1

Engineering Design Methodology

MCEN4011

Taguchi Methods
Pawel Podsiadlo
Tribology Laboratory
School of Civil and Mechanical Engineering
Curtin University
Western Australia
Email: [email protected]
 2

Overview

 Methods for Parametric Design

 Taguchi Methods
 Quality through Design
 Design of Experiments (DOE)
 Signal-to-Noise Ratio
16336 ISO - Robust Parameter Design (RPD)
 3

Methods for Parametric Design

For parametric design we use

 Guided Iteration Methodology

 Optimization Methods

 Taguchi Methods
 4

Guided Iteration Methodology

Input Information

Formulate
Problem

Generate Analyses
Alternatives Functionality
DFM
Evaluate Cost
Alternatives Reliability
Not Acceptable Robustness
Guided Other
Redesign

Acceptable
 5

Optimization Methods

 Optimization by differential calculus

 Lagrange multipliers
 Lattice search
 Univariate search
 Steepest ascent
 Linear programming
 Simplex algorithm
 Geometric programming
 6

Taguchi Methods

Basic and
Performance Needs
are optimized by
Taguchi Methods
 7

Taguchi Methods

 Genichi Taguchi (1924 – 2012) was an engineer and statistician

born in Japan.
 He worked during 1950’s to improve communication systems.
 He is father of the “Taguchi Methods” and “Robust*
Engineering”
 Taguchi methods were widely used by Japanese manufacturers
in the 1970s but were not introduced to the US until the 1980s.

*The degree to which a product’s quality of performance is

insensitive to variations in manufacture, in-service wear, and
in-service environmental variations is called robustness.
 8

Taguchi Methods
 Taguchi has made an influential contribution to industrial
statistics. The key elements of his quality philosophy are
1. Taguchi loss function that is used to measure financial loss
to society resulting from poor quality,

2. The philosophy of off-line quality control, i.e., designing

products and processes such a way that they are insensitive
("robust") to parameters outside the design engineer's
control, and

3. Innovations in the statistical design of experiments, notably

the use of an outer array for factors that are uncontrollable
in real life, but systematically varied in the experiment.
 9

Main Concepts

 Quality should be designed into the product and not inspected

into it.

 The product should be designed in such a way that it is immune

to causes of variation.

 The loss of quality should be measured as function of deviation

from the nominal value of the quality characteristic.

 Quality is best achieved by minimizing the deviation from

target (i.e., minimizing variations).
 10

Meanings of Quality
 The traditional method of calculating the cost of quality is based
on the number of parts rejected and reworked.

 This method of quality evaluation is incapable of distinguishing

between two samples; both within the specification limits, but
with different distributions of targeted properties.

 Quality is defined as the loss incurred by society due to

 failure to meet performance characteristics,
 failure to meet customer expectations, and
 harmful side effects (e.g., pollution).
 11

Meanings of Quality
 Traditional approach: Quality loss is defined as conformance
to specification and it is represented by a step function.

 Taguchi philosophy: Quality loss is defined as deviation from

nominal and it is represented by a quadratic function.

Traditional Taguchi
 12

Meanings of Quality

 Taguchi introduced the concept of the quality loss of a

product.
 All critical parameters (including dimensions) of a
product should be at their target values, ensuring the best
performance of a product.
 If parameters deviate from their target values the
product’s performance will deteriorate and the product
will lose quality.
 Taguchi proposed a quadratic function to quantify the
quality loss.
 13

Taguchi’s Quality Loss Function

 𝐿𝐿 𝑦𝑦 = 𝑘𝑘 𝑦𝑦 − 𝑚𝑚 2 is the quality loss function

y – the quality characteristic

m - the target value for y
k – a constant, the quality loss coefficient
A – the cost of replacement or repair if the critical
parameter does not meet tolerance
requirements
∆ – the tolerance value of the critical parameter

 y exceeds the tolerance ∆ when 𝐿𝐿 𝑦𝑦 = 𝐴𝐴. In this case 𝑦𝑦 = 𝑚𝑚 + ∆

and we get
𝐿𝐿 𝑚𝑚 + ∆ = 𝐴𝐴 = 𝑘𝑘 𝑚𝑚 + ∆ − 𝑚𝑚 2 = 𝑘𝑘 ∆ 2 and 𝑘𝑘 = 𝐴𝐴⁄∆2 .
 14

Taguchi’s Quality Loss Function

 This formula calculates loss for an individual part in
nominal-is-the best situation:

2
𝐿𝐿 𝑦𝑦 = 𝑘𝑘 𝑦𝑦 − 𝑚𝑚

 This formula calculates the average loss per part in nominal-

is-the best situation:

