Journal of Loss Prevention in the Process Industries 16 (2003) 918 www.elsevier.

com/locate/jlp

Economic-statistical design of multivariate control charts for monitoring the mean vector and covariance matrix
C.-Y. Chou a,, C.-H. Chen b, H.-R. Liu c, X.-R. Huang a
a

Department of Industrial Engineering and Management, National Yunlin University of Science and Technology, Touliu 640, Taiwan, ROC b Department of Industrial Management, Southern Taiwan University of Technology, Yung-Kang 710, Taiwan, ROC c Department of Food and Nutrition, Hung-Kuang Institute of Technology, Shalu 433, Taiwan, ROC

Abstract When a control chart is used to monitor a process, three test parameters should be determined: the sample size, the sampling interval between successive samples, and the control limits or critical region of the chart. In this paper, we present the procedure to conduct the economic-statistical design of multivariate control charts for monitoring the process mean vector and covariance matrix simultaneously; i.e. to economically determine the optimum values of the three test parameters such that the statistical constraints (including the type I error probability and power) of the control chart can be satised. The test statistic 2nL is applied to develop this procedure and the cost model is established based on the function given by Montgomery and Klatt. A numerical example is provided to illustrate the solution procedure of the design and then the effects of cost parameters on the optimal design are examined. 2003 Elsevier Science Ltd. All rights reserved.
Keywords: Quality control; Control chart; Economic-statistical design; Mean vector; Covariance matrix

1. Introduction Statistical process control is an effective approach for improving product quality and saving production costs for a rm. Since 1924, when Dr Shewhart presented the rst control chart, various control chart techniques have been developed and widely applied as a primary tool in statistical process control. The major function of control charting is to detect the occurrence of assignable causes so that the necessary corrective action can be taken before a large quantity of nonconforming product is manufactured. The control charting technique may be considered as the graphical expression and operation of statistical hypothesis testing. When a control chart is used to monitor a process, three test parameters should be determined: the sample size, the sampling interval between successive samples, and the control limits or critical region of the chart. Duncan (1956) proposed the rst economic model for determining the three test parameters for the X-bar con-

Corresponding author. E-mail address: choucy@pine.yuntech.edu.tw (C.-Y. Chou).

trol charts that minimizes the average cost when a single out-of-control state (assignable cause) exists. Duncans cost model includes the cost of sampling and inspection, the cost of defective products, the cost of false alarm, the cost of searching assignable cause, and the cost of process correction. Since then, considerable attention has been devoted to the optimal economic determination of the three parameters of X-bar charts (Duncan, 1971; Gibra, 1971; Goel, Jain, & Wu, 1968; Knappenberger & Grandage, 1969). Montgomery (1980) gave a thorough review of the literature in economic designs of various control charts. A bibliography of related papers is also available in works by Ho and Case (1994) and Vance (1983). Since the solution from an economic design of control charts may result in poor statistical properties, Saniga (1989) presented the economic-statistical design for a control chart in which the cost function is minimized subject to the constrained minimum value of power and maximum value of the type I error probability. Although many efforts have been put on economic designs of control charts that measure single characteristic, some industrial products and processes are characterized by two or more measurable characteristics, and their joint effect describes product quality. For example,

0950-4230/03/$ - see front matter 2003 Elsevier Science Ltd. All rights reserved. doi:10.1016/S0950-4230(02)00089-X

10

C.-Y. Chou et al. / Journal of Loss Prevention in the Process Industries 16 (2003) 918

in the production of synthetic ber, the tensile strength and diameter may be equally important quality characteristics. These characteristics are jointly distributed random variables and cannot appropriately be controlled by independently applying a control chart to each variable. Some authors (Ghare & Torgersen, 1968; Hotelling, 1947; Jackson, 1956) have developed quality control procedures for several related random variables. Among these procedures, Hotelling T2 control chart is probably the most widely known. Montgomery and Klatt (1972) presented a cost model to economically design the T2 control charts. Chen (1995) conducted the economicstatistical design of the T2 control charts by adding the statistical constraints to the design procedure. Since the T2 control charts monitor only the process mean vector and the process covariance matrix may also have an impact on product quality, in this paper we apply the test statistic 2nL (which will be reviewed in Section 3) to develop the economic-statistical design of the multivariate control charts for monitoring the process mean vector and covariance matrix simultaneously. The cost function used in this paper is based on the model given by Saniga (1977) and Montgomery and Klatt (1972). An example is provided to illustrate the solution procedure of the design and then some sensitivity analyses are conducted to investigate the effects of cost parameters on the solution of the design.

