance between the
process and sensor and V
0
is the measured temperature. Show that
V
0
I
n
Z
t
R Z
t
,Z
p
1
(iii) Show that the Laplace transform of the equation above is
"
VV
0
s
"
II
n
s

1
C
p
C
t
Rs
2
C
p
C
t
s
where C
p
and C
t
are the thermal capacitances of the process and
thermometer respectively.
(iv) Use the Matlab
1
toolbox Simulink
1
to model the transfer function
in part (iii) above. (Hint: use the Pulse source with an appropriate
height and width to deliver some heat Q.)
Chapter 2
Resonator sensors
2.1 Introduction
Resonator sensors are mechanical structures driven at their resonant fre-
quency. It is arranged in the design of the sensor structure that the parameter
to be measured alters the resonant frequency or the frequency spectrum or
both. Unfortunately, the resonant frequency is generally a nonlinear function
of the measurand and is subject to temperature drift. Also, mathematical
modelling of all but the simplest structures is not easy. However, the
dominant feature of resonator sensors that outweighs these disadvantages
is their very high sensitivity. Having the output signal as a variable frequency
can also be an advantage in environments subject to electromagnetic inter-
ference. But this does require more complex signal processing to recover
the measured value.
Although the concept of resonator sensors is not new, with the advent of
new materials and advances in modelling and fabrication techniques, smaller
and more elegant designs are being developed. Sensors have been reported to
measure mass, uid ow, density, viscosity, rotation rate, force, pressure and
acceleration. Microsensors using resonant techniques are in particular
attracting a lot of interest because of the well-developed semiconductor
fabrication techniques. Such sensors not only oer miniaturization, but
also batch production, which lowers costs. Some of these devices are
described in section 3.1.
Commonly used structures are wires, beams or cylinders. Because the
output is a frequency, signal processing is straightforward and problems
with noise and interference are signicantly reduced. Also, since the device
is being operated at resonance, power consumption is low. The main
disadvantages are that the resonant frequency can be signicantly altered
by temperature uctuations and that the output frequency is not linearly
proportional to the measured parameter.
The oscillation process can be designed as either
.
harmonic: energy is transferred from one stored form to another during a
cycle; losses are made good by a power supply, or
.
relaxation: only one form of energy storage exists; energy is regularly
supplied in bursts and is dissipated by the oscillatory process between bursts.
The sensing part of the oscillatory structure may be at a point or small region
(lumped) as in the case of the spring, mass and damper, or spread over the
whole structure (distributed) as with the longitudinally vibrating cylinder.
This chapter describes some of the structures that have been produced in
prototype form or as a commercial product during the past 25 years. They
are not microsensors, although it is likely that some of the structures
below will be adapted for use in miniature form as fabrication techniques
develop.
2.2 The damped harmonic oscillator
Often, resonating systems can be described by a linear, second order, dier-
ential equation over some if not all of their range. The standard equation
commonly used in engineering for any linear, second order, damped
system is of the form
1
.
2
n
d
2
x
dt
2

2
.
n
dx
dt
x y 2.1
where yt is the forcing term, xt is the resulting displacement, .
n
is the
undamped natural frequency, and is the damping ratio (if < 1 the
system response to an impulse will be oscillatory).
The format of equation (2.1) has become established because of the
convenience of having .
n
and as coecients, which are easily identied.
When comparison is made with the equation for some real system then the
undamped natural frequency and whether oscillation will occur can
quickly be determined. The simple spring, mass and damper system, for
example, is modelled by
m
d
2
x
dt
2

dx
dt
kx F 2.2
where m is the eective oscillating mass, is the damping constant, k is
the spring stiness and F is the applied force. By comparing coecients of
equations (2.1) and (2.2) it can be seen that for this system
.
n

k
m
r
2.3

2
km
p 2.4
24 Resonator sensors
The general solution to equation (2.1) can be either of two types depending
on whether the system is being driven by y, or whether the energy it has at
t 0 is being allowed to dissipate.
2.2.1 The non-driven oscillator
Three types of solution to equation (2.1) exist depending on whether the
system is,
underdamped. < 1 critically damped. 1 or overdamped. 1.
The solution for the underdamped case, with x x
0
at t 0, and x 0 at
t 1is
xt x
0
exp.
n
t
cos .t

1
2
1,2
sin .t
2.5
where . .
n
1
2
1,2
.
For resonator sensors, designs are such that (1 allowing the approx-
imation
xt x
0
exp.
n
t cos .
n
t 2.6
xt x
0
expt,2t cos .
n
t 2.7
where t 1,2.
n
and is sometimes referred to as the relaxation time. This
is not to be confused with the relaxation oscillator where the relaxation time
of the dissipation of the stored energy determines the oscillation frequency.
The form of equation (2.7) is shown in gure 2.1.
2.2.2 The driven oscillator
It is important to realize that when the oscillator is driven sinusoidally the
only solution allowed for equation (2.1) is that the oscillation frequency is
the same as the driving frequency (this is not true in a nonlinear system).
This means that, by using feedback, the frequency of the drive circuitry
can be made to follow the resonant frequency of the sensor as it changes
and so be used as the output signal. The general solution for a maintained
oscillator driven by
y y
0
sin .t 2.8
can be expressed in the form
x x
0
sin.t c 2.9
where x is the displacement of the point in the structure where y is input. The
phase c is the angle by which the displacement leads the force. To determine
the full description of equation (2.9), its derivatives together with equation
The damped harmonic oscillator 25
(2.8) are substituted into equation (2.1):
dx
dt
.x
0
cos.t c 2.10
d
2
x
d
2
t
.
2
x
0
sin.t c. 2.11
When substituted equation (2.1) becomes
.
2
x
0
.
2
n
sin.t c
2
.
n
.x
0
cos.t c y
0
sin .t. 2.12
Gathering terms,
1
.
2
.
2
n
x
0
sin.t c
2
.
n
.x
0
cos.t c y
0
sin .t. 2.13
This can be further simplied by using the trigonometric relations
sin.t c sin .t cos c cos .t sin c 2.14
cos.t c cos .t cos c sin .t sin c 2.15
Figure 2.1. Response of a non-driven oscillator.
Equation (2.12) then becomes
1
.
2
.
2
n
cos c
2.
.
n
sin c
x
0
sin .t
1
.
2
.
2
n
sin c
2.
.
n
cos c
x
0
cos .t y
0
sin .t. 2.16
Equation (2.16) can only be satised if the coecient of x
0
cos .t is zero:
1
.
2
.
2
n
sin c
2.
.
n
cos c 0. 2.17
This gives the result
tan c
sin c
cos c
2..
n
.
2
n
.
2
2.18
from which can be obtained
cos c
.
2
n
.
2
.
2
n
.
2
2
2..
n
1,2
2.19
sin c
2..
n
.
2
n
.
2
2
2..
n
1,2
. 2.20
Equation (2.16) can be re-written
1
.
2
.
2
n
cos c
2.
.
n
sin c
x
0
sin .t y
0
sin .t 2.21
and substituting equations (2.19) and (2.20) into equation (2.21) gives
x
0

y
0
.
2
n
.
2
n
.
2
2
2..
n
1,2
. 2.22
The full solution for x is therefore
x
y
0
.
2
n
.
2
n
.
2
2
2..
n
1,2
sin
.t tan
1
2..
n
.
2
n
.
2
. 2.23
The variation in c with . is as shown in gure 2.2. The driving force always
leads the displacement.
Resonator sensors are usually lightly damped so that the resonant
frequency approximates to .
n
. Some form of feedback is used to keep the
system operating at its resonant frequency ensuring that the drive frequency
. will always be .
n
. Consequently, the value of c will be maintained at ,2.
However, in a complex structure, parts will be vibrating with the same
frequency but dierent phases. It may be possible to use these relative
phases as an indication of the value of the measurand.
2.2.3 Q factor
A useful parameter when dealing with resonant systems is the Q factor. The
denition of Q is
Q
2 energy stored
average energy dissipated per cycle
.
For a mechanical resonator the stored energy will oscillate between potential
and kinetic forms. The stored energy will be entirely of the potential form
when the velocity is zero and the displacement x is a maximum. From equa-
tion (2.6), for the lightly damped, non-driven oscillator, setting the cosine
term to a maximum gives
xt x
0
exp.
n
t. 2.24
The potential energy E, stored in a linear system of elasticity constant K, is
1
2
K(displacement)
2
giving
total stored energy
1
2
Kx
2
0
exp2.
n
t. 2.25
The rate of dissipation of energy is dE,dt and dierentiating equation
(2.25) gives
dE
dt
.
n
Kx
2
0
exp2.
n
t. 2.26
Using this with the denition of Q from above gives
Q
2
1
2
Kx
2
0
exp2.
n
t
.
n
Kx
2
0
exp2.
n
t period
2.27
Figure 2.2. Variation of phase with angular frequency.
Q

.
n
2,.
n
2.28
Q
1
2
. 2.29
This useful result enables Qto be easily determined from the system equation
via . It can also be seen using equation (2.22) that
the amplitude at resonance
the amplitude as the frequency tends to zero
y
0
,2
y
0
Q. 2.30
This illustrates that, for a high-Qsystem, the amplitude of the oscillations are
Q times larger at resonance than at low frequencies.
It is also instructive to examine the way power is absorbed by the
oscillator as the driving frequency is varied. Monitoring the power absorbed
by the oscillator when used as a sensor may be useful in monitoring its
operational state in order to determine the value of the measurand. The
relationship conveniently used to determine the power absorbed, P, is the
product of the force and velocity:
P y
dx
dt
. 2.31
Substituting equation (2.8) and the dierential of equation (2.23) into
equation (2.31) gives
P y
0
sin .t
y
0
.
2
n
.
2
n
.
2
2
2..
n
1,2
.cos.t c. 2.32
The cosine term can be expanded using the trigonometric relation
cos.t c cos .t cos c sin .t sin c 2.33
so that equation (2.32) becomes
P
y
2
0
.
2
n
.
2
n
.
2
2
2..
n
1,2
.sin .t cos .t cos c sin
2
.t sin c. 2.34
The time average of sin .t cos .t over a period is zero and of sin
2
.t is
1
2
.
P

1
2
y
2
0
.
2
n
.sin c
.
2
n
.
2
2
2..
n
1,2
. 2.35
Substituting for sin c using equation (2.20)
P
y
2
0
.
3
n
.
2
.
2
n
.
2
2
2..
n
2
. 2.36
The variation in absorbed power, P, with . is shown in gure 2.3.
Engineers often nd the frequency spread, from where the peak occurs to
the points where the power has dropped by half, a useful concept to describe
the narrowness of the resonance peak. Using equation (2.36) to determine the
half-power points the denominator .
2
n
.
2
2
2..
n
2
is doubled.
To evaluate this, rst consider the power at the peak where . .
n
. This
reduces the denominator .
2
n
.
2
2
2..
n
2
to 2.
2
n
2
. Thus, for half
power, the denominator must be equal to twice 2.
2
n
2
:
.
2
n
.
2
2
2..
n
2
22.
2
n
2
. 2.37
Re-arranging,
.
2
n
.
2
2
22.
2
n
2
2..
n
2
2.38
which can be re-written as
.
n
.
2
.
n
.
2
22.
2
n
2
2..
n
2
. 2.39
Putting . .
n
. and making the assumption that .
n
. gives
.2.
n
2.
2
n
2.40
. .
n
2.41
and, using equation (2.29),
.
.
n
2Q
. 2.42
This provides a further description of Q,
Q
.
n
2.
2.43
Figure 2.3. Variation in absorbed power with angular frequency in a high-Q resonator.
where 2. is the full bandwidth of the frequency spread at the half-power
points. This indicates that a high-Q system has a narrow resonant peak.
To conclude this section, the advantages of high Q are summarized as:
.
lowdissipation of energy per cycle so that little power is needed to maintain
oscillation;
.
a narrow frequency band over which resonance occurs giving the sensor
high resolution;
.
greater immunity to loading eects from sense and drive circuitry.
It should be noted that some oscillatory sensors operate over a small range of
frequencies around resonance. For example, the sensor may sweep through a
range of frequencies in order to collect information about the measurand. In
such systems low Q is preferable so that constant amplitude exists over the
working frequency range.
Sections 2.3 and following illustrate some resonant sensor designs and
their applications.
2.2.4 Driving techniques
As described above the measurand modies the natural frequency of the
resonator sensor and this can be output as a representation of the measured
value. To maintain the oscillation at this variable resonant frequency a feed-
back loop is required. A detector unit must measure some parameter that
indicates the status of the sensor. A control unit then decides whether the
sensor is at resonance and the appropriate output is then fed to a power oscil-
lator. The power oscillator supplies the energy required to drive the sensor
and also provides the sensor output.
There are three ways by which feedback may be employed.
(a) Monitoring of the amplitude of the resonator output and adjusting the
drive frequency so as to maintain this at a maximum.
(b) Monitoring the phase dierence between the driving signal and the dis-
placement of the sensor. The drive signal leads the displacement by ,2
at resonance.
Figure 2.4. Block diagram of a resonator sensor drive system.
(c) Monitoring the power absorbed by the resonator, which is a maximum
at resonance.
Readily available integrated circuits can be used to construct the system.
Typically, these might include comparators, voltage controlled oscillators,
phase-locked loops and gating circuits.
Matlab
1
programs
The following M-le was used to draw gure 2.1.
clear, clf
wn=3000; dampratio=0.01; T=1/(2*wn*dampratio);
c=(1-dampratio^2)^0.5;
w=wn*c;
time= [0:0.0001:0.1];
for j=1:length(time);
t=time(j);
A=exp(-dampratio*wn*t);
cosine=cos(w*t);
sine=sin(w*t);
B=dampratio/c;
G(j,:)=A*(cosine+B*sine);
end
figure(1), plot(time,G,k)
The following M-le was used to draw gure 2.2 (labelled gure (1) in the le)
and gure 2.3 (labelled gure (5) in the le).
clear, clf
wn=3000; dampratio=0.01; T=1/(2*wn*dampratio);
FREQ= [2900:1:3100];
for j=1:length(FREQ);
w=FREQ(j);
A=(wn^2-w^2);
B=w/T;
D1=A.^2;
D2=B^2;
D3=(D1+D2);
D=D3^0.5;
cosine=A./D;
sine=-B./D;
tangent=-B./A;
G(j,:)=acos(cosine);
H(j,:)=asin(sine);
K(j,:)=atan(tangent);
M(j,:)=cosine;
power(j,:)=-0.199*wn^2*sine./D;
end
figure(1), plot(FREQ,-G*180/pi)
figure(2), plot(FREQ,H*180/pi)
figure(3), plot(FREQ,K*180/pi)
figure(4), plot(FREQ,M)
figure(5), plot(FREQ,power)
2.3 Vibrating wire sensors
This device is the simplest of all the resonator sensors. Assuming that the
rigidity of the wire to bending is negligible, the frequency of transverse oscil-
lations is
.
n

T
j
s
2.44
where T is the tension, j is the mass per unit length and l is the length of wire.
Sensors have been developed using wires made from the metals
tungsten, indium and high tensile steel as well as alloys such as Elinvar.
Excitation is achieved by using an a.c. current through the wire, which is
mounted in a permanent magnetic eld. The terminal impedance varies
strongly through resonance so the drive circuitry is designed to produce
the required frequency by tuning to this condition.
The areas of application are in measuring force, mass, strain and pres-
sure. With good design practice and protection of the wire from contamina-
tion, zero point stability, and hysteresis can be as low as 0.02%. Expansion
of the wire due to temperature produces errors of order 10 ppm/8C. Marshall
and Hunter (1980) reported using the vibrating wire to measure pressure. A
commercial device was available from Foxboro. Figure 2.5 shows a double
diaphragm arrangement.
Figure 2.5. Dierential pressure cell.
Vibrating wire sensors 33
2.4 Torsionally vibrating rod
A liquid density sensor (gure 2.6) was developed by Langdon (1980) at GEC
for use in leadacid batteries. The head unit contains a piezoelectric driver to
vibrate the rod, which is made of glass. Power comes from the battery itself,
making the unit self-contained. Density measurements are claimed to be as
accurate as oat hydrometers, having an error of 0.2%.
For a rotational second order system equation (2.2) can be re-written as
I
d
2
0
dt
2

d0
dt
c0 t 2.45
where I is the moment of inertia of the rotating component, is the rota-
tional damping constant, c is the spring restoring torque constant and t is
the applied torque. By comparing terms with equation (2.3) the undamped
natural frequency can be seen to be
.
n

c
I
r
. 2.46
2.5 Longitudinally vibrating plate
An established, non-resonating, method for on-line viscosity measurement
uses a vibrating sphere driven at a xed frequency of 200 Hz (Matusik and
Scarna 1981).
A more compact sensor for on-line viscosity measurement has been
devised using a thin rectangular plate clamped at its centre (Borovikov
1976) (gure 2.7). It may be driven and monitored by magnetostrictive
devices to produce longitudinal resonance. The damping, measured either
by switching o the power or measuring the power loss, is then used to
determine the viscosity. Alternatively, the phase shift c between driving
Figure 2.6. Torsion rod liquid density sensor.
force and displacement can be used where
c tan
1
2.
n
c
.
2
.
2
n
2.47
with c approximately proportional to the square root of the dynamic
viscosity. The most signicant disadvantage of the sensor is that penetration
of the wave motion into the liquid is very poor. At the typical 20 kHz drive
frequencies penetration is <10 mm, so probe cleanliness is essential. The non-
resonating device operating at the much lower frequency does not suer from
poor penetration.
2.6 Flexural resonators
A number of designs have been put forward for ow measurement using a
exing structure.
Flexing vane
MacDonald (1983) produced a device (gure 2.8) having a small vane (metal
or ceramic can be used) placed longitudinally in the pipe. It was driven into
exural resonance at about 1 kHz by a centre-mounted transducer. Phase
Figure 2.7. Viscosity measurement using a longitudinally resonating plate.
Figure 2.8. Flexing vane mass ow rate sensor.
Flexural resonators 35
dierences between pickups at the two ends of the vane were used to indicate
mass ow rate. Prototypes were built by GEC using a vane size of
150 mm20 mm which gave linear mass ow rate measurement for velocities
of 04 0.02 ms
1
. As is common in such systems, protection of the sensing
elements from contaminants that adhere is important and entrained gas can
also give false readings.
Flexing pipe
In this design the whole pipe section is set into resonance. The resonance
process can be likened to the vibrating wire sensor, the expression for the
undamped natural frequency .
n
being of the form of equation (2.3). The
resonant frequency is determined by the total mass in the device (including
the pipe itself). This makes the sensor useful in multi-phase or multi-compo-
nent applications. The Sarosota company have developed an on-line density
meter (gure 2.9(a)), in which the expression for the density is given as
,
A
.
2
n
B
.
n
C. 2.48
A dierent approach is to use the Coriolis eect. Coriolis showed that a
moving object in an accelerating reference frame experiences a side thrust.
The accelerating reference frame in this application is the oscillating pipe.
The vector equation relating the Coriolis force is
F
C
2m!
dr
dt
2.49
where m is the mass of the object moving at an angular velocity ! whilst
changing its radius vector r. F
C
and r are perpendicular. The device measures
mass ow. Because large lengths of pipe are needed the pipe is often bent into
a U-tube (gure 2.9(b)). Schlumberger and Micro Motion are among a
number of companies now marketing straight and U-shaped systems.
The Coriolis force causes a tiny twist to the pipe which can be detected
by a pair of position sensors or strain gauges.
Figure 2.9. Examples of vibrating pipe ow rate sensors.
Flexing tube
A liquid level sensor has been devised using a exing tube (gure 2.10)
300 mm long and 6 mm in diameter (Langdon 1983). The eective exural
mass of the tube is altered by the liquid surrounding it. This causes the
resonant frequency of about 5 kHz to decrease as the depth of liquid
increases. The response is linear apart from when the liquid passes through
a tube nodal point. This can be overcome by either
1. driving the device sequentially at dierent resonant frequencies, or
2. by plane polarizing the vibration pattern using vanes on the tube and
driving it at two dierent resonant frequencies at the same time.
In either method the nodes appear at dierent points. By combining the
output responses the nonlinearities can be removed.
These devices are mass level sensors and as such the density of the liquid
aects the output and may need to be measured separately. However, in
some applications, such as aircraft re-fuelling, this is the desired quantity.
2.7 Vibrating cylinder
A thin walled metal cylinder mounted axially within a pipe can be used as an
on-line densitometer (gure 2.11) (Agar 1967, 1976). It is held rigidly at one
Figure 2.10. Flexing tube liquid level sensor.
Vibrating cylinder 37
end and can be excited electromagnetically by external coils if made of ferro-
magnetic material. If the liquid is corrosive a ceramic tube is used instead
with piezoelectric sense/drive devices bonded to it. Being a thin tube, it
does not restrict the ow to any extent. Liquid passes over both surfaces
so radial pressure dierences are minimal. Sensor dimensions should be
much smaller than the wavelength so that acoustic coupling is low. Under
this restriction the uid behaves as incompressible and .
n
is unaected.
A further application of the vibrating cylinder is in the construction of a
rate gyro (gure 2.12). G H Bryan in 1890 observed that the vibration pattern
in a vibrating wineglass lagged the angle of rotation. This is a further example
of the Coriolis eect. The development of a sensor using this phenomenon
has been reported by Langdon (1982). The cylinder, which can be of metal
or ceramic, is driven at its fundamental exural mode by piezoelectric
transducers producing four nodes in symmetrical positions around the
cylinder. When rotated these nodes move and their angular change, moni-
tored by the piezoelectric transducers, is used to give an output proportional
to rotation rate.
Figure 2.11. Vibrating cylinder on-line densitometer.
Figure 2.12. Vibrating cylinder rate gyro.
A commercial device has been developed by GEC Marconi Avionics with
a sensitivity of 0.018 s
1
operating up to 4808 s
1
. An automotive version has
been developed from this by Bosch (see Reppich and Willig 1995). Johnson
(1995) has reported a similar device developed by British Aerospace for use
in the automotive industry and is in commercial production.
2.8 The tuning fork
An invasive liquid density sensor is available from Schlumberger using
tuning fork-type tines. It will operate in static tanks or in pipelines where
ow rates are below 3 ms
1
. The tines are driven and sensed by piezoresistive
devices and the resonant condition is maintained by monitoring the phase.
The relationship between the density and the natural frequency is given by
K
0
K
1
K
2
2
2:50
where K
0
, K
1
and K
2
are constants and is the time period of the resonant
frequency.
Ichinose and Terada (1995) have reported the development of an
angular rate sensor for automotive application. A miniature tuning fork
structure was designed using nite element techniques. The resonant
frequency is approximated as
!
n

2
t
l
2
E
r
2:51
where is the eigenvalue, t is the thickness, l is the length, E is Youngs
modulus and is the density. The eects of the Coriolis force on the resonat-
ing tines were measured using piezoelectric sensors. Signal recovery was
enhanced by synchronous detection.
2.9 The double tuning fork structure
Structures of this type have been developed for on-line density measurement
(gure 2.13). The ow pipe bifurcates to form the double tuning fork con-
guration. The contra motion of the side-sections tends to cancel any
bending forces that would otherwise be presented at the mountings.
However, axial stresses at the mountings can aect the resonant frequency
so exible couplings have to be used. The resonant frequency is related to
the density by
f f
0
r
2:52
The double tuning fork structure 39
where ,
0
is a constant. Errors of 0.1% or better can be achieved (Dress
1983).
A commercial pressure sensor using the double tuning fork structure is
now available from Yokogawa of Japan. The vibrating element is of solid
Elinvar alloy. In this conguration the modulus of elasticity E changes
proportionally with axial force. The force comes from a diaphragm at one
end of the vibrator. The resonant frequencypressure relationship is non-
linear so a microprocessor is used with a look-up table to output a linear
response. The value of Q is 10
4
and errors are typically 0.02% of the
reading. With suitable temperature compensation (most error is said to be
due to thermal expansion of the diaphragm) zero and sensitivity drift can
be as low as 0.001% per 8 C of full scale.
Interest in tuning fork-type structures still continues. Randall et al (1997)
have reported a pressure sensor using a metallic, triple-beam structure. It
is stated that loss of energy due to clamping is minimized in the symmetrical
structure used. Beeby et al (2000) describe the fabrication and testing of a
double tuning-fork, resonator microsensor constructed in silicon.
2.10 Bulk acoustic wave resonator sensors
Since the early 1970s quartz crystal systems have been commercially
available for monitoring thin lm deposition in vacuum evaporators. The
exposed quartz is part of an oscillating circuit and usually operates in thick-
ness-shear-mode. The resonant frequency is modied by the gain in mass.
Such devices are sometimes referred to as microbalances. Unfortunately,
temperature also aects the resonant frequency. Indeed, the technique can
be used for temperature measurement and commercial thermometers are
available with sensitivities of 0.001 8 C. By combining both sensors to
produce a temperature compensated thin-lm monitor it is possible to
monitor deposition of material to atomic layer resolution (gure 2.14).
Figure 2.13. On-line density measurement using a bifurcated pipe.
A typical commercial AT-cut quartz oscillator is a disk about 1016 mm
in diameter or a square or rectangle of similar size. The Sauerbrey equation
relates the frequency shift to the change in mass,
f
f
2
0
m
,kA
2.53
where , is the density, k is the frequency constant and A is the surface area of
the quartz crystal. Substituting these constants for quartz gives
f
2.25 10
2
f
2
0
m
A
2.54
with A expressed in m
2
and m in kg. Operating frequencies are usually 5, 9 or
10 MHz. At 10 MHz a sensitivity of 0.3 Hz per nanogram is achievable.
Various surface coatings have been developed for adsorbing a specic
gas or vapour. In this way the microbalance becomes a qualitative and quan-
titative sensor for a particular chemical species. Finding highly selective coat-
ings is not easy.
Biologically active coatings have also been produced to create specic
biosensors for use in the gas/vapour phase as well as in solution. The attrac-
tive feature with biological reagents is that they are extremely specic.
However, the construction methodology for biological systems is much
more dicult and often the sensor can only be used once.
2.11 Thick lms
A number of the sensors described above have piezoelectric drive and sense
transducers. Commonly, fabrication of these components is by some form of
thick lm process. Films can be classied by order of magnitude of their
thickness: ultra-thin <0.1 mm, thin 0.11.0 mm, thick 1.0100.0 mm.
Thin lm production often uses techniques developed by the integrated
circuit industry. As described later in chapter 3, this can be ideal when sensor
and electronics are combined on a single chip. However, some coating
materials cannot withstand the harsh production processes used. Also,
manufacture in silicon foundries is expensive and may not be viable for
Figure 2.14. Thin lm thickness-monitor using quartz resonators.
Thick lms 41
small-scale production. In these cases other techniques are more appropriate
for both thin and thick lm production. These include electrochemical
deposition, the LangmuirBlodgett technique, spin casting and screen
printing.
Of these, screen printing was one of the earliest techniques adapted
for thick lm production. Sensors made by screen printing have been in
development for some time (see Brignell et al 1987). The piezoresistance
strain gauge has enjoyed commercial success, rst marketed about two
decades ago.
Screen printing is an adaptation of silkscreen printing used in ne art for
pattern layering in textile and paper design and in some paintings. Here, a
synthetic fabric (silk is obsolete now) is stretched on to a rigid frame. A
pattern is laid down on the fabric by the artist and hot wax poured on to
make parts of the fabric impermeable to the printing ink. Additionally,
once the wax has set patterns can be carved forming permeable pathways
for the ink to reach the workpiece. The screen is then held rmly against
the workpiece using a press. The ink is spread uniformly over the fabric by
a roller and so penetrates through to the workpiece forming the desired
design. The process can be repeated with dierent designs and ink colours
to produce a composite.
For sensor application the desired layout for the active material is
formed on the screen fabric using photolithography so as to create a mask.
This might be, for example, a bridge circuit conguration. The ink is
squeezed through the mask to form a lm having a thickness typically in
the range 25100 mm. The lm is then dried by low temperature heat treat-
ment. Following this stage high temperatures are used to cause sintering of
the ingredients of the paste and wetting of the substrate so as to form
good adhesion. The ring prole is critical to the production quality.
The screen printing technique is not as highly specied as that of the
semiconductor foundry systems and consequently these sensors tend to
have poorer manufacturing tolerances. Researchers and manufacturers
produce recipes, sometimes of secret formulation, for these printable
pastes. The general formulation is for the active material to be suspended
in an organic solvent together with some form of matrix material (see
Baudry 1987 and Morton et al 1991). Here a piezoelectric lm based on
lead zirconate titanate (PZT) was used. Some materials which are not natu-
rally piezoelectric can be polarized to make them so. This usually involves
gently heating the lm and allowing it to cool in the presence of a strong elec-
tric eld.
Interest in thick lm sensors is still high, principally because of the lower
costs in making small, hybrid devices and because operating temperatures
can be higher than for comparable thin lm sensors. As with other sensor
types described in chapter 3, advances are dependent on organic and physical
chemistry for the development of lm materials.
References
Agar J 1967, 1976 Apparatus for Measuring the Density of a Fluid UK Patents 1175586,
1542564
Baudry H 1987 Screen printing piezoelectric devices Proc. 6th European Microelectronics
Conference (Bournemouth) pp 456463
Beeby S P, Ensell G and White N M 2000 Microengineered silicon double-ended tuning-
fork resonators Engineering Science and Education J. December
Borovikov A P 1976 Measurement of viscosity of media by means of shear vibration of
plane resonators Inst. Exp. Tech. 19(2) 223224
Brignell J E, White N M and Cranny A W J 1987 Sensor applications of thick lm
technology Proc. IEE I 135(4) 7784
Dress W B 1983 A high resolution ultrasonic densitometer IEEE Ultrasonics Symposium
pp 287290
Ichinose T and Terada J 1995 Angular rate sensor for automotive application Sensors and
Actuators SP-1066 pp 4956
Johnson B 1995 Vibration rotation sensors Sensors and Actuators SP-1066 pp 4147
Langdon R M, 1980 Liquid level measurement UK Patent GB 2067756A
Langdon R M 1982 Marconi Review Fourth Quarter pp 231249
Langdon R M 1983 Vibrating transducers in measurement Chem. Engineer 397 3335
MacDonald 1983 A vibrating vane mass ow meter 1st Euro. Conf. Sensors and their
Applications, UMIST pp 5859
Marshall J K and Hunter P 1980 Strain (GB) 1 3740
Matusik F J and Scarna P C 1981 Latest instrument makes on-line viscosity control of
slurries possible Control Engineering 28 116118
Morton B, De Cicco G, Gandol A and Tonelli C 1991 Advances in ferrelectric thick lm
materials and sensors Proc. 8th European Microelectronics Conference (Rotterdam)
pp 392399
Randall D S, Rudkin M J, Cheshmehdoost A and Jones B E 1997 A pressure transducer
using a metallic triple-beam tuning fork Sensors and Actuators A: Physical 60(13)
160162
Reppich A and Willig R 1995 Yaw rate sensor for vehicle dynamics control system
Sensors and Actuators SP-1066 pp 6776
Sarosota Automation Instrumentation Division 1985 Technical Information on FD 800
series density meters
Exercises
1. (i) State two advantages and two disadvantages to resonator sensors.
(ii) Distinguish between relaxation and harmonic oscillators.
(iii) Give three advantages to having high Q in resonant systems.
2. (i) Describe three approaches by which the resonant frequency might be
tracked.
(ii) Using functional blocks, design systems for these three dierent
approaches to tracking.
Exercises 43
3. The general equation for a damped second order system is
1
.
2
n
d
2
x
dt
2

2
.
n
dx
dt
x y.
Show, by using the Laplace transform or similar technique, that the
solution for a lightly damped, non-driven system with zero starting
conditions is
xt x
0
exp.
n
t
cos .t

1
2
1,2
sin .t
.
4. For a certain resonator sensor when the power is turned o the dis-
placement x declines as
xt x
0
exp.
n
t cos .
n
t.
Given that the stored potential energy is obtained using
1
2
Kx
2
determine
the maximum potential energy as a function of t and hence show that
Q 1,2.
5. A certain resonator sensor for measuring liquid density on-line can be
modelled by the equation
1.12 10
7
d
2
x
dt
2
7.4 10
7
dx
dt
x y.
(i) Determine the undamped natural frequency .
n
, the damping ratio
and the quality factor Q.
(ii) If the density of the liquid is related to the resonant frequency by
,
c
.
2
n
u
.
n
where c, u and are all constants, derive an expression for the

sensitivity.
6. A certain pressure sensor is constructed using a vibrating wire. Given that
the frequency f of the wire goes as f 1,2l
T,j
p
, determine the fre-
quency shift if the temperature decreases by 15 8C. The original frequency
is 50 kHz and the coecient of linear expansion for the wire material is
1.4 10
5
K
1
. Assume that the change in the wire tension is linearly
related to the change in length and that the actual length of the wire
remains xed.
7. (i) With the aid of a diagram explain the Coriolis eect, indicating the
relative orientations of the parameters involved.
(ii) Coriolis liquid owmeters can be constructed using either straight or
U-shaped pipes. Draw a diagram for each type indicating the force
direction. How might this force be measured?
Chapter 3
Semiconductor-based sensors
3.1 Introduction
The development of the silicon-based technologies in the latter half of the
twentieth century has undoubtedly had a major impact on societies through-
out the world. Components fabricated from silicon can be found in every
type of electronic device. Such devices have themselves enabled the sharp
increase in the rate of change of technological development and its
implementation. Although new semiconductor materials are coming into
more widespread use the vast majority of the devices produced are still
silicon based. The maturity of silicon foundry technology provides relatively
cheap routes to manufacturing the now vast range of products from the
simplest of discrete components to complex VLSI structures. As well as
the proliferation of components for digital applications, the use of power
electronic devices increases steadily and the explosive growth in the use of
cellular phones has caused rapid development of microwave-frequency
components.
The development of digital component technology is no longer the fore-
front of research. The new challenge is the development of microsensors to
interface the computer directly to the real world. The use of microprocessors
with sensors is already well established in the so-called smart sensors (see
section 5.5). This trend is set to continue with the combining of sensor and
circuitry on a single chip. For many applications the concept of a sensor
and its associated signal conditioning/processing circuitry being made
together in a single miniature package is very attractive.
Although single-crystal silicon structures can be used to sense light,
strain, temperature, magnetic and electric elds the sensitivities may not be
as high as that of more traditional sensors. An example of this is the applica-
tion of the Hall eect for the measurement of magnetic eld. The discrete
sensor can be constructed from a small plate of a IIIV compound
(formed from elements taken from groups III and V in the periodic table)
such as indium arsenide, indium antimonide or gallium arsenide, whereas a
magnetometer fabricated in silicon using bipolar technology is less sensitive.
However, a recent development is to use NMOS technology to produce
an ultra-thin slab of around 10 nm so that a larger potential dierence is
generated. Even though carrier mobility is lower than in a bipolar device,
sensitivity is increased by a factor of 20.
The design problems for fabricating semiconductor sensors can be sig-
nicant. For example, provision may be needed for an input lter, crucial
in keeping all but the desired measurand away from the exotic sensing
element. Harsh environments of temperature, vibration or pressure, or the
presence of substances that would poison the sensor, can also pose diculties
for design engineers.
Interdisciplinary teams are needed for the design, development and fab-
rication processes. If the sensor can be fabricated entirely in silicon it can be
manufactured at the same time as all the other components on the chip using
the same masks and stages. However, if, rather than adopting available
silicon foundry plant, non-standard manufacturing techniques are required
to produce a hybrid, the costs may mean that a proposed device will only
be economically viable if high volumes can be sold.
3.2 Mechanical microsensors
A number of new materials are nding applications in mechanical micro-
sensors. Silicon has been used for several decades and is attractive because
the manufacturing technology is well developed to produce large, extremely
pure, single crystals. These can be doped, etched or machined as desired. The
mechanical properties of silicon are also very attractive to sensor design
engineers. It has
.
high tensile strength
.
good fatigue strength
.
high Youngs modulus, similar to steel
.
good strength/weight
.
good hardness
.
good resistance to deforming plastically so creep and hysteresis eects are
minimal.
Construction is usually in single-crystal silicon. However, polycrystalline
silicon is more exible and so is sometimes preferred for deecting structures
which are in turn constructed on single-crystal substrates. A disadvantage is
that polycrystalline silicon is more prone to creep and hysteresis.
Sensor developments using bulk micromachining began in the early
1960s. A single crystal wafer of silicon is selectively processed throughout
its bulk. This may involve architectural dimensions of less then 1 mm or up
46 Semiconductor-based sensors
to the full thickness of the wafer of 500 mm. Most commercial microsensors
currently are of the bulk-machined type.
Since the early 1980s surface micromachining has attracted a lot of
attention. As the name suggests, the sensing element is a thin layer within
the surface. The advantages are a reduction in sensor size of an order of
magnitude compared with a bulk micromachined device and the relative
ease of adding integrated circuitry adjacent to the sensing element. The
main disadvantage is that, because of their smaller size the sensitivities
may not be as high as can be achieved using bulk micromachining. Con-
sequently, additional circuitry has to be added to the substrate to amplify
the response.
Mechanical microsensors for the measurement of strain, pressure,
acceleration and gas ow are commercially available. The general method
of operation is that the measurand aects some mechanical feature of the
sensor and a secondary sensor using capacitance, piezoresistance or, to a
lesser extent piezoelectricity, provides the electrical signal.
3.2.1 Using silicon to measure strain
pn junction devices
The physics of the pn junction is known to be sensitive to stress. The band-
gap energy changes proportionally with stress. It has been suggested that
diodes, bipolar transistors and MOSFETs might provide miniature stress
sensors. However, stresses need to be high and, as yet, reproducibility has
been dicult to achieve. No commercial application is available at the
time of writing.
Piezoresistance devices
Unlike metallic strain gauges, where the change of the bulk shape causes
change in the resistance, in semiconductor gauges it is the change of the
resistivity by the strain that is the dominant term. The gauge factor G of a
strain gauge is given by
G 1 2i
,
,e
L
3.1
where i is Poissons ratio, , is the resistivity, and e
L
is the longitudinal strain.
The fractional change in resistivity per strain, ,,,e
L
, is the piezo-
resistivity. In metals the piezoresistivity is approximately 0.3 and, together
with a typical value for i of 0.5, gives G a value of about 2. However, in
silicon based semiconductors the piezoresistivity can be 100. This
becomes the dominating term in the gauge factor equation. Semiconductor
strain gauges are available with G values of up to 150. Some other advantages
of semiconductor strain gauges are as follows.
Mechanical microsensors 47
.
The sensing element(s) can be tailored to the application easily in order to
produce maximum sensitivity.
.
Matching sets of sensors required for bridge circuitry techniques are easily
produced.
.
Miniaturization is easily achieved.
.
Mass production is straightforward using the silicon industry techniques.
.
n- and p-type doped devices are available having opposite sign piezoresis-
tive coecients. It should also be noted that dierent values of coecients
exist for the dierent crystal orientations.
The main disadvantage to semiconductor, piezoresistive devices is their signi-
cant temperature sensitivity. The piezoresistive coecient is temperature
sensitive. This is reduced if doping is increased but this also reduces the
piezoresistive coecient. It must also be remembered that the semiconductor
material itself has a high temperature coecient of resistance, making the
zero-setting resistance uctuate. These two eects can be minimized by using
a compensation technique such as a bridge network. The zero value of a
device is also subject to temperature-dependent drift due to mechanical stresses
incorporated during manufacture. Careful design is required to minimize these.
3.2.2 Pressure gauges
The growth rate in the use of miniature pressure sensors has been phenom-
enal in recent years. The automotive industry alone has seen an increase of
20% per annum in the past decade. The earliest application was probably
in 1989 for monitoring exhaust gas re-circulation. More recently devices
have been incorporated for measuring manifold, tyre, fuel and hydraulic
uid pressures. Gauge types are classied according to the pressure reference:
.
absolute: compared with vacuum;
.
dierential: pressure is measured between two points in the system;
.
gauge: compared with atmospheric pressure.
The piezoresistive phenomenon described above is commonly used as the
basis of the sensing process. Bulk-machining is used on a single-crystal
silicon plate or circular diaphragm of order 10 mm thick. Piezoresistors are
constructed within the diaphragm inner surface which is of order 1 mm
2
.
Their layout is governed by the strain geometry of the diaphragm as it
deforms under the applied pressure. For a circular diaphragm two strain
directions need to be considered, one along the radial direction and the
other normal to the radius resulting from hoop stress (gure 3.1). Both
stress relationships are a function of radius. Dalley and Riley (1978)
present the strain equations for the diaphragm:
radial stress
r

3P1 i
2
8Et
2
R
2
0
3r
2
3.2
tangential stress
t

3P1 i
2
8Et
2
R
2
0
r
2
3.3
where E is the Youngs modulus, i is Poissons ratio, t is the thickness of the
diaphragm material, R
0
is the diaphragm radius, r is some point along the
radius, and P is the dierential pressure.
It can be seen that
r
changes sign at the point where R
2
0
3r
2
, whereas
t
is always positive. Thus by careful positioning of four strain gauges, two in
compression and two in tension, high sensitivity and temperature compensa-
tion can be achieved.
Piezoresistive pressure gauges have been available commercially since
the early 1980s. Typical parameter values are
range 0100 kPa
sensitivity 3 mV/kPa with a 10 V power supply
frequency response 100 kHz
nonlinearity and hysteresis 0.5%
They cannot be used in high temperature or pressure systems directly and,
with certain noxious materials, without protection.
Kress et al (1995) describe an integrated silicon pressure sensor for
automotive applications such as manifold intake pressure, fuel tank and
brake-line hydraulic pressure. Not only does the device contain on-board
signal conditioning but also an electronic trimming system. This trimming
process corrects for production tolerances including sensitivity and is
performed by sending a digital compensation code word to the chip. The
applied voltage is such as to irreversibly store the sensor characteristic.
Previous designs have used the more expensive laser trimming technique.
The pressure range is 20115 kPa (absolute) and the output is 0.44.65 V
using a 5 V supply. An isolation technique is used enabling the device to
operate over the full automotive temperature range of 40 8C to 125 8C.
A pressure sensor having higher sensitivity has been reported using
resonance, Greenwood (1988). A shuttle structure was machined out of
single-crystal silicon. The supporting struts for the shuttle were connected to
the diaphragm. As the diaphragm deformed the changes in stress on the
shuttle supports changed the mechanical resonant frequency. High long-term
stability was achieved. Recent work by others on similar devices has shown
that high precision is possible, making it attractive for commercialization.
Figure 3.1. Strain orientations in a simple diaphragm.
However, it is likely to be marketed as a high-cost device due to the complexity
of fabrication.
Pressure gauges are also manufactured using capacitance as the second-
ary sensor. This approach has been used in non-microsensors since the
emergence of the instrumentation industry. In these traditional devices the
whole of the diaphragm acts as one of the plates of a capacitor. Shown in
gure 3.2 is this type of capacitive pressure sensor. One plate is a xed
metal base, the other is a exible circular diaphragm which is clamped
around its circumference. The diaphragm is deformed as shown by the
applied pressure P.
The capacitance at zero pressure can be determined using the formula
for a parallel plate capacitor,
C
0
R
2
0
d
. 3.4
The deformation of the diaphragm means that the average separation of the
plates is reduced and the resulting increase in capacitance C is given by
Neubert (1975) as
C
C

1 i
2
R
4
0
16E dt
3
P 3.5
where, E is the Youngs modulus, i is Poissons ratio and t is the thickness of
the diaphragm material.
The microsensor versions of this device use a silicon diaphragm and eva-
porated metal lm to form the upper and lower capacitor plates. Some ver-
sions do not use the whole diaphragm as one of the capacitor plates. Instead
sets of miniature capacitors are fabricated within the gap, those at the edges
being unaected by the pressure providing a reference to give temperature
compensation.
3.2.3 Accelerometers
The growth in the commercial application of the miniature accelerometer has
been almost as great as that of the miniature pressure sensor. There has
always been a demand for accelerometers but in recent years their usage
has soared. Again the automotive industry has shown rapid take-up for
Figure 3.2. Conventional capacitance pressure gauge.
installation in air-bag, active suspension and breaking systems (see Gardener
et al 2001).
An accelerometer is a device to measure acceleration or, in some
congurations, vibration (gure 3.3). It consists of an inertial mass m,
which undergoes the acceleration a. According to Newtons Second Law a
force F is required to cause this acceleration:
F ma. 3.6
This force is delivered to the mass via the coupling that connects the mass
to the vehicle. If this coupling is some form of elastic element, which is
operating in its linear region, then the displacement is proportional to the
acceleration. Measuring this displacement therefore provides an estimate
of the acceleration. A miniature cantilever made in silicon is a simple
structure that has been used as the basis for such a device.
If the elastic properties of the material remain constant then the
mathematical model of the system is a linear equation and the general
relationships described in section 2.2 are appropriate. The damping ratio
needs to be approximately 1 so that the system does not oscillate and that
the maximum deection is reached speedily. This can be achieved by appro-
priate sensor geometry and utilizing the viscosity of the surrounding uid
(air). The solution to equation (2.1) for a step input then becomes
x 1 1 .
n
t e
.
n
t
l
3
3EI
F 3.7
where E is Youngs modulus, l the length of the beam and I the second
moment of area bd
3
,12, where b and d are the beam width and thickness
respectively. For a cantilever it can be shown that
.
n

