Вы находитесь на странице: 1из 11

Int J Adv Manuf Technol (2005) 25: 1180–1190

DOI 10.1007/s00170-003-1962-x

ORIGINAL ARTICLE

Kun-Lin Hsieh · Lee-Ing Tong

Process quality improvement of a qualitative response


with dynamic characteristic

Received: 4 July 2003 / Accepted: 18 September 2003 / Published online: 29 March 2004
 Springer-Verlag London Limited 2004

Abstract To achieve quality improvement in manufacturing


environments, the design of experiment (DOE) and Taguchi 1 Introduction
methods are two efficiency approaches to address such prob-
lems. Applying those methods to resolve quality improvement Fierce market competitiveness has driven manufacturers to focus
frequently focuses on the measurable quality response (or quan- on the work of enhancing their product’s quality. Off-line quality
titative response). However, the complexity of products or pro- control is a well-known cost-effective means of optimizing the
cesses is gradually increased and due to the limitation of measur- product or process quality. Design parameters and noise param-
ing equipment, the quality response could not be directly meas- eters are the factors heavily influencing quality of responses for
ured. Visual inspection or measurement is then used to judge the product or manufacturing processes. Herein, experimental
the quality for nonmeasurable response (or qualitative response). designers can control design parameters and the noise param-
Basically, for assessing the quality of a qualitative response, it is eters cannot be directly controlled by them. The technique of
initially divided into several classes or ordered categories. As for experimental design is frequently applied to study the relation-
customers’ requirements gradually changing and applications’ ship between quality response and design factors or noise factors.
flexibility gradually increasing, a dynamic structure for this pro- The primary goal is to improve quality via a robust design. A ro-
cess will be another important consideration for manufacturers. bust design is desired to obtain the optimum setting of design
That is, it causes the work of quality improvement to be dynami- parameters for a product or a manufacturing process in such
cally analyzed. Dr. Taguchi had proposed a dynamic method to a manner that the product attains its desired target with minimum
analyze such issues. However, only several studies focused on variation.
quality improvement of a qualitative response during the static For most products, the measurable (or quantitative) qual-
characteristic. Parameter optimization approaches were seldom ity response is frequently considered due to the inherent na-
proposed to address quality improvement of a qualitative re- ture of the quality response. Conventional experimental design
sponse with the dynamic characteristic. Therefore, in this study, techniques were used to investigate the relationship between
we proposed an integrated parameter-optimization approach to a quantitative quality response and the design dimensions. Ad-
resolve quality improvement of a qualitative response with dy- ditionally, it had been applied well for many regions: e.g., op-
namic characteristics. An illustrative example, quality improve- timization of chemical processes or manufacturing processes.
ment of lead twist during the stamping process for lead frame Taguchi’s method [10] combines experimental design techniques
manufacturing at Science-based park in Taiwan, is employed to with quality loss considerations, which will make it be an ef-
demonstrate the effectiveness of the proposed approach. ficient approach for process improvement. In some cases, the
quality response of interest may be a nonmeasurable response
(qualitative). The nonmeasurable response can also be described
Keywords Dynamic characteristic · Parameter optimization · linguistically [15, 16]. For instance, in assessing the uniformity
Quality improvement · Qualitative response of quality of the ion implanting process on a wafer for semi-
conductor manufacturing, linguistic descriptions, such as perfect
K.-L. Hsieh (u) · L.-I. Tong uniformity, acceptable uniformity, and worst uniformity, could
Department of Information Management, define efficiently the performance of quality. Using a linguis-
Nanhua University, tic description allows practitioners to obtain more information
32 Chung Keng Li, Dalin Chiayi 622, Taiwan, R.O.C. hidden in engineering philosophy. Taguchi [12] had proposed
Assistant professor
E-mail: klhsieh2644@mail200.com.tw an accumulation analysis (AA) to achieve the parameter opti-
Tel.: +886-5-2721001, ext. 56430 mization of an ordered categorical response. In a later work,
1181

