What is Robust Design or
Taguchi’s method?
• An experimental method to achieve product
and process quality through designing in an
insensitivity to noise based on statistical
principles.
History of the method
• Dr. Taguchi in Japan: 1949-NTT
– develops “Quality Engineering”
– 4 time winner of Demming Award
• Ford Supplier Institute, early 1980s
• American Supplier Institute, ASI
– Engineering Hall of Fame
• Statistics Community
– DOE
– S/N Ratio
Who uses Taguchi’s Methods
• Lucent • Toyota
• Ford • TRW
• Kodak • Chrysler
• Xerox • GTE
• Whirlpool • John Deere
• JPL • Honeywell
• ITT • Black & Decker
Documented Results from Use
• 96% improvement of NiCAD • $900,000 annual savings in the
battery on satellites (JPL/ NASA) production of sheet-molded
• 10% size reduction, 80% compound parts (Chrysler)
development time reduction and • $1.2M annual savings due to
20% cost reduction in design of a reduction in vacuum line
choke for a microwave oven (L.G. connector failures (Flex
Electronics) Technologies)
• $50,000 annual cost savings in • 66% reduction in variability in
design of heat staking process arrival time and paper
(Ann Arbor Assembly Corp) orientation (Xerox)
• 60% reduction in mean response • 90% reduction in encapsulation
time for computer system (Lucent) variation (LSI Corp)
Insensitivity to Noise
• Noise = Factors which the engineer can not or
chooses not to control
– Unit-to-unit
• Manufacturing variations
– Aging
• Corrosion
• UV degradation
• wear
– Environmental
• human interface
• temperature
• humidity
How Noise Affects a System
Noise
Useful Energy
Energy Ideal Function of Quality Characteristic, y
Signal Factor, M Product or Process
Harmful Energy
Caused by Noise
Control
Factors
Step 1: Define the Project Scope 1/2
• A gyrocopter design is to be published in a Sunday Comics
section as a do-it-yourself project for 6-12 year old kids
• The customers (kids) want a product they can easily build
and have a long flight time.
| WW |
---
WL
--- ---
---
1/4”
BL
----
Step 1: Define the Project Scope 2/2
• This is a difficult problem from an engineering standpoint
because:
– hard to get intuitive feel for effect of control variables
– cant control materials, manufacturing or assembly
– noise factors are numerous and have strong effect on
flight.
Step 2: Identify Ideal Function
• Ideally want the most flight time (the quality characteristic
or useful energy) for any input height (signal or input
energy)
• Minimize Noise Effect
• Maximize Slope
Time of Flight
Drop Height
Step 3: Develop Noise Strategy 1/2
• Goal is to excite worst possible noise conditions
• Noise factors
– unit-to-unit
– aging
– environment
Step 3: Develop Noise Strategy 2/2
• Noise factors
– unit-to-unit
Construction accuracy
Paper weight and type
angle of wings + many, many others
– aging
damage from handling
– environment
angle of release
humidity content of air
wind
Step 4: Establish Control Factors and Levels
1/4
• Want them independent to minimize interactions
– Dimensionless variable methods help
– Design of experiments help
– Confirm effect of interactions in Step 7
• Want to cover design space
– may have to guess initially and perform more
than one set of experiments. Method will help
determine where to go next.
