You are on page 1of 9

This article was downloaded by: [ ] On: 25 October 2012, At: 12:17 Publisher: Taylor & Francis Informa

Ltd Registered in England and Wales Registered Number: 1072954 Registered office: Mortimer House, 37-41 Mortimer Street, London W1T 3JH, UK

IIE Transactions
Publication details, including instructions for authors and subscription information: http://www.tandfonline.com/loi/uiie20

On constructing T monitoring
a b

control charts for on-line process


b

GUNABUSHANAM NEDUMARAN & JOSEPH J. PIGNATIELLO JR.

Oracle Corporation, 1710 Valley View Lane, Irving, TX, 75061, USA E-mail:

FAMU-FSU College of Engineering, Florida State University, Florida A & M University, Tallahassee, FL, 32310-6046, USA E-mail: Version of record first published: 27 Jul 2007.

To cite this article: GUNABUSHANAM NEDUMARAN & JOSEPH J. PIGNATIELLO JR. (1999): On constructing T control charts for on-line process monitoring, IIE Transactions, 31:6, 529-536 To link to this article: http://dx.doi.org/10.1080/07408179908969856

PLEASE SCROLL DOWN FOR ARTICLE Full terms and conditions of use: http://www.tandfonline.com/page/terms-and-conditions This article may be used for research, teaching, and private study purposes. Any substantial or systematic reproduction, redistribution, reselling, loan, sub-licensing, systematic supply, or distribution in any form to anyone is expressly forbidden. The publisher does not give any warranty express or implied or make any representation that the contents will be complete or accurate or up to date. The accuracy of any instructions, formulae, and drug doses should be independently verified with primary sources. The publisher shall not be liable for any loss, actions, claims, proceedings, demand, or costs or damages whatsoever or howsoever caused arising directly or indirectly in connection with or arising out of the use of this material.

IIE Transactions (1 999) 31, 529-536

On constructing T~ control charts for on-line process

GUNABUSHANAM NEDUMARAN' and JOSEPH J. PIGNATIELLO JRA2


Oracle Corporation, 1710 Valley View Lane, Irving, T X 75061, USA E-mail: gnedumar @us.oracle.com *FAMU-FSU College of Engineering. Florida State University, Florida A & M University. Tallahassee, FL 32310-6046, USA E-mail: pigna@engzfsu.edu

'

Received May 1997 and accepted December 1997

Downloaded by [ ] at 12:17 25 October 2012

X2 control limits on T~ control charts during on-line process monitoring. We make recommendations on the minimum number of subgroups necessary for control charts based on estimated parameters to perform similar to control charts based on true parameters when the X2 control limit is used. We discuss an exact procedure for constructing T~ control charts using estimated parameters that perform similar to control charts based on true parameters ;egardless of the number of preliminary subgroups. Furthermore, we suggest an implementation approach for using lilt's control limit on r2 control charts un ti1 the accumulation of recommended minimum number of subgroups.

