Download as pdf or txt
Download as pdf or txt
You are on page 1of 4

October 22, 2001

To: Todd Minnick


From: Paul F. Jackson,
Position Feature Bonus Position Residual Percent-of
Todd Minnick, Measured Size Tolerance Adjusted Tolerance Tolerance
Todd I recently found the Power-point presentation that you contributed to the 2001
Quality and Productivity Research Conference, Austin, TX, May 23-25, 2001 on the Web. I TPm Xf BT TPa TPr TPp
have done a little digging into the subject since I thought that I had discovered a solution to the 1 0.0045 0.1000 0.0030 0.0015 0.0035 0.5625
problem of predicting process capability with geometric bonus tolerance last year. The
2 0.0045 0.1015 0.0015 0.0030 0.0020 0.6923
"Residual Tolerance Process Capability Model" was published in Quality magazine in February
2001. This summer I analyzed it in comparison to other models that I found, "Calculation of Cpk 3 0.0035 0.1003 0.0027 0.0008 0.0042 0.4545
Under Conditions of Variable Tolerances" by Glen Gruner, Quality Engineering Magazine 1991 4 0.0049 0.1004 0.0026 0.0023 0.0027 0.6447
and "Calculating MMC Cpk" by Marty Ambrose, 1999. I discovered that all three methods were 5 0.0042 0.1010 0.0020 0.0022 0.0028 0.6000
flawed because the practice of combining the variables to produce a value for the "Adjusted True
Position, Percent-of-tolerance, or Residual tolerance" could either amplify or moderate the total 6 0.0048 0.1002 0.0028 0.0020 0.0030 0.6154
variation of the combined variances for size and geometric deviation. 7 0.0044 0.1003 0.0027 0.0017 0.0033 0.5714
8 0.0046 0.1006 0.0024 0.0022 0.0028 0.6216
That discovery led me to the equation (shown on the next page) for "Variable Tolerance Process 9 0.0042 0.1007 0.0023 0.0019 0.0031 0.5753
Capability." It is essentially the same as "Method #2" shown in your presentation because the 10 0.0045 0.1001 0.0029 0.0016 0.0034 0.5696
difference between the USL and the Mean-TPa is the same as the difference between Mean-TPm
11 0.0040 0.1007 0.0023 0.0017 0.0033 0.5479
and (USL + Mean-Bonus). Also, the variation that the margin is compared to is the combined
variances of Bonus and TPm. Glen Gruner, in his paper in 1991 stated the equation for the 12 0.0040 0.1009 0.0021 0.0019 0.0031 0.5634
combined variation of size and geometric deviation but then he used (or he assumed equivalent) 13 0.0041 0.1001 0.0029 0.0012 0.0038 0.5190
the variation of the values for the surrogate variable that he created "TPa" (adjusted true 14 0.0042 0.1002 0.0028 0.0014 0.0036 0.5385
position). 15 0.0040 0.1005 0.0025 0.0015 0.0035 0.5333
In my investigation of the various methods I found that the variation of the surrogate (USL, 16 0.0040 0.1005 0.0025 0.0015 0.0035 0.5333
bonus and geometric deviation combined) does not always equal the sum of the variation
17 0.0042 0.1004 0.0026 0.0016 0.0034 0.5526
from each component of the surrogate variable. I fell into the same trap with my "Residual
Tolerance" by assuming that the separate variations were combined without bias. Marty 18 0.0042 0.1003 0.0027 0.0015 0.0035 0.5455
Ambrose's method determines the ratio between the TPm and (USL + Bonus) for each measured 19 0.0045 0.1005 0.0025 0.0020 0.0030 0.6000
feature and then compares the mean of those ratios to (1). The Percent-of-Tolerance method not 20 0.0041 0.0996 0.0034 0.0007 0.0043 0.4881
only changes the separate variances by combining them in the ratio, it then expands that ratio to 21 0.0040 0.1004 0.0026 0.0014 0.0036 0.5263
a common scale further modifying the variance. The weaknesses of all three methods can be 22 0.0040 0.1005 0.0025 0.0015 0.0035 0.5333
demonstrated with patterned data, Big bonus - big deviation and small bonus - small deviation, 23 0.0040 0.1005 0.0025 0.0015 0.0035 0.5333
compared to small bonus – big deviation and big bonus – small deviation.
24 0.0042 0.1013 0.0017 0.0025 0.0025 0.6269
I could not understand the justification for subtracting 1.0 (three standard deviations of the
combined variance) from the capability ratio in your equation "Method #3." 25 0.0041 0.1006 0.0024 0.0017 0.0033 0.5541
STDEV 0.000302 0.000401 0.000401 0.000498 0.000498 0.050882
MEAN 0.004228 0.100484 0.002516 0.001712 0.003288 0.564120
"Variable Tolerance Process Capability" Jackson, 2001 Bonus = (Feature Size – (the "MMC" or "LMC" specified size)) For each measured feature.
TPm = True Position "Measured" = 2 ((Xmeasured – Xbasic)2 + (Ymeasured – Ybasic)2 )1/2 For each measured feature.
 
