Variations in a process are classified into two distinct parts called common and special cause variations. In the presence of

Transcription

1 Research Article (wileyonlinelibrary.com) DOI: /qre.1385 Publishe online 29 February 2012 in Wiley Online Library Mixe Exponentially Weighte Moving Average Cumulative Sum Charts for Process Monitoring Nasir Abbas, a * Muhamma Riaz a,b an Ronal J. M. M. Does c The control chart is a very popular tool of statistical process control. It is use to etermine the existence of special cause variation to remove it so that the process may be brought in statistical control. Shewhart-type control charts are sensitive for large isturbances in the process, whereas cumulative sum (CUSUM) type an exponentially weighte moving average (EWMA) type control charts are intene to spot small an moerate isturbances. In this article, we propose a mixe EWMA CUSUM control chart for etecting a shift in the process mean an evaluate its average run lengths. Comparisons of the propose control chart were mae with some representative control charts incluing the classical CUSUM, classical EWMA, fast initial response CUSUM, fast initial response EWMA, aaptive CUSUM with EWMA-base shift estimator, weighte CUSUM an runs rules base CUSUM an EWMA. The comparisons reveale that mixing the two charts makes the propose scheme even more sensitive to the small shifts in the process mean than the other schemes esigne for etecting small shifts. Copyright 2012 John Wiley & Sons, Lt. Keywors: average run length (ARL); control charts; cumulative sum; exponentially weighte moving average; statistical process control 1. Introuction Variations in a process are classifie into two istinct parts calle common an special cause variations. In the presence of common cause variation only, the process is sai to be statistically in control, but once the process inclues both common an special cause variations, it is eeme out of control. Statistical process control (SPC) is the application of the statistical tools to istinguish between common an special cause variations (cf. Montgomery 1 ). The most important of those statistical tools is control chart. Cumulative sum (CUSUM) charts by Page 2 an exponentially weighte moving average (EWMA) charts by Roberts 3 are the two most commonly use types of control charts for etecting the smaller an moerate shifts in the process, whereas Shewhart-type charts are goo at etecting larger shifts. An effective measure use for comparing the performance of the control charts is the average run length (ARL). If we efine a ranom variable R L equal to the number of samples until the first out of control signal occurs, then the probability istribution of this ranom variable R L is known as the run length istribution. The average of this istribution is the ARL. The in-control ARL of a control chart is enote by ARL 0, whereas out-of-control ARL is enote by ARL 1. Lucas 4 propose the use of a combine Shewhart CUSUM quality control scheme in which the CUSUM limits help in etecting the smaller shifts whereas the Shewhart limits increase the sensitivity of the chart for the larger shifts. Similarly, the use of the combine Shewhart EWMA scheme was recommene by Lucas an Saccucci 5 to make the EWMA chart more sensitive for the larger shifts. Another alternative was presente by Jiang et al. 6 to use the aaptive CUSUM proceure with EWMA-base shift estimators in which a range of shifts is targete an the reference value of the CUSUM is upate using the EWMA estimate. In this article, we propose the use of a mixe EWMA CUSUM control chart with the motivation to further enhance the sensitivity of the control chart structure, particularly for the shifts of smaller magnitue in the process. The organization of the rest of the article is as follows. In the next section, we present the basic esign structure of CUSUM- an EWMA-type control charts. The etails regaring the esign structure of the propose scheme are provie in Section 3. Section 4 consists of the comparison of the propose scheme with its counterparts. An illustrative example of the propose scheme is presente in Section 5, an finally, Section 6 summarizes the finings of this article. a Department of Statistics, Quai-i-Azam University Islamaba, Islamaba, Pakistan b Department of Mathematics an Statistics, King Faha University of Petroleum an Minerals, Dhahran, 31261, Saui Arabia c Department of Quantitative Economics, IBIS UvA, University of Amsteram, Plantage Muiergracht 12, 1018 TV, Amsteram, The Netherlans *Corresponence to: Nasir Abbas, Department of Statistics, Quai-i-Azam University Islamaba, Islamaba, Pakistan. nasirabbas55@yahoo.com 345