In this paper we study the performance of

variate X control charts. He showed that when control limits are constructed using parameters estimated with Statistical Process Control (SPC) charts are tools that data from a small number of subgroups, false alarm rates are used to monitor the state of a process by distin- and Average Run Lengths (ARLs,) of the control chart guishing between common causes and special causes of increased from their nominal values that are based on variability. When several characteristics of a manufac- true parameter values. Quesenbermry [2] further showed tured component are to be monitored simultaneously, that m should be at least 100 when n = 5 in order for the multivariate Shewhart-type X2 o r T~ control charts can estimated control limits to perform similar to true control be used [I]. As long as the points plotted on the X 2 or T~ limits with respect to the in-control run length distribucontrol chart fall below the Upper Control Limit (UCL) tion and the ARL. of the chart, the process is assumed to operate under a In this paper, we extend Quesenberry's [2] study 1.0 stable system of common causes, and hence, in a state of multivariate T~ control charts. In particular, we consider control. When one or more points exceed the UCL, the the issue of the minimum number of subgroups necessary process is deemed out of control due to one or more for the control chart constructed using estimated paramspecial causes and an investigation is carried out to eters to perform similar to the control chart constructc:d using true parameters during the on-line monitoring stage. detect these special causes. When the in-control values of the process mean vector Further, we suggest an implementation procedure so that and covariance matrix are known, these known parame- on-line monitoring with T~ control charts can begin at the ter values are used to find the statistic plotted on a X 2 crucial start-up stages of the process. control chart. The UCL of this chart is based on the chiThis paper is organized as follows. In the next section square distribution. However, in most cases these pa- we describe the multivariate process model and review rameter values are unknown, and hence are estimated some pertinent literature. In the following section, we from some m initial subgroups of size n taken when the present the results of some simul;3tion experiments pe,rprocess is believed to be stable. When a future subgroup formed along with our recommendatio~l for the minimum is drawn from the process, Hotelling's T~ statistic is cal- number of subgroups necessary for the control cha-rt culated using parameter estimates and is plotted on a T~ based on parameter estimates to perform similar to the control chart. The UCL of this chart is based on the F chart based on true parameters. We then describe an exact distribution. method and an implementation p~:ocedurefor constructQuesenberry [2] has studied the effects of using esti- ing control charts based on parameter estimates that mated parameters to construct the control limits of uni- perform similar to the chart based on true parameters.

1. Introduction

0740-8 17X

O 1999 " I I E

Nedumaran and Pignatiello Jr .


2. Multivariate process model and literature review We let Xij = (Xvl,Xi,2,.. . ,+)' denote a p x I vector that represents the p observatsons on the jth component in the ith subgroup, i = 1,2,. . . and j = 1,2,. . . ,n. We assume that Xij's are independent and identically distributed normal random variables with mean p and co2 when the process is in control. We let variance matrix I X i denote the average vector for the ith subgroup, and we let Si denote the unbiased estimate of the covariance matrix for the ith subgroup. That is, This statistic is plotted on a T~ control chart. Alt 13) has given the UCL of this control chart as

where Ct(m', n , p ) =
p(ml

+ I)(n - 1)
1)'

(m'n - m t - p +

F",,,, is the ( 1 - a)th percentile point of the F distri bution with v l and vz degrees of freedom, and a is the specified acceptable false alarm probability for each subgroup plotted on the T~ control chart. If the process parameters are estimated from a reasonably large number of initial subgroups, the usual practice for.constructing T~ control charts is to use UCLzl instead of the exact UCLp. However, according to Montgomery [ l ] "we must be careful" when following this practice. In this paper, we show that X2 control limits issue a relatively large number of false alarms even when the process parameters are estimated from relatively large number of subgroups. Lowry and Montgomery [4] have compared the numerical values of UCL,2 and UCLp for several values of m, n, and p, with a = 0.001. T o select the minimum m for which UCLp can be replaced by the approximation UCL,,, they used the relative error between the two control limits as the criterion. They have recommended minimum rn values for several combinations of p and n for which the relative error is less than 0.1. However, they did not consider the average run length or the run length distribution performance of the respective control charts. Quesenberry [2j has addressed the issue of the minimum number of initial subgroups necessary while using univariate control charts in order for the chart based on estimated parameters to perform in a similar manner to the chart based on true parameters during the-on-line process monitoring stage. Let B, denote the event that the ith future subgroup average (i > m) plots outside the control limits when the univariate normal process parameters are estimated from m initial subgroups. Quesenberry [2] showed that the events Biand Bj(i, J > m ) are positively cbrrelated. This dependence causes both the A R L and the Standard Deviation of the Run Length (denoted by SDRL) to increase from their nominal values, despite the increase in false alarm rates. Quesenberry [2] used simulation to study the ARL performance of the univariate X chart for several values of m, with n = 5, to find the minimum value of m for which the estimated control limits perform like the true, but unknown, control limits. His simulation results showed that for small m (i.e., m < 50) both the A RL and the SDRL exceeded their nominal values. For instance, when m = 20 and n = 5, the ARL increased from 370.4 (for true parameters) to 434, and the corresponding SDRL increased from 369.9 to 754 for the in-control process. Both the A R L and the SDRL got reasonably

