Documente Academic
Documente Profesional
Documente Cultură
520 Õ Vol. 122, AUGUST 2000 Copyright © 2000 by ASME Transactions of the ASME
兺 ␦ ⭐⌬
i⫽1
i k (4)
兺 t ⭐T
i⫽1
i k (5)
1
A process is centered when its process mean is identical to the nominal value of
Fig. 3 The process-tolerance-cost relationship the corresponding dimension.
Journal of Manufacturing Science and Engineering AUGUST 2000, Vol. 122 Õ 521
冑兺
I
␦ i2 ⭐⌬ k (6)
i⫽1
3.2 Illustrative Examples
For symmetric tolerances
冑兺
I Example 1. Solve the tolerance synthesis problem 共TSP兲 in
Figs. 1 and 3 with the DOE approach. Minimize the manufactur-
共 t i 兲 2 ⭐T k (7) ing cost subject to the WC and RSS constraints. As five dimen-
i⫽1
sions are to be designed, the number of factors is five, i.e., A, B, C,
In the Monte Carlo simulation approach tolerances are accumu- D, and E. Since two process alternatives for dimensions A, B, D,
lated statistically from the random data of component tolerances. and E are available 共see Fig. 3兲, the number of levels for these
2.7 The Setup Constraint. Setup reduction is key to imple- factors is two. Similarly, the number of levels for factor C is three.
menting JIT production 关34兴, and it can be achieved by process T serves as the tolerance stackup constraint 共see Table 2兲 with
improvement, e.g., by the methods presented in Shingo 关34兴 and T WC computed using the WC model and T RSS using the RSS
Feng and Kusiak 关35,36兴. The conventional models for tolerance model 共see Table 3兲. The cost is the response to be minimized.
synthesis, e.g., Ostwald and Huang 关17兴, Monte and Datseris 关37兴, Equation 共8兲 is used to compute the manufacturing cost for the
and Chase et al. 关19兴, do not consider setup constraints. five dimensions in the assembly chain:
I
y⫽ 兺 c ⫽c
i⫽1
i A ⫹c D ⫹c C ⫹c E ⫹c B (8)
3 Designing the Experiment
3.1 The Methodology. Table 2 shows the full 2 4 ⫻3 1 mixed design data for Example 1
with 48 runs 共no replicates兲. Three alternative optimal solutions
Factors and Levels. In tolerance synthesis, a component di- are found in Table 2:
mension is considered as a factor. The process alternatives applied
关Solution 1兴 Run number 12 共y min⫽610, t A* ⫽t A2 ⫽4, t D * ⫽t D2
to generate the corresponding dimensional tolerances are consid-
ered as different levels. Depending on the number of process al- ⫽3, t C* ⫽t C2 ⫽5, t E* ⫽t E1 ⫽2, t B* ⫽t B2 ⫽4, and the tolerance
ternatives, the number of levels might be two, three, mixed, and stackup is t A ⫹t D ⫽t C ⫹t E ⫹t B ⫽18⭐18兲.
so on. 关Solution 2兴 Run number 14 共y min⫽610, t A* ⫽t A2 ⫽4, t D * ⫽t D1
Replicates. In the DOE approach applied to tolerance synthesis, ⫽2, t C* ⫽t C2 ⫽5, t E* ⫽t E2 ⫽3, t B* ⫽t B2 ⫽4, and the tolerance
first the component tolerances are generated from the Monte Carlo stackup is t A ⫹t D ⫹t C ⫹t E ⫹t B ⫽18⭐18兲.
simulation, then the tolerance stackup is computed using either the
WC or RSS model 共constraints兲. For the tolerance data obtained 关Solution 3兴 Run number 32 共y min⫽610, t A* ⫽t A2 ⫽4, t D * ⫽t D2
from the Monte Carlo simulation, the WC model is used, other- ⫽3, t C* ⫽t C1 ⫽3, t E* ⫽t E1 ⫽3, t *S ⫽t B2 ⫽4, and the tolerance
wise, the RSS model should be applied. stackup is t A ⫹t D ⫽t C ⫹t E ⫹t B ⫽17⭐18兲.