 USL + X − X   .005 + .002516 − .004228 
Ppk =  Bonus TPm
=  = 2.184
 3 σˆ S2 + σˆ S2   3 × .0005018 
 Bonus TPm 
"Cpu Calculations for True Position with MMC Modifier" Minnick, 2001 TPa = True Position "Adjusted" = (TPm – Bonus) For each measured feature
 USL − X + 3 σˆ 2 + σˆ 2   
 S S   USL − X   .005 − .001712 
− 1  = (2.184 − 1) = 1.184
TPa Bonus TPm
Ppk =  = TPa
− 1 = 
 3 σˆ S2 + σˆ S2   3 σˆ S + σˆ S
2 2
  3 × .0005018 
 Bonus TPm   Bonus TPm 
"Residual Tolerance Process Capability" Jackson, Quality Magazine 2001
 X − LSL 
   .003288 − 0 
Ppk =  TPr =  = 2.201 TPr = True Position "Residual" = (USL + Bonus – TPm) For each measured feature.
 3σˆ S   3 × .000498 
 TPr 
"Calculating MMC Cpk" Ambrose, 1999
1− X 
 TPp   1 − .564120 
Ppk =  =  = 2.855 TPp = True Position "Percent of Tolerance" = ((USL + Bonus) / TPm) For each measured feature.
 3 σˆ S   3 × . 050882 
 TPp 
"Calculation of Cpk Under Conditions of Variable Tolerances" Gruner, Quality Engineering Magazine 1991
 USL − X 
 TPa   .005 − .001712 
Ppk =   =  = 2.201 TPa = True Position "Adjusted" = (TPm – Bonus) For each measured feature.
 3σˆ S   3 × .000498 
 TPa 
Current Method (Bonus ignored)
 USL − X 
 TPm   .005 − .004228 
Ppk =   =  = .852
 3σˆ S   3 × .000302 
 TPm 
USL LSL
.103 .097
 
 USL + X − X  LB USL
Ppk =  Bonus TPm
 = 2.184
 3 σˆ S2 + σˆ S2  X X New distribution from
 Bonus TPm  combined variances.

In the "Variable Tolerance Process Capability" equation there is no need to create a


new variable so that the mean and standard deviation of the distribution can be
compared to a constant limit. The USL is replaced with the mean variable
tolerance (constant USL .005 plus the mean bonus). The difference between it and
the mean of the geometric deviation is then compared to the combined variances of
both (geometric deviation and bonus). If the shaft size and position specification
was changed from Dia .103-.097 TP Dia. .005 @ MMC to Dia .108-.097 TP Dia. 0 .005 +/-.003
@ MMC, the value for Ppk would be the same. The USL would be zero and all of
the geometric tolerance would be in the bonus value. Process owners can improve Position MMC Size LMC
capabilities for the geometric tolerance by shifting the mean for size nearer to the
opposite material condition limit specified in the feature control frame. Of course the bilateral capability for size will determine how near the mean can approach that limit.

USL LSL
  Method #2
 USL − X  .103 .097
Ppk =  TPa
 =2.184 LB USL
 3 σˆ S + σˆ S2
2
 TPa
 Bonus TPm 
X X X
The only reason to create the variable TPa "Adjusted true position " was to produce
a mean respective of bonus tolerance that could be compared to a constant USL. Glen Bonus
Gruner calculated the TPa for each measured feature size and position but he used the
standard deviation of the values for TPa not the combined variances of size and position.
3σˆ S ≠ 3 σˆ S2 + σˆ S2 Your Method #2 yields identical results to my variable
TPa Bonus TPm
tolerance equation because the mean bonus is subtracted from the mean geometric
deviation before it is compared to the constant USL (.005). If the specification were
changed to Zero @ MMC or with a big bonus and small deviation the mean TPa will .005 +/-.003
be negative.
Position MMC Size LMC
Method #3 USL LSL
.103 .097
 
 USL − X  LB USL
Ppk =  TPa
−1 = 1.184 TPa
 3 σˆ S2 + σˆ S2  X X X
 Bonus TPm 
By subtracting 1 from the equation for unilateral tolerance process capability the
distribution is shifted three standard deviations toward the upper specification 3 Sigma shift
limit effectively comparing the difference between the USL and the mean TPa Why?
to six standard deviations of the distribution rather than three. Why?

.005 +/-.003
Position MMC Size LMC

You might also like