0HPRU\W\SH&RQWURO&KDUWVLQ6WDWLVWLFDO
3URFHVV&RQWURO
'LWSURHIVFKULIWLVPHGHPRJHOLMNJHPDDNWGRRUHHQILQDQFLsOHELMGUDJHYDQKHW,QVWLWXXWYRRU
%HGULMIVHQ,QGXVWULsOH6WDWLVWLHNYDQGH8QLYHUVLWHLWYDQ$PVWHUGDP
2PVODJRQWZHUS1DVLU$EEDV
ISBN: 978-94-6108-318-0
0HPRU\W\SH&RQWURO&KDUWVLQ6WDWLVWLFDO
3URFHVV&RQWURO
$&$'(0,6&+352()6&+5,)7
WHUYHUNULMJLQJYDQGHJUDDGYDQGRFWRU
DDQGH8QLYHUVLWHLWYDQ$PVWHUGDP
RSJH]DJYDQGH5HFWRU0DJQLILFXV
SURIGU'&YDQGHQ%RRP
WHQRYHUVWDDQYDQHHQGRRUKHWFROOHJHYRRUSURPRWLHVLQJHVWHOGH
FRPPLVVLHLQKHWRSHQEDDUWHYHUGHGLJHQLQGH$JQLHWHQNDSHO
RSGRQGHUGDJVHSWHPEHUWHXXU
GRRU
1DVLU$EEDV
JHERUHQWH5DZDOSLQGL3DNLVWDQ
3URPRWLHFRPPLVVLH
3URPRWRU
3URIGU5-00'RHV
&R3URPRWRU
'U05LD]
2YHULJHOHGHQ3URIGU3%RVZLMN
3URIGU&*+'LNV
3URIGU(5YDQGHQ+HXYHO
3URIGU-GH0DVW
3URIGU&%5RHV
3URIGU-(:LHULQJD
'U06FKRRQKRYHQ
)DFXOWHLW(FRQRPLHHQ%HGULMIVNXQGH
7RP\SDUHQWV
WƌĞĨĂĐĞ
ǀŝŝ
3UHIDFH
7KLVWKHVLVLVWKHUHVXOWRIUHVHDUFKDWWKH,QVWLWXWHIRU%XVLQHVVDQG,QGXVWULDO6WDWLVWLFVRIWKH
8QLYHUVLW\RI$PVWHUGDPWKH1HWKHUODQGV,WZDVUHDOO\DSOHDVXUHIRUPHWRZRUNRQLQWKH
ILHOGRIVWDWLVWLFDOTXDOLW\FRQWUROZKLFKLVDYHU\LQWHUHVWLQJILHOGIURPERWKVFLHQWLILFSRLQWRI
YLHZ DQG UHOHYDQFH LQ SUDFWLFH 1RZ , ZRXOG OLNH WR WKDQN VRPH RI WKH SHRSOH WKDW KDYH
VXSSRUWHGPHLQZULWLQJWKLVWKHVLV
,ZRXOGOLNHWRWKDQN$OPLJKW\$OODK7KHFRPSDVVLRQDWHDQGWKHPRVWPHUFLIXOZKR
EHVWRZHGWRPH+LVEOHVVLQJVWRFRPSOHWHWKLVWDVN$OOUHVSHFWWR+LV3URSKHW 3HDFHEHXSRQ
KLP IRUHQOLJKWHQLQJWKHHVVHQFHRIIDLWKLQ$OODKFRQYHUWLQJDOOKLVNLQGQHVVDQGPHUFLHV
XSRQPH
)LUVW RI DOO , ZRXOG DOVR OLNH WR WKDQN P\ SURPRWHU 3URI GU 5RQDOG - 0 0 'RHV
ZKR UHDOO\ HQKDQFHG P\ UHVHDUFK VNLOOV DQG SROLVKHG P\ DELOLW\ WR FRQYHUW VRPH UHVHDUFK
ZRUNLQWRDSURSHUZULWWHQIRUP,WZDVDQKRQRUIRUPHWRVKDUHWKHUHVHDUFKLGHDVZLWKKLP
DQGWDNHJXLGDQFHLQWKHPDWKHPDWLFDOHYDOXDWLRQVDQGZULWHXSRIDUWLFOHVDQGWKHVLV
,ZRXOGOLNHWRWKDQNP\FRSURPRWHU'U0XKDPPDG5LD]+LVFRQWULEXWLRQVLQWKLV
WKHVLVDUHQRWQHJOLJLEOH+HUHDOO\KHOSHGPHE\SURYLGLQJPHZLWKPDQ\XVHIXOVXJJHVWLRQV
DQG FRPPHQWV , ZRXOG OLNH WR VD\ KHUH ³7KDQN \RX 6LU´ IRU \RXU SDVVLRQ IRU \RXU
FRQVWUXFWLYHFULWLFLVPDQGIRUEHLQJVXFKDQLFHIULHQGO\VXSSRUWHU,UHDOO\HQMR\HGZRUNLQJ
ZLWK\RXDQG,DPSURXGWREH\RXUVWXGHQW
7KDQNV DOVR WR DOO RI P\ WHDFKHUV ZKR DOZD\V JXLGHG PH WKURXJKRXW P\ DFDGHPLF
FDUHHU , ZRXOG OLNH WR PHQWLRQ HVSHFLDOO\ 'U 0XKDPPDG $VODP 'U -DYDLG 6KDELU 'U
=DZDU+XVVDLQ'U=DKLG$VJKDU0U-DPVKDLG,TEDO0UV0DU\DP$VLPDW4XDLGL$]DP
8QLYHUVLW\ ,VODPDEDG DQG 0U 6KDITDW +XVVDLQ 0U 6DLIXU5HKPDQ DW )* 6LU 6\HG
&ROOHJH 7KH 0DOO 5DZDOSLQGL , DP WKDQNIXO WR 0V $WLH %XLVPDQ 2IILFH 0DQDJHU DW
,QVWLWXWH IRU %XVLQHVV DQG ,QGXVWULDO 6WDWLVWLFV IRU KHU FRRSHUDWLRQ LQ DOO NLQG RI
DGPLQLVWUDWLYH PDWWHUV DQG LVVXHV , ZRXOG DOVR OLNH WR WKDQN 0UV ,QH] =ZHWVORRW IRU KHU
HIIRUWVDWWKHSULQWLQJVWDJHRIWKHWKHVLV
7KLV WKHVLV LV D FRPELQDWLRQ RI PXOWLSOH UHVHDUFK DUWLFOHV WKDW DUH SXEOLVKHG RU
VXEPLWWHGLQGLIIHUHQWMRXUQDOV,ZRXOGOLNHWRWKDQNDOOWKHHGLWRUVDQGDQRQ\PRXVUHIHUHHV
IURPWKRVHMRXUQDOV IRUWKHLUYDOXDEOHFRPPHQWVDQGFULWLFLVPWRLPSURYHWKHLQLWLDOIRUPV
RIWKHUHVHDUFKDUWLFOHV
$WODVWEXWQRWOHDVWWKDQNVDUHGXHWRP\IDPLO\IRUWKHLUVXSSRUWGXULQJDOOWKH\HDUV
RIP\VWXG\,HVSHFLDOO\WKDQNP\IDWKHUIRUDOZD\VEHOLHYLQJLQPHDQGP\PRWKHUIRUKHU
QRQHQGLQJ SUD\HUV IRU PH :ULWLQJ WKLV WKHVLV ZLWKRXW P\ IDPLO\¶V VXSSRUW ZDV UHDOO\
XQWKLQNDEOHIRUPH
1$6,5$%%$6
ŽŶƚĞŶƚƐ
ŝdž
WƌĞĨĂĐĞ͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘ǀŝŝ
ŚĂƉƚĞƌϭ͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘ϭ
/ŶƚƌŽĚƵĐƚŝŽŶ͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘ϭ
ϭ͘ϭ
^ƚĂƚŝƐƚŝĐĂůƉƌŽĐĞƐƐĐŽŶƚƌŽů͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘ϭ
ϭ͘Ϯ
ƵŵƵůĂƚŝǀĞƐƵŵĐŽŶƚƌŽůĐŚĂƌƚƐ͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘ϯ
ϭ͘ϯ
džƉŽŶĞŶƚŝĂůůLJǁĞŝŐŚƚĞĚŵŽǀŝŶŐĂǀĞƌĂŐĞĐŽŶƚƌŽůĐŚĂƌƚƐ͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘ϱ
ϭ͘ϰ
KƵƚůŝŶĞŽĨƚŚĞƚŚĞƐŝƐ͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘ϲ
ŚĂƉƚĞƌϮ͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘ϵ
ZƵŶƐƌƵůĞƐďĂƐĞĚh^hDĂŶĚtD͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘ϵ
Ϯ͘ϭ
/ŶƚƌŽĚƵĐƚŝŽŶ͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘ϵ
Ϯ͘Ϯ
dŚĞƉƌŽƉŽƐĞĚƐĐŚĞŵĞƐĨŽƌƚŚĞh^hDĐŚĂƌƚƐ͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘ϭϮ
WĞƌĨŽƌŵĂŶĐĞĞǀĂůƵĂƚŝŽŶŽĨƚŚĞƉƌŽƉŽƐĞĚƐĐŚĞŵĞƐ͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘ϭϰ
ŽŵƉĂƌŝƐŽŶƐ͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘ϭϴ
/ůůƵƐƚƌĂƚŝǀĞĞdžĂŵƉůĞ͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘Ϯϭ
Ϯ͘Ϯ͘ϭ
Ϯ͘Ϯ͘Ϯ
Ϯ͘Ϯ͘ϯ
Ϯ͘ϯ
dŚĞƉƌŽƉŽƐĞĚƐĐŚĞŵĞƐĨŽƌƚŚĞtDĐŚĂƌƚƐ͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘Ϯϰ
WĞƌĨŽƌŵĂŶĐĞĞǀĂůƵĂƚŝŽŶŽĨƚŚĞƉƌŽƉŽƐĞĚƐĐŚĞŵĞƐ͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘Ϯϲ
ŽŵƉĂƌŝƐŽŶƐ͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘ϯϬ
/ůůƵƐƚƌĂƚŝǀĞĞdžĂŵƉůĞ͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘ϯϲ
Ϯ͘ϯ͘ϭ
Ϯ͘ϯ͘Ϯ
Ϯ͘ϯ͘ϯ
Ϯ͘ϰ
ŽŶĐůƵĚŝŶŐƌĞŵĂƌŬƐ͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘ϯϴ
ŚĂƉƚĞƌϯ͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘ϯϵ
DŝdžĞĚtDͲh^hDĐŚĂƌƚƐ͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘ϯϵ
ϯ͘ϭ
DŝdžĞĚtDͲh^hDĐŚĂƌƚĨŽƌůŽĐĂƚŝŽŶ͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘ϰϬ
ĞƐŝŐŶƐƚƌƵĐƚƵƌĞŽĨƚŚĞƉƌŽƉŽƐĞĚĐŚĂƌƚ͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘ϰϭ
ŽŵƉĂƌŝƐŽŶƐ͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘ϰϰ
/ůůƵƐƚƌĂƚŝǀĞĞdžĂŵƉůĞ͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘ϱϬ
ϯ͘ϭ͘ϭ
ϯ͘ϭ͘Ϯ
ϯ͘ϭ͘ϯ
ϯ͘Ϯ
DŝdžĞĚtDͲh^hDĐŚĂƌƚĨŽƌĚŝƐƉĞƌƐŝŽŶ͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘ϱϯ
ͲtDĐŽŶƚƌŽůĐŚĂƌƚ͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘ϱϰ
h^hDͲ ĐŽŶƚƌŽůĐŚĂƌƚ͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘ϱϲ
ĞƐŝŐŶƐƚƌƵĐƚƵƌĞŽĨƚŚĞƉƌŽƉŽƐĞĚĐŚĂƌƚ͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘ϱϳ
ŽŵƉĂƌŝƐŽŶƐ͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘ϲϯ
/ůůƵƐƚƌĂƚŝǀĞĞdžĂŵƉůĞ͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘ϲϳ
ϯ͘Ϯ͘ϭ
ϯ͘Ϯ͘Ϯ
ϯ͘Ϯ͘ϯ
ϯ͘Ϯ͘ϰ
ϯ͘Ϯ͘ϱ
ϯ͘ϯ
ŽŶĐůƵĚŝŶŐƌĞŵĂƌŬƐ͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘ϳϬ
ƉƉĞŶĚŝdžϯ͘ϭ͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘ϳϮ
ƉƉĞŶĚŝdžϯ͘Ϯ͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘ϳϮ
DĞŵŽƌLJͲƚLJƉĞŽŶƚƌŽůŚĂƌƚƐŝŶ^ƚĂƚŝƐƚŝĐĂůWƌŽĐĞƐƐŽŶƚƌŽů
dž