and

Downloaded by [ ] at 12:17 25 October 2012

When the process is in control and the in-control process parameter values are known, the statistic plotted on the X2 control chart for the ith subgroup is

When the process is in control this statistic has a chisquare distribution with p degrees of freedom [l]. It is plotted on a X 2 control chart with a UCL given by

where is the ( 1 - a)th percentile point of chi-square distribut~on with p degrees of freedom and a is the probability of a false alarm for each subgroup plotted on the x2 control chart. If the process parameter values are not known, data from m initial subgroups are collected when the process is believed to be in control. Then, pooling data from these m subgroups and assuming that the process was in control, unbiased estimates of the mean vector and the covariance matrix are given by

respectively. A T~ control chart is constructed and the m initial subgroups are tested retrospectively to ensure that the process was in control when these initial subgroups were drawn. In this paper, we consider the case where the process is being monitored on-line and where the in-control values of the process parameters are estimated from m initial subgroups. When a future subgroup (i > m) is drawn from the process for on-line monitoring purposes, the statistic plotted on the control chart is

Constructing

75 control charts
q

Downloaded by [ ] at 12:17 25 October 2012

close to their nominal values for the univariate X chart when m was at least 100. T o study the efkct of m on the A-control Run Length (RL) distribution for the univariate X chart, Quesenberry [2] used Monte Carlo simulation to estimate Pr[RL 5 x ] for several values of x when the control limits are estimated from m subgroups. The estimated probabilities exceeded the nominal values for small m, and get reasonably close to the nominal values when m is at least 100. For instance, when m = 20 and n = 5, Pr[RL 5 201 increased from 0.0526 (for true parameters) to 0.073 for the in-control process. Based on these two considerations, namely, the ARL and the in-control RL distribution, Quesenberry 121 concluded that m should be at least 100 for n = 5 in order for the estimated X control chart limits to perform like the true limits. For other subgroup sizes, he conjectured that m should be at least 400/(n - 1) based on the speculation that the degrees of freedom of the variance estimator should be approximately 400 for this to work. In the next section, we study the performance of T~ control charts with X2 control limits. We compare the performance of this chart with the case where the process parameter values are known. We call the latter the "true parameters case" and refer to the characteristics of this chart as the "nominal" characteristics. We recommend a minimum m for which the control chart based on estiperforms similar to the control chart mated based on true parameters. We further suggest a control chart implementation program that can be used until the accumulation of the recommended minimum number of subgroups.

ARL = -

1 -j'

and

SDRL = fl - JG:(ARL-

1-fl

I),

where

/? = 1 - Pr(Ai) = ~ r ( d < LICL;,).


However, when parameter estimates from the m initial subgroups are used this run length characterization is are correlated for i # j and not valid since q2 and i,j > m. T o verify the last statement, we consider -cov(Xi - X, X, - h). It can be shown by direct evaluation that

