IE380 Unit 10

You might also like

Download as pdf or txt
Download as pdf or txt
You are on page 1of 33

IE380

Quality Control and


Improvement
Unit 10
DOE-Fractional Factorial Modeling, Experimentation, Calculation
Confounding, resolution, and mirror images
Summary of Last Time
• DOE- 2𝑘 full factorial experiments
a) Treatment of experimental variation d) Modeling issues
o Blocking. o Which factors are significant?
o Randomization. o Replication, noise estimation, t-tests.
o Inclusion into experiments. o Residual analysis.
b) Setting up experiments: o Recovering process settings from experimental
o Choice of model. values.
o Choice of variable levels. o Validation
c) Calculation of effects.
o Location effects.
o Cross effects.
• To determine whether terms are
o Calculation of 𝑏𝑖 ’s. necessary: (a) t-test (b) F-test
Fractional Factorial Experimentation
If I have four factors (location, temperature, fixative, duration) which I think may be
important, I must run 16 experiments to completely investigate their effects.
But what if I only have the time/people/budget to perform eight experiments?
• I could guess the three significant. Problems?
Might guess wrong!
• I could run two separate 22 full factorial. Problems?
Location vs temperature fixative vs duration
22 = 4 22 = 4 4+4=8
miss cross effects!
• I could run a 24−1 fractional factorial experiment? What is this??
Fractional Factorial Experimentation
Before we define 24−1 fractional factorial experiment, let’s consider the following
question:
If we can actually run 16 experiments instead of 8, should do a full 24 experiments?
Yes! Complete information
No! The we have no replications.
• Ok, what is 24−1 fractional factorial experiment?
4: The total number of variables being tested. In this case location,
temperature, fixative, and duration.
1: The reduction in power of the test. Therefore the format for this test is
actually that 24−1 = 23 full factorial experiment. I will perform 23 = 8 tests.
4−1
2 Fractional Factorial Experiments
• Is this for free? Why not do this all the time?
No. Lose information
Confounding
• How are we going to set it up? We start with a standard 23 ful factorial
experiment.
This is the standard 23 calculation matrix.

What is the least useful column?


(This is misleading! We will talk about this
more later)
𝑿𝟏 𝑿𝟐 𝑿𝟑
4−1
2 Fractional Factorial Experiments
• So will assign this to variable 4.
• This yields a design matrix: 24−1 𝑑𝑒𝑠𝑖𝑔𝑛 𝑚𝑎𝑡𝑟𝑖𝑥

Test I 𝑿𝟏 𝑿𝟐 𝑿𝟑 𝑿 𝟏 𝑿𝟐 𝐗 𝟑 = 𝑿𝟒
1 1 -1 -1 -1 -1
2 1 +1 -1 -1 +1 This tells operators where to
3 1 -1 +1 -1 +1 set the values
4 1 +1 +1 -1 -1
5 1 -1 -1 +1 +1
6 1 +1 -1 +1 -1
7 1 -1 +1 +1 -1
8 1 +1 +1 +1 +1
4−1
2 Fractional Factorial Experiments
• And a calculation matrix

How did we get these columns?


Multiply! 123 = 1 ∙ 2 ∙ 3 (component-wise)
4−1
2 Fractional Factorial Experiments
• What do you notice about this matrix? In particular the columns?

