You are on page 1of 10

C A PA B I L I T Y S T U D I E S

How To Perform
A Machine Capability Study
F
ew topics have suffered more debate and misunderstanding in industry
than machine capability studies.
With no known standards for analysis, capability study methods vary
PLAN, EXECUTE, greatly from company to company and even plant to plant within the same cor-
poration. Automotive standards, such as QS-9000, ISO/TS 16949 and VDA (the
ANALYZE AND German Automobile Industry Association), provide us requirements for capa-
AVOID PITFALLS. bility studies and facilitate limited understanding, but they fall short in the how
to category.
Nevertheless, every Green Belt and Black Belt should know how to perform a
machine capability study and thoroughly understand several pitfalls associated
with it.

Definition

The best place to start with machine capability is to define it. In recent time,
a number of definitions have evolved:
• Inherent variation of a machine.
By Gregor y • Calculation of Cp and Cpk.
• Assessment of variations in a manufacturing process compared to a given
Roth II tolerance.
• A measure of the inherent ability of a production machine to reproduce its
results with accuracy and precision.
• A measure of performance.
• Analysis of the random variability in a process.
• Inherent reproducibility of parts produced by the process.
• Machine capability equaling 6 sigma.
• Inherent machine capability.
None of these completely defines machine capability. Determination of the
inherent variation of a machine is the most popular definition, but while this is
an excellent theoretical definition, it is rarely achieved in industry.
The preceding may seem like a harsh statement, but examine the facts. Data
from a machine (current running condition) is collected and analyzed, and the
results are published. If the results of the analysis are unacceptable, something
is done to make the machine better. If the results of the analysis are acceptable,
no action is taken on the machine.
The reality of the situation is acceptable machine performance can be
improved: A worn feed arm or bearing can be replaced. Either action can
improve the capability of the machine.
Yet, the inherent variation of the machine still has not been determined. To
determine the inherent variation, as much play (variation) as possible must be
removed from every moving part in the machine, and the optimum raw mate-

22 I N O V E M B E R 2 0 0 3 I W W W . A S Q . O R G
H o w To P e r f o r m a M a c h i n e C a p a b i l i t y S t u d y

rial properties for input must be found. A new, ues .501, .500, .497, .503, .496, .504, .501, .502 and
debugged machine running for the first time is about .499 are a sample from a machine. This data yields an
as close to the inherent variation of the machine as average of .5003. If .504 (the next data point) is added
you can get. to the data set, the average changes to .5007. The aver-
No one in industry determines the inherent varia- age has now shifted .0004 with the addition of one
tion of older, existing equipment. It is not cost effec- data point. This .0004 shift can change a Cpk value
tive or necessary to producing acceptable product. An from 1.96 to 1.18. (Cpk = .753 - .752 divided by .00051
operating level (variation output) of the machine that = 1.96. Shifting the average .0004 yields .753 - .7524
is satisfactory is found by meeting expectations for divided by .00051 = 1.18.)
quality, quantity, cost, cycle time and maintenance The sample average may be adjusted by machine set-
cost. For academic purposes, let’s call this “acceptable ting, but the population average is fixed. If you study a
operational level machine spread.” given machine over a long period (several different
Most companies or plants are comfortable with this studies), the results will show the sample average is
study method, but it does not determine the inherent constantly moving. But remarkably, the variation (stan-
variation of a machine. This level of determination dard deviation) remains reasonably constant.
provides optimum production output, maintenance
cost and quality of product by studying various speeds, Steps for Machine Capability Study
feeds and other aspects to get the optimum settings to
maximize yield. This machine spread level represents Based on the previous information, it is important
the delicate balance between output and cost factors. that planning is completed before embarking upon a
A machine capability study is performed to answer machine capability study. The following describes the
questions about the machine. The study can be cate- seven basic steps necessary to perform a capability study:
gorized as capability or diagnostic in origin. The Step one—planning. This involves the selection of
analysis of the data will provide some insight into the variables to be studied and requires asking several
operation of the machine. Usually, it will determine questions:
the ability of the machine to hold a given tolerance by • Is the study to determine capability or for diag-
comparing the 6-sigma spread (commonly called nostic reasons?
process spread) to the tolerance spread.
• For what specific question is the study being per-
Other helpful information includes adjustments for
formed?
manufacturing and location of a machine or tool head
with respect to part station. Additional considerations • Is there an essential dimension at the heart of the
may be prediction of percentage out of tolerance for problem?
scrap or rework, cost estimations, interrelationship of • Is there a key dimension that will cover machine
sequential processing steps and determination of the guidance or movement?
location of the sample average. Any one or all of these
• Is the study to predict percentage out of toler-
may be needed to direct corrective actions for a
ance?
machine not performing up to expectations.
The next consideration is sampling the population to • Will more than one dimension be required?
determine process parameters. In theory, if all the meas- • Is the sample average needed from multiple cavi-
urements from a population (machine ran until it could ties or stations?
not produce any more parts) were collected, the popu-
• From which cavity, spindle, fixture or machine
lation average and standard deviation could be calculat-
are data required?
ed. This is one number, and it would never change.
In reality, not all the measurements from the popu- Step two—measuring. This entails selecting the
lation are ever collected because machines are retired measuring instrument. The gage should follow the 10
from service before running the entire population. to one rule for product measurement to minimize
Every time a sample is taken from the population, a measurement error. If part tolerance is .XXX, then
different estimate of the population (process) average gage increment should be .XXXX. Another way of
and the population standard deviation is obtained. stating this is, if the part tolerance is .02, then the
In fact, the estimated average changes slightly with smallest gage increment is at least .002.
each consecutive part produced; thus, the estimated The gage on the production floor may or may not be
average is a dynamic entity. For example, the data val- adequate for the study. Ensure the gage is calibrated

