Spring 2000 AE 3145 Measurements Lab 1

Mechanical Measurements
Measurements
Direct/Indirect comparison (rulers, balance scale, interferometer
Calibrated system (odometer, spring scale, pressure gage)
Measurement System
Issues:
Detector/Sensor: device which detects and responds to measurand
Transducer: converts measurand to an analog more easily measured
(force-displacement-resistance-voltage)
Signal Cond.: amplify, filter, integrate, differentiate, convert freq. to
voltage, etc.
Computer: widely used today
Detector-
Sensor
Transducer
Signal
Conditioner
Indicator,
Recorder,
Controller,
Computer,
etc.
Physical
Measurand
Spring 2000 AE 3145 Measurements Lab 2
Example of Measurement System
These are simple mechanical systems
Issues
mechanical vs electrical output
analog vs digital
calibration
accuracy, precision, resolution, sensitivity, linearity, drift, backlash?
detector
transducer
readout
detector
transducer
readout
Tire Pressure Gage Bourdon Type Pressure Gage
Spring 2000 AE 3145 Measurements Lab 3
Computers Readout Systems
S/H ADC
Control
Voltages digital words
to computer
Computer
MUX
S/H ADC
Control
Voltages
Computer
S/H
S/H
S/H
S/H
S/H
S/H
MUX
Sequential Sampling Data Acquisition
Simultaneous Sampling Data Acquisition
Key:
MUX =multiplexer (switch)
S/H =sample & hold (hold voltage while ADC reads)
ADC =analog to digital converter (voltmeter)
Volt
t
Chan #1
Chan #2
skew
sample
Readings are not
recorded simultaneously
digital words
to computer
Volt
t
Chan #1
Chan #2
No skew
sample
Readings are
recorded simultaneously
Spring 2000 AE 3145 Measurements Lab 4
Design of an Experiment
Not always a sequential process!
Proof of Concept Experiment (POCE) is VERY important
uncover problems before full scale experiments begin
learn about how experiment will work
minimize risks!
Conceptual
Design
Preliminary
Design
Experiment
Evaluate-
Analyze-
Report
Problem
Literature
Physics
Facilities Experience
Subsystems
Hardware
POCE POCE
Signal
Conditioning
& DAQ
Purpose?
Style
Spring 2000 AE 3145 Measurements Lab 5
Experimental Data
All measurements have uncertainty to some degree
error = = x
meas
- x
exact
Objective: minimize error so that
-u +u at N:1 certainty
or
x
m
- u x
exact
x
m
+ u at N:1 certainty
Types of error:
Accuracy &
Precision
Precision
(can calibrate)
Accuracy
(can average)
Neither
Spring 2000 AE 3145 Measurements Lab 6
Terms Used to Define Performance
Accuracy: difference between measured and true values;
typically specified by a maximum value.
Precision: difference between measured values during
repeated measurements of the same quantity.
Resolution: smallest increment of change in the measured
value that can be determined reliably (repeatedly).
Sensitivity: change of an instruments output per unit change in
the measured quantity; may or may not be related to above
quantities for a given instrument.
Spring 2000 AE 3145 Measurements Lab 7
More Types of Error
x
y
Ideal
Actual
x
y
Ideal
Actual
time
y
Ideal
Actual
x
y
Ideal
backlash
hysteresis
x
y
Ideal
Actual
Can be calibrated out...
Backlash & Hysteresis Drift
Quantization Error Nonlinearity
(may be able to calibrate)
Spring 2000 AE 3145 Measurements Lab 8
Statistical Approaches
Probability & Statistics provides a way to deal with uncertainty
we will cover only a VERY limited introduction
this is the subject of courses in Math & ISyE (good electives!)
F
r
e
q
u
e
n
c
y
bias error
precision (random) error
x
m
X
avg
X
exact
total error
X
m
number of times
reading falls in
this range of x
theoretical distribution
Bias error: systematic errors that can be
removed by calibration
(e.g., gage is 6% high)
Precision error: random error that is
not directly controllable
(by definition) but we seek
ways to minimize by careful
design and statistical analysis
Bias:
calibration
consistent human error
loading error
Precision:
disturbances
noise
variable conditions
Blunders:
human error
software!
Either:
backlash, hysteresis, friction
damping
drift
variations in test procedure
Spring 2000 AE 3145 Measurements Lab 9
Some Basic Concepts
Sample Space: all possible events or outcomes of an experiment;
also can be referred to as a population.
Event: subset of sample space
Sample: finite number taken from population (e.g., 15 turbine
blades taken from 8600 produced for XX-300 engine)
sample must be randomly selected from population
sample may or may not be returned to population before resampling
Probability: likelihood of an event (measured as % of successful
events in sample space or population). We can often analytically
compute this likelihood based on mathematics applied to the
population. 0 P(event) 1.
There are many rules for computing probabilities for different kinds
of events; these are the subject of courses and books on
probability theory.
Spring 2000 AE 3145 Measurements Lab 10
Examples
These happen to be binomial distributions (repeat experiment N
times; each trial is independent of others; each trial is successful
or not; probability of success, p, is same for each trial).
0 50 100
Number
N
u
m
b
e
r

o
f

H
e
a
d
s
Heads in 100
coin tosses
1 2 3 4 5 6
Number
1
N
u
m
b
e
r

o
f

E
v
e
n
t
s
Numbers on one die
2 3 4 5 6 7 8 9 10 11 12
1
N
u
m
b
e
r

o
f

E
v
e
n
t
s
6
Number (sum of dice)
Sum of numbers on 2 dice
)! ( !
!
) 1 ( ) (
x N x
N
x
N
p p
x
N
trials N in successes x P
x N x

