Sleeper Robust Design Presentation

Successful Statistics LLC
Robust Design Simplified
Andy Sleeper
Andy@SuccessfulStatistics.com
1 (c) 2007 Successful Statistics LLC
Outline
What is robust design?
Some simple examples
How variation flows through a system
Three ways to reduce variation
– Noise reducers
– Interaction effects
– Nonlinear effects
DOE for robustness
Comparison with Taguchi inner / outer arrays
(c) 2007 Successful Statistics LLC

What is Robust?
A system (product or process) is robust if it performs
properly in a wide range of conditions
Robust products Products that are not robust

A pen that writes until the A pen that stops writing
ink is empty after a few months
A car that starts at -20° A car that does not start
A vacuum cleaner that A vacuum cleaner that
maintains suction levels loses suction
A TV that lasts 10 years A TV that needs repair
with no need for repair after 2 years
What is Robust Design?

In the design of a new system, any activity can
be called robust design, if it causes the
system…
– To have longer life (higher reliability)
– To be more consistent from use to use
– To be more consistent from product to product
– To perform consistently as temperature and other
conditions change
General rules for robust design
– Always identify critical characteristics (CTQs) that
quantify customer satisfaction
– Always look for ways to reduce variation in CTQs

Example: Supplier Selection

Two suppliers, X and Y,
make the same part to
the same specifications
A CTQ of the part must
be within spec limits of
34.0 – 38.0
A sample of 30 parts
from each supplier are
tested:
Which supplier is the
more robust design
choice?
Example: Process Design

A driver’s license renewal Series process – each
process has five steps: worker handles one task
data entry, skill testing, only
eye test, photograph, and
cashier
Customers vary widely, Parallel process – each
some requiring much customer sees one worker
more time to process than only
other customers
Waiting time is CTQ
Is a serial or parallel
process the more robust
design choice?

Example: Baking Cookies

In an experiment to find
the best recipe for baking
cookies, several batches
375
are made at 2 levels each
of time and temperature
Temp(F)
Batches are graded by
tasters on a Likert scale
Too raw Too cooked
325
8 Time(min) 12
Based on the distributions

shown, what is the best
time and temperature?
Where Does Variation Come From?
X Y
System
Input Output
Y = f(X, N)
variables variables
Input variables are controlled Noise variables include many

by design choices types of variation, such as
N •Environmental variables
Noise variables are not Noise •Tolerances on X variables
controlled by design choices, variables •Variations in the system
and will vary randomly •Noise sources we choose not
to control
•Noise sources we can not
control
•Unknown noise sources

Variation Flows Through the System

Variation caused by
noise variables (N) )
Y ( X, N
combines with input Y= f
variables (X), reflects
through the transfer
function f(X,N), and
appears as variation in Noise
output variables (Y)
To reduce variation in Y, X
our choices are either
– Reduce the variation of
noise variables
– Change the transfer
function to have a flatter
slope
Three Ways to Reduce Variation in Y

There are only three Y
ways to reduce variation
in Y:
Noise reducers: design
choices that reduce the
variation of noise X
Y
Interactions: design
choices that reduce the
slope of the transfer
function
Nonlinear effects: when X
Y
the transfer function is
curved, use a flatter
portion of the curve
X

Noise Reducers
The two-sample examples shown earlier are examples
of noise reducers
– Between suppliers X and Y, supplier X ships parts with less
variation. Choosing supplier X will make our product more
robust
– In the driver’s license process, my waiting time depends on
the variation of customers who came before me. Having
parallel paths reduces the impact of earlier customers on my
waiting time
Always examine the variation of many parts or units
flowing through a process
Y
11 (c) 2007 Successful Statistics LLC X
Finding Noise Reducers in an

Experiment
Replicated experiments have multiple trials at each
run, or set of factor settings
– Note: Replication requires more experimental units.
Repeating measurements on the same units is not the
same as replication.
Suggested analysis method
– Calculate the log of the standard deviation across
replicated measurements, for each run
– Analyze log(std dev) as a separate response variable
– Find optimal settings which both optimize the mean
response and minimize log(std dev)
– Many software products have multiple response
optimizers capable of this task
– BUT, calculating log(std dev) is usually a separate step