S I X S I G M A F O R U M M A G A Z I N E I N O V E M B E R 2 0 0 3 I 23
H o w To P e r f o r m a M a c h i n e C a p a b i l i t y S t u d y

and functioning correctly before any measurements example. from shop floor statistical process control
are made for the study. Don’t forget to verify calibra- (SPC)—will provide long-term data. This will certain-
tion after the study is accomplished to ensure the gage ly demonstrate what the machine is sending to the
did not drift during use. next operation over time. It also suggests the question,
Step three—personnel. This involves selection of can the customers live with this output?
personnel to operate the machine and measure the Step five—approval. You must obtain approval from
parts. Ideally, use the person who normally does these the area (machine) supervisor before the study is
things. If that is not possible, use personnel who are started. Beyond courtesy, the supervisor may have
knowledgeable in the operation to be studied. Don’t good rationale on why the study should not be per-
use personnel who are temporary or do not have in- formed at this time or provide information that might
depth knowledge of how to operate the required need to be included in the study.
machine. No one wants to be in the embarrassing position of
Step four—sample size. This covers determination reporting results of the study to management only to
of the sample size and selection of the samples. have the area supervisor discredit the results. For
Factors for consideration are: example, the coolant chiller was being repaired at the
• Consecutive pieces—short-term results. time of the study, and thus variation was doubled.
Before data collection begins, the machine should
• Nonconsecutive pieces over time—longer-term be set up properly (in other words, to the center one-
results. third of the control limits) for capability study and the
• Process drift. setup approved by the area supervisor.
• Tool life. Step six—recording. This involves collection of the
raw data in an unbiased manner. Order of manufac-
• Operator shift changes. ture must be preserved for analysis. Odd readings can-
• Material (batch) changes. not be discarded from the study results without solid
• Supplier changes in incoming material (both statistical rationale. Just because something doesn’t fit
internal and external suppliers). with the rest of the data is not sufficient rationale to
discard it. In addition, machine settings should be
• Environmental changes—for example, ambient recorded along with the raw data:
temperature, humidity, pollution from neighbor-
• Feed rate.
ing plant, and day and night operation.
• Speed.
• Time for unforeseen circumstances to affect the
process (Murphy’s law). • Cycle times.
Which is the correct method to use? Surprisingly, • Head pressure.
the correct answer is not in any book or a technical • Coolant type.
paper. The answer leads what the person performing
the study is trying to learn: • Coolant temperature.
• Are you looking for a short-term or long-term • Tool changes.
answer? • Tool type.
• Is the capability study performed for diagnostic • Ambient temperature.
reasons or to determine capability? • Barometric pressure.
• How much risk in a wrong answer is acceptable to • Relative humidity.
you?
• Machine location.
The answers to these questions drive the timing and
quantity of samples required for the capability study. • Others factors as required by a responsible au-
Fifty consecutive pieces will provide useful short-term thority such as a quality management or a manu-
data (this is a suggested minimum quantity for manu- facturing standard.
facturing factories). This represents the best the Product configuration should also be recorded by
machine can produce at a point in time. part number, revision level, configuration, product
Pieces taken over time (for example, twice a shift, supplier, shift, time and operator. This recorded infor-
each shift for two weeks) will provide long-term data. mation will be helpful in the future for making realis-
Historical data covering several months or years—for tic comparisons between old and new studies.