ŚĂƉƚĞƌϰ͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘ϳϱ
ƵdžŝůŝĂƌLJŝŶĨŽƌŵĂƚŝŽŶďĂƐĞĚh^hDĂŶĚtDĐŽŶƚƌŽůĐŚĂƌƚƐ͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘ϳϱ
ϰ͘ϭ
ŽŶƚƌŽůĐŚĂƌƚƐƵƐŝŶŐĂƵdžŝůŝĂƌLJŝŶĨŽƌŵĂƚŝŽŶ͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘ϳϱ
ϰ͘Ϯ
tDĐŽŶƚƌŽůĐŚĂƌƚƐƵƐŝŶŐĂƵdžŝůŝĂƌLJŝŶĨŽƌŵĂƚŝŽŶ͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘ϳϳ
ŽŵƉĂƌŝƐŽŶƐ͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘ϴϮ
dŚĞĐĂƐĞŽĨƚǁŽĂƵdžŝůŝĂƌLJǀĂƌŝĂďůĞƐ͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘ϴϰ
/ůůƵƐƚƌĂƚŝǀĞĞdžĂŵƉůĞ͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘ϴϲ
ϰ͘Ϯ͘ϭ
ϰ͘Ϯ͘Ϯ
ϰ͘Ϯ͘ϯ
ϰ͘ϯ
h^hDĐŽŶƚƌŽůĐŚĂƌƚƐƵƐŝŶŐĂƵdžŝůŝĂƌLJŝŶĨŽƌŵĂƚŝŽŶ͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘ϴϴ
ŽŵƉĂƌŝƐŽŶƐ͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘ϵϭ
/ůůƵƐƚƌĂƚŝǀĞĞdžĂŵƉůĞ͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘ϵϮ
ϰ͘ϯ͘ϭ
ϰ͘ϯ͘Ϯ
ϰ͘ϰ
ŽŶĐůƵĚŝŶŐƌĞŵĂƌŬƐ͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘ϵϯ
ƉƉĞŶĚŝdžϰ͘ϭ͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘ϵϰ
ƉƉĞŶĚŝdžϰ͘Ϯ͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘ϵϰ
ŚĂƉƚĞƌϱ͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘ϵϳ
WƌŽŐƌĞƐƐŝǀĞĐŽŶƚƌŽůĐŚĂƌƚŝŶŐ͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘ϵϳ
ϱ͘ϭ
dŚĞƉƌŽƉŽƐĞĚƉƌŽŐƌĞƐƐŝǀĞŵĞĂŶĐŽŶƚƌŽůĐŚĂƌƚ͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘ϵϳ
ŽŵƉĂƌŝƐŽŶƐ͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘ϭϬϬ
/ůůƵƐƚƌĂƚŝǀĞĞdžĂŵƉůĞ͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘ϭϬϰ
ϱ͘ϭ͘ϭ
ϱ͘ϭ͘Ϯ
ϱ͘Ϯ
&ůŽĂƚŝŶŐĐŽŶƚƌŽůĐŚĂƌƚƐĨŽƌƉƌŽĐĞƐƐĚŝƐƉĞƌƐŝŽŶ͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘ϭϬϴ
&ůŽĂƚŝŶŐ ĐŽŶƚƌŽůĐŚĂƌƚ͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘ϭϭϬ
&ůŽĂƚŝŶŐ ĐŽŶƚƌŽůĐŚĂƌƚ͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘ϭϭϮ
ŽŵƉĂƌŝƐŽŶƐ͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘ϭϭϱ
/ůůƵƐƚƌĂƚŝǀĞĞdžĂŵƉůĞ͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘ϭϭϴ
ϱ͘Ϯ͘ϭ
ϱ͘Ϯ͘Ϯ
ϱ͘Ϯ͘ϯ
ϱ͘Ϯ͘ϰ
ϱ͘ϯ
ŽŶĐůƵĚŝŶŐƌĞŵĂƌŬƐ͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘ϭϮϭ
ƉƉĞŶĚŝdžϱ͘ϭ͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘ϭϮϮ
ƉƉĞŶĚŝdžϱ͘Ϯ͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘ϭϮϮ
ZĞĨĞƌĞŶĐĞƐ͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘ϭϮϯ
^ĂŵĞŶǀĂƚƚŝŶŐ͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘ϭϮϳ
ƵƌƌŝĐƵůƵŵsŝƚĂĞ͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘͘ϭϮϵ
DĞŵŽƌLJͲƚLJƉĞŽŶƚƌŽůŚĂƌƚƐŝŶ^ƚĂƚŝƐƚŝĐĂůWƌŽĐĞƐƐŽŶƚƌŽů
ŚĂƉƚĞƌϭ
ͳ
This chapter provides an introduction to statistical process control () and its main
technique: the control chart. A brief description of the so called memory control charts,
including cumulative sum (CUSUM) and exponentially weighted moving average (EWMA)
control charts, is also given. Finally, a synopsis of the thesis containing the inspiration
towards the proposals and an outline are presented.
ͳǤͳ
Production processes are subject to variations, e.g. in the process of filling bottles
with cooking oil the amount of oil filled will not be exactly same; in the process of making
tube light rods the diameter or length of any two rods will not be the same. These variations
are mainly classified into two types, namely common cause variation and special cause
variation. Common cause variation always exists even if the process is designed very well
and maintained very carefully. This variation should be relatively small in magnitude and is,
uncontrollable and due to many small unavoidable causes. A process is said to be in statistical
control if only common cause variation is present. The variations outside this common cause
pattern are called special cause variations. These variations are subject to some problem in
the system, like poor tuning of equipment, controller fell asleep or got absent, computer
stopped working, poor lot of raw material, machine break down. A process working under
both types of variation is said to be out of control. The increase of variation (or the inclusion
of special cause variation) in the process generally changes the process parameters like
DĞŵŽƌLJͲƚLJƉĞŽŶƚƌŽůŚĂƌƚƐŝŶ^ƚĂƚŝƐƚŝĐĂůWƌŽĐĞƐƐŽŶƚƌŽů
ϭ
Ϯ
location or/and dispersion parameters. Change in the process location can lead to a greater
number of nonconforming items. Similarly, a change in the process dispersion is also
important to be detected as an increase in the process dispersion shows a straightforward drop
in the quality of process, as a larger spread in the data leads to lower uniformity in the
process. On the other hand, the detection of the decrease in the process dispersion may
improve the quality of the process, if the underlying special cause can be detected as early as
possible.
possesses some of the most extensively used techniques to detect the presence of
special cause variation in processes. The control chart is one of those techniques and it started
with Shewhart control charts containing the mean () chart for process location and the range
( ), the standard deviation () and the variance ( ) charts for process dispersion. The
structure of these control charts is based on a statistic plotted against three additional lines:
the center line ( ), the upper control limit ( ) and the lower control limit ( ). The two
control limits (i.e. and ) are basically the parameters of a control chart which are
selected in such a way that there is a very small probability (generally referred as False Alarm
Rate (
) in the quality control literature and denoted by ) of the in control data points
falling outside these limits. Similarly, the probability of the out of control data points falling
outside the control limits is called the power (used as a performance measure) of a control
chart. Another performance measure for the control charts is the average run length
we define a random variable
. If
equal to the number of samples until the first out of control
signal occurs then the probability distribution of this random variable
is known as the run
length distribution. The average of this distribution is called average run length and is
denoted by
control
. The in control
is denoted by
of a control chart is denoted by
.
