You are on page 1of 5

Two Alternati ves to the Shewhart X Co ntrol C hart

MORTON KLEIN
Columbia University, New York, NY 10027
Averag erun length (ARL) valuesare calcu la tedfor two X con trol chart schem es and com p ared w ith th ose of a stan d ard Shew h chart. B oth con trolcharts are basedon run s rules an d are ea sily art im p le m e n te dAn out-of-control . conditionfor o n e of the charts is a run of two of two succe ssiv e points beyo nd a special con tro llim it. T he oth er chart uses a run of two of th re e su cc e ssive points beyo nd a di eren tcontro llim it. B oth sch em es are show n to h a ve better, th at is, lower, ARL values than the stand ard S hew h chart for proce ss a ve ra g e art shifts a s large a s 2.6 sta nd ard de via tio ns m fro the m e an .

Introduction
HE Shew hartX control chart, w h an out-of-co th conditionw hen a hicbe signals singlepoint falls beyond ntrolree-sig m a for variables and a control chart limit, has e n the standard attributes since the rst quarter of this veness is century. rooted in its simplici y andItskly. However, it t its attracti largeprocess average shifts quic ability to detect is also known to be relati ely insensiti sustained v ve to sm allprocess ave rageshifts.

Th e rst proposal ore m ake the Shew hartchart m(S hew hart(1941)) totestsin ve in the p roc essruns rules.sensitiaddition al shifts averag esu g gThree suchsm all are ested to tests the form in We s t- e rn Ele ctric (19 56), N elson descri and of bed (1984), ntrol condition if k of n descri an They successi be ve out-of-co Montgom ery (1997). or three-sig m a points fall e re 2 one -, n. limits, w h beyond achievetwo-, goal, they ese k While th do simultaneous tests of signica t increases in their so at out-of- cotrol signals,as show n in Champ the expense false Woodallsn(1 9 8 7 )importantn study. Indeed, and M ontgom e rulesshouldbe used cautiousl be c a u s e ry that th e additional sensitizing (1997 )sugge ste d eects of false y, of thes.potentially delete- rious alarm Alternati control for m eth odologies have beenve su g g ested , chartexam ple, the cumulative sum (CUSUM) and exponentially weighted moving ave rDr. Klein is a Professor in the Departme nt of Industrial Engineering and Operations Resear ch. He is a Member of ASQ. His email address is klein@ieor.colu mbia.edu.

a g e (EWMA)all sche m es.averageof th ese a ve h excelle sm descri ed in M ontgom ery (1997). nt process Both shift capabilities,asfar, th ey d o n o t se em detection b However, widespread appli- cation beyondhave so to achieved process industries.This may be due to a the chem ical that the required calcuperception typical shop oor work and/or to too lations are the com plexfor usual organizational inertia in s ociated r, we procedural changes. Therefore,a s this pape with considerhich may ofbe m ore acceptableto sim pler traditional m ethods, w variations practitioners. For(199 Ross of pedagog icalrep urposes, Derman and es, each three7) conside specially designated(lower which used dtwo additional scheman than -sigm a) control limits. two su ccessi In the rst, ve out-of-cotrol signal is givensuch control limits. n if pointssecond ,an out- of-co fall outside In any two of three either ofve points is obtained ntrol signal fall if th ent (less than suc- cessi a) control beyond dierein their rst schemsigm th reelimits. Thus, an out of control e,given twoobtainedve points,point is above an upper controlsuccessiif signal is either limit both and the other is below a lower limit. In their second points e, an beyond one control limit, or if schem are three successi signal is obtainedor out-of-co trol points are above, if n ve any two of control limit. In brief, they showed below,that both schem es provided increased that either mod- erate p rocess average shifts sensitiviy of a standardShewhartcontrol chart. t over that to parisonswill be given in the next Detailed com section. H ere, we conside r two slightly nt alterna- tives to the Shew hart X chart.diere In the rst, either