2 2. Classical CUSUM an EWMA charts Shewhart-type control charts use only the current observation or sample to monitor the process. In this section, we consier control charts that also use previous observations along with the current observation. These mainly inclue CUSUM an EWMA schemes, an we provie here the etails regaring their usual esign structures (also known as classical CUSUM an EWMA control charts) Classical CUSUM control charts The CUSUM chart was originally introuce by Page 2 an is suite to etect small an sustaine shifts in a process. These charts measure a cumulative eviation from the mean or a target value. The two versions of the CUSUM chart, use to evaluate an out of control conition, are the V-mask CUSUM an the tabular CUSUM. The V-mask proceure, which is not very common in use, normalizes the eviations from the mean an plots these eviations. As long as these eviations are plotte aroun the target value, the process is sai to be in control, otherwise out of control. The tabular metho of evaluating a CUSUM chart is commonly use an is similar to a Shewhart-type control chart. In this metho, we plot a function of the subgroup average against the control limits, which are set accoring to a prefixe ARL 0 value. Execution of the tabular CUSUM scheme for controlling the location parameter of the process is performe using two statistics calle C + an C, which are efine as Ci þ ¼ max 0; ðx i m 0 Þ k þ Ci 1 þ (1) Ci ¼ max 0; ðx i m 0 Þ k þ Ci 1 þ where X i enotes the ith observation (e.g. the sample mean of sample i), m 0 is the target mean, an k is the reference value, which is usually chosen equal to half of the shift (in stanar units) to be etecte. The quantities C + an C are known as upper an lower CUSUM statistics, respectively, which are initially set to zero. These two statistics are plotte against the control limit h. As long as the values of Ci þ an Ci are plotte insie the control limit h, the process is sai to be in control, otherwise out of control. If the statistic Ci þ is plotte above h, the process mean is sai to be shifte above the target value, an if the statistic Ci is plotte above h, the process is sai to be shifte below the target value. The quantities k an h are the parameters of the CUSUM control chart, an their proper selection is very important because it greatly influences the ARL performance of the CUSUM chart. Hawkins an Olwell 7 provie a complete ARL stuy for the tabular CUSUM charts for the mean of a normally istribute process with ifferent choices of k an h. Some of their results are given in Table I, with k =0.5 an representing the amount of shift (in stanar units) in the process mean Classical time-varying EWMA control charts The EWMA control chart was introuce by Roberts. 3 Like the CUSUM scheme, EWMA also uses the past information along with the current, but the weights attache to the ata are exponentially ecreasing as the observations become less recent. An EWMA control chart for monitoring the mean of a process is base on the statistic Z i ¼ lx i þ ð1 lþz i 1 (2) where i is the sample number an l is the constant such that 0 < l 1. The quantity Z 0 is the starting value, an it is taken equal to the target mean m 0 or the average of the initial ata in case when the information on the target mean is not available. The control structure of the EWMA chart (which inclues the upper control limit (UCL), centre line (CL) an lower control limit (LCL)) is efine as rffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi l LCL ¼ m 0 Ls X 1 1 l 2 l ð g Þ2i CL ¼ m 0 (3) rffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi l UCL ¼ m 0 þ Ls X 1 1 l 2 l ð Þ2i X i s, an L is the control limit coefficient that etermines where s X is the stanar eviation of the inepenent observations the with of the control limits. Like CUSUM charts, EWMA control charts also have two parameters (l an L). l etermines the ecline of weights, whereas L etermines the with of the control limits, so jointly these two parameters etermine the ARL performance of the EWMA charts. Steiner 8 provie a etaile stuy on the ARL performance of EWMA control chart with these Table I. ARL values for the classical CUSUM scheme with k = h = h =

