STAT 4007 Statistical Quality Control

Chapter 3 How the Statistical Process

Control (SPC) Works

Dr. Philip Lee

2014/15 Term 2

Reference: Chapter 5 of Statistical Quality Control , Douglas C. Montgomery, 6th

Edition
1
 Statistical Process Control (SPC) [From Ch1]
• Monitor on-line processes
• Control chart: Popular tool to determine if a process is in a
state of statistical control (i.e. in-control).
• Example: Production line producing N items per hour:
Batch #1 (9am):
Summary statistics
N items produced of the n items
n items Control chart of X
being x1 , s1 ,... a
sampled s
f
Batch #2 (10am): d
N items produced s
n items f
being x2 , s2 ,... d
sampled
x3
…

2
 Chapter Outline
• Section 3.1: Hypotheses: In Control and Out-of-control
Process
• Section 3.2: Control Charts
• Section 3.3: Choice of Control Limits, Sample Size and
Sampling Frequency
– Average Run Length (ARL)
– Average Time to Signal (ATS)
• Section 3.4: Phase I and Phase II of Control Chart Application

3
 Section 3.1
Hypotheses: In Control and Out-of-control Process

4
 In Control and Out-of-control Process
Variations in a process can be classified by either of the following:
1. Common Cause of Variation:
– Driven by natural variability (i.e. background noise)
– Cumulative effect of many small, unavoidable causes in each
production step (inherent in all process)
2. Special Causes of Variation:
– Arises from (i) improperly controlled machines, (ii) wrong tool,
(iii) defective raw material, (iv) operating errors, …
– Usually LARGE compared with the background noise
=> unacceptable to the process performance.
Definition: An in control process is a process that is operating with
common causes of variation ONLY.
Definition: An out-of-control process is a process that is operating
in the presence of special causes. 5
 Common and Special Causes of Variation
Special causes of variation would cause (1) a shift in mean,
and/or (2) an increase in variability of the quality characteristic
Process Shift in mean and
Out-of-Control increase in
variability:

Increase in
variability
Shift in mean:
Special cause of
Process in variation is present
control: Only
common cause
of variation

6
 Hypothesis Testing in SPC
1. Samples of products are obtained periodically
2. At each time point,
– Measure the quality characteristics of the sample
– Obtain the summary statistics of the quality characteristics
– Test the following hypotheses:
H0: Process in control
H1: Process out-of-control
Challenge: Too complicated to perform hypothesis testing one-
by-one over time
Solutions: Construct Acceptance Regions over time, and
compare them with the summary statistics obtained from the
sample => Control Chart
7
 Section 3.2
Control Charts

8
 Statistical Basis of the Control Chart
• Control Chart: Plot of summary statistic of a quality
characteristic for different samples over time
• Typical control chart: Contains (1) a center line (CL), (2) an
upper control limit (UCL) and (3) a lower control limit (LCL)
Typical Control Chart
quality characteristic
Sample Mean of a

Sample number (over time)

9
 Hypothesis Testing using Control Chart
Want to test H0:  = 0 vs H1:  ≠ 0
Consider x to be the y-value of the control chart, with
1. Center Line CL= 0 , and
2. UCL and LCL as the range of acceptance region for x .
Significance level α: We reject H0 at α, where
α = Pr(Type I error) = Pr(a data point outside the control limits)
quality characteristic
Sample mean of a

Acceptance
0 region with
significance
level α

Sample number 10
 Identification of In Control and Out-of-control Processes
If a point plots outside the control limits, we reject H0 and
conclude that the process is out-of-control.
Action: Investigation of special cause + corrective action to
eliminate the special cause
If all the points plot within the control limits, we accept H0
and conclude that the process is in control.
** However, if the points behave in a systematic or non-
random manner over time (next 5 pages), it could be an
indication that the process is out-of-control in the near future -
May want to stop the process and figure out the possible
causes.

11
 Systematic Pattern from the Control Charts
1. Cyclic Pattern: May due to the systematic change in
environment, such as
(a) temperature
(b) regular rotation of operators/machines
(c) fluctuation in voltage or pressure.