Vol. 32, No. 4, October 2000

427

Journal of Quality Technology

428

MORTON KLEIN

twotwo successi points below a lower control ve points above an or successi edto obtain an oupper control limit ve limit are need the out-of co trol signalis given al. n In the sec- ond, n ut-o f-cotrol sign if two of three successiof points are above an upper ve three are below a lower control limit or two choice of such limits is control limit. appe ndices. discussed the The in Note that a two-of- wo rule obser sim plest t runs- rule requiringm oresignal. is thevationthe to obtain an out-of-co a standardruns-rule,except ntrol than one Also, two-of-threer and lowe r lim its sugge ste d re rule is that diere than the usual two-sigm a lines. the uppe he are nt Equally e m e sare sim plethed sh op to or, plem eof o both these sch important for anare m ore sensiti nt. easy It will beave ra g e that boththan the im ve to shown changes process control chart (with no supplem ntary Shewhart an important rang eof such stan-gdard tests)over ch an ees. We evaluate all of the sch em es by withparin their avera g e run length (ARL) prolescomfor a g of the Shew hart schem e. Recall and a given that that specied con-gtrol chart scheme the expected process of pointsshift, an on the chart until an ave ra e plotted ARL is number ntrol signal is obtained.An ARL prole out-of-coq u e n c e ARL values as- sociated with is the u e n c eof proc e ss averag e shifts. The ARL a se qas eociatedof value s s with a zero- alued shift is v n the in-control value. It represe ts the ex- called

pected number of points until a false out-ofcontrol signalis obtained. For variables giving rise to the points e that the randomall calcu lations, we assumand plotted on a a standarddeviationequalto nt The m ean are independeone. norm al with control chartis zero w hen th e process is inof the distribution e s s is consideredto be out-ofcontrol. w henproc proc e s s average is non -zero. The the control the no rm al distribution is sym m etric Because m ean , the ARL prole valu es are the about for equal positi e and negati sam e its v proc e ss average shifts. Changes in deviation ve ave rage , m e a- sured in standard the proce ss assum ed be suddenand sustained. units, are to In brief, alterna-thanour results show that both of ARL ti the standard Shewhart better our ve have proles averacontrol schem es aboutscheme For for proc e s proc e s sg e shifts shifts, the Shewhartchart s up to 2.6. largerprole isaverage ARL applications, eith er of our two altern ati m However, ve in many essh ould be arginally better. the Shewhart schem when the latter is used w ithout adjunct m ore useful than chart, runs-rules.

Results
Table 1 three-sigm a contains X values for for process Shew - e shifts from ARLcontrolntrol) atostandard ave ra g hart es zero (in-co chart out-ofcontrol valu

TABLE 1 . ARL Proles

Shewhart

D-R:2 of 2

D-R:2 of 3 Control Limits

2 of 2

2 of 3

Shift 0 0 .2 0 .4 0 .6 0 .8 1 1 .2 1 .4 1 .6 1 .8 2 2 .2 2 .4 2 .6 2 .8 3 4 5

3 370 308 200 120 72 44 28 18 12 8 .7 6 .3 4 .7 3 .6 2 .9 2.4 2.0 1.2 1.0

1.9 3 2 2 370 313 204 116 65 37 23 15 10 7 .2 5 .5 4 .4 3 .6 3 .1 2 .8 2 .5 2 .1 2 .0

2.0 6 9 8 370 308 193 107 58 33 20 13 8 .9 6 .6 5 .1 4 .1 3 .4 3 .0 2 .7 2 .5 2 .1 2 .0