Step 4: Establish Control Factors and Levels
2/4
• Methods to explore the design space
– shot-gun
– one-factor-at-a-time
– full factorial
– orthogonal array (a type of fractional factorial)
Step 4: Establish Control Factors and Levels
3/4
Control factor array for the paper gyrocopter parameter optimization
experiment
1 2 3 4 5 6 7 8
Run WL WW BL Size B_Fold Gussets
1 1 1.0/ww 0.50 1.33 x WL 100% 1 0 None
2 1 1.0/ww 0.75 1.67 x WL 75% 2 15% 45deg
3 1 1.0/ww 1.00 2.00 x WL 50% 3 30% 45deg
4 1 1 .5/ww 0.50 1.33 x WL 75% 2 30% 45deg
5 1 1.5/ww 0.75 1.67 x WL 50% 3 0 None
6 1 1.5/ww 1.00 2.00 x WL 100% 1 15% 45deg
7 1 2.0/ww 0.50 1.67 x WL 100% 3 1 5% 45deg
8 1 2.0/ww 0.75 2.00 x WL 75% 1 30% None
9 1 2.0/ww 1.00 1.33 x WL 50% 2 0 45deg
10 2 1.0/ww 0.50 2.00 x WL 50% 2 15% None
11 2 1.0/ww 0.75 1.33 x WL 100% 3 30% 45deg
12 2 1.0/ww 1.00 1.67 x WL 75% 1 0 45deg
13 2 1.5/ww 0.50 1.67 x WL 50% 1 30% 45 deg
14 2 1.5/ww 0.75 2.00 x WL 100% 2 0 45deg
15 2 1.5/ww 1.00 1.33 x WL 75% 3 15% None
16 2 2.0/ww 0.50 2.00 x WL 75% 3 0 45deg
17 2 2.0/ww 0.75 1.33 x WL 50% 1 15% 45deg
18 2 2.0/ww 1.00 1.67 x WL 100% 2 30% None
Step 4: Establish Control Factors and Levels
4/4
Step 5: Conduct Experiment and Collect
Data
3 feet 6 feet 9 feet
20# paper 24# paper 20# paper 24# paper 20# paper 24# paper
1 0.68 s 0.55 s 1.48 s 1.48 s 2.31s 2.38 s
2 0.74 0.58 1.19 1.58 2.25 2.44
3 0.68 0.45 1.35 1.03 1.48 1.96
4 0.58 0.71 1.25 1.22 2.34 1.75
5 0.71 0.68 1.58 1.41 2.28 2.41
6 0.67 0.55 1.64 1 .51 2.44 2.08
7 0.65 0.7 1.16 1.21 2.68 2.7
8 0.71 0.6 1.93 1.75 2.61 2.73
9 0.84 0.63 1.83 1.64 2.09 2.5
10 0.74 0.61 1.7 1.22 2.09 2.31
11 0.61 0.45 1.22 1.03 1.48 1.96
12 0.61 0.58 1.38 1.22 2.28 2.3
13 0.87 0.68 1.64 1.19 2.02 2.41
14 0.81 0.65 2.09 1.51 2.27 2.67
15 0.84 0.63 1.7 1.22 1.51 2.5
16 0.68 0.68 1.54 1.64 2.44 2.5
17 0.71 0.68 1.54 1.51 2.6 2.6
18 0.61 0.84 1.96 1.64 2.73 3.05
Data for Runs 5 and 15
2.5
2
Time (sec)
1.5 Run 5
1 Run 15
0.5
0
0 2 4 6 8 10
Height (ft)
Step 6: Conduct Data Analysis 1/7
• Calculate signal-to-noise-ratio (S/N) and Mean
• Complete and interpret response tables
• Perform two step optimization
– Reduce Variability (minimize the S/N ratio)
– Adjust the mean
• Make predictions about most robust configuration
Step 6: Conduct Data Analysis 2/7
• Calculate signal to noise ratio, S/N, a
metric in decibels S/N gain
variability
reduction
Useful output 3 27%
S/N =
Harmful output 6 50%
12 75%
Effect of Mean
= Variability around mean
y 2
= 10 log 2 Note: This is one of many
s
forms of S/N ratios.