3 time-varying limits use to monitor the mean of a normally istribute process. These ARLs are reprouce for ARL 0 = 500 an are given in Table II. The results of Table II inicate that the smaller values of l result in better ARL 1 performance for the smaller shifts but the larger values of l provie more shelter against larger shifts. Now for a user of the EWMA chart, it is the balance of ARL 0 an ARL 1 that etermines the parameters of the EWMA chart Moifications of the classical CUSUM an EWMA control charts After the evelopment of CUSUM an EWMA charts, several moifications of these charts have been presente to further enhance the performance of these charts. Lucas 4 presente the combine Shewhart CUSUM quality control scheme in which Shewhart an CUSUM limits are use simultaneously. Lucas an Crosier 9 recommene the use of the fast initial response (FIR) CUSUM, which gives a hea start to the CUSUM statistic by setting the initial values of the CUSUM statistic equal to some positive value (nonzero). This feature gives better ARL 1 performance but at the cost of a ecrease in ARL 0. Yashchin 10 presente the weighte CUSUM scheme, which gives ifferent weights to the previous information use in CUSUM statistic. Riaz et al. 11 applie the runs rules schemes to the CUSUM charts an showe that the runs rules base CUSUM performs better than the classical CUSUM for small shifts. Similarly, on the EWMA sie, Lucas an Saccucci 5 presente the combine Shewhart EWMA quality control scheme, which gives better ARL 1 performance for both small an large shifts. Steiner 8 provie the FIR EWMA, which gives a hea start to the initial value of the EWMA statistic (like FIR CUSUM) an hence improves the ARL 1 performance of the EWMA charts. Abbas et al. 12 applie the runs rules schemes to the EWMA charts an showe that the runs rules base EWMA performs better than the classical EWMA for small shifts. In the next section, we present a mixe EWMA CUSUM quality control scheme for monitoring the mean of a normally istribute process. The inspiration is to get an improve ARL performance by combining the features of EWMA an CUSUM charts in a single control structure. 3. Design structure of the propose mixe EWMA CUSUM scheme In this section, we propose an assortment of the classical EWMA an CUSUM schemes by combining the features of their esign structures. The sai proposal mainly epens on two statistics calle M þ i an M i, which are efine as ) M þ i ¼ max 0; ðq i m 0 Þ a i þ M þ i 1 (4) M i ¼ max 0; ðq i m 0 Þ a i þ M i 1 where a i is a time-varying reference value for the propose charting structure, an the quantities M þ i an M i are known as the upper an lower CUSUM statistics, respectively, which are initially set to zero (i.e. M þ i ¼ M i ¼ 0) an are base on the EWMA statistic Q i, which is efine as Q i ¼ l q Y i þ 1 l q Qi 1 (5) In Equation (5), l q is the constant like l in Equation (2) such that 0 < l q 1 an the initial value of the Q i statistic is set equal to the target mean, that is, Q 0 = m 0. Now the mean an variance of statistic Q i is given as MeanðQ i Þ ¼ m 0 ; VarðQ i Þ ¼ s 2 l 2i Y 1 1 l q (6) 2 l q an this will be use later in the calculation of the parameters of the propose chart. m 0 an s Y are the population mean an stanar eviation, respectively. If these two population parameters are unknown, these can be estimate from preliminary samples. In Equations (4) an (5), we are consiering the case of iniviual observations (n = 1), which may be extene easily for the subgroups. Now the statistics M þ i an M i are plotte against the control limit, say b i. As long as the values of M þ i an M i are plotte insie the control limit, the process is sai to be in control, otherwise out of control. It is to be note here that if the statistic M þ i is plotte above b i, the process mean is sai to be shifte above the target value, an if the statistic M i is plotte above b i, the process is sai to Table II. ARL values for the classical EWMA scheme at ARL 0 = 500 l ¼ 0:1 l ¼ 0:25 L ¼ 2:824 L ¼ 3 l ¼ 0:5 L ¼ 3:072 l ¼ 0:75 L ¼ 3:

4 be shifte below the target value. The control limit b i is selecte accoring to a prefixe ARL 0. A large value of the prefixe ARL 0 will give a larger value of b i an vice versa. The two quantities a i an b i are efine as sffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi pffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi a i ¼ a VarðQ i Þ ¼ a l 2i s Y 1 1 l q 2 l q g sffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi pffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi b i ¼ b VarðQ i Þ ¼ b l 2i s Y 1 1 l (7) q 2 l q where a* an b* are the constants like k an h, respectively, in the classical set up for the CUSUM. The time-varying reference values a i an b i are ue to the variance of the EWMA statistic in Equation (6). For a fixe value of a*, we can select the value of b* from the tables (that are given later in this section) that fix the ARL 0 at our esire level. In general, a i is chosen equal to half of the shift (in units of the stanar eviation of Q i ). Hence, we choose a* = 0.5 because it makes the CUSUM structure more sensitive to the small an moerate shifts (cf. Montgomery 1 ), to which memory charts actually target. To evaluate the ARL performance of a control scheme, there are ifferent approaches use in the literature, incluing Markov chains, integral equations, Monte Carlo simulations an ifferent types of approximations. We have use the Monte Carlo simulation approach in this article to evaluate the ARLs of the propose chart. An algorithm in R language has been evelope to calculate the run lengths. The algorithm is run 50,000 times to calculate the average of those 50,000 run length. A etaile stuy on the ARL performance of the propose EWMA CUSUM control chart to monitor the mean of a normally istribute process is provie in Tables III V for some selective choices of, l q an b*. For this purpose ARL 0 sarefixe at 168, 400 an 500, which are the commonly use choices. For other values of ARL 0 s one may easily obtain the results on similar lines. Table III. ARL values for the propose EWMA CUSUM scheme with a* = 0.5 at ARL 0 = 168 l q ¼ 0:1 l q ¼ 0:25 l q ¼ 0:5 b ¼ 21:3 b ¼ 13:29 b ¼ 8:12 l q ¼ 0:75 b ¼ 5: Table IV. ARL values for the propose EWMA CUSUM scheme with a* = 0.5 at ARL 0 = 400 l q ¼ 0:1 b ¼ 33:54 l q ¼ 0:25 b ¼ 18:7 l q ¼ 0:5 b ¼ 10:52 l q ¼ 0:75 b ¼ 6: Table V. ARL values for the propose EWMA CUSUM scheme with a* = 0.5 at ARL 0 = 500 l q ¼ 0:1 l q ¼ 0:25 l q ¼ 0:5 b ¼ 37:42 b ¼ 20:18 b ¼ 11:2 l q ¼ 0:75 b ¼ 7:

5 The relative stanar errors for the results provie in Tables III V are also calculate an foun to be less than 1.2%. Moreover, we have also replicate the ARL results of the classical CUSUM an the classical EWMA using our simulation algorithm an foun almost similar results as by Hawkins an Olwell 7 an Lucas an Saccucci, 5 respectively, ensuring the valiity of the simulation algorithm use. The main finings about our propose EWMA CUSUM quality control scheme for monitoring the mean of a normally istribute process are given as follows: 1. Mixing the EWMA an CUSUM schemes really boosts the ARL performance of the resulting combination of the two charts especially for small an moerate shifts in the process (cf. Tables III V). 2. For etecting small shifts in the process, the performance of the propose scheme is better with smaller values of l q an vice versa (cf. Tables III V). 3. The propose scheme is ARL unbiase, that is, for a fixe value of ARL 0, the ARL 1 ecreases with a ecrease in the value of an vice versa (cf. Tables III V). 4. For a fixe value of, the ARL 1 of the propose scheme ecreases with a ecrease in ARL 0 (cf. Tables III V). 5. For a fixe value of ARL 0, the control limit coefficient b* ecreases with the increase in l q (cf. Tables III V). 