3EI
ml
3
r

1
2
Ebd
3
ml
3
s
3.8
where m is the eective mass which can be taken as that of the end block if
this is greater than that of the beam by a factor of about 30.
In addition to the simple cantilever, other designs have been developed
and include the hammock, crab-leg and folded exures (Gardner et al 2001).
The methods developed to detect this displacement are described below, the
most popular being the piezoresistance and capacitance types.
Figure 3.3. A simple cantilever accelerometer.
Piezoresistive
Accelerometers employing piezoresistance have been commercially available
from the mid 1980s. Early designs were adaptations of the pressure gauge
described above with an added mass in the centre of the diaphragm. In
more recent designs the micro-inertial mass is suspended from struts(s) of
material containing the piezoresistive elements. If silicon is used the
spring is normally nonlinear and so careful design of the geometry is
essential if a linear output is desired. As described above, piezoresistive
devices are prone to contamination and suer from temperature drift.
Encapsulation is needed and temperature compensation can be achieved
using a dummy reference element.
As with the piezoresistance pressure gauge, using multiple sensors is a
better alternative to having a dummy reference. This is easily illustrated in
the simple cantilever. Sensors placed on the upper and lower surfaces
suer opposite stresses and when appropriately congured in a bridge
network, as shown in gure 3.4, can multiply the sensitivity as well as
compensating for temperature drift. E
Theevenin
is the open circuit potential
dierence.
Capacitive
The underside of the inertial mass is coated with a thin metallic lm which
acts as one of the plates of a parallel plate capacitor. The substrate is similarly
treated to form the other plate. The capacitance C of a parallel plate
capacitor is given by
C
A
0
r
d
3.9
where A is the plate area,
0
is the permittivity of free space,
r
is the dielectric
constant of the material between the plates, and d is the plate separation.
Figure 3.4. Deection bridge with four active sensors.
Variation in d is used to alter C. This is a nonlinear relationship which
may be linearized using a bridge circuit or microprocessor (see Smart sensors,
section 5.5). The capacitance is very temperature sensitive due to thermal
expansion aecting d. This may be reduced by careful design incorporating
the dierential twin capacitor technique shown in gure 3.5. A further
advantage of this dierential system is that it can be incorporated into two
arms of an a.c. bridge which linearizes the output signal,
E
The evenin
V
s
,2dx, where 2d is the xed separation of the outer plates
and x is the displacement of the centre plate.
These devices became commercially available shortly after the piezo-
resistive types. Reidemeister and Johnson (1995) report an application in
automobiles. Thermal stability was achieved by matching the coecients
of thermal expansion of the beam, support base-pin and substrate in this
single capacitor design. A guard electrode was also used to improve
performance. The structure was enhanced using nite element analysis so
as to measure low values of acceleration.
Analog Devices produced one of the rst commercial capacitance-based
devices described in more detail below.
Piezoelectric
Unfortunately silicon has no piezoelectric eect. Consequently, silicon-based
sensors need to be coated in some piezoelectric thin lm such as zinc oxide.
The advantages of piezoelectric sensing are a wider working temperature
range and smaller size. However, a signicant disadvantage is charge
leakage. This has the eect that the measurement of a steady acceleration
appears to be zero after a few seconds. This restricts their application to
vibration measurement. They are also more expensive than the above.
Resonance
A further design of accelerometer is based on the resonator technique. This is
a similar concept to that described above for pressure measurement. Rather
than the deection of the inertial mass being measured, the change in the
modulus of elasticity of the strut material due to varying stress is used.
Changing the elasticity has the eect of changing the natural frequency of
the micro-exing structure. Piezoelectric drive and sense elements are used
Figure 3.5. Dierential capacitor system.
in a feedback loop to maintain resonance. The resonant frequency is the
output signal indicating the acceleration.
Force-balance
Designs have become established using the force-balance technique. The
concept is to use a magnetic or electric eld to counter the acceleration
force so that the inertial mass suers virtually no displacement. Con-
sequently, a linear response is achieved regardless of the geometrical
structure of the mass suspension system. The output signal is derived from
the feedback signal that generates the magnetic or electric eld.
A commercial device using this is the ADXL-50 from Analog Devices,
rst available in 1991 and able to measure accelerations of 50g (Patel
1992). It is a surface-machined device and uses capacitance to monitor the
movement. The xed capacitor plates are a double comb-like structure
mounted 1 mm above the surface on either side of the shuttle section.
The shuttle has projections extending from either side. These form the
moving capacitor plates which mesh with the combs. The total capacitance
of the whole double-comb structure is only 1 pF and is less than 1 mm
2
. It
is congured to operate in dierential mode as described above. The
remainder of the chip contains an amplier, oscillator demodulator and
lter which provide a d.c. voltage directly proportional to acceleration.
The complete chip is some 9 mm
2
. Rather than allow the shuttle to move
signicantly a force-balance approach is adopted. A voltage is applied to
the capacitor plates so that an electrostatic force is produced which
opposes the acceleration force constraining the shuttle movement. This
has the added advantage that the device can be tested without applying an
acceleration.
3.2.4 Flow sensors
Thermal
The transfer of heat from a heated element to a owing uid was rst
proposed by C C Thomas in 1911 as a means of measuring ow. A simple
implementation is in the so-called hot-wire anemometer used to measure
wind speeds. Two self-heated sensors are incorporated in a deection
bridge network. One is shielded from the wind so as to act as a reference.
In this way compensation is possible for changes in the ambient temperature
of the wind.
For uid ow in pipes, using non-miniature sensors and for microsensor
implementation, a more robust approach is to have up-stream and down-
stream sensing elements separated by a heater. Microsensors have been
developed where the whole system is made on one silicon substrate.
Johnson and Higashi (1987) produced a device for gas ow measurement
where two identical resistance thermometers were mounted on two micro-
bridges above the substrate. The thermal conductivity of silicon is too high
and would allow heat energy to drain into the substrate and so raise the
power consumption of the device. Silicon nitride has lower conductivity
and so was used to construct the bridges. A heating element was mounted
between the two resistors, half on each microbridge, and gas owed in a
channel underneath. The tiny size of this microbridge structure resulted in
low power consumption (typically 10 mW raised the temperature by
100 8 C) and a fast response of about 3 ms. Various renements have been
reported measuring ow velocities of up to 30 ms
1
. Devices are available
commercially from Honeywell for air mass ow sensing. Although in
principle the technique should be applicable to liquids the risk of contamina-
tion is a signicant problem. Any deposit on the surfaces would change the
thermal conductivity and upset the calibration.
Resonance
Bouwstra et al (1990) have reported a silicon nitride, microbridge structure
measuring 600 200 21 mm. The resonant frequency of the structure was
85 kHz. Excitation was achieved thermally using thin lm resistors. The
elevated temperature was 20 8 C above ambient. Piezoresistive sensors on
the structure provided the output signal and a frequency shift of 800 Hz
was observed for ows of 010 ml min
1
.
The advantages of resonant systems, as mentioned in chapter 2, are their
high sensitivity, fast response and good stability. The major drawback in this
device is the potential change in mass if contaminated. Only clean uids
may be used.
Capacitance
Cho and Wise (1993) reported a gas ow sensor based on dierential
pressure. The design geometry of the dome structure was such that the
mass ow rate was proportional to the dierential pressure between inlet
and outlet pipes. This dierential pressure was used to vary the plate separa-
tion of an internally-mounted silicon micro-capacitor. Typical values of
capacitance were 515 pF with 1 fF resolution measured using a CMOS
switched capacitance circuit. It was reported that 1 fF corresponded to
0.13 Pa dierential pressure.
The disadvantages of these devices, as of capacitance sensors generally,
are their temperature drift and leakage current. Improved designs may
reduce these.
The application of micro-owmeters is set to increase over the years
ahead due to the growing importance of micro-uidics in MEMS (micro-
electrical-mechanical systems) (Gardener et al 2001).
3.2.5 Angular rate sensors
In section 2.7 the rate gyro was introduced as an angular rate sensor employing
the Coriolis eect (see section 2.6). Semiconductor designs are mostly either
some form of driven tuning-fork structure or a resonating cylinder or ring.
A tuning fork-type design was reported by Bernstein et al (1993). This
was a micro-machined device that used a comb-drive structure. Microsensors
of this type have evolved from the miniature tuning fork rate-gyro described
in section 2.8. Whereas, the operating principle of the cylinder/ring devices is
that of the vibrating tube rate-gyro described in section 2.7.
Johnson et al (1995) have reported the development of a ring-type, surface
micro-machined, rate gyro. Ametal ring was fabricated on a p-type silicon sub-
strate by electroplating and photoresist technologies. A total of 32 electrodes
were formed on the substrate around the ring (shown diagrammatically in
gure 3.6) for sense and drive purposes. The substrate also contained CMOS
circuitry for buering and amplication. A second integrated circuit contained
control loops, programmable DACs to correct for oset and gain errors, and a
reference voltage source used to electrostatically polarize the ring.
There is a signicant and expanding market for angular rate sensors. A
number of manufacturers now produce rate-gyros, many of which have a
similar external size of about 25 mm25 mm in section and about 18 mm
high. However, there are a variety of constructional styles and performances.
Those described above can be classed as truly solid-state devices. However,
some are hybrid structures the primary sensor being a discrete ring, cylinder
or, in one design an equilateral Elinvar prism, which is packaged with an
integrated circuit.
3.3 Surface acoustic wave sensors (SAWS)
In section 2.10 sensors were described using a quartz crystal made to resonate
mechanically throughout its bulk. Such oscillations are described as
thickness-shear-mode (TSM). Surface acoustic waves (originally known as
Rayleigh waves) are also mechanical waves but as the name suggests only
propagate through the surface of the device. The disturbance is greatest at
the surface but declines with depth and eectively only penetrates about a
Figure 3.6. Miniature angular rate sensor.
wavelength. Although the velocity of the waves is similar to that in TSM
devices, being in the range 3.84.2 10
2
ms
1
, frequencies are much
higher. Devices can be made to operate at up to 5 GHz. Since the sensing
ability is usually dependent on the fractional change in frequency f ,f ,
such high frequencies are clearly an advantage. A further advantage of oper-
ating at microwave frequencies is that remote sensing by wireless operation is
possible. The tiny sensor with antennae can be powered by the incoming
electromagnetic wave and the returned wave transmits the data, see
Varadan in Gardener et al (2001).
The operating principle of SAW sensors is that the measurand alters the
velocity of wave propagation which in turn causes a change in the time interval
between the injected and recovered signals. Devices typically have an area of a
few mm
2
. The piezoelectric eect is used for driving and sensing. Construction
may be of a homogenous design. The drive and sense components and sub-
strate are the same piezoelectric material. Alternatively, a semiconductor
substrate can be used such as Si or GaAs. Since these semiconductors are
not piezoelectric a coating of piezoelectric material is needed to generate the
waves and sense their arrival. The most commonly used coatings are ZnO,
AlN or Pb(Zr,Ti)O
3
(PZT) at about 5 10
7
m thick (0.05`) for best ZnO
SiO
2
Si coupling. Using a semiconductor substrate means that signal condi-
tioning circuitry may be built on to the device using monolithic fabrication
technology. This is clearly an attractive option. However, the physics of the
wave propagation changes. The wave velocity may no longer be independent
of frequency. Also, the coating processes such as vapour deposition or sputter-
ing mean that the piezoelectric layer will be polycrystalline and rough, which
tends to disrupt wave coupling into the substrate.
The electrode pattern on the surface of the piezoelectric material is com-
monly the inter-digital style, a comb-like structure, with up to 20 electrodes
per set. The electrode patterns are fabricated using photolithography. The
spacing between electrodes of the same polarity is one acoustic wavelength
and is about 10 mm. The surface of the device is constructed as either a
delay line or as a standing wave resonator. The structures of the two types
are shown below.
Resonator saw sensor
The resonator SAW sensor (gure 3.7) has an echelle-type grating at each
end of the active surface. These have the eect of reecting the wave back
towards the centre of the device so as to develop a standing wave. The
sense and drive electrodes together with an amplier, having automatic
gain control, provide a positive feedback loop so as to maintain the resonant
condition.
Changes in the propagation time between the transmit and receive
electrodes will alter the resonant frequency, .
n
. The resonant frequency is
Surface acoustic wave sensors (SAWS) 57
given by
.
n

2n
t
3.10
where is the ampliers phase shift, n is the excited mode number, and t is
the time delay, which is dependent on the measurand.
Delay line saw sensor
In this type of SAW sensor (gure 3.8) the oscillating frequency is xed and
the phase shift caused by the delay is measured. The delay line type lends
itself well to a dual format enabling a reference device to be constructed
on the adjacent area of the same substrate. Of the two operating types it is
the more commonly used when a reference is required.
The phase shift c between the transmitted and received signals caused by
the propagation time of the SAW is given by
c 2f
0
t 3.11
where t is the time to travel over the surface between the electrodes. This can
be expressed as
c 2f
0
l
v
3.12
Figure 3.7. Surface acoustic wave sensor: resonator type.
Figure 3.8. Surface acoustic wave sensor: delay line technique.
where l is the path length and v is the variable velocity altered by the
measurand. For a dual delay line device f
0
is xed. The dierence in the
phases of the two paths is given by
c
sense
c
reference
2f
0
l
sense
v
sense
l
reference
v
reference
c. 3.13
An analogue multiplier can be conveniently used to combine the outputs of
the reference and sense lines. Using the trigonometrical identity
sin Asin B
1
2
cosA B
1
2
cosA B 3.14
gives
sin2f
0
t c sin 2f
0
t
1
2
cosc
1
2
cos4f
0
t c. 3.15
A low-pass lter leaves just the
1
2
cosc term as the output. Some delay line
sensors operate singly, without a reference section. In this type the r.f.
oscillator signal can be multiplied with the received signal from the delay
line to give the
1
2
cosc term.
3.3.1 Mass and gas/vapour sensors
A simple mass sensor, or microbalance, can be made using the SAW principle.
Changes in mass, caused by some evaporation process for example, alters
the wave velocity which is then output as a variable frequency. The change
in frequency is given by
f k
1
k
2
f
2
0
j 3.16
where j is the mass per unit area and k
1
, k
2
are constants of the material. For
Y-cut quartz k
1
is 9.8 10
8
and k
2
is 4.18 10
8
m
2
s/kg. Measure-
ments as low as 1 picogram have been observed.
The design concept for a gas sensor is to have a coating that will selec-
tively adsorb the molecules of the gas to be detected. This sorption changes
the properties of the interface between coating and substrate which results in
a change in the propagation time. This eect has been used in a resonator
sensor to generate a change in the operating frequency. Wohltjers and
Dessy reported the rst SAW gas sensor in 1979. Research and development
in this eld has intensied. The advantages of SAW gas sensors compared
with the bulk acoustic wave devices are
.
the two orders of magnitude higher operating frequency means greater
sensitivity,
.
mounting the back of the planar substrate to a rigid surface gives greater
stability than is possible with bulk resonators,
.
miniature size means several can be built on one chip for multi-sensing and
temperature compensation.
Surface acoustic wave sensors (SAWS) 59
The design eort is now concentrated on the development of the chemical
coating needed to detect the desired species. Vapour sensors have also
been developed using a similar approach although here it is often the
change in mass that changes the resonant frequency.
Sensors have been reported for measuring vapour levels of explosives,
drugs, water and of the gases SO
2
, H
2
, NH
3
, H
2
S, NO
2
, CO
2
and CH
4
.
3.3.2 Temperature sensor
A LiNbO
3
substrate SAW temperature sensor has been reported by Neumeis-
ter et al (1990). This is a resonator-type device where the oscillation frequency
.
n
is a function of temperature. This is due to the thermal expansion of the
substrate causing an increase in the SAW propagation time. The temperature
dependence is given by
.
.
n
% c
1
c
2
T 3.17
where c
1
, c
2
are the thermal coecients of expansion and velocity respec-
tively. The device operates at about 40 MHz between 50 8C and 150 8C
with a constant sensitivity of 4 Hz/8C. Resolution is high at 0.001 8C but
repeatability is 0.2 8C.
3.3.3 Strain sensor
The SAW delay line technique can be used to detect strain. If the propagation
speed of the SAW wave is not aected by strain in the material then the
resulting change in phase is only due to the change in the physical path
length caused by the strain. Adapting equation (3.12),
c
c

l
l
. 3.18
But the tensile strain is dened as l,l so c is proportional to .
Placing the strain sensor on a miniature cantilever or similar structure
enables the construction of an accelerometer. Similarly, a miniature pressure
sensor can be constructed using a quartz diaphragm. The strain sensor is
placed on the non-contact side of the diaphragm so that uid cannot reach
the SAW region since this would also cause a change in the frequency.
3.3.4 Magnetic sensor
The magneto-elastic eect, or magnetostriction, was observed by Joule in
1842. Magnetoelastic materials change their length when subjected to a
longitudinal magnetic eld. The eect is small, typically 10 ppm in moderate
magnetic elds.
If such a magnetoelastic thin lm is placed on a SAW device any
magnetic elds present will cause stress to occur across the boundary layer
due to the magnetoelastic material changing its length. This stress will alter
the velocity of the SAW.
Hanna (1987) has reported a resonator-type SAW magnetic sensor. The
relationship between resonant frequency and magnetic eld was nonlinear
but high sensitivity was achieved.
3.3.5 Conclusion
A number of SAW sensors have been developed and this is set to continue.
To summarize, the attractive features of these devices are
.
their high resolution,
.
the facility to provide on-board electronics,
.
the potential for wireless operation,
.
the potential for high volume production using monolithic technology.
Their disadvantages are
.
poor long-term stability,
.
high temperature and humidity sensitivities.
Compensation techniques may help towards overcoming some of these.
3.4 Chemical microsensors
Traditionally, solutions have been analysed using glass electrodes for pH and
ion-selective electrodes (ISEs) for ions such as K
, Ca
2
, Cl
and Br
.
Enzyme sensors and biosensors are also available commercially. However,
during the research and development of semiconductor materials over the
past 40 years it was noticed that surface properties were aected by a
variety of impurities. In Japan in the early 1960s semiconductors were rst
used to detect gases in the atmosphere and later for the analysis of solutions.
A great deal of eort has been put into the development of such devices and is
continuing.
In recent years the world-wide growth in sales of chemical sensors has
not suered to the same extent as other areas of the global economy. The
environmental and medical elds are expected to show the largest increase.
Because of their limited initial use, chemical microsensors have greater pre-
dicted annual growth than the traditional sensors. Estimates of chemical
microsensor growth rates are put at between 9.5 and 10.5% per annum for
the years to 2006.
Although some developments have involved diode structures such as the
Schottky, GCD (gate controlled device) and ICD (ion controlled device), the
Chemical microsensors 61
majority are based on the eld eect transistor (FET). The functional part
of the FET is the gate region. FET sensors can be classied according to
the type of species to be detected. The following mnemonics have been
coined:
ISFET ion-sensitive FET
REFET reference FET
CHEMFET chemically sensitive FET
ENFET enzyme FET
IMFET an ISFET with an immunological reagent
BIOFET an ISFET with a biologically active surface coating.
Most commonly, inorganic devices are referred to as ISFETs with their eld
of application specied. To appreciate their operating principles it is helpful
to know a little chemistry.
3.4.1 Ions in solution
When an electrode is placed in an ionic solution a reaction takes place
between the two so that a resulting equilibrium potential dierence is pro-
duced. This potential dierence between electrode and solution depends
mainly, but not entirely, on the activity a of a particular ionic species. (In
some situations the presence of a similar ion may modify the behaviour of
the desired electrodeion interaction.) The activity a of an ion in solution
depends on the concentration c,
a Kc. 3.19
However, K is not constant but is a function of c and the concentrations of
other ions present. At low concentrations, <10
4
mole litre
1
, K 1. The
modied Nernst equation relates the The venin emf to a,
E
Theevenin
E
0
RT
zF
lna
x
ka
y
3.20
where E
0
is a constant depending on electrode composition, R is the universal
gas constant (8.314 J K
1
), z is the number of charges per ion (1, 2 or 3), T is
the absolute temperature, F is the Faraday number (96 493 Coulombs), a
x
is
the activity of the measurand ion x, a
y
is the activity of ion y if present, and k
is the selectivity of the electrode to y; 0 k 1.
E
Theevenin
is the potential dierence produced between the single active
electrode and the solution in which it is immersed. Another electrode also
has to be used in the solution in order to complete the circuit. The emf pro-
duced by this second electrode must be constant irrespective of the solution
under test. It behaves in eect as a reference electrode delivering a xed
potential dierence to the circuit.
3.4.2 Measurement of pH
One of the ionic activities most frequently measured, particularly by the
process industries and when monitoring the environment, is that of the
hydrogen ion. This is a measure of the acidity/alkalinity of the solution.
The dening equation for pH is
pH log
10
a
H
3.21
where a
H
is the activity of the hydrogen ion. In pure water dissociation of the
water molecule occurs to a small extent,
H
2
O , H
OH
3.22
resulting in a hydrogen ion concentration of 10
7
moles per litre at 25 8C. If
the temperature rises, more dissociation occurs, increasing the hydrogen ion
concentration. This would give the impression that the solution had become
more acidic. However, as long as the H
and OH
ions remain in equal

concentrations the solution stays neutral. Consequently, for the pH meter
to output the value 7, indicating neutrality, a temperature-dependent correc-
tion must be applied. Generally, therefore, pH meters have some form of
temperature compensation system. Table 3.1 shows the pH values for three
types of ionic solution.
The pH of a solution can be determined using electrodes and applying
the Nernst equation. Assuming the sensing electrode is highly selective and
only responds to a
H
, equation (3.20) can be written as,
E
Theevenin
E
0
RT
zF
ln a
H
for the hydrogen ion z is 1 3.23
E
Theevenin
E
0
RT
F
ln a
H
3.24
E
Theevenin
E
0
RT
F
ln 10
pH
3.25
E
Theevenin
E
0
2.303
RT
F
pH 3.26
which at 25 8C gives
E
Theevenin
E
0
0.0592 pH volts 3.27
Table 3.1. Hydrogen ion concentration and pH values.
Ionic solution Hydrogen ion concentration
(moles per litre)
pH
Pure water at 25 8C 10
7
7
Strongly acidic Tends to 10
0
Tends to 0
Strongly alkaline Tends to 10
14
Tends to 14
Equation (3.27) has been in use for a number of years for measuring pH. The
traditional system uses a sensing electrode housed in a permeable glass
membrane. As described above, to measure the potential dierence
between electrode and solution a further electrode is required whose
E
Theevenin
does not change with the constitution of the solution. The output
of this reference electrode E
ref
is included in the calibration. These glass
electrode instruments generally work well once conditioned but are bulky,
fragile and expensive.
A new type of pH microsensor is now commercially available based on
the MOSFET. Similar devices are also available for the measurement of
other ions in solution, the family of sensors are the ion specic FETs or
ISFETs.
3.4.3 The ISFET
The ISFET was the rst attempt at using FETs as sensors and was initially
proposed to measure pH by Bergveld in 1970. To help appreciate the
operation of the ISFET it is useful to consider the operation of the
MOSFET from which it was developed.
The acronym MOS describes the material sandwich in the construction
of the FET, metal-oxide-semiconductor (gure 3.9). The semiconductor may
be p- or n-type doped, to produce a PMOS or NMOS device. NMOS devices,
using n-channel enhancement, have a faster switching action due to the holes
in the p-type substrate being more mobile. This is not usually of any
signicance in sensor technology but is important in high-frequency
electronic devices. NMOS is also preferred in very large-scale integrated
circuit devices (VLSIs), since packing densities can be higher.
The constructional details of the n-channel, enhancement mode, NMOS
FET are summarized below.
Figure 3.9. The NMOS FET.
.
The p-type substrate is lightly doped so its resistance is high.
.
The source-substrate np junction is of order 10
10
unless forward
biased.
.
Normally the source and substrate connections shown on the diagram are
connected internally.
When a positive potential is applied to the gate relative to the substrate,
electrons are attracted to the surface of the p-type substrate providing a
low resistance channel of order 1 k. The gate potential thus controls the
current between drain and source. The resistance between gate and substrate
is very high so that there is negligible gate current. The drain-source current
I
DS
is given by
I
DS

j
m
C
0
wU
DS
L
U
GSu
U
t

1
2
U
DS
3.28
where U
GSu
is the potential dierence between gate and substrate, U
t
is the
threshold potential dierence of the FET, j
m
is the eective mobility of
electrons, C
0
is the capacitance per unit area of the gate structure, w is the
width of the gate, L is the length of the gate, and U
DS
is the potential
dierence applied between drain and source.
Figure 3.10 shows how the NMOSFET is adapted to form the
ISFET.
Positive ions (cations) from the solution are absorbed into the porous
membrane, which functions as the gate. These induce a negative charge
surface layer on top of the p-type silicon thus providing a conducting
Figure 3.10. The ISFET structure.
channel between the source and the drain. As with the FET, the ISFET
substrate and source are at the same potential so that the np junction
does not become forward biased. As described above, the FET structure
can have a p- or n-type substrate and, depending on the type of ion, cation
() or anion (), the current I
DS
either increases or decreases with ion
concentration.
Initially it was thought that an ISFET could be constructed by simply
omitting the metal gate electrode. In this way the SiO
2
insulating layer
would absorb hydrogen ions that would in turn change the biasing of the
FET and so produce, eectively, a pH-sensitive device. However, using
SiO
2
alone, sensitivity was found to be only 30 mVpH
1
and not suciently
stable when exposed to electrolytes. Consequently, a layer of silicon nitride
(Si
3
N
4
) is now commonly used on top of the SiO
2
to act as the absorbing
membrane. This device yields 50 mVpH
1
linearly from 1.8 to 10 pH. The
sensitivity can vary depending on the thickness and the manufacturing
process. Other oxides, such as those of the more exotic metals, have been
examined as the active surface layer and some give 50 mVpH
1
. Alumi-
nium oxide, Al
2
O
3
, has also been used but gives a rather slower response
compared with the silicon compounds.
In the early 1970s it was thought that a reference electrode was not
essential. However, this is not the case. The reference electrode ensures the
device is biased at its optimum operating point.
The theoretical relationship for the drain-source current is obtained by
combining the Nernst equation with the FET equation
I
DS

j
m
C
0
wU
DS
L
U
GSu
U
RT
zF
lna
x
ka
y
E
ref
1
2
U
DS
3.29
where E
ref
is the potential dierence of the reference electrode. A practical
conguration of the structure shown in gure 3.10 is depicted in gure 3.11.
Figure 3.11. A practical ISFET structure.
The depositing of dierent electro-chemically active materials on top of
the silicon nitride enables other ions to be detected. For example, sodium
alumino-silicate produces a sensor for detecting Na
ions.
3.4.4 FET gas sensors
MOSFETs having a palladium or platinum coated gate can be used to
construct a hydrogen gas sensor. Palladium causes the hydrogen molecule,
H
2
, to dissociate on its surface. These atoms then diuse through the
palladium and at the metalinsulator boundary become polarized. This
provides an electric eld that shifts the threshold voltage of the FET.
Unfortunately, the rate of diusion is slow and an integral heater is often
employed to increase this. The metallic layer is operated at temperatures
up to 150 8C.
Devices for sensing more complex species are continuing to be devel-
oped based on the construction of sophisticated adaptations to the pH and
hydrogen ISFETs. For example, the CO
2
sensor shown in gure 3.12 is a
pH sensor coated with a membrane which is permeable to CO
2
. A small
cavity dissolves the diused CO
2
to form the weak carbonic acid, H
2
CO
3
,
and the resulting acidic gel is measured for pH to determine the quantity
of acid.
By the judicious choice of materials, their thicknesses, the metal gate
structure and even including measurement of the operating temperature,
gas sensors have been constructed for
hydrogen H
ammonia NH
3
hydrogen sulphide H
2
S
carbon dioxide CO
2
hydrocarbons C
x
H
y
where y 2x 2
alcohol vapours C
x
H
yz
OH
z
Figure 3.12. The carbon dioxide ISFET sensor.
3.4.5 Biosensors
The design strategy is to take an established ISFET and surround it with
a selective membrane that allows through the desired species. The
production of these highly selective membranes is by the LangmuirBlodgett
technique. First a complex organic or polymer molecule is developed that
is permeable to the species to be detected and has attached hydrophobic
and hydrophilic radicals. These molecules are oated on pure water as
mono-molecular layers. The hydrophilichydrophobic radicals ensure
that the molecules are correctly orientated as they oat. The FET is
dipped into the surface and a single molecular layer is deposited across its
surface as it is withdrawn. To aid the process a moving boom is used to
maintain constant surface tension so that gaps or overlaps in the coating
do not occur. Multiple layers can be deposited and may even be of dierent
materials or orientations. In this way a highly selective molecular sieve is
produced.
This membrane forms a small chamber over the ISFET. The chamber
contains a reagent chosen to react with the biological species to be measured
so as to produce a specic by-product. The by-product, which might for
example be NH
3
, CO
2
or a change in pH, is then detected by the ISFET. Bio-
sensors of this type have been developed for measuring urea, glucose and
penicillin. It is interesting to note that biosensors are far more selective
than ordinary chemical sensors. This is entirely due to the superiority of
natural systems over the engineered. Unfortunately, the typical life spans
of such sensors are measured only in days (an ENFET for penicillin has
been produced which remains active for up to 2 months).
An exciting application of the biosensor is its use in vivo. However, a
design consideration when sensors are used within the blood stream is that
proteins may aect the specication of the device. A further, more serious,
concern is that the presence of the sensor may also encourage the formation
of thromboses.
3.4.6 Conclusion
A small number of FET based sensors are available commercially. This is
certain to be the beginnings of a major industry, the limitations to the
variety of sensors being that of designing appropriate membrane barriers
with reasonable longevity. The advantages of FET sensors over the
traditional glass ISEs are as follows.
.
Their low output impedance simplies the signal conditioning.
.
Their small size enables localized measurement, the use of small amounts of
reagents, and use with catheters in medical applications.
.
Their fabrication using well-established, semi-conductor industry, tech-
niques enables multiple sensing on one chip for dierent species, the
reference electrode to be fabricated within the package, and temperature or
light sensing on chip, for compensation.
.
Faster response.
.
Batch fabrication.
3.5 Optical sensors
3.5.1 The waveparticle duality
It is normal practice to classify electromagnetic radiation in terms of its
wavelength. Light is the visible part of the electromagnetic spectrum, having
wavelengths in the range 400700 nm. The near infrared is taken as 0.7
1.7 mm, the main infrared as 1.730 mm and the far infrared as 301000 mm.
Ultraviolet light has wavelengths in the range 10400nm. The near visible
and the shorter wavelengths are sometimes referred to as soft and hard or
vacuum ultraviolet respectively. However, depending on the circumstances,
electromagnetic power may be considered as either the arrival of a continuous
wave or of a streamof particles. This dichotomy applies to all particles whether
they have mass or not and is known as the waveparticle duality.
Isaac Newton (16421727) was a believer in the particle or corpuscle
theory of light, but diraction and interference experiments, such as those
of Thomas Young (17731829), reinforced the wave concept. However, in
1899 Lummer and Pringsheim conducted some experiments that could not
be explained using a classical wave-model. The power spectral density of
radiation from a perfectly emitting, or black-body, was measured at
various temperatures (gure 3.13).
The theoretical studies of Wien (1896) and Rayleigh (1900), based on the
classical statistical physics laws of Maxwell and Boltzmann, failed to com-
pletely account for these curves. It was Planck in 1901 who postulated that
energy could only be absorbed/emitted by the black-body in discrete
amounts he called quanta. He suggested that the transferred energy
quanta, E, was determined by
E hf 3.30
where h is Plancks constant and f is the frequency of the radiation. It was Ein-
stein in 1907 who extended the concept further to say that energy in any system
went in discrete intervals. He also gave the name photons to light quanta.
3.5.2 Denitions and units
It would, at rst, seem sensible to use equations like the one above and
Poyntings vector (see section 4.4.4),
P E ^ H Wm
2
3.31
Optical sensors 69
where E is the electric eld strength in Vm
1
and H is the magnetic eld
strength in Am
1
, to dene light units. However, the subject is complicated
because the eye, like the ear, does not have a at spectral response. Thus, for
example, light from an argon ion (green) laser will appear to be brighter than
light from a neon (red) discharge lamp of the same intensity. Consequently, a
comparison between intensities is only meaningful if it is made in a restricted
range of the spectrum. Historically, photometry has relied on standards
based on a dened light source. No account is taken of the spectral details
of the source: it is simply accepted as being spread throughout the visible
spectrum in a standard way. The SI denitions and units for photometry
given below have developed from these historical techniques.
The SI photometry units
Luminous intensity (I)the candela (cd)
The candela is the luminous intensity perpendicular to the surface of a tiny,
black-body emitter. The emitter is dened as having a surface area of
1
6
10
5
m
2
and being at a temperature at which platinum freezes under a
pressure of 101 325 Nm
2
(this corresponds to a temperature of 2045 K).
Luminous ux ()the lumen (lm)
Flux is an old English word meaning ow or discharge. The ow of light
from a hypothetical point source is said to be isotropic, that is, it emits
Figure 3.13. Power spectra of the radiation from a perfect black-body.
light uniformly in all directions. (Real light sources are not isotropic
emitters.) The amount of light owing in a given direction is the radiant
power and is termed the luminous ux. The lumen is that ux owing
from the dened point source of one cd within a cone having a solid angle
of one steradian (). Generally,

I d lm 3.32
For an isotropic point source this would give the total ux as
total
4I lm 3.33
Illuminancethe lux (lx)
This is the irradiance, the amount of light falling on a surface. One lux of
illumination is produced by one lumen falling on an area of one square
metre, i.e.
1 lx 1 lmm
2
3.34
Typical values recommended for lighting installations are 500600 lx for
oces down to 100 lx for entrance porches when measured horizontally
1 m above oor level.
Luminance (cd m
2
)
The brightness or radiance of a surface, either because it is self-luminous or
illuminated, is termed its luminance. The SI unit of luminance is the candela
per square metre.
The sensitivity of the eye
As mentioned above, the sensitivity of the eye is wavelength dependent and
varies slightly between individuals. Photometric matching experiments have
been carried out with numbers of subjects in order to relate the luminous
ux perceived , to the actual spectral power distribution received
p
. The
relationship can be expressed as
K
m
1
0
p
V` d` 3.35
where V` is the relative luminous eciency of the eye and K
m
is a constant.
The Commission International de lEclairage (CIE) in 1971 recommended an
agreed set of values for V`. These values are sometimes referred to as the
standard eye-ball. The form of V` is shown in gure 3.14.
Optical sensors 71
Practical photometry units
Researchers in the various optoelectronic elds tend not to refer to the SI
photometry units because of the inconvenience of having to have standard
sources and detectors of dened spectral response. Also, since much
optical research is based on narrow spectral bands often in invisible parts
of the spectrum, using photometry units is inappropriate. Consequently,
researchers working at a particular narrow band of wavelengths will
simply measure the electrical output of the detector. Then, knowing its
active surface area and its eciency at these wavelengths, the light
intensity in Wm
2
can be determined. Researchers commonly use the
following units which are based on the cm
2
as a unit area. This of course is
not an SI unit.
Power spectral intensity of a source S`
This is the power emitted per unit surface area of a source between ` and
` c`, into a cone of unit solid angle. In practice the units Wcm
2
mm
1
sr
1
are commonly used for S`.
Brightness B
This is the total power emitted by a source over all wavelengths, per unit
surface area, into a cone of unit solid angle:
B
1
0
S` d` Wcm
2
sr
1
. 3.36
Figure 3.14. The agreed relative spectral response of the eye.
Radiant ux R
This is the total power emitted from a source into a cone of unit solid angle
between ` and ` c`
R A
s
S` Wmm
1
sr
1
3.37
where S` is assumed constant over the surface area A
s
of the source.
The diuse source
An extended, diuse source is one that behaves like an aggregate of point
sources and so has an equal brightness when viewed from any angle. Such
extended sources are sometimes said to be Lambertian. However, because
of this aggregate of point sources, the luminous intensity is no longer
equal in all directions. Instead a cosine dependence occurs and is termed
Lamberts cosine law,
I
c
I
0
cos c cd. 3.38
Also, if a detector D is placed at some angle 0 to this direction, then the
illumination it receives is further reduced by a factor of cos 0, as shown in
gure 3.15.
3.5.3 Light sources
A large number of optical sensors are available each having a particular
spectral sensitivity, speed of response, physical size and cost. Before
examining light sensing devices it is worthwhile considering the types and
nature of light emitters available. In this way an appropriate choice of
sensor can be made.
Incandescent objects
Bodies such as the sun, lament lamps and molten metals give out a
continuous spectrum of radiation similar to that of the perfect black-body
Figure 3.15. Lamberts cosine law.
Optical sensors 73
(gure 3.13). However, the closeness of the emission to that of the black-body
can be poor, especially at lower temperatures. Consequently, a correction
factor called the emissivity can be dened, which is both wavelength and
temperature dependent,
`. T
actual radiation
black-body radiation
. 3.39
The emissivities for some materials are illustrated in gure 3.16.
Filaments in lament lamps are usually made of tungsten. The lamps
glass or quartz envelope will lter the emission spectrum of the hot tungsten.
Also, the tungsten slowly evaporates during its working life and forms a
coating on the inside of the envelope. This coating also absorbs light. It
has become common practice to add a small trace of a halogen, such as
iodine, into high performance lamps. Tungstenhalogen compounds are
volatile and re-evaporate from the glass surface whenever it is switched on.
These materials also aect the transmitted spectrum.
The following relationships may be useful in determining ideal lament
performance.
(a) Plancks equation gives the power emitted per unit area of source,
per unit wavelength from a black-body at an absolute temperature T,
G
C
1
expC
2
,`T 1`
5
Wm
2
,m 3.40
where C
1
is 3.7418 10
16
Wm
2
and C
2
is 1.4388 10
2
mK. This relation-
ship generates the black-body power spectral density curves of gure 3.13.
The power, P, emitted per unit surface area from the whole spectrum is
the integral of G over all wavelengths,
P
1
0
Gd` 3.41
Figure 3.16. The emissivities of some materials.
which can be shown to give
P oT
4
Wm
2
. 3.42
The term oT
4
, is the StefanBoltzmann law for a black-body.
(b) The StefanBoltzmann law describes the radiated power, P, per
unit area of emitter surface. The presence of another body, above absolute
zero, near the radiator will cause the radiator to absorb power. The
emissivity constant also applies to absorption by the non-black-body. The
total, net, radiated power is
P oT
4
B
T
4
0
Wm
2
3.43
where `. T is the emissivity, o is the StefanBoltzmann constant
(5.670 10
8
Wm
2
K
4
), T
B
is the absolute temperature of the body, and
T
0
is the surrounding absolute temperature.
(c) Wiens displacement law describes the shift of the spectral peak `
p
of the function G with the absolute temperature T,
`
p
a,T 3.44
where a is a constant of value 2.892 10
3
mK
1
.
The black-body function G for some incandescent source can be related
to the power spectral intensity function S`. If the source is emitting
isotropically into a hemisphere then
S
G
2
10
10
Wcm
2
mm
1
sr
1
. 3.45
Discharge lamps
When light is emitted from isolated ions, atoms or molecules, then line or
band spectra are produced. These are precise characteristic frequencies and
have little temperature dependence. Common examples are neon and
sodium lamps. In a discharge lamp excitation of the gas is by collision
with free electrons. Ions and atoms while excited have their outer electron(s)
in higher orbitals. Molecules in addition to this can have rotational and
vibrational energy. When the ions, atoms or molecules return to their
ground states, quanta are released of characteristic wavelength. In high-
pressure arc lamps a far greater number of excited states are possible,
resulting in a spectral distribution graph rather than just line spectra.
These distributions approach the black-body spectrum to a greater or
lesser extent with the additional line spectra. The ultraviolet content is
higher in high-pressure arc lamps than in the black-body spectrum. The
exact nature of these spectral distribution graphs depends on the gas
composition, its pressure, and the type of discharge and lamp geometry.
Optical sensors 75
Light-emitting diodes (LEDs)
These are pn junction diodes which, when forward biased, emit optical
radiation. Holes and electrons in the junction region recombine, thus
dropping from the conduction band to the valence band. The photon
released corresponds to this energy reduction. The process is called injection
luminescence since it is caused by the injection of charge carriers into the
junction region when forward biased. Schockley derived the equations for
an ideal diode: the current i
d
is given by
i
d
i
s
exp
eV
d
kT
3.46
where i
s
is the reverse saturation current, e the charge of the electron, k
Boltzmanns constant, T the absolute temperature and V
d
the bias voltage.
When forward biased, equation (3.34) approximates to
i
d
i
s
exp
eV
d
kT
. 3.47
The luminous intensity is approximately proportional to the forward current.
A number of devices are available constructed from combinations of
aluminium (Al), gallium (Ga), arsenic (As), phosphorus (P) and indium
(In), each having its own spectral distribution peaking at the wavelength `
p
:
GaP `
p
0.560 mm green
GaAsP `
p
0.655 mm red
GaAlAs `
p
0.80.9 mm near infrared
InGaAsP `
p
1.31.6 mm near infrared
There are two basic geometries giving rise to surface emitting and edge emit-
ting types. The edge emitters give a much narrower beam. Turn on/o times
are typically 10 ns giving modulation bandwidths of up to 10
8
Hz. Luminous
intensities are of the order of 1 mcd. Infrared is preferred in optical bre
applications because of its better transmission through silica, compared
with visible light. This is described further in section 4.2.2. Typical operating
conditions have i
d
set at about 20 mA, and V
d
about 1.6 V; total light output
is between 1 and 10 mW. Figure 3.17 shows the emission spectrum of
a GaAlAs LED. The brightness B is the area under the curve and is
approximately 10
2
Wcm
2
sr
1
.
LASERs (Light Amplication by the Stimulated Emission of Radiation)
The lasing medium used in the rst laser was a crystal of ruby in rod form. It
was pulse-energized from a ash lamp. Lasers are now available with lasing
media made of gas, dye-solution or semiconductor. The ions, atoms or
molecules of the lasing medium have to undergo a population inversion.
This means that they are put into an unstable high-energy state. They can
spontaneously return to their ground state by emitting a photon. But they can
be triggered or stimulated to do so by photons of the same type as to be
emitted. Consequently, an avalanche eect can occur. Highly reecting
mirrors are mounted at each end of the lasing cavity to ensure that this
inecient process continues. Photons emerge in-phase and usually highly
collimated providing an intense, monochromatic, coherent beam. In gas
lasers bandwidths may be less than 10
5
nm. Semiconductor laser band-
widths are not as narrow. Also, the total light output may only be about
the same as the ordinary LED due to the smaller emitting area. The bright-
ness B in gure 3.18 is approximately 10
5
Wcm
2
sr
1
.
It is important to remember that all LED devices are adversely aected
by temperature. The intensity variation is
I I
0
expc0 3.48
where c is the temperature coecient of intensity, and 0 the temperature in
degrees Celsius. Also, the spectral peak `
p
shifts with temperature according
to
`
p
0 `
p
0 u0 3.49
where `
p
0 is the spectral peak at 0 8 C. In GaAsP devices, for example, c is
10
3
K
1
and u is 0.193 nmK
1
. In addition, laser LEDs suer from changes
to the optical cavity caused by temperature variation. This can result in
changes to `
p
as well as mode hopping, a change of the standing wave
pattern inside the resonant cavity.
Figure 3.17. Spectral output of a GaAlAs LED.
Optical sensors 77
3.5.4 Sensor types
When designing an optical measurement system the following criteria may
need to be considered:
.
sensing at the appropriate wavelength,
.
inherent noise in the sensor,
.
linearity,
.
speed of response for the desired bandwidth,
.
stability,
.
compatibility with the overall system, particularly the optical geometry.
Photodetectors can be broadly classied into two groups: thermal detectors
and quantum detectors. In thermal detectors the absorption of photons
causes bulk warming of the sensor. These sensors have a very wide operating
range of wavelengths, are relatively slow and have a low sensitivity (see
section 3.6). In quantum detectors individual photons must deliver enough
energy to overcome some threshold to release charge, the photoelectric
eect. There are two groups of quantum detectors:
(a) external, or surface, photoelectric eect devices, the prime example of
which is the photomultiplier, and
(b) internal photoelectric eect devices; the two types of semiconductor
photodetector are the photoconductor and the junction photon detec-
tors, the photodiode and the phototransistor.
Figure 3.18. Spectral output of a GaAlAs injection laser diode.
(a) Surface photoelectric detectors
The photomultiplier
These are extremely sensitive devices used principally in research. They are
vacuum tube structures (see gure 3.19) requiring a stable, high voltage
d.c. supply of a few kV. A photon, passing through the glass window,
strikes the cathode and may liberate an electron from the surface (photo-
electric eect). There are a series of electrodes along the inside of the tube,
called dynodes, at progressively higher voltages. The photo-electron
accelerates towards the rst dynode and on striking the surface has sucient
kinetic energy to liberate several more electrons. These then accelerate to the
next dynode and the process repeats. Avalanche multiplication occurs. The
resulting photocurrent is collected at the anode. The photo-active surface
is made from exotic alloys of the reactive metals. The dynodes are also
coated in similar materials. Photomultipliers are made with diering
overall gains and spectral responses and are usually too sensitive to be
operated in daylight. Some are sensitive enough to detect the arrival of a
single photon. They have high linearity, very high bandwidth and low dark
current. Their advantages when comparing with semiconductor devices are
their ultraviolet response and large detecting area.
(b) Semiconductor photodetectors
These are devices in which a photon from the incident light beam causes an
electron to be excited from the valence band to the conduction band. The
detector will only respond to photons whose energy is
E
G
hf or hc,` 3.50
where E
G
is the energy gap between the valence and conduction bands. The
two main types of detector are described below.
Photoconductive detectors
As the name suggests, these devices behave like variable resistors and are
sometimes called light-dependent resistors. When the material is irradiated,
if the incident photon has an energy E
G
, then an electronhole pair is
Figure 3.19. Schematic structure of a photomultiplier.
Optical sensors 79
produced. The electron and hole drift to the appropriate electrodes under the
inuence of the applied electric eld. The resulting photocurrent aects
the conductivity of the material which magnies the device current. The
relationship between resistance R, and total incident light power is extremely
nonlinear
log
10
R a b log
10
PA 3.51
where P is the intensity, A is the sensing area and a, b are constants. The three
most common devices and the wavelengths of their response peaks `
p
are
cadmium sulphide CdS `
p
0.53 mm visible
lead sulphide PbS `
p
3.0 mm infrared
indium sulphide InSb `
p
5.3 mm infrared.
Response times are long, typically 50 ms. Dark resistance is of order 1 M
and falls nonlinearly with increase in illuminance to approximately 1 k.
Temperature stability is poor and depends on composition, fabrication
method and light level. Also, the zero value drifts with time by up to 20%
over 1000 hours. The application of these devices is therefore limited, since
they are not suitable for fast response or precision measurement. They are,
however, reasonably sensitive. CdS light-dependent resistors, for example,
have a gain of 10
3
. The gain is dened as the ratio of
the number of electrons flowing
number of electrons generated, per unit time
(holes have a short half-life and so can be ignored). They are used in optical
switching systems, light alarms and photographic exposure meters.
Junction photon detectors
There are two classes of semiconductor junction device: the photodiode and
the phototransistor. The photodiode is the most popular type with the PIN
photodiode being the most widely used device.
Photodiodes
If the photon has sucient energy, and if an interaction occurs, an electron
hole pair is produced. The pn junction eld sweeps these apart to the
appropriate electrodes. The ideal diode equation for any pn junction
diode was given above and is repeated here:
i
d
i
s
exp
eV
d
kT
. 3.52
The equivalent circuit for the photodiode is shown in gure 3.20; internal
noise generation sources have been omitted. The photocurrent i
p
created
by the incoming photons can be expressed as
(number of photons arriving/sec) (conversion efficiency)
(charge of the electron)
This can be re-expressed in terms of the incident light power as
(incident power) (conversion efficiency) (charge of the electron)
energy per photon
.
Expressing this algebraically gives
i
p
ejPA
`
hc
3.53
where j is the quantum eciency, the number of electrons collected per
incident photon, e is the charge of the electron, P is the incident light intensity
in Wm
2
, A is the surface area in m
2
, h is Plancks constant, and c is the speed
of light.
Not all of the generated photocurrent i
p
reaches the diode terminals.
From the diagram of the diode equivalent circuit (gure 3.20), using
Kirchhos current law gives
i
p
i
d
i
R
j
i
C
j
i
ext
. 3.54
Note that the forward biased terminal current is i
ext
so that increasing
the light intensity, which causes an increase in i
p
, results in a reduction in the
forward terminal current. The illumination has the eect of lowering the
currentvoltage characteristic as shown in gure 3.21.
The advantages of photodiodes when compared with photoconductive
cells are their higher sensitivity, faster response time, smaller size, better
stability and, when biased appropriately, excellent linearity.
A photodiode is made to operate at a particular point on its character-
istic by the choice of biasing. Generally, diodes may be operated with or
without an external bias voltage. A dierent performance results for each
of the three possible biasing congurations. Biasing is usually achieved via
Figure 3.20. Photodiode equivalent circuit.
Optical sensors 81
the signal conditioning circuit. These are described below with their relative
merits.
Photodiode signal conditioning circuits
Photovoltaic mode
Without an external bias the device is operating in the photovoltaic mode.
Figure 3.22 shows a simple and buered signal conditioning circuits.
The incident radiation generates a voltage across the junction, the
electronhole pairs being swept apart by local elds. The value of R
L
is
chosen to be very high, so that i
ext
is practically zero. For low frequencies
leakage through C
j
will be negligible so that i
C
j
can be assumed to be zero.
i
p
i
d
i
R
j
3.55
i
p
i
s
exp
eV
d
kT
V
d
R
j
3.56
Figure 3.21. The photodiode characteristic.
Figure 3.22. Signal conditioning circuits for the photovoltaic mode.
and re-arranging gives
exp
eV
d
kT
1
i
p
i
s
V
d
i
s
R
j
. 3.57
Typically, the photocurrent will be of the order of 10 mA, whereas the reverse
saturation current is of order 10
8
A, so that i
p
)i
s
. Also, V
d
will be
approximately the same size as the product i
s
R
j
, so equation (3.57) can be
written
exp
eV
d
kT
i
p
i
s
3.58
eV
d
kT
ln
i
p
i
s
3.59
V
d

