Vous êtes sur la page 1sur 26

Statistical Tolerancing

Fritz Scholz
Mathematics & Computing Technology
MS: 7L-22, Phone 425-865-3623
e-mail: fritz.scholz@boeing.com
http://www.rt.cs.boeing.com/MEA/stat/tolerance.html
http://www.rt.cs.boeing.com/MEA/stat/
Statistical Tolerancing11-11-991
Why Tolerancing?
Perfect/nominal parts = perfect assembly.
Perfect parts = interchangeable parts
= mass production, low cost.
Nothing is ever perfekt, manufacturing variation is inevitable.
May have almost perfect parts.
How almost can it be? When do several almost perfect parts
cause problems at assembly? = tolerances
What is not designed with proper tolerancing
will have to be custom xed at assembly.
Costly, not interchangeable, not acceptable.
Statistical Tolerancing11-11-992
Tolerance Chain for Crank Case
G
L
2
L
1
L
3
L
4
L
5
L
6
Linear clearance criterion C = L
1
L
2
L
3
L
4
L
5
L
6
> 0
Statistical Tolerancing11-11-993
General Smooth Functions: Linearization
Y = f(X
1
, . . . , X
n
)
= f(
1
, . . . ,
n
)
variation propagation box

(X
n
,
n
)
(X
1
,
1
)
detail parts
actual, nominal

(Y, )
assembly t
criterion
actual, nominal
Y = f(X
1
, . . . , X
n
) f(
1
, . . . ,
n
) +
n

i=1
(X
i

i
)
a
i
..
f(
1
, . . . ,
n
)

i
Y a
0
+a
1
X
1
+. . . +a
n
X
n
, a
0
= f(
1
, . . . ,
n
) (a
1

1
+. . . +a
n

n
)
= f(
1
, . . . ,
n
) a
0
+a
1

1
+. . . +a
n

n
Statistical Tolerancing11-11-994
Tolerancing and Goalpost Mentality

nominal
nominaltol nominal+tol
tol

i

i
+T
i

i
T
i

X
i
Statistical Tolerancing11-11-995
Worst Case or Arithmetic Tolerancing
X
i
within [
i
T
i
,
i
+T
i
], X
i
= realized detail part dimension,

i
= nominal dimension, T
i
= tolerance.

A
= a
0
+a
1

1
+. . . +a
n

n
nominal assembly t criterion,
a
i
= 1 or other known coecients
Y
A
= a
0
+a
1
X
1
+. . . +a
n
X
n
realized assembly t criterion
Y
A
within [
A
T
A
,
A
+T
A
] assembly t criterion variation range.
T
A
= |a
1
|T
1
+. . . +|a
n
|T
n
arithmetic stacking or worst case tolerancing
Statistical Tolerancing11-11-996
Statistical Tolerancing
No goalpost mentality, aim for the middle
View part to part variation as random, manufacturing variation
Normal distribution is often a suciently reasonable choice
Other distributions are possible (uniform, triangular, . . .)
Distributions typically span tolerance range.
The t criterion Y
A
= a
0
+a
1
X
1
+. . . +a
n
X
n
normal
Central Limit Theorem (CLT).
Take advantage of variation cancellation
Statistical Tolerancing11-11-997
Normal Variation over Tolerance Interval
T T
i i i i
- +
i
Statistical Tolerancing11-11-998
RSS Method: Normal Distribution
99.73% of all realized assembly t criteria Y
A
fall within
[
A
RSS,
A
+ RSS] where RSS =
_
a
2
1
T
2
1
+. . . +a
2
n
T
2
n
whence Root Sum of Squares or RSS.
Basic advantage over arithmetic tolerancing (|a
i
| = 1, T
i
= T)
_
a
2
1
T
2
1
+. . . +a
2
n
T
2
n
=