DĞŵŽƌLJͲƚLJƉĞŽŶƚƌŽůŚĂƌƚƐŝŶ^ƚĂƚŝƐƚŝĐĂůWƌŽĐĞƐƐŽŶƚƌŽů
,
while out of
/ŶƚƌŽĚƵĐƚŝŽŶ
A shortcoming of a Shewhart control chart is that its conclusion is merely based on
the present sample which means that it pays no attention to the past data resulting into a
relatively bad performance for small disturbances in the process. In contrast, the CUSUM
control charts and the EWMA control charts (also referred as memory control charts) are
based (in different) ways on past information along with current. Due to this feature these
charts are more efficient to detect small and moderate shifts. The present study is based on
providing new memory control charting techniques (by modifying the existing structures and
also by designing some new structures) that perform relatively better than the existing ones,
especially for small and moderate shifts in the process parameters. In the pursuing sections,
we provide the detailed structures of CUSUM and EWMA charts for monitoring the process
parameters.
ͳǤʹ
The CUSUM chart was originally introduced by Page (1954) and is suited to detect
small and sustained shifts in a process. The chart measures a cumulative deviation from the
mean or a target value. There exist two versions of the CUSUM chart, used to monitor the
process location: the V-mask CUSUM and the tabular CUSUM. The V-mask procedure,
which is not very common in use, normalizes the deviations from the mean (or target) and
plots these deviations. As long as these deviations are plotted around the target value the
process is said to be in control, otherwise out of control. The tabular method of evaluating a
CUSUM chart works by accumulating the deviations up and down from a target value for
which we use the notations and , respectively. The quantities and are known as
upper and lower CUSUM statistics, respectively, and these are defined as:
",
!
"
!
(1.1)
DĞŵŽƌLJͲƚLJƉĞŽŶƚƌŽůŚĂƌƚƐŝŶ^ƚĂƚŝƐƚŝĐĂůWƌŽĐĞƐƐŽŶƚƌŽů
ϯ
ϰ
where # is the sample number, is the study variable, is the target mean of study variable
,
is the reference value of CUSUM scheme, often taken equal to the half of the amount of
shift which we are interested to detect (cf. Ewan and Kemp (1960)). The starting value for
both plotting statistics is taken equal to zero, i.e. . Now we plot these two
statistics against the control limit $ and it is concluded that the process mean has moved
upward if % $ for any value of # whereas the process mean is said to be shifted
downwards if % $ for any value of #. The CUSUM chart is defined by two parameters i.e.
and $. These two parameters are used in the standardized manner (cf. Montgomery
(2009)) given as:
&' ,
$ ('
(1.2)
where ' is the in control standard deviation of the study variable and & and ( are the two
constants which have to be chosen very carefully because the
performance of the
CUSUM chart is very sensitive to these constants.
Several CUSUM structures are also recommended for monitoring the process
dispersion. Page (1963) introduced the CUSUM chart for monitoring an increase in process
dispersion using sample ranges. Following him, Hawkins (1981), Tuprah and Ncube (1987),
Chang and Gan (1995), Acrosta-Mejia et al. (1999) and Castagliola et al. (2009) proposed
several improved versions of CUSUM charts for process dispersion. These charts are based
on transforming the sample variance such that the new transformed form may be closely
approximated by a normally distributed variable and hence applying the usual CUSUM
structures (recommended by Page (1954)) on it.
DĞŵŽƌLJͲƚLJƉĞŽŶƚƌŽůŚĂƌƚƐŝŶ^ƚĂƚŝƐƚŝĐĂůWƌŽĐĞƐƐŽŶƚƌŽů
/ŶƚƌŽĚƵĐƚŝŽŶ
ͳǤ͵
The EWMA control chart was introduced by Roberts (1959) to particularly address
the shifts of small and moderate magnitude. Like the CUSUM scheme, EWMA also utilizes
the past information along with the current, but the weights attached to the data are
exponentially decreasing as the observations become less recent. An EWMA control chart for
monitoring the location of a process is based on the statistic:
) * ! + *)
(1.3)
where # is the sample number and * is a constant such that , * - +. The quantity ) is the
starting value and it is taken equal to the target mean or the average of initial data in case
when the information on the target mean is not available. The control limits for the EWMA
statistic given in (1.3) are given as:
.
where
' /
0
0
+ + * 4
2
3
0
+ + * 2
! ' /
0
1
(1.4)
is the control limit coefficient. Like CUSUM charts, EWMA control charts also have
two parameters (* and ). * determines the decline of weights, while
of the control limits, so jointly these two parameters determine the
determines the width
performance of the
EWMA charts. The above mentioned limits given in (1.4) are called time varying limits of
the EWMA charts. For large values of # these limits converge to constant limits which are
given as:
' /
0
0
,
,
! ' /
0
0
(1.5)
DĞŵŽƌLJͲƚLJƉĞŽŶƚƌŽůŚĂƌƚƐŝŶ^ƚĂƚŝƐƚŝĐĂůWƌŽĐĞƐƐŽŶƚƌŽů
ϱ
ϲ
Hence, the factor 5+ + * 6 in (1.4) tends to 1 if the sample number tends to infinity.
For monitoring the process dispersion, EWMA charts are also based on normalizing
the sample variance. Wortham and Ringer (1971) suggested an EWMA control chart for
monitoring the process dispersion. Ng and Case (1989), Crowder and Hamilton (1992),
Castagliola (2005) and Huwang et al. (2010) followed them and proposed improved versions
of EWMA chart for monitoring process variance.
ͳǤͶ
After the development of Shewhart, CUSUM and EWMA charts by Shewhart (1931),
Page (1954) and Roberts (1959), respectively, several modifications of these charts have been
presented in order to further enhance the performance of these charts. Klein (2000), Khoo
(2004), Koutras et al. (2007) and Antzoulakos and Rakitzis (2008) proposed the application
of different runs rules with the Shewhart structure. Riaz (2008a) and Riaz (2008b) proposed
the auxiliary based control charts for monitoring the process variability and location
respectively, where both of these charts are based on regression-type estimators. Lucas
(1982) presented the combined Shewhart-CUSUM quality control scheme in which Shewhart
limits and CUSUM limits are used simultaneously. Lucas and Crosier (1982) recommended
the use of the fast initial response (FIR) CUSUM which gives a head start to the CUSUM
statistic by setting the initial values of the CUSUM statistic equal to some positive value
(non-zero). This feature gives better
performance but at the cost of a decrease in
.
Yashchin (1989) presented the weighted CUSUM scheme which gives different weights to
the previous information used in CUSUM statistic. Similarly, on the EWMA side, Lucas and
Saccucci (1990) presented the combined Shewhart-EWMA quality control scheme which
gives better
performance for both small and large shifts. Steiner (1999) provided the
DĞŵŽƌLJͲƚLJƉĞŽŶƚƌŽůŚĂƌƚƐŝŶ^ƚĂƚŝƐƚŝĐĂůWƌŽĐĞƐƐŽŶƚƌŽů
/ŶƚƌŽĚƵĐƚŝŽŶ
FIR EWMA which gives a head start to the initial value of the EWMA statistic (like FIR
CUSUM) and hence improves the
performance of the EWMA charts.
It is hard to find an application of the runs rules schemes with the CUSUM and
EWMA charts in the literature. Chapter 2 proposes the application of some runs rules
schemes with the control structure of CUSUM and EWMA charts for process location. The
performance of these runs rules based CUSUM and EWMA charts is evaluated in terms of
. Comparisons of the proposed schemes are made with some existing representative
CUSUM- and EWMA-type counterparts used for small and moderate shifts. The findings
reveal that the proposed schemes are able to perform better than the other schemes under
investigation. The work of Chapter 2 has been published in Quality and Reliability
Engineering International as Riaz, Abbas and Does (2011) and Abbas, Riaz and Does (2011).
Chapter 3 introduces a new control structure named as mixed EWMA-CUSUM
control chart for monitoring the process location. The core of this idea is to mix the effects of
EWMA and CUSUM charts into a single structure such that the resulting mixed chart
perform better than the classical ones (i.e. CUSUM by Page (1954) and EWMA by Roberts
(1959)). This is done by applying the CUSUM structure over the EWMA statistic. An
obvious counterpart of this mixed chart is also developed for monitoring the process
dispersion which is named as CS-EWMA chart as its plotting statistic is based on
cumulatively summing the exponentially weighted moving averages. Some additional
material in the form of comparisons and illustrative examples are also provided. From this
chapter, an article on the mixed EWMA-CUSUM chart for location is published in Quality
and Reliability Engineering International as Abbas, Riaz and Does (2012a), while another
article on CS-EWMA chart for process dispersion has been accepted for publication in
Quality and Reliability Engineering International as Abbas, Riaz and Does (2012b).
DĞŵŽƌLJͲƚLJƉĞŽŶƚƌŽůŚĂƌƚƐŝŶ^ƚĂƚŝƐƚŝĐĂůWƌŽĐĞƐƐŽŶƚƌŽů
ϳ
ϴ
Following the approach of Riaz (2008a) and Riaz (2008b), chapter 4 proposes several
new control structures using the information from auxiliary variable(s). These charts include
the CUSUM and EWMA charts (for monitoring the process location) based on the
information of one or more auxiliary variables. The regression estimation technique for the
mean is used in defining the control structure of the proposed charts. Comparisons with univariate as well as bi-variate EWMA and CUSUM charts are provided. An article on auxiliary
based EWMA chart for location has been accepted for publication in Communications in
Statistics - Theory and Methods as Abbas, Riaz and Does (2012c).
