Sie sind auf Seite 1von 7

Ch.

7 Process Capability Analysis (PCA)

• PCA - an engineering assessment of process variability


• PCA output:
◦ a probability distribution (e.g. histogram)
Or
◦ proportions of measurements not meeting specifications. Capability indices.
• Natural Tolerance Limits

◦ These identify current process performance


◦ Assumptions:
1. Process is in control
2. Measurements are normally distributed
◦ UNTL = µ + Kσ; LNTL = µ − Kσ
◦ If K = 3, then the natural tolerance limits contain 99.73% of the measurements

• Some Questions that PCA can help answer


1. How well will the process hold tolerances?
2. How should the process be modified?
3. What sampling frequency should be used when monitoring a process?
4. If equipment is replaced, what are the performance requirements on the new
equipment?
5. Which supplier should be used?
6. What is the best order of production when there is an interactive effect of processes
on tolerances?
7. How can process variability be reduced?
• Main PCA techniques
1. Histograms or Probability Plots
2. Control Charts
3. Designed Experiments
• 7.2 PCA with a Histogram

◦ Example A. Sugar content in frosted flakes is an important quality characteristic.


◦◦ A histogram of 100 sugar content measurements is given in the link Frosted
Flakes Sugar Content under Plots.
◦◦ The measurement distribution appears to be close to normal.
◦◦ x̄ = 37.6 and s = 2.5
◦◦ Crude Estimate of Process Capability:

x̄ ± 3s = 37.6 ± 7.5

i.e. Approximately 99.73% of sugar content measurements will lie between 30.1
and 45.1%.

• 7.3 Capability Indices


◦ Measuring Potential Capability
USL − LSL
Cp =

1
◦◦ This compares the width of the specifications to the width of the interval
containing approximately 99.73% of the measurements.
◦◦ if Cp < 1.0, then the process is not potentially capable of meeting specifications.
◦◦ if Cp > 2.0, then the process is potentially capable of meeting specifications at
the 6σ level
◦◦ A potentially capable process might not be actually capable if it is not centered
on or near the target value.
◦◦ Example B.
Suppose the mean of a process is 15.7, and the process standard deviation is
4.5, when the process is in control. If USL= 38 and LSL= 3, then compute Cp
and interpret. Is the process capable?
38−3
Cp = 6(4.5) = 1.296. The process is potentially capable.
◦◦ The percentage of the specification interval used by the process is
µ ¶
1
P = 100%
Cp

e.g. P = 1/1.296 = .772 so 77.2% of the specification band is used up the process.
◦◦ Cp can be estimated from large samples, using σ b = R̄/d2 or S.
◦◦ Example A (cont’d). If sugar content is required to be between 28 and 45%, then
we estimate Cp as
bp = U SL − LSL = 1.15
C
6s
Potentially capable
◦ One-Sided Tolerances

U SL − µ
Cpu =

µ − LSL
Cpl =

◦ Example B (cont’d). Compute Cpu and Cpl and interpret.

U SL − µ
Cpu = = 1.65

The process is sufficiently far from the upper specification limit.
µ − LSL
Cpl = = .94

The process is too close to the lower specification limit. Not Capable.
◦ Measuring Actual Capability with One Index

Cpk = min {Cpu , Cpl }

◦ Example B (cont’d). Find Cpk and interpret.

Cpk = min {1.65, .94} = .94


The process is not actually capable.
Since the process is potentially capable, this index is telling us that the process is
not centered close enough to the target value.

2
◦ Example A (cont’d). Estimate Cpk for the Frosted Flakes sugar content data and
interpret.

Cpk = min {Cpu , Cpl } =

min {1.0003, 1.2964} = 1.0003


The process is capable. Since Cpk < Cp , we know that the process could be
improved by centering on the target.
◦ Non-normal measurements
◦◦ 99.73% of normal measurements are within specification limits when Cpk = 1
◦◦ Simplest approach to non-normal measurements:
Transform
Often, a log, reciprocal or square root are sufficient.
Look at a histogram of the transformed data to check adequacy.
e.g. Suppose a histogram of the measurements x1 , x2 , . . . , xn shows skewness to
the right, and a histogram of the logs of these measurements looks more
normal. An estimate of Cp for this process is

bp = log(U SL) − log(LSL)