1.7 8 1 4 370 277 150 79 44 26 16 11 7.8 5.9 4.6 3.8 3.2 2.8 2 .6 2 .4 2 .0 2 .0

1.9 3 0 7 370 271 142 73 40 23 15 10 7.1 5.4 4.3 3.6 3.1 2.8 2 .5 2 .4 2 .0 2 .0

Journal of Quality Technology

Vol. 32, No. 4, Octo ber 2000

TWO ALTERNATIVES TO THE SHEWHART X CONTROL CHART

429

up to ve-sigm a. These ARL values are adjace to com parable values for the Derm annt Ross runs-rules our two-of- wo and the two-of(D-R) and t three for each of thm es. sche eseschem es are also listed. param usedm eth used to d e - termThe valuesof eters The etersis descri ed in the the od ineappendices. For the param ave rageshifts from 0 to b process 2 .6 sta e s a for d eviatio n s, we se e that the rules val- u n d rd both than those of either ARL of (boldfaced) are lower sch eour s . suggested the Shew hartor the shifts g reaterthan addition, for m e In 2.6, the process ave rage D- Rthe Shewhartscheme ARL prole advantages ofn e , and for m any practical are, at m ost equal to o applications, negligible. we have proles forAlthough 8 a n d 1 .9 3not included ARL control our rules, using respecti values are esse limits o f e .7 thoseshow,n in vely, rounded ntially their the sam 1 as Table 1.

(1 9 9eShewhart-EWMA X e and chart schem es. in the nt s o m 7 ), study, itdevelopmd a s a device to red uce control evaluation of In that ance of wa s distribution of the in-co the length. varithe u se ntrol run

Ap pendix A: Computational Details For the Two of Two Scheme


The ARL prole of a standard is to be com paredwith that for this scheme Shewhart chart. appropriate found so that the in-controllimits lower control Hence, have to be upper andwill be equ al to 3 ARL , valuesa m ethis scheme that of the standard for value a s 7 0 .4 the hart control chart. Shew can be viewed as consistin of A control chartone above the upper controlg three below limit, onere gions:between the two limit and a center the gion the lower of a single point re- prob- ability control limits. We d enote in the upper re - gion by pU , in the fallingregion by pL, and in the center region by lower p. The values of th e se probabil-ities determ ine the location of the control limits. Consider an states {1, 2, 3}: three transie no absorbing Markov chain with state {1} represe nting nt limit, (UCL), and points {3} a theeither control state {2} a point beyond upper co the ntrol limit control limitstate above pointc e ss re ach e s lower b so rb in g sta te {4} wThe two succes- sive (LCL). pro- below th e a are beyond just one ofh en control limits. points the The Table A1, with transition probabilities probabilities of chain are given intransition den ed above. the The pU , pL and p as of the rst p a ss a g etime from expected value to the absorbing (out-of-co trol) starting is equal to the in-control avera g e run n state 4 state 1 length. The expectedfrom each of the times rst passage states (number of transitions) to the absorbing state can give n eter- mined ,by solvingle, e rm a n ,G elser,a n dbe d be low 3 ): e xa m p the linear system Olkin (1 9 7 , see fo r D
TABLE A1. TransitionProbabilities Markov for Chain W ith 3 Transient States

Concluding

Remarks

In of-three to its prole, the two- addition ho havefavorable ARLsuch ve rule rule attracti to practitionersw to the may be Shewharta been using a s an adjunct despiteits slight advantage over standard control chart. However, the two-of- wo rule, we nd the sim pliciy of the t t latter m ore appealing. We note are our proposed ARL eterm s,thatnot as go od ascontrol charts, in m e s. For e x - amprole values andan EWMA schegave som parample, Saccucci for Lucas (1990) This ARL eEWMA scheme with = scheme hasters L4 1 , 2.9 and = 0.2 5 . 3 7 0 an and ARL values ofin-control ARL value of 1 0 a n d 3 .5 for p rocess averag e shifts of 0.5, 1.0 and 2 .0 . T h e se ARL values dom inatethose of our two of three control chart of 1 0 1 , 2 3 a n d 4 .3 for the sam e process average shifts. To the best of our know ledge, a two (1992). two rule was rst used by Hurw itz and M athur ofrunsThey usedan to replace all setting because of rules in itdi cul- ties experienced in using industrial three standard operationalas adjuncts to the standardShew hart three rules used two of two control lim its of 1.5, chart. They ately match the in- control ARL (9 2 ) to apprShew hart chart with three adjunct runsoxim of the It is interesting that the rela- tively high rules. incidenceof false out-of-cotrol signals, n obtainedfrom the originaland from this rep lacemnt e schem e,was not reported as troublesom e. A two of two runs-rulewas also u sed , by Klein

