4 Quality Engineering:

The Taguchi Method

4.1. Introduction 57
4.2. Electronic and Electrical Engineering 59
Functionality Evaluation of System Using Power (Amplitude) 61
Quality Engineering of System Using Frequency 65
4.3. Mechanical Engineering 73
Conventional Meaning of Robust Design 73
New Method: Functionality Design 74
Problem Solving and Quality Engineering 79
Signal and Output in Mechanical Engineering 80
Generic Function of Machining 80
When On and Off Conditions Exist 83
4.4. Chemical Engineering 85
Function of an Engine 86
General Chemical Reactions 87
Evaluation of Images 88
Functionality Such as Granulation or Polymerization Distribution 91
Separation System 91
4.5. Medical Treatment and Efficacy Experimentation 93
4.6. Software Testing 97
Two Types of Signal Factor and Software 97
Layout of Signal Factors in an Orthogonal Array 97
Software Diagnosis Using Interaction 98
System Decomposition 102
4.7. MT and MTS Methods 102
MT (Mahalanobis–Taguchi) Method 102
Application of the MT Method to a Medical Diagnosis 104
Design of General Pattern Recognition and Evaluation Procedure 108
Summary of Partial MD Groups: Countermeasure for Collinearity 114
4.8. On-line Quality Engineering 116
References 123

56 Taguchi’s Quality Engineering Handbook. Genichi Taguchi, Subir Chowdhury and Yuin Wu
Copyright © 2005 Genichi Taguchi, Subir Chowdhury, Yuin Wu.
4.1. Introduction 57

4.1. Introduction
The term robust design is in wide spread use in Europe and the United States. It
refers to the design of a product that causes no trouble under any conditions and
answers the question: What is a good-quality product? As a generic term, quality
or robust design has no meaning; it is merely an objective. A product that functions
under any conditions is obviously good. Again, saying this is meaningless. All en-
gineers attempt to design what will work under various conditions. The key issue
is not design itself but how to evaluate functions under known and unknown
conditions.
At the Research Institute of Electrical Communication in the 1950s, a telephone
exchange and telephone were designed so as to have a 40- and a 15-year design
life, respectively. These were demanded by the Bell System, so they developed a
successful crossbar telephone exchanger in the 1950s; however, it was replaced 20
years later by an electronic exchanger. From this one could infer that a 40-year
design life was not reasonable to expect.
We believe that design life is an issue not of engineering but of product plan-
ning. Thinking of it differently from the crossbar telephone exchange example,
rather than simply prolonging design life for most products, we should preserve
limited resources through our work. No one opposes the general idea that we
should design a product that functions properly under various conditions during
its design life. What is important to discover is how to assess proper functions.
The conventional method has been to examine whether a product functions
correctly under several predetermined test conditions. Around 1985, we visited the
Circuit Laboratory, one of the Bell Labs. They developed a new circuit by using
the following procedure; first, they developed a circuit of an objective function
under a standard condition; once it could satisfy the objective function, it was
assessed under 16 different types of conditions, which included different environ-
ments for use and after-loading conditions.
If the product did not work properly under some of the different conditions,
design constants (parameters) were changed so that it would function well. This
is considered parameter design in the old sense. In quality engineering (QE), to
achieve an objective function by altering design constants is called tuning. Under
the Taguchi method, functional improvement using tuning should be made under
standard conditions only after conducting stability design because tuning is im-
provement based on response analysis. That is, we should not take measures for
noises by taking advantage of cause-and-effect relationships.
The reason for this is that even if the product functions well under the 16
conditions noted above, we cannot predict whether it works under other condi-
tions. This procedure does not guarantee the product’s proper functioning within
its life span under various unknown conditions. In other words, QE is focused not
on response but on interaction between designs and signals or noises. Designing
parameters to attain an objective function is equivalent to studying first-order
moments. Interaction is related to second-order moments. First-order moments
involve a scalar or vector, whereas second-order moments are studied by two-
dimensional tensors. Quality engineering maintains that we should do tuning only
under standard conditions after completing robust design. A two-dimensional ten-
sor does not necessarily represent noises in an SN ratio. In quality engineering,
noise effects should have a continual monotonic tendency.
58 4. Quality Engineering: The Taguchi Method

Quality engineering says that we should adjust a function to a target value or

curve under standard conditions only after reducing variability. The idea is that,
first, robust design or functional improvement for certain conditions of use needs
to be completed, following which tuning (a method of adjusting a response value
to a target value or curve) is done. Yet people in charge of designing a receiver
sometimes insist that unless tuning is done first, output characteristics cannot be
attained. In this case, what can we do? This is the reason that quality engineering
requires us to acquire knowledge of measurement and analysis technologies. Sev-
eral different measurement methods should be investigated.
A study to attain an objective value or function under 16 types of noise con-
ditions by altering parameters can be regarded as one parameter design method;
however, this is different from quality engineering. Indeed, it is true that quality
characteristics have been studied by changing system parameters in a number of
design of experiments cases, utilized broadly in the Japanese engineering field,
and consequently, contributed to improving product quality. However, in terms of
efficacy, these methods have not been satisfactory. We do not describe them in
detail here because many publications do so. Improving objective characteristics
by parameters, termed design of experiments (DOE), has the following two purposes:
(1) to check out effects on objective characteristics and improve through tuning,
and (2) to check out effects on output characteristics and conduct tolerance de-
sign. The objective of this chapter is to compare quality engineering and conven-
tional methods.
Quality engineering has been used in the United States since the 1980s; how-
ever, it differs from the quality engineering detailed here, which is called the
Taguchi method, the Taguchi paradigm, or Taguchi quality engineering in Europe
and the United States. Quality engineering is robust design based on the following
three procedures: (1) orthogonal array, (2) SN ratio, and (3) loss function. These
three procedures are not robust design per se because they are used to evaluate
technical means or products. To design robustness, we are required to understand
the meaning of robust design before considering technical or management means.
Those who consider practical means are engineers and production people.
Robust design (including product and process design) generally means design-
ing a product that can function properly under various conditions of use. If this
is its lone objective, we can study reliability or fraction nondefective, whose goal
is 100%. Yet a corporation is merely a means of making a profit. Increased cost
caused by excessive quality design leads to poor profitability and potentially to
bankruptcy. The true objective of design and production is to earn a corporate
profit.
The essence of the Taguchi method lies in measurement in functional space
(to calculate a signal’s proximity to an ideal function of output using the reciprocal
of the average sum of squared deviations from an ideal function), overall mea-
surement in multidimensional space from a base point, and accuracy. Quality en-
gineering involves a research procedure for improving functionality and accuracy
of diagnosis or prediction by taking advantage of measurement methods in func-
tional space and in multidimensional space. Each improvement is an individual
engineer’s task. In the Taguchi method for hardware design, downstream repro-
ducibility is important. For this purpose, it is desirable to collect data whose input
signal and output characteristic are energy-related. Also, analysis can be conducted
using the square root of the data. On the other hand, in case of diagnosis and
4.2. Electronic and Electrical Engineering 59

prediction, as long as multidimensional data have sufficient accuracy, the equation

of Mahalanobis distance in homogeneous or unit space is used. To measure the
validity of use, an SN ratio is used by calculating the Mahalanobis distance from
objects that obviously do not belong to the homogeneous space. The SN ratio is
used to examine the effect of control factors. In the case of hardware study, control
factors include design constants, and in the case of multidimensional data, control
vectors include the selection of items. The reproducibility of control factor effects
is checked using orthogonal arrays.
In this chapter we explain the Taguchi method in the field of quality engi-
neering by explaining how to use it in several typical technological fields. Since
the concepts apply for general technological fields, the Taguchi method is classi-
fied as a generic technology. In this chapter we briefly explain its application to
each of the following fields: (1) electronic and electrical engineering, (2) me-
chanical engineering, (3) chemical engineering, (4) medical treatment and effi-
cacy experimentation, (5) software testing, (6) MT and MTS Methods, and (7)
on-line quality engineering.
Sections 4.2 to 4.5 cover quality engineering for hardware. Section 4.4 includes
engine function as a generic function. In Section 4.6 we discuss debugging tests
in terms of software quality. Section 4.7 highlights a new quality engineering
method, which includes pattern recognition in regard to diagnosis and prediction.
Section 4.8 addresses the system design of feedback quality management in the
category of the on-line quality engineering. All formulas used in this chapter are
given in Chapter 3 of Reference 1.

4.2. Electronic and Electrical Engineering

In general, the electronic and electrical engineering field emphasizes the following
two functions: to supply power and to supply information. Power includes not only
electric power but also output, such as current and voltage. However, these func-
tions are almost the same as those in mechanical system design, as discussed in
the next section. Thus, as a unique function in electronic and electrical engineer-
ing (and in mechanical engineering), we explain experimentation and evaluation
of functionality in cases where both input signal and output data are expressed in
complex numbers. The electronic and electrical technology regarding information
systems generally consists of the following three types:
1. Systems using power (amplitude): power supply or AM (amplitude modulation)
system
2. Systems using frequency: measurement or FM (frequency modulation) system
3. Systems using phase: PM (phase modulation) system
After elucidating the significance of two-stage optimization, which is common
to all categories, we provide details. Even a system classified as 2 or 3 above needs
stability and functionality to power. In contrast, a receiving system (or transmitting
and receiving system) focuses more on accuracy in receiving information than on
stability of power. Therefore, we should develop a system by considering each
functionality carefully. Energy difference for frequency difference is too small to
measure directly. As for phase, energy difference for phase difference is so small
that it cannot be measued. We need to calculate the SN ratio and sensitivity by
60 4. Quality Engineering: The Taguchi Method

measuring frequency or phase itself for 2 or 3. Therefore, a procedure is needed

for studying functionality by computing the frequency or phase data in system 2
or 3, as explained later.
Quality engineering proposes that for all three categories above, after creating
as complicated a circuit or system as possible, we implement parameter design to
determine optimal levels of design parameters in the first stage and adjust them
to target values in the second stage. This is called two-stage optimization.
An ideal function shows a relationship between a signal and an output char-
acteristic under certain conditions of use created by engineers. Indeed, no ideal
function exists in nature; however, the process of studying and designing a function
so as to match it with an ideal function is research and development. In quality
engineering, we conduct research by following two steps:
1. Reduce the variability of a function under various conditions.
2. Bring the function close to an ideal function under standard conditions.
The first step is evaluation by the SN ratio; the second, called adjustment or
tuning, brings a function close to an ideal by using traditional methods such as
least squares. Most conventional research has been conducted in reverse, that is,
the second step, then the first. In 1984, we visited the Circuit Laboratory, one of
the Bell Labs, for the first time. They were developing a new circuit. As a first step,
they developed it such that its function as a signal could match an objective func-
tion under standard conditions. Next, they tested whether its function could work
under 16 different types of conditions (environmental conditions or noise condi-
tions, including a deterioration test). If it functioned well under the 16 conditions,
its development was completed. However, if not, they altered design constants
(parameters) so that the objective function would work under all 16 conditions.
Now, setting design constants to A, B, ... , and the 16 noise conditions to N1, N2,
... , N16, we obtain
ƒ(A, B, ... , Ni) ⫽ m (i ⫽ 1, 2, ... , 16) (4.1)
In this case, m is a target value (or target curve). If we have 16 or more control
factors (called adjusting factors in quality engineering), we can solve equation (4.1).
If not, we determine A, B, ... , to minimize a sum of squared differences between
both sides of equation (4.1). In other words, we solve the following:

冘 [ƒ(A,B, ... , N ) ⫺ m]
16
2
min i (4.2)
A,B,... i⫽1

This method is termed least squares. Although functional improvement based

on this method has long been used, the problem is that we cannot predict how
the function behaves under conditions other than the preselected 16 conditions.
It should be impossible to adjust the function to be objective function under a
variety of conditions of use.
Tuning or adjustment in quality engineering, adjust a certain value to a target
value or curve, prescribes that tuning must be done only under standard condi-
tions. What type of characteristic or target curve a product should have under
standard conditions is a user’s decision and an ideal function in design. Although
a target value or curve of a product can be selected by a designer before shipping,
it is predetermined by customers of the product or parent companies. In both
cases, the target value or curve needs to be studied under standard conditions.
4.2. Electronic and Electrical Engineering 61

This is because a production engineering or manufacturing department should

produce products with a target value or function when shipped. Even if a product
meets the target function at the point of shipment, as conditions of use vary or
the product itself deteriorates when used, it sometimes malfunctions. This market
quality issue is regarded as a functionality problem for which a design department
should take full responsibility. The main objective of quality engineering is to take
measures for functional variability under various conditions of use in the market.
Indeed, we can study an ideal function for signals under standard conditions;
however, we need to improve functionality for noises in the market when designing
a product. When there is trouble in the market, no operator of a production plant
can adjust the functions in that market. Thus, a designer should not only maintain
effects for signals but also minimize functional variability caused by noises.

EFFECTIVE AND INEFFECTIVE ELECTRIC POWER Functionality

In quality engineering, to assess functionality, we regard a sum of squared mea- Evaluation of System
surements as total output, decompose it into effective and harmful parts, and im- Using Power
prove the ratio of the effective part to the harmful part by defining it as an SN (Amplitude)
ratio. To decompose measurements properly into quadratic forms, each measure-
ment should be proportional to the square root of energy. Of course, the theory
regarding quadratic form is related not to energy but to mathematics. The reason
that each measurement is expected to be proportional to a square root of energy
is that if total variation is proportional to energy or work, it can be decomposed
into a sum of signals or noises, and each gain can be added sequentially.
In an engineering field such as radio waves, electric power can be decomposed
as follows:

apparent power ⫽ effective power ⫹ reactive power

The effective power is measured in watts, whereas the apparent power is measured
in volt-amperes. For example, suppose that input is defined as the following si-
nusoidal voltage:

input ⫽ E sin(␻t ⫹ ␣) (4.3)

and output is expressed by the following current:

output ⫽ I sin(␻t ⫹ ␤) (4.4)

In this case, the effective power W is expressed as

EI
W⫽ cos(␣ ⫺ ␤) (4.5)
2

cos(␣ ⫺ ␤) is the power factor. On the other hand, since the

EI
apparent power ⫽ (4.6)
2

the reactive power is expressed by

EI
reactive power ⫽ [1 ⫺ cos(␣ ⫺ ␤)] (4.7)
2
62 4. Quality Engineering: The Taguchi Method

Since in the actual circuit (as in the mechanical engineering case of vibration),
phases ␣ and ␤ vary, we need to decompose total variation into parts, including
not only effective energy but also reactive energy. This is the reason that variation
decomposition by quadratic form in complex number or positive Hermitian form
is required.

OUTPUT DECOMPOSITION BY HERMITIAN FORM

We set the signal to M, measurement to y, and the proportionality between both
to

y ⫽ ␤M (4.8)

If we know an ideal theoretical output, y should be selected as loss. If M, y, and ␤

are all complex numbers, how can we deal with these? When y is a real number,
defining the total variation of measurement y as total output ST, we decompose ST
into useful and harmful parts:

ST ⫽ y 12 ⫹ y 22 ⫹ 䡠䡠䡠 ⫹ y n2 (4.9)

Using the Hermitian form, which deal with the quadratic form in complex num-
ber, we express total variation as

ST ⫽ y1y1 ⫹ y2y2 ⫹ 䡠䡠䡠 ⫹ yn yn (4.10)

Now y is a complex conjugate number of y. What is important here is that ST in

equation (4.10) is a positive (strictly speaking, nonnegative) real number which
reflects not only real number parts but also imaginary number parts.
The proportionality coefficient ␤ is estimated by

M1 y1 ⫹ M2 y2 ⫹ 䡠䡠䡠 ⫹ Mn yn
␤⫽ (4.11)
M1M1 ⫹ M2M2 ⫹ 䡠䡠䡠 ⫹ MnMn

In addition, the complex conjugate number of ␤ is expressed by

M1y1 ⫹ M2y2 ⫹ 䡠䡠䡠 ⫹ Mn yn

␤⫽ (4.12)
M1M1 ⫹ M2M2 ⫹ 䡠䡠䡠 ⫹ MnMn

The variation of proportional terms is calculated as

S␤ ⫽ ␤␤(M1M1 ⫹ M2M2 ⫹ 䡠䡠䡠 ⫹ MnMn)

(M1 y1 ⫹ 䡠䡠䡠 ⫹ Mn yn)(M1y1 ⫹ 䡠䡠䡠 ⫹ Mn yn)

⫽ (ƒ ⫽ 1) (4.13)
M1M1 ⫹ M2M2 ⫹ 䡠䡠䡠 ⫹ MnMn

Therefore, the error variation is computed as

Se ⫽ ST ⫺ S␤ (4.14)

Since, in this case, a compounded noise factor is not included,

Se
VN ⫽ Ve ⫽ (4.15)
n⫺1
4.2. Electronic and Electrical Engineering 63

The SN ratio and sensitivity are calculated as

(1/r) (S␤ ⫺ Ve)
␩ ⫽ 10 log (4.16)
Ve

1
S ⫽ 10 log (S ⫺ Ve) (4.17)
r ␤
Now r (effective divider), representing a magnitude of input, is expressed as
r ⫽ M1M1 ⫹ M2M2 ⫹ 䡠䡠䡠 ⫹ MnMn (4.18)
In cases where there is a three-level compounded noise factor N, data can be
tabulated as shown in Table 4.1 when a signal M has k levels. L1, L2, and L3 are
linear equations computed as
L1 ⫽ M1 y11 ⫹ M2 y12 ⫹ 䡠䡠䡠 ⫹ Mk y1k

L2 ⫽ M1 y21 ⫹ M2 y22 ⫹ 䡠䡠䡠 ⫹ Mk y2k

L1 ⫽ M1 y31 ⫹ M2 y32 ⫹ 䡠䡠䡠 ⫹ Mk y3k (4.19)

Based on these conditions, we can find the variation and variance as follows:
Total variation:
ST ⫽ y11y11 ⫹ y12y12 ⫹ 䡠䡠䡠 ⫹ y3k y3k (ƒ ⫽ 3k) (4.20)
Variation of proportional terms:
(L1 ⫹ L2 ⫹ L3)(L1 ⫹ L2 ⫹ L3)
S␤ ⫽ (ƒ ⫽ 1) (4.21)
3r
Effective divider:
r ⫽ M1M1 ⫹ M2M2 ⫹ 䡠䡠䡠 ⫹ Mk Mk (4.22)
Variation of sensitivity variability:
L1L1 ⫹ L2L2 ⫹ L3L3
SN ␤ ⫽ ⫺ S␤ (ƒ ⫽ 2) (4.23)
r
Error variation:
Se ⫽ ST ⫺ S␤ ⫺ SN ␤ (ƒ ⫽ 3k ⫺ 3) (4.24)

