STAT 4007 Statistical Quality Control

Chapter 6 Process Capability

Dr. Philip Lee

2014/15 Term 2

Reference: Chapter 8 of Statistical Quality Control , Douglas C. Montgomery,

6th Edition
1
 Introduction
• Chapter 4: Variables Control Charts
Chapter 5: Attributes Control Charts
- Use of Shewhart control charts to (hopefully) remove the
variability due to special causes (Phase I).
• Suppose that we are now in phase II,
– the process is assumed to be relatively stable.
– We would like to compare the output of an in control
process with the specifications (e.g. LSL and USL).
=> Process Capability
• Process capability measures the effectiveness of a process
in creating a product/item.

2
 Capable Process
• Definition: A capable process is a process that
– is stable and not changing
– fits the specifications with a little extra room to spare
(typically 25%)
• (Intuition): The smaller the variation existing in a process,
the more capable the process would be in meeting the
specifications.
• Process Capability Index (PCI) measures the ability of the
process to meet the specification.
Process Specifications
Process variability σ e.g. LSL, USL
(and perhaps Process mean µ)
3
 Attribute Quality Characteristic
• Process capability is typically measures by the number or
proportion of nonconformities being produced.
• Measures on Process Capability:
1) Process Fallout: Number of nonconformities being
produced. Unit: parts per million (ppm)
2) Process Fallout Rate: Proportion of nonconformities being
produced. Unit: percentage

E.g. Suppose that Pr(defect |Process in control) =  = 0.0027,

then Process Fallout = 2,700 ppm
Process Fallout Rate = 0.27%
Not much to evaluate => Not the focus of the chapter
4
 Chapter Overview
The chapter focuses on Variable Quality Characteristics:
• Section 6.1: Natural Tolerance Limits and Specification
Limits
• Section 6.2: Process Capability Index (PCI):
– The Cp Index
– The Cpk Index (and one-sided indices: Cpu and Cpl)
– The Taguchi Capability Index Cpm
• Section 6.3: Comparisons between PCIs

5
 Section 6.1
Natural Tolerance Limits and Specification Limits

6
 Natural Tolerance Limits and Specification Limits
• Natural Tolerance Limits:
– The stable control limits obtained in Phase I, which is the
control limits for Phase II operation
– Represents the natural variability of the quality
characteristic existing in an in control process
Upper Natural Tolerance Limit (UNTL):   3 or ˆ  3ˆ
Lower Natural Tolerance Limit (LNTL):   3 or ˆ  3ˆ
Question: What is the difference between Natural Tolerance
Limits and Control Limits (in Ch4 and 5)?

• Natural Tolerance Limits identifies the actual process

condition 7
 Natural Tolerance Limits and Specification Limits
• Specifications are set by producer/need from customers
– Upper Specification Limits (USL)
– Lower Specification Limits (LSL)
(Ch1): Nonconformities (or defects) = items with quality
characteristic outside the specification limits
Unless specified, assume Target Value T = (USL+LSL)/2 for
the rest of the chapter

• Definition: Process Spread = UNTL - LNTL

• Definition: Specification Spread = USL - LSL
• Simple Comparison between the two spreads
=> Indication on whether a process is good or not (next page)
8
 Specifications and Process Capability

Specification limits outside the natural tolerance limits

=> most of the items produced are conforming
9
 Example on Off-centered Process
 Centering is important when the process spread is
compared with the specification spread (Example below)
Example 6.1: Suppose that the diameter is the quality characteristic
of a screw (螺絲), with its specification limits 5.000 ± 0.015cm.
Based on a sample from an in control process,
x = 4.990cm and s = 0.004cm.
a) Compute the Process Fallout Rate.
b) If the process is centered at its target value, does the Process
Fallout Rate decrease?

LNTL LSL `x Target UNTL USL

4.978 4.985 4.99 5.00 5.002 5.015 10
Quality Characteristic (x)
 Example on Off-centered Process
Solutions: Process Fallout Rate:
Pr( X  LSL )  Pr( X  USL )  Pr( X  4.985)  Pr( X  5.015)
 4.985  4.990   5.015  4.990 
 Pr Z    Pr  Z  
 0.004   0.004 
 ( 1.25)  ( 6.25)  0.1056
which is quite large!

