Documente Academic
Documente Profesional
Documente Cultură
Control charts have two general uses in an improvement project. The most common application is as a tool to monitor process
stability and control. A less common, although some might argue more powerful, use of control charts is as an analysis tool. The
descriptions below provide an overview of the different types of control charts to help practitioners identify the best chart for any
monitoring situation, followed by a description of the method for using control charts for analysis.
Identifying Variation
When a process is stable and in control, it displays common cause variation, variation that is inherent to the process. A process is in
control when based on past experience it can be predicted how the process will vary (within limits) in the future. If the process is
unstable, the process displays special cause variation, non-random variation from external factors.
Control charts are simple, robust tools for understanding process variability.
Every process falls into one of these states at any given time, but will not remain in that state. All processes will migrate toward the
state of chaos. Companies typically begin some type of improvement effort when a process reaches the state of chaos (although
arguably they would be better served to initiate improvement plans at the brink of chaos or threshold state). Control charts are robust
and effective tools to use as part of the strategy used to detect this natural process degradation (Figure 2). 3
2.
A central line (X) is added as a visual reference for detecting shifts or trends this is also referred to as the process
location.
3.
Upper and lower control limits (UCL and LCL) are computed from available data and placed equidistant from the central
line. This is also referred to as process dispersion.
Control limits (CLs) ensure time is not wasted looking for unnecessary trouble the goal of any process improvement practitioner
should be to only take action when warranted. Control limits are calculated by:
1.
2.
3.
Adding (3 x ? to the average) for the UCL and subtracting (3 x ? from the average) for the LCL
Mathematically, the calculation of control limits looks like:
(Note: The hat over the sigma symbol indicates that this is an estimate of standard deviation, not the true population standard
deviation.)
Because control limits are calculated from process data, they are independent of customer expectations or specification limits.
Control rules take advantage of the normal curve in which 68.26 percent of all data is within plus or minus one standard deviation
from the average, 95.44 percent of all data is within plus or minus two standard deviations from the average, and 99.73 percent of
data will be within plus or minus three standard deviations from the average. As such, data should be normally distributed (or
transformed) when using control charts, or the chart may signal an unexpectedly high rate of false alarms.
Controlled Variation
Controlled variation is characterized by a stable and consistent pattern of variation over time, and is associated with common
causes. A process operating with controlled variation has an outcome that is predictable within the bounds of the control limits.
Uncontrolled Variation
Uncontrolled variation is characterized by variation that changes over time and is associated with special causes. The outcomes of
this process are unpredictable; a customer may be satisfied or unsatisfied given this unpredictability.
Please note: process control and process capability are two different things. A process should be stable and in control before
process capability is assessed.
2.
3.
4.
Xbar-Range Charts
Another commonly used control chart for continuous data is the Xbar and range (Xbar-R) chart (Figure 8). Like the I-MR chart, it is
comprised of two charts used in tandem. The Xbar-R chart is used when you can rationally collect measurements in subgroups of
between two and 10 observations. Each subgroup is a snapshot of the process at a given point in time. The charts x-axes are time
based, so that the chart shows a history of the process. For this reason, it is important that the data is in time-order.
The Xbar chart is used to evaluate consistency of process averages by plotting the average of each subgroup. It is efficient at
detecting relatively large shifts (typically plus or minus 1.5 ? or larger) in the process average.
The R chart, on the other hand, plot the ranges of each subgroup. The R chart is used to evaluate the consistency of process
variation. Look at the R chart first; if the R chart is out of control, then the control limits on the Xbar chart are meaningless.
Table 1 shows the formulas for calculating control limits. Many software packages do these calculations without much user effort.
