You are on page 1of 15

Research Article

(wileyonlinelibrary.com) DOI: 10.1002/qre.1208


Published online 22 June 2011 in Wiley Online Library

A Comparison Study on Effectiveness and Robustness of Control Charts for Monitoring Process Mean and Variance
Yanjing Ou,a Di Wen,b Zhang Wua and Michael B. C. Khooc
This article compares the effectiveness and robustness of nine typical control charts for monitoring both process mean and variance, including the most effective optimal and adaptive sequential probability ratio test (SPRT) charts. The nine type, CUSUM type and SPRT type) and three versions (the basic version, charts are categorized into three types (the X optimal version and adaptive version). While the charting parameters of the basic charts are determined by common wisdoms, the parameters of the optimal and adaptive charts are designed optimally in order to minimize an index average extra quadratic loss for the best overall performance. Moreover, the probability distributions of the mean and standard deviation shift are studied explicitly as the inuential factors in a factorial experiment. The shift main ndings obtained in this study include: (1) From an overall viewpoint, the SPRT-type chart is more effective than -type chart by 15 and 73%, respectively; (2) in general, the adaptive chart outperforms the CUSUM-type chart and X the optimal chart and basic chart by 16 and 97%, respectively; (3) the optimal CUSUM chart is the most effective xed sample size and sampling interval chart and the optimal SPRT chart is the best choice among the adaptive charts; charts and the CUSUM charts are always equal to one. Furthermore, and (4) the optimal sample sizes of both the X this article provides several design tables which contain the optimal parameter values and performance indices of 54 charts under different specications. Copyright 2011 John Wiley & Sons, Ltd.
Keywords: extra quadratic loss; quality control; statistical process control; control charts; adaptive charts

1.

Introduction

ince it was introduced as a statistical process control (SPC) tool to monitor processes and ensure quality, the control chart has been increasingly adopted in modern industries and non-manufacturing sectors1--5 . When dealing with a quality characteristic x which is a variable, it is usually necessary to monitor both the mean and variability6 . The combination of a Shewhart X chart and a R chart (or an S chart) has been employed widely for this purpose. As online measurement and distributed computing systems become a norm in todays SPC applications, more sophisticated CUSUM, EWMA and sequential probability ratio test and standard deviation shift . This includes the omnibus (SPRT) schemes have also been developed to detect mean shift EWMA chart7 , the MaxEWMA chart8 , the WLC chart9 , the ABS CUSUM chart10 , and some combined CUSUM or EWMA schemes which is composed of several individual CUSUM or EWMA charts11, 12 . As more and more new charts have been developed over the decades, it is essential to conduct a systematic study that evaluates and compares the performance of different charts in a quantitative and analytical manner under various conditions, so that the SPC users not only roughly know which charts are better than the others, but are also aware of the degree of difference among the performance of the charts. This kind of information will greatly facilitate the SPC users to select a suitable chart for their applications. Usually, a user may not be interested to adopt a complicated chart just for a performance improvement of 1 to 2%. But if the achievable improvement resulting from an advanced chart is more than, say, 20%, one would be willing to put and standard deviation in effort and spend resources to adopt the advanced chart. The probability distributions of mean shift will be different for different processes. Their inuence on the chart performance should also be investigated carefully. shift Some comparison studies of chart performance are available in the literature, such as the comparisons of the multivariate EWMA , CUSUM charts and a special SPRT chart14 , the economic control charts15 , the Shewhart and CUSUM charts13 , the adaptive X

a School b Wee

of Mechanical and Aerospace Engineering, Nanyang Technological University, Singapore 639798, Singapore Kim Wee School of Communication and Information, Nanyang Technological University, Singapore 639798, Singapore c School of Mathematical Sciences, Universiti Sains Malaysia, Penang, Malaysia 11800, Malaysia Correspondence to: Yanjing Ou, School of Mechanical and Aerospace Engineering, Nanyang Technological University, Singapore 639798, Singapore. E-mail: m080004@ntu.edu.sg

Copyright 2011 John Wiley & Sons, Ltd.

Qual. Reliab. Engng. Int. 2012 , 28 3--17

Y. OU ET AL. charts for monitoring process mean and variance12 , the non-parametric charts for detecting defect shifts16 , the prole monitoring approaches17 , the GLR chart18 and the CUSUM charts for detecting mean shifts19 or both mean and standard deviation shifts10 . However, few of the comparative studies have covered all major types and versions of the control charts for monitoring process mean and variance. Specically, seldom are the very effective SPRT charts and adaptive charts included. Many of these studies are conducted in a qualitative or descriptive manner. Such qualitative approaches are unable to tell the degree of the superiority of one chart over another or to rank the charts performance properly when several charts are involved in the comparison. Moreover, almost none of these comparisons have either considered the robustness of chart performance or the probability distributions and standard deviation shift . In most of the studies, the charting parameters have not been optimized, of the mean shift that is, a chart used for comparison may not be the best one of its type and version. This article studies the overall performance of nine typical control charts for monitoring process mean and variance in a -type, CUSUM-type and SPRT-type charts, and three versions (the basic quantitative and analytical manner. It includes the X version, optimal version and adaptive version) for each type. The optimal and adaptive charts will be designed by an optimization procedure in which the objective function to be minimized is the average extra quadratic loss (AEQL), which is a measure of the overall performance of the charts. A factorial experiment will be used to compare the nine charts under different specications on false alarm rate, mean shift range and standard deviation shift range. Moreover, the probability distributions of the the mean and standard deviation shift will be handled explicitly as one input factor. The results and ndings obtained in this shift study, together with a set of design tables, will provide the SPC users with useful aids to select or design a suitable chart for their application. In this study, the quality characteristic x is assumed to follow an identical and independent normal distribution with known in-control 0 and standard deviation 0 ( 0 and 0 may be estimated from the eld records or historical data). When process shift occurs, the mean and standard deviation of x will change. = 0+
0,

0.

(1)

Here, and are the mean shift and standard deviation shift, respectively, in terms of 0 . When a process is in control, = 0 and = 1. For convenience of discussion, x will be converted to z which follows a standard normal distribution when the process is in control. z= x 0 .
0

(2)

Since control charts are often used to detect deterioration in product quality in most of the applications20, 21 , this article only handles the increasing variance shift. The remainder of this article proceeds as follows. The nine control charts will be introduced in the next section, followed by a discussion of the general methodologies of the comparative study. The chart performance will be rst compared for a general case, and then further studied in a factorial experiment. Subsequently, the design tables are introduced and explained, and an illustrative example is presented. The discussions and conclusions are drawn in the last section.

2.

Nine control charts

It is impossible to include all the control charts with different types and versions in any comparative study. In this study, nine , CUSUM, SPRT) and three versions typical and representative charts will be investigated. They are categorized into three types (X (basic, optimal, adaptive). Basic X CUSUM SPRT Basic X Basic CUSUM Basic SPRT Optimal Optimal X Optimal CUSUM Optimal SPRT Adaptive VSSI X VSSI CUSUM VSI SPRT

-type Each type or version has a distinctive level of performance complexity and a different way of implementation. The X charts22 decide process status based on the observations in the latest sample. The CUSUM-type charts23 consider the data in all sample points by checking the cumulative sum. The SPRT charts are developed based on the SPRT24 . A typical model of the SPRT chart proposed by Stoumbos and Reynolds25 will be adopted in this article. This SPRT chart allows the sampling rate to vary based on the data observed at the current sample. Its statistical properties are evaluated based on an assumption that the time required to obtain an individual observation is short enough to be neglected, relative to the sampling interval. The EWMA-type charts are not included as their performance is quite similar to that of the CUSUM-type charts26 . The basic version of each type of charts serves as the baseline for the comparison and its charting parameters are determined loosely according to common wisdoms and conventions. On the contrary, the parameters of the optimal and adaptive charts will be determined analytically aiming at the best overall performance. In view of implementation, the basic and optimal charts use xed sample size and sampling interval (FSSI). But the adaptive charts (such as the variable sampling intervals (VSI) chart and
Copyright 2011 John Wiley & Sons, Ltd. Qual. Reliab. Engng. Int. 2012 , 28 3--17