n T , whereas |a
1
|T
1
+. . .+|a
n
|T
n
= n T.
= tighter assembly tolerance RSS or relaxed part tolerances T
i
.
Statistical Tolerancing11-11-999
RSS Method: Basic Assumptions
Each part dimension X
i
varies according to a normal distribution.
The distribution is centered over the range
i
T
i
.
T
i
= 3
i
, with
i
= standard deviation of X
i
,
i.e., 99.73% of all X
i
fall within
i
T
i
.
The part variations are independent from dimension to dimension.
Statistical Tolerancing11-11-9910
Current Problems in Statistical Tolerancing
RSS may be too optimistic = c
_
a
2
1
T
2
1
+. . . +a
2
n
T
2
n
the ination factor c can be motivated by
dierent distributions
mean shifts from nominal center (ATA patent)
other ad hoc rationales (Benderizing)
Not all tolerance stacks are (approximately) linear
Then RSS wont work, not addressed much in literature
757 cargo door hinge line, 747 keel beam, coordination hole pinning,
etc.
Simulations (e.g. VSA, or other tools) provide exible solutions, but
they are of black box type (sensitivities?)
Statistical Tolerancing11-11-9911
Various Distributions over Tolerance Range
normal density
c = 1
uniform density
c = 1.732
triangular density
c = 1.225
trapezoidal density: a = .5
c = 1.369
elliptical density
c = 1.5
half cosine wave density
c = 1.306
beta density: a = 3, b = 3
c = 1.134
beta density: a = .6, b = .6
c = 2.023
beta density: a = 2, b = 2 (parabolic)
c = 1.342
DIN - histogram density: p = .7, f = .4
c = 1.512
Statistical Tolerancing11-11-9912
RSS Method for Nonnormal Variation
Assumptions
distributions symmetric and centered on
i
distributions span tolerance range
independence of part to part variation
= RSS(c) = c

_
n

i=1
a
2
i
T
2
i
where the factor c varies from distribution to distribution.
More generally RSS(c
1
, . . . , c
n
) =

_
n

i=1
c
2
i
a
2
i
T
2
i
if the distributions change from part to part dimension.
Statistical Tolerancing11-11-9913
Why Mean Shifts And Why Bound Them?
Mean shifts from nominal are a fact of life.
A manufacturing process unable to hit the nominal
will also be unable to center the process on nominal.
The center of the process distribution only becomes known
well into production. Corrections may only increase variation.
Mean shifts all the way to the goalposts
i
T
i
, while C
pk
1,
lead back to worst case tolerancing.
Statistical Tolerancing11-11-9914
Normal Variation with Mean Shift
Statistical Tolerancing11-11-9915
Mean Shifts From Nominal Center
Focus on normal part variation that is not centered on nominal,
but still mostly within the tolerance range.
The mean shift from center is controlled by
|
i
| = |
i

i
| =
i
T
i

0
T
i
, where 0
i
1 .
Here
i
is the mean of X
i
with mean shift
i
=
i

i
.
The variation mostly within the tolerance range is controlled by
the capability index C
pk
=
T
i
|
i
|
3
i
, e.g., C
pk
1 .
Statistical Tolerancing11-11-9916
RSS Stacking With Mean Shifts
Stack mean shifts arithmetically

1
|a
1
|T
1
+. . . +
n
|a
n
|T
n
Stack reduced part to part variation statistically
reduced variation T
i

i
T
i
= (1
i
)T
i
3
i
if C
pk
1 .
statistical stack
_
a
2
1
T
2
1
(1
1
)
2
+. . . +a
2
n
T
2
n
(1
n
)
2
Combine the two stacks in worst case fashion.

