Control Charts and

Process Capability
Prof. Sayak Roychowdhury
Sampling
• It is not always possible to measure quality characteristics
of each item in a population.
• Samples are used to provide information about process or
product characteristics at a fraction of cost.
• Necessary for destructive tests
• A sampling design is a procedure by which the
observations in a sample are chosen from the population
• An element is an object for which data are gathered
• A sampling unit is an individual element or a collection of
elements from a population
• A sampling frame is a list of sampling units
Sampling Errors

Random • Inherent sampling variability, e.g. due to

Variation instrument, people etc.

• Happens in opinion polling, customer satisfaction

Misspecification
survey, incorrect listing of sampling frame

• Happens in sample surveys, cases where

Non responses
measurements not possible
Sampling Methods

Sampling

Simple Random Stratified Cluster

Sampling Sampling Sampling
Simple Random Sampling
• A sample of size 𝑛 is chosen from a fixed population of
size 𝑁. In SRS, each possible sample of size 𝑛 has equal
chance of being selected.
• Random number tables may be used for sampling
• In estimating population mean 𝜇 by the sample mean 𝑋, ത
the variance of the estimator is given by
𝑠2 𝑁−𝑛
𝜎ො𝑥2ҧ = (𝑠 2 sample variance)
𝑛 𝑁
(𝑁 − 𝑛)/𝑁 is called finite population correction factor
Precision is inverse of variance.
Stratified Random Sample
• Useful when the population is heterogeneous, e.g.
production from multiple machines, under multiple
operators, samples from different geographical regions.
• Stratified random samples are obtained by separating the
elements of the population in nonoverlapping groups
(strata).
• Proportional allocation of sample size, for 𝑘 strata, let 𝑁𝑖
be the population size of the 𝑖 𝑡ℎ strata, and σ𝑘𝑖=1 𝑁𝑖 = 𝑁.
Sample size from each strata:
𝑛𝑁𝑖
𝑛𝑖 = i = 1,2, . . k
𝑁
Stratified Random Sample
• Sample mean and variance of estimator are given by
𝑘
1
𝑥ҧ𝑠𝑡 = ෍ 𝑁𝑖 𝑥ҧ𝑖
𝑁
𝑘 𝑖=1
1 𝑁𝑖 − 𝑛𝑖 𝑠𝑖2
𝑉𝑎𝑟 𝑥ҧ𝑠𝑡 = 2 ෍ 𝑁𝑖2 , 𝑖 = 1,2, . . 𝑘
𝑁 𝑁𝑖 𝑛𝑖
𝑖=1
2
1 𝑛𝑖 𝑛𝑖 𝑥𝑖𝑗 −𝑥ҧ𝑖
Where xത 𝑖 = σ𝑖=1 𝑥𝑖𝑗 , 𝑠𝑖2 = σ𝑗=1
𝑛𝑖 𝑛𝑖 −1
Cluster Sampling
• When a sampling frame is not available or obtaining
samples from all segments of the population is not
feasible due to geographical reasons, cluster sampling is
used.
• Population is divided into groups of elements, called
clusters
• Clusters are randomly selected and a census data is
obtained.
• Sampling error maybe reduced by choosing many small
clusters rather than choosing a large cluster
Example of Cluster Sampling
A researcher wants to conduct a study to judge the performance
of sophomore’s in business education across the India.
By using cluster sampling, the researcher can club the
universities from each city into one cluster (North, South, East,
West regions).
These clusters then define all the sophomore student
population in India.
Next, either using simple random sampling or systematic
random sampling, randomly pick clusters for the research
study.
Subsequently, by using simple or systematic sampling, the
sophomore’s from each of these selected clusters can be
chosen on whom to conduct the research study.
How to choose sample size?
• Bound on error estimation on population mean:
Let there is (1 − 𝛼) probability that the difference
between the estimated mean and the actual mean is not
greater than 𝐵 (tolerable error bound).
𝜎
𝐵 = 𝑧𝛼 𝜎𝑥ҧ = 𝑧𝛼 ⇒
2 2 𝑛
𝜎 2
𝑛 = 𝑧𝛼2 2
2 𝐵
An analyst wishes to estimate the average bore size of a large casting. Based on historical
data, it is estimated that the standard deviation of the bore size is 4.2 mm. If it is desired to
estimate with a probability of 0.95 the average bore size to within 0.8 mm, find the
appropriate sample size.
How to choose sample size?
• Bound on error estimation on population proportion:
Let there is (1 − 𝛼) probability that the difference
between the estimated proportion 𝑝Ƹ and the actual
proportion 𝑝 is not greater than 𝐵 (tolerable error bound).
E.g. proportion of satisfied customers, prop of non
conforming etc.
𝑝(1 − 𝑝)
𝐵 = 𝑧𝛼 𝜎𝑝ො = 𝑧𝛼 ⇒
2 2 𝑛
2 𝑝(1 − 𝑝)
𝑛 = 𝑧𝛼 2
2 𝐵
Either put 𝑝 = 𝑝Ƹ (sample proportion) or 𝑝 = 0.5 for
conservative estimate
Example
We want to estimate with a probability of 0.90 the
proportion of nonconforming tubes to within 4%. How large
a sample should be chosen if no prior information is
available on the process?
How to choose sample size?
• Estimating difference between 2 population means:
2 2
2 𝜎1 + 𝜎2
𝑛 = 𝑧𝛼
2 𝐵2
Where B is the tolerance of error for estimating the
difference in population means with sample means

• Estimating difference between 2 population proportions:

2 𝑝1 (1 − 𝑝1 ) + 𝑝2 (1 − 𝑝2 )
𝑛 = 𝑧𝛼
2 𝐵2
7 QC Tools
• Cause and Effect Diagram
• Check Sheet
• Control Chart
• Histogram
• Pareto Chart
• Scatter Diagram
• Stratification / Defect Concentration Diagram
Control Charts
Utility of Control Charts
• Control charts are proven techniques to improve
productivity
• Effective in defect identification and prevention
• Control charts prevent unnecessary process adjustments
• Diagnostic information
• Process capability information
 Rational Subgroups (Section 5.3.4
Montgomery)
• Sampling procedure to ensure that the variation within the
group is only due to chance causes.
• Lots from which the subgroups are chosen should be
homogeneous, e.g. same machine, same operator , same mold
cavity etc.
• Items of any one subgroup should be produced under essentially
same conditions.
• Instant time method: Parts in the subgroup are chosen in the
same time-instant. The next subgroup is picked after a certain
time interval. (Maximum variation among subgroup)
• Period of time method: Subgroups are sampled from parts
produced since the last sample was taken. It is used to make
decisions about acceptance of products produced since the last
inspection.
27.08.2024
In-control or Out of Control
Causes of Variation
Samples maybe out-of-control due to:
• Special Cause or Assignable Cause: Not inherent in the
process, does not affect all the time. It could be the use of
a wrong tool, tool damage, operator mistake, incorrect
measurement etc. Control charts are used to detect the
presence of special causes as soon as possible. Special
causes need to be removed to get the process back to
normality.
• Common Cause: Variability due to common or chance
causes, inherent to a process. It is inherent part of process
design and affects all time. Process need not be changed
due to common cause variations.
Xbar & R Charting
Subgp. X1 X2 X3 X4 Xbar R
• Step 1. (Startup) Collect data for 25. 1 20.50 3.10 2.10 4.00 7.43 18.40

Xbarbar is grand average, Rbar is... 2

3
1.20
5.40
2.40
2.20
2.40
2.30
1.40
0.20
1.85
2.53
1.20
5.20

• Step 2. (Startup) “Trial” limits: 4

5
1.10
1.40
11.00
6.50
3.10
2.20
1.50
6.50
4.18
4.15
9.90
5.10
6 2.30 2.30 0.30 19.40 6.08 19.10
• Step 3. (Startup) Find out of control 7 13.10 3.10 2.40 0.40 4.75 12.70

signals. Remove if assignable causes 8

9
0.60
1.60
2.10
1.60
3.30
13.10
5.30
0.50
2.82
4.20
4.70
12.60
are found 10 8.10 1.00 2.20 0.10 2.85 8.00
11 3.20 5.30 9.20 1.30 4.75 7.90
• Step 4. (Startup) Revise limits. 12 4.50 5.40 4.50 14.70 7.28 10.20
13 2.40 1.30 0.20 10.30 3.55 10.10

• Step 5. (Steady State) Plot and local 14

15
1.60
3.20
5.40
9.30
3.10
4.00
7.20
2.10
4.32
4.65
5.60
7.20
authority investigates if out-of-control 16 0.20 0.60 1.30 2.60 1.17 2.40

signals occur (can act). 17

18
1.30
2.00
1.30
6.10
5.10
5.10
0.30
5.20
2.00
4.60
4.80
4.10
19 2.30 1.10 6.10 5.20 3.67 5.00
20 2.40 4.60 0.60 6.20 3.45 5.60
21 3.00 1.60 6.50 1.20 3.07 5.30
22 5.60 14.10 6.50 2.30 7.12 11.80
23 0.50 1.10 2.10 1.10 1.20 1.60
24 1.40 3.00 4.40 2.20 2.75 3.00
25 2.30 2.20 1.40 2.40 2.08 1.00
Xbarbar= 3.86 Rbar=7.3
 Trial Limits
d2(n=4) = 2.059
est = 7.300/2.059 = 3.55

D2 = 4.698
10
D1 = 0.000
0 8
UCLXbar = Xbarbar + 3.0 ×
n 6
UCL
Signals on the R chart. Do 4 Xbar
detective work. Overnight
stays. Not fair to keep. 2