Y. OU ET AL. variable sample sizes and sampling intervals (VSSI) chart) vary the sampling rate as a function of the observed data. They take samples at a higher rate when there is an indication of a change in the process. In this study, while the adaptive versions of and CUSUM charts adopt the VSSI model (i.e. adapting both sample sizes and sampling intervals), the adaptive SPRT both the X chart will use the VSI model (adapting sampling intervals only), as the SPRT chart has already incorporated the adaptive sample size feature intrinsically. In fact, Ou and Wu27 has found that the VSSI SPRT chart is only slightly more effective than the VSI SPRT chart. Brief introductions to the design and implementation of each of the nine charts are given as follows: chart: This chart detects the two-sided mean shift and increasing variance shift by checking the absolute value (1) The basic X of the sample mean | z| of z (the standardized x in Equation (2)). It has three parameters: the sample size n, sampling interval d, and control limit H. When | z|>H, the process is thought to be out of control. According to the common practice chart reasonably sensitive to in SPC, the sample size n usually takes a value between four and six in order to make the X chart. The sampling interval d depends on both small and large process shifts. In this study, n is set to ve for the basic X the allowable inspection rate, and the control limit H is determined by the specied false alarm rate. chart: This chart has the same three parameters n, d and H as the basic X chart. However, the sample size n, (2) The optimal X as an independent design variable, will be determined by an optimization design. The other two parameters d and H are dependant design variables and are determined by the two constraints on inspection rate and false alarm rate. chart: This chart has six parameters: the sample size n1 and n2 (n1 <n2 ), sampling interval d1 and d2 (d1 >d2 ), (3) The VSSI X warning limit w and control limit H (w<H). As recommended by many researchers28 , the shorter sampling interval d2 is xed at a minimum allowable value (dmin ). Then the sample sizes n1 , n2 and the longer sampling interval d1 are optimized as the independent design variables. The warning limit w and control limit H are determined by constraints. During implementation, if | z|<w, the smaller sample size n1 and longer sampling interval d1 are employed for the next sample; otherwise (i.e. w | z | H), the larger sample size n2 and shorter sampling interval d2 are to be used. An out-of-control signal is produced when | z|>H. (4) The basic CUSUM chart: This chart has four parameters: the sample size n, sampling interval d, reference parameter k and control limit H. According to the recommendation given in most articles12 , the sample size n is set to one, because (n = 1) makes the CUSUM chart much more sensitive to large process shifts and also more effective from an overall viewpoint. Like chart, the sampling interval d and control limit H of this CUSUM chart are determined by the two constraints for the basic X pertaining to inspection rate and false alarm rate. The CUSUM chart keeps on examining the cumulative sum Ct 10 . C0 = 0, Ct = max(0, Ct1 +|zt | k). (3)

It is recalled that z has a mean value equal to zero when the process is in control. Thus, whenever there is an increasing or decreasing mean shift or an increasing variance shift, |zt | is likely to become larger than the reference value k, and the statistic Ct will increase gradually. Sooner or later, Ct is larger than the control limit H and the process is thought to be out of control. As a result, this single CUSUM chart handling the absolute value of zt is able to detect both two-sided mean shift and increasing variance shift. The reference parameter k must be larger than the in-control mean value E(|z|) of |z| (E(|z|) is approximately equal to 0.8), so that Ct is always shrunk toward zero when the process is in control. The actual value of k depends on the targeted shift size, which is difcult to predict and differs for different cases. In this general study, the k value for the basic CUSUM chart is xed as one. (5) The optimal CUSUM chart: Like the basic CUSUM chart, this chart has four parameters n, d, k and H. While d and H depend on the two constraints, the sample size n and reference parameter k will be optimized as the independent design variables. (6) The VSSI CUSUM chart: This chart has seven parameters: the sample sizes n1 and n2 (n1 <n2 ), sampling intervals d1 and d2 (d1 >d2 ), reference parameter k, warning limit w and control limit H (w<H). The shorter sampling interval d2 is xed at dmin . The independent variables n1 , n2 , d1 and k will be determined optimally. The warning limit w and control limit H are determined by constraints. During implementation, if the cumulative sum Ct <w, the smaller sample size n1 and longer sampling interval d1 are employed; otherwise (i.e. Ct w), the larger sample size n2 and shorter sampling interval d2 are to be used. (7) The basic SPRT chart: This chart has four parameters: the sampling interval d, reference parameter k, lower control limit g and upper control limit H (g<H). A sample or an SPRT is taken at the end of each sampling interval d. Within a sample, the observation zi is taken sequentially and the cumulative sum Ci of |zi | is examined. C0 = 0, Ci = Ci1 +|zi | k. (4)

If Ci >H, the process is considered as out of control. If g Ci H, sampling will be continued. Finally, if Ci <g, the process is considered as in control and the current sample is terminated, and the next sample will be taken after a time period of d. To design a basic SPRT chart, an operating characteristic average sample number (ASN) also has to be determined. It is actually the average sample size of the SPRT chart. In this article, ASN is set to 3 and k as 1.2 for the basic SPRT chart, with reference to the cases and examples presented in Reynolds and Stoumbos12 and Ou et al.29 . The sampling interval
Copyright 2011 John Wiley & Sons, Ltd. Qual. Reliab. Engng. Int. 2012 , 28 3--17

Y. OU ET AL. d is determined by the constraint on inspection rate which is equal to ASN / d. Finally, the two control limits g and H are adjusted simultaneously so that the constraints on false alarm rate and ASN are satised. (8) The optimal SPRT chart30 : Like the basic SPRT chart, this chart has four parameters d, k, g and H. The sampling interval d and reference parameter k will be optimized. The two control limits g and H are again adjusted simultaneously in order to satisfy the constraints on ASN and false alarm rate. The optimal SPRT chart is implemented exactly in the same manner as the basic SPRT chart. (9) The VSI SPRT chart27 : This chart has six parameters: the sampling interval d1 and d2 (d1 >d2 ), reference parameter k, warning limit w, and lower and upper control limits g and H. The independent variables d1 , d2 , k and w will be determined optimally. The dependent variables g and H are determined so that the constraints are satised. During implementation, if the statistic | z| of a sample is smaller than w, the longer sampling interval d1 is adopted for the next sample; otherwise, the shorter sampling interval d2 is to be used.

3.

Methodologies of the investigation

3.1. Specications To carry out the design of a control chart, all or part of the following four specications have to be determined: (1) The allowable inspection rate (R). It is equal to the ratio between the average sample size and the average sampling interval. The value of R depends on the available resources such as manpower and instrument. Usually, only the in-control (or long run) value of R is considered, because a process often runs in an in-control condition for a long period and only occasionally falls into an out-of-control status for a short time period. The inspection rate in the short out-of-control period has little inuence on the long run value of R and is of much less concern31 . (2) The minimum allowable value ( ) of the in-control ATS0 . The value of is decided based on the requirement on false alarm rate. is usually (3) The maximum value ( ,max ) of the mean shift . The corresponding ( ,max ) of the standard deviation shift set as ( ,max + 1)12 . The values of ,max and ,max may be chosen based on the knowledge about a process (e.g. the maximum possible shift in a process) or on the shift domain, the users are interested to investigate. (4) The minimum allowable value (dmin ) of sampling interval d. The value of dmin is determined by some practical considerations, such as the amount of inspection that can be nished in a short time period or the assumption for an SPRT chart that d must be signicantly larger than the time for taking an observation. In this study, the time unit is made equal to the time period in which one unit (or one product) can be inspected. For example, if the available resource allows ve units to be inspected per hour, the time unit is 12 (= 60 5 ) min. This setting has a benet that the inspection rate R always equals one. It only inuences the scaling of the sampling interval d, but has no effect on the results of comparison. In addition, the minimum sampling interval dmin is xed at 0.3, in terms of the standard time unit. As R and dmin are xed, only the remaining specications , ,max and ,max have to be determined. 3.2. Performance measures The in-control and out-of-control performance of a control chart is usually measured by the average time to signal (ATS). The in-control ATS0 must be large enough so that false alarm occurs infrequently. On the other hand, the out-of-control ATS should be short enough in order to detect the process shifts quickly. In this study, out-of-control ATS is computed under the steady-state mode. It assumes that the process has reached its stationary distribution at the time when the process shift occurs. While the ATS charts are calculated by simple formulae, the ATS values of other seven charts are determined values of the basic and optimal X by Markov chain procedures. For example, for a VSSI CUSUM chart, the zone between zero and control limit H is divided into m in-control transient states (m is set to 100 in this study). The width of the interval of each state is equal to H / (m 0.5), and the center Oi of state i is equal to i (i = 0, 1,. . . , m 1). The ATS of a VSSI CUSUM chart is calculated by: ATSm1 = (Imm Rmm )1 Dm1 where ATS and D are the ATS vector and sampling interval vector, respectively, I is an identify matrix and R is the transition matrix. If the center Oi of state i is larger than the warning limit w, the corresponding element di in the sampling interval vector D is equal to the shorter sampling interval d2 and the sample size used to calculate the transition probability pij (from state i to state j in matrix R) is equal to the larger sample size n2 . Conversely, if Oi is smaller than w, di is equal to the longer sampling interval d1 and the sample size takes the smaller one n1 . Since it is difcult to predict the actual magnitudes and types of process shifts for most of the applications, many researchers12, 26, 32 pointed out that a control chart must have an excellent overall performance in a broad shift domain, that is, and . being effective for detecting both small and large process shifts of different combinations of In this article, the overall effectiveness of the control charts will be measured by the AEQL and average ratio of ATS (ARATS)33 . The index AEQL stems from the quadratic loss function34 .
Copyright 2011 John Wiley & Sons, Ltd. Qual. Reliab. Engng. Int. 2012 , 28 3--17