(Note: For an I-MR chart, use a sample size, n, of 2.) Notice that the control limits are a function of the average range (Rbar). This
is the technical reason why the R chart needs to be in control before further analysis. If the range is unstable, the control limits will
be inflated, which could cause an errant analysis and subsequent work in the wrong area of the process.
d2
D3
D4
1.128
3.268
1.693
2.574
2.059
2.282
2.326
2.114
2.534
2.004
2.704
0.076
1.924
2.847
0.136
1.864
2.970
0.184
1.816
10
3.078
0.223
1.777
11
3.173
0.256
1.744
12
3.258
0.283
1.717
13
3.336
0.307
1.693
14
3.407
0.328
1.672
15
3.472
0.347
1.653
Can these constants be calculated? Yes, based on d2, where d2 is a control chart constant that depends on subgroup size.
The I-MR and Xbar-R charts use the relationship of Rbar/d2 as the estimate for standard deviation. For sample sizes less than 10,
that estimate is more accurate than the sum of squares estimate. The constant, d2, is dependent on sample size. For this reason
most software packages automatically change from Xbar-R to Xbar-S charts around sample sizes of 10. The difference between
these two charts is simply the estimate of standard deviation.
u-Chart
Similar to a c-chart, the u-chart is used to track the total count of defects per unit (u) that occur during the sampling period and can
track a sample having more than one defect. However, unlike a c-chart, a u-chart is used when the number of samples of each
sampling period may vary significantly.
np-Chart
Use an np-chart when identifying the total count of defective units (the unit may have one or more defects) with a constant sampling
size.
p-Chart
Used when each unit can be considered pass or fail no matter the number of defects a p-chart shows the number of tracked
failures (np) divided by the number of total units (n).
Notice that no discrete control charts have corresponding range charts as with the variable charts. The standard deviation is
estimated from the parameter itself (p, u or c); therefore, a range is not required.
A number of points may be taken into consideration when identifying the type of control chart to use, such as:
Variables control charts (those that measure variation on a continuous scale) are more sensitive to change than attribute
control charts (those that measure variation on a discrete scale).
Variables charts are useful for processes such as measuring tool wear.
Use an individuals chart when few measurements are available (e.g., when they are infrequent or are particularly costly).
These charts should be used when the natural subgroup is not yet known.
In a u-chart, the defects within the unit must be independent of one another, such as with component failures on a printed
circuit board or the number of defects on a billing statement.
Use a u-chart for continuous items, such as fabric (e.g., defects per square meter of cloth).
A c-chart is a useful alternative to a u-chart when there are a lot of possible defects on a unit, but there is only a small
chance of any one defect occurring (e.g., flaws in a roll of material).
When charting proportions, p and np-charts are useful (e.g., compliance rates or process yields).
The R chart displays change in the within subgroup dispersion of the process and answers the question: Is the variation within
subgroups consistent? If the range chart is out of control, the system is not stable. It tells you that you need to look for the source of
the instability, such as poor measurement repeatability. Analytically it is important because the control limits in the X chart are a
function of R-bar. If the range chart is out of control then R-bar is inflated as are the control limit. This could increase the likelihood of
calling between subgroup variation within subgroup variation and send you off working on the wrong area.
Within variation is consistent when the R chart and thus the process it represents is in control. The R chart must be in control to
draw the Xbar chart.
If the Xbar chart is in control, the variation between is lower than the variation within. If the Xbar chart is not in control, the
variation between is greater than the variation within.
This is close to being a graphical analysis of variance (ANOVA). The between and within analyses provide a helpful graphical
representation while also providing the ability to assess stability that ANOVA lacks. Using this analysis along with ANOVA is a
powerful combination.
Conclusion
Knowing which control chart to use in a given situation will assure accurate monitoring of process stability. It will eliminate erroneous
results and wasted effort, focusing attention on the true opportunities for meaningful improvement.
Examples:
If we look at the quality measurements over time we may find that they may be
mostly consistent, with only occasional random variation. In other situations, the
measurements may change with a pattern that results from a fundamental flaw in
the service delivery process. Unfortunately, it can be quite difficult to tell the
difference between a flaw in the service delivery process and a mere random
variation.