kT
e
ln
ejPA`
i
s
hc
3.60
where V
d
is the output potential dierence across R
L
and has a logarithmic
response with P and `. Figure 3.21 gives the output relationship when an
op-amp is used.
In the photovoltaic mode of operation the frequency response is only
moderate but the noise is low and is mainly thermal noise, which is spectrally
at. This biasing approach is thus preferred for low noise applications at low
frequencies. Below 100 kHz the photovoltaic mode provides a better signal to
noise ratio than the photoconductive biasing mode described below and, at
frequencies of less than 1 kHz, is far superior due to lack of 1,f noise. A
disadvantage is that the output voltage varies logarithmically with light
intensity as equation (3.60) shows. However, this may be put to good use,
Figure 3.23. Simple signal conditioning circuits for biased photodiodes.
Optical sensors 83
for example in the measurement of liquid depth. Absorption is an exponen-
tial relationship so having a logarithmic sensor response would give a linear
output. A further most important application of diodes in the photovoltaic
mode is as a solar cell. However, the physical structure of solar cells is
dierent since large areas are needed and power transfer is required.
Photoconductive/photoamperic modes
With an external reverse bias the device is operating in the photoconductive
mode. The junction eld is increased, reducing charge carrier transit times
and junction capacitance because of the increase in the width of the depletion
region. Thus high-frequency performance is improved. Thermal noise and
shot noise are present with 1,f noise becoming dominant below 1 kHz.
If a bias of zero volts is applied, this is referred to as the photoamperic
mode. This mode has very low photodiode leakage current but is slower than
photoconductive. Photodiode sensitivity is linear in both of these modes.
For reverse bias operation the diode current i
d
is the leakage current i
s
,
and is typically 10
8
A. (If the bias is zero then i
d
is zero.) The reverse bias
voltage applied will be of the order of 10 V and, since r
s
is small, this will
be the same as V
d
. R
j
is typically in the region of 10
8
. Consequently, i
R
j
,
which is V
d
,R
j
, will be approximately 10
7
A. The current i
C
j
will be small
because of the decreased capacitance so to a good approximation equation
(3.54) becomes
i
p
10
8
10
7
i
ext
. 3.61
Figure 3.24. Spectral response of photodiodes when operated in the photoconductive
mode.
Thus, for a typical photocurrent of 10 mA, the external current will eectively
be equal to the photocurrent,
i
ext
ejPA
`
hc
3.62
which shows a linear response with P and ` and is sometimes written as
i
ext
cPA 3.63
where c ej`,hc and is termed the responsivity with units of AW
1
. The
spectral response is shown in gure 3.24.
There are four main types of photodiode. These are briey described
below.
The pn junction photodiode
The descriptions in the previous paragraphs
relate to the pn junction diode. The
semi conductor material may be of silicon or
of GaAsP.
PIN photodiode
This diode has an intrinsic layer of silicon
between the extrinsic p and n layers. This
lowers the device capacitance and produces a
much faster response than for the pn junction
photodiode, being typically 0.4 ms. Another
reason for its popularity is that the intrinsic
region can be tailored to optimize the desired
quantum eciency and bandwidth.
Schottky photodiode
An extremely thin metal lm (of order 10
8
m
thick) forms a Schottky barrier with an n-
type semiconductor. The thin metal lm
enhances the response to the ultraviolet end
of the spectrum. An anti-reection coating is
essential to reduce losses. With the right
choice of metal and coating quantum ecien-
cies can reach 70% and response times in the
region of 1 ns.
Optical sensors 85
Avalanche photodiode
This device is operated under high reverse bias
voltage. The liberated electronhole pairs
accelerate to such kinetic energies that, when
collisions with other atoms occur, secondary
charge carriers are produced. Avalanche
photodiodes have ultra-fast response times
and can operate up to GHz frequencies. The
breakdown voltage for these devices when fabricated in silicon is in the
region of 100 V. Photocurrents of up to 1 mA can occur when operated
near this potential dierence. Avalanche gain may be between 100 and
10 000 and can substantially reduce the noise equivalent power because the
gain is achieved within the device and not subsequently in some signal con-
ditioning circuit. This gives these photodiodes a signicant advantage over
non-avalanche types. However, the power supply voltage needs to be
stable to 20 mV, in order to maintain constant gain.
Phototransistors
Phototransistors detect photons in the same way as photodiodes, the enlarged
collectorbase junction acting as the sensing region. For an npn transistor, if
the base is oating then the holes created forward bias the baseemitter
junction. The bipolar-transistor, common-emitter mode, current-gain para-
meter h
FE
amplies the photocurrent. The overall current is 1 h
FE
i
p
.
Thus, phototransistors are more sensitive than pn photodiodes, but their
response is less linear. Also, the capacitance of the forward-biased, emitter
junction causes a slow response of typically 15 ms and so frequency perfor-
mance is poor. Also, because the collector current is temperature-sensitive,
phototransistors are less temperature stable than photodiodes. Photo-
transistors manufactured as three-terminal devices can be biased and oper-
ated in the normal bipolar fashion giving greater exibility in system
design. For this reason the phototransistor may be a more appropriate
optical sensor than the photodiode in some applications. They are often
used in reading, writing and encoding systems. Photodarlington congura-
tions are available. Here the phototransistor emitter current is cascaded
into the base of a normal bipolar transistor, giving a multiplicative current
gain. Typically, photodarlington devices are ten times more sensitive but
ten times slower.
Commercially available devices
Table 3.2 lists some devices with an indication of their prices as advertised at
the beginning of the year 2002. Price depends on a number of characteristics
other than the few listed. These include the size of the detecting area,
eciency, additional optical lters or lenses, the angle of acceptance, level
of dark current and specic spectral response. All of these vary between
manufacturers. Purchasers need to study individual data sheets carefully.
The charge-coupled device optical sensor
Charge-coupled devices (CCDs) are a family of silicon semiconductor
devices capable of performing the general functions of analogue data
handling and digital or analogue memory. The CCD concept was invented
in 1969 at Bell Labs by George Smith and Willard Boyle whilst researching
into new forms of computer memory. The combination of CCD with an
optical sensing array was fabricated soon after to produce a very versatile
optical sensor that has enjoyed rapid growth over the intervening years.
The two discrete parts, the optical sensing process and the charge-coupling
action, are described below.
Sensing
Sensing is achieved by the photoelectric eect. Free electrons are created at
sites in the silicon illuminated by photons in the spectral range 300 to
1100 nm. The response peak is around 800 nm. The device is constructed
having an array of these photosites which are typically 10 mm across. The
number of free electrons generated in each cell is directly proportional to
the intensity of the incident radiation on that cell. If the incident radiation
intensity pattern on the array of cells is a focused image then the amounts
of charge, or charge-packets, created in each cell will be a representation of
Table 3.2. A selection of semiconductor photodetectors.
Device / manufacturer Type Spectral
range
Response
time
Cost
()
NORP12 Photoconductor
(CdS)
`
p
530 nm rise 18 ms
fall 120 ms
1.51
VT935G / E G & G Vactec Photoconductor `
p
550 nm rise 35 ms
fall 5 ms
0.67
SFH213 / Siemens Si pn `
p
850 nm 5 ns 0.47
AEPX65 / Centronic Si pn `
p
800 nm 1 ns 14.49
BPX65 / Centronic Si PIN `
p
850 nm 3.5 ns 6.11
BPW34 / Siemens Si PIN `
p
850 nm 20 ns 0.73
OP993 / Optek Si PIN `
p
890 nm 5 ns 1.57
C30902E / EG & G Si avalanche 400100 nm 0.5 ms 120.36
SDP8405 / Honeywell Phototransistor `
p
850 nm 15 ms 0.71
BPV11 / TFK Phototransistor `
p
950 nm 3.8 ms 0.35
SFH309FA-4 / Inneon Phototransistor `
p
900 nm 10 ms 0.42
BPW38 / Fairchild Photodarlington `
p
850 nm 250 ms 1.30
Optical sensors 87
the viewed object. Each charge-packet corresponds to a picture element or
pixel. The array of contiguous cells may be one- or two-dimensional.
Charge coupling
Specic locations, called storage elements, are created in p-type silicon by
asymmetric doping in conjunction with the eld from a pair of gate elec-
trodes that are close to the surface of the silicon. Figure 3.25 illustrates a
Figure 3.25. A two-phase CCD shift register. (a) Physical structure and changing potential
well patterns. (b) Clock voltage waveforms.
two phase-type device. An insulating layer of silicon dioxide separates the
electrodes and storage elements. The storage elements form a contiguous,
linear array. The dierent concentrations of n-type dopant create the
required asymmetric electric eld.
The potentials on the gate electrodes can be alternately raised and
lowered so as to cause the individual charge-packets beneath them to be
passed from one storage element to the next. This is illustrated in gure
3.25(a) for the three cells A, B, C within the array. The two complementary
clock voltage waveforms c
1
and c
2
(gure 3.25(b)) are connected to the
alternate, closely spaced, gate electrodes on the surface of the thin insulating
layer of silicon dioxide.
At the start of the cycle c
1
voltage is HIGH and the charge packet of
electrons in storage element A is trapped by the potential well under gate
electrode 2. In the next half-cycle the potential well under gate 2 is reduced
due to c
1
having gone LOW. At the same time the adjacent electrode 3
connected to c
2
has gone HIGH. Consequently, the electron charge packet
is attracted to the new potential well under electrode 3. A further half-
cycle later the potential well under electrode 3 is reduced due to c
2
going
LOW and the electron packet moves to the new well under electrode 4
which has gone HIGH.
The line of storage elements thus operates as an analogue shift register
and, because the transfer of charge from one storage element to the next is
very ecient, the amount of charge in each charge-packet stays substantially
the same. Shift registers are available with up to 4096 contiguous elements.
The optical CCD device
Figure 3.26 illustrates the two processes of sensing and charge-coupling
within the optical CCD sensor. During an appropriate exposure time the
incident light generates electrons in each photosite, the number being propor-
tional to the exposure time and the light intensity. A feedback process sets the
exposure time. In low light levels this integration time is increased. In strong
light conditions the amount of charge generated can be so large as to cause
Figure 3.26. Schematic layout of an optical CCD sensor.
Optical sensors 89
leakage into adjacent photosites. When such saturation occurs the exposure
time is automatically reduced.
After an appropriate exposure time the charge-packets in each of the
photosites are all transferred in a single action to a CCD shift register
mounted alongside the photosites. This transfer action is also accomplished
by the charge-coupling process. A single, long electrode is used to provide the
eld switching-action to facilitate the transfer. This long gate-electrode is
called the transfer gate. After a transfer has taken place the transfer gate is
immediately returned to the non-transfer state so that the photosites can
begin integrating the next image frame. During the light integration period
the CCD transport register, now loaded with pixel data from the previous
frame, is rapidly clocked in order to shunt the information to the device
output circuitry. A serial set of analogue voltage pulses is output (gure
3.27).
The useful characteristics of optical CCDs are
.
able to operate in low light levels,
.
wide working temperature range of 55 to 70 8C,
.
wide spectral response,
.
low power consumption,
.
low operating voltage,
.
an output data rate of up to 20 MHz,
.
automatic exposure system.
Devices are available from a number of companies in one- or two-
dimensional format and with various numbers of cells and inter-cell spacings.
For project work a small development board can be bought at low cost
containing a one-dimensional array of 256 photosites, a cheap lens and
additional circuitry. Currently, a mid-range two-dimensional device having
512 512 pixels can output 25 frames per second. Higher resolutions of
4096 4096 can only produce 0.5 frames per second. There are some very
Figure 3.27. An optical CCD serial output signal.
high frame-rate devices (1000 fps) available with correspondingly poor
resolution (80 80 pixels). Colour imaging can be achieved in several ways.
Three-CCD system
The light beam is split using prisms so as to form three separate images. Each
beam has either a red, green or blue lter in its path before reaching its
particular CCD. Cameras using this approach produce rich, high resolution
images of high quality but are bulky, expensive and have high energy
consumption.
Light beam chopper
A single CCD chip is used. The image passes through a rotating wheel
containing the red, green and blue lters. This is not a popular option
since the colour information is produced sequentially and so high speed
imaging is not possible.
Pixel lters
The individual photosites each have their own red, green or blue lter
covering the entrance window. The lter layout is a mosaic pattern, the
most common style being the Bayer pattern (gure 3.28).
Here photosites in alternate rows contain red and green lters and blue
and green lters. The extra number of greens follows the spectral response of
the eye. A so-called de-mosaicing algorithm is used on the data stream that
renders an output colour tone for each pixel by averaging nearest neigh-
bours. In this way resolution is not lost. This colour imaging system is the
most popular in domestic quality cameras.
Variable depth photosites
A recent development employs the phenomenon that photons of dierent
colour penetrate to dierent depths in silicon. Blue is the least penetrating
and red the most. Photosites are stacked.
Figure 3.28. The Bayer lter pattern.
Optical sensors 91
This variety of resolution quality and corresponding costs has meant
that optical CCDs now enjoy a wide range of applications. These include
.
facsimile machines using one-dimensional devices,
.
high speed mail sorting,
.
rapid non-contact inspection for quality control,
.
computer-controlled material-handling systems,
.
small, rugged, low power TV cameras for security use and remote
inspection systems,
.
digital photography,
.
high performance video cameras.
3.6 Temperature sensors
The measurement of temperature is an important function in its own right.
However, it is also possible to measure other variables by using heat ow.
Such devices are termed thermal sensors. The primary sensor has its heat
content reduced or increased by the variable and the secondary sensor
measures the resulting temperature. An example of such a device is the
vacuum pressure sensor. The cooling of a heated element by conduction to
the enveloping residual gas causes a drop in its temperature. Element
temperatures are calibrated against pressures for a specic gas. Other
thermal sensors have been developed to measure ow, infrared radiation,
gas analysis, true-rms power, humidity and dew point. Two types of infrared
device are described below.
3.6.1 Thermal infrared radiation sensors
These devices operate on the warming eect of the incident radiation. This
distinguishes them from devices such as infrared photodiodes where the
quantum eects of individual photons is detected. However, they are
orders of magnitude slower and less sensitive than the quantum devices
but have a at spectral response. They are usually used to remotely determine
the temperature of incandescent objects. The energy absorbed causes a rise in
the temperature of the sensor that can be related, using Plancks law, to the
temperature of the hot object. Two common types of thermal detectors are
described below.
Thermopiles and thermocouples
A thermopile is produced by mounting a number of thermocouples in series.
All the hot junctions are tightly packed into the front or centre of the sensing
head and the cold junctions to the cooler rear or rim. A black coating is the
primary sensing element. Originally, bulk devices were not very precise and
expensive. Miniature, semiconductor versions have been developed. Com-
mercially available devices use a bismuth/antimony thermocouple (Elbel
et al 1985), or a silicon/aluminium thermocouple (from Xensor Integration).
Silicon-based thermocouples are particularly attractive for inclusion in
integrated circuit designs because standard integrated-circuit process tech-
nology can be used. Single-crystal, p-type silicon with aluminium generates
approximately 1 mVK
1
. A distinguishing feature of the thermoelectric
eect compared with other temperature measuring techniques is that a
temperature dierence is being measured. The cold junction will need to be
carefully located on the chip. However, temperature dierence is a variable
that might be useful as part of another sensor type, for example gas ow
as described in section 3.2.4.
Bolometers
These are devices that use a temperature-sensitive resistance. As above, a
black coating is the primary sensor. The resistive material can be a thin
lm or ake of a metallic or semiconductor material, which is mounted on
a non-conducting substrate such as a ceramic. The metals platinum,
bismuth and telluride have been used. Values of order 100 are typical.
To achieve this, thin coatings of the material are evaporated on to the sub-
strate in a pattern designed to give maximum track length. For high speci-
cation individual sensors are measured and trimmed to the correct tolerance.
A glass envelope is applied over the surface for protection.
3.6.2 Thermistors
These are thermally sensitive resistors. Devices rst became available in the
1940s made from combinations of metal oxides. The proportions of the
components are not a xed ratio, as is the case with compounds, and conse-
quently are referred to as non-stoichiometric substances. A negative tem-
perature coecient (NTC) is exhibited that can be extremely high
(5%K
1
) compared with metals (see gure 3.29). The value of the
coecient, like the ambient temperature resistance, depends on the
mixture of oxides used. Fabrication is by heating the compressed metal
oxide powders at a sucient temperature so that they sinter into a homo-
genous, polycrystalline solid. Oxides of manganese, nickel, copper, cobalt,
vanadium, chromium, titanium and tungsten are used as well as various
oxides of iron. Manufacturing limitations mean that the tolerances given
to thermistor specications are no better than a few percent. An alternative
type of NTC device uses a single crystal of boron-doped silicon carbide.
These have a lower sensitivity (2%K
1
) but better tolerances.
Generally, thermistors are not used for accurate thermometry because
of their poor tolerance and high nonlinearity. They are used to make
Temperature sensors 93
thermometers of limited span and, because of their high sensitivity, for
temperature control. The nonlinear relationship was rst obtained by
Steinhart and Hart in 1968 by curve-tting. A polynomial of the form
1,T A
0
A
1
ln R
T
A
n
ln R
T
n
was found to model the behaviour
of a particular thermistor, where n is an integer, A
n
are a set of constants
and T is the absolute temperature. To a good approximation higher order
terms can be ignored to give
R
T
K exp
u
T
3.64
where K and u are constants for the particular specimen. The form of
equation (3.64) can also be obtained from solid-state physics using the
FermiDirac distribution on an ideal, uniform crystal.
For convenience equation (3.64) can be re-expressed to eliminate K
using the thermistors resistance at some ambient temperature R
a
,
R
T
R
a
exp u
1
T

1
T
a
. 3.65
PTC devices were introduced more recently following the discovery that the
ceramic barium titanate when doped showed a distinct positive temperature
coecient. Other materials were developed using compounds of barium
strontiumlead titanate. Silicon PTCs are also available with somewhat
better performance. The response curves are not as well behaved as the
NTC devices so that an analytical solution can only be formulated for a
restricted range. At some temperatures PTCs may operate as NTCs.
Generally, PTCs are used for temperature and over-current protection.
Temperature protection is achieved by choosing a PTC whose R
min
value
occurs at a temperature just below the over-heat condition. It is mounted
so as to be in good thermal contact with the system being monitored. For
temperatures above T
ref
the resistance increase can be as much as
100%8C
1
. The characteristics published by manufacturers of PTCs
Figure 3.29. Thermistor characteristics.
often include the variation in resistance going through the values centred
on T
ref
.
An op-amp circuit can be constructed for buering and enhancing
the switching action. Alternatively, for low voltage applications the PTC
can be placed directly into the power line of the system it is protecting.
Consequently, heat transferred to the PTC from overheating of the system
causes a severe current restriction. After the PTC has cooled normal opera-
tion resumes.
Over-current protection is achieved by placing the device in the power
line of the system to be protected. Self-heating of the PTC occurs if the
system demands excess current. Again the resulting increase in resistance
causes the current to be restricted. Barium titanate is commonly used for
this application in the form of a ceramic disc. Combinations of PTCs with
resistors can be used to modify the characteristic.
For a more detailed description of thermistor characteristics and
applications see Hyde (1971) and manufacturers data sheets.
3.6.3 Integrated circuit temperature sensors
The electrical characteristic of the pn diode junction is temperature
sensitive. The ideal diode equation presented above in section 3.4.3 is
reproduced below.
i
d
i
s
exp
eV
d
kT
. 3.66
Re-arranging gives
V
d

kT
e
ln
i
d
i
s
1
3.67
which, in principle, indicates that if a constant current source is used to
provide i
d
then V
d
is proportional to T. However, it is found from empirical
data on the temperature sensitivity of the turn-on voltage (approximately
0.7 V at 25 8C for a silicon device) that the actual response diers from
that predicted. From equation (3.67) the temperature sensitivity should go as
dV
d
dT

k
e
ln
i
d
i
s
1
. 3.68
Taking the forward current as 10
3
larger than the reverse saturation current
gives a theoretical temperature sensitivity of about 0.6 mVK
1
. The
empirical result is a constant of 2 mVK
1
over the range 50 K to 400 K.
This anomaly is attributed to the temperature dependence of the reverse
saturation current.
The base-emitter junction in the bipolar transistor can also be used in a
similar way to the diode but tends to give better performance. So in practice
Temperature sensors 95
the transistor, with its collector and base terminals joined, provides the better
device. Generally, these pn junction devices are referred to as PTATs,
proportional to absolute temperature.
For the best PTAT performance two identical transistors are fabricated
on the same chip and made to operate at a constant ratio r of their collector
current densities. Under these conditions the dierence in their base-emitter
voltages is given by
V
BE

kT
e
ln r. 3.69
(See Sze 1994 for a typical circuit conguration). The name PTAT is
normally associated with this type of construction.
A commercially available integrated circuit is the AD590 from Analog
Devices. Low temperature-coecient, thin-lm resistors convert the PTAT
voltage to a PTAT current. A linear output current of 1 mAK
1
is output
over the range 55 8C to 150 8C. National Semiconductor also market an
integrated-circuit temperature sensor, the LM135. It is described as operat-
ing as a 2-terminal Zener with a breakdown voltage directly proportional to
absolute temperature at 10 mVK
1
.
Errors in PTAT devices can be as low as 0.25 8C at room temperature
and 0.75 8C over the full operating range. They come in transistor package
format, metal or plastic and 8-pin chip. This makes them well suited to
circuit board mounting to provide temperature compensation or alarm.
They can also be fastened to objects such as pipes and housings using ther-
mally conducting cement. They have also been used in ow measurement,
anemometry and liquid level measurement. Protection from moisture can
easily be achieved using shrink-wrap sleeving. For more hostile environments
sheathing or shrouds may be needed.
3.7 Detectors of ionizing radiation
3.7.1 Radioactivity and historical developments
Ionizing radiation occurs naturally from cosmic rays and whenever the
structure of nuclei change. Disintegration, or ssion, of a nucleus is termed
radioactive decay and generally occurs with the heavier elements. Some
unstable nuclei form part of a series. A series is a cascade of unstable,
parentdaughter isotopes. There are four such series: three occur naturally
and one is articial. Three of the series end in stable isotopes of lead and
the fourth in bismuth. During the decay process an c-particle or a u-particle
is emitted and any excess energy is released as a -ray photon. Other particles
which are much less ionizing may also be emitted, such as the neutron and the
neutrino. Neutrons can be detected by secondary events such as nuclear
scattering or transmutation. Changes to nuclei can also occur by the fusion of
two lighter nuclei to form a heavier one. This process occurs in the stars, the
hydrogen bomb and perhaps will be produced one day in a controlled way
for electric power production. Particle accelerators and x-ray machines can
also create ionizing radiation.
Ionizing radiation can be simply classied as one of two types: photons,
which are part of the electromagnetic spectrum; and particles, such as the u-
particle, a high energy electron (e
), and the c-particle, a helium nucleus

(
4
2
He
2
).
The ionizing ability, and therefore the attenuation/absorption, is
greatest with the heavy particles and least with the photons. Generally, this
means that the detection and measurement of photons is the more
dicult.
Radioactivity was discovered in Paris by Becquerel in 1896 and since
then many techniques have been developed for the measurement of ionizing
radiation. The discovery itself was made because samples of salts of uranium,
being studied for their luminescent properties, exposed photographic plates
stored in a lightproof container. Photographic lm is still in use for x-ray
imaging and dosimetry badges. The ionization chamber (1898) monitors
current ow through an air or gas cell ionized by the radiation and can be
calibrated to measure the energy of the particle or photon. The Geiger
Muller tube is a low-pressure variant of the ionization chamber. It is an
event counter and is made more sensitive because of the avalanche produced.
However, it gives no indication as to the energy of the particle or photon. The
detection eciency for photons is only 1%. The detected photons are
absorbed by the wall material, which in turn produces secondary electrons
that are detected in the usual way. In the Wilson cloud or expansion
chamber (1911) super-saturated vapour condenses on air molecules ionized
by the radiation path. The hydrogen bubble chamber (1952) contains super-
heated liquid hydrogen in a pressurized tank having a glass top. Bubbles
of hydrogen gas form around the ions produced by the radiation revealing
the path. The original scintillation counter (c. 1900) was a screen coated
with zinc sulphide (ZnS) used by Rutherford and others. The ashes of
light emitted when struck by c-particles were counted manually in a dar-
kened room over many hours. A modern, more convenient, device is the
sodium iodide (NaI) detector activated with thallium. A photomultiplier is
used to detect the light ashes produced by -ray photons. Caesiun iodide
can also be used; it is less hygroscopic so can be used to detect c-particles
by using an appropriate optically opaque cover. It has a higher absorption
coecient than NaI so thin layers can be used. It is also physically stronger,
making it useful in space vehicles. A number of materials have been devel-
oped for scintillation use, including plastic sheets. A more recent develop-
ment is the use of lithium uoride (LiF) which, when absorbing high-
energy photons, develops metastable states within the lattice. When heated
Detectors of ionizing radiation 97
later this energy is released as light (the thermoluminescent eect). An
application is in re-useable dosimetry badges.
3.7.2 Counting and spectra
Generally, the higher the density of the material the more ecient it is at
stopping the radiation. If the energy of an incident particle or photon is to
be measured, rather than simply counting the number of events per second
(the activity), then the radiation must be completely absorbed. Measuring
the number of ion pairs produced by an event can then be used to calculate
the initial energy of the particle or ray. Unless the activity is extremely
high there will be no overlap of arrivals as the time within the detector is
only 10
10
s. Such a procedure is called spectroscopy and the spectrum in
gure 3.30 is typical of the time integration of arrivals. The broadening of
the spectral lines is an artefact of the detection system.
3.7.3 Absorption processes
Charged particles
Kinetic energy is lost by collisions (which may produce ionization) or radia-
tion. Energy loss is expressed as a function of path:
Loss of kinetic energy per unit path travelled
dT,dx J m
1
3.70
dT,dx
collision
dT,dx
radiation
. 3.71
Figure 3.30. Energy spectrum of the c-particle source
241
Am.
The collision loss term is dependent on particle type, its instantaneous energy
and the medium. Radiation loss is also energy dependent. The radiation type
is usually Bremsstrahlung (breaking radiation) but for very high-energy
particles Cherenkov radiation is also produced. This occurs when the
speed of the particle is greater than the speed of light in the medium and is
caused by the medium being polarized.
Photons
Numbers decline exponentially with path length. Intensity I, goes as
I I
0
expjx 3.72
where j, the attenuation coecient, is composed of
j j
es
j
pe
j
cs
j
pp
3.73
where es is the elastic scattering out of the beam, pe is the photoelectric eect
(the photon liberates an electron), cs is Compton scattering (the photon is
scattered inelastically so that its energy decreases resulting in an increase in
wavelength) and pp is pair production (the photon is annihilated to
produce an electron and a positron).
Assuming that all the photon energy is lost within the detector, the
elastic scattering can be ignored and j can then be described as the absorp-
tion coecient. Each of the absorption processes (photoelectric eect,
Compton scattering and pair production) is energy and medium dependent.
3.7.4 Solid-state detectors
Particles, x-rays and -rays are commonly detected by PIN photodiodes. To
produce an electronhole pair in silicon requires an energy of about 3.6 eV.
This is much lower than the typical 30 eV required to ionize a gas atom.
Also, being a solid the high density means a higher probability of collision
and hence detection. The concept of quantum eciency j, as discussed in
optical detection in section 3.4, again has to be considered. The eciency
depends on both the performance of the device in stopping the particle or
photon and whether the electrons produced reach the electrodes. The fully
reverse biased semiconductor diode has an excellent ability at separating
and collecting the electronhole pairs produced. Under these conditions j
may be considered to be only dependent on the stopping ability of the
device. Consequently, the structure may be modied in some applications
to provide a larger sensing volume (gure 3.31).
High resistivity or compensated semiconductor material is essential,
otherwise the number of free charge carriers already in the material will dom-
inate those released by the radiation. Also, to ensure that all the charge
carriers produced reach the electrodes the material should be damage-free
and contain minimal trapping centres.
Particle detection
Incident high-energy electrons of hundreds of keV produce typically 85 elec-
tronhole pairs per mm in silicon. Each electronhole pair created requires
3.6 eV, resulting in an energy loss of 300 eVmm
1
. Consequently, a silicon
particle detector made 300 mm deep would produce about 4 fC of charge.
For lower energy incident electrons detection is less ecient. This is
because electrons generally suer severe deections at collisions, especially
when colliding with heavy atoms. Thus, for lower incident speeds, the prob-
ability of backscatter out of the detector rises. Conversely, heavy ions are
much more eciently stopped so standard photodiodes can be used. These
particles are stopped in the surface layers but this may mean that particles
go undetected because of absorption in the entrance window. The detection
of neutrons is not achieved directly but by secondary events following
nuclear interactions. Neutron detectors may have a surface coating to gener-
ate secondary particles that are then detected by the diode.
Photon detection
High-energy photons interact with the detector material in one of the three
ways described above in section 3.83. A typical response is shown in gure
3.32. For a more detailed description see Malm et al (1973).
For a photodiode in high reverse-bias as mentioned above all the charge
carriers produced will be collected. Consequently, the quantum eciency is
set by the absorption eciency so that equation (3.72) can be re-written using
j
I
0
I
I
0
3.74
j 1 expjx. 3.75
Figure 3.31. Photodiode structure for detecting radiation.
As an example the value for j, for a device 380 mm deep, on an n-type silicon
substrate of resistivity 4000
.
cm, irradiated by 15 keV X-rays, is 55%.
Silicon and germanium devices have been in use since the 1960s and the
technology for large, pure crystals has been well developed. Germanium has
to be operated below 150 K to eliminate noise due to the eects of thermally
produced leakage current. Research continues on material development and
processing techniques. More recently cadmium telluride (CdTe) and mercury
iodide (HgI
2
) devices have become attractive. Their higher atomic number
gives higher absorption coecients but large crystals are dicult to
produce without defects.
The energy resolution of photodiodes exceeds that of scintillation detec-
tors by up to a factor of 10 and so produces more accurate energy spectra.
However, scintillation detectors respond much faster and consequently can
give lower counting errors.
CCDs for X-ray use
These are now being developed for satellite and medical use. Their high sensi-
tivity makes them attractive, particularly in reducing patient radiation dose.
Their small size means that they are gaining in popularity with orthodontists.
They are best used for low energy photons of order 10 keV. Their
structure is similar to that described in section 3.5.4. JFETs are mounted
on the CCD substrate so as to minimize the connection capacitance. Their
Figure 3.32. Photon absorption processes.
high input impedance means there is good matching to the small output
currents from the CCD. JFETs are also relatively immune to radiation
damage.
3.8 Magnetic sensors
Magnetometers have been in use from the early days of electricity and mag-
netism. Early instruments were based on either the deection of a compass
needle, or the deection of a ballistic galvanometer from the emf generated
by a search coil, or more recently and still in common use, the ux-gate
magnetometer. More exotic and sensitive devices have become available in
recent years such as nuclear magnetic resonance magnetometers, capable
of measuring nT, and the super-conducting interference device, or SQUID,
used in biomagnetic research for measuring elds of pT.
For general laboratory use magnetometers are readily available for
measurements of elds from 0.1 mT to several teslas. Table 3.3 indicates
orders of magnitude for various sources of magnetic eld.
Of the semiconductor based devices described below those based on the
Hall eect are the most popular.
3.8.1 Hall eect
This phenomenon was discovered in 1879 by E H Hall. At the time con-
duction in metals was not understood (the electron itself was not discovered
until 1897 by J J Thomson) and this eect proved to be very instructive.
Hall noted that a current-carrying conductor perpendicular to a
magnetic eld generates an emf in the mutually perpendicular direction, as
shown in gure 3.33. The emf must be measured by a high-impedance
voltmeter.
The explanation for the eect can be derived from the Lorentz force
equation,
F qE v ^ B. 3.76
Table 3.3. Magnitudes of some common sources of magnetic
elds.
Magnetic eld source Order of magnitude
Geomagnetism 10 mT
Data storage 1 mT
Current carrying conductor 1 mT
Small permanent magnet 10 mT
Superconducting magnet 10 T
In the derivation it is assumed that the length of the plate is much bigger than
d. The magnetic eld causes the moving charge carriers to move to the top of
the plate of material. If the charges are not allowed to leave through the sides
of the plate an equilibrium is reached where F is zero, giving
E
H
v ^ B. 3.77
A dynamic equilibrium is reached where
electrostatic repulsion electromagnetic force
E
H
being the Hall electric eld across the sides of the plate. The Hall voltage
V
H
can be obtained from
E
H

V
H
d
. 3.78
Combining equations (3.77) and (3.78) gives
V
H
vBd. 3.79
Now, the current owing in a material can be written in terms of the number
density n of the charge carriers, the charge of a carrier q, the cross-sectional
area A and the drift velocity v to give
i nAqv. 3.80
Substituting for v in equation (3.79) and replacing A as wd gives
V
H
B
i
nAq
d 3.81
V
H
B
i
ndwq
d 3.82
V
H
B
i
nwq
3.83
V
H
R
H
B
i
w
3.84
where R
H
is called the Hall coecient. In practice correction factors have
to be added to equation (3.84) due to scattering of the charge carriers and
Figure 3.33. The Hall phenomenon.
Magnetic sensors 103
geometrical eects. See Baltes and Castagnetti in Sze (1994) for a more
detailed theoretical analysis.
At the time of the discovery it was found that most metals gave the
expected result but some gave the opposite sign of V
H
. Following the
development of quantum mechanics and the band theory of conduction it
was realized that these anomalous results were due to the charge carriers
being holes.
In metals V
H
is tiny and consequently has little practical use other than
for materials research. However, the eect in semiconductors is orders of
magnitude higher because of the lower value of n. Sensitivities can be
between 5 and 500 V/AT, InSb being more sensitive than Si. Following the
development of semiconductor fabrication techniques, Hall eect semi-
conductor probes began to be marketed as magnetometers.
There are a number of useful applications of the Hall eect:
.
Semiconductor research and development. The Hall eect can be used to
measure n and give the sign of the charge carriers to denote whether
these are holes or electrons. Also, since the conductivity o is determined
by nqv, by measuring the conductivity v can be determined.
.
Magnetometer. By making the connections as shown in gure 3.33 and
driving a constant current through the plate the Hall voltage gives a
direct reading of an unknown magnetic eld. The side contact points
need to be small and well aligned since a potential gradient exists along
the plate, causing the current ow. Misalignment would generate a poten-
tial dierence without a magnetic eld being present. The magnetometer is
small, compact, sensitive and can operate from d.c. up to 100 kHz. Cur-
rently, commercial products can be obtained with high sensitivity that
measure down to 10 nT or, for high eld strengths, alternative instruments
can measure as high as 15 T. Various geometries are available. GaAs
devices can operate at higher temperatures because intrinsic conduction
only begins at about 400 8C in this IIIV compound.
.
Proximity switch. This is simply triggered by the presence of a magnetic
eld of sucient strength. The plate of material has with it some TTL
circuitry within the package. It is readily available from component
suppliers. It is used for detecting metallic objects, for burglar alarm
switches, and in some paddle wheel-type owmeters.
3.8.2 Magnetoresistance
First discovered by W Thomson (later to become Lord Kelvin) in 1856, this
phenomenon is also a result of the Lorentz force but it is the deection of the
current path by the magnetic eld that is directly utilized here. In this device
the length of the material is much shorter than its width so that the Hall eld
can be assumed to be negligible.
The magnetic eld causes deection of the current lines as shown in
gure 3.34. This results in a current increase i
L
through one side of the
plate and a corresponding decrease in the other. The change in resistance
of a specimen goes as
R R
0
1 tan
2
0
L
3.85
where 0
L
is the Lorentz angle denoting the deection of the current lines. It
can be shown that tan 0
L
jB where j is the eective carrier mobility; see
Baltes and Castagnetti in Sze (1994) for a fuller description.
A number of manufacturers produce magnetoresistive sensors. The
usual fabrication is as a thin lm of Permalloy (20% iron, 80% nickel)
strips on a silicon substrate, which are connected so as to form a bridge
network. This facilitates the production of bespoke congurations and
integrated circuitry. Sensor range is typically 5 mT with a resolution of
10 nT and developing a resistance change of about 3% at maximum eld.
Frequency response is similar to Hall eect devices at tens of kHz.
Materials exhibiting the so-called giant magnetoresistive or GMR eect
discovered by Baibich et al (1988) are currently attracting a lot of attention.
This eect is based on spin-dependent scattering of electrons within very thin
sandwiched layers of materials. Alternate layers of FeCr or CuCo are
fabricated having thicknesses of about ten atoms. Resistance drops by up
to 15% in elds of 30 mT and bandwidth is high going from d.c. to
MHz. Sensors based on GMR became available in the mid 1990s. Many
magnetic storage media now employ GMR materials.
3.8.3 Magnetodiode
The magnetodiode structure is that of a long, thin PIN diode (gure 3.35).
Double injection occurs from the n
and p
regions to give a high injection

current into the intrinsic region. The carriers drift under the inuence of the
Figure 3.34. The magnetoresistive eect.
Magnetic sensors 105
electric eld E along the long axis. The presence of the magnetic eld B,
orthogonal to E, causes bunching of the charge carriers at the bottom or
top faces of the plate depending on the direction of B. This is the so-called
magnetoconcentration eect discovered by Suhl in 1949, which causes a
gradient in the carrier concentration across the plate. The recombination
rates at the two sides of the plate are arranged to be dierent. This has
been achieved by polishing one side and rough grinding the other. The
overall eect is to have the diode currentvoltage characteristic modied
by the magnetic eld. Unfortunately this is a nonlinear relationship that is
sensitive to temperature drift. Sensitivities of 5 V/mAT have been achieved
at bias currents of around 5 mA and later versions, using CMOS technology,
have produced 25 V/mAT. Frequency response is less than 10 kHz.
In production the quality control of the nishes to the plate surfaces is
poor, meaning that carrier recombination rates vary so that reproducibility is
poor. This together with the nonlinearity and temperature sensitivity has
meant that there has been little commercial interest in magnetodiodes.
3.8.4 Magnetotransistor
These are transistors whose physical structure has been designed so that the
collector current is optimally sensitivity to an applied magnetic eld. A
number of distinct phenomena are said to contribute to magnetotransistor
operation. These are
.
carrier deection caused by the Lorentz force,
.
magnetoconcentration,
.
modulation of carrier injection.
The relative signicance of these depends on the particular design.
An original device was the drift-aided, lateral, double-collector device
developed by Davies and Wells (1970). The structure illustrated in gure
3.36 is of a bipolar pnp transistor.
The dominant process occurring here is carrier deection. The base has
current owing through between the two end connections reminiscent of a
Hall plate. The generated Hall eld deects the current carriers so that the
two collectors receive diering amounts of current. The dierence in the
Figure 3.35. The magnetodiode PIN structure.
collector currents i
c
is given by
i
c
kj
n
j
p
Bi
e
3.86
where k is a geometry factor, j
n
and j
p
are the electron and hole mobilities
and i
e
is the emitter current.
There have been many designs of magnetotransistors reported.
Roumenin (1990) has given a classication and review of the various
kinds. Some devices have the current owing vertical to the surface unlike
the lateral design shown above. These measure magnetic eld in the plane
of the structure.
Using equation (3.86) to dene the sensitivity as i
c
,Bi
e
, a value of
0.25 T
1
has been reported recently for a bipolar device employing both
carrier deection and injection modulation.
In addition to the standard bipolar construction technology, CMOS
fabrication can be used. A typical sensitivity for a lateral CMOS device is
1.5 T
1
.
The MOSFET structure can be adapted so that the channel acts like a
Hall plate or employs carrier deection. These devices are termed
MAGFETs. A recent report gave a sensitivity of 640 V/AT for a NMOS,
Hall MAGFET operating at 5 V with a drain current of 0.5 mA.
Despite the advantage of higher sensitivity, magnetotransistors have not
provided commercial competition to the inexpensive and simple Hall
integrated circuit structure. For the various magnetotransistor designs this
can be due to higher levels of noise, nonlinearity, temperature dependence
and large o-set values that occur due to small manufacturing misalignment.
Bibliography
In this chapter various designs of novel semiconductor sensor have been
described. However, there is little information on fabrication techniques.
For those readers who would like to know more of this highly specialized
foundry technology the following texts might be helpful.
Figure 3.36. The drift-aided, lateral, double-collector, magnetotransistor.
107
Gardener J W, Varadan V K, Awadelkarim O O 2001 Microsensors, MEMS and Smart
Devices (Wiley)
Kovacs G T A 1998 Micromachined Transducers Sourcebook (McGraw-Hill)
Maluf N1999 An Introduction to Microelectromechanical Systems Engineering (Artech House)
Sze S M 1994 Semiconductor Sensors (Wiley)
References
Baibich R et al 1988 Phys. Rev. Lett. 61(21) 24722475
Bernstein J et al 1993 A micromachined comb-drive tuning fork rate gyroscope Proc.
IEEE Micro Electromechanical Systems Workshop, Fort Lauderdale, FL, February
1993, pp 143148
Bouwstra S, Legtenberg R, Tilmans H A C and Elwenspeek M 1990 Resonating micro-
bridge mass ow sensor Sensors and Actuators A21A23 332335
Cho S T and Wise K T 1993 A high-performance microowmeter with built-in self-test
Sensors and Actuators A 36 4756
Dalley J W and Riley W F 1978 Experimental Stress Analysis (New York: McGraw-Hill)
Davies L W and Wells M S 1970 Magnetotransistor incorporated in a bipolar IC Proc.
ICMCST, Sydney, Australia, pp 3443
Elbel T, Muller J E and Volklein F 1985 Miniaturisierte thermische Strahlungsensoren:
die neue Thermosaule TS-50.1 Feingeratetechnik (Berlin) 34 113115
Fairchild Weston CCD Sensors 1989 Systems and Development Technology, Databook
Gardner J W, Varadan V K and Awadelkarim O O 2001 Microsensors, MEMS and Smart
Devices (Wiley)
Gardner J W 1994 Microsensors: Principles and Applications (Wiley)
Greenwood J C 1988 Silicon in mechanical sensors J. Phys. E: Sci. Instrum. 21 11141128
Hanna S M 1987 Magnetic eld sensors based on SAW propagation in magnetic thin
lms IEEE Trans. UFFC 34 191194
Johnson J D, Zarabadi S R and Sparks D R 1995 Surface micromachined angular rate
sensor Sensors and Actuators SP-1066 7783
Johnson R G and Higashi R E 1987 A highly sensitive chip microtransducer for air ow
and dierential pressure sensing applications Sensors and Actuators 11 6372
Kovacs G T A 1998 Micromachined Transducers Sourcebook (McGraw-Hill)
Kress H-J, Marek J, Mast M, Schatz O and Muchow J 1995 Integrated silicon pressure
sensor for automotive application with electronic trimming Sensors and Actuators
pp 3540
Malm H L, Randorf T W, Martini M and Zanio K R 1973 Gamma ray eciency compar-
isons for Si(Li), Ge, CdTe and HgI
2
detectors IEEE Trans. Nucl. Sci. 20 500509
Maluf N 1999 An Introduction to Microelectromechanical Systems Engineering (Boston:
Artech House)
Neubert H K P 1975 Instrumentation Transducers: An Introduction to their Performance
and Design 2nd edition (London: Oxford University Press) pp 237238
Neumeister J, Thum R and Lu der E 1990 A SAW delay-line oscillator as a high-resolution
temperature sensor Sensors and Actuators A 2123 670672
Patel H 1992 Silicon that moves IEE Review July/August pp 268, 269
Reidemeister E P and Johnson L K 1995 Capacitive acceleration sensor for vehicle
applications Sensors and Actuators SP-1066 pp 2934
Roumenin C S 1990 Bipolar magnetotransistor sensors. An invited review Sensors and
Actuators A24 83
Spark D R, Zarabadi S R, Johnson J D, Jiang Q, Chia M, Larsen Q, Higdon W and
Castillo-Borelley P 1997 A CMOS integrated surface micromachined angular rate
sensor: its automotive applications Proc. Transducers 97, Int. Conf. on Solid State
Sensors and Actuators, Chicago, IL, 1619 June, pp 851854
Sze S M 1981 Physics of Semiconductor Devices 2nd edition (Wiley)
Sze S M 1994 Semiconductor Sensors (Wiley)
Wilson J and Hawkes J F B 1983 Optoelectrics: An Introduction (Prentice-Hall)
Xensor Integration, Delft, the Netherlands, Data sheet XI-IR3774 Infrared sensor,
91.083774
Exercises
1. Explain why semiconductor-based sensors are at the centre of many
research and development programmes. Describe some of the diculties
in applying such semiconductor devices to the various measurement
elds.
2. Discuss the advantages and disadvantages to using silicon as a con-
struction material for micro sensors.
3. A miniature gas ow sensor is constructed using a silicon nitride bridge-
structure operating as a resonant system.
(i) Describe the operating principle.
(ii) A piezoresistive strip is attached to the structure to monitor the
oscillations. Draw a functional block diagram to illustrate how
resonance might be maintained.
4. Pressure sensors are now available manufactured from a silicon disc.
The disc acts as a diaphragm on which piezoresistive tracks have been
formed.
(i) State two restrictions that limit the applications of such devices.
(ii) The radial and circular stresses occurring in the disc are given by
radial stress
r

3P1 v
2
8Et
2
R
2
0
3r
2
tangential stress
t

3P1 v
2
8Et
2
R
2
0
r
2
.
For a xed pressure, sketch the forms of these stress relationships as
a function of the radius. Where on the diaphragm is the point of
zero stress?
(iii) Draw an appropriate signal conditioning circuit and explain how
the piezoresistive elements need to be connected to provide tem-
perature compensation and give maximum sensitivity.
Exercises 109
5. (a) Explain the phenomenon of the surface acoustic wave giving typical
numerical values of the associated parameters. Describe how a
surface acoustic wave sensor might be constructed to measure
(i) pressure
(ii) magnetic eld.
(b) Acertain surface acoustic wave mass sensor has a sensitivity given by
df
dj
k
1
k
2
f
2
0
Hz kg
1
m
2
Given that the resonant frequency is 10

9
Hz and that
k
1
9.8 10
8
and k
2
4.18 10
8
m
2
s kg
1
determine the resolution of the mass measurement if the signal proces-
sing system can measure to a resolution of 5 in 10
8
.
6. Explain the terms in the modied Nernst equation
E
Theevenin
E
0
RT
zF
lna
x
ka
y
.
Given that
pH log
10
a
H
combine these two relationships to show that
E
Theevenin
E
0
0.0592 pH volts
7. Describe, with the aid of a diagram, the operating principles of a FET
adapted for use as a biosensor. Give some advantages and disadvantages
of the invasive use of these sensors with patients.
8. (a) Draw a circuit diagram for a photodiode and its signal conditioning
so that it functions in the photovoltaic mode. Show that in this
mode
V
d

kT
e
ln
ejPA`
i
s
hc
.
What are the advantages and disadvantages to this mode of
operation?
(b) (i) Photodiodes can also be used to detect x-rays. The absorption
of the rays can be described by
I I
0
expjx.
Show that
j 1 expjx.
(ii) A certain photodiode used to detect x-rays has a sensor depth of
300 mm and the value of j is 2 10
3
m
1
. Calculate the photo-
current given that 85 electronhole pairs are produced per mm
and that the incident photon ux is 10
5
s
1
.
9. (a) Describe the fabrication and performance of the NTC and PTC
thermistors. Include sketch graphs and applications.
(b) Using the ideal diode equation show that the forward bias voltage
when a diode is conducting responds PTAT.
(c) The pn junction in a PTAT device is usually a transistor base-
emitter junction rather than a diode. Why is this?
(d) One design of a PTAT device uses two identical transistors mounted
on the same substrate. Their collector currents i
c1
and i
c2
are
arranged to be in a xed ratio . Using your expression from (b)
above show that the dierence in their base currents V
be
goes as
V
be

kT
e
ln .
10. A certain magnetoresistor is commercially available having a resistance
R
0
of 500.0 and a sensitivity of 2.5 kT
1
.
(i) Draw the physical layout of a bridge circuit similar to that of gure
3.4 such that four of these magnetoresistors might be used to
measure a magnetic eld.
(ii) Show that for small changes in resistance cR and a supply voltage
V
s
, the optimum The venin output potential dierence is given by
E
Theevenin

cR
R
0
V
s
.
(iii) Assuming that noise levels in the circuitry might be in the region of
50 mV estimate the minimum magnetic eld that could be measured.
Exercises 111
Chapter 4
Optical bre sensors
4.1 Introduction
The rst patents for optical bre sensors appeared in the mid-1960s. Since
then signicant development has taken place from these rst simple, light-
guide devices to the present intricate and sophisticated systems. The
advance has been facilitated by the rapid progress made in the use of
optical bres by the telecommunications industry. This has resulted in the
availability of high-quality, cheap bres. In addition, the growth in the use
of compact disc players, laser printers and infrared remote control units
has made optical components readily available at low cost.
The establishment of optical bres for telecommunication links has
encouraged their use for telemetry in some industrial elds. The chemical,
nuclear, petrochemical and other process industries have begun to use
them for supervisory control and data acquisition (SCADA) principally
because of their
.
lower installation costs,
.
intrinsic safety due to the electrical isolation and low-power light levels,
.
robustness and reliability, leading to lower maintenance costs,
.
immunity to electromagnetic interference,
.
chemical inertness,
.
absence of electromagnetic transmission,
.
high temperature resilience,
.
high bandwidth.
Industries such as electricity, automobile and avionics have also shown
interest, as have robotic and medical engineers.
The use of optical bres in instrumentation systems can be grouped into
three categories,
.
light guides for illumination in remote monitoring and in obscure locations,
.
imaging systems, such as endoscopes, face-plates and magniers,
.
sensors that use the bre itself or some small artefact coupled to the bre.
It is this last category that is attracting most eort currently and is the theme
of this chapter. The concept of combining an optical bre sensor and the
data link as a complete system is very attractive, especially to industrial
instrumentation engineers. The advantages listed above for the optical
bre data link should then become true for the whole instrumentation
system. In addition, the low weight and the small size of optical bre
sensors are attractive features in many measurement elds.
The common parameters to be measured in the process industries are
temperature, ow, pressure, displacement/position and levels in tanks and
hoppers. It is unlikely that sophisticated and expensive devices will be
taken up initially. The industry is traditionally very conservative and needs
much convincing. The harsh environment of process plant will make delicate
devices impractical. It is most likely that the technology will gain acceptance
through the introduction of simple sensor systems that are relatively cheap
and reliable or highly specied and unique. For multi-sensing situations a
simple radial network of links to the control room will be used, although a
data highway ring is envisaged for future, more extensive systems.
4.2 The optical waveguide
Telemetry based on metallic wire pairs, although by far the most common
technique currently, is facing a growing challenge from optical bres in all
applications. Light is, without doubt, one of the most important physical
quantities in the universe. The Biblical creation story records the creation
of light as one of the earliest acts of God. There is also little doubt that
light was the original historic method for transmitting information over a
distance. This would have been achieved by reection of sunlight, lamps or
re beacons. At the end of the 6th century BC Aeschylus recorded that the
news of the downfall of Troy was relayed from Asia Minor to Argos using
a chain of re signals. In the 2nd century BC Polybius described a method
by which the Greek alphabet could be transmitted with re signals using a
two-digit, ve-level code. In 1870 John Tyndall demonstrated that light
could follow a curved jet of water and later J L Baird led patents for the
transmission of light in glass rods. In the 1950s bre optic devices were
developed for inspection instruments. However, it was in 1966 that Charles
Kao and George Hockham at Standard Telecommunication Laboratories,
Harlow, proposed the principle of information transmission via a trans-
parent dielectric. It was the subsequent development of high quality glass
bres that led to their current widespread use in telecommunications.
Recent developments in data compression techniques have made
possible the transmission of real-time, digital, video images in serial
The optical waveguide 113
format. This advance, with the intrinsic benets of optical bres, has accel-
erated the phenomenal growth of an international, information transmission
industry that struggles to keep up with the increasing demand for capacity. A
recent development, as a means of increasing capacity without installing
extra lines, is the use of light of dierent wavelengths down a bre to
provide extra channels, in eect frequency domain multiplexing. It is
usually referred to as dense wavelength division multiplexing or DWDM.
The current maximum in commercial use is 100 dierent wavelengths per
bre. With a transmission rate of 10 gigabits s
1
per channel this yields a
total single-bre rate of 1 terabit (10
12
) s
1
. Gupta et al (2002) have reported
a laboratory rate of 3.2 terabits s
1
using 40 lasers in the range 1531
1563 nm. The channel bit-rate was 80 gigabits s
1
over 100 km.
The cell network used by cell phones employs optical bres for interlink-
ing. Mobile cell phones themselves are only short range and register with a
GHz antenna system in the local cell. The use of mobile phones for games
and video imaging is further fuelling demand for capacity. Clearly, the
level of utilization of optical bre telecommunications is set to increase
well into the decades ahead.
4.2.1 Principles of propagation
An optical bre is a waveguide usually of circular cross-section. The bre is
made from some dielectric material such as a polymer or a glass. To under-
stand the propagation process requires an appreciation of classical electro-
magnetic wave theory. It was James Clerk Maxwell in 1864 who showed
conclusively that light was an electromagnetic phenomenon. This was an
idea originally proposed by Michael Faraday. Maxwell was able to show
that by combining the known magnetic and electrostatic eld equations,
with an important modication, a pair of dierential equations could be
produced whose solutions were transverse waves. Moreover, the velocity of
these waves was the same as the free-space speed of light. This conrmed
that what is now called the electromagnetic spectrum is indeed composed
of electric and magnetic elds. This derivation of Maxwells is found in
many texts and is illustrated below.
Gausss theorem
In dierential equation format, Gausss electric eld equation for free space
can be expressed in cartesian coordinates as
0E
x
0x

0E
y
0y

0E
z
0z

,
0
4.1
where , is the free charge density. For convenience this is often expressed
more generally in terms of a vector dierential operator r, pronounced
114 Optical bre sensors
del. Del can be interpreted in any three-dimensional co-ordinate system. In
the cartesian system it is expressed as
r
_
i
0
0x
j
0
0y
k
0
0z
_
4.2
where i, j and k are unit vectors in the x, y and z directions respectively. Del
can operate on a scalar or a vector function. The non-vector operation on the
scalar c is
rc
_
i
0
0x
j
0
0y
k
0
0z
_
c. 4.3
This scalar operation is the gradient of c and is sometimes written grad c.
When operating on a vector function the multiplication can be one of two
forms. These are the dot product , or cross product . The dot product
on some vector Y in cartesian coordinates is written
r Y
_
i
0
0x
j
0
0y
k
0
0z
_
Y. 4.4
This is sometimes referred to as the divergence of Y and written div Y. Dot
product multiplication produces a scalar. The alternative cross product
vector multiplication is
rY
_
i
0
0x
j
0
0y
k
0
0z
_
Y. 4.5
rY is also-called curl Y. The resulting quantity is a vector normal to the
plane formed by the original vectors.
Following this notation equation (4.1) can be written
div E
,
0
. 4.6
If a dielectric medium is present then the presence of induced charge has to be
included in the calculation of E. The introduction of the vector D, called the
electric displacement, overcomes this problem. D is only dependent on the
free charges present and is dened as
D
r
0
E or E. 4.7
Thus equation (4.6) becomes
div D ,. 4.8
If there are no free charges present, as is the case for an optical bre, then
div D 0. 4.9
A similar expression can be written for the magnetic eld B as
div B 0. 4.10
Div B is always zero because there are no free magnetic poles. The magnetic
eld B is related to the magnetization force H by
B j
0
j
r
H or jH. 4.11
H unlike B is not altered by the presence of some material. This is similar to
the vector D, which is only dependent on the presence of real charges and not
on any induced charge in the medium.
Faradays law
Faradays law of electromagnetic induction can be expressed in this vector
calculus form as
rE
0B
0t
4.12
the negative sign coming from Lentzs law. Using the description of curl
given after equation (4.5), equation (4.12) can be written
curl E
0B
0t
. 4.13
Ampe`res theorem
Ampe` res theorem relates the magnetization force H to the actual current
owing I, taken around some closed path s,
_
H ds I. 4.14
This can be expressed in dierential equation form in terms of the free
current density J as
curl H J. 4.15
However, as mentioned above, Maxwell introduced an extra term to incor-
porate any current due to the electric displacement D. Equation (4.15)
becomes
curl H
0D
0t
J. 4.16
Because there are no free charges in an optical bre equation (4.16) reduces to
curl H
0D
0t
. 4.17
Maxwells equations
The four equations that have become know as Maxwells equations are
obtained using equations (4.7) and (4.11) to substitute for D and B so that
equation (4.9) becomes
div E 0 4.18
div H 0 4.19
curl E j
0H
0t
4.20
and equation (4.17) becomes
curl H
0E
0t
. 4.21
The nal stage in the derivation involves combining these four equations to
solve for E and H. To do this the following mathematical identity is required:
rrY rr Y r
2
Y. 4.22
r
2
is the second derivative of rand is sometimes referred to as the Laplacian
operator. Thus, taking the curl of curl E gives
curl curl E grad div E r
2
E 4.23
curl
_
j
0H
0t
_
grad 0 r
2
E 4.24
j
0curl H
0t
r
2
E 4.25
j
00E,0t
0t
r
2
E 4.26
r
2
E j
0
2
E
0t
2
. 4.27
Now the general expression for a wave of amplitude Y in dierential
equation format is
r
2
Y
1
v
2
p
0
2
Y
0t
2
4.28
where v
p
is the phase velocity of the wave, the velocity at which the wave
crests move forward. By comparing terms it can be seen that equation
(4.27) is a wave and that
v
p

