Design for Six Sigma

Central Composite Design

DFSS BB314 Central Composite Design.ppt

314-1
 Purpose

Understand the difference between designs that are used

to select KPIVs and designs that are used to optimize the
settings of already confirmed KPIVs
Design and analyze a Central Composite Design (CCD)
Apply CCD to an exercise

Copyright © 2001-2004 314-2

Six Sigma Academy International LLC All Rights Reserved
 DFSS Roadmap
DEFINE
DEFINE Rally
Rally Points
Points 0-1
0-1
Failure modes
Establish Business Needs & Priority Analyze/Minimize Product & and analyses
Process Risks
Develop Project Charter & Plan
Determine Model & Analyze Tolerances & Find critical
Identify Target Markets & Segments
Customers Sensitivities Y=f(x)
& Y’s: Identify Customer Needs/Wants Develop/Evaluate Measurement relationships
Establish Critical to Quality Systems
Quality, Characteristics (CTQC’s)
Initial Robust Product Design
Price &
Establish Technical Features
Product
OPTIMIZE
OPTIMIZE Rally
Rally Points
Points 4-6
4-6
Demand CONCEPT Optimize
CONCEPT Rally
Rally Points
Points 1-2
1-2
Minimize Product Complexity Y’s &
Identify CTQC Metrics/Measures
Maximize Product Velocity critical X’s,
Identify Key Measurement Systems RMI, WIP,
Optimize Critical Inputs - Final
Initial Develop Optimal Design Concept Robust Design FGI &
Product Supply
Develop Business Case & Schedule Optimize, Simulate Processes
Platform Chain
Final Robust Product Design
key X’s, DESIGN
DESIGN Rally
Rally Points
Points 3-4
3-4
product Design System, Subsystems &
VERIFY
VERIFY Rally
Rally Points
Points 5-8
5-8
velocities Components
Set Initial Control Systems/Plans
Set critical
Design Processes X’s, kan-
Measure
Verify Product Performance bans
the X’s Model & Assess Critical Parameters

and Y’s Verify Process Performance

Check Key Y’s
Test Plans & Reports

Copyright © 2001-2004 314-3

Six Sigma Academy International LLC All Rights Reserved
 Why Central Composite Design?

Develops Quadratic Models

Relatively inexpensive
Supports sequential experimentation strategy
Optimizes process knowledge
Commonly used with Response Surface Methods

Copyright © 2001-2004 314-4

Six Sigma Academy International LLC All Rights Reserved
 What is Central Composite Design (CCD)?

Advanced regression application

Active analysis
Used to extend existing data
- Quadratic terms
- Uses continuous data
Used with Response Surface Methods
Keep small (2-6 factors), gets large fast

Copyright © 2001-2004 314-5

Six Sigma Academy International LLC All Rights Reserved
 Important Formulas (Review)
First-Order Linear Model
Y = 0 + iXi + 
Y = 0 + 1X1 + 2X2 +…+ kXk + 
Linear Model With 2-Way Interactions
Y = 0 + iXi + ijXiXj + 
Y = 0 + 1X1 + 2X2 +…+ kXk + 12X1X2 +…+ 
Second-Order Quadratic Model
Y = 0 + iXi + iiXi2 + ijXiXj + 
Y = 0 + 1X1 + 2X2 +…+ kXk + 12X1X2 +…+ 11X12 +
22X22 +…+ kkXk2 + 
2-Level
2-Levelfactorial
factorialdesigns
designscan
canbe
beused
usedfor
forthe
thefirst
firsttwo
twomodels,
models,
Central
CentralComposite
Compositedesigns
designsare
areneeded
neededfor
forthe
thethird
thirdmodel
model
Copyright © 2001-2004 314-6
Six Sigma Academy International LLC All Rights Reserved
 What is a Central Composite Design?
Two-Level Central
Factorial Composite
Cube Points
Center Points
Axial points

Each factor run at two If the axial points are outside the square, as
settings (Lo and Hi) –Two here, each factor has 5 settings
points define the equation
If the axial points are on the sides of the
for a straight line
square, each factor has 3 settings
Center points are only used
Either way, 3 or more points can be used to
to test for curvature
define a quadratic equation, i.e. – curved line

Copyright © 2001-2004 314-7

Six Sigma Academy International LLC All Rights Reserved
 Properties of CC Designs

Alpha = distance (in coded units) from

center of design to axial points
Alpha
Alpha > 1, axial pts outside corner pts
Default alpha set to ensure Rotatability
(prediction error is equal in all directions)

OR to ensure that blocks are orthogonal

(block coefficient(s) have no impact on other
coefficients in the model)

Alpha = 1, axial pts on the faces

Face centered design
Use only when it is too dangerous to go
outside the corner points

