MP4001 / MP4001P

Statistical Quality Control (SQC) concerns the use of

statistical methods and other problem-solving techniques to
improve
p
the q
qualityy off the p
products in the manufacturing
f
g
industry and service section.

QUALITY ASSURANCE AND MANAGEMENT

STATISTICAL QUALITY CONTROL (SQC)

CONTENTS
Dr Wu Zhang
MAE NTU
MAE,

1 QC seven tools
2 Control charts

Office: N3.2-02-14,
N3 2 02 14 Tel: 67904445,
67904445 email:
[email protected]

3 Process capability

TEXT
Montgomery,
g
y, D. C.,, ((2009),
), Introduction to Statistical
Quality Control, John Wiley & Sons.
1

4 Gage repeatability and reproducibility

5 Acceptance sampling plan

2 Control Charts
1 QC Seven Tools

2.1 Introduction

2-1

2 2 Control Charts for Variables

2.2

2 25
2-25

1.1 Histogram (Bar Chart)

1-2

2.2.1 x & R Charts

2-26

1.2 Check Sheet

1-12

2.2.2 Rational Subgroups

2-69

2 2 3 Cusum (Cumulative Sum) Chart

2.2.3

2-72

1.3 Pareto Chart

1-14

2.2.4 Runs Rules

2-88

1.4 Cause and Effect Diagram

1-16

2.3 Control Charts for Attribute

2-94

1.5 Defect Concentration Diagram

1-18

1.6 Scatter Diagram

1-20
3

2 3 1 np Chart
2.3.1

2-94
2
94

2.3.2 c Chart

2-115
4

3 Process Capability

5 Acceptance Sampling Plan

3.1 Process Capability Ratio Cp

3-11

3 2 Process
3.2
P
C
Capability
bili R
Ratio
i Cpk

3 34
3-34

3 3 Process Capabilit
3.3
Capability Ratio Ckm

3 47
3-47

4 Gage Repeatability and Reproducibility

4.1 Gage (Measurement) Error

44-11

4.2 The Gage

g Capability
p
y Cgage

4-19
5

5.1 Single Sampling Plan

5-7

5.2 Double Sampling Plan

5-29

5.3 Rectifying Inspection

5-51

5.4 Sampling Plan Standard

5-65

1.1 Histogram
g
((Bar Chart))
1 QC (Quality Control) SEVEN TOOLS
QC seven tools
tools is a powerful collection of problemproblem
QC
solving tools useful in achieving product quality
improvement. Most of them (except the control chart) are
very easy to use, but quite effective for many applications.
(1)
(2)
(3)
(4)
(5)
(6)
((7))

Histogram
Check sheet
Pareto chart
Cause and effect diagram
Defect concentration diagram
Scatter diagram
Control chart

It is a very useful graphical technique for summarizing

and presenting data.
Table 1-1 presents the observations of the inside
diameter for the piston-ring (individual observations
are different
diff
t from
f
eachh other).
th )
Number of samples:

25

Number of observations per sample (sample size):

Total
o a number
u be of
o observations:
obse va o s:

25
5 5 = 1255

Figure 1-1 is the corresponding histogram, which consists

of 13 bars of equal width, between the minimum (73.967)
and maximum (74.030) values of the diameter.

Table 1-1 Forged

g ppiston-ring
g inside diameter (mm)
(
)

Usually, the number of bars in a histogram is between 4

and 20.
The height of a bar is equal to the number of the diameter
values inside the interval under that bar. For example,
1st bar: 73.965 73.970,
2nd bar:73
bar:73.970
970 73.975,
73 975
4th bar: 73.980 73.985,

one diameter value

zero diameter value
eight diameter values

The higher
Th
hi h the
h bbar, the
h more lik
likely
l that
h a value
l will
ill fall
f ll
in the corresponding interval.
3

The diameter is a random variable. Three properties of

the probability distribution of is critical to the quality
of the product:
1.Shape of the distribution (normal or uniform ?)
2. Location (or average, mean value ())
3. Scatter (or spread, variability, standard deviation ())
Such information cannot be easily perceived from Table
1-1,, but can be clearlyy illustrated byy the histogram
g
in Fig
g
1-1. Because the histogram is similar to the probability
distribution curve.
Figure 1-1 Histogram for piston-ring diameter data

(F2-4 p39)
5

1. The histogram shows that the distribution of ring diameter

i roughly
is
hl symmetric
t i andd unimodal
i d l (a
( single
i l maximum
i
value),
l )
similar to a normal distribution.

Example
p 1-1
The specification for the piston-ring diameter is: 74 0.020 mm

2. The average is close to 74 mm.

Upper specification limit: USL = 74 + 0.020 = 74.020

3. The variability
y is qquite high,
g as the min. diameter value is
as small as 73.967 mm, and the max. value is as large as
74.030 mm. The range is (74.030 73.967) = 0.063.

Lower specification limit: LSL = 74 - 0.020 = 73.980

From the viewpoint of product quality, the small the

variabilityy (or
( the range),
g ), the better the quality
q
y is.

From the histogram in Fig 1-2, we can easily find that

One piston-ring is defective due to excessively small .

This information is very useful for determining whether the

manufactured diameter can satisfy the design specifications
and whether a process improvement program should be
carried out.

Three piston-rings are defective due to excessively large .

7

The defective rate is (1 + 3) / 125 = 3.2%

If we can improve the process by

1 Training operators for better skill, or
2 Purchasing machines with higher precision
The variation of will be reduced (Fig 1-3), that is, the
observed diameter values cluster more closely
y to the
nominal value of 74mm.
As a result, the defective rate may be reduced to zero.
Figure 1-2 Histogram with specification limits

10

1.2 Check Sheet

Check sheet is also very useful in collecting either historical or
current operating data about a process. Figure 1-4 is a check
sheet
h t for
f investigating
i
ti ti the
th various
i
types
t
off defects
d f t that
th t occurredd
on a tank.
The time-oriented summary at the bottom is particularly
valuable in looking for trends or other patterns of defects.
For example, extraordinarily more defects are found in June and
July of 1988 (also in Jan and Feb of 1989). People are alarmed
to conduct investigation, and find out and remove the roots of
the problems in these months.

Figure 1-3 Histogram for improved process

11

Also, the category-oriented summary at the right-hand-side

can tell which types of defects occur more frequently, or more
dominant.
12

1.3 Pareto Chart

Pareto chart is a graph of the frequency distribution of
defects arranged by category. The defect type occurring
most frequently is plotted at the most left-hand side.
side
Through this chart the user can quickly and visually
identify
id
tif the
th mostt frequently
f
tl occurring
i types
t
off defects.
d f t
Thus the causes of these defect types will be identified and
attacked first.
first
In Figure 1-5, 129 out of the 166 defects (78%) are
attributable to the five most dominant types.
(36 + 34 + 29 + 17 + 13) / 166 = 129 / 166 = 78%

Figure 1-4 A check sheet to record defects on a tank

13

(f4-15 p151)

So, if the causes of these 5 types of defects can be found

So
14
and removed, majority of the defects are eliminated.

1.4 Cause and Effect Diagram

The cause andd effect
Th
ff diagram
di
is
i a systematic
i approachh to
discover the potential causes of the defects.
Figure 1-6 shows the cause and effect diagram for the tank
defect problem. The diagram consists of the defect box, the
center line and multi-level potential cause categories.
y, a team is formed to build the cause and effect
Usually,
diagram through brainstorming. The team members include
engineers and experts from all of the relevant departments.
After the cause and effect diagram has been built, we may
have to rank the causes in order to identify those that are
most likely to impact the problem, and then take the
corrective actions correspondingly.
p
gy
Figure 1-5 Pareto chart of the tank defect data

15

(f4-16 p153)

16

1.5 Defect Concentration Diagram

A defect concentration diagram is the picture of a unit, showing
all relevant views. Then the various types of defects are drawn on
the picture, and the diagram is analyzed to determine the
potential causes of the defects from the location of the defects.
Figure 1-7 presents a defect concentration diagram of a
refrigerator. Surface-finish defects are identified by the dark
shaded areas on the refrigerator. It is found that all the defects
appear at the middle of the refrigerator. A follow-up investigation
discloses that some problem during the transportation of the unit
is responsible for the defects. Because, when moving the unit, a
b lt is
belt
i used
d around
d the
th middle
iddl off the
th refrigerator,
fi
t and
d this
thi belt
b lt is
i
made of abrasive material!
Figure 1-6 Cause and effect diagram for the tank defect problem
17

Using other belts with proper materials will solve this problem.
18

(f4-18 p156)

1.6 Scatter Diagram

The scatter diagram is a useful plot for identifying a potential
relationship between two variables Y and X. Usually, X indicates
the cause of a quality problem; and Y is the quality result.
, y (yi, xi), for i
Data are collected in ppairs on the two variables,say
= 1, 2, , n. Then, each yi is plotted against the corresponding xi
in the scatter diagram.
Figure 1-8 shows a scatter diagram relating defective rate Y
against the temperature X in a process. The scatter diagram
indicates a strong positive correlation between Y and X; that is, as
X is increased,, Y also increases.

Figure 1-7 Surface finish defects on a refrigerator

In a quality improvement program, X should be decreased as

much as possible in order to cut down the defective rate
rate.
(f4-19 p156)

19

20

Figure 1-8 A scatter diagram

(f4-21 p158)

21

2 CONTROL CHARTS
2.1 Introduction

A process that is operating with only the unavoidable natural

variability is said to be in control. An in-control
in control status is
considered as a normal and acceptable status.

Natural variability
Any production process, regardless of how well designed or
how carefully maintained, will generate a certain amount of
i h
inherent
or natural variability.
i i i This
hi type off variability
i bili may
be reduced, but cannot be completely eliminated.
The natural variability is the cumulative effect of many small,
unknown and essentially unavoidable causes,
unknown,
causes i.e.,
i e the chance
causes. Examples: ambient temperature, humidity, dust,
sunshine, vibration of a nearby machine, nearby high way,
oscillation of electrical network.
1

It means that when a process is in control, there is still

variability, but very small.
It is inappropriate to interfere or adjust an in-control process
with only natural variability. Because it may lead to overcorrection and introduce even larger variability and, thus, make
things worse.
thi
We should let the in-control
in control process continue to run.
run
2

System variability
Another kind of variability
variability, the system variability
variability, is generally
large in magnitude, and is incurred by the assignable causes

mean off x (or

( the
h average off x in
i engineering
i
i
terms). x is a random variable or a quality
characteristic (e.g.
(e g a dimension).
dimension) Each random
variable has a probability distribution as in Fig 2-1.

standard deviation of x (approximately equal to R/6,

where R is the range of the probability distribution
off x, = R/6).
R/6) is
i a measure off the
th variability
i bilit off
x.

the value of when the process is in control

(usually, the nominal or target value of x).

the value of when the process is in control (0 =

R0 /6,, where R0 is the value of R when the p
process
is in control).

It usually arises from three sources:

(1) Operator errors: Using incorrect cutter, setting wrong speed.
(2) Defective raw material: erroneous or low quality material
(3) Improper
I
adjustment
dj
ffor machines:
hi
Adj i an in-control
Adjusting
i
l
process when there is only natural variability. It will turn the small
natural variability to large system variability
variability.
System variability represents an abnormal and unacceptable outof-control status. System variability should be signaled, identified
and removed as soon as possible. Then, the process goes back to
an in-control
i
l status with
i h only
l minor
i
naturall variability.
i bili
3

When = 0 and
Wh
d = 0, the
h variability
i bili off x is
i small,
ll it
i is
i
the natural variability. The process is considered in
control.
control
When 0 and/or 0, the variability of x is large,
large it
is the system variability. The process is considered out of
control.
Usually,
y a pprocesses will operate
p
in the in-control status
for most or a long time period.

Figure 2-1 In-control and out-of-control processes

Eventually, however, assignable causes will occur,

seemingly at random, resulting in a "shift" to an out-ofcontroll status (
( 0 and/or
d/ 0, or both).
b h)
5

(f5-1 p180)

In Figure 2-2, a machine tool is cutting a shaft.

Until time t1, the process is in control. There is only small natural
variability.
y As a result,, both the mean and standard deviation of
the process are at in-control levels, 0 and 0.
After time t1, the process is out of control.
control
At time t1, assignable cause 1 occurs (e.g. the breakage of
th cutter
the
tt off the
th machine
hi tool),
t l) its
it effect
ff t is
i to
t shift
hift the
th
process mean to a new value1 > 0.
At time t2, assignable cause 2 occurs, resulting in a process
standard deviation change, 1> 0 (even though, = 0).
At time t3, assignable cause 3 presents, resulting in 2 < 0
and 1 > 0.
7

It is desired to detect the out-of-control status at the first

time in order to minimize the number of defectives. Here,,
the control chart comes to play.

Figure 2-2 Chance and assignable causes of variation

(f4-1 p131)

10

A control chart has the following three usages:

A control chart is a graphical display of the sample points versus

the sample number or time.

1. On-line process control. Distinguishing between the incontrol and out-of-control status.

A sample is a group of n observed values x1, x2, , xn of a quality

characteristic x ((e.g.,
g , a diameter,, weight,
g , velocity).
y)

(1)If the process is in control, avoid over-interference,

leave the process alone.
alone
(2)If the process is out of control, a prompt signal
is given and the process is shut down immediately
before many defects are manufactured. Then, the
investigation
g
and corrective action mayy be undertaken
to resume the process to an in-control status.

A statistic is a function in terms of the n observed xi (i = 1, 2, ,

n).
n)

2.Estimating the parameters (e.g., 0 and 0) of the process.

3D
3.Determining
i i the
h process capability.
bili
11

A widely used statistic is the sample mean, denoted by x. Sample

mean is
i the
th average off the
th n values
l
off xi in
i a sample.
l It is
i a
measure of the central tendency of x, or the estimate of the mean
value of x.
x

x1 x2 xn n
x
xi / n
n
i1

(2-1)
(2
1)

In a control chart
chart, each sample point carries the value of the
12
statistic (e.g. x ) of a corresponding sample.

Table 2-1 Diameter of a shaft (mm) (n = 5)

Sample
No. x1
1
2
3
.

x2

x3

x4

x5

74.030 74.002 74.019 73.992 74.008 74.010

73 995 73.992
73.995
73 992 74.001
74 001 74.011
74 011 74.004
74 004 74.001
74 001
73.988 74.024 74.021 74.005 74.002 74.008
.
.
.
.
.
.

A control chart that is used to inspect the sample mean x is called

the x chart. On a x chart, each sample point carries or indicates
the x value of a sample.
Since, x, this control chart is actually to monitor the mean
(central tendency of the process distribution),
distribution) or to insure =
13
0.

