Int J Adv Manuf Technol (2004) 23: 896–902
DOI 10.1007/s00170-003-1733-8
O R I GI N A L A R T IC L E
C.-M. Hsu Æ C.-T. Su Æ D. Liao
Simultaneous optimisation of the broadband tap coupler optical
performance based on neural networks and exponential
desirability functions
Received: 14 October 2002 / Accepted: 16 January 2003 / Published online: 18 February 2004
Springer-Verlag London Limited 2004
Abstract This study presents an integrated procedure
using neural networks and exponential desirability
functions to resolve multi-response parameter design
problems. The proposed procedure is illustrated through
optimising the parameter settings in the fused bi-conic
taper process to improve the performance and reliability
of the 1% (1/99) single-window broadband tap coupler.
The proposed solution procedure was implemented on a
Taiwanese manufacturer of fibre optic passive components and the implementation results demonstrated its
practicability and effectiveness. A pilot run of the fused
process revealed that the average defect rate was reduced
to just 2.5%, from a previous level of more than 35%.
Annual savings from implementing the proposed procedure are expected to exceed 0.5–1.0 million US dollars.
This investigation has been extensively and successfully
applied to develop optimal fuse parameters for other
coupling ratio tap couplers.
Keywords Parameter design Æ Multi-response problem Æ
Neural network Æ Exponential desirability function Æ
Single-window broadband tap coupler
1 Introduction
Optical performance in a coupler manufacturing process
is usually influenced by more than one variable. These
variables include machine parameters, raw materials, the
C.-M. Hsu (&)
Department of Business Administration,
Minghsin University of Science and Technology,
Hsinchu, Taiwan R.O.C.
E-mail: cmhsu@must.edu.tw
C.-T. Su Æ D. Liao
Department of Industrial Engineering and Management,
National Chiao Tung University,
Hsinchu, Taiwan R.O.C.
process followed, environmental conditions and so on.
From the perspective of cost or feasibility, some variables cannot be precisely controlled. Furthermore, even
when these variables are controllable, the optimal combination of parameter levels that maximises product
quality may be unknown. The Taguchi method is a
conventional approach to resolving this problem. This
method allows engineers to determine a feasible combination of design parameter levels such that the variability of a product’s response is reduced and the mean
is close to the target. However, optimising a multi-response problem using the standard Taguchi method is
difficult. The usual recommendation for the optimisation
of a process/product with multiple responses is left to
engineering judgment and is verified by experiments [1].
However, the introduction of human judgment increases
uncertainty in the decision-making process. Logothetis
and Haigh [2] applied the multiple regression technique
and linear programming approach to optimise a fiveresponse process by the Taguchi method. Their method
was limited when the t-values of the regression coefficients were insignificant or when the coefficient of
determination was low. Pignatiello [3] presented a quadratic loss function for multi-response problems and
established a predictive regression model using controllable variables. Following the descent direction and
repeatedly establishing a new local experimentation region, this method minimised the expected loss. However,
it was difficult to determine the cost matrix and additional experimental observations were required. Tong
et al. [4] proposed a procedure to determine the multiresponse signal-to-noise (MRSN) ratio through integrating the quality loss of each response. However,
determining the weight ratios for responses was difficult,
and the optimal combination of factor levels was likely
to be dominated by the ‘‘maximum quality loss’’ in the
total of the trials. Antony [5] proposed an approach
using the Taguchi loss function and principal component analysis to optimise a submerged arc-welding process. In this study, it was difficult to determine the
optimal parameter settings if two or more eigenvalues
897
greater than one were obtained according to Kaiser’s
criterion [6]. Superimposing the response contour plots
and finding an optimal solution by visual inspection is a
simple and intuitive approach to resolving multi-response problems [7]. However, such a method is severely
limited by the number of input variables and/or responses [8]. The use of a dimensionality reduction
strategy has thus become a popular means of simultaneously optimising (compromising) multiple responses.
This method converts a multi-response problem into a
single-response problem with an aggregated measure,
which has often been defined as a desirability function
[9, 10] or as an estimated distance from the ideal design
point [11]. Kim and Lin [8] developed a modelling approach based on maximising exponential desirability
functions for optimising a multi-response system. Their
approach aimed to identify the settings of the input
variables to maximise the degree of overall minimal level
of satisfaction with respect to all the responses. Furthermore, the method required no assumptions regarding the form or degree of the estimated response models
and was sufficiently robust to handle the potential dependences between response variables.
