Ledolter, J. and Hogg, R. V. Applied Statistics for Engineers and Physical Scientists, Third Edition Prentice Hall, 2009

80371 05 293-335 r2 sr

7/26/08

1:45 PM

Page 293

Chapter

STATISTICAL PROCESS CONTROL
5.1 Shewhart Control Charts

5

All processes have some variation. When we manufacture a product, measurements on the ﬁnal product inevitably show variation from unintentional changes to the process as well as random variation. Many different factors enter into a production process, and a change in each will cause some variation in the ﬁnal product.This variation may come from differences among machines, lot-to-lot variation, differences in suppliers and incoming raw materials, changes in the plant environment, and so on. Despite the fact that considerable effort is expended in attempting to control the variability in each of these factors, there will still be variability in the ﬁnal product. In the end, this variability has to be controlled. Statistical control charts or, more generally, statistical process control methods are procedures for monitoring process variation and for generating information on the stability of a process. It is important to check the stability of processes, because unstable processes will result in lost production, defective products, poor quality, and, in general, loss of consumer conﬁdence. For example, in the production of integrated-circuit boards, which involves several welding procedures, it may be the weld strength that is of importance. Selecting a small sample of such boards at regular intervals and measuring the weld strength by a certain pull test to destruction will provide valuable information on the stability of the welding process. In the production of concrete blocks, it is the compressive strength that is of importance and that needs to be controlled. Measurements on a small number of concrete blocks—say, twice during each production shift—can give us valuable information on the stability of the production process. In the production of thin wafers for integrated-circuit devices by high-temperature furnace oxide growth processes, it is the thickness of the wafers that needs to be controlled. Measurements on the thickness of a few selected wafers from every other furnace run can indicate whether the thickness of the product is stable. Here we have given only three examples. Many others can be found, and we encourage the reader to think of still others.

293

80371 05 293-335 r2 sr

7/26/08

1:45 PM

Page 294

294 Chapter 5 Statistical Process Control
UCL

90

86 Centerline Average x

82

LCL 78

2

4

6 Sampling period

8

10

12

Figure 5.1-1 x-chart for the sample means in Table 5.1-1 (see p. 297)

5.1-1 x-Charts and R-Charts
A control chart is a plot of a summary statistic from samples that are taken sequentially. Usually, the sample mean and a measure of the sample variability, such as the standard deviation or the range, are plotted on control charts. Figures 5.1-1 and 5.1-2 are two examples. Figure 5.1-1 shows the average compressive strengths of concrete blocks from samples of size n = 5. Twice during each shift, ﬁve concrete blocks are taken from the production line, their compressive strengths are determined, and the average is entered on the chart. Since we plot averages, we call this an x-chart 1where “x-bar” stands for average2. In Figure 5.1-2, we display the variability within the samples over time and plot the ranges from successive samples; we call such a plot an R-chart 1where R stands for range2.
20 16 Range R 12 Centerline 8 4 UCL

2

4

6 Sampling period

8

10

12

Figure 5.1-2 R-chart for the sample ranges in Table 5.1-1 (see p. 297)

80371 05 293-335 r2 sr

7/26/08

1:45 PM

Page 295

Shewhart Control Charts

295

Control charts also include bands, or control limits, that help us determine whether a particular average 1or range2 is “within acceptable limits” of random variation. Through these limits, control charts try to distinguish between variation that can normally be expected 1variation due to common causes2 and variation that is caused by unexpected changes 1variation due to special causes2. One should not tamper with the process if the measurements on these charts fall within the control limits. However, if one notices shifts in the process level on the x-chart, and if plotted averages are outside the control limits, one must conclude that something has happened to the process. Similarly, if the process variation, as measured on the R-chart, changes by more than what could be expected under usual circumstances, one must conclude that the process variability is no longer stable. In these circumstances, steps have to be taken to uncover the special causes and keep them from recurring. The theoretical explanation of the control limits in the x-chart is as follows: Let us suppose that our observations are from a stable distribution with mean m and variance s2.Then the mean X of a random sample of size n from this distribution has an approximate normal distribution with mean m and variance s2>n. If averages are plotted on a chart, the probability that a sample average X falls within the bounds m - 3s> 1n and m + 3s> 1n is 0.9974. If m and s2 are known, we call m - 3s> 1n and m + 3s> 1n the lower control limit 1LCL2 and the upper control limit 1UCL2, respectively. Thus, with a stable process, the probability of X falling between the LCL and the UCL is very large. Consequently, it is rare that a sample average X from such a stable process would fall outside the control limits. Unfortunately, we usually do not know whether our process is stable, nor do we know the values of m and s2. Thus, we begin by taking several samples, each of size n. Let us say that there are k such samples. Let x1, x2, Á , xk and s1, s2, Á , sk be the means and standard deviations of these samples. Then we can estimate s by the average of the standard deviations, namely, s = 1s1 + s2 + Á + sk2>k. Moreover, if we are reasonably satisﬁed with the overall average of the nk observations in these k p samples, we can use x = 1x1 + x2 + Á + xk2>k as the centerline in our control p chart; that is, we use x as an estimate of m. p With these estimates, the control limits around the centerline x are given by p p x - A3s and x + A3s, where the constant A3 is chosen in such a way that A3s is a good estimate of 3s> 1n. The constant A3 depends on the sample size n and is given in Table C.1. The development behind the selection of A3 is beyond the scope of this book. Note, however, that A3 in Table C.l is always slightly larger than 3> 1n; this is because we use an estimate s instead of the unknown standard deviation s. Note also that, for moderately large sample sizes, A3 and 3> 1n are quite similar; for example, for n = 5, A3 = 1.43, and 3> 1n = 1.34; for n = 10, A3 = 0.98, and 3> 1n = 0.95; for larger n, the difference disappears almost completely. However, the calculation of the standard deviations involves a considerable amount of computation. An alternative, and simpler, procedure for estimating 3s> 1n proceeds as follows: Let R1, R2, Á , Rk be the ranges of the k samples. Calculate the average range R = 1R1 + R2 + Á + Rk2>k. It can be shown that A2R is a good estimate of 3s> 1n when one is sampling from a normal distribution. The constant A2 depends on the sample size n and is also given in Table C.1. With this modip ﬁcation, the x-chart consists of a centerline at x, which is an estimate of m, and p p control limits x - A2R and x + A2R.

The constants D3 and D4 are also given in Table C. every hour. The control limits in the x-chart are given by p p LCL = x . However. xn2 . and R.1 shows that the constant D3. the sample size n is 4 or 52 at various times. In this book. we calculate the average x = g i = 1xi>n and the range R = max1x1.A2R and UCL = x + A2R. it is rare that a sample average or sample range will fall outside the control limits. provided that the process has remained under control 1which means that the level has not shifted and the variability has not changed2.Table C. k ja j =1 These quantities form the respective centerlines in the x-chart and the R-chart. p x = and the average of the ranges. from every tenth batch. D3. for a stable process. and thus also the lower control limit D3R.80371 05 293-335 r2 sr 7/26/08 1:45 PM Page 296 296 Chapter 5 Statistical Process Control There are several control charts for measures of variability. Á . R. the more stable the process. we compute the grand average 1the average of the averages2. From the k sample averages and ranges. The frequency also depends on the potential loss that is caused when deteriorations of the process are not recognized on time and. the probability of an individual R falling between the LCL and the UCL is extremely large. The constants A2. It is often recommended that k = 10 to 20 such samples be obtained before constructing the control limits. In sum. From each sample. respectively. we concentrate on R-charts.min1x1. We should also point to a limitation of the R-chart. namely. The lower and upper control limits are taken to be LCL = D3R and UCL = D4R. The frequency of the sampling depends on the stability of the process. are zero for sample sizes smaller . We approximate the centerline of the R-chart with the average of the k values of R. on the cost of the sampling inspecn tion. they depend on the sample size n. twice a shift 1compressive strength2. and thus the limits. and we enter these quantities on the x-chart and R-chart.1 in Appendix C. Either the range R or the standard deviation s of samples of size n can be used to measure variability. k ja j =1 1 k x. these samples can be taken every four hours 1see the weld strength example2. Depending on the application. and so on. xn2. are chosen so that. Á . then the averages or the ranges 1or both2 are likely to exceed the limits and generate an alarm. If the process is stable. R = 1 k R.1. if there are shifts and drifts in the process. These constants are chosen such that almost all future averages x and ranges R will fall within the respective control limits. on which we plot ranges of successive samples. from every other furnace run 1wafer example2.and s-charts can be constructed. the construction of the control charts is very simple. The control limits in the R-chart are given by LCL = D3R and UCL = D4R. We take samples of a few observations 1usually. of course. the longer is the time between samples. and D4 can be found in Table C. The constants D3 and D4.

the control limits for the x-chart are p LCL = x .38 and The limits on the R-chart are LCL = D3R = 10218. one would also like to know whether certain actions have led to a reduction in variability. We could have expected this. called a special cause. because we were told that the process was in control when the observations were taken.6 83. . But let us plot the results of the next two samples. discussions with workers on the production line.92 = 79. This implies that. checking whether there were changes in raw materials. we ﬁnd from Table C. also given in Table 5.0 83. We see that the averages and ranges of all 10 samples are within these limits. With n = 5 observations in each sample.80371 05 293-335 r2 sr 7/26/08 1:45 PM Page 297 Shewhart Control Charts 297 Table 5.1-1 Compressive Strength of Concrete 1kg>cm22 Sample 1 2 3 4 5 i 6 7 8 9 10 Mean 11 12 91 84 93 76 83 84 83 78 82 88 79 72 Compressive Strength 88 89 90 84 85 84 89 79 81 90 87 79 88 80 87 82 81 90 80 90 87 83 82 76 90 79 89 79 80 79 82 81 86 84 85 77 83 87 85 82 86 83 91 85 79 87 83 78 x 88. on these charts. in quality improvement applications.52 + 10. or at least thought to be under control. but not about reductions. which are the limits shown in Figures 5.4 is smaller than the lower control limit on the x-chart.10. we ﬁnd that the twelfth average x = 76.82. any causes found should be eliminated. and D4 = 2.0 85.1-1 and 5.8 88.2 76. In that table.4 R 8 10 8 8 6 11 11 12 8 7 R = 8.0 82. because. D3 = 0. looking for any other unusual condition2 that will identify a cause that can be assigned to this event. The process was sampled twice during each production shift.6 83.8 80. Consider the data in Table 5.A2R = 84.1-2.66.4 x = 84.115218.. Thus.e. p UCL = x + A2R = 84.0 84.577218. Such a ﬁnding should lead to an investigation 1i.1 that the constants are A2 = 0. the R-chart can warn about increases in variability.0 86. This fact should alert the user to the possibility that this particular sample represents the result of an unusual event.9 8 7 Samples used to determine the control limits than 7.92 = 18. That is unfortunate.577.577218.92 = 89.1-1. which lists the compressive strength measurements on concrete blocks from k = 10 samples of size n = 5. for small n.52 83.1-1.115. Finally.52 .92 = 0 and UCL = D4R = 12. and the observations were taken while the process was under control.