Three Elements of a Control Chart

Figure 2-3 A control chart

(f5-3 p188)

14

Basic Operational Rules for a control chart:

(1)Sample size (n ) the number of observations in a sample

(2)S
(2)Sampling
li interval
i t
l (h) the
h time
i interval
i
l between
b
two samples
l
(3)Control limits:
Central line (CL):

target value of the statistic when a

process is in control (e
(e.g.,
g the statistic
of a x chart is x , and the target of x is 0
when a pprocess is in control. So,, for x
chart, CL = 0).

( ) If all sample
(1)
p p
points fall within the control limits LCL
and UCL, the process is thought in control (there is only
small, natural variability), see Fig. 2-3.
(2) If a sample point falls beyond the control limits
(below LCL or above UCL), the process is considered as
out of control (there is large, system variability), see Fig
24
2-4.

Upper control limit (UCL) upper limit of the sample points

Lower control limit (LCL) lower limit of the sample points
15

It is how a control chart distinguishes between the inin

control and out-of-control status --- depending on the
positions of the sample points relative to the control
limits LCL and UCL.

16

The values of the control limits LCL and UCL are very
critical
iti l to
t the
th operating
ti characteristics
h
t i ti off the
th control
t l
chart.
If the control limits LCL and UCL are too tight (Fig 2-5),
some sample points may fall beyond the control limits
even when the process is in control. It means that the
control chart will p
produce signals
g
indicating
g pprocess out
of control, when process is actually in control.

17
Figure 2-4 An out-of-control process detected by a x control chart

These signals are called False alarms. Because they

mislead the operators by telling them the process is out
of control when the process is in fact in control. False
alarms waste time and effort to search for a problem (an
assignable
i bl cause)) that
th t does
d
nott exist.
i t False
F l alarms
l
also
l
make the operator lose confidence on the control chart.

18

On the other hand,, if the control limits LCL and UCL are
too wide (Fig 2-6), the sample points will not fall beyond
the control limits even when the process is already out of
control.
Or in other words, the control chart tells the operators
that the process is still in control when the process is in
f t outt off control.
fact
t l The
Th failure
f il
or delay
d l off detecting
d t ti outt
of-control status will result in a great number of
defective products.
products

Figure 2-5 Too tight control limits result in false alarms

19

20

3-sigma Control Limits

3-sigma control limits are the best and most widely used
trade-off. They are neither too tight nor too wide.
For a x chart
UCL CL 3
x
x
x

(2-2)

LCL CL 3
x
x
x

The control limits LCLx and UCLx are below or above the
central line CLx by 3 times of x (see Fig
Fig. 2-3)
2 3).

Since it is for a x chart, sigma means the standard deviation x

of x . It is noted that x of x is different from the standard
deviation x (or simply ) of x. The calculation of x will be
discussed later
later.
Figure 2-6 Too tight control limits result in false alarms 21

22

Two main categories of control charts

By using the 3-sigma control limits,

((1))

(1) When a process is in control, the possibility of false alarm

is very low,
low or almost impossible (see Fig 2-3)
2 3).
( ) When a process
(2)
p
is out of control,, the capability
p
y of detectingg
the out-of-control status is quite high. A signal is likely to be
produced very soon after the problem occurs (see Fig 2-4).
The general idea of the 3-sigma control limits can be applied to
the x chart, as well as many different types of control charts.
23

Control charts for variables

To control the quality characteristic that can be expressed in

terms of a continuous,
continuous numerical measurement
(e.g.,dimension, weight, volume).
E
Example:
l the
th diameter
di
t x (=
( 74 00.02mm)
02
) off a shaft.
h ft
(2)

Control charts for attributes

To control the quality characteristics that can only take

integer
g values
Example: the number d of defectives in a box of integrated
circuits.

24

2.2.1

& R charts

Suppose that x is a quality characteristic (e.g., a diameter). If x1,

x the
x2, , xn are the observations of a sample
p of size n,, then
sample mean is also a random variable.

2.2 Control Charts for Variables

x x1 x2 xn / n

T controll charts
Two
h
are usedd to monitor
i a variable
i bl x:
chart
x chart,

monitoring the mean value

al e of x (ensure
(ens re = 0)

R chart,
chart

monitoring the standard deviation of x

(ensure = 0)

The two x &R charts should be used together in order to monitor

both and ((ensure both = 0 and = 0)).

(2 3)
(2-3)

x is the value of individual product. Its probability

distribution can be built based on the data under columns 1 to
5 on Table
bl 2-2.
i the
th average value
l off n products.
d t The
Th probability
b bilit
x is
distribution of x is built based on the data under column 6 on
Table 2-2.
22

25

Table 2-2 Preliminary samples collected in phase I

26

If x has a normal distribution with mean x and standard

deviation x, then the sample mean x also follows a normal
distributed with
(2-4)

x x

x x / n

The mean x of x is equal to the mean xof x, but the standard

deviation xof x is smaller than x of x by a factor of n.
Conventionally, x and xare denoted as and , respectively,
for simplicity
simplicity, thus

x / n

(2-5)

For example, if = 74, = 0.01, and the sample size n = 5.

Then
27

x 74 x / n 0.01/ 5 0.004

28

According
g to the Central Limits Theorem,, even if the
distribution of x is non-normal, the distribution of is
approximately normal as long as n 3 or 4. Moreover, eq (25) is still valid.
Since n is usually equal to or larger than 3, therefore, when
designing a x chart, we always assume that the distribution
off x is
i normall regardless
dl off th
the di
distribution
t ib ti off x.
Since engineers are usually very familiar with the normal
distribution, the design of the x chart becomes easier.

Figure 2-7 How the control chart works

(f4-4 p135)

29

30

Sample range R
The range of a random variable x is the difference between the
maximum and minimum of all the possible values of x.
Usually,
y, the range
g is approximately
pp
y equal
q to 6 of the
probability distribution of x. Or in other words, the range is
different from only by a constant of six. As a result, the
range can be used as a measure of of the process distribution
of x.

Figure 2-8 Central limits theorem

31

The statistic used by the R chart is the sample range R. The

sample
l range is
i an estimate
ti t off the
th reall range, as sample
l mean x
is an estimate of . The sample range R is equal to the
difference between the maximum and the minimum of the n
32
values of xi in a sample.

For example,
p , if the five observed diameters in a sample
p of size 5
are
x1 = 2, x2 = 4, x3 = 1, x4 = -2, x5 = 0
Then, the sample range for this sample is

R x x = x2 - x4 = 4 (-2)
( 2) = 6
max

min

(2 6)
(2-6)

The value of R is always equal to or larger than zero.

zero
A R chart is used to monitor the sample range.
range A sample point on
the R chart carries the value of the sample range R of a sample.
The R chart displays
p y these sample
p points
p
versus the sample
p
number or time (see Figure 2-9).
33

Figure 2-9 A R chart

34

Since the calculation of sample range R is much simpler than

the calculation of sample standard deviation, the R chart is used
widely in industry to ensure = 0.
h
x &R chart

When process is in control:

= 0, R = R0 = 60.

A
Actual
l design
d i procedure
d
off the
h

When process is out of control:

= 1 0, R = R1 60.

(1) Sample size n --- decided by the available manpower and

instrument. n is usually equal to 4, 5 or 6, for the x &R
charts
charts.

Conversely,
If R = R0 = 60, it means = 0, the process may be in control
(we still have to check if = 0).

(2) Sampling interval h --- decided by the production shift

and rate.
(3) Central line CL and control limits LCL, UCL.

If R R0, it means 0, the process must be out of control.

35

36

Design of the x chart

In Eq (2-7), both LCL and UCL depend on 0 and 0.

However, in practice, 0 and 0are usually unknown and
must be estimated from the observed data.

3-sigma control limits of the x chart

UCL x CL x 3 x 0 3 0

Phase I (Build the chart): In this phase, m (at least 20 to

25)) preliminary
p
y samples
p are taken (see
(
Table 2-2).
) Each
preliminary sample also contains n observations of the
quality characteristic x.

CL x 0

(2-7)

LCL x CL x 3 x 0 3

Note, x ,

0
n

The purpose of taking the preliminary samples in phase I is

not to monitor the p
process,, but to collect sufficient data,, so
that, at the end of phase I, 0 and 0 can be estimated from
these data, and the control charts can be built.

x / n

When process is in control, x 0 ,

x 0 / n

Here, 0 and 0 are the mean and standard deviation of the

individual observation x when the process is in control.
37

Phase II (Use the chart): The control charts (built at the

end of p
phase I)) now can be used to monitor or control the
forthcoming production.

38

Table 2-2 shows the 25 preliminary samples available at the end

off a phase
h
I example.
l The
Th quality
lit characteristic
h
t i ti x is
i the
th diameter
di
t
of a piston ring.
We first calculate the sample mean and sample range for each of
the preliminary samples as listed at the right most 2 columns in
Table 2-2. For example,
74.030 74.002 74.019 73.992 74.008
74.010
5
R1 74.030 73.992 0.038
x1

73.995 73.992 74.001 74.011 74.004

74.001
5
R2 74.011 73.992 0.019
x2

Figure 2-10 Phase I and phase II (m = 25)

39

......

(2-8) 40

Estimation of 0

Estimation of 0

The estimate of 0 is the grand average x (average of averages)

First, calculate average range R (average of sample ranges)

0 x

x1 x2 xm
m

(2-9)

In Table 2-2,
x

R1 R2 Rm
m

(2 11)
(2-11)

In Table 2-2,

(2-10)
74.010 74.001 73.998 1850.028

74.001
25
25

is an even more accurate and reliable estimate of 0,

compared to each individual x .

0.038 0.019 0.035 0.581

0.023
25
25

(2-12)
(2
12)

0 off 0 is
Then,
h the
h estimate
i
i calculated
l l d by
b
(2 13)
(2-13)

0 R / d 2

41

The constant d2 is dependent

p
onlyy on the sample
p size n and can
be found from Table 2-3. For example, if n = 5, d2 = 2.326. 42

After 0 and 0 have been estimated, we can build the x &R

charts at the end of phase I. For the x chart, the basic formulae
are (see Eq (2-7))

Table 2-3
2 3 Factors for constructing variable control charts

UCL x 0 3

0
n

0 3

0
n

x 3

R / d2
3R
x
n
d2 n

CL x 0 0 x

(2-14)

R / d2
3R
LCL x 0 3
0 3
x 3
x
n
n
n
d2 n

43

44

The simplified formulae:

UCL
CL

LCL

3R
d

x A R

The constant

A
2

x A R

3
d

If all sample points fall around the central line CL of the

cchart
a t (Fig
( g 2-11))

(2-15)

3R
d

x chart carries the value of the sample

x
x

A sample point in a
mean of a sample.

(2-16)

0 0

It means that the mean value or the average of the quality

characteristic x is equal or close to 0. The process is likely
to be in control (we still have to check if = 0).

A2 also only depends on n and can be directly found in

Table 22-33. For example
example, when n = 5,
5 A2 = 0.577
0 577
Eqq (2-15)
(
) makes the computation
p
slightly
g y simpler,
p , but
loses some underlying information.

45

46

If a sample point falls above the upper limit UCL of the

chart (Fig 2-12 (a))

> 0 > 0

It means that the mean value or the average of x has

increased and process is out of control.
increased,
control
If a sample point falls below the lower limit LCL of the x
chart (Fig 2-12 (b))

x
Figure 22-11
11 An in-control
in control process monitored by a x chart
47

< 0 < 0

It means that the mean value or the average of x has

decreased, and process is out of control.

48

Design of the R chart

The R chart plots the points of sample range R versus the sample
number or time ((see Figure
g
2-9).
)
The sample size n and sampling interval h of the R chart are the
same as for the x chart. The n observations of x in each sample
are used to calculate both sample mean x and sample range R.
The 3-sigma control limits for the R chart:

UCL CL 3 R 3
R

CL R

(2-17)

LCL CL 3 R 3
R

Figure 2-12 How x chart detecting 0

(2) The standard deviation R of R is calculated as

R
d
3R
d2 d2

49

(1) R is the average value of the sample ranges obtained

from the m preliminary samples. R is taken as the target CLR
of R.

R d 3 0 d 3 0 d 3

50

The basic formulae for calculating the 3-sigma control limits

for the R chart (see Eq (2-17)):

UCL R R 3 R R 3
CLR R

(2-18)
(2
18)

The constant d3 is also dependent on the sample size n and

can be found from Table 2-3
(3) If the calculated value of LCLR is smaller than zero,
LCLR is pushed up to zero. Because the sample range will
never be smaller than zero.

d3
3d R
RR 3
d2
d2

51

d
3d R
LCL R R 3 R R 3 3 R R 3
d2
d2

(2-19)

52

A sample point in a R chart carries the value of a sample

range R.
R

The simplified
p
formulae for the R chart:
UCL R R

3d 3 R 3d 3
1
R D4 R
d2
d2

CLR R
LCL R R

If all sample points fall around the central line CL of the R

chart (Fig 2-13 (a))

(2-20)

3d 3 R 3d 3
1
R D3 R
d2
d2

R R0 6 60, 0

where,

D 1 3
3

d
,
d
3

D 1 3
4

d
d

(2-21)
3

It means that the standard deviation or the spread of the

quality characteristic x is equal or close to 0. The process is
likely to be in control (we still have to check if = 0).

D3 and D4 depend on sample size n and can be found in Table

22-33. For example
example, when n = 5,
5 D3 = 0,
0 D4 = 2.115
2 115
53

54

If a sample point falls above the upper limit UCL of the R chart
(Fig 2-13 (b))
R > R0 6 > 60, > 0
It means that the standard deviation or the spread of x has
increased. The p
process is out of control and the pproduct qquality
y
deteriorates.
If a sample point falls below the lower limit LCL of the R chart
(Fig 2-10 (c))
R < R0 6 < 60, < 0

Figure 2-13 How R chart works

55

It means that the standard deviation or the spread of x has

decreased. It indicates that the product quality is improved.
However the process is still considered as out of control
However,
because of 0

56

Retrospective Check
The control limits of thex and R control charts depend on the
values of the in-control 0 and 0 that are estimated from the
data contained in the m preliminary samples in phase I (Fig 210) It iimplies
10).
li that,
th t during
d i phase
h
I,
I in
i order
d to
t accurately
t l
estimate 0 and 0, the process should be in an in-control
status How can we ensure it?
status.
Customarily, the control limits of the x and R control charts
Customarily
determined based on the data of the preliminary samples are
treated onlyy as the trial control limits. Theyy allow us to
check whether the process was in control when the m
preliminary samples were taken.