=

Spring 2000 AE 3145 Measurements Lab 11
Properties of Probability Distribution Functions
0 ) (
0 . 1 ) (
6 . 0 ) (
1
=
=
=
x P
x P
also
x x P
CDF =Cumulative Distribution Function =P(x x
1
) PDF =Probability Distribution Function =f(x )
( )
=
=
=
=

2
1
1
) ( ) (
) ( ) (
) (
) (
2 1
1
x
x
x
dx x f x x x P
and
dx x f x x P
So
dx x f CDF
or
CDF
dx
d
x f
x
1
x
Area =0.6
F(x)
PDF
x
1
x
0.6
1.0
CDF
Spring 2000 AE 3145 Measurements Lab 12
Normal or Gaussian Distribution
Uses:
When N for a binomial
distribution (e.g., for very large
populations)
For events that are made up
from many independent events
each with any kind of
distribution
For properties of samples when
# samples is very large (e.g.,
sample means)
many measurement problems
others

x
F(x)
2
1

=
2
2
2
) (
exp
2
1
) (

x
x f
=mean

2
=variance
=standard deviation
2
2
2
1
) (
z
e z f

=

Books show tabulated Z-distribution for

mean () = 0 and =1

=
x
z
where
To use tabulated Z values, change variable to z
To use tabulated Z values, change variable to z
MS Excel:
(a) =NORMDIST(x, mean, std-dev, CDF/PDF) where x=value and CDF/PDF=TRUE for CDF or FALSE for PDF)
(b) =NORMSDIST(z) this is the Z-distribution form (normalized for mean=0 and std-dev=1)
(c) =NORMINV() and =NORMSINV() are inverses (see Excel Help)
You can use your
spreadsheet!
You can use your
spreadsheet!
Spring 2000 AE 3145 Measurements Lab 13
Examples of Normal Probability Calculations
997 . 0 ) 3 3 ( Pr
954 . 0 ) ( ) ( ) 2 2 ( Pr
683 . 0 ) ( ) ( ) ( Pr
2
2
2
2
1
1
= +
= = = +
= = = +
+

ob
dz z f dx x f ob
dz z f dx x f ob
One Sigma:
Two Sigma:
Three Sigma:
Six Sigma is used for
many electronic
manufacturing processes
Error Level Name Error Level Prob. that
Error is
Smaller
Prob. that Error
is Larger
Probable Error
0.67
50% 1:2
One Sigma

68% ~1:3
90% error
1.65
90% 1:10
Two Sigma
1.96
95% 1:20
Three Sigma 3 99.7% 1:370
Maximum Error 3.29 99.9 1:1000
Four Sigma
4
99.994% 1:16000
Six Sigma 6 99.9999999% 1:1.01e9
Use MS Excel functions
to compute Probs.
Spring 2000 AE 3145 Measurements Lab 14
Sample Statistics
Previous results assumed that we know and for the
distribution. What if this is unknown (as is often the case)?
Note: we use N=1 for s
2
because we have N independent x
i
but
if we also know x
mean
we could need only know N-1 x
i
and could
then compute the last remaining x
i
result. There are only N-1
degrees of freedom for this calculation!
1
1
) (
1
2
2
1
2
2
1

= =
= =
= =
=
N
N
x
N
x
N
x x
s variance Sample
N
x
x mean Sample
N
i
i
N
i
i
N
i
i
Mean Square Square of mean
Spring 2000 AE 3145 Measurements Lab 15
Central Limit Theorem
This is one of the most important statistical theorems.
Is the basis for much of the statistical methods we can
commonly apply to measurements
Suppose we test 100 bolts and the average (mean) failure
strength is 75 ksi. We repeat this test with 100 more bolts and
get 78 ksi. We can repeat many times and ask what the
distribution for the mean average values will be.
ANSWER: Gaussian (normal) with
This is also the distribution for any random variable which is the
result of many independent random variables no matter what
the underlying distributions may be.
N
x

=
Spring 2000 AE 3145 Measurements Lab 16
Application: Confidence Intervals
A common question is to ask if one takes N readings and
computes the sample average, how confident can you be that
the average is really close to the true mean ()?
Confidence intervals are a way to describe this; we know that
the sample mean for many different samples will be normal.