24 I N O V E M B E R 2 0 0 3 I W W W . A S Q . O R G
H o w To P e r f o r m a M a c h i n e C a p a b i l i t y S t u d y

Table 1. Data Collection Figure 1. Histogram

2.8347 2.8365 2.8359 2.8350 2.8377 LSL Study USL

2.8345 2.8350 2.8350 2.8363 2.8330


2.8345 2.8350 2.8332 2.8340 2.8337
2.8355 2.8355 2.8340 2.8342 2.8377
2.8332 2.8354 2.8355 2.8339 2.8338
2.8340 2.8350 2.8350 2.8340 2.8353
2.8357 2.8350 2.8360 2.8370 2.8352
2.8356 2.8348 2.8320 2.8335 2.8345
2.8337 2.8352 2.8340 2.8345 2.8355
2.832 2.833 2.834 2.835 2.836 2.837 2.838
2.8344 2.8332 2.8339 2.8330 2.8345
LSL=lower specification limit
Note: Numbers in bold are the minimum and maximum
USL=upper specification limit

Step seven—analysis of the data. Analysis is divided The dotted line is a computer interpretation of how
into three different parts: the raw data actually appears. The histogram itself is
in bars.
1. Histogram. The width of a bar, therefore, is equivalent to one
2. X bar and R chart. dot on the black line. Because of this, the dot and the
bar will never look exactly alike, but it does offer an
3. Descriptive statistics. approximation of a distribution curve.
Results of data collection from an actual machining A histogram provides the shape of the curve indi-
capability study will serve as an example. Data is shown cating whether the data appears normal. Some manu-
in Table 1. facturing processes do not yield a normal (or
Analysis can begin with just looking at the raw data. Gaussian) distribution. The correct type of distribu-
The difference between the maximum (2.8377) and tion should be determined for each machine study.
minimum (2.8320) yields a range (.0057), which is a A skewed distribution, shown in Figure 2, can be
measure of variation. If the range exceeds the tolerance readily recognized from a normal histogram. If the
spread (upper tolerance minus lower tolerance), it is distribution is skewed but the data is taken from a
obvious the machine is not capable of meeting tolerance. process that is known to produce a normal curve, this
The next step is to draw a histogram to provide a may suggest a bias in sampling, and the data gathering
picture of the process (computers are helpful with technique should be reviewed.
this), as shown in Figure 1. The solid line on the his- If the mean (arithmetic average), mode (most fre-
togram shows how a normal process should appear. quent value) and median (data midpoint) are close in
value, this indicates the data
Figure 2. Skew One is normal. As these values
move away from one another,
Comparison of distributions Comparison of distributions it suggests a skewed or non-
300 normal distribution of some
300
type. It needs to be clearly
200 understood that tools
Frequency

Frequency

200
designed for use on a normal
curve will not provide correct
100 100 answers on nonnormal data.
The next tool to use is an
0
Negative skew 0 Positive skew X bar and R chart (see Fig-
0 10 20 30 40 50 60 70 80 60 70 80 90 100 110 120 130 ure 3, p. 26). It will answer
C3 C2 whether the process is in sta-
tistical control.

S I X S I G M A F O R U M M A G A Z I N E I N O V E M B E R 2 0 0 3 I 25
H o w To P e r f o r m a M a c h i n e C a p a b i l i t y S t u d y