12
 Systematic Pattern from the Control Charts
2. A Mixture – Points mostly around the two control limits,
with relatively few points near the center line
• The pattern could be generated from two (or more)
distributions with different mean values
• Causes: (a) frequent process adjustment by the operator
(b) output products from several sources (e.g. parallel
machines) feeding into a common stream before sampling

13
 Systematic Pattern from the Control Charts
3. Shift in Process Level: May due to
(a) the introduction of new workers
(b) changes in raw material/machines
(c) change in inspection method/standard
(d) change in skill/attentiveness/motivation of the operator

14
 Systematic Pattern from the Control Charts
4. A Trend: May due to
(a) gradual deterioration of a tool/process component
(b) human causes. E.g, operator fatigue without the
presence of supervision

15
 Systematic Pattern from the Control Charts
5. Stratification – Points cluster around the center line.
• Lack of natural variability
• May due to incorrect calibration of control limits.
E.g. During Phase I, a sample of size 5 is obtained by taking
one observation from each of the five parallel processes –
which takes the between-sample variability into a sample.

16
 Shewhart Control Charts
Walter Shewhart (pronounced like "shoe-heart”)
• an American statistician, known as the father
of statistical quality control.
• worked in Bell Telephone Laboratories to
improve the reliability of the transmission
system – introduced control charts
• Shewhart control charts: The first and the
simplest kind of control charts.

17
 Shewhart Control Charts
Let w be a sample statistic (e.g. x ) of a quality characteristic.
Let μw and σw be the mean and standard deviation of W
Shewhart Control Charts: The center line (CL), upper control
limit (UCL) and lower control limit (LCL) can be expressed by
the following over all samples:

UCL  w  L w
CL  w Lσw

LCL  w  L w Lσw

where L is the “distance” of the control limits from the center

line, expressed in standard deviation units (e.g. L=3).
18
 Examples of Non-Shewhart Control Charts
LCL and UCL vary over time:

19
 Section 3.3
Choice of Control Limits, Sample Size
and Sampling Frequency

20
 Choice of Control Limits
Choice of Control Limits: Critical decisions in designing a
control chart:
• If the control limits are too far away from the center line,
– decrease the probability of a type I error α
i.e. Pr(out-of-control signal | Process in control)
– increase the probability of a type II error β
i.e. Pr(in control signal | Process out-of-control)
• If the control limits are too close to the center line,
– increase the probability of type I error α
– decrease the probability of type II error β

21
 Choice of Control Limits
• Standard Practice in the US: 3-sigma control limits
Assume that the quality characteristic is normally
distributed(*).
α =Pr(Type I error) = 2( 3) = 2(0.00135)= 0.0027
Interpretation: We expect 27 incorrect out-of-control
signals generated in every 10,000 points, sampled from an
in control process
Unless specified, we assume that all the control limits
are 3-sigma control limits for the rest of this course.

(*)  would increase (decrease) substantially if the

distribution is fat-tailed (thinner tailed). 22
 Choice of Sample Size and Sampling Frequency
Sample Size: related with the shape of the OC curve.
• If the shift in process mean (i.e. δ in the horizontal axis) is
small, then larger sample is needed for easier detection.
• If the process shift is relative large, smaller sample would
be sufficient
Operating Characteristic Curves
Pr(LCL< x-bar < UCL)

Easier to
n=5 detect small
shifts using
large samples
n=10
n=15

Deviation in Process mean δ 23

 Choice of Sample Size and Sampling Frequency
Sampling Frequency: Function of Sample Size and resources
• Ideal Scenario: High frequency with large sample
=> Too Expensive to implement!!!
• In Practice: Either
(1) small samples at higher frequency, or
(2) large samples at lower frequency.
Industry favors small samples at higher frequency:
Especially high-volume manufacturing process with many
possible types of special causes.
Quantities to calibrate sample size and sampling frequency
Average Run Length (ARL) and Average Time to Signal (ATS)
24
 Average Run Length (ARL)
Definition: Run length is the number of points required to obtain
an out-of-control indication (i.e. outside the control limits).
Let X be the random variable for the run length, then
X ~ Geometric(p)
where p is the probability that a point is outside the control
limits.

Probability Mass Function of X:

Pr( X  x)  p(1-p) x-1 , x  1, 2,

1 1 p
where  X  , and  X2 
p p2
25
 Average Run Length (ARL)
• Definition: Average Run Length (ARL) is the expected
number of points required to obtain an out-of-control
indication. That is, 1
ARL  E ( X ) 
p

where p = Pr(A point is outside the control limits)

• Example: Consider a x-bar chart with 3 limits.