Chapter 5 proposes an alternative to the CUSUM and EWMA charts, named as the
progressive mean (7) control chart. This newly developed control chart is not only
outperforming the existing memory charts, but also, its control structure is very simple as
compared to the CUSUM and EWMA charts. An article on 7 control chart has been
published in Quality and Reliability Engineering International as Abbas, Zafar, Riaz and
Hussain (2012). Using the idea of a progressive statistic, two new control charts are also
developed (named as floating control charts) for monitoring the process dispersion. These
floating charts also surpass the existing CUSUM and EWMA charts for monitoring the
process standard deviation. An article on the floating charts has been submitted for
publication in International Journal of Production Research as Abbas, Riaz and Does
(2012d).
DĞŵŽƌLJͲƚLJƉĞŽŶƚƌŽůŚĂƌƚƐŝŶ^ƚĂƚŝƐƚŝĐĂůWƌŽĐĞƐƐŽŶƚƌŽů
ŚĂƉƚĞƌϮ
ʹ
The control chart is an important statistical technique that is used to monitor the quality of a
process. One of the charting procedures is the Shewhart-type control charts which are used
mainly to detect large shifts. Two alternatives to the Shewhart-type control charts are the
cumulative (CUSUM) control charts and the exponentially weighted moving average
(EWMA) control charts which are especially designed to detect small and moderately
sustained changes in quality. Enhancing the ability of design structures of control charts is
always desirable and one may do it in different ways. Runs rules schemes are generally used
to enhance the performance of Shewhart control charts. In this chapter we propose the use of
runs rules schemes for the CUSUM and EWMA charts and evaluated their performance in
terms of the
. Comparisons of the proposed schemes are made with some existing
representative CUSUM and EWMA-type counterparts used for small and moderate shifts.
The comparisons revealed that the proposed schemes perform better for small and moderate
shifts while they reasonably maintain their efficiency for large shifts as well. This chapter is
based on two articles for monitoring the process location i.e. Riaz, Abbas and Does (2011)
and Abbas, Riaz and Does (2011).
ʹǤͳ
The CUSUM and EWMA chart structures discussed in Chapter 1 are known as the
classical CUSUM and EWMA charts. The detailed study on the
performance of
DĞŵŽƌLJͲƚLJƉĞŽŶƚƌŽůŚĂƌƚƐŝŶ^ƚĂƚŝƐƚŝĐĂůWƌŽĐĞƐƐŽŶƚƌŽů
ϵ
ϭϬ
CUSUM chart was done by Hawkins and Olwell (1998), whereas Steiner (1999) evaluated
the
8 of the classical EWMA. These
respectively.
@
8 are provided in Table 2.1 and Table 2.2,
TABLE 2.1: 9:; Values for the classical CUSUM chart with < => ?
(A
(B
0
0.25
0.5
0.75
1
1.5
2
2.5
3
168
74.2
26.6
13.3
8.38
4.75
3.34
2.62
2.19
465
139
38.0
17.0
10.4
5.75
4.01
3.11
2.57
TABLE 2.2: 9:; values for the classical EWMA chart at 9:;= ?==
C
0
D => E
; F> GFH
499.89
D => F?
;I
500.81
D => ?
; I> =JF
499.36
D => J?
; I> =GG
0.25
102.99
169.49
255.96
321.3
0.5
28.86
47.5
88.75
139.87
0.75
13.56
19.22
35.55
62.46
1
8.22
10.4
17.09
30.57
1.5
4.17
4.77
6.27
9.8
2
2.66
2.94
3.4
4.46
499.36
The results of Tables 2.1 and 2.2 are based on the test that one point falling outside
the limits indicates an out of control situation (the classical scheme of signaling). This test
may be further extended to a set of rules named as sensitizing rules and runs rules schemes
which help to increase the sensitivity of the charts to detect out of control situations. The
common set of sensitizing rules are (cf. Nelson (1984)): one or more points outside the
control limits; two out of three consecutive points outside the 2 sigma warning limits but still
inside the control limits; four out of five consecutive points beyond the 1 sigma limits but still
inside the control limits; a run of eight consecutive points on one side of the center line but
still inside the control limits; six points in a row steadily increasing or decreasing but still
inside the control limits; fourteen points in a row alternating up and down but still inside the
DĞŵŽƌLJͲƚLJƉĞŽŶƚƌŽůŚĂƌƚƐŝŶ^ƚĂƚŝƐƚŝĐĂůWƌŽĐĞƐƐŽŶƚƌŽů
ZƵŶƐƌƵůĞƐďĂƐĞĚh^hDĂŶĚtD
control limits. The basic principle underlying these runs rules is twofold. Firstly, specific
patterns of out of control conditions might be detected earlier, such as a small but persistent
trend. Secondly, the decision rules are designed to have roughly the same (marginal) false
alarm probability.
To enhance the performance of control charts, many researchers have used the idea of
using different sensitizing rules and runs rules schemes with the Shewhart-type control
charts, e.g. see Klein (2000), Khoo (2004), Koutras et al. (2007), and Antzoulakos and
Rakitzis (2008). The application of sensitizing rules causes an increase in false alarm rates,
whereas the runs rules schemes take care of this issue. Klein (2000), Khoo (2004), and
Antzoulakos and Rakitzis (2008) suggested different runs rules schemes, namely K out of L
and modified K out of L, to be used with the Shewhart-type control charts. They studied their
performance and found that these runs rules schemes perform better as compared to the usual
Shewhart-type control charts.
There is a variety of literature available on CUSUM and EWMA charts. e.g. see
Lucas and Crosier (1982), Yashchin (1989), and Hawkins and Olwell (1998) for CUSUM
and Lucas and Saccucci (1990), Steiner (1999), and Capizzi and Masarotto (2003) for
EWMA. All the existing approaches use only the usual scheme of signaling an out of control
situation. It is hard to find an application of the runs rules schemes with the CUSUM and
EWMA charts in the literature. However, Westgard et al. (1977) studied some control rules
using combined Shewhart-CUSUM structures. They proved superiority of this combined
approach on the separate Shewhart’s approach but ignored any comparison with the separate
CUSUM application. Also their control rules considered only one point at a time for testing
an out of control situation. The false alarm rates of their control rules kept fluctuating and no
attempt was made to keep them fixed at a pre-specified level which is very important for
valid comparisons among different control rules/schemes.
DĞŵŽƌLJͲƚLJƉĞŽŶƚƌŽůŚĂƌƚƐŝŶ^ƚĂƚŝƐƚŝĐĂůWƌŽĐĞƐƐŽŶƚƌŽů
ϭϭ
ϭϮ
In this chapter we analyze some of the K out of L runs rules schemes (like MNM, MOP and
modified MOP schemes) with CUSUM and EWMA charts, following Klein (2000), Khoo
(2004), and Antzoulakos and Rakitzis (2008), and compared their performance (in terms of
the
) with some other schemes meant particularly for small shifts. Section 2.2 contains
the detailed discussion about the proposed runs rules schemes applied on the CUSUM chart,
while the discussion on the schemes applied on the EWMA chart are included in Section 2.3.
ʹǤʹ
A process is called to be out of control when a point falls outside the control limits.
Specific runs rules or extra sensitizing rules can be used in addition to enhance the power of
detecting out of control situations. The CUSUM charts can also take benefit out of these runs
rules schemes if properly applied with the CUSUM structures. Following Klein (2000), Khoo
(2004), and Antzoulakos and Rakitzis (2008), we propose here two runs rules schemes to be
used with the CUSUM charts to monitor the location parameter. The proposed schemes are
based on the following terms and definitions.
Action Limit (9;): This is a threshold level for the value of CUSUM chart statistic. If some
value of CUSUM statistic exceeds the
of the
, the process is called to be out of control. The value
would be greater than the classical CUSUM critical limit $ for a fixed
.
Warning Limit (Q;): This is a level for the value of the CUSUM chart statistic beyond
which (but not crossing the
) some pattern of consecutive points indicate an out of control
situation. The value of the R would be smaller than the classical CUSUM critical level $
for a fixed
.
DĞŵŽƌLJͲƚLJƉĞŽŶƚƌŽůŚĂƌƚƐŝŶ^ƚĂƚŝƐƚŝĐĂůWƌŽĐĞƐƐŽŶƚƌŽů
ZƵŶƐƌƵůĞƐďĂƐĞĚh^hDĂŶĚtD
Using the above definitions we propose the two runs rules schemes for the CUSUM chart as:
Scheme I: A process is said to be out of control if one of the following four conditions is
satisfied:
1.
2.
3.
4.
One point of falls above the
One point of falls above the
.
.
Two consecutive points of fall between the R and the
.
Two consecutive points of fall between the R and the
.
Scheme II: A process is said to be out of control if one of the following four conditions is
satisfied:
1.
2.
3.
4.
One point of falls above the
One point of falls above the
Two out of three consecutive points of fall between the R and the
given shift; i.e. the
are proportional to the value of the
is higher if the values of the R and the
There are infinite pairs of R and
The objective is to find those pairs of
level and at the same time minimize the
The
.
Two out of three consecutive points of fall between the R and the
Note that the values of the R and the
versa.
.
.
for a
are higher and vice
which fix the in control
and R that maintain the
.
at a desired level.
value at the desired
value.
computations may be carried out using different approaches, like integral
equations, Markov chains, approximations and Monte Carlo simulations. Details regarding
the first two may be seen in Brook and Evans (1972) and Lucas and Crosier (1982) and the
references therein. An
lower-sided CUSUM (
approximation for the upper-sided CUSUM (
) and the
) is given as (cf. Siegmund (1985)):
DĞŵŽƌLJͲƚLJƉĞŽŶƚƌŽůŚĂƌƚƐŝŶ^ƚĂƚŝƐƚŝĐĂůWƌŽĐĞƐƐŽŶƚƌŽů
ϭϯ
ϭϰ
and
The
S TUV>WW ! M@ &( ! +>+XX +
M@ &
S TUV>WW ! M@ &( ! +>+XX +
M@ &
8 for a two-sided CUSUM can be obtained by the following relation:
+Y
+
! +Y
Monte Carlo simulation is also a standard option to obtain approximations for the
and
we have adopted this approach in our study. For that purpose we have developed a simulation
algorithm using an add-in feature of Excel software which helps calculating the
8.