This chart indicates the process Table 2. Calculation of Descriptive Statistics


is in statistical control. That
means no plots (or points) are Part tolerance Raw data Statistical analysis
outside the control limits, plots
Upper tolerance = 2.839 Maximum = 2.8377 Average + 3 sigma = 2.83825
are evenly distributed about the
centerline, and there are no runs, Nominal = 2.835 Mode = 2.8350 Mean = 2.83474
trends or shifts. Lower tolerance = 2.831 Minimum = 2.8320 Average – 3 sigma = 2.83124
It is not well-publicized that for Tolerance spread =.008 Range = 0057 6-sigma spread = 00701
an X bar and R chart to be effec-
tive, the data must be plotted in Note: Each line is to be read horizontally across and each parameter is to be com-
pared to the other parameters on that line.
the order of manufacture. Control
limits are statistical limits and can-
not be interchanged with tolerance limits, which are set the process in control and repeat the seven steps.
by engineering. Control limits represent an average of Again, remember that statistical tools designed for use
a subgroup size, while tolerance limits are individual on a normal curve will not provide correct answers on
data points. Thus, they should not be shown on the nonnormal data.
same graph. The next part is the calculation of descriptive statis-
A process in control means all the variation is the tics. The following method may be new to you, but it will
result of the laws of chance—the natural variation of provide a logical approach to analysis. Using Table 2 will
the process. This is referred to as a stable or in control. facilitate analysis and ensure all aspects are covered.
A process not in control has variation from the laws of Starting with the top line of Table 2, the upper tol-
chance plus variation from an assignable cause. This is erance (2.839) is compared to maximum of actual
referred to as an unstable process or out of control. data (2.837) and then to the theoretical (average + 3
When a process is out of control and the assignable sigma) maximum (2.83825) of the data. (Note the
cause is repaired or removed as appropriate, standard maximum in each case is a logical approach.) If the
deviation of the machine or process will become raw data is above the upper tolerance limit, there is
smaller. This is one of the few absolutes in statistics. always nonconforming product (ideally, data will be
If the process is not in statistical control, stop the less than the tolerance limit). When raw data is within
analysis at this point, take the steps necessary to bring tolerance and the average plus 3 sigma (theoretical

Figure 3. Xbar/R Chart for study

2.837
3.0 SL = 2.836
Sample mean

2.836

2.835 X = 2.35

2.834
-3.0 SL = 2.833
2.833

0 1 2 3 4 5 6 7 8 9 10

0.006
3.0 SL = 0.005540
0.005
Sample range

0.004
0.003
R = 0.002620
0.002
0.001
0.000 -3.0 SL = 0.000

Note: SL = statistical limit

26 I N O V E M B E R 2 0 0 3 I W W W . A S Q . O R G
H o w To P e r f o r m a M a c h i n e C a p a b i l i t y S t u d y

Figure 4. Curve One Figure 5. Curve Two

Lower tolerance limit


Upper tolerance limit

Area out of tolerance Area out of tolerance

2.839 2.831

Nominal = 2.835 Mode = 2.8350 Mean = 2.83474 Tolerance spread =.008 Range = 0057 6-sigma spread =.00701

maximum) is above the upper tolerance, noncon- process. If the range is larger than the tolerance
forming product should be expected at some time spread, the process is not capable of meeting engi-
during the production run. neering tolerances, and appropriate corrective action
If the process is producing nonconforming product, steps are needed or nonconforming product should
you can utilize the Z score statistical tool and a table of be expected at some time during the production run.
area under a normal curve to determine the extent of The same holds true if the 6-sigma spread (.00702)
the out of tolerance conditioning. Figure 4 shows an is larger than the tolerance spread .(008). In this
example of a process exceeding the upper tolerance example, 6 sigma is not larger. Ideally, the 6-sigma
limit. spread (process spread) should be 75% or less than
The second line of Table 2, comparing the nominal the tolerance spread (.008). At this point, the
to the mode (most frequent value), yields an indica- required information to calculate Cp and Cpk is avail-
tion of how well the process is centered. Comparison able.
of the mean to the nominal shows exactly how far the This example yields a Cp (1.14) and Cpk (1.03).
process is off center. The difference between the two Remember, Cpk calculates for the position of the mean
numbers (.00026) is precisely how far the process in the tolerance zone, not process centering. This
needs to be adjusted to bring it back to the nominal of example is off center by .00026 as discussed in the
the tolerance. nominal line.
It is possible to get a different perspective by substi- Process centering and process variations are two sep-
tuting the median for the mode for comparison. If the arate and distinct process characteristics. As such, they
mean, mode and median are close together in value, it must be dealt with individually. Would your company
is an indication the data is producing a normal curve. run a machine with a Cp of 1.13 if correcting it meant
The third line of Table 2 addresses the opposite end serious dollars or lost schedule? Is the control system
of the process curve from the first line but is read in (operator) sufficient to keep the process adequately
the same manner. If the raw data is below the lower centered? Obviously, this is a personal decision.
tolerance limit (2.831), there is always nonconform-
ing material. When raw data is within tolerance and A Snapshot in Time
the average minus 3 sigma (2.83124) is below the
lower tolerance, nonconforming product should be A machine spread study is equivalent to taking a pic-
expected at some time during the production run. ture of a car race. It is a snapshot in time, and the
If the process is producing nonconforming product, process continues to move with time just as the car
you can utilize the Z score statistical tool and a table of race continues after a picture is taken.
area under a normal curve to determine the extent of The snapshot in time yields valuable information
the out of tolerance conditioning. Figure 5 shows an about the process but is only as current as when the
example of the process exceeding the lower tolerance data was gathered. By the time the data is evaluated, it
limit. is in the past when compared to the process time.
The last line of Table 2 provides a summary of the The problem is how to use this information to