Table 4.1
Input/output data
Signal
Noise M1 M2 䡠䡠䡠 Mk Linear Equation
N1 y11 y12 䡠䡠䡠 y1k L1
N2 y21 y22 䡠䡠䡠 y2k L2
N3 y31 y32 䡠䡠䡠 y3k L3
64 4. Quality Engineering: The Taguchi Method

Total noise variation:

SN ⫽ Se ⫹ SN ␤ (ƒ ⫽ 3k ⫺ 1) (4.25)
Error variance for correction:
Se
Ve ⫽ (4.26)
3k ⫺ 3
Noise magnitude including nonlinearity (total error variance):
SN
VN ⫽ (4.27)
3k ⫺ 1
SN ratio:
(1/3r) (S␤ ⫺ Ve)
␩ ⫽ 10 log (4.28)
VN
Sensitivity:
1
S ⫽ 10 log (S ⫺ Ve) (4.29)
3r ␤
We have thus far described application of the Hermitian form to basic output
decomposition. This procedure is applicable to decomposition of almost all quad-
ratic forms because all squared terms are replaced by products of conjugates when
we deal with complex numbers. From here on we call it by a simpler term, decom-
position of variation, even when we decompose variation using the Hermitian form
because the term decomposition of variation by the Hermitian form is too lengthy.
Moreover, we wish to express even a product of two conjugates y and y as y2 in
some cases. Whereas y2 denotes itself in the case of real numbers, it denotes a
product of conjugates in the case of complex numbers. By doing this, we can use
both the formulas and degrees of freedom expressed in the positive quadratic
form. Additionally, for decomposition of variation in the Hermitian form, decom-
position formulas in the quadratic form hold true. However, as for noises, we need
to take into account phase variability as well as power variability because the noises
exist in the complex plane. Thus, we recommend that the following four-level
noise N be considered:
N1: negative condition for both power and phase

N2: negative condition for power, positive for phase

N3: positive condition for power, negative for phase

N4: positive condition for both power and phase

Change of phase is equivalent to change of reactance. By studying noises altering
reactance, we set up positive and negative levels.
Except for noise levels, we can calculate the SN ratio in the same manner as
that for the quadratic form. After maximizing the SN ratio, we aim at matching
sensitivity ␤ (a complex number) with a target value. To adjust both real and
imaginary parts to target values, instead of sensitivity ␤2 we need to consider con-
trol factor effects for ␤ itself.
4.2. Electronic and Electrical Engineering 65

If we compute a theoretical output value in case of no loss, by redefining a

deviation from a theoretical output value as an output datum y, we analyze with
SN ratio and sensitivity. Although the calculation procedure is the same, sensitivity
with loss should be smaller. In some cases, for a smaller-the-better function, we use
a smaller-the-better SN ratio using all data.

HOW TO HANDLE FREQUENCY Quality Engineering

Setting frequency to ƒ and using Planck’s constant h, we obtain frequency energy of System Using
E as follows: Frequency

E ⫽ hf (4.30)

When E is expressed in joules ( J), h becomes the following tiny number:

h ⫽ 6.62 ⫻ 10⫺34 J/s (4.31)

In this case, E is too small to measure. Since we cannot measure frequency as part
of wave power, we need to measure frequency itself as quantity proportional to
energy. As for output, we should keep power in an oscillation system sufficiently
stable. Since the system’s functionality is discussed in the preceding section, we
describe only a procedure for measuring the functionality based on frequency.
If a signal has one level and only stability of frequency is in question, measure-
ment of time and distance is regarded as essential. Stability for this case is nominal-
the-best stability of frequency, which is used for measuring time and distance. More
exactly, we sometimes take a square root of frequency.

STABILITY OF TRANSMITTING WAVE

The stability of the transmitting wave is considered most important when we deal
with radio waves. (A laser beam is an electromagnetic wave; however, we use the
term radio wave. The procedure discussed here applies also for infrared light, vis-
ible light, ultraviolet light, x-rays, and gamma rays.) The stability of radio waves is
categorized as either stability of phase or frequency or as stability of power. In the
case of frequency-modulated (FM) waves, the former is more important because
even if output fluctuates to some degree, we can easily receive phase or frequency
properly when phase or frequency is used.
In Section 7.2 of Reference 2, for experimentation purposes the variability in
power source voltage is treated as a noise. The stability of the transmitting wave is
regarded as essential for any function using phase or frequency. Since only a single
frequency is sufficient to stabilize phase in studying the stability of the transmitting
wave, we can utilize a nominal-the-best SN ratio in quality engineering. More spe-
cifically, by selecting variability in temperature or power source voltage or deteri-
oration as one or more noises, we calculate the nominal-the-best SN ratio. If the
noise has four levels, by setting the frequency data at N1, N2, N3, and N4 to y1, y2,
y3, and y4, we compute the nominal-the-best SN ratio and sensitivity as follows:

–14 (Sm ⫺ Ve)

␩ ⫽ 10 log (4.32)
Ve

S ⫽ 10 log –14 (Sm ⫺ Ve) (4.33)

66 4. Quality Engineering: The Taguchi Method

Now

(y1 ⫹ y2 ⫹ y3 ⫹ y4)2
Sm ⫽ (ƒ ⫽ 1) (4.34)
4
ST ⫽ y 21 ⫹ y 22 ⫹ y 23 ⫹ y 24 (ƒ ⫽ 4) (4.35)

Se ⫽ ST ⫺ Sm (ƒ ⫽ 3) (4.36)

Ve ⫽ –13 Se (4.37)

After determining the optimal SN ratio conditions, we perform a confirmatory

experiment under those conditions and compute the corresponding frequency.
When dealing with a clock, its frequency can be set to 1 second. If we have a target
value for a frequency, we can bring the frequency close to the target by using
control factors affecting sensitivity S greatly or changing the effective inductance
or capacitance. Since tuning is relatively easy for radio waves, sensitivity data are
often of no importance.
To generalize this procedure, we recommend using multiple frequencies. We
select K levels from a certain band and measure data at two compounded noises,
N1 and N2, as illustrated in Table 4.2. For signal M, k levels can be selected arbi-
trarily in a certain band. For example, when we obtain k levels by changing the
inductance, we do not need to consider an ideal relationship between inductance
and M. If an equation representing such a relationship exists, it does not matter
so much, and we do not have to include a deviation from the equation in errors.
Therefore, we can decompose total variation as shown in Table 4.3. The effect of
signal M is the effect of all k levels.

(y11 ⫹ y12)2 ⫹ 䡠䡠䡠 ⫹ (yk1 ⫹ yk2)2

SM ⫽ (ƒ ⫽ k) (4.38)
2
[(y11 ⫹ 䡠䡠䡠 ⫹ yk1) ⫺ (y12 ⫹ 䡠䡠䡠 ⫹ yk2)]2
SN ⫽ (4.39)
2k
Se ⫽ ST ⫺ SM ⫺ SN (4.40)

ST is now the total variation of output.

Table 4.2
Cases where signal factor exists
Noise
Signal N1 N2 Total
M1 y11 y12 y1
M2 y21 y22 y2
⯗ ⯗ ⯗ ⯗
Mk yk1 yk2 yk
4.2. Electronic and Electrical Engineering 67

Table 4.3
Decomposition of total variation
Factor f S V
M k SM VM
N 1 SN
e k⫺1 Se Ve
Total 2k ST

The SN ratio and sensitivity are as follows:

–12 (VM ⫺ Ve)

␩ ⫽ 10 log (4.41)
VN

S ⫽ 10 log –21 (VM ⫺ Ve) (4.42)

Now

SM
VM ⫽ (4.43)
k
Se
Ve ⫽ (4.44)
k⫺1
SN ⫹ Se
VN ⫽ (4.45)
k

Since we consider stabilizing all frequencies in a certain range of frequency,

this procedure should be more effective than a procedure of studying each
frequency.

THE CASE OF MODULATION FUNCTION

When we transmit (transact) information using a certain frequency, we need fre-
quency modulation. The technical means of modulating frequency is considered
to be the signal. A typical function is a circuit that oscillates a radio wave into
which we modulate a transmitting wave at the same frequency as that of the trans-
mitting wave and extract a modulated signal using the difference between two
waves. By taking into account various types of technical means, we should improve
the following proportionality between input signal M and measurement charac-
teristic y for modulation:

y ⫽ ␤M (y and M take both positive and negative numbers) (4.46)

Thus, for modulation functions, we can define the same SN ratio as before.
Although y as output is important, before studying y we should research the fre-
quency stability of the transmitting wave and internal oscillation as discussed in
the preceding section.
68 4. Quality Engineering: The Taguchi Method

When we study the relationship expressed in equation (4.46) by setting a dif-

ference between a transmitting wave and an internally oscillated radio wave to
output, it is necessary for us to measure data by changing not only conditions
regarding the environment or deterioration as noises but also frequency (not a
modulated signal but a transmitting wave). If we select two levels for noise N, three
levels for frequency F, and k levels for modulated signal M, we obtain the data
shown in Table 4.4. Frequency F is called an indicative factor. According to this, we
decompose total variation as shown in Table 4.5.
Now r represents the total variation of the input modulated signal. The calcu-
lation procedure is as follows:
(L1 ⫹ L2 ⫹ 䡠䡠䡠 ⫹ L6)2
S␤ ⫽ (ƒ ⫽ 1) (4.47)
6r
(L1 ⫹ L4)2 ⫹ 䡠䡠䡠 ⫹ (L3 ⫹ L6)2
SF ␤ ⫽ ⫺ S␤ (ƒ ⫽ 2) (4.48)
2r
L21 ⫹ 䡠䡠䡠 ⫹ L26
SN(F )␤ ⫽ ⫺ S␤ ⫺ SF ␤ (ƒ ⫽ 3) (4.49)
r
Se ⫽ ST ⫺ (S␤ ⫹ SF ␤ ⫹ SN(F )␤) (F ⫽ 6k ⫺ 6) (4.50)
Using these results, we calculate the SN ratio and sensitivity:
(1/6r)(S␤ ⫺ Ve)
␩ ⫽ 10 log (4.51)
VN

1
S ⫽ 10 log (S ⫺ Ve) (4.52)
6r ␤
Now
SN(F )␤ ⫹ Se
VN ⫽ (4.53)
6k ⫺ 3

FUNCTIONALITY OF SYSTEM USING PHASE

In the case of analog modulation, modulation technology has evolved from AM
and FM to PM. Indeed, digital systems also exist; however, in quality engineering

Table 4.4
Case where modulated signal exists
Modulated Signal
Linear
Noise Frequency M1 M2 䡠䡠䡠 Mk Equation
N1 F1 y11 y12 䡠䡠䡠 y1k L1
F2 y21 y22 䡠䡠䡠 y2k L2
F3 y31 y32 䡠䡠䡠 y3k L3
N2 F1 y41 y42 䡠䡠䡠 y4k L4
F2 y51 y52 䡠䡠䡠 y5k L5
F3 y61 y62 䡠䡠䡠 y6k L6
4.2. Electronic and Electrical Engineering 69

Table 4.5
Decomposition of total variation
Factor f S V
␤ 1 S
F␤ 2 SF␤
N(F)␤ 3 SN(F)␤
e 6(k ⫺ 1) Se Ve
Total 6k ST

we should research and design only analog functions because digital systems are
included in all AM, FM, and PM, and its only difference from analog systems is
that its level value is not continuous but discrete: that is, if we can improve SN
ratio as an analog function and can minimize each interval of discrete values and
eventually enhance information density.

❒ Example

A phase-modulation digital system for a signal having a 30⬚ phase interval such as
0, 30, 60, ... , 330⬚ has a ␴ value that is expressed in the following equation,
even if its analog functional SN ratio is ⫺10 dB:

␤2
10 log ⫽ ⫺10 (4.54)
␴2

By taking into account that the unit of signal M is degrees, we obtain the fol-
lowing small value of ␴:

␴ ⫽ 31.6␤ ⫽ 3.16⬚ (4.55)

The fact that ␴ representing RMS is approximately 3⬚ implies that there is almost
no error as a digital system because 3⬚ represents one-fifth of a function limit re-
garded as an error in a digital system. In Table 4.6 all data are expressed in radians.
Since the SN is 48.02 dB, ␴ is computed as follows:

␴ ⫽ 兹10⫺4.802 ␤ ⫽ 0.00397 rad

⫽ 0.22⬚ (4.56)

The ratio of this ␴ to the function limit of Ⳳ15⬚ is 68. Then, if noises are selected
properly, almost no error occurs.

In considering the phase modulation function, in most cases frequency is han-

dled as an indicative factor (i.e., a use-related factor whose effect is not regarded
as noise.) The procedure of calculating its SN ratio is the same as that for a system
70 4. Quality Engineering: The Taguchi Method

Table 4.6
Data for phase shifter (experiment 1 in L18 orthogonal array; rad)
Voltage
Temperature Frequency V1 V2 V3
T1 F1 77 ⫹ j101 248 ⫹ j322 769 ⫹ j1017
F2 87 ⫹ j104 280 ⫹ j330 870 ⫹ j1044
F3 97 ⫹ j106 311 ⫹ j335 970 ⫹ j1058
T2 F1 77 ⫹ j102 247 ⫹ j322 784 ⫹ j1025
F2 88 ⫹ j105 280 ⫹ j331 889 ⫹ j1052
F3 98 ⫹ j107 311 ⫹ j335 989 ⫹ j1068

using frequency. Therefore, we explain the calculation procedure using experi-

mental data in Technology Development for Electronic and Electric Industries [3].

❒ Example

This example deals with the stability design of a phase shifter to advance the phase
by 45⬚. Four control factors are chosen and assigned to an L18 orthogonal array.
Since the voltage output is taken into consideration, input voltages are selected as
a three-level signal. However, the voltage V is also a noise for phase modulation.
Additionally, as noises, temperature T and frequency F are chosen as follows:
Voltage (mV): V1 ⫽ 224, V2 ⫽ 707, V3 ⫽ 2234

Temperature (⬚C): T1 ⫽ 10, T2 ⫽ 60

Frequency (MHz): F1 ⫽ 0.9, F2 ⫽ 1.0, F3 ⫽ 1.1

Table 4.6 includes output voltage data expressed in complex numbers for 18 com-
binations of all levels of noises V, T, and F. Although the input voltage is a signal
for the output voltage, we calculate the phase advance angle based on complex
numbers because we focus only on the phase-shift function:
Y
y ⫽ tan⫺1 (4.57)
X
In this equation, X and Y indicate real and imaginary parts, respectively. Using a
scalar Y, we calculate the SN ratio and sensitivity.
As for power, we compute the SN ratio and sensitivity according to the Hermitian
form described earlier in the chapter. Because the function is a phase advance (only
45⬚ in this case) in this section, the signal has one level (more strictly, two phases,
0 and 45⬚). However, since our focal point is only the difference, all that we need
to do is to check phase advance angles for the data in Table 4.6. For the sake of
convenience, we show them in radians in Table 4.7, whereas we can also use
degrees as the unit. If three phase advance angles, 45, 90, and 135⬚, are selected,
we should consider dynamic function by setting the signal as M.
4.2. Electronic and Electrical Engineering 71

Table 4.7
Data for phase advance angle (rad)
Voltage
Temperature Frequency V1 V2 V3 Total
T1 F1 0.919 0.915 0.923 2.757
F2 0.874 0.867 0.876 2.617
F3 0.830 0.823 0.829 2.482
T2 F1 0.924 0.916 0.918 2.758
F2 0.873 0.869 0.869 2.611
F3 0.829 0.823 0.824 2.476

Since we can adjust the phase angle to 45⬚ or 0.785 rad (the target angle),
here we pay attention to the nominal-the-best SN ratio to minimize variability. Now
T and V are true noises and F is an indicative factor. Then we analyze as follows:

ST ⫽ 0.9192 ⫹ 0.9152 ⫹ 䡠䡠䡠 ⫹ 0.8242 ⫽ 13.721679

(ƒ ⫽ 18) (4.58)
15.701 2
Sm ⫽ 13.695633
18
(ƒ ⫽ 1) (4.59)
5.5152 ⫹ 5.2282 ⫹ 4.9582
SF ⫽ ⫺ Sm ⫽ 0.0258625
6
(ƒ ⫽ 2) (4.60)
Se ⫽ ST ⫺ Sm ⫺ SF ⫽ 0.0001835

(ƒ ⫽ 15) (4.61)
Se
Ve ⫽ (0.00001223) (4.62)
15

Based on these results, we obtain the SN ratio and sensitivity as follows:

18 (13.695633 ⫺ 0.00001223)
––
1
␩ ⫽ 10 log
0.00001223
⫽ 47.94 dB (4.63)
S ⫽ 10 log ––
18 (13.695633 ⫺ 0.00001223)
1

⫽ ⫺1.187 dB (4.64)
72 4. Quality Engineering: The Taguchi Method

The target angle of 45⬚ for signal S can be expressed in radians as follows:

45
S ⫽ 10 log (3.1416) ⫽ ⫺1.049 dB (4.65)
180

As a next step, under optimal conditions, we calculate the phase advance angle
and adjust it to the value in equation (4.65) under standard conditions. This process
can be completed by only one variable.
When there are multiple levels for one signal factor, we should calculate the
dynamic SN ratio. When we design an oscillator using signal factors, we may select
three levels for a signal to change phase, three levels for a signal to change output,
three levels for frequency F, and three levels for temperature as an external con-
dition, and assign them to an L9 orthogonal array. In this case, for the effective
value of output amplitude, we set a proportional equation for the signal of voltage
V as the ideal function. All other factors are noises. As for the phase modulation
function, if we select the following as levels of a phase advance and retard signal,

Level targeted at ⫺45⬚: M⬘1 ⫺ M⬘2 ⫽ m1

Level targeted at 0⬚: M⬘2 ⫺ M⬘2 ⫽ m 2

Level targeted at ⫹45⬚: M⬘3 ⫺ M⬘2 ⫽ m3

the signal becomes zero-point proportional. The ideal function is a proportional

equation where both the signal and output go through the zero point and take
positive and negative values. If we wish to select more signal levels, for example,
two levels for each of the positive and negative sides (in total, five levels, M1, M2,
M3, M4, and M5), we should set them as follows:

M1 ⫽ ⫺120⬚

M2 ⫽ ⫺60⬚

M3 ⫽ M4 ⫽ 0 (dummy)

M5 ⫽ ⫹ 60⬚

M6 ⫽ ⫹ 120⬚

Next, together with voltage V, temperature T, and frequency F, we allocate them

to an L18 orthogonal array to take measurements. On the other hand, by setting M1
⫽ ⫺150⬚, M2 ⫽ ⫺90⬚, M3 ⫽ ⫺30⬚, M4 ⫽ ⫹30⬚, M5 ⫽ ⫹90⬚, and M6 ⫽ ⫹150⬚,
we can assign them to an L18 array. After calibrating the data measured using M0
⫽ 0⬚, they are zero proportional.