Nair [9] also developed a scoring scheme (SS) and, Jean and ditions. Taguchi also recommended using techniques of Omega
Guo [4] modified Nair’s SS to propose a weighted probabil- (Ω) transformation. Omega transformation, which transfers the
ity scoring scheme (WPSS) for addressing the parameter op- accumulation probability of factor level to a corresponding Ω
timization of an ordered categorical response. Fuzzy set the- value. Then it can yield the predicted accumulated probability.
ory [5] is a well-known approach used to manage the uncer- Next, the optimum parameter setting can be screening on the fac-
tainties of the qualitative type or the linguistic description of tors’ effect diagram. Subjective judgment, while attempting to
response, especially for decision making for weight for several determine the optimum parameter setting from the factors’ effect
items [1]. In this study, the fuzzy set theory is also applied to as- diagram, will be frequently used.
sess the weight value of the linguistic category for a qualitative Nair [9] presented two scoring schemes (SS) to separately
response. study the dispersion and location effects. The mean square is rec-
Requirements of customers gradually change and the flexi- ommended as evidence of the important effect. Correspondingly,
bility of applications gradually increase, causing the work of the optimal condition solutions of both effects can be obtained
quality improvement to be dynamically analyzed. Restated, the according to the contribution of the dispersion effect and the lo-
quality of response may be varied with some conditions and it cation effect of each control factor. The final optimal control fac-
can be viewed as a dynamic system. For instance, a plating pro- tor level combination is obtained by making an adjustment action
cess can be viewed as a dynamic system – the thickness of plat- between the dispersion effect and the location effect. This ap-
ing varies with the amount of passing current. Taguchi [10, 14] proach is computationally complex for practitioners. Moreover,
also developed a solution for the dynamic system. The variabil- the final adjustments are limited to well-experienced engineers,
ity and sensitivity of the response are simultaneously considered who must use their subjective judgment to compromise disper-
in Taguchi’s dynamic system. In this study, we extend Taguchi’s sion and location effects.
dynamic method to a robust structure. Not only the variabil- Jean and Guo [4] proposed a weighted probability scoring
ity and sensitivity of a qualitative response are considered, but scheme (WPSS) to reduce Nair’s SS drawbacks. Their approach,
the performance of response is also included in the integrated which is simpler and more straightforward than Nair’s SS, com-
approach. Until now, only several studies focused on quality bines the dispersion and location effects into a single mean
improvement of a qualitative response during the static charac- square deviation (MSD). According to the ordered categorical
teristics. Parameter optimization approaches were seldom pro- data definition, the appropriate characteristic is employed and
posed to address quality improvement of a qualitative response the expected mean square deviation can be solved. The optimal
with dynamic characteristics. Therefore, in this study, we pro- control factor level combination is obtained by selecting the min-
pose an integrated parameter optimization approach to resolve imum mean square deviation. Although simpler than Nair’s SS,
quality improvement of a qualitative response with dynamic their approach suffers two drawbacks: (1) Their approach must
characteristics. still use subjective judgment to compromise the dispersion and
The rest of this study is organized as follows. Section 2 re- location effect and (2) the weight given might confuse an inex-
views pertinent literature involving optimization techniques of perienced engineer.
a qualitative response, Taguchi’s dynamic system, and the fuzzy
set theory. Subsequently, Sect. 3 systematically describes the de- 2.2 Taguchi’s dynamic system
tailed procedure of the proposed parameter optimization. Sec-
tion 4 provides an illustrative case involving the lead twist prob- Parameter design is also commonly referred to as robust de-
lem during a stamping process from a lead frame manufacturer sign [10]. Taguchi’s parameter design can be divided into two
in Taiwan’s Science-Based Park to demonstrate the effectiveness classes, depending on the system’s characteristics: static and
of the proposed approach. Concluding remarks will be made in dynamic. These two classes differ primarily in that the latter
Sect. 5. employs the signal factor to analysis and the former does not.
As is well recognized, the dynamic characteristic is an import-
ant class in Taguchi’s parameter design. It can describe a real
problem well by using Taguchi’s dynamic system. The design
2 Literature review
parameters and noise parameters can influence a product and op-
2.1 Parameter optimization for the qualitative (or linguistic) erational process. The design parameters, as controlled by the
response designer, can be divided into (1) signal factor, which only influ-
ences the average of the quality response, and (2) control fac-
To resolve the parameter optimization of a qualitative response, tors, which will influence the variability of quality response. The
Taguchi [12] had developed the accumulation analysis (AA) to noise parameters cannot be controlled or they are too expensive
address it. Four steps are constructed in Taguchi’s AA: (1) defin- to control. Parameter design of Taguchi’s method concentrates
ing the categories and the number of categories for a qualitative primarily on selecting parameters settings to minimize the effect
response, (2) drawing the effects of the factor levels upon the of the noise factor. Restated, the design parameters for a prod-
probability value, (3) plotting the accumulated probabilities, and uct or process should be set up so that the response is close to
(4) deciding the optimum parameter setting and predicting the the desired target with the minimum variability [10, 14]. Experi-
accumulated probabilities of each category under optimum con- mental design is frequently employed to explore the relationship
1182