So remove. 0
1 6 11 16 21

21

16

11 UC
LR

1
1 6 11 16 21

27.08.2024
 Control Limits of Shewhart Control
Chart

• R chart
• 𝑈𝐶𝐿𝑅ത = 𝑅ത + 3𝜎𝑅ത ≈ 𝐷4 𝑅ത
• 𝐿𝐶𝐿𝑅ത = 𝑅ത − 3𝜎𝑅ത ≈ 𝐷3 𝑅ത
Derivation
1 𝑚 1
• 𝑋ത = σ𝑖=1 𝑋ത𝑖 = σ𝑚
𝑖=1 σ𝑛
𝑗=1 𝑋𝑖𝑗 Centre Line
𝑚 mn
• UCL and LCL these are 3 sigma
𝜎
ത
• 𝑋~𝑁 𝜇, 𝜎 , 𝑋~𝑁(𝜇, )
𝑛
𝜎
• UCL = 𝑋ത + 3𝜎𝑋ത = 𝑋ത + 3
𝑛
𝜎
• LCL = 𝑋ത − 3𝜎𝑋ത = 𝑋ത − 3
𝑛
Derivation
• Relative Range 𝑊 = 𝑅/𝜎 a random variable
• Parameters of distribution of 𝑊 depend on sample size 𝑛
• Mean of 𝑊 is 𝑑2
ത 2 as 𝑅ത is
• Estimator of 𝜎 is 𝜎ො = 𝑅/𝑑2 , we may use 𝜎ො = 𝑅/𝑑
the average range of 𝑚 preliminary samples
• 𝑈𝐶𝐿 = 𝑥ҧ + 3𝜎/ො 𝑛 = 𝑥ҧ + 3𝑅/(𝑑ത ഥ ഥ
2 𝑛) = 𝒙 + (𝑨𝟐 )𝑹
• 𝐶𝐿 = 𝑥ҧ
• 𝐿𝐶𝐿 = 𝑥ҧ − 3𝜎/ො 𝑛 = 𝑥ҧ − 3𝑅/(𝑑 ത ഥ ഥ
2 𝑛)) = 𝒙 − (𝑨𝟐 )𝑹
Derivation
• Standard deviation of 𝑊 is 𝑑3
• Estimator of 𝜎 is 𝜎ො = 𝑅/𝑑2
• Standard deviation of 𝑅 can be written as 𝜎𝑅 = 𝑑3 𝜎 as
𝑅 = 𝑊𝜎
ത 2
• 𝜎ො𝑅 = 𝑑3 𝑅/𝑑
• For R chart
ത
3𝑑 𝑅
• 𝑈𝐶𝐿 = 𝑅ത + 3𝜎ො𝑅 = 𝑅ത + 3 = 𝑫𝟒 𝑹ഥ
𝑑2
• 𝐿𝐶𝐿 = 𝑅ത − 3𝜎ො𝑅 = 𝑅ത − 3𝑑3 𝑅/𝑑
ത 2 = 𝑫𝟑 𝑹
ഥ
• 𝐶𝐿 = 𝑅ത
 Revised Control Limits
Discard out of control samples with assignable causes.

Revised control limits are calculated as below, for total

number of samples 𝑚 and number of defective samples 𝑑:

ത 𝑚𝑋ത −σ𝑑 𝑋ത 𝑑
𝑋𝑛𝑒𝑤 = = 𝑋ത0
𝑚−𝑑
𝑚𝑅ത −σ𝑑 𝑅𝑑
𝑅ത𝑛𝑒𝑤 = = 𝑅ത0
𝑚−𝑑
𝑅ത0
𝜎0 = ( 𝑑2 can be found in table)
𝑑2
𝑈𝐶𝐿𝑋ത = 𝑋ത0 + 𝐴 𝜎0 ; 𝐿𝐶𝐿𝑋ത = 𝑋ത0 − 𝐴 𝜎0
𝑈𝐶𝐿𝑅ത = 𝐷2 𝜎0 ; 𝐿𝐶𝐿𝑅ത = 𝐷1 𝜎0 ;

27.08.2024
X-bar and R chart
• 𝑋ത chart monitors between sample variability, 𝑅 chart
monitors within sample variability
• To design 𝑋ത − 𝑅 chart, the following must be specified:
• Sample size
• Control limit width
• Frequency of sampling
• 𝑋ത chart is capable to signal moderate to large process
shifts (2𝜎 or larger)
• 𝑅 chart is relatively insensitive to shift in process standard
deviation for small samples 𝑒. 𝑔. 𝑛 = 5
Error in Making Inference
• Type I Error: This error results from inferring a process is
out of control when it is not. It is denoted by 𝛼. This
happens due to chance causes, when a control charts falls
outside control limits. For 3𝜎 limits, probability of type I
error is 0.0027.
• Type II Error: This error results from inferring a process is
in control when it is out of control. It is denoted by 𝛽. This
can happen when the process mean or the process
variability or both have changed.
Xbar & R Charting
Subgp. X1 X2 X3 X4 Xbar R
• Step 1. (Startup) Collect data for 25. 1 20.50 3.10 2.10 4.00 7.43 18.40

Xbarbar is grand average, Rbar is... 2

3
1.20
5.40
2.40
2.20
2.40
2.30
1.40
0.20
1.85
2.53
1.20
5.20

• Step 2. (Startup) “Trial” limits: 4

5
1.10
1.40
11.00
6.50
3.10
2.20
1.50
6.50
4.18
4.15
9.90
5.10
6 2.30 2.30 0.30 19.40 6.08 19.10
• Step 3. (Startup) Find out of control 7 13.10 3.10 2.40 0.40 4.75 12.70

signals. Remove if assignable causes 8

9
0.60
1.60
2.10
1.60
3.30
13.10
5.30
0.50
2.82
4.20
4.70
12.60
are found 10 8.10 1.00 2.20 0.10 2.85 8.00
11 3.20 5.30 9.20 1.30 4.75 7.90
• Step 4. (Startup) Revise limits. 12 4.50 5.40 4.50 14.70 7.28 10.20
13 2.40 1.30 0.20 10.30 3.55 10.10

• Step 5. (Steady State) Plot and local 14

15
1.60
3.20
5.40
9.30
3.10
4.00
7.20
2.10
4.32
4.65
5.60
7.20
authority investigates if out-of-control 16 0.20 0.60 1.30 2.60 1.17 2.40

signals occur (can act). 17

18
1.30
2.00
1.30
6.10
5.10
5.10
0.30
5.20
2.00
4.60
4.80
4.10
19 2.30 1.10 6.10 5.20 3.67 5.00
20 2.40 4.60 0.60 6.20 3.45 5.60
21 3.00 1.60 6.50 1.20 3.07 5.30
22 5.60 14.10 6.50 2.30 7.12 11.80
23 0.50 1.10 2.10 1.10 1.20 1.60
24 1.40 3.00 4.40 2.20 2.75 3.00
25 2.30 2.20 1.40 2.40 2.08 1.00
Xbarbar= 3.86 Rbar=7.3
 Trial Limits
d2(n=4) = 2.059
est = 7.300/2.059 = 3.55