𝐿𝐿� (𝑦𝑦) = 𝑘𝑘 𝜎𝜎 2 + 𝑦𝑦� − 𝑚𝑚 2

where
𝜎𝜎 2- the population variance on y due to common causes in the process and it
is usually approximated by the sample variance
𝑦𝑦� - the mean of all observations 𝑦𝑦𝑖𝑖 in the sample
𝑚𝑚 - the target value
 15

Taguchi’s Quality Loss Function

 This formula calculates loss for an individual part in smaller-

the better situation (e.g. pollution from exhaust pipe):

𝐿𝐿 𝑦𝑦 = 𝑘𝑘𝑦𝑦 2

 This formula calculates loss for an individual part in larger-

the better situation (e.g., mechanical power, strength, wear
resistance):

1
𝐿𝐿(𝑦𝑦) = 𝑘𝑘 2
𝑦𝑦
 16

Definition of Engineered Quality

“The engineered quality of a product is assessed by estimating the

loss imparted by the product to the society from the time product is
shipped.”
Dr. Genichi Taguchi
 17

Taguchi’s Quality Loss

 This loss is a loss to the customer and therefore to the public in
general (loss to the society).

 Taguchi’s definition of quality places more emphasis on

customer satisfaction whereas previously all definitions were
related to the producer.

 Optimum customer satisfaction can be achieved by developing

the product which meets the target value on a consistent basis.

 The single most important aspect of Taguchi’s quality control

philosophy is the minimization of variation around the target
value.
 18

Sony Case Study

 The desired level for color

density of televisions is
denoted as T, while T ± 8
defines the tolerance limits.

Distribution of color density in TV sets

LAL – Lower Action Limit UAL – Upper Action Limit
 19

Sony Case Study

 In Sony-U.S. the attention was

focused on the “adequate quality”
defined by true tolerance limits. As
result we see a more or less uniform
distribution.

 In Sony-Japan the attention was

focused on producing televisions with
color densities as close as possible to
T. As a result we see a bell-shaped
distribution.
Distribution of color density in TV sets
 20

Sony Case Study

 L = kσ2 k = 0.16
 Calculations show that the expected loss of Sony-
U.S. sets is about three times higher than the
expected loss of Sony-Japan sets!

Manufacturer Variance Expected loss per unit

U.S. plant 8.33 1.33

Japanese plant 2.78 0.44

Difference>> 0.89
 21

Bearing Case Study

 Two batches of main bearings for an internal

combustion engine were received from two different
sources, A and B, for a new engine development
program.
 Under laboratory conditions, bearings from source B
wore much faster than those from source A.
 To pin-point the cause of unequal wear, selected
performance characteristics of the bearings were
measured.
 22

Bearing Diameter Distribution

 23

Improving Quality
1. Inspection versus prevention
 No amount of inspection can put quality back into the
product, it merely treats symptoms.
 “Prevention is better than the cure.”

2. Sources of poor quality - noise

 85% of poor quality is attributable to manufacturing
system (observed by W. E. Deming).
 Noise is cause of variation in product function, whether
it originates in the manufacturing process, the
customer’s method of use or the environment.
 24

Improving Quality
3. Reduce variations by removing
 causes (on-line strategy),
 special causes, and
 common causes (could be difficult and expensive).

4. Reduce variations by making products and processes

immune to noise.
 A product or a process that can function well against a
background noise is said to be robust.
 25

Improving Quality
Product Sources of Variation
Development Environmental Product Manufacturing
Stages Variables Deterioration Variation
Product design   
Process design   
Manufacturing   
 - Countermeasures possible
 - Countermeasures impossible

Product development stages at which countermeasures against

various sources of variation can be built into the product.
 26

Experiments for Control Factors

How should we run experiments to determine the optimum design?

 Guided Iteration
 One-Factor-at-a-Time
 Full Factorial Experiment
 Fractional Factorial Experiment
 Orthogonal Array
 27

One-Factor-at-a-Time

 This strategy selects a starting point or uses a baseline set of

levels, for each factor.
 Then successively each factor is varied over its range with other
factors held constant at the baseline level.
 After all the experiments are performed, effect of each factor is
established.
 Main drawback of this strategy is that it fails to consider any
possible interaction between the factors.
 An interaction is the failure of one factor to produce the same
effect on the response at different levels of another factor.
 28

One-Factor-at-a-Time
Run No. A B C D E F G H
1: Current 2 2 2 2 2 2 2 2 Result 1
2 1 2 2 2 2 2 2 2 Result 2
3 2 1 2 2 2 2 2 2 Result 3
4 2 2 1 2 2 2 2 2 Result 4
5 2 2 2 1 2 2 2 2 Result 5
6 2 2 2 2 1 2 2 2 Result 6
7 2 2 2 2 2 1 2 2 Result 7
8 2 2 2 2 2 2 1 2 Result 8
9 2 2 2 2 2 2 2 1 Result 9

 Comparison of Result 1 and Result 2 will show the effect of parameter A

 Comparison of Result 1 and Result 3 will show the effect of parameter B
 For a 2 level, a 8 parameter experiment and 9 runs are needed.
 29

Design of Experiments

 Design of Experiments (DOE) is needed for factorial design.