between the design parameters and response. In a Taguchi ex- 2.3 Fuzzy set theory
periment, the design parameters and noise parameters are first
arranged in an orthogonal array (OA). Then a signal-to-noise The grade of representation in real life may have uncertainties [5].
ratio (SN ratio) is computed for each experimental combination. Some possible misunderstandings may occur when the uncertain-
The SN ratio, one of Dr. Taguchi’s major contributions to quality ties are represented in conventional binary form. The uncertainties
engineering, is an essential evaluation of parameter design [10]. among the grade of the representation can be represented well by a
The quality characteristic Y is generally regarded as a function linguistic description. There are few differences between fuzzy set
of signal factor M for a dynamic system. The relationship be- and traditional set: the traditional set takes a characteristic func-
tween signal factor and quality characteristics is assumed to be tion only with values 0 and 1 to describe a set, and the fuzzy set
linear, i.e., take a membership function (MF) with the interval [0, 1] to de-
scribe a set. Zadeh [15, 16] proposed a method based on fuzzy
Y = βM + ε, (1) set to formalize linguistic assessment. Basically, the fuzzy set can
be constructed by set elements with respect to the correspond-
where β denotes the sensitivity and ε represents the error term. ing problem. Herein, set elements can also be defined as possible
By considering the different level combination of the control fac- factors affecting the linguistic description. Employing the MF in
tors, Eq. 1 can be rewritten as the fuzzy set can possess the defuzzied activity. That is, the MF
can transform the linguistic assessment into a value in the interval
Y = β(d)M + ε(d), (2) [0, 1]. The magnitudes of the membership values are the member-
ship degree of a fuzzy term with respect to the set elements in the
where d denotes the level combination of control factor, β(d) fuzzy set. Accordingly, the possible misunderstandings using the
represents the system sensitivity under d, and ε(d) is the random conventional binary set can be avoided by using the concept of
error term under d. To resolve a system’s robustness, Taguchi the fuzzy set. Lai and Chang [6] proposed a fuzzy multiresponse
proposed the formula optimization procedure to derive an appropriate combination or
process of parameter setting. A strategy of optimizing the most
β2
η = 10 log (3) possible response values and minimizing the deviation from the
MSE most possible values is used because one must consider not only
for evaluating the SN ratio in decibels (dB). The mean square the most possible values but also the imprecision of the predicted
of error (MSE) represents the mean square of the distance be- response. Tong and Su [13] also proposed a procedure, capable
tween the measured response and the best-fit line. SN ratio and of applying fuzzy set theory to multiple-attribute decision making
sensitivity (S) are utilized as the index for evaluating variability (MADM) [2], to optimize a multiresponse problem. That investi-
and sensitivity. For understanding the detailed contents about SN gation also applied a similar technique for ordering performance
ratio and S value, readers can refer to [10]. by the similarity to an idea solution index to determine the opti-
Wasserman [14] presented a case study of parameter de- mum parameter setting. For detailed content of the fuzzy set the-
sign with dynamic characteristics by using multiple regression ory, readers can refer to [5, 15, 16]. Chen [1] had applied fuzzy
models. However, practitioners with limited statistical training set theory to study the parameter optimization for multiple re-
have difficulty in comprehending such a method. Lunani et sponse problems for considering the weight values of the multiple
al. [7] proposed two graphical methods, the sensitivity-standard responses. Then Hsieh [3] proposed an integrated methodology
deviation plot and the gamma plot, for identifying suitable to achieve quality improvement for processes with the qualita-
measures of dispersion and for performing data analysis. The tive response. His approach applied fuzzy set theory to resolving
two plots provide different views of the data and, they recom- the uncertainties between linguistic categories for a qualitative
mended that two plots be used jointly to supplement the dif- response. However, the significance of design factors cannot be
ferent sources of information. However, the statistical theory studied in this approach. Restated, the flexibility for engineering
for constructing two plots is more complicated. McCaskey and application or engineering analysis will be limited.
Tsui [8] discussed that Taguchi’s dynamic model can be di-
vided into a multiplicative model and an additive model. They
also recommended that two-phase optimization in Taguchi’s dy- 3 Procedure for parameter optimization of
namic system can only be used for the multiplicative model.
a qualitative response with dynamic characteristic
A new procedure was proposed to resolve the additive model
in Tauchi’s dynamic system. However, the complicated com- Five phases are included in the proposed procedure. Phase 1 will
putation of their procedure will be difficult for those having define initially the category and the number of the category. Then
limited statistical training. Su and Hsieh [11] proposed a novel the weight of the corresponding category for a qualitative re-
means of applying neural networks to achieve quality improve- sponse incorporating the engineer’s knowledge and experience
ment for a response with dynamic characteristics. Although the into decision making will be computed. In Phase 2, a summa-
application of their methodology is flexible for others, the en- rized evaluation value is computed by using the weight value
gineering analysis for design factors cannot be included in their derived from Phase 1. The difference between the summarized
methodology. evaluation value and the target for each category can be de-
1183