that we simply need to judge whether a manufactured item is satisfactory. half-day. here we just check an item on a pass–fail basis: that it is within or outside the speciﬁcations. where n is the size of each sample. That we have a stable process 1or. they can be applied to virtually all situations in which data are taken sequentially in time. although we prefer to take more accurate measurements. In fact. The requirement to identify assignable causes and eliminate them forces management and workers to take an aggressive attitude toward maintaining the quality of the work. Assume that an inspector on the production line checks a sample of n items at certain stated periods 1every hour. most items must be within speciﬁcap tions too. 1nA2R are outside the speciﬁcation limits.The situation should be reviewed carefully. If a stable process is not capable of producing items within the speciﬁcation limits. A point outside the control limits forces us to ﬁnd an assignable cause of this unusual event and. however. with questions such as “How many items are outside the speciﬁcations?” and “Were the speciﬁcations determined correctly?” addressed.80371 05 293-335 r2 sr 7/26/08 1:45 PM Page 298 298 Chapter 5 Statistical Process Control Control charts are useful methods that help us assess whether a process is stable. Remarks: The use of control charts and a strategy of investigating and eliminating special causes will lead to stable processes. almost all items should be between x . That is. to say it differently. Suppose.1nA2R and x + 1nA2R. but in other areas as well. such as the service industry. but also for proportions. The ﬁrst step in improving processes is to bring them under control. say. 5. to make certain changes in the process that prevent such conditions from happening again. we must think about making changes to our process. is still present and may lead to products that are outside the speciﬁcation limits. depending on the numbers of items produced each day2 and observes the number of defectives. more important. for example. unusual causes. Once we have eliminated special causes and have made a process stable. Their graphical simplicity makes them a very valuable instrument for process control. However.1-2 p-Charts and c-Charts Control charts are useful not only for averages and ranges. Thus. etc. we want to make it clear at this point that control limits and speciﬁcation limits are not the same. 3s. d. a process that is under control2 implies that we have been successful in eliminating special. day. This can be done in the following way: Because A2R is an estimate of 3s> 1n. If these two bounds are within the speciﬁcations. if one or both of the bounds x . 1nA2R is an estimate of m . Control charts are also useful not just in manufacturing applications. we will learn how to decide which changes are most promising. such as proportions of defectives. it follows p that 1nA2R is an estimate of 3s. it is highly likely that some of the items will be outside the speciﬁcation limits. The variability that is due to common causes. They alert the user to situations in which something has shifted.. we can check whether the process also satisﬁes the required speciﬁcation limits. However. If the underlying distribution is approximately normal 1unimodal and fairly symmetric withp p out long tails2. x . among . In later chapters on the design of experiments. Control charts will uncover many external sources that lead to shifts in the mean level and in the variability of the process.

9732 50 = 0. we have assumed that 2. we should look for possible reasons for the sudden increase in the number of defectives. and 5>50 = 0. or percentages.000.p2 n and UCL = p + 3 B p11 . 2.027 is the average fraction of defectives. in a stable process 1i.3 B p11 . If this is done for k periods. However. 2>50 = 0. d>n. Thus. this may not be the case for other items. we ﬁnd that the sixth additional fraction defective is above the UCL.027210. n = 50 fuses are tested. together with the centerline at p. we plot the LCL at zero or omit it entirely. In such a situation.04. we call it a p-chart.1-3. For the ﬁrst k = 20 hours. a process that produces defectives at the rate p2. The control limits LCL and UCL. because we are plotting fractions of defectives. 2>50 = 0.04.9732 50 = -0. a point exceeding the upper control limit indicates that the process has deteriorated. In particular. 4>50 = 0. and 5.02. The average fraction of defectives is 1d1 + d2 + Á + dk2 p = . Additional values would also be plotted as long as the process is under control.1-1 number of defectives: 1 1 3 0 2 4 0 0 1 2 3 2 0 1 1 1 3 0 0 2. together with 6 more recent ones 1those with d values of 1.We must ﬁrst decide whether this fraction is representative for our particular process.027210.e. 2. given that nk = 1. 2. can never be less than zero. are plotted on a chart. Example Each hour.027 . .p2 n . 2>50 = 0. d>n.80371 05 293-335 r2 sr 7/26/08 1:45 PM Page 299 Shewhart Control Charts 299 the n items.3 and UCL = 0.. These 3s limits are obtained from the sampling distribution of a proportion which has variance p11 .08. d2. and fraction of defectives 1>50 = 0. This suggests that the process has become unstable and that corrective action should be taken.7 percent defective is acceptable and that we are willing to produce at that level. we have plotted the preceding 20 values of the fraction defective.096. will be between the lower and upper control limits: LCL = p . we ﬁnd the following 5. In this example. 4. If it is. then LCL = 0. nk Statistical theory implies that.04. Á .042 B 10.027 + 3 B 10. almost all of the future fractions of defectives.p2>n. dk. we obtain the number of defectives d1. and because the fraction defective. Because LCL 6 0. In Figure 5. it follows that p = 27>1000 = 0.102. Fractions outside these limits suggest that the process has gone out of control and that the fraction of defectives has changed.

rather than the number of defectives among n items.31c and UCL = c + 31c.1-3 p-chart for the data in Example 5. resulting in c1. Á . The average is c = 26>15 = 1. These are approximations to the 3s limits l . of the Poisson distribution2 with the sample mean. and the respective lower and upper control limits for the c-chart are LCL = c . cj. except that now we count the number of ﬂaws or defectives on a certain unit 1a bolt of fabric.00 5 10 Hour 15 20 25 Figure 5. 2.12 UCL Fraction defectives. This is done each hour for k hours. with an average of c = 1c1 + c2 + Á + ck2>k.73. as well as the variance.1-2 2 1 1 0 5 2 3 1 1 2 0 0 4 3 1.73 = -2. together with the 10 additional observations 5 0 1 2. The Poisson distribution with parameter l is appropriate in this context. 31l of a Poisson distribution.73 = 5. Suppose that we determine the number c of blemishes in 50 feet of a continuous strip of tin plate. a length of wire. LCL = 1.04 Centerline 0.80371 05 293-335 r2 sr 7/26/08 1:45 PM Page 300 300 Chapter 5 Statistical Process Control 0. p 0.1-1 The c-chart is similar to the p-chart. . j = 1. and we also have UCL = 1.68.3 21. Its centerline is given by c. k. ck.73 . c2.08 0. Á .22 1=02 and 3 1 1 0 2 2 These 15 points. and so on2. Example We observe k = 15 50-foot tin strips and obtain the following numbers of blemishes: 5. The control limits are obtained after replacing the parameter l 1which is the mean.73 + 321. because it approximates the distribution of the number of defectives. The c-chart is a time-sequence plot of the number of defectives.

the control limits may change slightly.and s-charts. c UCL 4 2 Centerline 5 10 15 20 25 Sampling period Figure 5. the probability of such an arrangement is 1>4? Certainly not. as it is with these 10 additional points. new control limits are calculated if the points continue to fall within the control limits. to claim improvement when.” Minitab constructs control charts for measurement variables. this probability is 1>8. or x. indicate that the process has become unstable.and c-charts. even if the process has stayed unchanged. The commands in “Stat 7 Control Charts 7 Attribute Charts” construct control charts for attribute data. are known to misinterpret their graphs. they are usually referred to as Shewhart control charts. In the United States. in the desire to improve. Remarks: The control charts that we have been discussing in this section were developed in the late 1920s by Shewhart in the United States and by Dudding and Jennet in Great Britain. future points are plotted on this c-chart. Of course. For example. Two consecutive sales periods with sales above the stable level are often taken as evidence of improvement. Indeed. such as p. especially.80371 05 293-335 r2 sr 7/26/08 1:45 PM Page 301 Shewhart Control Charts 301 6 Number of blemishes. as long as the process is under control. however. Through the commands in “Stat 7 Control Charts 7 Variables Charts for Subgroups. Is that enough. Points outside the control limits. a supervisor may believe that three successive points below the centerline of a p-chart are an indication that the process has improved. Sales managers.and R-charts. as there is only one arrangement with all ﬁve points above the stable level among the 25 = 32 equally likely ones. thus. Because this probability is rather small. you should guard against reading too much into short sequences of points.1-4. Occasionally. such as x. Causes of these unusual events must be found and eliminated. we recognize that the probability of getting three points below the centerline is fairly large. Perhaps ﬁve observations above the stable level in succession would be more reason for celebration. Of course. The probability of such an event is 1>32. however. . with no actual changes. it seems more likely that some signiﬁcant improvement has taken place.1-2 are plotted on the c-chart in Figure 5. as there is only one arrangement with all three points below the current level among 23 = 8 equally likely ones.1-4 c-chart for the data in Example 5. When viewing time-sequence plots such as Shewhart charts.

4 138. we consider deviations xi . by contrast. g = x. Is this process under control? If so. 2d ed. are more responsive to small changes in the mean level. Montgomery. or cusums as they are often abbreviated. The averages and the ranges 1in units of 0.2 139. i=1 r The mean of the in-control process is usually taken as the reference value g.g of the observations 1or sample averages2 from a reference value g and calculate the cumulative sums Sr = a 1xi . such as D. The cumulative sums.8 140.] .0 141. R 9 8 15 6 10 8 10 8 7 13 Hour 11 12 13 14 15 16 17 18 19 20 Average.0 141.2 141.4 140. that p is. L. Five knobs from each hour’s production were selected and measured. Grant notes that a particular dimension determines the ﬁt of a molded plastic rheostat knob in its assembly.1 5.2 Range. In cusum charts. Exercises 5. 1New York: McGraw-Hill. 23.2 140.0 142. 1New York: Wiley. x 137. Grant. The dimension was speciﬁed by the engineering department as 0.8 Range. They create an atmosphere in which the quality of the process is checked on a regular basis. p. They enhance our awareness of the present state of the process and make us “listen to” the process. is their relative insensitivity to small or moderate changes in the mean value.6 141. Statisticians have developed rules that help us decide whether a trend in the cusum path comes from a change in the level or whether it is due to random ﬂuctuations in the process.0 139.140 . Such charts require that measurements on the process be taken on a regular basis and that the results be prominently displayed. 19522.001 inch2 for the ﬁrst 20 hours were as follows: Hour 1 2 3 4 5 6 7 8 9 10 Average. are then plotted against r.8 140. 5th ed. Cumulative sum 1cusum2 charts. A disadvantage of these charts. x 139.003 inch. [Interested readers may consult books on process control.80371 05 293-335 r2 sr 7/26/08 1:45 PM Page 302 302 Chapter 5 Statistical Process Control 5. 20042].6 141.1-1 E. L.6 140. Introduction to Statistical Quality Control. Statistical Quality Control.4 141.1-3 Other Control Charts Shewhart control charts provide a useful display of the data and give us a simple rule for making decisions as to whether a process has started to become unstable. does it satisfy the speciﬁcation limits? [E.0 139.2 140.g2 = 1xr .1. however. A rising cusum path is an indication that the level of the process may have increased.8 143.g2 + Sr . 0. A special gauge was designed to permit quick measurement of the actual value of this dimension. R 12 9 3 8 9 13 6 8 8 6 Construct an x-chart and an R-chart.

