7/29/2019

POLYTECHNIC UNIVERSITY
OF PUERTO RICO - SAN JUAN
FLUID MECHANICS LABORATORY
ENGI 2421

EXPERIMENTAL DATA ANALYSIS

Instructor: Dr. Moises Angeles

1
 7/29/2019

UNCERTAINTY CONCEPTS

• You make a number of repeated measurements of the same

measurand
– e.g. to measure the air temperature of a room (measurand), you can repeat the
same measurement many times

• Ideally all these repeated measurements should give exactly the

same value and this value should be equal to the true value of the
measurand

• In reality the results of the repeated measurements always

differ to some extent and their mean value also usually differs from
the true value

2
 7/29/2019

What is the uncertainty?

3
 7/29/2019

UNCERTAINTY CONCEPTS
• Error: difference between the actual value
and the measured value measurements
true value of the same
measurand
• Uncertainty: Probable value of the error in
the experimental measurement (deviation
degree)

• Random error: inevitable error, impossible

to control during the measurement process
(personal fluctuations, random electronic
fluctuations in the device and instruments,
friction, etc.

• Systematic errors: error that is generated in

the same amount in all measurements. It
can be originated by instrument defect or Source: [Link]
bad calibration

4
 7/29/2019

What are the Statistics?

5
 7/29/2019

STATISTICS
• “n” repeated measurements xi of the  Let to have “n” random measurements
same measurand
 Eg. the data of ages of a randomly selected population of 10
• Arithmetic mean people: 8, 19, 58, 35, 45, 12, 6, 13, 18, 47
n: number of  Class interval: a range of values open in the right hand side
= sample data
=5 k: number of classes, n: number of data

• Population standard deviation For n= 10, k=5. Select 6 intervals of [ ⟩

Bin Class interval
N: total number Relative Cumulative
= − of population Frequency Frequency
 Frequency (Fi): Class Frequency
data number of observations 0-10 2
that fall within each
• Sample standard deviation (unbiased) (events) 10-20 4
20-30 0

n: number of 30-40 1
= −
− sample data 40-50 2

In a sample, “N-1” gives unbiased estimate of the 50-60 1

population standard deviation 60 and
When n  ∞ (large data), ⟶ above 0

6
 7/29/2019

FREQUENCY

 Relative Frequency: frequency in any class interval Relative Frequency Histogram

or event (Fi) divided by the total number of
observations(n)

 Cumulative Frequency: is the sum of the all

relative frequency
Relative Cumulative
Frequency Frequency
Class Frequency Ages
0-10 2 0.2 0.2
Cumulative Histogram
10-20 4 0.4 0.6
20-30 0 0 0.6
30-40 1 0.1 0.7
40-50 2 0.2 0.9
50-60 1 0.1 1

0 1
60 and above 0
Ages
Total 10
7

7
 7/29/2019

STATISTICS
standard error is the standard deviation of the
sample arithmetic averages (every measurement is
assumed as the mean value for a sample)

= Population standard error

≈ Sample standard error

Standard error for small samples tend to systematically

underestimate the population standard error.

For large data set, the underestimate is reduced

8
 7/29/2019

How I can calculate the confidence

interval?

9
 7/29/2019

STATISTICS

• Confidence interval: it is a range of values we are fairly sure our true value
lies within the interval ̅ − ≤ ≤ ̅+ for some confidence level (1-α)

= ± is the average of the repeated measurements

is the uncertainty
best estimate of the margin of error
actual value for a given or uncertainty
probability (confidence
level, − )

= ∗ Uncertainty for large number of data set

= ∗ Uncertainty for small number of data set (n≤10)

10

10
 7/29/2019

CONFIDENCE INTERVAL

Example: For a given confidence level of 95%, determine the confidence interval if the
average value is 175cm and the uncertainty is 6.2 cm

Average
uncertainty
measured
data

Confidence interval with

95% of confidence level

95% of experiments will include the

true mean, but 5% won't

11

11
 7/29/2019

How is calculated the z and t

score?

