Professional Documents
Culture Documents
P =
[ ]
1/2
2/3
i
ASM
D
T
(7)
Then compute the component tolerances:
T
1
= P D
1/3
1
, T
2
= P D
1/3
2
, etc.
Example 2. Statistical Allocation by Precision Factor.
Compute the assembly tolerance for the shaft/housing assembly by a statistical sum:
T
ASM
2
= T
A
2
+ T
C
2
+ T
F
2
+ T
B
2
+ T
D
2
+ T
E
2
+ T
G
2
.015
2
= (.0015
2
+.0025
2
+.0025
2
) + P
2
( 8.0
2/3
+
.400
2/3
+ 7.711
2/3
+.400
2/3
)
Again, the fixed tolerances are subtracted from the assembly tolerance before computing
the precision factor.
Solving for the precision factor: P = .004836
9
Re-allocating:
T
B
= .004836 (8.00)
1/3
= .00976 T
E
= .004836 (7.711)
1/3
= .00955
T
D
= .004836 (.400)
1/3
= .00356 T
F
= .004836 (.400)
1/3
= .00356
The Precision Factor method is similar to the Proportional Scaling method, except there is
no initial allocation required by the designer. Instead, the tolerances are initially allocated
according to the nominal size of each component dimension, then scaled to meet the
specified assembly tolerance. This procedure could also be followed assuming a worst
limits sum for the assembly tolerance (equation 1). The results are summarized in Table
1.
Table 1. Comparison of Allocation Methods
Proportional Precision Factor
Original Worst Stat Worst Stat
Part Tolerance Case 6 Case 6
.0015 .0015 .0015 .0015 .0015
.008 .00378 .01116 .00312 .00967
C .0025 .0025 .0025 .0025 .0025
D .002 .00094 .00279 .00115 .00356
E .006 .00283 .00837 .00308 .00955
F .002 .00094 .00279 .00115 .00356
G .0025 .0025 .0025 .0025 .0025
ASSEMBLY TOL. .0150 .0150 .0150 .0150
PROP. FACTOR .472221.39526 .00156 .004836
* Fixed tolerances
5. Tolerance Allocation Using Optimization Techniques
A promising method of tolerance allocation uses optimization techniques to assign
component tolerances such that the cost of production of an assembly is minimized. This is
accomplished by defining a cost-vs.-tolerance curve for each component part in the
assembly. The optimization algorithm varies the tolerance for each component and
searches systematically for the combination of tolerances which minimizes the cost.
Figure 6 illustrates the concept simply for a three component assembly. Three cost-vs.-
tolerance curves are shown. Three tolerances (T
1
, T
2
, T
3
) are initially selected. The
corresponding cost of production is C
1
+ C
2
+ C
3
. The optimization algorithm tries to
increase the tolerances to reduce cost, however, the specified assembly tolerance limits the
tolerance size. If tolerance T
1
is increased, then tolerance T
2
or T
3
must decrease to
keep from violating the assembly tolerance constraint. It is difficult to tell by inspection
which combination will be optimum. The optimization algorithm is designed to find it with
a minimum of iteration. Note that the values of the set of optimum tolerances will be
different when the tolerances are summed statistically than when they are summed by worst
limits.
10
a. Cost-vs.-Tolerance Functions.
The key factor in optimum tolerance allocation is the specification of cost-vs.-tolerance
functions. Several algebraic functions have been proposed, as summarized in
Cost
Tolerance
[Worst Case]
[Statistical]
Total Cost:
Fig. 6. Optimal tolerance allocation for minimum cost.
Table 2. The constant coefficient A may include setup cost, tooling, material, prior
operations, etc. The B term determines the cost of producing a single component
dimension to a specified tolerance.
Table 2. Proposed Cost-of-Tolerance Models
Cost Model Author Ref
Reciprocal Squared A + B/tol
2
Spotts [3]
Reciprocal A + B/tol
Chase&Greenwood [4]
Exponential A e
-B(tol)
Speckhart [5]
Little has been done to verify the form of these curves. Manufacturing cost data are not
published since they are so site-dependent. Even companies using the same machines
would have different costs for labor, materials, tooling and overhead.