D2 = 4.698
10
D1 = 0.000
0 8
UCLXbar = Xbarbar + 3.0 ×
n 6
UCL
Signals on the R chart. Do 4 Xbar
detective work. Overnight
stays. Not fair to keep. 2

So remove. 0
1 6 11 16 21

21

16

11 UC
LR

1
1 6 11 16 21

27.08.2024
 Revised Limits
Phase I Phase II

d2(n=4) = 2.059
9
est = 6.300/2.059 = 3.062 8
D2 = 4.698 7 Range of Times (4
patients, not including one Revised
D1 = 0.000 6 "extra-long" overnight UCLXbar
5 Xbar
60 = the process capability 4 CLXbar
= 18.4 hours 3 LCLXbar

(range for hospital) 2

1 Evidence of
No specs. so no Cpk
0
1 6 11 16 21 26 31
Subgrou
16
Range of Times (4 patients, not including one "extra-

11
UCLR
R
CLR
6

1
1 6 11 16
Subgroup 21 26 31

27.08.2024
 Revision Formula
Xbarbar,revised = [25 Xbarbar,trial – (removed)] ÷ (25 – # removed)]
(25*3.86 – 7.43 – 6.08) ÷ 23 = 3.6 hours
Makes a small difference but it is fair as long as we clarify we are not
considering overnight stays
Only remove if an assignable cause was found and eliminated.
Otherwise leave data in (common or chance causes).

Process capability: Measurement of the Common Cause variation/system

quality:
→6 0 derive fairly measures common cause variation
→18.4 hours (you can pretty much count on range less than that)

27.08.2024
X bar and S chart
• It is occasionally desirable to monitor process standard
deviation directly, rather than indirectly as done in 𝑅
chart.
• 𝑋ത and 𝑆 chart are preferable when
• The sample size 𝑛 is moderately large for 𝑛 > 10 or 12.
• The sample size 𝑛 is variable
• The unbiased estimator of population variance 𝜎 2 is
sample variance 𝑠 2
σ𝑛 2
2 𝑖=1 𝑖 𝑥 − 𝑥ҧ
𝑠 =
𝑛−1
• The sample sd 𝑠 estimates 𝑐4 𝜎, sd of 𝑠 is 𝜎 1 − 𝑐42
*
Xbar and S Chart
• Since 𝐸 𝑠 = 𝑐4 𝜎 the center line is 𝑐4 𝜎. The 3 sigma limits
of the s-chart is given by
𝑈𝐶𝐿 = 𝑐4 𝜎 + 3𝜎 1 − 𝑐42
𝐶𝐿 = 𝑐4 𝜎
𝐿𝐶𝐿 = 𝑐4 𝜎 − 3𝜎 1 − 𝑐42
X bar and S chart with Sample Estimators
• Consider 𝑠/𝑐4 as an unbiased estimator of 𝜎
• For 𝑚 preliminary samples with sd 𝑠𝑖 , the average of m
1 𝑚
standard deviation is given by 𝑠ҧ = σ𝑖=1 𝑠𝑖
𝑚
3𝑠 ҧ
𝑈𝐶𝐿 = 𝑠ҧ + 1 − 𝑐42 = 𝐵4 𝑠ҧ
𝑐4
𝐶𝐿 = 𝑠ҧ
3𝑠ҧ
𝐿𝐶𝐿 = 𝑠ҧ − 1 − 𝑐42 = 𝐵3 𝑠ҧ
𝑐4

3 3
Where 𝐵3 = 1 − 1 − 𝑐42 and 𝐵4 = 1 + 1 − 𝑐42
𝑐4 𝑐4
X bar and S chart with Sample Estimators
• Control limit for corresponding 𝑋ത chart is given by
3 𝑠ҧ
𝑈𝐶𝐿𝑋ത = 𝑥ҧ + = 𝑥ҧ + 𝐴3 𝑠ҧ
𝑐4 𝑛
𝐶𝐿𝑋ത = 𝑥ҧ
3𝑠 ҧ
𝐿𝐶𝐿𝑋ത = 𝑥ҧ − = 𝑥ҧ − 𝐴3 𝑠ҧ
𝑐4 𝑛
 Probability of an Xbar False Alarm
Assume that the subgroups are rational (skipping and
representative of a homogeneous group) and the
system is under control (no assignable causes so
distribution is the same), what is probability of a false
alarm on the next subgroup from the Xbar chart?
CLT → Xbar ~ N[,/sqrt(n)]
1-Pr{LCLXbar =  – 3/sqrt(n) ≤ Xbar ≤ UCLXbar =  +
3/sqrt(n)} =1- Pr{-3 ≤ Z ≤ 3} = 1-2 Pr{Z ≤ -3} = 0.0027
Even if you are doing everything correctly, you have a
0.0027 chance of a false alarm.
(Average Run Length (ARL) in control 1÷0.0027= 370.4)

27.08.2024
Average Run Length (ARL)
• To measure the performance of a control chart, ARL is used.
• ARL denotes the number of samples, on average, required to
detect and out of control signal.
• If 𝑃𝑑 is the probability that a process is out of control then run
length is 1 with probability 𝑃𝑑 , 2 with probability 1 − 𝑃𝑑 𝑃𝑑 ,
3 with 1 − 𝑃𝑑 2 𝑃𝑑 . Hence
∞ 𝑗−1 𝑃𝑑 1
𝐴𝑅𝐿 = σ𝑗=1 𝑗 1 − 𝑃𝑑 𝑃𝑑 = 2 =
1− 1−𝑃𝑑 𝑃𝑑
• For a process in control, 𝑃𝑑 is 𝛼 (probability of type I error)
• For an in-control process, ARL should be as large as possible.
• For an out of control process 𝑃𝑑 = 1 − 𝛽. 𝛽: Probability of
1
Type II error. 𝐴𝑅𝐿 =
1−𝛽
• For out of control process, ARL should be as small as possible.
ARL Curve