 DOE was introduced by R. A. Fisher in England in the 1920’s.
 In his early applications, Fisher wanted to find out how much
rain, water, fertilizer, sunshine, etc. are needed to produce the
best crop.
 As a researcher in Electronic Control laboratory in Japan,
Taguchi carried out research with DOE techniques in the late
1940’s.
 Taguchi made the techniques user-friendly (easy to apply) and
used to improve the quality of manufactured products.
 30

Full Factorial Experiment

 In a full factorial experiment strategy factors are varied
together instead of one at a time.
 This strategy is used in experiments involving several
factors where it is necessary to study the joint effect of the
factors on a response.
 This strategy is widely used in research work.
 It considers all possible combinations for a given set of
factors and their levels.
 The number of experiments required is nk where n is the
number of levels and k is the number of factors.
 31

Full Factorial Experiment

 Main advantage: it takes into account all the main and

interaction effects, providing a full picture.

 Disadvantage: it requires a large number of experimental

runs.
 32

Example: Welded Joints

 We want to study the effect of three factors, i.e., ambient

temperate, wind velocity, and bar size on the strength of welded
joints used in railway tracks.
 A statistical experimental design that was formulated for this
study was a two-level, three-variable factorial design, i.e., a 23
factorial design.
 The two level means each variable will have only two values, i.e.
high (the maximum value) and low (the minimum value).
 The levels were chosen for each variable based on desired field
conditions to be simulated.
 33

Example:Welded Joints

Data for the two-Level, three-variable factorial design

 34

Example:Welded Joints (Codes)

Coded test Conditions Actual Test Conditions

T V B
o
Test Number X1 X2 X3 F mph 1/8''
1 -1 -1 -1 0 0 4
2 +1 -1 -1 70 0 4
3 -1 +1 -1 0 20 4
4 +1 +1 -1 70 20 4
5 -1 -1 +1 0 0 11
6 +1 -1 +1 70 0 11
7 -1 +1 +1 0 20 11
8 +1 +1 +1 70 20 11

Test Conditions for the two-level, three-variable factorial design

 35

Example: Welded Joints (Results)

Ambient Wind Bar Ultimate Tensile
Number Temperature Velocity Size Strength, ksi
(i) X1 X2 X3 Yi
1 -1 -1 -1 87.5
2 1 -1 -1 87.3
3 -1 1 -1 77.8
4 1 1 -1 87.0
5 -1 -1 1 79.1
6 1 -1 1 97.6
7 -1 1 1 78.6
8 1 1 1 87.7

Results of welding experiments

 36

Example: Welded Joints

(Geometric Representation)

[0, 20, 11] 87.7 (+1, +1, +1)

78.6
(-1, +1, +1) 7 8 [70, 20, 11]

(-1, +1, -1) 3 4 (+1, +1, -1)

87.0
[0, 20, 4] 77.8 [70, 20, 4] Wind velocity (X2)

79.1 (-1, -1, +1) Bar size (X3)

6 (+1, -1, +1)
5
[0, 0, 11] [70, 0, 11]
87.5 97.6 Ambient temperature (X1)
1 2
(+1, -1, -1)
(-1, -1, -1) 87.3
[70, 0, 4]
[0, 0, 4]

Geometric representation of the 23 factorial design

 37

Analysing the Experimental Results

 There are different methods of analyzing the data

obtained from the sets of experiments.

 Two of these methods are informal, giving approximate

but often useful information.

 The third, called Analysis of Variations (ANOVA) is more

formal and more precise and requires more work.
 38

Example: Welded Joints

(main, temperature)

 The main effect of the ambient temperature on the

ultimate tensile strength of the welds:
 Y2 + Y4 + Y6 + Y8   Y1 + Y3 + Y5 + Y7 
E1 
=   −  
4  4 
   
 87.3 + 87.0 + 97.6 + 87.7   87.5 + 77.8 + 79.1 + 78.6 
E1 =   −   = 89.90 − 80.75 = 9.15 ksi
 4   4 
 An alternative way to view the main effect:

E1 = [(Y − Y )+ (Y − Y )+ (Y − Y )+ (Y − Y )]/ 4
2 1 4 3 6 5 8 7
 39

Example: Welded Joints

(main, temperature)

78.6
87.7
77.8 Wind Bar Effect of
87.0 Velocity Size Temperature
0 mph 4/8" 87.3 - 87.5 = - 0.2 ksi
Wind velocity