Symmetric about 14|23


Every column has a ‘twin’ or an ‘alias’. E.g. 𝐼 = 1234 and 1 = 234
4−1
2 Fractional Factorial Experiments
• What does this mean? How will this effect our estimates?
For example, to find 𝐸1 we would multiply result data by column one and divided by four.
In doing this, am I getting 𝐸1 , or 𝐸234 ?
𝐸1 =< 𝑋1 ∙ 𝑦𝑖𝑒𝑙𝑑 >÷ 4 = 17.6
𝐸234 =< 𝑋234 ∙ 𝑦𝑖𝑒𝑙𝑑 >÷ 4 = 17.6
17.6 is the main effect of location PLUS the cross effect of temp, type, duration.
𝐸1 + 𝐸234 = 17.6 ⇒ 𝐸1 = −100, 𝐸2 = 117.6
𝑜𝑟 𝐸1 = 17.6 , 𝐸234 = 0
• This effect is called confounding, the mixture of different effects in the same statistics.
• In this particular case the main effect of variable 1 and the three-way effect of variables
2, 3, and 4, but there is other confounding in the above matrix as well.
• This is because certain columns are aliases.
4−1
2 Fractional Factorial Experiments
• If you do the standard calculations (using 𝑙′s instead of 𝐸's, to denote the fact that we
are "fractional"):
𝑙0 = 𝑚𝑒𝑎𝑛 + 1234 2 1234 =< X1234 ∙ 𝑦𝑖𝑒𝑙𝑑 >÷ 4
< X1 ∙ 𝑦𝑖𝑒𝑙𝑑 >÷ 4 𝑙1 = 1 + 234
𝑙2 = 2 + 134
𝑙3 = 3 + 124
𝑙12 = 12 + 34
𝑙13 = 13 + 24
𝑙23 = 23 + 14
𝑙123 = 123 + 4
• What now?
1. Remove higher order effects -3 factor or greater
2. Process knowledge –Know 𝐸14 = 0
3. Run more experiments
4−1
2 Fractional Factorial Experiments
• So here the trade-off of fracturing becomes apparent: We get estimates of a
larger set of different variables, but we pay the price of confounding, which
clouds our estimates.
• Why do we see confounding?
• How many unique combinations of the four variables are there if we are testing each at 2
levels?
4
2 = 16
• How many test did we do?
8
• So in effect we have eight equations and sixteen unknowns. This cannot
completely specify the system.
Confounding Relations and Resolution
• Before we examine how to mitigate the effects of confounding, we will help
ourselves notationally, by introducing the concept of a confounding relation.
• To determine the confounding structure:
1. Start with the duplicate columns in the Design Matrix
In the above example, 4 = 123
2. Square one side. (Note that 4x4 = I Why?)
This yields 𝐼 = 1234
3. Take all products of relations above.
For this case this still yields 𝐼 = 1234
4. Write a general expression for 𝐼. Again, for this case this is 𝐼 = 1234
This gives the resolution of this design
Confounding Relations and Resolution
5. Multiply all sides of this relation by each column in full factorial calculation
matrix (if a column has two different aliases in the design matrix, for example 4 =
123, you may use either one):
𝑙0 = 𝐼 ⇒ 𝑚𝑒𝑎𝑛 + 1234 2
𝑙1 = 1 ⇒ 1 + 11234 = 1 + 234
𝑙2 = 2 ⇒ 2 + 12234 = 2 + 134
𝑙3 = 3 ⇒ 3 + 12334 = 3 + 124
𝑙12 = 12 ⇒ 12 + 112234 = 12 + 34
𝑙13 = 13 ⇒ 13 + 112334 = 13 + 24
𝑙23 = 23 ⇒ 23 + 122334 = 23 + 14
𝑙123 = 123 ⇒ 123 + 1122334 = 123 + 4
Is it clear how we did this??
Confounding Relations and Resolution
6. Eliminate higher order effects:
𝑙0 = 𝑚𝑒𝑎𝑛
𝑙1 = 1
𝑙2 = 2
𝑙3 = 3
𝑙12 = 12 + 34
𝑙13 = 13 + 24
𝑙23 = 23 + 14
𝑙123 = 4
Not so fast! Is this valid?
It depends! Can I drop 3 factors? Maybe, maybe not.
Process knowledge.
Confounding Relations and Resolution
7. Determine resolution. The resolution is the length of the shortest term in
the confounding relation given in step 4.
III: Some main and 2 factor confounding.
IV: Some main and 3-factor, 2-factor and 2-factor confounding.
V: Some main and 4-factor, 2-factor and 3-factor confounding.
What is the resolution of the example? You can see this above, or from the
least number of digits in the relation in number 4. (𝐼 = 1234)
5−2
Example:A 2 fractional factorial
• How many experiments will this system have? How many variables?
5 variables. 25−2 = 8 experiments
5−2
Example:A 2 fractional factorial
• We must add variables 4 and 5. For now we will add them into columns 12 and 13. We
will determine the confounding structure:
1. Start with the duplicate columns
In the above example, 4 = 12, 5 = 13
2. Square one side.
This yields 𝐼 = 124 = 135.
3. Take all products of relations above.
For this case this yields 𝐼 = 112345 = 2345.
4. Write a general expression for 𝐼.
Again, for this case this is 𝐼 = 124 = 235 = 2345; resolution III.
5−2
Example:A 2 fractional factorial