termined in Phase 3. Next, in Phase 4, we will determine the Table 2. The reference table of crisp number
optimum parameter setting by screening those factors in the
sequence affecting the variability, sensitivity, and quality with SCALE 1 2 3 4 5 6 7 8
respect to the result obtained from Phase 2 and Phase 3. Fi- No. of terms Two There Five Five Six Seven Nine Eleven
nally, depending on the predecided criterion and carrying out
confirmed experiments, one will verify the effectiveness of the 1. Excellent 0.954
optimum parameter setting. The detailed procedure will be given 2. Very high 0.909 0.917 0.909 0.917 0.864
3. High–very high 0.875 0.701
as follows. Figure 2 graphically depicts the flowchart of the pro- 4. high 0.750 0.833 0.717 0.885 0.750 0.773 0.750 0.667
posed approach. 5. Fairly high 0.7 0.584 0.630
6. Mol high 0.637 0.590
Phase 1: define the categories, the number of categories for 7. Medium 0.583 0.500 0.500 0.500 0.500 0.500 0.500
a qualitative response, and compute their corresponding weight 8. Mollow 0.363 0.410
9. Fair low 0.300 0.416 0.370
values 10. Low 0.166 0.283 0.115 0.250 0.227 0.250 0.333
11. Low–very low 0.125 0.299
For analyzing engineering problems, engineers’ knowledge or 12. Very low 0.091 0.083 0.091 0.083 0.136
experience can provide more available information about it. Es- 13. None 0.046
pecially, the knowledge and experience are available for decision
making about weight value for several items. The fuzzy set the-
ory will be employed to incorporate the engineers’ knowledge
being chosen are employed to represent such three
and experience into analysis; two steps are included in this phase.
categories. However, if the different scales are consid-
Step 1: The definition of categories and the numbers of cat- ered, the scale with the minimum categories should
egories will be determined with respect to engineers’ be used, e.g., two engineers choose three linguistic
knowledge or the judgment of the experienced engineers. descriptions (high, medium, low) to represent quality
Step 2: To decide the weight value of each category. of response, six scales (SCALE 2, 3, 4, 6, 7, 8) from
(i) Incorporating the concept suggested by Chen [1] Table 1 can be the choice. Herein, we should choose
with engineers’ knowledge or experience, the impor- scale 2 for including the minimum linguistic terms.
tance of each category can be represented as several (iii) After choosing the linguistic descriptions from
linguistic descriptions. Thirteen descriptions are fre- Table 1, the crisp number can be determined by
quently used. Table 1 lists the definition of linguistic screening out Table 2, e.g., scale 2 is chosen from
descriptions. Table 1 and three linguistic descriptions (high,
(ii) In Table 1, eight corresponding scales are also pro- medium, low) are also determined. Then three crisp
vided. The linguistic term and the corresponding scale numbers 0.833, 0.500, and 0.166 are screened out
will be denoted as yes in Table 1. The scale can be from Table 2. The summation of the crisp numbers
regarded as the maximum numbers for including lin- being screened out does not equal 1. Hence, they
guistic terms, e.g., scale 3 denotes that we can describe should be standardized into a weight value lying be-
a response by the maximum categories of five (very tween 0 and 1. The weight value will be viewed
high, high, medium, low, very low). If a response just as the weight of each category, denoted as (Lw) K ,
has three categories, the three linguistic descriptions where k = 1, 2,. . . , s and k means the number of
categories and sk=1 (Lw)k = 1.
(iv) Assuming a engineers assess the weight value of the
Table 1. The reference table of eight scale category, n weight values can be obtained (n ≤ a).
The ratio of the nth is an /a, where an denotes the
SCALE 1 2 3 4 5 6 7 8 numbers of engineers in the nth term. If the weight
No. of terms Two There Five Five Six Seven Nine Eleven value of the kth category in the nth term is (Lw)nk ,
the weight value Wk of the kth category will be given
1. Excellent Yes as follows:
2. Very high Yes Yes Yes Yes Yes
3. High–very high Yes Yes 
n
4. High Yes Yes Yes Yes Yes Yes Yes Yes Wk = (Lw) pk × (a p /a). (4)
5. Fairly high Yes Yes Yes p=1
6. Mol high Yes Yes
7. Medium Yes Yes Yes Yes Yes Yes Yes
8. Mollow Yes Yes
9. Fair low Yes Yes Yes Phase 2: design experiments and compute the summarized eval-
10. Low Yes Yes Yes Yes Yes Yes Yes uation value for each trial
11. Low–very low Yes Yes
12. Very low Yes Yes Yes Yes Yes Step 1: Engineers utilize brainstorming to screen out the possible
13. None Yes
design factors. Then the level of each design factor can
1184

also be determined and an orthogonal array (OA) L x will denotes the target value.
be chosen, where x denotes the number of experimental Step 2: Integrate the difference values of all g levels for signal
trial (x = 1, 2, . . . , h). factor into an assessment index for the xth experimental
Step 2: Compute the accumulatively weighted ratio value Cxuk trial. Such an assessment index will be given as follows:
for the xth experimental trial for the kth category under  g 
the uth signal factor. Assuming m xuk (k = 1, 2, . . . , s) 
Dx = − log10 d xu .
2
(8)
denotes the number of the kth observation for the xth
u=1
(x = 1, 2, . . . , h) experimental trial
 under the uth (u =
1, 2, . . . , g) signal factor, where sk=1 m xuk = m and m
denotes the number of observations. The accumulatively Phase 4: determine the optimum factor/level combination
weighted ratio value can be computed as follows:
Step 1: Compute the estimated slope value β̂ and variability σ̂ 2