Y. OU ET AL. AEQL = 1
,max ( ,max 1) 0
,max ,max

( 2 + 2 1)ATS(

)d

d .

(5)

It is actually a weighted average ATS over the shift domain of (0< ,max , 1< ,max ), using the extra loss ( 2 + 2 1) as the weight. If a chart has a small AEQL value, its out-of-control ATS value is generally short. The second index ARATS is a more heuristic measure of the overall performance. It directly calculates the average of the ratios between the out-of-control ATS( , ) of a chart to be evaluated and the ATS( , )benchmark of a benchmark chart. ARATS = 1 ( ,max ,max 1) 0
,max ,max

ATS( , ) d ATS( , )benchmark

d .

(6)

The method of LegendreGauss Quadrature can be employed to calculate the integrations in (5) and (6) quickly and accurately. In this article, AEQL will be used as the objective function for the optimization designs of the control charts because the computation of AEQL does not require a predetermined benchmark chart and therefore is relatively more tractable. Furthermore, the use of AEQL can avoid some possible bias incurred by the selection of the benchmark. and , because the data of It is very difcult, if not impossible, to estimate the probability distributions of process shifts out-of-control cases are not only sparse, but also vary with time. As a result, it is generally assumed explicitly33 or implicitly12, 26 that all process shifts occur with equal probability, and a uniform distribution with a density function of (1 / ( ,max ( ,max 1))) and (as in Equations (5) and (6)). is used for 3.3. Optimization model Based on the specications, the charting parameters of the optimal and adaptive charts will be determined by the following optimization model: Objective function: Subject to Constraint for ATS0 : Constraint for inspection rate: ATS0 = . r0 = R. (8) (9) Minimize AEQL. (7)

Independent or dependent design variables: Charting parameters. In Equation (9), r0 is the resultant or actual inspection rate. As discussed in the last section, different control charts use different charting parameters as the independent or dependent design variables. The optimal values of the independent design variables are searched so that the objective function AEQL is minimized. The dependent design variables are adjusted to meet the two chart, when the constraints (8) and (9). For some charts, these two constraints can be met easily. For example, for an optimal X sample size n (the independent design variable) is given, the sampling interval d is made equal to n / R so that constraint (9) is satised. Then the control limit H is determined to make ATS0 equal to (constraint (8)). However, for some other charts, a recursive procedure is required in order to satisfy both constraints simultaneously. For example, each of the warning limit w and control limit H of a VSSI CUSUM chart has effect on both r0 and ATS0 . But w mainly affects the value of r0 and the satisfaction of constraint (9), and H largely inuences the value of ATS0 and constraint (8). For a given set of the values of all independent variables, the two dependent variables w and H will be determined as follows: (1) The warning limit w is rst adjusted to make r0 = R. (2) Then the control limit H is adjusted to ensure ATS0 = . These two steps will be repeated a few times until both constraints (8) and (9) are met within the predetermined tolerances. In design practice, the two constraints (8) and (9) are treated as equality constraints rather than inequality ones. This helps to fully utilize the available resources and chart capacity.

4.

A general case

The nine control charts are rst compared for a general case in which specications are set as follows: = 370,
,max = 4.5, ,max = ,max + 1 = 5.5.

(10)

Based on the above specications, the nine control charts are designed and the resultant charting parameters and results are all listed in Table I. The ATS values of the nine charts are displayed in Table II. In Table II, there are a total of 35 out-of-control and ) and one in-control cell ( = 0 and = 1). The smallest ATS in each cell cells (combinations of the discrete values of has been bolded. For most of the cases, a chart is unable to produce smaller ATS than other charts for all different shifts35 . However, as long as one chart has smaller ATS at more cells and/ or to a larger degree, this chart is thought to be more effective
Copyright 2011 John Wiley & Sons, Ltd. Qual. Reliab. Engng. Int. 2012 , 28 3--17

Y. OU ET AL.

Table I. The nine control charts in the general case ( = 370,


Parameters Chart Basic x Optimal x VSSI x Basic CUSUM Optimal CUSUM VSSI CUSUM Basic SPRT Optimal SPRT VSI SPRT n or n1 5 1 1 1 1 1 n2 1 1 d or d1 5.0000 1.0000 1.1419 1.0000 1.0000 1.1419 3.0000 1.1500 1.2500 d2 0.3000 0.3000 0.8081 k 1.0000 1.6500 1.4081 1.2000 1.4338 1.4418 g

,max = 4.5,

,max = 5.5)

w 1.3735 0.1148 1.1500

H 1.1046 2.9997 2.9990 3.8811 1.4878 1.9338 2.3913 1.8694 1.8536

ASN 2.9986 1.1497 1.1527

AEQL 70.8180 26.0932 21.9421 30.3176 24.9708 20.8559 31.8586 19.4424 19.2033

PCI 3.6878 1.3588 1.1426 1.5788 1.3003 1.0861 1.6590 1.0125 1.0000 (9) (6) (4) (7) (5) (3) (8) (2) (1)

ARATS 3.5827 1.4458 1.1885 1.6450 1.3561 1.1026 1.6052 1.0170 1.0000 (9) (6) (4) (8) (5) (3) (7) (2) (1)

0.8671 0.2102 0.1904

than others. It is observed that the differences among the ATS values of the nine charts are very substantial. Specically, the -type SPRT-type charts signicantly outperform the CUSUM-type charts and the latter are substantially more effective than the X charts. On the other hand, within each type of charts, the adaptive version generally outperforms the optimal version which in turn surpasses the basic version. It is noted that the VSI SPRT chart always produces the lowest, or nearly the lowest, ATS across the entire shifts domain compared with other charts. However, the superiority of the VSI SPRT chart over the optimal SPRT chart is generally minor. The values of the two overall performance indices AEQL (Equation (5)) and ARATS (Equation (6)) are shown in Table I, together with another index, Performance Comparison Index (PCI). The index PCI is the ratio between the AEQL of a chart and the AEQL of the best chart under the same conditions (i.e. the same specied values of , ,max and ,max , as well as the same probability and that will be discussed shortly). This index PCI facilitates the performance comparison and ranking based distributions of on AEQL. Under a given condition, the chart with the lowest AEQL has a PCI value equal to one, and the PCI values of all other charts are larger than one. Since the VSI SPRT chart has the smallest AEQL in this general case, it stands as the best chart and PCI = AEQL AEQL = . AEQLbest AEQLVSI SPRT (11)

