Ie466.robust.handout

IE 466: Concurrent Engineering T. W. Simpson

32.3 Taguchi’ Robust Design Method
s
Since 1960, Taguchi methods have been used for improving the quality of Japanese
products with great success. During the 1980’ many companies finally realized that the old
s,
methods for ensuring quality were not competitive with the Japanese methods. The old methods
for quality assurance relied heavily upon inspecting products as they rolled off the production
line and rejecting those products that did not fall within a certain acceptance range. However,
Taguchi was quick to point out that no amount of inspection can improve a product; quality must
be designed into a product from the start. It is only recently that companies in the United States
and Europe began adopting Taguchi’ robust design approaches in an effort to improve product
s
quality and design robustness.
What is robust design? Robust design is an “engineering methodology for improving
productivity during research and development so that high-quality products can be produced
quickly and at low cost” (Phadke, 1989). The idea behind robust design is to improve the quality
of a product by minimizing the effects of variation without eliminating the causes (since they are
too difficult or too expensive to conrol). His method is an off-line quality control method that is
instituted at both the product and process design stage to improve product manufacturability and
reliability by making products insensitive to environmental conditions and component variations.
The end result is a robust design, a design that has minimum sensitivity to variations in
uncontrollable factors.
Dr. Genichi Taguchi bases his method on conventional statistical tools together with
some guidelines for laying out design experiments and analyzing the results of these
experiments. Taguchi's approach to quality control applies to the entire process of developing
and manufacturing a product— from the initial concept, through design and engineering, to
manufacturing and production. Taguchi methods are used to specify dimension and feature
detail and normally follow DFM activities. In the next section we discuss Taguchi’ concept of a
s
quality loss function. This is then followed by a detailed description of Taguchi’ approach to
s
parameter design.

32.3.1 Taguchi’ Quality Loss Function
s
Consider the popular story of Sony and two of their television production facilities, one
in the USA, the other in Japan. The color density of the televisions manufactured by Sony-USA
were uniformly distributed and fell within the tolerance limits, m ± 5 ( where m is the target for
color density), while the televisions from Sony-Japan followed a normal distribution, more
televisions were on target but about 0.3% fell outside of the tolerance limits. The color density
distributions are illustrated in Figure 1. The differences in customer perceptions of quality
resulted from Sony-USA paying attention only to meeting the tolerances whereas in Sony-Japan,
the focus was on meeting the target and minimizing the variance around that target. If we assign
a grade to each television based on its color density as done in Figure 12, then we see that Sony-
Japan produced many more Grade A sets and fewer Grade C sets in comparison to Sony-USA.
Overall, a much larger portion of Sony-Japan's televisions receive higher grades then those made
by Sony-USA; hence, the customer’ preferred the televisions sets produced by Sony-Japan over
s
those produced by Sony-USA.

1


Figure 1 Color Density Distribution of Television Sets (Phadke, 1989)

To measure quality, Taguchi defines a Quality Loss Function. The quality loss function
is a continuous function that is defined in terms of the deviation of a design parameter from an
ideal or target value, see Figure 2b. Taguchi’ view on the nature of the quality loss function
s
represents a fundamental paradigm shift (particularly in the U.S.) in the way in which
manufacturers consider whether or not a product is good. The traditional approach employed by
U.S. manufacturers (as evidenced by Sony-USA) has been to use a “step function” that ensures
that performance fell within the upper and lower specification limits as shown in Figure 2a.

Figure 2 Quality Loss Function (Phadke, 1989)

Taguchi’ loss function can be expressed in terms of the quadratic relationship:
s
L = k (y - m)2 [32.1]
where y is the critical performance parameter value, L is the loss associated with a particular
parameter y, m is the nominal value of the parameter specification, k is a constant that depends
on the cost at the specification limits (can be determined conservatively by dividing the cost of
scrap in $, by the square of the lower or higher tolerance values). This function penalizes the
deviation of a parameter from the specification value that contributes to deteriorating the
performance of the product, resulting in a loss to the customer. The key here is that a product
engineer has a good understanding of what the nominal size of the specification is. The usual
lower and upper limits for the tolerance of a given design parameter are changed to a continuous
function that presents any parameter value other than the nominal as a loss. The loss function
given in Eq.32.1 is referred to as “nominal is best,” but there are also expressions for cases when
higher or lower values of parameters are better (Phadke, 1989).

