Professional Documents
Culture Documents
Use of Orthogonal Arrays in Optimising Wax Coat Thickness For Hearing Aid Shell Manufacture
Use of Orthogonal Arrays in Optimising Wax Coat Thickness For Hearing Aid Shell Manufacture
Page 1 of 13
Conference on the Implementation of New Advances in Engineering Design 2010
to avoid, or take account of, interaction effects between factors. Avoiding interaction affects is known as achieving
additivity, which infers that the main factor effects can be combined in a predictable way.
A
Design Design Design Results
Factor A Factor B Factor C
Exp 1 0 0 0
Exp 2 0 1 1
Exp 3 1 0 1 B
Exp 4 1 1 0
(a) Simple Orthogonal array (L4 OA) C (b) 3D design space
Fig. 1: Principle of orthogonality illustrated for three two-level design factors via an L4 Orthogonal Array
The net result of implementing the simple ‘L4’ OA shown in Figure 1 is that four experiments will be sufficient
to select the best levels for the three design factors as opposed to the eight experiments required with the
conventional one-factor-at-a-time (full-factorial) approach. This saving increases dramatically with more design
factors (Table 1). The four experiments used are not the only orthogonal combination, as can be seen in Fig. 1(b).
Having outlined the search mechanism of an OA, results obtained under ‘compound noise’ conditions CN1 and
CN2 are presented in Table 2 (note that design factor levels are now arbitrarily coded as ‘1’ and ‘2’), along with the
calculated mean.. These compound noise conditions are those due to other factors, typically the uncontrollable non-
design factors such as ambient conditions (e.g. temperature, humidity, wind), which also affect results and cause
common variation. In practice these noise conditions are promoted through repetitions over a period of time, i.e.
repeated cycles through the four experiments.
Page 2 of 13
Conference on the Implementation of New Advances in Engineering Design 2010
From the OA results data, the mean value for each factor level is displayed in a convenient format for
comparison purposes - the response table (Table 3). For example, the average value of the results when factor A is
at level 1 is the average of Exp 1 (2.75) and Exp 2 (3.25), which equals 3.
In Table 3 the optimum level for each factor is that with the highest mean response (underscored) and a ranking
is made on the basis of the difference between the mean responses at each level. The differences are the effect a
factor level has when changed despite the other factor levels changing too. It is apparent from the difference values
that factor A and factor B individually have considerably greater main effects on the response than factor C. These
main factor effects can be represented graphically as a response graph (Figure 2), in which these effects are more
readily compared.
5
4
3
2 factor A
factor B
1
factor C
0
level 1 level 2
Fig. 2: Response graph
Assuming additivity has been maintained in this example, the optimum factor settings for the highest response
will be A1B1C2, which may or may not be a combination that was tested in the experiment (a 50% chance for this
L4). Each mean design factor effect and the overall mean are then be used to predict the expected optimum value of
y, E(y) using Equation 1:
E ( y ) = A1 + B1 + C 2 − 2 y() (1)
Therefore E(y) will be 3 + 3.75 + 4 – 2(2.9375) = 4.875, which is a better mean performance than observed in the
experiments.
As the L4 OA employs two levels per design factor then only linear relationships have been identified, which
means that if we attempt to interpolate within the design factor levels (Fig. 2) there is a risk of making erroneous
predictions. Thus the above example has been used to decide which design factor levels are the best to use. In
practice, ‘cheap’ two-level OA are most often used as a decision tool, also known as a screening experiment, where
good (i.e. significant) design factors are screened out for more in-depth investigation with higher order OA that can
handle design factors with three-levels or more.
Page 3 of 13
Conference on the Implementation of New Advances in Engineering Design 2010
Page 4 of 13
Conference on the Implementation of New Advances in Engineering Design 2010
The ‘saturated’ design of an L9 OA means that all its degrees of freedom are used up on main effects if
factors are allocated to all four columns. It is possible to observe an interaction between two columns if required but
this will employ two columns and render them unavailable for observing the main effects of two other factors and so
tends to be avoided. Therefore, the L9 OA will be used here on the assumption that the factors to be chosen for
experimentation will not have significant interaction, as was described in the introduction.