In this study, an integrated procedure based on
neural networks and exponential desirability functions
was proposed to optimise the parameter settings in the
fused bi-conic taper (FBT) process that fabricates 1%
(1/99) single-window broadband tap couplers. The proposed optimisation procedure can help manufacturers of
fibre-optic passive components by greatly improving the
performance and reliability of 1% (1/99) single-window
broadband tap couplers at minimum cost.
are being used in numerous commercial applications,
such as optical fibre communications, optical fibre
amplifiers and lasers and so on. The FBT technology is
used to produce both WDMs and couplers. This technology relies on bringing bare fibre into contact, then
melting and drawing the cross-section to produce a tapered region, as illustrated in Fig. 1a [12]. This procedure produces a very thin tapered region, which must be
processed extremely carefully, and must be packaged to
protect the components during shipping, handling and
installation. In a typical package, as illustrated in
Fig. 1b [13], the fused section of the fibres is suspended
above the quartz substrate, and positioned between two
epoxy supports for mechanical stability. This assembly is
then enclosed in a metal tube and sealed. The FBT
process has been used for over a decade to fabricate
most of the coupler components used in various fibre
optic telecommunication, instrumentation, and sensor
systems. The FBT process is used extensively not only
because of its ready availability and relatively low cost,
but also because of its inherent environmental stability
and versatility.
3 Optimisation methodologies
The optimisation methodologies, neural networks and
desirability functions needed for developing the proposed solution procedure are briefly introduced in this
section.
3.1 Back-propagation neural networks
2 Characteristics and construction of couplers
Branching components (sometimes given the synonyms
couplers and splitters) are passive components with
more than two ports that distribute optical power
among fibres in a predetermined fashion. Wavelength
insensitive couplers are branching components in which
power is routed independently of the wavelength composition of the optical signal. Each component may
combine and divide optical signals simultaneously, as in
bi-directional (duplex) transmission over a single fibre.
However, the wavelength-division multiplexers/de-multiplexers (WDMs) are branching components in which
power is routed based on the wavelength composition of
the optical signal. Passive optical branching components
Fig. 1 a Fabrication of a FBT
device. b Metal tube package
for a FBT device
Neural networks mimic the way by which biological
brain neurons generate intelligent decisions. Numerous
neural network models exist that simulate various aspects of intelligence. To resolve parameter design problems with multiple responses, the back-propagation (BP)
neural networks are applied to construct the functional
relationship between control factors and output responses in an experiment. A standard BP neural model
consists of three or more layers, including an input layer,
one or more hidden layers and an output layer. The
theoretical results revealed that a single hidden layer is
sufficient to allow a BP neural model to approximate
any continuous mapping from the input patterns to the
output patterns to an arbitrary degree of freedom [14].
The training of a BP neural network involves three
898
stages: (1) feed-forward the input training pattern, (2)
associated error calculation and back-propagation, and
(3) weight and bias adjustments. Once network performance is satisfactory, the relationships between input
and output patterns are determined and then the weights
are used to recognise new input patterns. The two
parameters with the greatest effect on the training performance of a BP neural network are the learning rate
and momentum. The learning rate controls the degree of
weight change during training. The momentum avoids
significantly disrupting learning direction when some
training data differ markedly from the majority from
most of the data (and may even be incorrect). A smaller
learning rate and larger momentum reduce the likelihood of the network finding weights that are only a local
minimum, but not a global one [14]. The detailed algorithm of the BP neural network and the guidelines for
selecting appropriate training parameters can be found
in Fausett [14] and Hagan et al. [15].