2


If a large number of parts are considered, say N, the average loss per part is equal to the
summation of the losses given by Eq. 32.1 for each part, divided by the total N. The average
quality loss results from deviation around the average value of y from the target and the mean
square deviation of y around its own mean. The average quality loss can be expressed as:
L = k[S2 + (µ - m)2] [32.2]
where µ is the average value of y for the set of parts, and S is the variance around the average.
2

To minimize loss, the traditional approach is to monitor the process variables during
production and adjust the process to reduce manufacturing imperfections so that response
parameters fall within the specified tolerances. This method adds cost to the manufacturing
process does not improve the quality of the product. Using Taguchi’ approach the average
s
response has to be adjusted, and the variance must be reduced in order to minimize loss.
Reducing the variation is accomplished by product and process engineers who use off-line
quality control techniques; adjustments to the average response are realized by process and
production engineers during production using on-line quality control techniques. Within the
Taguchi philosophy both quality improvement methods are considered; however, building
quality into the product during the design stage (i.e., off-line) is the ultimate goal.
To achieve desirable product quality by design, Taguchi suggests a three-stage process:
system design, parameter design, and tolerance design. System design is the
conceptualization and synthesis of a product or process to be used. The system design stage is
where new ideas, concepts and knowledge in the areas of science and technology are utilized by
the design team to determine the right combination of materials, parts, processes and design
factors that will satisfy functional and economical specifications. To achieve an increase in
quality at this level requires innovation, and therefore improvements are not always made. In
parameter design the system variables are experimentally analyzed to determine how the product
or process reacts to uncontrollable “noise” in the system; parameter design is the main thrust of
Taguchi’ approach. Parameter design is related to finding the appropriate design factor levels
s
to make the system less sensitive to variations in uncontrollable noise factors, i.e., to make the
system robust. In this way the product performs better, reducing the loss to the customer.
The final step in Taguchi’ robust design approach is tolerance design; tolerance design
s
occurs when the tolerances for the products or process are established to minimize the sum of
the manufacturing and lifetime costs of the product or process. In the tolerance design stage,
tolerances of factors that have the largest influence on variation are adjusted only if after the
parameter design stage, the target values of quality have not yet been achieved. Most engineers
tend to associate quality with better tolerances, but tightening the tolerances increases the cost of
the product or process because it requires better materials, components, or machinery to achieve
the tighter tolerances as we discussed in earlier chapters. Taguchi’ parameter design approach
s
allows for improving the quality without requiring better materials or parts and makes it possible
to improve quality and decrease (or at least maintain the same) cost. Parameter design is
discussed in detail in the next section.

32.3 Taguchi’ Parameter Design Approach
s
In parameter design, there are two types of factors that affect a product’ functional
s
characteristic: control factors and noise factors. Control factors are those factors which can
easily be controlled such as material choice, cycle time, or mold temperature in an injection
molding process. Noise factors are factors that are difficult or impossible or too expensive to
control. There are three types of noise factors: outer noise, inner noise, and between product

3


noise. Examples of each type of noise factor and controllable factors in product and process
design are listed in Table 1. Noise factors are primarily response for causing a product’ s
performance to deviate from its target value. Hence, parameter design seeks to identify settings
of the control factors which make the product insensitive to variations in the noise factors, i.e.,
make the product more robust, without actually eliminating the causes of variation.

Product Design Process Design
Outer Noise Consumer’ usage conditions
s Ambient Temperature
Low temperature Humidity
High temperature Seasons
Temperature change Incoming material variation
Shock Operators
Vibration Voltage change
Humidity Batch to batch variation
Inner Noise Deterioration of parts Machinery aging
Deterioration of material Tool wear
Oxidation (rust) Deterioration
Between Piece to piece variation where they are Process to process variation where they are
Product supposed to be the same, e.g., supposed to be the same, e.g.,
Noise Young’ modulus
s variations in feed rate
shear modulus
allowable stress
Controllable Factors All design parameters, e.g., All process design parameters
• dimensions All process setting parameters
• material selection
Table 1 Examples of Noise and Control Factors (adapted from Byrne and Taguchi, 1987)

Design of experiments techniques, specifically Orthogonal Arrays (OAs), are employed
in Taguchi’ approach to systematically vary and test the different levels of each of the control
s
factors. Commonly used OAs include the L4, L9, L12, L18, and L27, several of which are listed in
Table 2. A complete listing of OAs can be found in text such as (Phadke, 1989). The columns
in the OA indicate the factor and its corresponding levels, and each row in the OA constitutes an
experimental run which is performed at the given factor settings. For instance, in Fig. 2
experimental run #3 has Factor 1 at Level 2, Factor 2 at Level 1, and Factor 3 at Level 2. It is up
to the experimental designer to establish the appropriate factor levels for each control factor;
typically either 2 or 3 levels are chosen for each factor. Selecting the number of levels and
quantities properly constitutes the bulk of the effort in planning robust design experiments.