q2

Let Bi(i > m) denote the event that q2 exceeds UCLZ. Events Bi and B, are not independent for i # j, i,j > m since and are correlated. Consequently, the sequence of trials comparing q2 with UCL? is not a sequence of Bernoulli trials, and hence, the run length until the occurrence of event Bi does not follow a geomet~nic distribution. As suggested by Quesenberry I:!], one reasonable way to decide when m is large enough for the control cha-rt based on estimated parameters to behave essentially similar to the control chart based on true parameters is to determine when the run length distribution between events Bi is close to that of Ai.Further, note from (6) that the ARL and the SDRL are app~:oximatelyequal when the true process parameters are known and when / I is 3. Characterizing the run length distribution relatively close to 1. To study the effect of m on t h in-control ~ run length We suppose that data on m initial subgroups of size n distribution when n = 5, we simlilated 10 000 observafrom a stable N p ( p , Z ) distribution are availabk. The tions from the run length distribution for several values of process mean vector and covariance matrix are esti- m and p. We considered three .process dimensions of mated using (3). Future subgroups are drawn from the p = 3, 6 and 10. We selected a = 0,0027, the value usually process at regular intervals for on-line monitoring used in univariate X control charts. purposes. Suppose that there has been a step-change in For each combination of p and m, we generated m the process mean from the in-control value of p = CI, to subgroups from a stable multivariate normal distribution. p = p, where p, # k. We let Ai denote the event thatv The process parameters were estimated based on these m the control chart statistic X: of a future subgroup in-control initial subgroups. Then, further subgroups (i > m) exceeds UCL,I. We let Y denote the run length, were generated from the same stable normal distribution which is the number of points plotted on the chart and the T~ statistic was calculated for each subgroup. until a point exceeds the UCL. If the process parameter This statistic was plotted on a co,ntrol chart with UCL,,I values are known and the X 2 statistic (1) is plotted on a until a signal was issued. The number of subgroups control chart with control limit (2), then the sequence drawn until a signal was given constituted one observaof trials that consist of determining if Xf exceeds UCLx2 tion from the run length distribution. This procedure was would be a sequence of Bernoulli trials. Consequently, replicated 10 000 times and an empirical run length disthe run length Y until event A i occurs wouId have a tribution was found. These empirical run length distrigeometric distribution with mean A RL and standard butions are shown in Figs. 1-3 along with the nominal deviation SDRL given by [2]: run length distribution (denoted by m = inf.). The curves

q2

532

Nedumaran and Pignatiello Jr .

Fig. 1. Run length empirical distribution for p = 3, n = 5 , a = 0.0027:

Fig. 3, R u n length empirical distribution for p = 10, or = 0.0027.

n = 5,

Downloaded by [ ] at 12:17 25 October 2012

of m. Three process dimensions of p = 3, 6 and 10 were considered. Both in-control and out-of-control situations . were considered. The false alarm probability for each subgroup was set at 0.0027. For each combination of p, m and shift magnitude I, m subgroups were generated from a stable multivariate normal distribution and the process parameters were estimated. Starting with subgroup m + 1, the process mean was changed from ~ r to , p = p, by introducing a shift of magnitude 2 where

Fie. 2. Run length e m ~ i r i c a ldistribution for p = 6, n = 5 , a = 0.0027.

appear to have non-differentiable points due to the scaling of the k-axis. To study further the effect of m when n = 5, values of t h e A R L a n d theSDRLweresimulatedforseveralvalues
Tnbtc I . ARL and SDRL for p = 3, n = 5 , or = 0.0027

Subgroups were generated from the out-of-control process and the T~ statistic (4) was calculated for each subgroup using estimated parameters. This statistic was plotted on a control ,-hart with (ICLZ2until a signal was issued. The number of subgroups until a signal was given constituted one observation from the run length distribution. This procedure was replicated 10 000 times and the ARL and the SDRL were obtained. The results are showngin Tables 1-3. Several observations can be made from Figs. 1-3 and Tables 1-3. When the process parameyers are estimated