Step 6: Conduct Data Analysis 3/7
Results of the parameter optimization experiment
1 2 3 4 5 6 7 8 slope S/N
Run WL WW BL Size B_Fold Gussets (sec/ft)
1 1 1.0/ww 0.50 1.33 X WL 100% 1 0 None 0.25 6.94 dB
2 1 1.0/ww 0.75 1.67 X WL 75% 2 15% 45deg 0.25 2.67 dB
3 1 1.0/ww 1.00 2.00 X WL 50% 3 30% 45deg 0.19 -0.24 dB
4 1 1.5/ww 0.50 1.33 X WL 75% 2 30% 45deg 0.22 0.69 dB
5 1 1.5/ww 0.75 1.67 X WL 50% 3 0 None 0.26 9.04 dB
6 1 1.5/ww 1.00 2.00 X WL 100% 1 15% 45deg 0.25 3.81 dB
7 1 2.0/ww 0.50 1.67 X WL 100% 3 15% 45deg 0.26 -1.95 dB
8 1 2.0/ww 0.75 2.00 X WL 75% 1 30% None 0.29 4.73 dB
9 1 2.0/ww 1.00 1.33 X WL 50% 2 0 45deg 0.26 2.64 dB
10 2 1.0/ww 0.50 2.00 X WL 50% 2 15% None 0.24 2.81 dB
11 2 1.0/ww 0.75 1.33 X WL 100% 3 30% 45deg 0.19 0.76 dB
12 2 1.0/ww 1.00 1.67 X WL 75% 1 0 45deg 0.24 3.87 dB
13 2 1.5/ww 0.50 1.67 X WL 50% 1 30% 45deg 0.24 1.62 dB
14 2 1.5/ww 0.75 2.00 X WL 100% 2 0 45deg 0.28 0.87 dB
15 2 1.5/ww 1.00 1.33 X WL 75% 3 15% None 0.23 -3.96 dB
16 2 2.0/ww 0.50 2.00 X WL 75% 3 0 45deg 0.27 9.04 dB
17 2 2.0/ww 0.75 1.33 X WL 50% 1 15% 45deg 0.28 4.88 dB
18 2 2.0/ww 1.00 1.67 X WL 100% 2 30% None 0.31 2.99 dB
Step 6: Conduct Data Analysis 4/7
Response Table
Factor response averages table for the
parameter optimization experiment
Factor Time Time
Level (slope) (S/N)
1.0/ww 0.23 2.80
WL 1.5/ww 0.25 2.01
2.0/ww 0.28 3.72
0.50 0.25 3.19
WW 0.75 0.26 3.82
1.00 0.25 1.52
1.33 X WL 0.24 1.99
BL 1.67 X WL 0.26 3.04
2.00 X WL 0.25 3.50
100% 0.26 2.23
Size 75% 0.25 2.84
50% 0.25 3.46
0% 0.26 5.40
B_Fold 15% 0.25 1.38
30% 0.24 1.76
Gussets None 0.26 3.76
45deg 0.25 2.39
Step 6: Conduct Data Analysis 5/7
Response plot
Step 6: Conduct Data Analysis 6/7
Two Step Optimization
• Reduce Variability (minimize the S/N ratio)
– look for control factor effects on S/N
– Don’t worry about mean
• Adjust the mean
– To get desired response
– Use “adjusting factors”, those control factors
which have minimal effect on S/N
Step 6: Conduct Data Analysis 7/7
• For gyrocopter
– wing width = .75in
– wing length = 2.00/0.75 = 2.67 in
– body length = 2.00 x 2.67 = 5.33 in
– size = 50%
– no body folds Predicted Performance
– no gussets S/N = 9.44 dB
Slope = .31 sec/ft
Step 7: Conduct Conformation Run
• To check validity of results
• To check for unforeseen interaction effects
between control factors
• To check for unaccounted for noise factors
• To check for experimental error
Predicted Confirmed
S/N 9.44 dB 9.86
Slope .31sec/ft .32 sec/ft
How Taguchi’s Method Differs from an
Ad-hoc Design Process
• Organized Design Space • Concurrently Addresses
Search Manufacturing Variation
• Clear Critical Parameter • Concurrent Design-Test
Identification Not Design-Test-Fix
• Minimize Development
• Focus on Parameter Time (Stops Fire Fighting)
Variation (Noise)
• Corporate Memory
• Clear Stopping Criteria Through Documentation
• Robustness centered not • Encourages Technology
Failure Centered Development Through
• Reusable Method System Understanding
How Taguchi’s Method Differs from
Traditional Design of Experiments
• Focused on reducing the • Tries to reduce interaction
impact of variability rather between control factors
than reducing variability rather than study them
• Focused on noise effects Requires little skill in
rather than control factor statistics
effects • Usually lower cost
• Clearly focused cost
function - maximizing the
useful energy
How Taguchi’s Method Differs from
Shainin’s Method
• Focused on both Product • Widely Used
and Process Design rather Internationally
than Primarily on Process
• Fire prevention rather than
• Oriented to developing a fire fighting
robust system not finding
a problem (Red X). • Accessible
Taguchi tells what • Many Case Studies
parameter values to set to Available
make system insensitive to
parameter Shainin
identifies as needing
control.
Plan for Application at Tektronix
• Select a parameter design problem
• Design the experiment
• Perform the experiment
• Reduce data
• Report results to Company
• Assuming success
– design more experiments
– train more engineers
– Plan for student-run experiments