Response. In deterministic tolerance design, a frequently used
Solution 1 and 2 of Example 1 imply that one should order two
objective is to minimize the machining 共assembly兲 cost of com-
different types of bearings. Using one bearing type reduces the
ponents 共assemblies兲. Thus, the response is the total machining
inventory holding cost, the setup cost, and variation due to differ-
共assembly兲 cost. However, in probabilistic tolerance synthesis, the
ent processes or suppliers used. For one bearing type a setup
objective is often to minimize the variation of the tolerance
constraint is imposed as t D ⫽t E . Similarly, it is not cost effective
stackup. Therefore, the response should be the variance of the
to select two different types of manufacturing processes 共machine
longest dimensional chain. Based on the tolerance stackup con-
tools兲 to produce dimensions A and B if the two similar compo-
straint, two responses, the worst case response and the root sum
nents can be produced on one machine. The setup constraint in
square response, are considered. Furthermore, the process shift is
this case is t A ⫽t B . The design and data in Table 3 incorporate
incorporated into each response.
setup constraints. The optimal solution with the WC constraint is
Constraints. The standard DOE approach does not consider any
marked with*, while with the RSS constraint with**. The design
constraints. In order to apply the DOE approach to tolerance syn-
matrix in Table 3 includes those runs of Table 2 that satisfy the
thesis, additional columns need to be incorporated into the facto-
two setup constraints. Therefore, the setup constraint should be
rial design matrix. Figure 4 illustrates the case with multiple di-
considered before the experiments are conducted, which will help
mensional chains. Table 1 shows the constraint columns for the
reduce the number of experimental runs. Selecting run number 32
problem in Fig. 4. Since T 2 is the tolerance stackup in the longest
as the optimal solution with the two setup constraints considered
dimensional chain in Fig. 4, the standard variation of this toler-
corresponds to the following processes:
ance stackup is used as a response 共see Table 1兲.
A ⫽t A2 ⫽4,
t* * ⫽t D2 ⫽3,
tD t C* ⫽t C1 ⫽3,
t E* ⫽t E2 ⫽3, t B* ⫽t B2 ⫽4,
with the tolerance stackup t A ⫹t D ⫹t C ⫹t E ⫹t B ⫽17⭐18, and cost
y min⫽610.
Example 2. For the data in Example 1, minimize the variation
of the assembly tolerance stackup with the DOE approach subject
to setup and cost constraints.
Table 4 shows the process capability data for the process alter-
Fig. 4 Two dimensional chain natives provided in Fig. 3. The mean value C of each centered
Table 3
Journal of Manufacturing Science and Engineering AUGUST 2000, Vol. 122 Õ 523
T2 ⫽ 兺
i⫽1
2
i (9)
The cost constraint is
I
兺 c ⭐Y
i⫽1
i (13)
Table 5
Table 6
achieves a minimum variation design at a lower cost than the where l 1 is the ‘‘as-is’’ run number, and l 2 is the ‘‘to be’’ run
latter 共600 vs. 610兲. If one wants to further reduce the variation, number. For example, from run number 16 to 39 共by tightening
the value of Y should be relaxed. The analysis of variance the variation of dimensions A and B兲 for the centered process, the
共ANOVA兲 is an effective tool in identifying and reducing the variation reduces from 2.82 to 2.22, while the cost increases from
variation. 600 to 640, i.e., VR(16,39)⫽0.6/40⫽0.015. While from run num-
ber 10 to 16 共by tightening the variation of factors D and E兲,
4 Analysis of Experiments VR(16,10)⫽(2.82⫺2.69)/20⫽0.0065. Comparing the two ratios,
tightening the variation of dimensions A and B reduces the overall
4.1 ANOVA for the Centered Process. Table 8 shows the variation 0.015/0.0065⫽2.31 times compared to tightening the
results of the analysis of variance for the simulated tolerance variation of dimensions D and E based on 99 percent confidence
stackup in Table 6. The effect of the process on dimensions A and interval. Similarly, for the skewed process, (VR(16,39)
B in Table 6 is significant at 1 percent, which implies that the ⫽0.01775)/(VR(16,10)⫽0.006)⫽2.9583 based on 95 percent
variation of assembly processes for dimensions A and B should be confidence interval. Therefore, one should select run number 39 if
tightly controlled in order to cost effectively reduce the variation the cost is relaxed to 640 共see Table 7兲.
