Measurement: 2.1 The Sensor Model
Measurement: 2.1 The Sensor Model
Measurement: 2.1 The Sensor Model
2.1 TheSensorModel
As depicted in Figure 2.1, a generic process sensor determines the current state of the
process from one or more physical observables, which by defnition are physical quanti-
ties that can be directly measured. Physical observables include temperature, frequency,
light intensity, speed, and so forth. Te sensor uses these inputs to produce a measure-
ment, which may be a simple number (e.g., the local temperature) or a more complex
data set such as an image.
If the measurement is to have any validity it must be true that a given set of physical
observable values must always produce the same measurement value; otherwise, the
measurement would be ambiguous. By taking a process measurement, we are assuming
that the quantity to be measured can be measured and that it depends uniquely on the
physical observables. Tis requirement is similar to the so-called zeroth law of thermody-
namics (which states that temperature exists as a measurable quantity), and we could even
call it the zeroth law of measurement. Tis general concept is described mathematically
by a function S, whose domain is the set of all possible values of the physical observables
and range is the set of all possible measurement values:
measurement S observables ( )
(2.1)
Te fact that a sensors output can be described by a function (at least in theory) follows
directly from the causal relationships linking the quantity of interest in the process, the
physical observable, and the resulting measurement.
Another requirement for making process measurements is that the function
(equation 2.1) must be continuous, which suggests that infnitesimal changes in the
physical observable can only cause infnitesimal changes in the measurement. Tis
requirement ensures that the measurement data is diferentiable and that the rate of
change in the measurement remains fnite.
A fnal requirement is that the measurement reported by the sensor must be linearly
dependent on the process quantity to be measured. Ideally, the output of the sensor
2
Neasuremenf
44168_C002.indd 7 8/27/07 2:52:52 PM
Copyright 2008 by Taylor and Francis Group, LLC
S Industrial Process Sensors
(denoted as s) gives the exact value of the quantity (q) to be measured, but in practice
there is ofen a small discrepancy given by
s Aq b + (2.2)
where A is a gain factor and b is a constant ofset. An important function of calibration
is to adjust the sensors electronics so that A 1 and b 0, thus ensuring that s q. In
most process control applications it is not necessary to get the calibration exact, because
the controller usually bases its corrective actions on changes in the measurement rather
than on the absolute value of the measurement.
Te sensor model S represented by equation 2.1 is a set of equations that describe the
physical mechanism of the interaction between the physical observable and the process. It
is possible to formulate a mathematical model for any process sensor; whether or not the
model captures the physics of the measurement is another matter. For instance, a proxim-
ity sensor that is based on electrical capacitance may be sensitive to changes in humidity
or barometric pressure due to its design. If the model used by the sensor fails to include
the environmental efects on the capacitance, then it will fail to describe the output of the
sensor under all operating conditions. Te sensor will always generate a signal, but the
signal could be wrong. If however the model adequately describes the sensor design, then
the sensitivity to these extraneous efects can be determined or even corrected.
2.2 UnitsofMeasure
Te sensor, together with whatever support circuitry is required, produces a measure-
ment value that is usually represented by a voltage or current level, serial port signal, or
other type of computer interface signal. Te signal may be displayed on a panel if there is
no direct computer connection. When measurements are represented by a voltage or cur-
rent level (as in the case of 420 milliamp [mA] loop control systems), the pre-established
range (minimum and maximum) of possible readings is used to interpret the sensor
output. Tus a current output of 5 mA from one sensor might mean 1 meter per second
(1 m/s), whereas 5 mA from a diferent sensor might mean 30C. Terefore, not only the
range but also the unit of measure is assumed. Sensors that communicate directly with
a computer system generally do not have this ambiguity, since the units can be included
in the string of characters that convey the measurement result.
Te International System of Units include fundamental quantities such as the meter,
kilogram, and second, and derived units such as the newton (the unit of force equal to
1 kgms
2
). Since it is ofen necessary to express a reading that is either very large or very
Figure 2.1 A generic process sensor.
Physical
Observable(s)
Measurement
Process
Sensor
44168_C002.indd 8 8/27/07 2:52:55 PM
Copyright 2008 by Taylor and Francis Group, LLC
Measurement 9
close to zero, the International System of Units also defnes prefxes that multiply the
base unit by a power of 10 (see Table 2.1).