Factors Factors Factors
Run A B C Run A B C D E F G Run A B C D
1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1
2 1 2 2 2 1 1 1 2 2 2 2 2 1 2 2 2
3 2 1 2 3 1 2 2 1 1 2 2 3 1 3 3 3
4 2 2 1 4 1 2 2 2 2 1 1 4 2 1 2 3
5 2 1 2 1 2 1 2 5 2 2 3 1
6 2 1 2 2 1 2 1 6 2 3 1 2
7 2 2 1 1 2 2 1 7 3 1 3 2
8 2 2 1 2 1 1 2 8 3 2 1 3
9 3 3 2 1
(a) L4 (23) array (b) L8 (27) array (c) L9 (34) array
Table 2 Some Commonly Used Orthogonal Arrays

4


To implement robust design, Taguchi advocates the use of an “inner array” and “outer
array” approach. The “inner array” consists of the OA that contains the control factor settings;
the “outer array” consists of the OA that contains the noise factors and their settings which are
under investigation. The combination of the “inner array” and “outer array” constitutes what is
called the “product array” or “complete parameter design layout.” Examples of each of these
arrays are given in the case study in the next section. The product array is used to systematically
test various combinations of the control factor settings over all combinations of noise factors
after which the mean response and standard deviation may be approximated for each run using
the following equations.
1 n
• Mean response: y = ∑ yi [32.3]
n i=1
n (y i − y )2
• Standard deviation: S= ∑ n− 1
[32.4]
i =1
The preferred parameter settings are then determined through analysis of the “signal-to-noise”
(SN) ratio where factor levels that maximize the appropriate SN ratio are optimal. There are
three standard types of SN ratios depending on the desired performance response (Phadke,
1989):
• Smaller the better (for making the system response as small as possible):
1 n 
SN S = − 10 log ∑ y i2 
n  [32.5]
 i=1 
• Nominal the best (for reducing variability around a target):
y2 
SN T = 10 log 2  [32.6]
S 
 
• Larger the better (for making the system response as large as possible):
1 n 1 
SN L = − 10 log ∑ 2  [32.7]
n y 
 i=1 i 
These SN ratios are derived from the quadratic loss function and are expressed in a decibel scale.
Once all of the SN ratios have been computed for each run of an experiment, Taguchi
advocates a graphical approach to analyze the data. In the graphical approach, the SN ratios and
average responses are plotted for each factor against each of its levels. The graphs are then
examined to “pick the winner,” i.e., pick the factor level which (1) best maximize SN and (2)
bring the mean on target (or maximize or minimize the mean, as the case may be). Using this
information, the control factors can also be grouped as follows.
1. Factors that affect both the variation and the average performance of the product.
2. Factors that affect the variation only.
3. Factors that affect the average only.
4. Factors that do not affect either the variance or the average.
Factors in the first and second groups can be utilized to reduce the variations in the system,
making it more robust. Factors in the third group are then used to adjust the average to the target
value. Lastly, factors in the fourth group are set to the most economical level. Finally,

5


confirmation tests should be run at the “optimal” product settings to verify that the predicted
performance is actually realized. A demonstration of Taguchi’ approach to parameter design
s
serves as our case study in the next section.

32.4 Case Study: Parameter Design of an Elastometric Connector
The following case study is taken from “The Taguchi Approach to Parameter Design,” by
D. M. Byrne and S. Taguchi, Quality Progress, Dec. 1987, pp. 19-26. The case uses Taguchi’ s
parameter design approach to integrate product and process design decisions for elastometric
connector used in an automotive engine application.

32.4.1 The Problem
The experiment that is being conducted seeks to determine a method to assemble an
elastometric connector to a nylon tube while delivering the requisite pull-off performance
suitable for an automotive engineering application. The primary design objective is to maximize
the pull-off force while secondary considerations are made to minimize assembly effort and
reduce the cost of the connector and assembly.
Four control factors and three noise factors have been identified for the connector and
tube assembly. The control factors consist of the (A) interference, (B) connector wall thickness,
(C) insertion depth, and (D) percent adhesive in connector pre-dip; a sketch of the control factors
for the connector and tube is given in Figure 3.