20 mph 4/8" 87.0 - 77.8 = + 9.2 ksi

79.1 97.6 0 mph 11/8'' 97.6 - 79.1 = + 18.5 ksi
20 mph 11/8'' 87.7 - 78.6 = + 9.1 ksi
Bar size ETemp 89.90 - 80.75 = +9.15 ksi

87.5 Temperature 87.3

Ultimate tensile strength, ksi

89.90

80.75

Temperature

Main Effect of Temperature

 40

Example: Welded Joints

(main effect, velocity)
 The main effect of the wind velocity on the ultimate
tensile strength of the welds:
Y +Y +Y +Y  Y +Y +Y +Y 
E2 =  3 4 7 8  −  1 2 5 6 
 4   4 

 77.8 + 87.0 + 78.6 + 87.7   87.5 + 87.3 + 79.1 + 97.6 

E2 =   −   = 82.78 − 87.88 = −5.1 ksi
 4   4 

 An alternative way to view the main effect:

E2 = [(Y − Y )+ (Y − Y )+ (Y − Y )+ (Y − Y )]/ 4
3 1 4 2 7 5 8 6
 41

Example: Welded Joints

(main effect, velocity)

78.6
87.7
77.8 Ambient Bar Effect of
87.0 Temp. Size Wind Velocity
0oF
Wind velocity

4/8" 77.8 - 87.5 = - 9.7 ksi

70oF 4/8" 87.0 - 87.3 = - 0.3 ksi
79.1 97.6 0oF 11/8'' 78.6 - 79.1 = - 0.5 ksi
70oF 11/8'' 87.7 - 97.6 = - 9.9 ksi
Bar size
EWV 82.78 - 87.88 = - 5.1 ksi

87.5 Temperature 87.3

Ultimate tensile strength, ksi

87.88

82.78

Wind Velocity

Main effect of wind velocity

 42
Example: Welded Joints
(main effect, bar size)

 The main effect of the bar size on the ultimate tensile

strength of the welds:
 Y + Y6 + Y7 + Y8   Y + Y2 + Y3 + Y4 
E3 =  5  −  1
 


 4   4 
 79.1 + 97.6 + 78.6 + 87.7   87.5 + 87.3 + 77.8 + 87.0 
E3 =   −   = 85.75 − 84.90 = 0.85 ksi
 4   4 

 An alternative way to view the main effect:

E3 = [(Y5 ) ( ) ( ) (
− Y1 + Y6 − Y2 + Y7 − Y3 + Y8 − Y4 / 4 )]
 43

Example: Welded Joints

(main effect, bar size)

78.6
87.7
77.8 Ambient Wind Effect
Effectofof
87.0 Temp. Velocity Wind
BarVelocity
Size
0oF
Wind velocity

0 mph 79.1 - 87.5 = - 8.4 ksi

70oF 20 mph 97.6 - 87.3 = + 10.3 ksi
79.1 97.6 0oF 0 mph 78.6 - 77.8 = + 0.8 ksi
70oF 20 mph 87.7 - 87.0 = + 0.7 ksi
Bar size
EBS 85.75 - 84.90 = + 0.85 ksi

87.5 Temperature 87.3

Ultimate tensile strength, ksi

84.90 85.75

Bar Size

Main effect of bar size

 44

Interpretation of Interactions
 45

Example: Welded Joints

(temperature & velocity)
[0, 20, 11] 87.7
78.6
7 8 [70, 20, 11]

[0, 20, 4] 92.45

3 4
87.0
77.8 [70, 20, 4] (-) WV

Wind Velocity
Wind Velocity

87.35
79.1 83.30
97.6
5 6 (+) WV
[0, 0, 11] [70, 0, 11]
1 2 Bar size
78.20
87.5 Temperature 87.3
[70, 0, 4]
[0, 0, 4]
Temperature

(77.8+78.6)/2 = 78.20
87.35 = (87.0+87.7)/2 (87.35-78.2) = 9.15
Wind velocity

ETxWV = (9.15 - 9.15)/2 = 0

Effect of temperature
92.45 = (87.3+97.6)/2 (92.45-83.30) = 9.15 does not depend on
(87.5+79.1)/2 = 83.3 Temperature level of wind velocity

Interaction effect of temperature and wind velocity.

 46
Example: Welded Joints
(velocity & bar size)
[0, 20, 11] 87.7
78.6
7 8 [70, 20, 11]

[0, 20, 4] 3 4
87.0 88.35
77.8 [70, 20, 4] (+) BS
Wind Velocity

83.15

Bar Size
87.40
79.1 97.6 (-) BS
5 6 82.40
[0, 0, 11] [70, 0, 11]
1 2 Bar size
87.5 Temperature 87.3
[70, 0, 4]
[0, 0, 4]
Wind Velocity

83.15 = (87.7+78.6)/2

(83.15-88.35) = - 5.2
(87.0+77.8)/2=82.4
EWVxBS = (-5.2 - (- 5.0))/2 = - 0.1
Wind velocity

EffectEffect of wind
of wind
(82.40-87.40) = - 5.0 velocity mirginally
88.35 = (97.6+79.1)/2 velocity marginally
depent on level of
depends baron level of
size
Bar size
(87.3+87.5)/2 = 87.40 bar size.

