2

Chapter 2

MEASUREMENT UNCERTAINTY

29
 2.1. MEASUREMENT UNCERTAINTY

The term uncertainty is used to describe variation in the measurement or control of any quantity
associated with an experiment. Thus, uncertainty is a measure of the accuracy of the
experimental data. Specifically, it is an estimate of the maximum error which might reasonably
be expected from a set of measurements. It is important to estimate the uncertainty associated
with a measurement and to determine ways to reduce it.

Similar to an experiment, numerical analysis of any problem also contains errors associated with
the assumption(s) made, the method of analysis, the type of grid and the number of iterations that
are used, etc. It is also very important that results from any numerical analysis also presented
with their uncertainty.

Measurement of any property contains errors. The errors are the differences between the
measured values and the true value. The true value is either given or estimated from previous
repeated experiments, or assumed to be the reasonable human error associated with the
limitations of the measuring instrument being used. For example, the uncertainty of an
instrument such as a digital pressure transducer will typically be given in the operating manual.
If it is not, then a set of experiments that compare the reading from the transducer to a known
input can be devised that allows estimation of the uncertainty of the transducer. In the case of a
simple measurement device, such as a ruler, the uncertainty can be estimated by determining
how closely an individual can reasonably read the graduations on the ruler.

Error in a measurement has two components: fixed or bias error and random or precision error
(repeatability). Bias error is the difference between the mean value, x , of the measurements and
the actual value of the quantity being measured. The mean value is estimated from the following
equation:

1 N
x   xi (2-1)
N i 1

In this equation, N is the number of samples and xi is the value for each measurement. With
thirty or more samples, the mean is often used as an estimate of the true value of a measurement

Bias is the systematic error. In repeated measurements at the same conditions, bias error is the
same for all of them. The bias can not be determined unless we know the true value. Many
experimental apparatus can operate within a special range of experimental conditions. If the
environment changes, the accuracy of the equipment will affected, and thus the bias will change.
Bias errors can be eliminated by calibration if they are large and known. Small unknown biases
contribute to the bias limit.

Large unknown biases usually come from human errors in data processing, incorrect handling
and installation of instrumentation, and unexpected environmental disturbances such as changes

30
in humidity, temperature, or Barometric pressure. A well controlled environment should not have
large unknown biases.

The major sources of bias error are: 1) Nonlinearity, 2) poor calibration, 3) Hysteresis, and 4)
Drift.

The output of a device should be linearly proportional to its input. However, all devices deviate
from linearity. Nonlinearity can always be described as a combination of an offset from the ideal
linear response and a wobble about the ideal linear response. The departure from linearity may
be classified in one of two ways: Independent of reading and dependent on reading. When an
independent of reading classification is specified, it means that the maximum deviation from
linearity is less than a certain percentage of full scale. The a dependant on reading classification
is specified, it means that the maximum deviation from the measurement reading is less than a
certain percentage of that particular reading.

Calibration is used to determine the response of a device to a known input. Once this response is
known, the device can be used to determine the value of an unknown input. The accuracy of the
relationship of the response determined during calibration directly affects the accuracy of the
measurement.

The response of a measurement device has hysteresis when the output of the device is dependent
on the direction from which the corresponding input is approached. Specifically, a reading near
the midrange of an instrument will have two different values depending on whether the reading
is achieved from descending from full scale or achieved from ascending from zero. The usual
cause of hysteresis is friction or inertia.

Drift is used to describe an error that manifests itself as a change in output for a fixed input over
time. Drift can be observed by maintaining a steady input while monitoring the output.
Generally, drift is due to the effect of temperature change during the warm up of the electronic
components in a device.

Precision error is the variation between repeated measurements. The estimated standard
deviation from the mean, S, is used as a measure of the precision error. A large standard
deviation means a large scatter in the measurements. The standard deviation is defined as:
1
 N  x  x 2  2
S   i  (2-2)
 i 1 N  1 
Here, xi , x and N are respectively the individual, the estimated average, and the number of data
points. Figure 2-1 shows precision and bias errors of a set of measurements.

If the number of data points is large and they follow a normal distribution curve, and if there is
no bias, the mean is the estimated average of the data points and 2S is the uncertainty interval
with 95% confidence level or 20:1 odd. The 20:1 odd means that if the experiment is repeated 20
times, there is a possibility that 19 times the data falls within the range x  2S and there is only
once that it will be out of this range.

31
Experiments are either single sample or multiple samples. In single sample experiments, test
point is run once. In multiple sample experiments, each test point is run more than once.
Research in fluid mechanics and heat transfer are usually single sample experiments.