1
j
p . 4.29
For free space equation (4.29) becomes
c
1
j
0
0
p . 4.30
Inserting the values of j
0
and
0
in equation (4.30) gives the free-space speed
of light, which added conrmation to Maxwells theory.
If the derivation is carried out for H a similar expression to equation
(4.27) is obtained,
r
2
H j
0
2
H
0t
2
. 4.31
The general solution to equation (4.28) is the sinusoidal wave equation
Y Y
0
exp j.t k s 4.32
where k is the propagation vector and s is the displacement. The magnitude
of k is given by
k j j k
2
`
0
. 4.33
k is also-called the free-space wave number. From equation (4.32) the sinu-
soidal wave equations for E and H are
E E
0
exp j.t k s 4.34
H H
0
exp j.t k s. 4.35
Orientation of the elds in an electromagnetic wave
Inserting equations (4.34) and (4.35) into the four Maxwell equations yields
k E 0 4.36
k H 0 4.37
k E j.H 4.38
k H .E. 4.39
Equations (4.36) and (4.37) reveal that the electric and magnetic elds are
orthogonal to the direction of travel. Equation (4.38) shows that H is
normal to the plane of k and E and equation (4.39) that E is normal to the
plane of k and H. For these conditions all to be true k, H and E must be
mutually orthogonal as shown in gure 4.1.
The laws of reection and refraction
The equations derived above can be used to analyse the phenomena that occur
as a wave crosses the boundary between two dierent dielectric materials. This
is a signicant issue with optical bres since the bre must have a dielectric
outer cladding to support and protect it. Figure 4.2 illustrates a wave travelling
with velocity v
1
in the rst medium and v
2
in the second.
The three waves in the two media have the propagation vectors k
i
for the
incident wave, k
re
for the reected wave and k
ra
for the refracted wave. The
wave equations are
Y Y
0
exp j.t k
i
s 4.40
Y Y
0
exp j.t k
re
s 4.41
Y Y
0
expj.t k
ra
s 4.42
where Y represents E or H.
In the plane of the boundary all three components of the waves must be
in phase otherwise there would be a discontinuity. This is true for any wave
phenomenon. Thus,
k
i
s k
re
s k
ra
s 4.43
which becomes
k
i
sin c k
re
sin u k
ra
sin 4.44
The magnitudes of k
i
and k
re
are the same since they are in the same material
and so will have the same wavelength `. Consequently,
\c \u 4.45
one of the laws of reection. Also, writing equation (4.33) as
k j j k
2
`
0
2f
c
4.46
Figure 4.1. The mutual orthogonality of the electric and magnetic elds and the direction
of propagation.
Figure 4.2. Transmission of a wave across a dielectric boundary.
equation (4.44) can be re-written as
2f
v
1
sin c
2f
v
2
sin . 4.47
The frequency does not alter across the boundary otherwise a discontinuity
would occur as stated above. Consequently, equation (4.47) becomes
sin c
sin

v
1
v
2
1
n
2
4.48
which is Snells law,
1
n
2
being the refractive index going from medium 1 to
medium 2. If medium 1 is free space then equation (4.48) becomes
sin c
sin

c
v
2
n
2
4.49
where n
2
is the absolute refractive index of medium 2. It is often more
convenient to deal with absolute refractive indices. Equation (4.48) can be
re-written
sin c
sin

c,v
2
c,v
1
n
2
n
1
. 4.50
If the wave is to travel down the optical bre without loss then the angle of
incidence at the walls c must be large enough for total internal reection to
occur. The least angle of incidence for this to occur is when
sin 1 4.51

2
. 4.52
From equation (4.50), it can be seen that this particular value of c, called the
critical angle, is sin
1
n
2
,n
1
.
The requirement to have waves strike the inside of the core bre at
greater than the critical angle has implications on the entrance and exit of
light at the bre ends. This is indicated in gure 4.3.
Figure 4.3. The numerical aperture.
At the entrance to the bre, Snells law, equation (4.48), becomes
sin 0
sin
n
1
n
3
4.53
where n
3
is the refractive index in the medium adjacent to the bre end.
Re-arranging equation (4.53) gives
sin 0
n
1
n
3
sin . 4.54
But from gure 4.3 it can be seen that sin cos c, so equation (4.54)
becomes
sin 0
n
1
n
3
cos c. 4.55
If c is the critical angle then sin c n
2
,n
1
. Using the trigonometrical
relationship cos
2
c 1 sin
2
c equation (4.55) becomes
sin 0
n
1
n
3
1 sin
2
c
_
4.56
sin 0
n
1
n
3
1
_
n
2
n
1
_
2
4.57
sin 0
1
n
3
n
2
1
n
2
2
_
4.58
n
3
sin 0
n
2
1
n
2
2
_
. 4.59
The expression n
3
sin 0 is called the numerical aperture, NA.
NA
n
2
1
n
2
2
_
. 4.60
NA becomes smaller as the value of n
2
approaches that of n
1
. In commercial
bres the dierence c between n
1
and n
2
is only about 1%, making NA about
0.2. Increasing c has the disadvantage of increasing inter-modal dispersion, a
phenomenon described later.
For light to propagate along the bre it must enter within a cone of half-
angle 0 as shown in gure 4.3. This is sometimes referred to as the cone of
acceptance. Also, because in geometrical ray diagrams rays are reversible,
an identical cone must exist at the emerging end of the bre within which
the light is conned. These two cones indicate the optimum position relative
to the ends of the bre for the light source and detector (gure 4.4). In
addition, the emitting area of the light source should be less than the bre
cross-sectional area and be close to the bre end. Also, the emission half-
angle of the source should be less than that of the bre. The sensitive
surface of the detector should be large enough and placed such that all of
the emergent light is captured.
The concepts described above of reection, refraction, critical angle and
cones of acceptance and emergence, apply to bres of core diameter 10mm.
The transmission paths through such bres are numerous, each suering a
dierent number of reections. Such bres are classed as multimode. For
smaller bres, whose diameters are approaching the wavelength of the light
being transmitted, interference eects begin to occur. This has the result of
reducing the number of modes of transmission. In the smallest of bres used
only one mode of transmission is possible. The wave travels along the
central axis of the bre and is called the meridional ray. These monomode
bres are preferred in long distance telemetry because a light pulse will be
elongated in time if the light travels along multiple paths of various lengths.
Pulse broadening, or dispersion, sets an upper limit on the bit-rate that can
be serially transmitted. There are other causes of dispersion so this form is
called intermodal or mode dispersion. Graded refractive index optical bre is
an alternative to monomode. It is of a similar diameter to multimode but
light travelling near the edges of the rod travels faster than the axial rays and
so compensation is achieved. It is the less commonly used type.
Graded refractive index (GRINrod) lenses have been developed for
optical bre systems and are often used in sensor research. These are rod-
shaped, similar to the bres, but are usually between 1 and 5 mm in diameter
whereas multimode and graded index bres can be up to 100 mm in diameter
although 50 mm is the more typical.
The evanescent wave
Although there is no refracted wave once the critical angle has been exceeded
this does not mean that there is no disturbance in the cladding medium.
Considering again the refracted wave of gure 4.2 given by equation (4.42)
and assuming it to be plane polarized in the x direction, resolving into the
x and z directions gives
Y Y
0
exp j.t k
ra
x cos k
ra
z sin . 4.61
Now, re-arranging equation (4.50) gives an expression for sin ,
sin
n
1
n
2
sin c. 4.62
Figure 4.4. Placement of light source and detector.
But equation (4.62) implies that, for angles of c greater than the critical
angle, sin 1, which is not possible. To continue this analysis it is therefore
expedient to replace both sin and cos in equation (4.61):
Y Y
0
exp j .t k
ra
x
1 sin
2
_
k
ra
z
n
1
n
2
sin c
_ _ _ _
. 4.63
But since sin 1, the root here is imaginary so the complex form is
substituted:
1 sin
2
_
j
sin
2
1
_
. 4.64
Equation (4.61) becomes
Y Y
0
exp j
_
.t
_
jk
ra
xsin
2
1
1,2
k
ra
z
n
1
n
2
sin c
__
. 4.65
The physically more meaningful solution is obtained by taking the negative
root to give an exponential decline with x,
Y Y
0
expk
ra
xsin
2
1
1,2
exp j
_
.t k
ra
z
n
1
n
2
sin c
_
. 4.66
This can be expressed as
Y Y
0
expx, exp j
_
.t k
ra
z
n
1
n
2
sin c
_
4.67
where

1
k
ra
sin
2
1
1,2

1
k
ra
n
2
1
,n
2
2
sin
2
c 1
1,2
4.68
or

`
2
n
2
2n
2
1
sin
2
c n
2
2
1,2
. 4.69
The term `
2
is the wavelength of the wave in medium 2. This is better
expressed in terms of the free space wavelength `
0
using
f
c
`
0
. 4.70
Similarly for medium 2,
f
v
2
`
2
c,n
2
`
2
. 4.71
Combining equations (4.70) and (4.71) gives
c
`
0
c,n
2
`
2
4.72
so
`
2
n
2
`
0
. 4.73
Equation (4.69) can then be expressed as

`
0
2n
2
1
sin
2
c n
2
2
1,2
. 4.74
The value of is a measure of the depth of penetration of the electromagnetic
elds into the forbidden region.
If this derivation is carried out separately for E and H, including the
boundary conditions for the normal components to the boundary, a phase
shift of ,2 appears between the two waves. Now the Poynting vector N
for energy ow is given by N E H. Consequently because of the phase
shift there is no net ow of energy into medium 2. This protruding wave is
termed evanescent.
4.2.2 Degradation of transmission in optical bres
Currently, the transmission limit for an optical bre telecommunications link
between repeater stations is in excess of 100 km. This achievement has been
made possible by the development of
.
high quality manufacturing techniques and processes, and
.
the materials from which the core and cladding are manufactured.
Transmission performance is degraded by attenuation and dispersion.
Attenuation
There are three mechanisms by which the intensity of the propagated light
can be reduced.
Absorption
The bre material can absorb photons. The electrons in the molecules from
which the bre is made have discrete energy levels. An electron can move to a
higher energy state by absorbing a quantum of energy. Thus, if a photon
having the same energy as that needed for the transition passes through
the region of the molecule then it may be absorbed. This energy puts the
molecule into an excited state. The energy is released some short time later
as the molecule reverts to its ground state. The overall eect is to increase
the thermal energy of the material. These electron energy levels in silicon-
based bres correspond to photons in the ultraviolet region.
In addition, the molecules of the bre material may vibrate about their
lattice points. The kinetic energies of these vibrations also have discrete
values. Again, if a photon is in the vicinity, having the same quantum of
energy as needed for an increase in the kinetic energy, then it may be
absorbed. In this process the energies correspond to photons in the infrared
region of 812 mm.
The absorption spectra in gure 4.5 illustrate these two processes. A
minimum occurs at around 1.550 mm providing a low absorption window.
The theoretically estimated attenuation level in silicon-based bres has
been virtually reached at 0.2 dBkm
1
(Ohashi et al 1990).
The biggest absorption losses can come from minute traces of metals
such as iron and copper in the glass. These may be present in the raw
materials or even be introduced during the manufacturing process.
Modern processing techniques have virtually eliminated this problem.
However, it has proved very dicult to overcome the eects of the presence
of the radical OH. Also, if the bre is exposed to hydrogen gas during its
working life the number of OH radicals can increase, causing an increase
in attenuation. Attenuation peaks occur at wavelengths of 0.725, 0.825,
0.875, 0.950, 1.380 and 2.73 mm. See Keck et al (1973).
Exposure of the bre to ionizing radiation or ultraviolet light is known to
cause an increase in absorption. Ions produced in the glass act as absorption
centres. It may be possible to employ these phenomena to create sensors.
Mode leakage
Attenuation may occur due to light leaking from the core into the cladding.
As described in a previous section, light entering within the cone of
Figure 4.5. Representation of the absorption spectra for a silicon-based optical bre.
acceptance and central to the axis will totally internally reect. However,
non-axial rays called skew rays may enter and propagate by reections at
just below the critical angle. They may travel hundreds of metres in a
helical manner but are ultimately lost. When developing optical bre
sensors this mode of transmission may need to be accounted for.
If the bre is bent, rays will strike the inner wall at less than the critical
angle and so leak away. Manufacturers generally specify a minimum bend
radius for their bres that must be observed during installation to minimize
losses. Microbends can occur during manufacture due to mechanical
irregularities in the core-cladding interface.
Scattering
During the manufacturing process, as the bre solidies, minute local
irregularities in refractive index occur. These are due to uctuations in
density and the concentrations of any dopants present. These irregularities
are of the order of a tenth of the wavelength of the light being transmitted.
The resulting eect is to cause light to be scattered out of the bre. According
to Rayleighs law, scattering is proportional to `
4
. So short wavelengths
suer signicantly more scattering. For the currently available high-
performance bres this is the most signicant loss mechanism. For silicon-
based bres it causes an attenuation of 1.7 dB per km at 820 nm.
In summary, the above attenuation mechanisms can be described using
I I
0
expcx 4.75
where c a
atomic
a
molecular
a
bending
a
scatter
.
Optical bre joints and terminations also cause scattering and absorp-
tion. When a wave passes through a refractive index boundary some reec-
tion occurs. For normal incidence this Fresnel reection causes a 4% loss
in transmission. This amounts to 8% or 0.35 dB for a joint. Optical coupling
liquid or cement can be used for index matching. However, to achieve the
lowest losses, fusing the components together whilst clamped in a special
alignment jig is the standard approach. Losses are reduced to between 0.1
and 0.3 dB. Demountable connectors are available. These are particularly
useful for research but require a high degree of mechanical precision for
best results. Lenses can also be useful for coupling since they reduce the
eects of axial misalignment and dust. Their disadvantage is that angular
misalignment becomes more signicant. Losses of around 1 dB per connector
are typical.
Dispersion
Sir Isaac Newton is reputed to have discovered that a prism could disperse
white light into its constituent colours and a further prism recombine the
colours to restore the white light. This illustrates that refractive index is
wavelength dependent. It follows that this is because the speed of an electro-
magnetic wave within a medium is wavelength dependent.
Consequently, because no practical light source can be of a single wave-
length, if a pulse of light is injected into an optical bre the pulse length will
increase during transmission. This is referred to as chromatic (colour) disper-
sion. It occurs within any mode so is classed as an intramodal dispersion. The
group velocity of the pulse is dependent on the second derivative of the
refractive index d
2
n,d`
2
. Optical bre engineers sometimes refer to materials
in which this term is non-zero as showing material dispersion; see Senior
(1985) page 81. A further, less signicant, form of dispersion can occur in
optical bres due to the wavelength dependence of the angle a ray in a
given mode makes with the bre axis. This is sometimes referred to as wave-
guide dispersion. Again pulse broadening occurs but in multimode bres the
eect is negligible compared with material dispersion.
Pulse broadening during the propagation down a long bre has similar
consequences to intermodal dispersion described above. To keep this
problem to a minimum, for long telecommunication links quasi-mono-
chromatic light sources (laser diodes) are used.
Yamada et al (1980) found that the combined eect of material and
waveguide dispersions is a minimum at a wavelength of 1.32 mm in a single
mode bre.
4.2.3 Manufacturing techniques
Early techniques used high quality glasses of the type used in lens manufac-
ture. For a number of years the rod and tube method was used. In this process
the core material, in the form of a rod, is inserted into a tube made of the
cladding material. The two are heated to form a rod called the preform,
which is then drawn to form the thin bre. This technique is still employed
for producing specialist bres having a refractive index prole that is not
symmetrical. Multiple rods of diering diameter and refractive index are
placed within the tube of cladding material before being heated to produce
the preform. Monomode bres made in this way can preserve the orientation
of polarization during light propagation. Another early development was the
double crucible technique. Here two heated, concentric, platinum crucibles
containing the molten core and cladding glasses share a common drain
from which the preform is delivered. The advantage of this process is that
it is continuous if the glasses are regularly replenished.
The bre drawing system is common to all of the manufacturing techni-
ques that have been developed. It is a high precision process. The preform is
heated in the extrusion oven and the diameter of the drawn bre output is
monitored. A negative feedback system controls the draw speed. The bre
is also coated with its outer jacket before being wound on to the take-up
reel. Proof testing of the nished product may also be carried out by applying
a known tension via the take-up reel mechanism.
More recent bre manufacturing procedures involve various forms of
chemical vapour deposition techniques derived from the semiconductor
industry. In the modied chemical vapour deposition process (MCVD),
gas-phase reagents such as germania-tetrachloride and oxygen are burned
together inside an externally heated silica glass tube. The resulting powder,
under the inuence of the heating, diuses into the glass inner wall. The
tube is later collapsed to produce the preform. By varying the types and
concentrations of the reagents both core and cladding regions can be
formed as well as graded refractive index proles. A similar technique is
the outside chemical vapour deposition process (OCVD). Here a rotating
substrate or bait rod is used on which to deposit the materials produced by
the MCVD method. The core materials are deposited rst, followed by the
cladding. The bait rod is removed and the glass layers collapsed to provide
the solid preform. A renement of this technique enables continuous
rather than batch manufacture. Here the rotating bait rod is drawn
through the burner region. The burner is mounted around the rod so as to
produce the coating powders axially.
Table 4.1. Comparison of telemetry lines.
Characteristic Twisted
metallic pair
Coaxial cable Optical bre
Attenuation (dB/km) 3 2 1 (glass)
150 (polymer)
Distance-bandwidth
product (MHz-km)
1 10
2
10
5
Repeater spacing (km) 1 5 100
System cost Low Medium Medium
Lifetime (years) Similar life spans of 2540
Crosstalk High Low None
Noise immunity Low Medium High
Electrical isolation None None Total
Weight and size High High Low
Termination
technique
Solder Solder with
standard
connections
Splicing by fusion
of well-aligned
ends
Fabrication control Loose Moderate Precise
4.2.4 Comparison of optical bres with electrical lines
Table 4.1 compares various transmission media characteristics for the
common telemetry lines.
4.3 Extrinsic optical bre sensors
Extrinsic optical bre sensors have some external artefact connected to the
bre that is modied by the measurand. The bre takes no part in the
sensing. The optical bre merely acts as the telemetry link and, on some
occasions, as the power line. The majority of the optical bre sensors
developed are of this type. Udd (1991) gives a detailed catalogue of all of
the types of optical bre sensors. Some of the reported sensing processes
used in extrinsic systems are described below.
4.3.1 Displacement sensors
A simple optical switch can be made using a shutter to interrupt the beam
(gure 4.6) (see Jones 1985). S and D are the light source and detector
bres. This has been developed further to provide a continuous positional
sensor using a Moire -type fringe grating. The Moire grating is moved
across the gap between the faces of the transmit and receive bres and the
fringes generated are subsequently counted.
A spherical lens has been used to move the light beam across the front of
the bre (see Dotsch et al 1983). The received intensity varies with lens
displacement (gure 4.7). Two receiving bres mounted side by side have
been used to provide a dierential change in light signals as the lens (or
source bre) moves across. This overcomes the problem of variation in
source brightness.
In section 4.2.1 above it was shown that the light entering at the front
face of an optical bre was only able to remain within the bre if the angle
the ray made to the bre axis was within the cone of acceptance. As described
previously, this also means that a transmitting bre emits light within the
Figure 4.6. Binary optical switch.
Extrinsic optical bre sensors 129
cone. Changing the separation of the bres results in a change in the intensity
of the received light (gure 4.8).
A strain gauge system and a hydrophone have been developed using this
conguration. Kissinger (1967) led a patent for a displacement sensor using
this technique with a bundle of bres in reection mode. Frank (1966) led a
patent for a pressure sensor using just two bres. The general arrangement
for the transmit and receive beams for a two-bre system is shown in
gure 4.9.
With the reector set at a distance of less than x
min
from the front faces
of the bres the collecting bre receives no light. This condition can be
related to the radii r
S
, r
D
and the bre separation y,
y r
S
r
D
2x
min
tan 0
A
. 4.76
If the reector is moved so that the separation increases beyond x
min
the
intensity of the light received increases in a near linear manner and
becomes a maximum when
y r
S
r
D
2x
max
tan 0
A
. 4.77
Increasing the separation further would cause a decrease in the received
intensity tending towards an inverse square law relationship. The range
x
min
to x
max
is set by the choice of r
S
, r
D
and their separation y. The shape
of the transfer function was developed by Jones and Philp (1983). Figure
4.10 shows the received power for two dierent receiving bres set at dierent
distances from the reector.
Jones and Philp (1983) have developed this concept further using three
bres. By the judicious choice of radii and separation one receiver bre can
Figure 4.7. Spherical lens displacement sensor.
Figure 4.8. Fibre displacement sensor.
function as a reference, with the other providing the data (gure 4.11). The
signal bre gives a near linear response over the working range whereas
the reference bre is of almost constant output as indicated in gure 4.10.
In this way any variation in light source performance can be accounted for.
(a)
(b)
Figure 4.9. The reectance mode displacement sensor: (a) minimum physical displace-
ment. (b) Maximum physical displacement.
Figure 4.10. Received power as a function of displacement for two dierent geometries.
This sensor was designed to measure pressure. Generally devices that
measure small displacement can be adapted for the measurement of pressure,
temperature, strain, vibration and acceleration.
4.3.2 Positional encoders
Precision translational or rotational position sensors can be made using a
transparent encoded plate or disc. The Gray code described in section
5.3.1 and table 5.1 is used to minimize reading errors. The binary pattern
is usually made photographically. For an 8-bit system, giving a resolution
of 1 part in 256, eight optical bres are required. Figure 4.12 shows a transla-
tional encoder plate covering the range 015 for an 8-bit system.
Multiplexing systems have been developed that only have a single
transmit and receive bre pair. This can be done by using a white light
source separated into colours by using either lters, a diraction grating or
a prism. This is frequency-domain multiplexing although in bre optics it
is usually referred to as wavelength-division multiplexing. Figure 4.13(a)
illustrates a prism conguration.
Alternatively, time-domain multiplexing can be used. Here short pulses
of monochromatic light are transmitted into the system. Each receive bre is
made to be of dierent length so that a time delay is created between received
data pulses. Figure 4.13(b) gives a schematic diagram. A length of bre 0.2 m
long corresponds to a delay of about 1 ns. Each bre would need to be at least
this much longer than its neighbour to ensure separation of pulses in the
telemetry link.
The disadvantages are that the wavelength multiplexing technique
requires precision optical engineering whereas time-domain multiplexing
requires sophisticated, high-speed, electronic circuitry.
Figure 4.11. Stepped, three-bre conguration.
Figure 4.12. Gray code positional encoder.
Diraction gratings have also been successfully used to monitor
displacement using either intensity variation or colour change. Prototypes
have been developed for measuring displacement and pressure and for
hydrophonessee chapter 7 of Udd (1991).
Figure 4.13. (a) Multiplexing by wavelength. (b) Time division multiplexing.
4.3.3 Absorption sensors
A number of dierent sensor types can be constructed based on light
absorption. An input and an output bre are required. Single-ended cell
structures are the preferred physical design as this produces a more
compact unit suitable for use as a probe.
The opacity of some liquid crystals is temperature dependent. Using an
arrangement such as shown in gure 4.14 provides a simple temperature
sensor. Unfortunately liquid crystals are only stable at moderate tem-
peratures so the useful range is limited. This arrangement is best used as a
temperature switch since the opacitytemperature relationship is nonlinear
and of short span. A sandwich of crystals, each switching at a dierent
temperature, would provide a thermometer, albeit with a stepped response.
The absorption of light passing through a gas or a solution can be used
as a means of measuring its concentration. The absorption is described by the
LambertBeer law,
absorptivity log
10
_
I
0
I
_
cx 4.78
where I
0
is the intensity of light received with no solute in solution, I is the
intensity of light received with solute present, is the absorption coecient
of the species expressed in square metres, x is the distance between the
transmit and receive bres path in metres, and c is the concentration of the
solution in moles per cubic metre.
Re-arranging equation (4.78) gives
I I
0
10
cx
4.79
and assumes parallel light passing through the test cell (gure 4.15). The
dimensions of the test cell and light source intensity need to be chosen for
a particular application. GRINrod lenses could be used to provide a
compact unit.
The dynamic range of the received light should be about two orders of
magnitude. At the low end of the range the output signal needs to be greater
than the noise level. It is common in this type of sensor to use two dierent
wavelengths, one sensitive to the species to be measured and one not. In this
way variation in the performance of the light source, or the presence of a
species other than the one under test, can be accounted for.
Figure 4.14. Single-ended absorption sensor using a mirror.
If the chemical species itself is not a good absorber it is sometimes
possible to use a reagent whose absorptivity changes with the concentration
of the test species. An example of this is the use of a pH sensitive indicator
such as phenol red for measuring acidity levels; see Dakin and Culshaw
(1988). Reagent colour change can also be used although then the optical
system becomes more involved, requiring spectral analysis.
In applications where a reagent is required it may be preferable to have
the reagent absorbed in a membrane that is permeable to the species under
test. One approach is to have an optically white (meaning 100% reectivity
over all relevant wavelengths) diuse reector (gure 4.16) (see Grattan and
Meggitt 1995). The membrane is, in-eect, a coating on the white surface.
The KubelkaMunk relationship, equation (4.80), is used to relate
the reectance with the concentration of the test species and assumes a
semi-innite scattering medium,
FR
1 R
2
2R

c
S
4.80
where R is the reectance, S is a scattering coecient and FR is the
KubelkaMunk or remission function. This nonlinear relationship between
Figure 4.15. A test cell interrogated with parallel light.
Figure 4.16. Absorption sensor using diuse reectance.
c and R means extensive characterization procedures are required otherwise
interpolation between calibration points would yield high errors.
Finally, a common approach in absorption sensors is to use light
scattering within the test cell rather than some form of reector to increase
performance (gure 4.17). Although backscattering is not as ecient as a
reector at returning light to the detector bre, the level of source light
can easily be increased to compensate. Scattering does ensure thorough
irradiation of the test cell. Particles of about 1 mm are used. A commonly
used source of such particles is polystyrene latex. It is also possible to
mount the reagent on scattering particles if it is not practical for it to be in
solution. Polyacrylamide microspheres have been used to support phenol
red to produce a pH probe for physiological applications; see Kirkbright
et al (1984).
Each of the test cells in gures 4.144.17 could be constructed with a
selective membrane that permits the entry of the test species but retains the
reagents within the cell. Cellulose dialysis tubing is used for aqueous
solutions whereas silicone rubber is more appropriate for gas systems.
Multiple reagent probes have been produced that detect specic species
through two or more chemical reactions. The only limit on the development
of these probes is in obtaining appropriate reagents that, ideally, are re-
usable. The name optrode is becoming popular for optical sensor probes.
4.3.4 Luminescence sensors
Luminescence is the emission of light from a previously excited molecule as it
reverts to a lower energy state. The intensity of the emitted light declines
exponentially with time. It can be shown that, for a large sample, the
decay time constant t is equal to the average time a molecule exists in its
excited state. This is usually referred to as the lifetime. There are two
classes of luminescence, uorescence and phosphorescence. Fluorescence
refers to emission where lifetimes are short, of the order of 10
9
s, and
occurs in various solids, liquids and gases. Phosphorescence only occurs in
solids and, because the transition to the lower energy state is classed as
forbidden, the light emission may be over many minutes. Emission intensities
are much lower with phosphorescence and generally require more sensitive
detectors or time integration of the signal or both. For a text on uorescence
see Lakowicz (1983).
Figure 4.17. Absorption sensor using back-scatter with inert particles.
Fluorescence sensors
Light is admitted from the transmitting bre and is absorbed by the medium
according to the absorption relationship given above in equation (4.79). The
response of the uorescent material is monitored. There are several mechan-
isms by which uorescence may occur; see Scheggi and Baldini (1993).
1. The species to be measured is itself a uorescent material.
2. Molecules of the test species are made to attach to molecules of a
uorescent reagent called a uorophore making them uorescent. This
type of procedure is sometimes referred to as labelling.
3. The species reacts with a reagent to form a new complex or compound
that is uorescent.
4. The species interacts with a uorophore so that the uorescence is
modied in intensity or lifetime.
One form of this latter mechanism is the phenomenon of quenching. The test
species has the eect of either preventing the uorophore from absorbing the
light transmitted to it (static quenching) or altering the emission process of
the excited uorophore (dynamic quenching). With either phenomenon the
uorescent intensity is reduced. For both types of quenching the relationship
between the intensity of the uorescence I and the quencher Q present in
concentration [Q] is given by
I
I
0
1
1 KQ
4.81
where I
0
is the uorescent intensity in the absence of the quencher. For static
quenching K is the dissociation constant of the reaction between the uoro-
phore and the test species. For dynamic quenching it is the SternVolmer
constant which involves the lifetime of the uncombined uorophore and
the bimolecular quenching constant.
For dynamically quenched uorophores measurement of the lifetime
may be an alternative to the measurement of intensity. The lifetime of the
uorophore without the presence of the test species is a reference value, t
0
.
The lifetime is reduced to t because the excited state of the uorophore is
depopulating non-radiatively in addition to the normal, light-emitting
process. The lifetimes are related by a similar expression to equation (4.81),
t
0
t
1 K
SV
Q 4.82
where K
SV
is the SternVolmer constant; see Dakin and Culshaw (1997).
An advantage of this technique is that the intensity of the light input I
0
is not part of the determination. This makes the system immune to variations
in the performance of the light source or to tarnishing of the optical
components.
With luminescence techniques the emitted light is of a dierent wave-
length from that of the exciter. This can give up to 10 greater sensitivity
compared with the straightforward absorption technique. However, absorp-
tion techniques can often be more immune to the presence of other species.
Commercial devices have become available. An interesting example is in
blood analysis where partial pressures of O
2
and CO
2
can be monitored as
well as pH.
Phosphorescence sensors
Certain rare earth compounds phosphoresce with an exponential decline that is
related to their temperature. This phenomenon has been used toproduce a tem-
perature sensor. The structure of the device is simply a bi-directional bre
having the active material coated over the front end. This end is in thermody-
namic equilibrium with the environment to be measured. These materials func-
tion in the range 200 to 400 8C. The light emission L can be expressed as
L L
0
exp
_
t
t
_
4.83
where L
0
is the initial phosphorescent intensity as the energizing light pulse is
switched o. The value of the time constant t is measured and from calibra-
tion tables the temperature is determined. The value of t is reduced as the
temperature is increased, typically ranging from 5 s at 200 8C, to about
1 s at 400 8C. As with the uorescence case, because the time constant is
being measured to determine the value of the measurand, variation in
source intensity or attenuation in the system does not aect calibration. A
commercial device is available from the Luxtron company.
4.3.5 Total internal reection-based sensors
A liquid level sensor has been constructed using 908 prisms (gure 4.18). The
angle of incidence is 458, which is larger than the critical angle for a glass to
Figure 4.18. Liquid level sensor.
air interface. However, if the liquid is allowed to surround the prism light will
not totally reect and the detector will have a much-reduced signal. An array
of such sensors down the side of the tank provides discrete level measurement
(Jones 1981).
A more sophisticated application of the total internal reection is in the
use of the evanescent eld. This was described toward the end of section
4.2.1. Following equations (4.67) and (4.74), the amplitude of the electric
eld component can be written as
E E
0
expx, 4.84
where

`
0
2n
2
1
sin
2
c n
2
2
1,2
4.74
and c is the angle of incidence, which must be greater than the critical angle,
and `
0
is the free space wavelength.
Equation (4.74), using typical values of n
1
, n
2
and 0, gives as about one
wavelength. There is no energy transfer as the H and E vectors are out of
phase. The result is that energy ows in during one half-cycle but is returned
in the next.
To make use of this phenomenon in sensing it is necessary for power to
be coupled out of the rst medium via the evanescent wave. Thus the
coupling component has to come within about a wavelength of the rst
medium. When this occurs the totally internally reected wave is reduced
in power. There are two modes of construction for evanescent wave sensors.
Front end sensor
Here the evanescent wave penetrates into the material placed within a wave-
length of the obliquely cut front end of the bre. Quantitative chemical
analysis has been achieved by having the test solution in contact with the
end of the bre. Changes in the concentration of the solution alter the
value of its refractive index which aects .
The measurement of temperature and strain has been reported using two
bres. One bre is xed and the other moves in response to the measurand so
that the coupling of the light into the second bre is aected (gure 4.19).
This eect is sometimes referred to as frustrated total internal reection.
Spillman and McMahon (1980) have used this approach to produce a
hydrophone.
Figure 4.19. Frustrated total internal reection.
Side-coupling sensor
In this mode of operation the cladding is removed from a section of the bre
allowing the evanescent wave to form around the sides. A similar bre is
mounted alongside in parallel, about a wavelength away, in order to
receive the wave. By allowing the measurand to alter the separation of the
bres a temperature or strain sensor can be produced similar to the front-
end system above.
A more sophisticated alternative is to coat the exposed bre with some
chemical or biological reagent whose refractive index will vary with the
concentration of the species to be measured. A number of patents have
been established for biological sensors using this technique. The attractive
features are as follows.
.
The evanescent wave is so localized that only the behaviour of the bonded
reagent aects the sensor and not the general solution, thus providing a
high selectivity.
.
Using the disposable capillary structure shown in gure 4.20 means that a
xed volume of the analyte (test solution) is drawn inside. The sensitized
Figure 4.20. A disposable, capillary-ll, evanescent wave sensor.
region contains an appropriate reagent dye for monitoring the level of
some chemical species or an antibody to monitor the level of the
corresponding antigen. In either case the evanescent wave is variably
absorbed by the response of the reagent and absorption spectroscopy
can be performed. Alternatively, the reagent may be a uorescent material,
enabling uorescent spectroscopic analysis of the reected signal.
A number of evanescent wave sensors have been built in hybrid, miniature
form on a substrate. A typical structure involves a MachZehnder
interferometer described in section 4.4.1. An example is the hydrogen
sensor reported by Nishizawa et al 1986. These and similar structures are
sometimes referred to as integrated optic devices or optical integrated circuits
(Nishihara 1989).
4.3.6 Pyrometers
Pyrometry is the measurement of the temperatures of incandescent objects
from their colour. The emission spectrum of a perfect black-body is given
in gure 3.13. The traditional pyrometer is a remote, hand-held device
that produces an image of the hot object on a translucent screen. This is
viewed through the eyepiece alongside a set of coloured templates that are
calibrated for temperature. The user obtains the best match of colour
and then reads the temperature from the template. Because of the variation
of emissivity in objects which are not perfect black-bodies the templates
are calibrated for a particular application, such as molten iron in a
blast-furnace.
To produce an optical bre pyrometer a material capable of withstand-
ing furnace temperatures is needed for the probe section. A single crystal of
sapphire has proved suitable. This is coated with a high melting-point metal
so as to provide a shield to any visible light and also act as the black-body
cavity walls (gure 4.21). Iridium is a good choice, despite being hard to
work, as it is the most resistant material to chemical attack. The sapphire
is optically coupled to a silica bre outside the hot environment.
Figure 4.21. An optical bre pyrometer.
Since the black-body cavity is integral to the instrument a correction
for the non-perfect emissivity can be incorporated. Also, coupling and
absorption losses in the system can be accounted for by using the relative
intensities of two dierent parts of the spectrum using narrowband lters.
Resolutions of 0.5 8 C are possible with a tolerance of 0.0025% in the
range 5002000 8 C. A commercially available optical bre pyrometer has
been produced by Luxtron.
4.4 Intrinsic optical bre sensors
A strict denition of intrinsic sensing is that the measurand modies some
parameter or property of the optical bre. Measurement occurs within the
bre. The benets of intrinsic optical bres in addition to those of the
extrinsic type are
.
greater geometrical versatility,
.
wider dynamic range,
.
much higher sensitivity,
.
high frequency response.
The disadvantages of intrinsic sensors compared to the extrinsic types are
generally
.
greater complexity,
.
more sophisticated signal recovery and processing is required,
.
greater precision and stability of optical components is required,
.
a more highly specied and stable light source is needed,
.
greater sensitivity to ambient parameters reduces selectivity and causes
drift, making the measurement of steady or slowly varying measurands
dicult.
The optical phenomena used for measurement are
.
attenuation or phase shift due to physical distortion of the bre,
.
rotation of polarization,
.
scattering.
4.4.1 Physical distortion
Microbending
In section 4.2.1 the propagation of light down an optical bre was described
and the need for total internal reection at the bre walls. The calculations
assumed an axially straight bre. With microbend sensors the curvature of
the bre axis is altered by the measurand so that losses in transmission do
occur. The arrangement is shown in gure 4.22.
The microbend technique has been used to produce position, force,
pressure, strain, acceleration and vibration sensors. Nonlinearities of less
than 1% have been reported.
4.4.2 Change in optical-path length
In section 4.2.1 the wave equation was written as
Y Y
0
exp j.t k s. 4.32
The term .t k s completely describes the phase of the wave at any
position and instant in time. For simplicity, assuming that the wave travels
axially in the z direction, the phase term can be re-written as .t kz,
where k is the free-space wave number 2,`
0
. During transmission
through some material this free-space wavelength is shortened since v
1
< c.
c f `
0
4.85
v
1
f `
1
4.86
dividing
c
v
1
n
1

`
0
`
1
. 4.87
Consequently, the phase term in the medium is expressed as .t n
1
kz
where n
1
is the absolute refractive index. The term n
1
z, the physical path
multiplied by the refractive index, is known as the optical path length.
If the wave were to divide, be sent along two dierent paths, and then
recombine there would be some phase dierence c where
c n
1
kz
1
n
2
kz
2
. 4.88
If the two paths are within the same material,
c nkz
1
z
2
nkcz. 4.89
Since these phase shifts are fractions of the wavelength of light the potential
exists for measuring tiny displacements or changes in refractive index. The
construction of the sensor must be such that the measurand generates some
change in path dierence. These phase shifts are measured by interference
Figure 4.22. A microbend sensor.
Intrinsic optical bre sensors 143
eects using some form of interferometer. Unfortunately, the light intensity
variation with path dierence is not a linear function.
There are three historic designs of interferometer that have been adapted
for optical bre sensing. These are the MachZehnder (gure 4.23), the
FabryPe rot and the Michelson types and are illustrated below in original
and optical bre forms. The sensor element can be constructed on a substrate
as an optical waveguide. This substrate-based technology is often referred to
as integrated optics; see Go pel et al (1992). The most commonly used for
optical bre sensing is the MachZehnder interferometer.
The FabryPe rot interferometer (gure 4.24) has the advantage that the
light passes many times through the test region which is advantageous in
some applications.
Figure 4.23. MachZehnder interferometer.
Figure 4.24. FabryPe rot interferometer.
The Michelson interferometer (gure 4.25) has some similarities to the
MachZehnder design. For a comparative analysis of the interferometer
types see Udd (1991).
A signicant problem faced by designers of interferometer-based
sensors is that steady state or slowly varying measurands are dicult to
measure accurately due to thermal drift causing changes in the optical
path dierence. Such drifting has a time scale of 0.1 s. Consequently,
interferometric sensors are best suited to measurands varying at 10 Hz.
A further diculty lies in the measurement of lengths greater than one
wavelength since identical interference intensities are reproduced through-
out every 2 phase-shift. The zero setting also becomes ambiguous. Thus
small, relative displacements can only be measured unless a more
sophisticated auto-zeroing construction and additional signal analysis is
employed.
Optical bre interferometers have been developed to measure refractive
index, acceleration, electric and acoustic elds, electric current, temperature,
strain and pressure. Refractive index can be used for chemical analysis
measured either at the end of the bre in a Michelson or FabryPe rot
integrated optic structure, or through the side of the bre via the
evanescent-wave using a MachZehnder integrated optic interferometer.
Described below are three devices that have achieved some commercial
success.
Figure 4.25. The Michelson interferometer.
Magnetostriction magnetometer
Ferro-magnetic materials exhibit magnetostriction and below saturation the
longitudinal strain produced e, which is cz,z, generally goes as
e CH
2
4.90
where C is a material constant; see Livingston (1982).
A magnetometer can be constructed by directly bonding a material
having magnetostrictive properties to the surface of an optical bre. In the
presence of a magnetic eld the bre will either be stretched or compressed
axially, thus causing a change in the phase of the light transmitted. From
equation (4.89),
c nkjzCH
2
4.91
where j is the strain transfer eciency and is a correction factor for the
change occurring in k due to axial strain in the bre core. Driving the
system at mechanical resonance improves sensitivity. An appropriate
interferometer structure is the MachZehnder type. Dagenais et al (1988)
reported a resolution of 20 pT
Hz
p
at 1 Hz operating at a mechanical
resonance of 26.5 kHz.
Hydrophone
Rapid developments in hydrophones have taken place since the 1970s driven
particularly by the military services. Optical bre based hydrophones are
attractive due to their structural versatility and relatively low cost. Noise
levels of at least 10 dB below ambient underwater noise have been achieved
with sensitivities of 300 dBmPa
1
.
There are two approaches to construction. The bre is either simply
given a exible coating and mounted perpendicular to the vibration axis as
in gure 4.26(a) or wound around a compliant mandrel to increase sensitivity
(gure 4.25(b)).
Sagnac eect
A great deal of eort has been expended over the past thirty years on the
development of an optical bre gyroscope based on the Sagnac eect
reported in 1913. See Post (1967) for an account of the Sagnac eect.
Figure 4.26. Optical bre hydrophones.
An optical arrangement for observation of the Sagnac eect is illustrated
in gure 4.27. The optical medium is a ring of radius R rotating at some
angular velocity together with the detector. The time of travel from A to
B is longer than that from B to A due to the rotation. Consequently, if a
beam of light is split into two and one is made to propagate in the direction
of rotation while the other is in the counter direction, a phase dierence will
occur. This phase dierence is eectively due to the dierence in path
length. A simple classical approach assumes two counter-rotating light paths.
The time t taken for the light to travel one circuit around the ring,
when the ring is stationary, is
t
2R
v
4.92
where v is the speed of light in the glass.
If the ring rotates then during t it will have turned through t
radians. Consequently, the path AB will have increased by Rt and for
BA reduced by Rt. Therefore, the path dierence is
z 2Rt 2R
2R
v
4.93
z 4A
v
4.94
where A is the area of the ring. The phase dierence c, from equation (4.89),
is nk z,
c nk4A
v
4.95
which would suggest the result
c n
2
8A

c`
0
. 4.96
Figure 4.27. The Sagnac eect.
However, Post (1967) reported that experimental observation by various
workers yielded the result that, for the case where the propagation medium
is co-moving with the interferometer, the presence of the medium has no
eect on the interference fringes observed. So in equation (4.96) the refractive
index is eectively unity giving
c 8A

c`
0
. 4.97
Post also reports that the Sagnac eect does not depend on the shape of
the surface area and does not depend on the location of the centre of
rotation.
From equation (4.97) the phase shift is directly proportional to and
hence the usefulness of the technique for inertial guidance systems. The
best devices currently available can detect rotations as low as 0.00001
degrees per hour.
4.4.3 Rotation of polarization
It is useful rst to review the phenomena of polarization. It was shown in
section 4.2 that the electric and magnetic elds in an electromagnetic
wave are orthogonal to each other and the direction of propagation
(gure 4.1). As a convention the electric eld direction is used to describe
the orientation of polarization. A wave propagating with some xed orien-
tation of E is said to be plane polarized, the electric eld lying in a plane
through the axis of propagation. Photons are emitted from an ordinary
light source having all possible orientations of E around the propagation
axis. This light is described as unpolarized. The superposition of beams
of polarized light was investigated by Fresnel and Arago (see Longhurst
1970). The three laws they discovered are summarized here to help appreci-
ate the nature of polarized light.
1. Orthogonally, plane-polarized beams of light cannot produce interference
eects.
2. Two plane-polarized beams of light can produce interference eects pro-
vided that the orientation of their planes of polarization is the same and
that the beams were derived from the same plane-polarized source or
the same plane of polarization of an unpolarized source.
3. Two plane-polarized beams of light derived from perpendicularly polar-
ized components from unpolarized light never produce interference
eects even when their planes of polarization are rotated so as to have
the same orientation.
A number of techniques are available that can be used to separate or lter
light waves according to their polarization.
Polarization by glass plate
Polarization can be achieved using an optical boundary. When an unpolar-
ized beam crosses an optical boundary at an oblique angle partial polariza-
tion of the reected and refracted waves takes place. The waves have
polarizations normal to the plane of incidence and in the plane of incidence
respectively. The amount of polarization that occurs is a function of the angle
of incidence. The reected wave is completely polarized when the angle of
incidence is Brewsters angle. Brewsters angle is tan
1
n
refraction
,n
incidence
;
see Bleaney and Bleaney (1965). This is sometimes made use of in lasers.
Laser light is polarized but unless there is some mechanism in the lasing
cavity, such as a glass plate inclined at Brewsters angle, the orientation of
polarization randomly uctuates with a timescale of the order of 1 ms.
Polarization by birefringence
One of the phenomena that can occur in optically active materials is that the
speed of propagation of the wave is dependent on its polarization. This
eectively means that the material has diering refractive indices for dierent
orientations of polarization. This phenomenon is termed birefringence. Natu-
rally occurring examples of birefringent materials are crystals of quartz and
calcite (Iceland spar). If an unpolarized beam of light falls on a single crystal
at the correct orientation two waves emerge, each plane polarized and whose
planes of polarization are orthogonal. (It is easy to see why the original name
for this phenomenon was double refraction.) To understand this eect it
should be remembered that, as stated above, an unpolarized beam contains
all possible orientations of polarization in equal measure. It is therefore pos-
sible to resolve the electromagnetic elds of a beam into two orthogonal axes
in a plane perpendicular to the direction of propagation. If these axes are
appropriately inclined to the optically asymmetric axes of the crystal then
the two components of the beam will refract dierently and so separate.
Polarization by selective absorption
Polarization can also be achieved using dichroic materials. Dichroism is the
selective absorption of one of the orthogonal polarizations within an
unpolarized light wave. The phenomenon, discovered by Biot in 1815,
occurs in some minerals, the most well-known being tourmaline, and in
some organic compounds. The development of polarizing lm was pioneered
by Land and rst patented in 1932; see Land (1951) for a review of early
developments. The Polaroid Corporation was formed by Land and Wheel-
wright in 1933. The ratio of the polarized components after passing
through a polarizing lm can be as high as 10
5
.
Light propagating through optical bre can suer change of polariza-
tion due to various external inuences. Such phenomena can be used to
provide a sensing mechanism.
It is helpful to consider the various states of polarization that the output
wave might develop due to birefringence. The presence of birefringence may
be intrinsic to the glass or induced by the measurand. A plane-polarized wave
input to the bre can be resolved into two identical orthogonal components
having planes of polarization, say E
x
and E
y
, travelling along a birefringent
bre in the z direction. Assuming that there is no attenuation, when the
waves emerge into the detector system the electric eld, formed from the
vector addition of E
x
and E
y
, will have three possible forms (gure 4.28).
If the retardation of one wave relative to the other is zero or a multiple of
, the resultant electric eld vector will be of xed orientation in time. This
state is termed linear polarization. (Linear polarization and plane polariza-
tion are in eect the same state but the use of the word linear groups the
phenomenon with the other phenomena described next.) If the retardation
of one wave relative to the other is an odd multiple of ,2 then the tip of
the resultant electric eld vector will trace a circle. This is termed circular polar-
ization. All other retardations, and therefore the most common, cause the tip
of the electric eld vector to trace an ellipse hence the term elliptical polariza-
tion. Both of these vector rotations occur at the optical frequency and will not
be seen visually. Also, if as stated above there is no attenuation the linearly
polarized beam will emerge with the same angle of polarization as the plane
wave input. If attenuation does occur, as well as aecting this angle, circular
polarization will not be possible. Instead an ellipse will be generated having
major and minor axes aligned to the x and y axes.
Clearly, for sensor use the intrinsic state of polarization within the bre
needs to be stable so that change caused by the measurand can be recovered.
To achieve this the bre must be monomode and have one of the following
criteria.
Figure 4.28. States of polarization (E
x
leads E
y
).
1. Have no birefringence. This is dicult to achieve, requiring special
materials and high precision manufacturing techniques.
2. Have high birefringence along the axis by deliberately creating stress
during the manufacturing process. The diering refractive indices for
the two orientations means that it is possible to arrange for one mode
to be lost in the cladding so only one plane of polarization propagates.
The cladding mode is stripped away at the receiver.
3. Have a tiny core of elliptical cross-section. Polarised light injected having
its polarization aligned with the major axis of the ellipse will propagate
without change of orientation.
Some of the polarizing phenomena that have been used in sensor design are
described below.
The Faraday magneto-optic eect
In 1845 Michael Faraday discovered that, when plane-polarized light passes
through glass subjected to a strong magnetic eld aligned with the direction
of propagation, the plane of polarization is tilted. This eect is as a result of
the interaction of the magnetic elds of the light and electron orbitals. The
angle of tilt 0 is given by
0 V
_
H dl 4.97
where dl is the closed loop path increment along the eld line H and V is the
Verdet constant for the level of the eect in a given material. V is wavelength
dependent but temperature has little eect on the process. Its value is
typically 5 10
6
rad A
1
for near infrared light in silica. Response times
are much faster than in sensors involving magnetostriction. Bandwidths of
1 GHz have been achieved.
A most useful application of this technique has been the development of
an optical bre current sensor (gure 4.29). Pioneering work was carried out
for the CEGB research centre by Rogers (1979) at Fawley power station in
England. The sensor is formed by wrapping the optical bre around the
Figure 4.29. An optical bre current sensor.
busbar. From Ampe` res law,
I
_
H dl 4.98
so equation (4.97) becomes
0 nVI 4.99
for n turns of bre. Since the size and position of the busbar is not part of
the calculation, vibrations of the conductor are not important. However,
vibration of the bre was found to aect polarization.
Polarization can be measured using a Wollaston prism. This produces two
orthogonal, linearly polarized output beams. The two beamintensities I
1
, I
2
are
measured and the current determined using the fractional relationship
0 K
I
1
I
2
I
1
I
2
4.100
where K is a constant of the particular bre. This use of a fractional intensity
relationship overcomes problems of light source drift and path change.
Research has also been carried out by Siemens in Munich and Kema in
the Netherlands.
Kerr eect
In 1875 Kerr discovered that when an electric eld is applied to an isotropic,
dielectric material it becomes birefringent. A plane-polarized light wave
having its electric eld aligned with the applied eld is slowed by this eld.
In the Faraday magneto-optic eect described above it is the magnetic
eld from the light wave that interacts with the rotary motion of the electron
orbitals. In the Kerr eect it is the light electric eld that interacts with the
linear component of the electron orbital motion. The Kerr eect is quantied
by the change in refractive index caused and is dependent on the square of the
electric eld
n K`
0
E
2
4.101
where K is the Kerr coecient. The resulting phase dierence c suered by
the wave due to the applied eld is determined using equation (4.89) giving
c 2KE
2
dz. 4.102
The response time of the Kerr eect is of the order of 1 ps which has led to
applications in high speed optical shuttering.
The measurement of electric eld strength can be achieved by obser-
vation of changes in polarization. A plane-polarized wave inclined at 458
to the eld to be measured is injected into the bre and becomes elliptically
polarized because of the slowing of its component aligned with the external
eld. The form of the ellipse is then analysed to measure the eld strength.
Both current and voltage sensors were investigated by the CEGB. For
voltage measurement the electric eld must be applied transversely to the
bre. This was achieved by mounting the bre in spiral form and sending a
light pulse up to the high-voltage line and reecting it back to the detector
at ground potential. The state of polarization depends on the square of the
electric eld as indicated in equation (4.102). Because of this nonlinearity
the voltage measuring device proved more dicult to realize than that of
the current due to local changes in the electric eld occurring across the
bre. This was due to dust, moisture and vibration. It was suggested that
development of an appropriate crystalline bre, having a linear response
rather than squared, might overcome this problem. Alternatively, the use
of polarization, optical-time-domain reectometry, as described in section
4.5.2, could overcome the integration of the nonlinearity error.
The Kerr eect can also occur due to interaction with the electric eld
form another light wave within the dielectric. This is referred to as the
optical Kerr eect.
The photoelastic eect
A material is said to have photoelastic properties if its refractive index can be
made to alter by some applied stress o or strain . Photoelasticity was
discovered by Brewster in 1816. A wave having its plane of polarization in
the same direction as the stress travels faster than a wave with the orthogonal
polarization. Common materials exhibiting photoelasticity are gelatin, glass
and polycarbonate. The phenomenon has been used for several decades by
engineers involved in stress analysis. Usually a model is built of the structure
to be investigated in polycarbonate or similar photoelastic material. White
light is polarized, passed through the model and viewed through a further
polarizer. Stresses in the photoelastic material are said to cause retardation
between orthogonal polarizations of the components of the wave. This
results in destructive interference of a particular wavelength. The observer
sees secondary colours where the stresses are located. The compactness of
these fringe patterns indicates the levels of stress.
The following paragraphs show how quantitative analysis of stress or
strain can be achieved. Figure 4.30 illustrates a wave travelling in the z
direction through a photoelastic material being stressed in the x direction.
The retardation can be described by the relation
retardation stress stress optic coefficient optical path.
Retardation is usually quoted as a physical length, such as 150 nm. This can
be expressed as a phase shift,
phase shift stress stress optic coefficient optical path
2
`
or
c o
x
C
o
nz
2
`
. 4.103
In some applications working with strain is preferred. Retardation is then
given as
retardation strain strain optic coefficient optical path.
c
x
C
nz
2
`
. 4.104
The usual arrangement is to have a polarizing lter placed in front of the
photoelastic material. The plane of polarization is set at 458 to the x axis.
In this way light from an unpolarized source is made to have equal compo-
nents in the x and y directions. The emergent light passes through a further
polarizer placed in its path and nally a detector simply measures the optical
power.
Using equations for the electric and magnetic energy stored in an elec-
tromagnetic wave, it can be shown that the instantaneous rate of energy
ow per unit area P is the cross product of E and H,
P
1
2
E H. 4.105
P is known as the Poynting vector and its direction is that of the energy ow.
The average rate of energy ow per unit area, usually referred to as the
intensity, is determined from the rms values of E and H. For a plane wave
the ratio of E to H is a useful relationship and has been called the intrinsic
impedance of the medium Z
0
. For a plane wave polarized in the x direction
Z
0