ARL

SDRL

ARL

SDRL

ARL

SDRL

ARL

SDRL

ARL

SDRL

Constructing

control charts
=

533
0.0027
1.0 1.5

Table 2. ARL and SDRL for p = 6, n = 5, a


0.0

0.5

SDK'L
ARL

2.0
SDRL

ARL

SDRL

ARL

SDRL

ARL

SDRL

ARL

Table 3. ARL and SDRL for p = 10, n = 5, a = 0.0027


0.0 0.5

Downloaded by [ ] at 12:17 25 October 2012

l .O

1.5

SDRL ARL

2.0

ARL

SDRL

ARL

SDRL

ARL

SDRL

ARL

SDRL

from a small number of subgroups, there is an increase in that the Lowry and Montgomery [4] recommended minthe rate of false alarms for the in-control process and the imum rn would result in a large number of false alarms ARLs are shorter for both in-control and out-of-control and very short ARLs. In general, we speculate that a minimum m in the range processes. Also, for small values of m, the SDRLs are larger than the A RLs, whereas they are approximately from 800p/3(n- I ) to 400p/(n- 1) is necessary for the equal for the case where the true process parameters are control chart based on estimated parameters to behave known. As m increases, both the run length distribution like the chart based on true parameters. and the ARLs approach their nominal values. T o test this conjecture, we considered two subgroup Following Quesenberry's [2] approach, we make the sizes of n = 4 and 10 when p = 8 and a = 0.0027.For (,he m is in the range following recommendations for the minimum value of m n = 4 case, our recommended mir~imum necessary for the control chart based on estimated pa- 700-1000,and for the n = 10 cast:, it is in the range 300rameters to behave similar to the control chart based on 450. We selected m values in the middle of each recomtrue parameters. For p = 3 and n = 5, we recommend mended range for validation purposes. That is, for lhe m = 200. For p = 6 and n = 5 , a minimum of m = 400 is n = 4 case we selected m = 900,iind for the n = 10 c;ise recommended. For p = 10 and n = 5, we recommend we selected m = 400. The criterion used by Lowry and m = 600. By contrast, the criterion used by Lowry and Montgomery [4] would lead to the following recommenMontgomery [4], i.e., maintaining relative error of less dations for minimum m for the case where u = 0.0027. than 0.1 between the x2 control limit and Alt's control The Lowry and Montgomery [4] recommended minimum limit, would lead to substantially lower recommendations m would be 70 and 30 for n = 4 .and 10, respectively. for the minimum m for the case where n = 5 and For each case we found the in-control run length emu = 0.0027. In particular, their recommended minimum pirical distribution and the in-control and out-of-control value of m would be 40, 50 and 70 for p = 3, 6 and 10, ARL and SDRL following the: procedures described respectively. It can be seen from Figs. 1-3 and Tables 1-3 above. The results presented in Ta.bIes 4 and 5 and Figs,. 4

534
Tnblc 4. Validation test: ARL and SDRL for p = 8, n = 4,CI = 0.0027 m 1
0.0
0.5
1.O

Nedumaran and Pignut iello Jr .

1.5

2.0

ARL

SDRL

ARL

SD RL

ARL

SDRL

ARL

SDRL

ARL

SDRL

TnbIe 5. Validation test: A RL and SDRL for p = 8, n = 10, a = 0.0027


0.0

0.5

1.0

1.5

2.0

ARL

SDRL

ARL

SDRL

ARL

SDRL

ARL

SDRL

ARL

SDRL

Downloaded by [ ] at 12:17 25 October 2012

and 5 show that our recommended minimum m provides adequate performance, whereas the Lowry and Montgomery [4] recommended minimum m would result in a large number of false alarms. A number of additional computer runs were carried out for different values of n and p to test our speculation on minimum m. The results (not reported here) suggest that the speculation is reasonable for other values of n and p.

4. Chart implementation during start-up stage


Fig. 4. Validation test: run length empirical distribution for p = 8, n = 4,a = 0.0027.

In this section, we first describe an exact method for constructing uCL using estimated parameters that performs similar to the chart based on true parameters with respect to specific percentile points of the in-control run length distribution. Next, we describe how Alt's [3] control limits can be used instead of the exact method. We suppose that data from m initial in-control subgroups are available. We let % and (i,j > m) denote two future subgroups. Then,

x,

and

Fig. 5. Validation test: run length empirical distribution for p = 8, n = 10, a = 0.0027.