1
|a
1
|T
1
+. . . +
n
|a
n
|T
n
+
_
a
2
1
T
2
1
(1
1
)
2
+. . . +a
2
n
T
2
n
(1
n
)
2
Statistical Tolerancing11-11-9917
Tolerance Stack with Mean Shift
X1
part dimension 1
X2
part dimension 2
X3
part dimension 3
Y
assembly criterion
Statistical Tolerancing11-11-9918
Some Comments
This hybrid worst case mean shift stack and RSS stack for
remaining part variation was introduced by Mansoor (1963)
and studied by Greenwood and Chase (1987) as a compromise
between worst case and statistical tolerancing.
The assembly tolerance of this hybrid method grows like O(n),
reduced by the
i
factors.
The above hybrid tolerance stacking formula is increasing in the
i
thus we can replace the
i
by the bounds (
0
) placed on them.
Statistical Tolerancing11-11-9919
Mean Shifts And Other Part Distributions
The remaining part variation can also be described or modeled
by other than normal distributions.
This then leads to the following hybrid tolerance stacking formula

1
|a
1
|T
1
+. . . +
n
|a
n
|T
n
+
_
c
2
1
a
2
1
T
2
1
(1
1
)
2
+. . . +c
2
n
a
2
n
T
2
n
(1
n
)
2
This is no longer increasing in the
i
.
Statistical Tolerancing11-11-9920
Correlated or Dependent Variation
A crucial assumption for the RSS method is independence or zero
correlation for detail dimensions
Method can be generalized using the general variance formula
var(a
0
+a
1
X
1
+. . .+a
n
X
n
) = a
2
1
var(X
1
)+. . .+a
2
n
var(X
n
)
+

i=j
a
i
a
j
_
var(X
i
)

var(X
j
)
ij
where
ij
is the correlation between X
i
and X
j
, 1
ij
1.
This generalizes the RSS formula to
RSS

a
2
i
T
2
i
+

i=j
a
i
a
j

ij
T
i
T
j
By proper combination of the signs of
ij
and the signs of the
sensitivity coecients a
i
one could have
RSS

< RSS =
_

a
2
i
T
2
i
Statistical Tolerancing11-11-9921
Historical Perspective
The RSS method is based on the CLT and the long known fact that
variance of a sum of independent random variables
= sum of the variances of these variables
It is dicult to trace the roots of statistical tolerancing.
Boeings (1990) proprietary Tolerancing-Design Guide attributes
the RSS method to Backhaus and Fielden via Wades (1967) book
which points to an in-house IBM article (not able to trace).
Statistical Tolerancing11-11-9922
Some Early References
W.B. Rice (1944) argues (without RSS) that chances are small
for jointly extreme dimensions (by multiplying probabilities).
C.A. Gladman (1945) employs the same reasoning (no RSS).
J. Gilson (discussion to Gladman) points in the same direction.
However, in Gilsons (1951) book we nd the RSS formula
plus a lot of muddled stu without proof.
E.E. Bates (1947/1949) appears to use the RSS method,
refers to it as statistics, keeps specics hidden (trade secret?)
Statistical Tolerancing11-11-9923
The Earliest Reference (so far)
W. A. Shewhart (1931)
Economic Control of Quality of Manufactured Product
long out of print, reprinted 1980 as
50
th
Anniversary Commemorative Reissue by ASQC
Chapter XVII: Design Limits on Variability
discusses statistical tolerancing, but not RSS by name.
does not pretend to be new, but no references
statistics was in its infancy at that time
Statistical Tolerancing11-11-9924
What Do The Standards Say? Not Much!
ANSI Y14.5M-1994 just introduced a statistical tolerancing symbol.
It still needs meaning (mean shifts, distributions, process control).
How should/can it be used in tolerance stacking?
DIN 7186 Part 1 and Part 2 (Draft) took a stab at it
(DIN = Deutsche Industrie Norm)
That was more than 20 years ago, results in German only.
Statistical Tolerancing11-11-9925
Grodes Comments on German Tolerancing Eort
Eort failed, because industry felt that contents
were too dicult to be understood by designers
who generally do not like statistical applications
but try to avoid it.
Meanwhile all the old famous experts have died.
A company which uses statistical tolerancing can save a lot of money.
Statistical Tolerancing11-11-9926

Vous aimerez peut-être aussi