E
x
H
y
_
. 4.106
Figure 4.30. The photoelastic eect.
Combining equations (4.105) and (4.106) for a plane wave polarized in the x
direction gives
P
x

1
2
j
_
E
2
x
. 4.107
Generalizing, equation (4.107) indicates that for a wave of any plane of
polarization the intensity is proportional to the square of the amplitude of
the electric eld in that plane.
Now, since the detector in gure 4.30 will be measuring the intensity
resulting from the combination of two orthogonal components, the separate
amplitudes must be combined before calculating the intensity. The variation
in electric eld for a wave polarized in the x-plane can be expressed as
E
x
E
ax
sin.t c 4.108
where E
ax
is the peak amplitude. A similar expression is used for E
y
but
without the phase shift c.
Finally, if the second polarizer has its orientation set at some angle 0 to
the x axis, the resultant electric eld will be
E
total
E
ax
sin.t c cos 0 E
ay
sin .t sin 0. 4.109
But, as stated above, the instantaneous intensity measured by the detector
will be related to the square of the resulting electric eld, so squaring
equation (4.109):
E
2
total
E
ax
sin.t c cos 0 E
ay
sin .t sin 0
2
. 4.110
Averaging over one cycle and assuming E
ax
E
ay
, equation (4.110) gives the
mean as
"
EE
2
total

E
2
ax
2
1 sin 20 cos c. 4.111
The above derivation has been performed for two orthogonally polarized
waves. But, as described following the diagram above, a wave plane-
polarized at 458 to the x axis is used to generate the two orthogonal waves.
If the intensity of this input wave is I
0
each orthogonal component will
have an intensity I
0
,2. (This is veried from resolving the input wave electric
eld E
0
into its x and y components; E
x
E
y
E
0
cos 45 and remembering
that intensity goes as E
2
.) So expressing equation (4.111) in terms of intensity
gives the intensity at the detector as
I
I
0
4
1 sin 20 cos c. 4.112
The usual practice is to set 0 to 458 so that for zero retardation I is zero. A
further subtle adaptation is to pre-retard one of the orthogonal components
by the use of a quarter-wave plate. Increasing the path dierence by ,2 turns
the cosine function in equation (4.112) into sine. Substituting for c gives
I
I
0
4
_
1 sin
_
o
x
2
`
C
o
nz
__
. 4.113
Since c is small this has two advantages. Firstly, near zero the sine function
changes faster than cosine thus making the system more sensitive. Secondly,
for small values the sine value of the phase and the phase expressed in radians
are virtually the same. Equation (4.113) becomes
I
I
0
4
_
1 o
x
2
`
C
o
nz
_
. 4.114
This means that o
x
becomes a linear function of the sensor output, which is
preferable. Re-arranging equation (4.114) gives
o
x

2`I
C
o
nzI
0
`
2C
o
nz
. 4.115
The development of photoelastic optical bre sensors began a generation
ago. Spillman (1982) for example reported an application as a pressure
sensor. Hydrophones and accelerometers were also developed in various
research establishments. The US Naval Research Laboratory was interested
in such devices in the early eighties. However, few systems came to commer-
cial fruition and interest has somewhat declined currently.
4.5 Distributed sensing
Measuring some variable over an extensive physical range is an attractive
concept to a number of areas of science and engineering. It is not a novel
development. Electrical engineers have for several decades been using short
voltage pulses to detect breaks in underground cables. A reection occurs
at the break and timing the return of the pulse enables an estimate to be
made of the distance to the break. Commercial instruments are readily
available for this. However, performance can be upset by the variation in
the dielectric constant of the cable due to damage or ageing which aects
the pulse speed. This procedure has been adopted by the telecommunications
industry to locate breaks or poor joints in optical bre lines. Radar and ultra-
sonic imaging are further examples of distributed sensing although not
usually thought of as such. Optical bre sensors are an attractive choice
for distributed sensing not only because of the features listed in section 4.1
but also for the higher performance and, for the case of intrinsic sensing,
the simplicity and compactness of the sensing line. A number of manu-
facturers now produce distributed optical bre sensors. The earliest available
and probably the most popular are for temperature measurement.
Optical bre systems are classed as distributed when sensing takes
place continuously along the bre length. For systems where sensing
occurs discretely at various locations the term quasi-distributed is used. The
following applications have been identied as potential beneciaries from
distributed optical bre sensors.
.
Temperature measurement in process plant, power stations, marine
systems and civil engineering structures such as long tunnels.
.
Strain measurement particularly in civil engineering and aviation struc-
tures.
.
Leak detection for pipelines, in basements and ducting, and for marine
applications.
.
Monitoring of composite materials and cements by embedded distributed
sensors for either the forming (curing or setting) stages or in later use.
.
Monitoring of electrical power engineering systems.
.
Hydrophone arrays for submarine use.
In some applications, such as a re alarms, a simple switching action is all
that is required. Knowledge of how hot or where along the line the re is
may not be needed. A cryogenic leak detector has been marketed using the
variation of refractive index with temperature to cause leakage of light
from the bre core into the cladding. The bre is simply wrapped around
the base of the cryogenic vessel or pipeline duct. When the temperature at
any point along the bre drops below 60 8C the alarm is triggered.
Complex multiplexing circuitry is not neededonly continuous intensity
monitoring of the light transmitted.
The problems facing designers of truly distributed optical bre sensing
systems are broadly twofold:
1. Identication of an appropriate sensing mechanism for the desired
measurand.
2. Choosing a suitable multiplexing style so that data values returning from
points along the line are separable.
4.5.1 Sensing mechanisms
A number of developments have been reported of sensing in distributed
systems but not all have reached commercial exploitation. Several are
described below. Fuller accounts can be found in Udd (1991) and Grattan
and Meggitt (1995).
Rayleigh scattering
This form of scattering and its origins are described in section 4.2.2. Back-
scatter within the internal cone of acceptance is only a small fraction of
the total scatter. Also, light travelling along the bre in either direction is
Distributed sensing 157
attenuated exponentially. Consequently, the backscattered signal (gure
4.31) is weak and signal recovery techniques are needed to improve the
signal-to-noise ratio.
The gradient discontinuities indicate increased loss at the two points
illustrated. These are induced by the measurand and can be derived from a
number of mechanisms such as
.
the temperature dependence of backscattering,
.
microbending,
.
evanescent wave losses,
.
temperature dependent absorption caused by a dopant in the glass,
.
variation in birefringence in a monomode bre due to the photoelastic
eect or Kerr eect,
.
rotation of polarization due to the Faraday eect.
Raman scattering
In 1928 Raman discovered that scattering by molecules not only produced
Rayleigh scattering, where there is no change in wavelength, but also
weaker, wavelength-shifted components. The eect was most easily observed
in liquids, less so in solids and gases. These other wavelengths arise because
of molecular vibrational and, in gases, rotational motions. An incoming
photon causes the molecule to switch into an excited, virtual state. The
word virtual is used here to mean an excited state that is not one of the
specic energy states within the molecule. Rayleigh scattering occurs when
the molecule returns to the ground state directly. The collision is elastic,
the outgoing photon has the same energy as the incoming photon, and
emission is isotropic. If, however, the molecule drops to an intermediate
excited state, corresponding to a vibrational or rotational mode, the
photon released has less energy. This is referred to as a Stokes photon. It
has a longer wavelength (lower frequency) than the incoming photon. The
probability of Raman scattering is orders of magnitude less likely than
Figure 4.31. Attenuation of the back-scattered pulse.
Rayleigh scattering and is an inelastic process. A further, even less likely, pos-
sibility is that the incoming photon strikes a molecule already in a vibrational
or rotational excited state. When the molecule reverts to its ground state a
photon of higher energy than the input photon is released. This is described
as an anti-Stokes photon. Figure 4.32 illustrates these transitions in an
energy diagram. The probability of the molecule being in its excited state
before the collision is governed by the MaxwellBoltzmann distribution
with a value of expfE,kTg, where E is the energy interval E
1
E
0
.
The vibrational molecular frequency is modulated by the electric eld of
the incoming photon. The Raman frequencies of the emitted photons can be
derived using the trigonometric identity given in chapter 3. If the molecular
vibrational frequency is f
m
and the photon frequency is f
0
then equation
(3.14) gives the Stokes photon as having a frequency f
0
f
m
and the anti-
Stokes photon as f
0
f
m
. Figure 4.33 illustrates this as a wavelength
diagram, more commonly used than frequency by optical engineers.
Figure 4.32. Energy level diagram for a molecule having a single (vibrational) excited state.
Figure 4.33. Representation of a scattering spectrum.
The shift in wavelength is sucient in germanium-doped, silica bres for
high-quality, narrow-band lters to separate the light into its component parts.
It is found that the emitted power from the anti-Stokes band is the most
sensitive to temperature at about 0.8%K
1
at 300 K. The usual practice is to
measure the levels of Stokes and anti-Stokes backscatter using time-
averaging techniques on the very weak signals received. The ratio is then
determined as a referencing procedure to enable temperature measurement.
Commercial devices are available giving a reasonably linear response in
the range 200400 K. Temperature resolution is about 1 8C with a spatial
resolution of a few metres in 1 km.
Brillouin scattering
Brillouin scattering is a similar phenomenon to Raman scattering except that
the scattering is from groups of molecules in motion rather than individuals.
The groups are moving in bulk because of acoustic vibrations in the material.
The energy transitions are signicantly smaller than those in the Raman
phenomena, making spectroscopic resolution of the resulting wavelengths
very dicult. Some sensor developments have occurred but commercializa-
tion of Brillouin-based systems looks less promising than for the other
forms of scattering.
Polarization
In section 4.4.2 some of the phenomena aecting the state of polarization of
light within bres were described. In principle, any such phenomena could be
useful in producing a distributed sensor. More sophisticated optical compo-
nents and types of bre are needed than a simple intensity-based system.
Quasi-distributed sensing
In principle any of the extrinsic or intrinsic sensing techniques described in
previous sections could be used with an optical bre line. The disadvantages
to this approach are the possible bulk, fragility and costs of having an
extensive bespoke system. Also, any multiple splicing or jointing will
reduce the light intensity being propagated. The advantages are that, for a
bespoke system, sensing for dierent measurands is possible as well as
having dierent ranges or sensitivities at dierent locations. Also, the
spatial resolution can be high, only limited by the multiplexing system.
Having only a few discrete sensors means that the bandwidth of the bre
is available for precisionquality rather than quantity.
4.5.2 Multiplexing
Signal processing techniques in the well-established eld of radar have been
adapted for use in distributed optical bre sensing. As with radar, the
returned signal can be very weak and may be comparable with noise in the
system. Use of a variable gain amplier may be appropriate. The amplier
gain is made to be time-dependent so as to correct for attenuation loss
with distance.
OTDR
Optical, time-domain reectometry in its simplest form is a pulse-echo tech-
nique. A light pulse that is backscattered will return to the input end after a
time delay proportional to the distance of the scattering region from the
input. Eectively, gure 4.31 also illustrates the returned power versus
time from an input pulse. For good spatial resolution short interval input
pulses are required but this limits the pulse power which means a poor
signal-to-noise ratio for the returned pulse. Averaging techniques, such as
described in section 6.4.2, can be used to improve the signal-to-noise ratio.
A strong reection can be produced from the end of the bre if this is suitably
polished and coated, which could be useful in synchronizing the trigger for
the averaging process.
A more sophisticated signal processing technique uses a random stream
of binary pulses known as a pseudo-random binary sequence or PRBS. Such
sequences are easily generated using various congurations of a shift register,
an exclusiveOR gate and a clock. An example is shown in gure 4.34. The
sequence is called pseudo-random because it is not truly random but repeats
after a time dependent on the clock speed and the number of bits in the shift
register. However, if the shift register is large enough the sequence will not
repeat for years!
The maximum possible sequence length before repetition is 2
p
1 but to
achieve this the correct value of n has to be chosen for a given p. Random
code generation has received signicant theoretical attention because of its
usefulness in a number of signal processing applications and in analysing
negative feedback systems. Commercial PRBS generators are available and
are also often incorporated in signal analysers.
In the optical bre distributed-sensor the PRBS is used to determine a
time-correlation between the outgoing and returning signals. In chapter 6
the formal mathematical relationships for correlation are described. The
Figure 4.34. A simple PRBS generator.
correlation algorithm in section 6.3.4 adapted for cross-correlation is
R
xy
rt
1
N
N
j 1
x
j
y
j r
. 6.44
The outgoing pulse sequence x is stored in memory and the returned pulse
sequence y is amplied and stored. Figure 4.35 shows a model of the
system using the Matlab
1
toolbox Simulink. The integer delay between
signals represents the transit time. Noise has been added, as seen in the
scope trace of gure 4.36, to simulate the real bre.
According to equation (6.44), the stored outgoing and returned sequences
are multiplied together point-by-point to give a set of values which are then
added together to give a particular value of R
xy
. A time-shift of one clock
pulse t is then made between the two stored pulse sequences and the multi-
plicationaddition process repeated. A correlation is said to occur when one of
the calculated values of R
xy
r t is a maximum. Knowing the number of
clock shifts r needed to produce the correlation gives the time delay of the
returned signal. Figure 4.37 illustrates the type of response.
A delay of 250 ms was inserted between signals. A correlation can be seen
at the point 1.25 ms in the diagram. (The length of the time window is set by
the correlator parameters whereas the start point is set by the simulation run-
time.)
Polarization, OTDR
As mentioned in section 4.5.1, phenomena that cause changes to the state of
polarization of light propagating along the bre can be used for distributed
sensing. Using pulses of light, as with time-domain reectometry, the local
Figure 4.35. Cross-correlation simulation.
value of a measurand can be determined. The state of polarization is meas-
ured rather than the intensity. The Faraday and Kerr eects could facilitate
the construction of distributed current and voltage sensors.
A more complex approach was reported in Dakin and Culshaw (1997).
This strategy was devised to overcome the main disadvantage to using back-
scatter, namely the low intensity of the backscatter compared with the
forward travelling wave (10
5
). Here the optical Kerr eect is used as part
of the interrogation strategy. A continuous light wave is sent along the
bre in counter-propagation to a high-power, pulsed, light wave. The opera-
tional scheme is that the continuous wave has two modes of propagation
which may be polarization states or waveguide modes, or even dual paths
Figure 4.37. Output of the cross-correlator.
Figure 4.36 Scope traces.
in two closely fabricated cores. The measurand must cause coupling to occur
between modes. The optical Kerr eect also causes coupling of light between
the modes. At the test point the combination of the eect of the measurand
and the strong electric eld from the pulse cause coupling nonlinearly. In this
way the measurand and spatial information can be determined by intensity
analysis of the two continuous output modes as a function of time. The
system requires sophisticated optical and signal processing components.
Variations of the strategy are still under development.
OFDR
Optical, frequency-domain reectometry usually employs some form of
chirp. A chirp is a varying frequency sinusoid. They have long been used
by signal processing engineers in various applications including radar. The
variation in chirp frequency can be any desired function such as a linear
sweep, logarithmic sweep or a stepped sequence. In optical bre applications
a continuous-working, laser diode is usually used, the amplitude of which is
modulated by the chirp. Because the outgoing pulse is frequency swept, the
modulation frequency of the return pulse will be low for the nearer regions
of the bre and higher for the more distant. If a stepped chirp is used a
Fourier transform of the detector output signal will deliver amplitude and
discrete distance (frequency) information.
4.6 Referencing techniques
The majority of optical bre sensors use the variation of light intensity to
determine the value of the measurand. Consequently, for accurate values
of the measurand precise measurement of the light intensity is required.
However, the value of the output signal from the detector can vary for
reasons other than the eect of the measurand. In order of signicance
these are as follows.
1. The emitted intensity of the source varying due to temperature or ageing.
LEDs have lifetimes of over ten years but intensity variation with tem-
perature is signicant, typically 0.7%8 C
1
. Halogen lamps have a
total operating time of about one year if used continuously, but this is
reduced if the lamp is used intermittently or if subject to mechanical
shock. Ageing of the lament was described in section 3.4.3. Also, a
10% change in supply voltage will cause a 4% change in the so-called
colour temperature, i.e. the spectral distribution of gure 3.13, and a
corresponding reduction in emission at the spectral peak of about 3%.
This would not only have an eect on intensity measurement but would
also change the relative intensities of the dierent colours present
making ratiometric colour referencing unreliable. Finally, attenuation in
the bre is dependent on the wavelength used, and is approximately
0.015 dB/nm
1
(gure 4.5). A change in wavelength of 1.0 nm would
cause a 0.2% change in intensity received.
2. The components in the optical paths being aected by the environment.
Introducing high curvature on bres during installation can aect the
modal distributions of the light and light leakage. Attenuation in connec-
tors can vary by 0.2 dB. This could cause a variation of up to 5% of the
full-scale deection.
3. The sensitivity of the detector varying due to temperature uctuation or
component tolerance. PIN photo-diodes have lifetimes of over ten
years. Their sensitivity to light varies at 0.4% 8 C
1
and is also wave-
length dependent.
(The performance of sources and detectors was described in sections 3.5.3
and 3.5.4.)
Any combination of these sources of error can occur. Consequently, for
high performance, some form of referencing is required if regular calibration
is to be avoided. Some elaborate systems have been developed to minimize
these eects. Generally, the more sophisticated the system the more bulky
and expensive it becomes. Some of the more common techniques are outlined
below. A list of references is given in Grattan and Meggitt (1995). Systems
generally involve some form of ratiometric measurement for referencing.
Double detection
Two detectors can be used to monitor the behaviour of the source as
illustrated in the simple system of gure 4.38.
The received intensity at the detectors can be expressed as
D
1
cMS. D
2
uS 4.116
where c and u are calibration constants involving the optical lines. Eliminating
S gives
M
uD
1
cD
2
. 4.117
This method relies on c, u and the sensitivity of the detectors being xed. To
reduce the eects of having dierent optical paths and detectors Y-connectors
Figure 4.38. Referencing to compensate for source drift.
Referencing techniques 165
can be used. However, the use of Y-connectors can introduce further errors,
add to the cost and may require the use of some form of multiplexing as
described below.
Delay line
One simple multiplexing technique involves an optical delay-line as shown
in gure 4.39. The source emits a short pulse that arrives as two separate
pulses at the detector. High-speed signal conditioning circuitry is needed to
measure the amplitude of each pulse. This technique is, in-eect, time-
division multiplexing.
Wavelength modulation
If the eect the measurand has on the intensity of the light received is
wavelength dependent, this can be used as a form of multiplexing. In tele-
communications engineering this would be referred to as frequency-division
multiplexing. The ratio of the wavelengths detected provides normalization.
A single, broad-band source can be used as illustrated in gure 4.40(a), or
two separate light sources as shown in gure 4.40(b).
Figure 4.39. Multiplexing using a delay-line.
Figure 4.40. (a) Broad-band source, dual-wavelength referencing system. (b) Dual-source,
dual-wavelength referencing system.
Amplitude modulation
When using two sources an alternative approach avoiding the use of optical
lters is to drive the sources at dierent frequencies (gure 4.41). Electronic
lters are then used to separate the signals after the detector.
The optical bridge
A more optically sophisticated approach is the so-called optical bridge
(gure 4.39). This takes its name from the electrical bridge circuit and
involves two sources and two detectors. If the two sources are pulsed
sequentially then
D
1
cMS
1
. D
2
S
1
4.118
and, a short time later,
D
1
uS
2
. D
2
cS
2
4.119
where c, u, and c are the calibration constants of the optical paths.
Combining these equations to eliminate S
1
, S
2
, D
1
and D
2
D
1
D
2
cM
.
D
1
D
2
u
c
4.120
giving
M
u
cc
4.121
so that M only depends on the stability of the optical paths.
Figure 4.42 shows sources of dierent wavelength. With the TDM
approach described above, identical sources and detectors would be appro-
priate. Also, of course, frequency modulation of two identical sources
could be used with electrical ltering of the signals from two identical
detectors.
Finally, it should be noted that the presence of ambient light might also
contribute to the signal(s). This may be easily overcome by having the output
of the source amplitude-modulated. Using a frequency of the order of 1 kHz
will enable sunshine, lamplight at the power supply frequency and its harmo-
nics to be removed. The desired signal is recovered by using a narrow-band,
Figure 4.41. Dual wavelength referencing system using electronic lters.
Referencing techniques 167
electronic lter in the detector signal processing circuitry set at the modula-
tion frequency.
References
Bleaney B I and Bleaney B 1965 Electricity and Magnetism 2nd edition (Oxford University
Press)
Dagenais D M, Bucholtz F and Koo K P 1988 Elimination of residual signals and reduc-
tion of noise in a low-frequency magnetic bre sensor Appl. Phys. Lett. 53 14171476
Dakin J and Culshaw B 1988 Opical Fiber Sensors: Principles and Components vol 1
(Boston: Artech House)
Dakin J and Culshaw B 1997 Opical Fiber Sensors: Applications, Analysis and Future
Trends vol 4 (Boston: Artech House)
Dotsch H, Martens G and Meyer W 1983 Fibre Optic Components for Industrial Control,
IEE Conf. pub. 221, Optical Fibre Sensors (London: IEE) pp 6771
Frank W E 1966 Detection and measurement devices having a small, exible ber trans-
mission line US Patent 3,273,447 September
Go pel W, Hesses J and Zemel J N 1992 Sensors: a Comprehensive Survey vol 6 Optical
Sensors (New York: VCH
Grattan K T V and B T Meggitt (eds) 1995 Optical Fiber Sensor Technology (Chapman and
Hall)
Gupta G C 2002 Optical Fiber Communication Conf., Anaheim, CA, paper TuA5
Jones B E 1981 Simple optical sensors for the process industries using incoherent light
Symp. Optical Sensors and Optical Techniques in Instrumentation, London, UK
(London: IMC) p 15
Jones B E 1985 Optical bre sensors and systems for industry J. Phys. E: Sci. Instrum. 18
770782
Jones B E and Philp G S 1983 A bre optic pressure sensor using reective techniques
Proc. Int. Conf. Optical Techniques in Process Control, The Hague, Netherlands
(Craneld: BHRA Fluid Engineering) pp 1125
Figure 4.42. The optical bridge referencing system.
Keck D B, Maurer R D and Schultz P C 1973 Appl. Phys. Lett. 22 307
Kirkbright G F, Narayanaswamy R and Welti N A 1984 Fibre-optic pH probe based on
the use of an immobilised colorimetric reagent Analyst 109 1025
Kissinger C D 1967 Fibre optic proximity probe US Patent 3,327,584, June
Lakowicz J R 1983 Principles of Fluorescence Spectroscopy (New York: Plenum Press)
Land H E 1951 Some aspects of the development of sheet polarizers J. Opt. Soc. Am.
41(12) 957
Livingston J D 1982 Magnetomechanical properties of amorphous metals Phys. Status
Solidi 70 591596
Longhurst R S 1970 Geometrical and Physical Optics 2nd edition (Longmans)
Nishihara H, Haruna M and Suhara T 1989 Optical Integrated Circuits (New York:
McGraw-Hill)
Nishizawa K, Sudo E, Yoshida M and Yamasaki T 1986 High sensitivity waveguide type
hydrogen sensor 4th Int. Conf. Optical Fibre Sensors, Tokyo p 131
Ohashi M, Shiraki K and Tajima K 1992 Optical loss property of silica-based single mode
bers IEEE J. Lightwave Technol. 10(5) 539543
Pe rot A and Fabry C H 1898 Sur un voltme` tre electrostatique interfe rentiel pour etalon-
nage J. de Phys. 7 650659
Post E J 1967 Sagnac eect Rev. Mod. Phys. 39 475
Rogers A J 1979 Optical measurement of current and voltage on power systems IEE
J. Electronics and Power Appl. 2(4) 120124
Scheggi A M and Baldini F 1993 Chemical sensing with optical bres Int.
J. Optoelectronics 8(2) 133156
Senior J M 1985 Optical Fibre Communications, Principles and Practice (Prentice-Hall)
Spillman W B jnr 1982 Multimode bre-optic pressure sensor based on the photoelastic
eect Opt. Lett. 7 388
Spillman W B jnr and McMahon D H 1980 Frustrated-total-internal reection multimode
bre-optic hydrophone Appl. Opt. 19 113117
Udd E (ed) 1991 Fiber Optic Sensors, An Introduction for Engineers and Scientists (Wiley-
Interscience)
Yamada J I, Saruwatari M, Asatani K, Tsuchiya H, Kawana A, Sugiyama K and Kimura
T 1980 High speed optical pulse transmission at 1.29 mm wavelength using low-loss
single-mode bers IEEE J. Quantum Electronics QE-14 791
Exercises
1. Compare the performance of metallic telemetry lines with those of optical
bre. List the reasons why optical bre sensors are attractive.
2. Explain the signicance of the numerical aperture when dealing with
optical bres and derive the formula.
3. (a) Attenuation in optical bres can be caused by absorption and scatter-
ing. Describe the nature of these phenomena and include a sketch
graph.
(b) A test facility is to be built to investigate wavelength related attenua-
tion in optical bres.
Exercises 169
(i) Draw a diagram of an appropriate arrangement suggesting a
suitable light source and detector.
(ii) Draw a signal conditioning circuit for your detector and explain
its operation. (Hint: see chapter 3.)
(iii) How might referencing be achieved to overcome drift in the
system performance?
4. Explain how black-body radiation can be used to measure temperature by
means of an optical bre.
(a) Draw a diagram naming the materials and optical components of
such a device.
(b) Drift in the detector and optical couplings can be accounted for by
using ratiometric measurement. Using the Planck formula given in
section 3.5.3, determine the intensity ratio for light of wavelengths
550 nm and 450 nm at a temperature of 700 8 C and then at 725 8C.
(You may assume that the lters used are non-attenuating and
have the same very narrow bandwidth.)
(c) How might the problem of the sensor probe tip not being a perfect
black-body be accounted for?
5. Two optical bres are placed side-by-side to form an intensity-based,
reectance-mode, displacement sensor. Their core radii are a
1
and a
2
and the separation of their centres is D.
(a) Explain, with the aid of a diagram, how movement of the reector
can produce a near-linear variation in intensity with displacement.
Derive two expressions l
min
and l
max
for the limits of this near-
linear, displacement range.
(b) How might such a device be used to measure temperature? Draw a
sketch.
6. An optical-bre chemical sensor is to be developed based on the absorp-
tion principle.
(a) State the LambertBeer absorption law.
(b) With the aid of a sketch describe a suitable test-cell structure.
(c) By judicious choice of the signal conditioning circuit a photodiode
detector could be made to output a linear response with concentra-
tion. Explain this quoting any appropriate formulae. (Hint: see
photodiode section in chapter 3.)
7. Design an evanescent, front-end displacement sensor after the style of
Figure 4.19. Determine an appropriate oblique angle for the front face
using typical values for the refractive indices etc. Determine the sensitivity
for your design as the intensity-displacement gradient, i.e. dI,dx.
8. The Faraday magneto-optic eect and the Kerr eect can be used to
measure electrical current and eld. Sketch an arrangement whereby
both of these phenomena could be implemented in one optical bre sensor
so that, in principle, an electrical power sensor could be built for use on a
high-voltage overhead power line. What would be the major diculty to
be overcome?
Exercises 171
Chapter 5
Intelligent sensing
5.1 Introduction
The development of the microprocessor and the subsequent IBM personal
computer (PC) led to a growth in the processing of the data from sensors
during the 1980s. Many data acquisition boards or cards have become
available for plugging into the IBM-type bus with sophisticated support soft-
ware for data logging, processing, analysis and display. Such activity has
been described as intelligent instrumentation. More recently sensors have
become available having a microprocessor within the sensor head enabling
processing to be carried out in situ. These devices have come to be termed
smart sensors and are outlined in section 5.5. Articial intelligence (AI) soft-
ware is also nding application in instrumentation systems for interpreting
the output of one or more sensors. Some of the developments of AI are
briey described in section 5.6.
Before considering the application of intelligence to sensing it is impor-
tant to appreciate the restrictions placed on the empirical data caused by the
interfacing to a digital system. The output of a sensor will generally be an
analogue, smoothly-varying function with time. However, to incorporate
this signal into a digital system it has to be limited in two ways:
.
It must be time-sampled. This is sometimes called discrete time-sampling.
During the processing of the sampled value of the measurand the system
is o-line and so empirical data will be lost.
.
Because a digital system has a prescribed number of bits, resolution of
the empirical data will be reduced. The continuous span of the input
measurand becomes discrete in value.
Time-sampling in PC-based systems may be deliberately limited to moderate
rates so that signicant processing and outputting can be carried out between
the taking of the samples. Such a system would be described as operating in
real time because there is no need for archiving along the data acquisition
pipeline. Provided that the run-time of the algorithm doing the processing is
not too involved, sample rates of up to 10
5
per second are possible. However,
graphical interfacing to the monitor is slow and when visual presentations of
the data are required this will severely reduce the acquisition speed if the
system is to remain real-time.
The sampled data must be changed to digital format using an analogue-
to-digital converter (ADC). In a simple system this will be under the control
of the PC. However, in systems working at speeds beyond the limit of the PC,
a separate board with its own microprocessor is used. The ADC and any
extra memory are controlled locally on the board. The board may also
process the data. Specialist digital signal processing (dsp) microprocessors
are commonly used with algorithms for digital ltering or Fourier analysis
etc. (see section 5.4). In this way large volumes of data, sampled at high
speed, can be stored and processed. The processed information can then be
sent to the PC at a slower rate for further analysis and display. Such a
system would be described as real-time if the information was owing
through each successive section of the system at the same rate.
Figure 5.1. The time-sampling of a continuous waveform.
Introduction 173
The problems related to discontinuous and discrete sampling are
described in the following sections.
5.2 Discontinuous-time sampling
When a waveform f t is time-sampled at some rate a new signal is produced
which is, in eect, an amplitude modulated version of the original:
f
ptf t. 5.1
The function pt is a train of narrow unit pulses at a frequency .
s
t. This
process has the eect of producing sidebands in the frequency spectrum as
illustrated in gure 5.2.
The sidebands occur centred at multiples of the sampling frequency as
shown. To avoid the sidebands overlapping with the original, the sampling
frequency must be chosen with reference to the maximum frequency .
max
within the function f t. From the diagram it can be seen that for the
centre spectrum not to touch one of the sidebands,
.
s
! 2.
max
. 5.2
This result was discovered independently by Shannon and Nyquist. It
can be formally veried as shown below. Using complex notation for the
Fourier series, pt can be written as
pt
X
1
k1
c
k
exp jk.
s
t 5.3
where
c
k

t
T
sinkt,T
kt,T
5.4
or
c
k

t
T
sinc
kt
T
. 5.5
Figure 5.2. The frequency spectrum of an ideally sampled function.
174 Intelligent sensing
T is the sampling period and t is the width of the sampling pulse. The time-
sampled function of f t is
f
s
t f tpt 5.6
f t
X
1
k1
c
k
exp jk.
s
t. 5.7
The Fourier transform of f
s
t is
= f
s
t
1
1
X
1
k1
c
k
exp jk.
s
t f t exp j.t dt 5.8
X
1
k1
c
k
1
1
f t exp j. k.
s
t dt 5.9
X
1
k1
c
k
= f . k.
s
. 5.10
The interpretation of equation (5.10) is that the Fourier transform of f
s
t is
the Fourier transform of f t but is repeated at multiples of .
s
and that these
multiples are attenuated by diering amounts according to the function c
k
.
This conrms the relationship stated in equation (5.1) and in addition
shows that for a sampling function that is not made from innitely narrow
pulses, the sidebands are attenuated by a sinc function.
If a system is designed for input frequencies up to .
max
, having
.
s
2.
max
, and the signal increases to .
max
c., then a lower frequency
is produced at the system output of .
max
c.. This is known as aliasing.
Aliasing is the production of low frequencies in the output by the sampling
system itself. In systems that have a xed sampling rate an anti-aliasing
lter may be needed to limit the input frequencies to <
1
2
.
s
. Because realis-
able lters do not have sharp cut-o frequencies it is necessary to allow a
greater frequency interval between the spectrum and the sidebands. High
delity, real systems may have sampling frequencies of 6.
max
.
5.3 Discrete sampling of the measurand
Any digital word is binary weighted with the most signicant bit (MSB)
having the highest weighting and the least signicant bit (LSB) the lowest
weighting. It follows therefore that the best resolution that can be achieved
when converting from an analogue value to a digital word is
1
2
LSB. This
is sometimes referred to as the quantization error and can be up to
1
2
LSB. Figure 5.3 illustrates the eect.
Consequently, when designing a data acquisition system the required
resolution is used to determine the number of bits needed in the digital
Discrete sampling of the measurand 175
word. For example, an 8-bit ADC able to accept analogue inputs within the
range 010 V will have a resolution of
1
2
10
255
V or 0.01961 V.
As well as the
1
2
LSB quantization occurring across the range it also creates a
zero error of
1
2
LSB as shown in gure 5.3.
5.3.1 Digital-to-analogue conversion
Digital-to-analogue converters (DACs) are described here before ADCs
since some ADCs use a DAC as part of their structure. A common use of
the DAC is for the conversion of the digital output of a computer system
into an analogue value. Another use in instrumentation is with digital
sensors when analogue processing/transmission of the output is required.
Examples of digital sensors are the Gray code disc, for encoding the position
of a rotating shaft and the multi-probe conductivity sensor for liquid level
measurement.
A digital word can be converted into an analogue current or voltage by a
variety of switching circuits. The most commonly used circuit is the R2R
ladder network. The circuit shown in gure 5.4 has an op-amp as an
output buer summing the switched currents to provide a voltage output.
Without the op-amp the circuit will act as a current source.
The R2R conguration ensures that the current is halved at each
junction. Currents are switched between ground and the virtual ground of
the op-amp. Thus, provided the reference voltage source V
ref
and the
resistors remain constant, the fractional weightings of the currents will be
correct and their summation in the op-amp will provide the correct analogue
Figure 5.3. An ADC transfer curve.
output. The output will, of course, be stepped rather than smoothly varying.
If this is undesirable then placing a low pass lter on the output will remove
the step eect.
There are two main types of error that may occur with DACs. These can
best be appreciated using the graph in gure 5.5. If the staircase curve were
not to be properly aligned with the straight line (perhaps due to a drift in the
internal voltage source V
ref
) then there would be a calibration error across
the range. The error could be up to 1 LSB at full scale. The other error is
to do with nonlinearity (this might be due to poor tolerances or drift of
the resistor values). Errors occur across the range and could be
1
2
LSB or
greater.
Figure 5.4. The R2R ladder network.
Figure 5.5. Types of error with DACs.
Settling time
The switching time for the components in the DAC is not innitely fast. The
settling time is that time required, following an input change, for the
analogue output to settle to within
1
2
LSB of the nal value.
In current output DACs the output is converted to a voltage by simply
connecting an external load resistor to the current line. In this type of DAC
output settling-time is governed by the time constant associated with the
external load resistance, the DAC output capacitance and any stray
capacitance in parallel with the external load resistor.
In voltage output DACs the op-amp buer gives the advantage of low
output impedance but at the cost of reduced speed. The slew rate of the
op-amp (the maximum rate of change of output voltage) has a major
inuence on the system settling time. Consequently, the LSB will have the
shortest settling time and the MSB will have the longest. Also, the external
circuitry has a more pronounced eect on the settling time than that for
the current output DAC and may contribute signicantly to it.
Glitches
Glitches in digital systems are transient voltage spikes, positive or negative
going, occurring because of shortcomings in the switching process. The
time span of a glitch is typically of the order of a nanosecond. Because of
this, and the somewhat random nature of their occurrence, their detection
can be extremely dicult. This makes faultnding in a working system,
using a conventional oscilloscope or analyser, an unenviable task.
Glitches can occur at the analogue output of a DAC when a transition
is occurring in the MSBs. Figure 5.6 illustrates the possible outcome with a 3-
bit DAC.
Figure 5.6. An example of a glitch on a 3-bit DAC.
On being incremented by 1 LSB the input value 011 should change to
100,
011
000 glitch
100
but a large output glitch may occur of 000 if the 1s become 0s before the 0
becomes a 1.
Binary codes have been developed to minimize the transitional error
that can occur. The Gray code, mentioned above in the context of a rota-
tional sensor, is one such code. At each transition the bit is changed that
would give the least error if mis-read. Some values are shown in table 5.1.
Glitches are less likely to occur in fast-switching systems. If necessary,
glitches may be removed using a zero-order-hold (ZOH) circuit. The
expression zero order is used to denote that the device does not modify the
value it is holding. Other hold devices attempt to correct for changes that
are occurring while o-line. They apply a correction related to a rst or
higher order polynomial. Figure 5.7 illustrates the schematic diagram of a
ZOH. The DAC is connected to the capacitor until just before a conversion
is ready. The switch is then opened to isolate the DAC. During this time the
capacitor provides the output value. After a suitable settling time the switch
is closed again and the capacitor value and output up-dated.
Table 5.1. The Gray code.
Deanery number Standard binary code value Gray code value
0 0000 0000
1 0001 0001
2 0010 0011
3 0011 0010
4 0100 0110
5 0101 0111
6 0110 0101
7 0111 0100
8 1000 1100
9 1001 1101
10 1010 1111
11 1011 1110
12 1100 1010
13 1101 1011
14 1110 1001
15 1111 1000
This process introduces an extra delay and so slows the overall rate of
conversion.
5.3.2 Analogue-to-digital conversion
There are a large number of ADCs on the market using a variety of tech-
niques. Prices vary from a few pounds to over 100. The highest prices are
paid for ADCs with high resolution and high speed. Prices continue to fall
and this is particularly the case for those enjoying mass application. ADCs
used in video systems are an example of a higher performance device being
reasonably cheap. The following sections illustrate three broad groupings of
ADCs according to their operating technique, performance and relative cost.
1. The parallel or ash conversion ADC
These devices are the fastest and most expensive ADCs. If high resolution is
also required cost is high. The technique involves a reference voltage source
which is connected across a potential divider chain to provide a set of refer-
ence voltages. The analogue input is simultaneously compared with all of
these values using a set of comparators. Comparator circuits operate at
high speed giving a fast result, typically <30 ns. The comparator outputs
are encoded into an n-bit digital word. For an 8-bit ADC the converter
would require 255 comparators. The schematic circuit for a 3-bit device is
shown in gure 5.8.
For very high rate data acquisition systems, where even this device is not
fast enough, the signal may be multiplexed to several such ADCs so that the
conversion time does not limit the acquisition rate.
2. The dual slope ADC
The dual slope conversion technique is a slow procedure. However, high-
resolution devices can be purchased cheaply. (See gure 5.9.)
The integrator is initially connected to the analogue input for a xed
time period, t
1
. The control unit then switches the integrator input to a
reference voltage of opposite sign and the time taken t
2
, to discharge the
Figure 5.7. The zero-order-hold device.
capacitor is determined. The timings (gure 5.10) are carried out using a
clock, a logic unit and a counter. During the time t
1
, the capacitor acquires
the potential dierence V, where
V t
1
V
in
RC
5.11
V
N
0
f
c
V
in
RC
5.12
Figure 5.8. The parallel conversion ADC.
where f
c
is the clock frequency and N
0
is the number of clock cycles that sets
the charging time t
1
.
During the integration of the reference source the capacitor must suer
the same change in potential but of opposite sign so as to become discharged.
That is
V t
2
V
ref
RC
5.13
V
n
f
c
V
ref
RC
5.14
where n is the variable number of clock cycles required to discharge the
capacitor. Setting equation (5.12) equal to (5.14) for complete discharge,
N
0
f
c
V
in
RC

n
f
c
V
ref
RC
5.15
or
n N
0
V
in
V
ref
. 5.16
This shows that the value of the count n is directly proportional to the
analogue input V
in
. High precision of the values of R and C and other
components is not required. However, component stability is essential.
Also, variation in switch resistance and the reference voltage source will
Figure 5.9. Simplied schematic for the dual slope ADC.
Figure 5.10. Dual slope ADC timing sequence.
signicantly aect accuracy. The main disadvantage is that the conversion
time t
1
t
2
is long, typically !40 ms and is dependent on the input level.
This relatively simple technique provides high resolution chips at low
cost. They are widely used in digital electronic instruments such as digital
multimeters and phase meters. Because of this well-established application
area these ADC chips often contain additional features such as auto-
zeroing. Also, in applications where interference may be a problem, the
value of t
1
can be set to an integer number of cycles of the interfering
frequency. Integrating V
in
over this period would tend to average the
interference amplitude to zero.
3. Feedback converters
In terms of performance, feedback converters fall between the ash con-
verters and the dual slope devices. They are the most popular ADCs, with
a broad range of specications available from many manufacturers. As the
name implies a conversion is performed which is then fed back through a
DAC and a comparison made with the original analogue input. The many
types of feedback converter can be classied according to the sequence by
which the output is generated.
3a. The successive approximation converter
This is one of the most popular of all of the types of ADC (gure 5.11). The
operating procedure is sometimes called a binary search. The bits in the
generated binary word are set high in sequence and at each change their
value is compared with the input.
At the start of conversion the logic circuitry sets the MSB high and then
waits for the decision from the comparator. If the digital output is larger than
the analogue input the MSB is set to zero. If the digital output is less than the
analogue input the MSB is retained at 1 and the process is repeated on
the next lower MSB. Each new test eectively is evaluating half of the
range of the previous one. Because of this the process is also referred to as
Figure 5.11. The successive approximation ADC.
an interval-halving technique. For an n-bit digital output there will be n com-
parisons made. This means that the conversion time for this type of feedback
converter is xed regardless of the behaviour of the analogue input.
3b. The tracking converter
In this ADC an up-down counter (gure 5.12) provides the digital output. A
comparator again monitors the input and feedback signals and controls the
up/down incrementing as appropriate.
If the digital output signal is too high the counter will decrement its
output by 1 LSB and, if too low, the counter will increment its output by
1 LSB. Usually, the comparator has thresholds so that the digital output
does not change if the error is less than
1
2
LSB. The time to generate the
conversion is one clock cycle.
As long as the input is not changing by more than 1 LSB during a clock
cycle the converter will correctly output the digital equivalent of the analogue
input. An illustration is given in gure 5.13.
A summary is given below in table 5.2 of the ADC types and their per-
formance. There are a number of other types of ADC available. Some, like
the dual slope converter, are based on charge balancing; examples are the
switched capacitor and the delta-sigma systems. Others employ single-slope
Figure 5.12. The tracking ADC.
Figure 5.13. Performance of a tracking ADC.
integration or voltage-to-frequency conversion. See Horowitz and Hill
(1989) for descriptions.
Conversion speed and frequency response of ADCs
ADCs are designed to accept either bipolar or unipolar input, such as 5 V
or 010 V and come with resolutions of 6, 8, 10, 12, 16 or 18 bits. If the full
resolution of an ADC is to be utilized the signal input level should be close to
the maximum input voltage that the ADC can accept. Also, to maintain reso-
lution to 1 LSB the amplitude of the analogue input signal must not change
by more than the equivalent of
1
2
LSB during the time in which the ADC
performs the conversion. To determine the input frequency limit that can
be oered to the device an analysis is performed using a sinusoidal input
V V
0
sin .t. 5.17
Dierentiating to determine the slew rate,
dV
dt
.V
0
cos .t. 5.18
The maximum slew rate is therefore
dV
dt
.V
0
. 5.19
The implications of the criteria described above can be seen in the analyses
given below. The ADCs are assumed to be unipolar. Consequently,
because a sinusoid has a maximum amplitude swing of 2V
0
for best
performance,
2V
0
V
max
5.20
where V
max
is the maximum permitted input voltage.
Table 5.2. Comparison of some ADC types.
Analogue-to-digital
converter
Cost Speed Comments
Parallel or ash High Very fast
(30 ns)
Used in fast, high-performance systems.
Up to 10-bit versions available, but 8-bit
is the most common.
Dual slope Low Slow
(40 ms)
Variable conversion time. Can be
congured to reduce interference. Good
accuracy, 10-bit to18-bit versions are
available.
Feedback
converters
Moderate
to high
Fast, 25 ms
to 400 ns
Up to 16-bit resolution is available.
Generally, ADCs such as the ash converter and the tracking converter
can change their output at the rate of 1 LSB per clock cycle. But, as already
stated the correct digital output will only be delivered as long as the rate of
change of the analogue input does not exceed
1
2
LSB per clock cycle. Hence
the maximum acceptable slew rate is given by
dV
dt

1
2
LSB
clock period
. 5.21
Now,
1 LSB
V
max
2
n
5.22
2
n
being the number of increments that V
max
is divided into. Also, the clock
rate f
c
can be substituted for the clock period,
clock period
1
f
c
. 5.23
So equation (5.21) becomes
dV
dt

1
2
V
max
2
n
f
c
5.24
dV
dt

V
max
2
n1
f
c
. 5.25
Equating the two slew rates of equations (5.19) and (5.25) gives
.V
0

V
max
2
n1
f
c
. 5.26
Substituting for V
max
using equation (5.20),
.V
0

2V
0
2
n1
f
c
. 5.27
This sets the maximum value of the input frequency .
max
as
.
max

f
c
2
n
. 5.28
For an 8-bit resolution ADC with a clock rate of rate of 10
6
s
1
, equation
(5.28) gives
f
max
621.7 Hz. 5.29
In this example the digital output is updated every 1 ms. But the analogue
input frequency must not exceed the 621.7 Hz. Also, of course, no fast
transients or steps must occur. This is in stark contrast to the frequency
limit set by the ShannonNyquist sampling theorem described in section 5.2.
The sampling theorem states that f
max
< 2f
s
, the term f
s
being the sampling
rate. In this example the sampling rate is the clock rate of 10
6
s
1
which
sets the limit of f
max
as 500 kHz.
A similar analysis can be performed for the successive approximation
converter. As explained above an n-bit device would take n clock cycles to
complete a conversion. As before, while the ADC is performing the con-
version the analogue value must not change by more than the equivalent
of
1
2
LSB. Again, considering the slew rates as previously,
dV
dt

1
2
LSB
n clock cycles
5.30
so the maximum frequency reduces to
.
max

f
c
n2
n
. 5.31
Using the same numerical values as the previous example, a resolution of
8-bits and clock period of 1 ms then
f
max

10
6
82
9
or 77.71 Hz. 5.32
The conversion time is 8 ms, so the sampling rate can only be up to
125 ksamples s
1
. For this sampling rate the ShannonNyquist theorem
sets the maximum frequency input as 62.5 kHz.
Clearly, restricting the movement of the analogue input to
1
2
LSB
during conversion to provide good resolution is limiting the frequency
performance of the ADCs. This restriction may be overcome by having a
device before the ADC which samples the analogue signal and inputs this
as a xed value to the ADC for the length of time it takes to perform the
conversion. This holding technique is described below.
The sample-and-hold device
The zero-order-hold device of gure 5.7 and the sample-and-hold device are
functionally very similar. The need to use a sample-and-hold device will
depend on the frequencies present in the analogue input signal and the
resolution required. Sample-and-hold devices are not innitely fast at
sampling the signal but this aperture time, as it is known, can be typically
as low as 0.3 ps. This eectively improves the acceptable input slew rate to
dV
dt

1
2
LSB
aperture time
. 5.33
Equating this with the maximum slew rate of a sinusoid given in equation
(5.19) gives
.V
0

1
2
2
n
V
max
aperture time
. 5.34
As before, if the full input range of the ADC is used 2V
0
V
max
giving
.
max