If the center of the process is the target value, Process Fallout Rate:
Pr( X  LSL )  Pr( X  USL )  Pr( X  4.985)  Pr( X  5.015)
 4.985  5.000   5.015  5.000 
 Pr Z    Pr Z  
 0.004   0.004 
 ( 3.75)  ( 3.75)  0.00017

Note: Compared with α=2Φ(-3)= 0.0027, Process Fallout Rate << α,

as Process spread =0.024cm < 0.030cm = specification spread
11
 Example on Off-centered Process
Solutions (Cont’d):
Original Process
process mean shifts
(sample) to the target
mean value

0.1056

LNTL LSL `x Target UNTL USL

5.00 5.002
4.978 4.985 4.99 5.015

Quality Characteristic (x)

12
 Specifications and Process Capability
Assume process mean  = Target value T

Possible corrective actions for Case 3:

1) Reduce the process spread (Ch4 and 5)
2) Increase the specification limits
3) Perform 100% inspection
13
 Section 6.2
Process Capability Index (PCI)

14
 Process Capability Index (PCI)
• Process Capability Index (PCI) measures the ability of the
process to meet the specification.
– What the process ‘should do’ vs what the process is
‘actually doing’.
• General Form of PCI:
Allowable range of measuremen ts
General form of PCI 
Actual range of measuremen ts
** Want PCI to be as large as possible **
• Three Common PCIs:
– The Cp Index
– The Cpk Index
– The Cpm Index (Taguchi (田口) Capability Index)
15
 Quick Summary of PCIs
• PCIs compare the specifications (i.e. LSL, USL) with the actual
process (i.e. µ,) => PCIs are functions of (LSL, USL, µ,)

USL  LSL  USL     LSL  USL  LSL

Cp  C pk  min  ,  C pm 
6  3 3  6  2  (   T )2

USL  LSL  USL  x x  LSL  Cˆ pm 

USL  LSL
Estimates Cˆ p  Cˆ pk  min  , 
6s  3s 3s  6 s 2  ( x  T )2
s R sR USL  LSL  USL  x x  LSL  USL  LSL
ˆ  , , Cˆ p  Cˆ pk  min  ,  Cˆ pm 
c4 d 2 d 3
6̂  3ˆ 3ˆ  6 ˆ 2  ( x  T )2

16
 The Cp Index
USL  LSL Specification Spread
Cp 
6 Process Spread
The CP Index indicates the potential of a process
• If CP ≥ 1, the process has potential to generate low process
fallout rate.
• If CP < 1, the process does not have the potential to generate
low process fallout rate

• Note: High Cp DOES NOT imply low process fallout rate

– If µ is different from Target T=(USL+LSL)/2, process fallout
rate could be large even though Cp ≧ 2.0 (next page)
– If µ = T =(USL+LSL)/2, then
High Cp  Low Process Fallout Rate
17
 Problem with the Cp Index

Process
Specification spread
LSL USL
spread

LNTL UNTL

USL  LSL 1
• Example 6.2: From the above, C p  6 
6(0.08333)
 2 1
The process has potential, but process fallout rate = 50%
Conclusion: Cp is not able to describe the actual performance
of the process, but instead a minimum requirement (potential)
for a process to be good (in terms of low process fallout rate)
=> Need Cpk index to address the centering issue. 18
 Cpk Index Based on Cpu and Cpl Indices
Cpk is constructed based on two one-sided capability indices:
USL  
Upper capability index: C pu 
3
  LSL
Lower capability index: C pl 
3
 USL     LSL 
Cpk index: C pk  min( C pu , C pl )  min  , 
 3 3 
• If Cpk ≥ 1 => both Cpu ≥ 1 and Cpl ≥ 1
=> items within specifications ≥ 1-2(-3)= 99.73%
(Assume the quality characteristic is normally distributed)
• Interpretation: Cpk reflects the actual performance of a
process (via maximum fallout rate) instead of potential
performance for the Cp index.
• Note: One-sided Problem (without either the LSL or USL): Use
Cpu or Cpl as the PCI (Example 6.3)
19
 Example on One-sided Process Capability Index
Example 6.3: Table 6.1 contains the bursting strength (破裂強度)
of 100 glass containers. Assume that the strength is normally
distributed, with lower specification limit given by 200 psi,
(a) Evaluate a one-sided process-capability index.
(b) Compute the process fallout rate.
Table 6.1: Bursting Strengths for 100 Glass Containers
265 197 346 280 265 200 221 265 261 278
205 286 317 242 254 235 176 262 248 250
263 274 242 260 281 246 248 271 260 265
307 243 258 321 294 328 263 245 274 270
220 231 276 228 223 296 231 301 337 298
268 267 300 250 260 276 334 280 250 257
260 281 208 299 308 264 280 274 278 210
234 265 187 258 235 269 265 253 254 280
299 214 264 267 283 235 272 287 274 269
215 318 271 293 277 290 283 258 275 251 x  264.06, s  32.02
** No USL in this example because higher bursting strength
means better container! 20
 Example on One-sided Process Capability Index
Solutions: (a) From the data,`x =264.06, s=32.02. Therefore,
ˆ ˆ  LSL 264.06  200
C pl    0.667
3ˆ 3(32.02)
(Note that both Cp and Cpu are undefined because of no USL)
(b) Process fallout rate:
 LSL - ˆ   200  264.06 
Pr( X  LSL)  Pr Z    Pr Z  
 ˆ   32.02 
 Pr( Z  2.001)  0.0227
under the standard normal distribution. The process fallout
rate is 2.27%, or process fallout of 22,700 ppm.