The VSI SPRT chart is also used, in this general case, as the benchmark for calculating ARATS in Equation (6). These three indices, AEQL, ARATS, and PCI, provide fairly comprehensive information regarding the comparison of the overall performance of the charts. In Table I, the numbers within the parentheses under columns PCI and ARATS indicate the performance ranking of the charts, with (1) for the most effective chart and (9) for the least effective one. The ranking based on PCI is as follows: optimal CUSUM, optimal X, basic CUSUM, basic SPRT, basic X chart . VSI SPRT, optimal SPRT, VSSI CUSUM, VSSI X, (12)

It is interesting to note that the ranking based on ARATS is almost the same except one exchange of the ranking positions between the basic CUSUM and basic SPRT charts. Since both PCI and ARATS usually deliver similar results about the relative performance of the charts, only PCI will be pursued in the following discussions. From the chart rankings, it is also noted that, without the optimization design, a basic SPRT chart or a basic CUSUM chart chart. may be less effective than an optimal X and must exist in any process. However, as aforementioned, it is usually unknown and A joint probability distribution of difcult to be estimated. As a result, one may have to design the control charts based on an assumed uniform distribution over and is quite different the shift domain. A concern arises on how well these charts will work if the actual distribution of from a uniform one. A test has been carried out for this general case. The control charts designed based on the assumption of and uniform distribution (as listed in Table I) are now applied to three different circumstances. In each circumstance, each of follows a beta probability distribution. f ( )= f ( )= (a + b) (a) (b) (a + b) ( (a) (b)
a1 ( b1 ,max ) , a+b1 ,max )

1)a1 ( ,max )b1 . ( ,max 1)a+b1

(13)

These beta distributions serve as the representatives of different types of non-uniform marginal probability distributions of and (Figure 1). The skewness of a beta distribution depends on the parameters a and b. Here a and b are the parameters for distribution f ( ), whereas a and b are for f ( ). If (a<b), the probability distribution is skewed to the right (Figure 1(a)). This represents the situations where most of the shifts cluster to the lower end within the shift range. If (a>b), the probability distribution is skewed to the left (Figure 1(c)). This arises when most of the shifts cluster to the upper end. Finally, if (a = b), the
Copyright 2011 John Wiley & Sons, Ltd. Qual. Reliab. Engng. Int. 2012 , 28 3--17

Y. OU ET AL.

Table II. ATS values of the nine control charts in the general case ( = 370,

,max = 4.5,

,max = 5.5)

Chart 1.0 Basic x Optimal x VSSI x Basic CUSUM Optimal CUSUM VSSI CUSUM Basic SPRT Optimal SPRT VSI SPRT Basic x Optimal x VSSI x Basic CUSUM Optimal CUSUM VSSI CUSUM Basic SPRT Optimal SPRT VSI SPRT Basic x Optimal x VSSI x Basic CUSUM Optimal CUSUM VSSI CUSUM Basic SPRT Optimal SPRT VSI SPRT Basic x Optimal x VSSI x Basic CUSUM Optimal CUSUM VSSI CUSUM Basic SPRT Optimal SPRT VSI SPRT Basic x Optimal x VSSI x Basic CUSUM Optimal CUSUM VSSI CUSUM Basic SPRT Optimal SPRT VSI SPRT Basic x Optimal x VSSI x Basic CUSUM Optimal CUSUM VSSI CUSUM Basic SPRT Optimal SPRT VSI SPRT

0 370.0 370.0 369.7 369.2 369.7 369.4 369.9 369.9 369.9 23.32 8.242 6.770 7.034 7.054 5.218 3.889 4.796 4.540 10.74 3.021 2.420 3.304 2.789 2.142 2.243 1.938 1.866 7.412 1.895 1.552 2.238 1.818 1.469 1.918 1.347 1.317 5.956 1.444 1.217 1.742 1.412 1.189 a1.788 1.108 1.094 5.153 1.208 1.046 1.460 1.194 1.038 1.719 0.981 0.976

0.9 12.95 55.28 48.39 22.23 40.05 27.74 15.85 27.23 25.91 9.577 5.967 4.815 5.517 5.195 3.838 3.108 3.487 3.308 7.708 2.741 2.200 3.056 2.551 1.976 2.161 1.790 1.728 6.371 1.819 1.495 2.158 1.751 1.422 1.896 1.307 1.280 5.496 1.415 1.195 1.708 1.385 1.170 1.779 1.092 1.080 4.914 1.194 1.035 1.442 1.180 1.028 1.714 0.973 0.968

1.8 2.819 8.186 5.293 4.363 4.909 2.690 1.943 2.192 2.009 3.799 3.208 2.497 3.405 2.881 2.138 2.183 1.905 1.826 4.436 2.150 1.738 2.503 2.038 1.614 1.981 1.469 1.430 4.618 1.627 1.351 1.951 1.578 1.302 1.839 1.203 1.183 4.529 1.333 1.136 1.612 1.310 1.118 1.755 1.048 1.039 4.348 1.153 1.006 1.391 1.142 1.002 1.702 0.950 0.947

2.7 2.501 2.116 1.237 2.099 1.613 1.029 1.529 0.800 0.779 2.656 1.780 1.368 2.142 1.661 1.275 1.756 1.132 1.106 3.056 1.591 1.308 1.935 1.538 1.260 1.807 1.158 1.139 3.429 1.389 1.173 1.686 1.360 1.150 1.767 1.073 1.061 3.649 1.219 1.053 1.476 1.204 1.044 1.721 0.985 0.980 3.729 1.092 0.962 1.314 1.084 0.962 1.684 0.917 0.916

3.6 2.500 0.878 0.678 1.354 0.847 0.694 1.502 0.605 0.615 2.508 1.102 0.887 1.458 1.070 0.881 1.590 0.801 0.798 2.618 1.183 1.003 1.485 1.164 0.995 1.680 0.929 0.924 2.839 1.162 1.006 1.422 1.149 1.001 1.697 0.945 0.942 3.065 1.093 0.962 1.323 1.085 0.961 1.682 0.915 0.914 3.229 1.019 0.910 1.222 1.014 0.913 1.662 0.876 0.877

4.5 2.500 0.572 0.572 0.973 0.583 0.587 1.500 0.577 0.593 2.500 0.774 0.684 1.066 0.773 0.697 1.529 0.660 0.668 2.517 0.913 0.809 1.161 0.910 0.817 1.599 0.779 0.783 2.601 0.973 0.869 1.193 0.970 0.874 1.637 0.837 0.840 2.742 0.972 0.874 1.171 0.969 0.879 1.644 0.846 0.849 2.89 0.941 0.855 1.123 0.940 0.861 1.638 0.832 0.836

1.9

2.8

3.7

4.6

5.5

Copyright 2011 John Wiley & Sons, Ltd.

Qual. Reliab. Engng. Int. 2012 , 28 3--17

Y. OU ET AL.

Figure 1. Three sets of Beta probability distributions of (

): (a) Set 1; (b) Set 2; and (c) Set 3

probability distribution is symmetrical (Figure 1(b)). In this article, three different circumstances or different combinations of the and will be studied. Table III lists these three values of (a , b , a and b ) for the marginal probability distributions of circumstances together with the uniform distribution one. With these given probability distributions f ( ) and f ( ), the AEQL produced by the control charts is calculated by the following formula: AEQL =
,max ,max

( 2 + 2 1)ATS(

)f ( )f ( ) d

d .