4. Comparisons In this section, we present a comprehensive comparison of the propose mixe EWMA CUSUM scheme with some existing representative EWMA an CUSUM control charts available in the literature. The performance of the control chart is compare in terms of ARL. The set of the schemes consiere for the comparison consist of the classical CUSUM, the classical EWMA, the FIR CUSUM, the FIR EWMA, the aaptive CUSUM with EWMA-base shift estimator, the weighte CUSUM an the runs rules base CUSUM an EWMA. The ARLs of these charts are given in Tables I, II an VI XIII Propose versus classical CUSUM: The ARL values for the classical CUSUM control scheme propose by Page 2 are given in Table I. Comparison of the classical CUSUM with the propose schemes reveals that the propose scheme is performing really goo for all the values of l q, particularly for the smaller values of l q. We can see that, for all the values of l q, the propose scheme has better ARL performance for small shifts, that is, < 1, as compare with the classical CUSUM (cf. Table I versus Table III) Propose versus classical time-varying EWMA: The ARL values for the classical EWMA with time-varying limits, given by Steiner, 8 are provie in Table II. Comparing the classical EWMA with the propose scheme, we observe that the propose scheme have better ARL 1 performance for small shifts, that is, < 0.75, with its respective values of l q (cf. Table II versus Table V) Propose versus FIR CUSUM: The FIR CUSUM presente by Lucas an Crosier 9 provies a hea start to the CUSUM statistic. The ARLs of the CUSUM with FIR feature are given in Table VI in which hea start is represente by C 0. The FIR feature ecreases the ARL 0 of the CUSUM chart, an more importantly this ecrease ARL 0 becomes very small for the larger values of C 0 (for C 0 =2, ARL 0 = 149), Table VI. ARL values for the FIR CUSUM scheme with k = h =4, C 0 = h =5, C 0 = Table VII. ARL values for the FIR EWMA scheme an propose chart FIR EWMA EWMA CUSUM l = 0.1, L =3 l = 0.1, L =3 l q = 0.1, a* = 0.5 l q = 0.1, a* = 0.5 f = 0.4 f = 0.5 b* = b* =

6 Table VIII. ARL values for aaptive CUSUM with þ min an l = 0.3 at ARL 0 = 400 g ¼ 1:5 g ¼ 2 g ¼ 2:5 g ¼ 3 h ¼ 5:05 h ¼ 4:73 h ¼ 4:505 h ¼ 4:394 N. ABBAS, M. RIAZ AND R. J. M. M. DOES g ¼ 4 h ¼ 4:337 g ¼1 h ¼ 4: Table IX. ARL values for the symmetric two-sie weighte CUSUM scheme at ARL 0 = 500 k = 0.5 g h Table X. WL, AL an ARL values for the runs rules base CUSUM scheme 1 at ARL 0 = 168 Limits WL AL Table XI. WL, AL an ARL values for the runs rules base CUSUM scheme 2 at ARL 0 = 168 Limits WL AL Table XII. ARL values for the runs rules base EWMA scheme 1 at ARL 0 = 500 l ¼ 0:1 l ¼ 0:25 L s ¼ 2:556 L s ¼ 2:554 l ¼ 0:5 L s ¼ 2:36 l ¼ 0:75 L s ¼ 2: which is not recommene in case of sensitive processes like in health care (cf. Bonetti et al. 13 ). Comparing the propose scheme with the FIR CUSUM, we foun that for smaller values of l q, the propose scheme has a better ARL performance for small shifts than the FIR CUSUM, even if the FIR CUSUM oes not have the fixe ARL 0 at 168 but has smaller ARL 0 value, that is, 149 (cf. Table VI versus Table III).

7 Table XIII. ARL values for the runs rules base EWMA scheme 2 at ARL 0 = 500 l ¼ 0:1 l ¼ 0:25 L s ¼ 2:3 L s ¼ 2:345 l ¼ 0:5 L s ¼ 2:202 l ¼ 0:75 L s ¼ 1: Propose versus FIR EWMA: FIR EWMA presente by Steiner 8 is similar to the FIR CUSUM as it also gives a hea start to the EWMA statistic. The control limits for the FIR-base EWMA chart are given as r ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi 1þa i 1 m 0 Ls X 1 ð1 f Þ ð Þ l 1 1 l 2 l ð Þ2i where a =( 2/log(1 f) 1)/19. The ARLs for the FIR EWMA with l = 0.1 an the propose chart with l q = 0.1 are given in Table VII. Comparing the propose scheme with the FIR EWMA, we foun that the propose scheme has a better performance than the FIR EWMA for smaller shifts, that is, < 0.5. For moerate an larger shifts, FIR EWMA seems superior as compare with the propose chart Propose versus aaptive CUSUM with EWMA-base shift estimator: Jiang et al. 6 propose the use of aaptive CUSUM with EWMA-base shift estimator. They use the concept of aaptively