For example, imagine a hotel that measures customer satisfaction with a survey
that records a rating from 1 (worst quality) to 10 (best quality). Every week the
hotel calculates average ratings for the week. The manager desires to know if
quality is consistent or changing over time. Imagine that weekly ratings for a series
of weeks are 8.8, 8.1, 8.9, 9.2, 8.2, 8.5, and 7.6. The manager is concerned about
the last weeks rating of 7.6. Does it represent a statistical anomaly, or is it simply
a result of random variation. The fact is, customers assign different meaning to
different points on the scale, and quality could have actually improved that week
but the customers were simply frugal in giving high ratings. So, one thing the
manager may want to know is if that 7.6 fits the random pattern of the other data,
or if it does not.
How
One way to investigate this is with Statistical Process Control, or SPC. SPC allows us
to make assumptions about data to help separate simple random variationalso
called natural variationfrom variation that is caused by changes in the process
called assignable variation.
The foundation of Statistical Process Control is the central limit theorem, which
states that the sum (or average) of a number of measurements from any single
given probability distribution will approximate a normal distribution. This holds true
regardless of the distribution of the individual measurements. How close the sums
(or averages) approximates the normal distribution depends on how many
measurements are in a sample, which number is called the sample size.
The
two
general
types
of
quantitative
quality
measures
are attributes and variables. An attribute is dichotomous, meaning that it takes on
the values of acceptable or unacceptable. Unacceptable represents a quality
defect. An example is a bank transaction, which is either correct or is not correct.
X-bar charts
First, we will consider variables, which are continuously scaled. In theory, a variable
can take on any value within a given range, although they are often measured at
specific intervals. The average weekly ratings at the hotel are variables, since they
can take on any value between 1 and 10.
There are two things we are concerned about with our sample measurements. One
is that the mean of the measurements is statistically incontrol. To be
in control means that the process merely contains random variation, not assignable
variation. One way to analyze the control of a process is with a control chart. To
test for statistical control of the variable mean we would use an x-bar control chart.
The term x-bar, denoted
, represents the mean of a sample of measurements.
Each measurement in a sample is an X value, i.e. X1, X2,X3, etc. If the sample size
is n, then n of the X value measurements are averaged to calculate an x-bar value.
(The bar denotes mean of values.)
An x-bar chart is a line graph with each plotted point being the x-bar value for a
sample of measurements. We also plot a line on the chart which represents the
mean of our x-bar values. The mean of x-bar values is, as you might imagine,
called x-bar-bar, denoted
(X X-bar
48
45
45
45
50
46.6
45
48
46
50
46
47.0
47
51
46
49
45
47.6
41
40
42
41
42
41.2
47
46
48
48
44
46.6
46
44
45
43
42
44.0
43
40
52
48
49
46.4
12
mean values
45.63
5.43
which is
X-barbar
R-bar
Note that the sample size is five. The sample size is NOT seven, which is the
number of samples.
The x-bar values are calculated as the mean of the measurements in each sample.
The column labeled R contains the sample range values. The range of a sample is
simply the maximum measurement in the sample minus the minimum
measurement in the sample. These R values will be used later.
We can calculate x-bar-bar to be 45.63 as the average of the x-bar values. Of
course, we would expect some of the x-bar values to be above the x-bar-bar value,
and some to be below it. If there is just natural variation, then we would expect a
random distribution above and below the mean. If we plot the data we see the
following:
We observe that the first three points are above the mean line, but that otherwise,
the data seems to be randomly dispersed above and below the mean. The
movement around the mean line in this example appears to be natural variation.
Another thing it would be good to know about the x-bar values is how dispersed
they are. In particular, we would like to know if a particular x-bar value was outside
of a reasonable range. The x-bar values are going to vary somewhat simply due to
natural variation. Unusual variation may indication that there is a special cause, or
a specific reason, for variation. That special cause may be a problem that needs to
be addressed.