(14)

It is similar to Equation (5) except that the uniform density function (1 / ( ,max ( ,max 1))) in Equation (5) is replaced by f ( )f ( ). Now each of the nine control charts that are designed based on the uniform assumption is applied to three processes with different f ( ) and f ( ) as shown in Figure 1. That is, the AEQL value will be calculated by Equation (14). The resultant AEQL values under the three beta distributions, as well as the AEQL values under the uniform distribution, are displayed in the top half of Table III. As expected, the AEQL values produced by all nine charts become smaller when the probability distributions f ( ) and f ( ) are skewed to the right, and they become larger when f ( ) and f ( ) are skewed to the left. The bottom half of Table III shows the PCI (the ratio of AEQL / AEQLbest ) values for each circumstance. For example, each PCI value under the column skew to left is obtained by dividing each AEQL value under that column by the AEQL value of the best chart (i.e. the VSI SPRT chart) that also appears in the same column. It can be seen that the VSI SPRT chart outdoes all other charts in terms of
Copyright 2011 John Wiley & Sons, Ltd. Qual. Reliab. Engng. Int. 2012 , 28 3--17

10

Y. OU ET AL.

Table III. Effect of the probability distributions of

and

( = 370, AEQL Beta

,max = 4.5,

,max = 5.5)

Uniform Basic x Optimal x VSSI x Basic CUSUM optimal CUSUM VSSI CUSUM Basic SPRT Optimal SPRT VSI SPRT 70.8180 26.0932 21.9421 30.3176 24.9708 20.8559 31.8586 19.4424 19.2033

Skew to right (a = 2, b = 4) (a = 2, b = 4) 46.5188 21.7717 17.5720 23.3255 19.9381 15.4960 18.0428 14.1193 13.6797

Symmetrical (a = b = 3) (a = b = 3) 62.4804 24.8483 20.6543 29.6059 23.9713 19.8134 28.9390 18.3339 18.0570 PCI Beta

Skew to left (a = 4, b = 2) (a = 4, b = 2) 84.3568 30.6235 26.3897 37.1621 30.1687 26.1067 43.2999 24.6068 24.4740

Uniform Basic x Optimal x VSSI x Basic CUSUM Optimal CUSUM VSSI CUSUM Basic SPRT Optimal SPRT VSI SPRT 3.6878 (9) 1.3588 (6) 1.1426 (4) 1.5788 (7) 1.3003 (5) 1.0861 (3) 1.6590 (8) 1.0125 (2) 1.0000 (1)

Skew to right 3.4006 (9) 1.5915 (7) 1.2845 (4) 1.7051 (8) 1.4575 (6) 1.1328 (3) 1.3190 (5) 1.0321 (2) 1.0000 (1)

Symmetrical 3.4602 (9) 1.3761 (6) 1.1438 (4) 1.6396 (8) 1.3275 (5) 1.0973 (3) 1.6027 (7) 1.0153 (2) 1.0000 (1)

Skew to left 3.4468 (9) 1.2513 (6) 1.0783 (4) 1.5184 (7) 1.2327 (5) 1.0667 (3) 1.7692 (8) 1.0054 (2) 1.0000 (1)

Average 3.4988 1.3944 1.1623 1.6105 1.3295 1.0957 1.5875 1.0163 1.0000 (9) (6) (4) (8) (5) (3) (7) (2) (1)

AEQL under all circumstances. The optimal SPRT chart is also very competitive and always has an AEQL value very close to that of the VSI SPRT chart. The numbers inside the parentheses in the bottom half of Table III indicate the rankings of the charts based on PCI in each column. It is found that the rankings in each column are similar to the ranking under the uniform assumption with no or only one or two exchange of chart ranking positions. This implies that, if a chart is designed using a uniform assumption, the relative performance of this chart (or its ranking position) will be nearly the same even if f ( ) and f ( ) are non-uniform. and have some, but quite limited, inuence on the The above discussions show that the probability distributions of relative performance of the control charts. Usually, if a chart that is designed by using uniform f ( ) and f ( ) has a better (or worse) performance compared with other charts, it also has better (or worse) performance under different non-uniform f ( ) and f ( ).

5.

A factorial experiment

Next, a factorial experiment is carried out in order to further study and compare the performance of the nine charts. Three factors will be considered including (1) ; (2) ,max ( ,max = ,max + 1); and (3) the parameters (a , b , a , b ) of the probability distributions. Each of these factors will be studied in two levels: Factor 1 2 3 Level 1 200 3 4 a = 2, b = 4, skew to right a = 2, b = 4, skew to right Level 2 600 6 7 a = 4, b = 2, skew to left a = 4, b = 2, skew to left
Qual. Reliab. Engng. Int. 2012 , 28 3--17

,max ,max

f ( ) f ( )

11

Copyright 2011 John Wiley & Sons, Ltd.

Y. OU ET AL.

Table IV. PCI values of the charts in a factorial experiment


Case 1 2 = 200 ( ,max =3, a =2 b =4 Chart Basic x Optimal x VSSI x Basic CUSUM Optimal CUSUM VSSI CUSUM Basic SPRT Optimal SPRT VSI SPRT a =2 b =4 2.9824 1.6354 1.3276 1.5474 1.4506 1.1259 1.1103 1.0000 1.0013
,max =4)

6 = 600

a =4 b =2 a =4 b =2

( ,max =6, a =2 b =4 a =2 b =4 3.3807 1.2949 1.1034 1.4452 1.2524 1.0613 1.5471 1.0066 1.0000

,max =7)

a =4 b =2 a =4 b =2

( ,max =3, a =2 b =4 a =2 b =4 3.5811 2.1550 1.6733 1.9008 1.7548 1.2115 1.0936 1.0324 1.0000

,max =4)

a =4 b =2 a =4 b =2

( ,max =6, a =2 b =4 a =2 b =4 3.6288 1.4881 1.2020 1.7545 1.3945 1.0992 1.5098 1.0000 1.0042

,max =7)

a =4 b =2 a =4 b =2

PCI 3.4292 1.4727 1.2207 1.6029 1.3670 1.1081 1.5262 1.0126 1.0007 (9) (6) (4) (8) (5) (3) (7) (2) (1)

SPCI 0.2557 0.3227 0.2046 0.2121 0.1958 0.0523 0.3367 0.0116 0.0015

3.1536 1.3380 1.1269 1.5145 1.3019 1.0970 1.4995 1.0205 1.0000

3.6203 1.1322 1.0429 1.2924 1.1262 1.0511 1.9944 1.0174 1.0000

3.3785 1.5257 1.2178 1.8540 1.4541 1.1445 1.4708 1.0044 1.0000

3.7083 1.2119 1.0718 1.5143 1.2019 1.0743 1.9843 1.0194 1.0000

chart; (b) CUSUM chart; and (c) SPRT chart Figure 2. The PCI values for nine charts under different cases: (a) X

It results in eight different cases or combinations of the three factors. The general case discussed in the last section is nearly at the center of this experiment space. The data (the PCI values) obtained from this experiment are displayed in Table IV. As before the PCI values are calculated to indicate the AEQL values of the charts relative to the AEQL value of the best chart, for each case or under each column. For example, in case 2, the AEQL value of each chart is rst determined under ( = 200, ,max = 3, ,max = 4, (a = 4, b = 2), (a = 4, b = 2)). Since the VSI SPRT chart has the smallest AEQL in this case, then the PCI value of each chart is obtained by dividing its AEQL by the AEQL of the VSI SPRT chart. In this factorial experiment, the VSI SPRT chart usually outdoes all other charts, in terms of AEQL under many cases. However, the optimal SPRT chart is also very competitive. Especially in cases 1 and 7, the AEQL of the optimal SPRT is slightly smaller than that of the VSI SPRT chart. In these two cases, PCI = AEQL / AEQLoptimal SPRT , as the best chart here is the optimal SPRT chart rather than the VSI SPRT chart. The rightmost two columns in Table IV display the sample mean PCI and sample standard deviation SPCI of the PCI values of each chart over the eight cases. The sample mean PCI is the most holistic measure of the effectiveness of a chart, considering not only different specications , ,max and ,max , but also various probability distributions f ( ) and f ( ). According to PCI, the charts are ranked as follows: optimal CUSUM, optimal X, basic SPRT, basic CUSUM, basic X chart. VSI SPRT, optimal SPRT, VSSI CUSUM, VSSI X, (15)

The sample standard deviation SPCI of PCI is a measure of the variability of the PCI values of a chart under different cases. If a chart has a smaller SPCI , its PCI value may vary less under different situations, and this chart is more robust in performance. For example, the SPCI of the VSI SPRT chart is very small, because the PCI values of this chart are always equal or very close to one. On the contrary, the basic SPRT chart has a very large SPCI , because its PCI value is relatively small for some cases (e.g. 1.0936
Copyright 2011 John Wiley & Sons, Ltd. Qual. Reliab. Engng. Int. 2012 , 28 3--17