Jamieson [6] reported a government study in which relative costs were determined for
actual parts for several metal-removal processes. This seems to be the same data presented
11
as a case study by Trucks [2]. Jamieson correlated the process/cost results with the
tolerance-vs.-size chart of figure 4. This data was curve fit by regression analysis using
each of the models shown in Table 2. Typical results are shown in figure 7. The
reciprocal tolerance curve appears to fit the machining process data the best.
0
1.0
2.0
3.0
4.0
5.0
6.0
0 0.001 0.002 0.003 0.004 0.005 0.006 0.007 0.008
Tolerance(in.)
Grinding
Reaming
Turning
A + B
tol
2
1
A + B
tol
1
A e
B (tol)
Machine Processes
CurveFits:
Fig. 7. Comparison of cost-vs.-tolerance models.
b. Tolerance Allocation by Lagrange Multipliers
A closed-form solution for the least-cost component tolerances was developed by Spotts
[3]. He used the method of Lagrange Multipliers, assuming a cost function of the form
C=A+B/tol
2
. Chase and Greenwood [4] extended this to cost functions of the form
C=A+B/tol as follows:
function) (Cost
T
i
+ t) (Constrain
T
i
= 0 (i = 1, . . n)
)) /T B (A (
T
j i i
i
+
+ ) T (
T
2
j
i
= 0 (i = 1, . . n)
= 1,..n) (i
2T
/T B
i
2
i i
12
Eliminating by expressing it in terms of T
1
:
T
i
=
_
B
i
B
1
1/3
T
1
(8)
Substituting into the assembly tolerance sum:
T
2
ASM
= T
1
2
+
_
B
i
B
1
2/3
T
1
2
T
1
=
1/2
2/3
1
i
ASM
B
B
1
T
1
1
1
1
1
]
1
,
_
+
(9)
Substitute this result in equation 8 to obtain the minimum cost tolerances. The numerical
results for the example problem are shown in Table 3. The Setup Cost is coefficient A in
the cost function. The Reference Cost and Reference Tolerance are used to compute
coefficient B.
Table 3. Minimum Cost Tolerance Allocation
Allocated Tolerances
Tolerance Cost Data A+B/tol A+B/tol
2
Setup Reference Reference Worst Stat. Worst Stat.
Part Cost Cost Tolerance Case 6 Case 6
A $ .15 - .0015* .0015 .0015 .0015 .0015
B 4.75 $1.20 .008 .00241 .00789 .00305 .00936
C 2.50 - .0025* .0025 .0025 .0025 .0025
D 2.80 2.75 .002 .00207 .00711 .00159 .00576
E 4.50 0.88 .006 .00194 .00683 .00227 .00750
F 2.80 2.75 .002 .00207 .00711 .00159 .00576
G 2.50 - .0025* .0025 .0025 .0025 .0025
Assembly Cost $27.58 $28.98 $26.05 $43.11 $22.10
Acceptance Fraction 1.000 .9973 1.000 .9973
True Cost $28.98 $26.12 $43.11 $22.16
*Fixed tolerance
Parts A, C and G are vendor-supplied. Since their tolerances are fixed, their cost cannot be
changed by re-allocation, so no reference cost was required.
The "True Cost" is just the total cost of the assembly divided by the acceptance fraction or
yield. Thus, the cost is adjusted to include a share of the cost of the rejected assemblies. It
does not include any parts which might be saved by re-work or the cost of rejecting
13
individual component parts. Vendor-supplied parts may be excluded by assigning them a
zero setup cost. An interesting exercise is to calculate the optimum acceptance fraction,
that is, the rejection rate which would result in the minimum True Cost. This requires an
iterative solution. For the example problem, the results are shown in Table 4:
Table 4. Minimum True Cost
Cost Model True Cost Accept Fract Z
A + B/tol 26.10 .99477 2.789
A + B/tol
2 21.71 .97066 2.178
The corresponding component tolerances would increase by the constant factor Z-
i
/Z- .