3.2 Desirability functions
Suppose that there are r output responses y=(y1, y2, ...,
yr), determined by a set of input variables x=(x1, x2, ...,
xp). The general multi-response problem can be defined
as
yj ¼ fj x1 ; x2 ; :::; xp þ ej for j ¼ 1; 2; :::; r
ð1Þ
where fj denotes the response function between the jth
response and the input variables; and j represents the
error term. Usually, the exact form of fj cannot be
known, but can be estimated over a limited experimental
region by using model-building techniques, such as
regression and neural networks. Integrating all the
different responses simplifies such a complicated multiresponse problem as a single objective optimisation
problem. The desirability function approach transforms
an estimated response (e.g. the jth estimated response ^yj )
to a scale-free value dj (0 £ dj £ 1), called desirability.
The larger value of dj increases as the desirability of the
corresponding response increases. Hence, the multi-response problem can be stated as [8]:
maximize k
x
subject to
dj ^yj ðxÞ >k for j ¼ 1; 2; :::; r
x2X
ð2Þ
where t is a constant ()¥<t<¥), called exponential
constant, and z denotes a standardised parameter representing the distance between the estimated response
and its target in units of the maximum allowable deviation. For example, for the nominal-the-best (NTB),
smaller-the-better (STB), and larger-the-better (LTB)
type responses [16], the parameter z can be defined,
respectively, as [8]:
^yj ðxÞ Tj ^yj ðxÞ Tj
z ¼ max
; for yjmin 6^yj ðxÞ6yjmax
ð6Þ
¼
yj Tj
Tj yjmin
z¼
z¼
^yj ðxÞ yjmin
yjmax yjmin
yjmax ^yj ðxÞ
yjmax yjmin
; for yjmin 6^yj ðxÞ6yjmax
ð8Þ
3.3 Proposed optimisation procedure
The proposed procedure for resolving a multi-response
parameter design problem comprises seven steps and is
summarised as below:
Step 1
Step 2
Step 3
ð3Þ
where W denotes the experimental region.
The exponential desirability function is suggested as
follows [8]:
(
exp ðtÞexp ðtj zjÞ
; if t 6¼ 0
exp ðtÞ1
d ð zÞ ¼
ð5Þ
1 j zj; if t ¼ 0
ð7Þ
where the bounds on a response (yminj and ymaxj) should be
specified in advance. The bounds may be determined
according to the specification limits of the product or
process, the regulations or standards of the organisation,
the physical range of the response or the subjective judgments of the decision makers. z ranges between -1 and 1
for an NTB type response, and otherwise ranges between 0
and 1. In either case, the desirability function value d(z)
achieves its maximum value of 1 when z=0. The function
d(z) given in Eq. 5 has been proven to provide a reasonable and flexible representation of human perception [17,
18] and is convenient to handle analytically [8].
Step 4
ð4Þ
; for yjmin 6^yj ðxÞ6yjmax
Step 5
Step 6
Step 7
Identify the quality characteristics (responses),
major control factors, noise factors and exponential constant for each response.
Assign control and noise factors to the
orthogonal arrays; conduct the experiment and
collect the experimental data.
Design a BP neural network to represent the
relationship between input control factors and
output responses.
Present all possible factor level combinations to
the developed network (in step 3) and compute
the estimated responses.
Apply the exponential desirability functions to
transform the multiple responses into an
aggregated performance measure.
Optimise the parameter settings by selecting the
combination that maximizes the overall satisfaction (k).
Conduct the confirmation experiment, and if
the result is unsatisfactory, return to step 1 and
repeat the proposed procedure.
899
Table 1 The specifications of
1% (1/99) single-window
broadband tap couplers, the
exponential constants and
values of yminj and ymaxj
Grade
Premium
A
B
Exponential
constant
yminj
ymaxj
CR (%)
EL (dB)
IL-A (dB)
IL-B (dB)
PDL-A (dB)
PDL-B (dB)
99±0.2
99±0.2
99±0.2
2.5
£ 0.20
£ 0.40
£ 0.60
2
£ 21.50
£ 22.00
£ 23.00
)1
£ 0.20
£ 0.30
£ 0.60
1.5
£ 0.30
£ 0.35
£ 0.40
1
£ 0.30
£ 0.35
£ 0.40
3
98.8
99.2
0.00
0.60
18.00
23.00
0.00
0.60
0.00
0.40
0.00
0.40
4 Case study
may depend on the following process-related control
factors:
This section demonstrates the effectiveness of the proposed procedure using a case study, which was undertaken to optimise the fused process parameters and
hence improve the performance and reliability of the 1%
(1/99) single-window broadband tap coupler.