C
6slog x

◦ An Alternative to Cpk

U SL − LSL
Cpm = p
6 σ 2 + (µ − T )2
where T is the target value.
e.g. A: USL = 45, LSL = 28, x̄ = 37.6, s = 2.47, T = 36.5
Cpm = 1.05 ≥ 1. Process is capable.
e.g. B: USL = 38, LSL = 3, µ = 15.7, σ = 4.5, T = 20.5
Cpm = .89. Process is not capable.
◦ Confidence Intervals and Tests on Cp and Cpk
bp and C
◦◦ C bpk are estimates of process capability
◦◦ Confidence intervals for the true values can be set up.
A 1 − α confidence interval for Cp is
 s s 
χ21−α/2,n−1 χ2α/2,n−1
Cbp bp
,C 
n−1 n−1

The χ2 values are tabulated in Table III.


bp = 1.15. Find a 95% confidence interval for the Cp .
◦◦ Example A. n = 100 C
α = .05.
.
χ2.975,99 = 74.22 χ2.025,99 = 129.56
so the required interval is
Ãr r !
74.22 129.56
1.15 , = (.995, 1.32)
99 99

◦◦ The 1 − α confidence interval for Cpk is given as equation 7-18 on page 369 of the
textbook.
◦◦ Exercise: Find a 95% confidence interval for Cpk for the sugar content
measurements.

3
◦◦ Hypothesis Testing.
Does Cp take on a particular value Cp0 ?
Simplest approach: construct a 1 − α confidence interval for Cp .
If Cp0 is not in the interval, we conclude (at the α level, that Cp0 is not the true
value of Cp .
e.g. For the Frosted Flakes example, we can conclude (at level α = .05) that
Cp 6= .9. We cannot conclude (at the same level) that Cp 6= 1.0.
• 7.4 PCA Using Control Charts

◦ PCA is only appropriate when process is in control


◦ An in-control process is not necessarily capable
◦ σ and µ are often estimated using control charts ⇒ capability index estimates

• 7.5 PCA Using Designed Experiments (DOE)

◦ DOE is a system for analyzing the effects of controllable inputs or factors on quality
characteristics of interest
◦ It helps to identify which factors have an effect, and at what levels these inputs
should be set at for optimal performance
◦ DOE can help isolate and estimate sources of variability in a process.
◦ Popular model
2 2 2
σtotal = σFactor 1 + · · · + σFactor m

◦ Frosted Flakes Example. Sugar content measurement variability is possibly due to


variability in the initial amount of added sugar I, milling temperature T , and the
measuring device M :
2
σtotal = σI2 + σT
2 2
+ σM
If this model holds, variability in sugar content can be reduced by reducing
variability in one or more of these factors.
◦ DOE provides a framework for setting up experimental trials to determine
estimates of the above σ’s.

• 7.6 Measuring Gauge Capability

◦ Measuring devices or gauges contribute to measurement variability:


2 2 2
σtotal = σgauge + σproduct

◦ By taking repeated measurements on the same part, one can assess gauge capability.
◦ Control Chart Method:
1. Take n measurements on each of m parts.
2. Construct an R-chart
3. Points plotting out of control represent operator difficulties and should be
removed.
4. Estimate σgauge using R̄/d2
2
5. Estimate σtotal by calculating S 2 for entire set of n × m measurements
2
6. Estimate σproduct by subtraction:
2
σ
bproduct = S 2 − (R̄/d2 )2

Gauge capability – percentage of product characteristic variability:

σgauge /σproduct × 100%

4
Estimate:

p × 100%
d22 S 2 − (R̄)2

◦ Example. Data are from Exercise 7-18. Two measurements are taken using the
same gauge on each of 20 different parts. Assess the capability of the gauge.
part measurement
no. 1 2

[1,] 19 23
[2,] 22 28
[3,] 19 24
................
[19,] 25 23
[20,] 17 16

◦◦ R̄ = 2.8 and n = 2
◦◦ R-chart: UCL = 2.8(3.267) = 9.1; LCL = 0
◦◦ All ranges plot in control. No need to revise.
◦◦ d2 = 1.128 and S 2 = 15.4
◦◦ Capability Estimate:

p × 100%
2
d2 S − (R̄)2
2

2.8
=p × 100%
1.1282 (15.4) − 2.82
= 90.2%
This is unacceptably high. (Compare with Example 7-7 which contains data
from a different gauge having capability 29.2%)

• 7.7 Setting Specification Limits on Discrete Components

◦ tolerance stack-up: when the final product fails to meet specifications, even though
individual parts meet specifications
◦ When a product is composed of many parts, the probability of failing to meet
specifications depends on the behaviour of the individual components:
◦◦ Final product: n components with quality characteristics x1 , x2 , . . . , xn ,
◦◦ Quality characteristic of final product:

y = a1 x1 + a2 x2 + · · · + an xn

where the a’s are fixed constants


◦◦ If the x’s are independent and normally distributed with means µ1 , . . . , µn , and
variance σ12 , . . . , σn2 , then y is normally distributed with mean

a1 µ1 + · · · + an µn

and variance
a21 σ12 + · · · + a2n σn2

◦ Example. Gear Set - Gear, Clip and Washer. Each component is produced
independently of the others.
◦◦ The total width of these three parts must be kept between 29 and 31 mm.