ʹǤʹǤͳ
The performance of the two proposed schemes for the CUSUM chart has been
evaluated in terms of
under different in control and out of control situations. To meet the
desired objective, we have used our simulation algorithm to find the
for the pairs of R and
size Z from [ \ ! @
]^
_`
and
values
. We have generated for different values of @ 100,000 samples of
' a and we have calculated the statistics and for all
samples. Here and ' refer to the mean and the standard deviation of the process under
study and @ is the amount of shift in . Here ' is assumed to be in control i.e. ' ' . The
value of @ indicates the state of control for our process mean, i.e. @ implies that the
process mean is in control (i.e. ) and @ b that the process mean is out of control
(i.e. ). Without loss of generality we have taken and ' + in our simulations.
After obtaining 100,000 samples we have applied all four conditions of the two proposed
runs rules schemes (i.e. Schemes I & II). In this way the run lengths are found for the two
DĞŵŽƌLJͲƚLJƉĞŽŶƚƌŽůŚĂƌƚƐŝŶ^ƚĂƚŝƐƚŝĐĂůWƌŽĐĞƐƐŽŶƚƌŽů
ZƵŶƐƌƵůĞƐďĂƐĞĚh^hDĂŶĚtD
proposed schemes. This procedure is repeated 5,000 times and each time the run lengths are
computed for both schemes. By taking the average of these run lengths we obtain the
for the two schemes.
8
To evaluate the performance of the two proposed schemes I and II, we will report the
results for the values of the
equal to 168, 200 and 500. Other values of the
can be
easily obtained. The choices made will show the performance of the two schemes and enable
us to make comparisons with the results of other schemes and approaches from the literature.
By fixing the
pairs of R
obtained the
the
8
and
at a desired level for the proposed schemes I and II, we are able to obtain
using our algorithm. Then for these pairs of R and
, we have
at different values of @ for both the schemes. The results of the R and
along with their corresponding
values for the above mentioned pre-specified
are provided in Tables 2.3 – 2.8 for both schemes. In Tables 2.3 – 2.8 the first two
columns contain the R and
pairs which fix the
and the remaining columns give the corresponding
value at a specified desired level
values.
Some researchers (e.g. Antzoulakos and Rakitzis (2008)) suggest also to report the
standard deviations of the run lengths along with the
values to describe more about the
run length behavior. Moreover, Palm (1990) and Shmueli and Cohen (2003) highlighted the
importance of percentile points of the run length distribution and suggested to report them for
the interest of practitioners. Therefore the standard deviations (denoted by c
) and the
percentile points (denoted by ) of the run length distribution are also computed for
proposed schemes I and II. The results of c
and (for # + MB B dB e) are
provided in Riaz, Abbas and Does (2011) as Tables VIII – XI for the two proposed schemes
at
+Xf. For the other values of
similar tables can be easily obtained.
DĞŵŽƌLJͲƚLJƉĞŽŶƚƌŽůŚĂƌƚƐŝŶ^ƚĂƚŝƐƚŝĐĂůWƌŽĐĞƐƐŽŶƚƌŽů
ϭϱ
ϭϲ
TABLE 2.3: Q;, 9; and 9:;E values for the proposed scheme I at 9:;= EgG
Q;h
Limits
0.25
0.5
0.75
3.42
9;h
4.8
71.872
25.564
3.44
4.6
72.258
3.48
4.4
3.53
4.2
C
1
1.5
2
13.539
8.66
5.078
3.679
25.653
13.5
8.568
5.013
3.607
71.936
25.593
13.496
8.516
4.936
3.525
71.399
25.3
13.332
8.404
4.828
3.423
TABLE 2.4: Q;, 9; and 9:;E values for the proposed scheme II at 9:;= EgG
Q;h
Limits
0.25
0.5
0.75
3.5
9;h
4.44
71.489
25.379
3.6
4.19
72.938
25.368
3.7
4.08
73.11
25.369
3.8
4.03
73.589
25.403
C
1
1.5
2
13.398
8.462
4.941
3.541
13.352
8.383
4.83
3.424
13.306
8.344
4.777
3.376
13.277
8.316
4.75
3.347
TABLE 2.5: Q;, 9; and 9:;E values for the proposed scheme I at 9:;= F==
Q;h
Limits
0.25
0.5
0.75
3.9
9;h
4.24
82.952
28.697
3.8
4.29
84.537
3.7
4.4
3.6
4.77
i
3.57
C
1
1.5
2
13.801
8.904
4.909
3.492
28.736
14.065
8.661
5.002
3.523
82.115
28.496
13.822
8.812
4.993
3.561
84.15
28.716
13.961
8.981
5.228
3.725
79.474
28.94
14.262
9.213
5.51
4.076
TABLE 2.6: Q;, 9; and 9:;E values for the proposed scheme II at 9:;= F==
Q;h
Limits
C
9;h
0.25
0.5
0.75
1
1.5
2
3.9
4.23
82.975
28.206
13.77
8.869
4.978
3.435
3.8
4.28
81.403
28.356
13.903
8.699
4.966
3.503
3.7
4.6
82.732
28.824
14.078
8.874
5.14
3.687
81.518
29.138
14.266
9.114
5.457
4.12
3.64
i
DĞŵŽƌLJͲƚLJƉĞŽŶƚƌŽůŚĂƌƚƐŝŶ^ƚĂƚŝƐƚŝĐĂůWƌŽĐĞƐƐŽŶƚƌŽů
ZƵŶƐƌƵůĞƐďĂƐĞĚh^hDĂŶĚtD
TABLE 2.7: Q;, 9; and 9:;E values for the proposed scheme I at 9:;= ?==
Q;h
Limits
0.25
0.5
0.75
4.8
9;h
5.12
141.111
38.599
4.7
5.2
150.372
4.6
5.39
i
4.49
C
1
1.5
2
17.392
10.518
5.905
4.057
38.594
17.529
10.599
5.898
4.14
145.189
38.195
17.468
10.558
6.007
4.237
146.564
38.492
17.725
10.857
6.333
4.689
TABLE 2.8: Q;, 9; and 9:;E values for the proposed scheme II at 9:;= ?==
Q;h
Limits
0.25
0.5
0.75
4.8
9;h
5.11
139.705
38.856
4.7
5.19
142.159
4.6
5.5
4.54
i
C
1
1.5
2
17.459
10.506
5.822
4.078
37.975
17.267
10.583
5.872
4.104
145.787
38.334
17.394
10.734
6.053
4.273
149.035
39.904
17.568
10.966
6.451
4.873
The standard errors of the results reported in Tables 2.3 – 2.8 are expected to remain
around 1% (in relative terms) as we have checked by repeating our simulation results. We
have also replicated the results Table 2.1 for ( A of the classical CUSUM scheme using
our simulation routine and obtained almost the same results which ensures the validity of the
algorithm developed in Excel and the simulation results obtained from it.
We have observed for the two proposed schemes I and II that:
i)
many pairs of R and
may be found which fix the
but the optimum choice helps minimizing the
ii)
at a desired level
value (cf. Table 2.3 – 2.8);
the two proposed schemes perform very good at detecting small and moderate
shifts while maintaining their ability to address the large shifts as well (cf. Tables 2.3
– 2.8);
DĞŵŽƌLJͲƚLJƉĞŽŶƚƌŽůŚĂƌƚƐŝŶ^ƚĂƚŝƐƚŝĐĂůWƌŽĐĞƐƐŽŶƚƌŽů
ϭϳ
ϭϴ
iii)
the two proposed schemes I and II are almost equally efficient for the shifts of
small, moderate and large magnitude and hence may be used as a replacement of each
other at least for normally distributed processes;
iv)
with an increase in the value of @ the
;
schemes, at a fixed value of
the
decreases rapidly for both the
v)
with a decrease in the value of
vi)
the proposed schemes I and II may be extended to more generalized schemes
schemes for a given value of @ (cf. Tables 2.3 – 2.8);
decreases quickly for both the
(as given in Klein (2000), Khoo (2004), and Antzoulakos and Rakitzis (2008));
vii)
the c
decreases if the value of @ increases for both the schemes I and II
(cf. Riaz, Abbas and Does (2011));
viii)
the run length distributions of both the schemes are positively skewed (cf.
Riaz, Abbas and Does (2011)).
ʹǤʹǤʹ
In this section we compare the performance of the proposed schemes I and II with
some existing schemes for detecting small, moderate and large shifts. The
is used as a
performance measure for all the schemes under discussion. The existing schemes we have
considered for comparison purpose include the classical CUSUM scheme of Page (1954), the
weighted CUSUM scheme of Yashchin (1989), the EWMA scheme given in Steiner (1999)
and the fast initial response (FIR) CUSUM scheme of Lucas and Crosier (1982). The
results for the above mentioned schemes are provided in the Table 2.1 – 2.2 and Tables 2.9 –
2.10 at some selective values of
which will be used for the comparisons.
DĞŵŽƌLJͲƚLJƉĞŽŶƚƌŽůŚĂƌƚƐŝŶ^ƚĂƚŝƐƚŝĐĂůWƌŽĐĞƐƐŽŶƚƌŽů
ZƵŶƐƌƵůĞƐďĂƐĞĚh^hDĂŶĚtD
Now we present a comparative analysis of the proposed schemes with the existing schemes
one by one.
Proposed versus the classical CUSUM: The classical CUSUM scheme of Page (1954)
accumulates the up and down deviations from the target and is quite efficient at detecting
small shifts. Table 2.1 provides the
Tables 2.3 – 2.4 provide the
performance of the classical CUSUM scheme.
performances of the two proposed schemes. The results of
these tables advocate that the proposed schemes are better compared with the classical
CUSUM scheme for small shifts while for moderate and large shifts their performances
almost coincide.