Statesat tim e t + 1 1 2 3 4

States at tim e t

1 2 3 4

p p p 0

pU 0 pU 0

pL pL 0 0

0 pU pL 1

Vol. 32, No. 4, October 2000

Journal of Quality Technology

430

MORTON KLEIN

M14 = 1 + (p)M14 + (pU )M24 + (pL)M34 M24 = 1 + (p)M14 + (pL)M34 M34 = 1 + (p)M14 + (pU )M24 . Here, M14 is th e e xpe cted to the ssa g e entrol, p tim the starting state {1} rst it a isout-of-cofro m absorbing, state {4}, that is, the solution, in-control run length. Onew ing formula,average form of of this system ,is the (1992) for th ethe due to Hurwitz and Mathur follo average run length: 1 M14 = . (A1) pU pL 1 p 1 + pU 1 + pL

= Pr{X < 1.7 8 1 4 b}

(A2)

w he= X N (0, 1). 8 1 4 = if b0 2 7 0 6 . we have re Thus, 0.0 1 , S imilarl , pL = Pr{X < 2.7 1 4 }= 0.2 1 7 =7 8 ,and using, y pU Pr{X > the } 2 Equa- tion (A1), 0.7 8ARL, for b = 1, is 25.78.

Ap pendix B: Computational Details For the Two of Three Scheme


ch ex The approa is the sa m eMarkov use d above - cept that the absorbingas that chain incre ases in size to eight states{1, 2, . . . , 8} with the rst s even of th em as transie (see Table B1). The s ta te s nt a re : State (OO) State (OU) has twoboth control points ve betwe e n su ccessi limits; has a rst point between both control limits and the secondabove the UCL;

Since we use sy m m e trical control lim its, se ttingpL = , it= pea sy to show that the above reducesto and noting that 1 p = pL + pU = 2p pU is 1+ p M14 = 2p . Now , set M14 = ARL = 3 7 0 .4to m a t h the inc control ARL of thestandard Shew hart schem e,and solve for p to obtain p = pL = pU = 0.0 3 7 4 2 2Given, . pU = pL = 0.0 3 7 4 2 2 using standardNorm al tables, , the lower control limit is 1.7 8 1 4and the upper control limit is +1.7 8 1 4 . For process average out-of-co trol average run ay, the calculation of the shift values of, s ofb > 0 , lengths requires pU mod n d to the probabilitiespL the the , recalculation accou e and nor- mal distribution relatint ie for the above in the v to use d shift control limits. Then Equation (A1) is to determForthe ARL values associated ine exam ple, for a proce ss with the shift. standarddeviations, have averageshift of b we pL = Pr{X + b < 1.7 8 1 4 }

State (OL) has a rst point between both control limits and the secondbelow the LCL; has its rst point above the UCL and its secondbelow the LCL; has its rst point above the control and its secondbetween the UCL limits; State (LO) has its rst point below the control and its secondbetween the LCL limits; State (LU) has its rst point below the LCL and its secondabove the UCL; State (OOC) three absorbing state, hasthe LCL the points either two of or above the UCL. below As in our earlier calculation, the expectedeach rst pas- sage tim es (number of can be determ inedby of the thesolving the state transitions)from states to absorbing linear system given below . The expected State (UL) State (UO)

TABLE B1. TransitionProbabilities Markov Chain W ith 7 Transient for States

Statesat tim e t + 1 1 (OO) 1 2 3 4 5 6 7 8 (OO) (OU) (OL) (UL) (UO) (LO) (LU) (OOC) p 2 (OU) pU 3 (OL) pL 4 (UL) pL 5 (UO) p 6 (LO) 7 (LU) 8 (OOC) pU pL pL + pU pU pL pL + pU 1