2.3 SimpleStatistics
Consider a new thermocouple that has been installed in the process and is now gener-
ating temperature data. Suppose the frst reading is 30.0C and the second reading is
30.2C. Two possibilities can be considered: either the temperature is increasing or the
thermocouple reading is uncertain by at least 0.2C. All sensor readings are subject to
a certain amount of variability, so in order to know whether or not the temperature is
actually increasing, we must know whether or not a change of 0.2C is signifcant in this
case. In other words, if the temperature is truly constant, how much variability can be
expected from the thermocouple readout?
Simple statistical descriptors can be used to assess whether or not a given variation
in a reading is noteworthy. Te frst of these is the mean, or numerical average. Given a
sequence of N readings {r
1
, r
2
, , r
N
}, then the mean M is defned as
M
N
r
i
i
N
1
1
(2.3)
Te standard deviation s of the readings is defned to be
1
2
1
N
r M
i
i
N
( )
(2.4)
Te signifcance of the standard deviation is that it provides an indication of the vari-
ability of the data. Small values of s result from readings that are uniformly near the
Table 2.1 Prefxes Used with the International System of Units
Symbol Prefx Multiplier
P peta 10
15
T tera 10
12
G giga 10
9
M mega 10
6
k kilo 10
3
c centi 10
2
m milli 10
3
micro 10
6
n nano 10
9
p pico 10
12
f femto 10
15
44168_C002.indd 9 8/27/07 2:52:58 PM
Copyright 2008 by Taylor and Francis Group, LLC
10 Industrial Process Sensors
mean, and large values of s are caused by large variations from one reading to the next.
Te ratio (s/M) is the coefcient of variation, which can be used to compare the vari-
ability between two sets of readings with diferent mean values.
As an example, consider the hypothetical thermocouple mentioned earlier, and sup-
pose that its output is recorded once a minute for a period of 50 minutes, as shown in
Figure 2.2a. Te mean of these 50 measurements is 30.04C, and the standard deviation
is 0.30C; therefore, the coefcient of variation is only 1%. Although there are random
fuctuations from one reading to the next, there does not appear to be either an upward
or a downward trend in the data. Since the standard deviation is 0.30C and the dif-
ference between the frst two readings is only 0.20C, it seems likely that the initial
increase in the reading was not signifcant (i.e., no action should be taken by the process
controller).
Te measurement values in this hypothetical example are normally distributed,
which is to say that the probability P that the reading equals r is given by
P r
r M
( ) exp
( )
1
]
1
1
2 2
2
2
(2.5)
29.0
29.5
30.0
30.5
31.0
31.5
0 10 20 30 40 50
Time (min)
T
e
m
p
e
r
a
t
u
r
e
(
C
)
29.0
29.5
30.0
30.5
31.0
31.5
T
e
m
p
e
r
a
t
u
r
e
(
C
)
0 10 20 30 40 50
Time (min)
(a)
(b)
Figure 2.2 Te output from a hypothetical thermocouple as a function of time when (a) the
temperature is steady and (b) the temperature is increasing. Te dashed lines indicate the plus and
minus 2s levels for the baseline reading.
44168_C002.indd 10 8/27/07 2:53:01 PM
Copyright 2008 by Taylor and Francis Group, LLC
Measurement 11
where the mean M and standard deviation s of the measurement are defned above. Te
diference (r M) is usually interpreted as instrumental error. It is known from statisti-
cal analysis that in the special case of normally distributed data, about 68% of the values
are within one standard deviation of the mean, and about 95% are within two standard
deviations.
1
Only one in a million readings will deviate farther than 5s from the mean
(assuming that the physical quantity being measured actually remains constant). In this
example s 0.3C, so 95% of the readings are expected to lie in the range of 29.4C to
30.6C as indicated by the dashed lines in Figure 2.2a. In fact, the fgure shows that 49
of the 50 readings (i.e., 95%) do lie within this region, which means that the temperature
is stable according to this hypothetical thermocouple. If it is known that the standard
deviation is 0.3C, then it becomes clear that the apparent increase of 0.2C in the sec-
ond temperature reading is not a signifcant change.
Figure 2.2b shows the output of the same hypothetical thermocouple during a heat-
ing cycle, with the plus and minus 2s limits indicated by dashed lines as before. Te
frst two readings just happen to be 30.0C and 30.2C, as before. As argued above, the
increase in the second reading is not large enough to indicate that the temperature is
increasing. However, afer about 20 minutes the temperature readings frequently exceed
the limit indicated by the upper dashed line. At that point it is clear that the temperature
has increased by an amount that is statistically signifcant.