e by using summarized evaluation value Yx and the signal
Cxue = Rxuk Wk , e = 1, 2, . . . , s, (5) factor. Then plot the response diagram by utilizing the
k=1 SN ratio and the S ratio. Screen out those factors affect-
where Rxuk = m xuk /m denotes the ratio value of the xth ing variability and sensitivity.
experimental trial for the kth category under the uth sig- Step 2: Utilizing Eq. 8 to screen out the adjustable factors to
nal factor; Wk denotes the weight value of the kth cate- adjust the fitted regression line to the target line. That
gory. is, the performance can be enhanced after doing the
Step 3: The summarized evaluation value Yxu for the xth (x = work.
1, 2, . . . , h) experimental trial under the uth signal factor The adjustable factors, which are the factors that can
will be given as follows: be changed without affecting variability or sensitivity,
only have effect on the performance. Besides, the con-

s cept of the qualitative response can be viewed as the
Yxu = Cxuk . (6) larger-the-better (LTB) and the goal of the adjustable
k=1 factor is to move the fitted regression line to the tar-
get line. Figure 1 depicts the concept of adjustable
factor.
Phase 3: compute the difference between the ratio of each cate- Step 3: Determine the optimum parameter setting in the se-
gory and the target for each experimental trial quence of variability reduction, sensitivity, and perform-
2 between the ratio value R ance enhancement.
Step 1: Compute the difference d xu xuk
of the xth experimental trial for the kth category and the
target value Tk under the uth signal factor: Phase 5: Confirm and compare with the given criteria


s The comparison criteria can be predecided according to the real
2
d xu = [Wk (Rxuk − Tk )]2 (7) requirements. To verify the effectiveness of the result, we must
k=1 conduct a confirmed experiment. If the result cannot be accepted
where or it cannot arrive at the criteria on which we predecided, it must
 go back to phase II to reanalyze the design factors and signal
1, k = 1 factor or redesign The experiment or repeat the entire procedure
Tk =
0, k = 2, 3, . . . , s until the criteria can be met.

Fig. 1. The concept diagram of the ad-


justable factor
1185

Fig. 2. The flowchart of the proposed parameter optimization

Lead (or finger) could be twisted for external torque or external


4 Illustrative case study – quality improvement of force and, it would lead to a cost loss and lower yield. Hence,
lead twist in Taiwan’s lead frame manufacturer the lead frame manufacturer decided to institute quality improve-
4.1 Problem description ment for the defect status of lead twist. By the way, the effects
of design factors are also studied. For the stamping process, the
An illustrative example, from a lead frame manufacturer at bare will be injected initially into the stamping station. Then the
Science-based park in Taiwan, is employed to demonstrate the upper mode will press downward to cut the unnecessary part bare
detailed procedure for our proposed approach. Lead frame is to locate it correctly on the guide. Next, the prototype of the lead
a necessary material for conventional integrated circuit (IC) frame will be injected out of the stamping station after a given
packaging. A basic manufacturing flowchart of lead frame can operating time (or cutting time) and a given operating pressure.
be divided into four processes: stamping, plating, downsetting, Different cutting speed, cutting force, torque of upper and lower
and taping processes. The purpose of the stamping process is to mode, planarity of punch and dies, or thickness of bare will af-
cut the unnecessary part from a patterned bare and the proto- fect the quality of stamping. If lead twist occurs, the gold wire
type of lead frame can be formed. Lead twist, lead planarity, lack line cannot be mounted well on leads during the bonding process.
Ag plating, Ag nonplating, over downset depth or under downset Restated, it will lead to wire breaking and will increase yield loss
depth are the defective statuses frequently being inspected dur- for the packaging process.
ing lead frame manufacturing. Among those defect statuses, lead In this improvement project, quadratic flat package (QFP)
twist has been recognized as an important source for causing bro- lead frame is initially chosen. The reason is that this type of
ken wire during the wire bonding in the later packaging process. QFP gradually has a huge amount of production and the cost is
1186

higher than other types. The chosen QFP has 100 leads, where
each side has 24 leads. Figure 3 depicts the architecture diagram.
The quality response we focused on is lead twist, and no equip-
ment can directly measure the twist angle. The manufacturer had
self-designed a line-scanning matcher to evaluate the twist angle
now. That is, in such equipment, there are several ordered cate-
gories with different angle ranges. The scanning signal obtained
from measuring is then compared with those categories in the
laser scanning matcher and the matching category can be deter-
mined. The result obtained by the line scanning matcher will be
a range, the definition of twist angle being that angle between
the horizontal line and the maximum twist side. Figure 4 de-
picts the definition of twist angle. From the definition of twist Fig. 5. The relationship between ratio of width and thickness and twist angle
angle, the quality response can be viewed as an ordered categor- less than 30
ical response.
To satisfy the customer’s requirements, the design of the lead
frame could provide several ratios of width-thickness. Under dif-