Interaction effect of wind velocity and bar size

 47
Example: Welded Joints
(bar size & temperature)
[0, 20, 11] 87.7
78.6
7 8 [70, 20, 11]
92.65
[0, 20, 4] 3 4
87.0 (+) T

Temperature
[70, 20, 4]
Wind Velocity

77.8 87.15
79.1 97.6
5 6 82.65
[0, 0, 11] [70, 0, 11] (-) T
Bar size 78.85
1 2
87.5 Temperature
87.3
[70, 0, 4]
[0, 0, 4]
Bar Size

(78.6+79.1)/2 = 78.85 92.65 = (87.7+97.6)/2

Bar size
EBSxT = (5.5 - (-3.8))/2 = 4.65
87.15 = (87.3+87.0)/2 Effect
(77.8+87.5)/2 = 82.65
Temperature Effectofofbar size
temperature
strongly
strongly depend on
depends on
(78.85-82.65) = -3.8 (92.65-87.15) = 5.5 levellevel
of temperature.
of bar size

Interaction effect of bar size and temperature

 48
Example: Welded Joints Process
response
(Algebraic Representation)
Test X1 X2 X3 X1X2 X1X3 X2X3 X1X2X3 Y
1 -1 -1 -1 +1 +1 +1 -1 Y1
2 +1 -1 -1 -1 -1 +1 +1 Y2
3 -1 +1 -1 -1 +1 -1 +1 Y3
4 +1 +1 -1 +1 -1 -1 -1 Y4
5 -1 -1 +1 +1 -1 -1 +1 Y5
6 +1 -1 +1 -1 +1 -1 -1 Y6
7 -1 +1 +1 -1 -1 +1 -1 Y7
8 +1 +1 +1 +1 +1 +1 +1 Y8
Divisor 4 4 4 4 4 4 4
Effect E1 E2 E3 E12 E13 E23 E123

Design Matrix

Calculation Matrix
Matrix used for algebraic estimation of variable effects
 49

Algebraic Representation

 The X1, X2, and X3 columns are referred to as design

matrix. They define the test conditions in the 23 factorial
design.

 The next four columns (X1X2 through X1X2X3) are obtained

by forming all possible cross-product combinations of the
first three columns.

 The seven columns are referred to as calculation matrix.

Each column in the matrix is used together with the last
column data (process response) to estimate the factor
effects.
 50

Algebraic Representation

 To obtain the estimate of the main effect X1, the inner

product of column X1 and Y column is formed and then
divided by N/2. The calculation is given by

𝐸𝐸1 = −1𝑌𝑌1 + 1𝑌𝑌2 − 1𝑌𝑌3 + 1𝑌𝑌4 − 1𝑌𝑌5 + 1𝑌𝑌6 − 1𝑌𝑌7 + 1𝑌𝑌8 ⁄4

Since
N = 8 divide
the sum by 4

 The two-factor interaction effects and the three factor

interaction effects would be obtained in a similar way.
 51

Full Factorial Experiment

 In general, for a 2k factorial design:

 There are 2k unique test conditions and hence the calculation

matrix will have 2k rows.

 The calculation matrix will have 2k – 1 columns representing the

k main effects, k!/(k-1)!1! two-factor interactions, k!/(k-2)!2!
three factor interactions, …, and one k-factor interactions.

 The i-th effect estimate is given by

Ei = (2/n)[+Y1 +Y2 … + Yn]

where n = 2k and i = 1, 2, …, 2k -1.

 52

Full Factorial Experiment

 Although the class of two-level factorial designs appears

to be an efficient way to deal simultaneously with several
factors, this efficiency seems to disappear as the number
of variables to be studied grows.

 For example, a two-level, ten-variable factorial design

requires 210= 1024 experiments to estimate the following
effects:
 53

Full Factorial Experiment

(two-levels, ten-variables)

 10 main effects
 45 two-factor interaction effects
 120 three-factor interaction effects
 210 four-factor interaction effects
 252 five-factor interaction effects
 210 six-factor interaction effects
 120 seven-factor interaction effects
 45 eight-factor interaction effects
 10 nine-factor interaction effects
 1 ten-factor interaction effect
 54

Fractional Factorial Experiment

 A fractional factorial design is an alternative to a full

factorial design.

 Only a small set of experimental runs is selected from a

full factorial design.

 As a result, the interaction effects are often disregarded.

 The fractional factorial method is problematic as there are

no guidelines for its application and subsequent analysis.
 55

Fractional Factorial Experiment

 From a practical standpoint, it is reasonable to assume

that most of these variable effects, particularly higher
order interactions, will be quite small.