12

12
 7/29/2019

NORMAL DISTRIBUTION
For any μ and σ

−
= z-score

μ=0; σ=1

Standard Normal Distribution, where the Cumulative

Distribution Function (CDF) is the Probability that a random
value is within a z-score interval (confidence level − )

−1 < <1
=
−
/ /

: significance
level

What is the z-score for a probability (1 − ) of

− < < = 68.27%
Answer: the z-score becomes z=1.0
13

13
 7/29/2019

t-STUDENT DISTRIBUTION ,
−
, = t-score with n-1 degree of . .
freedom

n-1: number of data sample minus 1

sample standard deviation
− : confidence level

(1- ) is the probability that a random value is

within a t-score interval

− = , < < ,

/ / : significance level

What is the t-score for n=10 and a confidence level

of 90%
For n-1=9 and 90% confidence level
 , % = .

14

14
 7/29/2019

SUMMARY CONFIDENCE INTERVAL CALCULATION

Procedure
1. With all data calculate the average and the
standard deviation and the standard error Confidence level
−

= = − = Level of
−
significance = Level of
α/2 significance =
2. Choose a confidence level 1-α (for errors you α/2
can use the normal distribution to calculate 1-
α)

3. In the table of probability, for a given 1-α find

the corresponding z-score or t-student

4. Calculate For a confidence level of 95%, the z-

score is 1.96
= ∗ or = ∗

15

15
 7/29/2019

DETERMINE IF THE DATA FOLLOW A NORMAL DISTRIBUTION

 Denote by x the “n” repeated measurements
 Determine if your data have a normal distribution
 In any experimental measurement, you must confirm you have a normal
distribution
A normal distribution must satisfy simultaneously the following three conditions

N1  0.68n  1.41* n
N 2  0.95n  0.654 * n
N 3  0.997 n  0.164 * n
N1 = data number between − and +
N2 = data number between − and +
N3 = data number between − and +
n = total data number and is the arithmetic average of all data

16

16
 7/29/2019

OUTLIERS
 Outliers are measurements that differ too much from  Chauvenet’s criterion =
expectation, which are commonly removed to avoid
Eliminate the measurement if
damaging impact on processing.
−
 Outliers detection: calculate how much a measurement >
xi differs from the average x of the sample
measurements. The difference is normalized by dividing it
with the standard deviation σ (z-score)
−
=
 The simplest outlier
detection uses a
fixed threshold θ to
determine whether or
not a measurement is
an outlier

the outlier criterion is

>

17

17
 7/29/2019

EXPERIMENTAL CALCULATION OF UNCERTAINTY

Try to minimize the systematic

error

−
>

Calculate standard error and

uncertainty propagation

Typically for z = 1, 2, 3
68%, 95%, 99%

18

18
 7/29/2019

EXPERIMENTAL CALCULATION OF UNCERTAINTY

•Experience with instrumentation

•Manufacturer information
•Estimate the least count of the
instrument

Example: using a thermometer with a scale of 0.50C a

temperature of 270C was measured
T = 27 ± 0.250C

19

19
 7/29/2019

UNCERTAINTY PROPAGATION

= , , , ∗
=
∗

Volumetric flow is calculated indirectly

measuring mL, ρ, T, t, Lr

Errors in: mL, ρ, T, t, Lr

There are partial uncertainties for every
variable measured

Total uncertainty in Q:
Uncertainty for every variable propagates to
the final result Q

20

20
 7/29/2019

UNCERTAINTY PROPAGATION
Let to Q be a function of several variables

= , , ,…,
Multiplication, Division or a combination
Addition and Subtraction …
=
= ± ± ±⋯ …
Relative uncertainty
= + + +⋯
= + + ⋯+ + +⋯
General function of multivariables
= , , ,…
Where function Q is
= + + +⋯ known

= ±
21

21
 7/29/2019

UNCERTAINTY PROPAGATION

For the equation

∗
Q=Q(mp, ρ, t, Lr) =
∗
– Weight relative uncertainty
um= 0.1%

– Beam length (Lr) uncertainty

uL= 0 (neglected)

– Relative uncertainty for the volumetric flow

−
= + + + = ≈
−
and are closer to the mean
density ̅

22

22