7 0.0 0. and 30 defectives. *5.68 5.5 1.1 0.1-4 A company that produces certain bolts considers their quality adequate as long as the proportion of defectives is not larger than 2. thus.1-3 A one-month record of daily 100 percent inspection of a single critical quality characteristic of a part of an electrical device led to an average fraction defective of p = 0. the company takes a random sample of 100 bolts each hour and counts . the inspector found 8. To monitor the quality.06 5.6 1.52 5.3 0.Weld strength is measured by a pull test to destruction.5 percent.40 5.42 5.0145. 45.6 1. 30132: 38–43 119852.32 5.After this one month of 100 percent inspection.7 1. 7.6 1. Actually. 15. uses statistical quality control tools to ensure proper weld strength in several of their welding procedures.1-2 Astro Electronics.94 4. 12.6 0.82 4. Shecter.22 Range 1.06 4.5 1.28 4.6 1. ﬁve2 is taken periodically throughout the production process.86 4.] Construct the x.54 5. *5. Shecter reports that initially the samples were taken every hour. the company switched to a sampling plan under which a sample of 500 units was selected each day. A sample of a small number of items 1in this case.34 5. but the data showed a relatively stable process.0145 as a centerline.58 5. 20. an RCA division. 10.4 1.2 0. are listed as follows: Date 1>3 Time 7:30 11:25 16:20 20:00 23:30 7:20 11:25 16:20 20:20 23:00 7:40 10:20 14:00 20:20 23:00 8:00 11:20 14:00 9:00 7:40 11:00 14:00 Average 5.4 0.90 4.80371 05 293-335 r2 sr 7/26/08 1:45 PM Page 303 Shewhart Control Charts 303 *5.58 4. 8.7 1. 13. plot this information on a p-chart and interpret your ﬁndings. “Process Control for High Yields. Interpret your results and comment on the stability of the process. the middle.48 5. and the end of each shift.6 1.96 5.and R-charts.4 1. sampling about every 4 hours 1or less2 was thought appropriate.” RCA Engineer.68 4.8 1>4 1>7 1>8 1>9 1>10 [E.42 5.42 5.9 1. During the ﬁrst 10 days. Using p = 0. usually at the beginning.The averages and the ranges of 22 such samples.9 1.8 0. Note that the frequency of sampling should depend on the stability of the process. taken from an article by Shecter.

10. To monitor the quality. 1.1-7 The following data give the results of an inspection of 100-yard pieces of woolen textile:The numbers of defects among the last 12 samples are 3. and 7. 5. calculate the centerline and the lower and upper control limits of the appropriate control chart. Calculate the control limits of the c-chart. Hint: The average run length in (a) is given by ARL = 112P1T = 12 + 122P1T = 22 + 132P1T = 32 + Á = 112P1X1 7 h2 + 122P1X2 7 h and X1 6 h2 + 132P1X3 7 h and X2 6 h and X1 6 h2 + Á . 1. If the average number of defects shifts to 2. 7. our only concern is whether we exceed the upper control limit2 is given by ARL = 1 . What conclusions would you draw from this information? *5. 2. What conclusions would you draw from this information? 5. 0.e. 1a2 Using p = 0.0. and 4.1-9 We are concerned that the level of a process might increase from a speciﬁed acceptable level m0.. 8.1-6 In the production of stainless steel pipes the number of defects per 100 feet should be controlled.ks and upper control limit m0 + ks. 1b2 Suppose that the numbers of defectives in the last six samples were 2. Suppose that one uses the Shewhart chart for individual observations 1n = 12 with centerline at m0 and upper control limit of h = m0 + ks. 1c2 Calculate the average run lengths in part 1b2 for a two-sided chart with lower control limit m0 .1-5 Past experience has shown that the number of defects per yard of material follows a Poisson distribution with l = 1.025. 8. what is the probability that it will be detected by the c-chart on the ﬁrst observation following the shift? What is the probability that this shift is not recognized for the next 10 1202 observations? 5. 0. 10. 3. and 8.2. The expectation of this random variable. This information was used to establish the control limits of the associated c-chart. P1X 7 h2 1b2 Assume that the observations X come from a normal distribution. Assume that successive observations of X are independent with standard deviation s. Á .0. Construct the appropriate control chart. 4. T is a random variable that can take on integer values 1. 2.80371 05 293-335 r2 sr 7/26/08 1:45 PM Page 304 304 Chapter 5 Statistical Process Control the number of defective bolts. E1T2. 5. Calculate the average run length for k = 2. 1. is called the average run length 1ARL2. With average proportion p = 0. 0.5. From 15 randomly selected pipes of length 100 feet. 9. Deﬁne the run length T as the time at which the process exceeds the control limit for the ﬁrst time. 5. 3.02.0. calculate the lower and upper control limits of the appropriate control chart. 0. 4. 3. 0. Is this process under control? 5. 7. 2. Now suppose that the numbers of defectives in the last six samples were 3. 1a2 Show that the ARL for the one-sided Shewhart chart 1i. 4. 4. 2. we obtained the following data on the number of defects: 6. 1. Is this process under control? *5. the company takes a random sample of 80 components each hour and counts the number of defective items.1-8 A company that produces electronic components considers their quality adequate as long as the proportion of defectives is not larger than 2 percent. 3. 9. where one is concerned about increases as well as decreases in the level. 6. 7. 3.

9 25.4 24. and the dimension of the containers 1in centimeters.4 25.M.80371 05 293-335 r2 sr 7/26/08 1:45 PM Page 305 Process Capability Indices 305 You can show the result in 1a2 by using the independence assumption 1which implies that the probability of an intersection of events is the product of the individual probabilities2 and properties of geometric sums.1 25.3 25. 1:30 2:30 3:30 4:30 5:30 6:30 25. from front to back2 is of interest.1 25.M.2 25.0 25.2 25.0 24.2 Process Capability Indices 5.8 25.2 25.8 25.0 25.2 25.2 25.M.0 25.1 25. 5.3 25.2 Construct x.8 25.5 25.1 25. 5.3 Measurements 25.4 24. 1:30 2:30 3:30 4:30 5:30 6:30 7:30 8:30 9:30 10:30 11:30 12:30 A.2 25.2 25.0 25. Measurements for the last 25 hours are listed as follows: Time 6:30 A.4 24.6 25.0 25.4 24.1 24.2 24.1 25.1 24.8 25.2 25.0 25.8 25.7 25.1 25.1 25.2 25.9 25.3 25. the customer requires that certain product .3 25.5 25.0 24.3 25.0 25.2 25.2 25.0 25.2 25.1 25.3 25.9 25.1 25.9 25.0 25.2 24.3 25.1 25.2 25.2 25.3 25.9 25.0 25.1 25.1 25.2-1 Introduction One must check whether processes are capable of producing products that satisfy required speciﬁcations.4 25.2 25.2 25.9 25. Each hour.3 25.1 25.4 25.1 24.0 25.3 25.1 24.1 24. 7:30 8:30 9:30 10:30 11:30 12:30 P.1 25.2 25.5 25.1 25.1 25.9 25. and check whether the level and the variability of the process are under statistical control.1 25.1 25.0 25. Typically.and R-charts.1 25.1 25.2 25.2 25.2 25.1-10 A company manufactures paper containers for a detergent.1 25. four cartons are selected from the production run and their dimensions are measured.8 24.2 25.

with speciﬁcation limits LSL = 3.968 3.252 .35 2. Table 5. Of course.254 Date May 19.51 18.244 .000 4.252 .30 16.247 .002 3.35 15. The capability measures we examine are expressed in terms of the speciﬁcations 1the target value and the lower and upper speciﬁcation limits2 and the process characteristics 1the process mean m and the process standard deviation s2.273 .247 Date .00 17. and USL = 4. For the gauge. it is 0.1 percent2 outside the speciﬁcation limits. Engineering considerations and the intended use of the product play important roles in setting the speciﬁcations.248 .988 4.989 3. Once speciﬁcations are set.27 16.256 .2-1.257 . the production process must be monitored to ensure that products meet the speciﬁcations.993 3. we introduce several process capability indices.00 2.994 3. We notice 2 of the 95 width measurements 1or 2.10 16. adding the target value and the speciﬁcation limits to this graph and calculating the proportion of values that are outside these limits. Estimates of capability indices are obtained by taking samples from the process under study and replacing the process characteristics by their sample estimates. thickness2 of steel ﬂats.30 18.998 4.989 3. while 1 of the 95 gauge measurements 1or 1 percent2 is outside the speciﬁcation limits.15 1.00 18.990 3. no 1or very few2 values should be outside the limits. We illustrate this approach with data on the width and gauge 1i.004 3.989 Gauge . 1990 Time 12. Dot diagrams for the 95 width and gauge measurements in Table 5.2-1 Width and Gauge Measurements on 95 Steel Flats Time 16.25 1..985 Gauge .265 in.988 3.e.57 19.248 .242 .15 15.247 . The ﬁgures also show that the process is slightly off target.2-1 are shown in Figure 5. and USL = 0.97 in.20 1.00 16.252 .247 .00 Width 3. then we say that the process is capable of producing to the required speciﬁcations.992 3.00 1.250 .262 .250 .257 . with process means for both width and gauge below their target values.992 3. examine their importance as well as their shortcomings.. the target is 4 inches. they are also called the tolerances of the product.80371 05 293-335 r2 sr 7/26/08 1:45 PM Page 306 306 Chapter 5 Statistical Process Control speciﬁcations be met.993 3. a lower speciﬁcation limit 1LSL2.247 .03 in.25 inch. The speciﬁcations are determined by translating customer requirements into suitable product requirements.992 3. A simple approach in checking conformance is to construct a dot diagram of the measurements 1or a histogram if the data set is large2.990 3.990 3.32 17. Speciﬁcations are usually given in terms of a target value 1Tg2. If they do.992 3.249 . In this section.257 . and an upper speciﬁcation limit 1USL2. For the width.995 3. respectively.247 .235 in. and discuss their implementation.00 15.00 19.21 16.992 3.35 20.998 3.248 . with lower and upper speciﬁcation limits LSL = 0.30 Width 3.

999 4.998 3.30 1.30 7.997 3.995 .994 3.250 .249 .251 .249 .996 3.248 .249 .30 2.249 .30 21.986 3.248 .05 7.245 .244 .250 .012 4.00 5.30 22.002 3.994 3.990 3.00 22.00 9.40 24.250 .994 3.246 May 21.003 3.249 .991 3.249 .20 9.998 3.009 4.243 .30 4.991 3.990 3.990 4.250 .00 2.991 4.80371 05 293-335 r2 sr 7/26/08 1:45 PM Page 307 Process Capability Indices 307 20.987 3.00 21.246 .248 .249 .246 .00 23.993 3.984 3.246 .30 8.30 7.249 .00 18.30 11.990 3.246 .00 7.993 4.989 3.20 7.250 .009 4.50 10.994 3. 1990 .996 3.15 7.50 20.995 3. 1990 17.988 3.00 4.990 3.30 5.00 5.00 0.00 1.00 22.000 4.994 3.969 3.987 .30 23.00 2.012 4.30 19.30 6.239 .245 .251 .984 4.997 3.00 10.250 .006 4.00 1.249 .30 3.245 .994 3.248 .250 .250 .990 3.30 8.40 9.005 4.245 .30 3.998 3.000 4.00 0.252 .989 4.246 .250 .30 9.994 3.00 6.30 3.991 3.992 4.249 .994 3.20 8.15 1.249 .249 .249 .30 5.245 .246 .253 .30 3.246 .244 .000 3.246 .248 .245 .00 17.253 .251 .249 .247 .249 .246 .30 22.248 .249 .247 .250 .05 8.00 11.10 3.45 1.251 .006 4.254 .30 21.997 3.996 3.005 4.00 20.30 18.246 .000 3.245 May 20.006 4.000 4.30 9.987 3.35 8.30 24.00 21.989 3.30 4.990 3.250 .997 3.002 3.00 3.248 .003 4.250 .988 3.30 8.10 7.988 3.009 3.252 .990 3.30 23.006 4.00 23.12 19.40 2.00 4.021 3.250 .246 .988 3.247 .00 8.