DESIGN BY SIMULATION
In electronic and electric systems, theories showing input/output relationships are
available for analysis. In some cases, approximate theories are available. Instead of
experimenting, we can study functionality using simulation techniques on the com-
puter. About half of the electrical and electronic design examples in References 1
4.3. Mechanical Engineering 73

and 4 are related to functionality (reliability) design by simulation, even though

there are not as many electrical and electronic examples. Even if a theory is im-
perfect or partial, functionality design often goes well. This is because in quality
engineering we study basic functionality first and then adjust the outputs to target
values with a few parameters on an actual circuit or machine.
Chapter 16 of Reference 4 illustrates Hewlett-Packard’s simulation design using
complex numbers, in which real and imaginary parts are dealt with separately for
parameter design. Ideally, it would be better to design parameters for power ac-
cording to the procedure discussed earlier and to study phase stability based on
the technique shown earlier. It was necessary to allocate all control factors selected
to an L108 orthogonal array. In place of analyzing real and imaginary parts sepa-
rately, parameter design should be performed using the data of effective amplitude
and phase angle. However, decomposition of total variation by the Hermitian form
was not known at that time.
If simulation calculation does not take long, after creating three noise factor
levels before and after each control factor level, we should assign them to an
orthogonal array. Then, to compute the SN ratio, we measure data for the noise-
assigned orthogonal array (called the outer array) corresponding to each experi-
ment in the orthogonal array.
In the twenty-first century, simulation will play a major role in R&D. In this
situation, partially established theories or approximation methods suffice for stud-
ying functional variability because such techniques can help improve robustness.
Tuning the outputs to the objective functions may be performed using hardware.

4.3. Mechanical Engineering

A research way of reducing variability only because an objective function is varying Conventional
is regarded as whack-a-mole, or unplanned. Although this problem-solving method Meaning of Robust
should not be used, engineers have been trying to prevent problems with a new Design
product caused by current research and design. Ideally, we should prevent quality
problems in the manufacturing process or the market through optimization at the
design stage.
Although the Nippon Telegraph and Telephone Public Corporation (NTT) de-
signed exchangers or telephones in the 1950s, those who produced these products
were private companies such as NEC, Hitachi, Fujitsu, and Oki. Therefore, without
knowing how to produce them, NTT designed them in such a way that they could
prevent quality problems in the market.
While the system realization department conceptualized future transmission sys-
tems and requested devices, units, materials, and parts with new functions needed
at that time, the device realization department developed new products and tech-
nologies to meet the requests. That is, the telecommunication laboratory did not
have a design department. The results from research and development were draw-
ings and specifications, and the manufacturers of communication devices noted
above produced products according to these specs.
A design department is responsible for avoiding problems in the manufacturing
process or the marketplace before mass production or shipping is begun. There-
fore, in the design phase, manufacturers had to improve quality (functional vari-
ability) in the manufacturing process or the marketplace without knowing the type
74 4. Quality Engineering: The Taguchi Method

of manufacturing process that would be used or how quality would be controlled.

In fact, no problems occurred. Researchers knew the objective functions of prod-
ucts they had developed themselves. In fact, they knew not only the objective
functions but also the generic functions that were the means to attaining the ob-
jective functions. The mistake that researchers in Europe and the United States
made in evaluating objective or generic functions was not related to the creation
of technical means.
Although these researchers are creative and talented, their research procedure
was first to develop and design a function under standard conditions, then test
functionality under those conditions and study countermeasures against the prob-
lems through design changes. Then, by modifying product design, they take coun-
termeasures for functional variability (functionality) caused by conditions of
production or use after discovering functional problems through tests under var-
ious conditions. Needless to say, this can be regarded as ‘‘whack-a-mole’’ research.
Quality engineering recommends that we modify product design using an L18
orthogonal array before finding problems. Researchers quite often say that they
cannot consider technical measures without looking at actual problems. The con-
ventional design procedure comprises the following steps:
1. Under standard conditions, we design a product or system that each of us
considers optimal.
2. To find design problems, we test the product or system under various
conditions.
3. We change the design of the product or system (called tuning, a counter-
measure based on cause-and-effect relationships) such that it functions well
even under conditions where the original design does not work properly.
Quite often the second and third steps are repeated. We call this procedure
‘‘whack-a-mole.’’ Quality engineering insists that such research methods be
stopped.

New Method: Quality engineering recommends improving functionality before a quality problem
Functionality Design (functional variability) occurs. In other words, before a problem happens, a de-
signer makes improvements. Quite a few researchers or designers face difficulties
understanding how to change their designs. Quality engineering advises that we
check on changes in functionality using SN ratios by taking advantage of any type
of design change. A factor in design change is called a control factor. Most control
factors should be assigned to an L18 orthogonal array. A researcher or designer
should have realized and defined both the objective and generic functions of the
product being designed.

❒ Example

Imagine a vehicle steering function, more specifically, when we change the orien-
tation of a car by changing the steering angle. The first engineer who designed such
a system considered how much the angle could be changed within a certain range
4.3. Mechanical Engineering 75

of steering angle. For example, the steering angle range is defined by the rotation
of a steering wheel, three rotations for each, clockwise and counterclockwise, as
follows:

(⫺360)(3) ⫽ ⫺1080⬚
(360)(3) ⫽ ⫹1080⬚

In this case, the total steering angle adds up to 2160⬚. Accordingly, the engineer
should have considered the relationship between the angle above and steering cur-
vature y, such as a certain curvature at a steering angle of ⫺1080⬚, zero curvature
at 0, and a certain curvature at ⫹1080.
For any function, a developer of a function pursues an ideal relationship between
signal M and output y under conditions of use. In quality engineering, since any
function can be expressed by an input/output relationship of work or energy, es-
sentially an ideal function is regarded as a proportionality. For instance, in the
steering function, the ideal function can be given by

y ⫽ ␤M (4.66)

In this equation, y is curvature (the reciprocal of a turning radius) and M is a

steering angle range. European or U.S. researchers (especially inventors of func-
tions) who develop a product or system know that a relationship exists between its
signal and output characteristic y. They make efforts to design functions and invent
a number of techniques in order to match a function with an ideal function under
standard conditions. This matching procedure known as tuning, utilizes the least-
squares method in mathematical evaluation.
Quality engineering maintains that we first improve functionality (functional var-
iability) under conditions of use by deferring until later the adjustment of a function
to the ideal function under standard conditions. Conditions of use are various en-
vironmental conditions under which a product or system is used, including both
environmental conditions and deterioration. For the case of the steering function,
some environmental conditions are road conditions and a tire’s frictional condition,
both of which are called noises. In quality engineering we compound all noises into
only two levels, N1 and N2, no matter how many there are.
Now, we consider compounding noises for the steering function. For the output,
curvature of radius, we define two levels of a compounded error factor N:

N1: conditions where the proportionality coefficient ␤ becomes smaller

N2: Conditions where the proportionality coefficient ␤ becomes larger

In short, N1 represents conditions where the steering angle functions well, whereas
N2 represents conditions where it does not function well. For example, the former
represents a car with new tires running on a dry asphalt road. In contrast, the latter
is a car with worn tires running on a wet asphalt or snowy road.
By changing the vehicle design, we hope to mitigate the difference between N1
and N2. Quality engineering actively takes advantage of the interaction between
control factors and a compounded error factor N to improve functionality. Then we
76 4. Quality Engineering: The Taguchi Method

measure data such as curvature, as shown in Table 4.8 (in the counterclockwise
turn case, negative signs are added, and vice versa). Each Mi value means angle.
Table 4.8 shows data of curvature y, or the reciprocal of the turning radius for
each signal value of a steering angle ranging from zero to almost maximum for both
clockwise and counterclockwise directions. If we turn left or right on a normal road
or steer on a highway, the range of data should be different. The data in Table 4.8
represent the situation whether we are driving a car in a parking lot or steering in
a hairpin curve at low speed. The following factor of speed K, indicating different
conditions of use, called an indicative factor, is often assigned to an outer orthog-
onal array. This is because a signal factor value varies in accordance with each
indicative factor level.
K1: low speed (less than 20 km/h)

K2: middle speed (20 to 60 km/h)

K3: high speed (more than 60 km/h)

The SN ratio and sensitivity for the data in Table 4.8 are computed as follows.
First, we calculate two linear equations, L1 and L2, using N1 and N2:
L ⫽ M1 y11 ⫹ M2 y12 ⫹ 䡠䡠䡠 ⫹ M7y17 (4.67)
L2 ⫽ M1 y21 ⫹ M2 y21 ⫹ M2 y22 ⫹ 䡠䡠䡠 ⫹ M7y27 (4.68)
Next, we decompose the total variation of data in Table 4.8.
ST ⫽ total variation (f ⫽ 14) (4.69)
Variation of proportional terms:
(L1 ⫹ L2)2
S␤ ⫽ (4.70)
2r
where
r ⫽ M 21 ⫹ M 22 ⫹ 䡠䡠䡠 ⫹ M 27

⫽ (⫺720)2 ⫹ (⫺480)2 ⫹ 䡠䡠䡠 ⫹ 7202

⫽ 1,612,800 (4.71)

Table 4.8
Curvature data
M1 M2 M3 M4 M5 M6 M7
ⴚ720 ⴚ480 ⴚ240 0 240 480 720
N1 y11 y12 y13 y14 y15 y16 y17
N2 y21 y22 y23 y24 y25 y26 y27
4.3. Mechanical Engineering 77

Although we choose 1⬚ as a unit of magnitude of signal r, we can choose 120⬚

and eventually obtain r ⫽ 112. That is, designers can define a unit of signal as
they wish. On the other hand, the difference of proportionality coefficient for signal
M (difference between linear equations N1 and N2), SN␤, is calculated
(L1 ⫺ L2)2
SN ␤ ⫽ (4.72)
2r
Therefore, the error variation
Se ⫽ ST ⫺ (S␤ ⫹ SN ␤) (4.73)
As a result of decomposing total variation, we obtain Table 4.9.
Finally, we can compute SN ratio and sensitivity as follows:
(1/2r) (S␤ ⫺ Ve)
␩ ⫽ 10 log (4.74)
VN

L1 ⫹ L2
␤ˆ ⫽ S ⫽ (4.75)
2r
Se
Ve ⫽ (4.76)
12
SN ␤ ⫹ Se
VN ⫽ (4.77)
13
The reason that sensitivity S is calculated by equation (4.75) instead of using
S␤ is because S␤ always takes a positive value. But sensitivity is used to adjust ␤
to target, and ␤ may be either positive or negative, since the signal (M), takes both
positive and negative values.

A major aspect in which quality engineering differs from a conventional re-

search and design procedure is that QE focuses on the interaction between signal
and noise factors, which represent conditions of design and use. As we have said

Table 4.9
Decomposition of total variation
Source f S V
␤ 1 S␤
N␤
e
1
12
SN␤
Se
冎
Ve VN

Total 14 ST
78 4. Quality Engineering: The Taguchi Method

previously, in quality engineering, as a first step, we improve functionality, and

second, do tuning (adjustment) on a difference between a function at an optimal
SN ratio condition and an ideal function under standard conditions.

ADVANTAGES AND DISADVANTAGES OF THE NEW METHOD

Since our new method focuses on functions, we do not need to find problems
through a number of tests, and setting up noise conditions does not have to be
perfect. The new method has five major advantages:
1. Before product planning, it is possible to study subsystems (including over-
looked items) or total system functionality. In other words, the study of
functional variability at the product design stage is unnecessary, leading to
earlier delivery.
2. When we improve functionality using generic functions, such developed
common technologies are applicable to a variety of products. This helps not
only the current project plan, but the next and consequently streamlines the
research process.
3. Instead of considering all possible noise conditions such as in the traditional
reliability study, only a few noise conditions selected from conditions of use
are needed for analysis.
4. Since only the type of functionality is sufficient for each signal factor type,
we can attain optimal design simply and quickly. Even if there are multiple
types of signal factors, we can study them at the same time.
5. When systems are going to be connected, only timing is needed, without the
study on variability.
The disadvantages of our new method are as follows:
1. It is necessary to persuade engineers who have never thought about the
concept of signal factors to acknowledging the importance of selecting them.
However, this might also be regarded as an advantage or key role of our new
method because we can teach them engineering basics. That may stop en-
gineers using easy ways to measure something related to an objective and
let them look back to the essence of engineering. Some engineers work
extremely hard and long hours but still often experience unexpected prob-
lems in the marketplace. The necessity of preventing such problems forces
engineers to consider signal factors and to evaluate functionality using SN
ratios under a variety of noise conditions because this method of improving
design is quite effective.
2. Engineers need to calculate SN ratios, but this does not mean that they must
do the actual calculation. Calculation is not a burden to engineers, since
computer software is available. What they must do is to think about the
engineering meaning of an SN ratio, then determine what type of output
generated by the signal factor level must be collected. Engineers must un-
derstand the new evaluation method: Use only one SN ratio to express a
function.
3. In quality engineering, when we consider a generic function, both input
signal and output characteristic are not objective related but are work and
energy related. However, since we have difficulties measuring work and en-
ergy in many cases, we have no other choice but to select an objective func-
4.3. Mechanical Engineering 79

tion. For example, the steering function discussed earlier is not a generic
but an objective function. When an objective function is chosen, we need to
check whether there is an interaction between control factors after allocating
them to an orthogonal array, because their effects (differences in SN ratio,
called gain) do not necessarily have additivity. An L18 orthogonal array is
recommended because its size is appropriate enough to check on the ad-
ditivity of gain. If there is no additivity, the signal and measurement char-
acteristics we used need to be reconsidered for change.
4. Another disadvantage is the paradigm shift for orthogonal array’s role. In
quality engineering, control factors are assigned to an orthogonal array. This
is not because we need to calculate the main effects of control factors for
the measurement characteristic y. Since levels of control factors are fixed,
there is no need to measure their effects for y using orthogonal arrays. The
real objective of using orthogonal arrays is to check the reproducibility of
gains, as described in disadvantage 3.
In new product development, there is no existing engineering knowledge most
of the time. It is said that engineers are suffering from being unable to find ways
for SN ratio improvement. In fact, the use of orthogonal arrays enables engineers
to perform R&D for a totally new area. In the application, parameter-level intervals
may be increased and many parameters may be studied together. If the system
selected is not a good one, there will be little SN ratio improvement. Often, one
orthogonal array experiment can tell us the limitation on SN ratios. This is an
advantage of using an orthogonal array rather than a disadvantage.

We are asked repeatedly how we would solve a problem using quality engineering. Problem Solving and
Here are some generalities that we use. Quality Engineering
1. Even if the problem is related to a consumer’s complaint in the marketplace
or manufacturing variability, in lieu of questioning its root cause, we would
suggest changing the design and using a generic function. This is a technical
approach to problem solving through redesign.
2. To solve complaints in the marketplace, after finding parts sensitive to en-
vironmental changes or deterioration, we would replace them with robust
parts, even if it incurs a cost increase. This is called tolerance design.
3. If variability occurs before shipping, we would reduce it by reviewing process
control procedures and inspection standards. In quality engineering this is
termed on-line quality engineering. This does not involve using Shewhart control
charts or other management devices but does include the following daily
routine activities by operators:
a. Feedback control. Stabilize processes continuously for products to be pro-
duced later by inspecting product characteristics and correcting the
difference between them and target values.
b. Feedforward control. Based on the information from incoming materials or
component parts, predict the product characteristics in the next process
and calibrate processes continuously to match product characteristics with
the target.
c. Preventive maintenance. Change or repair tools periodically with or without
checkup.
80 4. Quality Engineering: The Taguchi Method

d. Inspection design. Prevent defective products from causing quality problems

in downstream processes through inspection.
These are daily routine jobs designed to maintain quality level. Rational design
of these systems is related to on-line quality engineering and often leads to consid-
erable improvement.

Signal and Output In terms of the generic function representing mechanical functionality,
in Mechanical
Engineering y ⫽ ␤M (4.78)

both signal M and output characteristic y should be proportional to the square

root of energy or work. That is, we need to analyze using its square root because
total output can be expressed as follows:

ST ⫽ y 12 ⫹ 䡠䡠䡠 ⫹ y n2 (4.79)

By doing so, we can make ST equivalent to work, and as a result, the following
simple decomposition can be completed:

total work ⫽ (work by signal) ⫹ (work by noise) (4.80)

This type of decomposition has long been used in the telecommunication field.
In cases of a vehicle’s acceleration performance, M should be the square root of
distance and y the time to reach. After choosing two noise levels, we calculate the
SN ratio and sensitivity using equation (4.80), which holds true for a case where
energy or work is expressed as a square root. But in many cases it is not clear if it
is so. To make sure that (4.80) holds true, it is checked by the additivity of control
factor effects using an orthogonal array. According to Newton’s law, the following
formula shows that distance y is proportional to squared time t 2 under constant
acceleration:

b 2
y⫽ t (4.81)
2

From a quality engineering viewpoint, it is not appropriate to use (4.81). Instead,

taking ease of calculation into account, we should use

t⫽ 冪2b 兹y
冪b
2
t ⫽ ␤M ␤⫽ 兹y ⫽ M (4.82)

Generic Function of The term generic function is used quite often in quality engineering. This is not an
Machining objective function but a function as the means to achieve the objective. Since a
function means to make something work, work should be measured as output. For
example, in case of cutting, it is the amount of material cut. On the other hand,
as work input, electricity consumption or fuel consumption can be selected. When
4.3. Mechanical Engineering 81

Ford Motor Co. experimented on functionality improvement of an engine during

idling in 1995, they chose a signal, output, and ideal function as follows:

Signal M: amount of fuel flow

Output y: indicated mean effective pressure (IMEP)
Ideal function: y ⫽ ␤M

Now, output y indicates average pressure to rotate a crankshaft, which is regarded

as a force for work. However, the generic function of an engine is a chemical
reaction, discussed in the next section.
Quality engineering maintains that what is picked up as signal, what is mea-
sured, and what types of noises are chosen depends on the theme selected by
researchers. Since the output of work is a request from users (consumers or op-
erators in manufacturing), work should be selected as signal, and the amount
needed for the work (in case of machinery, used power) should be measured as
data. In the experiment on cutting conducted by JR in 1959 (known as National
Railway at that time), the power needed for cutting was measured as net cutting
power in the unit kilowatthours.
In an actual experiment, we should take measurements as follows: Selecting test
pieces and cutting them for different durations, such as T1 ⫽ 10 s, T2 ⫽ 20 s, and
T3 ⫽ 30 s, we measure the amount of cut M (grams) and electricity consumption
y, as shown in Table 4.10. N indicates a two-level noise. According to what these
data show, we perform two different analyses.