S I X S I G M A F O R U M M A G A Z I N E I N O V E M B E R 2 0 0 3 I 27
H o w To P e r f o r m a M a c h i n e C a p a b i l i t y S t u d y

realistically predict the future. Table 3 will demon- can be rendered useless by measurement error.
strate one method of process prediction. The 50 5. The user must change hats and look at the
pieces from Table 1 represent short-term data. This process from the process point of view.
can be utilized to estimate long-term data and predict
the future. 6. People believe many major myths associated with
Refer back to the second line in Table 2 (where the machine capability studies.
mean was calculated) and substitute the confidence 1. They provide short term vs. long term analysis. A
interval for the mean (95% in this example). Then machine capability study is usually performed with
calculate the average plus and minus three standard short-term data. Sometimes it is necessary to review
deviations from the confidence interval limits to long-term data (as with SPC data).
achieve the results shown in Table 3. Compare these In this case, the analysis approach needs to change.
numbers to the tolerances. Long-term data consists of many short-term averages
This will demonstrate whether the process is truly added together to yield one long-term overall average.
capable of meeting specifications over time because Depending on the user’s circumstances, an overall
the predicted movement of the sample averages has average may not be what is required to diagnose a
been included. Read Table 3 in the same manner as problem with the process. Understanding the range
Table 2 was read. If out of tolerance conditions are through which the process average (short-term aver-
predicted, the percentage out of tolerance may be cal- ages) migrates may be more beneficial to problem
culated as a Z score solving or process control.
Unfortunately, traditional SPC methods do not pro-
vide the information required for modern, high vol-
ume production lines.1 Control charts illustrate how
and a table for area under a normal curve. This the process average migrates, but they calculate stan-
method with the example in Table 3 yields a Cp of 1.03 dard deviation for each subgroup.
and Cpk of 0.87. Over time, the process is not capable After years of study, I have found standard deviation
of meeting a goal of 1.33 Cp/Cpk, and nonconforming for a specific process remains reasonably constant
product (and possible scrap and rework) will occur over time, but the process average migrates in a range
sometime in the production run. around some central point (central limit theorem).
If you chose to run the process using a machine The migration of the average inflates the apparent
spread study as indicated, nonconforming product value of standard deviation for calculation of long-
would be produced. Could your company (your cus- term data as shown in Figure 6.
tomer) tolerate occasional nonconforming product? When percentage out of tolerance is calculated, it is
inflated because sigma is inflated due to continual
Pitfalls to Capability Studies sample average shifting. Figure 6 shows the average
shifting, moving the whole curve out of tolerance for
There are six pitfalls to capability studies: a brief period. This calculation will yield a smaller per-
1. They provide short-term vs. long-term analysis. centage out of tolerance than will long-term data.
2. Process potential software can be misleading. The irony of this approach is the customer is
shipped the accumulation of all the curves including
3. The formula used to calculate standard deviation
any product that is not conforming to specification
could change the results of a capability study.
(assuming it is not sorted out). A short-term study will
4. Good work on a correct machine capability study show the process capable of meeting specifications, but
the process will generate noncon-
Table 3. Process Prediction Results formance in the long run if the
process average is not carefully
Part tolerance Raw data Statistical analysis
controlled.
(with confidence inter val data added) Bottom line, for the same basic
Upper tolerance = 2.839 Maximum = 2.8377 Average + 3 sigma = 2.8386
calculations required to determine
capability indices, you can com-
Nominal = 2.835 Mode = 2.8350 Mean = 2.83474/2.8350
plete the information needed for
Lower tolerance = 2.831 Minimum = 2.8320 Average – 3 sigma = 2.8308 Table 2 or Table 3 and gain more
Tolerance spread =.008 Range = 0057 6-sigma spread = .0078 knowledge about the process.