upating the reference value of the CUSUM chart using the EWMA estimator an then using a suitable weighting function. The ARL values for the aaptive CUSUM are given in Table VIII in which þ min, l, g an h are the parameters of the chart. Comparing the performance of the propose scheme, we notice that the propose scheme is outperforming the aaptive CUSUM for small values of. For moerate an large values of, both the propose scheme an the aaptive CUSUM have almost the same ARL performance (cf. Table VIII versus Table IV). There is also an aaptive EWMA chart (cf. Capizzi an Masarotto 14 ), but its performance is inferior to the aaptive CUSUM, so the results of our proposal are superior to the aaptive EWMA as well Propose versus weighte CUSUM: Weighte CUSUM presente by Yashchin 10 gives weights to the past information in the CUSUM statistic. The ARLs for the weighte CUSUM are given in Table IX in which the weights given to the past information are represente by g. The comparison of the propose scheme with the weighte CUSUM shows that the propose scheme is performing better than the weighte CUSUM for the small an moerate shifts (e.g. < 1.5). For larger values of, the weighte CUSUM almost coincie with the propose scheme (cf. Table IX versus Table V) Propose versus runs rules base CUSUM: Riaz et al. 11 propose the use of the runs rules schemes with the esign structure of the CUSUM charts. The ARLs for the two runs rules base CUSUMs are given in Tables X an XI, in which WL an AL are representing the warning limits an action limits, respectively. The comparison of the propose scheme with both the runs rules base CUSUM schemes shows that the propose scheme has the ability to perform better than the runs rules base CUSUM for all the choices of l q (cf. Tables X an XI versus Table III) Propose versus runs rules base EWMA: Abbas et al. 12 propose the use of the runs rules schemes EWMA structure. The ARLs for the two runs rules base EWMAs are given in Tables XII an XIII. The comparison of the propose schemes with both the runs rules base EWMA schemes shows that the propose scheme is performing better as long as l > 0.1 for the runs rules base EWMA schemes. For l = 0.1, runs rules base EWMA scheme 2 becomes a bit superior to the propose chart (cf. Tables XII an XIII versus Table V) Overall view: To provie an overall comparative view of the propose scheme with the other existing counterparts, we have mae some graphical isplays in the form of ARL curves. Three selective graphs of ifferent charts/schemes (iscusse in Tables I V an VI XIII) are given in Figures 1 3. In Figures 1 3, RR CUSUM (EWMA) stans for the runs rules base CUSUM (EWMA) schemes an the other terms/ symbols use are self-explanatory. By examining the graphs of ARL curves of ifferent schemes uner stuy, we foun that the ARL curves of the propose schemes are on the lower sie, which shows evience for the omination of the propose scheme over the other schemes. For the smaller values of, the ifference between the ARL of the propose scheme an the other schemes is larger, whereas for the moerate values of, this ifference almost isappears. For larger values of, the ARL curve of the propose chart seems above the ARL curves of some other charts, showing the poor performance of the propose chart for larger shifts. To sum up, we may infer that in general the propose scheme is goo to etect small an moerate shifts whereas for larger shifts its performance is inferior to some of the other schemes uner investigation. 351

8 Propose (λ=0.5) FIR CUSUM (Co=1) RR CUSUM II (WL=3.5, AL=4.44) Classical CUSUM (k=0.5, h=4) RR CUSUM I (WL=3.53, AL=4.2) ARLs Figure 1. ARL curves for the propose scheme, the classical CUSUM, the FIR CUSUM an the runs rules base CUSUMs at ARL 0 = 168 Propose (λ=0.1) Propose (λ=0.5) Aaptive CUSUM (γ=2.5) ARLs Figure 2. ARL curves for the propose scheme an aaptive CUSUM at ARL 0 = 400 ARLs Propose (λ=0.5) Classical EWMA (λ=0.5) Weighte CUSUM (γ=0.9) RR EWMA I (λ=0.5) RR EWMA II (λ=0.5) Figure 3. ARL curves for the propose scheme, the classical EWMA, the FIR EWMA, the runs rules base EWMA an the weighte CUSUM at ARL 0 = Illustrative example 352 Besies exploring the statistical properties of a metho, it is always goo to provie its application on some ata for illustration purposes. Authors generally provie these types of examples using real or simulate ata sets, for example, Lucas an Crosier, 9 Khoo, 15 Antzoulakos an Rakitzis 16 an Riaz et al. 11 Following this inspiration, we present here an illustrative example to show how the