1.
2.
3.
4.
5.
6.
4.1 Problem encountered
Table 2 lists the experimental levels of the critical
process control factors mentioned above.
The problems encountered in a factory’s mass-production of versatile couplers are machine instability, environmental influences (such as temperature, humidity,
and airflow) and product diversity. In addition, each
machine must be sufficiently stable to copy the optimal
parameter and mass production is ineffective without the
optimal parameter. To apply the proposed procedure to
optimise the parameter settings in the FBT process, the
quality characteristics of interest must be identified first.
Discussion with the personnel managing quality and
reliability engineering identified six crucial quality
characteristics (responses), and these characteristics were
selected herein to enhance quality performance, as follows:
1.
2.
3.
4.
CR (%)
EL (dB)
IL-A (dB)
IL-B (dB)
5. PDL-A (dB)
6. PDL-B (dB)
Coupling ratio (NTB)
Excess loss (STB)
Insertion loss at 1% tap port (STB)
Insertion loss at 99% through port
(STB)
Polarization dependent loss (at 1%
tap port) (STB)
Polarization dependent loss (at 99%
through port) (STB)
The engineering management agreed that convex
exponential desirability functions should be employed
for the responses IL-A, while concave exponential
desirability functions should be employed for the
responses, CR, EL, IL-B, PDL-A and PDL-B. Table 1
lists the specifications of different grades of 1% (1/99)
single-window broadband tap couplers. The table also
lists the exponential constants, and values of yminj and
ymaxj in Eqs. 6, 7, and 8.
Several variables influence the performance of the tap
coupler. Discussion with the product engineer revealed
that tap coupler optical performance in the fused process
DS
PRL
HMF
TH
PHT
HP
Drawing speed
Pre-drawing length
Hydrogen (H2) mass flow
Torch height
Pre-heating time
Hydrogen (H2) pressure
4.2 Experiments and data collection
Six control factors at three levels require 36=729 trials
for a full factorial experiment, a lengthy process. The
main effects of control factors could be accurately estimated by conducting 18 experimental trials arranged
according to a Taguchi L18(21·37) orthogonal array [19].
Hence, the six control factors were assigned to columns
3 to 8 in the Taguchi L18 orthogonal array and Table 3
lists the collected experimental data. Notably, the four
responses, CR, EL, IL-A and IL-B, were collected at
three wavelength levels, namely 1510 nm, 1550 nm, and
1590 nm. Table 3 lists the data for the worst case in the
three wavelength conditions for further analysis.
4.3 Data analysis
The experimental results presented in Table 3 were
analysed using the proposed procedure. Randomly
Table 2 Critical process control factors and their experimental
levels
Control factor
Drawing speed
Pre-drawing length
Hydrogen (H2) mass flow
Torch height
Pre-heating time
Hydrogen (H2) pressure
Level 2 is the existing level
Code
A
B
C
D
E
F
Level
1
2
3
DS1
PRL1
HMF1
TH1
PHT1
HP1
DS2
PRL2
HMF2
TH2
PHT2
HP2
DS3
PRL3
HMF3
TH3
PHT3
HP3
900
Table 3 Collected experimental data
Trial Factor
Response
A B C D E F CR
EL
IL-A
IL-B
PDL-A
PDL-B
Rep. 1 Rep. 2 Rep. 1 Rep. 2 Rep. 1 Rep. 2 Rep. 1 Rep. 2 Rep. 1 Rep. 2 Rep. 1 Rep. 2
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
1
2
3
1
2
3
1
2
3
1
2
3
1
2
3
1
2
3
1
2
3
1
2
3
2
3
1
3
1
2
2
3
1
3
1
2
1
2
3
2
3
1
1
2
3
3
1
2
3
1
2
2
3
1
1
2
3
2
3
1
3
1
2
2
3
1
1
2
3
3
1
2
1
2
3
3
1
2
2
3
1
2