5. The time the process remains in the in-control state before going out of control is assumed to follow an exponential distribution with mean l1 hours. 6. When the process goes out of control, it stays out of control until detected and corrected. 7. During each sampling interval, there exists at most one assignable cause which makes the process out of control. The assignable cause will not occur at sampling time. 8. When the multivariate control chart indicates that the process is out of control, the process is stopped for investigating the assignable cause. 9. The cost for investigating real and false alarms is the same. 10.Considering the statistical properties of the multivariate control chart, the upper bound of the type I error probability is set to be 0.1 and the lower bounds of the powers for the two out-of-control states are all set to be 0.9.

3. The test statistic Suppose that the output of a process can be described by p quality characteristics, and Y is a p 1 random vector whose jth element is the jth quality characteristic and is multivariate normally distributed. Let E(Y) m be the p 1 mean vector of the characteristics and Cov(Y) be the p p covariance matrix of Y. Generally, m and are unknown. For a random sample of size n from Y, say Y1, Y2, , Yn, the sample mean vector and sample covariance matrix may be computed by Y S 1 Y, ni 1 i 1 )(YiY )T , (Y Y n1i 1 i

2. Model assumptions To simplify the mathematical manipulation and analysis of the control chart, the following assumptions are made: 1. The quality of the process can be described by the mean vector and covariance matrix of p characteristics and is monitored by a multivariate control chart using the test statistic 2nL. 2. The p quality characteristics monitored by the multivariate control chart follow a multivariate normal distribution with mean vector m and covariance matrix . 3. In the beginning of the process, the process is assumed to be in control; that is, m m0 and 0, where m0 is a given vector and 0 is a given positive denite matrix. 4. There are two out-of-control states caused by two assignable causes, respectively. After the rst assignable cause occurs, the process mean vector shifts to m1 m0 m and the process covariance matrix remains unchanged, where the p 1 vector m is known. After the second assignable cause occurs, the process covariance matrix shifts to 1 0 and the process mean vector remains unchanged, where the p p matrix is known.

(1)

(2)

where superscript T denotes the transpose operation. The likelihood ratio criterion of testing the hypothesis H0: m m0 and 0 against alternatives H1: m m0 or 0 may be expressed as (see Anderson, 1958) L

e n

np/2

1 1 n/2 1 |(n1)S exp Tr 0 | 0 [(n1)S 2

m0)(Y m0)T] , n(Y where Tr denotes the trace operation of a matrix. The value of L is between 0 and 1. If L ca, the null hypothesis H0 is rejected, where ca is the lower 100ath percentile of the distribution of L. However, under the null hypothesis, the exact distribution of L is unknown. Therefore, statisticians (Davis, 1971; Nagarsenker & Pil-

(3)

C.-Y. Chou et al. / Journal of Loss Prevention in the Process Industries 16 (2003) 918

11

lai, 1974; Sugiura, 1969) transform L to the statistic 2nL, which may be expressed as a chi-square series, as the test criterion by obtaining its asymptotic distribution. Thus, the null hypothesis H0 should be rejected when 2nL 2nca UCL, where UCL is the upper 100 ath percentile of the distribution of 2nL, and is the upper control limit (UCL) of the multivariate control chart in this paper, which indicates that the assignable cause may exist in the process. To consider the statistical constraints (i.e. the type I error probability and the power) of the chart, the distribution functions of 2nL under the null and alternative hypotheses should be evaluated. 3.1. The distribution function of 2nL under H0 According to Sugiura (1969), the distribution function of 2nL under the null hypothesis, which is related to the type I error probability of the multivariate control chart, may be obtained through the following ve steps: Step 1 The characteristic function of 2nL under the null hypothesis is f(t) E[eit(2nL)] n(12it)g 2 ng (12it)np(12it)/2p g=1 2 (2e / n)npitp g=1

yields the characteristic function of 2nL, i.e. f(t) enf(t), which can be expressed as the summation of a chi-square series (Nagarsenker & Pillai, 1974). Step 4 According to Theorem 2.6.3 in Anderson (1984), if Z 2nL, then the density function of Z is 1 f(z) 2

eitzf(t) dt.