21
 Process fallout for Various Values of PCIs
Assume that (1) the quality characteristic is normal distributed
(2) the process is centered at its target value (µ=T)
Process Fallout (in ppm) for Various Values of the Cp, Cpl and Cpu

Example: When Cpl =0.7. Process fallout (in ppm) is

LSL -  
  10  PrZ  3C pl   10  Pr( Z  2.1)  10  17,864

PrX  LSL   106  Pr Z  6 6 6

  
I.e., we expect 17,865 defects per 1 million of items produced.
22
 Capable Process and Desirable Process
• Definition: A process is said to be capable if the PCI of interest
is greater than or equal to 1.33.
• Example: Assume µ = T = (LSL+USL)/2 and Cp=1.33,
Process Range 1 1
  100%   100%  75%
Specificat ion Range C p 1.33
=> A buffer of (specification range)/4 in mean shift is allowed
to maintain low process fallout (< 2700ppm)
• Definition: A process is said to be desirable if the PCI of
interest is greater than or equal to 1.00.
Table 6.3 Appropriate Responses to Cp Values
QC Assessment Response
Sufficient to inspect at the beginning of operation.
Cp ≥1.33 Pass
May consider speed up the process.
1 ≤ Cp < 1.33 Needs Watching Danger of producing defects. Needs watching
Need to modify the procedures/equipment/tolerance.
Cp <1 Fail 23
100% inspection of items produced
 Biasness of Ĉ p and Ĉ pk
USL  LSL
• Point Estimators for C p  6
USL  LSL USL  LSL USL  LSL
(1) Cˆ p  , (2) Cˆ p  and (3) ˆ
Cp 
6s 6( R / d 2 ) 6( s / c4 )

• Point Estimators for C pk  min  USL   ,   LSL 

 3 3 

(1) Cˆ pk  min  USL  x , x  LSL  , (2)  USL  x x  LSL 

Cˆ pk  min  , 
 3s 3s   3( R / d 2 ) 3( R / d 2 ) 
ˆ  USL  x x  LSL 
(3) pk
C  min  , 
 3( s / c4 ) 3( s / c4 ) 
Question: Are they unbiased estimators for Cp/Cpk ?
Fact: E[S ]  c4 but E  1   1
 S  c4

24
 Confidence Intervals for the Cp Index
USL  LSL
Point Estimator for Cp: Cˆ p  , Fact: (n  1) S 2  2 ~  n21
6s
 2 ( n  1) S 2 
CI construction: Pr  1 / 2,n 1     / 2 ,n 1   1  
2