 Therefore, it is likely that fewer than 1024 tests would be

used to study effects of 10 factors.
 56

Fractional Factorial Experiment

 The reasons for these assumptions are:

 Although many variables (10) are considered to be potentially
important, in the end only a few of these (perhaps no more than
three or four at the most) will prove to be important.
 Because the relationship between the process response Y and the
independent variables Xi is expected to be smooth, interactions of
higher orders will probably be very small in magnitude, and
therefore can be neglected.

 Usually, interactions of three or more factors can be

neglected without too much loss of information.
 57
Example: Welded Joints
(fourth variable)

 After studying the welded joints for railway track let us

consider a fourth variable, i.e., the type of welding flux.

 Now there are four variables and a two level full factorial
design needs 24 = 16 experimental runs.
 58
Example: Welded Joints
(fourth variable)

 Suppose the investigators can perform only 8 runs.

 Based on the assumption about the negligible importance

of third and fourth order interaction effects, we can use
the column of plus and minus signs associated with the
123 interaction to assign the fourth variable (see slides 48
and 59).
 59
Example: Welded Joints
(fourth variable)

Fourth variable is assigned to this column (see slide 48)

 60
Example: Welded Joints
(fourth and fifth variables)

*This column is used to estimate the mean response.

Design matrix for a two-level fractional design

 61

Taguchi’s Orthogonal Arrays

 Taguchi popularized the use of orthogonal arrays as

efficient design structure.
 Orthogonality is the characteristic of the designed
experiment that allows the user to separate the effect of
each factor from those of other factors.
 Taguchi formalised the fractional factorial DOE method
and published a library of orthogonal arrays.
 Choosing the most appropriate array from the library of
arrays is the key step in applying the Taguchi methods.
 62

Example: Orthogonal Array

Test Factor X1 Factor X2 Product
xi xj xi.xj
1 +1 +1 +1
2 +1 -1 -1
3 -1 +1 -1
4 -1 -1 +1
Σxi.xj 0
 Orthogonality is represented as Σxi.xj=0 (the dot product is zero) and
factors are perpendicular to each other.
 Orthogonality allows to evaluate effects of each factor separately.
 All the level settings appear an equal number of times (i.e., 4 times (-1)
and 4 times (+1)) and all the level values are used (i.e., +1 and -1).
 Number of runs for each factor is the same (i.e., 4 runs).
 Each factor has equal weights and hence, it is a balanced design.
 63

Notation for Orthogonal Arrays

Latin square is an n × n array filled with n different symbols, each

occurring exactly once in each row and exactly once in each column.
L18 is the most preferred orthogonal array
 64

Taguchi’s Orthogonal Arrays

Orthogonality is evidenced by cancelling

combinations between factors
Factors 1 and 2 Factors 2 and 3???

11 11
12 22
21 12
22 21
 65

Taguchi’s Orthogonal Arrays

 66

Taguchi’s Orthogonal Arrays

L18(21x37) 1 2 3 4 5 6 7 8 Data
1 1 1 1 1 1 1 1 1
2 1 1 2 2 2 2 2 2
3 1 1 3 3 3 3 3 3
4 1 2 1 1 2 2 3 3
5 1 2 2 2 3 3 1 1
6 1 2 3 3 1 1 2 2
7 1 3 1 2 1 3 2 3
8 1 3 2 3 2 1 3 1
9 1 3 3 1 3 2 1 2
10 2 1 1 3 3 2 2 1
11 2 1 2 1 1 3 3 2
12 2 1 3 2 2 1 1 3
13 2 2 1 2 3 1 3 2
14 2 2 2 3 1 2 1 3
15 2 2 3 1 2 3 2 1
16 2 3 1 3 2 3 1 2
17 2 3 2 1 3 1 2 3
18 2 3 3 2 1 2 3 1
 67

Common Orthogonal Arrays

Orthogonal Number of Number of Number of Number of
array factors levels per trials required trials in a
factor by orthogonal traditional
array full factorial
experiment
3
L4(2 ) 3 2 4 8
L8(27) 7 2 8 128
L9(34) 4 3 9 81
L12(211) 11 2 12 2048
L16(215) 15 2 16 32768
L16(45) 5 4 16 1024
L18(21X37) 1 2 18 4374
7 3
The choice of which orthogonal array to use depends on the
number of (control + noise) factors and the number of levels.
 68

Quality through Design

 Quality through design increases robustness (off-line
strategy).
 Flow of information is required between designers,
manufactures and marketing people.
Design Production Use

 Pushing information back to the designer from the

downstream life of a product requires:
 Barriers between design and manufacturing to be broken
down, and
 Sales and market research expertise to be brought to the
product development team.
 69

Off-line Quality Control

Custom Requirements

 According to Taguchi there are

System Design
three stages of engineering
design process:
 System design
Parameter Design
 Parameter design
 Tolerance design
Tolerance Design

Robust Product
 70

System Design
 The overall structure of a product and product features
are defined.