(7)

Step 5 The distribution function of 2nL may be obtained using its denition, i.e. F(z) P(2nL z) f(z) dz.
0

(8)

As the value of w in Eq. (6) is equal to 3, the distribution function of 2nL under null hypothesis may be expressed as (see Sugiura, 1969)
1 2 F(z) P(c2 [P(c2 f z) B2n f+2 z)P(cf z)]

(4)

1 2 2 2 2 2 n2[(3B2 24B3)P(cf+4 z)6B2P(cf+2 z) (3B2 4B3)P(cf z)] 6 1 2 2 2 n3[(4B44B2B3 B3 2)P(cf+6 z) B2(4B33B2)P(cf+4 z) 6

2 3 2 4 B2(4B3 3B2 ), 2)P(cf+2 z)(4B4 4B2B3 B2)P(cf z)] O(n

(9) where fp B2 B3 and B4 p(6p4 45p3 110p2 90p 3) . 480 p(p 1) , 2

Step 2 The approximation formula for the gamma function in Eq. (4) is n(x h) n2 (x h1 / 2)nx x (1) Br+1(h) w1 ), r O(x r ( r 1) x r1
w r

(5)

p(2p2 9p 11) , 24 p(p 1)(p 2)(p 3) 32

where Br(h) is the Bernoulli polynomial of degree r. Taking the logarithmic operation on Eq. (4) and substituting Eq. (5) into (4) results in p(p 1) 2p n(12it) nf(t) 4 (2)rBr+1 [(12it)r1] r(r 1)nr r1
w

(6)

O(nw1), where Br+1

g1

Br+1(g / 2).

. Step 3 Applying the exponential operation on Eq. (6)

The distribution function in Eq. (9) is sufciently accurate only for large sample size (say, n50). In practical operation of a control chart, the sample size is usually small. Therefore, Eq. (9) cannot be used to obtain the type I error probability for the chart. In this paper, to nd the type I error probability for the multivariate control chart, we apply the abovementioned ve steps directly by increasing the value of w in Eq. (6) and run these mathematical operations in the software mathematica 4.0. (2000). In most of the cases, as w30, the chi-square series would converge and consequently the type I error probability for the chart can be numerically

12

C.-Y. Chou et al. / Journal of Loss Prevention in the Process Industries 16 (2003) 918

obtained. The values obtained from mathematica 4.0 are consistent with those given in the study by Nagarsenker and Pillai (1974). 3.2. The distribution function of 2nL under H1 In order to evaluate the powers of the multivariate control chart, the distribution function of 2nL under alternative hypotheses should be studied. Sugiura (1969) developed an asymptotical distribution function of 2nL under H1. However, this function is, again, appropriate only for large sample size. When this function is expanded as what we do in Section 3.1, it cannot converge to a certain probability value. Therefore, in this paper, we apply the simulation and regression approaches to evaluate the distribution function of 2nL under H1. The simulation and regression procedures for the rst out-of-control state (caused by the shift of mean vector) are described as follows: Step 1 Select a value for UCL (called z value) as the critical value of the chart. The selection of z value must meet the statistical requirements. For example, in the case of n 4 and p 2, after the ve steps in Section 3.1 are run, we can obtain the upper 10th percentile of the distribution of 2nL under H0 which is 14.386. Since the upper bound of the type I error probability of the chart is set 0.1 in this paper, the UCL would be greater than or equal to 14.386. When we select a z value, the possible range is z14.386 and, say, 14.40 can be a starting value. Step 2 Generate n p 1 random vectors from the multivariate normal distribution with mean m1 and covariance matrix 0. This step may be done using the software mathematica 4.0. Step 3 From the output of Step 1, calculate the value of 2nL using Eqs. (1), (2) and (3) Step 4 Let mj 1 if the calculated value of 2nL is greater than the selected z value. Otherwise, let mj 0. The subscript j denotes the run of simulation. Step 5 Repeat Steps 24 by 10,000 times. Step 6 Compute the simulated power for the selected z value as follows:
10000