For many distributions.3s.80371 05 293-335 r2 sr 7/26/08 1:45 PM Page 308 308 Chapter 5 Statistical Process Control LSL USL (a) LSL USL Figure 5. to the process spread.27 percent defectives. and we recommend their use.2-1 Dot diagrams of width and gauge measurements 5. m + 3s2 covers virtually all of the distribution—in fact. the interval 1m .2 that. the better. the conformance of the process to the required speciﬁcations. However. For capable processes.73 percent if the distribution is normal. Cp = 1 corresponds to 0. 6s Process Spread where LSL and USL are. the probability beyond three sigma limits is 0. the lower and upper speciﬁcation limits and s is the process standard deviation. or 2. we expect that the process spread is smaller than the allowable spread and Cp 7 1.2-2 Process Capability Indices Dot diagrams and histograms are effective graphical summaries of process capability. A large Cp indicates small process variability compared with the width of the speciﬁcation interval. 1The speciﬁcation limits are three standard deviations from the target value. 99.An interval of length 6s measures the extent of the process variability and it expresses the process spread.0027. The larger this index is. for a normal distribution.700 defective parts per million. USL – LSL. and many companies require their suppliers to document the capability of their processes through such calculations. For normal distributions centered at the target.2 Many companies require that . The Cp Capability Index A commonly used capability index is Cp = Allowable Spread USL . Capability indices quantify the capability of a process—in other words. it is common practice to calculate capability indices. Cp relates the allowable spread.LSL = . respectively. and we learned in Section 3.

0 1implying no more than 0.33. The mean and standard deviation for the width are 3.80371 05 293-335 r2 sr 7/26/08 1:45 PM Page 309 Process Capability Indices 309 Cp 7 1.LSL N . as illustrated in Figure 5. as calculations with the normal distribution show2 or Cp 7 1.0.235 N and Cp1Gauge2 = = 1. they are 0. despite the fact that the actual process spread. the process is off target. In the second illustration. The estimated Cp is given by USL .2-2. 6s. We estimate Cp by replacing the process standard deviation s by its estimate s.19. Because Cp makes no reference to the target value. even though the process is not capable of meeting the speciﬁcations. respectively. Motorola2 require that Cp be at least 2. Caution: Cp makes no reference to the target value. 4. is small in comparison to the allowable spread. However. we use the width and gauge measurements in Table 5.5 1no more than 7 defective parts per million2. Some companies 1for example. causing a considerable fraction of defectives.00421. We would have preferred values at least as large as 1. Case 3 3s 3s Case 2 Case 1 3s 3s 3s 3s LSL Target USL Figure 5.33 1which implies no more than 63 defective parts per million.0080. It provides a good description of capability only when the process is on target and the process mean and the target value are the same. which we obtain from past process data. USL .97 N Cp1Width2 = = 1.9947 and 0. it is misleading when the process is off target. 610.1 defective part per million2.03 .2-1. Cp is deceptively large in this case.2-2 Illustrating problems with Cp .265 .24894 and 0. we do not recommend its use.25 610. Cp = 6s As an illustration.3. Hence.LSL.00802 0. For the gauge.004212 These values are somewhat smaller than what we would like to see.

For capable processes.9947 3. except that s* = 2E1X .LSL N Cpk = mine . 1.Tg2. we expect that the smaller of these two standardized differences is larger—hopefully. The distance of a measurement X from the target Tg can be written as X . the sum of the distance of the observation from the process Cpm = USL .47. 3s 3s For the width and gauge measurements in Table 5.Tg = 1X .235 N Cpk1Gauge2 = mine .03} = 1. In practice. we need to replace the process characteristics m and s by their estimates x and s. and we recommend its use. One of them is the Cpk capability index. f.2-2 is unacceptable.9947 .27.97 N Cpk1Width2 = mine .80371 05 293-335 r2 sr 7/26/08 1:45 PM Page 310 310 Chapter 5 Statistical Process Control The Cpk Capability Index Processes with small variability. 310. f = min{1. We expect the same for the difference between the process mean and the lower speciﬁcation limit.Tg22 measures the variability of process measurements around the target Tg.24894 . given by Cpk = mine USL .24894 0.10} = 1. Cpk is a much better measure of capability than Cp.33 is taken as an indication that the process fails to satisfy the required speciﬁcations. 1.3.2-1. f = min{1.00802 310. have sparked the development of several indices similar to Cp. 6s* . because the process mean is larger than the upper speciﬁcation limit. it is negative.10. we have 4. we expect that the difference between the upper speciﬁcation limit and the process mean is at least three standard deviations. much larger—than 1. These indices take into account both the process variability and the deviation of the process mean from the speciﬁcation limits 1or the target2.3. f.03 . in fact.265 . The Cpm Capability Index The Cpm capability index is deﬁned as Cpm is similar to Cp. not around the process mean m.0.x x . For example.0. The estimated Cpk is given by USL .00802 and 0.m m .LSL .004212 310. A Cpk that is smaller than the usual cutoff of 1. 3s 3s Cpk relates the distances between each speciﬁcation limit and the process mean to three standard deviations.m2 + 1m . Cpk for illustration 2 in Figure 5.004212 These capability indices are also somewhat smaller than what one would have hoped for. Therefore.LSL . but poor proximity to the target.03 310.

= 1.265 .5 and 0. We can estimate Cpm from N Cpm = 62s2 + 1x . Cpm = Cp. . The second component expresses the bias.0.0042122 + 10.97 = 1. Many companies require that Target-Z lie between -0. we use the width and gauge measurements in Table 5. CR = Process Spread 1 .4.LSL = 2 6s 1 + 1m .Tg22 USL .04.15.2-1: N Cpm1Width2 = N Cpm1Gauge2 = 6210. which means that the process mean must be within one-half of a standard deviation from the target. USL . = Cp Allowable Spread This measure is known as the capability ratio. Target-Z is estimated from sample data by replacing m and s with the sample statistics x and s. Note that this is equivalent to requiring that Cp 7 1.0022 0.Tg22. Target-Z is the standardized difference between the target value Tg and the process mean m.80371 05 293-335 r2 sr 7/26/08 1:45 PM Page 311 Process Capability Indices 311 mean and the distance of the process mean from the target.Tg22 USL .3. 6210. namely.9947 . then the denominator in the expression for Cpm increases and Cpm itself decreases. It must be supplemented by a measure that assesses whether the process mean is close to the target.m s .LSL we ﬁnd that Cpm = 62s2 + 1m . If the process mean is different from the target.5.75.0.008022 + 13. it represents the proportion of the allowable spread that is taken up by the process spread. Target-Z = Tg . With s* = 2s2 + 1m .33. Multiplied by 100.Tg 22 s2 .2522 Capability Ratio and Target-Z Several other indices are in use.Tg2 A s2 = A Cp 1 + 1m . For processes that are on target with m = Tg. The capability ratio has the same drawbacks as Cp.LSL . many companies require that the capability ratio of most of their processes be less than 0.03 . Small values of CR are desirable. As an example. Some companies use the reciprocal of Cp.24894 .235 4.

0. One could argue that Motorola’s tolerated 1. also realizes that the process is not always exactly on target.5s. variation due to slowly changing factors such as operators. One could argue that the process depicted is a good one. Motorola’s Six Sigma Concept In announcing the achievement of Total Customer Satisfaction as the corporation’s fundamental objective.4 ppm? Obviously. we ﬁnd the following estimates CR1Width2 = 0.6s. Case 3 in Figure 5. tolerating drifts in the process mean is quite reasonable.Pa Tg . capable processes satisfy both CR 6 0.52 = 0.00 . Motorola introduced the concept of Six Sigma as a statistical way of measuring quality. Motorola requires highly capable processes.75 and -0. and operating instructions2.6s 6 X 6 Tg + 6s2 = 1 .3. and assuming that the observations are from a normal distribution with 1shifted2 mean Tg + 11.5 6 Target-Z 6 0. However.24894 = 0.e. one can calculate the probability that an individual observation falls outside the interval 1Tg .5s2 s 6 Z 6 Tg + 6s .25. Motorola’s Six Sigma goal is a 3.00421 Neither the width nor the gauge would satisfy the capability standards. however. but fails the requirements on Target-Z.66. the company requires that processes be on target and that the speciﬁcation limits be at least 6s away from the target—hence the name “Six Sigma.0.P1Tg .25 . Cpk = 1.5 6 Z 6 4. Motorola views its failure rates in terms of parts per million 1ppm2.” Such a requirement on the process translates into a capability index Cp = 2. mean levels shift dynamically over time.5 standard deviations below the upper speciﬁcation limit. materials. the process is clearly off target. In fact.0000034. For the width and gauge measurements. as quite often it is difﬁcult to keep the mean exactly on target. hence. Note that the shifted process mean is 4..4-ppm defect level! How does the Six Sigma requirement translate into a defect level of 3.9947 = 0. Under the Six Sigma goal.52s and standard deviation s.2-2 illustrates a process that leads to acceptable Cp and Cpk.5.52 L 1 .5s2 s b = 1 . Motorola.84 and and Target-Z1Width2 = Target-Z1Gauge2 = 4. Tg + 6s2. Allowing for drifts also serves as a proxy for the long-term variation in a process 1i.P1-7.0080 0.80 CR1Gauge2 = 0. as the process variability is well within the acceptable range.0.1Tg + 1. 0. Nevertheless. and a level adjustment would make things even better. better than the standard usually adopted.6s . which is probably larger than the standard .80371 05 293-335 r2 sr 7/26/08 1:45 PM Page 312 312 Chapter 5 Statistical Process Control In sum.P1Z 6 4.1Tg + 1. Motorola allows drifts away from the target of at most 1.4 parts per million. In practice.5. This probability is given by P1observation outside specification limits2 = 1 .5-sigma drift of the process mean from the target is somewhat on the large side—certainly larger than the one-half standard deviation that is allowed under Target-Z. or 3.

there are potential problems. For example.52 = 2. In a situation where one-sided speciﬁcation 1or tolerance2 is prescribed. products are part of a complicated assembly involving many components. We refer to this as tolerance “stack-up. assume that the process characteristics are given by m = 53 mg>liter and s = 0.” However. Speciﬁcation Limits Capability indices depend on the speciﬁcations. As with any other single summary statistic. Take. However. Assumption of Normality Capability indices are designed for normal distributions.2-3 Discussion of Process Capability Indices Capability indices summarize process information in a succinct manner. Capability measures are also useful tools for monitoring capability over time. as well as deteriorations. speciﬁcations should not be taken too narrow either. which usually are determined by customers and engineers.e. Process Capability Indices for One-Sided Speciﬁcations Consider a situation in which the tolerances on a product are one sided. in the process.. It is questionable whether capability indices are meaningful for . Unnecessary expenses are incurred with narrow speciﬁcation limits if tight tolerances are not needed. and processes with high 1or low2 capability can be identiﬁed. One needs to be aware of these problems in interpreting such statistics. 5. If there is too much variability in the door and>or the frame. 3s For example. indicating a highly capable process.5 mg>liter. They establish a common language that is dimensionless 1i. Engineering and manufacturing can communicate through these measures. the parts may not ﬁt together as they should. one must keep in mind that a capability index is just a single summary statistic and can never bring out all features of a distribution. Usually. It would be wrong to select them too wide. which reﬂects the shortterm variation in the process. the product requirements may specify that the concentration of a certain ingredient be at least 50 mg>liter 1so that LSL = 502 or that the weight of an item be at most 80 grams 1so that USL = 802. it does not depend on the particular units of the observations2 and that compares the desired and the actual performance of production processes. we calculate one-sided Cpk indices: Cpk1lower2 = m .Then the capability index for the process. because that may allow inferior products to be manufactured. If there is too much variability in one component 1because speciﬁcation limits or tolerances are set too wide2. It is important that careful thought go into the selection of the speciﬁcations. the door will not close correctly. we assumed a normal distribution when we related the Cp index to the proportion of unacceptable parts. for example. is Cpk1lower2 = 153 . they indicate improvements.m .0. car doors and car frames.80371 05 293-335 r2 sr 7/26/08 1:45 PM Page 313 Process Capability Indices 313 deviation used in the denominator of the capability index.LSL 3s and Cpk1upper2 = USL .502>310. relative to a one-sided lower tolerance LSL = 50 mg>liter.