Statistically, we can calculate a range of reasonable variation in the x-bar chart.
That reasonable range is bounded by control limits. Theupper control limit (UCL)
indicates the maximum value that is statistically reasonable, and the lower control
limit (LCL) indicates the minimum reasonable value. By reasonable we mean
being likely to occur given natural variation. This is different from acceptable in
terms of acceptable quality. It is up to management, employees, and customers to
decide what is acceptable in terms of quality measurements. SPC control carts
instead just look at statistical variation, to help identify if the variation follows a
usual probability distribution.
The upper and lower control limits are calculated from some equations based on the
central limit theorem, which was mentioned previously. We need to know the
standard deviation of the x-bar values, which we can calculate over time. It might
be easier to determine the standard deviation of individual measurements, which
allows us to calculate the standard deviation of sample means (x-bar values)
according to the following equation:
where the z value is the number of standard deviations (sigmas) from the mean to
put the control limits. A three-sigma control chart uses z=3, which considers
measurements within three standard deviations of the mean to be natural
variation. For the normal distribution, 99.7 percent of the distribution occurs within
three standard deviations of the mean, implying that values will be outside that
range only 0.3 percent of the time. In other words, it is quite unlikely that natural
variation will cause recurring values outside of three-sigma control limits.
A more sensitive control chart has z=2, which is called a two-sigma control chart.
Only 95.4 percent of the normal distribution occurs within two standard deviations
of the mean, implying that 4.6 percent of the time the values will be outside of that
range. This means that a two-sigma control chart is more sensitive than a threesigma control chart, meaning that it is more likely to detect when a process changes
(but also more likely to have a false alarm).
For our example above, imagine that we determine that the standard deviation of
measurements is 2.3. We calculate the standard deviation of x-bar values as 3.
For a three-sigma control chart, the LCL is 45.63-3x1.029=42.54 and the UCL is
45.63+3x1.029=48.71. These control limits can be plotted on a control chart as
follows:
In this data we conclude that something peculiar might have happened in period 4,
since the x-bar value is below the lower control limit. Keep in mind that it
is possible that the x-bar value for period 4 is a random occurrence. We need to be
careful about jumping to conclusions that something is definitely wrong with the
process. However, such a x-bar value is not very likely to occur without a special
cause. (Since only 0.3 percent of truly random values will be outside of the threesigma control limits.) Therefore, it would be good to investigate the situation
further.
If we do not have the standard deviation of measurements or sample means, we
can use the R (range) values to estimate the standard deviation by using tables that
contain control limit factors. The following is a three-sigma factor table.
1.880
3.268
1.023
2.574
0.729
2.282
0.577
2.114
0.483
2.004
0.419
0.076
1.924
0.373
0.136
1.864
0.337
0.184
1.816
10
0.308
0.233
1.777
where r-bar is calculated as the average of the sample range values, as discussed
previously. The lower control limit (LCL) for an x-bar chart is similarly calculated as
Since the factors were from a three-sigma factor table, we expect that three
standard deviations of x-bar values will be between the LCL and the UCL. The
central limit theorem shows that the x-bar values tend to follow a normal
distribution. Three standard deviations from the mean of normal distribution
includes 99.7 percent of the probability density. This means that 99.7 percent of the
time, a random number drawn from a normal distribution will be within three
standard deviations from the mean. It is quite unlikely that values would fall
outside of that range on a regular basis.
From the data above, we can calculate control limits for the example x-bar chart.
We have n=5, X-bar-bar=45.63, R-bar=5.43, and A=0.577. Therefore we have
LCL=45.630.5775.43=42.50 and UCL= 45.63+0.5775.43=48.76 . Note that
these control limits are almost the same as were calculated above using the
standard deviation of measurements.
R-charts
An x-bar chart tells us if the central tendency (i.e. mean) of the samples appears to
be in control over time. Another chart, the R-chart, tells us if the variance within
each sample tends to be in control over time. It is certainly possible for the sample
means to be in control over time, yet the sample variance is getting worse and
worse.
