Seventh Lecture

Error Analysis
Instrumentation and Product Testing
1. Basic Concept
F
r
e
q
u
e
n
c
y
True value, X
T
Mean of measurement
output, X

Variations, i.e.
random error
Systematic error = X

X
T
Repeatability (that describes
precision) Standard deviation of
measurement data
Systematic error: Shift in the mean value
Random error: Standard deviation
Repeat measuring the same physical quality a number of time,
we shall obtain the following:
2. Random errors

They are represented by repeatability described in earlier lectures.

Repeatability (R) is numerically equal to the half range random
uncertainty (U
r
) of the measurement.

For normal distribution (n), at 95%
confidence level
x mean, population
-1.96o +1.96o
Repeatability, R = Zo
=1.96o
z
x mean, population
For normal distribution (n), at o% confidence
level
- Z
o
o + Z
o
o
Repeatability, R = Z
o
o
z
To estimate the population standard deviation from the sample
deviation, say, n = 10 or n = 20, we shall use the Students t
distribution. The repeatability may be expressed as

Repeatability, R = Z
o
o

= t s

where the value of t can be found from the t distribution table based on
the sample size n and the confidence level o,






s is usually referred to as the sample standard deviation.

( )
1
1
2

=

=
n
x x
s
n
i
Example. A balance is calibrated against a standard
mass of 250.0g. The difference in grams from the
250.0g scale reading were determined as

+10.5, -8.5, +9.5, +9.0, -10.0, +9.0, -9.5, +10.0,
+10.0 and -8.0

Find the repeatability R of the balance at 95%
confidence level.
You have to use a Students t distribution table to
determine the repeatability.

Probability = (1 o)/2 = 0.025,
Degree of freedom = (n 1) = 9.

t = 2.262
Apply the formula,

( )
1
1
2

=

=
n
x x
s
n
i
s = 9.66
The repeatability,
R = t s = 21.85g.
By calculating the mean of the differences (2.2g),
it is noted that the mean value lies at
250.0 + 2.2 = 252.2g
A positive bias (systematic error) of 2.2g, which
should be adjusted to zero.
Alternatively, the calculation can be conducted by using MS Excel.
We can use MicroSoft Excel to determine repeatability
Quite often, the repeatability of an instrument varies from
time to time by a considerable amount. This does not
necessarily indicate that the instrument is faulty but rather
that repeatability is a somewhat variable quantity.

Some authorities advocate that three repeatability tests be
carried out on three similar but not identical specimens in
quick succession. If the ratio between the highest and
lowest value is not greater than 2:1, then the root mean
square value of the three results should be regarded as the
repeatability of the instrument.

If the ratio obtained is greater than 2, then the instrument
should be examined for faults, and on rectification further
tests should be made.
Example. Three repeatability tests were carried out on the
balance introduced in last example. The results obtained
were as follows:

R
1
= 22g, R
2
= 24g, and R
3
= 28g

Find the repeatability of the balance.

Solution:
R
3
/ R
1
= 28/22 = 1.27 < 2
79 . 24
3
28 24 22
3
2 2 2 2
3
2
2
2
1 1
2
. . .
=
+ +
=
+ +
= =

=
R R R
n
R
R
n
i
i
s m r
Rounding up, the repeatability, R
r.m.s.
= 25
Random errors can largely be eliminated by calculating the mean of
the measurements,



since in statistical analysis of data




where is the standard deviation of the sample mean, o is the
standard deviation of the population, and n is the sample size (i.e.
number of repeated measurements for the same measurand). The
standard error of the mean is usually expressed by .

n
x
x
i
=
n
x
o
o =
x
o
x
o
For example, a set of length measurements with n = 20, =
18.18mm, and o = 0.318mm,




The length can be expressed as

18.18 0.07 mm (for 68% confident limit)
18.18 0.14 mm (for 95% confident limit)

It should be noted that the sample standard deviation s is used as an
estimator of the standard deviation of the population o in some
textbooks for the calculation of .

x
mm
x
071 . 0
20
318 . 0
= = o
x
o
3. Error reduction using intelligent instruments

It also includes

- a microcomputer, and
- one or more transducers (secondary transducers)

The secondary transducers monitor the environmental conditions
(modifying inputs). By reading the outputs of the primary
transducer and the secondary transducers, the microcomputer
processes the signals based on a pre-loaded programme.

It can be programmed to take a succession of measurements of a
quantity within a short period of time (sampling frequency)
-1
and
perform statistical calculations on the readings before displaying an
output measurement. This is valid for reducing random errors.
The ability of intelligent instruments to reduce systematic errors
requires the following pre-conditions be satisfied:

The physical mechanism by which a measurement transducer
is affected by ambient condition changes must be fully
understood and all physical quantities which affect the
transducer output must be identified.

The effect of each ambient variable on the output
characteristic of the measurement transducer must be
quantified.

Suitable secondary transducers for monitoring the value of all
relevant ambient variables must be available for input to the
intelligent instrument.

The accuracy of a measurement system is a function of its ability
to indicate the true value of the measured quantity under specific
conditions of use and at a defined level of confidence.

The accuracy A is expressed by




where R is the instrument repeatability and U
s
is the systematic
error.

2 2
s
U R A + + =
4. Calculations of accuracy and errors

4.1 Accuracy of a measuring instrument
Systematic Error
Random Error
( Zo)
Example. The systematic error of a balance is estimated to be 5g
and the random error of its measurements is 25g. State the
repeatability and calculate the accuracy of the instrument.