1. Analysis of cutting performance. This analysis is based on the following ideal

function:

y ⫽ ␤M (4.83)

Since in this case, both signal of amount of cut M and data of electricity con-
sumption y are observations, M and y have six levels and the effects of N are not
considered. For a practical situation, we should take the square root of both y and
M. The linear equation is expressed as

L ⫽ M11 y11 ⫹ M12 y12 ⫹ 䡠䡠䡠 ⫹ M23y23 (4.84)

r ⫽ M 211 ⫹ M 12
2
⫹ 䡠䡠䡠 ⫹ M 23
2
(4.85)

Table 4.10
Data for machining
T1 T2 T3
N1 M11 M12 M13
y11 y12 y13
N2 M21 M22 M23
y21 y22 y23
82 4. Quality Engineering: The Taguchi Method

Regarding each variation, we find that

ST ⫽ y 211 ⫹ y 212 ⫹ 䡠䡠䡠 ⫹ y 223 (ƒ ⫽ 6) (4.86)

L2
S␤ ⫽ (ƒ ⫽ 1) (4.87)
r
Se ⫽ ST ⫺ S␤ (ƒ ⫽ 5) (4.88)

Se
Ve ⫽ (4.89)
5

Using these results, we obtain the SN ratio and sensitivity:

(1/r) (S␤ ⫺ Ve)

n ⫽ 10 log (4.90)
Ve

1
S ⫽ 10 log (S ⫺ Ve) (4.91)
r ␤

Since we analyze based on a root square, the true value of S is equal to the
estimation of ␤2, that is, consumption of electricity needed for cutting, which
should be smaller. If we perform smooth machining with a small amount of elec-
tricity, we can obviously improve dimensional accuracy and flatness.

2. Improvement of work efficiency. This analysis is not for machining performance

but for work efficiency. In this case, an ideal function is expressed as

y ⫽ ␤T (4.92)

The SN ratio and sensitivity are calculated as

ST ⫽ y 211 ⫹ y 12
2
⫹ 䡠䡠䡠 ⫹ y 23
2
(ƒ ⫽ 6) (4.93)

(L1 ⫹ L2)2
S␤ ⫽ (ƒ ⫽ 1) (4.94)
2r
(L1 ⫺ L2)2
SN ␤ ⫽ (ƒ ⫽ 1) (4.95)
2r

Now

L1 ⫽ T1 y11 ⫹ T2 y12 ⫹ T3y13 (4.96)

L2 ⫽ T1 y21 ⫹ T2 y22 ⫹ T3y23 (4.97)

r ⫽ T 21 ⫹ T 22 ⫹ T 23 (4.98)

(1/2r) (S␤ ⫺ Ve)

␩ ⫽ 10 log (4.99)
VN

1
S ⫽ 10 log (S ⫺ Ve) (4.100)
2r ␤
4.3. Mechanical Engineering 83

Then

Ve ⫽ –41 (ST ⫺ S␤ ⫺ SN ␤) (4.101)

VN ⫽ –51 (ST ⫺ S␤) (4.102)

In this case, ␤2 of the ideal function represents electricity consumption per unit
time, which should be larger. To improve the SN ratio using equation (4.99) means
to design a procedure of using a large amount of electricity smoothly. Since we
minimize electricity consumption by equations (4.90) and (4.91), we can improve
not only productivity but also quality by taking advantage of both analyses.

The efficiency ratio of y to M, which has traditionally been used by engineers,

can neither solve quality problems nor improve productivity. On the other hand,
the hourly machining amount (machining measurement) used by managers in
manufacturing cannot settle noise problems and product quality in machining.
To improve quality and productivity at the same time, after measuring the data
shown in Table 4.10 we should determine an optimal condition based on two
values, SN ratio and sensitivity. To lower sensitivity in equation (4.91) is to minimize
electric power per unit removal amount. If we cut more using a slight quantity of
electricity, we can cut smoothly, owing to improvement in the SN ratio in equation
(4.90), as well as minimize energy loss. In other words, it means to cut sharply,
thereby leading to good surface quality after cutting and improved machining
efficiency per unit energy.

Since quality engineering is a generic technology, a common problem occurs for When On and Off
the mechanical, chemical, and electronic engineering fields. We explain the func- Conditions Exist
tionality evaluation method for on and off conditions using an example of
machining.
Recently, evaluation methods of machining have made significant progress, as
shown in an experiment implemented in 1997 by Ishikawajima–Harima Heavy
Industries (IHI) and sponsored by the National Space Development Agency of
Japan (NASDA). The content of this experiment was released in the Quality En-
gineering Symposium in 1998. Instead of using transformality [setting product
dimensions input through a numerically controlled (NC) machine as signal and
corresponding dimensions after machined as measurement characteristic], they
measured input and output utilizing energy and work by taking advantage of the
concept of generic function.
In 1959, the following experiment was conducted at the Hamamatsu Plant of
the National Railway (now known as JR). In this experiment, various types of cut-
ting tools ( JIS, SWC, and new SWC cutting tools) and cutting conditions were
assigned to an L27 orthogonal array to measure the net cutting power needed for
a certain amount of cut. The experiment, based on 27 different combinations,
revealed that the maximum power needed was several times as large as the mini-
mum. This finding implies that excessive power may cause a rough surface or
variability in dimensions. In contrast, cutting with quite a small amount of power
means that we can cut sharply. Consequently, when we can cut smoothly and power
effectively, material surfaces can be machined flat and variability in dimension can
be reduced. In the Quality Engineering Symposium in June 1998, IHI released
84 4. Quality Engineering: The Taguchi Method

research on improvement in machining accuracy of a space rocket engine, entitled

‘‘Optimization of Ti High-Speed Machining Process’’ by Kenji Fujmoto and two
others from Ishikawajima–Harima Heavy Industries, Kazuyuki Higashihara from
the National Space Development Agency of Japan, and Kazuhito Takahashi from
the University of Electro-Communications.
This research did not adopt transformality, which defines an ideal function as
the following equation by setting the input signal of the NC machine N to signal
M, and a dimension of a machined test piece y to output:
y ⫽ ␤M (4.103)
The paradigm of the testing method in quality engineering maintains that we
should select energy or work as input signal M and output y. The ideal function
in equation (4.103) is not expressed by an equation proportional to energy, as IHI
clearly enunciated in the forum. Since cutting itself is work in the cutting process,
the amount of cut is regarded as work. Therefore, output is work. To make some-
thing work, we need energy or power. IHI originally wished to input electricity
consumption. However, since it changes every second, they considered it inappro-
priate for use as signal. They conducted an experiment based on Table 4.11, in
which T, M, and y indicate the number of tool paths, measured amount of cut,
and electricity consumption, respectively. In addition, as noise factor levels, two
different types of thickness of a flat plate N1 and N2, each of which also had
different Young’s modulus, were chosen.
The reason that they chose the amount of cut as signal M (as did the JR ex-
periment in 1959) is that the amount of cut is regarded as required work. The
ideal function was defined as follows, after they took the square root of both sides
of the original equation:
兹y ⫽ ␤ 兹M (4.104)
The only difference between (4.103) and (4.104) is that the former is not a
generic but an objective function, whereas the latter is a generic function repre-
senting the relationship between work and energy. Since y represents a unit of
energy, a square of y does not make sense from the viewpoint of physics. Thus, by
taking the square root of both sides beforehand, we wish to make both sides equiv-
alent to work and energy when they are squared.
On the other hand, unless the square of a measurement characteristic equals
energy or work, we do not have any rationale in calculating the SN ratio based on
a generic function. After taking the square root of both sides in equation (4.104),

Table 4.11
Cutting experiment
T1 T2 T3
N1 M: amount of cut M11 M12 M13
y: electricity consumption y11 y12 y13
N2 M: amount of cut M21 M22 M23
y: electricity consumption y21 y22 y23
4.4. Chemical Engineering 85

we compute the total variation of output ST, which is a sum of squares of the square
root of each cumulative electricity consumption:
ST ⫽ (兹y11)2 ⫹ (兹y12)2 ⫹ 䡠䡠䡠 ⫹ (兹y23)2 ⫽ y11 ⫹ y12 ⫹ 䡠䡠䡠 ⫹ y23 (ƒ ⫽ 6)
(4.105)
This is equivalent to a total sum of electricity consumptions. Subsequently, total
consumed energy ST can be decomposed as follows:
ST ⫽ (Electricity consumption used for cutting)
⫹ (electricity consumption used for loss or variability in cutting)
(4.106)
While an ideal relationship is defined by energy or work, we use the square
root of both sides of the relationship in analyzing the SN ratio. After we take the
square root of both sides, the new relationship is called a generic function.
We can compute the power used for cutting as a variation of proportional terms:
L1 ⫹ L2
S␤ ⫽ (4.107)
r1 ⫹ r2
Now L indicates a linear equation using square roots and is calculated as
L1 ⫽ 兹M11 兹y11 ⫹ 兹M12 兹y12 ⫹ 兹M13 兹y13

L2 ⫽ 兹M21 兹y21 ⫹ 兹M22 兹y22 ⫹ 兹M23 兹y23 (4.108)

r1 ⫽ (兹M11)2 ⫹ (兹M12)2 ⫹ (兹M13)2

⫽ M11 ⫹ M12 ⫹ M13

r2 ⫽ M21 ⫹ M22 ⫹ M23 (4.109)

Variation of sensitivity fluctuations N ␤,
L1 ⫺ L2
SN ␤ (4.110)
r1 ⫹ r2
According to Table 4.11, a general performance evaluation, or test of work
quantity and quality, is simple to complete. All that we need to do in designing a
real product is tuning, regarded as a relatively simple procedure. For any machine
function, after defining an ideal function of signal and output based on work and
energy, we should calculate the SN ratio and sensitivity by taking the square root
of both sides of the functional equation.

4.4. Chemical Engineering

As a function regarding chemical reaction or molecules, often we can measure
only percentage data of molecular conditions. Although we conduct R&D based
on experiments on as small a scale as possible (e.g., flasks), the ultimate goal is
optimal design of large-scale production processes. Quality engineering in
86 4. Quality Engineering: The Taguchi Method

chemical reactions deals with a technical method of studying functional robustness

on a small scale. Robustness of chemical reactions revolves around control of re-
action speed.

Function of an The genetic function of an engine is a chemical reaction. Our interest in oxygen
Engine aspirated into an engine is in how it is distributed in the exhaust.
❏ Insufficient reaction. We set the fraction of oxygen contained in CO2 and CO
to p.
❏ Sufficient reaction. We set the fraction of oxygen contained in CO2 to q.
❏ Side reaction (e.g., of Nox). We set the fraction of oxygen contained in side-
reacted substances to 1 ⫺ p ⫺ q.
Based on this definition, ideally the chemical reaction conforms to the follow-
ing exponential equations:
p ⫽ e⫺␤1T (4.111)
p ⫹ q ⫽ e⫺␤2T (4.112)
Here T represents the time needed for one cycle of the engine and p and q are
measurements in the exhaust. The total reaction rate ␤1 and side reaction rate ␤2
are calculated as
1 1
␤1 ⫽ ln (4.113)
T p
1 1
␤2 ⫽ ln (4.114)
T p⫹q
It is desirable that ␤1 become greater and ␤2 be close to zero. Therefore, we
compute the SN ratio as
␤21
␩ ⫽ 10 log (4.115)
␤32
The sensitivity is
S ⫽ 10 log ␤21 (4.116)
The larger the sensitivity, the more the engine’s output power increases, the
larger the SN ratio becomes, and the less effect the side reaction has.
Cycle time T should be tuned such that benefits achieved by the magnitude of
total reaction rate and loss by the magnitude of side reaction are balanced
optimally.
Selecting noise N (which has two levels, such as one for the starting point of
an engine and the other for 10 minutes after starting) as follows, we obtain the
data shown in Table 4.12. According to the table, we calculate the reaction rate
for N1 and N2. For an insufficient reaction,
1 1
␤11 ⫽ ln (4.117)
T p1

1 1
␤12 ⫽ ln (4.118)
T p2
4.4. Chemical Engineering 87

Table 4.12
Function of engine based on chemical reaction
N1 N2
Insufficient reaction p p1 p2
Objective reaction q q1 q2
Total p1 ⫹ q1 p2 ⫹ q2

For a side reaction,

1 1
␤21 ⫽ ln (4.119)
T p1 ⫹ q1

1 1
␤22 ⫽ ln (4.120)
T p2 ⫹ q2

Since the total reaction rates ␤11 and ␤12 are larger-the-better, their SN ratios are
as follows:

␩1 ⫽ ⫺10 log
1
2 冉1
␤211
1
⫹ 2
␤12 冊 (4.121)

Since the side reaction rates ␤21 and ␤22 are smaller-the-better, their SN ratio is

␩2 ⫽ ⫺10 log –21 (␤21

2
⫹ ␤22
2
) (4.122)

Finally, we compute the total SN ratio and sensitivity as follows:

␩ ⫽ ␩1 ⫹ ␩2 (4.123)

S ⫽ ␩1 (4.124)

If side reactions barely occur and we cannot trust their measurements in a chem- General Chemical
ical reaction experiment, we can separate a point of time for measuring the total Reactions
reaction rate (1 ⫺ p) and a point of time for measuring the side reaction rate (1
⫺ p ⫺ q). For example, we can set T1 to 1 minute and T2 to 30 minutes, as
illustrated in Table 4.13. When T1 ⫽ T2, we can use the procedure described in
the preceding section.

Table 4.13
Experimental data
T1 T2
Insufficient reaction p p11 p12
Objective reaction q q11 q12
Total p11 ⫹ q11 p12 ⫹ q12
88 4. Quality Engineering: The Taguchi Method

If the table shows that 1 ⫺ p11 is more or less than 50% and 1 ⫺ (p12 ⫹ q12)
ranges at least from 10 to 50%, this experiment is regarded as good enough. p11
and (p12 ⫹ q12) are used for calculating total reaction and side reaction rates,
respectively:

1 1
␤1 ⫽ ln (4.125)
T p11

1 1
␤2 ⫽ ln (4.126)
T2 p12 ⫹ q12

Using the equations above, we can compute the SN ratio and sensitivity as

␤21
␩ ⫽ 10 log (4.127)
␤21

S ⫽ 10 log ␤21 (4.128)

Evaluation of Images An image represents a picture such as a landscape transformed precisely on each
pixel of a flat surface. Sometimes an image of a human face is whitened compared
to the actual one; however, quality engineering does not deal with this type of case
because it is an issue of product planning and tuning.
In a conventional research method, we have often studied a pattern of three
primary colors (including a case of decomposing a gray color into three primary
colors, and mixing up three primary colors into a gray color) as a test pattern.
When we make an image using a pattern of three primary colors, a density curve
(a common logarithm of a reciprocal of permeability or reflection coefficient)
varies in accordance with various conditions (control factors). By taking measure-
ments from this type of density curve, we have studied an image. Although creating
a density curve is regarded as reasonable, we have also measured Dmax, Dmin, and
gamma from the curve. In addition to them (e.g., in television), resolution and
image distortion have been used as measurements.
Because the consumers’ demand is to cover a minimum density difference of
three primary colors (according to a filmmaker, this is the density range 0 to
10,000) that can be recognized by a human eye’s photoreceptor cell, the resolving
power should cover up to the size of the light-sensitive cell. However, quality en-
gineering does not pursue such technical limitations but focuses on improving
imaging technology. That is, its objective is to offer evaluation methods to improve
both quality and productivity.
For example, quality engineering recommends the following procedure for tak-
ing measurements:
1. Condition M1. Create an image of a test pattern using luminosity 10 times as
high and exposure time one-tenth as high as their current levels.
2. Condition M2. Create an image of a test pattern using as much luminosity
and exposure time as their current levels.
3. Condition M3. Create an image of a test pattern using luminosity one-tenth
as high and exposure time 10 times as high as their current levels.
At the three sensitivity curves, we select luminosities corresponding to a per-
meability or reflection coefficient of 0.5. For density, it is 0.301. For a more prac-
4.4. Chemical Engineering 89

tical experiment, we sometimes select seven levels for luminosity, such as 1000,
100, 10, 1, 1/10, 1/100, and 1/1000 times as much as current levels. At the same
time, we choose exposure time inversely proportional to each of them.
After selecting a logarithm of exposure time E for the value 0.301, the SN ratio
and sensitivity are calculated for analysis.
Next, we set a reading of exposure time E (multiplied by a certain decimal
value) for each of the following seven levels of signal M (logarithm of luminosity)
to y1, y2, ... , y7:

M1 ⫽ ⫺3.0, M2 ⫽ ⫺2.0, ... , M7 ⫽ ⫹3.0 (4.129)

This is a procedure for calculating a necessary image-center luminosity from a

sensitivity curve by taking into account a range of consumer uses. This procedure
is considered appropriate from the standpoint not only of quality engineering but
also the performance evaluation of photo film.
The ASA valve of a photo film indicates a necessary luminosity at the center of
an image for each film. For a permeability or reflection coefficient, a point with
value 0.5 is easiest to measure and the most appropriate for evaluation of an image,
whereas we do not create images at Dmax and Dmin. Quality engineering is a strategic
technology used to evaluate performance by taking consumer uses into account.
Consumer uses are all signals and noises. In the case of an image, a signal is color
density. To study an image by breaking down each of the miscellaneous color
densities into a pattern of three primary colors is a rationalization of the mea-
surement method.
The reason for measuring data at the center (at the point where the reflection
or permeability coefficient is 50%) is that we wish to avoid inaccurate measure-
ment, and this method has long been used in the telecommunication field. In
addition, the reason for taking luminosity and exposure time in inverse proportion
to this to see the density range is that we wish to check the linearity of color density
of an image for luminosity.
This linearity is totally different from the study of images by pixels. That is, in
dealing with digital information, we can select only the accuracy of each dot as a
measurement because total image performance improves as the accuracy of each
dot increases.
Among consumer uses are not only signals but also noises, which should be
smaller to enhance performance. For a photographic film, conditions regarding
film preservation and environment when a picture is taken are included in noises.
We select the following levels of film preservation conditions as two noise levels:

N1: good preservation condition, where the density change is small

N2: poor preservation condition, where the density change is large

In this case, a small difference between the two sensitivity curves for N1 and N2
represents a better performance. A good way to designing an experiment is to
compound all noise levels into only two levels. Thus, no matter how many noise
levels we have, we should compound them into two levels.
Since we have three levels, K1, K2, and K3, for a signal of three primary colors,
two levels for a noise, and seven levels for an output signal, the total number of
data is (7)(3)(2) ⫽ 42. For each of the three primary colors, we calculate the SN
ratio and sensitivity. Now we show only the calculation of K1. Based on Table 4.14,
90 4. Quality Engineering: The Taguchi Method

Table 4.14
Experimental data
Linear
M1 M2 䡠䡠䡠 M7 Equation
K1 N1 y11 y12 䡠䡠䡠 y17 L1
N2 y21 y22 䡠䡠䡠 y27 L2
K2 N1 y31 y32 䡠䡠䡠 y37 L3
N2 y41 y42 䡠䡠䡠 y47 L4
K3 N1 y51 y52 䡠䡠䡠 y57 L5
N2 y61 y62 䡠䡠䡠 y67 L6

we proceed with the calculation. Now, by taking into account that M has seven
levels, we view M4 ⫽ 0 as a standard point and subtract the value of M4 from M1,
M2, M3, M5, M6, and M7, each of which is set to y1, y2, y3, y5, y6, and y7.
By subtracting a reading y for M4 ⫽ 0 in case of K1, we obtain the following
linear equations:

L1 ⫽ ⫺3y11 ⫺ 2y12 ⫺ 䡠䡠䡠 ⫹ 3y17 (4.130)

L2 ⫽ ⫺ 3y21 ⫺ 䡠䡠䡠 ⫹ 3y27 (4.131)

Based on these, we define r ⫽ (⫺3)2 ⫹ (⫺2)2 ⫹ 䡠䡠䡠 ⫹32 ⫽ 28.