Example: Solenoid Design

This experiment has three factors at two
levels, with five replications
Column L calculates the log of the standard

deviation

Analysis of log(std dev) and Mean
Analysis of log(std dev): Analysis of mean response


Optimized Mean and Std. Deviation
In STATGRAPHICS software, the multiple response
optimizer finds the best design
Interactions
An interaction occurs when one factor
alters the response of the system to
changes in a different factor
If an interaction makes the slope of the
transfer function flatter, this is an
opportunity to make the system more
robust
Y X2 at level 2
X2 at level 1
16 (c) 2007 Successful Statistics LLC X1

Finding Interactions in an Experiment

Many experiments, if designed properly, will
find significant interaction effects
– To estimate all two factor interactions, look for
experiments that are Resolution V or higher
– To estimate only a few interactions, a Resolution IV
experiment may work
– Resolution III experiments will not work
If interaction effects are significant, always
look at the interaction plot to see what the
interaction means
Look for ways to use interactions to flatten the
transfer function
Example: Cookies

Example: Cookies
Interaction Plot
Ideal Taste
is 4
There are many ways to get the ideal taste of 4 …

Which time and temperature combination is the most robust choice?
Note: Finding the Robust Design

Requires a Little Thinking!
If you ask the optimizer
how to get Taste = 4, you
will find a range of time-
temperature solutions
By default, optimizers
only predict the mean,
not variation
Options:
– Use log(std dev) as an
additional response and
optimize both
– Observe the changes in
slope on the interaction
plot. Choose the design
that makes the transfer
function flatter

Nonlinear Effects
Many systems are naturally nonlinear
If a nonlinear effect creates a flatter spot in
the transfer function, this is an opportunity
to make the system more robust
Sometimes a second factor can be used to
adjust the mean back to its target value
Y
21 (c) 2007 Successful Statistics LLC X
Finding Nonlinear Effects in an

Experiment
3-factor BBD
An experiment requires more
than two levels per factor to
find nonlinear effects
Many experimental designs
are available
– Three-level factorials
2-factor CCD
– Central composite designs
– Box-Behnken designs
If nonlinear effects are
significant, look for ways to
use them to flatten the
transfer function

Example: More Cookies
Example: Cookies
Interaction Plot
For ideal taste (4),
here are two
options:
– Temp = 325,
Time = 10.7 min.
– Temp = 375,
Time = 8.6 min.
Which choice is
more robust?

Designing Experiments
for Robustness
Factors for the experiment should include
– Control factors (X)
– Noise factors (N) that you choose to control, just for the
experiment
Choose an appropriate experimental design
– To find noise reducers, replicate the experiment
– To find interactions, use Resolution V or higher
– To find nonlinear effects, use three levels
Note: with more than 3 factors, it is generally not
practical to meet ALL these criteria. Based on
knowledge of the process, guess what kind of robust
opportunity is most likely, and design the
experiment accordingly
Analyzing Experiments
for Robustness
To use noise reducers for robustness
– Calculate log(std. dev.) across replications, and analyze
log(sd) as an additional response
– Choose factor levels that decrease log(sd)
To use interactions for robustness
– If the analysis shows significant interactions, look at
interaction plots
– Choose factor levels where the function is flatter
To use nonlinear effects for robustness
– If the analysis shows significant nonlinear effects, look
at main effects plots
– Choose factor levels where the function is flatter
Without doing extra work, optimizers will optimize
the mean, but not minimize the standard deviation

Taguchi’s Inner and Outer Arrays

Taguchi made popular a different approach to
experiments for robust design
– The Inner Array is a designed experiment for the
control factors (X)
– The Outer Array is a designed experiment for noise
factors (N) that are controlled in the experiment
– The experiment includes one run at every
combination of inner and outer array runs
If the inner array has A runs, and the outer array has B
runs, the total experiment has A x B runs
– To analyze the experiment, calculate a measure of
variation (called a signal-noise ratio) and analyze
this as the response to the inner array experiment
Example Worked Two Ways