Figure 3 Control Factors for Connector and Nylon Tube Experiment
(Byrne and Taguchi, 1987)

The noise factors in the experiment are (E) conditioning time, (F) conditioning
temperature, and (G) conditioning relative humidity. Each control factor is to be tested at three
levels while each noise factor is tested at two levels. The factors and levels of concern for the
experiment are listed in Table 3. In terms of product and process design, Factors A and B
represent product design parameters while Factors C-G represent process design parameters.
However, during routine operation the noise factors (E-G) are uncontrollable are thus taken as
“noise” which can adversely affect product performance. Fortunately, these noise factors can
able controlled for the purposes of this experiment. In this regard, Taguchi’ parameter design
s
approach can be used to help make product and process design decisions to improve the
robustness of a system.

6


Controllable Factors Levels
A. Interference Low Medium High
B. Wall thickness Thin Medium Thick
C. Insertion depth Shallow Medium Deep
D. Percent adhesive Low Medium High
Noise Factors Levels
E. Conditioning time 24 h 120 h
F. Conditioning temperature 72ºF 150ºF
G. Conditioning relative humidity 25% 75%
Table 3 Factors and Levels for Connector and Tube

32.4.2 The Experiment
Following Taguchi’ method, two experimental designs are selected to vary (i) the
s
control factors and (ii) the noise factors. An L9 orthogonal array is selected for the controllable
factors while an L8 orthogonal array is chosen for the noise factors, see Table 4. The ones, twos,
and threes in the L9 array in Table 4a correspond to the low, medium, and high levels identified
for each control factor and listed in Table 3. Similarly, the ones and twos in the L8 in Table 4b
corresponding the low and high levels for each of the noise factors. Note that only the columns
labeled E, F, and G in Table 4b are actually used in the experiment. Since there are only three
noise variables, the remaining columns in the L8 array are used to estimate the interactions
between certain noise factors (e.g., ExF represents the interaction between conditioning time, E,
and temperature, F). Finally, the last column in the L8 array is used to estimate the variance in
the experiment.

Factor Factor
Run A B C D Run E F ExF G ExG FxG e
1 1 1 1 1 1 1 1 1 1 1 1 1
2 1 2 2 2 2 1 1 1 2 2 2 2
3 1 3 3 3 3 1 2 2 1 1 2 2
4 2 1 2 3 4 1 2 2 2 2 1 1
5 2 2 3 1 5 2 1 2 1 2 1 2
6 2 3 1 2 6 2 1 2 2 1 2 1
7 3 1 3 2 7 2 2 1 1 2 2 1
8 3 2 1 3 8 2 2 1 2 1 1 2
9 3 3 2 1
(a) L9 Orthogonal Array (b) L8 Orthogonal Array
for the Control Factors for the Noise Factors
Table 4 Designs for the Control and Noise Factors

The total set of experiments that are performed is obtained by combining the L9 array of
control factors (the outer array) with the L8 array of noise factors (the inner array). The total
number of experiments is the product of the number of runs of each array, i.e., 9 x 8 or 72
experiments. For each experiment, the pull-off force is measured using the specified settings for
each control factor level and noise factor level. The average pull-off force for each combination
of the control factors A-D. Since the objective in the experiment is to maximize the pull-off
force, the signal-to-noise ratio for “Larger is Better” is also computed for each set of runs. These
results are summarized in Table 5 and discussed in the next section.

7


Outer E 1 1 1 1 2 2 2 2
Array (L8) F 1 1 2 2 1 1 2 2
G 1 2 1 2 1 2 1 2
Inner Array (L9) Responses
Run A B C D y SNL
1 1 1 1 1 15.6 9.5 16.9 19.9 19.6 19.6 20.0 19.1 17.525 24.025
2 1 2 2 2 15.0 16.2 19.4 19.2 19.7 19.8 24.2 21.9 19.475 25.522
3 1 3 3 3 16.3 16.7 19.1 15.6 22.6 18.2 23.3 20.4 19.025 25.335
4 2 1 2 3 18.3 17.4 18.9 18.6 21.0 18.9 23.2 24.7 20.125 25.904
5 2 2 3 1 19.7 18.6 19.4 25.1 25.6 21.4 27.5 25.3 22.825 26.908
6 2 3 1 2 16.2 16.3 20.0 19.8 14.7 19.6 22.5 24.7 19.225 25.326
7 3 1 3 2 16.4 19.1 18.4 23.6 16.8 18.6 24.3 21.6 19.850 25.711
8 3 2 1 3 14.2 15.6 15.1 16.8 17.8 19.6 23.2 24.2 18.838 24.852
9 3 3 2 1 16.1 19.9 19.3 17.32 23.1 22.7 22.6 28.6 21.200 26.152
Table 5 Pull-Off Force for Connector and Tube Parameter Design Experiment