If all m points in phase I plot inside the trial control limits

(Fig. 2-10), then we conclude that the process was in control
when the m p
preliminary
y samples
p were taken,, and the trial
control limits can be used as the final control limits in phase
II for controlling future production.
If one or more of the m sample points plot out of the trial
control limits,
(1) E
Examining
i i each
h off th
the out-of-control
t f
t l points,
i t looking
l ki for
f
the possible problems (assignable causes). If found, the
assignable causes must be removed,
removed and the out-of-control
out of control
points are discarded. The trial control limits are recalculated,
using only the remaining sample points.
points

57

58

Steps for setting up the x & R charts

(2) In some cases, it may not be possible to find out a
responsible
p
assignable
g
cause for the out-of-control
points, then we may either
(i) Eliminate the out-of-control points, or

(1)

Decide the sample size n (based on the available

resources) and the sampling interval h (based on the
working shifts).

(2)

In phase I, take m preliminary samples, and calculate

the sample mean xi and sample range Ri for each
sample (eq (2-8)).

(3)

Calculate the grand average x (eq (2-9)) and the

average
g R of sample
p ranges
g (eq
( q (2-10)).
(
))

(4)

Find out the constants d2 and d3 from Table 2-3 based

on sample size n.
60

(ii) Retain
i the
h out-of-control
f
l points
i
If m is
i large,
l
andd there
th are only
l one or two
t out-oft f
control points, either treatment (i) or (ii) will not
distort the control limits significantly.
significantly The control
chart will work satisfactorily.
59

((5))

Calculate the central line CL and the trial control limits

LCL a nd UCL for both the x and R charts (eqs (2-14)
and (2-19)).
(Sometimes, the simplified formulae eqs (2-15) and (220) also can be used to calculate the control limits).

(6)

Pl t the
Plot
th two
t charts.
h t

(7)

Carry out retrospective check

check. If it is passed,
passed the trial
control limits are launched as the final control limits.

EXAMPLE 2-1:
(1)

Inside diameter x of the piston rings

Decide the sample size n and the sampling interval h.

n = 5, h = 1 hour

(2)

In phase I, take 25 preliminary samples (Table 2-2), and

calculate the sample mean x i and sample range Ri for
each sample.

(8)
Start phase II, using the two charts to monitor the
pprocess.
61

Calculate the grand average x and the average R of

sample ranges.

(3)

62

(5)
Calculate the central line CL and the control limits
LCL and UCL for both the x and R charts.

25

i 1

25

74.010 74.001 73.998 1850.028

74.001
25
25

UCLx x

25

(4)

R
i 1

25

chart:

0 . 581
0 . 023
25

3R
3 0.023
74.001
74.014
d2 n
2.326 5

CLx x 74.001

Find out the constants d2 and d3 based on sample size

n = 5.

LCLx x

3R
3 0.023
74.001
73.988
d2 n
2.326 5

From Table 2-3, d2 = 2.326, d3 = 0.864.

63

64

R chart
UCL R
R

3d R
3 0.864 0.023
0.023
0.049
d
2.326
3

CL R 0.023
R

LCL R
R

3d R
3 0.864 0.023
0.023
0.026 0
d
2.326
3

((6))

Plot the x and R charts ((Figure

g
2-14).
)

(7)

Carry out retrospective check.

65

(8) After the retrospective check has been passed, we can start
phase II
II, monitoring the process.
process
[1] chart is to monitor the central location of x (ensure
= 0);
and the R chart is to monitor the spread of x (ensure =
0)). The two charts are runningg side byy side.
[2] Take a sample of n (five) observations for every one hour (h
= 1 hr), calculate the sample mean x and sample range R.
For example, the five observations for a sample are
x1=74.030, x2=74.002, x3=74.019, x4=73.992, x5=74.008,
74.030 74.002 74.019 73.992 74.008
74.010
5
R 74.030 73.992 0.038
67

Figure 2-14 x and R charts

66

[3] Plot sample meanx on the x chart; and plot sampling

range R on the R chart.
[4] If (the sample point x falls inside the control limits of thex
chart) and (the sample point R falls inside the control limits of
the R chart), the process is in control ( = 0, = 0).
[5] If (the sample point x falls beyond the control limits of x
the chart) and/or (the sample point R falls beyond the control
li it off the
limits
th R chart),
h t) the
th process is
i outt off control
t l (
( 0,
and/or 0).

68

Concentrated sampling

2.2.2 Rational Subgroups

When the primary purpose of using the control chart is to

detect out-of-control status,, each sample
p should consist of
units that are produced at the same time (or as closely as
possible). Concentrated samples are usually taken at the
end of each sampling interval.

The rational subgroup concept means that samples (also

called
ll d subgroups)
b
) should
h ld be
b selected
l t d in
i a time
ti order
d
according to the working shift of production, so that the
control chart is more effective in detecting the out-of-control
out of control
status. Time order subgrouping also helps to identify the
source of the problem
problem.

Random sampling
i
When th
Wh
the control
t l chart
h t is
i employed
l d to
t make
k decision
d i i about
b t
the overall quality level of a product, samples should
consist of units that are randomly picked up over the whole
sampling interval h and, therefore, are representative of all
units.
69

Concentrated sampling is used most widely.

70

2.2.3 Cusum (Cumulative Sum) Chart

In this section, we only consider the process mean and assume
that is always equal to 0. So,
So as long as 0 , the process is
in control.
The x chart
Th
h t is
i quite
it effective
ff ti for
f detecting
d t ti large
l
scale
l mean
shifts. However, it may become ineffective to small and
medium mean shifts.
shifts
For the x chart, the sample mean shift is

(2-22)

When the process is in control

Figure 2-15 Concentrated and random sampling

71

,
0

When the process is out of control

(2 23)
(2-23)
= 0.
0.

72

So, the cusum chart comes to play, because it is much more

effective than the x chart to detect small and medium mean
shifts.
But,, the design
g and operation
p
of the cusum chart are more
complicated, and its fundamental algorithm is more difficult
to understand.
In this section, in order to simplify the discussion, we use
the x chart,, that is a special
p
p size n = 1.
x chart with sample
In each sample of the x chart, there is a single observation
x. Thus, x = x/1 = x. From Eq (2-23),

(2-24)

In this section, xi represents the single observation of the ith

sample. The estimate of the mean shift for this sample is

Figure
g
2-16 Mean shift

73

i x i 0

(2-25)
74

Figg 2-17 displays

p y a x chart for monitoring
g the diameter x of a
shaft. The in-control parameters are 0 = 10 and 0 = 1, and the
control limits are
LCL = 7, CL = 10, UCL = 13

(2-26)

In Table 2-4, the values of xi of the first 20 samples are

observed
b
d when
h the
th process is
i in
i control.
t l All off these
th
20 xi
values fall within the control limits.

75

Figure 2-17 An x control chart

76

Table 2-4: Data for the cusum chart example (0 = 10)

77

Then, the cutter of the machine tool is broken suddenly after

the 20th sample. It increases the process mean to = 11. The
process becomes out of control.
= - 0 = 11 10 = 1 > 0

(2-27)

The subsequent observations xi (i = 21.

21 22,
22 ..)) are taken under
the out-of-control status. But, the corresponding sample points
still fall within the control limits of the x chart. It means that
the x chart fails to detect the out-of-control status, because it is
not veryy effective for detectingg the mean shift of small
magnitude (for example, = 1).
The reason is that the x chart (orx chart) makes decisions only
based on the data of the current (or last) sample point without
considering
id i the
h series
i off the
h previous
i
sample
l points
i or the
h
79
operational history of the process.

78

Telling a story: In a company, employees are required to

arrive and start working a 8:00am in the morning. However,
a lateness of no more than 5 minutes will not be punished.
An operator is later every day, but always before 8:05am
(i i < 5) Then
(i.e.,
Th he
h can never be
b caught.
ht

80

Later the company changes the rule. It will penalize any

employee whose cumulated lateness time is more than 10
minutes. Since the sum C4 of the first four i (i.e., the sum of
th lateness
the
l t
time
ti in
i the
th first
fi t four
f
days).
d )
C4 = 1 + 2 + 3 + 4 = 11 > 10

(2 28)
(2-28)

the operator is caught in just four days with the new rule

being effective.

Ci is the sum of the values from the first sample to the

ith (current) sample.
C 1 = 1
C 2 = 1 + 2
C 3 = 1 + 2 + 3

This example
Thi
l ill
illustrate
t t the
th power off the
th cumulative
l ti sum, or
the capability of the CUSUM chart. This chart incorporates
all the information in the sequence of the sample points by
monitoring the cumulative sums Ci of the mean shifts .

C
i1

i1

C
i

i1

(2-29)

IIt iis why

h the
h CUSUM chart
h iis more effective
ff i than
h the
h x chart
h
81
(or x chart) for detecting small process shifts.

Or,

C ( ) C
i

i1

i1

For example, in Table 2-4,

((2-30))

C1 = 1 = -0.55

It means:

C2 = C1 + 2 = -0.55 + (-2.01) = -2.56

C 1 = 1

C3 = C2 + 3 = -2.56 + (-0.71) = -3.27

C 2 = 1 + 2 = C 1 + 2
C 3 = 1 + 2 + 3 = C 2 + 3

82

C4 = C3 + 4 = -3.27 + 1.66 = -1.61

(2-31)

(2-32)

A CUSUM chart
h t plots
l t th
the sample
l points
i t off C1, C2, C3, .
That is, the ith sample point of a CUSUM chart carries the
value of Ci.
83

84

If = 0 (process in control),
The x values of some samples are larger than 0 (see Table 24),
4)
xi > 0 = (xi - 0) > 0,
0 Ci > Ci-1 Ci
i

(Note: Ci = Ci-1
i 1 + i )
But,, the x values of some other samples
p are smaller than 0,
xi < 0 = (xi - 0) < 0, Ci < Ci-1 Ci
i

Figure 2-18 Plot of the cumulative sum

As a result, Ci sometimes increases, and sometimes decreases,

always wanders around zero.
85

86

If > 0 (process is out of control with increasing mean shift > 0)

The x values of almost all samples are larger than 0,

2 2 4 Runs Rules
2.2.4

xi > 0 = (xi - 0) > 0 Ci > Ci-11 Ci

In this section,, again,

g , we only
y consider the pprocess mean and
assume that is always equal to 0.

Most of the time, Ci increases point by point until going

beyond an upper limit UCLcusum.

A control chart can signal an out-of-control condition by

In Fig 22-18,
18 the 26th sample point goes above UCLCUSUM,
and the CUSUM chart signals the out-of-control status.

(1)Basic rule

when one sample point falls beyond the

control limit LCL or UCL;

(2)Runs rules

when several sample points exhibit some

nonrandom pattern.
pattern

The control limit UCLcusum of the CUSUM chart is calculated by

Markov chain method or Monte Carlo simulation.
A similar CUSUM chart can be used to detect the out-of-control
status with decreasing mean shift ( < 0 or < 0).
87

88

A control chart may

y presents
p
some nonrandom p
patterns of
the sample points when the process is out of control.
Such nonrandom patterns are seldom found when the process
is in control. Therefore, their occurrence is a strong
indication that the process is out of control (even if all of the
sample points are within the control limits).
When the basic rule is used together with the runs rules, the
capability of the control charts will be enhanced.
enhanced However,
However
using the runs rules increases the difficulty of operation.

89

(2)Four out of five consecutive points plot at a distance of

11-
or beyond from the center line on one side (points 16,
17, 18, 19, 20 on Fig 2-20).

Figure 2-20 Runs rule 2

91

Figure 2-19 Runs rule 1

90

(3)Eight consecutive points plot on one side of the center

line (points 13
line.
13, 14
14, 15
15, 16
16, 17,
17 18,
18 19,
19 20 in Fig 2-21).
2 21)

Figure 2-21 Runs rule 3

92

The out-of-control status in Fig 2-17 can be detected if the 3rd

runs rule is applied to the last eight sample points.
points

2.3 Control Charts for Attribute

To control the quality characteristics (e.g., d and C) which can
only take integer values.
values
2.3.1 np Chart
The np control chart is used to monitor the number d of
nonconforming
g items found in a sample
p of size n. A
nonconforming item is a defective product, which fails to meet
the quality requirement on the product.
For example, if five defectives are found in a sample
containing
i i 100 integrated
i
d circuit
i i boards
b d (n
( = 100),
100) then
h d = 5.
5

93

Fraction nonconforming p is the ratio of the number D of

nonconforming items in a population to the total number N of
items in that population.
p=D/N

Example 2-2
Last year, a company produces 10000 TVs, among which 295
are nonconforming. A sample of 100 TVs are inspected and 3
TVs in the sample are found nonconforming. Therefore

(2-33)

p can be estimated by

p d / n

94

N = 10000, D = 295, p = D / N = 0.0295

(2-34)
(2
34)

n = 100, d = 3, p = d / n = 0.03
The sample size n is usually much smaller than N. Therefore,
estimating
g p by
y d/n ((inspecting
p
g onlyy n items in a sample)
p ) is
much easier than calculating p by D/N (inspecting the whole
population of N items).
95

96


d is a random number and may have any integral
value i
(0 i n). Moreover, d follows a binomial distribution,
which is determined by two parameters n and p.

n
Probd i Pd (i) pi (1 p)ni
i
n!

pi (1 p)ni
i!(n i)!

n!1 2 n,

i!1 2i,

4!
0 . 1 0 (1 0 . 1 ) 4 0 0 . 6561
0! ( 4 0 )!
4!
Pd (1 )
0 . 11 (1 0 . 1 ) 4 1 0 . 2916
1! ( 4 1 )!
4!
Pd ( 2 )
0 . 1 2 (1 0 . 1 ) 4 2 0 . 0486
2! ( 4 2 )!
4!
Pd ( 3 )
0 . 1 3 (1 0 . 1 ) 4 3 0 . 0036
3! ( 4 3 )!
4!
Pd ( 4 )
0 . 1 4 (1 0 . 1 ) 4 4 0 . 0001
4! ( 4 4 )!
Pd ( 0 )

(2-35)

(n i)!1 2(n i)

Mean and standard deviation of d are

np ,
d

np (1 p )

For example,
p , if the cardboards are inspected
p
with sample
p
size of four, and 10% of the cardboards are nonconforming
(n = 4 and p = 0.1), d may be equal to 0, 1, 2, 3, or 4

(2-36)
(2
36)
97

Note: Pd(0) + Pd(1) + Pd(2) + Pd(3) + Pd(4) = 1 = 100%

98

Cumulative Binomial probabilities

It means that d cannot take any value smaller than zero for
larger than four.
four

npp 4 0 .1 0 .4
d

(2-38)
0.05

0.10

0.20

0.30

0.40

0.50

0.358

0.122

0.012

0.001

0.000

0.000

0.736

0.392

0.069

0.008

0.001

0.000

0.925

0.677

0.206

0.035

0.004

0.000

0.984

0.867

0.411

0.107

0.016

0.001

It means that, on average, there will be 0.4 nonconforming

items in each sample.

np (1 p ) 4 0 .1 (1 0 .1) 0 .6
d

n = 20

(2 39)
(2-39)

99

100

np Control Chart: monitoring the number d of

nonconforming
f
i items
it
in
i a sample
l off size
i n. The
Th name off the
th
np chart reflects the fact that the mean of d is equal to np.
Each sample point on the np chart carries the value of d of a
sample.
When using the np chart, the process is considered as in
control if p = p0, where p0 is the in-control ((or acceptable,
p
, or
nominal) value of fraction nonconforming p. The value of p0
is usually very small, but cannot be zero (p0 = 0 is just a
dream). When process is in control,
p = p0,

np ,
d

np (1 p )
d

(2-40)

101

UCL CL 3 npp 3 npp (1 p )

d

CL np
d

(2-43)

LCL CL 3 np 3 np (1 p )
d

p = d / n, or d = np

(2-41)

Or in other words, d is different from p only by a constant of n.