5. Multiply all sides of this relation by each column in full factorial calculation matrix:

𝑙0 = 𝐼 ⇒ 𝑚𝑒𝑎𝑛 + 124 2 + 135 2 + (2345)/2


𝑙1 = 1 ⇒ 1 + 24 + 35 + 12345
𝑙2 = 2 ⇒ 2 + 14 + 1235 + 1345
𝑙3 = 3 ⇒ 3 + 1234 + 15 + 1245
𝑙12 = 12 ⇒ 12 + 4 + 235 + 1345
𝑙13 = 13 ⇒ 13 + 234 + 5 + 1245
𝑙23 = 23 ⇒ 23 + 134 + 125 + 45
𝑙123 = 123 ⇒ 123 + 34 + 25 + 145
5−2
Example:A 2 fractional factorial
6. Eliminate higher order effects:
𝑙0 = 𝑚𝑒𝑎𝑛
𝑙1 = 1 + 24 + 35
𝑙2 = 2 + 14
𝑙3 = 3 + 15
𝑙12 = 12 + 4
𝑙13 = 13 + 5
𝑙23 = 23 + 45
𝑙123 = 34 + 25
Question: Does it matter where we insert the new variables?
Answer: Say we have a 26−2 fractional factorial.
6 variables. 16 exp. Reduction in power 2. Each column appears 22 = 4

Option 1:
Insert 5 = 123, 6 = 1234.
This implies
𝐼 = 1235 = 12346 = 456.
Resolution III.

Option 2:
Insert 5 = 123, 6 = 124.
This yields
𝐼 = 1235 = 1246 = 3456.
Resolution IV!
• This is why we use confounding relations - to find the greatest
resolution possible given a fixed number of variables and experiments
Question: Do we have to try all ten combinations above?
No!
Any two 3 factors would give resolution 4
Any two 3+4 factors would give resolution 3
Reducing Confounding via Mirror Images
• How can we remove, or at least reduce, the effects of confounding?
We do so through the use of mirror images.
• Given our original fractional factorial experiment, where 4 = 123

We calculate its mirror image:


• The generator for the above is 4 = 123
• We will do a second experiment with
the generator 4 = −123

How we will design this experiment?

What does 4 = −123 mean?


Reducing Confounding via Mirror Images
• The design matrix for this experiment is: Our output will be then:
(using 𝐼 = −1234,
𝑙1′ = 1+ < 1 ∙ −1234 >= −234)
𝑙0′ = 𝐼 ⇒ 𝑚𝑒𝑎𝑛 − 1234 2 = 𝑚𝑒𝑎𝑛
𝑙1′ = 1 ⇒ 1 − 234 = 1
𝑙2′ = 2 ⇒ 2 − 134 = 2
𝑙3′ = 3 ⇒ 3 − 124 = 3

𝑙12 = 12 ⇒ 12 − 34

𝑙13 = 13 ⇒ 13 − 24

𝑙23 = 23 ⇒ 23 − 14

• It has the generator 4 = −123. 𝑙123 = 123 − 4 = −4
• Therefore I = −1234
Reducing Confounding via Mirror Images
If we combine this with the original output, we see that: 𝑙0 + 𝑙0′
= 𝑚𝑒𝑎𝑛
2
𝑙1 + 𝑙1′
≈ 𝐸1
4=123 4=-123 2
𝑙2 + 𝑙2′
𝑙0 = 𝑚𝑒𝑎𝑛 𝑙0′ = 𝑚𝑒𝑎𝑛 ≈ 𝐸2
2
𝑙1 = 1 𝑙1′ = 1 𝑙3 + 𝑙3′
𝑙2 = 2 𝑙2′ = 2 2
≈ 𝐸3

𝑙3 = 3 𝑙3′ = 3 𝑙12 + 𝑙12
≈ 𝐸12
𝑙12 = 12 + 34 ′ 2
𝑙12 = 12 − 34 ′
′ 𝑙13 + 𝑙13
𝑙13 = 13 + 24 𝑙13 = 13 − 24 ≈ 𝐸13
′ 2
𝑙23 = 23 + 14 𝑙23 = 23 − 14 𝑙23 + 𝑙23′

𝑙123 = 4 ′
𝑙123 = −4 ≈ 𝐸23
2

𝑙123 + 𝑙123
≈ 𝐸123
2
Reducing Confounding via Mirror Images
• But what does this mean numerically?

