Professional Documents
Culture Documents
142
Teaching the Taguchi method to industrial engineers Work Study
Jiju Antony and Frenie Jiju Antony Volume 50 . Number 4 . 2001 . 141±149
Step1: formulation of the problem – the the analysis were discussed in the classroom
success of any experiment is dependent as part of the process of gaining an
on a full understanding of the nature of understanding of experimental objectives and
the problem. process.
Step 2: identification of the output The paper helicopter experiment is quite
performance characteristics most relevant well known among engineers and statisticians
to the problem. in both the academic and industrial worlds.
Step 3: identification of control factors, Many industrial training programmes on Tm
noise factors and signal factors (if any). use it in some form. However they often focus
Control factors are those which can be on the design and analysis of the experiment
controlled under normal production without providing guidance to engineers on
conditions. Noise factors are those which the interpretation of results from the analysis.
are either too difficult or too expensive to Moreover, many courses do not cover the
control under normal production importance of careful experimental planning
conditions. Signal factors are those which for the success of any industrially designed
affect the mean performance of the experiment.
process. The purpose of this experiment was to
Step 4: selection of factor levels, possible provide undergraduate engineering students
interactions and the degrees of freedom with an understanding of the role of
associated with each factor and the Taguchi’s ‘‘parameter design’’ (sometimes
interaction effects. called ‘‘robust design’’) in tackling both
Step 5: design of an appropriate product and process quality-related problems
orthogonal array (OA). in real-life situations. Parameter design is a
Step 6: preparation of the experiment. well established methodology for improving
Step 7: running of the experiment with product and process quality at minimal cost
appropriate data collection. by reducing the effect of undesirable external
Step 8: statistical analysis and influences which cause variation in product or
interpretation of experimental results. process performance (Phadke, 1989).
Step 9: undertaking a confirmatory run of The objective of the exercise was to identify
the experiment. the optimal settings of control factors which
would maximise the flight time of paper
helicopters (with minimum variation). Here
Paper helicopter experiment control factors refer to those which can be
easily controlled and varied by the designer or
In many academic institutions within the UK, operator in normal production conditions. A
the focus of engineering statistics is on the brainstorming session by a group of students
theory of probability (for example, card identified six control factors which were
shuffling, dice rolling, etc.), the mathematical thought to influence the time of flight (refer to
aspects of probability and probability Table II). Brainstorming should be
distributions (e.g. normal, exponential, considered an integral part of the Taguchi
binomial, Poisson, log-normal, etc.), methodology – it is a useful technique in
hypothesis tests, etc. Quality improvement identifying the most influential factors in an
techniques (DOE, Tm, SPC, etc.) are often experiment.
not covered. Understandably, graduates are In order to simplify the experiment, each
not confident about using such techniques at factor was studied at two levels. The ‘‘level’’
their place of work. of a factor here refers to the specified value of
As part of an exercise to increase the
awareness of Tm amongst industrial Table II Control factors and their range of settings for the experiment
engineers, the authors used a simple paper Control factor Labels Level 1 Level 2
helicopter experiment, readily used in
Paper type A Regular Bond
academic institutions. Due to a limited
Body length B 8cm 12cm
amount of time, one member from each
Wing length C 8cm 12cm
group in the class was involved with the
Body width D 2cm 3cm
experimental work. However, the students
Number of clips E 1 2
were all asked to analyse and interpret the
Wing shape F Flat Angled
data (on an individual basis). The results of
143
Teaching the Taguchi method to industrial engineers Work Study
Jiju Antony and Frenie Jiju Antony Volume 50 . Number 4 . 2001 . 141±149
a setting. For example, in the experiment, Figure 1 Template for paper helicopter design
body width was studied at 2cm and 3cm.
Factors at three (and higher) levels make
analysis more complicated – and are therefore
not used in awareness-raising sessions.
Having identified the control factors, it is
important to list the interactions which are to
be studied for the experiment. Interaction
exists when the effect of one factor is not the
same at different levels of the other factor. An
effect refers to the change in response due to
the change in level of a factor (Antony et al.,
1998b). Consider, for example, the factors
wing length and body length of the paper
helicopter. Assume each factor was kept at
two-levels for the study. Time of flight is the
response (or quality characteristic) of interest.
itself taking about 90 minutes). In this case,
Interaction between wing length and body
the statistical analysis was executed as a
length exists when the effect of wing length on
homework assignment though the results
time of flight at two different levels of body
were discussed in the classroom in detail.
length is different.