5
◦◦ Gear width: normally distributed with mean 27 and standard deviation .3
◦◦ Clip width: normally distributed with mean 2 and standard deviation .2
◦◦ Washer width: normally distributed with mean 1.1 and standard deviation .08
◦◦ ⇒ Total width W : normally distributed with mean 30.1 and standard deviation:

.09 + .04 + .0064 = .37

◦◦ Proportion of gear sets not meeting specifications:


P (W > 31) + P (W < 29) =
31 − 30.1 29 − 30.1
P (Z > ) + P (Z < )=
.37 .37
P (Z > 2.43) + P (Z < −2.97) =
.0075 + .0015 = .0090

◦ What are the specifications on gear width, clip width and washer width so that the
proportion of nonconforming gear sets is no more than .0001?
◦◦ Center the process: By decreasing mean gear width to 26.9 mm, we obtain a
mean total width of 30 mm.
◦◦ With no further changes, the fraction nonconforming would become
P (Z > 2.70) + P (Z < −2.70) = .007

◦◦ We require
P (W > 31) + P (W < 29) = .0001
or
1
P (Z > ) = .00005
σ

◦◦ Normal Table ⇒
1
= 3.90
σ
or
σ = .256 ⇒ σ 2 = .0655

◦◦ The current value of σ 2 is .09 + .04 + .0064 =.1364 so we will require the
component variances to be decreased substantially.
◦◦ Reducing washer width variance will not have much effect. The largest effect
will come from reducing gear width variance.
◦◦ There are many solutions; one possibility is
.04 + .02 + .0055 = .0655
i.e. standard deviation of gear width = .2, clip width .14, washer width .074
◦◦ If we set up the specification limits for each part so that the process limits
match exactly, then we have
U SL = 26.9 + 3(.2) = 27.5
LSL = 26.3 (gear width)
U SL = 2 + 3(.14) = 2.42
LSL = 1.58 (clip width)
U SL = 1 + 3(.074) = 1.22
LSL = 0.78 (washer width)

6
◦ General Principle: To set specification limits on components,
1. determine the variance allowable for the final product to meet specifications
with probability 1 − α:
µ ¶2
2 U SL − LSL
σ =
2Zα/2
2. determine component standard deviations σ12 , . . . , σn2 so that

a1 σ12 + · · · + an σn2 ≤ σ 2

3. Set component specification limits at process limits:

µi ± 3σi

4. Reduce component variability to achieve process capability in each case

• 7.8 Estimating Natural Tolerance Limits

◦ Natural Tolerance Limits:


µ ± Zα/2 σ
where µ = in-control mean, and σ = in-control standard deviation
◦ These limits contain (1 − α)100% of the process measurements (when the process is
in control)
◦ When Zα/2 = 3, (i.e. α = .0027)
Process Limits = Natural Tolerance Limits
These limits contain 99.73% of the process measurements when in control
◦ e.g. 2: The limits

µ ± 6σ

contain 99.9999998 % of the process measurements when in control


◦ When µ and σ are unknown, they must be estimated by x̄ and s.
◦ The limits
x̄ ± Zα/2 s
do not contain (1 − α)100% of the process measurements, even when in control,
because s 6= σ and x̄ 6= µ.
◦ Better Estimates: Replace Zα/2 by K (see Appendix Table VII):

x̄ ± Ks

◦ K depends on n and on α
◦ e.g. For a sample of size 10, take K = 4.433 in order to have 95% confidence that our
estimated natural tolerance limits will contain 99% of the measurements.
◦ Example. A sample of 12 measurements on a normally distributed quality
characteristic has a mean of 80 and a standard deviation of 7. Estimate 90%
natural tolerance limits for which you can have a confidence level of 95%.
n = 12 ⇒ K = 2.655
⇒ U N T L = 80 + 2.655(7) = 98.6
and
LN T L = 71.4
◦ What if the measurements are not normally distributed?
Try to determine the distribution of the measurements as well as possible. Use this
to determine these percentiles.
◦ A nonparametric method also exists - not practical

Das könnte Ihnen auch gefallen