-1 0 +1
Copyright © 2001-2004 314-8
Six Sigma Academy International LLC All Rights Reserved
 Possible Central Composite (CC) Designs

These
Thesearearethe
theCC
CCdesigns
designsavailable
availableininMINITAB
MINITAB––from
from22to to66
factors.
factors.Note
Notehow
howthe
thenumber
numberofofruns
runscan
canget
getlarge
large––ititisisalways
always
good
goodtotoreduce
reducethe
thenumber
numberofoffactors
factorsasasmuch
muchas aspossible.
possible.
Copyright © 2001-2004 314-9
Six Sigma Academy International LLC All Rights Reserved
 Properties of Default MINITAB CC Designs
Notes:
2 factors = 4 cube pts (Full)
3 factors = 8 cube pts (Full)
4 factors = 16 cube pts (Full)
5 factors = 16 cube pts
(Res V, half-fraction) or 32
cube pts (Full)
6 factors = 32 cube pts
(Res V, half-fraction) or 64
cube pts (Full)
Most designs rotatable
Orthogonal blocks

Copyright © 2001-2004 314-10

Six Sigma Academy International LLC All Rights Reserved
 Steps to Conduct a CCD Experiment (1)
Step 1: State the practical problem and objective of the experiment
Step 2: State the factors and levels of interest
Step 3: Screen initial factors – reduce to as few as possible
Step 4: Establish design strategy
Step 5: Create CC design in MINITAB; randomize the runs
Step 6: Conduct the experiment
Step 7: Fit the full quadratic model
Step 8: Reduce the model
 Review the ANOVA table and eliminate terms with p-values above .05.
 Eliminate non-significant blocks first, then highest order interactions terms,
then lower order, next remove insignificant quadratic effects, and finally main
effects
 Terms should be removed one at a time.
 You should not remove any term that is included in a higher-order term – this
is called preserving hierarchical integrity
 Run the final reduced model to include only those terms whose p-values are
deemed significant

Copyright © 2001-2004 314-11

Six Sigma Academy International LLC All Rights Reserved
 Steps to Conduct a CCD Experiment (2)
Step 9: Check model assumptions
• Analyze the residuals of the reduced model to ensure we have an
appropriate model
Step 10: Develop predictive model
Step 11: Determine optimal settings
• Contour plots
- Plot response (Y) vs. two most influential factors
- Plot response (Y) vs. next most influential pair of factors
- Repeat
• Surface plots
• Check stability of optimal solution

Step 12: Validate results by replicating optimum solution

Step 13: Final report

The highlighted steps are different for CC designs. The other steps
are the same as for Full Factorials and 2-Level Factorials.

Copyright © 2001-2004 314-12

Six Sigma Academy International LLC All Rights Reserved
 Step 3 – Screen Initial Factors
Reduce number of factors to manageable number
2-4 factors work well, 5 or 6 get to be large designs
Some screening designs
- 2-Level fractional or full factorials – recommended
 Support sequential experimentation if done right
- Res V Fractional Factorials or Full Factorials
- Center points
- L12 or P-B 12 run (models main effects only, not recommended)

Important Point:
Central Composite designs are used to model curvature. Period.
They should never be used to screen factors.

Copyright © 2001-2004 314-13

Six Sigma Academy International LLC All Rights Reserved
 Step 4 – Establish Design Strategy
All factors must be numeric, somewhat continuous (to make sense of
the quadratic model)
How will design be executed?
Run in blocks
- Corner points and some center points in first block (table of properties)
- Axial points and some center points in second block (table of properties)
- This strategy can save runs – conduct first half of experiment, fit linear model
with 2-way interactions, and test for curvature. If there is no curvature, there
is not much point in running the second block
- Caution: experimental runs performed in two separate and distinct groups –
process changes could have occurred, Blocking variable tests for differences
between groups
Run without blocks
- Reduces effects from other lurking variables

Copyright © 2001-2004 314-14

Six Sigma Academy International LLC All Rights Reserved
 Steps 7 & 8
Fit the Full Quadratic Model and Reduce
To review, the full quadratic model is given by:
Y = 0 + iXi + iiXi2 + ijXiXj + 
Y = 0 + 1X1 + 2X2 +…+ kXk + 12X1X2 +…+ 11X12 + 22X22 +…+ kkXk2
+
Main effect for each factor
Squared term for each factor
Every 2-way interaction
When reducing the model, consider blocks first, interactions second,
squared terms third, and main effects last
Drop terms one at a time – CC designs are not generally orthogonal,
so coefficients can change
Preserve hierarchy – keep main effect terms for factors with significant
squared terms, and for factors included in significant interactions