chart, repeat Step 7; otherwise, stop the simulation procedure. Step 8 For a set of selected z values and their corresponding simulated powers, treat the z value as an independent variable and the simulated power as the dependent variable, and obtain a polynomial regression equation by using forward selection (Montgomery & Peck, 1992) for a certain combination of n and p. This equation is used as a function to estimate the power for the corresponding UCL. For the second out-of-control state caused by the shift of covariance matrix, the same eight steps are applied to estimate and evaluate the power of the chart except that in Step 2, we generate n p 1 random vectors from the multivariate normal distribution with mean m0 and covariance matrix 1. The simulated power based on the second out-of-control state is denoted by Power2 in this paper.

4. The cost function The expected total cost per unit of product associated with the test procedure may be written as E(C) E(C1) E(C2) E(C3), (11)

where E(C1) is the expected cost per unit of sampling and carrying out the test procedure, E(C2) the expected cost per unit associated with investigating and correcting the process when the test procedure indicates that the process is out of control, and E(C3) is the expected cost per unit associated with producing defective products. The cost of sampling and testing is assumed to consist of a xed cost (denoted by a1) independent of the sample size n and a cost per unit sampled (denoted by a2), i.e. E(C1) (a1 na2) / k, (12)

Power1

. 10,000

j1

mj (10)

Step 7 Go back to Step 1, select another z value by increasing a step of 0.5 from the last selected z value, and obtain its corresponding simulated power. If the simulated power is greater than 0.9, which is the lower bound of the power for the

where k is the number of units produced between successive samples. To simplify the analysis procedure, we assume that there are only three states for the process, as what Saniga (1977) did. State 0 denotes the process is in control. State 1 denotes the process mean vector shifts such that the process is out of control. State 2 denotes the process covariance matrix shifts such that the process is out of control. The state that the process mean vector and covariance matrix shift simultaneously is not considered in this paper. Let a3 be the average cost of investigating and correcting an out-of-control process. Then E(C2) (a3

i0

riai) / k,

(13)

where ri is the conditional probability that the test pro-

C.-Y. Chou et al. / Journal of Loss Prevention in the Process Industries 16 (2003) 918

13

cedure indicates the process is out of control given that the process is in State i, for i 0, 1 and 2, and ai is the probability that the process is in State i at the time the test is performed. Let a4 represent the penalty cost of producing a defective unit of product. Thus E(C3) a4

i0

specication limits for each quality characteristic. Dene l and u as the lower and upper specication-limit vectors, respectively, whose elements j and uj (for j 1, 2, , p) represent, respectively, the lower and upper specication limits of the jth characteristic. Therefore, according to the denition, di may be determined by
u1

digi,

(14)

where di is the conditional probability of producing a defective unit given that the process is in State i, and gi is the probability that the process is in State i at any point of time. Therefore, by substituting Eqs. (12), (13) and (14) into (11), the expected total cost per unit of product can be expressed as E(C) (a1 na2) / k (a3 a4

1 d0 1 % [(2p)p/2|0|1/2]1 exp (Y 2
1 p 1 m0)T 0 (Ym0) dyp%dy1, u1

up

(17)

up

i0

riai) / k

(15)

1 d1 1 % [(2p)p/2|0|1/2]1 exp (Y 2
1 p

(18)

i0

digi.