3
1
3
1
2
1
2
3
1
2
3
3
1
2
3
1
2
1
2
3
2
3
1
2
3
1
98.644
98.733
98.798
98.689
98.748
98.747
98.797
98.617
98.738
98.612
98.954
98.779
98.720
98.791
98.662
99.105
98.682
98.775
98.775
98.791
98.728
98.830
98.783
98.817
98.831
98.709
98.783
98.720
98.768
98.759
98.632
98.811
98.793
98.731
98.758
98.613
0.053
0.011
0.060
0.049
0.025
0.059
0.025
0.134
0.045
0.039
0.075
0.038
0.068
0.070
0.190
0.051
0.060
0.061
0.047
0.021
0.084
0.034
0.097
0.017
0.160
0.024
0.056
0.109
0.100
0.022
0.075
0.086
0.083
0.058
0.059
0.390
19.715
20.464
20.287
20.005
20.367
20.211
20.326
19.960
20.135
19.951
20.302
20.227
20.350
20.048
19.772
20.410
19.687
20.314
selecting the training and testing data sets from the
experimental results, a BP neural network model was
constructed to model the functional relationship between input control factors and output responses. A
smaller learning rate and larger momentum are recommended for finding global minimum weights [14], and
thus the learning rate and momentum were set at 0.25
and 0.8, respectively. The candidate neural models were
obtained using the NeuralWorks Professional II/Plus
[20] software, as shown in Table 4. The 6-7-6 neural
Table 4 The candidate BP neural models
Structure
Training RMSEa
Testing RMSEa
6-4-6
6-5-6
6-6-6
6-7-6
6-8-6
6-9-6
6-10-6
6-11-6
6-12-6
0.0732
0.0732
0.0687
0.0642
0.0706
0.0565
0.0621
0.0652
0.0660
0.1208
0.0647
0.0573
0.0494
0.0795
0.0625
0.0828
0.0832
0.0687
a
20.239
20.271
20.201
20.379
20.458
20.584
20.440
20.208
19.964
20.515
20.205
20.173
19.735
20.389
20.094
20.060
20.245
20.128
0.104
0.050
0.103
0.097
0.079
0.101
0.066
0.194
0.100
0.100
0.145
0.091
0.117
0.130
0.248
0.095
0.114
0.106
0.090
0.061
0.139
0.085
0.151
0.054
0.211
0.067
0.109
0.158
0.146
0.071
0.056
0.138
0.136
0.113
0.101
0.443
0.180
0.240
0.310
0.180
0.200
0.490
0.200
0.340
0.270
0.170
0.210
0.360
0.210
0.320
0.290
0.170
0.210
0.300
0.170
0.230
0.280
0.190
0.270
0.410
0.220
0.280
0.250
0.170
0.240
0.390
0.220
0.290
0.280
0.180
0.240
0.280
0.010
0.030
0.020
0.020
0.020
0.030
0.030
0.020
0.020
0.100
0.010
0.030
0.020
0.020
0.030
0.020
0.010
0.030
0.010
0.020
0.020
0.010
0.010
0.020
0.020
0.010
0.010
0.020
0.010
0.020
0.010
0.030
0.030
0.030
0.020
0.020
network model with minimal training and testing
RMSEs was selected based on the table. Through the
well-trained BP neural model, the output responses under all possible control factor parameter combinations
can be accurately predicted. Meanwhile, by applying the
exponential desirability functions with pre-specified
exponential constants in Table 1, multiple responses are
transformed into a single response. Table 5 summarises
five combinations of control factor parameter settings
that produce larger values for the objective function (k)
and the corresponding desirability function (d(z)). Following consultations with engineers, the optimal levels
of control factors were set as A=DS2, B=PRL3,
C=HMF2, D=TH2, E=PHT1 and F=HP3.
4.4 Confirmation experiment
A confirmation was carried out by processing thirty (30)
pieces of 1% (1/99) single-window broadband tap couplers at the optimal parameter levels of control factors.
Table 6 lists the confirmatory results, and indicates
that all of the thirty trials conform to the specification of
1% (1/99) single-window broadband tap couplers.
RMSE: root mean squared error [20]
Table 5 Five combinations of control factor parameter settings that produce larger values for the objective function (k)
No.