12

Y. OU ET AL.

Figure 3. Overall comparison of the nine charts in terms of PCI

in case 5) and becomes very large for other cases (e.g. 1.9944 in case 4). This reveals that the degree of the inferiority of the basic SPRT chart to the best chart alters severely under different conditions. Figure 2 also shows the effectiveness and stability of the nine charts under different cases in terms of PCI. Within each type , CUSUM and SPRT), the VSSI version always outperforms the optimal version, whereas the optimal version of control charts (X chart are always outdoes the basic version. In addition, it is obvious that the PCI values of the basic SPRT chart and optimal X least robust under different conditions. Figure 3 illustrates the PCI values of the nine charts. It clearly shows that the SPRT-type charts outperform the CUSUM-type -type charts. The grand average PCISPRT of the SPRT-type charts (i.e. the average of the PCI charts and the latter outdo the X values of the basic SPRT, optimal SPRT and VSI SPRT charts) is equal to 1.180. Similarly, the grand average PCICUSUM is 1.359 for the CUSUM-type charts and PCIX is 2.041 for the X -type charts. This indicates that, from an overall viewpoint, the SPRT chart is more effective than the CUSUM and X charts by 15 and 73%, respectively. Within each type, the adaptive version outperforms the optimal version and the latter surpasses the basic version. The grand averages of the PCI values for the adaptive, optimal and basic versions are PCIadaptive = 1.110, PCIoptimal = 1.284 and PCIbasic = 2.186. This reveals that the adaptive chart outperforms the optimal chart and basic chart by 16 and 97%, respectively.

6.

Design tables

The earlier discussions highlight the importance of properly selecting and designing a control chart for an SPC application. This section provides three design tables (Tables VVII) for three specied values of ( = 200, 370, 600). In each table, ,max and ,max are specied at three levels ( ,max = 3, 4.5, 6, ,max = 4, 5.5, 7). The users can select a chart for which the tabulated values of , ,max and ,max are closest to the desirable values for their application. These three design tables should cover many applications in SPC practice. , CUSUM and For each set of specied , ,max and ,max , the charting parameters of the optimal and adaptive versions of the X SPRT charts are listed. The basic versions are not recommended, because they are substantially less effective than the optimal versions. As aforementioned, the sampling intervals, as well as the specication , in the design tables are expressed in terms of a standard time unit equal to the time period in which one product is inspected. This setting will be further explained in the next section of the example. It is found with surprise that the optimal sample sizes (including n1 and n2 ) of all charts in all three tables are always equal to charts. Moreover, and , which is quite different from some conventional settings, such as (n = 4) for X one for detecting both the fact that the two sample sizes n1 and n2 are always the same for the VSSI version reveals that a VSI chart is actually as and . effective as a VSSI chart for detecting both
Copyright 2011 John Wiley & Sons, Ltd. Qual. Reliab. Engng. Int. 2012 , 28 3--17

13

Y. OU ET AL.

Table V. Design table ( = 200)


,max ,max

Chart Optimal x VSSI x Optimal CUSUM VSSI CUSUM Optimal SPRT VSI SPRT Optimal x VSSI x Optimal CUSUM VSSI CUSUM Optimal SPRT VSI SPRT Optimal x VSSI x Optimal CUSUM VSSI CUSUM Optimal SPRT VSI SPRT

n1 n2 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1

d1

d2

ASN

PCIR

PCIS 1.4731 1.2041 1.3863 1.1200 1.0119 1.0000

PCIL 1.3380 1.1269 1.3019 1.0970 1.0205 1.0000 1.1991 1.0620 1.1869 1.0604 1.0118 1.0000 1.1322 1.0429 1.1262 1.0511 1.0174 1.0000

PCIAVE 1.4822 1.2195 1.3796 1.1143 1.0108 1.0004 1.3063 1.1186 1.2622 1.0770 1.0190 1.0000 1.2060 1.0690 1.1848 1.0569 1.0119 1.0000

1.0000 1.2500 0.3000 0.0000 1.0000 1.4419 1.1919 0.3000 1.2500 1.3081 1.3081 1.3000 0.7581 1.3419 1.0000 1.1419 0.3000 1.0000 1.6000 1.1419 0.3000 1.3500 1.1500 1.5419 1.2000 0.6500 1.5500

2.8070 1.6354 1.1122 2.8063 1.3276 1.5942 1.4506 0.2030 2.0411 1.1259 0.0428 1.8835 1.3087 1.0000 0.1958 1.1581 1.7945 1.1984 1.0013

4.5

5.5

2.8070 1.4302 1.2895 1.3682 2.8061 1.1887 1.1050 1.3313 1.34264 1.2570 0.1733 1.7834 1.0914 1.0791 0.0424 1.4199 1.1497 1.0275 1.0177 0.1038 1.3000 1.4075 1.1270 1.0000 1.0000 1.1907 1.0606 1.1758 1.0581 1.0118 1.0000

1.0000 2.8070 1.2949 1.1419 0.3000 1.3682 2.8061 1.1034 1.0000 1.7581 1.1153 1.2524 1.1419 0.3000 1.4000 0.1130 1.6738 1.0613 1.1500 1.7081 0.1656 1.1770 1.1492 1.0066 1.1419 0.9081 1.5500 0.2581 1.1081 1.4002 1.0873 1.0000

Table VI. Design table ( = 370)


,max ,max

Chart Optimal x VSSI x Optimal CUSUM VSSI CUSUM Optimal SPRT VSI SPRT Optimal x VSSI x Optimal CUSUM VSSI CUSUM Optimal SPRT VSI SPRT Optimal x VSSI x Optimal CUSUM VSSI CUSUM Optimal SPRT VSI SPRT

n1 n2 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1

d1

d2

ASN

PCIR

PCIS 1.6392 1.2961 1.5089 1.1525 1.0149 1.0000 1.3761 1.1438 1.3275 1.0973 1.0153 1.0000 1.2553 1.0880 1.2335 1.0760 1.0141 1.0000

PCIL 1.4423 1.1798 1.3922 1.1123 1.0266 1.0000 1.2513 1.0783 1.2327 1.0667 1.0054 1.0000 1.1753 1.0577 1.1673 1.0623 1.0174 1.0000

PCIAVE 1.6627 1.3263 1.5043 1.1515 1.0138 1.0023 1.4063 1.1689 1.3392 1.0989 1.0176 1.0000 1.2776 1.1014 1.2451 1.0745 1.0133 1.0000

1.0000 1.3000 0.3000 1.0000 1.4419 1.1919 0.3000 1.3081 1.3581 1.2581 1.3500 0.7000 1.3081 1.0000 1.1419 0.3000 1.0000 1.6500 1.1419 0.3000 1.4081 1.1500 1.4338 1.2500 0.8081 1.4419

2.9997 1.9065 1.0346 2.9983 1.5029 1.8604 1.6119 0.1194 2.1964 1.1896 0.0730 2.3552 1.3593 1.0000 0.2314 1.1338 2.1848 1.2164 1.0070 2.9997 1.5915 1.3735 2.9990 1.2845 1.4878 1.4575 0.1148 1.9338 1.1328 0.2102 1.8694 1.1497 1.0321 0.1904 1.1500 1.8536 1.1527 1.0000

4.5

5.5

1.0000 2.9997 1.4022 1.1419 0.3000 1.3735 2.9990 1.1585 1.0000 1.7581 1.3333 1.3346 1.1419 0.3000 1.4081 0.1148 1.9338 1.0851 1.1500 1.6081 0.0217 1.5525 1.1492 1.0083 1.1419 0.4919 1.6500 0.1696 1.4743 1.4854 1.0850 1.0000

