https://www.dept.aoe.vt.

edu/~aborgolt/aoe3054/

http://www.dept.aoe.vt.edu/~aborgolt/aoe3054/manual/index.html

http://www.dept.aoe.vt.edu/~aborgolt/aoe3054/manual/ch2/index.html

Chapter 2 - BASIC CONCEPTS IN EXPERIMENTS

W J Devenport and A Borgoltz
Last Modified January 19th, 2016

An experiment is essentially an organized set of observations of a device or

phenomenon. The observations may be qualitative (e.g. putting smoke in an air flow
to observe the flow pattern) or quantitative (e.g. measuring the defection of a beam
under load). Making quantitative observations requires the use of some kind of
measurement system. Such systems vary from the very simple (a thermometer) to
the very complex (an optical flow-speed sensor for a rocket nozzle). The purpose of
this chapter is to introduce some basic ideas about what a measurement system is,
how to analyze the results it produces, and how to estimate the accuracy of those
results.

1. Measurement systems

Most measurement systems may be broken down into 3 components; (i) a detector or
transducer that senses the physical quantity of interest and transforms it into a
mechanical or electrical signal, (ii) an intermediate stage that amplifies or otherwise
modifies the signal, and (iii) a final stage in which the signal is displayed or stored.

A simple example of such a measurement system is the mercury thermometer. The

mercury in the bulb senses a change in temperature through thermal expansion or
contraction. This mechanical signal is amplified by the capillary tube attached to the
bulb - a small change in the volume of the mercury producing a large change in its
height in the capillary. Finally the temperature is displayed by comparing the height of
mercury with a scale on the outside of the tube.
A slightly more complex example, perhaps better representative of those
measurements systems often seen in aerospace or ocean-related experiments, is the
pressure gage shown in figure 1. This particular gage uses a detector known as a
piezoelectric crystal to detect the fluid pressure at a pipe wall. When compressed such
a crystal produces a small voltage across its faces. An electrical circuit is then used to
amplify this voltage. The resulting signal is then displayed using a voltmeter or
perhaps read into a computer and stored.

Figure 1. A system for measuring pressures at a pipe wall.

It is obvious that to use a measurement system you have to know the relationship
between its output and the quantity it senses, called a calibration. You also have to
know the limitations of the system, specifically its range, its accuracy, precision and
repeatability, and how quickly it can respond to a change in the quantity being
measured, called its dynamic response.

1.1 Calibration
To calibrate a measurement system you have to compare its output with a known
input. Consider, for example, the pressure gage of figure 1. This device could be
calibrated by mounting the transducer in a closed vessel with a piston at one end,
as shown in figure 2. (Such a vessel is called a dead-weight tester.) Placing a
weight on the piston produces a known pressure in the vessel equal to the weight
divided by the piston area. After applying a series of weights and measuring the
corresponding output voltages of the pressure gage we can plot the relationship
between the output voltage and pressure (figure 3). Mounting the transducer back
in the pipe we now use this calibration curve to convert voltage readings into
pressures at the pipe wall.
Figure 2. Dead weight tester for calibration of pressure transducer.

As a general rule it is unwise, if you have the choice, to trust a calibration made by a
manufacturer or estimated through a theoretical calculation - measurement systems
rarely work exactly as they are supposed to. Repeating a calibration at regular
intervals is also a good idea since it may drift with time or ambient conditions. Having
a repeatable and accurate calibration is obviously a prerequisite for an accurate
measurement.

Figure 3. Transducer Calibration.

1.2 Range
The design of a measurement system obviously limits its range - the thermometer
cannot measure temperatures for which the mercury contracts into the bulb or blows
off the top of the capillary, the pressure gage cannot measure pressures for which the
voltage is greater than that which can be handled by the electrical amplifier. While
staying within the design limits is often important if the measurement system is not to
be damaged, staying within the calibration limits is important if accurate
measurements are required. For example, the calibration curve of figure 3 only
extends from 3 to 10 psi. Using the pressure gage to measure a pressure of 1 psi
almost certainly would not break it, but one would need to extrapolate the calibration
curve downwards. There is no guarantee that such an extrapolation would be accurate.