The method of calculating the mean response of each of the three levels for any factor is similar to that for a
two-level factor (as was shown in Table 3). Developing on from Equation 1, the predicted performance for an L9
will be:
E ( y ) = AL + BL + C L + DL − 3 y () (3)
SNR values can be used in Equation 3 if a prediction is desired in these terms.
Where L is the number of factor levels and Āj = mean response of factor A for level j.
n
ST = ∑ ( yi − y )
2
(5)
i =1
The variability of data that cannot be attributed to design factors will be assumed to be due to error (experimental,
measurement and other unexplained variability):
Sum of squares due to error, Se = ST-ΣSA (6)
The degrees of freedom (dof) for a design factor are equal to its number of levels minus one, i.e. 2, and the total
degrees of freedom for the experiment will be 9-1=8, and using this:
dof for error, fe = fT-ΣfA (7)
Variance is a measure of variability per degree of freedom:
VT = ST/fT (8)
VA = SA/fA (9)
Ve = Se/fe (10)
Note that Ve ≠ VT-ΣVA.
The ratio of the factor variance to error variance is then used as a test of the factor effect:
FA = VA/Ve (11)
An F value of less than one suggests that the associated design factor effect is less than the error of the
additive model. An F value greater than two means the design factor effect is not small, and F larger than four
means the design factor effect is large (Phadke, 1989).
Page 5 of 13
Conference on the Implementation of New Advances in Engineering Design 2010
To quantify the contribution of a design factor to the total response variability, the sum of squares must be
adjusted to remove the error contribution. Considering Equation 9, if VA is small then factor A is considered to have
no significant effect on the response variability, and the variability attributed to error is VA=Ve, which can be added
to that of other weak factors. Therefore Equation 10 becomes Ve=SA/fA, the error variation component, for a
‘saturated’ Orthogonal Array (on in which all columns are allocated a design factor) which means that SA=fAVe.
Equation 6 indicates how Ve is then pooled for more than one design factor. Pure sum of squares for A becomes:
S'A = SA - fAVe (12)
The percent contribution of design factor A to the total variability is then:
pc (%) = S'A/ST *100 (13)
If the percentage contribution due to error is below 15% it is unlikely that important design factors were omitted
from the experiment. If the percentage contribution due to error is over 50% then some important design factors
were likely to have been missed from the experiment.
Page 6 of 13
Conference on the Implementation of New Advances in Engineering Design 2010
Several samples were taken on different days for each experiment in order to include noise factors such as
ambient conditions and unknown operational factors. The results are shown in Table 6.
Table 6: L9 OA and results.
L9 OA N1 N2 N3
A B C D mean stdev SNR
1 1 1 1 0.15 0.20 0.18 0.25 0.23 0.18 0.13 0.21 0.30 0.20 0.05 11.85
1 2 2 2 0.22 0.18 0.20 0.18 0.17 0.19 0.22 0.19 0.19 0.19 0.02 20.95
1 3 3 3 0.17 0.16 0.16 0.17 0.16 0.16 0.14 0.12 0.18 0.16 0.02 18.92
2 1 2 3 0.17 0.16 0.28 0.16 0.19 0.21 0.13 0.23 0.19 0.19 0.04 12.65
2 2 3 1 0.37 0.35 0.38 0.23 0.24 0.22 0.42 0.29 0.30 0.31 0.07 12.65
2 3 1 2 0.25 0.16 0.23 0.23 0.16 0.19 0.15 0.14 0.19 0.19 0.04 13.52
3 1 3 2 0.14 0.16 0.15 0.14 0.13 0.15 0.14 0.16 0.10 0.14 0.02 17.73
3 2 1 3 0.46 0.45 0.47 0.66 0.53 0.53 0.51 0.57 0.65 0.54 0.08 16.83
3 3 2 1 0.10 0.12 0.09 0.11 0.11 0.10 0.14 0.12 0.17 0.12 0.02 13.68
Table 8 shows the factor levels for best SNR are D2 A1 B2 C3 in rank order.
Table 8: Response table for SNR.