As a result, d can be used as a measure of p of the product, and
the np chart can be used to ensure p = p0.
If d = np0,

p = np0 / n = p0,

(2-42)
(2
42)

Then, the process is in control. As such, np0 is the target of d.

In practice, if d is equal or close to np0, p will be equal or close
to p0, and the process is thought in control.
On the other hand, if d significantly deviates from np0, p must
differ from p0, and the process will be signalled as out of
control.
102

In Eq (2-43), both LCLd and UCLd depend on p0. However, in

practice, p0 is usually unknown. To estimate p0, we take m
preliminary samples. For the ith sample (see Table 2-5)

33-sigma
sigma np chart
d

The fraction nonconforming p is estimated by d/n, or

approximately,
approximately

Should
Sh
ld the
h calculation
l l i yield
i ld a negative
i LCL
C d, set LCL
C d = 0.
0
Because d cannot be a negative number.

p
i

d
n
i

i 1,2,,m

(2-44)

Even though each p i is an estimate of p0, the average p of p i is

an even more accurate and reliable estimate of p0.
m

p0 p

i 1

(2-45)

For example, from Table 2-5,

p p
103

0.240.300.160.12
0.2313
30

104

Table 2-5 Data of Cardboard (n = 50)

Then, p0 in equation (2-43) will be replaced by p .

UCLd np 3 np (1 p )
CLd np

(2 46)
(2-46)

LCLd np 3 np (1 p )

Sample size of the np chart is determined by the production &

inspection rate and the p0 value. If p0 is very small, sufficiently
large
g n should be chosen. An empirical
p
formula is as follows:
n = c / p0
where, c is a constant between 0.5 and 1.5. For example, if p0
= 0.01,
0 01 n should
h ld be
b set between
b
50 andd 150.
150
105

( )
(4)Calculate
the central line CLd and the control limits
LCLd and UCLd for the np chart (eq (2-46)).

EXAMPLE 2-3 Design

g an np
p Chart to Monitor the
number of Defective Cardboards (Table 2-5)

(1) Decide the sample size n and the sampling interval h.

UCLd np 3 np (1 p )
50
0 0.2313 3 50
0 0.2313 (1 0.2313) 20.510
10

n = 50, h = 0.5 hr
(2) In
I phase
h
II, ttake
k 30 preliminary
li i
samples,
l andd calculate
l l t
the fraction nonconforming pi for each sample (eq (2-44)).
(3) Calculate the average fraction nonconformingp (eq (245)) m
45)).
p

i 1

0.24 0.30 0.16 0.12

0.2313
30

106

107

CLd np 50 0.2313 11.565

LCLd np 3 np (1 p )
50 0.2313 3 50 0.2313 (1 0.2313) 2.620

108

(5) Round off the control limits

Or in other words, the fractional limits are exactly equivalent

to the following integral limits:

Fractional limits:
If 2.620
.6 0 d 20.510,
0.5 0,

tthee process
p ocess iss in co
control.
to.

If 3 d 20,,

the pprocess is in control.

If d < 2.620 or d > 20.510,

the p
process is out of control.

If d < 3 or d > 20,

the process is out of control.

But, since d is always an integer, the operational rules of this np

chart is
(i)

claiming
l i i in-control
i
l status if 3 d 20;
20

(ii)

signalling out-of-control
out of control status if d < 3 or d > 20.
20

The round-off rule: the fractional LCL is rounded up, and the
fractional UCL is rounded down.
Since th
Si
the integral
i t
l limits
li it are easier
i to
t handle
h dl than
th the
th fractional
f ti l
limits, the former will be used in the implementation of the np
chart
chart.

109

110

(6) Plot the np chart

Figure 2-23
2 23 Round off of the control limits
111

Figure 2-24 An np control chart

112

If a d sample point plots outside of the control limits, p has most

likely shifted to a new level and the process is thought out of
control.

(7) Start phase II,

II using the np chart to monitor d of
the process.
As long as the sample points remain within the control
limits,
ts, d iss equa
equal oor cclose
ose to C
CL ((=np
np0), oor p iss equa
equal or
o
close to p0, the process is thought in control. That is,
the defective rate of the Cardboard is kept
p around the
acceptable level of p0.

(d > UCL) indicates that d > np0 (or p > p0), the defective rate
becomes higher than p0. We must shut down the process
process, then
search and remove the problem.
(d < LCL) means that d < np0 (or p < p0). Even though defective
rate
a e gets
ge s lower,
owe , thee pprocess
ocess iss still
s thought
oug as out
ou of
o control
co o
(because p p0).

113

2.3.2 c chart
c chart
h
i to monitor
is
i the
h number
b C off nonconformities
f
i i in
i a
unit of n items. It is mainly used to handle some minor
(nonfatal) defects in the products.
products

We also have to look into the process to find out the reason for
improvement. The np chart may have to be redesigned, because
the value of p0 is changed and the control limits depends on p0.
114

Depending on the nature and severity of the nonconformities,

nonconformities
it is quite possible that a unit containing several
nonconformities is acceptable,
acceptable and not classified as
nonconforming.

A nonconformity is a minor defect in common language.

language

A nonconforming means that a unit as a whole is

unsatisfactoryy and must be rejected.
j

A unit is an entity for which it is convenient to keep

records. It could be a single item or a group of n items.

A nonconformity means a minor defect on one of the n

items in a unit. The unit may, or may not, be rejected.

115

116

Suppose ci is the number of nonconformities in the ith item within

a unit. And C is the total number of nonconformities in all of the n
items in a unit of n items.
n

C c1 c2 cn ci

(2 47)
(2-47)

i 1

In the example in Fig 22-25,

25 a unit is comprised of three cars
shipped by a container (n = 3). A nonconformity is a tiny spot on
the surface of a car due to some painting problem
problem.
So, ci is the number of tiny spots on the ith car in a unit, and C is
the total number of the tiny spots in all three cars of a unit. In Fig
2-25,,
c1 = 2, c2 = 1, c3 = 0
Figure 2-25 A unit of three cars

C c1 c2 c3 2 1 0 3
117

(2 48)
(2-48)
118

For this example,

p , based on the quality
q
y requirement,
q
, the
Quality Engineer specifies that a unit is nonconforming when
C is larger than 4. That is,
If C 4, the whole unit (three cars) is conforming.
Even though there may be a few nonconformities
(between 1 and 4) on some items in this unit, the unit
as a whole
h l will
ill be
b accepted.
t d
But, if C > 4,
But
4 the whole unit (three cars) is
nonconforming. Even though the impact of each
individual nonconformity is not severe,
severe the number of
nonconformities in a unit is intolerably high, the unit
as a whole will be rejected.
j
119

In unit 3, the third car contains no nonconformity (c3 = 0).

However, since unit 3 as a whole is nonconforming, this car will
b rejected
be
j t d together
t th with
ith other
th cars in
i this
thi unit.
it
120

Suppose, the average number of tiny spots (nonconformities)

in a unit is three
three. That is
is,

C is a random number and follows a Poisson distribution.

Pr ob (C i ) PC (i )

e i 2.718 i

i!
i!

= 3, C = 3, C 3 1.732

(2-49)

The Poisson distribution is uniquely determined by a single

parameter (> 0) which is the mean value C of C.
C Or in
engineering terms, is the average of the number of
nonconformities in each unit.
unit In the example in Table 22-66,
(3 + 1 + 5) / 3 = 3

(2-50)

The standard deviation C of C is equal to the square root of .

C = , C

(2-51)

e 3 30
PC (0)
0.0498
0!
e 3 31
PC (1)
0.14936
1!
e 3 32
PC ( 2)
0.22404
2!
e 3 33
PC (3)
0.22404
3!

121

(2-52)

122

3-sigma
g
c chart:
UCLC CLC 3 C 3
CLC

(2-53)

LCLC CLC 3 C 3

In Eq (2-53), both LCL and UCL depend on . However, in

practice is usually unknown.
practice,
unknown But
But, can be estimated from
the observed values of C in m preliminary units.

Figure 2-26 Poisson distribution

123

124

Table 2-7 Number of nonconformities in units of 100 printed

circuit boards
(m = 26, n = 100)

For example (Table 2-7), m = 26, n = 100. Cj is the total

number of nonconformities in the jth unit containing 100
printed circuit boards.
C1 = 21, C2 = 24, ., C26 = 15

C1 C2 C26 21 24 15

19.85
26
26

(2 54)
(2-54)

Now, replacing
p
g by
y ,
UCLC 3
CLC

(2 55)
(2-55)

LCLC 3

Should the calculation yields a negative value for LCLC, set

LCLC = 0.
0 Beca
Because
se the number
n mber of nonconformities cannot be
125
negative.

EXAMPLE 2-4 Design of a c chart to monitor the

number of nonconformities in units of n printed circuit
boards (Table 2-7)

(1)Decide the number n of items per unit and the sampling

i t
interval
l h.
h

126

((3)Estimate
)
the mean ((eqq (2-54)).
(
))

C1 C2 C26 21 24 15

19.85
26
26

(4)Calculate the central line CLC and the control

limits LCLC and UCLC for the c chart (eq (2-55)).

n = 100,
100 h = 0.5
0 5 hr
(2)In phase II, take 26 preliminary units,
units and calculate the
number Cj of nonconformities for each unit.

UCLC 3 19.85 3 19.85 33.22

CLC 19.85
LCLC 3 19.85 3 19.85 6.48

C1 = 21, C2 = 24, ., C26 = 15

127

128

(5)Round off the control limits

Since C is always an integer, the control limits of the c chart
should be rounded off. Particularly, LCLC should be rounded up
and
d UCLC should
h ld be
b rounded
d d down,
d
as for
f the
h np chart.
h
Thus,
Th in
i
this example, the final control limits are: LCLC = 7 and UCLC =
33.
33
(6)Plot the c chart.
(7)Start
II, using the c chart to monitor C of the process.

phase
If 7

33
33,

the process is
th
i thought
th
ht in
i control
t l (the
(th number
b
in19each
.85 unit is close to the
of nonconformities
average value of
)
).

If C < 7 or C > 33, the process is thought out of control

129

3 PROCESS CAPABILITY
Process capability refers to the capability that a manufactured
qquality
y characteristic x conforms to the design
g specifications
p
on the tolerance. The higher the process capability is, the lower
the defective rate.
Specification limits: decided by the design engineer based on
the functional requirements.
q
Theyy are fixed values in a process
p
capability study.

LSL = 74 0.05
0 05 = 73.95mm
73 95
USL = 74 + 0.05
0 05 = 74.05mm
74 05mm
Design tolerance = USL LSL = 74.05
74 05 73.95
73 95 = 0.1mm
0 1mm

Lower specification limit: LSL

Upper specification limit: USL
Design tolerance = USL - LSL

For example, the design specifications of the diameter x of

a shaft is decided by the design engineer as 74 0.05mm.
Then, the nominal or target value, T, of the diameter is
74mm, and

An individual shaft is classified as a defective,, if the

diameter x is smaller than 73.95mm or larger than
74.05mm.

(3 -1)

An individual
d v dua pa
partt iss cclassified
ass ed as a de
defective,
ect ve, if x iss ssmaller
a e
than LSL or larger than USL.

Natural tolerance limits: acquired from the actual

probability distribution of x (see Fig 3-1) varying along
with the changes of the manufacturing status.
Lower:
L
Upper:

LNTL
UNTL

((the
h minimum
i i
value
l off x))
(the maximum value of x)

The probability distribution of x, as well as LNTL and

UNTL can be changed by using different machines,
UNTL,
machines
processes, materials and so on.

Figure 3-1 Target and specifications

x is usually assumed to have a normal distribution. Therefore,

x may be infinitely small or infinitely large from a theoretical
viewpoint.
But from an engineering viewpoint, it is almost impossible
((with a probability
p
y of onlyy 0.0027)) that the x value will fall
below ( - 3) or above ( + 3). Thus, in engineering
practice, we can take ( - 3) as the minimum possible value
of x; and take ( + 3) as the maximum possible value of x.
LNTL - 3
UNTL + 3

(3-2)

Distribution range,
range R= UNTL - LNTL
= ( + 3) ( - 3) = 6 (3-3)
Figure 3-2
3 2 Lower and upper natural control limits
6
in the normal distribution

and can be estimated from the measured data (e.g., the data
obtained
bt i d dduring
i running
i a control
t l chart).
h t)
Suppose four observations are available
available, x1 = 2.0,
2 0 x2 = 2.1,
2 1 x3 =
1.9, x4 = 2.2, n = 4

x1 x2 x3 x4 2.0 2.1 1.9 2.2

2.05
n
4

x
i 1

Or
Or,

R R / 1 R 2.2 1.9

0.146
d2 d2 d2
2.059

(3 6)
(3-6)

d2 = 2.059,
2 059 for n = 4.
4
Then, if using 0.129 ,
LNTL 3 2.05 3 0.129 1.663

(3-4)

(f9-1 p432)

UNTL 3 2.05 3 0.129 2.437

(3-7)

n 1

( 2.0 2.05) 2 ( 2.1 2.05) 2 (1.9 2.05) 2 ( 2.2 2.05) 2

4 1
0.129
(3-5)
7

Applications of the process capability studies:

In order to achieve a high

g pprocess capability
p
y so that a
manufactured quality characteristic x conforms to the design
specification on the tolerance (or the value of x falls between
LSL and USL) with a high probability, we may

(1)Predicting how well the process will hold the tolerance.