1.3 Accuracy and Precision and Repeatability

Accuracy is merely an optimistic word for error, the difference between the output of
a measurement system and the true value. (When we say that a probe measures
velocity to an accuracy of 1 ft/s we mean that its measurements are in error by that
amount.) Much of the error in a measurement system can be eliminated by calibration.
However, drift, lack of repeatability and inaccuracy in the calibration mean that there
will always be some residual error.

Precision is the resolution with which a measurement may be made. For example if
the thermometer has graduations every 2 degrees it probably could be read with a
precision of about half that. Precision does not guarantee accuracy. The thermometer
scale may have been printed in the wrong place making it 10 degrees in error.

Repeatability is the difference between successive measurements of the same

quantity.

1.4 Static and Dynamic Response

Static response refers to the behavior of a measurement system when it is used to
measure a fixed quantity. Dynamic response refers to behavior when the quantity is
changing with time. The U-tube manometer (figure 4) is a measurement system that
gives a good physical feel for the difference between static and dynamic response.
Such a manometer senses the pressure difference between its two open ends through
the displacement of a column of fluid in a vertical U-shaped tube.

For steady pressure differences (p) the displacement of the manometer fluid (h) is
given by hydrostatics,

(1)
This equation defines the static response.

Figure 4. A U-tube manometer.

For pressures that are changing with time, equation (1) only holds if the changes are
of very low frequency - a fraction of a Hertz. This is because, as the frequency is
increased, inertia of the fluid column and friction between it and the tube start
becoming important. Thus, the motion of the fluid column begins to lag behind the
pressure change and the amplitude of this motion begins to differ from that given by
equation 1 (figure 5).

Figure 5. Manometer response to sinusoidal pressure fluctuation.

We refer to the ratio of the amplitude of the pressure fluctuations indicated by the
manometer to the actual amplitude as the "amplitude response". We refer to the lag of
the manometer, measured in terms of the angle of the sinusoidal fluctuation as the
"phase response". If we apply a series of sinusoidal pressure fluctuations of different
frequencies we can plot the amplitude and phase response as functions of frequency
(figure 6). These curves define the "dynamic response" of the manometer.

Figure 6. Possible amplitude and phase response curves for a U-tube manometer.

Most U-tube manometers tend to have dynamic response curves that look like A or B
of figure 6. Curve A is for a small diameter tube in which viscous friction between the
fluid and tube wall is sufficient to damp out much of the oscillatory motion of the
manometer fluid that fluctuations in the pressure being measured are trying to induce.
Curve B is for a manometer with a broad tube that will damp these oscillations to a
much smaller degree. At some frequencies the amplitude response is actually greater
than 1 (indicated pressure fluctuation greater than actual). These are frequencies close
to that at which the manometer fluid would naturally oscillate back and forth in the U
tube. Both curves show that the amplitude response of a manometer becomes very
small (and for the most part useless) when subjected to pressure fluctuations at
frequencies greater than a few Hertz.

The fact that we have used sinusoidal pressure fluctuations to define the dynamic
response of our manometer may seem somewhat artificial. In reality, however, it helps
us predict how the manometer will respond to any signal. Consider, for example, the
pressure signal shown in figure 7 (produced perhaps by an unsteady fluid flow). Such
a signal contains many frequencies. If the manometer operates as a linear device its
response is the (linear) sum of its phase and amplitude response to each of the
frequencies in the signal. i.e. it will respond mostly to the low frequency parts of the
signal.

Figure 7. Manometer response to typical pressure signal.

A manometer is obviously a measurement system with a poor dynamic response.

Many (usually electrically-based) measurement systems can respond much faster to
changes in the quantity they measure (at frequencies of kHz or even MHz). However,
they are all ultimately limited in much the same way.

2. Statistical analysis of measurements and related probability theory.

2.1 Mean, standard deviation and histograms

Once a series of measurements has been made the engineer is faced with the task of
interpreting them, i.e. deciding what they mean. This is often a lot more difficult than
it sounds. Imagine, for example, a test to examine the behavior of a wing spar. An
engineer places a position sensor on the spar and records the signal from it at cruising
conditions. What do the measurements (the recorded signal) mean? Do the say
anything about the average load on the spar? Is the spar vibrating excessively? Could
it be in danger of developing a fracture? Answering these questions does not only
require a physical and theoretical understanding of the situation but also requires some
analysis of the measurements made.