Te preceding example demonstrates that a discussion of the reproducibility of a
measurement is closely tied to a consideration of its statistics. In particular, the expected
variability in a measurement can be quantifed by the standard deviation. Although we
have not considered the root cause of the variability in the measurement, it is clear that
this variability must be reduced in order to improve the precision of the measurement.
2.4 SourcesofError
Te variability observed in multiple sensor readings is a combination of the real vari-
ability in the quantity being measured and the variability due to the sensor itself. Te
latter is usually called the instrumental error. Te goal of the measurement is to deter-
mine the process parameters (temperature, fow rate, etc.), and any variability in the
process should become evident from a statistical review of the process sensor data. As
demonstrated above, the interpretation of this data depends on the standard deviation
of the sensor output under quiescent conditions. Terefore, the instrumental error must
be assessed in order to determine the signifcance of changes in the sensor output.
Te sources of instrumental error can be classifed into systematic error and stochastic
(random) error. Systematic error is usually due to a faw in the design, installation, cali-
bration, or use of the process sensor. Tis type of error is manifested as a constant, or
slowly changing, ofset or multiplicative factor in the measurement. Common causes of
systematic error are sliding mechanical seals in force or position sensors, dark current in
infrared detectors and video cameras, and incorrect values of physical constants (such
as viscosity or index of refraction) in calculations. Te accumulation of process mate-
rial on the windows of optical sensors (fouling) will also lead to systematic errors as a
result of the decrease in light intensity.
44168_C002.indd 11 8/27/07 2:53:02 PM
Copyright 2008 by Taylor and Francis Group, LLC
12 Industrial Process Sensors
Systematic errors ofen lead to instrumental drif, wherein the amount of error
changes relatively slowly over time. Since electronic circuits are sometimes sensitive to
changes in temperature, a potential source of instrumental drif is the ambient tem-
perature in the process area. Depending on the operational design of the sensor, other
environmental conditions (such as barometric pressure) can contribute to drif.
Stochastic errors are rapid but usually small fuctuations in the sensor output, and
this random component of the signal is ofen called noise. Sensor noise comes from a
variety of sources including vibrations from the process equipment, ambient sound,
electrical ground loops, corroded electrical connections, and electronic circuitry.
Mechanical vibrations can interfere with the precise optical or mechanical alignment
in a sensor and cause an error in measurement. Sensors that are designed to measure
very small electrical signals contain amplifer circuits with a high gain; such devices
are ofen sensitive to vibration. Sensors that are very sensitive to vibration are micro-
phonic and therefore respond to ambient sound in the process area (ofen a very noisy
environment).
Ground loops are formed by connecting electrical equipment to ground via mul-
tiple paths or ground circuits. Tese connections create a continuous circuit or loop
in which alternating currents are induced by neighboring electrical equipment. Te
induced current leads to a fuctuating voltage ofset as a result of the resistance of the
wiring, and this ofset appears as a spurious signal in the sensors electronics.
2
In a
process environment, the current fowing through ground loops can be quite strong,
so the process sensors must be adequately shielded and properly grounded in order to
provide reliable results.
Mechanical vibrations and electrical ground loops contribute stochastic noise to the
process measurement, but it ofen happens that such noise is dominated by a few compo-
nent frequencies. Truly random electrical noise is caused by loose or corroded electrical
connections and electronic amplifcation. When an electrical connection becomes loose
or corroded, a slight voltage diference may develop across the junction. Tis voltage is
not well defned because the junction resistance is not well defned, and the fuctuating
voltage afects the sensor output in an unpredictable manner. Other sources of ran-
dom electrical noise include shot noise and thermal noise (see for example Fraden 1996,
pp. 212f). Shot noise is caused by the fact that electrical current is conducted by elec-
trons; the current appears to be continuous, but it is in reality a rapid and random
succession of charge transfer events. Termal noise is due to the random motion of elec-
trons in a conductor; this motion is dependent on the temperature of the conductor, so
an increase in temperature produces an increase in thermal noise.
Noise can ofen be mitigated by averaging the sensors output signal over a number
of readings. Stochastic error is as likely to cause an increase in signal as it is a decrease,
so the positive and negative fuctuations can be expected to cancel each other if a
sufciently large number of measurements are summed. An example is provided in
Figure 2.3, which shows a 10-point moving average of the data shown in Figure 2.2.