12 where n is the number of observations that are used to estimate the capability index. 2 1 1 + f. this interval extends from 0. and Hurley . and the standard deviation is 14>12 = 0. now consider a symmetric distribution with heavy tails 1i. P.2 Instead of a uniform distribution with light tails. N For the n = 95 width measurements of the steel ﬂats with Cpk = 1. it is capable of surprising you! The fact that the process is not under control means that some outside factors are causing it to be unstable. Because you don’t know what these factors are. Cpm = Cpk = Cp = 132>316210.87. and it is important to evaluate its margin of error. 242.. These two sets of limits must not be confused! A process should be under control before one assesses its capability.577. the calculated capability index is just an estimate of the unknown process capability index. for example.. 188–195. which is rather poor.87 to 1. R. and also rather simple. The capability index will look quite good. If the process is not under control. because. In this situation. and assuming that the measurements are normally distributed.] . the capability index raises unnecessary concerns. Assuming that the process is stable.57704 = 0. 1Keep in mind that it is always useful to supplement these indices by a histogram of the measurements. A capable process that is not in control is not very reassuring. whereas the probability of exceeding the speciﬁcations is large 1certainly larger than the probabilities implied by the normal distribution2. approximate 95 percent conﬁdence interval for Cpk is given by N Cpk e 1 .” Journal of Quality Technology.2 Suppose that the speciﬁcation limits are given by -a and +a. you don’t know how well you will produce in the future. Conﬁdence Limits for Capability Indices When calculating capability coefﬁcients. we replace the process parameters 1m and s2 by sample estimates. there is no guarantee that the unexplained variation will stay in a range such that the products are still satisfactory. “Conﬁdence Bounds for Capability Indices. gives an in-depth review of this and several other approaches to approximating conﬁdence intervals for capability indices. Hence.5. consider a = 1. The probability that this process exceeds the speciﬁcation limits is zero. The Difference Between Process Stability and Process Capability It is worthwhile to repeat earlier comments about the difference between control and speciﬁcation limits. A useful.80371 05 293-335 r2 sr 7/26/08 1:45 PM Page 314 314 Chapter 5 Statistical Process Control distributions that are very different from the normal distribution. [Kushler. Indeed. The mean of this distribution is given by 0. assume that the variability of a process is described by a uniform distribution on the interval from -1 to +1. To illustrate this point. 1992 1Vol. Speciﬁcation limits and control limits are not the same. N A 9n1Cpk22 21n . The former reﬂect the requirements of the customer. for the uniform distribution. it is possible to obtain approximate conﬁdence intervals for the various capability indices.. However. a result opposite to the one implied by a uniform distribution will occur. there is no probability beyond -1 and +1.H. For such distributions.21.e. especially if the sample size n from which these statistics are calculated is not very large.04. a distribution that makes it more probable for observations to fall far from the mean2. 1See Chapter 3. while the latter provide bounds on the commoncause variability of the process.

2-2 The target value for the pH content of a certain shampoo is 6.k2Cp. and the sample standard deviation 0.2 mm below the target.44.m ƒ >1USL . 5. the lower and upper control limits on the x-chart are 98. 1b2 A stable process 1i.and R-charts that use subgroups of size n = 5.165 and standard deviation s = 1. Assume that the target is 84 1kg>cm22.7419. There.1-10.2-4 Show that Cpk = 11 . minimum pH value of 5. and 1iii2 outside the speciﬁcation limits.85 and the upper speciﬁcation limit is USL = 6. A sample of 200 bottles was taken and the pH value for each bottle was determined. CR. Cpk. Use the latter expression to solve for s.2-5 The requirements on the dimensions of steel ﬂats specify a target of 100 mm.66. with a target of 100.8 mm and 100.LSL2.15. obtain an estimate of the probability of getting an observation 1i2 above the upper speciﬁcation limit.15. 1c2 The lower and upper speciﬁcation limits on a bakery good are 95 and 105 1decagrams2. and a standard deviation of 0.91. *1a2 Calculate and interpret Cp and Cpk.e.. Cpm. with lower and upper speciﬁcation limits given by 77 kg>cm2 and 91 kg>cm2. respectively.45. Cpk. The lower 5. Calculate and interpret Cp. 5.2-3 A sample 1n = 1502 from another production run of shampoo resulted in a 5.2-2 on this run of shampoo. Cpm.2 * 5. *1a2 Conﬁrm the sample mean and standard deviation. Hint: Use the fact that A2R in the control limits of the x-chart is an estimate of 3s> 1n. the dimensions of cartons and found that both the level and the variability of the process were under statistical control. Repeat Exercise 5. CR.2-1 Consider the 50 observations on the compressive strength of concrete listed in Table 5. with Tg the * 5. and estimate the capability indices Cp. the distribution of pH was well approximated by a normal distribution.2-7 Consider the data shown in Exercise 5.1-1 1samples 1 through 102. The stability of the process is well established and is being monitored with x.12. It was found that the minimum value was 5. a maximum of 7. In addition. or is there a problem with excessive variability? *1b2 Assuming that the distribution is normal. the maximum 6. with lower and upper speciﬁcation limits of 97 mm and 103 mm.8 mm. *1b2 Obtain a 95 percent conﬁdence interval for Cpk. and Target-Z. and Target-Z. and interpret your ﬁndings.2-6 Discuss the following statements and calculate the capability indices: 1a2 The x-chart is used to check whether process variability is stable. speciﬁcation limit is LSL = 5.52 and standard deviation 4. * 5.132. we studied the stability of . with sample mean 84. respectively.80371 05 293-335 r2 sr 7/26/08 1:45 PM Page 315 Process Capability Indices 315 Exercises 5. where k = 2 ƒ Tg .123. a process that is under statistical control2 is also capable. the sample average 6.89. Are you satisﬁed with the capability of your process? What could you do to increase the capability? Is there a problem with not being on target. speciﬁed target and LSL and USL the speciﬁcation limits. 1ii2 below the lower speciﬁcation limit. It turns out that the process mean is 0. Obtain Cp. Interpret your ﬁndings.212. an average of 6. A sample of n = 200 goods resulted in a sample mean x = 100. Cpk. and Cpm.

0 1the process mean is 1. that is.0 standard deviation off the target2. in the case of destructive inspection. If p is small. otherwise. we let p represent the fraction that is defective in the lot. 5. * 5. where 0 … p … 1. Note: This problem has been considered in Tadikamalla. but a large p suggests that the quality is unacceptable and that the lot should be rejected and not be shipped. in most instances. Example Let a lot of N = 1. and suppose that the mean misses the target by a multiple k 7 0 of the standard deviation. Items could be sent in boxes—or even boxcars—and a standard grouping of these items is often called a lot.5 1the process mean is 1. depending on the nature of the item. in which an item 1e. However. The measurements have a normal distribution with mean m and standard deviation s. we reject it. By the multiplication rule of probabilities.5 standard deviation off the target2. we accept the lot.We take n = 10 5.000 items contain Np defectives.. we resort to sampling to determine whether the lot should be accepted or rejected. calculate the proportion of defectives if *1b2 Cp = 1. that is. In particular. 272. Calculate Cp and Cpk.Thus. Relate the proportion of defectives to the capability index and the multiple k 7 0.g.R. we may accept certain lots that we would reject if we looked at every item. we should consider the probabilities of making these two types of errors.80371 05 293-335 r2 sr 7/26/08 1:45 PM Page 316 316 Chapter 5 Statistical Process Control The speciﬁcation limits for this process are 24 to 26 cm. *1c2 Cp = 1. That is. P. m = Tg + ks. with a target of 25 cm.3 Acceptance Sampling Manufactured parts 1items2 are often shipped from a supplier 1the producer2 to another company 1the consumer2 that uses the parts in the construction of some ﬁnal product. If all are satisfactory. some of these items are probably defective.2-8 Consider a situation with symmetric speciﬁcation limits around the target Tg. and we may reject others that we would accept in a 100 percent inspection.5 and k = 0. 83–85. Use the mean and the standard deviation of the 100 observations in your calculation. we do not know p or the number Np of defectives in the lot. We must ﬁnd an estimate of p to help us decide whether to accept or to reject the lot. for some c 7 0. To decide whether the acceptance sampling plan is a desirable one. *1a2 Cp = 2 and k = 1. a fuse2 is destroyed in the testing. Assume a capability index of Cp.5 standard deviations off the target2. the supplier is providing acceptable quality. 100 percent inspection is usually extremely expensive 1sometimes more costly than the production of the item itself2 and sometimes impossible—for example. here. “The Confusion Over Six-Sigma Quality.” in Quality Progress. Of course.c and USL = Tg + c. and check whether the process is capable. Unfortunately. we could inspect all of the items in the lot.. LSL = Tg . there are Np defective items among the N parts.5 and k = 1.3-1 items at random and without replacement and test them. Let each lot consist of N items. However. the probability of obtaining n good items and thus accepting the lot is . November 1994 1Vol. where N can range from 10 or so to several thousands. By observing just a sample 1only a fraction of the items in the lot2.5 1the process mean is 0.

Np .0. Nevertheless.00010.f under P1X … 02 with n = 10.1n .2 N . if N is considerably larger than n.000 items in the lot. and we frequently use it.6 0.06. and its graph is given in Figure 5. with p = 0.0 0.p210. The OC curve is a function of p. The reader may ﬁnd this OC curve highly undesirable because there is a probability of 0. say.p2a1 Np Np Np b a1 b Á a1 b. The procedure of sampling with replacement is not too practical. OC(p) 1.4 0.1 N .15 0.1n .12 With n = 10 and N = 1.1n .052 = 0.1n .5. OC10. The probability of accepting the lot is called the operating characteristic 1or OC2 curve.35 of accepting the lot. there is a 35 percent chance of accepting a seemingly undesirable lot that has 10 percent defectives.30 0. In addition.000.11. however. 1See the related discussion in Section 4.302 = 0.2 In this example.2 N .10 0.12 = 11 . even though there are 1.02210 = 10.20. This is the probability that we would have obtained if we had taken the n = 10 items at random and with replacement—that is.3-1.2 p 0.1 N . the probability of accepting the lot is approximately equal to the binomial probability 11 .2 ba ba bÁa b N N .252 = 0. replacing each selected item before selecting the next.3-1 OC curve.80371 05 293-335 r2 sr 7/26/08 1:45 PM Page 317 Acceptance Sampling 317 a N . N . Probability of accepting a lot as a function of the proportion of defectives.03 can be found from the binomial c.Np .8 0. That is. it is easy to see that N N N L L Á L L 1.25 0. OC10.1 N .12 Hence.35. OC10.2 N . The values OC10.60.p210.d.Np N . the OC curve is approximated by OC1p2 L 11 . N . p . and OC10.202 = 0.02.Np . the probability of accepting a desirable lot.102 = 100 defective items among the N = 1.152 = 0. is 11 .1 N .35 Figure 5.82.98210 = 0.12 N .102 = 0. OC10. this binomial probability provides an excellent approximation to the true probability.05 0.20 0.