In single sample experiments, the precision error is dominant. In contrast, in multiple samples
experiments, the fixed error is dominant since in multiple measurements of a point, the random
errors are averaged out.

Figure 2-1

2.2. UNCERTAINTY IN EXPERIMENTAL RESULTS

In general, the results of experiments are calculated from a set of measurements:

R  R( x1 , x2 ,.., xN ) (2-3)

The uncertainty in experimental results should have the same odd or confidence level as the
measurements. The uncertainty in the result is expressed as:

1

 N  R 
2
2

R    xi   (2-4)
x
 i 1  i
  


This method is called Root Sum Square (RSS). The partial derivative of the result, R, with
R
respect to the measurement xi , , is called the sensitivity coefficient. xi is the uncertainty in
x i
the variable xi . The sensitivity coefficient weights the value of the uncertainty of the variable
corresponding to the dominance of the variable in the fundamental equation.

The RSS method can be used to estimate uncertainty in experimental results if: (a) measurements
are independent and repeated measurements of each quantity follow a normal distribution and (b)
the same odds are used for estimating the uncertainty of each measurement. The RSS equation is
generally used in the normalized form as:

32
 1
 2 2 2 2
 x  x    
u   1 R 
u 

 2 R u   ..   x N R u  
(2-5)
R  R x 1   R x 2   R x N 
 1   2   N  
 

Here, uR and ui are, respectively, the uncertainties in the result and in the measurement of xi .
Example: The following equation is used in an orifice plate to calculate the volume flow rate of
the fluid:

P
Q  CA

Here Δ𝑃 is pressure differential, ρ is density and Q is the volume flow rate. If ΔP = 250±5 Pa
and ρ=1.195± 0.01 kg/m³, find the volume flow rate and its variations. C and A are constants at
0.7 and 0.01m² respectively.
Solution:
The volume flow rate is calculated using the average or nominal values of the given data.
P 250 N m 2
Q  CA  (0.7) * (0.01m ) * 2
 0.101 m 3 s
 1.195 kg m 3

In this problem, the uncertainty for volume flow rate is a function of the pressure and density
where these two variables have its measurement uncertainty. The uncertainty equation for Q is:

1
 P Q 2
   Q  2
2

U Q    U P    U   
 Q P   Q   
Taking the partial derivative and then substitution yields
1

Q CA  P  2 1 P Q P CA 1
   →  
P 2     Q P  P  2  2
2

CA 
  
1

Q  CA  P  2 P  Q  CAP 1
   →  
 2     2
Q   P  2 
2 2
2
CA 
  
5
U P   0.02 (2%)
250
0.01
U   0.0084 (0.84%)
1.195
Thus, the uncertainty equation for Q becomes:

33
 1
 1 2
  1  2
2

U Q    U P     U     0.01084 or 1.08%
 2   2  

If both bias and precision errors are considered, first the root sum square of the biases and the
precision are calculated separately as:

1
 N  x R  2
2

S R    i
 xi  
 i 1  R xi  
1
(2-6)
 N
 x i R  
2 2

 R     xi  
 i 1  R xi  

Then, depending on the confidence level, one of the following equations is used to estimate the
total uncertainty, uR .

uR  (  R  tS R ) for 95% confidence level (2-7)

Here, S R and  R are respectively the precision and bias errors in the final result, and t is from
the student t distribution. The student t distribution is used because the distribution of the
precision error does not have a normal distribution. It has a distribution that is dependent on the
number of samples and, therefore, the degrees of freedom. Thus, t in Equation 2-7 is dependent
on the number of samples collected. The shape of the student t distribution is very similar to the
normal distribution, but has a greater width to its bell shape.

For samples higher than 30, t is assumed to be 2. For samples less than 30, t is obtained from the
student t distribution table, using the required confidence level and the number of degrees of
freedom.,  . For single sample experiments, the sample number is the number of degrees of
freedom. Table 2-1 presents the values for t in terms of the degrees of freedom for the 95%
confidence level.

34
 Figure for Table 2-1

Table 2-1

Equation 2-7 can be used to estimate the overall uncertainty, provided that the bias limit is
symmetrical about the measurement. If the bias limit is nonsymmetrical, then the upper and
lower limits of the uncertainty interval should be specified separately. Figure 2-2 shows the
measurement uncertainty for symmetrical and nonsymmetrical biases.