In this, and many other, situations a good first step would be to calculate the mean and
standard deviation of the measurements. The definitions of these quantities depend on
whether the data from which they are to be calculated are discrete (i.e., a sequence of
measured points) or continuous (e.g., an electrical signal). For a set of N discrete
samples of a quantity x the mean and the standard deviation are given by the
x

relations
 (2)
Note that the standard deviation is calculated dividing by N-1 rather than N. This is
because only N-1 of the samples are independent of the mean. For a continuous
signal measured as a function of time x(t) over a period from t = 0 to T the mean and
standard deviation are defined as

(3)
Note that the standard deviation squared is called the variance.

In our example the mean of each signal would be the average deflection of the spar
which presumably could be used to estimate its average load. The standard deviation
or variance are measures of how widely the measurements are spread around the mean
(i.e., how large, on average, (x - ) or (x(t) - ) are). In this case (assuming the
i
2 2

position sensor had a good dynamic response) they could be taken as indications of
the intensity of vibrations in the spar. For obvious reasons, the standard deviation is
often referred to as the "root mean square" or r.m.s., especially when dealing with
electrical signals. Since such signals are often sinusoidal it is useful to remember that
the r.m.s. of a sinusoid is 0.7071 of its amplitude, verify this for yourself using
equation 3.

There are many other ways of presenting data in a statistical way. Perhaps the most
revealing is a histogram. To construct a histogram the range of the quantity being
measured (deflection in our example) is divided up into a number of equal intervals,
or 'bins'. For discrete data we then merely add up the number of samples falling in
each bin. For continuous signals we add up the time for which the signal lies in each
bin. The result is a graph showing how the data are distributed. In our example, such a
histogram could show whether our spar had ever deflected sufficiently to yield. The
mean and standard deviation may be represented on a histogram, as shown in figure 8.
The mean, being a measure of the average, lies near the center of the histogram. The
standard deviation, being a measure of the spread, is usually about one quarter to one
sixth of the width of the histogram.
Figure 8. Histogram of spar deflections.

2.2 Probability density functions - the normal distribution

There is a close connection between histograms and probability. Consider, for
example, measurements of the heights of waves hitting an oil rig during a season. A
histogram of these measurements (see figure 9) can be used to estimate the probability
of waves of a certain size hitting the oil rig in the future. For example, the probability
of a wave with a height (H) between 20 and 30 feet hitting the rig may be estimated
by adding up the number of samples in bins in the range 20 < H < 30 and then
dividing by the total number of samples in the histogram. This is equivalent to taking
the ratio of the area of the histogram between H = 20 and H = 30 and dividing by the
total area, see figure 9.
Figure 9. Histogram of heights of waves hitting an oil rig.

The natural extension of this is the probability density function, which is defined in
terms of its area as follows.

"Consider a probability density function p(x) of a quantity x (in the present example,
wave height). The area under p(x) between two values x and x is the probability P
0 1

that a given sample of x will fall between x and x ."

0 1

Mathematically this may be written as

(4)
Obviously the total area under p(x) (the probability of a sample having any value)
must be unity, i.e.

(5)
Depending on the physical process controlling the quantity represented, probability
density functions may have any form. However, it is a matter of experience that the
vast majority of random processes (including random experimental error) produce
the same shape of probability density function, namely the normal (or Gaussian)
distribution. Figure 10 is a graph of the normal distribution, given by the
mathematical relation

(6)

Figure 10. The normal distribution.

As you can see the normal distribution is a function of the mean and standard
deviation of the quantity x. Using equation (4) we see above the probability P of a
given value x of a quantity governed by a normal distribution falling within a
range x to x is
0 1

(7)
Making the substitution
 (8)
we obtain

(9)
which can be rewritten as

(10)
where

(11)
The value of the function in equation (11) is tabulated in table 1. Note that a decimal
point should be placed ahead of each of the numbers in this table. The table works
for both negative and positive values of . For negative , values read from the table
should be negated. Some typical uses of the normal distribution are illustrated in the
example below.

