Professional Documents
Culture Documents
Maximizing Product,
Minimizing Costs
n many textbook scenarios, an engineer will study a process via a
A CASE STUDY
ON OPTIMIZING I designed experiment and use the results to optimize the product
response according to a loss function. Responses might include yield,
weld or seal strength, or perhaps, the diameter of a drilled hole.
MULTIPLE PRODUCT However, in real-world situations, the quality practitioner often has sev-
eral product responses to consider simultaneously. For example, the
RESPONSES WHILE practitioner might be interested in adjusting not only a products length,
CONTROLLING but also its weight to respective target values.
Often, this optimization must be performed under competing require-
OPERATING COSTS ments. For example, increasing factor A might move product length clos-
er to its target value, but it also might negatively impact the weight
response. Add to this already murky mix a management imposed
requirement to hold operating costs to a minimum, and suddenly the
textbook scenarios are out the window.
In this article, we present a case study adapted from an actual experi-
ment performed on a molding process in which several product charac-
teristics were optimized simultaneously, while the operating cost of the
machine was minimized. Data and some of the variable names have been
By Marina L. Lopez, changed to protect confidentiality.
.. The techniques used in the case study were fractional factorial design,
INSIC and TU V regression modeling and simultaneous optimization using Derringer
Rheinland Argentina and Suichs desirability function.1 Analysis was performed using Minitab
statistical software and the solver function in Excel.
S.A., and Mar y
Process Overview
McShane-Vaughn,
The plastic molding machine used in the experiment consist of a mold
Southern Polytechnic with 100 cavities. The mold cavities were filled with raw material powder
and then vibrated for a certain period of time. Next, the mold was closed,
State University and the heating and cooling cycle began according to the set soak times.
After the parts were cooled down, they were unloaded from the machine
and measured. The product characteristics of interest included length and
weight of the plastic part.
Experimental Design
18 I F E B R U A R Y 2 0 0 8 I W W W . A S Q . O R G
Maximizing Product, Minimizing Costs
S I X S I G M A F O R U M M A G A Z I N E I F E B R U A R Y 2 0 0 8 I 19
Maximizing Product, Minimizing Costs
time, pressure and amplitude main effects. From the Weight Model
constant term, the average expected length using the
current factor settings equalled 0.366 inches. From the constant term, the average expected
The vibration amplitude x material density interac- weight achieved at the current operating conditions
tion coefficient showed that when these two factors was equal to 0.8568 grams. The interactions vibration
are set at higher-than-current or lower-than-current pressure x density and vibe amplitude x density were
levels, expected length will increase. In addition, both significant at the alpha = 0.05 level. When both
increasing the main effects of vibration time and factors in each of these interactions were set to lower-
pressure increases expected length. The final analysis than-current or higher-than-current settings, expect-
of variance table and fitted model are shown in ed weight increased. In addition, as the main effects
Figure 5. of vibration time and quantity increased, expected
weight increased. The final analysis results are shown
in Figure 6.
20 I F E B R U A R Y 2 0 0 8 I W W W . A S Q . O R G
Maximizing Product, Minimizing Costs
S I X S I G M A F O R U M M A G A Z I N E I F E B R U A R Y 2 0 0 8 I 21
Maximizing Product, Minimizing Costs
Finally, for s > 1, t > 1, there was a more severe desir- $90 corresponds to the cost at current operating con-
ability penalty as the response moved from the target ditions. The resulting desirability functions for
value. length, weight and cost, respectively, are shown in the
If the target for the response was a minimum value following formulas:
(smaller-the-better objective), as in the case of the
cost function, the desirability function was calculated 0 if y 1 (x ) < 0.350
y 1 (x ) 0.350
as: 0.05 if 0.350 y 1 (x ) < 0.400
d1 (y 1 ) = 0.450 y 1 (x )
1 if y i (x ) < Ti
0.05 if 0.400 y 1 (x ) < 0.450
s
U i y i (x )
di (y i ) = U i Ti if Ti y i (x ) < Ui 0 if y 1 (x ) 0.450
0 if y i (x ) > =Ui
0 if y 2 (x ) < 0.820
y 2 (x ) 0.900
in which T represented a value small enough to be 0.08 if 0.820 y 2 (x ) < 0.900
acceptable for the response, and U was a value above d2 (y 2 ) = 0.980 y 2 (x )
0.08 if 0.900 y 2 (x ) < 0.980
which the response was considered unacceptable.