• If 𝑙12 = 16.23 and 𝑙𝑙2 = −6.19, then:
First 24−1 exp, 4 = 123, 𝑙12 = < 12 ∙ 𝑦𝑖𝑒𝑙𝑑 >÷ 4 = < 34 ∙ 𝑦𝑖𝑒𝑙𝑑 >÷ 4 = 16.23

Second 24−1 exp, 4 = −123, 𝑙12 = < 12 ∙ 𝑦𝑖𝑒𝑙𝑑 >÷ 4 = < −34 ∙ 𝑦𝑖𝑒𝑙𝑑 >÷ 4 = −6.19

16.23 + (−6.19)
𝐸12 = = 5.02
2

16.23 − (−6.19)
𝐸34 = = 11.21
2
Question: Through the original and the mirror image, we did 16
experiments, the same number as if we would have originally done a
24 full factorial. So what do we gain by doing this? Why will we ever
want to fracture our factorial?
1) Larger ‘reduction of power’ -> Greater savings
25−2 ⇒ we would have 16 experiments instead of 32
27−3 ⇒ we would have 24 + 24 = 32 experiments instead of 27 = 128
2) Iterative experimentation
Maybe after first 24−1 we could drop variables
Reducing Confounding via Mirror Images
• Let's consider the second example we had:

We had added variables 4 and 5; but now


we add them into columns -12 and -13.
Reducing Confounding via Mirror Images
We will determine the confounding structure:
1. Start with the duplicate columns in the Design Matrix
In the above example, 4 = −12, 5 = −13
2. Square one side.
This yields 𝐼 = −124 and 𝐼 = −135
3. Take all products of relations above.
For this case this still yields 𝐼 =− −112345 = 2345
4. Write a general expression for 𝐼.
Again, for this case this is 𝐼 = −124 = −135 = 2345; resolution III.
Reducing Confounding via Mirror Images
5. Multiple all sides of this relation by each column in full factorial
calculation matrix:
Reducing Confounding via Mirror Images
6. Eliminate higher order effect

𝑙0′ = 𝑚𝑒𝑎𝑛
𝑙1′ = 1 − 24 − 35
𝑙2′ = 2 − 14
𝑙3′ = 3 − 15

𝑙12 = 12 − 4

𝑙13 = 13 − 5

𝑙23 = 23 + 45

𝑙123 = −34 − 25
Reducing Confounding via Mirror Images
• So we can now combine our measurements from both experiments to
isolate out all of the location effects and some of the second order effects.
• Note that for this system, we are able to obtain estimates of all 5 main
effects, and some of the second order effects in 16, instead of 25 = 32
experiments.
• Got most of what I want using half of the cost of full factorial
• It should be clear that by taking mirror images of different columns you can
isolate different variables.
• Practice and trial and error are the only sure ways o develop a talent for
setting up the proper sequence of experiments.
Summary
So for fractional factorial experimentation:
1.Determine the number of experiments and variables. For example 27−2 implies
2.Take the design matrix for the appropriate powered full factorial experiment.
In this case it would be a 25 .
3.Decide where to insert the extra variables. (How?)
Confounding relations to max resolution.
4. Isolate factors in some way (removing higher order effects, guessing,
experiments with mirror images) in the equations for the 𝑙's. (Again, these
correspond to the 𝐸's in full factorial experimentation.)
5. If using a mirror image, combine the 𝑙's to get 𝐸's?
Add or subtract and divide by 2.
Summary
6. Change these factors into b's (How?)
𝑏𝑖 = 𝐸𝑖 /2
7. Isolate significant factors. (How?)
Replicate & hypothesis testing & t-test with replication
8. (Possibly) Do further, full factorial experimentation on significant
factors.
9. Optimize model and prescribe settings.
10. Validate settings

You might also like