Example
A sensor is used to detect the flow rate of fuel to a jet engine. Due to electrical
interference in the instrumentation used, however, successive readings from the
sensor fluctuate. The following are 21 such readings (in arbitrary units),

Reading Flow rate Reading Flow rate Reading Flow rate

1 .512 8 .734 15 .627
2 .477 9 .771 16 .701
3 .794 10 .486 17 .573
4 .672 11 .559 18 .721
5 .713 12 .614 19 .802
6 .588 13 .687 20 .553
7 .621 14 .722 21 .605
(a) Determine the mean and standard deviation.
From equation 2 we have

and,

(b) Assuming the readings are distributed normally calculate the probability that a
reading taken at random will have a value between .5 and .6.

Applying equation 10 we have,

(c) What percentage of a large number of readings are likely to lie above a value
of .8?

Applying equation 10 again,

i.e. 5.6%

(d) What percentage of a large number of readings are likely to lie within two
standard deviations from the mean?

We have

i.e. 95%. Alternatively we could say that there is a 95% probability that a reading will
lie within two standard deviations of the mean.
2.3 Linear regression
Linear regression is the process of fitting the best possible straight line through a
series of points. Linear regression is often used to reduce a set of calibration points to
a simple mathematical relationship (that can, for example, be implemented on a
computer) or deduce the underlying trends from a set of measurements that are
expected, on theoretical grounds, to follow a straight line.

Consider a series of measured points (x , y ) describing a relationship between two

i i

quantities x and y. (A good example here is the measurements of voltage and pressure
we would have got in our calibration in section 1.2). We wish to find a straight line of
the form

(12)
that passes as closely as possible through all the points. This is done by choosing the
constants A and B to minimize the differences between the straight line and points.
The difference at the "ith point" is,

(13)
The sum of the squares of the differences at all the points is

(14)
(we take the square so that the positive and negative differences don't cancel each
other out). To find the values of A and B for which S is a minimum we take the
derivatives of S with respect to A and B and set them to zero. This gives

(15)
and

(16)
Rearranging we obtain explicit expressions for A and B,
 (17)

(18)

To express how good the fit of the straight line is we usually use the correlation
coefficient, defined as,

(19)
where,

(20)
and

(21)
Verify for yourself that r=1 for a perfect fit (y always equal to Ax + B) and that r<1
i i

otherwise. Using r to decide how good or bad a fit might be depends on the
application and even then is a matter of experience. In general, if r>0.99 the
differences between the points and the line will be barely noticeable. If r<0.95 then
the fit will appear poor when plotted on graph paper and may be of little use. Many
calculators have built in routines for performing linear regression.

3. Uncertainty analysis
Almost all engineers in the "real world" deal in one way or another with data derived
from experiments or tests. They may have obtained the data. They may be trying to
use the data in a design. They may wish to use the data to test the results of their
computation. They may want to use the data to make a production or marketing
decision. Whatever the situation, it is essential that the engineer have a good idea of
the likely accuracy of the data. He or she may be required to state the likely accuracy
of his or her own data, or will be required to interpret the accuracy estimates of others.
In either case a knowledge of how accuracy is calculated is needed.

Estimates of experimental accuracy are usually referred to as 'Uncertainty Estimates' ,

'Uncertainty Intervals' (the interval around a measurement within which the true value
should lie) or just 'Uncertainties'. The techniques used to obtain them are collectively
named 'Uncertainty Analysis'. Both Kline and McClintock (1953) and Holman (1989)
describe uncertainty analysis. The former of these is the most original and probably
the best reference here. The latter, however, may be easier to get hold of.

An uncertainty interval defines a symmetrical band around a measurement. Ideally it

should be chosen so that there is a 95% probability that the true value lies within it.
Another way of saying the same thing is that, if 20 successive unbiased measurements
are made of the same quantity 19 of them should fall within the uncertainty interval of
the true value.