28 I N O V E M B E R 2 0 0 3 I W W W . A S Q . O R G
H o w To P e r f o r m a M a c h i n e C a p a b i l i t y S t u d y

Capability indices are tricky in correct interpreta- mula or method for calculation of standard deviation
tion and controversial in their application. The user to bring consistency to its organization.
must clearly understand the use of each statistical tool It is accepted by industry that the sum of the squares
to obtain correct analysis. method provides consistent results. Before the advent
2. Process potential software can be misleading. of computers, the shortcut (R bar divided by d2
Many statistical software companies are now providing method) calculation for standard deviation was not
process potential as part of their analysis of data. only needed, but also welcomed with open arms
Unfortunately, this software misleads the user since it because of the simplicity of calculation. Today, howev-
fails to warn that its computer analysis displays only er, almost everyone has access to a computer. It seems
what the process is capable of performing when all a bit illogical to have a computer calculate standard
problems are corrected—not what the process per- deviation using any shortcut method. However, the cal-
formance actually is. culation method remains a personal choice of the user.
Actually, the software shows all assignable causes Formulas are available to calculate long-term standard
removed or repaired, which produces a standard devia- deviation. They tend to fall into two categories of
tion smaller than the actual process standard deviation. error: 1. Calculated standard deviation is smaller than
The result is the user mistakenly thinks the process the true standard deviation of the process, which
is capable of meeting the specification and is puzzled masks the true percentage out of tolerance.
as to why all the scrap and rework is being generated. 2. Most processes do not conform to the constant
Figure 7 is an example. In it, data from an actual used in the formula. For example, in the formula
machining process provides a view of misrepresenta-
tion of process potential.
Some software packages are taking the data strings the d2 value is a constant. Constants in formulas are
from A, B and C in the figure, forcing a common aver- estimators, not predictors of the process. It is difficult
age between them and then calculating standard devi-
ation from the forced average. If A, B, and C contain
the same average, the computer estimated standard Figure 6. Migration of Average
deviation will be as small as calculated.
However, reality is A, B, and C do not share a com-
mon average, and the customer of this machine is see-
ing the full extent of the variation, including sample
average shifts and any associated nonconformance.
While Figure 7 is extreme, it is visually apparent some Short-term capability
software will show a process capable of meeting toler-
ance while it is actually producing parts out of toler-
ance. This is a misleading situation.
3. The formula used to calculate standard deviation
could change the results of a capability study.
Inconsistencies are not limited to the brand of soft- Long-term capability
ware used. Users may utilize any number of methods
for computing standard deviation, such as but not lim-
Figure 7. Process Standard Deviation Example
ited to the following:
A
1.1
Predicted variation
Sample mean

1.0 C Actual
variation

0.9 B
This fact alone will cause variation in indices and
add to the confusion and misunderstanding. It would 0 5 10 15
be prudent for management to standardize one for- Sample number

S I X S I G M A F O R U M M A G A Z I N E I N O V E M B E R 2 0 0 3 I 29
H o w To P e r f o r m a M a c h i n e C a p a b i l i t y S t u d y