(1)make the standard deviation of x small. Because it makes

the distribution range R (= 6) of x likely to fall within
specification
ifi ti limits
li it LSL andd USL.
USL
Compare Fig 33-1
1 (a) and (c)
(2)make the mean of x close to the target T (usually,
(usually T is set
at the center between LSL and USL).
Compare Fig 3-1 (a) and (b)

3.1 Process Capability Ratio Cp

Cp is a measure of the distribution range R with respect to
design
g tolerance ((USL LSL).
)
Cp

USL LSL
USL LSL
USL LSL

R
UNTL LNTL
6

(3-8)

(3)Selecting between competing vendors.

vendors
The process capability is usually measured by one of the three
process capability ratios: Cp, Cppk and Ckm.
For any of Cp, Cpk and Ckm, the greater its value is, the higher
the process capability, and the lower the defective rate.
10

If Cp is great, the distribution range R (UNTL LNTL) of

the quality characteristic x is relatively small compared to
the design tolerance (USL LSL), or x is less likely to fall
beyond the specification limits and is more likely to conform
to the design tolerance requirement.
The greater the Cp, the better the product quality. The
minimum requirement is Cp = 1,
1 otherwise,
otherwise a substantial
amount of products will fail to meet the design specifications
(indicated by the darkened area in Fig 33-3
3 (c)).
(c))

Example 3-1: x =740.05mm is the diameter of a shaft.

A suppose 0.0099 mm
Also
USL = 74.05
74 05 mm, LSL = 73.95
73 95 mm,
Cp

(2)Assisting product developers/designers in selecting or

modifying a process
process.

USL LSL 74.05 73.95

1.68
6
6 0.0099
11

12

Table 33-1
1 Recommended values of the Cp

Figure 3-3 Different values of Cp

13

One-sided specifications
U
Upper
specification
ifi ti
In some applications, x is constrained only by an USL, but has
no LSL, i.e.,
x USL

(3-9)
(3
9)

For example, x is the deflection of a cantilever beam under

certain
t i load
l d F.
F The
Th design
d i engineer
i
specifies
ifi that
th t x mustt be
b no
larger than 0.01mm. i.e.,
x 0.01
USL = 0.01. LSL is not needed,, because,, we hope
p the deflection
to be as small as possible.
15

14

For upper specification, we calculate upper Cp ---- CPU

Suppose there is a virtual LSL that is symmetrical with USL about
(Fig 3-4 (a)).
CP

USL LSL 0.5 (USL LSL) USL

6
0.5 6
3

(3-10)

This Cp is taken as CPU ((Note, it is independent

p
of LSL),
) so
CPU

USL
3

(3 11)
(3-11)

16

Lower specification
In some other applications, x is constrained only by a LSL, but
h no USL,
has
USL ii.e.,
x LSL

(3 12)
(3-12)

For example
example, x is the power of an engine.
engine The design engineer
specifies that x must be no smaller than 1000 watt, i.e.,
x 1000
LSL = 1000. USL is not needed, because, we hope the power of
the engine to be as high as possible.
Figure 3-4 One-sided specifications
17

Another quality index that is directly related to the process

capability is PPM. It is the number of defective (nonconforming)
Parts Per Million. PPM is actually a measure of the defective rate
p ((or the fraction nonconforming).
g)

For lower specification, we calculate lower Cp ---- CPL

Suppose there is a virtual USL that is symmetrical with LSL
about (Fig 3-4
3 4 (b)).

C
P

USL LSL 0.5(USL LSL) LSL

6
0.56
3

(3-13)
(3
3)

This Cp
C is taken as CPL (Note,
(Note it is independent of USL),
USL) so

CPL

LSL
3

18

PPM = 1,000,000 p

(3-15)

p = Prob (x < LSL or x > USL)

(3-16)

Table 3-2 shows several PPM values along with the associated
values
l
off the
th Cp. Here,
H
th
the assumptions
ti
are:

(3-14)

(1)The quality characteristic x has a normal distribution

distribution.

19

(2)The process mean is centered between the upper and lower

specification limits ( = T).

20

Table 3-2 Cp versus PPM

For example,
F
l if Cp = 1.00
1 00 (see
(
Fig
Fi 3-3),
3 3) the
h probability
b bili p
that x falls beyond LSL or USL is 0.0027. Then,
PPM = 1,000,000 0.0027 = 2700
It is obvious that, the larger the Cp, the smaller the
defective rate p and the smaller the PPM.

21

22

If the qquality
y characteristic x has a normal distribution,, but the
process mean is not centered between the upper and lower
specification limits (i.e., T), PPM can be calculated easily.
Suppose the mean and standard deviation of the random
variable x are and , respectively. Then, the probability that x
is smaller than a fixed value X is equal to
X
Pr ob( x X )
z

where,

(3-17)

The random variable z follows a standard normal

distribution with = 0 and = 1. ( ) is the cumulative
probability function of z.
23

24

Furthermore,
F
h
many tables
bl di
display
l the
h values
l
only
l for
f z>
0. However, if z < 0,
(z) = 1 - (-z)

(3-19)

For example,
(-0.54) = 1 - (0.54) = 1- 0.7054 = 0.2946

Table 3-3
33
25

26

At the lower end, the number, PPML, of defective Parts Per

Million is calculated as follows (Fig 3-5).
PPML = 1,000,000 prob(x < LSL)

= 1,000,000
1 000 000

LSL

= 11,000,000
000 000 z L
where
where,

zL

LSL

(3-21)
(3
21)

Figure 33-5
5 Calculation of PPM
27

28

At the upper end, the number, PPMU, of defective Parts Per

Milli is
Million
i calculated
l l t d as follows.
f ll
PPMU= 1,000,000 prob(x > USL)

prob(x < LSL) = prob(x > USL)

= 1,000,000 [ 1 - prob(x < USL) ]

If the pprocess mean is centered between the upper

pp and
lowerspecification limits ( = T), then

USL

= 1,000,000
1 000 000 1

PPML = PPMU = 0.5 PPM

(3-20)

= 1,000,000 1 zU
where, zU USL

(3-22)
(3
22)

The overall PPM

PPM = PPML + PPMU

(3-23)
29

Example 3-2 (Figure 3-5)

The weight x of a product follows a normal distribution with
= 10g and = 2g. It is required that x be maintained between 5g
and 14g, i.e.,

30

At the lower end,

LSL 5 10
zL

2.5
2

( L) = (-2.5)
(zL)
( 2 5) = 11- (2.5)
(2 5) = 0.00621
0 00621
PPML = 1,000,000
,000,000 z L

= 1,000,000
,000,000 0.006
0.00621 = 6210
6 0

At the upper end,

LSL = 5, USL = 14, T = 0.5 (5 + 14) = 9.5

zU

Here, T, and Table 3-2 cannot be used.

USL 1410

(zU) = (2) = 0.97725

PPMU = 1,000,000
1 000 000 1 zU
= 1,000,000 11 0.97725 = 22750
31

32

The overall PPM

3.2 Process Capability Ratio Cpk

PPM = PPML + PPMU = 6210 + 22750 = 28960

In Fig 3-6, LSL and USL are the same for all six processes.
The standard deviation of all p
processes is equal
q to 2.

It can be seen that most of the defectives occur at the

upper
pp end, as the pprocess mean has shifted to the
right side.

Therefore, all of the six processes have the same Cp (Eq (3-8)).
Cp

USL LSL 62
62
38

=2
6
62

(3-26)

That is, if judged by Cp, the six processes have the same
process capability
capability. It is somewhat misleading
misleading.
33

34

However, obviously, the process capability degenerates from top

to bottom,
bottom because more and more defective units will be
produced (PPM becomes larger and larger from the top to the
bottom).
Whyy cant Cp tell this difference of the p
process capabilities
p
of the
six processes? Because, Cp only considers the ratio between (USL
LSL) and R (= 6), and does not take into account the centering
of process distribution between the specification limits.
As a conclusion, Cp is a valid measure of process capability only
when = T.
When T, other more appropriate measures should be used for
the process capability
capability.
Figure 3-6 Relationship of Cp and Cpk

35

36

Cpk is another process capability ratio that takes into account

not only the ratio between (USL LSL) and R (6), but also
the location of the process mean relative to the
specification center T.
CPU
C pk minCPU, CPL
CPL

if CPU CPL
if CPU CPL

(3 27)
(3-27)

USL
3

LSL
3
USL LSL
Cpk min
i CPU
C , CPL
C min
i
,

(3-29)

The more moves to the upper side, the smaller the (USL - )
and the smaller the Cpkk.

CPL

USL LSL

3
3
Cpk = CPU = USL
3

When > T, the upper side is more critical, most of the

d f ti
defectives
will
ill be
b produced
d d att the
th upper side.
id Therefore,
Th f
CPU
is smaller than CPL. Cpk is equal to CPU.

Since (Eqs (3-11) and (3-14))

CPU

If > T (Fig 3-7(a)),

(3 28)
(3-28)
37

38

If < T (Fig 3-7(b)),

USL LSL

3
3

Cpk = CPL =

LSL
3

(3-30)

When < T, the lower side is more critical, most of the

defectives will be produced at the lower side. Therefore, CPL is
smaller than CPU. Cpk is equal to CPL.
The more moves to the lower side, the smaller the ( - LSL)
andd the
h smaller
ll the
h Cpk.

Figure 3-7 Cpk for different cases

39

40

If = T (Fig 3-7(b)),

USL LSL

or CPU = CPL
3
3
USL
Cpk = CPU =
3
Or Cpk = CPL =

LSL
3

In Figure 3-6,
For process (a)
(3-31)

F process (b)
For
(3-32)
(3
32)

From Eq (3-31),
(3 31),
USL 2 (USL ) USL LSL

Cp
Cpk =
3
2 3
6

62 50 50 38
C pk min
,
min 2,2 2
3 2
3 2
62 53 53 38
,
C pkk min
min1.5,2.5 1.5
3 2
3 2

For process (c)

((3-33))

62 56 56 38
,
C ppk min
min 1,3 1
3 2
3 2

It indicates that,
that if = T,
T the two process capability ratios,
ratios Cpk
41
and Cp, are equivalent.

For process (d)

42

The value of Cppk depends on the difference ( - T). It means

that Cpk takes into account the location of the process mean
relative to the specification center T.

62 62 62 38
C pk min
,
min 0 , 4 0
3 2
3 2

As a conclusion, when T, Cpk is a more appropriate

measure off process capability
bilit than
th Cp.

F process ((e))
For
62 65 65 38
C pk min
,
min 0.5,4.5 0.5
3 2
3 2

The Cpk value becomes smaller and smaller from pprocess (a)
( ) to
process (e). It exactly indicates the fact that the process
capability degenerates from process (a) to process (e).
43

44

It can be shown that Cpk is always smaller than or equal to Cp.

Cpk < Cp

when T

Cpk = Cp

only when = T.

(3-34)

While Cpk is called the actual capability of a process, Cp is

called the p
potential capability
p
y. Cp is the maximum p
possible
value of the process capability that may be achieved for a
process when = T.

Figure 3-8 Increasing Cpk to Cp.

When the process mean is far away from the T ( < T or >
T), Cpk is very small. Cpk becomes larger and larger when is
moved towards T from either side. Finally, when is
coincident
i id t with
ith T,
T Cpk = Cp.
45

( T )
2

T=

LSL + USL)

(2)Reduce , ( is the denominator in Eq. (3-28)).

Ckm

(
(3-35)
)
where, and are the mean and standard deviation of a
qualityy characteristic x;; and T is the target
q
g of x.
1
(
2

(1)Move towards T, in order to make the actual capability Cpk

closer to Cp.
46

Ckm is calculated by the following formula.

3.3 Process Capability Ratio Ckm

Ckm is defined based on Taguchis Loss
2Function
2

How to improve the actual capability (increasing Cpk)?

USL LSL
6

(3-37)

Ckm is
i greater when
h is
i smaller.
ll

(3-36)

Professor Taguchi uses 2 to measure the loss incurred by a

d t i t productquality.
deteriorate
d t lit From
F
Eq
E (3-35),
(3 35) th
the smaller
ll th
the
value or the difference ( - T), the smaller the value of 2(or

) Conversely
).
Conversely,
a small
(or 2 ) indicates
better
product quality or higher process capability.

Since Ckm depends on , and takes into account the location

of the process mean relative to the specification center T (i.e.
- T),
T) so,
so like Cpk, Ckm can be used to handle the off-center
process.
Same as for Cpk, there are two ways to increase Ckm (or to
reduce the loss function 2 ), see Eq
q ((3-35).
)
(1)Reduce | - T|, i.e., centering the process mean at T.

47

(2)Reduce .

48

Example 3-3 (Figure 3-9)

The weights of three products A, B and C follow normal

distributions. For all three products, the target value T of the
weight is equal to 50g, and LSL = 43g, LSL = 57g.
For product (A),

= 50, = 2

2 22 (50 50)2 4
Ckm

=2

57 43
1.167
6 2
6

For product (B),

= 52, = 2

2 2 2 (52 50) 2 8
C km

= 2.828

57 43
0.825
6 2.828

For product (C),

= 50, = 3

2 32 (50 50)2 9 = 3
Ckm

Figure 3-9 Two processes

49

57 43
0.778
63

Product (A) has smallest loss function value, or highest

process capability ratio Ckm.
The quality of product A is better than the quality of product
B, because the mean of the former is coincident with the
target
ta
get T..
The q
qualityy of product
p
A is also better than the quality
q
y of
product C, because the former has smaller standard deviation.
51

(f9-10 p445)

50

Summary

(1)Cp is the simplest process capability ratio. It is a valid

measure off th
the process capability
bilit when
h the
th process mean is
i
coincident with the center between the specification limits. It
indicates the potential capability of a process.
process
(2)Cpkk measures process capability based on PPM (Defective
Parts Per Million). It can handle off-center process capability.
(3)Ckm measures process capability based on Taguchis loss
function. It is also able to handle off-center process capability.

52

For the following error components (see Figure 4-1):

4 GAGE REPEATABILITY AND REPRODUCIBILITY

4.1 Gage (Measurement) Error

An iimportant aspect off SQC iimplementation

A
l
i is
i to ensure
adequate accuracy and precision for gage and measuring system.

(11) = (10) + (1)

In any problem involving measurements, some of the observed

variability will be due to production itself, and some will be due
to measurement or gage error. Gage error includes all the
variabilities related to the gage, as well as the measuring system.
I may be
It
b caused
d bby the
h gage iitself,
lf or by
b the
h inspectors,
i
or by
b the
h
measuring conditions and procedures.

production is the deviation of a produced dimension from its

target
g value. gage isthe measuring
g deviation from the
produced dimension. total is the total deviation from the
target.
Both production and gageare random variables and usually
follow thenormal distribution. Since total is a function in
terms of production and gage . Therefore, total is also a
normally distributed random variable.

total

Statistical methods can be used to separate the components of

variability, as well as to give an assessment of gage capability, so
that corresponding actions can be taken to reduce the overall
1
variability (errors).