9 Table XIV. Application example of the propose scheme using l q = 0.25, a* = 0.5 an b* = at ARL 0 = 500 Sample no. Xi = Yi Qi ai M þ i M i b Sample no. Xi = Yi Qi ai M þ i M i b a a a a a a a a a a Propose scheme giving out-of-control signal. 353

10 propose scheme can be applie in the real situation. For this purpose, a ata set is generate containing 40 observations. The first 20 observations are generate from the in-control situation (i.e. N(0, 1) so that the target mean is 0), an the remaining 20 observations are generate from an out-of-control situation with a small shift introuce in the process (i.e. N(0.5, 1)). The classical CUSUM, the classical EWMA an the propose scheme are applie to this ata set, an the parameters are selecte to be k = 0.5 an h = 5.09 for the classical CUSUM scheme, l = 0.25 an L = for the classical EWMA scheme, an l q = 0.25, a* = 0.5 an b* = for the propose scheme to guarantee that ARL 0 = 500. The calculations for the propose scheme are given in Table XIV, an the graphical isplay of all three control structures are provie in Figures 4 6, with the statistics C þ i an C i plotte against the control limit h for 6 C+ C- h 5 4 C i Sample Number Figure 4. The classical CUSUM chart for the simulate ata set using k = 0.5 an h = 5.09 at ARL 0 = 500 Zi Control Limits Z i Sample Number Figure 5. The classical EWMA chart for the simulate ata set using l = 0.25 an L = at ARL 0 = M+ M- bi 10 8 EC i Sample Number 354 Figure 6. The propose scheme for the simulate ata set using l q = 0.25, a* = 0.5 an b* = at ARL 0 = 500

11 the classical CUSUM scheme, Z i plotte against the control limits L for the classical EWMA scheme an M þ i an M i plotte against the control limit b i for the propose scheme. From Tables XIV an Figure 6, it is obvious that out-of-control signals are receive at samples 32, 33, 34, 35, 36, 37, 38, 39 an 40 by the propose scheme (giving eight out-of-control signals). Figures 4 an 5 show that the separate applications of the classical CUSUM an EWMA schemes fail to etect any out-of-control situation for the given ata set. This clearly inicates superiority of the propose scheme over the classical CUSUM an EWMA schemes, an it is exactly in accorance with the finings of Section Summary an conclusions Control charts are wiely use in processes to etect any special cause variations. Mainly, they are categorize into memory control charts an memoryless control charts. Memoryless control charts (like Shewhart control charts) are esigne to etect the larger shifts in the process, whereas the esign structure of the memory control charts is mae such that they are goo at etecting the small an moerate shifts in the process. CUSUM control charts an EWMA control charts are the two most commonly use memory control charts in the literature. These control schemes o not only use the current observation but also accumulate the information from the past to give a quick signal if the process is slightly off-target. In this article, we have combine the CUSUM an EWMA control schemes into a single control structure an propose a mixe EWMA CUSUM control scheme. The performance of the propose scheme is compare with other CUSUM- an EWMA-type control charts, which are meant to etect small an moerate shifts in the process. The comparisons reveale that the propose scheme is really goo at etecting the smaller shifts in the process as compare with the other schemes uner stuy. Acknowlegements The authors are thankful to the reviewer(s) an eitor for the useful comments to improve the initial version of the article. The author Muhamma Riaz is inebte to King Fah University of Petroleum an Minerals Dhahran Saui Arabia for proviing excellent research facilities through project SB REFERENCES 1. Montgomery DC. Introuction to Statistical Quality Control. 