1
2
3
4
5
Control factor
k
d(z)
A
B
C
D
E
F
CR
EL
IL-A
IL-B
PDL-A
PDL-B
DS2
DS2
DS3
DS2
DS2
PRL3
PRL3
PRL3
PRL3
PRL3
HMF2
HMF3
HMF3
HMF2
HMF1
TH2
TH3
TH3
TH3
TH2
PHT1
PHT1
PHT1
PHT1
PHT1
HP3
HP3
HP3
HP3
HP2
0.5636
0.4596
0.4972
0.6118
0.5027
0.9860
0.9965
0.9961
0.9787
0.9754
0.4535
0.4440
0.4407
0.4396
0.4392
0.9411
0.9571
0.9555
0.9294
0.9238
0.4658
0.5361
0.6191
0.5741
0.4731
0.9951
0.9951
0.9947
0.9940
0.9945
0.4535
0.4440
0.4407
0.4396
0.4392
901
Table 6 Confirmatory results
Tube no.
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
Response
Quality grade
CR (%)
EL (dB)
IL-A (dB)
IL-B (dB)
PDL-A (dB)
PDL-B (dB)
99.1454
99.1467
98.8475
98.8470
99.1564
98.8429
99.1583
99.1608
98.8367
99.1639
99.1640
98.8763
99.1655
98.8343
99.1657
99.1669
99.1681
98.8297
99.1703
99.1731
98.8259
98.8165
98.8144
98.8123
98.8116
98.8089
98.8084
99.1930
98.8059
98.8053
0.0750
0.0693
0.0397
0.0483
0.1351
0.0931
0.0590
0.0338
0.0917
0.0340
0.0500
0.1817
0.1420
0.1174
0.0499
0.0410
0.0872
0.0925
0.0591
0.0421
0.0592
0.0588
0.0941
0.0245
0.0842
0.0468
0.0325
0.1654
0.0859
0.0521
20.7492
20.7476
20.2982
20.3336
20.8581
20.2487
20.8036
20.7816
20.3342
20.8063
20.8259
20.6937
20.9280
20.6461
20.8220
20.8273
20.8694
20.2815
20.8585
20.8661
20.3308
20.2778
20.3265
20.3064
20.4651
20.1464
20.3034
21.0803
20.3388
20.2461
0.1170
0.1189
0.0850
0.0986
0.1779
0.1396
0.0998
0.0800
0.1378
0.0800
0.0899
0.2308
0.1783
0.1561
0.0912
0.0843
0.1339
0.1343
0.1067
0.0855
0.1002
0.1015
0.1407
0.0764
0.1361
0.0988
0.0789
0.2062
0.1335
0.1019
0.1821
0.1389
0.1533
0.1176
0.1809
0.1023
0.0725
0.1762
0.0977
0.1918
0.1060
0.1483
0.1057
0.1513
0.1174
0.1037
0.1275
0.1266
0.1426
0.1929
0.1553
0.1320
0.1142
0.1239
0.1737
0.1211
0.1616
0.0778
0.1348
0.0889
0.1186
0.0070
0.0105
0.0101
0.0075
0.0074
0.0100
0.0110
0.0114
0.0106
0.0110
0.0094
0.0222
0.0147
0.0038
0.0116
0.0211
0.0104
0.0134
0.0146
0.0090
0.0090
0.0062
0.0153
0.0170
0.0055
0.0116
0.0112
0.0189
0.0076
Moreover, 28 of the 30 couplers are graded as ‘‘Premium’’ and the others are graded as ‘‘A’’. The authors
are confident that the obtained optimal combination of
process control factor parameters can be directly applied
to mass producing fused optical couplers.
4.5 Implementation
The optimal levels of process control factors were
implemented in a pilot run of the fused process in a
phase over 20 days. Evaluation of 200 couplers revealed that the average defect rate was reduced to
2.5%, from over 35% previously. Meanwhile, the
monthly device output from the factory is approximately 10,000 pieces this year. The demand is expected
to grow rapidly in the coming months, with annual
growth of over 50% being assumed. Consequently, this
valuable investigation to optimise the fused process
parameters can not only increase throughput by 30%
through increasing the yield ratio, but can also
increase the price by 25% through producing more
reliable high performance couplers. Given these
achievements, annual savings are expected to reach
USD 500,000–1,000,000, well above the cost of the
experiment, at only around USD 3,000.