000 5.3-2. so we begin by using the Poisson approximation.10 is called the consumer’s risk.2 a 0. and that is undesirable.05 0.18 and 0. so the probability of rejecting a lot of acceptable quality is quite close to the desired a. Furthermore. p .35.06 Figure 5.35.02 and p1 = LTFD = 0. are frequently called acceptable quality level 1AQL2 and lot tolerance fraction defective 1LTFD2. The only way we can correct the situation is by taking a larger sample size and redesigning the acceptance sampling procedure. With l = 1100210.02 0. we see from the Poisson distribution that if Ac = 4. for the particular item that we are manufacturing.062 = 6 and Ac = 4 together imply.02 and p1 = 0.To achieve such an OC curve.947 = 0. we want the probabilities of errors at those two values of p to be about a = 0. the producer’s risk of rejecting a satisfactory lot and the consumer’s risk of accepting an undesirable lot.1 p 0.022 = 2. that OC10. However.04 0.4 0.3-2 OC curve for Example 5.10. Example Suppose that we wish to design an acceptance sampling plan for lots of N = 5. in this example.06.02 is called the producer’s risk.8 0. which here are a = 0. The probabilities of errors at AQL = 0.6 0.01 0. because the consumer loses if a bad lot is accepted. according to the Poisson distribution with l = 6.285.The probability a = 0.3-2.18. p0 = AQL = 0. 1 . because the producer is hurt if a good lot is rejected.18 of rejecting a good lot of AQL = 0. we desire an OC curve such as that in Figure 5.80371 05 293-335 r2 sr 7/26/08 1:45 PM Page 318 318 Chapter 5 Statistical Process Control This means that the probability of rejecting a lot with only 2 percent defectives is 0. say. The two values of p that were used in this example.022 = P1Y … 42 = 0.03 0.0 0.05 b 0.053. That is. Now. then OC10. too. Let us begin with an initial guess of n = 100 as the sample size. The probability b = 0. OC(p) 1.10. p0 = 0. Ac. the probabilities of the two types of error.18 and b = 0. In other words. respectively.35 of accepting a bad lot of LTFD = 0. Probability of accepting a lot as a function of the proportion of defectives. are too high: 0. have special names.3-2 items.02 and LTFD = 0. Suppose also that.05 and b = 0. To lower this probability. 1100210.0.10. The scheme is to take a sample of size n and accept the lot if the number of defectives is less than or equal to an acceptance number. which is much larger than the desired b value of 0. it seems as if the sample size must be fairly large.947.10.062 = P1Y … 42 = 0.

Example Suppose that we have a lot of N = 1. Also.0.034 is the probability of rejecting a lot of acceptable quality with AQL = 0. a highly desirable sampling plan is described by n = 200 and Ac = 7: Take a sample of n = 200 items from each lot.10 or l = 8. 5 OC1p2 = P1Y … 52 L a 180p2y e-80p y! . OC10.207 is still too high and not desirable. we ﬁnd that we should use n = 80.02 = 0.80371 05 293-335 r2 sr 7/26/08 1:45 PM Page 319 Acceptance Sampling 319 we must increase the sample size.964. Ac = 5.966 = 0. or 2.000 items and we desire an AQL = 0.022 = 4 and 1n210. which gives both l = 15010. y=0 From the Poisson distribution. This probability of 0.5 5.062 = P1Y … 62 = 0.06>0. The OC curve . engineers do not need to construct acceptance sampling plans from ﬁrst principles because the appropriate plans can be looked up in manuals provided by organizations such as the American National Standards Institute 1ANSI2 and the American Society for Quality 1ASQ2.The approximate operating characteristic curve of this sampling plan is 7 OC1p2 L a where n = 200. otherwise. OC10. For many years. however. OC10. which means that we should reject the lot if six or more defectives are found in the sample of n = 80 items. and the rejection number Re = 6. y=0 Usually.062 = 12 imply that n = 200.090. we obtain OC10. We can help him or her in this regard by constructing an OC curve. We believe that the engineer should understand the implications of such a plan.62 = 0.949 and 0. probabilities that are close to the desired 1 . Trying various l values so that the ratio between them is 0.025.05.02 = 0.a = 0.062 = 9.966 means that 1 . reject it.02 = 3. 1np2ye-np y! .785.05 or l = 4. we see that l = 4 and l = 12 with Ac = 7 provide the respective OC values of 0. accept the lot. That is. we calculate the probability of accepting the lot.42 = 0.02 or l = 1.020.02 = 0.02. MIL-STD-105E2.02 = 0. namely. OC10. the federal government had required the use of certain Military Standard plans 1MIL-STD-105D and its 1989 revision.994. From a MIL-STD-105D table that is found in older books on quality control. If there are no more than 7 defectives.983.025 or l = 2. Using the Poisson approximation.95 and b = 0. 1n210. and OC10.10.3-3 percent.022 = P1Y … 62 = 0. Choosing Ac = 6 so that OC10. Let us try n = 150.191. However.022 = 3 and l = 15010.15 or l = 12. OC10.03 or l = 2. the government replaced the mandate with acceptable non governmental standards on sampling procedures.034 is close to the desired a = 0. recently.

8 0.3-4.0 0.Thus.02 = 0. with weights OC1p2 and 1 .3-3 OC curve for MIL-STD-105D example through these points is given in Figure 5. we must average the values p and zero.125 0.AOQ is the expected value AOQ1p2 = 1p23OC1p24 + 10231 .1912 if the lot is 10 percent defective.10 0. it is zero because.02 = 0.15 Figure 5.OC1p24 = p3OC1p24. the fraction defective entering the process is p. yields OC10.983 = 0.OC1p2.025 0. . in the second case.That is. The maximum of the AOQ curve is called the average outgoing quality limit 1AOQL2.To get an average outgoing quality 1AOQ2.3-3. we may want to reduce 1if possible2 the acceptance number to Ac = 4.4 0. If a lot with fraction defective p is accepted. which are their respective probabilities of occurring.10 or l = 8. when n = 80.05 0. Let as now consider one ﬁnal concept associated with an acceptance sampling plan. it is allowed to continue on into the production process. the AOQL is about 4 percent. In the ﬁrst case. after replacing the bad items. If it is rejected. Although the probability a = 1 0. This seems to us a better plan.102 = 0.947 and OC10. an agreement is usually made with the supplier that the lot is to be 100 percent inspected and that bad items are to be replaced with good ones before the lot is sent on into the production process.2 p 0.100. all of the items are good.025 or l = 2.6 0.075 0.017 is desirably small at AQL = 0. we may be somewhat concerned about the high probability of accepting the lot 1given by OC10. It tells us about the worst possible average of outgoing quality and is usually determined by calculus or empirical means. The AOQ curve associated with the MIL-STD-105D plan of Example 5. which.3-3 is plotted in Figure 5.025.80371 05 293-335 r2 sr 7/26/08 1:45 PM Page 320 320 Chapter 5 Statistical Process Control OC(p) 1.

calculate the respective probabilities of accepting lots of N = 50 items that are 2. 11 11 Acceptance sampling plans are important because we want to avoid letting too many defectives into the consumer’s production process.03 0. 1a2 Sketch the OC curve.05 0.p210. and 20 percent defective.035. 1b2 The sampling plan is used for large lots. However.3-4 AOQ1p2 L p11 . we must realize that quality cannot be inspected into products. Hence.02 0. Exercises 5.04 p 0. acceptance sampling plans will deﬁnitely be needed.10. *5. we found that OC1p2 L 11 . Reliable suppliers can be trusted and there is no need to inspect their products.02.3-4 AOQ curve for MIL-STD-105D example Example In Example 5. Sketch the OC curve.01 AOQL 0.075 0.3-1 A sampling plan has a sample size of n = 15 and an acceptance number Ac = 1.04 0. 1a2 Using hypergeometric probabilities. Use binomial probabilities to compute an approximation to the probabilities in part 1a2. 1c2 Calculate the consumer’s risk at LTFD = 0.10 0. which has the maximizing solution p = 1>11.15 Figure 5. it is possible to eliminate the acceptance sampling procedures altogether. .p210. Products have to be built properly in the ﬁrst place! Therefore.3-2 A sampling plan uses n = 100 and Ac = 3. 6. 10. *1b2 Calculate the producer’s risk at AQL = 0. 5.3-1.125 0.025 0. The lot size N is large in comparison to the sample size n.3 5.80371 05 293-335 r2 sr 7/26/08 1:45 PM Page 321 Acceptance Sampling 321 AOQ(p) 0. if a supplier has a good quality control program. thus. But until all suppliers can be classiﬁed in this manner. AOQL = AOQ1p = 1>112 = a 1 10 10 b a b = 0.

and observe the producer’s risk a at AQL = 0. What is the probability that we accept a batch that contains 5 percent defectives? Sketch the OC curve of this double-sampling plan. *5. the Poisson approximation2. and . If the lot is p = 0. Note: This sampling scheme is called a double-sampling plan. we reject it. furthermore.3-6 The following sampling plan is used: 112 Select a sample of size 2 from a lot of 20. If one fails. we accept the lot. If all four articles meet the quality speciﬁcation.05 and b = 0. once the reason is found.3-7 Speciﬁcations require that a product have certain quality characteristics. Engineers face such problems constantly. They indicate whether something has gone wrong and needs to be ﬁxed.000. what is the probability that we accept the lot? Repeat your calculations for several other values of the fraction defective p. *1b2 Plot the AOQ curve and determine the approximate value of AOQL. If two or more fail. 122 If the item in the second sample is good. The product is made in batches of N = 1. Suppose that the lot size N is large enough to permit us to use the binomial distribution 1and. it is often quite difﬁcult to determine exactly what has gone wrong and. otherwise.4 Problem Solving 5.2 fraction defective. 5. The plan stipulates that rejected lots are 100 percent inspected and that defective items are replaced by good ones.010.05 and the consumer’s risk at LTFD = 0. 5.10. otherwise. If both are defective.10.000 items. The acceptance number is Ac = 1.4-1 Introduction The statistical methods that we have discussed so far in this chapter help us monitor processes. reject it. The current inspection scheme is to select n = 4 items from each batch. how to ﬁx it. If one is good and one defective. If both items are good. LTFD = 0. However.02 and * 5. * 5. and plot the OC curve. we reject the lot. If both pass the inspection. Control the probabilities of errors at those two values of p at about a = 0. *1a2 Plot the OC curve. The sampling schemes that we have discussed in this section are singlesampling plans—plans according to which we accept or reject a lot on the basis of a single sample.80371 05 293-335 r2 sr 7/26/08 1:45 PM Page 322 322 Chapter 5 Statistical Process Control *1d2 Assume that a rejected lot is 100 percent inspected and that defectives are replaced by good items. Calculate the producer’s risk at AQL = 0. accept the lot. Plot the AOQ curve and ﬁnd 1approximately2 the average outgoing quality limit AOQL. we take another sample of two.3-4 A sampling plan consists of sampling n = 300 items from a lot of N = 5. reject it.3-3 A sampling plan for the fraction of defectives consists of sampling n = 50 items and rejecting the lot if there are more than two defectives. which 5.08. because it allows the possibility of delaying the decision on the lot until a second sample is chosen. accept the lot. can be determined only by a destructive test.3-5 Design an acceptance sampling plan for large lots with AQL = 0.002 and the consumer’s risk b at LTFD = 0. because p is small and n relatively large. take a second sample of one item. we accept the lot.