The advantages of the Lagrange multiplier method are:
1) It eliminates the need for iterative solutions,
2) It can handle either worst limit or statistical assembly models,
3) It allows alternative cost-tolerance models.
The limitations are:
1) Tolerance limits cannot be imposed on the processes. Most processes are only
capable of a specified range of tolerance.
2) It cannot readily treat the problem of simultaneously optimizing interdependent
tolerance stacks. That is, when an assembly has more than one tolerance stack,
with common component dimensions appearing in each stack, an iterative solution
of a nonlinear set of equations is required.
Problems which exhibit these characteristics may be optimized using nonlinear
programming techniques.
6. Limitations of Common Assembly Models
The common models for assembly tolerance accumulation have distinct limitations when
applied to tolerance allocation:
1) The Worst Limit model results in component tolerances which are tight and
costly to produce.
2) Statistical models allow looser tolerances, but often predict higher assembly
yields than actually occur in production.
3) For assemblies with small numbers of components, or which have one
component tolerance which is much greater than the remaining components, the
common Statistical model can give component tolerances which are tighter than if
computed by a Worst Limit model.
14
4) Statistical models assume manufacturing variations follow a Normal or classic
bell-shaped distribution, symmetrically positioned at the midpoint of the tolerance
limits. They do not take into account possible skewness or bias which are common
in manufactured parts. Figure 8 illustrates the unexpected rejects which can occur
when skewness and bias are not accounted for.
In an earlier paper by the authors, these limitations were discussed with several examples
[7].
Bias results in a shift in the nominal dimension. It is particularly harmful, since it can
accumulate in an assembly and cause unexpectedly high rejection rates. Virtually all
manufacturing processes exhibit bias. Some processes are more prone to bias than others.
Bias can occur from tooling or fixture errors, setup errors or tool wear. It may be
deliberately introduced during setup to compensate for tool wear or to allow for re-work.
It may naturally occur in a process as in thermal shrinkage of molded parts. Bias is as
critical a factor in an assembly model as is the process capability or variance.
Assumed
RSS
Reality
Mean Shift
Component
1
Component
2
Assembly
Sum
Component
3
Unexpected
Rejects
Fig. 8. Ideal vs. actual assembly distributions.
15
7. Estimated Mean Shift Model
A new model for assembly tolerance accumulation has been proposed which includes an
estimate of expected bias . It is called the Estimated Mean Shift method because the
designer must estimate the bias for each component in an assembly. This is done by
defining a zone about the midpoint of the tolerance range as shown in figure 9, which is
the probable location of the mean of a typical batch of parts. The midpoint tolerance zone
is expressed as a fraction of the specified tolerance range for the part dimension, (a number
between 0 and 1.0). If the process to be used to produce the part is closely controlled, a
low mean shift factor may be selected, say 0.1 to 0.2. For less well-known processes, such
as a part supplied by a new vendor, a larger factor, say 0.7 or .8, could be selected to
account for the uncertainty. For common processes the factor could be selected on the
basis of prior history from quality assurance data.
Probable Location
of the Mean
MIDPOINT LTL UTL
TOLERANCE
Fig. 9. The location of the mean is not known precisely.
Once the range of mean shift has been estimated for each component the assembly
tolerance is calculated using the following model:
T
ASM
= m
i
i
x
f
T
i
+
[
,
_
2
i
i
2
2
i
T
x
f
) m (l
]
1/2
(10)
where m
i
is the mean shift factor for the ith component. The assembly tolerance in
equation 10 has been split into two summations. The first term is the sum of the mean
shifts added as a worst limit. The second term is the sum of the component tolerances
added statistically. Thus we have the contributions of both mean shift or bias and part
tolerance or variance on the resulting assembly tolerance. A more complete discussion of
this model, with examples, may be found in reference 7.
16
Note the special case when all the mean shifts are chosen as zero. The resulting assembly
tolerance reduces to the simple statistical model (equation 4). If all mean shifts are chosen
equal to 1.0, it reduces to a worst limit model (equation 2). Thus, the Estimated Mean
Shift model can simulate the entire continuum between these two extremes, as shown in
figure 10.