(L1 ⫹ L2)2
S␤ ⫽ (4.132)
2r
(L1 ⫺ L2)2
SN ␤ ⫽ (4.133)
2r
ST ⫽ y 11
2
⫹ y 12
2
⫹ 䡠䡠䡠 ⫹ y 27
2
(4.134)

Ve ⫽ ––
12 (ST ⫺ S␤ ⫺ SN ␤)
1
(4.135)

VN ⫽ ––
13 (ST ⫺ S␤)
1
(4.136)

(1/2r) (S␤ ⫺ Ve)

␩1 ⫽ 10 log (4.137)
VN

1
S1 ⫽ 10 log (S ⫺ Ve) (4.138)
2r ␤

As for K2 and K3, we calculate ␩2, S2, ␩3, and S3. The total SN ratio ␩ is computed
as the sum of ␩1, ␩2, and ␩3. To balance densities ranging from low to high for
K1, K2, and K3, we should equalize sensitivities for three primary colors, S1, S2, and
S3, by solving simultaneous equations based on two control factors. Solving them
such that

S1 ⫽ S 2 ⫽ S3 (4.139)

holds true under standard conditions is not a new method.

4.4. Chemical Engineering 91

The procedure discussed thus far is the latest method in quality engineering,
whereas we have used a method of calculating the density for a logarithm of lu-
minosity log E. The weakness of the latter method is that a density range as output
is different at each experiment. That is, we should not conduct an experiment
using the same luminosity range for films of ASA 100 and 200. For the latter, we
should perform an experiment using half of the luminosity needed for the former.
We have shown the procedure to do so.
Quality engineering is a method of evaluating functionality with a single index
␩. Selection of the means for improvement is made by specialized technologies
and specialists. Top management has the authority for investment and personnel
affairs. The result of their performance is evaluated based on a balance sheet that
cannot be manipulated. Similarly, although hardware and software can be designed
by specialists, those specialists cannot evaluate product performance at their own
discretion.

If you wish to limit granulation distribution within a certain range, you must classify Functionality Such
granules below the lower limit as excessively granulated, ones around the center as Granulation or
as targeted, and ones above the upper limit as insufficiently granulated, and pro- Polymerization
ceed with analysis by following the procedure discussed earlier for chemical re- Distribution
actions. For polymerization distribution, you can use the same technique.

Consider a process of extracting metallic copper from copper sulfide included in Separation System
ore. When various types of substances are mixed with ore and the temperature of
a furnace is raised, deoxidized copper melts, flows out of the furnace, and remains
in a mold. This is called crude copper ingot. Since we expect to extract 100% of the
copper contained in ore and convert it into crude copper (product), this ratio of
extraction is termed yield. The percentage of copper remaining in the furnace slag
is regarded as the loss ratio p. If p ⫽ 0, the yield becomes 100%. Because in this
case we wish to observe the ratio at a single point of time during reaction, cali-
bration will be complicated.
On the other hand, as the term crude copper implies, there exist a considerable
number (approximately 1 to 2% in most cases) of ingredients other than copper.
Therefore, we wish to bring the ratio of impurity contained in copper ingot, q*,
close to zero. Now, setting the mass of crude copper to A (kilograms), the mass
of slag to B (kilograms) (this value may not be very accurate), the ratio of impurity
in crude copper to q*, and the ratio of copper in slag to p*, we obtain Table 4.15
for input/output. From here we calculate the following two error ratios, p and q:
Bp*
p⫽ (4.140)
A(1 ⫺ q*) ⫹ Bp*
Aq*
q⫽ (4.141)
Aq* ⫹ B(1 ⫺ p*)
The error ratio p represents the ratio of copper molecules that is originally
contained in ore but mistakenly left in the slag after smelting. Then 1 ⫺ p is called
the yield. Subtracting this yield from 1, we obtain the error ratio p. The error ratio
q indicates the ratio of all noncopper molecules that is originally contained in the
furnace but mistakenly included in the product or crude copper. Both ratios are
calculated as a mass ratio. Even if the yield 1 ⫺ p is large enough, if the error
92 4. Quality Engineering: The Taguchi Method

Table 4.15
Input/output for copper smelting
Output
Input Product Slag Total
Copper A(1 ⫺ q*) Bp* A(1 ⫺ q*) ⫹ Bp*
Noncopper Aq* B(1 ⫺ p*) Aq* ⫹ B(1 ⫺ p*)
Total A B A⫹B

ratio q is also large, this smelting is considered inappropriate. After computing the
two error ratios p and q, we prepare Table 4.16.
Consider the two error ratios p and q in Table 4.15. If copper is supposed to
melt well and move easily in the product when the temperature in the furnace is
increased, the error ratio p decreases. However, since ingredients other than cop-
per melt well at the same time, the error ratio q rises. A factor that can decrease
p and increase q is regarded as an adequate variable for tuning. Although most
factors have such characteristics, more or less, we consider it real technology to
reduce errors regarding both p and q rather than obtaining effects by tuning. In
short, this is smelting technology with a large functionality.
To find a factor level reducing both p and q for a variety of factors, after making
an adjustment so as not to change the ratio of p to q, we should evaluate func-
tionality. This ratio is called the standard SN ratio. Since gain for the SN ratio
accords with that when p ⫽ q, no matter how great the ratio of p to q, in most
cases we calculate the SN ratio after p is adjusted equal to q. Primarily, we compute
p0 as follows when we set p ⫽ q ⫽ p0:
1
p0 ⫽ (4.142)
1 ⫹ 兹[(1/p) ⫺ 1][(1/q) ⫺ 1]
Secondarily, we calculate the standard SN ratio ␩ as
(1 ⫺ 2p0)2
␩ ⫽ 10 log (4.143)
4p0(1 ⫺ p0)
The details are given in Reference 5.
Once the standard SN ratio ␩ is computed, we determine an optimal condition
according to the average SN ratios for each control factor level, estimate SN ratios

Table 4.16
Input/output expressed by error ratios
Output
Input Product Slag Total
Copper 1⫺p p 1
Noncopper q 1⫺q 1
Total 1⫺p⫹q 1⫹p⫺q 2
4.5. Medical Treatment and Efficacy Experimentation 93

for optimal and initial conditions, and calculate gains. After this, we conduct a
confirmatory experiment for the two conditions, compute SN ratios, and compare
estimated gains with those obtained from this experiment.
Based on the experiment under optimal conditions, we calculate p and q. Unless
a sum of losses for p and q is minimized, timing is set using factors (such as the
temperature in the furnace) that influence sensitivity but do not affect the SN
ratio. In tuning, we should gradually change the adjusting factor level. This ad-
justment should be made after the optimal SN ratio condition is determined.
The procedure detailed here is even applicable to the removal of harmful el-
ements or the segregation of garbage.

4.5. Medical Treatment and Efficacy Experimentation

A key objective of quality engineering is to rationalize functional evaluation. For
example, for medical efficacy, it proposes a good method of evaluating great effect
and small side effects in a laboratory. Evaluation in a laboratory clarifies whether
we can assess both main effects and side effects without clinical tests on patients
at the same time. We discuss monitoring of individual patients later in this section.

❒ Example

First, we consider a case of evaluating the main effects on a cancer cell and the
side effects on a normal cell. Although our example is an anticancer drug, this
analytic procedure holds true for thermotherapy and radiation therapy using super-
sonic or electromagnetic waves. If possible, by taking advantage of an L18 orthogonal
array, we should study the method using a drug and such therapies at the same
time.
Quality engineering recommends that we experiment on cells or animals. We
need to alter the density of a drug to be assessed by h milligrams (e.g., 1 mg) per
unit time (e.g, 1 minute). Next, we select one cancer cell and one normal cell and
designate them M1 and M2, respectively. In quality engineering, M is called a signal
factor. Suppose that M1 and M2 are placed in a certain solution (e.g., water or a
salt solution), the density is increased by 1 mg/minute, and the length of time that
each cell survives is measured. Imagine that the cancer and normal cells die at the
eighth and fourteenth minutes, respectively. We express these data as shown in
Table 4.17, where 1 indicates ‘‘alive’’ and 0 indicates ‘‘dead.’’ In addition, M1 and

Table 4.17
Data for a single cell
T1 T2 T3 T4 T5 T6 T7 T8 T9 T10 T11 T12 T13 T14 T15
M1 1 1 1 1 1 1 1 0 0 0 0 0 0 0 0
M2 1 1 1 1 1 1 1 1 1 1 1 1 1 0 0
94 4. Quality Engineering: The Taguchi Method

M2 are cancer and normal cells, and T1, T2, ... indicate each lapse of time: 1, 2,
... minutes.
For digital data regarding a dead-or-alive (or cured-or-not cured) state, we cal-
culate LD50 (lethal dose 50). In this case, the LD50 value of M1 is 7.5 because a
cell dies between T7 and T8, whereas that of M2 is 13.5. LD50 represents the quan-
tity of drug needed until 50% of cells die.
In quality engineering, both M and T are signal factors. Although both signal and
noise factors are variables in use, a signal factor is a factor whose effect should
exist, and conversely, a noise is a factor whose effect should be minimized. That
is, if there is no difference between M1 and M2, and furthermore, between each
different amount of drug (or of time in this case), this experiment fails. Then we
regard both M and T as signal factors. Thus, we set up X1 and X2 as follows:

X1 ⫽ LD50 of a cancer cell (4.144)

X2 ⫽ LD50 of a normal cell (4.145)

In this case, the smaller X1 becomes, the better the drug’s performance becomes.
In contrast, X2 should be greater. Quality engineering terms the former the smaller-
the-better characteristic and the latter larger-the-better characteristic. The follow-
ing equation calculates the SN ratio:

X2
␩ ⫽ 20 log (4.146)
X1

In this case, since there is no noise, to increase ␩ is equivalent to enlarging the

ratio X2/X1.
For a more practical use, we often test N1, N2, ... , Nk as k cancer cells, and
N⬘1, N⬘, k as k⬘ normal cells. Suppose that k ⫽ 3 and k⬘ ⫽ 3. N1, N2, and
2 ... , N⬘
N3 can be selected as three different types of cancer cells or three cancer cells of
the same type. Or we can choose two cancer cells of the same type and one cancer
cell of a different type.
Next, we obtain dead-or-alive data, as illustrated in Table 4.18, after placing
three cells for each of M1 and M2. Table 4.19 shows the LD50 values of N1, N2,
and N3 and N⬘, 1 N⬘2, and N⬘
3.

Table 4.18
Dead-or-alive data
T1 T2 T3 T4 T5 T6 T7 T8 T9 T10 T11 T12 T13 T14 T15 T16 T17 T18 T19 T20
M1 N1 1 1 1 1 1 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0
N2 1 1 1 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0
N3 1 1 1 1 1 1 1 1 1 1 1 0 0 0 0 0 0 0 0 0
M2 N1⬘ 1 1 1 1 1 1 1 1 1 1 1 1 1 1 0 0 0 0 0 0
N2⬘ 1 1 1 1 1 1 1 1 0 0 0 0 0 0 0 0 0 0 0 0
N3⬘ 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 0
4.5. Medical Treatment and Efficacy Experimentation 95

Table 4.19
LD50 data
M1 5.5 3.5 11.5
M2 14.5 8.5 19.5

Since the data for M1 should have a smaller standard deviation as well as a
smaller average, after calculating an average of a sum of squared data, we multiply
its logarithm by 10. This is termed the smaller-the-better SN ratio:

␩1 ⫽ ⫺10 log –13 (5.52 ⫹ 3.52 ⫹ 11.52) ⫽ ⫺17.65 dB (4.147)

On the other hand, because the LD50 value of a normal cell M2 should be infinites-
imal, after computing an average sum of reciprocal data, we multiply its logarithm
by 10. This is larger-the-better SN ratio:

␩2 ⫽ ⫺10 log
1
3 冉 1
14.52
⫹
1
8.52
⫹
1
19.52 冊 ⫽ 21.50 dB (4.148)

Therefore, by summing up ␩1 and ␩2, we calculate the following ␩, which is the

total SN ratio integrating both main and side effects:

␩ ⫽ ␩1 ⫹ ␩2

⫽ ⫺17.65 ⫹ 21.50
⫽ 3.85 dB (4.149)

For example, given two drugs A1 and A2, we obtain the experimental data shown
in Table 4.20. If we compare the drugs, N1, N2, and N3 for A1 should be consistent
with N⬘,
1 N⬘2, and N⬘3 for A2. Now suppose that the data of A1 are the same as those
in Table 4.19 and that a different drug A2 has LD50 data for M1 and M2 as illustrated
in Table 4.20. A1’s SN ratio is as in equation (4.149). To compute A2’s SN ratio,
we calculate the SN ratio for the main effect ␩1 as follows:

␩1 ⫽ ⫺10 log –31 (18.52 ⫹ 11.52 ⫹ 20.52) ⫽ ⫺24.75 dB (4.150)

Table 4.20
Data for comparison experiment
A1 M1 5.5 3.5 11.5
M2 14.5 8.5 19.5
A2 M1 18.5 11.5 20.5
M2 89.5 40.5 103.5
96 4. Quality Engineering: The Taguchi Method

Next, as a larger-the-better SN ratio, we calculate the SN ratio for the side effects
␩2 as

␩2 ⫽ ⫺10 log
1
3 冉 1
89.52
⫹
1
40.52
⫹
1
103.52 冊 ⫽ 35.59 (4.151)

Therefore, we obtain the total SN ratio, combining the main and side effects, by
adding them up as follows:
␩ ⫽ ⫺24.75 ⫹ 35.59 ⫽ 10.84 (4.152)
Finally, we tabulate these results regarding comparison of A1 and A2 in Table
4.21, which is called a benchmarking test. What we find out from Table 4.21 is
that the side effect of A2 is larger than A1’s by 14.09 dB, or 25.6 times, whereas
the main effect of A2 is smaller than A1’s by 7.10 dB, or 1/5.1 times. On balance,
A2’s effect is larger than A1’s by 6.99 dB, or 5 times. This fact reveals that if we
increase an amount of A2, we can improve both main and side effects by 3.49 dB
compared to A1’s. That is, the main effect is enhanced 2.3 times, and the side
effect is reduced 1/2.3 times. By checking the SN ratio using the operating window
method, we can know whether the main and side-effects are improved at the same
time.

When we study therapy such as using heat or electromagnetic waves, we select

as many control factors as possible, assign them to an orthogonal array L18, increase
the power at a certain interval, and measure data as shown in Table 4.18. Then,
after preparing a table such as Table 4.19 for each experiment in the orthogonal
array, we calculate the SN ratios. As for remaining calculations, we can follow the
normal quality engineering procedures.
When we wish to measure temperature itself when dealing with thermotherapy,
we should choose a temperature difference from a standard body temperature.
For example, suppose that we set the starting temperature to a body temperature
(e.g., 36⬚C) when we raise the temperature of a Schale or petri dish containing
one cancer cell M1 and one normal cell M2 by 1⬚ each 10 minutes (or 5 minutes).
Indeed, it is possible to raise the temperature quickly; however, by increasing it
slowly, we should monitor how many times we need to raise the temperature until
a cell dies. In this case, the temperature T is regarded as a signal whose standard

Table 4.21
Comparison of drugs A1 and A2
Main Effect Side Effect Total
A1 ⫺17.65 21.50 3.85
A2 ⫺24.75 35.59 10.84
Gain ⫺7.10 14.09 6.99
4.6. Software Testing 97

point is 36⬚C. If the cancer cell dies at 43⬚C, the LD50 value is 6.5⬚C, the difference
from the standard point of 36⬚C. If the normal cell dies at 47⬚C, the LD50 value is
10.5⬚C. When a sonic or electromagnetic wave is used in place of temperature, we
need to select the wave’s power (e.g., raise the power by 1 W/minute.

4.6. Software Testing

Quality engineering classifies a signal type into two categories, the quality of an
active signal (in the case of software that a user uses actively) and the quality of
the MTS (Mahalanobis–Taguchi system), dealing with a passive signal. In this sec-
tion we detail the former, and in the next section we discuss the case for the latter.

Quality engineering supposes that all conditions of use necessarily belong to a Two Types of Signal
signal or noise factor. Apart from a condition where a user uses it actively or Factor and Software
passively, the effects of a signal factor are those that are essential. When we plan
to design software (software product) using a system with computers, the software
is considered a user’s active signal factor. On the other hand, when we conduct
inspection, diagnosis, or prediction using research data (including various sensing
data), the entire group of data is regarded as consisting of passive signal factors.
Signal factors should not only have a common function under various conditions
of use but also have small errors.
In this section we explain how to use the SN ratio in taking measures against
bugs when we conduct a functional test on software that a user uses actively. Soft-
ware products have a number of active signal factors and consist of various levels
of signal factors. In this section we discuss ways to measure and analyze data to
find bugs in software.