3 control factors (A, B, C) and 2 noise factors (D, E), at 2 levels each
Taguchi Approach A Simpler Way
L8 inner array (8 runs) Half-fractional factorial
L4 outer array (4 runs) 25-1 design, Resolution V
32 total runs 16 total runs
Finds ways to reduce noise Finds ways to reduce noise
caused by D-E caused by A-E
Detects noise reducers if Detects noise reducers if
replications added replications added
Detects interaction effects Detects interaction effects
Can be hard to understand Easy to understand using
simple graphs

Robust Design
Robust design can be complicated … but why?
Measure variation at every opportunity
– Measure quantitatively, not pass-fail
– Replicate experiments
– View graphs of the sample distribution, and
calculate measures of variation
– Look for opportunities to reduce variation
Design experiments for robustness
Look at graphs and think – don’t just rely on
the optimizer
Design systems for the flatter portions of the
transfer function
Summary
What is robust design?
Some simple examples
How variation flows through a system
Three ways to reduce variation
– Noise reducers
– Interaction effects
– Nonlinear effects
DOE for robustness
Comparison with Taguchi inner / outer arrays

Sleeper Robust Design Presentation

Uploaded by

Document Information

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Sleeper Robust Design Presentation

Uploaded by

Copyright:

Available Formats

Successful Statistics LLC

Robust Design Simplified

1 (c) 2007 Successful Statistics LLC

Successful Statistics LLC

(c) 2007 Successful Statistics LLC

Robust products Products that are not robust

3 (c) 2007 Successful Statistics LLC

Successful Statistics LLC

What is Robust Design?

4 (c) 2007 Successful Statistics LLC

(c) 2007 Successful Statistics LLC

Example: Supplier Selection

5 (c) 2007 Successful Statistics LLC

Successful Statistics LLC

Example: Process Design

6 (c) 2007 Successful Statistics LLC

(c) 2007 Successful Statistics LLC

Example: Baking Cookies

Based on the distributions

7 (c) 2007 Successful Statistics LLC

Successful Statistics LLC

Where Does Variation Come From?

Input variables are controlled Noise variables include many

8 (c) 2007 Successful Statistics LLC

(c) 2007 Successful Statistics LLC

Variation Flows Through the System

9 (c) 2007 Successful Statistics LLC

Successful Statistics LLC

Three Ways to Reduce Variation in Y

(c) 2007 Successful Statistics LLC

11 (c) 2007 Successful Statistics LLC X

Successful Statistics LLC

Finding Noise Reducers in an

(c) 2007 Successful Statistics LLC

Example: Solenoid Design

Column L calculates the log of the standard

Successful Statistics LLC

Example: Solenoid Design

14 (c) 2007 Successful Statistics LLC

(c) 2007 Successful Statistics LLC

Example: Solenoid Design

15 (c) 2007 Successful Statistics LLC

Successful Statistics LLC

16 (c) 2007 Successful Statistics LLC X1

(c) 2007 Successful Statistics LLC

Finding Interactions in an Experiment

17 (c) 2007 Successful Statistics LLC

Successful Statistics LLC

18 (c) 2007 Successful Statistics LLC

(c) 2007 Successful Statistics LLC

There are many ways to get the ideal taste of 4 …

Successful Statistics LLC

Note: Finding the Robust Design

(c) 2007 Successful Statistics LLC

21 (c) 2007 Successful Statistics LLC X

Successful Statistics LLC

Finding Nonlinear Effects in an

(c) 2007 Successful Statistics LLC

Example: More Cookies

23 (c) 2007 Successful Statistics LLC

Successful Statistics LLC

24 (c) 2007 Successful Statistics LLC

(c) 2007 Successful Statistics LLC

25 (c) 2007 Successful Statistics LLC