1 2 3 diff
A 17.24 12.94 16.08 4.30
B 14.07 16.81 15.37 2.74
C 14.07 15.76 16.43 2.36
D 12.73 17.40 16.13 4.67
The response graphs, Figures 4 and 5, present the results of Tables 7 and 8 graphically:
Page 7 of 13
Conference on the Implementation of New Advances in Engineering Design 2010
0.40
0.35
0.30
0.10
0.05
0.00
1 2 3
Factor Level
20.00
18.00
16.00
Wax coat thickness SNR (dB)
14.00
12.00 A
B
10.00
C
8.00 D
6.00
4.00
2.00
0.00
1 2 3
Factor Level
Aiming to minimise variance about a target thickness of 0.20mm and exploring predictions for various factor level
settings (Eq. 3) finds a suitable design of A1 B2 C3 D2:
Predicted mean for A1 B2 C3 D2 = 0.18 + 0.35 + 0.20 + 0.17 – 3(0.23) = 0.21mm
Predicted SNR for A1 B2 C3 D2 = 17.24 + 16.81 + 16.43 + 17.40 – 3(15.42) = 21.62 dB
Conducting ANOVA (Table 9) highlights that the mean effects of A and D are indistinguishable from error.
Table 9: ANOVA.
ANOVA-means f S V F S' pc(%)
A 2 0.01 0.00
B 2 0.07 0.03 4.02 0.05 38
C 2 0.03 0.02 2.00 0.02 12
D 2 0.02 0.01
(e) 4 0.03 0.01 50
T 8 0.13 0.02 100
Page 8 of 13
Conference on the Implementation of New Advances in Engineering Design 2010
0.00
-1.00 -0.80 -0.60 -0.40 -0.20 -0.200.00
-0.40
y = 1.0393x - 0.7439
-0.60
log stdev
-0.80
-1.00
-1.20
-1.40
-1.60
-1.80
-2.00
log mean
From Fig. 6 the slope of the line is very close to 1, which is a strong endorsement (Box et al, 1978) of a log-
based transformation like the SNR for this problem. The use of the SNR has been questioned in the literature (Box,
1988; León et al, 1987; Logothetis, 1990), particularly in regard to its theoretical foundations, e.g. its effect on
additivity. The Box-Cox transformation (Box, 1988) has been proposed as a more rigorous approach that determines
the transformation with the minimal error sum of squares (Eq. 6) via an iterative process. However, it seems to be
Page 9 of 13
Conference on the Implementation of New Advances in Engineering Design 2010
used rarely in practical engineering problems presumably because its greater analytical complexity is too
burdensome.
ANOVA has identified factor B as contributing 38% of the observed effects on coating thickness and factor C as
contributing 12%. The commensurate error contribution of 50% obtained by the Taguchi pooling up strategy
indicates that other significant but unobserved factors have affected results. As only four design factors are observed
by an L9 then pooling two into error in this way oversimplifies the prediction equation (Eq. 3), which means it is
difficult to choose design factor settings that produce an acceptable coating thickness. Hence it has been abandoned
for this case. Graphical methods such as probability plots should have been used for comparison. Furthermore, the
dynamic output characteristic (Taguchi, 1987) would potentially produce results that mean optimum settings for any
desired thickness would be identified.
The original process settings A1 B1 C1 D1 produced a mean wax coat thickness of 0.20mm and SNR of 11.85
dB, compared to 0.21mm and 21.62 dB for the optimal process design A1 B2 C3 D2. The equivalent values of
standard deviation can be found by rearranging the SNR equation (Eq 2):
y2
σ= (14)
anti log⎛⎜
SNRNB ⎞
⎝ 10 ⎟⎠
Thus the optimal process settings produce a standard deviation of 0.019mm compared to 0.051mm for the
original settings, and if we assume tolerance limits of ± 0.1mm then the conventional ± 3σ process limits will be ±
0.057mm for the optimal settings, i.e. well within tolerance, compared to ± 0.153mm for the original settings, i.e.
well outside tolerance. Furthermore, if we look at the potential effect on production of this improved capability by
calculating the probability density under a Normal distribution beyond the tolerance limits in each case, then the
original process settings produce in excess of 9% whereas the optimal settings reduce this to less than 0.002%. This
is potentially a major process improvement identified by this optimisation method.