Figure 4-1 Error components

production

gage

(4-1)

Figure 4-2 Components of variability

3

The gage error cannot be determined based on an individual

gage , instead,
d i ti
deviation
i t d it mustt be
b measuredd by
b the
th standard
t d d
deviation gage of the probabilitydistribution of all values of gage .
Similarly the production error cannot bedetermined based on an
Similarly,
individual deviation production , it must be measured by the standard
deviation production
.
d ti

total standard deviation of total .

total

indicates the total

error

productiostandard deviation of production . production indicates the

production
d i error.

ggage
g

standard deviation of gage .

ggage
g

indicates the
gage error.

Like Eq (4-2),

2
gage

2
repeatability

Because of Eq (4-1), the variances (squares of the standard

deviations ) satisfy :
2
2
2
total
production
gage

(4-2)

Further Decomposition of Gage Error

The gage error can be further decomposed into the error
relating
l ti to
t repeatability
t bilit andd the
th error relating
l ti to
t reproducibility.
d ibilit

ggage
g repeatabil
p
ityy reproducib
p
ilityy

(4-3)
(4
3)

Repeatability
regards the inherent error of the gage
i lf (inadequate
itself
(i d
precision).
ii )
Reproducibility
regards the variability due to different
inspectors, or the error made by the inspectors.

EXAMPLE 4-1 (Table 4-1)

2
reproducibility

(4-4)

Each of the 20 produced parts has been measured twice

b each
by
h off th
the three
th inspectors.
i
t
All off them
th use a same
gage.

O
Overally,
ll
total production gage
production repeatability reproducib ility

2
2
2
total
production
gage
2
2
2
production
repeatabil
ity reproducibility

(4-5)

(4-6)

(1)The production
error

production

T bl 4-1
Table
4 1 Data
D ffor E
Example
l 4-1
41

The production error is reflected by the differences among the

x (j = 1,
average readings ~
1 2,
2 , 20) of the 20 different parts
parts.
Because if there is no production error, all of the 20 parts have
the exactly same dimension,
dimension then the difference between the
~
20 x j values should be equal or very close to zero.
j

~
Each x is the average of the three x values from the three
inspectors for the jth part. For example,
j

~
x1 (20.5 20.0 20.0) / 3 20.17

~
x2 (23.5 24.0 23.5) / 3 23.67
9

To evaluate production, we set up the following sample based on the

20 ~x j values.
values The sample size is (n = 20) and number of samples
is m = 1.

The sample range R is equal to (~x15 - ~x20 = 29.67 18.17=11.5) and

is attributable to the production error production
d ti .
R R / 1 R 11.5

production

R 11.5

3.079
d 2 3.735

(4-7)
(4
7)
(4-8)
(4
8)

10

(2)The error relating to repeatability repeatability

The error relating to repeatability is reflected by the
diff
difference
off two
t readings
di
taken
t k by
b a same inspector
i
t for
f a
same part. Because if the gage is perfectly precise, the two
readings will have the exactly same value,
value then the
difference between them should be equal or very close to
zero.
zero
To evaluate repeatability, we consider the ppairs of the two
readings from a part taken by an inspector, and use each of
the pairs as a sample. There are totally 60 such samples in
Table 4-1 (20 samples for each of the three inspectors).

where, d2 = 3.735 can be found from Table 22-33 for n = 20.

11

12

Each sample contains two readings. The sample size is (n = 2)

and number of samples is m = 60.
The sample
Th
l range R off eachh off these
th
60 samples
l are listed
li t d att
the right hand side. R is attributable to the inherent error of the
gage repeatibility.
R

(1 1 0) (0 0 1) ( 2 1 2)
1.15
3 20

repeatability

R 1.15

1.02
d2 1.128

(4 9)
(4-9)
(4-10)

where, d2 = 1.128 for samples of size n = 2.

13

(3)The error relating to reproducibility reproducibility

The error relating to reproducibility is reflected by the difference
of the three grand average ~xi (i = 1, 2, 3) from the three different
inspectors. Because if there is no variability among the three
inspectors, the three ~xi will be equal or very close to each other.
To evaluate reproducibility, we set up the following sample based on
the xi values
al es shown
sho n at the bottom of Table 4-1.
41

14

The sample range R is attributable to the variability relating

to the inspectors or the reproducibility reproducibility.

0.32
0.32
1

reproducibility

R 0.32

0.19
d 2 1.693

(4-11)
(4-12)

where d2 = 1.693
where,
1 693 for sample size (n = 3).
3)
This single sample contains three ~xi . The sample size is (n = 3) and
number of samples
p is m = 1.
15

16

Next, the gage error

2
2
2
gage
repeatabil
ity reproducib ility

gage

Finally, the total error

2
2
2
total
production
gage

((4-13))

1 .02 2 0 .19 2 1 .08

1 .039

gage

Here, the inherent gage error repeatibility is much larger than the
Here
error
reproducibility due to the variability of the inspectors.

3.0792 1.0392 10.560

3.250

(4-14)

The majority of the total error is attributable to the

production error in this example.
example The contribution from the
gage error is negligible.
Or in other words, the gage error will not significantly
influence the total error and the g
gage
g is considered suitable
for this measuring application.

17

4.2 The Gage capability Cgage

The distribution of gage is usually well approximated by a normal
distribution (Fig 44-2).
2). Then, 6 gage is a good estimate of the
distribution range.
While the process capability ratio Cp for a quality characteristic x
is
Cp

USL LSL
6 total

(4-15)

USL LSL
6 gage

A large value of Cgage indicates that the gage error is

negligible
li ibl compared
d tto th
the design
d i tolerance,
t l
andd will
ill
not significantly influence the total error.
Suppose, in example 4-1, the part has USL = 1077 and
LSL = 1002,
1002 and gage = 1.039
1 039
C gage

1077 1002
12.03
6 1.039

((4-17))

Here, the
h distribution
di ib i range off gage is
i only
l one twelfth
lf h off
the design tolerance. The gage error is negligible.

the ggage
g capability
p
y is
C ggage
g

18

((4-16))
19

As a generally rule, Cgage 10 often implies adequate gage

capability and the gage is suitable for the measuring
capability,
20
application.

Summary

production

gage

production
p

repeatabil
p
ityy

gage

repeatability

total

total

production
2
production

production

reproducib
p
ilityy

reproducibility

After the components of errors have been separated,

corresponding
di actions
ti
can be
b taken
t k to
t reduce
d
the
th mostt
dominant component(s).

attributable to the variability of the

manufacturing process producing the part.

repeatability

attributable to the inherent error of the gage

used
d tto measure the
th part.
t

reproducibility

attributable to the variability of the inspector

21
measuring the part.

production

can be reduced by improving the process and

using machines with higher precision
precision.

repeatability

can be reduced by
y usingg more precise
p
gage.
g g

reproducibility
p
y

can be reduced by training the inspectors or

setting more uniform regulation for using the
gage.
22

5 ACCEPTANCE SAMPLING PLAN

The number N of units in a lot is called lot size.

Acceptance sampling
A
li plan
l is
i concernedd with
i h inspection
i
i andd
decision making regarding the lots of products. Usually, this
decision is either to accept or to reject the lot.
lot

If every unit in a lot is inspected, it is the 100% inspection.

Acceptance sampling plan is used primarily for incoming or

receiving inspection. But, frequently a manufacturer will apply
the acceptance sampling to its own product between various
stages
t
off production.
d ti

If only a sample of n units (n < N) are inspected, it is the

sampling inspection
inspection. For example
example, suppose lot size N =
100 and sample size n = 10. For every lot, 10 units will be
randomly
y selected and,, then,, inspected.
p
The sampling inspection is used widely to replace the 100%
inspection, because it can save the amount of inspections,
and is especially useful in destructive testing.

Since the sampling inspection makes a conclusion about the

quality of the whole lot based on the result of inspecting a
sample of only n units, it may make two types of errors.

The fraction nonconforming of a lot is equal to

P=D/N

In an example,
example N = 500,
500 n = 5.
5

(5-1)
where,
h
D iis th
the total
t t l number
b off nonconforming
f
i units
it in
i the
th lot.
l t
Meanwhile, the fraction nonconforming can be estimated by

p = d / n

(5-2)

where, d is the total number of nonconforming units found in a

sample
p of size n.

(1) Suppose a lot contains 5 nonconforming units (D = 5),

p = D/N = 5/500 = 0.01, the quality is quite good.
Now,, 5 units are randomlyy picked
p
for sampling
p g inspection.
p
Accidentally, they may be just the 5 nonconforming units in the
lot (d = 5). The estimate fraction nonconforming is
.

p d / n 5 / 5 100 %

It must mislead the inspector to reject the lot, he may believe

that the lot quality is very bad and reject the lot.
This case implies a type of error called the producers risk.

(2) Suppose another lot contains 250 nonconforming units

(D = 250), p = D/N = 250/500 = 0.50, the quality is quite
poor.
However, if the 5 units randomly picked for the sampling
However
inspection are, by chance, all conforming ones (d = 0), the
estimate fraction nonconforming is p d / n 0 / 5 0 . The
inspector will also be misled and believes that the lot
quality is extremely good and accept the lot.

(1)Producers risk: a lot is rejected when

hen quality
q alit is good.
good The
probability of this risk is denoted by .
(2)Customers risk: a lot is accepted when quality is bad. The
probability of this risk is denoted by .

This case implies

p
another type
yp of error called the
customers risk.
5

The pperformance of a sampling

p g plan
p is usuallyy measured byy and
. An effective acceptance sampling plan should minimize both6
and .

5.1 Single Sampling Plan

A single
i l sampling
li plan
l makes
k an accept or reject
j decision
d i i based
b d
on the inspection of a single sample.
Two parameters determine a single sampling plan.
Sample size:

n (number of units in a sample)

Acceptance number:c
Let d be the number of nonconforming units found in a sample,
If d c,
c

accept the lot;

If d > c,

reject the lot.

(5-3)

The performance of a single sampling plan is totally determined

by n and c. The lot size N is not an influential factor, as long as n
<< N.
7

Example 5-1
5 1 A single sampling plan (n = 89,
89 c = 2) is used
to check the incoming lots, each of them contains 1000
washers (N = 1000,
1000 N >> n).
n)
From
o eac
each lot
ot of
o 1000
000 units,
u ts, a random
a do sample
sa p e oof 89 uunits
ts
are inspected and the number d of nonconforming units is
counted. If d is less than or equal
q to 2, the lot is accepted;
p
if
d is greater than 2, the lot is rejected.
Suppose the values of d are equal to 2, 0, 3, 1, 5,
respectively, for the samples from the first five lots, then
l 1,
lots
1 2 andd 4 are accepted,
d andd lots
l 3 andd 5 are rejected.
j
d

Operating-Characteristic
p
g
((OC)) curve ((Figure
g
5-1))

It is noted that,
that during the use of a sampling plan
plan, the two
parameters n and c are fixed, but the number d is different
from sample
p to sample.
p d is a random number and
follows a binomial distribution.

The OC curve is a very useful tool to show the overall

performance characteristic of a sampling plan. It plots Pa
versus the fraction nonconforming p. Pa is the probability
that a lot is accepted.

For given n and p (p is fraction nonconforming), the

probability for d equal to an integer i is
n
n!
Prd i Pd (i) pi (1 p)ni
pi (1 p)ni
i!(n i)!
i

(5-4)

Example:
E
l A truck
t k delivers
d li
1000 llots
t off washers
h to
t
Company A. The company carried out a single sampling
inspection on each lot.
lot If 98 lots are rejected by the
sampling plan,
Pa

1000 98
0.902
1000

(5-5)

10

Pa = Prob(a lot is accepted) = Prob(d c)

If p increases d is likely to increase

it is less likely (d c) Pa is likely to decrease.

So, if p increases, Pa decreases; if p decreases, Pa

increases.
increases

Figure 5-1 OC curve of the single sampling plan

(f13-2 p614)

11

12

Example 5-2 For the single sampling plan where n = 89 and,

c = 2.
Under a single sampling plan (n, c), the value of Pa for a
given p value is calculated as follows:

When p = 0.01, using Eq (5-6),

Pa Pr obd c

Pd (0) Pd (1) Pd (c)

c

(5-6)

n!
p i (1 p ) ni
i 0 i! ( n i )!

Pd (i )
i 0

89!
( 0.01) i ( 0.99 ) 89 ii
i 0 i! (89 i )!
89!
89!
89!

( 0.01) 0 ( 0.99 ) 89
( 0.01)1 ( 0.99 ) 88
( 0.01) 2 ( 0.99 ) 87
0!89!
1!88!
2!87!
0.9397

Pa

Pr
P ob
bd 0 Pr
P obbd 1 Pr
P obbd c

The entire OC curve is drawn by connecting several points (Pa, p).

13

14

When p = 0.05,
2

89!
(0.05)i (0.95)89i
)!
i 0 i!(89 i )
89!
89!
89!

(0.05)0 (0.95)89
(0.05)1 (0.95)88
(0.05)2 (0.95)87
0!89!
1!88!
2!87!
0.1721

Pa

After more points of (Pa, p) are calculated and plotted, the OC

curve can be drawn.
drawn
Different sampling plans have different performance, and
therefore, different OC curves, depending on the parameters n
and c.
15

Fi
Figure
5-2
5 2 OC curves ffor diff
differentt sampling
li plans
l
16

Acceptable Quality Level AQL

AQL is a small value for p, representing a high quality level. If p
= AQL, the product quality is considered good and the lots should
be accepted.
For example,
p , suppose
pp
AQL
Q is set as 0.01. If p = AQL
Q = 0.01,, the
product quality is considered good and the lots should be
accepted.
Producers risk is the probability that a lot is rejected when
qquality
y is ggood. Here,, quality
q
y is ggood means (p = AQL).
Q ) So,,
= pprob(lot
( is rejected)
j
) | quality is good = pprob(lot
( is rejected)
j
) | p = AQL
= 1 - prob(lot is accepted) | p = AQL = 1 - Pa | p = AQL

(5-7)
17

is smaller when Pa | p = AQL is larger.