In the exact approach, control limits are constructed for a specified number (k) of future subgroups. That is, we establish control limits to be used for monitoring the

Constructing

control charts

535
ulated 10 000 observations from the run length distribution for several values of m and p with n = 5. We considered three process dimensions of p = 3, 6 and 10. We selected a = 0.0027, the value usually used in univariate X control charts. For each combination of p and m, we generated m subgroups of size n = 5 from a stable multivariate normal distribution. The process paramet.ers were estimated based on these rn in-control initial subgroups. Then, further subgroups were generated from the same stable normal distribution and the T* statistic was calculated for each subgroup. This statistic was plotted on a control chart with an uppe;i control limit of UCLr> until a signal was issued. The nu-mber of subgroups until a signal was given constituted one observation from the run length distribution. This procedure was replicated 10 000 times and an empirical run length distribution was found. These empirical run 1t:ngth distributions are shown in Figs. 6-8 along with the nominal run length distribution (denoted by m = inf.). Figures 6-8 show that when A:it's control limit is based on 20 subgroups, the empirical run lengih distribution

process for the next k subgroups. The value of k may be chosen to represent the daily or weekly production volume. The T~ statistic plotted for the ith subgroup is given by

Let

4 be the desired probability such that

We let UCLE denote the exact upper control limit that satisfies (7). Then, That is where

P~LT;,,
1

5 U C L ~= ] 1 -#I,

~i~~ = I I I ~ X {n(Xi - P'3-l(Xi - F) },

Downloaded by [ ] at 12:17 25 October 2012

Here, T : , , is a %,,,,-type statistic discussed by Siotani [5] with Siotani's y = (m + 1)lmn and 6 = llmn. Thus, UCLE is the upper (1 - 4)th percentile point of the distri bution of For the control chart based on estimated control limits to perform in a similar manner to the control chart based on true parameters, 4 is set equal to the run length distribution of the true parameters chart. That is,

4 = Pr[RL 5 k] = 1 - (1 - a)',
where a is the probability of a false alarm for one subgroup. Siotani [5] pointed out that the sampling distribution of ?,,,,,-type statistic is "extremely difficult" to find. Siotani 1:5] suggested a two-s tage procedure for finding approximations to the upper percentile points of the distribution of which was investigated further by Seo and Siotani [6]. However, this two-stage procedure involves extensive computations and gives only approximate percentile points. As an alternative, we suggest an implementation procedure for using the simpler control limit as given by Alt [3]. Alt's control limit (5) is exact for any given future subgroup. However, it is not exact when several future subgroups are plotted on a chart with the same control limit sinceit - ignores the dependence between (Xi - Sf) and (ft, - x). Hence, one implementation method would be to update the control limit after every future subgroup plotted on the chart. But, such an approach is not recommended since it involves extensive computations for finding the control limit for each future subgroup. More importantly, such an approach will have a low power in detecting gradual mean shifts since the estimated process mean is updated after every subgroup. T o study the effect of m on the in-control run length distribution when Alt's control limit (5) is used, we sim-

Fig. 6. Empirical run length distl-ibution for p = 3, n := 5 ,


or = 0.0027.

Fig. 7 . Empirical run length dist:iibution for p = 6, n = 5 ,


a = 0.0027.