Proposed versus the weighted CUSUM: Yashchin (1989) presented a class of weighted
control schemes that generalize the basic CUSUM technique by assigning different weights
to the past information used in the classical CUSUM statistic. The
performance of the
terms as defined earlier in this article. Tables 2.7 and 2.8 provide the
performance of the
weighted CUSUM scheme is given in Table 2.9 where j represents the weight and the other
proposed schemes at
B so these tables can be used to compare the proposed
schemes with the weighted CUSUM scheme.
TABLE 2.9: 9:; values for the symmetric two-sided weighted CUSUM scheme at 9:;= ?==
k
< => ?
0.5
1
0.7
l
3.16
86.30
0.8
3.46
0.9
1.0
C
1.5
2
15.90
6.08
3.52
70.20
13.30
5.66
3.50
3.97
54.40
11.40
5.50
3.60
5.09
39.00
10.50
5.81
4.02
By comparing the results of Tables 2.7, 2.8 and 2.9 we can see that the proposed schemes
perform better than the weighted CUSUM for small and moderate shifts. Particularly, when @
DĞŵŽƌLJͲƚLJƉĞŽŶƚƌŽůŚĂƌƚƐŝŶ^ƚĂƚŝƐƚŝĐĂůWƌŽĐĞƐƐŽŶƚƌŽů
ϭϵ
ϮϬ
is small the performance of our proposed schemes is significantly better than that of the
weighted CUSUM scheme. However for j + the weighted CUSUM scheme is the same as
the classical CUSUM so the comments of the proposed versus the classical CUSUM scheme
hold here as well.
Proposed versus the classical EWMA: Steiner (1999) gave a simple method for studying
the run length distribution of the classical EWMA chart. Table 2.2 presents some selective
8 of the EWMA chart where * is the weighting constant and
coefficient. As the proposed schemes have
is the control limits
B in Tables 2.7 and 2.8 we use these
tables for a comparison with the EWMA chart. From the Tables 2.2, 2.7 and 2.8 we see that
for @ >MB the classical EWMA chart has an
schemes are minimizing the same
value of +d+>e whereas the proposed
value around +A. This shows that the proposed
schemes perform better than the classical EWMA scheme for @ >MB. The same superiority
also holds for all >MB - @ - +. However, for @ % + the proposed schemes and the classical
EWMA scheme have almost the same behavior as can be easily seen from the corresponding
tables.
Proposed versus the FIR CUSUM: Lucas and Crosier (1982) presented the Fast Initial
Response (FIR) CUSUM which gives a head start value, say , to the classical CUSUM
statistic. A standard CUSUM has while an FIR CUSUM sets and to
8 for the FIR CUSUM at ( A and +
some nonzero value. Table 2.10 presents the
for discussion and comparison purposes. The FIR CUSUM scheme decreases the
values as compared to those of the classical CUSUM scheme at the cost of reduction in
value from +Xf to +XP (see Table 2.1 vs. Table 2.10) which is generally undesirable in
sensitive processes (e.g. those directly related to intensive care units which are highly time
sensitive, cf. Bonetti et al. (2000)). The
results given in Tables 2.3 and 2.4 of the
DĞŵŽƌLJͲƚLJƉĞŽŶƚƌŽůŚĂƌƚƐŝŶ^ƚĂƚŝƐƚŝĐĂůWƌŽĐĞƐƐŽŶƚƌŽů
ZƵŶƐƌƵůĞƐďĂƐĞĚh^hDĂŶĚtD
proposed schemes are also obtained for
purposes here. Looking at the
same amount of reduction in
+Xf and hence can be used for comparison
results of Tables 2.3 and 2.4 we can see that almost the
may be achieved, as obtained by FIR CUSUM, using the
proposed schemes without paying any cost in terms of a decrease in
value and the need
of a head start value.
C
TABLE 2.10: 9:;m for FIR CUSUM scheme with n= lOH and < => ?
l H, n= E
0
0.25
0.5
0.75
1
1.5
2
163
71.1
24.4
11.6
7.04
3.85
2.7
In brief the proposed schemes have shown better performance for the smaller values of @ (i.e.
small shifts), which is the main concern of CUSUM charts, while for larger values of @ the
proposed schemes can perform equally well as the other schemes. The better performance can
be further enhanced with the help of other runs rules schemes of Khoo (2004) and
Antzoulakos and Rakitzis (2008).
ʹǤʹǤ͵
To illustrate the application of the proposed CUSUM schemes we use the same
method as in Khoo (2004). Two datasets are simulated consisting of some in control and
some out of control sample points. For dataset 2.1 we have generated 50 observations in total,
of which the first 20 observations are from [+ (showing the in control situation) and the
remaining 30 observations are generated from [>MB+ (showing a small shift in the mean
level) while for dataset 2.2 we have generated 30 observations in total, of which the first 20
observations are same as for dataset 2.1 and the remaining 10 observations are generated
from [++ (showing a moderate shift in the mean level).
DĞŵŽƌLJͲƚLJƉĞŽŶƚƌŽůŚĂƌƚƐŝŶ^ƚĂƚŝƐƚŝĐĂůWƌŽĐĞƐƐŽŶƚƌŽů
Ϯϭ
ϮϮ
The two proposed CUSUM schemes of this study (i.e. schemes I and II) are applied to
the above mentioned two datasets. Additionally, the classical CUSUM scheme is also applied
to these two datasets for illustration and comparison purposes. The CUSUM statistics are
computed for the two datasets and are plotted against the respective control limits used with
the three CUSUM schemes by fixing the
( B>d is used as the control limit to have
R A>f
and
at B. For the classical CUSUM scheme,
B. For the proposed scheme I,
B>+M are used, while for the proposed scheme II R A>f and
B>++ are used to have the
value equal to B for both schemes. The graphical
displays of the three CUSUM schemes for the two datasets are given in the following two
figures.
Figure 2.1 exhibits the behavior of dataset 2.1, where a small mean shift was
introduced. Figure 2.2 illustrates the behavior of dataset 2.2, where a moderate mean shift
was introduced.
Figure 2.1: CUSUM chart of the classical scheme and the proposed scheme I and II for
dataset 2.1
н
ϲ
,;ĐůĂƐƐŝĐĂůͿ
t>ĂŶĚ>;ƐĐŚĞŵĞ/Ϳ
t>ĂŶĚ>;ƐĐŚĞŵĞ//Ϳ
ϱ
ϰ
ϯ
Ϯ
ϭ
Ϭ
ϭ
ϯ
ϱ
ϳ
ϵ ϭϭ ϭϯ ϭϱ ϭϳ ϭϵ Ϯϭ Ϯϯ Ϯϱ Ϯϳ Ϯϵ ϯϭ ϯϯ ϯϱ ϯϳ ϯϵ ϰϭ ϰϯ ϰϱ ϰϳ ϰϵ
^ĂŵƉůĞEƵŵďĞƌ
DĞŵŽƌLJͲƚLJƉĞŽŶƚƌŽůŚĂƌƚƐŝŶ^ƚĂƚŝƐƚŝĐĂůWƌŽĐĞƐƐŽŶƚƌŽů
ZƵŶƐƌƵůĞƐďĂƐĞĚh^hDĂŶĚtD
Figure 2.2: CUSUM chart of the classical scheme and the proposed scheme I and II for
dataset 2.2
ϲ
н
,;ĐůĂƐƐŝĐĂůͿ
t>ĂŶĚ>;ƐĐŚĞŵĞ/Ϳ
t>ĂŶĚ>;ƐĐŚĞŵĞ//Ϳ
ϱ
ϰ
ϯ
Ϯ
ϭ
Ϭ
ϭ Ϯ ϯ ϰ ϱ ϲ ϳ ϴ ϵ ϭϬ ϭϭ ϭϮ ϭϯ ϭϰ ϭϱ ϭϲ ϭϳ ϭϴ ϭϵ ϮϬ Ϯϭ ϮϮ Ϯϯ Ϯϰ Ϯϱ Ϯϲ Ϯϳ Ϯϴ Ϯϵ ϯϬ
^ĂŵƉůĞEƵŵďĞƌ
In Figure 2.1 we see that an out of control signal is received at sample points # 49 and 50 by
the proposed scheme I (i.e. two out of control signals) and at sample points # 49 and 50 by
the proposed scheme II (i.e. two out of control signals). Similarly, in Figure 2.2 the out of
control signals are received at sample point # 26 by proposed scheme I (i.e. one out of control
signal) and at sample points # 25, 26 and 27 by the proposed scheme II (i.e. two out of
control signal). For both datasets, the classical CUSUM scheme failed to detect any shift in
the process mean.
It is evident from the above figures that the proposed schemes have detected out of
control signals, which are not spotted by the classical CUSUM scheme for the dataset 2.1,
where a small mean shift was present, while the situation is almost identical for dataset 2.2,
where a moderate shift we introduced. It is to be noted that these signaling performances of
the proposed schemes versus the classical CUSUM scheme are in accordance with the
DĞŵŽƌLJͲƚLJƉĞŽŶƚƌŽůŚĂƌƚƐŝŶ^ƚĂƚŝƐƚŝĐĂůWƌŽĐĞƐƐŽŶƚƌŽů
Ϯϯ
Ϯϰ
findings of subsection 2.2.2, where we found that the proposed schemes are more efficient
than the classical CUSUM scheme for small shifts, while almost equally good for other shifts.
ʹǤ͵
Shewhart control charts are good in detecting large disturbances in the process, but it
takes too long for Shewhart-type charts to detect a small or moderate shift. To overcome this
problem some sensitizing rules are designed but their implementation inflates the prespecified false alarm rate. This issue may be resolved by the introduction of the runs rules
schemes as we have mentioned in Section 2.2. Klein (2000), Khoo (2004) and Antzoulakos
and Rakitzis (2008) presented runs rules schemes applied on the Shewhart control charts to
enhance their performance for small and moderate shifts, keeping the false alarm rate at the
pre-specified level. The application of these runs rules schemes is not commonly used with
the CUSUM and EWMA control charts. Taking inspiration from the application of runs rules
on CUSUM charts, we propose two runs rules schemes in this section for the design structure
of the EWMA control chart named as “simple MOM EWMA scheme” and “modified MOP
EWMA scheme”. The procedural and conceptual framework of these two proposed schemes
is defined as:
Simple FOF EWMA scheme: According to the simple MOM EWMA scheme a process is said
to be out of control if two consecutive points are plotted either below a lower signaling limit
( ) or above an upper signaling limit ( ).
