Download as ppt, pdf, or txt
Download as ppt, pdf, or txt
You are on page 1of 17

Useful Tips for Presenting Data and

Measurement Uncertainty Analysis

Ben Smarslok

University of Florida 1
Mechanical and Aerospace Engineering
Outline
• Why is presenting data properly important?

• Explain important terminology and definitions


– NIST vs. ISO vs. ASME/ASTM
– Oberkampf definitions of model uncertainty (not included)

• Experimental scenarios and corresponding methods


– Uncertainty propagation
– Crossed vs. nested factors (ANOVA vs. VCA)
– p-values
– Interlaboratory Studies (not included)

University of Florida 2
Mechanical and Aerospace Engineering
m = 75 ± 5 g
What is the meaning of ± 5 ?
• Best guess by experimenter
• Half the smallest division of measurement
• Standard deviation: 
• Standard error: m = /n
• Expanded uncertainty of ± 2 or ± 3 (95% or 99%
confidence interval)
• Standard uncertainty: u
• Combined standard uncertainty: uc

*(Courtesy of Duane Deardorff presentation from UNC)

University of Florida 3
Mechanical and Aerospace Engineering
What does x ± u mean?
• Engineers think in terms of ±2 (95%)
• Physicists generally report ±1 (68% CI)
• Chemists report ±2 or ±3 (95% or 99% CI)
• Survey/poll margin of error is 95% CI
• Accuracy tolerances are often 95% or 99%
• NIST Calibration certificate is usually 99%

• Conclusion: The interpretation of ± u is not consistent


within a field, let alone between fields
– It is very important to explain the statistical relevance of the
uncertainty bounds!!!

University of Florida 4
Mechanical and Aerospace Engineering
Presenting Uncertainty Precisely
• Choose a standard for presenting uncertainty (I prefer
NIST), and reference the standard
• Explain the source of the uncertainty
– Type A – calculated by statistical methods (it is useful to explain
the design of experiments and the number of samples involved)
– Type B – determined by other means, such as estimate from
experience or manufacturers specifications
• Use terms carefully!
– Error vs. Uncertainty: Error is the deviation from the true value and
measured value (never known), which is estimated as uncertainty
– Bias vs. variability (will explain later)
• Avoid use of ambiguous ± notation without explanation
sx
• Pet peeve: CV  coefficient of variation 
– COV = covariance x

University of Florida 5
Mechanical and Aerospace Engineering
NIST Classification of Measurement Uncertainties
• Uncertainty classification:
– Random uncertainty /
variability – scatter in the
measurements (v)
– Systematic uncertainty / bias –
systematic departure from the
true value (b)
• Type of evaluation:
– Type A – calculated by sx
v  s  xt = true value of specimen
statistical methods x x N  = experimental population average
– Type B – determined by other
x = experimental sample average
means, such as estimate from
experience vx = random error of sample
B Range is at 95% (2) level x = systematic error of sample
b
2 of a normal distribution

University of Florida 6
Mechanical and Aerospace Engineering
Uncertainty Analysis Example
• Consider our transverse modulus work (E2)

• Hooke’s Law:
P = Load
P 2

E2  A = Area A  w  t
A 2
1

= transverse strain

• We will work through this problem backwards

Total Bias &


Components Contributors
Uncertainty Variability

University of Florida 7
Mechanical and Aerospace Engineering
Level 1: Total Uncertainty
• In general,
U X 95   t95 (v X )2  (bX ) 2
– where, vX and bX were propagated from component uncertainties
t95 = Student’s t distribution at 95% confidence level
(depends on # of DOF)

• Total uncertainty of E2 at 1 (68%) confidence for


comparison to experimental results
U 68 E 2   (v E 2 ) 2  (bE 2 ) 2

• Or, at the commonly accepted 95% level


E2,95%  9.01  0.12GPa

University of Florida 8
Mechanical and Aerospace Engineering
Level 2: Uncertainty Propagation
• Law of Propagation of Uncertainties (LPU):
2
n
 q  2 n n
 q   q 
u (q )   
2
 u ( pi )       u ( pi , p j )
i 1  pi  i 1 j 1  pi   p j 
– where, p are the inputs (components) and q is the output
• E2 Example:
– Uncertainty contributors were analyzed for each of the components
of E2
– Random and systematic effects propagated separately
– Only systematic uncertainties can have correlated effects
• Thickness and width are correlated
2 2 2 2
 E   E   E   E 
v ( E 2 )   2  v T 2 ( P )   2  v T 2 (  2 )   2  v T 2 ( t )   2  vT 2 ( w )
2
 P    2   t   w 
2 2 2 2
 E   E   E   E   E   E 
b2 ( E2 )   2  bT 2 ( P )   2  bT 2 ( 2 )   2  bT 2 (t )   2  bT 2 ( w)  2  2   2  bT (t )bT ( w)
 P    2   t   w   t   w 