The rightmost four columns in each design table enumerate the values of PCIR , PCIS , PCIL and PCIAVE for each chart. Here, PCIR carries the PCI value of the chart when the real probability distributions of f ( ) and f ( ) are the beta distributions that are skewed to the right. Similarly, PCIS or PCIL is related to the beta distributions that are symmetrical or skewed to the and . left. Finally, PCIAVE is the average of PCIR , PCIS and PCIL . Usually, the users do not know the accurate distributions of However, they may have some general and rough idea about them. For example, they may be able to tell that most of the mean shifts and standard deviation shifts cluster to the lower end, that is, having f ( ) and f ( ) skewed to the right. Under such a circumstance, they may make use of the data under the column of PCIR . On the other hand, if the users have no idea at all about f ( ) and f ( ), they can refer to PCIAVE which takes all different distributions into consideration. In addition to detection effectiveness, simplicity in implementation is also an important consideration when selecting a control , VSSI CUSUM, optimal SPRT and VSI SPRT charts) generally have higher detection effectiveness chart. The adaptive charts (the VSSI X and optimal CUSUM charts), because these adaptive charts allow the sampling rate to vary than the FSSI charts (the optimal X at each sampling point. However, it will in turn increase the difculty to run the charts. This operational complexity cannot be overcome just by means of a computer, because changes in sample sizes and sampling intervals must be carried out by the
Copyright 2011 John Wiley & Sons, Ltd. Qual. Reliab. Engng. Int. 2012 , 28 3--17

14

Y. OU ET AL.

Table VII. Design table ( = 600)


,max ,max

Chart Optimal x VSSI x Optimal CUSUM VSSI CUSUM Optimal SPRT VSI SPRT Optimal x VSSI x Optimal CUSUM VSSI CUSUM Optimal SPRT VSI SPRT Optimal x VSSI x Optimal CUSUM VSSI CUSUM Optimal SPRT VSI SPRT

n1 n2 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1

d1

d2

ASN

PCIR

PCIS 1.7850 1.3785 1.6065 1.1875 1.0157 1.0000 1.4608 1.1852 1.3964 1.1216 1.0157 1.0000 1.3057 1.1089 1.2771 1.0896 1.0132 1.0000

PCIL 1.5257 1.2178 1.4541 1.1445 1.0044 1.0000 1.3019 1.1037 1.2787 1.0807 1.0106 1.0000 1.2119 1.0718 1.2019 1.0743 1.0194 1.0000

PCIAVE 1.8219 1.4232 1.6051 1.1812 1.0175 1.0000 1.5047 1.2202 1.4127 1.1244 1.0161 1.0000 1.3352 1.1276 1.2912 1.0877 1.0109 1.0014

1.0000 1.3000 0.3000 1.0362 1.0000 1.4919 1.2419 0.3000 1.2500 0.1178 1.2581 1.2581 0.2337 1.3581 0.8081 1.2581 0.2209 1.1757 1.0000 1.1919 0.3000 1.2389 1.0000 1.6500 1.1419 0.3000 1.4000 0.1291 1.2000 1.3919 0.1536 1.2500 0.6000 1.4419 0.1332 1.3081 1.0000 1.1419 0.3000 1.3759 1.0000 1.7581 1.1419 0.3000 1.4000 0.1291 1.1500 1.5081 0.1208 1.1419 0.4919 1.7081 0.1323 1.4743

3.1440 2.1550 3.1435 1.6733 1.9569 1.7548 2.6551 1.2115 2.6093 1.2639 1.0324 2.6104 1.2711 1.0000 3.1440 1.7512 3.1433 1.3715 1.6633 1.5631 2.1722 1.1711 2.1893 1.1990 1.0221 2.0677 1.1777 1.0000 3.1440 1.4881 3.1434 1.2020 1.4989 1.3945 2.1722 1.0992 1.9201 1.1492 1.0000 1.5717 1.0809 1.0042

4.5

5.5

operators or an automatic inspection system. Among the four adaptive charts, the implementation of the optimal SPRT chart is relatively easier, because this chart uses a xed sampling interval and only adapts the sample size. On the other hand, the FSSI chart is especially easy for implementation. However, charts use FSSI; therefore, they are simpler for operation. The optimal X if a computer-aided SPC system is in use, there is almost no difference in the difculty of running the two FSSI charts (i.e. the and optimal CUSUM charts). optimal X The four types of PCI in the design tables facilitate the users to consider both effectiveness of performance and simplicity in operation when selecting a chart. If the PCI value shows that the improvement gained by a complicated chart in effectiveness is insignicant, it may not be worthwhile to replace a simpler chart by this complicated one. On the contrary, if the achievable improvement in terms of PCI is substantial, the users may believe that the gain in effectiveness is sufcient to outweigh the increase in complexity. As a general guideline, the optimal CUSUM chart and optimal SPRT chart are recommended for most of the SPC applications. chart The optimal CUSUM chart is the most effective and robust FSSI chart and can be implemented as easily as the optimal X when an on-site computer is available. This chart should be selected when the users prefer to using FSSI. The optimal SPRT chart is the best adaptive chart and should be selected when detection speed is critical and the users would like to adopt the adaptive feature for better performance. This chart is almost as effective as the VSI SPRT chart and is considerably more effective than the and VSSI CUSUM charts. Meanwhile, the optimal SPRT chart enjoys the ease of using a xed sampling interval. In all other VSSI X , VSSI CUSUM and VSI SPRT charts), both sample sizes and sampling intervals will have to be three adaptive charts (i.e. the VSSI X changed during the operation.

7.

An example

This example involves a polymerization process in a silicon-ller manufacturing factory36 . The density x of the silicon ller is a key quality characteristic. The probability distribution of x is approximately normal and the in-control standard deviation 0 is estimated to be 0.001 g / cm3 . The in-control mean 0 can be easily adjusted to the nominal value of 1.145 g / cm3 . In order to standardize the design and operation, the density x is converted to z, conforming to a standard normal distribution when the process is in control. z= x 1.145 . 0.001

The inspection rate is 1 unit/ h and the in-control ATS0 is specied as 350 h by the Quality Assurance (QA) engineer. Thus, the standard time unit is equal to 1 h or 60 min, and the tabulated value for in the design tables should be 350. The QA engineer is interested to study the chart performance in a mean shift range (0< 4.5) and a standard deviation shift range (1< 5.5), i.e. ,max = 4.5, ,max = 5.5. Based on the above specications, a design can be selected from Table VI (where is equal to 370, close to 350) and within the catalogue of ( ,max = 4.5, ,max = 5.5). Since the QA engineer has no idea about the probability and standard deviation shift , he checks PCLAVE to gure out the effectiveness of the control distributions of the mean shift charts. The PCLAVE values show that the optimal SPRT chart is only about 1% less effective than the most effective VSI SPRT chart. Furthermore, the former uses a xed sampling interval and thus can reduce some operational difculty. The QA engineer nally
Copyright 2011 John Wiley & Sons, Ltd. Qual. Reliab. Engng. Int. 2012 , 28 3--17

15

Y. OU ET AL. decides to adopt the optimal SPRT chart for monitoring the density x. The parameters of the control charts can be obtained from Table VI. It is noted that the sampling interval for the optimal SPRT chart is d = 1.1500 60 = 69 minutes. In order to facilitate the implementation, d is rounded off to 70 min. This adjustment will change the value of ATS0 to 375 (1.35% bigger than the original value of 370) and decrease the value of r0 to 0.9855 (1.45% smaller). Both changes are minor and tolerable. The values of AEQLVSSI x / AEQLoptimal SPRT and AEQLVSSI CUSUM / AEQLoptimal SPRT are equal to 1.1487 and 1.0799, respectively. This means that the average speed of the optimal SPRT chart for detecting out-of-control cases is higher than that of the and VSSI CUSUM charts by 14.87 and 7.99%, respectively. Meanwhile, the optimal SPRT chart is easier for implementation VSSI X as it uses xed sampling interval.

8.