An R-chart is created similar to how we create an x-bar chart except for the
following:
the points which are plotted are the R, or range, values calculated above.
the central line is the R-bar value, which is the mean of the R values.
the upper control limit, LCLR , is simply the B-factor times R-bar.
the lower control limit, UCLR , is simply the C-factor times R-bar.
Again, since we are using three-sigma factors, we would expect the R values to fall
within the control limits 99.7 percent of the time.
For the data from the example above we calculate UCL R =2.1145.43=11.48, and
LCLR =05.43=0. Note that lower control limits for R-charts are bounded by zero,
since it is impossible to have R values less than zero.
We see that in period 7 the range of values appears to have gone up beyond what is
statistically common. It is interesting to note that the x-bar value for that sample
was fine, meaning that on average the measurements in the sample were in
control. However we had more variance in that sample that we might have usually
expected, so we should investigate for special causes or changes in the process.
Do we care if an R value is at or below the LCL? (Assuming we have a non-zero
LCL) It seems that less variance would be better. In fact it might be good to
investigate improvements in the process so that they can be assured to continue in
the future.
P-charts and C-charts
Next we will look at control charts for attributes, which are not continuous variables
but are things that can be counted. A p-chart considers the portion of a sample that
is defective, where each item in the sample is either defective or not. For example,
an airline might track on-time arrivals of flights on each given day. If there are 50
flights being tracked each day (the sample size), we could determine a p value as
the portion of flights that arrive late. On a given day if 10 flights arrive late then the
p value is 10/50=0.2.
The airline might plot the p value over time on a p-chart. The center line of the pchart is the average of a series of p values, which is p-bar. The control limits will be
above and below the center line the appropriate number of standard deviations (z).
A standard deviation for a p-chart is calculated according to the following equation:
where p-bar is the average of a series of p values and n is the sample size (the
number of items in each sample). The control limits are simply:
and
However, note that a p value can never be negative, implying that the LCL should
never be negative. If the LCL calculates to be negative then use LCL=0.
For example, the airline might have the following late arrival data for the past seven
days:
day
late
10
0.20
0.10
0.16
12
0.24
0.18
11
0.22
12
0.24
p-bar
0.19
It looks like the p values are well within the control limits, with no unusual patterns.
We might conclude that the late arrival process, while not ideal, is in statistical
control.
Finally, a c-chart considers attributes that can be counted, without a specific sample
size. Instead we have a sample frame which defines a range of defects to be
counted. For example, the airline might count the number of complaint letters that
come in day to day. There is no sample size since any number of letters could
arrive (and an individual might even send multiple letters).
The plot values for a c-chart are the c values, which are the counts for each sample
frame. The center line is c-bar, the average of a series of c values. We often
assume that the count values follow a Poisson distribution, which has the standard
deviation as follows:
and
For example, if the airline might count the number of complaint letters that arrive
over the past seven days as follows:
day
15
11
18
14
c-bar
10.00
This data is in statistical control. Note, however, that if we used two-sigma control
limits (z=2) then points would be outside of the control limits, suggesting the counts
of complaints is not in statistical control.
Summary of Things to Look For
What are we looking for in a control chart that might cause us to suspect the
process has changed? The following are some examples:
An unusual tendency for the sample values to be above or below the x-barbar or R-bar line. Unusual might mean five or more in a row.
An unusual tendency for the sample values to be near a control limit. Here,
unusual might mean two or more values that are very near the control
limit.
Values that appear outside of the control limits. If three-sigma control limits
are used, then it is quite unusual for even one value to appear outside of the
control limits.
Other peculiar patterns in the data, such as erratic fluctuations above and
below the mean, or patterns that repeat over a fixed number of samples.
Again, with each of these occurrences we investigate for special causes. Such
patterns of behavior can happen with mere natural variation, but they are not likely.