Uncertainty intervals may be represented in one of several ways. When listed along
with a measurement the '±' sign is used to indicate an uncertainty, e.g. "the flow speed,
U, was 20 ± 1 m/s" indicates an uncertainty of 1 m/s. When presented in isolation the
symbol () is used to indicate uncertainty, the quantity to which the uncertainty
interval refers going inside the parentheses, e.g. " (U) = 1 m/s". Obviously (U) has
the same units as U. Occasionally uncertainties are also presented as percentages of a
measured quantity e.g. "there is a 5% uncertainty in velocity" means that (U)/U =
0.05.

In general, uncertainty analysis may be divided into two parts, (1) determining the
uncertainty in primary measurements, (2) determining the uncertainty in a result
derived from those measurements.

3.1 Determining the uncertainty in primary measurements

A primary measurement is one that is not derived from any other, e.g. voltage from a
voltmeter, temperature from a thermometer, head from a manometer, distance from a
dial gage etc. Estimating an uncertainty interval in a primary measurement essentially
involves making an educated guess based on several sources of information, namely;
 (a) Digital resolution, size of smallest divisions in scale. These clearly limit the
accuracy with which an instrument can be read. The lowest possible
uncertainty is half the digital resolution. If you can't accurately read between
the finest divisions on a scale the lowest possible uncertainty is half the size of
the smallest division. (Note that these are lowest estimates. Quite often a
meter reading fluctuates over a range of values much larger than its
resolution. In this case a much larger uncertainty estimate is in order, see (c)
below.)
 (b) Manufacturers information, calibration information. Quite often a reading
may be very precise (say to six significant digits) but very inaccurate (say 15%
in error). Reading manufacturers specs to find out just how accurate an
instrument is, or looking over a recent calibration of the instrument, or
calibrating the instrument yourself, are ways of checking up on this.
 (c) Repeated measurements of the same quantity. Sometimes the output of
an instrument or meter will vary randomly with time due to such things as
electrical noise, unsteadiness in the quantity being measured, or changes in
ambient conditions. Assuming the average indicated quantity is the true value
(it may not be, see (b) and (e)), the uncertainty may be estimated by taking
several successive readings and calculating their standard deviation. The
uncertainty is then taken as twice the standard deviation. This in effect
assumes that the random variations follow a normal distribution. As we saw
above, an interval with a size of ±2 standard deviations from the mean
encloses 95% of the area of a normal distribution. Thus, on average, it would
contain 19 out of 20 readings.
 (d) Comparison with other independent measurements of the same quantity.
Occasionally two independent devices are available for making the same
measurement (e.g. a voltmeter and an oscilloscope for measuring voltage).
The difference between measurements made by such devices is a useful guide
to estimating uncertainties since the uncertainty clearly cannot be smaller
than the difference.
 (e) Other factors, validity of the measurement scheme. Other factors include
anything else the engineer may know of that could have affected the accuracy
of the measurement e.g., operating an instrument outside its design range,
unusual environmental conditions, blunders in taking the data. Another
important consideration which will often outweigh all others is the validity of
the measurement scheme itself. For example, Pitot-static tubes are often used
to measure velocities in turbulent flows, despite the fact that the assumptions
made in using this technique do not hold in such flows. As a result, no matter
 how accurate and precise the rest of the measurement system is, this
measurement will have a large uncertainty.
 (f) Experience. Your gut feelings and honest confidence in what you are doing.

3.2 Determining the uncertainty in a result

In general experimental data is processed to generate results. The connection
between the raw primary measurements and the results is always a mathematical
function of some kind, i.e. if R is a result and a, b, c.... are mathematically
independent primary measurements of known uncertainty used in calculating R then
we have
(22)
To calculate the uncertainty in R resulting from the uncertainties in a, b, c ... the
equation

(23)
is used. This equation follows the differential calculus of small changes and, strictly
speaking assumes that uncertainty intervals are small and that the partial derivatives
of R do not become infinite. Use of this equation, and the procedure described in
section 3.1 above is best explained using an example.

Example: Power dissipated in a resistor

The power dissipated by a resistor in an electrical circuit is being estimated using the
voltage across it measured using a digital voltmeter. The resistance R has a nominal
value of 100 . The voltmeter reads 28.0 volts, with a resolution of 0.1V. Estimate the
uncertainty in the power measurement.