Discussions and conclusions

Promptly detecting both mean shift and standard deviation shift is the most challenging task in SPC for variables. This article , CUSUM and SPRT charts), each carries out a systematic and analytic comparison among three main types of control charts (X with three versions (basic, optimal and adaptive versions). The comparisons are conducted under various conditions, considering not only different in-control ATS0 , mean shift ranges and standard deviation shift ranges, but also various probability distributions and . of This study ranks the charts in line of both types and versions according to detection effectiveness. The main ndings include -type chart by 15 and 73%, respectively; the adaptive that the SPRT-type chart is more effective than the CUSUM-type chart and X chart outperforms the optimal chart and basic chart by 16 and 97%, respectively; the optimal CUSUM chart is the most effective FSSI chart and the optimal SPRT chart is the best choice among the adaptive charts; the optimal sample sizes (including n1 and and . Such information will facilitate the users to design or n2 ) of all charts are always equal to one for detecting both select a chart based on both superiority in performance and simplicity in implementation. Both the optimal and adaptive charts are designed by an optimization procedure using AEQL as the objective function. The optimization design of the control charts is strongly recommended, as it can signicantly improve the overall performance of the charts and meanwhile does not increase the difculty in implementation. This study also provides several design tables containing 54 charts for different design specications. These tables will aid the users to select a chart by considering both performance and simplicity in implementation, as well as the probability distributions and . of

References
1. Costa AFB, De Magalhaes MS. An adaptive chart for monitoring the process mean and variance. Quality and Reliability Engineering International 2007; 23:821--831. 2. Zhang LY, Chen GM, Castagliola P. Ont and EWMAt charts for monitoring changes in the process mean. Quality and Reliability Engineering International 2009; 25:933--945. 3. Shi LH, Zou CL, Wang ZJ, Kapur KC. A new variable sampling control scheme at xed times for monitoring the process dispersion. Quality and Reliability Engineering International 2009; 25:961--972. 4. Khoo MBC, Xie M. A study of Time-between-Events control chart for the monitoring of regularly maintained systems. Quality and Reliability Engineering International 2009; 25:805--819. 5. Runger GC, Testik MC. Multivariate extensions to cumulative sum control charts. Quality and Reliability Engineering International 2004; 20:587--606. 6. Montgomery DC. Introduction to Statistical Quality Control. Wiley: New York, 2009. 7. Domangue R, Patch SC. Some omnibus exponentially weighted moving average statistical process monitoring schemes. Technometrics 1991; 33:299--313. 8. Chen G, Cheng SW, Xie HS. Monitoring process mean and variability with one EWMA chart. Journal of Quality Technology 2001; 33:223--233. 9. Wu Z, Tian Y. Weighted-loss-function CUSUM chart for monitoring mean and variance of a production process. International Journal of Production Research 2005; 43:3027--3044. 10. Wu Z, Yang M, Khoo MBC, Yu FJ. Optimization designs and performance comparison of two CUSUM schemes for monitoring process shifts in mean and variance. European Journal of Operational Research 2010; 205:136--150. 11. Gan FF. Joint monitoring of process mean and variance using exponentially weighted moving average control charts. Technometrics 1995; 37:446--453. 12. Reynolds MR Jr, Stoumbos ZG. Should observations be grouped for effective process monitoring?. Journal of Quality Technology 2004; 36:343--366. 13. Margavio TM, Conerly MD. A comparison of multivariate moving average control charts for the process mean. International Journal of Production Research 1995; 33:1313--1321. 14. Stoumbos ZG, Reynolds MR Jr. The SPRT control chart for the process mean with samples starting at xed times. Nonlinear Analysis 2001; 2:1--34. 15. Tannock JDT. An economic comparison of inspection and control charting using simulation. International Journal of Quality and Reliability Management 1997; 14:687--699. 16. Das N. A comparison study of three non-parametric control charts to detect shift in location parameters. International Journal of Advanced Manufacturing Technology 2009; 41:799--807. 17. Colosimo BM, Pacella M. A comparison study of control charts for statistical monitoring of functional data. International Journal of Production Research 2010; 48:1575--1601. 18. Renolds MR Jr, Lou JY. An evaluation of a GLR control chart for monitoring the process mean. Journal of Quality Technology 2010; 42:287--310. 19. Ryu JH, Wan H, Kim S. Optimal design of a CUSUM chart for a mean shift of unknown size. Journal of Quality Technology 2010; 42:311--326. Copyright 2011 John Wiley & Sons, Ltd. Qual. Reliab. Engng. Int. 2012 , 28 3--17

16

Y. OU ET AL.
20. Reynolds MR Jr, Glosh BK. Designing control charts for means and variances. ASQC Quality Congress Transaction 1981; 400--407. &S charts for monitoring process capability. Journal of Manufacturing Systems 2002; 21. Wu Z, Xie M, Tian Y. Optimization design of the X 21:83--92. 22. Shewhart WA. Economic Control of Quality of Manufacturing Product. Van Nostrand: New York, 1931. 23. Page ES. Continuous inspection schemes. Biometrika 1954; 41:100--115. 24. Wald A. Sequential Analysis. Wiley: New York, 1947. 25. Stoumbos ZG, Reynolds MR Jr. Control charts applying a sequential test at xed sampling intervals. Journal of Quality Technology 1997; 29:21--40. 26. Reynolds MR Jr, Stoumbos ZG. Control charts and the efcient allocation of sampling resources. Technometrics 2004; 46:200--214. 27. Ou YJ, Wu Z. An SPRT control chart with variable sampling intervals. Technical Report, School of Mechanical and Aerospace Engineering, Nanyang Technological University, Singapore, 2010. 28. Runger GC, Montgomery DC. Adaptive sampling enhancements for Shewhart control charts. IIE Transactions 1993; 25:41--51. 29. Ou YJ, Wu Z, Goh TN. A new SPRT chart for monitoring process mean and variance. Technical Report, School of Mechanical and Aerospace Engineering, Nanyang Technological University, Singapore, 2010. 30. Ou YJ, Wu Z, Lee KM, Chen SL. A highly effective SPRT control chart for monitoring process mean. Technical Report, School of Mechanical and Aerospace Engineering, Nanyang Technological University, Singapore, 2010. 31. Arnold JC, Reynolds MRJR. CUSUM control charts with variable sample sizes and sampling intervals. Journal of Quality Technology 2001; 33:66--81. 32. Sparks RS. Cusum charts for signalling varying location shifts. Journal of Quality Technology 2000; 32:157--171. 33. Wu Z, Jiao JX, Yang M, Liu Y, Wang ZJ. An enhanced adaptive CUSUM control chart. IIE Transactions 2009; 41:642--653. 34. Taguchi G, Wu Y. Introduction to Off-Line Quality Control. American Supplier Institute: Romulus, MI, 1980. 35. Reynolds MR Jr, Stoumbos ZG. Comparisons of some exponentially weighted moving average control charts for monitoring the process mean and variance. Technometrics 2006; 48:550--567. 36. Chen KS, Yu KT, Sheu SH. Process capability monitoring chart with an application in the silicon-ller manufacturing process. International Journal of Production Economics 2006; 103:565--571.

Authors biographies Yanjing Ou is a PhD student in the School of Mechanical and Aerospace Engineering of Nanyang Technological University, Singapore. She received her Bachelor degree and Co-Bachelor degree at the Wuhan Polytechnic University in 2007 and received her Master degree at the City University of Hong Kong in 2008. Her current research work focuses on control chart. Di Wen is a Master student in the Wee Kim Wee School of Communication and Information of Nanyang Technological University, Singapore. He received his Bachelor degree in the Renmin University of China. Zhang Wu is an Associate Professor in the School of Mechanical and Aerospace Engineering of Nanyang Technological University, Singapore. He obtained his PhD in 1988 and MEng in 1984 from the McMaster University, Canada, and obtained his BEng in 1982 from the Huazhong University of Science and Technology, China. His current research interests include the development of algorithms in quality control, reliability, nonlinear optimization and geometrical tolerance. He is a senior member of the American Society for Quality. Michael B. C. Khoo is a Professor in the School of Mathematical Sciences, Universiti Sains Malaysia (USM). He received his PhD in Applied Statistics in 2001 from the Universiti Sains Malaysia. His research interests include statistical process control and reliability analysis. He is a member of the American Society for Quality and serves as a member of the editorial board of Quality Engineering.

17
Qual. Reliab. Engng. Int. 2012 , 28 3--17

Copyright 2011 John Wiley & Sons, Ltd.

You might also like