Source: QCI, Amitava

Mitra
 Analysis of Pattern in Control Chart
• Process should be investigated when there is non-random
pattern in control chart
• Most sample averages
are below centre line !
• Continuous rise (run-up)
or fall (run-down)

• A run length of 8, or consecutive

8 points above or below centre
line may indicate out-of-control
• Cycles are another type of
pattern 
 Analysis of Pattern in Control Chart

• A mixture pattern when

most points are near
control limits, result of two
or more distributions

• A shift in the process may occur

due to introduction of new
operator, material, machine,
inspection method etc.

Analysis of Pattern in Control Chart
• A trend occurs when there
is continuous deterioration
of tools. In chemical
processes they occur due
to settling or separation of
components.

• Stratification is when points

cluster artificially near centre
line. This may happen when
rational subgrouping is not
done.
Rules of Identifying Out of Control Points
1. If single point plots are outside control limits
2. If 2 out of 3 consecutive points plots fall outside 2𝜎
warning limits on the same side of centre line
3. If 4 out of 5 consecutive points fall beyond 1𝜎 limit on
the same side of centre line
4. If 9 or more consecutive points fall on one side of centre
line
5. If 6 or more consecutive points steadily increases or
decreases
Process Capability Analysis
1. Predicting how well the process will hold the tolerances
2. Assisting product developers/designers in selecting or
modifying a process
3. Assisting in establishing an interval between sampling for
process monitoring
4. Specifying performance requirements for new equipment
5. Selecting between competing suppliers and other aspects of
supply chain management
6. Planning the sequence of production processes when there is
an interactive effect of
processes on tolerances
7. Reducing the variability in a process
Process Capability
• Assumptions:
• The quality characteristic has normal distribution
• Process is in statistical control
𝑈𝑆𝐿−𝐿𝑆𝐿
෢
• Process Capability Ratio (PCR) : 𝐶𝑝 =
6ෝ
𝜎
1
•𝑃= 100 gives the percentage of the specification band
𝐶𝑝
used by the process.
𝑈𝑆𝐿−𝜇 𝜇−𝐿𝑆𝐿
• 𝐶𝑝𝑘 = min 𝐶𝑝𝑢 = , 𝐶𝑝𝑙
= is used to determine
3𝜎 3𝜎
if the process is centered.
• Generally, if 𝐶𝑝 = 𝐶𝑝𝑘 , the process is centered at the
midpoint of the specifications, and when 𝐶𝑝𝑘 < 𝐶𝑝 the
process is off center
Process Capability
Recommended Values
Is 𝐶𝑝𝑘 Adequate?
Is 𝐶𝑝𝑘 Adequate?
• The process capability ratio 𝐶𝑝𝑘 was initially developed
because 𝐶𝑝 does not adequately deal with the case of a
process with mean 𝜇 that is not centered between the
specification limits.
• However, 𝐶𝑝𝑘 alone is still an inadequate measure of process
centering.
• Both processes A and B have 𝐶𝑝𝑘 = 1.33, yet their centering is
clearly different.
• For process A, 𝐶𝑝𝑘 = 𝐶𝑝 = 1.33, implying that the process is
centered, whereas for process B, 𝐶𝑝 = 2.67 > 𝐶𝑝𝑘 = 1.33,
implying that the process is off-center.
• For any fixed value of 𝜇 in the interval from LSL to USL, 𝐶𝑝𝑘
depends inversely on 𝜎 and becomes large as 𝜎 approaches 0.
Process Capability 𝐶𝑝𝑚
• A process capability ratio that is a better indicator of centering
𝑈𝑆𝐿−𝐿𝑆𝐿
𝐶𝑝𝑚 =
6𝜏
• 𝜏 is the squared root of expected squared deviation from
target
1
𝑇 = (𝑈𝑆𝐿 + 𝐿𝑆𝐿)
2
𝜏 = 𝐸[ 𝑥 − 𝑇 ] = 𝐸[ 𝑥 − 𝜇 2 ] + 𝜇 − 𝑇 2
2 2
= 𝜎2 + 𝜇 − 𝑇 2

𝑈𝑆𝐿 − 𝐿𝑆𝐿 𝐶𝑝
𝐶𝑝𝑚 = =
6 𝜎2 + 𝜇 − 𝑇 2 1 + 𝜉2
𝜇−𝑇
Where 𝜉 =
𝜎
Process Capability 𝐶𝑝𝑚
• A necessary condition for 𝐶𝑝𝑚 ≥ 1 is
1
𝜇−𝑇 < (USL − LSL)
6

Thus, a given value of 𝐶𝑝𝑚 places a constraint on the

difference between 𝜇 and the target value T.
Type I and Type II Error in CC
• Type I Error (𝛼): Detecting a shift in process when there is
no shift (False Alarm)
• Wider the control limits, lower is the probability of Type I Error
• Type II Error (𝛽): Not detecting a shift in process when
there is a shift
• Closer the control limits, smaller is the probability of Type II
Error
• Decreases with increase in sample size
• OC-curve is used to as a visual tool to analyze the change in
probability of Type II Error with the change in process
parameter.
• Sample size should be chosen judiciously, so that a
balance in probabilities of Type I and Type II Error is
maintained.
OC-curve for 𝑥ҧ chart
OC-curve for 𝑥ҧ chart
• Consider the process mean has shifted from 𝜇0 to
• 𝜇0 + 𝑘𝜎
• 𝛽 = 𝑃 𝑈𝐶𝐿 ≤ 𝑥ҧ ≤ 𝐿𝐶𝐿 𝜇 = 𝜇1 = 𝜇0 + 𝑘𝜎)
𝑈𝐶𝐿−(𝜇0 +𝑘𝜎) 𝐿𝐶𝐿− 𝜇0 +𝑘𝜎
•𝛽=Φ 𝜎 −Φ 𝜎
𝑛 𝑛
𝐿𝜎 𝐿𝜎
𝜇0 + −(𝜇0 +𝑘𝜎) 𝜇0 − − 𝜇0 +𝑘𝜎
𝑛 𝑛
=Φ 𝜎 −Φ 𝜎
𝑛 𝑛