1
2
n
aperture time
5.35
which is similar to previous expressions for .
max
but the aperture time has
replaced the clock period. Using the previous example of an 8-bit device
and assuming an aperture time of 0.3 ps, equation (5.35) gives f
max
as
2072 MHz! Clearly, a marked improvement has occurred in the maximum
acceptable input frequency.
The maximum permissible input frequency is now only limited by the
ShannonNyquist sampling theorem. All ADCs take some nite time t to
perform the conversion. During this time they are o-line and so this
eectively sets the sampling rate. If a sample-and-hold device is used then
the sampling time is slightly lengthened. The ShannonNyquist theorem
gives the maximum acceptable input frequency as
f
max
<
1
2t aperture time
. 5.36
To conclude, the use of a sample-and-hold device enables the theoretical limit
set by the sampling theorem to be realized without deterioration in ADC
resolution.
5.4 PC-based data acquisition boards
There are a large number of boards now commercially available and con-
sequently it is worthwhile considering the criteria for choosing a particular
product. Boards range in price from less than 100 to several thousand
pounds. The large price range is due to the variation in speed, resolution
and the sophistication of the data handling. A simple board would typically
have an ADC of 12-bit resolution, with a conversion time of 1 ms and no data
handling capability other than simply transferring information to the PC
bus. As well as having a well-known brand name, a board at the top of
the price range will have some memory and processing power of its own
and perhaps be able to write directly to the hard drive. However, the most
signicant feature to increase cost is the sampling speed.
A high speed board might have an ADC of 16-bit resolution operating
in <10 ms. Such ADCs alone can cost in excess of 100. There may be one
such ADC on each channel since multiplexing of input channels through
one ADC converter slows the sampling rate. For this reason the highest
speed boards usually only have a few input channels. When reading
manufacturers specications it is important to distinguish between the
conversion rate of the ADC and the actual sampling rate of a particular
input channel. Clearly, a board with a single ADC and 16 input channels
will be sampling at least 16 times slower than the ADC. Some boards will
allow the multiplexer to be turned o so that the one functioning channel
has the full ADC speed.
To appreciate the way the PC microprocessor communicates with these
boards a little knowledge of microprocessor system architecture and the
handling of data input/output is needed. Broadly speaking a microprocessor
uses two designated regions of memory, one for storing the program and the
other for the data being processed. When the computer system boots up, the
program is loaded to the microprocessors electronic memory from a non-
volatile memory (not lost at power down) such as a hard disk. The electronic
memory is provided within the microprocessor itself and also externally on
arrays of memory chips. Electronic memory is used because of its far
higher access speed compared with that of memory media used for archiving.
The microprocessor has an addressing system for access to these electronic
memory locations. During the design of the microprocessor chip and its
peripherals the addresses of the various locations of memory are designated.
This is referred to as the memory map. Within the data memory mapped
locations, several addresses will be input/output ports and not actual
memory. When the program is running and data are collected from or sent
to these particular locations the microprocessor is eectively communicating
externally.
Software to transfer digital data via the PC ports can be written in a
high level language such as Pascal or some version of C. This is relatively
straightforward and when using commercially available data acquisition
and processing boards is usually supplied with the board by the manu-
facturers. The example given below illustrates the use of an inline routine
to access data presented to a PC input port from a data acquisition and
processing board. The expression inline refers to a section of code written
in assembly language within a program written in some high-level language.
This involves the direct use of the instruction set for the microprocessor as
devised by the chip designers. However, because every data movement and
processing action has to be specied individually it is a tedious activity
and time consuming. Its advantage is that after compilation, the conversion
of the instructions into a binary sequence for use in the processing unit,
the execution run-time is minimized. In the routine described below the
PC waits to retrieve a block of 128 data values which could, for example,
be the frequency spectrum produced from a fast Fourier transform.
These data values are generated by a separate microprocessor mounted on
PC-based data acquisition boards 189
the specialist processing board. (The board would also have it own
algorithms and memory.)
.........
pascal instruction
pascal instruction
pascal instruction
inline(
$BA/$0294/ { MOV DX,294H}
$B3/$20/ { MOV BL,32}
$EC/ {LOOP 1: IN AL,DX}
$84/$C3 { TEST AL,BL}
$74/$FB { JZ LOOP1}
);
N: =1;
REPEAT
inline (
$BA/$0294/ { MOV DX,294H}
$B3/$02/ { MOV BL,2}
$EC/ {LOOP 2: IN AL,DX}
$84/$C3 { TEST AL,BL}
$74/$FB { JZ LOOP2}
);
DATA[N]: = PORTW [$290];
CLEAR: = (PORT[$294] AND 2);
N: = N+1;
UNTIL N=129;
pascal instruction
pascal instruction
pascal instruction..........
In the rst loop the PC input port 294H is checked by the PC to see if the
board is ready to transmit data to it. The board will set a ag (a specic
binary value) at this port when it has completed the algorithm that generates
the 128 values. When this is the case the PC moves on to the second test (loop
2). Then each time the board moves a data value to its output port it again
sets the ag. This triggers the PC to retrieve this data value presented to it
at its input port address 290. After collection the PC then resets the location
294H and loop 2 is repeated awaiting the next data value. After all 128 values
have been transferred the PC returns to loop 1 waiting for the completion of
the next sampled window being processed by the board. Data transfer is done
in parallel format for maximum transfer speed.
Selection criteria
When selecting a data acquisition or processing board it is useful to consider
the following parameters.
.
The sampling rate required. This must be at least twice the frequency of the
highest frequency expected in the input signal.
.
The voltage resolution required and hence the number of bits.
.
The number of channels needed.
.
The range of the input analogue voltage expected.
.
Whether the input analogue voltage range of the board can be altered.
.
Whether the input analogue voltage expected is single or bipolar.
.
Whether dierential input may be required/is available.
.
Whether a choice between a.c. and d.c. input coupling is available.
.
Whether digital input/output is needed.
.
Whether analogue output is needed.
.
Noise levels.
.
Linearity.
.
Temperature dependency.
.
Cost.
5.5 Smart sensors
Many sensors are now sold which have a microprocessor contained within
the sensor unit and are described as smart sensors. This has enabled a
higher specication of the sensor units to be achieved because of the on-
board processing. A general scheme following the suggestion of Brignell
and Dorey (1983) is shown in gure 5.14.
Figure 5.14. A smart sensor architecture.
Smart sensors 191
The sensor itself is under microprocessor control, enabling the
excitation levels to be adjusted so that switching between ranges can be
done automatically. Similarly, ambient temperature and other parameters
can be monitored and any corrections to the sensor output made via the
analogue amplier or at the output using data stored in ROM. The most
valuable aspect of the microprocessor is the ability to produce a linear response
froma nonlinear sensor using a look-up table stored in ROM. This has enabled
the development of sensors with extremely low nonlinearity.
The output may be both digital and analogue. Analogue systems for
process plant applications feature live zeros so that a broken line is
not read by the controller as a value of zero. The standard working range
is 15 V or, more commonly for better interference immunity on long runs,
420 mA. Digital telemetry is becoming more established but, at the time
of writing, no standard protocol has yet been nally agreed for this Fieldbus.
The supervisory control input enables the device to be part of a hierarchical
control structure as well as for local operation. In addition to the provision of
alarm signals to indicate when set points have been exceeded, on-board
RAM can be used for data logging which can be poled later by the main con-
troller. When used on hazardous plant, where there may be risk of explosion,
a Zener diode safety-barrier device is often inserted into the electrical link.
These are commercially available devices. They are mounted at some
convenient place, usually at a common location for all such devices, and
are part of the plant bond; see Measurement Technology (1989). Their
function is to limit the ow of energy into the hazardous area, even under
fault conditions, to that below what is required for ignition. Care must
also be taken to ensure that plant instrumentation systems do not store
sucient energy to cause ignition. Plant safety is part of criminal law and
specialist advice is essential.
Many manufacturers now market transducer units with smart features.
An example is the Smart Deltapi KS Series of dierential pressure trans-
mitters by ABB Kent-Taylor. These are for use in the process industries.
The addition of microprocessing power to sensors will continue to accelerate
in the coming decade because of the competitive edge on sensor performance
that can be achieved. There have been government funded initiatives to
support SMEs (small-to-medium-sized enterprises having <250 employees)
wishing to go down this route with their more traditional products.
Microprocessor developments
In order to increase processing speed, the current trend in microprocessor
development has been to reduce the number of available program instruc-
tions that sequence the operations in the processor and to have the hardware
execute this reduced set very quickly and eciently. Such microprocessors
are described as having reduced-instruction-set or RISC architectures. A
RISC device will have typically fewer than forty instructions, whereas a more
traditional microprocessor might have several hundred. A further trend has
been to include more memory, ADC and DAC within the microprocessor
chip so that a small, stand-alone system can be easily fabricated requiring
minimal additional components. Such chips are referred to as microcontrol-
lers. Manufacturers provide development systems for their microcontrollers
to facilitate program design and to emulate the system performance. These
development systems run on a PC. Once the emulation is operating satis-
factorily the program instructions can be downloaded and burned into the
memory of a microcontroller chip. This is achieved using a small develop-
ment board unit that connects to the PC output port. The microcontroller
is then inserted into its own printed circuit board creating an independent,
task-specic, stand-alone system.
A number of manufacturers now market microcontrollers. Well-known
in the eld are Motorola Inc, who market several series, and Microchip
Technology Inc with their PIC family. Dedicated systems of this type have
a wide range of applications. For instrumentation use, as well as being
very suitable for inclusion in smart sensors, they can also be used to
control PC plug-in boards, data loggers and dedicated instrumentation
systems.
Despite the versatility and cost eectiveness of microcontrollers,
development work still continues in specialist microprocessors for digital
signal processing (dsp) applications. Texas Instruments Inc (TI), for
example, have a long pedigree with their TMS320 series. The instruction
sets of dsp microprocessors contain specic functions which are the back-
bone of dsp algorithms. With dsp microprocessors high performance
digital ltering, large-point, fast Fourier transforms (FFTs) and correlations
can be computed in real-time.
To conclude this section is a brief overview of the current trends in chip
design philosophy and technology. Purely digital logic chips have been
available for a decade that the user can customize directly. These are
termed programmable logic arrays (PLAs) and are used to create a relatively
simple device. The cheaply-produced blank chips contain of order 100 logic
elements or cells. The development environment operates on a small work-
station. The user creates some bespoke logic structure which is tested by
simulation and then burned down to hard-wire the blank. In this way a
small company can manufacture digital chips in-house. An example of an
application is that of a burglar alarm system chip. User-dened analogue
chips are, at the time of writing, beginning to be marketed. In a similar
way these have a number of analogue cells that can be interconnected by
burning down a design developed from the users workstation.
More complex structures using a mix of analogue and digital signal
technology can be built using ASICs (application-specic integrated
circuits). These are designed by the user using software supplied by the
Smart sensors 193
chip manufacturer. However, manufacture has to be carried out at the silicon
foundry. It is a much more expensive procedure and is only cost-eective
when large volumes of chips are to be produced. Generally, when producing
a traditional printed circuit board carrying analogue and digital signals it is
good practice to isolate the two areas using separate ground planes and to
route the connections so that high-frequency clock signals and other
digital pulses do not couple through into the analogue lines. This cross-
coupling can be a serious problem when designing mixed signal chips.
Also, the libraries provided by ASIC manufacturers will not contain highly
specialist microprocessors or operational ampliers. Radio frequency
analogue components are also unlikely to be available.
To summarize, currently the design of high-performance instrumen-
tation systems meaning high-speed, high-resolution, performing complex
signal processing with low noise and interference requires the use of discrete,
individually-specied chips.
5.6 Articial intelligence in instrumentation
There are many situations that require the processing and analysis of the
information from several sensors. This may be because the operator
cannot assess all the data provided at the desired speed. An example of
this could be that of the pilot of a ghter aircraft where some form of data
fusion is performed by an expert system. A dierent problem arises when
the desired analysis may involve an assessment of the data provided by a
number of sensors whose individual values may be acceptable within pre-
scribed interrelated tolerances. A good example of this is the assessment of
the quality of a food product where a number of parameters such as
colour, pH, smell and sugar content have to be within limits but in a
dependent way. For example, a higher pH may be acceptable if the sugar
content is higher. It is this latter type of application that has stimulated the
development of systems using articial intelligence based on fuzzy logic
and on articial neural networks.
The denition of the expression articial intelligence, rst coined by
John McCarthy in the 1950s, is like human intelligence somewhat vague.
Also, as in the study of human intelligence, with the growth in understanding
and the establishment of new developments, subsets of the eld have been
dened. The following sections briey describe some aspects of articial
intelligence appropriate to instrumentation applications.
5.6.1 Rule-based expert systems
An expert can be dened as a person who has extensive knowledge in some
eld or domain. This knowledge is in two forms, facts and rules. The rules
provide a structure as to how the conclusion is to be deduced from the facts.
This structure is incorporated in an inference engine. The inference engine in
its simplest format is a set of IFTHEN decision statements. The structure
and sequence of these statements denotes how the expert system reaches a
conclusion. The actual path taken through the structure in a given trial is
referred to as the inference chain. There are two types of inference chain,
forward and backward chaining. In backward chaining the inference structure
works backward from a given hypothesis and tries to verify its validity by
gathering appropriate information. An example of this procedure could be
the conrmation of the suspicions of a detective following the gathering of
appropriate evidence. Forward chaining begins with the information pre-
sented and attempts to draw a conclusion. Determining the cause of death
at a post-mortem from signs on the body and the results of tests performed
on specimens is an example of forward chaining. It is possible to have a
mixed system employing forward and backward chaining which is the
way much human reasoning is done. But for instrumentation applications
forward chaining is likely to be the more appropriate. An additional
feature required with forward-chaining is a strategy to deal with conict
resolution. A conict occurs when more than one rule res as the inference
engine works on the data. The resolution is achieved by using a strategy
such as prioritizing the rules, or not using a rule after it has scored a
hit, or using the rule that relates to the most recent addition to the data
memory.
The earliest commercial applications of AI came from the need to
incorporate the expertise of a renowned specialist or the combined skills of
a team into a knowledge base that could then be used either by a novice or
be available remotely or be processed at a much higher speed than is
humanly possible. It is easy to imagine examples of applications of such
expert systems.
.
A medical expert in the diagnosis and treatment options of some form of
cancer could have their lifetime of experience made available on-line,
world-wide.
.
The safe operation of a remote engineering installation such as a drilling
platform or space vehicle. The skills of the designers and technicians
could be stored on-board so that when certain sensors produced alarm
warnings the diagnosis and remedy of the situation could be obtained
even if communications to base were interrupted.
.
The fault-diagnosis of some complex device following its production or
for servicing. A good example of this is the growing use of computer
diagnostics for vehicle servicing. Long-range transportation companies
could log the state of their vehicles on the road for analysis on return to
the depot. In this way it should be possible to anticipate expensive break-
downs a long way from home.
Articial intelligence in instrumentation 195
.
The complexity of modern aircraft means that the ight crew may not be
able to react fast enough to the warnings of an attack or not have
enough skill to appreciate the signicance and appropriate remedy from
a set of alarm signals. The use of expert systems on commercial airliners
has enabled a reduction in the number of ight crew required.
Durkin (1994) contains a survey on the application elds of expert systems.
It is generally accepted that rule-based expert systems are the best
approach to producing a knowledge-based system, Negnevitsky (2002).
Expert systems are tailored to the application. But the general design, as
outlined above, is to have a large knowledge-base from which specic
conclusions can be inferred. This may involve prompting the user for more
data but will ultimately involve matching the input data with specic states
or conditions within the knowledge-base. A probability of match might
also be issued. In some applications an automatic response may be output.
For example, when the forward and ground radar of a high-speed, low-
ying, ghter aircraft detect a hill ahead. This is a control application
rather than instrumentation. Much work has been done and research still
continues apace on the applications of AI in control theory, particularly in
the eld of robotics.
Programming languages have been specically designed for expert
system use. The most well-known are LISP, PROLOG and OPS. LISP
was dened by John McCarthy and is the oldest. The cheaply available
computing power that became available in the 1980s with the PC and
other workstations has led a number of software houses to market packages
for expert system development. Consequently, the user does not need to
become a procient programmer in one of these languages. The development
environment is referred to as the shell. This is a software package containing
all the relevant utilities so that the user has only to input the facts and rules
and to design at the structural level.
5.6.2 Fuzzy logic
The description of a system, problem-solving and decision-taking is seldom
achieved using simple concepts of blackwhite but more of setting value
judgments to various shades of grey. For example, when describing the
temperature of a water tank expressions such as cold, cool, warm, hot,
very hot are used. This is in contrast to the simple machine approach of
dening a threshold temperature and describing the state of the water as
hot or cold depending on which side of the line it fell.
In an attempt to incorporate linguistic descriptors into a mathematical
representation the Polish logician and philosopher Jan Lukasiewicz in 1930
proposed the concept of multi-valued logic. The logic values could take
any real number between 0 and 1. The actual number assigned to a statement
was the possibility of it being true or false. This became known as possibility
theory. Max Black (1937) suggested that the actual numbers assigned to
states in a discrete, multi-valued logic system could be derived from the
opinions of a group of people. He was the rst to dene a simple fuzzy set.
In a paper Lot Zadeh (1962), Professor in Electrical Engineering at
Berkeley, stated that a new kind of mathematics was needed for these
fuzzy quantities. Further to this he argued that traditional probability
distributions were not appropriate when attempting to incorporate natural
language into system theory. Three years later Zadeh (1965) published a
paper dening a formal system of mathematical logic for fuzzy set theory.
Initially he believed that fuzzy logic would nd its home in the humanities
but in this latter paper he suggested that it would play an important role
in control engineering.
A set is a classication grouping. In natural language classication is
made according to whether something belongs to a particular set. A cow
belongs to the animal set. A Rolls-Royce belongs to the car set. In traditional
set theory membership would be in or out, TRUE or FALSE. Membership
of the set would be assigned binary logic 1. Non-membership would be
assigned logic 0. Sets dened in this stark way are sometimes described as
crisp. The dynamic range of the variable in question, such as the temperature
of a tank, is described as the universe of discourse. For a particular universe
of discourse there could be a family of sets. In a crisp set having a universe
of discourse X where a particular value of x may or may not belong, the
notation used is
if x belongs to X then, x 2 X
if x does not belong to X then, x , 2 X.
Zadehs concept was to have graded membership. A fuzzy set called A has a
membership function j
A
x which characterizes the degree of membership.
The dening statement for the membership function is
j
A
x : X ! 0. 1. 5.37
The membership value of a particular point x is in the range
0 j
A
x 1. 5.38
Figure 5.15 relates to the temperature of the tank described earlier. A crisp
set and a fuzzy set are shown for the descriptor hot.
For the tank example given above there is a family of fuzzy sets: cold,
cool, warm, hot and very hot. It is possible to have a fuzzy set contained
within another set on the same universe of discourse. These are sometimes
referred to as fuzzy subsets. For example very hot might contain the subset
very, very hot. A crisp subset must be entirely contained within the set to
which it belongs. However, fuzzy sets can spill over. When dealing with
subsets instead of describing the dynamic range as the universe of discourse
the notion of a reference superset is sometimes used. All values x
i
in the
reference superset must have degree of membership 1. In other words, a
reference superset is a crisp set where j
X
x
i
1.
To create a fuzzy set it is necessary to dene the membership function j.
If the universe of discourse is discrete then for each point in the range a value
is needed for its degree of membership to that set. An expert system can be
created by asking the expert to estimate the degree of membership of
values belonging to some universe of discourse to a fuzzy set. These
numbers can then be entered into a computer program as a vector. For
example, with the tank of gure 5.15 if the discrete temperature values are
X 10. 20. 30. 40. 50. 60. 70 8 C
then the fuzzy set hot might be estimated as having a membership
hot 0.15 0.17 0.19 0.5 0.9 0.93 0.95.
Commonly, fuzzy set vectors are written so that a particular membership
function value has with it that universe of discourse value to which it
belongs. The symbol / is used to simply separate the membership value
from the variable value. So, in the example given,
hot 0.15,10 0.17,20 0.19,30 0.5,40 0.9,50 0.93,60 0.95,70.
Membership functions can also be input into a fuzzy logic program as a
continuous mathematical function. Commonly used are the sigmoid,
Gaussian and pi functions. For a mid-range fuzzy set a triangle shape is
often used (gure 5.16). Commercial software for developing fuzzy logic
systems usually contains these functions as standard options. The user
simply enters some parameters to dene the exact shape of the function type.
It is possible to further shape a given fuzzy set to incorporate natural
language. So, descriptors such as very and extremely reduce the weighting
of the mid-range of the fuzzy set, whereas expressions such as moderately
and fairly enhance the mid-range. These fuzzy set modiers are known as
hedges. Figure 5.17 illustrates this concept on the fuzzy set cold.
Figure 5.15. Crisp and fuzzy sets.
The dashed line is obtained by squaring the function that generated the
sloping line. The dashed line would be described as very cold. The dotted line
is obtained by square-rooting the function that generated the sloping line.
This might be described as moderately cold. In general, changing the mem-
bership function to j
n
where n is a positive value will concentrate n 1
or dilate n < 1 the membership values of the fuzzy set by whatever
degree is required. Asymmetrical functions can also be easily created and
used such as a mid-range triangle but with one side hedged into a curve.
To generate the output of a fuzzy logic system logic operations are per-
formed on the data. Georg Cantor developed classical set theory towards the
end of the 19th century. In the fuzzy logic sense this is crisp set theory. Three
classical set operations are complement, intersection and union. Their per-
formance on fuzzy sets is described below. Whenever an operation is per-
formed a new set is created.
Complement
This is the logic inverse or NOT operation, so performing the complement
operation on the fuzzy set hot would generate a fuzzy set NOT hot. The
Figure 5.16. A triangular fuzzy set.
Figure 5.17. Hedging of a fuzzy set.
mathematical statement is
j
NOThot
1 j
hot
. 5.39
Intersection
In crisp set logic this is simply an AND operation. With fuzzy sets the
smallest values within the overlap have to be extracted. It is sometimes
called a min(imum) operation. In the example of gure 5.18 the dark triangle
is the intersection, which might be called tepid. The notation used to
represent this operation is \. So in the example given the new set is
j
tepid
j
cold
\ j
warm
. 5.40
Union
In this operation values are taken from one or both fuzzy sets. It is the logic
INCLUSIVE OR operation. The largest value is used and so it is sometimes
referred to as a max(imum) operation. Again, a new set is created. In the
example of gure 5.19 it might be called not tepid. The symbol used is [. So
j
not tepid
j
cold
[ j
warm
. 5.41
Figure 5.18. The intersection operation.
Figure 5.19. The union operation.
Finally, having described fuzzy sets and logic operations, the overall
procedure is examined. The task of the fuzzy system is to combine the data
input and produce a crisp output value. In fuzzy set theory the process of
interpreting various inputs into output(s) is termed inference. Mamdani
and Assilian (1975) published a paper on the use of a fuzzy system for the
control of a boiler and steam engine. The inference technique employed is
now the most frequently used and is referred to as the Mamdani method.
Figure 5.20 shows the strategic stages.
(1) Fuzzication
The rst stage is to take the specic input data values and to determine their
membership value within each appropriate fuzzy set. As previously
explained, an expert on the tank system has already estimated the degree
of membership values for the sets. For example, referring to gure 5.21,
the input temperature data value of 27 8C has a membership of 0.3 in the
fuzzy set cold, 0.0 in the fuzzy set warm and 0.05 in the fuzzy set hot. (In
fuzzy logic terminology numerical input data are said to be crisp.)
(2) Rule evaluation
The rules are the equivalent of IFTHEN statements in binary logic. The
logic operations described above are how the rules are implemented. Fuzzy
rules were developed in a further paper by Zadeh (1973). To illustrate the
Figure 5.20. The Mamdani inference method.
process, in the water tank example the pressure in the tank could be included
in the problem as a second data input. A new universe of discourse is now
introduced which could have its own fuzzy sets of low pressure, normal
pressure, high pressure. This is illustrated in gure 5.22. The water tank
expert who has a great deal of practical experience of the interrelation of
the temperatures and pressures and their control states that
Rule 1
If the temperature is COLD and the pressure is NORMAL then the heater is
ON.
Rule 2
If the temperature is HOT or the pressure is HIGH then the cooler is ON.
These two rules (more of course are possible) each have a corresponding set
operation and, or and each produces a new fuzzy set. In the example given the
inputs are that the temperature is 27 8C and the pressure is 7 atmos (gauge).
Figure 5.21. The tank temperature fuzzy sets.
Figure 5.22. The tank pressure fuzzy sets.
(3) Aggregation of the rule outputs
In gure 5.23 the grey triangle is the result of rule 1. Here the apex of the
triangle (degree of membership 0.30) is determined by the fuzzy set cold
for the temperature 27 8C. This is because the pressure of 7 atm in the
fuzzy set normal pressure has a degree of membership of about 0.75. The
and operation returns 0.3 as the common minimum value.
The black cusp shape below the curve is limited by the hot fuzzy set.
Here 27 8C has a membership of 0.05. Rule 2 is an or operation but the
pressure of 7 atmos has a membership of zero in the fuzzy set high pressure
and so has no eect on this result. The value 0.05 is the maximum degree of
membership from either of the two sets.
(4) Defuzzication
Once the logic operations have been performed and the results aggregated
the system needs to deliver a single, crisp output. A useful feature of gure
5.23 is that it gives a visual impression of the weightings of each rule
output in the aggregation. The most common way that a crisp output is
generated is from the centroid, shown as in the gure. The centroid is
the centre of mass for some three-dimensional object. The centre of mass
is that point around which the sum of all moments is zero. To nd the x
coordinate for this point x
cm
in an object having discrete elements of mass
m
i
equation (5.42) is used:
x
cm

P
i
m
i
x
i
M
. 5.42
Translating this into values of gure 5.23 gives
" vv
P
b
i a
j
i
v
i
P
b
i a
j
i
5.43
Figure 5.23. Aggregation of the two rule outputs for 27 8C at a pressure of 7 atm.
where " vv is the crisp output control signal and i is a particular value in the
universe of discourse v which extends over the range a to b.
Zadehs prediction about an important application of fuzzy logic being
in control engineering has in fact come true. Figure 5.24 shows a block
diagram for a simple control system.
Because the skills and knowledge of an expert operator can be
incorporated into the controller it has been found that nonlinearities, distur-
bances, drift and sensor degradation can be successfully accounted for.
Because of the nature of fuzzy logic the skills and knowledge of the expert
can be input in linguistic terms. A formal mathematical model is not
needed. Another advantage of fuzzy logic control is that modication to
the skills or rule-base can be made very easily while in use.
McNeill and Freiberger (1993) give a history of early industrial applica-
tions of fuzzy logic. The rst commercial application of fuzzy control was for
a Smidth-type cement kiln; see Holmblad and stergaard (1982). Cement is a
mixture containing the complexes calcium aluminate, calcium silicates and
calcium alumino-ferrite. Clay, limestone, sand and iron ore are input as
slurry. The kiln is a rotating drum approximately 5 m in diameter and
140 m long revolving at about 1 rpm. It operates at a temperature around
1300 8C and the through-time is in the region of 3 h. The variables to be con-
trolled are the fuel supply, slurry feed and drum rotation rate. Temperature
control is critical to forming the correct quality of granules, called clinker.
Previous attempts at computer control had failed. It was the discovery of
27 IFTHEN rules contained in a training manual for the kiln operators
that led the researchers to develop a fuzzy verbal rule-base. A typical rule
might be
if the oxygen percentage is rather high and the free-lime and kiln
drive-torque are normal decrease the ow of gas and slightly
reduce the fuel rate.
A successful trial was carried out in 1978. This was the rst industrial use of a
closed-loop, expert-system controller. A signicant proportion of cement
kilns world-wide now use fuzzy logic controllers.
Figure 5.24. A simple fuzzy logic-controlled process.
In principle any number of inputs and outputs can be incorporated into
a controller. However, with systems having many inputoutput variables the
rule structure can become very complex. For such systems it is often easier to
develop a number of fuzzy logic controllers to handle sub-systems in the
process.
Other areas where fuzzy logic has been applied are
.
information retrieval,
.
navigation systems,
.
timetabling for transportation systems,
.
a number of elds in robotics,
.
feature denition in imaging systems,
.
nancial planning,
.
meteorological analysis,
.
predictive control for the automatic operation of a train,
.
domestic electrical goods.
5.6.3 Articial neural networks
In section 5.6.1 the rule-based expert system was described. This is based on
formal logic using IFTHEN statements. The fuzzy logic of section 5.6.2 can
be seen as an extension of this using logic which is less binary by nature.
However, articial neural networks do not use logic. They require no prior
knowledge or skills from an expert to be programmed into the sotware.
They are systems that attempt to mimic the learning processes of the brain
and, as such, learn the relationships between data input and output. It
could be argued therefore that articial neural networks show true
intelligence. This is in contrast to the machine-intelligence of the expert
systems where intelligence is installed by the designers.
The fundamental building block is a mathematical model of the bio-
logical neurone. These articial neurones are grouped into some structure
that is then trained using sets of data. The software learns the relationships
needed to t the data and can then be used on unseen data to derive a
conclusion. Once trained a hard-wired version may be built if desired. The
following sections describe the essential features of how the articial
neural network is constructed, trained and used.
The biological neurone
The neurone is the fundamental cellular unit of the nervous system and the
building block of the brain (gure 5.25). Neurones occur in a number of
forms depending on their function and location in the body but have a
similar overall structure.
Each neurone is a processing unit that receives signals from many other
neurones through its input structure of dendrites. If the combined input
stimuli reach the threshold level the cell is triggered and a signal passes on
to the axon and through to the output structure leading on to many more
neurones. Its nearest logic equivalent is the AND gate but the system is
not simple binary.
The axon has a high-resistance outer membrane and contains a low-
resistance uid called axoplasm. It is the modication of cation concentra-
tions across the membrane that results in a voltage wave passing along the
neurone.
Table 5.3 shows an imbalance or ion dierential across the membrane
that is maintained against diusion and Coulomb forces by active transport.
Carrier molecules restore any disturbance to this imbalance using energy
supplied from the respiratory system. The net dierence in ion concen-
trations between inside and out, of 144.4116 mmol per litre, produces a
potential dierence of between 60 and 90 mV. The membrane is said to be
polarized with this resting potential. If the neurone receives a suciently
high stimulus from other input neurones or external stimuli such as heat,
physical stress, electric current or chemical intervention, the cell wall can
become permeable to the ions. An external stimulus to the neurones in the
brain is not the normal method of creating activity! Instead, the input is
derived from neurones in the rest of the body or from other neurones in
the massive neural tree-structure within the brain. Once the potential
dierence drops below the action potential in some region (the actual drop
Figure 5.25. A biological neurone.
Table 5.3. Ion concentrations in and around the neurone.
Concentration of ions Sodium ions (Na
)
(mmol per litre)
Potassium ions (K
)
(mmol per litre)
Outside cells (tissue) 140.0 4.4
Inside cells 16.0 100.0
in potential dierence may be as low as zero or even go negative) this triggers
adjacent parts of the membrane to become permeable and so cause a voltage
pulse-wave to travel down the axon. After the impulse has passed a recovery,
or refractory, period follows of about 1 ms during which the ion molecular
pumps return the cation concentration levels to the resting state of table 5.3.
It was Helmholtz in 1850 who rst measured transmission speeds along
nerve bres. Initially it was thought that these electrical pulses travelled at the
same speed as electrical signals along wires. But instead of the expected
2 10
8
ms
1
he measured speeds of only 25 ms
1
. Some neurones in the
body, particularly those involving the reactions, have an insulated sheath
around the axon called myelin. The sheath has gaps at various positions
along its length to aid re-polarization. In this type of neurone transmission
speeds may be up to 100 ms
1
.
The transmission between neurones is through the synapses. Bio-
chemical agents called neuro-transmitters are released by the axon of the
transmitting neurone and accepted by the dendrites of the receiving
neurones. If the incoming signal to a receiving neurone is strong enough
this neurone will trigger and subsequently send pulses to further neurones.
Neurones can have between 10
3
and 10
4
synapses.
The human brain contains of order 10
11
neurones. Recent investigations
have revealed that the doubling in size of the infant brain is due to growing
interconnectivity of neurones rather than the growth of more neurones.
Generally throughout life whenever some new information is learned by
the brain new links between neurones are formed. As part of the reinforce-
ment the synaptic eciency is modied during the learning process. If the
learning process is repeated several times with a suitable intervening period
the modication to the synapse is long-term. This phenomenon is responsible
for the dierentiation between short-term and long-term memory.
History of articial neural networks (ANN)
The beginning of the development of articial neural networks can be traced
back to work done in the 1940s; see Neuralware (1989). At this time research-
ers were attempting to understand the operation of the learning process in the
biological neurone. Donald Hebb suggested that when the synaptic input and
output of a neurone were both active the strength of the incoming connection
was reinforced. McCulloch and Pitts (1943) created a mathematical model of
a brain cell where the probability of triggering depended on the level of the
input signal and the threshold at the synapse. In 1951 Minsky and
Edmonds, inspired by the ideas of McCulloch and Pitts, constructed a
learning machine out of thermionic valves, motors, clutches and a gyropilot.
This device was able to learn suciently well to follow a maze. Following this
Frank Rosenblatt developed the concept of the perceptron. This was an
attempt to model the pattern recognition abilities of the eye. The outputs
from a matrix of sensors were hard wired into an array so as to recognize
certain features. In the early 1960s Widrow devised the adaline (adaptive
linear neuron), a threshold logic device and he and Ho developed a learning
rule for correcting the connecting strengths or weightings between the input
values and the summing unit. A multi-layer adaline, the madaline, was also
developed. Minsky and Papert (1969) published a book on the perceptron
and concluded that it would not be capable of learning by itself. Interest in
articial neural networks then waned for some time.
In 1978, through an association with MIT, John Hopeld of Bell
Laboratories became interested in the computational abilities of biological
systems. Hopeld devised a network from a study of the neuro-physiology
of the garden slug. As a physicist he realized that stability in the feedback
structure could be understood in terms of a minimum energy state. The
learning principle in the Hopeld network is that connections between the
articial neurones are reinforced if the output of a processing element is
the same as its input and, conversely, weakened if it diers. An application
of Hopeld networks that has been extensively investigated is in optical
pattern recognition systems.
The credit assignment problem
As mentioned in the introduction, an articial neural network system, unlike
an expert system, is able to learn the required expertise itself by training on
known data. Having been trained, it can then be used on unseen data to
generate novel output.
In the training process the strengthening or weakening of the inter-
connections between neurones uses a procedure involving credit assignment.
Devising a method for adjusting the synaptic eciencies of all the neurones
in an articial neural network during its learning stage was a major challenge.
Some among the early workers thought it an impossible task. The so-called
credit assignment problem is the problem of knowing how to determine the
signicance a particular interconnecting weight has on the system global
error. Credit assignment is not a problem when dealing with a single array
of neurones since input and output values and appropriate weightings are
easily handled. The problem arises when there are hundreds or thousands
of neurones all interconnected in a large network. Only those few inter-
connections at the input and output interfaces can be readily assessed for
their contribution to the global error. The major challenge is how to
change the weighting of a particular interconnection hidden deep within
the network having only the global error to work from.
Of the various techniques developed for solving the credit assignment
problem the back-propagation method has become the most popular. The
development has been attributed to Rumelhart and McClelland (1986) and
also to Parker (1985). The popularity of the technique is due to its success
in functioning well with nonlinearly separable systems. Its success did a lot to
revive interest in articial neural networks and, with the rapid developments
of cheap computing power, this is still continuing. The back-propagation
process is described below. Elaborations are described in Chauvin and
Rumelhart (1995).
Mathematical structure for the back-propagation technique
The multiple sets of equations needed to describe the structure requires some
rather cumbersome notation that is set out below. The articial neural
network structure shown in gure 5.26 gives the labelling for the articial
neurones and their inter-connecting weights w.
All the articial neurones are inter-linked from the input upwards
although for clarity these are not all shown. A route through the hidden
layers is shown as ith to jth to kth neurones with their interconnecting
weights w
ij
and w
jk
. Dierent numbers of neurones are shown in dierent
layers to indicate that no set pattern is required. Generally, it has been
found that three hidden layers are the useful upper limit. Adding more
layers does not tend to improve performance and increases the learning time.
The input to the jth neurone in level s is I
s
j
and is the sum of the
variously weighted outputs of all the neurones in the previous layer,
I
s
j

X
i
w
s
ij
X
s 1
i
. 5.44
The output of the jth neurone in level s is X
s
j
and is some function f of its
input I
s
j
,
X
s
j
f I
s
j
. 5.45
Various functions have been tried for f such as sine and tanh but the one
which has become the most popular closely resembles the biological
neurone and is called the sigmoid function,
f I
s
j

1
1 expI
s
j

. 5.46
Figure 5.26. A simple articial neural network structure.
A very useful feature of this function is the ecient way its derivative can be
calculated,
0f
0I
s
j
f 1 f . 5.47
The inputoutput terms for the articial neurone are shown in gure 5.27.
Initially, all the weights are set at random. For the learning process to
proceed a relationship is needed that relates the value of a particular
weight w to the global error, E. Then the weight can be adjusted by some
amount w. To achieve this the so-called gradient descent rule was dened,
w
s
ij
lcoef
0E
0w
s
ij
5.48
where lcoef is a learning coecient. Equation (5.48) is the essence of the
learning process in that it is the technique by which the global error is appor-
tioned to all of the connecting weights. It is a way of addressing the credit
assignment problem discussed above. The partial dierential term is a
measure of how sensitive the global error is to a particular weight. The
more inuential a weight is the more it will be modied during the training
process. The learning coecient is a value chosen, somewhat arbitrarily,
between 01. If it is too large the system may go into oscillation during
training. The changes to the weights are too large and the system cannot
converge to the minimum value of E. If it is too small the learning times
will be longer and there is a greater probability that the system will settle
at some minima that is not the lowest possible value of E.
From equation (5.48) it can be seen that it is a requirement that E should
be a function that is dierentiable with respect to each of the interconnecting
Figure 5.27. An articial neurone.
weights throughout the network. A squared function is used,
E
1
2
X
output
neurones
X
d
X
o
a

2
5.49
following the concept of a least squares analysis where X
d
and X
o
a
are the
presented desired output and the calculated actual output respectively for
each neurone in the output layer. (The superscript [o] signies the output
layer.) The factor
1
2
is simply for convenience of normalization and
becomes 1 after dierentiation. The successful development of the back-
propagation technique was in the concept of rst calculating the error in
the output layer and then assigning this appropriately to all the hidden
neurones throughout the network. The proportion of the error ascribed to
an individual neurone is determined using the local error of the neurone
calculated from the partial derivative of E. The local error c of the jth
neurone in level s is determined using
c
s
j

0E
0I
s
j
. 5.50
The use of the partial derivative means that c is the sensitivity of the global
error E to the input I of that particular neurone. The usefulness of this
denition for the local error can be seen by expanding the second term in
equation (5.48) using the chain rule
0E
0w
s
ij
0E
0I
s
j
0I
s
j
0w
s
ij
5.51
c
s
j
0I
s
j
0w
s
ij
. 5.52
Substituting for I
s
j
using equation (5.44),
0E
0w
s
ij
c
s
j
0
P
i
w
s
ij
X
s 1
i

0w
s
ij
. 5.53
The partial dierentiation reduces this to
0E
0w
s
ij
c
s
j
X
s 1
i
. 5.54
This means that the equation (5.48) can now be expressed as
w
s
ij
lcoef c
s
j
X
s 1
i
. 5.55
To implement equation (5.55) as an algorithm now only requires an
expression for c
s
j
.
For the output layer the determination of the c values for each neurone
is straightforward. Rewriting equation (5.50),
c
o
j

0E
0I
o
j
. 5.56
Using the chain rule this can be expressed as
c
o
j

0E
0X
o
j
0X
o
j
0I
o
j
. 5.57
But the actual error X
d
X
o
a
for each output neurone is known. Therefore,
the rst term in the right-hand side of equation (5.57) can be determined as
0E
0X
o
j
0
1
2
P
X
d
X
o
a

2
0X
o
j
5.58
X
jd
X
o
j
5.59
where X
jd
is the presented desired output for the jth output neurone.
Also, it can be seen from equation (5.45) that the second term in
equation (5.57) is the partial derivative of f:
0f
0I
o
j
0X
o
j
0I
o
j
. 5.60
By combining equations (5.59) and (5.60), c
o
j
in equation (5.57) can be
expressed as
c
o
j
X
jd
X
o
j

0f
0I
o
j
. 5.61
The partial derivative 0f ,0I
o
j
can be replaced by numerical estimates using
equations (5.46) and (5.47).
The determination of the local error values for neurones in the hidden
layers is more involved. A general expression is required for c
s
j
that involves
the s 1 layer. Once this is obtained errors occurring between the system
outputs and the training values presented can be determined as described
above. But then weightings can be adjusted for the interconnections between
neurones in the output and in the penultimate layer using c
s 1
j
. This procedure
can be repeated propagating backwards using c
s
j
, c
s 1
j
etc., towards the input
layer of neurones. The procedure is termed back-propagation because of this
working towards the input layer of neurones from the output layer.
Adapting the notation of equation (5.57) for some general level s gives
c
s
j

0E
0X
s
j
0X
s
j
0I
s
j
. 5.62
Again, using the chain rule, the rst term in equation (5.62) can be expanded
to give
0E
0X
s
j
X
k
0E
0I
s 1
k
0I
s 1
k
0X
s
j
. 5.63
The summation over k appears because all the neurones in level s 1, and
consequently E also, are aected by changes in X
s
j
. Using the denition
for c equation (5.63) can be rewritten as
0E
0X
s
j
X
k
c
s 1
k
0I
s 1
k
0X
s
j
5.63
X
k
c
s 1
k
0
P
j
w
s 1
jk
X
s
j
0X
s
j
. 5.64
Because the second term in equation (5.64) is a partial derivative in X
s
j
,
unless i j the dierential is zero, so equation (5.64) becomes
0E
0X
s
j
X
k
c
s 1
k
w
s 1
jk
. 5.65
Consequently, equation (5.62) becomes
c
s
j

0f
0I
s
j
X
k
c
s 1
k
w
s 1
jk
. 5.66
As before, the partial derivative 0f ,0I
s
j
can be replaced by numerical
estimates using equations (5.46) and (5.47).
All the components of equation (5.65) can now be determined so that
the adjustments to the weights can be made every time an inputoutput
pair of training vectors is presented to the network. In practice, les con-
taining training data pairs are constructed containing hundreds of values
of these vectors. These are fed into the system randomly during the learning
process. This random selection enables each training value to be used several
times but with reduced risk of having the system settle into less than the
minimum global error. A few hundred data values could typically be used
to provide tens of thousands of training cycles. Depending on the network
size and the computer used, training may take several hours. The training
algorithm is summarized in the ow chart shown in gure 5.28.
Once training is complete the weights are left in place and the system can
be used on new, unseen data. Errors in performance can be low, often down
to a few percent. This is clearly a poor performance when deriving values
for a system where a known, causal relationship exists or can be easily
determined. However, the advantage of articial neural networks is that
the causal relationship need not be known. The network is trained to
recognize inputoutput relationships and then provide an estimate of
an output for an unseen input. The unknown relationship may even be
nonlinear.
It is possible to build the trained system into a hard-wired structure
using op-amps and resistors. Articial neural network integrated circuits
have also recently become available. Because such systems use analogue tech-
niques, they can operate extremely quickly. They also have the advantage of
being fault tolerant. In a remote system such as a satellite, where radiation
damage is likely, the hard-wired system will degrade slowly as individual neu-
rones malfunction. A microprocessor-based system in similar circumstances
would fail catastrophically if the CPU or RAM were damaged.
Figure 5.28. Flow chart for the back-propagation training technique of a neural network
structure.
Development work is continuing in training strategies and in the design
of new network structures containing sub-loops and internal feedback. Some
software-based systems, such as for example in control applications, have
their weight values updated periodically as the plant characteristics age.
Combinations of articial intelligence techniques are also being investigated
such as with fuzzy logic and expert systems.
Some elds of application are listed below.
.
text to speech conversion,
.
image compression,
.
handwriting recognition,
.
object classication for image recognition and industrial inspection,
.
predictions of stock market and weather,
.
system modeling,
.
noise ltering,
.
parameter adjustment to improve tolerances in robotic systems,
.
functional synthesis.
References
Black M 1937 Vagueness: An exercise in logical analysis Phil. Sci. 4 427455
Brignell J E and Dorey A P 1983 Sensors for microprocessor-based applications J. Phys.
E 16(10) 952958
Chauvin Y and Rumelhart D E eds 1995 Backpropagation: Theory, Architectures, and
Applications (Lawrence Erlbaum Associates)
Durkin J 1994 Expert Systems Design and Development (Englewood Clis, NJ: Prentice
Hall)
Holmblad L P and stergaard J J 1982 Control of a Cement Kiln by Fuzzy Logic, Fuzzy
Information and Decision Processes (Amsterdam, North-Holland) pp 398399
Horowitz P and Hill W 1989 The Art of Electronics 2nd edition (Cambridge University
Press)
Mamdani E H and Assilian S 1975 An experiment in linguistic synthesis with a fuzzy logic
controller Int. J. ManMachine Studies 7(1) 113
McCulloch W S and Pitts W H 1943 A logical calculus of the ideas immanent in neural
nets Bull. Math. Biophys. 5 115133
McNeill D and Freiberger P 1993 Fuzzy Logic (New York: Simon and Schuster)
Measurement Technology Ltd 1989, Application Note AN9003, A Users Guide to Intrinsic
Safety, Luton, Bedfordshire, England
Minsky M L and Papert S S 1969 Perceptrons (Cambridge, MA: MIT Press)
Negnevitsky M 2002 Articial Intelligence: a Guide to Intelligent Systems (Addison Wesley)
Neuralware Inc. 1989 NeuralWorks Professional II Users Guide
Parker D B 1985 Learning Logic Report TR-47 (Cambridge, MA: MIT)
Rumelhart D E and McClelland J L (eds) 1986 Parallel Distributed Processing: Explora-
tions in the Microstructure of Cognition, 1, Foundations (MIT Press)
Zadeh L 1962 From circuit theory to system theory Proc. Radio Engineers 50
References 215
Zadeh L 1965 Fuzzy sets Information and Control 8(3) 338353
Zadeh L 1973 Outline of a new approach to the analysis of complex systems and decision
processes IEEE Trans. Systems, Man and Cybernetics SMC-3(1) 2844
Exercises
1. The analogue output from a certain sensor is to be converted to digital
format. If a resolution of 0.5% is required how many bit ADC is
needed? If the signal contains frequencies up to 1.5 kHz what should be
the sampling rate? Explain your choice.
2. Choose an ADC for the following applications and explain your choice:
(i) digital audio recording studio,
(ii) multimeter,
(iii) vibration sensor on an aircraft engine.
3. A manufacturer of portable audio mixing desks decides to replace the
heavy, multi-core cable that runs from the desk to the stage by an
optical bre. Calculate the serial bit-rate down the bre given that there
are 32 signal channels, each with 18-bit resolution and that the
maximum audio frequency is 20 kHz.
4. A certain monochrome, video-camera CCD sensor has an array of
512 512 pixels. If a brightness resolution of approximately 2% is
required what bit size must the analogue-to-digital converter be? Deter-
mine the output bit-rate if the camera generates 25 frames per second.
Suggest a suitable type of ADC.
5. A smart pressure gauge is to be designed based on a stainless-steel dia-
phragm and strain gauges; see equations (3.2) and (3.3) in section 3.2.2.
Draw a block diagram of the type gure 5.14 stating what features you
would incorporate in the device to achieve an overall error of <0.1%.
(The phase-sensitive detector of section 6.4.2 might be a useful inclusion.)
6. An automated hot-drink vending machine is to be controlled by a rule-
based expert system. A hot water tank is maintained at 70 8C by an
electrical heater and conventional thermostat. The tank is connected to
the water supply and a oat valve maintains the water level in the tank.
TEA, COFFEE, DRINKING CHOCOLATE and TOMATO SOUP
are normally available. SUGAR, EXTRA SUGAR and WHITENER
are options. Construct a ow chart of the logic sequences necessary to
incorporate the following features:
.
If all systems are operational but no action is under way an indicator
READY comes on.
.
The money-accept slot does not retain money when the system is shut
down (SYSTEM FAULT).
.
Awater level sensor shuts othe heater and closes down the operation if
the level is too low and an indicator light SYSTEM FAULT comes on
to inform the customers.
.
If the water temperature is low an indicator light asks the customers to
PLEASE WAIT. No other functions operate.
.
If the temperature rises to 100 8C the power to the immersion heater is
isolated and the system shuts down. The indicator light SYSTEM
FAULT comes on.
.
If there are no cups left the system shuts down. The indicator light
SYSTEM FAULT comes on.
.
The ingredient hoppers contain sensors. If any hopper is empty an
appropriate light informs the customers. If a drink is not available
the selection button for that drink becomes inoperative and the
money-accept slot does not retain money.
.
If the coin sensor detects that sucient coins have been entered they are
stored in a holding box until the drink has been dispensed. Otherwise
the coins are returned.
.
If the money available in the machine to dispense change has all been
used an indicator NO CHANGE AVAILABLE comes on. If excess
money is input it is returned.
.
If money is being stored in the holding box the selection procedure
becomes operative.
.
The customer can select one drink only. The selector buttons become
inoperative after a selection has been made.
.
Sugar and whitener buttons become operative after drink selection
unless soup is chosen.
.
If the cup sensor has detected that a cup is in place the dispensing begins.
If not the indicator SYSTEM FAULT comes on and money is returned
and then the system shuts down.
.
If the system has successfully dispensed then any change due is released.
.
The READY indicator comes back on.
7. An executive-car manufacturer decides to develop an electronically-con-
trolled, automatic, 4-speed gearbox. This is expected to simplify the com-
plexities of the gearbox hydraulics and provide greater exibility when
driven. Fuzzy logic is proposed as the best method of producing a control-
ler. Your task is to develop some fuzzy sets for
.
each gear versus road speed,
.
accelerator pedal position,
.
level of braking applied,
.
engine rpm.
The following rules are to be incorporated for the choice of gear to be
output to the selector mechanism:
(a) If the vehicle is stationary then select 1st gear.
Exercises 217
(b) If the road speed is between about 520 mph then select 2nd gear.
(c) If the road speed is between about 1540 mph then select 3rd gear.
(d) If the road speed is above about 35 mph then select 4th gear.
(e) If the accelerator pedal has been pressed beyond about midway then
changes in (a)(d) above should occur toward the higher road speed
ranges given.
(f ) If the accelerator pedal is depressed fully (kick down) then drop to the
next lower gear, excluding 1st gear, and only change up when the
engine speed exceeds about 4000 rpm.
(g) If the accelerator pedal is returned from kick down to midway then
move to the next higher gear.
(h) If the accelerator pedal is released and the brakes are applied gently
but the vehicle does not decelerate then, after several seconds, move
to the next lower gear, excluding 1st gear.
(i) If the ABS system has become active change to the next higher gear.
8. (a) By reference to the biological neurone explain the operation of the
articial neurone.
(b) Give an expression for the output of an articial neurone in terms of
its inputs.
(c) What is meant by the expression global error? Give an equation.
(d) The so-called gradient descent rule can be used to alter the weights.
Explain this process.
9. The combustion chamber and exhaust manifold struts on a certain air-
craft jet engine reach between 85 and 150 8C during ight. However, it
is known that the temperature pattern on these ten struts can indicate
an engine fault. It is decided to use an articial neural network to warn
of an engine malfunction.
(a) Assuming an analogue signal of 05 V is output from each of the ten
temperature measurement systems, suggest an articial neural
network structure that could be trained to give three possible
outputs NORMAL, WARNING and MALFUNCTION.
(b) With the aid of a ow chart describe the training process that would
need to be employed in order for the neural network to function
correctly.
Chapter 6
Data signal characterization and recovery
6.1 Introduction
When data signals from a sensor or from an instrumentation system are
weak, either because the variable to be measured is tiny, the sensor output
is poor, or the signal is attenuated in the telemetry link, recovery of the
information may be dicult. This is because the signal of interest is of a
similar magnitude to any interference or noise present. If the measurand is
a steady value then the recovery process is relatively simple and the
method adopted depends on whether it is noise or interference that is cor-
rupting the data. However, if the measurand is time varying its recovery
may be less straightforward. The most dicult situation is when the data
itself is random and the signal is weak.
Both the data signal and any unwanted signal can be classied as either
random or deterministic. The future value of a random signal cannot be
predicted. There is no way of knowing or calculating the forthcoming time
values. The only way the signal can be characterized is either by frequency
analysis or by using statistical parameters. The future values of a determinis-
tic signal can, in principle, be calculated. In practice the time-dependent
formula may not be fully understood or even known, but the concept of
the process being orderly and that the signal is following some pattern
classies it as deterministic. Table 6.1 gives some simple illustrations.
The presence of unwanted periodic signals is termed interference and
commonly comes from power supply lines. The origins of noise are more
varied and are described in section 6.4.
For a numerical description of the levels of both noise and interference
the expression signal-to-noise ratio is used. The power ratio is used in decibel
format,
10 log
10
_
data signal power
unwanted signal power
_
10 log
10
_
V
2
data
,R
V
2
noise
,R
_
dB 6.1
where R is the input resistance to the signal conditioning circuit. This is
more commonly expressed in terms of voltages since signal powers are not
normally measured. Assuming that R remains constant equation (6.1)
becomes
20 log
10
_
V
data
V
noise
_
dB. 6.2
6.2 Classication of data signals
Data signals can be sub-divided into two classications: deterministic and
random. Such data come from real processes which may themselves be
deterministic or random. Random processes are sometimes called stochastic.
The nature of the signal may not be identical with the process. For example
the movement of molecules in a gas is random but the pressure measured and
signal transmitted will be deterministic if the gas is in a steady state.
6.2.1 Deterministic data signals
As described above these signals come from sensor systems measuring
processes that have some time-dependent relationship. An example is the
temperature of a hot object as it is allowed to cool. This would be described
as a transient phenomenon. The resonating panel on a vehicle would be an
example of a complex periodic process, the signal from the vibration
sensor containing harmonics and perhaps other non-related frequencies.
The output of a resonator sensor, although measuring some steady variable,
would be a sinusoidal signal with some harmonics of lesser amplitude.
Deterministic data signals can be subdivided as shown in the diagram
(gure 6.1) and are described below, following Bendat and Piersol (1986).
Periodic sinusoidal data signal
In this signal a single frequency is present and can be easily described
mathematically as
vt v
0
sin.t c. 6.3
Table 6.1. Examples of random and deterministic signals.
Random Deterministic
Data signal Sounds from an engine bearing Output from a tachogenerator
Unwanted signal Noise Interference
220 Data signal characterization and recovery
Complex periodic data signal
This waveform is repetitive and is reproduced exactly every time period. The
time period is that associated with the fundamental frequency present ..
Fourier showed that such waveforms could be synthesized by summing a
set of sine and cosine terms whose frequencies are multiples of the
fundamental and that have appropriately weighted amplitudes. Multiples
of a frequency are also referred to as harmonics, the fundamental is the
rst harmonic. A steady value may also need to be added to account for
any oset:
vt
a
0
2