of tolerance stackup of the shaft assembly.
Example 3. In this example, deterministic data is considered
4.2 ANOVA for the Skewed Process. Table 9 shows the to obtain a set of processes with minimum variation subject to
ANOVA results for the skewed process based on the simulated RSS, setup, and cost constraints. Table 10 shows the experimental
data. The results in Table 9 indicate that the effect of factors A and design data, where the tolerance stackup T C is computed from the
left hand side of inequality 共7兲, cost from Eq. 共8兲, and T2 from Eq.
共9兲.
Table 8 Table 11 includes the experimental design data for minimum
variation subject to setup and cost constraints for a skewed pro-
cess. The tolerance data in each entry of Table 11 is computed
from 共15兲, which in turn can be derived from Eq. 共1兲 for C ⫽0,
and T S is computed from Eq. 共16兲.
t S i ⫽1.5t i (15)
T S l ⫽1.5T l (16)
where: t S l denotes the 3 value with 1.5 process shift for dimen-
sion i in Table 11, T S l is the tolerance stackup with 1.5 process
shift for each process in Table 11, T l is the tolerance stackup with
centered processes in Table 10, and l denotes the run number.
Journal of Manufacturing Science and Engineering AUGUST 2000, Vol. 122 Õ 525
Table 11
5.2 Discussion. The results in Table 10 and 11 concerned ventional three sigma design produces the variation shown in Fig.
with the RSS model based on the deterministic data support the 5共a兲. This design scheme results in a large number of defects and
results in Table 7 from the Monte Carlo simulation to obtain the a significant variation. In zero defect design, the variation is not
tolerance stackup for the random data. The same optimal process significantly reduced 共see Fig. 5共b兲 and Taguchi and Clausing
combination was selected in both cases. A question arises whether 关42兴兲. Two cases of the reduction of variation in six sigma design
to apply the Monte Carlo simulation approach to the random data are analyzed. One case assumes the functional tolerance can not
or to use the RSS model to the deterministic data to obtain a be modified, and one should select a suitable manufacturing pro-
tolerance stackup. For a large number of simulations, the results cess. By selecting a process with six sigma value that is half of
obtained from the Monte Carlo simulation should be more precise each design tolerance, the variation of a component tolerance and
and realistic than those from applying the RSS model to the de- the assembly tolerance stackup is reduced. This approach is rather
terministic data. Furthermore, using the RSS model to the deter- costly. The other case assumes a fixed process capability. One
ministic data 共see Tables 10 and 11兲 provides more conservative should select the full tolerance range of a component twice as
tolerance stackups for the centered process and especially for the large as the six sigma value of the corresponding process. In the
skewed process, than using the Monte Carlo simulation 共see Table latter case, only the number of defects is reduced while the varia-
7兲. Running a large number of simulations for a large number of tion remains unchanged.
dimensions is time consuming; however, it provides more realistic In robust design, both the number of defects and the variation
solutions. are reduced 共see Fig. 5共c兲兲. Since the reduction is based on
Zero defects vs. robust design. The zero defect design and ro- ANOVA, the design tolerance is not evenly relaxed or the process
bust quality design of tolerances are illustrated in Fig. 5. A con- is not evenly tightened. The emphasis is on the critical processes,
Journal of Manufacturing Science and Engineering AUGUST 2000, Vol. 122 Õ 527