𝛽 = Φ 𝐿 − 𝑘 𝑛 − Φ −𝐿 − 𝑘 𝑛
 Attribute Control
Charts, I-MR Chart
Prof. Sayak Roychowdhury
Control Charts
P-chart (fraction of non-conforming)
• Step 1. (Startup) Obtain the total fraction of nonconforming units or systems
using 25 rational subgroups each of size n.
• Step 2. (Startup) Calculate “trial” limits:
ҧ
𝑝(1− 𝑝)ҧ
• 𝑈𝐶𝐿 = 𝑝ҧ + 3 (Minitab>Stat>Control charts>Attribute Charts>P)
𝑛
ҧ
𝑝(1− 𝑝)ҧ
• 𝐿𝐶𝐿 = 𝑝ҧ − 3
𝑛

• Step 3. (Startup) Identify all the periods for which p = fraction nonconforming in
that period and p < LCLtrial or p > UCLtrial. Investigate, remove if unfair.
• Step 4. (Startup) Calculate the total fraction nc. using remaining. New 𝒑
ഥ=
“process capability”. Calculate revised limits.
• Step 5. (Steady State) Plot the fraction nonconforming, pj, for each period j and
alert designated local authority if out-of-control signals occur.
https://www.spcforexcel.com/knowledge/attribute-control-charts/p-control-charts#example

https://www.spcforexcel.com/knowledge/attribute-control-charts/p-control-
https://www.spcforexcel.com/knowledge/attribute-control-charts/p-control-charts#example

charts#example
P-chart

New Trainee Training implemented

P Chart of nc
P Chart of nc
1
0.30 0.12
UCL=0.1169

1 0.10
0.25
1
0.08

Proportion
0.20
0.06
_
Proportion

P=0.0509
0.15 UCL=0.1555 0.04

0.02
0.10
_
P=0.076 0.00 LCL=0

0.05 1 3 5 7 9 11 13 15 17 19 21
Sample
Results exclude specified rows: 14:16
0.00 LCL=0

1 3 5 7 9 11 13 15 17 19 21 23 25
Sample
Capability = 0.0909
np-chart (# of non-conforming)

• Subgroup size needs to be constant

• Calculate “trial” limits:
• 𝑈𝐶𝐿 = 𝑛𝑝ҧ + 3 𝑛𝑝(1 ҧ − 𝑝)ҧ
• 𝐶𝐿 = 𝑛𝑝ҧ
• 𝐿𝐶𝐿 = 𝑛𝑝ҧ − 3 𝑛𝑝(1 ҧ − 𝑝)ҧ
Variable Sample Size for Non conforming
Attribute Control Charts
1. Variable Control Limits:
ҧ
𝑝(1− 𝑝)ҧ ҧ
𝑝(1− 𝑝)ҧ
𝑈𝐶𝐿 = 𝑝ҧ + 3 , 𝐿𝐶𝐿 = 𝑝ҧ − 3
𝑛𝑖 𝑛𝑖
2. Control limits based on average sample size:
ҧ
𝑝(1− 𝑝)ҧ ҧ
𝑝(1− 𝑝)ҧ
𝑈𝐶𝐿 = 𝑝ҧ + 3 , 𝐿𝐶𝐿 = 𝑝ҧ − 3 where
𝑛ത 𝑛ത
σ𝑚
𝑖=1 𝑛𝑖
𝑛ത = is the average sample size.
𝑚

3. Standardized control chart:

𝑝𝑖 −𝑝ҧ
𝑍𝑖 = 𝑝ഥ 1−𝑝ഥ with control limits +3, and -3.
𝑛𝑖
OC-Curve for p-chart (7.2.4)
• Probability of Type II Error
• 𝛽 = 𝑃 𝑝Ƹ ≤ 𝑈𝐶𝐿 𝑝 − 𝑃 𝑝Ƹ ≤ 𝐿𝐶𝐿 𝑝 =
𝑃 𝐷 ≤ 𝑛𝑈𝐶𝐿 𝑝 − 𝑃 𝐷 ≤ 𝑛𝐿𝐶𝐿 𝑝
𝑛
• 𝑃 𝑥 ≤ 𝑛𝑈𝐶𝐿 𝑝 = σ𝑛𝑈𝐶𝐿
𝑥=0 𝑥
𝑝𝑥 1 − 𝑝 1−𝑥

𝑛
• 𝑃 𝑥 ≤ 𝑛𝐿𝐶𝐿 𝑝 = σ𝑛𝐿𝐶𝐿
𝑥=0 𝑥
𝑝𝑥 1 − 𝑝 1−𝑥
C-chart (count of non-conformities)
• Step 1. (Startup) Collect data
• Step 2. (Startup) Subgroup size is one inspected unit (e.g. 1 airplane, 1 case
of pencils)
• Step3.Calculate trial limits
• 𝑈𝐶𝐿𝑡𝑟𝑖𝑎𝑙 = 𝑐ҧ + 3 𝑐ҧ , 𝐶𝐿𝑡𝑟𝑖𝑎𝑙 = 𝑐ҧ (average count of non-
conformities)
• 𝐿𝐶𝐿𝑡𝑟𝑖𝑎𝑙 = 𝑀𝑎𝑥 {𝑐ҧ − 3 𝑐,ҧ 0}
• Step 3. (Startup) Find out of control signals. Remove if unfair.
• Step 4. (Startup) Revise limits Revised 𝒄ത is process capability
• Step 5. (Steady State) Plot and local authority investigates if out-of-control
signals occur (can act).