Copyright © 2001-2004 314-15

Six Sigma Academy International LLC All Rights Reserved
 Step 11 – Determine Optimal Settings
This step is different because of different tools
Graphical analysis works best
- Contour plots
 Select the most influential pair of factors
 Create contour plot of Y vs. the selected pair of factors
 Create contour plots for other pairs, in order of influence
 Use plots to highlight the region of optimality
 (for the case of 3 factors, you can create the contour plots of Y vs. the most
influential pair of factors, for low, middle and high settings of the 3’rd factor – see
Appendix)
- Surface plots
 Use to determine the stability of the optimal region
- Relatively flat surface – very stable
- Sharp peaked surface – not very stable
If optimal region found, set limits on X’s to maintain optimal condition

Copyright © 2001-2004 314-16

Six Sigma Academy International LLC All Rights Reserved
 Tear Strength Example (1)
You want to determine what effects sealing temperature, adhesive content
(%), and polyethylene additive (%) have on tear strength for consumer snack
packaging. The tear strength needs to be at least 10 grams/square inch, but
not more than 12.
Step 1: State the problem.
We want to determine settings of sealing temperature, adhesive content,
and polyethylene that will deliver the desired tear strength (10-12 grams/sq
inch) consistently. Optimum settings of the three variables should lead to a
stable control situation.
Step 2: State the factors and levels of interest
Note:
X1 Sealing temp Lo = 75 Hi = 85
The Lo and Hi
X2 Adhesive cont % Lo = 46 Hi = 64
settings define
X3 Polyethylene % Lo = 0.5 the corner points.
Hi = 1.7

Copyright © 2001-2004 314-17

Six Sigma Academy International LLC All Rights Reserved
 Tear Strength Example (2)
Step 3: Screen the factors.
We are assuming that this has already been done, and the three factors
listed here are a result of the screening. They are statistically significant
factors
Step 4: Establish design strategy
We will run the entire experiment in one run, i.e. – without blocks. Runs
are inexpensive, so we are not that interested in the sequential method of
experimentation.

Copyright © 2001-2004 314-18

Six Sigma Academy International LLC All Rights Reserved
 Tear Strength Example (3)
Step 5: Create the CC design in MINITAB
Stat > DOE > Response Surface > Create Response Surface Design

Select the design for 1 Block.

Copyright © 2001-2004 314-19

Six Sigma Academy International LLC All Rights Reserved
 Tear Strength Example (4)
Step 5: Create the CC design in MINITAB (cont’d)

Be sure to designate that the levels define

the Cube (Corner) points.
Enter factor names and levels.

Copyright © 2001-2004 314-20

Six Sigma Academy International LLC All Rights Reserved
 Tear Strength Example (5)
Step 5: Create the CC design in MINITAB (cont’d)

For teaching purposes, we unchecked the

“Randomize Runs” box. Ordinarily, we
would always randomize the runs.

Copyright © 2001-2004 314-21

Six Sigma Academy International LLC All Rights Reserved
 Tear Strength Example (6)
Step 5: Create the CC design in MINITAB (cont’d)

Copyright © 2001-2004 314-22

Six Sigma Academy International LLC All Rights Reserved
 Tear Strength Example (7)
Step 6: Conduct the experiment (we have done this for you)
Open MINITAB file DFSS BB314a CCD Tear Strength.MPJ to get responses

Step 7: Fit the full quadratic model

(Stat > DOE > Response Surface > Analyze Response Surface Design)

Default model is full quadratic.

Copyright © 2001-2004 314-23

Six Sigma Academy International LLC All Rights Reserved
 Tear Strength Example (8)
Step 8: Reduce the model
Response Surface Regression: Tear Strengt versus Seal Temp, Adhes Cont, ...

The analysis was done using coded units.

Estimated Regression Coefficients for Tear Str

Term Coef SE Coef T P Main effects

Constant 10.165 0.4443 22.880 0.000 are mostly
Seal Tem -1.104 0.2948 -3.744 0.004
Adhes Co 0.087 0.2948 0.296 0.773
significant.
Poly Con 1.020 0.2948 3.462 0.006
Seal Tem*Seal Tem -0.760 0.2869 -2.647 0.024
Squared terms
Adhes Co*Adhes Co -1.042 0.2869 -3.633 0.005 are all
Poly Con*Poly Con -1.149 0.2869 -4.003 0.003 significant.
Seal Tem*Adhes Co -0.350 0.3851 -0.909 0.385
Seal Tem*Poly Con -0.500 0.3851 -1.298 0.223 None of the
Adhes Co*Poly Con 0.150 0.3851 0.389 0.705 interactions
S = 1.089 R-Sq = 85.6% R-Sq(adj) = 72.6% are significant.

Remove the interactions, one at a time.