For this experiment, three interactions were
identified (from the brainstorming session) as
being of interest: Choice of orthogonal array design
(1) body length £ wing length (B £ C or
BC); The choice of a suitable orthogonal array
(2) body length £ body width (B £ D or (OA) design is critical for the success of an
BD); and experiment and depends on the total degrees
(3) paper type £ body length (A £ B or AB). of freedom required to study the main and
interaction effects, the goal of the experiment,
The following noise factors were identified (as resources and budget available and time
having some impact on the flight time but constraints. Orthogonal arrays allow one to
being difficult to control): compute the main and interaction effects via a
operator-to-operator variation; minimum number of experimental trials
draughts; (Ross, 1988). ‘‘Degrees of freedom’’ refers to
reaction time; and the number of fair and independent
ground surface. comparisons that can be made from a set of
One aim was to determine the control factor observations. In the context of SDOE, the
settings which would best dampen the effect number of degrees of freedom is one less than
of these noise factors. According to Taguchi, the number of levels associated with the
there is an optimal combination of factor factor. In other words, the number of degrees
settings which counters the effects of noise. In of freedom associated with a factor at p-levels
order to minimise the effect of these noise is (p-1). As the number of degrees of freedom
factors, the same student was responsible for associated with a factor at two levels is unity,
all timings – reducing the effects of variable in the present example the number of degrees
reaction times when hitting the stopwatch of freedom for studying the six main effects is
upon release of the helicopter and its hitting equal to six. The number of degrees of
the ground. freedom associated with an interaction is the
Figure 1 illustrates a template for the model product of the number of degrees of freedom
of a paper helicopter which can be made from associated with each main effect involved in
an A4 size paper. It forms the basis of a simple the interaction (Antony, 1998). In this simple
experiment, requiring only simple items such case, the number of degrees of freedom for
as paper, scissors and paper clips. It takes studying the three interaction effects is equal
about six hours to design the experiment, to three. Therefore the total degrees of
collect the data and then perform the freedom is equal to nine (i.e. 6 + 3). It is
statistical analysis (with the ‘‘experiment’’ important to notice that the number of
144
Teaching the Taguchi method to industrial engineers Work Study
Jiju Antony and Frenie Jiju Antony Volume 50 . Number 4 . 2001 . 141±149
experimental trials must be greater than the Column 1 – body width (D), column 2 –
total degrees of freedom required for studying wing length (C), column 4 – body length (B),
the effects. The standard OAs for factors with column 5 – body width £ body length (B £
two levels are L4 L8 , L1 6 , L32 and so on. Here D), column 6 – wing length £ body length (B
the notation ‘‘L’’ implies that the information £ C), column 7 – wing shape (F), column 8 –
is based on the Latin square arrangement of paper type (A), column 12 – body length £
factors. A Latin square arrangement is a paper type (AB) and column 14 – number of
square matrix arrangement of factors with clips (E).
separable factor effects. Here the numbers 4, The experimental layout showing all the
8, 12, 16, etc. denote the number of factors and interactions along with the flight
experimental trials. For the helicopter times (measured in seconds) is shown in
experiment, as the total degrees of freedom is Table III. As each factor was studied at two
equal to nine, the closest number of levels, coded level 1 represents the low level of
experimental trials that can be employed for a factor setting and level 2 represents the high
the experiment is 16 (i.e. L1 6 OA). Having level setting. Each experiment was replicated
identified the most suitable OA, the next step in order to capture variation in results due to
was to assign the main and interaction effects uncontrolled noise.
to various columns of the array. A standard
L16 OA (see Appendix) contains 15 columns
for either studying 15 main effects, or a Statistical analysis and interpretation of
combination of main and interaction effects results
so that the degrees of freedom will add up to
15. In the present example, there are only six In Taguchi’s parameter design, the basic
main and three interaction effects. This objective is to identify the conditions which
means that only nine columns out of 15 are optimise process/product performance. In
used. For example, factor D (refer to Table arriving at this optimal set of conditions,
III) was assigned to column 1 and factor C to Taguchi advocates the use of signal-to-noise
column 2. Column 3 is empty (see Table III) ratio (SNR) – the need is to maximise the
as the interaction between these factors was of performance of a system or product by
no interest in this experiment. Using the minimising the effect of noise while
standard linear graphs and OA (Ross, 1988), maximising the mean performance. The SNR
the remaining factors and interactions were is treated as a response (output) of the
assigned to the columns of an L1 6 in the experiment, which is a measure of variation
following manner. when uncontrolled noise factors are present in
145
Teaching the Taguchi method to industrial engineers Work Study
Jiju Antony and Frenie Jiju Antony Volume 50 . Number 4 . 2001 . 141±149
the system (Antony et al., 1999). Taguchi has The other main and interaction effects were
developed and defined over 60 different calculated in a similar manner (see Table V).
SNRs for engineering applications of Having obtained the average SNR values,
parameter design. For the present study, as the next step is the identification of significant
the objective was to maximise time of flight it main and interaction effects which influence
was decided to select the SNR related to the SNR. To achieve this, a powerful
larger-the-better (LTB) quality graphical tool called half-normal probability
characteristics. This is generally used for plots (HNPP) is useful.
quality characteristics such as strength, fuel A half-normal probability plot (HNPP) is
efficiency, process yield, life of a component obtained by plotting the absolute values of the
and so on. For LTB quality characteristics, effects (both main and/or interaction effects)
the SNR is given by the following equation: along the X-axis and the per cent probability
£1 1¤ along the Y-axis. The per cent probability
SNR ˆ ¡10log ¤ …1 †
n y2i can be obtained by using the following
equation:
where n = number of values at each trial …i ¡ 0:5†
condition (i.e., 2 from Table II) and yi = each Pi ˆ £ 100 …2 †
n
observed value.