Other advantages of the Estimated Mean Shift model include:
1) Full flexibility to mix shift factors in an assembly. Some parts may be nearly
worst case, while others nearly straight statistical. There is no need to penalize an
entire assembly with a worst limit analysis because of one poorly controlled
component.
2) Thermal expansion, which must be treated as worst limit, may now be included
in a statistical assembly analysis.
3) Early in the design stage, when little manufacturing data are available,
conservative shift factors may be assigned. Later, during production, as data
becomes available, manufacturing systems analysts may substitute more precise
values. This may allow tolerances to be loosened up so that production rates may
be increased.
17
Y = g(x1 ,x2 - - -xn )
Statistical
Worst Limit
Estimated
Mean
Shift
X - R
SPC
(Sampled)
100
%
0
%
SPC
X - R
Max Min
Mean -3 +3
100%
Inspection
(Go-No go)
Assembly Function
Probable Location
of the Mean
Fig. 10. Versatility of the Estimated Mean Shift Model
The Estimated Mean Shift model can provide a common ground for interaction between
engineering and manufacturing. The engineer's attention is focused on manufacturing
considerations and he must communicate with manufacturing to get the needed model data.
He can convey the critical design parameters to manufacturing in a form that permits
freedom to alter tolerances without violating design requirements. Manufacturing can
communicate meaningfully to engineering in terms of the two quality assurance parameters
most commonly used in statistical process control: mean and variance.
8. Advanced Statistical Tolerance Analysis Models
Advanced statistical methods are sometimes used for tolerance analysis because they
permit non-Normal distributions as models of component variation. These methods can
give much better estimates of the number of rejects than simple statistical analysis, when
the component distributions are well-known non-Normal functions. However, they are not
a convenient tool for tolerance allocation. A complete distribution for each component
must be input into the assembly equation before the advanced models can predict the
resulting assembly distribution and fraction of rejected assemblies. This is shown
18
graphically in figure 11. By contrast, in tolerance allocation, the reject fraction is selected
and the component tolerances are then scaled to meet the specified reject fraction.
y
x
1
x
2
x
n
y = g(x
1
,x
2
, . . . ,x
n
)
ASSEMBLY FUNCTION
COMPONENTS
ASSEMBLY
Fig. 11. Advanced Parametric Tolerance Analysis requires
that distributions be selected for each component
Two advanced statistical methods which have been applied to tolerance analysis are Monte
Carlo Simulation and the Method of Moments. These methods are explained excellently by
Hahn & Shapiro [8], Evans [9,10] and Shapiro & Gross [11]. The evaluation which
follows is taken from a study by Greenwood [12].
Monte Carlo Simulation uses pseudo-random number generators to describe a wide variety
of distribution shapes. A random dimension for each component is input into the assembly
function. The value of the resultant assembly variable is determined and compared to the
specified assembly limits. This process is then repeated many times and the count of
rejected assemblies is divided by the total trials to estimate the fraction of rejected
assemblies. In tolerance analysis, the permissible rejection fraction is usually quite small
and large samples on the order of 10,000 or 100,000 are required to give accurate
predictions of rejects. These large samples require significant computation time which is
not practical at present on personal computers. The computer programs and algorithms to
generate the random dimensions are not long, although caution must be exercised to use a
non-repeating sequence for the base random number generator.
The Method of Moments uses the statistical moments of the component distributions and
the first and second partial derivatives of the assembly function to find the first four
moments of the assembly distribution. These four moments are used to find the parameters
of a general distribution such as the Pearson system, the Johnson System or the Lambda
distribution. With the parameters of a distribution determined, the fraction outside of the
assembly limits is found from tables, numerical integration or in some cases by algebraic
equations. A computer program to do tolerance analysis by the Method of Moments will
be quite long and complex due to the need for numerical derivatives in most cases and the
many series summations to get the assembly moments. However, once the program is
complete, the computation time is moderate compared to Monte Carlo Simulation.