35
 Figure 2-2

If the number of independent sets of measurements for calculating a quantity is more than one,
then first the cumulative degrees of freedom for the measurements is calculated from the Welch-
Satterthwaite equation as:

36
 ( S i2 ) 2
vR  N 4 (2-8)
Si
i 1 vi

Then, from the student t distribution the t values are found for the 95% confidence level. The
overall bias and precision errors are found from the following equations:

S
 Si
N (2-9)
   i2
Here, S i ,  i , and N are respectively the overall precision and bias errors and the number of
independent sets of measurements. Assuming symmetrical bias, the uncertainty interval is:

u  (   tS ) (2-10)

37
 REFERENCES

1. Abernethy, et al, 1973, “Handbook, Uncertainty in Gas Turbine Measurements,” Arnold

2. Abernethy, R.B., Benedict, R.P., and Dowdell, R.B., 1985, “ASME Measurement
Uncertainty,” ASME Journal of Fluids Engineering, Vol. 107, pp. 161-164.

3. Kline, S.J., 1985, “The Purpose of Uncertainty Analysis,” ASME Journal of

4. Kline, S.J., and McClintock, F.A., 1953, “Describing Uncertainties in Single

5. Moffat, R.J., 1988, “Describing the Uncertainties in Experimental Results,”

38
 EXAMPLES

1. The Bernoulli‟s equation given below has been used extensively to obtain the mean
velocity from dynamic pressure measurement. What is the uncertainty in the U, given that the
uncertainties in U, P , and  are respectively u , p and  .

1
 2P  2
U   
  

Solution:

1
 P U    U  
2 2 2

uU    p      
U  P   U 
  

But,
1
 2P  2
1  
P U P 1  2  2 1    1
    
U  P U 2  P  2 U 2

1
 2P  2
 
1   
1
 U  1  2P  2 1 1
    
U  U 2     2 U 2

Then,

1
 1   1 
2 2
 2
uU    P      
 2   2  

2. What is the uncertainty in the Reynolds number given as:

UD
Re D 


The uncertainties in  ,U , D, and  are respectively  , U , D, and  .

39
 Solution:

1
   Re 2
  U  Re D   D  Re D     Re D  
2 2 2 2

u ReD    D
    U    D      
 Re D    Re D U   Re D D   Re D   

But,

  Re D UD 1
 1
Re D   Re D

U  Re D UD 1
 1
Re D U  Re D

D  Re D UD 1
 1
Re D D  Re D

  Re D UD  1
  1
Re D   Re D

then,


u ReD      U   D     
2 2 2 2

1
2

4. What is the uncertainty of the increase in length L, with a wire with diameter D, and a
load of F?

Uncertainty in diameter is d, the length is L, and the uncertainty in F is F. Assume
that the E modulus of elasticity has no uncertainty (E=0).

L
 e  L  L  e
L

40
 1/ 2
 L   L
2
 
2

L    L     e  
 L   e  


e
E

1
 e 
2
 e 
2
e
e         E     
 
   E  

F

A

1/ 2
    
2
 
2

    F     A  
 F   A  

Substitute partial derivatives and rewrite,

1/ 2
 1  F
2
 
2

    F    2  A  
 A   A  

1/ 2
1  1  F  
2 2

e     F    2  A  
E  A   A  

Hence,

1/ 2
 L2  F  F
2
 
L  e  L 2  2    F    2  A  
 E  A   A  

41
4. A hook carries a load of F pounds. What is the uncertainty of the calculated maximum
stress in the vertical portion of the hook? The hook has an a x a square cross-section and the
radius in the bend of the hook is R. The uncertainty in length measurement is a, in radius is R
and the uncertainty in force measurement is F.

Substituting into the uncertainty equation, :

F My
 
A I

1/ 2
    
2
  
2 2
     
2
 
2

    A     F     y     M     I  
 A   F   y   M   I  

A = a2

1/ 2
 A  2 
A      2  a  a
 a  

a
y
2

1/ 2
 y  
2
1
y    a     a
 a   2

M  FR

1/ 2
 M  
  M
 
2 2

 R  F   ( F  R) 2
1/ 2
M    F     R  
2

 F   R  

b  h3 a 4
I 
12 12

42
 1/ 2
 I  
2
4  a3
I    I     a
 a   12

2 1/ 2
 2  F  a 
    M a 4a 
2 2 2 2
 1  M 1   a
3
 R  F   F  R 
1/ 2
    a     F      a   
2 2
     a  
 A  A   I 2   2 I   2 I 
2 2
12 

Rearrange,

1
 2  F  a  
 
2 2
 1 M  a4
2 2 2 2
 1  M 1   a 
    a     F      a      R  F   F  R      a  
2 2

 A  A   I 2   2 I   
2
6 I2

43