University of Florida 9
Mechanical and Aerospace Engineering
Level 3: NIST Component Measurement Uncertainty Table

University of Florida 10
Mechanical and Aerospace Engineering
Level 4: Contributors of Component Uncertainty
(Further Analysis)
• Numerous different methods to analysis the significance of
uncertainty contributors
• It is important to use the appropriate analysis method
depending on the design of experiments (DOE)
– Either design the experiments properly or match the corresponding
method to the data you already have
• Most DOEs fall into one of these two categories:
Crossed Nested

Same patients in each hospital. Patients unique to each hospital.

University of Florida 11
Mechanical and Aerospace Engineering
Crossed Design: ANOVA
• Crossed (or factorial) DOEs correspond to analysis of
variance (ANOVA)
• Consider thickness in the E2 example
– Since the SAME specimens were measured in the SAME positions
with the SAME users, then the factors were crossed
– 3-way ANOVA with crossed, random variables was conducted

Uncertainty contributors:
Nominal: 0.09 x 1 in. Specimen – variability from specimen
to specimen
Position – variation across
measurement surface
User – error from user technique
Measurement repeatability

University of Florida 12
Mechanical and Aerospace Engineering
Thickness ANOVA
• 3-way ANOVA of crossed, random variables
– Statistical software available for ease of use:
Excel for 2 factors or SAS for 3 or more
• Factors:
– A = specimen a=4 Hypothesis Test for A:
– B = position b=3 H 0 :  2  0
– C = user c=4 H a :  2  0
– Repetitions: n=3
• ANOVA model:
yijkl    i   j   k  ( )ij  ( )ik  (  ) jk  ( )ijk   ijkl

• ANOVA results were not directly used in uncertainty analysis, but were
used to identify significant contributors and validate uncertainty
estimates

University of Florida 13
Mechanical and Aerospace Engineering
Results: Thickness ANOVA

• Use ANOVA to deterimine the


significance
~
of the
~
contributors
~

of uncertainty in thickness
• Position is most significant
factor with p-value = 0.006
• Not as interested in interactions
in this study
• Used to validate estimated
range of uncertainties of
thickness and width

University of Florida 14
Mechanical and Aerospace Engineering
Nested Design: VCA
• Nested DOEs correspond to variance component analysis (VCA)
• Consider a two-stage nested design of one specimen for thickness
– Relevant if positions and users were unique each time
– Specimens considered individually since the thickness does not have to be
the same from one specimen to the next
– Data was organized according to position
– y1, y2, and y3 refer to the repeated basic measurements

University of Florida 15
Mechanical and Aerospace Engineering
Variance Component Analysis of Thickness
• Goal: Develop a nested design to determine the
contribution of each factor in the overall variance
 process
2
   i2 where, i is a component in the process
i

• Variance of the measurement process for one specimen


 mp   bm   u   p
2 2 2 2

– Position – the three locations on the specimen where the


thickness was measured (unique to each specimen)
– User – four different users per position performed the
measurements
– Basic Measurement – three repeated measurements by each user
at each position
• Compare the weight of each contributor to determine
significance

University of Florida 16
Mechanical and Aerospace Engineering
Concluding Remarks
• Using proper statistical terminology and representation is
necessary to have meaningful results
• You can say your results are “pretty good”, but give what
your definition of “pretty good” is!
• Depending on the project, more or less uncertainty analysis
may be required
• It is important to design your experiments with the
statistical analysis in mind
• Age-old question: How many measurements do I need?
– Obviously depends on the circumstances, so there is no straight
forward answer
– Best recommendation: Feel comfortable enough with your results
that you can predict the next measurement within a desired range

University of Florida 17
Mechanical and Aerospace Engineering

You might also like