m1)T-1 0 (Ym1) dyp%dy1,

u1 up

The ais are cost coefcients independent of the test procedure. The probabilities ri, di, gi and ai in Eq. (15) will be examined in the following paragraphs. Further discussion of the general form of the cost function may be found in works by Knappenberger and Grandage (1969) and Montgomery and Klatt (1972). Economicstatistical design of multivariate control charts is to determine the three test parameters (i.e. n, k and UCL) such that E(C) in Eq. (15) is minimized and the statistical constraints (i.e. r00.1, r10.9 and r20.9) are satised. 4.1. Determination of ri The denition of ri is the conditional probability that the test procedure indicates the process is out of control given that the process is in State i, for i 0, 1 and 2. Note that r0 is the type I error probability of the chart, r1 the power due to the shift of process mean vector, and r2 is the power due to the shift of process covariance matrix. Therefore, if UCL z, then r0 1P(2nLz), (16) where P(2nLz) may be obtained through the ve steps in Section 3.1. Also, r1 and r2 can be estimated by Power1 (in Eq. (10)) and Power2, respectively, using the simulation and regression approaches given in Section 3.2. 4.2. Determination of di The denition of di is the conditional probability of producing a defective unit given that the process is in State i, for i 0, 1 and 2, whose values depend on the

1 d2 1 % [(2p)p/2|1|1/2]1 exp (Y 2
1 p

(19)

m0)T-1 1 (Ym0) dyp%dy1. 4.3. Determination of ai The element ai is dened as the steady-state probability that the process is in State i at the time the test is performed, for i 0, 1 and 2. To obtain ai, a transition probability matrix B is required. The elements in B, denoted by bij, are the probability of the process shifting from State i to State j during the production of k units, for i, j 0, 1 and 2. Suppose Q units are produced per hour and fractional units can be produced. The probability of remaining in State 0 (in-control state) while k units are produced is P0 exp(lk / Q). The probability assigned to States 1 and 2 (out-of-control states) while k units are produced is P1 P2 1 P0 1exp(lk / Q). Knappenberger and Grandage (1969) developed a formula to determine the values of P1 and P2 as follows: Pi 2![1exp(lk / Q)]qi(1q)2i , i!(2i)![1(1q)2] (20)

for i 1 and 2, where 0 q 1. Proper selection of the value of q can precisely describe the distribution of P1 and P2. Particularly, P1 P2 as q 0.667, P1 P2 as q 0.667, and P1 P2 as q 0.667. In practice, the values of P1 and P2 may also be determined by past experience. The elements of B may now be dened. The prob-

14

C.-Y. Chou et al. / Journal of Loss Prevention in the Process Industries 16 (2003) 918

ability of being in control at the mth sample and still in control at the (m 1)st sample is the probability of remaining in control during the production of k units, i.e. b00 P0. The probability of the process being in control at the mth sample and being in the ith out-ofcontrol state at the (m 1)st sample is the probability of shifting to the ith out-of-control state during the production of k units, i.e. b0i Pi, for i 1 and 2. The probability of the process being in the ith out-of-control state (for i 1 and 2) at the mth sample and still being in the same out-of-control state at the (m 1)st sample is the probability of detecting the out-of-control state at the mth sample times the probability of going the same out-of-control state again for the production of k units plus the probability of not detecting this out-of-control state at the mth sample, i.e. bii riPi (1ri), for i 1 and 2. The probability that the process is in the ith out-of-control state (for i 1 and 2) at the mth sample and is in control (or in another out-of-control state) at the (m 1)st sample is the probability that the ith out-of-control state is detected at the mth sample times the probability of remaining in control (or going another out-of-control state) for the production of k units, i.e. bij riPj, for i 1 and 2, j 0, 1 and 2, and i j. Therefore, the transition probability matrix B may be written as

Note that t is the conditional expectation of the occurrence of the assignable cause within an interval of sampling. In this paper we assume that during each sampling interval, there exists at most one assignable cause that makes the process out of control. Then, the probability gi (for i 1 and 2) depends on the probability that the process is in the ith out-of-control state when a sample is taken, and the probability that the process is in control when a sample is taken and shifts to this out-of-control state during the production of k units. That is gi ai a0Pi(1t), Consequently, we have g0 1g1g2 a0P0 a0P1t a0P2t. (26) for i 1 and 2. (25)

5. An example and solution procedure From examination of Eq. (15) and the probabilities in the preceding section, it can be seen that determining the optimal three test parameters is not straightforward. To illustrate the nature of the solutions obtained from economic-statistical design of multivariate control charts, a particular numerical example is presented. We assume that only two quality characteristics are of interest (i.e. p 2), that the in-control state is m0

B r1P0 r1P1 (1r1) r1P2 r2P0 r2P1 r2P2 (1r2)