After choosing s and t to determine shape, and put- 0 if y 2 (x ) 0.980
ting in the appropriate values for T, L and U, the
combination of input factor levels that maximizes the
1 if y 3 (x ) < 0
overall desirability function D = (d1 . d2 . d3 dm)1/m 90
could then be found. More information and exam- d3 (y3 ) = y 3 (x )
90 if 0 y 3 (x ) < 90
ples on the desirability function technique is avail- 0 if y3 (x ) 90
able in Response Surface Methodology.4
For this analysis, the exponents s and t were each
set equal to one. Lower, target and upper values for Overall Desirability Results
the length and weight desirability functions corre-
spond to the lower specification, spec nominal and We used the solver function in Excel to find a com-
upper specification values, respectively. bination of factor levels that produced a large value
For the cost desirability function, the target value for desirability D. Because the objective function, D,
was set to a stretch goal of $0, and the upper limit of was nonconvex, solutions to the optimization were
dependent on the starting values of the xs.
Figure 8. Comparison of Response Results After performing several trials with various
For Current and New Settings starting points, we found a combination of xs
that brought both weight and length close to
Current New New settings, New settings,
Factor
settings settings ignoring cost 1 ignoring cost 2
target at a small cost.
Vibration time 60 40 49 80 This value for D was achieved at the factor
Vibration pressure 45 50.5 46.5 32 settings of vibration time = 40 seconds, vibra-
Vibration amplitude 85 95 0.95 0.95 tion pressure = 50.5 pounds per square inch
Material density 0.40 0.41 0.41 0.40 (psi), vibration amplitude = 95%, material
Quantity 3.5 2.5 2.6 2.6 density = 0.41g/ml and quantity = 2.5 scoops.
Response
Using these settings, average expected length
Length 0.3660 0.3872 0.4000 0.4000
equalled 0.387 inches (target 0.400), average
Weight 0.8568 0.9000 0.9000 0.9000
Cost $90 $45 $67 $179 expected weight equalled its target of 0.90
grams, and the scrap and maintenance cost
22 I F E B R U A R Y 2 0 0 8 I W W W . A S Q . O R G
Maximizing Product, Minimizing Costs
due to vibration was cut in half, from $90 to $45. tion. By ignoring these costs, product targets might
It is interesting to note that if we had ignored the be met, but operating expenses might increase.
cost dimension of the problem, an optimization of
the length and weight functions would have achieved REFERENCES
target values for both length and weight, but would
1. George Derringer and Ronald Suich, Simultaneous Optimization of
have resulted in increased scrap and maintenance Several Response Variables, Journal of Quality Technology, Vol. 12, No. 4,
costs. 1980, pp. 214-219.
For example, we found various sets of factor com- 2. D.C. Montgomery, Design and Analysis of Experiments, sixth edition, John
Wiley & Sons, 2005.
binations that achieved the length and weight targets,
3. Derringer and Suich, Simultaneous Optimization of Several Response
with associated costs ranging from $67 to $179. A Variables, see reference 1.
comparison of optimal settings with the current oper- 4. R.H. Myers and D.C. Montgomery, Response Surface Methodology, John
ating conditions is shown in Figure 8. Wiley & Sons, 1995, pp. 244-260.
Eye on Costs
S I X S I G M A F O R U M M A G A Z I N E I F E B R U A R Y 2 0 0 8 I 23