Rejectable Quality Level RQL

Pa

pAQL
AQL

0.9397 , = 1 0.9397 = 0.0603 = 6.03%

Discussion: At p = AQL = 0.01, the product quality is

considered good. The producer hopes that all of the lots are
accepted. However, if 10000 lots are inspected by the
sampling
li plan,
l 6.03%
6 03% off th
them (603 llots)
t ) will
ill be
b rejected,
j t d on
average.
10000 = 10000 0.0603 = 603
A large means that the probability that lots of good quality
are rejected
j
is high.
g It is a risk (or
( loss)) suffered byy the
producer.
18

In Fig 5-1, RQL = 0.05,

RQL is a large value of p, representing a low quality level. If p =

RQL, the product quality is considered bad and the lots should be
rejected.
For example, suppose RQL is set as 0.05. If p = RQL = 0.05, the
product quality is considered bad and the lots should be rejected.
Customers risk is the probability that a lot is accepted when
quality is bad. Here, quality is bad means (p = RQL). So,
= prob(lot is accepted) | quality is bad
= prob(lot is accepted) | p = RQL = Pa | p = RQL

In Figg 5-1,, AQL

Q = 0.01,,

Pa

p RQL

0.1721

, = 0.1721 = 17.21%

Discussion: At p = RQL = 0.05, the product quality is

considered
id d bad.
b d The
Th customer
t
hopes
h
that
th t all
ll off the
th lots
l t are
rejected. However, if 10000 lots are inspected by the sampling
plan 17.21%
plan,
17 21% of them (1721 lots) will be accepted
accepted, on average.
average
10000 = 10000 0.1721
0 1721 = 1721
A large
g means that the p
probabilityy that lots of bad quality
q
y are
accepted is high. It is a risk (or loss) suffered by the customer.

(5-8)
(5
8)

is smaller when Pa | p = RQL is smaller.

19

20

(a) When n is fixed, increasing c implies that it is more likely that

d c (or that, Pa = Prob(d c) becomes higher). It will reduce ,
but increase (Fig 5-3 (a)).

For an effective
F
ff i sampling
li plan,
l its
i OC curve should
h ld
display (see Fig (5-1)):

(b) Co
Conversely,
ve se y, dec
decreasing
eas g c will
w reduce
educe , but increase
c ease ((Figg 5-3
53
(b)).

(1) A high Pa p AQL value. It means that lots of good

quality are very likely to be accepted,
accepted or the producer
producerss
risk is low.

(c) When n is allowed to be increased,

increased both and can be reduced
simultaneously. But the cost of the sampling inspection will
become higher.
higher (Fig 55-3
3 (c)).
(c))

(2) Also, a low Pa pRQL value. It means that lots of poor

qquality
y are veryy unlikelyy to be accepted,
p
or the customers
risk is low.
The values of and of a single sampling plan are
determined by the parameters n and c.

Figure 5-3 OC curves for different sampling plans

21

How is the lot formed ?

I can influence
It
i fl
the
h effectiveness
ff i
off the
h acceptance sampling
li
plan.
(1)
All lots inspected by a same sampling plan should be
homogeneous (the p values for all lots are almost identical). The
units should be produced by the same machines and the same
operators, from the common raw materials, and approximately at
th same time.
the
ti
An acceptance sampling plan is designed based on a
specified (or predetermined) p value and works most effectively
at this p level (or it produces least and ). If lots are
nonhomogeneous with different p values, the acceptance
sampling plan may not function as effectively as it could.
Furthermore, nonhomogeneous lots make it more difficult
23
to identify and eliminate the source of the problems.

(2)

22

Larger lots are preferred over smaller ones. It can be

shown that larger
g lot is usuallyy more effective in sense of
and .
For example
example, in a production line
line, 3000 units are produced
daily. The available resource allows 300 (10%) units to be
inspected. The following two schemes may be used. Both
schemes inspect 300 units per day. However, scheme 1
uses a larger
g lot size than scheme 2. Therefore, the former
is more effective (producing smaller and ) than the
latter.

24

(3) Lots should be comfortable to the material-handling systems.

For example
example, if each box contains 500 units
units, it is more
comfortable to have a lot size of 500 (i.e., each box is a lot) than
a lot size of 400.

Most of the acceptance sampling plans do not provide any direct

improvement of the product quality
quality. The quality of the accepted
lots are no better than the quality of the rejected ones.
For example, N = 1000, n = 100 and c = 9.
Lot A:

contains 120 nonconforming units, pA = 0.12

Randomly inspect 100 units, and d = 9

Lot B:

(accepted)

contains 120 nonconforming units, pB = 0.12

R d l iinspectt 100 units,
Randomly
it and
d d = 11 (rejected)
( j t d)

(4) The units selected for inspection from the lot should be chosen
at random, and should be representative of all the units in the lot.
25

The most effective way is to use an acceptance sampling plan

as an audit tool.
An effective sampling
p gp
plan accepts
p most of the lots if the
product quality is good, and rejects almost all lots if the
product quality is poor (see Fig 5-1). By this way, good
quality is rewarded, and poor quality is penalized. The
supplier is under high pressure to improve and ensure the
quality of his products.

27

Lot A is accepted and lot B is rejected.

rejected It happens just by chance.
chance
In fact, lot B has the same quality level as lot A, as pA = pB. 26

Example: Our company uses a single sampling plan (n = 89, c = 2)

to inspect the parts supplied by vendors A and B.
B One day
day, each of
the two vendors delivers 10000 lots of parts to our company.
The products
Th
d
off vendor
d A hhas a quality
li off p = 0.01,
0 01 then
h Pa =
0.9397 (Fig 5-1). Consequently, majority of the lots (9397) have
been accepted
accepted, only 603 lots are rejected.
rejected Vendor A is quite
satisfied, because he has been rewarded for the high quality of his
products.
products
The products of vendor B has a quality of p = 0.05, then Pa =
0 1721 A
0.1721.
As a result,
l most off the
h llots (8279) hhave bbeen rejected,
j
d
only 1721 lots are accepted. When the large amount of the rejected
lots are sent back,
back vendor B is badly disappointed.
disappointed He has to try
every way to improve the quality of his products, otherwise he
cannot stay in business.
business
It is the effect of the acceptance sampling plan.

28

5 2 Double Sampling Plan

5.2

The first sample of size n1 is drawn:

In a double-sampling
p g plan,
p , a second sample
p is sometimes
required before the lot can be sentenced (accepted or rejected).

Case 1:If d1 c1 the lot is accepted immediately by the first

sample, and the second sample is not needed.
Case 2:If
C
2 If d1 > c2, the
th lot
l t is
i rejected
j t d immediately
i
di t l by
b the
th first
fi t
sample, and the second sample is not needed.

A double-sampling plan is determined by four parameters.

nl = sample size of the first sample
cl = acceptance number of the first sample
n2 = sample
l size
i off the
th secondd sample
l
c2 = acceptance number for both samples (c2 > c1)

Case 3:If c1 < d1 c2, a second sample of size n2 is drawn.

After the second sample is inspected:

Let d1 and d2 be the numbers of the nonconforming

units found in the first and second samples
samples, respectively
respectively.
29

For cases 1 and

F
d 22, only
l one sample
l is
i needed;
d d for
f case 3,
3 two
samples are used.

(i)

If d1 + d2 c2, the lot is finally accepted.

(ii)

If d1 + d2 > c2, the lot is finally rejected.

30

Example
p 5-3 nl = 50,, cl = 1,, n2 = 100,, c2 = 3 ((Note,, cl < c2)

While the first sample of size n1 must be drawn

unconditionally the second sample may
unconditionally,
may, or may not,
not be drawn
drawn.
It is also noted that, in the second sample, c2 is used to check
(d1 + d2), rather than d2 only.

Fi
Figure
5-4
5 4 Double
D bl sampling
li plan
l
31

32

For a given double sampling plan (with parameters n1, n2, c1 and
c2), we may identify an equivalent single sampling plan (with
parameters n and c) which has an OC curve very close to the OC
curve of the double sampling
p g plan.
p
It means that the overall
performance of the two equivalent sampling plans are very much
the same.

(1)
If d1 1,
1 the lot is accepted by the first sample.
sample The
second sample is not needed.
(2)
If d1 > 3, the lot is rejected by the first sample. The
second sample is not needed.
(3)

It can be shown that, for a pair of equivalent double and single

sampling plans,

If 1 < d1 3, the second sample of size n2 is drawn.

(i)

If d1 + d2 3, the lot is finally accepted.

(ii)

If d1 + d2 > 3, the lot is finally rejected.

n1 < n < n1 + n2

It is noted that, after the second sample has been taken, there
must be a definite sentence (accept or reject) for every lot.
lot
33

(5-9)

The potential advantage of using a double sampling plan is that it

may reduce the average amount of inspections, compared to the
equivalent single sampling plan. It means, while the double
sampling plan performs as effectively as the equivalent single
sampling
li plan,
l the
th former
f
is
i less
l expensive
i to
t conduct
d t than
th the
th
34
latter.

For a single sampling plan, the number of inspected units in each

sample is consistently equal to the sample size n.
For a double sampling plan, if a lot is accepted or rejected on the
first sample,
sample the number of the inspected units is equal to the
sample size n1 of the first sample. Since n1 < n, using the double
sampling plan for this lot will save the amount of inspection.
inspection
On the other hand, if a decision cannot be made on the first sample
(
(second
d sample
l is
i needed),
d d) the
th number
b off the
th inspected
i
t d units
it is
i
equal to (n1+n2). Since (n1+n2) > n, using the equivalent single
sampling plan for this lot will save the amount of inspection.
inspection
It means that, by using a double sampling plan, the number of the
inspected units required by different lots will be different.
Consequently, we may have to consider the average number of
i
inspected
t d units
it per lot.
l t

Figure 5-5
5 5 Equivalent double and single sampling plans
35

36

For the agent, monthly income is not a constant. Suppose, there are
only two possible events:

A Story: Comparing the average incomes,

incomes AI,
AI of an
engineer and an insurance agent.

Event I (E1): the agent earns $10000 a month. The probability of

Event I is 0.3.
03

For the engineer, monthly income is a constant, equal to

$5000.
AI

5000

Event II (E2): the agent earns $2500 a month. The probability of

E t II is
Event
i (1 - 0.3).
0 3)

(5-10)

Therefore, the AI for the agent is

AI = (income for E1)(prob of E1) + (income for E2) (prob of E2)
= 10000 0.3
0 3 + 2500 (1 - 0.3)
0 3) = 4750

(5 11)
(5-11)

By comparing the AI values, it can be concluded that the engineer

makes
k more money, on average, than
h the
h iinsurance agent.

37

For the double sampling plan, sometimes n1 units are

inspected; and other times (n1+n2) units are inspected.

Average Sample Number, ASN --- the average number of

inspected units per lot.

Event I (E1): a decision about a lot is made on the first

sample and the number of inspected units is n1. Let the
pprobabilityy of Event I be PI.

For the equivalent single sampling plan, sample size n is a

constant For every lot
constant.
lot, n units are inspected.
inspected
ASN n

38

(5-12)

Event II (E2): a decision about a lot cannot be made on the

first sample (the second sample is needed) and the number
of inspected units is (n1+n2). The probability of Event II is
(5-13)
(5
13)
(1-P
(1
PI). Referring to Eq (5
(5-11).
11).
ASN = (number of inspected units for E1) (prob of E1) +
(number of inspected units of E2) (prob of E2)

n P (n n )(1 P )
1

39

n n (1 P )
1

40

How to determine PI ?
Example 5-3
53

PI = Pr{lot is sentenced on the 1st sample}

= Pr{lot is accepted on the 1st sample}
+ Pr{lot
{ is rejected
j
on the 1st sample}
p }
Prd1 c1 Prd1 c2

(5-14)

ASNsingle = n = 85

c1

n1!
pi (1 p)n1i

i
!
(
n
i
)!
i 0
1

Prd1 c1 Pd (0) Pd (1) Pd (c1 )

An equivalent double sample plan is also designed with

Prd1 c2 1 Prd1 c2 1 Pd (0) Pd (1) Pd (c2 )

c2

(5-15)

n1!
1
pi (1 p)n1i
i
!
(
n
i
)!

i 0
1

41

For the double sampling plan (see Eqs (5-13), (5-14) and (5-15)):
50!
Prd c
0.02 (1 0.02)
i!(50 i )!
1

i 0

Prd c 1
3

50i

i 0

50!
0.02 (1 0.02)
i!(50 i )!
i

0.7358
50i

A single sampling plan is used to inspect a can drink.

drink The
estimated defect rate (fraction nonconforming) is p = 0.02.
The sample size n = 85, and the acceptance number c = 2.
From Eq (5-12),

0.0177

PI Prd1 c1 Prd1 c2

nl = 50,
50 cl = 1,
1 n2 = 100,
100 c2 = 3

42

The comparison
p
in above example
p is conducted based on a
given value of p (p = 0.02). In fact, the ASN value of the double
sampling plan is a function in terms of p. Figure 5-6 compares
the ASN value of a double sampling plan with the ASN values
of the equivalent single sampling plan over different values of p.
(1)
When p is very small, the double sampling plan has
smaller
ll ASN th
than th
the equivalent
i l t single
i l sampling
li plan.
l Because,
B
most of the lots are accepted by the first sample and, therefore,
only have to inspect n1 units.
units

0.7358 0.0177 0.7535

ASN n1 n2 (1 PI )
50 100 (1 0.7535) 74.65

(2)
When p is very large,
large the double sampling plan also has
smaller ASN. Because, most of the lots are rejected by the first
sample
p and,, again,
g , only
y have to inspect
p n1 units.

Compare to (ASNsingle = 85), the double-sampling plan will

save the number of inspections (by 10.35 units per lot) in
thi example.
this
l
43

44

(3) When p is in the middle range, the double sampling plan

has larger ASN than the equivalent single sampling plan.
Because, many lots are neither accepted nor rejected by the
first sample, and the second sample has to be drawn. All of
these lots have to inspect (n1 + n2) units.
Above discussion means that using the double sampling plan
may not always save the amount of inspection.
inspection It is a better
choice only when p is either very small or very large.
But, a curtailment technique discussed below may ensure that
the double sampling
p g plan
p requires
q
less amount of inspection
p
than the equivalent single sampling plan.
Figure
g
5-6 ASN curve for single
g and double sampling
p g plans
p
45

46

Curtailment
As soon as the total number of observed nonconforming units
exceeds the second acceptance number c2, the inspection of the
double sampling plan is terminated and the lot is rejected. Because
further inspection will not change sentence of rejection.
Example 5-4, for a double sampling plan
nl = 50, cl = 1, n2 = 100, c2 = 3

Figure 55-7
7 Curtailment

During the inspection of a lot, in the first sample, units 4, 25 and

43 are nonconforming, d1 = 3. Since c1 < d1 c2, the second sample
has to be drawn.
In the second sample, units 2, 30, 45, 75 and 89 are
nonconforming, d2 = 5.

47

For an ordinary double sampling plan

Since d1 + d2 = 3 + 5 = 8 > c2, the lot is rejected. The
total number of inspections is equal to (n1 + n2 = 150).48

If the curtailed double sampling plan is used, the first sampling

proceeds as before.