32.4.2 Data Analysis
In this experiment, Taguchi’ graphical approach is used to plot the “marginal means” of
s
each level of each factor and “pick the winner” to determine the best setting for each control
factor. The average pull-off force and SN ratio for each level of each of the control factors are
plotted in Figure 4. These values are computed by averaging the mean pull-off force or SNL for
each factor for each level. For example, the average pull-off force for the shallow setting (Level
1) of the insertion depth (Factor C) is obtained by averaging Runs 1, 6, and 8 in Table 5, i.e.,
(17.525 + 19.225 + 18.838)/3 = 18.4. The same procedure is employed to compute the average
SNL for each level of each factor and the remaining pull-off force averages.

(a) Effect on SNL (b) Effect on Average Pull-off Force
Figure 4 Control Factor Effects

Figure 4 reveals that of the product design factors, interference (A) and wall thickness
(B), the interference has a larger impact on SNL and the average pull-off force. The medium
level for A (Amedium) is clearly the best choice for maximizing SNL and the average pull-off force.
As for the wall thickness, B, levels Bmedium and Bhigh are slightly better than Blow; however,
Bmedium is preferred to Bhigh in order to maximize the average pull-off force, see Figure 4b.

8


For the process design parameters, insertion depth (C) and percent adhesive (D), the
insertion depth has a much larger impact on SNL and the average pull-off force than does the
percent adhesive. From Figure 4a, Cmedium and Cdeep yield nearly equal SNL, but Cdeep should be
selected to maximize the average pull-off force. What is the implication of two levels of SNL
being nearly equal? The answer is that the average pull-off force increases when Cdeep is chosen,
and SNL increases beyond its value when Cmedium is chosen. However, the variance in the pull-
off force also increases when Cdeep is chosen and SNL decreases. The end result is that Cmedium
and Cdeep yield nearly equal SNL. Notice that the SNL values are nearly the same for Dmedium and
Dhigh as well. It appears that level Dlow should be selected for percent adhesive since it yields a
slightly higher SNL and average pull-off force compared to Dmedium and Dhigh.
Hence, the best settings to maximize SNL are Amedium, Cdeep, Bmedium, and Dlow based on
the experimental results for maximizing pull-off force. In the actual study, further analysis of the
data revealed that (1) the variance in the experiment was not constant and depended on the
specific levels of each control factor, and (2) there were several interactions between some of the
control factors and noise factors (Byrne and Taguchi, 1987). A summary of the results is given
in the next section for the full experiment.

32.4.3 Summary of Results
While it was easy to “pick the winner” from the effects plots in Figure 4, these factor
settings are only best for maximizing the pull-off force. In the complete set of experimental
results which also includes assembly effort and cost ratings, the choice of factor levels is not as
easy since many objectives are conflicting. A summary of the results is given in Table 6 where
insignificant effects are noted with a dash. Based on the full set of data, the best levels for each
factor are: Amedium, Blow, Cmedium, and Dlow. Only two of these choices are the same as those
obtained by examining Figure 4 since the most desirable settings for reducing cost rating and
assembly effort are not the same as those required to maximize pull-off force. The final choice
of factor level requires a trade-off between the multiple competing objectives which is often the
case in product and process design.

Assembly Pull-Off Cost Performance Rating Selected
Factor Level Effort Force Rating Assembly Pull-off Level
Interference 1 Low 8.1 18.7 Least Best Worst
(A) 2 Medium 8.3 20.7 – – Best X
3 High 8.7 19.8 Most Worst –
Wall 1 Thin – – Least – – X
thickness 2 Medium – – – – –
(B) 3 Thick – – Most – –
Insertion 1 Shallow 7.7 18.4 Least Best Worst
depth 2 Medium 8.3 20.3 – – – X
(C) 3 Deep 9.1 20.6 Most Worst Best
Percent 1 Low 8.3 20.5 Least Best Best X
adhesive 2 Medium 8.4 19.5 – Worst –
(D) 3 High 8.4 19.2 Most Worst Worst
Table 6 Summary of Parameter Design Results for Elastometric Connector

This chapter is an excerpt from Wysk, R. A., Niebel, B. W., Cohen, P. H., and Simpson, T. W.
Manufacturing Processes: Integrated Product and Process Design, McGraw Hill, New York, 2000.

9

Ie466.robust.handout

More Related Content

Similar to Ie466.robust.handout (20)

More from Ngoc Dep (20)

Recently uploaded (20)

Ie466.robust.handout