  2

 1 12 / 2,n 1 1 1 2 / 2,n 1 
 Pr      1
 S n 1  S n 1 

 USL  LSL 12 / 2,n 1 USL  LSL 2 / 2,n 1 
 Pr   Cp    1
 6S n 1 6S n 1 

   
/ 2 ,n 1 
2 2
 Pr C pˆ 1  / 2 , n 1 ˆ
 Cp  Cp 
  1
 n 1 n 1 


12 / 2,n 1 2 / 2,n 1

A 100(1-α)% CI for Cp: Cˆ p  C p  Cˆ p
n 1 n 1
** CI could be WIDE when n is small (Example 6.4)
25
 Example on Confidence Interval for Cp
Example 6.4: Consider a quality characteristic with USL=62 & LSL=38.
A sample of size n=20 with sample mean at the midpoint of the
specifications ( x  T  50 )and s=1.75.
Find (i) a point estimator for Cp and (ii) a 95% CI for Cp.
Solutions: A point estimator for Cp is
ˆ USL  LSL 62  38
Cp    2.29 (Very very good!)
6s 6(1.75)
Hence, a 95% confidence interval for Cp is:
 2
 2
Cˆ p 10.025,n 1  C p  Cˆ p 0.025,n 1
n 1 n 1
8.91 32.85
 2.29  C p  2.29
19 19
 1.57  C p  3.01
26
 Example to Support the Use of Cpk instead of Cp
Example 6.5: In Example 6.2, LSL=1, USL=2, μ=2 and σ=0.08333
(chart below) => Cp = 2.00 (> 1.33!) even though half of the
items failed to meet the specifications.
 USL     LSL   22 2 1 
Now, C pk  min  ,   min 
 ,   0  1,
 3 3   3(0.8333) 3(0.8333) 
suggesting that the process is not desirable.

Process
spread

LSL Specification spread USL

LNTL UNTL

27
 Cp vs Cpk
Magnitude of Cpk relative to
(a)
Cp is a direct measure on how
off-center the process is operating.
• Cp =2.0 for (a)-(e) (b)
[same USL-LSL and same σ]

 USL     LSL 
• C pk  min  ,  (c)
 3 3 

a) Process is centered, Cpk = Cp

b) Off-centered, Cpk < Cp (d)
c) Off-centered, Cpk < Cp
d) Process centered at a
specification limit, Cpk = 0 (e)

e) Process centered outside the

28
specification limits, Cpk < 0
 Problem with the Cpk index
• When µ≈LSL or USL, Cpk index may not be a good measure
because a shift in process mean may result in substantial
process fallout rate
• Example:

Given that LSL=35 and USL=65,

1. Cpk for process A = (USL-µ A)/(3A)=(65-50)/(3x5)=1.0
2. Cpk for process B = (USL-µ B)/(3B)=(65-57.5)/(3x2.5)=1.0
3. (Extreme case!) c=69.97 and c=0.01 => Cpk=1.0
If µ of the 3 processes shift to the right by 0.03.
Process fallout rate for C = 50%!! 29
 Taguchi Capability Index Cpm
USL  LSL USL  LSL
C pm  
6 6  2  (   T )2
τ2 is the mean squared error of X from the target T=(USL+LSL)/2,

 2  E[( X  T )2 ]  E{( X   )2  (   T )2 ]   2  (   T )2

 T
Cpm as a function of Cp: Define   to be the distance in

(standard deviation) of µ from the Target value T, then

USL  LSL USL  LSL Cp

C pm   
6   (  T ) 6 1    T  /   1  2
2 2 2

30
 Properties of the Taguchi Capability Index Cpm
USL  LSL Cp
 T
C pm  
6   (  T )
2 2
1  2 where   .

• C pm  0 for all  and σ
• As   , C pm  0.
• Cpm can address the centering issue which Cpk cannot:
Consider the case ≠T, then as   0 with LSL    USL,
USL  LSL USL  LSL 1
C pm   
6|  T | 6( USL  LSL ) / 2 3

However, C pk  min  USL   ,   LSL   

 3 3 
• In order for C pm  , we need BOTH   0 AND   T  0.
(Good!)
31
 Section 6.3
Comparisons between PCIs

32
 Cpk as a Function of Cp
• Recall that C pk  min(C pu , C pl )  min  USL   ,   LSL 
 3 3 
It can be shown that C pk  C p (1  k ) where
|  T | USL  LSL
k and T  .
( USL  LSL ) / 2 2
Hence C pk  C p with equality holds iff  = T. That is, when the
process mean is centered between the specifications.
• Proof:

33
 Cpm as a Function of Cp
USL  LSL Cp  T
• C pm   where   .
6  2  (   T )2 1  2 
Hence C pm  C p with equality holds iff  = T. That is, when
the process mean is centered between the specifications.
• How about Cpk vs Cpm? No straightforward comparisons
between the two.

34
 Comparisons between Cp , Cpk and Cpm
=38 N(59,12) =62

N(56,22)

N(50,42)

1. Cpk: 1.0 for all the 3 processes Table 6.1 PCIs for Various Process

2. Cp: Process C is the largest => Process Process μ σ δ Cp Cpk Cpm

C has the largest potential.
A 50 4 0 1 1 1.000
3. Cpm: Process A is the largest (High
penalty on the off-centering for B & C) B 56 2 3 2 1 0.632
4. Process A: Cp = Cpk = Cpm as  = T
C 59 1 9 4 1 0.442
(centered between USL & LSL).
35