When the process is in control (i.e. H0 is true),
p = Pr(Type I error) = α = 2Φ(-3) = 0.0027
=> ARL = 1/p = 1/0.0027 = 370
That it, even if the process is in control, we should expect
an out-of-control indication in every 370 samples (i.e.
false alarm) 26
 Average Run Length (ARL)
Criticism of using the ARL: The run length X~Geometric(p),
which is a very right-skewed distribution, and has a large
standard deviation when p is small
1 1 p 1
X  , and X   if p is small
p p p
Hence ARL itself is not a good quantity to summarize the run
length.
0.003 pmf of Geometric with p=0.0027

0.002 ARL=370

0.001

0
53
106

477
0

159
212
265
318
371
424

530
583
636
689
742
795
848
901
954
1007
1060
27
 Average Time to Signal (ATS)
• Definition: Average Time to Signal (ATS) is the expected
time (in hours) to obtain an out-of-control indication. That
is,
ATS  ARL  h
where h (in hours) is the sampling frequency.
• Example: Consider a normal 3-sigma control chart
(i.e. p=2(-3)=0.0027 and ARL=1/p=370). If the sampling
takes place every 30 minutes, then the average time to
signal when the process is in control is given by
ATS = 370x(0.5 hour)= 185 hours
In other words, we expect a false alarm (out-of-control
indication during an in control process) in every 185 hours.
28
 Average Time to Signal (ATS)
Example: In a production of wafers (晶圓), the target value of the
flow width is 1.5microns (10-6 metre). The process is currently out-
of-control, with the mean shifted to 1.725microns.
Suppose that sample size n=5 and sampling frequency = 1 hour.
(a) What is average run length and average time to signal of the
process?
Pr(LCL< x-bar < UCL)

n=5

n=10

n=15

Process mean (microns) 29

 Average Time to Signal (ATS)
Solutions: (a) From the OC curves (with n =5),
Pr( x between control limits) = 0.35
=> p = Pr( x exceeds control limits) = 1-0.35 = 0.65
 ARL = 1/p = 1/0.65 = 1.54
 ATS = ARL x h = 1.54 (1) = 1.54 hours
(b) We think that the above ATS is unacceptable because
mean flow width of 1.725 microns will result in excessive scrap
costs and possible manufacturing problems later on.
If we are allowed to either
(i) double the sampling frequency, or
(ii) double the sample size n.
Which one would be more effective in terms of the ATS?
30
 Average Time to Signal (ATS)
Solutions: (i) Double Sampling Frequency (i.e. sample every
30 minutes)

(ii) Double Sample Size (n increased to 10)

31
 Average Time to Signal (ATS)
Solutions (cont’d): By comparing (i) ATS = 0.77 hour and (ii)
ATS = 1.09 hour, we conclude that the increase in sampling
frequency * would be more effective in detecting an out-of-
control condition.

* Additional factors to be considered when the sampling

frequency is increased:
– the cost of sampling
– the losses associated with allowing the process to
operate out-of-control
– the rate of production

32
 Section 3.4
Phase I and II of Control Chart Application

33
 Phase I and Phase II of Control Chart Application
Phase I (Testing Phase)
• Large sample of process data is
gathered for retrospective (i.e.
backward looking) analysis,
Phase II
• Construct trial control limits
(Production Phase)
• If there are points outside the trial
control limits, • Stable Control Limits
from in Phase I to
– Investigate potential special
monitor the process
causes, and eliminate those
causes from the process
– exclude those points and
obtain a revised control limits
• Repeat the previous step until
stable control limits are obtained.

34
 Implementation of Phase I
During Phase I,
• Shewhart control charts are very effective because
– they are easy to be constructed and interpreted, and
– they are effective in detecting both large, sustained shifts
in the process parameters and outliers
– patterns on Shewhart control charts are easy to interpret
and have physical meaning.
• Average run length (ARL) is not a good performance
measure: More interested in probability of an special cause
to be detected, but not the occurrence of false alarms.

35
 Implementation of Phase II
• Special causes of variation are smaller, as most of the ugly
sources of variability have been removed during phase I.
• Average run length (ARL) is a good measure to evaluate
the performance of a control chart.
• Shewhart control charts are less likely to be effective in
phase II because they are not very sensitive to small and
moderate size process shifts (Chapter on Process Capability)

36