P0

P1

P2

0 0

and 0

10 0 0 15 0 3 10

and that the other necessary parameters are m1

It is easily shown that B is the transition matrix of an irreducible aperiodic positive recurrent Markov chain. Therefore, there exists a vector a such that aTB aT, (21) where aT [a0,a1,a2], a0 a1 a2 1, and ai is the steady-state probability that the process is in State i at the time the test is performed, for i 0, 1 and 2. It can be shown that the solution to Eq. (21) is a0 r1r2P0 / (r1P2 r2P1 r1r2P0), a1 r2P1 / (r1P2 r2P1 r1r2P0), a2 r1P2 / (r1P2 r2P1 r1r2P0). 4.4. Determination of gi The element gi is dened as the steady-state probability that the process is in State i at any point of time, for i 0, 1 and 2. Duncan (1956) has shown that given a shift between the mth and (m 1)st samples, the average fraction of time that elapses before the shift occurs is 1(1 lk / Q) exp(lk / Q) . t [1exp(lk / Q)](lk / Q) (22) (23) (24)

2 10 2 15 3 10 3 15

90 0 135 , 3 15

q 0.667, l / Q 0.001 (i.e. on an average, the process shifts out of control after every 1000 units are produced), a1 US$20 per sample, a2 US$0.2 per unit sampled, a3 US$10 per investigation, and a4 US$10 per defective unit discovered. The solution procedure is of two stages. In the rst stage, the feasible solution area of the upper control limit of the chart for a particular sample size is determined such that the search area can be reduced. In the second stage, a grid search is applied to nd the values of n, k, and UCL that minimize E(C). 5.1. The rst-stage solution procedure When the process is in control (i.e. State 0), the quality characteristics, Y, follow a multivariate normal distribution with mean m0 and covariance matrix 0. The type

C.-Y. Chou et al. / Journal of Loss Prevention in the Process Industries 16 (2003) 918

15

I error probability, r0, can be obtained through Eq. (16). For a particular sample size (n), to meet the statistical constraint of r00.1, the feasible solution area for the UCL of the chart may be found. Table 1 lists the feasible solution areas of UCL for some sample sizes under State 0. When the process mean vector shifts to m1 (i.e. State 1), the quality characteristics, Y, follow a multivariate normal distribution with mean m1 and covariance matrix 0. The power estimation function can be obtained through the simulation and regression approaches mentioned in Section 3.2. For a particular sample size (n), to satisfy the statistical constraint of r10.9, a power estimation function can be obtained and then the feasible solution area for the UCL of the chart may be determined. For example, as n 5, the power estimation function is Power1 1.040380.00887448z 0.000662788z20.0000171429z3, where z is a specic value of 2nL. Fig. 1 shows the simulated power value and the tted power estimation function in Eq. (27). The determination coefcient for this tted function is 0.999151. From this function, the feasible solution area for the UCL of the chart is z 30.492. Table 1 lists the feasible solution areas of UCL for some sample sizes under State 1. When the process covariance matrix shifts to 1 (i.e. State 2), the quality characteristics, Y, follow a multivariate normal distribution with mean m0 and covariance matrix 1. The power estimation function can be obtained through the simulation and regression approaches mentioned in Section 3.2. For a particular sample size (n), to meet the statistical constraint of r20.9, a power estimation function can be obtained and then the feasible solution area for the UCL of the chart may be determined. For example, as n 5, the power estimation function is Power2 0.996079 0.00137878z 0.000216846z ,
2

5.2. The second-stage solution procedure The grid-search approach is used to nd the values of n, k, and UCL that minimize E(C). A computer program is coded for this purpose. This program is able to interact with the software mathematica 4.0 and computes the probability elements in the cost function. The result from the rst-stage solution procedure greatly reduces the range of the search. From the output of the computer program, the optimum solution is n 6, k 62, Power1 0.998334, UCL 19.918, r0 0.009129, Power2 0.968012, and E(C) 0.77116. That is, the optimal control procedure is to take a random sample of size six every 62 units and conclude the process is out of control if 2nL 19.918. The expected cost per unit of the control procedure is US$0.77116.