+
N

Prob = c% that lies in shaded area defined by
x
c
a x =
x c
x a c Example 96 . 1 , 96 . 1 %, 95 : = = =
x
in fall readings x of Answer 96 . 1 % 95 : 1
x
x within fall will
true the confidence With Answer

96 . 1
, % 95 : 2

Spring 2000 AE 3145 Measurements Lab 17

Application: Confidence Intervals - contd (1)
Since where = true population standard
deviation, we can write:
For large sample sizes (large N), we can approximate with s
x
so that we have:
N
x
=
N x limits confidence Answer 96 . 1 : % 95 : 3
N S x limits confidence Answer
x
96 . 1 : % 95 : 4
EXAMPLE
Find 95% confidence limits for x
avg
=75 ksi when S
x
=8.3 ksi for N =50 samples.
ksi N S x limits Answer
x
3 . 2 75
50
3 . 8
96 . 1 75 96 . 1 : = = =
Find 99.9% confidence limits for previous case (use previous table):
ksi N S x limits Answer
x
9 . 3 75
50
3 . 8
29 . 3 75 29 . 3 : = = =
Spring 2000 AE 3145 Measurements Lab 18
Application: Confidence Intervals - contd (2)
Remarks
Above works only for N=large when is known.
When N=small, then we cannot approximate by S
x
due to N-1 in
the denominator and we must treat as unknown. This leads to
replacing the normal distribution with the Student T distribution.
This will not be discussed in these notes.
EXAMPLE (contd)
How many N are required to assume that m is in x
avg
5% with 95% confidence?
ksi
N
limits Answer 05 . 0 * 75 75
3 . 8
96 . 1 75 : = =
samples N or 19 8 . 18 : =
Spring 2000 AE 3145 Measurements Lab 19
Analysis of Uncertainty
Many times experimental results are the result of several
independent measurements combined into a single quantity
using a theoretical formula (e.g., w=PL
3
/3EI for a cantilever).
How do uncertainties in each variable contribute to the whole?
If y=y(x
1
, x
2
, x
N
) is a linear function, a statistical theorem states that:
2 / 1
2 2
2
2
1
2 1
1

|
|

\
|

+
|
|

\
|

+
|
|

\
|

=
N
x
N
x x y
x
y
x
y
x
y

For uncertainties, u
i
, that are small compared to x
i
we can use a Taylor Series expansion in u
i
:
2 / 1
2 2
2
2
2
1
1
(

=
N
N
y
u
x
y
u
x
y
u
x
y
u
N
N
N N N
u
x
y
u
x
y
u
x
y
x x x y u x u x u x y

+ = + + + ... ) ,... , ( ) ... , , (

2
2
1
1
2 1 2 2 1 1
so that y is now a linear function of the uncertainties. Applying this in the first equation yields:
Spring 2000 AE 3145 Measurements Lab 20
Bias and Precision Uncertainties
We noted earlier that errors will include both bias (offset) and precision
(random) components
These can usually be treated as independent and therefore the
uncertainties for each can be combined into a total:
Note: if they are not independent, other combining in other ways may
be necessary.
SINGLE SAMPLE: if we are considering a single calculation of the
dependent result (y), we must assume N=1 so that:
This affects only the precision error which is purely random so:
Note: for 95% confidence interval (1.96 ), a
c
=1.96
[ ]
2 / 1
2 2
bias y precision y total y
u u u

+ =
y c
y
c c
u a y
u
a y
N
a y y = = =
1

y y c sample single y
u u a u 96 . 1 = =

( ) [ ]
2 / 1
2 2
bias y precision y c sample single total y
u u a u

+ =
Spring 2000 AE 3145 Measurements Lab 21
EXAMPLE: Uncertainty Calculation
Consider a load transducer
made by attaching a strain
gage at the root of a
cantilever beam as shown.
A tip force, P, will produce a
bending strain,
x
, as shown.
Assume the electrical output
of the strain gage is e=K
x
e
0
where e
0
is the excitation
voltage and K is the calib.
factor to be determined.
The uncertainty in K is:
L
P
h
b
Strain gage
E h b
L P
E
I
y M
E
b
x x
2
6
/ / = = =
L P
E h b
e
e
e
e
K
6
2
0 0
= =

2 / 1
2 2 2 2 2
2
0
2
2
0
(

=
L
u
P
u
E
u
h
u
b
u
e
u
e
u
K
u
L P E h b
e
e K
Spring 2000 AE 3145 Measurements Lab 22
Variable Bias (u
x
/x) Precision (u
x
/x) Notes
e 1% 0.2% Digital DVM yields good precision
e
0
0.5% 0 Only single initial measurement made
b 0.5% 0 same
h 0.5% 0 same
E 2% 0 same
P 1% 2% Bias reflects calibration errors while
Precision is random error in readings
L 0.2 0 Only single initial measurement made
Uncertainty 2.65% 2.01%
Single sample 2.65% 3.94%
Example: Uncertainty Calculation - contd
See Example 3.12 in textbook for another case
Assume the following uncertainties:
Total uncertainty in the measurement of K is then:
[ ] % 75 . 4
2 / 1
2 2
= + =
bias precision
K
u u
K
u