1
n1
a
n
cos n .t b
n
sin n .t 6.4
where
a
n

2
T
_
T
0
vt cos n .t dt 6.5
and
b
n

2
T
_
T
0
vt sin n .t dt 6.6
where n is an integer 0, 1, 2, 3, . . . and T 1,f 2,.. These complex
signals are the most common.
Quasi-periodic data signal
The Fourier series (equation (6.4)) shows how combinations of harmonics
can produce any shape of repetitive, periodic wave. However, if three or
more sinusoids are combined that are not multiples of each other then the
resultant wave is not periodic. A general expression for such a waveform is
vt
1
n1
v
n
sin.
n
t 0
n
. 6.7
The frequency components .
n
will not be related by rational numbers.
Figure 6.1. Deterministic data signals.
Classication of data signals 221
Transient data signal
This type of data signal is not continuous in time. Such signals often occur at
the start or nish of some otherwise steady process. The relationships for
some common transient responses are shown below.
vt v
0
expat 6.8
vt v
0
expat cos bt. 6.9
These functions do not contain individual frequencies as in the previous
classications. The Fourier transform
V.
_
1
1
vt expj.t dt 6.10
illustrates that the signal contains a continuum of frequencies.
It should be remembered that whenever sampling of a steady process
has a start and nish the sample record will appear to have come from a
discontinuous process. Frequency components will be present in the
sample record that were not in the original. This is a similar problem to
aliasing described in section 5.2. To minimize the eect a long sample
record should be taken.
6.2.2 Random data signals
There are many occasions when the data produced from some phenomenon
is random. This may be simply because the variable being measured follows
no pattern. An example of this is the signal from a temperature sensor in
some environmental chamber that uctuates because occasionally the door
is opened for access. Another example of a random data signal is that of
the output from a vibration sensor on an engine bearing. For a full descrip-
tion of these phenomena sampling would have to be continuous to innity.
A further class of random phenomena is that of the chaotic system.
Much interest has been generated recently in chaos theory. The mathematical
relationship(s) of these systems can be written down and are often nonlinear,
time-dependent, high-order, dierential equations. Such equations can only
be solved numerically. The criterion that classies them as chaotic is that an
innitesimal change to the starting conditions signicantly alters the
outcome. In a real system this eectively means that, because of the
uncertainty of the starting conditions set by the natural limits on precision,
the outcome cannot be determined. Examples of chaotic systems are tur-
bulence in uid ow, laser output stability and the weather. To fully describe
the process of a chaotic phenomenon not only would sampling have to be
continuous to innity but all other possible outcomes have to be known
and recorded to innity as well.
The sub-groupings of types of random phenomena are described in
gure 6.2, again following Bendat and Piersol (1986).
The storing of a data signal from a random phenomenon is called the
sample record, vt. As mentioned above, the complete record would have
to run to innity and can then be justiably described as the complete
description of the random process. If a phenomenon is chaotic all possible
records have to be catalogued to give a full description of the random
process. Such a set of records is known as the ensemble, {vt}.
The stationary random process
Consider the collection of samples v
1
. . . v
n
shown in the graphs in gure 6.3.
The mean value " vv can be determined at some time t
1
assuming that there are
an innite number of sample records. This is said to be in the limit of N !1,
" vvt
1
lim
1
N
N
k1
v
k
t
1
. 6.11
The mean is also-called the rst moment.
Another statistical parameter called the correlation or joint moment can
be determined using values from two dierent times in a sample function and
then averaging over the entire ensemble:
R
vv
t
1
. t
1
t lim
1
N
N
k1
v
k
t
1
v
k
t
1
t. 6.12
Because the correlation is performed on two dierent parts of the same
signal, R
vv
is termed the auto-correlation function.
Note that both " vv and R
vv
are, strictly speaking, only truly representative
of the random process if the whole ensemble is used, i.e. N !1.
If the values " vv and R
vv
do not change as t
1
varies, the process is said to
be stationary. Under these conditions R
vv
is only a function of t. For truly
stationary events all higher order moments and joint moments are also
time invariant. If this is not the case the process may be referred to as
weakly stationary; see Baher (1991).
Figure 6.2. Classication of random processes.
Classication of data signals 223
An example of a stationary process is the noise from an engine bearing
when running at some steady speed. The ensemble of sensor signals would all
look very similar.
An example of a non-stationary process is the output from a pressure
sensor monitoring the sequential explosions in an internal combustion
engine. Each explosion would provide another set of data for the ensemble.
In this case both the values of " vv and R
vv
are dependent on the value of t
1
.
The ergodic random process
The above calculations can be performed on a single sample record say the
kth. Again, to obtain a complete description of the signal the calculations
must be performed for an innite time window, T !1.
" vvk )j
k
lim
1
T
_
T
0
v
k
t dt. 6.13
The symbol j is used to indicate that the mean value has been derived from a
theoretical function rather than a discrete set of empirical results.
R
vv
t. k lim
1
T
_
T
0
v
k
tv
k
t t dt. 6.14
Figure 6.3. A random process ensemble of data signals.
Now, if the random process is stationary AND in addition all properties such
as those derived in equations (6.13) and (6.14) are the same, whichever of the
sample records is used, then the process is said to be ergodic. Thus,
" v vk ! " vv and R
vv
t. k ! R
vv
t.
Most real life problems can be treated as ergodic. This is very fortunate since
parameters for describing a random process can then be derived from a single
sample function.
In the example above of an engine with vibration sensors, if every sensor
was generating waveforms that produce identical values of these statistical
parameters then the process is ergodic and clearly only one sensor is required.
However, if the sensors were located at dierent places around the engine so
that the waveforms were statistically dierent then each sensor is providing
unique data about the process. The process is stationary in that the " vv and
R
vv
values are the same when calculated across the sample records but do
not give the same values when a single record is used.
Non-stationary random processes
It is often dicult to extract any meaningful value from statistical data
obtained from ensembles of these processes. Impractically long ensembles
are often needed to provide sucient data.
As an example, supposing the engine-bearing problem described above
was for an aircraft. For a full description of the process the whole ight
would need to be recorded since signal levels and spectra would be varying
throughout the journey. However, Bendat and Piersol (1986) note that in
some investigations special categories of non-stationary behaviour exist
which, upon inspection, can be statistically analysed. For example, some
random process of ensemble {vt} could have a sample function of the form
v
k
t Atu
k
t 6.15
where {ut} is a stationary random process and At is a deterministic
factor. If such a process can be recognized, ensemble averaging may not be
necessarya single sample record may suce as for an ergodic process.
6.3 Characterization of random data signals
Since it is not possible to know exact time-domain descriptions due to the
unpredictability of random data some form of statistical approach must be
used. There are four commonly used statistical functions that can provide
quantitative characterization of random data:
.
mean values and variances,
.
probability density function,
Characterization of random data signals 225
.
auto-correlation function,
.
auto-spectral density function.
6.3.1 Mean values and variances
The most commonly used statistical parameter is the average or mean. For a
time-dependent signal such calculations need to be taken over the cycle. If
this is not easily achievable then clearly a large number of cycles is needed
so that errors caused by the inclusion of a part cycle are minimized. Calcula-
tion of the mean was described in equation (6.11). Using calculus notation,
j
v

1
T
_
T
0
vt dt. 6.16
Power in any electrical or mechanical system goes as the square of the ampli-
tude. Consequently, the mean square value may be more useful in some
applications.
2
v

1
T
_
T
0
v
2
t dt. 6.17
A further useful parameter is the variance o
2
. This determines the mean
square value about the mean.
o
2
v

1
T
_
T
0
vt " vv
2
dt. 6.18
The positive root of o
2
is the standard deviation.
6.3.2 Probability density functions
A probability density function is useful when wishing to determine the
likelihood of a particular value occurring. Consider the signal vt, shown
in gure 6.4.
The probability Pv that v will lie within the range shown (see gure
6.5), v
1
and v
1
v, can be estimated using the fractional time interval:
T
1
T
2
T
3
T
4
T

N
k1
T
k
T
. 6.19
Figure 6.4. The occurrence of a waveform within some voltage limits.
Such calculations could be performed for each interval of v across the range
to give a histogram of the probabilities.
In the limit v )0 this histogram becomes a distribution, the prob-
ability density function pv, shown in gure 6.6.
Now to calculate the probability P of v lying between v
1
, v
1
v an
integral is used:
P
_
v
1
v
v
1
pv dv. 6.20
Probability distributions are usually normalized meaning that the sum of all
probabilities is unity:
_
1
1
pv dv 1. 6.21
Other common statistical parameters can be calculated as follows:
mean value " vv
_
1
1
vpv dv 6.22
mean square value
"
v
2
v
2
_
1
1
v
2
pv dv. 6.23
The calculation for the mean square value can be adapted for higher orders,
m
n

_
1
1
v
n
pv dv 6.24
Figure 6.5. The probability of v lying within discrete intervals of v.
Figure 6.6. The probability density function of v.
where m
n
is referred to as the nth moment. If the mean is included then m
n
is
called a central moment,
m
n

_
1
1
v " vv
n
pv dv. 6.25
Moments have the following properties:
1. If pv is symmetrical in form then the 3rd, 5th, 7th etc. moments are zero.
This provides a useful test of symmetry.
2. Higher order moments give greater precedence to the extremes of a signal.
3. Setting n as 2 gives the variance.
The Gaussian or normal distribution
This is by far the most commonly encountered distribution and is worth
examining in more detail. It represents the distribution of some parameter
x, which could be a length measurement or signal value. Gauss is reputed
to have proposed this distribution when analysing astronomical data. His
hypothesis was that the probability of a particular error occurring decreased
with the magnitude of the error. It is found to be empirically correct. The
typical form of the spread of values around the mean is shown in gure 6.7.
The general mathematical relationship describing such a bell-shaped
curve is
px Aexp
_
x " xx
2
B
_
. 6.26
Figure 6.7. The Gaussian distribution.
Normalizing is achieved using a form of equation (6.21). Also, using
equation (6.25) relates B to the variance. Combining the two results gives
px
1
o
2
p exp
_
x " xx
2
2o
2
_
6.27
which is the Gaussian or normal probability density function. The curve of
gure 6.7 was generated using the Matlab
1
listing below:
clear, clf
mean=810; variance=5;
x= [800:.1:820]; A=1/(2*pi*variance).0.5;
B=(x-" xx).^2;
y=A*exp(-B/(2*variance));
plot(x,y);
As described above in equation (6.20), to nd the probability of x lying
between the values a and b, an integration is performed,
Px
a
< x < x
b

_
b
a
px dx 6.28
Fortunately, standard tables are available for evaluating such probabilities.
The table value is calculated integrating from 1 to the desired value of
x. Obtaining the table values for the two and then subtracting that of a
from b gives the probability.
A graph can be drawn of these tabulated Px values versus x to give a
further type of distribution known as the cumulative distribution function
(gure 6.8).
Variance and the standard deviation
In section 6.3.1 the variance was described as the square of the standard
deviation. This is not strictly true. As described in section 1.2.1, the standard
deviation s is a parameter determined from a limited set of empirical values
Figure 6.8. The cumulative distribution function.
not calculated using a probability distribution,
s
N
i 1
x
i
" xx
2
N
. 6.29
However, in the theoretical limit for N )1 then s )o.
Finally, it is worth noting that the point of inection on either side of the
bell-shaped curve is displaced from the mean by a value equivalent to the
standard deviation. This can be illustrated by nding the turning point of
the Gaussian. (The mean is taken as zero for convenience.)
px
1
o
2
p exp
_
x
2
2o
2
_
6.29
dp
dx
x
o
2
p 6.30
d
2
p
dx
2

p
o
2

x
2
o
4
p 6.31
which when set to zero gives x o.
The usefulness of the standard deviation is in the estimation of the
probability of an empirical value occurring at some way from the mean.
For x to occur between " xx and
1o. Px 0.683
2o. Px 0.954
3o. Px 0.9973.
Thus, the probability that a result will occur farther than 3o from the mean
is less than 2.7 per 1000.
6.3.3 Multi-variable probabilities, covariance and correlation
The use of correlation with time-dependent signals has many signicant
signal analysis applications. For example, it is used in echo-sounding,
system identication and instrumentation. To develop the mathematical
relationships for correlation it is helpful to rst consider two-variable
probability.
Assuming that the variables x and y are subject to probability distri-
butions px and py, then the probability of two independent events
happening is simply the product of their individual probabilities,
Px. y PxPy 6.32
Pa < x < b. c < y < d
_
b
a
px dx
_
d
c
py dy. 6.33
However, if the variables are related the probability functions are no longer
separable. An example of this is the weight and height distribution of a
population. The probability distribution function is now a composite
px. y. There is said to be a correlation between the two variables,
Pa < x < b. c < y < d
_
b
a
_
d
c
px. y dxdy. 6.34
Now, following the denition of the variance in equation (6.18) a co-variance
C
xy
can be determined,
C
xy

_
1
1
_
1
1
x " xxy " yypx. y dxdy. 6.35
If these variables are identical then equation (6.35) becomes
C
xy

_
1
1
_
1
1
x " xxx " xxpx dx 6.36
so that
C
xy
o
2
x
o
2
y
. 6.37
Alternatively, if the variables are totally unrelated then
C
xy

_
1
1
x " xxpx dx
_
1
1
y " yypy dy 6.38
which yields the result
C
xy
0 0. 6.39
These results can be summarized as
0 C
xy
o
x
o
y
6.40
depending on the level of correlation between x and y. A correlation
coecient has been dened,
,
C
xy
o
x
o
y
6.41
which lies between 1.
Independent random variables are uncorrelated. The converse, that
uncorrelated random variables are independent, is not necessarily true but
usually is in a physical system.
6.3.4 The correlation concept for time-dependent signals
As stated above the correlation of signals is an important tool in signal
analysis. If x and y are signals varying in time, xt and yt then the
correlation concept described above needs to be adapted. The changes are
1. Some suitable time window has to be chosen for the integral limits.
2. Being time-dependent signals the correlation becomes time dependent.
The two classes of correlation are auto-correlation,
R
xx
t
1
2T
_
T
T
xtxt t dt 6.42
and cross-correlation,
R
xy
t
1
2T
_
T
T
xtyt t dt. 6.43
The graph produced of Rt versus t is called the correlelogram (auto or
cross) and indicates whether the signal(s) involved, when time shifted by t,
show any correlation throughout the sampling window 2T. The function
can be easily determined with an algorithm,
R
xx
r t
1
N
N
j 1
x
j
x
j r
6.44
where r t represents the discrete t values, t being the sampling period and
r an integer beginning at zero.
If a correlation peak does occur at some value of t a similar expression
to equation (6.41) can be used to estimate the level of correlation.
6.3.5 Auto-correlation and spectral density functions
Power spectra and the correlation function
It is often useful in engineering to know how the power in a signal or an
oscillating system is distributed throughout the frequency spectrum. The
power spectral density function G indicates the power distribution in a
signal and has units of watts per radian or watts per hertz. The mean
power
"
WW within some bandwidth can be obtained from G,
"
WW
_
f
2
f
1
Gdf . 6.45
Now, the auto-correlation of a voltage signal vt has units of volts
2
which is
the power dissipated in a unit resistor. Intuitively, one might think that there
ought to be some relationship between the auto-correlation and the power
spectral density function. To determine this relationship it is necessary to
begin with the Fourier transform of the auto-correlation. For some signal
of voltage vt equation (6.42) is written as
R
vv
t
1
2T
_
T
T
vtvt t dt. 6.46
Because R
vv
is a function of t, expj.t dt is used to perform the Fourier
transform of equation (6.10),
=R
vv
t. T
_
1
1
_
1
2T
_
T
T
vtvt t dt
_
expj.t dt. 6.47
In order to align the limits of integration a large sampling interval is needed.
Making T !1 gives
=R
vv
t. T
1
2T
_
1
1
_
1
1
vtvt t expj.t dt dt. 6.47
Re-arranging,
=R
vv
t. T
1
2T
_
1
1
__
1
1
vt expj.t dt
_
vt t expj.t t dt
6.48
which are eectively two Fourier transforms. Writing Vj. as the trans-
form of vt in the square brackets above gives
=R
vv
t. T
1
2T
_
1
1
Vj.vt t expj.t t dt. 6.49
Substituting ` t t,
=R
vv
t. T
1
2T
Vj.
_
1
1
v` expj.` d` 6.50
revealing another Fourier transform. Thus,
=R
vv
t. T
1
2T
Vj.Vj. 6.51
which can be written as
=R
vv
t. T
1
2T
jVj.j
2
. 6.52
It is not immediately obvious what equation (6.52) means. To interpret this
relationship it is useful to begin with an expression for the total energy
dissipated in a unit resistor during the time interval T to T. For a signal
of voltage vt, the energy dissipated during the time window 2T would
normally be estimated using
E
_
T
T
v
2
t dt. 6.53
But another expression for E can be obtained using Parsevals theorem which
states that
the total energy in a signal is the sum of the energies in each of its
frequency components.
Again for a unit resistor this can be expressed as
E
_
1
1
jVj.j
2
d.. 6.54
Dividing equation (6.54) by 2T gives the average power
"
WW
1
2T
_
1
1
jVj.j
2
d.. 6.55
However, equation (6.55) is the integral of equation (6.52). These can be
equated if equation (6.55) is rst dierentiated with respect to .,
d
"
WW
d.

1
2T
jVj.j
2
S
vv
. 6.56
where d
"
WW,d. is the power spectral density function labelled here as S
vv
..
Comparing equation (6.56) with equation (6.52) gives
=R
vv
t. T S
vv
.. 6.57
So, the Fourier transform of the auto-correlation function produces the
power spectral density function. For this reason S
vv
. is also-called the
auto-spectral density function. Auto-spectral density is the fourth of the
descriptive techniques listed in section 6.3.
The Fourier transform of the auto-correlation shown above and its
inverse are known as the WienerKhintchine relations. The general expres-
sion of the transform is
S
xx
.
_
1
1
R
xx
t expj.t dt 6.58
which can be re-written as
S
xx
.
_
1
1
R
xx
t cos .t dt. 6.59
The j sin .t term disappears because S
xx
and R
xx
are real functions, not
complex. Also, since the cosine function is symmetrical the integral can be
limited to positive values of t,
S
xx
. 2
_
1
0
R
xx
t cos .t dt. 6.60
The inverse relationship is
R
xx
t
1
2
_
1
1
S
xx
. expj.t d. 6.61
or
R
xx
t
1
_
1
0
S
xx
. cos .t d.. 6.62
Weiner (USA) and Khintchine (USSR) independently derived these
relationships in the early 1930s.
The practical power spectral density function
The careful reader will have noticed that two separate symbols have been
used in this section to represent power spectral density, G and S
vv
. This
was an attempt to distinguish practical usage from theoretical derivation.
The function S
vv
. came from a Fourier transform and because the signal
was taken as continuous throughout time ranging from 1 to 1 the
resulting frequency range has the same limits. Negative frequencies,
however, are only of theoretical use when dealing with integrals. For practi-
cal use the single-sided power spectral density function G is employed. Now,
since R
vv
. is determined by multiplying vt with itself it must be an even
function, meaning symmetrical around the ordinate axis,
R
vv
t R
vv
t. 6.63
It follows that S
vv
. must be the same. Consequently, G. can be obtained
from S
vv
. using only positive frequencies, so
G. 2S
vv
. 2=R
vv
t. T for 0 . 1. 6.64
Also, from equation (6.56)
G.
1
T
jVj.j
2
for 0 . 1. 6.65
This can be a convenient method of performing a spectral analysis on a time-
domain signal. The auto-correlation and Fourier transform functions of
equations (6.64) and (6.65) can be eciently computed using algorithms in
a dsp microprocessor system. See Brook and Wynne (1988) for a fuller
description. As these authors point out, historically equation (6.64) was
used to determine frequency spectra. This was because the discrete Fourier
transform (DFT) is computationally intense requiring 4N
2
multiplications
and 2N
2
additions where N is the number of spectral coecients. However,
the publication of the fast Fourier transform (FFT) by Cooley and Tukey
(1965) meant that equation (6.65) can be computed directly. Interestingly,
because of the performance of the FFT some signal analysers now determine
the auto-correlation function by rst performing the FFT on a signal to
obtain S
vv
. and then, from equation (6.61), obtain R
vv
t using the
inverse FFT.
6.4 Noise
6.4.1 Origins of noise
As explained in section 6.1, noise is a random signal and is treated statis-
tically. There are three processes that generate noise.
(a) Thermal noise
The thermal motion of charge carriers in any circuitry produces tiny random
currents at the output terminals. This random motion has a mean of zero.
However, both the r.m.s. current and voltage are not zero and become
noticeable when the signal being recovered is weak. Thermal noise is also
sometimes referred to as Johnson or Nyquist noise. The integrated mean-
square-voltage in the frequency band f is given by
" vv
2
4kTRf 6.66
where k is Boltzmanns constant, T the absolute temperature and R the
resistance of the noise source. The spectrum is at and so, by analogy to
white light, is sometimes referred to as white noise. Equation (6.1) can be
re-written as a power spectral density function,
1
R
" vv
2
f
4kT watts per hertz. 6.67
The power spectral density for thermal noise is shown in gure 6.9.
(b) Shot noise
This form of noise is produced by the random arrival of charge carriers and
becomes noticeable when current ow is tiny. It can be shown using a Poisson
Figure 6.9. Power spectral density of thermal noise.
statistical analysis that the integrated mean-square-current is given by
"
ii
2
2ei
0
f 6.68
where e is the charge on the electron and i
0
is the mean current. The power
spectral density function is spectrally at, similar to gure 6.9.
(c) Flicker noise
A further type of noise occurs in semiconductors and some composition
materials. The power distribution declines with frequency and so is referred
to as one-over-f or icker noise in semiconductors (Nicollian and Brews
1982). Germanium devices tend to be worse than silicon. Flicker noise
does not occur in metallic components. The log of the power spectral
density is shown in gure 6.10. The integrated mean-square-voltage is
given by
" vv
2
Bi
2
f
f 6.69
where B is an empirical constant. It is more dominant than thermal noise
below 10 kHz. Flicker noise occurs in some composition resistors and so,
when dealing with low-level signals having frequencies of less than 10 kHz,
metallic thin-lm or wire-wound resistors should be used.
Amplication
When amplifying low-level signals, ampliers having low noise levels in
their input stages must be sought. Care must be taken when mounting
these, particularly when high input impedance devices are needed. A large
number of low-noise ampliers are commercially available and manufac-
turers sometimes include recommendations on circuit layout.
However, the measurement of a weak, steady variable poses further
problems to the signal conditioning. This is because the intermediate
Figure 6.10. Power spectral density of icker noise.
Noise 237
stages of the system have to be directly coupled and any amplier zero error
or drift will contribute to the measured value.
The commutating, auto-zeroing amplier may be a suitable choice to
reduce the eects of drift. This is a development following on from the
chopper-stabilized amplier that converts a d.c. signal to a.c. then uses a.c.
amplication and subsequently returns the signal to d.c. The advantage of
the commutating, auto-zeroing amplier is that it is constantly on-line.
There are a number of commercially available versions of this CMOS
integrated circuit package. As illustrated in gure 6.11, this device constantly
checks and corrects its own output drift. The input to the subsidiary amplier
B is grounded and its output then used to reset its own zero. At the clock
signal the output of the identical main amplier A is subtracted from the
output of B. Any discrepancy is then used to reset the zero of A. Note that
amplier A is always on-line. This helps to reduce clock-induced ripple in
the output. Zero stability is a factor of ve better than precision, bipolar-
technology, operational ampliers. The disadvantages of the device are
.
external capacitors are required for storing the zero-error correction voltage,
.
common mode voltage is restricted, 3.5 V to 5 V,
.
current output is limited, 5 mA,
.
if output saturation occurs recovery can be slow,
.
supply voltage V
to V
is limited to 18 V.
6.4.2 Noise reduction techniques
Whenever such noise phenomena occur a number of options are available to
enhance the signal-to-noise ratio.
Filtering
The noise spectra shown above are extended in frequency. If the signal and
noise are measured directly then a large contribution to the resultant value
Figure 6.11. Schematic diagram of an auto-zeroing amplier.
will come from the noise. A simple lter, placed in the signal conditioning
circuitry input, can be very eective if the signal is known to be in a restricted
part of the spectrum. Depending on the position of the data signal spectrum
low, high or band pass lters are used. See gure 6.12.
Simple time-averaging
Simple averaging can be used if the variable is static or slowly changing with
time, such as might occur with the temperature of a liquid in a large tank. A
low pass lter as described above is, in-eect, an analogue averaging device.
With the advent of cheap memory, averaging can be done more
sophisticatedly using digital techniques. For most types of noise the uctua-
tion of the amplitude of the current or voltage of the signal with time xt is a
Gaussian distribution around the mean. The determination of the mean
value of x and its standard deviation were described in section 6.3.1 from
equations (6.16) and (6.18). For practical use the discrete versions of these
functions are employed,
" xx
N
i 1
x
i
N
6.70
where the number of samples N taken within some time window T might be
typically 1024. The standard deviation of the samples is a measure of the
spread of the signal around the mean and represents the level of noise.
When " xx is zero, as is the case for pure noise, s is the root-mean-square of
x. It is determined using
s
N
i 1
x
2
i
N
. 6.71
As described in section 1.2.1, to gain greater condence in the value of " xx a
number of windows are sampled and the mean " xx
m
of all the " xx values
determined. It can be shown that the standard deviation s
m
for this ensemble
Figure 6.12. Noise ltering.
Noise 239
of mean values taken from p windows is determined using
s
m

s
p
p
N
i 1
x
2
i
pN
. 6.72
Clearly, increasing N or p reduces s
m
(the spread of the distribution) and so a
better estimate of the measurand is achieved. For example, increasing N or p
by a factor of 10 improves the signal-to-noise ratio by 10 dB. But obviously
the total sampling time pT needs to be short compared with the time for any
change in the variable to occur.
A renement to this technique is to perform a rolling average. Here the
mean value from the last window to be sampled is averaged with the stored
mean to give a new stored mean. This of course has the eect of awarding a
greater signicance to the last window sampled. Variations of the rolling
average technique have been developed where the mean values from the p
windows to be averaged are all weighted with progressively less signicance
the further back in time that they occurred. Various weighting functions have
been used.
These simple averaging procedures cannot be used when the noise
distribution is asymmetrical about zero since this would give a non-zero
value for " xx which would be incorporated with the measurand.
Sophisticated time-averaging
As indicated above there are limitations on the ultimate performance of
the simple averaging process. Firstly, the variable to be measured must be
virtually stationary and secondly there is the increase in 1/f noise at low
frequencies. The following techniques are more sophisticated averaging pro-
cesses that require that the signal is repetitive. This can often be arranged
either by driving the sensor with some form of periodic excitation or by
arranging that the measurand occurs repetitively. Such techniques are
eectively placing the data signal in a dierent part of the frequency
spectrum. A further advantage is that any amplication can be a.c.
coupled reducing the problems of drift.
The multi-channel scaler
For periodic or repetitive transient signals averaging can be performed with a
multi-channel scaler. This digital device consists of a number of channels,
1024 for example, which store the signal. If the repetition period is T a
time resolution t of T/1024 is achieved. Precision triggering is essential so
as to sample at the same point in the waveform at every scan. Up to 10
6
scans can be stored and averaged. Commercial systems often employ a
counting technique. Samples of the waveform are fed into a voltage-to-
frequency converter. This frequency is counted for the time t. The count
is then stored in the appropriate channel. Counts from subsequent scans are
either added or subtracted depending on the polarity of the sample. A
channel acts like the accumulator of an updown counter. After a large
number of scans the noise tends to average towards zero, as does any
unwanted repetitive signal having a period dierent from T. The bandwidth
is approximately equal to the inverse of the total sampling time.
The phase-sensitive detector
Phase-sensitive detection is a generic technique that can be implemented in a
number of forms. In all cases phase adjustment between a reference signal
and the data signal is available. Early commercial designs were entirely ana-
logue. Digital hybrids provide better stability. The basic principle is illus-
trated in gure 6.13. A deection bridge circuit driven sinusoidally, such as
might be the case for a strain gauge system, provides a useful example; see
Wilmshurst (1990). The phase-shifter is used to compensate for any
system-induced phase-shift. The output-smoothing lter performs an aver-
aging function. It has the eect of reducing the system bandwidth. It must
pass slowly varying data signals but reject the harmonics generated.
Frequency domain analysis is a useful approach in revealing the relative
merits of the dierent options when using phase-sensitive detection. To
conduct the analysis an analogue multiplier is substituted in place of the
switch. This provides a general model of the phase-sensitive detector, as
illustrated in the schematic of gure 6.14, enabling the performance of
dierent experimental congurations to be modelled.
Three multiplication types are described below with an appropriate
application to illustrate the concept.
Figure 6.13. Phase-sensitive detection.
Figure 6.14. Generalized phase-sensitive detector.
Noise 241
1. Bipolar square wave. This is the equivalent of the system in gure 6.13,
the example given being the strain gauge bridge. The bipolar waveform
performs the function of the inverting amplier. For the equivalent
performance the reference signal generator must have two outputs, one
sinusoidal the other bipolar. This square-wave action has the eect of
producing harmonics. The data signal appears as sidebands at these
harmonics, which also have the eect of increasing the noise.
2. Unipolar square wave. In some optical experiments, in order to use the
phase-sensitive detection technique, the beam of light is interrupted by a
chopper wheel rotating at high speed instead of having a signal generator.
The phenomenon under investigation is made to occur at a frequency set
by the light pulse repetition rate. A reference signal is taken from the
wheel using a separate photodiode. The waveform of the data signal is
of a unipolar format with data noise in one half cycle and noise alone
in the other. There is a signicant disadvantage to the unipolar waveform
over that of the bipolar. It has the same harmonics but, in addition, a d.c.
component. This results in the system accepting o-set drift and 1/f noise.
The reason for using the unipolar format in an optical experiment is the
convenience of modulating the light beam using the chopper wheel. A
more general advantage in other applications is that analogue multipliers
have an upper frequency limit of order MHz whereas switching circuits
can operate at GHz.
3. Sinusoidal waveform. The disadvantages described above when using
square waveforms do not occur when two pure sinusoids are multiplied
together. This is because no harmonics are produced. The multiplication
of two sinusoids results in two new sinusoids: one whose frequency is the
sum of the original frequencies and one whose frequency is the dierence
(equation (6.81)). If the two original frequencies are identical the result is
then a single sinusoid at twice the original frequency. This technique has
been successfully applied in spectroscopy. A range of wavelengths is used
to irradiate the material being investigated. At a particular wavelength
some peak (or trough) will occur. This is illustrated in gure 6.15(i). To
locate this peak the wavelength is increased steadily during the experiment
but, in addition, it is made to oscillate with small amplitude throughout
the much longer sweep cycle (gure 6.15(ii)). The result is to generate a
data signal modulated at the oscillating frequency the amplitude of
which depends on the steepness of the edge of the peak (gure 6.15(iii)).
Also, a phase-shift occurs between the waveforms on the up and down
sides of the peak. Following multiplication with the oscillation signal a
waveform is produced (gure 6.15(iv)) having an asymmetry centred on
the peak cross-over. With a smoothing lter the peak cross-over is
clearly seen in gure 6.15(v). This waveform is eectively the dierential
of the spectroscopy peak of gure 6.15(i). For the system to function
correctly the time constant of the lter must be long compared with the
time period of the oscillations but short compared with the sweep time. If
the sweep time is too fast for the smoothing circuit the cross-over point
will shift.
Generally, whatever waveform is used in a phase-sensitive detector the two
half-periods must be of exactly the same duration otherwise the level of
the recovered signal will be reduced.
The lock-in amplier
Phase-sensitive detectors can be built from discrete elements but it is usually
more cost-eective to use commercially available systems, which are
described as lock-in ampliers. See Meade (1983) for a detailed description.
Figure 6.15. Spectroscopy using phase-insensitive detection. Top to bottom: (i) a trans-
mission peak (shown scanned over the same time as the other results for comparison),
(ii) the scanning signal, (iii) the output of the optical sensor, (iv) the output of the multiplier
block, (v) the smoothed output.
Noise 243
Commercial devices contain a variable-frequency signal generator with
tracking lter, a low-noise wide dynamic range amplier, a high performance
linear phase detector and a post-detection lter with a choice of time con-
stants. Provision is made for the use of an external signal generator. Phase
adjustment is done manually by the operator in order to get the best signal.
Auto-correlation
One powerful application of correlation is to the recovery of signal infor-
mation using auto-correlation. Auto-correlation was described earlier in
sections 6.2.2 and 6.3.4. By correlating a signal with itself it is possible to
(i) derive some statistical information from a random signal,
(ii) separate sinusoids from a random signal,
(iii) determine the period of non-sinusoidal repetitive waveforms.
The statistical application was described in section 6.3. The following para-
graphs show how a sinusoid can be recovered when buried in noise and this
same concept can also be used to determine the period of a non-sinusoidal
waveform.
Auto-correlation of the sum of two signals
Supposing the signal zt consists of two parts,
zt xt yt 6.73
where xt is the desired data and yt is noise. Then, from equation (6.14),
reproduced here, an auto-correlation is determined,
R
vv
t
1
2T
_
T
T
vtvt t dt. 6.74
So the auto-correlation of equation (6.73) can be expressed as
R
zz

1
2T
_
T
T
xt ytxt t yt t dt 6.75
R
zz

1
2T
_
T
T
xtxt t xtyt t ytxt t ytyt t dt
6.76
R
zz
R
xx
R
xy
R
yx
R
yy
. 6.77
The terms R
xy
, R
yx
are cross-correlations terms of the correlation between x,
y. If there is no correlation then these two are both zero. If the unwanted
signal is true random noise then there will be no correlation between noise
and data. So
R
zz
R
xx
R
yy
. 6.78
Auto-correlation of a sine wave
The data signal xt can be expressed as
x
k
t x
0
sin.
0
t 0k. 6.79
The subscript k is used to indicate that the data could be part of some
ensemble {x(t)} with 0k being some unbiased random phase-shift. The
auto-correlation function can be performed across the ensemble if the data
is ergodic. The auto-correlation function then becomes
R
xx
t
x
2
0
2
_
2
0
sin.
0
t 0 sin.
0
t t 0 d0. 6.80
Using the trigonometrical identity,
sin Asin B cosA B cosA B 6.81
gives
R
xx
t
x
2
0
2
cos .
0
t. 6.82
Performing the auto-correlation across the ensemble of sample records,
rather than along a single sample record, gives the important result that
the auto-correlation function of the kth sample is independent of any
phase shift present and is the same frequency.
Auto-correlation of noise
The auto-correlation of noise is not calculated in the way used above. This is
because the time-dependent voltage signal is random and so no analytical
expression can be substituted into the integral for Rt. Instead the
value for the auto-correlation is determined using the WienerKhintchine
transformation, which converts the power spectral density function S f
into the auto-correlation directly,
R
xx
t
_
1
1
S
xx
f expjtf df . 6.83
In practice the frequency range is positive only so, writing G f for the real
power spectral density, equation (6.83) becomes
R
xx
t
_
1
0
G
xx
f cos 2ft df . 6.84
Note that the complex sine term is ignored since, as stated in a previous
section, auto-correlations are always even functions of t.
For a true white noise spectrum all frequencies are represented with
equal amplitude so G
xx
f a. Substituting in equation (6.83) with
Noise 245
S
xx
f
1
2
a gives
R
xx
t
a
2
_
1
1
expj2ft df . 6.85
But this can be evaluated easily by referring to the Fourier transform pair for
a delta function, ct. A delta function is an innitely tall spike at the origin
and innitesimally narrow such that the area is unity:
=ct
_
1
1
ct expj2ft dt 1. 6.86
But the inverse Fourier transform of unity is
=
1
1
_
1
1
expj2ft df ct. 6.87
Thus by comparing expressions it can be seen that
R
xx
t
a
2
ct. 6.88
Figure 6.16 is a sketch of this origin-centred delta function. Noise in practice
will not extend to innity due to the bandwidth limit of the system. So, to
evaluate equation (6.84) for a real system bandwidth limited noise is used.
The power spectral density function is shown in gure 6.17.
Substituting this function into equation (6.84) gives
R
xx
t
_
f
0
1,2B
f
0
1,2B
a cos2ft df 6.89
a
_
sin 2ft
2t
_
f
0
1,2B
f
0
1,2B
6.90
a
t
sin tB cos 2f
0
t 6.91
aB
_
sin tB
tB
_
cos 2f
0
t. 6.92
Figure 6.16. The auto-correlation of pure white noise.
For the special case where the frequency bandwidth ranges from 0 to B, so
that f
0

1
2
B, then using the trigonometrical identity,
sin 0 cos 0
1
2
sin 20 8.93
gives
R
xx
t aB
_
sin 2tB
2tB
_
. 8.94
Functions of the type sin x,x are sometimes called sinc functions. The form
of the function is shown in gure 6.18.
Finally, it can be seen that if a combined signal of a sine wave buried in
noise is auto-correlated then the autocorrelogram will separate the two. The
separation will increase the further from the origin.
Figure 6.17. Power spectral density function of bandwidth limited noise.
Figure 6.18. The sinc function.
Noise 247
6.5 Interference
6.5.1 Sources of interference and coupling mechanisms
As explained in section 6.1, interference is a deterministic phenomenon. Its
origins are virtually all from electromagnetic waves produced by
.
local power cables either external to the system or badly routed within it,
.
d.c. circuits which are being switched and consequently generating electro-
magnetic transients,
.
electrical discharges from high voltage terminals (a.c. lines produce inter-
ference but corona discharge from high voltage d.c. will generate noise),
.
arcing in electrical machinery,
.
uorescent lighting generating waves at twice the power line frequency,
.
digital circuitry,
.
radio transmissions,
.
microwave equipment.
Inductive coupling
From Faradays law, the induced emf in some secondary circuit element can
be expressed in terms of the current in the primary element,
emf M
dI
p
dt
. 6.95
The value of M is set by the geometry of the system. The general expression
for M was rst described by Neuman; see Bleaney and Bleaney (1965):
M
jj
0
4
__
ds
1
ds
2
r
6.96
where ds
1
and ds
2
are the current elements and r the separation. Its analytical
solution is dicult in all but the simplest geometrical arrangements. For
example, two co-axially mounted plane coils whose separation z is large
compared with their radii a and b and having n
1
and n
2
turns respectively is
M
jj
0
a
2
b
2
n
1
n
2
2z
3
. 6.97
A common occurrence of this coupling type is that of the magnetic eld
produced near a power transformer being coupled into some pre-amplied
circuitry containing weak signals.
Capacitive coupling
Alternating electric elds can induce charge movement on nearby conducting
objects eectively forming a capacitance across the gap. Again, the analytical
determination for the value of the capacitance can be dicult to derive in all
but the simplest cases. The simplest case is that of the two parallel plates,
C
r
A
d
6.98
where d is the plate separation and A is their common area of overlap.
Another useful expression is that of the capacitance per unit length
between two parallel, long wires,
C

0
r
log
e
d,a
Fm
1
6.99
where d is the separation of their centres and a their radii. When a wire is
close to and running parallel with a large conducting plane a similar
expression is obtained,
C
2
0
r
log
e
2h,a
Fm
1
6.100
where h is the separation. A simple example of how coupling might occur is
shown in gure 6.19.
The stray capacitances form a potential divider across the supply lines.
The signal line is eectively tapping into the electric eld across the gap
giving an additional signal,
X
2
X
1
X
2
V
s