19
An alternate scheme which requires a moderately complicated program with moderate
computation time is a blend of the Method of Moments and Monte Carlo Simulation. This
is the Hybrid method and uses Monte Carlo Simulation to generate a smaller number of
assembly values. The sample size is usually on the order of 1000 to 5000. The resultant
assembly dimensions are used to calculate the statistical moments of the assembly
distribution and then estimate the fraction of rejects. Most of the complexity of the
Method of Moments is eliminated since numerical derivatives and the series summations
are not needed to find the assembly moments from the component moments. Since the
sample size can be on the order of 1000, the computation is greatly reduced from the
simple Monte Carlo simulation.
In the following examples, the advanced statistical methods are applied to the shaft and
housing assembly previously shown in figure 3. To illustrate the non-Normal capability,
the component distributions are assumed to be Uniform rather than Normal. Trial values
must be selected for the unassigned tolerances in order to perform the analysis. The
rejection fraction or the fraction of assemblies outside the limits is estimated for each case.
Arbitrarily let T
B
=T
E
= .0060, T
D
= T
F
= .0020 in all of the following examples.
The specified assembly clearance is: .020 +/-.015
which gives assembly limits of: max. = .035, min. = .005
a. Method of Moments
The resulting assembly moments yield the following statistics:
Mean= .020, Standard Deviation = .00562, Skewness= 0.0, Kurtosis= 2.639
The Kurtosis of the sum of Uniform distributions is seen to be approaching 3.0, which is
the value for a Normal distribution, as predicted by statistical theory.
Two different empirical distributions are fit to the values which give the following
estimates of the number of rejected assemblies.
Table 5. Sample Results for Method of Moments
Rejects/1000
y > .035 y < .005
Johnson Bounded 1.44 1.44
Lambda 1.40 1.40
b. Monte Carlo Simulation
The random number generators require a starting or seed value. Different seeds and
varying sample size will give different values for the assembly rejects.
20
Table 6. Sample Results for Monte Carlo Simulation
Rejects/1000
Sample Size Seed y > .035 y < .005
100,000 84602 1.54 1.27
100,000 17805 1.54 1.50
1,000 1136 2 1
1,000 12091 2 3
1,000 12178 0 0
The large sample simulations give rejection fractions that agree well with the Method of
Moments. The small sample simulations do not agree as well and vary widely.
c. Hybrid
The Monte Carlo simulations of part (b) above were used to estimate the first four
assembly moments. These were fitted to the Lambda distribution to yield the fraction of
rejects.
Table 7. Sample Results for the Hybrid Method
System Moments: I II III IV Rejects/1000
Sample Mean St.Dev Skewne Kurtosis y>.035 y<.005
100,000 .02001 .005633 -.00275 2.630 1.49 1.48
100,000 .01998 .005614 -.00058 2.635 1.42 1.48
1,000 .01976 .005586 .05162 2.637 1.22 1.48
1,000 .01985 .005635 -.11791 2.643 0.99 2.55
1,000 .01992 .005493 -.03680 2.575 0.68 0.89
The various simulations give different estimates of the first four assembly moments and
rejection fraction. The wide variation in the predicted rejects indicates that the accuracy of
the Hybrid method also depends on sample size.
9. Design Limitations of the Advanced Methods
The following factors appear to seriously limit the application of the advanced statistical
tolerance methods as design tools:
1) In the early design stages, little information is available on distribution type.
Even for standard items, large variations between different batches are frequently
encountered. Quality control records of previous production usually just provide
the percent yield or mean and standard deviation and not distribution shape.
2) Advanced methods are computation-intensive, causing great difficulty when
doing allocation or synthesis work rather than just analysis. In the example
problem, tolerances for each component were selected and the fraction of rejects
was calculated. If several components were unassigned and a permissible rejection
fraction were to be matched, an iterative application of the advanced method would
21
be needed. If very many design tolerances were to be determined, a multi-
dimensional search in design space would result in many design iterations.
3) Finally, the quality control methods, which are still struggling to be accepted
and utilized in this country by many industries, are based on the Normal
distribution. These techniques do not utilize information on the third and fourth
moments because of the large sample size required. Even if distribution shape were
well-known by design, and this information were used to refine component
tolerances, the existing quality control techniques may not be sufficiently robust to
predict out-of-control conditions if only higher moments are changing.