3
Te
constant temperature case (Figure 2.3a) shows a fairly constant output, and the heating
cycle case (Figure 2.3b) clearly indicates an increase in temperature. Process sensors
ofen perform some internal signal averaging in order to produce a more stable signal.
44168_C002.indd 12 8/27/07 2:53:02 PM
Copyright 2008 by Taylor and Francis Group, LLC
Measurement 13
2.5 AnalysisofError
Te individual systematic and stochastic errors are involved in making a measurement
propagate through the process sensor and become compounded in the resulting output
signal. If the sensor model is known, it is possible to estimate the relative efect of each
of these contributions through an analysis of error. Afer one has identifed the relative
contributions of each source of error, it is possible to focus on reducing the most egre-
gious of them. A thorough example of error analysis is given in chapter 10, where the
efects of environmental conditions on a flm thickness sensor are considered.
Error analysis is essentially a calculation of the cumulative efect caused by either
small changes or systematic error; it is therefore a calculation of diferentials. A simple
example is the measurement of resistance, which is the basis of many simple sensors. If a
constant current I is passed through a resistance R, the voltage diference V between its
two terminals (see chapter 5) is given by
V IR
(2.6)
29.0
29.5
30.0
30.5
31.0
31.5
0 10 20 30 40 50
Time (min)
T
e
m
p
e
r
a
t
u
r
e
(
C
)
29.0
29.5
30.0
30.5
31.0
31.5
T
e
m
p
e
r
a
t
u
r
e
(
C
)
0 10 20 30 40 50
Time (min)
(a)
(b)
Figure 2.3 A 10-point moving average of the data shown in Figure 2.2 for a hypothetical ther-
mocouple: (a) the moving average of data shown in Figure 2.2a; (b) the moving average of data
shown in Figure 2.2b.
44168_C002.indd 13 8/27/07 2:53:05 PM
Copyright 2008 by Taylor and Francis Group, LLC
14 Industrial Process Sensors
Terefore, knowing the amount of current, one could measure the voltage and calculate
the resistance according to
R
V
I
(2.7)
Tus, two distinct quantities are involved in this measurement, and the total error in
measured resistance depends on the error in measured voltage and the error in the
assumed current.
Te absolute error (dR) in the measured resistance due to the error (dV) in the mea-
sured voltage is
R
R
V
V
I
V
I
_
,
1
(2.8)
where denotes partial diferentiation and the vertical bar signifes that the current is
held constant. Te relative or fractional error (dR/R) can be calculated by dividing both
sides of equation 2.8 by equation 2.7:
R
R
I
I
V
V
V
V
I
_
,
(2.9)
Similarly, it can be shown that the fractional error due to an error in the assumed value
of the current (which may have changed over time) is given by
R
R
I
I
V
(2.10)
Te negative sign indicates that the resistance is underestimated if the current is
overestimated.
Since the measurement of voltage is independent of the estimation (or prior mea-
surement) of the current, the contributions to the fractional error are also independent
and must be added in quadrature. Terefore, the total fractional error in the measured
resistance is
R
R
R
R
R
R
V
V
I V
_
,
_
,
_
,
2 2
2
++
_
,
I
I
2
(2.11)
Equation 2.11 shows that the relative error of the resistance measurement is greater than
the larger of the two relative errors in voltage and current. If those two error terms are
comparable to each other, then
R
R
V
V
_
,
1 414 .
(2.12)
44168_C002.indd 14 8/27/07 2:53:12 PM
Copyright 2008 by Taylor and Francis Group, LLC
Measurement 1S
Results similar to equations 2.11 and 2.12 can be derived for any process sensor if the
function S from equation 2.1 is defned. Te value of such calculations is that they pro-
vide a rigorous framework for understanding sensor error in terms of the error in each
of the physical observables involved in the measurement. Once the sensor variability is
defned, it is possible to address the variability of the process itself.
SuggestedReading
Bevington, P.R. and Robinson, D.K. (1992). Data Reduction and Error Analysis for the
Physical Sciences, Second Edition. New York: McGraw-Hill.
Triola, M.F. (2005). Elementary Statistics, Ninth Edition. Boston: Pearson/Addison-
Wesley.
44168_C002.indd 15 8/27/07 2:53:12 PM
Copyright 2008 by Taylor and Francis Group, LLC