(27)

6. Effect of cost parameters In this section, sensitivity analyses are conducted based on the preceding illustrative example to study the effect of cost parameters on the economic-statistical design of the multivariate control charts. The cost parameter a1 is the xed cost of taking a sample. Table 2 lists the optimal designs for different values of a1. It can be seen that sampling interval between samples increases as a1 increases. This is consistent with our reasoning. The cost parameter a2 is the inspection cost per unit. Table 3 lists the optimal designs for different values of a2. As a2 increases, the sample size decreases accordingly. This result is to be expected. In addition, increasing a2 leads to the decrease of the upper control limit. This tends to stabilize the power with the test. The cost parameter a3 is the cost of investigating and correcting the process. Table 4 lists the optimal designs for different values of a3. As a3 increases, both the sample size and upper control limit tend to increase. This is probably due to the expectation that increasing a3 may correspond to a decrease in type I error probability and an increase in power. The cost parameter a4 is the penalty cost of producing a defective unit of product. Table 5 lists the optimal designs for different values of a4. It is noted that as a4 increases, the sampling interval decreases, which means that test/sampling should be conducted more frequently such that a high penalty cost can be avoided.

(28)

where z is a specic value of 2nL. Fig. 2 shows the simulated power value and the tted power estimation function in Eq. (28). The determination coefcient for this tted function is 0.998880. From this function, the feasible solution area for the UCL of the chart is z 24.467. Table 1 lists the feasible solution areas of UCL for some sample sizes under State 2. In Table 1, the last column summarizes the feasible solution area of UCL of the chart for a particular sample size by simply examining the intersection of the feasible solution areas under States 0, 1 and 2.

7. Conclusions Although many authors have discussed the economic design of control charts for the last two decades, the design of multivariate control charts still receives relatively little attention in the literature. In this paper, we

16

Table 1 The feasible solution areas of the upper control limit for some sample sizes (Note that z indicates a specic value of 2nL.) Under State 1 z24.611 z30.492 z36.682 z42.950 z49.643 z56.367 z63.038 z18.634 z24.467 z31.076 z38.268 z46.193 z54.164 z62.041 Under State 2 Feasible solution area (intersection) 14.386z18.634 12.754z24.467 11.914z31.076 11.400z38.268 11.053z46.193 10.802z54.164 10.612z62.041

Sample size (n)

Under State 0

C.-Y. Chou et al. / Journal of Loss Prevention in the Process Industries 16 (2003) 918

4 5 6 7 8 9 10

z14.386 z12.754 z11.914 z11.400 z11.053 z10.802 z10.612

C.-Y. Chou et al. / Journal of Loss Prevention in the Process Industries 16 (2003) 918

17

present the procedure to conduct the economic-statistical design of multivariate control charts for monitoring the process mean vector and covariance matrix simultaneously. The test statistic 2nL is applied to develop this procedure and the cost model is established based on the function given in the study by Montgomery and Klatt (1972). A numerical example is provided to illustrate the design procedure and the effects of cost parameters on the design are studied. From the results of the study, the following observations are obtained:
Fig. 1. The simulated values of Power1 and its tted power estimation function.

1. As the xed cost of taking a sample increases, the sampling interval between samples also increases. 2. As the inspection cost per unit increases, both the sample size and upper control limit lead to decrease. 3. As the cost of investigating and correcting the process increases, both the sample size and upper control limit tend to increase. 4. As the penalty cost of producing a defective unit of product increases, the sampling interval decreases.

Acknowledgements This research was supported by the National Science Council of Taiwan, ROC under the grant NSC90-2218E-224-004.