Modified FOI EWMA scheme: According to the modified MOP EWMA scheme a process is
said to be out of control if one of the following two conditions is satisfied.
i.
At least two out of three consecutive points fall below an and the point above the
(if any) falls between the and the .
DĞŵŽƌLJͲƚLJƉĞŽŶƚƌŽůŚĂƌƚƐŝŶ^ƚĂƚŝƐƚŝĐĂůWƌŽĐĞƐƐŽŶƚƌŽů
ZƵŶƐƌƵůĞƐďĂƐĞĚh^hDĂŶĚtD
At least two out of three consecutive points fall above a and the point below the
ii.
(if any) falls between and the .
The signaling limits and mentioned above in the definitions of our two proposals
are especially set limits chosen for the two schemes separately, depending upon the desired
,
while
is the same as defined in (1.4). The control structure for the proposed
schemes is given as:
.
where
0
o ' /0 +
+ * 4
2
3
0
+ + * 2
! o ' /
0
1
o
(2.1)
is the signaling limit coefficient of the proposed schemes and the other terms are the
same as defined in Section 1.3. It has to be mentioned that the above mentioned signaling
limits coefficient
o
.
is set according to the pre-specified value of
Moreover, a
signaling limit on either side may be split into two lines (as is done in Section 2.2) to reach at
some optimum pair. We opted the choice where the outer split of the line is taken at infinity.
However, one may take some different appropriately chosen outer splits other than infinity.
The parameters of these two proposed schemes are the central line and two signaling
limits as given in (2.1) (i.e. , and ). The upper and lower signaling limits are
symmetric around the
positive relation between
.
Using the
and width of the signaling limits (depending upon
we fix
and will vary according to the pre-specified
at the desired level and find the corresponding pair of symmetric signaling limits.
Based on these especially set signaling limits we carry out our
o ),
study at the desired
values.
The calculation of
may be carried out using different approaches such as integral
equations, Markov chains, approximations and Monte Carlo simulations. We have chosen to
DĞŵŽƌLJͲƚLJƉĞŽŶƚƌŽůŚĂƌƚƐŝŶ^ƚĂƚŝƐƚŝĐĂůWƌŽĐĞƐƐŽŶƚƌŽů
Ϯϱ
Ϯϲ
use Monte Carlo simulations to obtain
calculation of
values. The simulation algorithms for the
values of the proposed schemes are developed in Excel using an Add-In
feature 7#L.
ʹǤ͵Ǥͳ
To investigate the performance of our proposed EWMA schemes we have considered
different in control and out of control situations. A suitable number of samples (say 100,000)
of a fixed size Z are generated from [ \ ! @
]^
_`
' a. The EWMA statistics for these
samples are then calculated and the conditions of the two proposed EWMA schemes (as
listed in Section 2.3) are applied on them using the signaling limits given in (2.1) through our
simulation algorithm. By executing this process repeatedly we obtain different run length
hvalues and other properties as well. It is to be
values which ultimately help computing
noted that the value of
@ , the
@ b , the
o
is worked out such that the desired
value is achieved. For
values are evaluated with the help of their corresponding
o
and then, for
values are computed by introducing different shifts in the process.
To evaluate the performance of the two proposals we fix the pre-specified
values, in this section, at 168, 200 and 500. These choices will suffice to exhibit the behavior
of our proposed schemes and will enable us to make valid comparisons with their already
existing counterparts. On similar lines other choices of
the
can also be obtained. By fixing
values at the above mentioned levels (using their corresponding
obtained the
at different values of @. These
2.16 for the aforementioned desired
values) at different choices of *͘
o)
we have
values are provided in Tables 2.11 –
preferences (along with their corresponding
DĞŵŽƌLJͲƚLJƉĞŽŶƚƌŽůŚĂƌƚƐŝŶ^ƚĂƚŝƐƚŝĐĂůWƌŽĐĞƐƐŽŶƚƌŽů
o
ZƵŶƐƌƵůĞƐďĂƐĞĚh^hDĂŶĚtD
TABLE 2.11: 9:; values for the simple FOF EWMA scheme at 9:;= EgG
0
* >+
M>+AB
169.8676
* >MB
o M>+fA
169.4769
* >B
M>PA
169.6763
169.7112
0.25
54.5771
73.4836
94.8246
110.7766
0.5
19.8026
26.6284
37.9883
49.063
0.75
10.5927
12.9547
17.408
23.2892
1
6.9435
7.9456
9.8833
13.0739
1.5
4.1117
4.3476
4.7293
5.5777
2
2.9796
3.0954
3.1231
3.3368
@
o
* >dB
o +>fP
o
TABLE 2.12: 9:; values for the modified FOI EWMA scheme at 9:;= EgG
0
* >+
+>fd
167.3173
* >MB
o +>ePX
169.9927
* >B
o +>fB
168.5416
170.9464
0.25
34.367
43.3236
55.4569
62.6484
0.5
14.0389
17.8611
23.0836
27.6195
0.75
8.1338
9.4968
11.9229
14.2207
1
5.7064
6.4798
7.6037
8.5804
1.5
3.7755
3.9823
4.2103
4.5232
2
3.2047
3.2708
3.3136
3.4072
@
o
* >dB
o +>Xd
TABLE 2.13: 9:; values for the simple FOF EWMA scheme at 9:;= F==
0
* >+
M>M++
200.5694
* >MB
o M>MA
199.8855
* >B
o M>e
200.8923
201.1229
0.25
60.9801
80.7515
107.709
126.1691
0.5
20.9561
28.6011
42.0774
55.7335
0.75
11.2452
13.7306
19.0848
25.862
1
7.1859
8.1962
10.6258
13.9457
1.5
4.198
4.4825
4.9002
5.7061
2
3.0577
3.1331
3.1899
3.4272
@
o
The standard deviation of the run lengths (denoted by c
by (# +, MB, B, dB and e) are also provided at
) and the # pV percentiles denoted
B in Abbas, Riaz and Does
(2011). Similar results can be easily obtained for other values of
with
* >dB
o +>fdB
.
These measures along
may help studying the behavior of the run length distribution.
DĞŵŽƌLJͲƚLJƉĞŽŶƚƌŽůŚĂƌƚƐŝŶ^ƚĂƚŝƐƚŝĐĂůWƌŽĐĞƐƐŽŶƚƌŽů
Ϯϳ
Ϯϴ
TABLE 2.14: 9:; values for the modified FOI EWMA scheme at 9:;= F==
0
* >+
+>feB
201.9206
* >MB
o M>f
200.5236
* >B
+>eM
200.5969
* >dB
o +>d+B
0.25
39.1578
50.0565
63.1482
70.9068
0.5
15.4204
19.5705
25.1144
30.4725
0.75
8.6061
10.1922
12.8642
15.3578
1
5.981
6.7621
7.7437
9.1272
1.5
3.8655
4.1115
4.3324
4.6841
2
3.2523
3.3186
3.3386
3.4238
@
o
o
200.886
TABLE 2.15: 9:; values for the simple FOF EWMA scheme at 9:;= ?==
@
* >+
M>BBX
o
* >MB
o M>BBA
* >B
o M>PX
* >dB
o M>++B
0
501.7558
505.5284
501.2598
502.0725
0.25
103.3109
169.1349
235.1138
280.6187
0.5
29.5748
47.0105
78.0771
108.8792
0.75
14.3216
19.2776
30.8742
45.3405
1
8.9561
10.5964
15.1992
22.1033
1.5
4.9197
5.2578
6.1014
7.7862
2
3.4498
3.5527
3.6815
4.0883
TABLE 2.16: 9:; values for the modified FOI EWMA scheme at 9:;= ?==
0
* >+
o M>P
502.883
* >MB
o M>PAB
499.6153
505.3564
501.9698
0.25
66.6864
97.0108
133.7117
155.7078
0.5
21.4251
31.2023
46.3541
57.7739
0.75
11.7427
14.4295
20.6223
26.0312
1
7.5539
8.6761
11.0991
13.8363
1.5
4.4676
4.7066
5.1336
5.7812
2
3.4534
3.549
3.6276
3.7787
@
* >B
M>MM
o
* >dB
o +>efM
The relative standard errors of the results reported in Tables 2.11 – 2.16 are also calculated
and are found to be around 1%. We have also replicated the results of the classical EWMA
DĞŵŽƌLJͲƚLJƉĞŽŶƚƌŽůŚĂƌƚƐŝŶ^ƚĂƚŝƐƚŝĐĂůWƌŽĐĞƐƐŽŶƚƌŽů
ZƵŶƐƌƵůĞƐďĂƐĞĚh^hDĂŶĚtD
chart and found almost the same results as Steiner (1999) which ensures the validity of our
simulation algorithm.
Mainly, the findings for the two proposed schemes are:
i.
the two proposed schemes are performing very well at detecting small and
moderate shifts while their performance for large shifts is not bad either (cf.
Tables 2.11– 2.16);
ii.
with an increase in the value of @ the
a given
iii.
iv.
v.
vi.
vii.
viii.
decreases rapidly for both schemes, at
(cf. Tables 2.11– 2.16);
with a decrease in the value of
the
for a given value of @ (cf. Tables 2.11– 2.16);
decreases quickly for both schemes
the modified MOP scheme is performing significantly better than the simple MOM
scheme for all choices of * (cf. Tables 2.11– 2.16);
performance of the two proposed schemes is generally better for smaller choices
of * (cf. Tables 2.11– 2.16);
the modified MOP scheme has the ability to perform well even for moderately large
values of *;
the application of both the schemes is quite simple and easily executable;
the performance of the EWMA type charts can further be enhanced by extending
the proposed schemes with the help of other runs rules schemes;
ix.
the c
decreases for both schemes as the value of @ increases (cf. Abbas, Riaz
and Does (2011));
x.
the run length distribution of both schemes is positively skewed (cf. Abbas, Riaz
and Does (2011)).