States at time t

p p

pU

pL p

p p

pU

Journal of Quality Technology

Vol. 32, No. 4, Octo ber 2000

TWO ALTERNATIVES TO THE SHEWHART X CONTROL CHART

431

value of absorbing(out-of-co trol) state 8 is equal passagetim 1 to the the rstaveragerunn e from starting state to the in-co ntrol length. M18 = 1 + (p)M18 + (pU )M28 + (pL)M38 M28 = 1 + (p)M58 + (pL)M48 M38 = 1 + (p)M68 + (pU )M78 (B1) M48 = 1 + (p)M68 M58 = 1 + (p)M18 + (pL)M38 M68 = 1 + (p)M18 + (pU )M28 M78 = 1 + (p)M58 . Here, M18 length. To ARL = 7 0 .4 ,we s ol e to an incontrolnd the 3control limits the obtain system v with the additional constraiThe pL above ofpL + = pU + p is pU M18pL= = 70.0 . 6 7 6 . solution = 1, = 3 0 2 nts: From pU, this .4 system tables, we nd that the control limits norm al 3 0 7 . New val- u e s of pL standard are 1.9 from process ave ra g eshifts are found and pU resulting using Equation normetc. Then, anduse the n ew standard (A2),al tables we calculations using of pL and pU to nd the ARL a s values with the process average shift specied, sociated by solving the above system ple for a procein of Equation (B1). b = 1 and a lower equations ss For average shift we have pLexam Pr{X + 1 < control limit of 1.9 3 0 w ,here X N (0, = Hence, pL = 7 of 1.91 6 9 1 . Similarl , pU = Pr{X > }, 1). 0.0 0 3 0 7 y 0.9 3 0 7 = 0.1 7 6 0 0 4and, solving the equations } in Equation (B1), the ARL, for b = 1, is 2 3 .3 7 4 7 . All ARL prole calculationswere done using Miis the average in-co ntrol run

crosoft Exce l. d on e by sim ilar m eth sche m e s Th e D erm an-R evaluationswe re explicit formulas oss the ARL that we use d thein D erm anand Ross ods, except for calculations given (1997).

References
Champ, C. W. and Woo da ll, W. H. (1987). Exact Results for Shewhart Control Charts with Suppleme ntary Runs Rules. Technometrics 29, pp. 393399. Derman, C.; G leser, L. J.; and O lkin, I. (1973). A Guide to Probability Theory and Appli cation . Holt, Rhinehart and Winston, New York, N.Y. Derman, C. and Ross, S. M. (1997). Statisti cal Aspects of Quality Cont rol . Academic Press, San Diego, CA. Hurwitz, A. M. and Math r, M. (1992). A Very Simple Set u of Process Control Rules. Quality Engin eering 5, pp. 2129. Klein, M. (1997). Modied Shewhart-EWMA Control Charts. IIE Transactions 29, pp. 10511056. M ontgome ry, D. C. (1997). Int roduction to Statisti cal Quality Cont rol, 3rd ed., John Wiley & Sons. New York, NY. Nelson, L. S. (1984). The Shewhart Control Chart Tests for Special Causes. Journal of Quality Technol ogy 16, pp. 237 239. Saccucci, M. S. and Lucas, J. M. (1990). Average Run Lengths for Expone ntially Weighted Moving Average Schemes Using the Mar kov Chain Approa ch. Journal of Quality Technol ogy 22, pp. 154162. Shewha rt, W. A. (1941). Co ntributions of Statistics to the Science of Engineering in Fluid Mechanics and Statisti cal Meth ods in Engin eering . Uni versi ty of Pennsyl vania Press, Philadelphia, PA, pp. 97124. W estern E lectric (1956). Statisti cal Quality Cont rol Handbook . Western Electric Cor poration, Indiana polis, IN.

Key Words: Average Run Length, Control Charts, Runs Test .

Vol. 32, No. 4, October 2000

Journal of Quality Technology

You might also like