Double sampling plan has two potential disadvantages.

disadvantages

However in the second sampling, after unit 2 has been inspected,

the total number of nonconforming units is (d1 + 1 = 3 + 1 = 4)
4).
It is already larger than c2 and the lot must be rejected even if all
of the remaining units are conforming.
conforming Thus,
Thus the second
sampling is terminated immediately and the lot is sentenced to be
rejected.

(1)Unless the fraction nonconforming p is very small or

very large (in which, most of the lots are accepted or
rejected
ejected in the
t e first
st sample),
sa p e), the
t e double
doub e sampling
sa p g plan
p a may
ay
require even more inspections on average than the
equivalent
q
single
g sampling
p g plan.
p

The total number of inspections is only equal to (n1 + 2 = 52). It

is much smaller than 150 units required by the double sampling
plan without curtailment.

However, the use of curtailment may significantly lower

ASN in the double sampling plan.

The curtailed double sampling plan always reaches the same

conclusion (accept or reject) as the double sampling plan without
curtailment.
t il
t But
B t the
th former
f
is
i able
bl to
t lower
l
ASN significantly
i ifi tl
49
(see Figure 5-6).

(2)It iis administratively

(2)I
d i i
i l more complex
l to use the
h double
d bl
sampling plan, compared to the single sampling plan.
50

5.3 Rectifying Inspection

Incoming fraction nonconforming:

p0

Rectifying inspection will improve the final quality of the

outgoing product. It is usually used for in-plant inspection, in
which the manufacturer wishes to ensure a required quality level
of the products before they are sent to the next stage.

Outgoing fraction nonconforming

(Average Outgoing Quality):

AOQ

AOQ < p0

(5-16)

IIn rectifying
tif i inspection,
i
ti
for
f eachh lot,
l t a single
i l sampling
li inspection
i
ti
is conducted first.
Event I:If the lot is accepted by the single sampling inspection, it
is sent to the output point without any further inspection.
Event II:If the lot is rejected by the single sampling inspection, a
100% inspection
p
is applied
pp
to the lot.
All discovered nonconforming units (either in the single sampling
inspection or in the 100% inspection) will be replaced by
51
conforming ones.

Figure 55-8
8 Rectifying inspection

(f13-10 p621 modify)

52

Average outgoing quality (AOQ)

Event I (E1): the lot is accepted by the single sampling

inspection.

Lot size N, incoming fraction nonconforming p0

The n inspected units are all conforming, because all discovered

nonconforming units have been replaced
replaced.

For the single sampling inspection:

But, the (Nn) non-inspected units may contain nonconforming

units. Since the rate of nonconforming is p0, the number of
nonconforming units contained in this portion is p0(N - n).

sample size n, acceptance number c.

Therefore, the average number of nonconforming units in an

accepted lot is

Let Pa be the probability of acceptance in the single

sampling
p g inspection.
p
Pa Prd c

d I = 0 + p0(N - n) = p0(N - n)

Pd (0) Pd (1) Pd (c)

c

n!
Pd (i)
p0i (1 p0 )ni
i 0
i 0 i!(n i)!

(5-18)

The probability of Event I is Pa.

(5-17)
53

54

Event II (E2): the lot is rejected by the single sampling

inspection.

Example: if N = 10,000,
10 000 n = 89,
89 c = 2,
2 p0 = 0.01
0 01
For an accepted lot,
lot 89 units have been inspected in the
single sampling plan. Any nonconforming units have been
replaced
ep aced by the
t e conforming
co o
g ones.
o es. So, after
a te the
t e single
s ge
sampling plan, there is no nonconforming units in these 89
units.

The whole lot contains no nonconforming units, because every

unit has been inspected either in the single sampling plan or in
th 100% iinspection,
the
ti
and
d all
ll discovered
di
d nonconforming
f
i units
it
have been replaced by the conforming ones. The average
number of nonconforming units in a rejected lot is,
is therefore,
therefore

9911 (= 10000 89) units have not been inspected in the

single sampling plan. The average number of nonconforming
units in this portion is 99.11 (= 0.01 9911).

d II = 0

(5-19)
(5
19)

The pprobabilityy of Event II is ((1 - Pa)).

Consequently, in this accepted lot, the average number of

nonconforming units is also equal to 99.11.
99 11
55

56

Now, combining both Events I and II, the overall average

number of nonconforming units in a lot is
d

E
Example
l 5-5
55

= ( d I for
f E1) (prob
( b off E1) + ( d II off E2) (prob
( b off E2)

A rectifying inspection plan is used to check the lots of

computer components. The incoming fraction nonconforming
is: p0 = 0.01
0 01

p0 ( N n) Pa 0 (1 Pa )
Pa p0 ( N n)

(5-20)

N = 10,000, n = 89, c = 2.

So, the outgoing fraction nonconforming is

d P p ( N n)
AOQ a 0
N
N

(5-21)

Pa p0 N
Pa p0
N

89!
0.01i (1 0.01)89i
i 0 i! (89 i )!
0.9397

If N is much larger than n, N n N,

AOQ

n!
p0i (1 p0 ) n i
i 0 i! ( n i )!

Pa Prd c

57

58

(5-22)

Using Eq (5-21)

AOQ is a function in terms of p0, (see Fig 55-9),

9) since

P p ( N n) 0.9397 0.01 (10000 89)

AOQ a 0

0.0093
10000
N

AOQ
Q

Pa p0 ( N n)
N

n!
p0i (1 p0 ) n i
i 0 i! ( n i )!

Pa

(5-23)

(1)
When p0 is very small, AOQ is also small. That means,
when
h th
the incoming
i
i quality
lit is
i very good,
d the
th average outgoing
t i
quality is very good too.

Since N >> n, using Eq (5-22),

AOQ Pa p0 0.9397 0.01 0.0094

Comparing AOQ with p0, it can be seen that the out-going fraction
nonconforming is smaller than the incoming one.
one
The reason is that some nonconforming
g units have been
discovered during either the single sampling inspection or the
100% inspection, and have been replaced by the conforming ones.
59

(2)
When p0 is very large, AOQ is again very small. Because,
when the incoming quality is very bad,
bad most of the lots are rejected
by the single sampling plan and subject to 100% inspection. Then,
duringg the 100% inspection,
p
, all nonconforming
g units are replaced
p
by the conforming ones. But, when p0 is large, the rectifying
inspection is very costly, because it requires a huge amount of
inspections and replacements.
60

(3)

IIn bbetween above

b
two extremes, the
h AOQ curve rises,
i
passes through a maximum, and then descends.
The maximum AOQ value is called the Average
Outgoing Quality Limit (AOQL),
(AOQL) which represents
the worst possible average quality (or the highest
possible value of AOQ) that would result from the
rectifying inspection. The outgoing lots are warranted to
have an AOQ
Q value no higher
g
than AOQL, regardless
g
of the incoming quality p0.
It is a very useful feature of the rectifying inspection.

Figure 5-9 AOQ curve

61

For example, in a production line, it is required that the incoming

fraction nonconforming
g to stage
g 2 be smaller than 0.016.
A rectifying inspection plan in designed: n = 89, c = 2, N =
10000. Correspondingly, AOQL = 0.0155 (Fig 5-9).

62

This rectifying
y g inspection
p
is inserted between stage
g 1 and
stage 2. Then, the outgoing quality of stage 1 is the incoming
fraction nonconforming p0 for the rectifying inspection.
Now, regardless of the value of p0, the AOQ of the rectifying
inspection (or the incoming p0 of stage 2) is always smaller
than 0.0155. As a result, the requirement that the incoming
f ti nonconforming
fraction
f
i to
t stage
t
2 be
b smaller
ll than
th 0.016
0 016 is
i
satisfied.

Figure
g
5-10 Application
pp
of AOQL
Q
63

64

5.4 Sampling Plan Standard

MIL STD 105E is a widely used acceptance sampling
system (standard). It is an overall strategy specifying the
way in which several sampling plans are to be used
together.
g
Particularly,
y, the standard will aid the QA
Q
engineers to determine the parameters (e.g., n, c) of the
sampling plan that will satisfy the inspection requirements.
The primary focal point of MIL STD 105E is the acceptable
quality level (AQL). Or in other words, the producers risk
is warranted at a specified low level. But the customers
risk
i k is
i nott guaranteed.
t d The
Th value
l off depends
d
d on the
th
sample size n. If the sample size is larger (i.e., more effort
will be spent on the sampling inspection),
inspection) the value will
65
be smaller.

The sample size n used in MIL STD 105E is mainly decided by

the two factors: the lot size N and the inspection
p
level.
(1) Lot size N. The higher the N, the larger the n should be.
(2) Inspection level. It is an indication of the effort spent on the
sampling inspection. The higher inspection level requires larger
sample
l size,
i andd will
ill result
lt in
i smaller
ll customers
t
risk
i k .

General inspection level

Figure 5-11 Inspection levels and customers risk

66

S i l inspection
Special
i
i level
l l
S 1 S-2,
S-1,
S 2 S-3,
S 3 and S-4.
S4
They use very small n (even smaller than in level I),
I) and
should be employed only when the small sample sizes are
necessary and high can be tolerated.

Level I: requires a sample size about one-half of that for

inspection Level II. It is used when higher can be tolerated.
g
as normal.
Level II is designated
Level III requires a sample size about twice as that for inspection
Level II.
II It is used when must be low
low.
67

68

Table 5-1 Sample size code letters

Table 5-1 associates a Sample Size Code (SSC) letter with the
lot size N and inspection level. The higher the alphabetical
order of the SSC letter, the large the sample size will be.
The alphabetical order of the SSC letter increases from top to
b tt
bottom
(along
( l
with
ith the
th increase
i
off the
th lot
l t size)
i ) andd from
f
left
l ft
to right (along with the increase of inspection level)
Suppose, if the lot size N = 2000 and a level II inspection is
used the SSC letter is K.
used,
K The actual value of the sample size
can be found based on the SSC letter from Tables 5-3 to 5-5.

69

70

A product can be inspected by one of the three types of

sampling plans.

Tightened
g
sampling
p gp
plan to be used when there is an
indication that the product quality becomes worse.

Normal sampling plan to be used as long as the supplier

is producing the product at AQL or better.

Example: n = 125, c = 1
(i) d c, accept

Example: n = 125, c = 2
(i) d c, accept

(ii) d > c, reject

The tightened
Th
ti ht d plan
l uses same n but
b t smaller
ll c comparedd to
t
the normal plan. Therefore, the condition of acceptance is
more stringent.
stringent

(ii) d > c, reject

For example
example, if d = 2 in a sample,
sample the corresponding lot will
be accepted if a normal sampling plan is in use; but the same
lot will be rejected
j
if a tightened
g
sampling
p g plan
p is used.
71

72

Reduced sampling plan to be used when there is an

indication that the product quality has been significantly
improved.
p
Therefore,, the effort of inspection
p
can be relaxed
(n is reduced).

(iii)

c < d < r,

(5-24)

Accept the current lot, but switch to the normal sampling

plan

Example: n = 50, c = 1
r = 3 (rejection number)
(i)

d c,

(ii)

r,r

acceptt
reject

73

74

Switch Rules

(3)Discontinuance of inspection.

Th sampling
The
li inspection
i
i system will
ill use different
diff
types off
sampling plans at different time according to the current quality
level of the products.
products

If ten consecutive lots remain on tightened inspection (i.e., cannot

be switched back to normal inspection), the entire inspection
should be stopped, and actions should be taken to improve the
quality of submitted lots.

The sampling inspection usually starts with the normal

sampling plan.
((1)Switch
)
from normal inspection
p
to tightened
g
inspection
p
When two out of five consecutive lots have been
rejected on the normal inspection.
inspection It is an indication that
product quality becomes worse.
(2)S i h ffrom tightened
(2)Switch
i h
d iinspection
i to normall inspection
i
i
p on the tightened
g
When five consecutive lots are accepted
inspection.
75

Figure
g
5-12 Switch between normal and tightened
g
plans
p
76

(4)Switch normal inspection to reduced inspection, when all of

the following four conditions are satisfied (an indication that
product quality has been significantly improved).
(i)The preceding 10 lots have been on normal inspection, and all of
them have been accepted.
(ii)The total number of nonconforming units in the preceding 10
samples is less than or equal to a specified limit D.
D
(iii)Production is at a steady condition; that is,
is no problems such
as machine breakdowns or material shortages.
(iv)Reduced inspection is considered desirable by the authority
responsible for the sampling inspection.
77

Table 5-2 Number of nonconforming units in the samples of 10 lots

Example for condition (ii): for a normal sampling plan,

plan
n = 89,, c = 2
In addition, D = 13 is specified as a limit for the sum of the
nonconforming units found in the preceding 10 samples.
It is found from Table 5-2 that, all of the 10 lots are accepted,
but, the total number of nonconforming units in the 10
samples
l from
f
these
th
10 llots
t is
i equall to
t 15,
15 larger
l
than
th D.
D That
Th t
means, condition (ii) has not been satisfied.
Consequently, the sampling inspection is not allowed to be
switched to the reduced sampling plan
plan, even if all other three
conditions are satisfied.
78

(5)

Switch from reduced inspection to normal inspection,

when any of the following four conditions is met.
(i)A llott is
i rejected.
j t d
(ii)Production is irregular or delayed.
delayed
(iii)When c < d < r,
r
Referring
g to Eqq (5-24),
(
), the current lot is accepted,
p ,
but normal inspection is reinstituted starting with the
next lot.
(iv)Other conditions warrant that normal inspection be
instituted.

79

80

Example 5-6 A step-by-step procedure of designing and using a

MIL STD l05E acceptance sampling system for an electronic
product.
(1)Specify Acceptable Quality Level AQL.[AQL = 0.65%]

Figure
g
5-13 Switch rules for three types
yp of inspections
p
81

(7)Determine the parameters of the sampling inspection plans.

[
Normal:
n = 125, c = 2.
Tightened:

n= 125, c = 1.

Reduced:

n = 50, c = 1, r = 3.

(2)Ch
(2)Choose
th
the iinspection
ti level.
l l

[l l II]
[level

(3)Decide the lot size N.

N

[N = 2000]

(4)Find the appropriate SSC letter

from Table 5-1.

[K]

(5)Decide the appropriate type

(single or double) of sampling plan.

[single]

(6)Check the appropriate tables

[Tables 5-3, 5-4, 5-5]

82

Table 5-3
5 3 Table for normal inspection

]
Note: on Tables 5-3,, 5-4 and 5-5,, AQL
Q = 0.65% is displayed
p y as
0.65
(8)Carry out the sampling inspections by following the switch
rules in Figure 5-13.
83

84

Table 5-5
5 5 Table for reduced inspection

T bl 5-4
Table
5 4 Table
T bl for
f tightened
i h
d inspection
i
i

85

86