In testing software, there is a multiple-step process, that is, a number of signals. Layout of Signal
In quality engineering, we do not critque software design per se, but discuss mea- Factors in an
sures that enable us to find bugs in designed software. We propose a procedure Orthogonal Array
for checking whether software contains bugs or looking for problems with (diag-
nosis of) bugs.
As we discussed before, the number of software signal factors is equivalent to
the number of steps involved. In actuality, the number of signal factors is tremen-
dous, and the number is completely different in each factor level. With software,
how large an orthogonal array we can use should be discussed even when signal
factors need to be tested at every step or data can be measured at a certain step.
To use an L36 orthogonal array repeatedly is one of the practical methods we can
use, as shown next.
❏ Procedure 1. We set the number of multilevel signal factors to k. If k is large,
we select up to 11 signal factors of two levels and up to 12 factors of three
levels. Therefore, if k ⱕ 23, we should use an L36 orthogonal array. If k is
more than 23, we should use a greater orthogonal array (e.g., if 24 ⱕ k ⱕ
59, L108 is to be used) or use an L36 orthogonal array repeatedly to allocate
98 4. Quality Engineering: The Taguchi Method

all factors. We should lay out all signal factors as an array after reducing the
number of each factor level to two or three.
❏ Procedure 2. We conduct a test at 36 different conditions that are laid out on
an orthogonal array. If we obtain an acceptable result for each experiment
in the orthogonal array, we record a 0; conversely, if we do not obtain an
acceptable result, we record a 1. Basically by looking at output at the final
step, we judge by 0 or 1. Once all data (all experiments in case of L36) are
zero, the test is completed. If even one datum of 1 remains in all experi-
mental combination, the software is considered to have bugs.
❏ Procedure 3. As long as there are bugs, we need to improve the software
design. To find out at what step a problem occurs when there are a number
of bugs, we can measure intermediate data. In cases where there are a small
number of bugs, we analyze interactions. Its application to interactions is
discussed in Section 4.7.

Software Diagnosis For the sake of convenience, we explain the procedure by using an L36 orthogonal
Using Interaction array; this also holds true for other types of arrays. According to procedure 3 in
the preceding section, we allocate 11 two-level factors and 12 three-level factors to
an L36 orthogonal array, as shown in Table 4.22. Although one-level factors are not
assigned to the orthogonal array, they must be tested.
We set two-level signal factors to A, B, ... , and K, and three-level signal factors
to L, M, ... , and W. Using the combination of experiments illustrated in Table
4.23, we conduct a test and measure data about whether or not software functions
properly for all 36 combinations. When we measure output data for tickets and
changes in selling tickets, we are obviously supposed to set 0 if both outputs are
correct and 1 otherwise. In addition, we need to calculate an interaction for each
case of 0 and 1.
Now suppose that the measurements taken follow Table 4.23. In analyzing data,
we sum up data for all combinations of A to W. The total number of combinations
is 253, starting with AB and ending with VW. Since we do not have enough space
to show all combinations, for only six combinations of AB, AC, AL, AM, LM, and
LW, we show a two-way table, Table 4.23. For practical use, we can use a computer
to create a two-way table for all combinations.
Table 4.23 includes six two-way tables for AB, AC, AL, AM, LM, and LW, and
each table comprises numbers representing a sum of data 0 or 1 for four condi-
tions of A1B1 (Nos. 1–9), A1B2 (Nos. 10–18), A2B1 (Nos. 19–27), and A2B2 (Nos.
28–36). Despite all 253 combinations of two-way tables, we illustrate only six. After
we create a two-dimensional table for all possible combinations, we need to con-
sider the results.
Based on Table 4.23, we calculate combined effects. Now we need to create
two-way tables for combinations whose error ratio is 100% because if an error in
software exists, it leads to a 100% error. In this case, using a computer, we should
produce tables only for L2 and L3 for A2, M3 for A2, M2 for L3, W2 and W3 for L2,
and W2 and W3 for L3. Of 253 combinations of two-dimensional tables, we output
only 100% error tables, so we do not need to output those for AB and AC.
We need to enumerate not only two-way tables but also all 100% error tables
from all possible 253 tables because bugs in software are caused primarily by com-
binations of signal factor effects. It is software designers’ job to correct errors.
 Table 4.22
Layout and data of L36 orthogonal array
A B C D E F G H I J K L M N O P Q R S T U V W
No. 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 Data
1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 0
2 1 1 1 1 1 1 1 1 1 1 1 2 2 2 2 2 2 2 2 2 2 2 2 1
3 1 1 1 1 1 1 1 1 1 1 1 3 3 3 3 3 3 3 3 3 3 3 3 1
4 1 1 1 1 1 2 2 2 2 2 2 1 1 1 1 2 2 2 2 3 3 3 3 0
5 1 1 1 1 1 2 2 2 2 2 2 2 2 2 2 3 3 3 3 1 1 1 1 0
6 1 1 1 1 1 2 2 2 2 2 2 3 3 3 3 1 1 1 1 2 2 2 2 1
7 1 1 2 2 2 1 1 1 2 2 2 1 1 2 3 1 2 3 3 1 2 2 3 0
8 1 1 2 2 2 1 1 1 2 2 2 2 2 3 1 2 3 1 1 2 3 3 1 0
9 1 1 2 2 2 1 1 1 2 2 2 3 3 1 2 3 1 2 2 3 1 1 2 1
10 1 2 1 2 2 1 2 2 1 1 2 1 1 3 2 1 3 2 3 2 1 3 2 0
11 1 2 1 2 2 1 2 2 1 1 2 2 2 1 3 2 1 3 1 3 2 1 3 1
12 1 2 1 2 2 1 2 2 1 1 2 3 3 2 1 3 2 1 2 1 3 2 1 0
13 1 2 2 1 2 2 1 2 1 2 1 1 2 3 1 3 2 1 3 3 2 1 2 0
14 1 2 2 1 2 2 1 2 1 2 1 2 3 1 2 1 3 2 1 1 3 2 3 1
15 1 2 2 1 2 2 1 2 1 2 1 3 1 2 3 2 1 3 2 2 1 3 1 0
16 1 2 2 2 1 2 2 1 2 1 1 1 2 3 2 1 1 3 2 3 3 2 1 0
17 1 2 2 2 1 2 2 1 2 1 1 2 3 1 3 2 2 1 3 1 1 3 2 1
18 1 2 2 2 1 2 2 1 2 1 1 3 1 2 1 3 3 2 1 2 2 1 3 1
19 2 1 2 2 1 1 2 2 1 2 1 1 2 1 3 3 3 1 2 2 1 2 3 0
20 2 1 2 2 1 1 2 2 1 2 1 2 3 2 1 1 1 2 3 3 2 3 1 1
21 2 1 2 2 1 1 2 2 1 2 1 3 1 3 2 2 2 3 1 1 3 1 2 1
22 2 1 2 1 2 2 2 1 1 1 2 1 2 2 3 3 1 2 1 1 3 3 2 0
23 2 1 2 1 2 2 2 1 1 1 2 2 3 3 1 1 2 3 2 2 1 1 3 1
24 2 1 2 1 2 2 2 1 1 1 2 3 1 1 2 2 3 1 3 3 2 2 1 1

99
100
Table 4.22 (Continued )
A B C D E F G H I J K L M N O P Q R S T U V W
No. 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 Data
25 2 1 1 2 2 2 1 2 2 1 1 1 3 2 1 2 3 3 1 3 1 2 2 0
26 2 1 1 2 2 2 1 2 2 1 1 2 1 3 2 3 1 1 2 1 2 3 3 1
27 2 1 1 2 2 2 1 2 2 1 1 3 2 1 3 1 2 2 3 2 3 1 1 1
28 2 2 2 1 1 1 1 2 2 1 2 2 1 3 3 3 2 2 1 3 1 2 1 1
29 2 2 2 1 1 1 1 2 2 1 2 2 1 3 3 3 2 2 1 3 1 2 1 1
30 2 2 2 1 1 1 1 2 2 1 2 3 2 1 1 1 3 3 2 1 2 3 2 1
31 2 2 1 2 1 2 1 1 1 2 2 1 3 3 3 2 3 2 2 1 2 1 1 0
32 2 2 1 2 1 2 1 1 1 2 2 2 1 1 1 3 1 3 3 2 3 2 2 1
33 2 2 1 2 1 2 1 1 1 2 2 3 2 2 2 1 2 1 1 3 1 3 3 1
34 2 2 1 1 2 1 2 1 2 2 1 1 3 1 2 3 2 3 1 2 2 3 1 1
35 2 2 1 1 2 1 2 1 2 2 1 2 1 2 3 1 3 1 2 3 3 1 2 1
36 2 2 1 1 2 1 2 1 2 2 1 3 2 3 1 2 1 2 3 1 1 2 3 1
4.6. Software Testing 101

Table 4.23
Supplemental tables
(1) AB two-way table
B1 B2 Total
A1 4 4 8
A2 6 8 14
Total 10 12 22

(2) AC two-way table

C1 C2 Total
A1 4 4 8
A2 7 7 14
Total 11 11 22

(3) AL two-way table

L1 L2 L3 Total
A1 0 4 4 8
A2 2 6 6 14
Total 2 10 10 22

(4) AM two-way table

M1 M2 M3 Total
A1 1 2 5 8
A2 6 4 4 14
Total 7 6 9 22

(5) LM two-way table

M1 M2 M3 Total
L1 0 0 2 2
L2 4 2 4 10
L3 3 4 3 10
Total 7 6 9 22

(6) LW two-way table

W1 W2 W3 Total
L1 1 0 1 2
L2 2 4 4 10
L3 2 4 4 10
Total 5 8 9 22
102 4. Quality Engineering: The Taguchi Method

Once they modify 100% errors, they perform a test again based on an L36 orthog-
onal array. If 100% error combinations remain, they correct them again. Although
this procedure sounds imperfect, it quite often streamlines the debugging task
many times over.

System When, by following the method based on Table 4.22, we find it extremely difficult
Decomposition to seek root causes because of quite a few bugs, it is more effective for debugging
to break signal factors into small groups instead of selecting all of them at a time.
Some ways to do so are described below.
1. After splitting signal factors into two groups, we lay out factors in each of
them to an L18 orthogonal array. Once we correct all bugs, we repeat a test
based on an L36 orthogonal array containing all signal factors. We can reduce
bugs drastically by two L18 tests, thereby simplifying the process of seeking
root causes in an L36 array with few bugs.
2. When we are faced with difficulties finding causes in an L36 array because
of too many bugs, we halve the number of signal factors by picking alter-
native factors, such as A, C, E, G, ... , W, and check these bugs. Rather than
selecting every other factor, we can choose about half that we consider im-
portant. The fact that the number of bugs never changes reveals that all
bugs are caused by a combination of about half the signal factors. Then we
halve the number of factors. Until all root causes are detected, we continue
to follow this process. In contrast, if no bug is found in the first-half com-
binations of signal factors, we test the second half. Further, if there happens
to be no bug in this test, we can conclude that interactions between the first
and second halves generate bugs. To clarify causes, after dividing each of the
first and second halves into two groups, we investigate all four possible com-
binations of the two groups.
3. From Table 4.22, by correcting 100% errors regarding A, L, and M based on
Table 4.23, we then check for bugs in an L36 orthogonal array. Once all bugs
are eliminated, this procedure is complete.
We have thus far shown some procedures for finding the basic causes of bugs.
However, we can check them using intermediate output values instead of the final
results of a total system. This is regarded as a method of subdividing a total system
into subsystems.

4.7. MT and MTS Methods

MT (Mahalanobis– We cannot describe human health with a single measurement. The MT method
Taguchi) Method represents human health by integrating measuring characteristics. We can apply
this method to a medical examination or patient monitoring. The concept of
distance expressing interrelationships among multidimensional information was
introduced by Mahalanobis. However, since his method calculates a distance for
each population (group) of samples, it is regarded as an application of deviation
value in a one-dimensional space to a multidimensional space. Taguchi proposes
introducing distance in a multidimensional space by prescribing the Mahalanobis
4.7. MT and MTS Methods 103

distance only in the unit cluster of a population and defining variable distance in
a population in consideration of variable interrelationships among items.
The MTS method also defines a group of items close to their average as a unit
cluster in such a way that we can use it to diagnose or monitor a corporation. In
fact, both the MT and MTS methods have started to be applied to various fields
because they are outstanding methods of pattern recognition. What is most im-
portant in utilizing multidimensional information is to establish a fundamental
database. That is, we should consider what types of items to select or which groups
to collect to form the database. These issues should be determined by persons
expert in a specialized field.
In this section we detail a procedure for streamlining a medical checkup or
clinical examination using a database for a group of healthy people (referred to
subsequently as ‘‘normal’’ people). Suppose that the total number of items used
in the database is k. Using data for a group of normal people—for example, the
data of people who are examined and found to be in good health in any year after
annual medical checkups for three years in a row (if possible, the data of hundreds
of people is preferable)—we create a scale to characterize the group.
As a scale we use the distance measure of P. C. Mahalanobis, an Indian statis-
tician, introduced in his thesis in 1934. In calculating the Mahalanobis distance in
a certain group, the group needs to have homogeneous members. In other words,
a group consisting of abnormal people should not be considered. If the group
contains people with both low and high blood pressure, we should not regard the
data as a single distribution.
Indeed, we may consider a group of people suffering only from hepatitis type
A as being somewhat homogeneous; however, the group still exhibits a wide de-
viation. Now, let’s look at a group of normal people without hepatitis. For gender
and age we include both male and female and all adult age brackets. Male and
female are denoted by 0 and 1, respectively. Any item is dealt with as a quantitative
measurement in calculation. After regarding 0 and 1 data for male and female as
continuous variables, we calculate an average m and standard deviation ␴. For all
items for normal people, we compute an average value m and standard deviation
␴ and convert them below. This is called normalization.

y⫺m
Y⫽ (4.153)
␴

Suppose that we have n normal people. When we convert y1, y2, ... , yn into Y1,
Y2, ... , Yn using equation (4.153), Y1, Y2, ... , Yn has an average of 0 and a standard
deviation of 1, leading to easier understanding. Selecting two from k items arbi-
trarily, and dividing the sum of normalized products by n or calculating covariance,
we obtain a correlation coefficient.
After forming a matrix of correlation coefficients between two of k items by
calculating its inverse matrix, we compute the following square of D representing
the Mahalanobis distance:

D2 ⫽
1
k 冉冘 冊
ij
aijYiYj (4.154)
104 4. Quality Engineering: The Taguchi Method

Now aij stands for the (i, j)th element of the inverse matrix. Y1, Y2, ... , Yn are
converted from y1, y2, ... , yn based on the following equations:
y1 ⫺ m1
Y1 ⫽
␴1

y2 ⫺ m 2
Y2 ⫽
␴2

⯗
yk ⫺ mk
Yk ⫽ (4.155)
␴k
In these equations, for k items regarding a group of normal people, set the
average of each item to m1, m 2, ... , mk and the standard deviation to ␴1, ␴2, ... ,
␴n. The data of k items from a person, y1, y2, ... , yk, are normalized to obtain Y1,
Y2, ... , Yk. What is important here is that ‘‘a person’’ whose identity is unknown in
terms of normal or abnormal is an arbitrary person. If the person is normal, D 2
has a value of approximately 1; if not, it is much larger than 1. That is, the Ma-
halanobis distance D 2 indicates how far the person is from normal people.
For practical purposes, we substitute the following y (representing not the SN
ratio but the magnitude of N):
y ⫽ 10 log D 2 (4.156)
Therefore, if a certain person belongs to a group of normal people, the average
of y is 0 dB, and if the person stays far from normal people, y increases because
the magnitude of abnormality is enlarged. For example, if y is 20 dB, in terms of
D 2 the person is 100 times as far from the normal group as normal people are. In
most cases, normal people stay within the range of 0 to 2 dB.

Application of the Although, as discussed in the preceding section, we enumerate all necessary items
MT Method to a in the medical checkup case, we need to beware of selecting an item that is derived
Medical Diagnosis from two other items. For example, in considering height, weight, and obesity, we
need to narrow the items down to two because we cannot compute an inverse of
a matrix consisting of correlation coefficients.
When we calculate the Mahalanobis distance using a database for normal peo-
ple, we determine the threshold for judging normality by taking into account the
following two types of error loss: the loss caused by misjudging a normal person
as abnormal and spending time and money to do precise tests; and the loss caused
by misjudging an abnormal person as normal and losing the chance of early
treatment.