(Minitab>Stat>Control charts>Attribute Charts>C)

U-chart (average count of non-conformities/unit)
• Step 1. (Startup) Collect data for m periods. 𝑢ത = σ𝑚 𝑚
𝑖=1 𝑐𝑖 / σ𝑖=1 𝑛𝑖 where 𝑐𝑖
is the count of defects (non conformities) in each subgroup
• Step 2. (Startup) Calculate “trial” limits for 𝑖 𝑡ℎ sample:
ഥ
𝑢
• 𝑈𝐶𝐿𝑡𝑟𝑖𝑎𝑙 = 𝑢ത + 3 , 𝐶𝐿𝑡𝑟𝑖𝑎𝑙 = 𝑢ത
𝑛𝑖
ഥ
𝑢
• 𝐿𝐶𝐿𝑡𝑟𝑖𝑎𝑙 = 𝑀𝑎𝑥 {ത
𝑢−3 , 0}
𝑛𝑖

• Step 3. (Startup) Find out of control signals. Remove if unfair.

• Step 4. (Startup) Revise limits ഥ is process capability
Revised 𝒖
• Step 5. (Steady State) Plot and local authority investigates if out-of-control
signals occur (can act).

(Minitab>Stat>Control charts>Attribute Charts>U)

U-chart
moving to a new
Average # nonconformities hospital one time
per patient (not usual or fair)
2.500
Average Demerits Per Patient

2.000

1.500 u
UCL
u-Bar
1.000 LCL

0.500

0.000
1 3 5 7 9 11 13 15 17 19 21 23 25
Subgroup Number
Demerit Chart
• Determine 4 classes of non-conformities, very serious (A), serious (B)
moderately serious (C), and minor (D). Assign weights 𝑤𝐴 , 𝑤𝐵 , 𝑤𝐶 and 𝑤𝐷 .
• For sample size 𝑛, let 𝑐𝐴 , 𝑐𝐵 , 𝑐𝐶 , 𝑐𝐷 denote the total number of defects of
each class.
• Determine standard non-conformities per unit 𝑢ത𝐴 , 𝑢ത 𝐵 , 𝑢ത 𝐶 , 𝑢ത 𝐷
• 𝐷 = 𝑤𝐴 𝑐𝐴 + 𝑤𝐵 𝑐𝐵 + 𝑤𝐶 𝑐𝐶 + 𝑤𝐷 𝑐𝐷
𝐷
• Demerits per unit is given by U =
𝑛
ഥ = 𝑤𝐴 𝑢ത𝐴 + 𝑤𝐵 𝑢ത 𝐵 + 𝑤𝐶 𝑢ത 𝐶 + 𝑤𝐷 𝑢ത 𝐷
• CL = 𝑈
2𝑢
𝑤𝐴 2𝑢
ഥ𝐴 +𝑤𝐵 ഥ𝐵 +𝑤𝐶2 𝑢 2𝑢
ഥ𝐶 + 𝑤𝐷 ഥ𝐷
• 𝜎0𝑢 =
𝑛
ഥ + 3𝜎0𝑢 𝐿𝐶𝐿 = 𝑈
• 𝑈𝐶𝐿 = 𝑈 ഥ − 3𝜎0𝑢
Chart Comparison
Method Advantages Disadvantages
Xbar & R Uses fewer inspections, gives Requires 2 or more
charting greater sensitivity charts for single type of
unit
p-charting Requires only go-no-go data, Requires many more
intuitive inspections, less sensitive
Demerit charting Addresses differences between Requires more
nonconformities inspections (but less than
p), less sensitive
u-charting Relatively simple version of Requires more
demerit charts inspections (but less than
p), less sensitive
c-charting Simple (u with n = 1) Requires more
inspections (but less than
p), less sensitive
Np-charting Simpler (equivalent to p just Same as p plus cannot
no ÷ n) have variable n
I-MR (Individual, Moving Range) Chart
• For some cases, rate of production is low, it is not feasible
for a sample size to be greater than 1.
• When testing process is destructive, and the cost of item
is high, then sample size might be chosen to be 1.
• If every unit is inspected, sample size is 1.
• Data comes slowly, so samples with elements produced at
long intervals creates problem for rational subgrouping.
I-MR chart (ch 6.4)
• Control chart for individual measurements
• Useful for detection of system stability
• 𝑅𝑖 = 𝑥𝑖 − 𝑥𝑖−1 (moving ranges)
• Trial Limits
𝑀𝑅
• 𝑈𝐶𝐿 = 𝑥ҧ + 3 ∗ (for n=2, 𝑑2 =1.128) 𝑈𝐶𝐿𝑅 = 𝐷4 𝑀𝑅, 𝐷4 = 3.267 𝑓𝑜𝑟 𝑛 = 2
𝑑2
• 𝐶𝐿 = 𝑥ҧ 𝐶𝐿𝑅 = 𝑀𝑅
𝑀𝑅
• 𝐿𝐶𝐿 = 𝑥ҧ − 3 ∗ 𝐿𝐶𝐿𝑅 = 0
𝑑2