6th e. John Wiley & Sons: New York, Page ES. Continuous Inspection Schemes. Biometrika 1954; 41: Roberts SW. Control Chart Tests Base on Geometric Moving Averages. Technometrics 1959; 1: Lucas JM. Combine Shewhart CUSUM Quality Control Schemes. Journal of Quality Technology 1982; 14(2): Lucas JM, Saccucci MS. Exponentially Weighte Moving Average Control Schemes: Properties an Enhancements. Technometrics 1990; 32: Jiang W, Shu L, Aplet DW. Aaptive CUSUM Proceures with EWMA-Base Shift Estimators. IIE Transactions 2008; 40(10): Hawkins DM, Olwell DH. Cumulative Sum Charts an Charting Improvement. Springer: New York, Steiner SH. EWMA control charts with time varying control limits an fast initial response. Journal of Quality Technology 1999; 31(1): Lucas JM, Crosier RB. Fast Initial Response for CUSUM Quality-Control Scheme. Technometrics 1982; 24: Yashchin E. Weighte Cumulative Sum Technique. Technometrics 1989; 31(1): Riaz M, Abbas N, Does RJMM. Improving the Performance of CUSUM Charts. Quality an Reliability Engineering International 2011; 27(4): Abbas N, Riaz M, Does RJMM. Enhancing the Performance of EWMA Charts. Quality an Reliability Engineering International 2010; 27(6): Bonetti PO, Waeckerlin A, Schuepfer G, Frutiger A. Improving Time-Sensitive Processes in the Intensive Care Unit: the Example of Door-to-Neele Time in Acute Myocarial Infarction. International Journal for Quality in Health Care 2000; 12(4): Capizzi G, Masarotto G. An Aaptive Exponentially Weighte Moving Average Control Chart. Technometrics 2003; 45(3): Khoo MBC. Design of Runs Rules Schemes. Quality Engineering 2004; 16: Antzoulakos DL, Rakitzis AC. The Moifie r out of m Control Chart. Communication in Statistics-Simulations an Computations 2008; 37(2): Authors' biographies Nasir Abbas got his M.Sc. in Statistics from the Department of Statistics Quai-i-Azam University Islamaba Pakistan in 2009; M.Phil in Statistics from the Department of statistics Quai-i-Azam University Islamaba Pakistan in 2011 an currently he is pursuing his Ph.D. in Statistics from the Institute if Business an Inustrial Statistics University of Amsteram The Netherlans. Aitionally, he is also serving as Assistant Census Commissioner in Pakistan Bureau of Statistics. His current research interests inclue Quality Control particularly control charting methoologies. His aress is: nasirabbas55@yahoo.com Muhamma Riaz earne his B.Sc. with Statistics an Mathematics as major subjects from the Government Goron College Rawalpini, University of the Punjab Lahore Pakistan in 1998; M.Sc. in Statistics from the Department of Mathematics an Statistics Quai-i-Azam University Islamaba Pakistan in 2001; M.Phil in Statistics from the Department of Mathematics an statistics, Allama Iqbal Open University Islamaba Pakistan in 2006 an Ph.D. in Statistics from the Institute if Business an Inustrial Statistics University of Amsteram The Netherlans in He serve as a Statistical Officer in the Ministry of Foo, Agriculture an Livestock, Islamaba, Pakistan uring , as a Staff Demographer in the Pakistan Institute of Development Economics, Islamaba, Pakistan uring , as a Lecturer in the Department of Statistics, Quai-i-Azam University, Islamaba, Pakistan uring , as an Assistant Professor in 355

12 the Department of Statistics, Quai-i-Azam University, Islamaba, Pakistan uring He is serving as an Assistant Professor in the Department of Mathematics an Statistics, King Faha University of Petroleum an Minerals, Dhahran 31261, Saui Arabia from 2010-Present. His current research interests inclue Statistical Process Control, Non-Parametric techniques an Experimental Designs. His contacts are: Ronal J.M.M. Does grauate (PhD 1982) in mathematical statistics from the University of Leien. Currently, he is professor of inustrial statistics at the University of Amsteram, managing irector of the Institute for Business an Inustrial Statistics an irector of the Grauate School of Executive Programmes at the Amsteram Business School. His research activities are the esign of control charts for nonstanar situations, the methoology of Lean Six Sigma an healthcare engineering. 356