❒ Example

The example shown in Table 4.24 is not a common medical examination but a
special medical checkup to find patients with liver dysfunction, studied by Tatsuji
4.7. MT and MTS Methods 105

Table 4.24
Physiological examination items
Examination Item Acronym Normal Value
Total protein TP 6.5–7.5 g / dL
Albumin Alb 3.5–4.5 g. / dL
Cholinesterase ChE 0.60–1.00 ⌬pH
Glutamate oxaloacetate transaminase GOT 2–25 units
Glutamate pyruvate transaminase GPT 0–22 units
Lactatdehydrogenase LDH 130–250 units
Alkaline phosphatase ALP 2.0–10.0 units
␥-Glutamyltranspeptidase ␥-GTP 0–68 units
Lactic dehydrogenase LAP 120–450 units
Total cholesterol TCh 140–240 mg / dL
Triglyceride TG 70–120 g / dL
Phospholipases PL 150–250 mg / dL
Creatinine Cr 0.5–1.1 mg / dL
Blood urea nitrogen BUN 5–23 mg / dL
Uric Acid UA 2.5–8.0 mg / dL

Kanetaka at Tokyo Teishin Hospital [6]. In addition to the 15 items shown in Table
4.24, age and gender are included.. The total number of items is 17.
By selecting data from 200 people (it is desirable that several hundred people
be selected, but only 200 were chosen because of the capacity of a personal com-
puter) diagnosed as being in good health for three years at the annual medical
checkup given by Tokyo’s Teishin Hospital, the researchers established a database
of normal people. The data of normal people who were healthy for two consecutive
years may be used. Thus, we can consider the Mahalanobis distance based on the
database of 200 people. Some people think that it follows an F-distribution with
an average of 1 approximate degree of freedom of 17 for the numerator and infinite
degrees of freedom for its denominator on the basis of raw data. However, since
its distribution type is not important, it is wrong. We compare the Mahalanobis
distance with the degree of each patient’s dysfunction.
If we use decibel values in place of raw data, the data for a group of normal
people are supposed to cluster around 0 dB with a range of a few decibels. To
minimize loss by diagnosis error, we should determine a threshold. We show a
simple method to detect judgment error next.
106 4. Quality Engineering: The Taguchi Method

Table 4.25 demonstrates a case of misdiagnosis where healthy people with data
more than 6 dB away from those of the normal people are judged not normal. For
95 new persons coming to a medical checkup, Kanetaka analyzed actual diagnostic
error accurately by using a current diagnosis, a diagnosis using the Mahalanobis
distance, and close (precise) examination.
Category 1 in Table 4.25 is considered normal. Category 2 comprises a group
of people who have no liver dysfunction but had ingested food or alcohol despite
being prohibited from doing so before a medical checkup. Therefore, category 2
should be judged normal, but the current diagnosis inferred that 12 of 13 normal
people were abnormal. Indeed, the Mahalanobis method misjudged 9 normal peo-
ple as being abnormal, but this number is three less than that obtained using the
current method.
Category 3 consists of a group of people suffering from slight dysfunctions. Both
the current and Mahalanobis methods overlooked one abnormal person. Yet both
of them detected 10 of 11 abnormal people correctly.
For category 4, a cluster of 5 people suffering from moderate dysfunctions, both
methods found all persons. Since category 2 is a group of normal persons, com-
bining categories 1 and 2 as a group of liver dysfunction – , and categories 3 and
4 as a group of liver dysfunction ⫹, we summarize diagnostic errors for each method
in Table 4.26. For these contingency tables, each discriminability (0: no discrimin-
ability, 1: 100% discriminability) is calculated by the following equations (see ‘‘Sep-
aration System’’ in Section 4.4 for the theoretical background). A1’s discriminability:
[(28)(15) ⫺ (51)(1)]2
␳1 ⫽ ⫽ 0.0563 (4.157)
(79)(16)(29)(66)
A2’s discriminability:
(63)(15) ⫺ (16)(1)]2
␳2 ⫽ ⫽ 0.344 (4.158)
(76)(16)(64)(31)

Table 4.25
Medical checkup and discriminability
A2: Mahalanobis
A1: Current Method Method
Category a Normal Abnormal Normal Abnormal Total
1 27 39 59 7 66
2 1 12 4 9 13
3 1 10 1 10 11
4 0 5 0 5 5
Total 29 66 64 31 95
a
1, Normal; 2, normal but temporarily abnormal due to food and alcohol; 3, slightly abnormal;
4, moderately abnormal.
4.7. MT and MTS Methods 107

Table 4.26
2 ⫻ 2 Contingency table
A1: Current method
Diagnosis:
Liver dysfunction Normal Abnormal Total
⫺ (Normal) 28 51 79
⫹ (Abnormal) 1 15 16
Total 29 66 95

A2: Mahalanobis method

Diagnosis:
Liver dysfunction Normal Abnormal Total
⫺ (Normal) 63 16 79
⫹ (Abnormal) 1 15 16
Total 64 31 95

Each of the equations above consists of a product of peripheral frequencies for

the denominator and a squared difference of diagonal products for the numerator.
Whereas the discriminability of the current method is 5.63%, that of the Mahal-
anobis distance is 34.4%, approximately six times as large as that of the current
method. This index is the concept of squared correlation coefficient and equivalent
to what is called contribution.
On the other hand, in expressing discriminability, the following SN ratios are
regarded as more rational. A1’s SN ratio:
␳1
␩1 ⫽ 10 log
1 ⫺ ␳1

0.0563
⫽ 10 log ⫽ ⫺12.2 dB (4.159)
0.9437
A2’s SN ratio:
0.344
␩2 ⫽ 10 log ⫽ ⫺2.8 dB (4.160)
0.656
Therefore, a medical examination using the Mahalanobis distance is better by
9.7 dB or 8.7 times than the current item-by-item examination. According to Table
4.25, both methods have identical discriminability of abnormal people. However,
the current method diagnosed 51 of 79 normal people to be or possibly be abnor-
mal, thereby causing a futile close examination. whereas the Mahalanobis method
judges only 16 of 79 to be abnormal. That is, the latter enables us to eliminate
such a wasteful checkup for 35 normal people.
108 4. Quality Engineering: The Taguchi Method

In calculating the SN ratio, we should use the dynamic SN ratio, as explained

below.

Design of General MT METHOD

Pattern Recognition There is a large gap between science and technology. The former is a quest for
and Evaluation truth; the latter does not seek truth but invents a means of attaining an objective
Procedure function. In mechanical, electronic, and chemical engineering, technical means
are described in a systematic manner for each objective function. For the same
objective function, there are various types of means, which do not exist in nature.
A product planning department considers types of functions. The term product
planning, including hardware and software, means to plan products to be released
in the marketplace, most of which are related to active functions. However, there
are quite a few passive functions, such as inspection, diagnosis, and prediction. In
evaluating an active function, we focus on the magnitude of deviation from its
ideal function. Using the SN ratio, we do a functional evaluation that indicates
the magnitude of error.
In contrast, a passive function generally lies in the multidimensional world,
including time variables. Although our objective is obviously pattern recognition
in the multidimensional world, we need a brand-new method of summarizing mul-
tidimensional data. To summarize multidimensional data, quality engineering
(more properly, the Taguchi method) gives the following paradigms:
❏ Procedure 1: Select items, including time-series items. Up to several thousand
items can be analyzed by a computer.
❏ Procedure 2: Select zero-point and unit values and base space to determine a
pattern.
❏ Procedure 3: After summarizing measurement items only from the base space,
we formulate the equation to calculate the Mahalanobis distance. The Ma-
halanobis space determines the origin of multidimensional measurement
and unit value.
❏ Procedure 4: For an object that does not belong to the base space, we compute
Mahalanobis distances and assess their accuracy using an SN ratio.
❏ Procedure 5: Considering cost and the SN ratio, we sort out necessary items
for optimal diagnosis and proper prediction method.
Among these specialists are responsible for procedures 1 and 2 and quality
engineering takes care of procedures 3, 4, and 5. In quality engineering, any qual-
ity problem is attributed to functional variability. That is, functional improvement
will lead to solving any type of quality problem. Therefore, we determine that we
should improve functionality by changing various types of design conditions (when
dealing with multiple variables, and selecting items and the base space).
Since the Mahalanobis distance (variance) rests on the calculation of all items
selected, we determine the following control factors for each item or for each
group of items and, in most cases, allocate them to a two-level orthogonal array
when we study how to sort items:
❏ First factor level. Items selected (or groups of items selected) are used.
❏ Second factor level. Items selected (or groups of items selected) are not used.
4.7. MT and MTS Methods 109

We do not need to sort out items regarded as essential. Now suppose that the
number of items (groups of items) to be studied is l and an appropriate two-level
orthogonal array is LN. When l ⫽ 30, L32 is used, and when l ⫽ 100, L108 or L124
is selected.
Once control factors are assigned to LN, we formulate an equation to calculate
the Mahalanobis distance by using selected items because each experimental con-
dition in the orthogonal array shows the assignment of items. Following procedure
4, we compute the SN ratio, and for each SN ratio for each experiment in or-
thogonal array LN, we calculate the control factor effect. If certain items chosen
contribute negatively or little to improving the SN ratio, we should select an op-
timal condition by excluding the items. This is a procedure of sorting out items
used for the Mahalanobis distance.

MTS METHOD
Although an orthogonalizing technique exists that uses principal components
when we normalize data in a multidimensional space, it is often unrelated to econ-
omy and useless because it is based too heavily on mathematical background. Now
we introduce a new procedure for orthogonalizing data in a multidimensional
space, which at the same time reflects on the researchers’ objective.
We select X1, X2, ... , Xk as k-dimensional variables and define the following as
n groups of data in the Mahalanobis space:

X11, X12, ... , X1n

X21, X22, ... , X2n

⯗
Xk1, Xk2, ... , Xkn

All of the data above are normalized. That is, n data, X1, X2, ... , Xn have a mean
of zero and a variance of 1. X1, X2, ... , Xk represent the order of cost or priority,
which is an important step and should be determined by engineers. In lieu of X1,
X2, ... , Xk, we introduce new variables, x1, x 2, ... , xk that are mutually orthogonal:

x1 ⫽ X1 (4.161)

x1 is equal to X1 and has a mean of zero and a variance of 1. Variable x 2 is the

part of variable X2, that remains after removing the part related to, except for the
part involving variable x1 (or X1). We consider the following equation, where X2 is
expressed by x1:

X2 ⫽ b21x1 (4.162)

In this case b12 is not only the regression coefficient but also the correlation co-
efficient. Then the remaining part of X2, excluding the part related to x1
(regression part) or the part independent of x1, is

x 2 ⫽ X2 ⫺ b21x1 (4.163)
110 4. Quality Engineering: The Taguchi Method

Orthogonality of x1 and x 2 indicates that a sum of products of x1j and x 2j amounts

to zero in the base space:

冘 x x ⫽ 冘 X (X
n n

1j 2j 1j 2j ⫺ b21x1j) ⫽ 0 (4.164)
j⫽1 j⫽1

Thus, x1 and x 2 become orthogonal. On the other hand, whereas x1’s variance is
1, x 2’s variance ␴22, called residual contribution, is calculated by the equation
␴22 ⫽ 1 ⫺ b 221 (4.165)
The reason is that if we compute the mean square of the residuals, we obtain the
following result:
1
n
冘(X 2j ⫺ b21x1j)2 ⫽
1
n
冘X 2
2j ⫺
2
b
n 21
冘 X x ⫹ n1 (⫺b ) 冘 x
2j ij 21
2 2
1j

2n 2 n
⫽1⫺ b ⫹ b 22l
n 2l n
⫽ 1 ⫺ b 22l (4.166)
We express the third variable, X3, with x1 and x 2:
X3 ⫽ b31x1 ⫹ b32x 2 (4.167)
By multiplying both sides of equation (4.167) by x1, calculating a sum of elements
in the base space, and dividing by n, we obtain

b31 ⫽冘X x 1
nV1 3j 1j (4.168)

V ⫽ 冘x
1 2
1 1j (4.169)
n
Similarly, by multiplying both sides of equation (4.167) by x 2, we have b32

b32 ⫽冘X x 1
nV2 3j 2j (4.170)

V ⫽ 冘x
1 2
2 2j (4.171)
n
The orthogonal part of x 3 is thus a residual part, so that X3 cannot be expressed
by both X1 and X2 or both x1 and x 2. In sum,
x 3 ⫽ X3 ⫺ b31x1 ⫺ b32x 2 (4.172)
In the Mahalanobis space, x1, x 2, and x 3 are orthogonal. Since we have already
proved the orthogonality of x1 and x 2, we prove here that of x1 and x 3 and that of
x 2 and x 3. First, considering the orthogonality of x1 and x 3, we have

冘 x (X
j
1j 3j ⫺ b31x1j ⫺ b32x 2j) ⫽ 冘x X
j
1j 3j ⫺ b31 冘x 2
1j

⫽0 (4.173)
4.7. MT and MTS Methods 111

This can be derived from equation (4.167) defining b31. Similarly, the orthogonality
of x 2 and x 3 is proved according to equation (4.170), defining b32 as

冘x 2j (X3j ⫺ b31x1j ⫺ b32x 2j) ⫽ 冘x X

2j 3j ⫺ b32 冘x 2
2j

⫽0 (4.174)

For the remaining variables, we can proceed with similar calculations. The var-
iables normalized in the preceding section can be rewritten as follows:

x1 ⫽ X1

x 2 ⫽ X2 ⫺ b21x1

x 3 ⫽ X3 ⫺ b31x1 ⫺ b32x 2

⯗
xk ⫽ Xk ⫺ bk1x1 ⫺ b2kx 2 ⫺ 䡠䡠䡠 ⫺ bk(k⫺1)xk⫺1 (4.175)
Each of the orthogonalized variables x 2, ... , xk does not have a variance of 1. In
an actual case we should calculate a variance right after computing n groups of
variables, x1, x 2, ... , xn in the Mahalanobis space. As for degrees of freedom in
calculating a variance, we can regard n for x1, n ⫺ 1 for x 2, n ⫺ 2 for x 3, ... , n ⫺
k ⫹ 1 for xk. Instead, we can select n degrees of freedom for all because n » k
quite often.

1 2
V1 ⫽ (x ⫹ x 12
2
⫹ 䡠䡠䡠 ⫹ x 1n
2
)
n 11
1
V2 ⫽ (x 2 ⫹ x 22
2
⫹ 䡠䡠䡠 ⫹ x 2n
2
)
n ⫺ 1 21
⯗
1
Vk ⫽ (x 2 ⫹ x 2k2 ⫹ 䡠䡠䡠 ⫹ x 2kn) (4.176)
n ⫺ k ⫹ 1 k1

Now setting normalized, orthogonal variables to y1, y2, ... , yk, we obtain

x1
y1 ⫽
兹V1

x2
y2 ⫽
兹V2

⯗
xk
yk ⫽ (4.177)
兹Vk

All of the normalized y1, y2, ... , yk are orthogonal and have a variance of 1. The
database completed in the end contains m and ␴ as the mean and standard devi-
ation of initial observations, b21, V2, b31, b32, V2, ... , bk1, bk2, ... , bk(k⫺1), Vk as the
coefficients and variances for normalization. Since we select n groups of data, there
112 4. Quality Engineering: The Taguchi Method

are k means of m1, m 2, ... , mk, k standard deviations of ␴1, ␴2, ... , ␴k, (k ⫺ 1)k/2
coefficients, and k variances in this case. Thus, the number of necessary memory
items is as follows:
(k ⫺ 1)k k (k ⫹ 5)
k⫹k⫹ ⫹k⫽ (4.178)
2 2
The correlation matrix in the normalized Mahalanobis space turns out to be an
identity matrix. Then the correlation matrix R is expressed as

冢 冣
1 0 䡠䡠䡠 0
0 1 䡠䡠䡠 0
R⫽ (4.179)
⯗ ⯗ ... ⯗
0 0 ⯗ 1
Therefore, the inverse matrix of R, A is also an identity matrix:

冢 冣
1 0 䡠䡠䡠 0
0 1 䡠䡠䡠 0
A⫽R ⫺1
⫽ (4.180)
⯗ ⯗ ... ⯗
0 0 ⯗ 1
Using these results, we can calculate the Mahalanobis distance D 2 as
1 2
D2 ⫽ (y ⫹ y 22 ⫹ 䡠䡠䡠 ⫹ y k2) (4.181)
k 1
Now, setting measured data already subtracted by m and divided by ␴ to X1, X2,
... , Xk, we compute primarily the following x1, x 2, ... , xk:
x1 ⫽ X1

x 2 ⫽ X2 ⫺ b21x1

x 3 ⫽ X3 ⫺ b31x1 ⫺ b32x 2

⯗
xk ⫽ Xk ⫺ bk1x1 ⫺ 䡠䡠䡠 ⫺ bk(k⫺1) (4.182)
Thus, y1, y2, ... , yk are calculated as
y1 ⫽ x1

x2
y2 ⫽
兹V2

x3
y3 ⫽
兹V3

⯗
xk
yk ⫽ (4.183)
兹Vk
4.7. MT and MTS Methods 113

When for an arbitrary object we calculate y1, y2, ... , yk and D 2 in equation
(4.181), if a certain variable belongs to the Mahalanobis space, D 2 is supposed to
take a value of 1, as discussed earlier. Otherwise, D 2 becomes much larger than 1
in most cases.
Once the normalized orthogonal variables y1, y2, ... , yk are computed, the next
step is the selection of items.
A1: Only y1 is used

A2: y1 and y2 are used

⯗
Ak: y1, y2, ... , yk are used
Now suppose that values of signal factor levels are known. Here we do not
explain a case of handling unknown values. Although some signals belong to the
base space, l levels of a signal that is not included in the base space are normally
used. We set the levels to M1, M2, ... , Ml. Although l can only be 3, we should
choose as large an l value as possible to calculate errors.
In the case of Ak, or the case where all items are used, after calculating the
Mahalanobis distances for M1, M2, ... , Ml by taking the square root of each, we
create Table 4.27. What is important here is not D 2 but D per se. As a next step,
we compute dynamic SN ratios. Although we show the case where all items are
used, we can create a table similar to Table 4.27 and calculate the SN ratio for
other cases, such as the case when partial items are assigned to an orthogonal
array.
Based on Table 4.27, we can compute the SN ratio ␩ as follows:
Total variation:
ST ⫽ D 21 ⫹ D 22 ⫹ 䡠䡠䡠 ⫹ D 21 (ƒ ⫽ l ) (4.184)
Variation of proportional terms:
(M1D1 ⫹ M2D2 ⫹ 䡠䡠䡠 ⫹ Ml Dl)2
S␤ ⫽ (ƒ ⫽ 1) (4.185)
r
Effective divider:
r ⫽ M 21 ⫹ M 22 ⫹ 䡠䡠䡠 ⫹ M 2l (4.186)
Error variation:
Se ⫽ ST ⫺ S␤ (ƒ ⫽ l ⫺ 1) (4.187)

Table 4.27
Signal values and Mahalanobis distance
Signal-level value M1 M2 䡠䡠䡠 Mk
Mahalanobis distance D1 D2 䡠䡠䡠 Dl
114 4. Quality Engineering: The Taguchi Method

SN ratio:
(1/r)(S␤ ⫺ Ve)
␩ ⫽ 10 log (4.188)
Ve
On the other hand, for the calibration equation, we calculate
M1D1 ⫹ M2D2 ⫹ 䡠䡠䡠 ⫹ Ml Dl
␤⫽ (4.189)
r
and we estimate
D
M⫽ (4.190)
␤
In addition, for A1, A2, ... , Ak⫺1, we need to compute SN ratios, ␩1, ␩2, ... , ␩k⫺1.
Table 4.28 summarizes the result. According to the table, we determine the num-
ber of items by balancing SN ratio and cost. The cost is not the calculation cost
but the measurement cost for items. We do not explain here the use of loss func-
tions to select items.