536

Nedumaran and Pignatiello Jr .


similar to control charts based on true parameters when X2 control limit i s used. We discussed an exact procedure for constructing T~ control charts based on estimated parameters that perform similar to control charts based on true parameters regardless of the number of preliminary subgroups. Further, we presented an implementation approach for using Alt's control limit on T2control charts until the accumulation of recommended minimum number of subgroups. Finally, it should be noted that the false alarm rates discussed in this paper are the rates for the entire charting program, and not for individual charts as is discussed by Quesenberry [2J.
Acknowledgments

Fig. 8. Empirical run length distribution for p = 10, n = 5, = 0.0027.

differs from the nominal run length distribution for a relatively large number of future subgroups, k > 150. It is essentially identical to the nominal run length distribution for a small number of future subgroups, k 5 50. Further, the empirical run length distribution approaches the nominal run length distribution for m 2 50. For the control chart with Alt's control limit to perform in a similar manner to the control chart based on true parameters, we propose the following implementation approach. When m < 50, construct the control chart for small k, e.g., k = 20 future subgroups, and update the coutrol limit after every k future subgroups. When process parameters are estimated based on at least 50 subgroups ( m 2 50), the resulting control limit can then be used for a large number of future subgroups. In fact, this control limit can be used until the recommended minimum number of subgroups, i.e., 800p/3(n - 1 ) to ,400pln - I ) , is accumulated, after which a permanent X2 control limit can be established. The implementation approach that we have described is based on the use of the T 2statistic. We showed that the false alarm rates increase when standard X 2 control limits are used on T~ control charts unless process parameters arc estimated from a relatively large number of subgroups. Our proposed implementation approach, which is based on Alt's control limit, can help alleviate this problem. The determination of Alt's control limit is only slightly inore difficult than that of the X 2 control limit. It should be pointed out that an alternative to the entire implementation approach described here is the use of Q charts developed by Quesenberry [7]. 5. Summary

This material is based in part upon the work supported by the Texas Advanced Research Program under Grant NO. 999903-120.

Downloaded by [ ] at 12:17 25 October 2012

References
[ I ] Montgomery, D.C. (1 996) Introduction to Statistical Quality Conrrol, John Wiley & Sons, New York, NY, pp. 362-367. [2] Quesenberry, C.P. (1993) The effect of sample size on estimated limits for and X control charts. Journal of Quality Technology, 25, 237-247. [3j Alt, F.B. (1976) Small sample probability limits for the mean of a mu1tivariate normal process in ASQC Technical Conference Transactions, pp. 1 70-1 76. [4] Lowry, C.A. and Montgomery, D.C. (1995) A review of multivariate control charts. IIE Transactions, 27, 800-810. [5] Siotani, M. (1959) The exact value of the generalized distances of the individual points in the multivariate normal sample. Annuls o f the Institute o f Starktical Mathematics, 10, 183-203. [6] Seo, T. and Siotani, S. (1993) Approximations to the upper percen tiles of T:,,-type statistics, in Statistical Science & Data Analysis, Matsusita, K . et al., (eds), VSP, Tokyo. [7] Quesenberry, C.P. (1 997) SPC Merho& for Qualisy Improvement, John Wiley & Sons, New York, NY.

Biographies
Dr. Gunabushanam Nedumaran is a Senior Consultant with Oracle Corporation. His research interests are in the area of applied statistics and statistical process control. He is an ASQ certified Quality Engineer.
Dr Joseph J. Pignatiello, Jr is an Associate Professor in the Department of Industrial Engineering in the FAMU-FSU College of Engineering at Florida State University and Florida A & M University. Dr Pignatiello's interests are in quality engineering and include statistical process control, process capability, design and analysis of experiments, robust design and engineering statistics. Dr Pignatiello won the 1994 Shewell Award from ASQC's Chemical and Process Industries Division and the 1994 Craig Award from ASQC's Automotive Division for his research on process capability. In 1990 and 1991 he won awards from the Ellis R. Ott Foundation for his research on SPC and the methods of Taguchi. He serves on the editorial boards of IIE Transacrions, Journal of Quality Technology. Quality Engineering and the In fernational Journal of Industrial Engineering.
Conrribured by the On-line Quaiily Engineering Deparrment.

In this paper we studied the performance of X 2 control limits on T~ control charts. We made recommendations on the minimum number of subgroups necessary for control charts based on estimated parameters to perform

You might also like