Fig. 2. The simulated values of Power2 and its tted power estimation function. Table 2 Effect of the xed cost of taking a sample (a1) on the optimal design a1 10 20 100 n 5 6 7 k 44 62 144 UCL 20.175 19.918 17.358

E(C) 0.58154 0.77116 1.55910

r0 0.013062 0.009129 0.015628

Power1 0.990333 0.998334 0.999955

Power2 0.935633 0.968012 0.989324

Table 3 Effect of the inspection cost per unit (a2) on the optimal design a2 0.1 0.2 1.0 n 7 6 4 k 61 62 66 UCL 21.190 19.918 17.111 E(C) 0.76112 0.77116 0.82831 r0 0.004431 0.009129 0.053275 Power1 0.999420 0.998334 0.983580 Power2 0.980800 0.968012 0.915114

Table 4 Effect of investigating and correcting the process (a3) on the optimal design a3 5 10 100 n 5 6 7 k 61 62 62 UCL 17.065 19.918 26.219 E(C) 0.76522 0.77116 0.86182 r0 0.031249 0.009129 0.000810 Power1 0.996753 0.998334 0.997110 Power2 0.956459 0.968012 0.964893

18

C.-Y. Chou et al. / Journal of Loss Prevention in the Process Industries 16 (2003) 918

Table 5 Effect of the penalty cost of producing a defective unit of product (a4) on the optimal design a4 10 20 100 n 5 6 7 k 41 62 153 UCL 20.021 19.918 17.276 E(C) 0.60567 0.77116 1.42832 r0 0.011537 0.009129 0.019836 Power1 0.998873 0.998334 0.999538 Power2 0.955834 0.968012 0.986430

References
Anderson, T. W. (1958). An Introduction to Multivariate Statistical Analysis. New York: Wiley. Anderson, T. W. (1984). An Introduction to Multivariate Statistical Analysis (2nd ed.). New York: Wiley. Chen, Y. K. (1995). Economic and economic-statistical design of Hotellings T2 control chart. MS thesis. Department of Statistics, National Tsing-Hua University, Hsinchu, Taiwan. Davis, A. W. (1971). Percentile approximations for a class of a likelihood ratio criteria. Biometrika, 58, 349356. Duncan, A. J. (1956). The economic design of X-bar charts used to maintain current control of a process. Journal of the American Statistical Association, 52, 228242. Duncan, A. J. (1971). The economic design of X-bar charts when there is a multiplicity of assignable causes. Journal of the American Statistical Association, 66, 107121. Ghare, P. M., & Torgersen, P. E. (1968). The multi-characteristic control chart. Journal of Industrial Engineering, 19, 269272. Gibra, I. N. (1971). Economically optimal determination of the parameters of an X-bar control chart. Management Science, 17, 635 646. Goel, A. L., Jain, S. D., & Wu, S. M. (1968). An algorithm for the determination of the economic design of X-bar charts based on Duncans model. Journal of the American Statistical Association, 63, 304320. Ho, C., & Case, K. E. (1994). Economic design of control charts: a literature review for 19811991. Journal of Quality Technology, 26, 178.

Hotelling, H. (1947). Multivariate quality control. In Eisenhart, Hastay, & Wallis (Eds.), Techniques of Statistical Analysis. New York: McGraw-Hill. Jackson, J. E. (1956). Quality control methods for two related variables. Industrial Quality Control, 12, 48. Knappenberger, H. A., & Grandage, A. H. (1969). Minimum cost quality control tests. AIIE Transactions, 1, 2432. MATHEMATICA 4.0. (2000). Champaign, IL: Wolfram Research, Inc. Montgomery, D. C. (1980). The economic design of control charts: a review and literature survey. Journal of Quality Technology, 12, 7587. Montgomery, D. C., & Klatt, P. J. (1972). Economic design of T2 control charts to maintain current control of a process. Management Science, 19, 7689. Montgomery, D. C., & Peck, E. A. (1992). Introduction to linear regression analysis. New York: Wiley. Nagarsenker, B. N., & Pillai, K. C. S. (1974). Distribution of the likelihood ratio criterion for testing 0, m m0. Journal of Multivariate Analysis, 4, 114122. Saniga, E. M. (1977). Joint economically optimal design of X-bar and R control charts. Management Science, 24, 420431. Saniga, E. M. (1989). Economic statistical control-chart designs with an application to X-bar and R charts. Technometrics, 31, 313320. Sugiura, N. (1969). Asymptotic expansions of the distributions of the likelihood ratio criteria for covariance matrix. The Annals of Mathematical Statistics, 40, 20512063. Vance, L. C. (1983). A bibliography of statistical quality control chart techniques, 19701980. Journal of Quality Technology, 15, 5962.