Summary of Partial To summarize the distances calculated from each subset of distances to solve col-
MD Groups: linearity problems, this approach can be widely applied. The reason is that we can
Countermeasure for select types of scale to use, as what is important is to select a scale that expresses
Collinearity patients’ conditions accurately no matter what correlation we have in the Mahal-
anobis space. The point is to be consistent with patients’ conditions diagnosed by
doctors.
Now let’s go through a new construction method. For example, suppose that
we have data for 0 and 1 in a 64 ⫻ 64 grid for computer recognition of hand-
writing. If we use 0 and 1 directly, 4096 elements exist. Thus, the Mahalanobis
space formed by a unit set (suppose that we are dealing with data for n persons
in terms of whether a computer can recognize a character ‘‘A’’ as ‘‘A’’: for example,
data from 200 sets in total if 50 people write a character of four items) has a 4096
⫻ 4096 matrix. This takes too long to handle using current computer capability.
How to leave out character information is an issue of information system design.
In fact, a technique for substituting only 128 data items consisting of 64 column
sums and 64 row sums for all data in a 64 ⫻ 64 grid has already been proposed.
Since two sums of 64 column sums and 64 row sums are identical, they have
collinearity. Therefore, we cannot create a 128 ⫻ 128 unit space by using the
relationship 64 ⫹ 64 ⫽ 128.
In another method, we first create a unit space using 64 row sums and introduce
the Mahalanobis distance, then create a unit space using 64 column sums and

Table 4.28
Signal values and Mahalanobis distances
Number of items 1 2 3 䡠䡠䡠 k
SN ratio ␩1 ␩2 ␩3 䡠䡠䡠 ␩k
Cost C1 C2 C3 䡠䡠䡠 Ck
4.7. MT and MTS Methods 115

calculate the Mahalanobis distance. For a few alphabets similar to ‘‘B,’’ we calculate
Mahalanobis distances. If 10 persons, N1. N2, ... , N10 write three letters similar to
‘‘B,’’ we can obtain 30 signals. After each of the 10 persons writes the three letters
‘‘D,’’ ‘‘E,’’ and ‘‘R,’’ we compute the Mahalanobis distances for all signals from a
unit space of ‘‘B.’’ This is shown in Table 4.29.
We calculate a discriminability SN ratio according to this table. Because we do
not know the true difference among M ’s, we compute SN ratios for unknown true
values as follows:
Total variation:
ST ⫽ D 211 ⫹ D 212 ⫹ 䡠䡠䡠 ⫹ D 23,10 (ƒ ⫽ 30) (4.191)
Signal effect:
D 21 ⫹ D 22 ⫹ D 23
SM ⫽ (ƒ ⫽ 3) (4.192)
10
SM
VM ⫽ (4.193)
3
Se ⫽ ST ⫺ SM (ƒ ⫽ 27) (4.194)

Se
Ve ⫽ (4.195)
27
By calculating row sums and column sums separately, we compute the following
SN ratio ␩ using averages of VM and Ve:

M ⫺ Ve)
1
––(V
␩ ⫽ 10 log 10
(4.196)
Ve
The error variance, which is a reciprocal of the antilog value, can be computed as
␴2 ⫽ 10⫺1.1␩ (4.197)
The SN ratio for selection of items is calculated by setting the following two levels:
❏ Level 1: item used
❏ Level 2: item not used
We allocate them to an orthogonal array. By calculating the SN ratio using equa-
tion (4.196), we choose optimal items.

Table 4.29
Mahalanobis distances for signals
Noise
Signal N1 N2 䡠䡠䡠 N10 Total
M1 (D) D11 D12 䡠䡠䡠 D1, 10 D1
M2 (E) D21 D22 䡠䡠䡠 D2, 10 D2
M3 (R) D31 D32 䡠䡠䡠 D3, 10 D3
116 4. Quality Engineering: The Taguchi Method

Although there can be some common items among groups of items, we should
be careful not to have collinearity in one group. Indeed, it seems somewhat un-
usual that common items are included in several different groups, even though
each common item should be emphasized; however, this situation is considered
not so unreasonable because the conventional single correlation has neglected
other, less related items completely.
A key point here is that we judge whether the summarized Mahalanobis dis-
tance corresponds to the case of using the SN ratio calculated by Mahalanobis
distances for items not included in the unit space. Although we can create Ma-
halanobis spaces by any procedure, we can judge them only by using SN ratios.

4.8. On-line Quality Engineering

None of us cares about causes of errors when we set a clock. We sometimes cali-
brate the time by comparing it with a time signal from radio or TV. This is the
case with almost all characteristics in the production process. Let’s say that ac-
cording to company S, the cost of comparing the time with the time signal is
approximately $1.50. In addition, assuming that calibration takes 5 seconds, it will
cost around 10 cents. A quality control system in the production process is re-
sponsible for handling eventual clock errors, given $1.50 for comparison of the
time with the time signal and 10 cents for calibration. If we predict such errors,
we can determine optimal checkup intervals or calibration limits using loss
functions.
A technical procedure for designing optimal checkup intervals and adjustment
limits is called on-line quality engineering or quality control in process. In on-line quality
engineering, we should clarify three system elements used to design a system:
process, checkup procedure (comparison with a standard such as a time signal
and generally called measurement), and correction method (in the case of calibra-
tion of a measuring machine called calibration, and in the case of process called
correction or adjustment). In this section we discuss a quality control system using a
simple example.

❒ Example

Although automobile keys are generally produced as a set of four identical keys,
this production system is regarded as dynamic because each set has a different
dimension and shape. Each key has approximately 9 to 11 cuts, and each cut has
several different dimensions. If there are four different dimensions with a 0.5-mm
step, a key with 10 cuts has 410 ⫽ 1,048,576 variations.
In actuality, each key is produced such that it has a few different-dimension
cuts. Then the number of key types in the market will be approximately 10,000.
Each key set is produced based on the dimensions indicated by a computer. By
using a master key we can check whether a certain key is produced, as indicated
by the computer. Additionally, to manage a machine, we can examine particular-
4.8. Online Quality Engineering 117

shaped keys produced by such machines at certain intervals. Now we consider a

case where we conduct process quality control by measuring differences between
dimensions of cuts of the last key in one set and those shown by a computer.
For the sake of convenience, the standard of key dimensions is Ⳳ30 ␮m for
each target value. Whether or not this Ⳳ30 ␮m tolerance is rational depends on
whether the safety factor of 8 is appropriate for the function limit of Ⳳ250 ␮m.
(In this case, since the height of a cut is designed using a 500-␮m step, if a certain
dimension deviates by 250 ␮m more or less than its design dimension, it can be
taken for other designs. Thus, the function limit is 250 ␮m. In this case, is a safety
factor of 8 appropriate? If we suppose the price of a key set shipped from a plant
to be $1.90, the eventual loss in the market when a key does not work becomes
64 times as great as the price, or $1.20. We consider the tolerance appropriate;
that is, when the malfunction of the key costs $120, Ⳳ30 ␮m makes sense.
A manufacturing department’s task is not to determine tolerances for a product
but to produce a product with dimensions that lie within a range of tolerances. To
achieve this, how should the manufacturing department design a quality control
system? For example, in an inspection room, once an hour we check the dimensions
of a key shortly after it has been machined. When we inspect every n products
produced, we call the number n the inspection interval (also called diagnosis,
checkup, or measurement interval). If we produce 300 sets in an hour, the current
inspection interval N0 is 300 sets. Now the inspection cost B when we pick up a
product in the production process and check it in the inspection room (including
labor cost) amounts to $12. In addition, the loss A for one defective set is already
assumed to be $1.90. On the other hand, the number of key sets produced while
the key selected is inspected for 10 minutes is called the time lag of inspection.
In this case we regard this as 50 sets. At present, we are controlling the production
process within a range of Ⳳ20 ␮m, which is two-thirds of the current tolerance,
⌬ ⫽ 30 ␮m. Then the current adjustment limit D0 is 20 ␮m. Currently, one of
19,560 products goes beyond the adjustment limit of Ⳳ20 (␮m). Since we stop
the production line and take countermeasures such as an adjustment or tool change
when a product exceeds the limit, we need an adjustment cost C of $58. In sum,
we enumerate all preconditions for designing a process control system:
❏ Dimensional tolerance: ⌬ ⫽ 30 ␮m
❏ Loss of defective product: A ⫽ $1.90 per set
❏ Inspection cost: B ⫽ $12
❏ Time lag between inspection and judgment: l ⫽ 50 sets
❏ Adjustment cost: C ⫽ $58
❏ Current inspection interval: n0 ⫽ 300 sets
❏ Current adjustment limit: D0 ⫽ 20 ␮m
❏ Observed mean adjustment interval: u0 ⫽ 19,560 sets
When we inspect a set in an hour and control the production process with the
adjustment limit D0 ⫽ 20 ␮m, the only observation is how many inspections hap-
pen from one to another adjustment. In this case, this value, u0, indicating process
stability, is 19,560.
118 4. Quality Engineering: The Taguchi Method

Design of Feedback Control System

To design all process control systems in an economical manner is the most crucial
task for managers. The cost of the current control system for one set is

B C
control cost ⫽ ⫹ (4.198)
n0 u0

In addition, we need a monetary evaluation of the quality level according to dimen-

sional variability. When the adjustment limit is D0 ⫽ Ⳳ20 ␮m, the dimensions are
regarded as being distributed uniformly within the limit because they are not con-
trolled in the range. The variance is as follows:

D 20
(4.199)
3

On the other hand, the magnitude of dispersion for an inspection interval n0 and
an inspection time lag of l is

冉 n0 ⫹ 1
2
⫹l 冊 D 20
u0
(4.200)

Consequently, we obtain the following loss function:

A
⌬2 冋 冉
D 20
3
⫹
n0 ⫹ 1
2
⫹l 冊 册 D 20
u0
(4.201)

Adding equations (4.198) and (4.201) to calculate the following economic loss L0
for the current control system, we have 33.32 cents per product:

L0 ⫽
B0
n0
⫹
C
u0
A
⫹ 2
⌬ 冋 冉
D 20
3 2冊 册
n0 ⫹ 1
⫹l
D 20
u0

⫽
12
300
⫹
58
19,560
⫹
1.90
302 冋 冉
202
3 冊
⫹
301
2
⫹ 50
202
册
19,560
⫽ 0.04 ⫹ 000.30 ⫹ 0.2815 ⫹ 0.87
⫽ 33.32 cents (4.202)

Assuming that the annual operation time is 1600 hours, we can see that in the
current system the following amount of money is spent annually to control quality:

(33.32)(300)(1600) ⬇ $160,000 (4.203)

If an error ␴m accompanies each measurement, another loss of measurement is

added to the loss above:

A 2
␴ (4.204)
⌬2 m
4.8. Online Quality Engineering 119

It is an optimal control system that improves the loss in equation (4.203). In short,
it is equivalent to determining an optimal inspection interval n and adjustment limit
D, both of which are calculated by the following formulas:

n⫽ 冪2uA B D⌬0

0
(4.205)

⫽冪 冉20冊
(2)(19,560)(12) 30
1.90
⫽ 745
⬇ 600 (twice an hour) (4.206)

D⫽ 冉3C
A
⫻
D 2
u0
⌬
2
0
冊1/2

(4.207)

⫽ 冋(3)(58)
1.90 冉 202
19,560
⌬2冊册 1/4

⫽ 6.4 (4.208)
⬇ 7.0 ␮m (4.209)

Then, by setting the optimal measurement interval to 600 sets and adjustment limit
to Ⳳ7.0 ␮m, we can reduce the loss L:

L⫽
B
n
C A
⫹ ⫹ 2
u ⌬ 冋 冉
D2
32
⫹
n⫹1
2
⫹l 冊 册 D2
u
(4.210)

Using this equation, we estimate the mean adjustment interval u:

D2
u ⫽ u0
D 20

⫽ (19,560) 冉 冊 72
202
⫽ 2396 (4.211)

Therefore, we need to change the adjustment interval from the current level of 65
hours to 8 hours. However, the total loss L decreases as follows:

L⫽
12
600
⫹
58
2396
⫹
1.80
302 冋 冉 72
3
⫹
601
2
⫹ 50 冊 册 72
2396
⫽ 0.02 ⫹ 0.0242 ⫹ 3.45 ⫹ 1.51
⫽ 9.38 cents (4.212)

This is reduced from the current level by

33.32 ⫺ 9.38 ⫽ 23.94 cents (4.213)

120 4. Quality Engineering: The Taguchi Method

On a yearly basis, we can expect an improvement of

(23.94)(300)(1600) ⫽ $114,912 (4.214)

Management in Manufacturing
The procedure described in the preceding section is a technique of solving a bal-
ancing equation of checkup and adjustment costs, and necessary quality level, lead-
ing finally to the optimal allocation of operators in a production plant. Now, given
that checkup and adjustment take 10 and 30 minutes, respectively, the work hours
required in an 8-hour shift at present are:

(10 min) ⫻ (daily no. inspections) ⫹ (30 min) ⫻ (daily no. adjustments)

(8)(300) (3)(300)
⫽ (10) ⫹ (30) (4.215)
n u

⫽ (10) 冉 冊 2400
600
⫹ (30) 冉 冊
2400
2396
⫽ 40 ⫹ 30.1 ⫽ 70 minutes (4.216)

Assuming that one shift has a duration of 8 hours, the following the number of
workers are required:

70
⫽ 0.146 worker (4.217)
(8)(60)

If there are at most six processes,

(0.146)(6) ⫽ 0.88 worker (4.218)

is needed, which implies that one operator is sufficient. As compared to the follow-
ing workers required currently, we have

(8)(300) (3)(300)
(10) ⫹ (30) ⫽ 83.6 minutes (4.219)
300 19,560

so we can reduce the number of workers by only 0.028 in this process. On the
other hand, to compute the process capability index Cp, we estimate the current
standard deviation ␴0 using the standard deviation of a measurement error ␴m:

冪D3 ⫹ 冉n 2⫹ 1 ⫹ l冊 Du
2 2
␴0 ⫽ ⫹ ␴2m
0 0 0
(4.220)
0

Now setting ␴m ⫽ 2 ␮m, we obtain ␴0:

␴0 ⫽ 冪3
202
⫹ 冉301
2
⫹ 50 冊 202
19560
⫹ 22

⫽ 11.9 ␮m (4.221)
4.8. Online Quality Engineering 121

Thus, the current process capability index Cp is computed as follows:

2⌬
Cp ⫽
6␴0

(2)(30)
⫽
(6)(11.9)
⫽ 0.84 (4.222)
The standard deviation in the optimal feedback system is

冪 冉 冊
72 601 72
␴⫽ ⫹ ⫹ 50 ⫹ 22
3 2 2396
⫽ 5.24 ␮m (4.223)
Then we cannot only reduce the required workforce by 0.028 worker (0.84 times)
but can also enhance the process capability index Cp from the current level to
(30)(2)
Cp ⫽ ⫽ 1.91 (4.224)
(6)(5.24)

Manufacturing Strategy: Balance of Production Measurements

and Quality Control System
The most essential strategy planned by manufacturing managers is not to balance
quality and cost but to improve productivity. The difference in profitability between
corporations is primarily the difference in production speed. Quite a few managers
insist that we should continue to produce only a certain volume because even if
the production speed is doubled and the production volume is doubled, the in-
creased volume tends to remain unsold in most cases. Nevertheless, we believe
that a key point is to develop a technology to double the production speed to prepare
for such a demand. As an extreme idea, after doubling the production speed, we
can stop after half a day. In this case, an R&D department could offer new jobs to
idle workers, or after lowering the retail price by two-thirds, could double sales.
Unless a company sells a product at a retail price several times as high as a
plant price (UMC or production cost), it cannot survive. This is because labor and
running expenses in sales, administration, and R&D departments account for a few
percent of its total cost. In other words, the retail price needs to include the total
running cost and profit. When a product is sold at a price three times as high as a
plant price, if the production volume is doubled, the plant price, excluding material
cost, will be cut in half. Therefore, if the material cost accounts for 30% of the
plant price, in cases of double production speed, the retail price can be reduced to
an amount 1.65 times as great as the plant cost:
[0.3 ⫹ –12 (0.7)] ⫹ –12 (2) ⫽ 1.65 times (4.225)
As compared to 3 times, 1.65 times implies that we can lower the retail price
0.55-fold. This estimation is grounded on the assumption that if we sell a product
122 4. Quality Engineering: The Taguchi Method

at 45% off the original price, the sales volume doubles. This holds true when we
offer new jobs to half the workers, with the sales volume remaining at the same
level. Now, provided that the mean adjustment interval decreases by one-fourth
due to increased variability after the production speed is raised, how does the cost
eventually change? When frequency of machine breakdowns quadruples, u de-
creases from its current level of 2396 to 599, one-fourth of 2396. Therefore, we
alter the current levels of u0 and D0 to the values u0 ⫽ 599 and D0 ⫽ 7 ␮m. By
taking these into account, we consider the loss function. The cost is cut by 0.6-
fold and the production volume is doubled. In general, as production conditions
change, the optimal inspection interval and adjustment limit also change. Thus, we
need to recalculate n and D here. Substituting 0.6A for A in the formula, we obtain

n⫽ 冪2u0.6AB D⌬
0

0.6 ⫻ 190 冉 7 冊
⫽冪
(2)(599)(12) 30

⫽ 481 → 600 (once in an hour) (4.226)

D⫽ 冋 (3)(58)
(0.6)(1.90) 冉 冊 册72
599
(302)
1/4

⫽ 10.3 → 10.00 (4.227)

On the other hand, the mean adjustment interval is as follows:

u ⫽ (599) 冉 冊
102
72
⫽ 1222 (4.228)

Hence, the loss function is computed as follows:

L⫽
B
n
C
⫹ ⫹
u
0.6A
302 冋 冉
D2
3
⫹
n⫹1
2 冊 册
⫹ 2l
D2
u

⫽
12
600
⫹
58
1222
⫹
(0.6)(1.90)
302 冋 冉
102
3
⫹
601
2
⫹ 100 冊 册
102
1222
⫽ 0.02 ⫹ 0.0475 ⫹ 4.22 ⫹ 4.15
⫽ 15.12 cents (4.229)

As a consequence, the quality control cost increases by 5.74 cents (⫽ 15.12 ⫺

9.38). However, the total loss including cost is

$1.90 ⫹ 9.38 cents ⫽ $1.99 (4.230)

(0.6)(190 cents) ⫹ 15.12 cents ⫽ $1.29 (4.231)
References 123

Suppose that the production volume remains the same, (300)(1600) ⫽ 480,000
sets. We can save the following amount of money on an annual basis:
(1.99 ⫺ 1.29)(48) ⫽ $33,600,000 (4.232)

References
1. Genichi Taguchi, 1987. System of Experimental Design. Dearborn, Michigan: Unipub/
American Supplier Institute.
2. Genichi Taguchi, 1984. Reliability Design Case Studies for New Product Development.
Tokyo: Japanese Standards Association.
3. Genichi Taguchi et al., 1992. Technology Development for Electronic and Electric Industries.
Quality Engineering Application Series. Tokyo: Japanese Standards Association.
4. Measurement Management Simplification Study Committee, 1984, Parameter Design
for New Product Development. Japanese Standards Association.
5. Genichi Taguchi et al., 1989. Quality Engineering Series, Vol. 2. Tokyo: Japanese Stan-
dards Association.
6. Tatsuji Kenetaka, 1987. An application of Mahalanobis distance. Standardization and
Quality Control, Vol. 40, No. 10.