Copyright © 2001-2004 314-24

Six Sigma Academy International LLC All Rights Reserved
 Tear Strength Example (9)
Step 8: Reduce the model (cont’d)
The final reduced model
Response Surface Regression: Tear Strengt versus Seal Temp, Adhes Cont, ...

The analysis was done using coded units.

Estimated Regression Coefficients for Tear Str

Term Coef SE Coef T P

Constant 10.165 0.4385 23.182 0.000
Seal Tem -1.104 0.2909 -3.794 0.002
Adhes Co 0.087 0.2909 0.300 0.769
Poly Con 1.020 0.2909 3.508 0.004
Seal Tem*Seal Tem -0.760 0.2832 -2.682 0.019
Adhes Co*Adhes Co -1.042 0.2832 -3.681 0.003
Poly Con*Poly Con -1.149 0.2832 -4.055 0.001

S = 1.075 R-Sq = 81.7% R-Sq(adj) = 73.3%

Adjusted R-Sq actually improved!!

Copyright © 2001-2004 314-25

Six Sigma Academy International LLC All Rights Reserved
 Tear Strength Example (10)
Step 9: Check model assumptions
Store the fits and residuals, use the four-in-one residual plot

Copyright © 2001-2004 314-26

Six Sigma Academy International LLC All Rights Reserved
 Tear Strength Example (11)
Step 9: Check model assumptions (cont’d)

Copyright © 2001-2004 314-27

Six Sigma Academy International LLC All Rights Reserved
 Tear Strength Example (12)
Step 9: Check model assumptions (cont’d)

What are the

conclusions?
What about the
histogram? Is it
normal?
Remember, we only
have 20 residuals – a
histogram is
somewhat unreliable.
The normal probability
plot looks okay –
normality should not
be an issue.

Copyright © 2001-2004 314-28

Six Sigma Academy International LLC All Rights Reserved
 Tear Strength Example (13)
Step 10: Develop predictive model
It’s right there – MINITAB already gave it to us.
Estimated Regression Coefficients for Tear Str using data in uncoded units
Term Coef
Constant -211.828
Seal Tem 4.64068
Adhes Co 1.42535
Poly Con 8.71940
Seal Tem*Seal Tem -0.0303838
Adhes Co*Adhes Co -0.0128696
Poly Con*Poly Con -3.19029

Write out the equation:

Tear Strength = -211.828 + 4.64*SealTemp + 1.425*AdhesCont +
…….

Copyright © 2001-2004 314-29

Six Sigma Academy International LLC All Rights Reserved
 Tear Strength Example (14)
Step 11: Determine optimal settings (Stat > DOE > Response Surface > Contour/Surface Plot)

Specify contour
values of: 8, 9, 10,
10.5, 11, 11.5, 12,
13, 14 (do not use
commas in input)

Which factor has the least

influence?
Adhesive Content
Which pair of factors have
the most influence?
Seal Temp and
Polyethylene Content

Copyright © 2001-2004 314-30

Six Sigma Academy International LLC All Rights Reserved
 Tear Strength Example (15)

Copyright © 2001-2004 314-31

Six Sigma Academy International LLC All Rights Reserved
 Tear Strength Example (16)

Note: we are holding

the factor we deem “OK” Again
least important at its
“Middle Settings”

Copyright © 2001-2004 314-32

Six Sigma Academy International LLC All Rights Reserved
 Tear Strength Example (17)

Note: we are holding

the factor we deem
least important at its
“Middle Settings”

Copyright © 2001-2004 314-33

Six Sigma Academy International LLC All Rights Reserved
 Tear Strength Example (18)

Copyright © 2001-2004 314-34

Six Sigma Academy International LLC All Rights Reserved
 Tear Strength Example (19)
Step 11: Determine optimal settings (cont’d)

Have we solved the

problem?
Do we have a region of
optimality?
What are the limits to
place on the X’s to
ensure that Tear
Strength has the right
characteristics?
Do we need to worry
about Tear Strength
being too high?
Spec is at least 10, not more than 12

Copyright © 2001-2004 314-35

Six Sigma Academy International LLC All Rights Reserved
 Tear Strength Example (20)
Step 11: Determine optimal settings (cont’d)

Is the region of
optimality stable? Surface Plot of Tear Str

Do we need to
worry about Tear
Strength being too 10

high?
Don’t forget: Tear Strength 5

Step 12: Validate 0

2

1 Poly Cont
results with 70

Seal Temp
80
90
0

confirmation runs.
Step 13. Final Hold values: Adhes Co: 55.0
report.

Copyright © 2001-2004 314-36

Six Sigma Academy International LLC All Rights Reserved
 End of Module

Discussion

Copyright © 2001-2004 314-37

Six Sigma Academy International LLC All Rights Reserved