 Initial settings of parameters are established by

evaluating customer requirements and manufacturing
limitations.

 Utilising the best available technology, a basic design is

developed.

 Functional, technological, and economical influences

must be taken into account.
 71

Parameter Design

 Search for design parameters that minimize the

product’s susceptibility to variation, the best
combination of the parameters.

 Select nominal values for design parameters.

 Adjust the process conditions to bring the product mean

and the process mean into coincidence.
 72
Parameter Design
(classification of factors)

 An important aspect of the Taguchi philosophy is the

classification of factors.

 Taguchi classifies the important factors that govern

product/process performance into signal, control, and
noise factors.

 Taguchi refers to those factors that affect average

response as signal factors and those that affect signal-
to-noise ratio as control factor.
 73
Parameter Design
(noise factors)

 Noise factors in the product/process environment

influence performance and they are generally
uncontrollable or very expensive to control.

 According to Taguchi, noise factors can only be described

in terms of statistical characteristics.
 74
Parameter Design
(noise factors)
 Noise factors can be broken down into outer, inner, and
variational noise.
 Outer noise (e.g., ambient temperature, humidity,
pressure, or people)
 Inner noise (e.g., deterioration, wear, fade of color,
shrinkage, and drying out)
 Variational noise*: part-to-part variation

* It is also known as product noise.

 75
Parameter Design
(signal-to-noise ratio)
 Taguchi suggests to use a signal-to-noise ratio as
a quantitative analysis tool:
Robustness ∝ Signal − to − Noise Ratio
Performance
Signal − to − Noise Ratio ∝
Variability of Performance
𝑦𝑦�
𝑆𝑆�𝑁𝑁 =
𝑠𝑠𝑦𝑦 2 where
𝑆𝑆⁄𝑁𝑁 - the signal-to-noise ratio
𝑦𝑦� - the average of response data
𝑠𝑠𝑦𝑦 2 - the variance of response data
 76
Parameter Design
(signal-to-noise ratio)

The table shows how Mean = 𝑦𝑦, � 𝜎𝜎 = 𝑠𝑠𝑦𝑦 2 and S/N = 20log(Mean/σ) [dB]
vary with variations in data. It can be concluded that S/N is a better
parameter than 𝜎𝜎 in assessing variability.
 77
Parameter Design
(signal-to-noise ratio)
 Signal-to-noise ratios (in dB) used for
𝑛𝑛
1
 the smaller-the-better: 𝑆𝑆�𝑁𝑁 = −10 log � 𝑦𝑦𝑖𝑖 2
𝑛𝑛
𝑖𝑖=1

𝑛𝑛
1 1
 the larger-the-better: 𝑆𝑆�𝑁𝑁 = −10 log � 2
𝑛𝑛 𝑦𝑦𝑖𝑖
𝑖𝑖=1

𝑦𝑦� 2
 the nominal-is-best: 𝑆𝑆�𝑁𝑁 = 10 log 2
𝑠𝑠𝑦𝑦

where n is the number of observations and y is the observed data.

 78
Example: Parameter Design of Wire-cut
Electrical Discharge Machining (WEDM)

 The aim is to investigate the dimensional accuracy of Wire-cut

Electrical Discharge Machining (WEDM) [6].

 The study has six input parameters at three levels: discharge

current (A), pulse duration (B), pulse gap frequency (C), wire speed
(D), wire tension (E), and dielectric flow rate (F).

 Three output parameters are dimensional error, flatness error, and

perpendicularity error.
(error = differences between WEDM and CNC end milling)

 The number of experiments that cover all combinations is 36 = 729.

 79
Input parameters and selection of
orthogonal array
Levels
Input parameters Unit Symbol
Level 0 Level 1 Level 2
Discharge current amp A 16.00 20.00 24.00
Pulse duration µsec B 3.00 6.00 9.00
Pulse gap frequency kHz C 40.00 50.00 60.00
Wire speed m/min D 7.00 8.00 9.00
Wire tension g E 1000.00 1100.00 1300.00
Dielectric flow rate MPa F 0.14 0.20 0.26

L27(313) orthogonal array (13 factors at three levels) was selected to

cover six parameters and their interactions and three levels.
 80

L27(313) Orthogonal Array

1
9 10 12 13

2 8, 11 5
Standard linear graph for L27

A
C D E

B F
 Six main (A to F) and two interaction (A×B) and
(A×F) effects are assigned to columns of the array Modified linear graph for
using a linear graph. column assignment

 The choice of the interactions is based on past studies.