Fig. 6. The location diagram of 20 leads

ferent ratios of width-thickness, the quality response could lead


to a significant result. To analyze the mechanism of the stamp-
ing process and the material feature of the lead frame by utilizing
engineering knowledge or engineering experience, the larger the
ratio of width-thickness, the smaller is the twist angle. From his-
torical data analysis, the relationship between the twist angle and
the ratio of width-thickness will be close to linear. It is depicted
in Fig. 5. After recognizing the condition, it is viewed as a dy-
namic system and the ratio of width-thickness can be extracted
to be the signal factor. Hence, this project can be viewed as
Fig. 3. The diagram of lead frame
the quality improvement of a qualitative response with dynamic
characteristics.
The inspected leads are determined as 20 leads and the cor-
responding location is depicted in Fig. 6. The inspection for the
20 leads primarily depends on historical inspection conditions
and engineering experience.

4.2 Analysis result

The analysis procedure is given as follows:


Phase 1: Determine the categories, the number of categories for
a qualitative response, and compute their corresponding
weights.
Step 1: Five ordered categories for the range of twist
Θ: angle are firstly determined by engineers’ know-
ledge and experience. The definition of five
: categories is given in Table 3.
Step 2: For involving the relative importance of five
: categories into analysis, five experienced en-
Fig. 4. The definition of twist angle gineers (marked as A1 , A2 , A3 , A4 and A5 )
1187

Table 3. The definition of category for the qualitative response Table 5. The weight value of each category

Category Range of twist angle Explanation Engineers I II III IV V

I 0–15◦ ← Best A1 0.379 0.280 0.209 0.095 0.038


II 15◦ –30◦ A2 0.435 0.239 0.174 0.109 0.044
III 30◦ –45◦ ← Moderate A3 0.380 0.235 0.199 0.133 0.054
IV 45◦ –60◦ A4 0.435 0.239 0.174 0.109 0.044
VI > 60◦ ← Worst A5 0.379 0.280 0.209 0.095 0.038

Table 4. The crisp number derived from transformation Table 6. The weight value (Wk ) of each category after integration

Engineers I II III IV V Engineers I II III IV V

A1 0.909 0.670 0.500 0.227 0.091 Weight 0.402 0.254 0.193 0.108 0.043
A2 0.909 0.500 0.363 0.227 0.091
A3 0.954 0.590 0.500 0.333 0.136
A4 0.909 0.500 0.363 0.227 0.091
A5 0.909 0.670 0.500 0.227 0.091 Step 2: Compute the accumulatively weighted ratio
value for different experimental trials under
each level of signal factor by using Eq. 5. Partial
brainstormed to determine the relative impor- results are listed in Table 7.
tance according to Table 1. Then a crisp number Step 3: Summarize the accumulatively weighted ratio
of each category for each engineer was derived value of each category under each experimen-
from the transformation of Table 2. The results tal trial for each level of signal factor according
of these crisp numbers are listed in Table 4. to Eq. 6. The obtained results are then regarded
Next, utilizing standardization of these crisp as the summarized evaluation value for each
numbers, the weight of each category deter- experimental trial under each level of signal fac-
mined from engineers can be obtained, and they tor. Table 8 lists the partial results.
are listed in Table 5. Substituting the result of Phase 3: Compute the difference between the ratio value of each
Table 5 into Eq. 4, the weight value of each category and the target for each experimental trial.
category integrating the experience for all engi- Step 1: Initially compute the difference value d2 be-
neers can be derived. Table 6 lists the result. tween the ratio value of each category and the
target for each experimental trial under each
Phase 2: Design experiments and compute the summarized eval- level of signal factor.
uation value of each trial. Step 2: Integrate the difference value d2 of each level
Step 1: Five control factors are chosen in this example. of signal factor into a distance index Dx ,
During those factors, factor A has two levels x = 1, 2, . . . , 18, for each experimental trial.
and the others have three levels and an orth- Table 9 lists the results.
ogonal array (OA) L 18 is then utilized. Phase 4: Determine the optimum parameter setting.

Table 7. The accumulatively weighted ratio value

SIGNAL 0.8T 1.0T 1.2T


FACTOR
Exp. No (I) (II) (III) (IV) (V) (I) (II) (III) (IV) (V) (I) (II) (III) (IV) (V)