• Revised Limits
• 𝜎0 = 0.8865𝑀𝑅0 , 𝑥0 = 𝑥ҧ𝑛𝑒𝑤
• 𝑈𝐶𝐿 = 𝑥0 + 3𝜎0 𝑈𝐶𝐿𝑅 = 3.686𝜎0
• 𝐿𝐶𝐿 = 𝑥0 − 3𝜎0 𝐿𝐶𝐿𝑅 = 0
• Minitab > Stat> Control charts > Variable charts for individuals > I-MR
• Researchers have indicated that MR chart cannot really provide additional
information on variability, MR values are not independent.
EWMA & CUSUM
Charts
Prof. Sayak Roychowdhury
Limitations of Shewhart Charts
• In Shewhart chart, the plotted point represents
information corresponding to last observation only.
• It does not use information from previous observations.
• This makes Shewhart charts insensitive to small shifts.
• They are less useful in phase II.
• Warning limits and patterns can be useful, but they
reduce the simplicity of the control charts
CUSUM Chart and EWMA Chart
• To detect small process shifts, CUSUM (Cumulative sum)
and EWMA (Exponentially weighted moving average)
charts are used as alternatives to Shewhart Control Chart
• These are good alternatives for phase II process
monitoring
• Sometimes process needs to be monitored when sample
size n =1, both CUSUM and EWMA charts work well in this
situation.
• EWMA charts are particularly robust against non-
normality (read section Montgomery 9.2.3).
CUSUM Chart
• First proposed by Page (1954)
• The CUSUM chart plots the quantity
𝑖

𝐶𝑖 = ෍(𝑥𝑗ҧ − 𝜇0 )
𝑗=1
Where 𝑥𝑗ҧ is the average of the 𝑗𝑡ℎ sample
𝜇0 is the target for process mean
Also applicable for 𝑛 = 1
• So CUSUM charts are particularly useful in chemical and
process industries and discrete part manufacturing,
where frequently subgroup size is 1.
CUSUM Chart (Ch 9.1 Montgomery)
• There are 2 ways to represent CUSUM charts, tabular method
and V-mask method. We will discuss tabular method.
• The tabular cusum works by accumulating derivations from 𝜇0
that are above target with one statistic 𝐶𝑖+ and accumulating
derivations from 𝜇0 that are below target with another
statistic 𝐶𝑖−
• 𝐶𝑖+ = max(0, 𝑥𝑖 − 𝜇0 + 𝐾 + 𝐶𝑖−1 +
) (Upper CUSUM)
• 𝐶𝑖− = max(0, 𝜇0 − 𝐾 − 𝑥𝑖 + 𝐶𝑖−1 −
) (Lower CUSUM)
• 𝐶0+ = 𝐶0− = 0
• 𝐶𝑖+ > 𝐻 or 𝐶𝑖− > 𝐻 indicate the process mean has shifted
• 𝐾 is called reference value, allowance or slack value, 𝐻 is
called decision interval
CUSUM Chart
• 𝐾 = 𝑘𝜎, typically halfway between 𝜇0 and out of control
mean 𝜇1 that we want to detect
𝜇1 −𝜇0
•𝐾= = 𝑘𝜎
2
• Note that 𝐶𝑖+ and 𝐶𝑖− accumulate deviations from the target
value 𝜇0 that are greater than K, with both quantities reset to
zero on becoming negative.
• If either 𝐶𝑖+ or 𝐶𝑖− exceed the decision interval H, the process
is considered to be out of control
• 𝐻 = ℎ𝜎
• 𝐾 and 𝐻 are chosen to provide good ARL performance
• Generally 𝑘 = 0.5, and ℎ = 4 𝑜𝑟 5 are chosen. Read (9.1.3)
CUSUM Chart
CUSUM Chart Parameter Values and ARL
Exercise
EWMA chart (ch 9.2 Montgomery)
• Control chart to detect small shift in the process, ideally used with individual observations.
• The exponentially weighted moving average is defined as
𝑧𝑖 = 𝜆𝑥𝑖 + 1 − 𝜆 𝑧𝑖−1 (𝑧0 = 𝜇0 𝑡𝑎𝑟𝑔𝑒𝑡 𝑜𝑟 𝑋ത )
• 𝜆 should be between 0.05 , 0.25 (use smaller 𝜆 for smaller shifts)
• Limits
• 𝑈𝐶𝐿 = 𝑋ത + 𝐿𝜎 𝜆/(2 − 𝜆)[1 − 1 − 𝜆 2𝑖 ]
• 𝐶𝐿 = 𝑋ത
• 𝐿𝐶𝐿 = 𝑋ത − 𝐿𝜎 𝜆/(2 − 𝜆)[1 − 1 − 𝜆 2𝑖 ]
• Use 𝜇0 (target mean) in place of 𝑋ത if given

• Usually 𝐿 is taken to be 3. For a steady state process

𝜆/(2 − 𝜆)[1 − 1 − 𝜆 2𝑖 ] becomes 𝜆/(2 − 𝜆)
ത 2 if 𝑅ത can be obtained. Sometimes
• 𝜎 can be estimated by process standard deviation, or 𝑅/𝑑
process history is used for estimation.
• EWMA is often used with Shewhart Chart, so that the combined chart can detect small shifts
and large shifts quickly enough.
• Minitab > Stat> Control charts > Time weighted charts> EWMA
EWMA chart
EWMA Parameters and ARL

• EWMA chart is fairly robust against non-normal distribution,

compared to Shewhart charts for individual measurement
(sec 9.2.3 Montgomery)