 Columns 11, 12, 13 are ignored.
 81
Results: Flatness Error (similar for dimensional and perpendicularity errors)

 Flatness error is a zone

defined by two parallel
planes between which the Pareto Anova analysis
entire surface must lie.

 Calculate contribution
of each factor and
interaction.
For example,

� 𝐴𝐴0 = −264.80 � 𝐴𝐴1 = −304.02

� 𝐴𝐴2 = −295.97
2 2

𝑆𝑆𝐴𝐴 = � 𝐴𝐴0 − � 𝐴𝐴1 + � 𝐴𝐴0 − � 𝐴𝐴2

2

+ � 𝐴𝐴2 − � 𝐴𝐴1 = 2574.54

-28.00

Contribution ratio (%) = -29.00

SA/Total=30.23 𝑛𝑛
1 1
𝑆𝑆�𝑁𝑁 = −10 log � 2 -30.00
Mean S/N ratio (dB)

This is the contribution of the 𝑛𝑛 𝑦𝑦𝑖𝑖

𝑖𝑖=1 -31.00
current towards the flatness error.
-32.00
The higher
 Check on significant S/N, the better -33.00

interaction between each the result is. -34.00

factor/interaction and -35.00

Mean S/N ratios are calculated to determine average effect of
flatness error using -36.00
each factor on flatness error at levels 0, 1 and 2.
ANOVA. A0 A1 A2 B0 B1 B2 C0 C1 C2 D0 D1 D2 E0 E1 E2 F0 F1 F2
Machining parameter level
 82

Tolerance Design

 Tolerance design is introduced using an example of

production of transformers.

 The main steps are the determination of the Taguchi

function and the calculation of tolerance.
 83
Example: Tolerance Design
of Transformer

In the production of transformers, any output voltage that

exceeds 115 ± 25 volts is unacceptable to the customer.
Exceeding these limits results in an estimated loss of
$300. The manufacturer, however, can adjust the voltage
in the plant by changing a resistor that costs $1.00.

 Determine the Taguchi loss function.

 At what tolerance should the transformer be
manufactured?
 84
Example: Design of Transformer
(Taguchi loss function)

𝑘𝑘 = 𝐴𝐴⁄∆2 = 300⁄252 = 0.48

2
𝐿𝐿 𝑦𝑦 = 𝑘𝑘 𝑦𝑦 − 𝑚𝑚

2
𝐿𝐿 𝑦𝑦 = 0.48 𝑦𝑦 − 115
 85
Example: Design of Transfomer
(Tolerance)

Since the loss is equal to the cost of the resistor ($1.00)

we get
1 = 0.48 (y- 115)2
(y- 115) = ± (1/0.48)0.5

The design specification can be then calculated as follows

y = 115 ± 1.4 [V]

where the tolerance is equal to ± 1.4 [V].

 86

Strengths of Taguchi Methods

 Engineering design is the centre of gravity of

product/process development process.

 Definition of the roles of factors that influence

product/process performance.

 Robust Design – Parameter design concept.

 Use of the loss function for establishing links between

variation and economic performance.
 87
ISO 16336 – Standard on
Taguchi Methods

 In 2014 ISO published a standard on Taguchi Methods.

 ISO 16336: Applications of statistical and related methods to new

technology and product development process – Robust Parameter
Design (RPD)

 This standard provides guidelines for applying the optimization

method of robust parameter design to achieve robust products.

 The approach of this standard can be applied to any products that are
designed and manufactured, including machines, chemical products,
electronics, food, consumer good, software, new material, and service.
 88

References

1. Taguchi, Genichi: Taguchi on Robust Technology

Development: Bringing Quality Engineering Upstream;
ASME Press, New York, 1993
2. Dieter, G. E., Engineering Design: A Materials and Processing
Approach, McGraw-Hill Book Co., Singapore, 2000.
3. Taguchi, Chowdhury, Taguchi: Robust Engineering: Learn
how to boost quality while reducing costs and time to market;
McGraw-Hill, New York, 2000
4. Roy Ranjit: A Primer on the Taguchi Method, Van Nostrand
Reinhold, New York, 1990
 89

References
5. Taguchi, S., Workshop Materials, Int. Conference on Robust
Quality Engineering, April 22-23, 2020, Kula Lumpur
6. Islam, M. N., Rafai, N. H., and Subramanian, S.S., “An
Investigation into Dimensional Accuracy Achievable in Wire-
cut Electrical Discharge Machining”, Proceedings of the World
Congress on Engineering 2010 Vol III, WCE 2010, June 30 -
July 2, 2010, London, U.K. Accessed through Internet
(12/2/2020),
http://www.iaeng.org/publication/WCE2010/WCE2010_pp24
76-2481.pdf
7. ISO 16336: Applications of statistical and related methods to
new technology and product development process – Robust
parameter design (RPD), ISO, 2014.