1 0.020 0.122 0.170 0.197 0.199 0.141 0.204 0.243 0.259 0.261 0.221 0.259 0.278 0.300 0.300
2 0.000 0.000 0.067 0.116 0.125 0.080 0.118 0.167 0.199 0.203 0.181 0.219 0.277 0.282 0.284
3 0.000 0.013 0.090 0.133 0.139 0.040 0.142 0.190 0.206 0.211 0.120 0.159 0.255 0.255 0.257
4 0.000 0.000 0.077 0.131 0.135 0.060 0.187 0.245 0.245 0.247 0.161 0.212 0.269 0.280 0.280
5 0.000 0.013 0.138 0.160 0.164 0.000 0.153 0.182 0.192 0.199 0.100 0.164 0.231 0.242 0.244
• • • • • • • • • • • • • • •
• • • • • • • • • • • • • • •
• • • • • • • • • • • • • • •
16 0.020 0.147 0.205 0.211 0.215 0.120 0.210 0.238 0.249 0.254 0.181 0.232 0.289 0.295 0.295
17 0.000 0.178 0.217 0.222 0.224 0.181 0.257 0.276 0.292 0.292 0.321 0.321 0.350 0.356 0.356
18 0.000 0.178 0.217 0.217 0.221 0.000 0.102 0.160 0.181 0.186 0.060 0.175 0.242 0.248 0.248
1188
20.000

Step 1: Integrating the concept of Taguchi’s dynamic


system and the results from Table 8, those fac- 18.000

tors having significant effect on variability or 16.000

sensitivity can be then determined. Herein, fac- 14.000

tors A and B significantly affect the variability 12.000

and they can be determined from Table 10 and


10.000
Fig. 7. Factors A, B, and D will be the factors
significantly affecting the sensitivity and they 8.000
A1 A2 B1 B2 B3 C1 C2 C3 D1 D2 D3 E1 E2 E3

are determined from Table 11 and Fig. 8. Fig. 7. Response diagram for S/N ratio
Step 2: By involving the concept mentioned on Fig. 1
and the distance index Dx in Phase 3, the fac-
tors that significantly move the fitted regression Table 11. Response table of sensitivity
line to the target line can be determined. Then
Level
Control factor 1 2 3 Max-min
Table 8. The summarized evaluation value of each experimental trial under
each signal factor’s level A −0.696 0.788 — 1.484
B 1.036 −0.527 −0.041 1.564
Signal factor C 0.297 0.255 −0.084 0.382
D 1.147 −0.693 0.014 1.840
Experimental trial 0.8T 1.0T 1.2T E 0.236 0.040 0.192 0.196

1 0.708 1.107 1.358


2 0.308 0.768 1.243
3 0.375 0.789 1.046
4 0.343 0.986 1.202
5 0.474 0.725 0.983
• • • •
• • • •
• • • •
16 0.215 0.254 0.295
17 0.224 0.292 0.356
18 0.221 0.186 0.248

Fig. 8. Response diagram for sensitivity

Table 9. The distance index Dx


factors B and C are initially determined as the
Exp. no A B C D E Dx
candidate of adjustable factors from Table 12
1 1 1 1 1 1 0.268
and Fig. 9. According to the definition we made,
2 1 1 2 2 2 0.330 factor C can meet the condition: with an effect
3 1 1 3 3 3 0.403 on the performance and without an effect on
4 1 2 1 1 2 0.370 variability and sensitivity.
5 1 2 2 2 3 0.463
• • • Step 3: We can separately find out the significant fac-
• • • tors for variability, sensitivity and performance
• • • by screening out the difference between the
16 2 3 1 3 2 0.308
Max and Min evaluation value on response
17 2 3 2 1 3 0.257
18 2 3 3 2 1 0.504 table. And, they will be marked on Tables 10

Table 10. Response table of S/N ratio Table 12. Response table of distance index

Level Level
Control factor 1 2 3 Max-min Control factor 1 2 3 Max-min

A 12.175 18.357 — 6.182 A 4.655 5.524 — 0.869


B 18.121 12.827 14.849 5.294 B 5.855 4.528 4.847 1.327
C 17.176 14.318 14.303 2.873 C 6.969 5.268 4.899 2.070
D 15.995 13.827 15.975 2.169 D 5.758 4.963 5.109 0.795
E 15.536 14.608 15.652 1.045 E 5.530 5.306 4.994 0.537
1189

Table 13. The accumulated probability of each category for the current set-
tings

Accumulated probability of category (%)