to imagine a process that follows a constant. A num- can be rendered useless by measurement error, as
ber of different methods are in practice today for determined by a gage reproducibility and repeatabili-
process prediction. ty (R&R) study or a gage drifting out of calibration.
One method is to calculate capability by subtracting Data is only as good as the measurement error associ-
the additional mean variation (calculated separately) ated with it. Thus, gage R&R studies are an integral
from the 6-sigma spread. This effectively yields short- part of performing a good capability study.
term data. Another method arbitrarily shifts the aver- If the effect the measurement error has had on the
age 1.5 sigma, then calculates a 3-sigma limit from this data (and process predictions) has not been evaluated,
shifted point. The problem with this approach is many take the following quote to heart: “It should be remem-
of today’s processes contain less than a 1.5 sigma shift. bered that the greatest designs of mankind are doomed
Perhaps a better solution to this problem may be to to failure if the component pieces cannot be measured
calculate a confidence interval for the mean and calcu- adequately and the measurement is always accom-
late ±3 standard deviation from the confidence interval plished by utilizing a sophisticated rubber ruler to esti-
(as shown earlier). This method will yield a better pre- mate dimensions.”2 I have seen measurement error (via
diction of the out of tolerance conditions and a realis- gage R&R) to be much as 150% of the part tolerance.
tic assessment of just where the process average will 5. The user must change hats and look at the
actually migrate. Another feature of this method is all process from the process point of view. It is important
the calculations utilize the actual current process sam- to remember that it is the machine itself that one is
ple data (data driven) and not a developed constant to attempting to understand.
fit all processes. Most of us view the machine as a nonentity. If that
All computer software uses different algorithms to thinking is reversed, a great deal of understanding can
calculate standard deviation. To illustrate this, one transpire. Think of the machine as a lower life form
standard set of data was selected and different soft- with limited intelligence. The machine only under-
ware packages were utilized to calculate standard devi- stands natural variation (laws of chance) and assigna-
ation from the same standard data set. Cp was then cal- ble causes. Since these just happen to be the only two
culated from the standard deviations and developed things that prevent us from producing a perfect part
into Table 4 (calculations and terminology in the table every time, it behooves us to learn from the machine.
are copied directly from the individual software pack- A machine does not understand:
ages). • Part tolerances.
This clearly shows different types of software calcu-
• Engineering changes.
late standard deviation differently, resulting in differ-
ent Cp index calculations. The standard deviation • People expectations.
from the table is also used to calculate Cp (tolerance • Latest ballgame scores.
spread held constant at .010) to demonstrate how an
• Company politics.
index will vary depending on the sigma calculation
method used. • Capability indices.
4. Good work on a correct machine capability study • What you want it to do.
At first, the preceding may seem arrogant; however,
none of these things is part of the natural world of the
Table 4. Sample Calculations
machine. Statistical tools are a way to ask the machine
Of Standard Deviations
a question and obtain an answer. If you ask the wrong
Method Standard Deviation Cp question (use the wrong statistical tool), you get the
Longhand .001167792 1.43 wrong answer. The user must recognize when this is
Rbar/d2 .001143 1.46
occurring.
The machine will react to a stimulus. A big hammer
Software company one against the side of a machine yields bad parts. Better
Descriptive statistics .0012 1.39 stated, change any variable (such as coolant, tempera-
Long term .0011738 1.42 ture or material hardness) and the machine will react
Six pack .0013143 1.27
to the change. The analysis of a machine may be sim-
plified by keeping one thought in mind: Use the right
Software company two .001168 1.43
tool to ask the right question, and you will get the
right answer.

30 I N O V E M B E R 2 0 0 3 I W W W . A S Q . O R G
H o w To P e r f o r m a M a c h i n e C a p a b i l i t y S t u d y

6. People believe many major myths associated with falls is to be aware of any errors discussed in this arti-
machine capability studies, so a discussion of them cle and to demonstrate consistency in the perform-
would not be complete without listing the major myths: ance of all studies using standardized methods and
• All the parts are well within tolerance. formulas.
• The process is perfectly centered.
REFERENCES
• The process needs to be improved.
1. Thomas A. Steward, “Reality-Based SPC,” Quality In Manufacturing,
• Capital is necessary to improve the process. March-April, 2000, p. 16.

• This is as good as the process can do. 2. Gregory Roth II, “Why Is My Measurement Different Than Yours?”
Quality Engineering, Vol. 5, No. 1.
• Measurement of the performance potential of
the process is needed.
BIBLIOGRAPHY
The problem with these myths is many people seri-
ously believe them, when the exact opposite can be Robert Bosch Corp., Technical Statistics, Machine and Process Capability of
Machining Facilities, booklet No. 9, second edition, July 1991.
true. One should not allow myths to stop the per-
formance of a capability study to obtain the truth
about a process.
Unfortunately, many of these pitfalls have become a WHAT DO YOU THINK OF THIS ARTICLE? Please share
way of life in many companies. Bad workplace habits your comments and thoughts with the editor by e-mailing
that are part of the company culture are hard to change.
The best advice to follow in overcoming these pit- godfrey@asq.edu.

S I X S I G M A F O R U M M A G A Z I N E I N O V E M B E R 2 0 0 3 I 31

You might also like