1,j.C
1
1,j.C
1
1,j.C
2
V
s

C
2
C
1
C
2
V
s
. 6.101
Loop circuits
Closed loop circuits can be inadvertently created by connecting leads, on
printed circuit boards or with earth connections as shown later in gure
6.21. Such loops can act as antennae and consequently introduce electro-
magnetic waves into the system.
The following two phenomena are not strictly coupling mechanisms but
rather injection processes.
Multiple earth connections
It is useful to rst consider some terminology. Many use the words earth and
ground synonymously. Earth is more commonly used in the United Kingdom
Figure 6.19. Coupling by stray capacitance.
Interference 249
than in North America. There are even two dierent circuit symbols (gure
6.20) employed, one for earth or ground the other for the equipment chassis
or frame. The British Standard is BS 3939, EN 60617-2.
Ott (1988) gives a standard denition of a ground as an equipotential
point or plane that serves as a reference potential for a circuit or system,
in that its voltage does not change regardless of the current applied to it
or drawn from it. He goes on to point out that this is not achievable in
practice and suggests that a better denition might be a low impedance
path for current to return to source.
There are two purposes for grounds: for safety and for the signal refer-
ence. Neither the real earth, earth cabling nor ground-planes have zero
potential dierence throughout. The resistance of soil can vary by several
orders of magnitude due to variation in composition and water content;
see Kodali (1996). The value of 1 per metre length could be taken as a
nominal order of magnitude estimate. By contrast, copper cable of 2 mm dia-
meter has a resistance of about 5 m per metre. Consequently, if a current is
passing through the earth or down an earthing cable a potential dierence
will exist between any sample points. Heavy electrical machines and faulty
equipment whose fault is below the trip level may cause earth leakage.
Also, stray magnetic elds may cause eddy currents by inductive coupling
into the earth.
These faults will have the eect of injecting an extra potential dierence
in any signal lines which are using earth cabling as part of the circuitry, such
as with ordinary co-axial cable, or as part of some bonding safety system
(gure 6.21).
Such leakage currents are at the power frequency. These earth loops can
be responsible for hum in studio audio systems. An interesting example is that
of the electric locomotive or tram that uses the track as the circuit return.
Figure 6.20. Earth-ground symbols: (a) earth or ground, (b) chassis or frame.
Figure 6.21. Injected interference from multiple earth connections.
Without the use of a Faraday cage shrouding the track foundation, tens of
volts may occur in the soil near the track as the vehicle passes.
It is also possible for an earth loop to act as an antenna in the way
described above under loop circuits. It is the higher frequencies that cause
interference via this mechanism rather than the power line frequency.
Printed circuit boards
A result similar to that caused by multiple earthing described above may
occur with the ground plane on a printed circuit board. The reason,
however, is that the signal and/or power line ground may not remain
everywhere at zero volts if the currents being returned along the track are
too large for the size of copper strip laid down. Any potential dierence
occurring between dierent parts of the ground plane will eectively intro-
duce a signal into the system.
It is a common experience of circuit board designers that signals
produced in some section of a system are carried through the supply rails
into other sections. This is a particular problem at high frequencies. It is as
a result of the supply rails not being able to deliver the power required by
a section so that the supply voltage is no longer steady but ripples at the
oscillating or switching frequency. This has the eect of causing other
sections of the system to be injected with the same frequency. CMOS
components are generally more prone to this problem than bipolar.
Co-axial cable
Early co-axial cable suered from being sensitive to movement; see Perls
(1952). Bending of the cable causes friction between layers that can create
charge on the layer surfaces. If, in the bending process, contact is lost
between some small part of the insulation and the conductor the charges
are not conducted away. This results in an additional emf at the terminals.
The initial solution was to coat the insulation in graphite powder so as to
maintain electrical continuity. However, modern manufacturing techniques
have enabled better physical bonding to occur between conductors and
insulation.
It is still good practice not to allow co-axial cable free movement when
carrying tiny signals.
6.5.2 Interference reduction techniques
Geometrical isolation
The inductive and capacitive mechanisms described above are signicantly
dependent on the physical parameters of the layout. Simply increasing the
physical separation of the components will decrease the coupling. The
interference will decrease with the inverse of the distance or better.
Interference 251
Shielding
The exclusion of steady or low frequency magnetic and electric elds can be
achieved by enclosing the sensitive circuitry within metallic shields.
Using Gausss theorem it can be shown that electrostatic charge
must reside on the outside of a conductor. Steady electric elds cannot be
sustained within metals since any potential gradient would cause charge to
move until the eld is zero. Consequently, the outer surface becomes an equi-
potential where any induced or transferred charge resides. These concepts are
also true for a hollow metallic conductor. This means that a metallic box is an
eective shield to steady, external electric elds. It is possible to use a metallic
mesh for shielding provided that the holes are small compared with the
distance between the screen and circuitry to be protected.
The ow of magnetic ux depends on the reluctance of the ux path. The
reluctance for a particular material specimen goes as
<
l
j
0
j
r
A
6.102
where A is the cross-sectional area and l the path length within the material.
The reluctance is lowest in a path containing material of high relative
permeability j
r
. If a closed container is constructed out of such a material
then the magnetic eld lines will remain within this and so shield the internal
circuitry (gure 6.22).
A commonly used material for the shield is the alloy mu-metal. It is often
employed for the construction of shields for cathode ray tubes.
Whenever a shielding enclosure is used it is good practice to maintain a
signicant gap between circuitry and enclosure. This helps to ensure
minimum risk of capacitive or inductive coupling.
Screening
The solution of the equation for an electromagnetic wave striking a conduc-
tor shows that the penetration of the wave into the conductor is attenuated
Figure 6.22. Magnetic shielding.
according to the expression
I I
0
exp
_
x
c
_
6.103
where c ofj
0
j
r
1,2
. c is a measure of the severity of the attenuation and
is referred to as the skin depth. For a wave of frequency 1 MHz impinging on
copper c is approximately 6.6 10
3
cm. Clearly, a thin layer of metal foil is
all that is required at these frequencies. Consequently, the use of conducting
lms or braiding, such as is found in co-axial cable, is a very ecient way
to provide screening from high frequency electromagnetic interference. If a
cage or mesh is used the size of the holes should be small compared with
the wavelength.
Earthing
The obvious remedy to avoiding the problems associated with having
multiple earths is to have a common earthing point to which all equipment
is joined. This may be done as either a line or star arrangement as shown
in gure 6.23.
Each technique has a disadvantage. The line approach could pose a
health and safety risk if the earth connections between units become discon-
nected while the units were still connected to the mains. The star approach
still has an antenna-type loop providing an opportunity for electromagnetic
waves to enter the system.
It is now a mandatory safety requirement that mains powered electrical
equipment has a connection to earth using an appropriate gauge of wire for
the anticipated fault condition and that the installation, as a whole, is earth
bonded. Equipotential bonding is the wiring together of the electrical mains
earthing point, the water and gas supply pipes and any other ducting or
piping systems at the points of entry to the facility and any exposed metallic
parts of the building structure. Because of these constraints it may not be
possible or desirable to adopt the strategies of gure 6.23. There are a
number of techniques such that, even though the sensor and receiver are
earthed, the signal lines are prevented from injecting any earth fault potential
dierence into the data.
Figure 6.23. Earthing strategies: (a) line, (b) star.
Interference 253
(a) The isolation transformer
For optimum performance there should be a screen (gure 6.24) between the
primary and secondary windings so as to reduce stray capacitance coupling.
It is also possible to limit power transfer by running the core at near magnetic
saturation, which may be an additional benet.
This solution is eective but disadvantages are cost, physical size, d.c.
isolation and poor frequency response.
(b) Balancing
The use of a centre-tap on the source signal and at the receiver circuitry eec-
tively produces a dierential system independent of earth potential. Because
both conductors are part of a balanced system (gure 6.25) any common
mode interference should be equal in both arms and so cancel at the receiver.
The drawbacks are that the source has to be in two parts, which may
only be achievable using a centre-tap transformer with its inherent limita-
tions, and that the component values and wiring impedances are identical
in the two sides of the system. This latter condition becomes increasingly
more dicult to achieve at higher frequencies due to stray capacitance.
(c) The dierential amplier
The dierential amplier (gure 6.26) has no input reference zero but uses its
supply ground as zero. The amplier power supply is not connected to the
Figure 6.24. The isolation transformer.
Figure 6.25. The balanced system.
external earthing. Special twin-core co-axial cable is used. In applications
where the sender unit has an internal ground that is required to be connected
to the receivers internal ground double screened, twin-core coaxial cable is
used.
(d) Optical coupling
Optical isolation (gure 6.27) is the ultimate in providing separation from
sender and receiver. Optical isolation is most easily accomplished with
digital signals. For analogue applications care has to be taken over the
electronic circuit design to ensure linearity.
Printed circuit board design
The problems highlighted in the previous section can be minimized by good
design practice.
.
Large tracks should be used for the power supply rails and any ground rail
or plane to minimize resistance and therefore any drop in potential.
.
Separate grounds should be used for analogue and digital sections for
systems containing mixed components.
.
Capacitors should be placed across the supply lines close to each integrated
circuit (special miniature types are available that can be tted below the
chip) to steady their supply. Alarge capacitor is placed across the incoming
Figure 6.26. The dierential amplier system.
Figure 6.27. Optical isolation.
Interference 255
supply lines to ensure fast charging of the smaller capacitors. Tantalum
electrolytic or metallized polycarbonate are the preferred type due to
their low internal inductance.
.
For high frequency applications miniature bead inductors of about 10 mH
should be placed in the power supply lines to the high frequency sections of
the system.
.
Short leads are everywhere preferable so as to minimize electromagnetic
reception and emission. Generally, all wire pairs should be twisted so as
to nullify antennae behaviour.
Filtering
Some sensors need to be driven by an a.c. supply or have an a.c. signal
conditioning circuit. For these systems it is possible to arrange for the a.c.
frequency to be well away from any interference. Then a suitable lter can
be used, similar to those described in the noise reduction section above, to
allow only the desired signal through.
When interference is at a very specic frequency, such as the power line,
a narrow band stop or notch lter is useful. A simple passive version is the
twin-T lter (gure 6.28). The two sections of the T each cause a 908 phase
shift, which are 1808 out of phase at the notch frequency theoretically
giving innite attenuation. In practice attenuations are variable due to
tolerances. See Horowitz and Hill (1989) for a comparison of lter types.
Analogue lters are available as integrated circuits. These have better
characteristics. The twin-T, for example, can provide a 60 dB notch.
The roll-o is only 10 dB at
1
2
f
notch
and 2f
notch
and 3 dB at
1
4
f
notch
and
4f
notch
.
If the system contains a microprocessor then a digital lter could
provide the optimum performance, removing tolerance and drift problems.
These more advanced lters can be made to follow any drift of the
interference frequency.
Figure 6.28. The twin-T lter.
References
Baher H 1991 Analog and Digital Signal Processing (Chichester: Wiley)
Beck M S and Plaskowski A 1987 Cross-correlation Flowmeters (Adam Hilger)
Bendat J S and Piersol A G 1986 Random Data Analysis and Measurement, Procedures 2nd
edition (Wiley)
Bleaney B I and Bleaney B 1965 Electricity and Magnetism 2nd edition (Oxford University
Press)
Brook D and Wynne R J 1988 Signal Processing: Principles and Applications (Edward
Arnold)
Cooley J W and Tukey J W 1965 An algorithm for the machine calculation of complex
Fourier series Math. of Computation 19 297330
Horowitz P and Hill W, 1989 The Art of Electronics 2nd edition (Cambridge University
Press)
Kodali V P 1996 Engineering Electromagnetic Compatibility (New York Institute of Elec-
trical and Electronics Engineers)
Meade M L 1983 Lock-in Ampliers; Principles and Applications (London: Peter Peregrinus)
Nicollian E H and Brews J R 1982 MOS Physics and Technology (New York: Wiley)
Ott H W 1988 Noise Reduction Techniques in Electronic Systems 2nd edition (WileyInter-
science)
Perls T A 1952 J. Appl. Phys. 23 674
Wilmshurst T H 1990 Signal Recovery from Noise in Electronic Instrumentation 2nd edition
(Adam Hilger)
Exercises
1. Distinguish between
(i) deterministic and random signals
(ii) noise and interference.
2. An ensemble of ten identical thermometers, each having a resolution of
0.1 8 C, is placed in a large, thermally insulated tank of water which is
continuously stirred. The tank is being steadily heated.
(i) What class of system is this?
(ii) If the tank were not being heated what would the classication be
then?
3. Asignal has a peak value of 2.0 V. Calculate its r.m.s. value if the signal is
(i) a sinusoid,
(ii) a saw-tooth,
(iii) a square-wave of 40% duty cycle.
In each case assume that the signal is unipolar and that the period is
20.0 ms.
4. The following diagram (gure E.2) shows the probability distribution
pv of the amplitude levels of a random signal vt.
Exercises 257
Figure E.2.
(i) Find the value of b in terms of V
0
so that pv is correctly nor-
malized.
(ii) Determine the mean and standard deviation of the signal.
(iii) Sketch the cumulative distribution function.
(iv) Estimate the probability that v
1
2
V
0
.
5. The auto-spectral density of a signal is shown below in gure E.5.
Determine the autocorrelation function.
Figure E.5.
6. One of the rotor bearings in a helicopter is suspected of having wear. The
normal output of a vibration sensor on the bearing housing is band-
limited white noise. If wear has occurred this will contain an additional
periodic signal of much smaller amplitude than the noise signal. Explain
how conrmation of wear might be achieved and an estimate obtained of
the signal amplitude.
7. A certain transducer is generating a sinusoidal signal of amplitude 0.1 V.
The signal is corrupted with band-limited white noise occupying the
band 0 to 200 rad s
1
and of amplitude 5 mW per rad s
1
. The frequency
of the sinusoidal signal is at the centre of the noise bandwidth.
Determine the auto-correlation of the signal and noise.
8. Why is twin core co-axial cable preferred to single core when connecting
microphones in public address systems?
9. A certain thermocouple produces 39.4 mVK
1
. Given that the signal
conditioning amplier has an input impedance of 10.0 M and a band-
width of 100.0 Hz estimate the minimum detectable temperature change
if the ambient temperature is 300 K.
10. Explain how averaging can increase the signal-to-noise ratio. Describe
both digital and analogue approaches to signal averaging. What is a
lock-in amplier?
11. The telemetry signal from a Hall eect probe is found to be corrupted
because of the arcing of an a.c. electric motor nearby. Describe the spec-
trum of this electromagnetic pollution and suggest some methods of
improving the relative signal strength.
12. A seismometer is placed by the track of a tram to monitor vibration
as the vehicle passes. The tram has an over-head a.c. supply and the
circuit return is through the metal rails. The seismometer consists of
an open-framed mass-spring unit and variable reluctance pick-up. A
single core co-axial cable carries the signal to a data logger placed on
the ground 10 m away. When the tram passes as well as the anticipated
seismic signal other random and deterministic signals are also present.
(i) Explain the origins of these unwanted signals.
(ii) Describe methods by which the relative signal strength might be
improved.
Exercises 259
Chapter 7
Novel developments in ow measurement
systems
7.1 Introduction
In the preceding chapters the design of instrumentation systems, new types of
sensors, signal recovery and processing have all been described. In this
chapter a specic eld of application is presented incorporating some of
these topics, particularly that of the processing of the recovered signal. It
is hoped that by examining a specic eld of measurement the reader will
appreciate the dierent approach taken when designing an overall system
for a specic task rather than the analysis and development of a component
element. This is the so-called top-down approach. Issues to be considered at
the outset of a design may include
.
the range of the variable to be measured,
.
the precision of the nal result needed,
.
the system block structure as described in section 1.3,
.
the numerical specication of each block in the structure,
.
the overall transfer function as described in section 1.4,
.
the contributory and overall error estimates and uncertainties,
.
the electromagnetic compatibility and other safety issues.
Flow measurement
The most commonly measured variables are mass, temperature, length, ow
rate, pressure and level in tanks and hoppers. Commercially, owmeasurement
is one of the most signicant. There are more than 100 dierent techniques used
to measure owrate. In the UKthere are over 200 suppliers of owmonitoring
equipment. The supply and maintenance of ow measuring devices is a major
industry. Excluding domestic owmeters, annual sales are typically,
UK 30 000 units
Western Europe 150 000 units
USA 220 000 units
Rest of the world 180 000 units
There are in the region of 15 million installed working meters worldwide.
Annual sales in the UK are worth 90 million, and worldwide sales are
estimated at 1000 million.
Fluid dynamics
To appreciate the need for the variety of signal processing techniques used in
modern owmeters it is helpful rst to understand the physical principles on
which these owmeters are based. The design of owmeters involves an
understanding of uid dynamics, a discipline founded in classical dynamics
and thermodynamics. The remainder of this section contains some terminol-
ogy and denitions used commonly in uid dynamics and the following
section briey describes the types of ow eld that are encountered.
Fluid particles
The concept of a uid particle is often used. This is not an atom or molecule
of the uid but rather some small uid element that remains identiably
intact along the ow.
Viscosity
Unlike a solid, there are no shear forces in a uid at rest. A uid can be
dened as that state of matter which cannot sustain a shear force when at
rest. It has been observed experimentally that all uids, however small
their viscosity and even those that do not wet the boundary wall, have a sta-
tionary surface or boundary layer adjacent to the wall. If another layer of
uid slides against this boundary layer, friction occurs because of the viscos-
ity of the uid. This friction causes a shear force across some uid element as
shown in gure 7.1.
The boundary wall need not be involved. Shear forces occur whenever
dierent layers of viscous uid slide against each other. For ow in one
dimension the shear force t between adjacent layers is given by
t j
du
dy
7.1
Figure 7.1. Shear force in one-dimensional ow.
Introduction 261
where du,dy is the velocity gradient at the point being considered and j is the
absolute or dynamic viscosity. If j is independent of du,dy the uid is said to
be Newtonian.
Pathline
This is the trajectory taken by a uid particle.
Streamline
This is a line drawn at an instant in time through the uid showing the local
direction of the ow at all points up and down the stream. A ne chord xed
at an up-stream point would indicate one streamline. Many such streamlines
can be constructed.
Stream tube
This is a closed surface formed at any instant in time by joining a set of adja-
cent streamlines. In a closed conduit the boundary wall is one of a set of
stream tubes that can be drawn. Fluid cannot, by denition, move through
the surface of a stream tube, making it a useful theoretical concept.
Steady ow
When the velocity and uid properties at any given location are independent
of time the ow is said to be steady. The streamlines no longer change posi-
tion with time. Pathlines and streamlines become identical.
Discharge Q
This is the volume of uid passing through a cross-section per second and is
measured in m
3
s
1
.
Mean velocity " uu
This is the discharge Q divided by the cross-sectional area A at any section
and is measured in ms
1
.
Uniform ow eld
If the velocity is constant throughout the ow, neither time-dependent nor
spatially-dependent, the ow is said to the uniform. It is possible to have
steady non-uniform ow. An example of this is the steady ow of a uid
through a tapering tube.
The equation of continuity
The equation of continuity is derived from the concept of the conservation of
mass. Once transients have died away the mass ow rate must be constant
262 Novel developments in ow measurement systems
everywhere so in simple terms,
,Q constant 7.2
or
,A" uu constant. 7.3
The general form of the equation of continuity for three-dimensional, time-
dependent analysis can be expressed as
0,
0t
0,u
0x

0,v
0y

0,w
0z
0 7.4
where u, v and w are the velocity components in the x, y and z directions
respectively. See Fox et al (2004). For the special case of an incompressible
uid , constant, so equation (7.4) becomes
0u
0x
0v
0y
0w
0z
0. 7.5
For steady ow there is no time-dependent variation of any uid property.
0,u
0x

0,v
0y

0,w
0z
0. 7.6
Bernoullis theorem
Daniel Bernoulli published his work on uid dynamics in 1738. Assuming
that the uid is ideal, meaning incompressible and having zero viscosity,
and making the further assumption that the uid ow is steady and irrota-
tional Bernoullis theorem can be stated as
The sum of the pressure, plus the kinetic energy per unit volume,
plus the potential energy per unit volume at a point on a streamline
is constant
The theorem can be conrmed by considering the mechanical energy changes
experienced by a uid particle owing within some stream tube and assuming
no energy loss. From gure 7.2, as the uid ows through cx
a
at position a
work is done,
F
a
cx
a
P
a
A
a
cx
a
P
a
A
a
u
a
ct 7.7
where ct is the time taken to traverse cx. Similarly, owing through cx
b
at
position b,
F
b
cx
b
P
b
A
b
u
b
ct. 7.8
P
b
is the pressure at the stream tube exit and so is acting to oppose the uid
motion hence the negative sign. The total work done is therefore
E P
a
A
a
u
a
ct P
b
A
b
u
b
ct. 7.9
Introduction 263
This net work done is equivalent to the change in potential and kinetic
energies of the uid elements.
Change in kinetic energy
1
2
m
b
u
2
b
1
2
m
a
u
2
a
1
2
,A
b
cx
b
u
2
b
1
2
,A
a
cx
a
u
2
a
7.10
Change in potential energy m
b
gh
b
m
a
gh
a
,A
b
cx
b
gh
b
,A
a
cx
a
gh
a
. 7.11
Adding equations (7.10) and (7.11) and equating to equation (7.9), with the
provision that for an incompressible uid A
a
u
a
A
b
u
b
, gives
P
a
,gh
a
1
2
,u
2
a
P
b
,gh
b
1
2
,u
2
b
. 7.12
This can be stated more generally that, for any point on a streamline,
P ,gh
1
2
,u
2
constant 7.13
which is the Bernoulli equation. A more rigorous treatment involves Eulers
equation; see Fox et al (2004).
P is sometimes referred to as the static pressure and the term
1
2
,u
2
as the
dynamic pressure. Contrary to popular intuition, equation (7.13) states that
for horizontal ow if the pressure decreases the velocity increases.
This simple relationship can be used to explain many common phenom-
ena such as the pressure drop inside a railway carriage on entering a tunnel,
aerodynamic lift, the trajectory of a spinning ball and the bow wave
produced in front of a moving ship. It is also the basis of the operating
principle of constriction-type owmeters such as the orice plate and
Venturi tube and of the Pitot-static tube; see Bentley (1995).
Figure 7.2. Energy changes for uid elements in a stream tube.
7.2 Types of ow eld
Sometimes the smoke from a cigarette appears to show two distinct types of
ow. Near the source the smoke column appears smooth and regular.
Farther away sections within the ow become separately identiable, the
column becoming irregular, twisting and turning.
It was Osborne Reynolds who investigated these two types of ow in
some detail. He did this by injecting dye into water owing in a transparent
pipe. At relatively low ow velocities the dye streak follows a line parallel to
the pipe wall. This is termed laminar ow. At higher ow velocities the
particles have random, uctuating velocities. This is termed turbulent ow.
In 1883, Reynolds noted that the change between ow types did not just
depend on the ow velocity, but on the empirically derived expression,
Re
ud
j
7.14
where d is the pipe diameter and j the kinematic viscosity (dened as
j j,,) and Re the Reynolds number. Although the Reynolds number
was devised for ow re gimes in pipes it can be adapted for other ow
applications. It is in eect a measure of whether inertial forces or viscous
forces dominate within a particular ow eld.
7.2.1 Laminar ow
Laminar owis said to occur when the uid particles move in layers or laminae.
Fluid particles close to a large plate will move in a plane parallel to the plate
surface and parallel to each other. Their velocity will depend on their distance
fromthe plate. Ina circular pipe suchlaminae forma series of co-axial cylinders
extending from the pipe axis to the wall. This is sometimes referred to as
Poiseuille ow. To obtain laminar ow the Reynolds number must be less
than about 2300. When viscous forces dominate the ow is usually laminar.
The velocity prole in a circular pipe can be determined from consider-
ing the shear forces. Assuming equilibrium exists then the pressure gradient
will have reached a steady value. The cylindrical section shown in gure 7.3
illustrates the relationship between pressure gradient and shear force.
At equilibrium the force provided by the pressure gradient across the
faces of the cylinder is equal to the force from the sum of the shear forces
on the cylinder sides,
Pr
2
2r x t. 7.15
Equation (7.15) can be combined with equation (7.1) to give
du
dr

r
2j
P
x
. 7.16
Types of ow eld 265
Assuming a short cylinder so that x !dx, on integration over the pipe
radius R with the boundary condition that u 0 at the wall, equation
(7.16) yields
u
R
2
4j
_
dP
dx
__
1
_
r
R
_
2
_
. 7.17
Since the ow prole is axially symmetric and the maximum velocity occurs
on the axis this can be re-written as
u U
max
_
1
_
r
R
_
2
_
. 7.18
For an axially symmetric ow re gime the mean velocity can be generally
determined using
Q A" uu
_
R
0
ur2r dr. 7.19
For laminar ow this yields the useful relationship " uu
1
2
U
max
.
7.2.2 Turbulent ow
If the Reynolds number for the pipe is in excess of about 2300 then turbulent
ow occurs. When inertial forces dominate the viscous forces the ow is
usually turbulent. Fluid particles move randomly. Streamlines constantly
change their conguration. Equation (7.1) does not apply when the ow is
turbulent. Theoretical analysis of turbulent ows is dicult and formulae
are often semi-empirical. For many turbulent ow re gimes within pipes the
time average velocity ^ uu at some point r from the tube axis is unchanging.
Fox et al (2004) state that ^ uu can then be represented by
^ uur U
max
_
1
r
R
_
1,m
7.20
where m depends on the Reynolds number. This generally models the
empirical behaviour except at the axis and close to the wall. Assuming the
Figure 7.3. Shear force and pressure gradient.
prole is valid from r 0 to R then equation (7.20) gives
" uu U
max
_
2m
2
m12m1
_
. 7.21
For fully developed turbulent ow often m 7 is taken as a representative
value. Equation (7.21) then gives " uu 0.82U
max
. Expressing equation (7.20)
in terms of " uu gives
^ u ur " uu
_
m12m1
2m
2
__
1
r
R
_
1,m
. 7.22
Generally, if the Reynolds number is close to 2300 the type of ow can uc-
tuate between laminar and turbulent. However, experiments have been
carried out where, using very smooth surfaces and isolating all disturbances,
laminar ow has been maintained even when Re has reached 100 000. But for
common use the value of 2300 is a reliable indicator for the expected ow
re gime in a pipe.
7.2.3 Non-steady ow
The theoretical modelling of non-steady ow is complex. Womersley (1957)
developed a theory for oscillatory ow and pulsed transmission. The study
was for the analysis of blood ow in mammalian arteries and was not the
rst attempt at a theoretical description for this type of ow.
Womersleys paper begins with an analysis of a viscous liquid owing
in a straight rigid circular tube. He assumed no radial motion so that u is
not a function of the position along the tube axis x but only of the radial
distance r from the axis. This infers that the pressure gradient 0P,0x is
also independent of x. The acceleration is given as
0u
0t

1
,
0P
0x
j
_
0
2
u
0r
2

1
r
0u
0r
_
. 7.23
In the rst section of his paper Womersley used an oscillatory pressure
gradient of the form
0P
0x
Acos.t c. 7.24
The solution to equation (7.23) then involves Bessel functions. Using
modulus and phase forms M
0
0
and
0
he obtained an expression for u of
the form
u
AR
2
j
M
0
0
c
2
sin.t c
0
7.25
Types of ow eld 267
where R is the tube radius and c is a non-dimensional parameter. The value
of c is said to be useful in characterizing the ow. It is given by
c R
.
j
_
. 7.26
The ow velocity is reduced by the viscosity but also depends on the value of
c.
In the following sections of his paper Womersley analysed ow in a
freely moving elastic tube and then went on to develop modications for
additional mass, elastic constraint and internal viscous damping. In the
nal section of the paper he discusses the issue of having a thin boundary
layer of lower viscosity. This was considered to be a signicant issue for
blood owing in small vessels. For a fuller description of the mechanics of
blood ow see Evans et al (1989).
Flow regimes and owmeters
Laminar ows have a lower resistance to ow and so are preferred in some
situations such as long pipelines in order to give greater pumping eciency.
In heat exchangers and with oxygen transfer in blood ow, turbulence gives
good mixing and so improves the transfer process. Some owmeters will not
function correctly in turbulent ows so precautions must be taken with these.
Others only work well in turbulent ows because turbulence reduces the
eects of viscosity. Also, for turbulent ow the variation in m only has a
small eect on the estimation of " uu from U
max
, whereas upsets to a laminar
ow can give errors of 30%. Flowmeters which can work in either type of
ow are usually calibrated for one or the other which means that if the
ow re gime changes erroneous results occur.
Many owmeters will not function correctly in swirling or pulsating
ows. For such owmeters manufacturers specify installation requirements
such as having long straight runs of pipeline before and after the meter
and inserting vane structures into the ow.
7.3 The cross-correlation owmeter
The correlation technique has been used in applications where the more
traditional owmeters cannot operate well. The concept of correlation
was described in section 6.3.4. In this context it is used to determine the
ow velocity. Cross-correlation owmeters have been developed for applica-
tions where conventional owmeters will not function or would suer
damage. Velocity measurement has been achieved during the conveying of
materials such as molten chocolate, slurries, powders, textiles and the ow
of sewage.
Two sensors are used, spaced at intervals along the velocity direction as
illustrated in gure 7.4. Any type of sensor can be used as appropriate for the
ow application.
The essential requirement is that the process has some randomly uc-
tuating parameter such as temperature, conductivity, opacity, electrostatic
charge, dielectric constant or density. Ideally, the local value of this
parameter should not change signicantly between the sensors so that the
same random signals v
a
and v
b
are produced at both sensors but simply
time-shifted. This frozen pattern is illustrated in gure 7.5.
Using the cross-correlation procedure on the signals v
a
and v
b
a peak in
the cross-correlation function R
v
a
v
b
occurs where the variable t is equal to the
time of travel T between sensors (gure 7.6).
The velocity is determined using L,T. Obviously, large values of T and
L would tend to give a more precise estimate of the ow velocity. However,
the height of the correlation peak is reduced as the correlation between the
two sensor signals deteriorates. This sets a limit on the sensor spacing L.
Early cross-correlation systems were analogue and consequently
prone to zero drift. The preferred approach is to use digital technology.
The correlation algorithm is derived from that given in section 6.3.4 repeated
below,
R
v
a
v
b
qt
1
N
N
j 1
v
aj
v
b j q
6.44
where t is the clock period. The time shift t is q t. The number of sampling
points is N for each of the time windows for v
a
and v
b
. Consequently, there
are N calculated values for the function R
v
a
v
b
. This number, in conjunction
with the clock speed, has to be sucient for the correlation to occur, i.e.
Figure 7.4. Sensors spaced along some velocity eld.
Figure 7.5. Sensor output signals.
The cross-correlation owmeter 269
N t T. This implies that a larger clock period is preferable for lower
velocities otherwise large allocations of memory are needed and the
algorithm will take longer to run. This conicts with system performance.
Beck and Plaskowski (1987) give a relationship for the fractional velocity
resolution as
u
u

t
2L
u. 7.27
Larger clock periods mean poorer resolution. Also, resolution is worse at
higher velocities. In designing the digital system the following issues need
to be considered:
.
What is the maximumvalue of Lthat can still give a correlation peak above
noise in the correlelogram?
.
What is the dynamic range of velocity likely to be encountered?
.
What resolution is desired for the maximum velocity?
It is interesting to note that the wider the bandwidth of the random para-
meter that is being recorded in the signals v
a
and v
b
the lower the level of
the background noise in the function R
v
a
v
b
. This means that the more
random the ow parameter providing the signals v
a
and v
b
the better the
performance of the system.
7.4 Ultrasound
Although ultrasound is dened as sound beyond the human range of about
20 kHz, for ow measurement frequencies in the region of 1 MHz are
commonly used. This is in order to achieve spatial resolution of about
1 mm. For any wave-based system resolution cannot be greater than the
wavelength of the wave employed. Because the speed of mechanical vibra-
tions in solids and aqueous liquids is of the order 3 kms
1
and 1.5 kms
1
respectively the required small wavelengths mean correspondingly high
frequencies. This is inconvenient in that air is opaque to ultrasound at
Figure 7.6. The cross-correlelogram of v
a
and v
b
.
these frequencies. Consequently, a coupling gel or compound has to be
used between the transducers and the target. Also, any large bubbles or
gas cavities in the beam path will obscure the measurement. (This is a
useful feature long-used for the detection of aws in welds and castings.)
Ultrasonic transducers
A number of materials have been used for transducers. Due to the high
frequencies involved materials having piezoelectric properties are found to
be the most appropriate. These include quartz, ceramics such as lead
zirconate-titanate, lead metaniobate and barium titanate and the plastic
lm polyvinylidene uoride (PVDF). Quartz is less commonly used now.
PVDF has a closer mechanical impedance match to aqueous media than
do the other materials enabling better coupling; see section 1.6. (The paper
by ODonnell et al (1981) includes data on various piezoelectric materials.)
Because of this and its exibility, PVDF could be used for direct contact
with the human body without the need for coupling gel. Unfortunately, it
is a relatively weak transmitter compared with ceramics operated at
resonance. Consequently, ceramics are the most commonly used for these
high-frequency transducers. The coupling medium used to exclude air
between the transducers and target can also act as an impedance matching
device.
For ultrasonic applications transducers are used in either pulsed mode
or continuous working (CW) mode. The equivalent circuit of a piezoelectric
transducer is shown in gure 7.7. In section 2.2.2 the concept of the driven
oscillator was described and in section 2.2.3 the Q factor. As mentioned
there, high-Q devices have the benets of narrow bandwidth, low loss per
cycle and greater immunity to loading eects. For CW systems, having a
stable narrow bandwidth is useful when performing frequency analysis on
echo signals as described later. However, in pulsed systems high Q causes
ringing of the transducer after the end of the electrical pulse because of
the low loss. This reduces the time resolution. It is possible to alter the
mechanical Q of the transducer by loading the output signal with electrical
Figure 7.7. The equivalent circuit of a piezoelectric transducer.
Ultrasound 271
impedance. In section 1.5 the analysis of the dynamic response of a system
was described. In this way the piezoelectric element of gure 7.7 with an
additional electrical impedance network could be modelled and the resultant
pulse response simulated in a software package such as Matlab
1
. A minor
disadvantage to reducing the Q would be some loss of signal amplitude.
Time-of-ight owmeters
The measurement of ow using ultrasound has become an established
technique in the process industries (gure 7.8). The time-of-ight approach
uses separate transmit and receive transducers within the ow mounted on
either side of the pipe. Acoustic coupling with the liquid is not a problem
and the use of preformed housings means that the angle of incidence is
precisely known.
Ultrasound is transmitted alternately P to Q and then Q to P. The
velocity u varies along the length l according to some prole. Consequently,
to correctly determine T the mean value of velocity " uu must be used.
Assuming a pencil ray of ultrasound, " uu across the pipe is determined using
" uu
1
R
_
R
0
ur dr. 7.28
The function ur chosen will depend on the ow type. Equation (7.18)
would be used for laminar ow or equation (7.21) for turbulent ow. The
propagation times are
T
PQ

l
c " uu cos 0
7.29
and
T
QP

l
c " uu cos 0
. 7.30
Combining gives
T
QP
T
PQ

l
c " uu cos 0
l
c " uu cos 0
7.31
T
2R
sin 0
_
1
c " uu cos 0
1
c " uu cos 0
_
7.32
Figure 7.8. The time-of-ight owmeter.
where 2R is the pipe diameter. Usually the speed of the ultrasound c ) " uu, so
T
4R" uu cot 0
c
2
. 7.33
For measurement of " uu to a resolution of 1% in a pipe of diameter 0.1 m with a
ow velocity of 1 ms
1
requires the measurement of T to a precision of
<1 ns. The development of circuitry to deliver this is not a trivial task
whether done as a true timing system, or as a phase-shift measurement of
the received signal compared with that transmitted, or as a correlation
system using a pseudo-random binary sequence. To avoid this diculty
the rst ultrasonic owmeters marketed were of the so-called sing-around
type; see Sanderson and Hemp (1981). This is based on frequency rather
than timing circuitry. Here the upstream and downstream signal times are
used to provide the time periods for two separate oscillators. The dierence
frequency of the two oscillators is
f
" uu sin 20
2R
. 7.34
A further advantage of this technique is that the value of c is not required.
Consequently, any drift due to temperature or change in liquid composition
is not an issue. However, bubbles in the ow can cause loss of signal forcing
the signal processing circuitry to re-start. Scott (1984) compares the
performance of the two types under calibration.
7.5 The Doppler eect
Relative motion between a transmitter and receiver gives the well-known
Doppler frequency-shift eect. It is possible to measure ow velocity non-
invasively by transmitting the ultrasound into the ow eld and then
examining the frequency shift of the echo. For the system to work well
there must be scattering particles within the uid. The frequency is
Doppler shifted twice. Firstly, the scatterers receive a Doppler-shifted fre-
quency f
1
as moving observers of a stationary transmitter of frequency f
0
,
f
1
f
0
_
1
v
c
_
7.35
where v is the velocity of the scatterer and c the velocity of the sound in
the medium. This is then Doppler-shifted to f
2
since the scatterers now
behave as moving transmitters emitting back to the stationary receiver
outside the pipe. The Doppler shift for a moving transmitter and stationary
receiver is given by
f
2

f
1
c
c v
7.36
The Doppler eect 273
which is slightly dierent from equation (7.35). This is because in the rst
case the transmitter and medium are both stationary. In the second situation
the transmitters are moving through the medium. Combining both Doppler
shifts gives
f
2

_
c v
c v
_
f
0
. 7.37
If the ultrasonic beam is inclined at some angle 0 to the ow then v is re-
written as v cos 0. Also, since usually c )v the dierence frequency can be
approximated to
f 2f
0
v cos 0
c
. 7.38
For example, a transmitter operating at 10 MHz, mounted at 308 to the ow
axis, in an aqueous ow with particulates moving at a velocity of
0.9 10
2
ms
1
would produce a frequency shift of about 100 Hz.
The signal processing of Doppler echo systems is often achieved using
techniques developed by telecommunication engineers. The Doppler echo
signal corresponds to that of a single-sideband transmission system of
carrier frequency 10 MHz modulated at 100 Hz. A simple way of recovering
the 100 Hz from the 10 MHz is to multiply the echo by the original transmis-
sion frequency. This can be readily achieved using analogue multiplication.
From the trigonometrical identity,
cos Acos B
1
2
cosA B
1
2
cosA B 7.39
cos 2f
1
t cos 2f
2
t
1
2
cos 2f
1
2
cos 2f
1
f
2
t. 7.40
The high-frequency term f
1
f
2
20 MHz can be easily removed by lter-
ing. Analogue multipliers are readily available in integrated circuit form.
Some early industrial systems using a non-invasive, Doppler approach
did not live up to the claims promised by the instrument manufacturers.
Users encountered two problems. Firstly, the attachment of the transducers
to the outside of the pipe using gel or compound meant that sometimes
the angle of incidence of the ultrasound beam with the pipe axis was not
that determined during calibration. This was due to either poor installation
or because of additional refraction caused by scaling on the internal pipe
walls. Also, transverse waves can be established in solids whose velocities
are dierent to those of longitudinal waves. Using the wrong velocity
aects the calculated angle of refraction of the wave entering the ow.
Secondly, if the ow re gime changed, as described at the end of section
7.2, estimation of the mean velocity from the maximum Doppler-shifted
frequency was incorrect. Gessner (1969) estimated the errors that might be
expected when proles changed. Cousins (1978) published calibration
results for various sizes of pipe and ow re gimes. Interrogation of the
prole by multiple ultrasonic beams has been proposed as a means of
obtaining a more reliable estimate of the mean velocity. Another approach is
the use of a pulsed or gated beam. In this way the mean velocity for a small
zone of the uid can be determined and hence the mean velocity along the
line of the beam for the whole. However, pulsed systems require more
power and are limited in spatial range and maximum ow velocity, see
Sanderson and Hemp (1981).
7.5.1 The echo spectrum
Another approach to the problem of uncertain velocity proles is to perform
a frequency analysis on the echo signal. In this way an estimate of the range
of velocities and, from the intensities of the spectral lines, the velocity distri-
bution can be made. This procedure has become very popular in clinical ow
measurement. Frequency spectra of the type shown below in gure 7.9 are
easily produced using the FFT on the low frequency (demodulated) echo
signals f . The symbol f
D
is used to denote f in the text from here.
WORKED EXAMPLE
A certain signal has a bandwidth of 010 kHz. A frequency analysis is
required and a FFT is to be used. If the frequency resolution is to be
about 80 Hz determine the required sampling rate of the input signal, the
length of the time window and the size of the FFT needed.
According to the sampling theorem described in section 5.2, the sam-
pling rate must be at least twice the highest frequency present to avoid alias-
ing. (A system designed to accept 10 kHz would have an analogue input lter
to prevent higher frequencies from entering. There would also be a lter for
enhancement to reduce the eects of windowing.) Therefore, the sampling
rate is 2 10
4
s
1
.
The length of the sampling window determines the frequency resolution.
In section 5.2 it was shown that the eects of sampling produced a sinc func-
tion in the frequency domain. A similar eect occurs with the whole sampling
window. The Fourier transform of the time window produces a sinc function
that determines the sharpness of the spectral lines. The width of the fre-
quency domain sinc function is approximately 1,T
W
where T
W
is the
window time. For the spectral lines to be separated so as to give a frequency
resolution of 80 Hz the time window must be about 12.5 ms or larger. The
sampling rate is 2 10
4
s
1
so during this time 250 samples will be recorded
for use by the FFT. However, the FFT algorithm requires that the number of
input points be a highly composite number of 2
N
; see Rabiner and Gold
(1975). So, in reality 256 (2
8
) samples must be taken. This 256-point FFT
will generate 128 unique frequency points in the output (a further 128
conjugate points are also produced that can be discarded).
The FFT sampling time will thus be 256,2 10
4
s giving a value for the
time window of 12.8 ms. From above, the constraint of choosing 2
N
means
that the actual frequency resolution will be about 78 Hz.
The mean frequency of gure 7.9 can be simply determined using
"
ff
D

_
1
0
, f
D
f
D
df
D
_
1
0
, f
D
df
D
N
i 0
w
i
f
Di
N
i 0
w
i
7.41
where , f
D
is the intensity distribution and w
i
are the weightings of the
discrete form of a spectrum of N frequency points. If the assumption is
made that the intensity of a received frequency component is simply propor-
tional to the number of scatterers travelling with the same velocity the mean
velocity can be estimated. Using equation (7.38),
"
ff
D

N
i 0
w
i
f
Di
N
i 0
w
i
2 f
0
cos 0
c
" uu. 7.42
Brody (1974) states that, under these idealized conditions, knowing the pipe
cross-sectional area A equation (7.42) should give the volume ow rate Q
using Q A" uu and that this does not assume a priori knowledge of the ow
prole.
However, estimation of the mean velocity is not quite as straightforward
as it might rst appear. Figure 7.10 for a laminar ow re gime illustrates the
diculty of the problem. In a more rigorous analysis the following issues
have to be considered:
.
the eect of having a non-parallel beam,
.
uncertainty in the angle of insonation,
.
the eects of multiple scattering,
Figure 7.9. A typical frequency spectrum from a Doppler owmeter.
.
the distribution of the scatterers,
.
the attenuation of the beam through the ow along the transmit and echo
paths.
Angelsen (1980) has developed a theoretical analysis. Evans et al (1989) also
describe work done in the eld.
However, notwithstanding the theoretical diculties of developing a
model for the scattering process, it has been found that the frequency
spectra can be empirically related to the mean velocity and good repeatability
has been reported.
7.5.2 Forward and reverse ow
A further very useful feature that can be obtained from the Doppler-shifted
echo spectrum is the identication of uid moving toward or away from the
transducer system. In the process industries this is not a common require-
ment. However, in the medical eld it can be extremely useful. For
example, reverse blood ow can occur around leaky heart valves.
The echo signal can be represented as
v
e

i
A
i
cos.
0
t c
i

j
F
j
cos.
0
.
F
j
t
k
B
k
cos.
0
.
B
k
t
7.43
where A
i
is the amplitude of an echo of the original transmitted frequency
.
0
, F
j
is the amplitude of the echo received from scatterers moving
towards the receiver and B
k
the amplitude of the echo from a scatterer
moving backwards. In practice the echo will be a continuous spectrum but
performing the FFT as described above will generate discrete spectral lines.
As already described, obtaining the frequency-shifted components is
relatively straightforward if only the frequency shift is required. In order
to determine whether the frequency is up- or down-shifted more complex
signal processing is necessary. Nippa et al (1975) list some options. These
are considered below. If the ow velocity is in the range 0.9 10
2
ms
1
to 0.9 ms
1
then for a 10 MHz system the frequency shift will be between
Figure 7.10. Insonation of a velocity prole.
100 Hz and 10 kHz. The spectrum is illustrated in gure 7.11, although not to
scale.
(1) Separation by ltering directly
It might be thought that simple ltering of the incoming echo spectrum
would be a suitable approach. The received components will be 10 MHz,
10.000110.001 MHz and 9.99999.99 MHz. However, as Nippa et al
(1975) point out the separation between frequencies in the region of
10 MHz and 10.0001 MHz to the required 40 dB is an impossible task
using lters especially when considering that the transmit frequency will drift.
(2) Frequency shifting
Down-shifting of the Doppler spectrum means that the performance
required of the lters becomes less demanding. Frequency shifting is a
common telecommunications procedure. For example, the composite
stereo signal in the UK, VHF-FM radio broadcasting uses frequency shifting
to make better use of the transmission bandwidth.
Frequency shifting is again achieved by a multiplication procedure. The
procedure used here is referred to as heterodyning by radio engineers. A
frequency .
m
, derived from the transmission frequency but slightly lower,
is multiplied with the echo signal and sum and dierence terms are generated
in the usual fashion. The frequency used for multiplication, .
m
, must be such
that after the multiplication the dierence term places the echo bandwidth in
a suitable range in the lower part of the frequency spectrum (gure 7.12).
A phase-locked loop technique can be used to generate .
m
; see Horowitz
and Hill (1989). The value of .
m
can be expressed as
.
m
.
0
.
het
7.44
where .
het
is generated by a xed, low-frequency oscillator. Because .
m
is
derived form .
0
any drift in .
0
will not cause the recovered signals to drift.
Clearly, .
het
must be greater than the highest Doppler frequency expected.
Figure 7.11. Echo spectrum for forward and reverse ow.
After ltering out the high frequencies there will be two spectral bands,
.
het
j
.
F
j
and .
het
k
.
B
k
and the spectral line .
het
.
A further very sharp, notch lter could be used to remove .
het
but a
modern implementation of this technique would be to use dsp at this stage
rather than analogue techniques. Performing a FFT on the now low-
frequency signal would give the spectrum directly and .
het
ignored.
(3) Phase rotation
Because of the demands placed on analogue lters by the two previous
methods the major part of the paper by Nippa et al (1975) reports work
on a phase-shifting system. The technique is similar to that of phase-
quadrature detection (gure 7.13) used in telecommunications engineering.
It involves two, precise, 908 phase-shifting elements as shown below.
For convenience a single velocity component from the echo spectrum of
equation (7.43) is used for illustration,
v
e
Acos.
0
t c F cos.
0
.
F
t Bcos.
0
.
B
t. 7.45
Figure 7.12. Echo spectrum for forward and reverse ow.
Figure 7.13. Phase-quadrature detection.
Multiplying the echo signal by the phase-shifted transmission frequency
Dcos.
0
t ,2 gives
V
A
Dcos
_
.
0
t
2
_
Acos.
0
t c Dcos
_
.
0
t
2
_
F cos.
0
.
F
t
Dcos
_
.
0
t
2
_
Bcos.
0
.
B
t. 7.46
Using the trigonometrical identity of equation (7.39) and ltering to remove
high frequency and d.c. components gives,
V
A

1
2
DF cos
_
2
.
f
t
_
1
2
DBcos
_
2
.
B
t
_
7.47
or
V
A

1
2
DF sin .
F
t
1
2
DBsin .
B
t. 7.48
But the signal V
A
is further phase-shifted by 908 so equation (7.48) becomes
V
0
A

1
2
DF sin
_
2
.
F
t
_
1
2
DBsin
_
2
.
B
t
_
7.49
which on simplifying gives
V
0
A

1
2
DF cos .
F
t
1
2
DBcos .
B
t. 7.50
Correspondingly, multiplying the echo signal by the transmission frequency
Dcos .
0
t gives
V
B
Dcos .
0
tAcos.
0
t c Dcos .
0
tF cos.
0
.
F
t
Dcos .
0
tBcos.
0
.
B
t. 7.51
Simplifying and ltering as before, this reduces to
V
B

1
2
DF cos .
F
t
1
2
DBcos .
B
t. 7.52
The outputs are then
V
0
A
V
B
DF cos .
F
t 7.53
V
B
V
0
A
DBcos .
B
t. 7.54
Two important constraints for the system to work well are:
.
The amplitudes DBin the signals V
0
A
and V
B
must be of the same magnitude
for the addition and subtraction procedures of equations (7.53) and (7.54)
to work correctly. Similarly for the amplitudes DF. This will require some
form of adjustable gain amplier within the system. The signals in the
system by Nippa et al (1975) diered by less than 0.2 dB.
.
The two, 908 phase-shifters must perform well over the range of frequencies.
The high frequency phase-shifter does not have a relatively large frequency
spread so is the less demanding to design. The low frequency second phase-
shifter covers a large range. According to Nippa et al (1975) the design used
in their system was an eight-pole, transistor-based lter which achieved
908 0.68 over their range of 50Hz to 7.5 kHz. A circuit published by
Dickey (1975) uses op-amps to generate a 908 phase-shift for the range
100 Hz to 10kHz.
Because of the advances of dsp, in a modern implementation the low fre-
quency sections of the system, ltering, gain-adjust, phase-shifting, addition
and subtraction would be processed digitally. Digital systems are more
prescriptive to design. They are also very stable in operation because
performance is not dependent on component values unlike analogue
systems which drift with age and temperature.
7.5.3 Blood ow measurement
The measurement of blood ow-rate is of signicant importance in a number
of clinical specialisms. However, making accurate, non-invasive measure-
ments of true volume ow-rate is not straightforward. Some clinical areas
where this would be useful are given below.
.
In order to assess the performance of the heart the blood volume ow-rate
is required. Currently, the dilution method is routinely used; see Khandpur
(1987). Here iced water is injected into the artery and the resultant tempera-
ture change used to determine the dilution and hence the volume. Clearly,
as with any invasive procedure, it is distressing to the patient and not
without risk.
.
In the development of the unborn baby, as the child grows the umbilical
supply has to increase to provide a greater supply of oxygen and nutrients.
If the chord is impaired the child causes the blood pressure of its mother to
increase in an attempt to obtain the desired supply rate. This high blood
pressure is a sign of the condition know as pre-eclampsia and can be dan-
gerous to the mother and baby. The use of a simple ultrasonic probe can
detect velocities but not total volume ow rate.
Some areas of blood ow measurement do not require total volume ow rate
but only some indication of change in ow velocity prole.
.
The partial blockage caused by a thrombosis results in much higher ow
velocities near the obstruction. A simple, hand-held ultrasonic transducer
with audio frequency output can be used to locate the site of the blood clot.
.
The growth of a tumour is marked by a stage where, in order to sustain
growth, a vascular system within the tumour has to develop. Wells et al
(1977) reported work on Doppler-shifted signals arising from microcircu-
lation within malignant breast tumours. The structure of new vessels in
tumours is dierent from normal tissue in that diameters are much
larger, walls are thinner and lack contractile elements. Burns et al (1982)
reported that Doppler-shifted spectra from blood owing in and near a
breast tumour are characteriztically dierent and that this might be a
useful diagnostic procedure.
Ultrasonic imaging systems are now very well developed. Duplex systems not
only produce an image but can also perform Doppler-shift measurements on
the imaging zone at a location set by a cursor superimposed on to the image
displayed on the monitor. Some duplex systems colour code the image so that
ow detected by the Doppler-shift appears as shades of red or blue in an
otherwise monochrome image. In addition green can be added as a function
of the variance of the signal as an indication of turbulence; see Namekawa
et al (1982). In this way clinicians can see whether ow is towards or away
from the probe and if turbulence is present green mixed with red or blue
gives shades of yellow or cyan respectively.
It might be thought that, with a sophisticated duplex imaging system, a
reliable estimate of the volume ow-rate should be possible by measurement
of blood vessel diameter and measurement of mean ow velocity using
Doppler-shift. Unfortunately, in addition to the problems of obtaining a
reliable mean velocity estimate from the echo as described above, a
number of other diculties exist:
.
The tube may not be circular.
.
The tube diameter may vary during systole and diastole.
.
The type of ow re gime may alter during the cardiac cycle so estimates of
mean velocity could be erroneous.
.
An estimate of the mean cross-sectional area and mean velocity over a
cardiac cycle will not give a true measurement of mean volume ow-rate
because the two are not linearly related. Attempts at instantaneous
measurement of both mean velocity and mean cross-sectional area are
dicult to coordinate due to signal processing limitations.
Many modern duplex scanners do have algorithms for calculating blood
volume ow rate and reasonable estimates can be achieved on vessels
between 4 and 8 mm in diameter (Evans et al 1989).
Alternatively, a number of estimators have become popular to quantify
ow in a way suitable to gain clinical information. The measurement of the
maximum frequency shift is relatively straightforward and can provide a
useful insight into ow abnormalities. Figure 7.14 shows the type of variation
that can occur over a cardiac cycle albeit for forward ow only. Mo et al
(1988) compared dierent methods of estimating the maximum frequency.
Although waterfall-type displays are sometimes used in research most
modern Doppler blood-ow analysers display the FFT frequency spectrum
of gure 7.9 as a sequence of vertically orientated frames. The location of
a sample frame is shown in gure 7.14. These displays are produced in
rolling format on the monitor and are referred to as sonograms. The intensity
information is in the z-axis direction (out of the page) and is shown as a
colour code by this type of analyser.
Understanding the data gathered becomes an issue of pattern recogni-
tion. Various algorithms have been invented over the years in an attempt
to automate the feature extraction process. Quantifying parameters in use
are:
.
the ratio S/D,
.
the pulsatility index,
S D
mean velocity
7.53
.
Pourcelots resistance index,
S D
S
. 7.54
To obtain a value for S some threshold has to be chosen above background.
Also, low frequency lters are used so that values of D are not inuenced by
extraneous vibrations. The mean velocity is estimated over the cardiac cycle.
This can be conveniently done as a rolling average using the mean frequency
from a FFT frame.
Conclusion
Although the measurement of ow has been carried out for millennia it still
attracts much research. This chapter has outlined some of the applications
of signal processing to the more challenging ow elds and illustrates the
Figure 7.14. Typical maximum Doppler-shifted frequency during a cardiac cycle.
expertise needed across the range of engineering physics to produce a
working device.
References
Angelsen A J B 1980 A theoretical study of the scattering of ultrasound from blood IEEE
Trans. Biomed. Engineering BME-27(2) February
Beck M S and Plaskowski A 1987 Cross-correlation Flowmeters (Adam Hilger)
Brody W R 1974 Theoretical analysis of the CW Doppler ultrasonic owmeter IEEE
Trans. Biomed. Engineering BME-21(3) May
Burns P N, Halliwell M, Wells P N T and Webb A J 1982 Ultrasonic Doppler studies of
the breast Ultrasound Med. Biol. 8(2) 127143
Cousins T 1978 The Doppler Ultrasonic Flowmeter: Flow Measurement of Fluids (Amster-
dam: North-Holland)
Dickey R K 1975 Outputs of op-amps have xed phase-dierence Designers Casebook,
Electronics, pp 8283, 21 August
Evans D H, McDicken W N, Skidmore R and Woodcock J P 1989 Doppler Ultrasound
Physics: Instrumentation and Clinical Applications (Wiley)
Fox R W, McDonald A T and Pritchard P J 2004 Introduction to Fluid Mechanics 6th
edition (Wiley)
Gessner U 1969 The performance of the ultrasonic owmeter in complex velocity proles
IEEE Trans. Biomed. Engineering BME-16(2)
Horowitz P and Hill W 1989 The Art of Electronics 2nd edition (Cambridge University
Press)
Khandpur R S 1987 Handbook of Biomedical Instrumentation (New Delhi: McGraw-Hill)
Mo L Y L, Yun L C M and Cobbold R S C 1988 Comparison of four digital maximum
frequency estimators for Doppler ultrasound Ultrasound Med. Biol. 14(5) 355363
Namekawa K, Kasai C, Tsukamoto M and Koyano A 1982 Real-time blood ow imaging
system utilizing autocorrelation techniques Ultrasound 82, ed Lerski and Morley
(New York: Pergamon)
Nippa J H, Hokanson D E, Lee D R, Sumner D S and Strandness D E jr. 1975 Phase
rotation for separating forward and reverse blood velocity signals IEEE Trans.
Sonics and Ultrasonics SU-22(5) September
ODonnell M, Busse L J and Miller J G 1981 Piezoelectric Transducers Methods of
Experimental Physics vol 19 (Academic Press)
Rabiner L R and Gold B 1975 Theory and Application of Digital Signal Processing
(Prentice-Hall)
Sanderson M L and Hemp J 1981 Ultrasonic owmetersa review of the state of the art
Int. Conf. Advances in Flow Measurement, Warwick University, 911 September,
sponsored by BHRA Fluid Engineering
Scott C 1984 Sounding out ultrasonic owmeters Control and Instrumentation August
Wells P N T, Halliwell M, Skidmore R, Webb A J and Woodcock J P 1977 Tumour
detection by ultrasonic Doppler blood-ow signals Ultrasonics September
Womersley J R 1957 An elastic tube theory of pulse transmission and oscillatory ow
in mammalian arteries Wright Air Development Centre, WADC Technical Report,
TR 56614
Exercises
1. When air is drawn into the lungs the pressure inside the trachea is found
to be about 2.5 Pa below atmospheric pressure. Given that the density of
air is about 1.25 kg m
3
estimate the velocity of air in the trachea. The
trachea divides into two, one to each lung. If each branch has 0.65 the
cross-sectional area of the trachea what is the velocity of air in a
branch? (Assume that the tube walls are rigid and that air is incompres-
sible under these conditions.)
2. A horizontal pipe has a cross-sectional area of A
1
and the liquid in this
section ows at a velocity v
1
. If the pipe reduces in diameter to A
2
with
a corresponding increase in ow velocity to v
2
, assuming ideal conditions
and that the liquid is incompressible show that
v
1

2P
,A
1
,A
2
2
1
where P is the pressure dierence between the two sections of pipe.

Hence show that the volume ow rate Q is given by
Q A
2
2P
,1 A
2
,A
1
.
3. Given that for laminar ow
u U
max
_
1
_
r
R
_
2
_
show that " uu
1
2
U
max
.
4. If the number of clock cycles necessary to reach a correlation in a cross-
correlation owmeter is q, show that
u
u

1
2q
.
5. In a certain correlation owmeter it is found that the maximum possible
separation of the sensors without losing the correlation is 0.45 m. If a
resolution of 2% is required in the velocity measurement, determine
(i) the clock period if the maximum velocity to be measured is about
0.35 ms
1
.
(ii) the lowest velocity that can be measured given that the memory
buers are all 256 locations long.
Exercises 285

Novel Sensors and Sensing

Uploaded by

Copyright:

Available Formats

Novel Sensors and Sensing

Uploaded by

Document Information

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

Novel Sensors and Sensing

Uploaded by

Copyright:

Available Formats

Chapter 1

The instrumentation system

where c, u and are all constants, derive an expression for the

ions remain in equal

), and the c-particle, a helium nucleus

regions to give a high injection

Given that the resonant frequency is 10

where P is the pressure dierence between the two sections of pipe.

You might also like