engineers. and so on. It is beneﬁcial to include all members who are associated. assess the relationships among variables. Because this diagram identiﬁes the main sources of defects.To identify the most prevalent defect in a manufacturing process. and we ourselves are not sure how best to do it. we collect information on the frequencies of the various defects. it has become a valuable tool in industry. The fuel efﬁciencies of various types of automobiles are measured to see how they differ. variation in the raw materials. we have to collect. with the project. or diameter of certain items to check whether they are within speciﬁcations. data can help us understand and solve problems. decide whether a process is under control. Information on the durability of a certain consumer product is collected to help estimate the amount of money a company will have to spend on warranty repairs. We might measure the width. weight. weekday–weekend differences. However. and much more. the workers with much knowledge about processes and machines must be included. the wind resistance of its chassis.We collect and display information on the percentage of defectives in successive lots to assess whether a production process is stable and to identify sources of its instability—for example. When an engineer is confronted with difﬁcult problems. Data are used routinely to assess the variability of measurements. reliable data are needed to help the team make wise decisions.4-2 Pareto Diagram Let us start our discussion with a simple Pareto diagram that displays the frequencies of various defects. of course. and knowledge of the subject are needed as much as anything. Reliable data provide the information that is needed to make decisions. 5. directly or indirectly. suggest various ways of improving a process. Information must be collected to reveal facts that help solve the problem at hand. Proper ways of obtaining and presenting information are extremely important. Water samples from different locations in a lake are analyzed to discover whether the levels of a particular pollutant depend on location. The size of the team is often determined by the magnitude of the problem: Frequently. compare the effectiveness of various methods. the design of its engine. common sense. more may be needed for a major one. Of course. and analyze information on their causes. The amounts of hydrocarbons that are emitted from car engines with and without catalytic converters are obtained to assess the effect of the catalytic converter on those emissions. we may want to check whether fuel efﬁciency is related to the weight of the car. we do advise that a team of experts familiar with the process be gathered to “brainstorm” the situation. two or three persons can handle a minor problem. and failure to maintain machines. Processes are run at various settings to determine which factors are important and to learn how we can reach the optimum.To be able to eliminate the most important defects. Here we give several examples. It helps to gain this experience by observing and working with experienced problem solvers. “problem solving” is a difﬁcult subject to teach. Experience. Other members of the team may be supervisors.80371 05 293-335 r2 sr 7/26/08 1:45 PM Page 323 Problem Solving 323 university education should prepare them for these kinds of situations. managers. . Naturally. Properly collected and analyzed. length. summarize. and. We take measurements on various types of products to determine which type is the best. Also. ﬁnd optimum conditions.

4-1 Major Causes of Lens-Coating Defects Before Lenses too thick or thin Scarred Cracked Unﬁnished Poorly coated Others 10 30 6 15 45 4 110 After 8 32 8 12 16 4 80 .4-1. we have arranged the bars in decreasing order of frequency: The most frequent cause is on the left and the least frequent on the right. Note that with qualitative variables such as type of defect. the items may show surface scars. the two most frequent defects—poorly coated lenses and scarred lenses—represent 145 + 302>110 = 0. The major contributors to the defects have to be identiﬁed. The three most frequent defects 1poorly coated lenses. consider a lens-polishing company which has found that the number of defectives has increased recently. because only then can an appropriate strategy be pursued to eliminate the most important defect1s2.3 percent. type of engineer 1civil. The data are given in the ﬁrst column of Table 5.9 percent. supplier Table 5. Rare defects are more or less inevitable. The line graph on the Pareto diagram connects the cumulative percentages of the k1k = 1. in this category. It is usually easier to reduce the occurrence of a frequent defect by half than it is to reduce the one occurrence of a rare defect to zero. it makes sense to order them in decreasing order of occurrence. A Pareto diagram is a bar graph that shows the frequencies of the various defects. in a lens-polishing process.68.4-1. or type of machine. However. For example. industrial. 52 most frequent defects. It turned out that a switch to a cheaper. 2. an item can be defective because the lenses are too thick or thin. the company started an effort to improve its lenscoating operation. For example. there are 45 defectives. yet less reliable. electrical. occurring every now and then. In Figure 5. A Pareto diagram is an important component of any quality improvement program because it focuses everyone’s attention on the one or two categories that lead to the most defects. Á . The second most important cause is scarred lenses. etc. In a rubber-molding process. or be incomplete.80371 05 293-335 r2 sr 7/26/08 1:45 PM Page 324 324 Chapter 5 Statistical Process Control Defects can arise from a number of sources. scarred lenses. unﬁnished. in the particular context in which we are looking for the most common type of defect. or 68 percent of all defective lenses. and unﬁnished lenses2 represent 145 + 30 + 152>110 = 0. contributing another 27. be misshapen. Pareto diagrams show that usually only two or three defects account for over 75 percent of the losses. scarred. As an example. or poorly coated. have cracks.2. and so on.82. the most important defect arises from poorly coated lenses. or 82 percent.The various sources can be identiﬁed through a Pareto diagram. the ordering of types is really arbitrary. The company has classiﬁed the total number of defectives from a day’s production 1N = 1102 according to several major causes. or 100145>1102 = 40. In our example. After viewing this information.

The company decided to return to the original supplier.80371 05 293-335 r2 sr 7/26/08 1:45 PM Page 325 Problem Solving 325 Number of defectives 110 100 Relative cumulative frequency (in percent) 100 Number of defectives 110 100 Relative cumulative frequency (in percent) Improvement 30 units or 30 100 27 110 percent 100 75 80 75 50 50 50 50 25 25 Figure 5. A summary of the distribution of defects after corrective action had been taken is given in the second column of Table 5. has developed certain cause-and-effect diagrams which depict the variables that may have affected the response. can be a lengthy project. a Japanese control engineer. Furthermore. Measures must be taken to correct the causes of low quality. which are discussed in the next section.4-3 Diagnosis of Causes of Defects A prime objective of many investigations is to improve quality—that is. because we can usually think of many factors that may have contributed to low quality. Possible causes of scarring should now be investigated. A comparison of the two Pareto displays in Figure 5. and in the process it reduced the number of defects due to poorly coated lenses by about two-thirds. these diagrams show that scarring is the next most important defect.4-1 also gives a nice summary of the improvements that can be attributed to the corrective action that was taken. Kaoru Ishikawa.4-1.These diagrams Sc arr ted oa Po Un fin Un Un To ot To Ot he rs Ot Cr ack arr he orl ot fin fin rs ed ed hic hic yc yc ish ed ish ish ed in oa ko ko ed ed ted r th r th in . Po orl Sc 5. such as poor coating of lenses or excessive wobble during machine rotation. We could start such an investigation by constructing Ishikawa cause-and-effect diagrams.4-1 Pareto diagrams before and after corrective action was taken of the lens-coating solution had caused an increase in poorly coated lenses. to take actions which lead to better products. Finding the dominant cause of a defect.

4-21a2 lists the main factors that affect wobble during Size Over-experienced Content Training Small Workers Materials Large Size Experienced Central axle Under-experienced Material quality Uneven G axle bearing Material quality Threads Nuts Loose Tight Wobble Personality F cover Knowledge Axle hole Interval tm jus Measurement Measuring tool Judgment Errors Training Off-center en Off-center Cover hole G axle bearing cover F cover Common aberration Bench movement Group movement Position placement Bead removing Steel pipe scars G axle cover Inspector Metal drill Axle stop Experience F axle cover Punch width Judgment method Plating Uneven Inspection Tools (a) Cause-and-effect diagram for wobbling during machine rotation Impurity Flare Pipe making Roll Conveyor Bench movement Bunching Dropping Loosening Material Correcting Planing Water pressure test Ad t Inspection Bench rolling Movement Polishing Weight Arrange bench drop Polish Bunching Difference Movement Carriage Valve W eig ht Position placement Bench rolling Surface paint Wire (b) Cause-and-effect diagram for steel pipe scarring. because they resemble the skeleton of a ﬁsh. second revised edition (Tokyo: Asian Productivity Organization. Figure 5. Figure 5.There are several different methods for constructing cause-and-effect diagrams. We reproduce two such diagrams in Figure 5. Guide to Quality Control.80371 05 293-335 r2 sr 7/26/08 1:45 PM Page 326 326 Chapter 5 Statistical Process Control are also called ﬁshbone diagrams.] .4-2 Two cause-and-effect diagrams [From Ishikawa. depending on how the information is organized and presented.4-2. 1982).. K.

local governments.2-2. Today. why does machine wobble occur? One possible factor is the variability among material 1components2.4 defects per 1 million opportunities. we have to study and quantify their effects in more detail. Some authors talk about DMAICT. Such a diagram shows why excess wobble may occur. Analyze opportunity. Motorola views Six Sigma as a metric. So. Motorola realized very quickly that it could be applied to other business processes for the same purpose. Thus. At the heart of the methodology is the DMAIC model for process improvement. and as a management system. Using this scale. aligning key business processes to achieve those requirements. the diagram follows the production process and lists the components that may affect the scarring at the various stages of production. and Control performance. Successfully composed diagrams usually enhance communication among those who attempt to improve the process. Six Sigma equates to 3. utilizing rigorous data analysis to minimize variation in those processes. Well-designed experiments have to be conducted to quantify the cause-andeffect relationships. To Motorola. hospitals. Six Sigma has developed into a comprehensive business performance methodology. Six Sigma has become a business improvement methodology that guides an organization in understanding and managing customer requirements. Cause-and-effect diagrams are useful ﬁrst steps for recognizing the factors that may be of importance. However. DMAIC is commonly used by Six Sigma project teams and is an acronym for Deﬁne opportunity. and driving rapid and sustainable improvement of many business processes. and so on. While Six Sigma started as a defect reduction effort in manufacturing. Strategies for efﬁcient experimentation will be discussed in the next two chapters.4-4 Six Sigma Initiatives In Section 5. as a methodology. almost 25 years later. Machine wobble is thought to be a major cause of production defects. Improve performance.4 defects per 1 million opportunities and on counting defects in products and processes. The origin of Six Sigma as a statistically based method for reducing variation in electronic manufacturing processes started at Motorola around the mid-1980s. we introduced the Six Sigma concept and showed how it relates to a 3. Today. Why does dispersion in the G axle bearing occur? It could be due to variability in the size of the G axle bearing. after a set of possible factors has been identiﬁed. Six Sigma is all three at the same time. Measure performance. and the military.2-2. Six Sigma as a metric is discussed in Section 5. Why does dispersion in the materials occur? It could be because of the variability in the G axle bearing. 5. The second diagram lists the causes that may affect the scarring of steel pipes. Here. It is used all over the world. there has been less emphasis on the literal deﬁnition of 3. “G axle bearing” becomes a twig on the branch.80371 05 293-335 r2 sr 7/26/08 1:45 PM Page 327 Problem Solving 327 machine rotation. in diverse organizations such as large and small corporations. where T stands for Transferring best practice to other areas of the .4-parts-per-million defect level. Therefore. “size” becomes a twig on the twig. banks. “materials” is written on the diagram as a branch. Hence. and it helps organize and relate the factors that inﬂuence the wobble. As Six Sigma has evolved. as well as the factors that inﬂuence those main factors. it establishes a scale for levels of goodness and quality.