Therefore, the advanced methods do not appear suited for initial design at present, but may
be applied during production as sufficient component data becomes available.
10. Conclusion -- Design Issues
The important design issues in mechanical tolerance analysis have been briefly described
and identified. The most important issue in mechanical tolerance analysis is to recognize
that design and manufacturing must become allies in producing competitive products to
ensure survival in international markets. This may appear trite, but the authors' experience
in practice and consulting indicates this is a major problem in many companies. The
conflict that arises between design and manufacturing illustrates the need to communicate
effectively.
The poor communication between design and manufacturing underscores the principal
technical issue in mechanical tolerance analysis: the lack of a common definition of a
tolerance. Debates over worst limit and simple statistical tolerance models creates conflict
between design and manufacturing. The Estimated Mean Shift is proposed as an enhanced
model for tolerance analysis that is a step toward a common definition of tolerance. Its use
by both engineering and manufacturing will permit communication of the needs of each in
terms of currently available quality parameters. It makes assumptions which are plausible to
designers and which can be monitored in production. This new method must be proven in
practice and modified or superseded as needed to obtain a single realistic meaning of
tolerance to both designers and manufacturers.
The allocation or synthesis of tolerances in a design is a difficult matter and results in many
iterations between design and manufacturing. The Proportional Scaling and Precision
Factor Methods are simple, but crude guides. Cost optimization is very desirable, but is at
present a subject of academic pursuit since cost-vs.-tolerance curves are almost universally
unavailable.
In order to make advanced tolerance analysis and optimization methods available for
tolerance allocation by designers, we must use quality control techniques to determine
process capabilities, track costs and then ultimately model machines and processes through
design of experiments. This is an enormous task which will require cooperation across
company and even industry boundaries.
22
11. Acknowledgment
This study was sponsored by the Association for the Development of Computer-Aided
Tolerancing Software, which is a joint effort by Brigham Young University and ten
industrial and government sponsors.
References
1. Fortini, E.T., Dimensioning for Interchangeable Manufacture, Industrial Press, 1967.
2. Trucks, H.E., Designing for Economical Production, 2nd ed., Society of Manufacturing
Engineers, Dearborn, Michigan, 1987.
3. Spotts, M.F., "Allocation of Tolerances to Minimize Cost of Assembly," Journal of
Engineering for Industry, Transactions of the ASME, Vol. 95, August 1973, pp. 762-
764.
4. Chase, K.W. and W.H. Greenwood, "Computer-Aided Tolerance Selection: CATS
User Guide," ADCATS Report No. 86-2, Brigham Young University, May 1986.
5. Speckhart, F.H., "Calculation of Tolerance Based on a Minimum Cost Approach,"
Journal of Engineering for Industry, Transactions of ASME, Vol. 94, May 1972, pp.
447-453.
6. Jamieson, A., Introduction to Quality Control, Reston Publishing, 1982.
7. Greenwood, W.H., and K.W. Chase, "A New Tolerance Analysis Method for
Designers and Manufacturers," Journal of Engineering for Industry, Transactions of
ASME, Vol. 109, May 1987, pp. 112-116.
8. Shapiro, S.S. and Gross, A.J., Statistical Modeling Techniques, Marcel Dekker, Inc.,
1981.
9. Evans, D.H., "Statistical Tolerancing: The State of the Art, Part 1," Journal of Quality
Technology, Vol. 7, No. 1, pp. 1-12, January 1975.
10. Evans, D.H., "Statistical Tolerancing: The State of the Art Part, 2," Journal of Quality
Technology, Vol. 7, No. 1, pp. 1-12, January 1975.
11. Hahn, G.J. and Shapiro, S.S., Statistical Models in Engineering, John Wiley & Sons,
Inc., 1967.
12. Greenwood, W. H., "A New Tolerance Analysis Method for Engineering Design and
Manufacturing", Ph.D. dissertation, Mechanical Engineering Department, Brigham
Young University, March, 1987.