Table IV illustrates the SNR values (based where: n = number of estimated effects
on equation 1) corresponding to each trial (n = 15) and i is the rank of the estimated
condition. effect when arranged in the ascending order of
Table V illustrates the average SNR values magnitude (e.g. for factor C, i = 15).
(SNR) at low (level 1) and high (level 2) levels Figure 2 illustrates the HNPP of the factor
and the effect of each main and interaction and interaction effects for the helicopter
effect on the SNR. experiment. The computer software package
Sample calculation for factor ‘‘C’’: ‘‘Design-ease’’ was used to construct the plot.
Average SNR at level 1 of factor ‘‘C’’ = Those effects which are active and real will
SNRC 2 = 1/8 * [8.93 + 6.71 + 6.41 + 6.62 fall off the straight line, whereas the inactive
+7.12 + 5.95 + 6.89 + 6.38] and insignificant effects will fall along the
= 6.88 straight line (Daniel, 1959). The figure
Similarly, average SNR at level 2 of factor reveals that main effects A, C, E and F are
‘‘C’’ = SNRC 2 = 1/8 * [7.78 + 8.05 + 10.06 + statistically significant. i.e. paper type, wing
7.95 + 9.01 + 8.07 + 8.80 + 7.47] length, number of clips and wing shape are
= 8.40 statistically significant. In order to support
Effect = SNRC 2 - SNRC 1 and justify this claim, another graphical tool
= 8.40 - 6.88 = 1.52 (main effects plot) is used. This shows the
average SNR values at low and high level
settings of each factor. Figure 3 illustrates the
Table IV SNR table main effects plot for the paper helicopter
Trial number SNR Trial number SNR experiment (using the values from Table V).
This graphical aid provides non-statisticians
1 8.93 9 7.12 with a better picture of the importance of the
2 6.71 10 5.95 effects of the chosen control factors. The
3 6.41 11 6.89 slope of the line is an indication of the
4 6.62 12 6.38 importance of a main or interaction effect.
5 7.78 13 9.01 The figure shows that the most dominant
6 8.05 14 8.07 factor is the wing length, followed by paper
7 10.06 15 8.80 type, wing shape and number of clips. As each
8 7.95 16 7.47 factor was chosen at two levels, the effect of
146
Teaching the Taguchi method to industrial engineers Work Study
Jiju Antony and Frenie Jiju Antony Volume 50 . Number 4 . 2001 . 141±149
Gunst, R.F. and Mason, R.L. (1991), How to Construct Ross, P.J., (1988), Taguchi Techniques for Quality
Fractional Factorial Experiments, ASQC Statistics Engineering, McGraw-Hill Publishers, New York, NY.
Division, ASQC Press, Milwaukee, MI. Rowlands, H., Antony, J. and Knowles, G. (2000), ``An
Lucas, J.M. (1992), ``Split plotting and randomisation in application of experimental design for process
industrial experiments’’, ASQC Quality Congress optimisation’’, The TQM Magazine, Vol. 12 No.2,
Transactions, Nashville, TN, pp. 374-82. pp. 78-83.
Morrison, J.M. (1997), ``Statistical engineering ± the key Schmidt, S.R and Launsby, R.G. (1992), Understanding
to quality’’, Engineering Science and Education Industrial Designed Experiments, Air Academy
Journal, pp. 123-7. Press, Washington, DC.
Phadke, M.S. (1989), Quality Engineering using Robust Taguchi, G. (1986), Introduction to Quality
Design, Prentice-Hall International, Englewood Engineering, Asian Productivity Organisation,
Cliffs, NJ. Tokyo.
Appendix
Table AI Coded design matrix of an L16 (21 5 ) orthogonal array
Column
Trial 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15
1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1
2 1 1 1 1 1 1 1 2 2 2 2 2 2 2 2
3 1 1 1 2 2 2 2 1 1 1 1 2 2 2 2
4 1 1 1 2 2 2 2 2 2 2 2 1 1 1 1
5 1 2 2 1 1 2 2 1 1 2 2 1 1 2 2
6 1 2 2 1 1 2 2 2 2 1 1 2 2 1 1
7 1 2 2 2 2 1 1 1 1 2 2 2 2 1 1
8 1 2 2 2 2 1 1 2 2 1 1 1 1 2 2
9 2 1 2 1 2 1 2 1 2 1 2 1 2 1 2
10 2 1 2 1 2 1 2 2 1 2 1 2 1 2 1
11 2 1 2 2 1 2 1 1 2 1 2 2 1 2 1
12 2 1 2 2 1 2 1 2 1 2 1 1 2 1 2
13 2 2 1 1 2 2 1 1 2 2 1 1 2 2 1
14 2 2 1 1 2 2 1 2 1 1 2 2 1 2 1
15 2 2 1 2 1 1 2 1 2 2 1 2 1 1 2
16 2 2 1 2 1 1 2 2 1 1 2 1 2 2 1
149