and Six Sigma became an in-house “branded” name for a performance improvement methodology. an improvement initiative2 that is focused on all processes necessary to meet customer expectations. Six Sigma becomes a performance-oriented system for executing business strategy. Six Sigma was effectively established as an industry in its own right. teaching the basic philosophy of quality improvement . Motorola engineers used Six Sigma as an informal name for an in-house initiative for reducing defects in production processes. Six Sigma processes are executed by Six Sigma Green Belts and Six Sigma Black Belts and are overseen by Six Sigma Master Black Belts. In the late 1980s. Then the company hires a Six Sigma trainer 1usually a statistician with experience in quality improvement methods who also understands the objectives of the ﬁrm2. People need to be trained in Six Sigma methods—especially the measurement and improvement tools 1i. notably in General Electric and other large manufacturing corporations. statistics and data-based problem solving2 and the communications> relationships skills that affect the information ﬂow among the various stakeholders of the organization. but also in organizations outside the manufacturing sector. the executive team has to decide on a strategy 1i. Act2. By 2000. In 1991.. Brief Outline of Six Sigma’s History In the mid-1980s. following the success of the original initiative. Motorola extended the Six Sigma methods to its critical business processes. which was a key paradigm during the Total Quality Management 1TQM2 movement of the late seventies and early eighties. Many excellent ﬁrms specialize in Six Sigma training. When an organization decides to implement the method. Most training programs use an approach something like this: Say a company wants 15 employees to acquire a good knowledge of a sound quality improvement program. Importance of Training Activities Although Six Sigma is capable of leveraging huge performance improvements and cost savings. The Sigma Belts are commonly seen as a parallel to the belts that exist in the martial arts. consultancy. Motorola has learned that the disciplined use of metrics and the application of the methodology are still not enough to drive desired breakthroughs and results that are sustainable over time. mobilizing teams to attack high-impact projects. Six Sigma had developed into a transferable “branded” corporate management initiative and methodology.e. helping organizations align their business strategy with critical improvement efforts. and implementation of Six Sigma methodology in organizations around the world. Motorola ensures that process metrics and Six Sigma’s structured methodology are applied to opportunities for improvement that are directly linked to the company’s organizational strategy.e. accelerating improved business results. General Electric was one of the ﬁrst large-scale adopters and advocates of Six Sigma and is considered by most experts to have been responsible for Six Sigma’s rapidly achieved high proﬁle. Motorola certiﬁed its ﬁrst “Black Belt” Six Sigma experts. Practiced as a management system. By the mid1990s. Do. The DMAIC model is a reﬁnement of the Deming–Shewhart PDCA wheel 1Plan. initiating an accredited training of Six Sigma methods. Check.80371 05 293-335 r2 sr 7/26/08 1:45 PM Page 328 328 Chapter 5 Statistical Process Control organization. Six Sigma starts at the top. For the greatest impact. involving the training. The trainer goes to the company for an initial week.. and governing efforts to ensure that improvements are sustained. none of this happens on its own.

and offering additional suggestions on the various projects. 122 Six Sigma is based on arbitrary standards. two classes of software support Six Sigma: analysis tools. this is an expensive training approach. giving further instruction on statistical methods and quality improvement.” 1This motto was made popular by the movie Jerry Maguire. but each company has its own ﬁgure2. providing lots of opportunities for charlatans. and further helpful suggestions on the selected improvement projects. Minitab 1which is used in this book2 has become an important analysis tool for Six Sigma. because good Six Sigma trainers are not cheap and 15 employees have given up at least four weeks of their normal work time. Six Sigma evolved as a quality initiative to reduce variance in the semiconductor industry and is about precision and accuracy. Lean Six Sigma New management technologies vie for the attention 1and the money2 of companies. Sometimes an employee who is participating earns a “Black Belt” by saving the company more than a certain amount per project 11 million dollars. used to perform statistical or process analysis. anyone can start up as a Six Sigma consultant. Seemingly. through their work on projects in need of improvement. and program management tools. The trainer then returns for a second week. while others may be tackled individually. the trainer leaves for about four weeks. respond to the corporate motto “Show me the money. Software for Six Sigma Generally. However. although 3. Jerry Maguire. 132 Many organizations and consulting ﬁrms of all sizes deliver Six Sigma training.4 defects per million might work well for certain products and processes. The training cycle may be repeated one more time. the savings produced by improvements on the various projects far exceed the costs of the Six Sigma training. Clearly. A recent development is the combination of Six Sigma with Lean production methods. noteworthy ﬁnancial results.2 Criticism of Six Sigma Advocates of Six Sigma believe that it can produce quick. How can companies afford to do this? In almost all cases 1General Electric is a good example2. used to manage and track a corporation’s entire Six Sigma program. Most CEOs are delighted with the program because trainer and employees. creating “Lean Six Sigma. for example. during which time the employees keep working on their projects. emphasizing the importance of the ultimate payback of various initiatives.80371 05 293-335 r2 sr 7/26/08 1:45 PM Page 329 Problem Solving 329 and statistical methods such as Shewhart control charts and principles of good experimental design 1discussed in the next two chapters2. and some companies that have embraced Six Sigma have done so poorly. there is also criticism: 112 Results are often exaggerated. At the end of the ﬁrst week. Then the trainer leaves for another four weeks. it may not be ideal for others. Lean . However. returning for a third week with some more training on statistics and quality improvement. Some of these projects may be studied by teams of employees. in which it was repeatedly used in a phone exchange between the athlete Rod Tidwell and his agent. with the trainer leaving for another four weeks before returning for a ﬁnal round of instruction. The trainer is aware of all of the projects and offers helpful suggestions. Even successful and established systems such as Six Sigma must market themselves to business.” Lean and Six Sigma are both process improvement methodologies. it is the company that selects the projects 1usually 10 or so2 that need improvement.

4 8.8 8. including the average number of pigs born alive per litter and the preweaning mortality. In addition.2 9.4 4.1 9.0 1.6 Percent Mortality 10.4 9.1 9. In this particular company.1 9.6 8.3 5. the female pigs 1sows2 of the breeding herd are artiﬁcially inseminated at carefully monitored times. called “services.3 6.3 9. Weekly records on the breeding herd are kept. Six Sigma programs have been more widely applied to service processes. about 200 inseminations.0 12. including many implementations in hospitals and other health care organizations.3 7.3 9.4 8.2 9.3 9.1 9.7 9.2 8.5 8.4 9.4 8.7 9.4 10.2 13.1 14.6 Heating1 Degree Days 3 0 0 0 0 0 0 0 0 9 10 72 45 41 48 75 118 196 201 241 203 226 234 336 Cooling1 Degree Days 26 121 80 88 84 114 119 71 105 49 21 5 5 0 9 2 0 0 0 0 0 0 0 0 Week 7>1>1995 7>8>1995 7>15>1995 7>22>1995 7>29>1995 8>5>1995 8>12>1995 8>19>1995 8>26>1995 9>2>1995 9>9>1995 9>16>1995 9>23>1995 9>30>1995 10>7>1995 10>14>1995 10>21>1995 10>28>1995 11>4>1995 11>11>1995 11>18>1995 11>25>1995 12>2>1995 12>9>1995 .6 10.3 9. Projects *Project 1 Pork producers keep detailed records on their operations. In large-scale breeding operations.7 12.1 9.8 6.” are carried out each week.0 9.5 14.6 9.7 9.6 9.6 15. An Iowa company is concerned about its breeding operation.2 11.9 9.80371 05 293-335 r2 sr 7/26/08 1:45 PM Page 333 Projects 333 quality activities into all areas of the organization and aligning those activities with the strategic goals of the ﬁrm.9 10.0 9.7 3. Weekly data for the period from July 1995 through June 1996 are listed as follows: Pigs Born per Litter 9.8 9.

4 9.3 9. use dot diagrams or histograms to display the variability of these two measures.5 9.8 6. and calculate the relevant summary statistics. a “cooling degree day” is the difference between the mean temperature and 65°F.9 9.3 9.1 9.5 9. Furthermore.4 4.1 9. *1a2 Use appropriate control charts 1in this case.2 8.5 9. and we graph individual observations against time.0 10.6 8.6 9.3 11.80371 05 293-335 r2 sr 7/26/08 1:45 PM Page 334 334 Chapter 5 Statistical Process Control Pigs Born per Litter 9. Note: The subgroup size is n = 1.3 5.6 11. individual observations charts and moving range charts2 to check whether the number of pigs born alive per litter and the preweaning mortality are under statistical control.3 10.7 9.9 9.4 11.3 9. The stability of the level of the series can be monitored by checking for trends in the time-sequence plot and by making sure that there are no unusual .2 10.4 9. Weekly heating and cooling degree days are obtained by summing the daily measures over the seven days of the week.5 9.9 2. Similarly.9 8.3 9.8 7.0 9.8 14.8 7.7 9.9 9.9 Heating1 Degree Days 261 280 301 309 270 346 466 323 255 215 247 330 171 209 199 160 130 107 96 125 61 74 15 64 29 2 0 0 Cooling1 Degree Days 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 4 0 0 0 1 17 26 0 1 43 61 71 Week 12>16>1995 12>23>1995 12>30>1995 1>6>1996 1>13>1996 1>20>1996 1>27>1996 2>3>1996 2>10>1996 2>17>1996 2>24>1996 3>2>1996 3>9>1996 3>16>1996 3>23>1996 3>30>1996 4>6>1996 4>13>1996 4>20>1996 4>27>1996 5>4>1996 5>11>1996 5>18>1996 5>25>1996 6>1>1996 6>8>1996 6>15>1996 6>22>1996 1 A “heating degree day” for a given day is deﬁned as the difference between the “balance point” temperature of 65°F (above which a building is assumed not to need any heating) and the mean temperature on that day.4 3.5 9.1 9.8 10.0 9.2 Percent Mortality 12.0 11. it is zero for days with mean temperature below 65°F.0 9.3 9.3 12.7 9.4 9.0 9.3 8.0 3.3 9. It is zero for days with mean temperature above 65°F.1 9.1 9.1 8.3 12.

normal with mean 2. Assume that the lower and upper speciﬁcation limits are -10 and +10. Interpret your ﬁndings. Consider the following processes: i. Use simulation. The stability of the variability of individual observations can be monitored with a moving range chart. 1a2 For each process.2 Both types of chart can be obtained with the Minitab command “Stat 7 Control Charts 7 Variables Charts for Individuals 1I-MR charts2.2.” 1Consult books on statistical process control or search the Web to learn more about moving range 1MR2 charts. iii. uniform on the interval from -12. uniform on the interval from -10 to +10.5 and standard deviation 2.5. normal with mean 0 and standard deviation 5. normal with mean 0 and standard deviation 2. v.80371 05 293-335 r2 sr 7/26/08 1:45 PM Page 335 Projects 335 observations.5 to +12. respectively. Repeat this 1.000 times and display the sampling distribution of the capability indices. Breeding barns are not air-conditioned in the summer and they are poorly heated during the winter.5. Discuss whether the mortality rate is affected by the weather. among them Cp and Cpk. iv.5. 1A “moving” range of individual observations is the absolute value of the difference between adjacent observations. Project 2 Capability indices were discussed in Section 5. calculate the proportion of nonconforming items and determine the capability indices Cp and Cpk.2 *1b2 The data include information on weekly heating and cooling degree days for Des Moines. Iowa. . 1b2 Study the behavior of these indices in samples of sizes n = 50 and n = 100. Draw samples of sizes n = 50 and n = 100 from the processes studied and calculate Cp and Cpk. ii.