Solution:


Repeatability, R = U
r
= 25g
Systematic error, U
s
= 5g


g A 495 . 25 5 25 , Accuracy
2 2
= + + =

As the systematic error and the repeatability are both stated in grams,
the accuracy of the instrument is 26g.
It is usual to express the accuracy in terms of its full scale deflection
(f.s.d.).
Accuracy = f.s.d.

% 26 . 0 % 100
10 10
26
3
=

The sum of the systematic and random errors of a typical

measurement, under conditions of use and at a defined level of
confidence.

Error assessment must also include the error of the calibrator itself
and take account of the confidence level upon which it is founded.

Example. The diameter of the setting gauge used to a sensitive
comparator was stated on its calibration certificate to be
60.0072mm and to have an accuracy of determination equal to
0.0008mm. Although not stated on the certificate, the level of
confidence was known to be better than 95%. The sample
standard deviation of 10 instrument readings yielded a value s =
0.37m. Estimate the error of the measurement at better than 95%
confidence level.
4.2 Estimation of total error
Solution:
From Students t distribution, at 95% confidence level, and sample
size n = 10, the value of t is 2.26.

U
r
= R = ts = 2.26 0.37 = 0.8338m



The half range error (uncertainty) of the setting gauge, U
l
= 0.8m.
Therefore, the half range error of the measurement





This result must be rounded up to the first decimal place as this is the
order of the observations. The error of measurement is 1.2m. This
estimate is given with better than 95% confidence level owing to the
effect of rounding up.
m 156 . 1 8 . 0 8338 . 0
2 2 2 2
= + + = + + =
l r t
U U U
4.3 Compound error/uncertainty

In many instances, the ultimate error (uncertainty) of a measurement
is dependent upon the errors of a number of contributory
measurements which are combined to determine the final quantity.
The measurement output

M=M(x
1
, x
2
, x
3
,)

is a function of a number of individual measurements x
1
, x
2
, x
3
, etc.
All of these measurements have individual error of Ax
1
, Ax
2
, Ax
3
,
etc. Then, the compound error of the measurement M, AM, can be
determined by substituting into the equation of M=M(x
1
, x
2
, x
3
,)
the maximum and minimum values of x
1
, x
2
, x
3
, etc., and thus
finding the maximum and minimum values for M.

This would obviously be a laborious process, and the problem is
better solved using partial differentiation:









Attentions must be paid to those high value terms that give
dominant contributions to AM.


i
i
dx
x
M
dM

c
c
=
i
i
x
x
M
M A
c
c
~ A

i
i
x
x
M
A
c
c
Example. Consider the measurement of electric power from

P = EI

where E is the e.m.f. and I is the current. They are measured as

E = 100V 1V
I = 10A 0.1A



The nominal value of the power is

P = 100 10 = 1000W


By differentiating the equation P = EI, it can be obtained that the error
(compound) of calculating the power


By taking the worst possible errors in voltage and current, AE =
1V and AI = 0.1A,



EdI IdE dI
I
EI
dE
E
EI
dP + =
c
c
+
c
c
=
) ( ) (
I E E I P A + A ~ A
( ) W 20 1 . 0 100 1 10 = + ~ AP
% 0 . 2 % 100
1000
20
= ~
A
P
P
Alternatively,

P
max
= (100 + 1) (10 + 0.1) = 1020.1W

P
min
= (100 1) (10 0.1) = 980.1W






In the above example, it is quite unlikely that the power would be
in error by these amounts because the chance of obtaining the
largest errors for both voltage and current measurements at the
same is very slim. It is especially true when Ax is used to specify
the tolerance of measurement.

% 0 . 2 ~
A
P
P
A more accurate method of estimating error has been derived as
follows:




If this relation is applied to the electric power relation of the
previous example, the expected error is







The expected error (or uncertainty) is 1.4% instead of 2.0%.

|
|
.
|

\
|
A
c
c
= A
2
i
i
x
x
M
M
( ) ( ) W 14 . 14 10 10
2 2
2 2
~ + = A + A = A I E E I P
% 4 . 1 % 100
1000
14 . 14
~ ~
A
P
P
Example. The resistance of a certain size of copper wire is given as




where
R
0
= 6O 0.3% is the resistance at 20C
o = 0.004C
-1
1% is the temperature coefficient of resistance
T = 30 1C is the temperature of the wire

Calculate the resistance of the wire and its uncertainty.

Solution:

The nominal resistance is


( ) | | 20 1
0
+ = T R R o
( ) | | O = + = 24 . 6 20 30 004 . 0 1 6 R
To determine the uncertainty, the partial differentiation is
performed:









The uncertainty of each component is:


( ) ( ) 04 . 1 20 30 004 . 0 1 20 1
0
= + = + =
c
c
T
R
R
o
( ) ( ) 60 20 30 6 20
0
= = =
c
c
T R
R
o
024 . 0 004 . 0 6
0
= = =
c
c
o R
T
R
O = = A 018 . 0
100
3 . 0
6
0
R
1 6
10 40
100
1
004 . 0

= = A C o
C T = A 1
Thus, the uncertainty in the resistance is:











It can be noted that the two components and

predominate the value of AR. For accuracy improvement, these
two terms must be of main concern.

( ) ( ) ( )
2
2
6
2
1 024 . 0 10 40 60 018 . 0 04 . 1 + + = A

R
( ) ( ) ( )
2
3
2
3
2
3
10 24 10 4 . 2 10 72 . 18

+ + = AR
O ~ A
3
10 53 . 30 R
% 49 . 0 % 100
24 . 6
10 53 . 30
3
~

~
A

R
R
0
0
R
R
R
A
c
c
T
T
R
A
c
c
Thank you