Premium
Premium
Premium
Premium
Premium
Premium
Premium
Premium
Premium
Premium
Premium
A
Premium
Premium
Premium
Premium
Premium
Premium
Premium
Premium
Premium
Premium
Premium
Premium
Premium
Premium
Premium
A
Premium
Premium
5 Conclusions
This investigation proposed an integrated procedure
based on neural networks and exponential desirability
functions to resolve the parameter design problem with
multiple responses. Effectiveness of the proposed procedure was demonstrated using a case study which was
undertaken to optimize the fused process parameters
that have been made in the development of FBT couplers to enhance the performance and reliability of the
1% (1/99) single-window broadband tap coupler. A
pilot run of the fused process over 20 days was
implemented and evaluation of 200 pieces of couplers
revealed that the average defect rate reduced to just
2.5%, from over 35% previously. Annual savings from
implementing the proposed procedure are expected to
exceed 0.5–1.0 million US dollars, whereas the expenditure for the experiment was below USD 3,000. This
investigation has been extensively and successfully applied to develop the optimal fuse parameters for other
coupling ratio tap couplers, such as 2/98, 3/97, 4/96, ...,
50/50.
Acknowledgements The authors would like to thank the National
Science Council, Taiwan, R.O.C. for partially supporting this research under Contract No. NSC 91-2213-E-159-013
902
References
1. Phadke MS (1989) Quality engineering using robust design.
Prentice-Hall, New Jersey
2. Logothetis N, Haigh A (1988) Characterizing and optimizing
multi-response processes by the Taguchi method. Qual Reliab
Eng Int 4:159–169
3. Pignatiello JJ (1993) Strategies for robust multi-response
quality engineering. IIE Trans 25:5–15
4. Tong L-I, Su C-T, Wang C-H (1997) The optimisation of multiresponse problems in Taguchi method. Int J Qual Reliab
Manag 14:367–380
5. Antony J (2000) Multi-response optimisation in industrial
experiments using Taguchi’s quality loss functions and principal component analysis. Qual Reliab Eng Int 16:3–8
6. Kaiser HF (1960) The application of electronic computers to
factor analysis. Educ Psychol Meas 20:141–151
7. Lind E, Goldin J, Hickman J (1960) Fitting yield and cost
response surfaces. Chem Eng Prog 56:62–68
8. Kim K-J, Lin DKJ (2000) Simultaneous optimisation of
mechanical properties of steel by maximizing exponential
desirability functions. Appl Stat 49:311–325
9. Harrington E (1965) The desirability function. Ind Qual Contr
21:494–498
10. Derringer G, Suich R (1980) Simultaneous optimisation of
several response variables. J Qual Tech 12:214–219
11. Khuri A, Conlon M (1981) Simultaneous optimisation of
multiple responses by polynomial regression functions. Technometrics 23:363–375
12. Hu A, Corke M, Curley J, Stowe DW (1989) Micro-fused
components for communication and sensor applications. In:
Proceedings of SPIE 1169:552–557
13. Dugan MP (1993) Passive optical branching components:
increasing their reliability. In: Tekippe VJ, Varachi JP (eds)
Passive fibre optic components and their reliability. Proceedings of SPIE 1973, pp 13–27
14. Fausett L (1994) Fundamentals of neural networks: architecture, algorithms, and applications. Prentice-Hall, New Jersey
15. Hagan MT, Demuth HB, Beale M (1996) Neural network design. PWS, Boston, Massachusetts
16. Ross PJ (1996) Taguchi techniques for quality engineering.
McGraw-Hill, New York
17. Kirkwood CW, Sarin RK (1980) Preference conditions for
multi-attribute value functions. Oper Res 28:225–232
18. Moskowitz H, Kim K (1993) On assessing the H value in fuzzy
linear regression. Fuzzy Set Sys 58:303–327
19. Phadke MS, Kackar RN, Speeney DV, Grieco MJ (1983) Offline quality control in integrated circuit fabrication using
experimental design. Bell Syst Tech J 62:1273–1309
20. NeuralWare (2000) NeuralWorks Professional II/Plus v5.4,
NeuralWare Inc, Pennsylvania