Page 10 of 13
Conference on the Implementation of New Advances in Engineering Design 2010
ANSYS DesignXplorer (DX) is a design optimization module within ANSYS Workbench, the framework that
guides the user through the entire simulation process. DX has a range of DoE tools but does not specifically use
Orthogonal Arrays, providing instead higher order experimental designs. It also uses Variational Technology (VT),
which can explore the design space more quickly than DX over the tolerance values of the design parameters
selected in order to evaluate the robustness of the artefact. Multiple objectives and constraints can be handled in
order to identify trade-off solutions.
Page 11 of 13
Conference on the Implementation of New Advances in Engineering Design 2010
Noise Factors, z
Design Factors, x
Figure 7: Product parameters
Most engineering products and processes are dynamic in nature, even though they may have been designed as
static systems. Active dynamic systems are those that require adjustment as part of their function, e.g. a carburettor.
Passive dynamic systems are those that receive a signal and reproduce it with minimum error, e.g. radio receiver.
For complicated products even if the overall system does not require direct adjustment, its sub-systems will need to
be adjustable to optimise the product. Similarly, processes need to be adjustable for adapting to new materials or
products.
A three-level signal factor can be included in the format of an L8 OA as shown in Table 8.
Table 8: L8 OA with 3-level signal factor and compounded noise.
OA Results S/N
factors M1 M2 M3 Ratio
A B C D E F G N1 N2 N1 N2 N1 N2
1 1 1 1 1 1 1
1 1 1 2 2 2 2
1 2 2 1 1 2 2
1 2 2 2 2 1 1
2 1 2 1 2 1 2
2 1 2 2 1 2 1
2 2 1 1 2 2 1
2 2 1 2 1 1 2
There is insufficient space here to explain how the analysis of a dynamic characteristic is carried out (see Fowlkes
and Creveling, 1995).
Page 12 of 13
Conference on the Implementation of New Advances in Engineering Design 2010
References
Journals:
Box, G.E.P. (1988) Signal-to-noise ratios, performance criteria and transformations, Technometrics, Vol 30(1),
February 1988, pp 1-40.
Coleman, D.E., Montgomery, D.C. (1993) A systematic approach to planning for a designed industrial experiment,
Technometrics, Vol 35(1), February 1993, pp 1-12.
Hamada, M. & Wu, C.F.J. (1995) The treatment of related experimental factors by sliding levels, Journal of Quality
Technology, Vol. 27(1), pp 45-55.
Huang, M.C., Tai, C-C. (2001) The effective factors in the warpage problem of an injection-moulded part with a
thin shell feature, Journal of Materials Processing Technology, Vol 110, pp 1-9.
León, R.V., Shoemaker, A.C., Kacker, R.N. (1987) Performance Measures Independent of Adjustment,
Technometrics, Vol 29(3), August 1987, pp 253-265.
Logothetis, N. (1990) Box-Cox transformations and the Taguchi Method, J. of Applied Statistics, Vol 39(1), pp 31-
48.
Books:
Box, G.E.P., Hunter, G.H. & Hunter, J.S. (1978) Statistics for Experimenters. John Wiley & Sons.
Fowlkes, W.Y., Creveling, C.M. (1995) Engineering Methods for Robust Product Design. Addison-Wesley.
ISBN:0-201-63367-1.
Hedayat, A.S., Sloane, N.J.A., Stufken, J. (1999) Orthogonal Arrays: theory and applications. Springer. ISBN: 978-
0-387-98766-8.
Montgomery, D.C. (2004) Design and Analysis of Experiments. 6th Edition. Wiley. ISBN: 978-0-471-48735-7.
Phadke, M.S. (1989) Quality Engineering Using Robust Design, Prentice Hall. ISBN: 013745167-9.
Taguch, G. (1987) System of Experimental Design (Vols 1 and 2), UNIPUB/Krauss International. ISBN: 978-0-527-
91621-3.
Taguchi, G, Chowdury, S., Taguchi, S. (2000) Robust Engineering, McGraw-Hill. ISBN: 0-07-134782-8.
Websites:
(Accessed 29th January 2010)
http://www.freepatentsonline.com/4652414.pdf
http://www.enotes.com/how-products-encyclopedia/hearing-aid
http://www.hearingreview.com/issues/articles/2004-03_03.asp
http://www.mathworks.com/
http://www.ansys.com/products/workbench/designxplorer.asp
Page 13 of 13