Uncertainty in primary measurements:

 Resistance: A glance through any manufacturers specifications will show you

that most often nominal resistances are only accurate to within ±5%. We shall
therefore take our primary uncertainty here as (R)=5 .
 Voltage: The uncertainty in the reading of a digital voltmeter is usually half the
resolution (the true voltage could lie anywhere between 27.95 and 28.05). We
therefore have (V)=0.05V.

Uncertainty in power:
 The relationship between power, voltage and resistance is simply

Preparing to apply equation (23) we have,

and so,

or,

about 5%. You will notice that apart from giving us an uncertainty estimate this
analysis also shows that the likely error in our power measurement is almost entirely
due to the uncertainty in resistance. To improve the accuracy we should therefore
concentrate on reducing the uncertainty of the resistance measurement, not on
improving the voltmeter. This kind of information can save a lot of time and money
(voltmeters are expensive, resistors are not).

4. References

1. Holman J P, 2001, Experimental Methods for Engineers, 7th Edition, McGraw

Hill, New York.
2. Kline S J and McClintock F A, 1953, "Describing Uncertainties in Single Sample
Experiments", Mechanical Engineering, vol. 75, pp. 3.
Table 1. Values of the function I in equation (11). Note that each value is preceded
by a decimal point that has been omitted in the table for brevity.

------------------------------------------------------------------------------
-------
0 0.01 0.02 0.03 0.04 0.05 0.06 0.07
0.08 0.09
------------------------------------------------------------------------------
-------
0.0 00000 00399 00798 01197 01595 01994 02392 02790
03188 03586
0.1 03983 04380 04776 05172 05567 05962 06356 06749
07142 07355
0.2 07926 08317 08706 09095 09483 09871 10257 10642
11026 11409
0.3 11791 12172 12552 12930 13307 13683 14058 14431
14803 15173
0.4 15554 15910 16276 16640 17003 17364 17724 18082
18439 18793
0.5 19146 19497 19847 20194 20450 20884 21226 21566
21904 22240
0.6 22575 22907 23237 23565 23891 24215 24537 24857
25175 25490
0.7 25804 26115 26424 26730 27035 27337 27637 27935
28230 28524
0.8 28814 29103 29389 29673 29955 30234 30511 30785
31057 31327
0.9 31594 31859 32121 32381 32639 32894 33147 33398
33646 33891
1.0 34134 34375 34614 34850 35083 35313 35543 35769
35993 36214
1.1 36433 36650 36864 37076 37286 37493 37698 37900
38100 38298
1.2 38493 38686 38877 39065 39251 39435 39617 39796
39973 40147
1.3 40320 40490 40658 40824 40988 41198 41308 41466
41621 41774
1.4 41924 42073 42220 42364 42507 42647 42786 42922
43056 43189
1.5 43319 43448 43574 43699 43822 43943 44062 44179
44295 44408
1.6 44520 44630 44738 44845 44950 45053 45154 45254
45352 45449
1.7 45543 45637 45728 45818 45907 45994 46080 46164
46246 46327
1.8 46407 46485 46562 46638 46712 46784 46856 46926
46995 47062
1.9 47128 47193 47257 47320 47381 47441 47500 47558
47615 47670
2.0 47725 47778 47831 47882 47932 47962 48030 48077
48124 48169
2.1 48214 48257 48300 48341 48382 48422 48461 48500
48537 48574
2.2 48610 48645 48679 48713 48745 48778 48809 48840
48870 48899
2.3 48928 48956 48983 49010 49036 49061 49086 49111
49134 49158
2.4 49180 49202 49224 49245 49266 49286 49305 49324
49343 49361
2.5 49379 49296 49413 49430 49446 49461 49477 49492
49506 49520
2.6 49534 49547 49560 49573 49585 49598 49609 49621
49632 49643
2.7 49653 49664 49674 49683 49693 49702 49711 49720
49728 49736
2.8 49744 49752 49760 49767 49774 49781 49788 49795
49801 49807
2.9 49813 49819 49825 49831 49836 49841 49846 49851
49856 49861
3.0
49865
3.5 4997674
4.0 4999683
4.5 4999966
5.0 4999997133
------------------------------------------------------------------------------
-------