25 pieces
(total 500 leads) I II III IV V

Current Settings
A1 B3 C2 D3 E1 43.6 52.4 78.6 91.8 100

Fig. 9. Response table of distance index


Table 14. The result of the optimum parameter settings
to 12 with bold entries (i.e. the difference
Optimum setting Accumulated probability of category (%)
value). Analyzing the results from Step 1 and A2 B1 C1 D1 E1
Step 2, factors A and B simultaneously affect (total 500 leads I II III IV V
the variability and sensitivity. And factor B each time)
has an effect on the performance. According
1 54.8 72.8 83.6 93.4 100
to the sequence of decision-making criteria for 2 52.4 74.6 82.8 95.2 100
variability, sensitivity, and performance, factor 3 56.2 73.2 84.2 93.2 100
A and factor B will be recognized as the fac-
tors affecting variability and, the optimum level
for factor A and factor B will be set as level Table 14. The accumulated probability of category (I
2 and level 1 for reducing variability. Not only and II) for each time significantly exceeds 70%; the
the variability can be reduced, but the sensi- criterion is met. Hence, the optimum parameter set-
tivity can be increased at the same time when ting for QFP production can be confirmed. By the way,
setting the same level for factors A and B. Fac- from the viewpoint of cost effectiveness, cost can be re-
tor D can be viewed as the factor affecting the duced by applying the optimum parameter setting we
sensitivity and its level is set as level 1. Al- derived.
though factors B and C will be regarded as the
candidates for adjustable factor, factor C is the
only one affecting the performance. So factor
C will be viewed as the adjustable factor and
it will be set as level 1. Finally, factor E does 5 Concluding remarks
not affect the variability, sensitivity, and per-
formance and its level setting can be determined Until now, most studies primarily focused on optimization of the
according to the considerations of cost or op- measurable response. As for quality improvement of a qualitative
eration. After discussions with senior engineers (or linguistic) response, only a few studies had been mentioned.
and the manager, factor E will be set as level Besides, for the limitation of measuring equipment, the increas-
1 according to the operation conditions. Hence, ing of a product’s complexity and flexible utilization of a user’s
in this quality improvement project, the opti- application, a qualitative response with dynamic consideration
mum parameter setting can be determined as will be necessary in the future. In this study, a parameter op-
A2 B1 C1 D1 E1 . timization approach addressing such case is proposed. After
Phase 5: Perform a confirmation and compare with the prede- demonstrating an illustrative example, several concluding re-
cided criteria. marks can be made:
To verify the effectiveness of the optimum parameter 1. The proposed approach does not require a huge amount of
setting obtained from the proposed approach, the se- statistical computation. It is an easy approach to use for those
nior engineers and manager initially decided the cri- practitioners having limited statistical training.
teria depending on the yield requirement (please re- 2. Incorporating engineers’ knowledge and experience into an-
fer to Table 13): the accumulated probability of the alysis it can provide more available engineering information
categories I and II must exceed 60%. After applying during the analysis. And the derived result will describe the
the optimum parameter setting to the QFP production real problem well.
line, the quality engineer randomly detected 25 pieces 3. In this study, a new schema of adjusting the fitted regres-
for three times from the production line after apply- sion line to the target line is proposed. By doing the work,
ing the optimum setting to QFP production. That is, the performance or quality can be studied and enhanced.
a total of 500 leads for each time are required to be Integrating the concepts of variability reduction, sensitivity
detected. The comparison result for the current set- increase, and performance enhancement will lead to an accu-
tings and the optimum parameter setting are listed in rate result.
1190

4. The proposed approach can be not only employed to quality 6. Lai Y-J, Chang T-H (1994) A fuzzy approach for multi-response opti-
improvement of a qualitative response with dynamic consid- mization: an off-line quality engineering problem. Fuzzy Set and Syst
63:117–129
eration, but it can also be employed to a qualitative response 7. Lunani M, Nair VN, Wasserman GS (1997) Graphical methods for ro-
with static consideration. bust design with dynamic characteristic. J Qual Technol 29:327-338
8. McCaskey SD, Tsui KL (1997) Analysis of dynamic robust design ex-
periment. Int J Prod Res 35:1561–1574
9. Nair VN (1986) Testing in industrial experiments with ordered categor-
ical Data. Technometrics 28:283–291
References 10. Phadke MS (1989) Quality Engineering Using Robust Design. Prentice-
Hall, Inc., New Jersey
1. Chen CK (1994) Applying fuzzy set theory to parameter optimization 11. Su CT, Hsieh KL (1998) Applying neural networks to achieve robust
for multiple response in Taguchi’s off-line quality control. Thesis, De- design for dynamics quality characteristic, Int J Qual and Reliabil Man-
partment of Industrial Engineering and Management, National Chaio age 15:509–519
Tung University 12. Taguchi G (1966) Statistical analysis (in Japanese). Maruzen, Tokyo
2. Chen SJ, Hwang CL (1992) Fuzzy multiple attribute decision making: 13. Tong LI, Su CT (1997) Optimizing multi-response problems in the
and Applications. Springer, Berlin Heidelberg New York Taguchi method by fuzzy multiple attribute decision making. Qual Re-
3. Hsieh KL (2001) Process improvement in the presence of qualitative re- liabil Eng Int 13:25–34
sponse by combining fuzzy sets and neural networks. Integr Manuf Syst 14. Wasserman GS (1996) Parameter design with dynamic characteristic:
12:449–462 a regression perspective, Qual Reliabil Eng Int 12:113–117
4. Jean Y-C, Guo S-M (1996) Quality improvement for RC60 chip resis- 15. Zadeh LA (1973) Outline of a new approach to the analysis of com-
tor. Qual Reliabil Eng Int 12:493–445 plex systems and decision processes. IEEE Trans Syst Man Cybernet
5. Klir GJ, Folger TA (1988) Fuzzy set, Uncertainty and Information. 3:28–44
Prentice-Hall, Inc., New Jersey 16. Zadeh LA (1965) Fuzzy sets. Inf Control 8:338–353

Вам также может понравиться