DĞŵŽƌLJͲƚLJƉĞŽŶƚƌŽůŚĂƌƚƐŝŶ^ƚĂƚŝƐƚŝĐĂůWƌŽĐĞƐƐŽŶƚƌŽů
Ϯϵ
ϯϬ
ʹǤ͵Ǥʹ
In this section we provide a detailed comparison of the proposed schemes with their
already existing counterparts meant for detecting small shifts, i.e. EWMA- and CUSUM-type
charts. The performance of all the control charting schemes is compared in terms of
. The
control schemes used for the comparison purposes include the classical EWMA, the classical
CUSUM, the FIR CUSUM, the FIR EWMA, the weighted CUSUM, the double CUSUM, the
distribution-free CUSUM and the runs rules schemes based CUSUM.
Proposed versus the classical EWMA: The classical EWMA is defined by Roberts (1959).
values for the classical EWMA are given in Table 2.2. The classical EWMA refers to
one out of one (+O+) scheme. The comparison of the three schemes (i.e. the classical EWMA
and the two proposed schemes) shows that both proposed EWMA schemes of Section 2.3 are
performing better than the classical scheme in terms of
(cf. Tables 2.15 & 2.16 vs. Table
2.2). Moreover the modified MOP scheme is outperforming the simple MOM scheme with a
great margin for the small shifts (i.e. >MB - @ - +>B). The performance of the two proposed
schemes almost coincide for larger values of @.
Proposed versus the classical CUSUM: The classical CUSUM is defined by Page (1954).
The
values of the classical CUSUM are given in Table 2.1 at
168 and 465. The
comparison of the classical CUSUM with the proposed schemes reveals that both schemes
are outperforming the classical CUSUM scheme at all the values of @ (cf. Tables 2.11 & 2.12
vs. Table 2.1). Particularly, comparing the three schemes at @ >MB, we observe that the
modified MOP scheme is performing the best with
scheme with
fixed at +Xf.
PA>A followed by the simple MOM
BA>X, whereas the classical CUSUM has
the modified MOP scheme is giving almost half
dA>M which mean that
than the classical CUSUM scheme with
DĞŵŽƌLJͲƚLJƉĞŽŶƚƌŽůŚĂƌƚƐŝŶ^ƚĂƚŝƐƚŝĐĂůWƌŽĐĞƐƐŽŶƚƌŽů
ZƵŶƐƌƵůĞƐďĂƐĞĚh^hDĂŶĚtD
Proposed versus the FIR CUSUM: The FIR CUSUM presented by Lucan and Crosier
(1982) gives a head start to the CUSUM statistic rather than setting it equal to zero. The
8 of FIR CUSUM with two different values for the head start ( ) are given in Table
2.10. Comparing the performance of FIR CUSUM with the proposed schemes we can see that
the modified MOP scheme is performing better than the FIR CUSUM even though that there is
a problem with the FIR CUSUM because its
value is less than 168 (the desired level).
Moreover, we see that if the value of Co increases than the value of
decreases, which is
not recommended in case of sensitive processes (cf. Bonetti et al. (2000)). The proposed
schemes are not only fixing the
at the pre-specified level (so that valid comparison can
be made) but also performing better in terms of
minimizing the
(with a fixed
)
8,
i.e. the proposed schemes are
without a decrease in
and without the need
of any head start value (cf. Tables 2.11 & 2.12 vs. Table 2.10).
Proposed versus the FIR EWMA: Lucas and Saccucci (1990) proposed the application of
the FIR feature with the EWMA control chart (especially with small values of *). The
values of the EWMA control chart with FIR feature are provided in Table 2.17.
@
TABLE 2.17: 9:; values for the FIR EWMA scheme
% Head
Start
* >+
M>f+A
* >MB
M>eef
* >B
P>d+
* >dB
P>fd
0
25
50
487
468
491
483
497
487
498
496
0.5
25
50
28.3
24.2
46.5
43.6
87.8
86.1
140
139
1
25
50
8.75
6.87
10.1
8.79
16.9
15.9
30.2
29.7
2
25
50
3.57
2.72
3.11
2.5
3.29
2.87
4.33
4.09
Comparing the FIR EWMA with the proposed schemes we observe that the proposed
schemes are not only having smaller
8
but they also fix the
value at desired level
DĞŵŽƌLJͲƚLJƉĞŽŶƚƌŽůŚĂƌƚƐŝŶ^ƚĂƚŝƐƚŝĐĂůWƌŽĐĞƐƐŽŶƚƌŽů
ϯϭ
ϯϮ
which is not the case with the FIR EWMA (cf. Tables 2.15 & 2.16 vs. Table 2.17). The other
comments made in favor of the proposed schemes versus the FIR CUSUM are also valid here
with the same spirit and strength.
Proposed versus the weighted CUSUM: Yashchin (1989) proposed a class of weighted
CUSUM charts which generalize the classical CUSUM charts by giving weights to the past
information and can be viewed as the EWMA version of the CUSUM charts. The
8 for
the weighted CUSUM are given in Table 2.9 where the weights given to the past information
are represented by j. Comparing the weighted CUSUM with the proposed schemes we notice
that the proposed schemes are performing better than the weighted CUSUM for all the values
of @, which shows the uniform superiority of the proposed schemes over the weighted
CUSUM (cf. Tables 2.15 & 2.16 vs. Table 2.9).
Proposed versus the double CUSUM: Waldmann (1995) has shown that the simultaneous
use of two classical CUSUMs improves the
performance of the CUSUM chart. This
simultaneous use of the two CUSUM charts is being given the name of double CUSUM. The
performance of the double CUSUM is given in Table 2.18 in which parameters of the 1st
CUSUM are
and $ and parameters of the 2nd CUSUM are
q
and $q .
TABLE 2.18: 9:;m for the double CUSUM with r I> I, sq g> G and rq I> Iat
9:;= ?==
C
$ M>X
0
0.5
1
1.5
2
507
27.1
9.85
5.55
3.57
Comparison of the double CUSUM with the proposed schemes shows that the double
CUSUM performs better than the simple MOM scheme for @ >B but the modified MOP
scheme performs better than both the simple MOM scheme and the double CUSUM. For all
other values of @, the modified MOP scheme is performing the best followed by the simple
MOM scheme (cf. Tables 2.15 & 2.16 vs. Table 2.18).
DĞŵŽƌLJͲƚLJƉĞŽŶƚƌŽůŚĂƌƚƐŝŶ^ƚĂƚŝƐƚŝĐĂůWƌŽĐĞƐƐŽŶƚƌŽů
ZƵŶƐƌƵůĞƐďĂƐĞĚh^hDĂŶĚtD
Proposed versus the distribution-free CUSUM: Chatterjee and Qiu (2009) proposed a
class of distribution-free CUSUM charts. The three non-parametric control charts named as
B1, B2 and B3 depend upon the variable which is defined as:
hhhuvh
t
{
whhhuvh b b x x y b z w +M x x Z
where is the number of samples since the last time the statistic was zero. The
performance of these non-parametric charts is given in Table 2.19.
TABLE 2.19: 9:; values for different distribution-free CUSUM schemes with nominal
9:;= F==
0
@
0.5
1
B1
B2
B3
178.43
173.78
201.86
25.31
18.37
27.38
12.54
7.94
9.14
30
B1
B2
B3
202.92
194.44
197.79
18.89
18.68
19.20
6.60
6.43
6.45
40
B1
B2
B3
195.04
198.98
201.87
22.40
20.52
21.36
5.66
5.70
5.77
50
B1
B2
B3
190.88
199.35
202.79
16.96
18.73
17.51
6.59
6.84
6.50
w|}~
Chart
5
For @ >B the best ARL performance is at w|}~ B by chart B1. In this case the
+d> whereas the
for the simple MOM and modified MOP schemes is M+> and +B>A,
respectively, which shows superiority of the modified MOP scheme. For @ + the
distribution-free charts perform slightly better for w|}~ A but for all other values of w|}~ ,
the modified MOP scheme is again performing better. This proves the dominance of the
modified MOP scheme as compared to the distribution-free CUSUM charts in general (cf.
Tables 2.13 & 2.14 vs. Table 2.19).
DĞŵŽƌLJͲƚLJƉĞŽŶƚƌŽůŚĂƌƚƐŝŶ^ƚĂƚŝƐƚŝĐĂůWƌŽĐĞƐƐŽŶƚƌŽů
ϯϯ
ϯϰ
Proposed versus the runs rules based CUSUM: In Section 2.2 we proposed two runs rules
schemes, namely CUSUM scheme I and CUSUM scheme II, on the CUSUM charts and
computed the
values for the two schemes which are given in Tables 2.5 and 2.6 for
M. Comparing the proposed EWMA schemes with these runs rules based CUSUM
schemes I and II we see that the two proposed EWMA schemes are performing better than
the CUSUM schemes of Section 2.1 (cf. Tables 2.13 & 2.14 vs. Tables 2.5 & 2.6).
Moreover, for an overall comparison of the proposed schemes with their existing
counterparts mentioned and compared above we have made some graphs showing
curves of different schemes.
It is evident from the Figures 2.3 – 2.5 that the
curves of the two proposed
EWMA schemes exhibit dominance in general as compared to all the other schemes covered
in this chapter.
Figure 2.3: 9:; curves for the simple FOF and modified FOI EWMA schemes, the classical
CUSUM and the FIR CUSUM at 9:;= EgG
tDƐŝŵƉůĞϮͬϮ;ʄсϬ͘ϭͿ
ĐůĂƐƐŝĐĂůh^hD;ŬсϬ͘ϱͿ
ϴϬ
tDŵŽĚŝĨŝĞĚϮͬϯ;ʄсϬ͘ϭͿ
&/Zh^hD;ŬсϬ͘ϱ͕ŽсϭͿ
ϳϬ
ϲϬ
Z>Ɛ
ϱϬ
ϰϬ
ϯϬ
ϮϬ
ϭϬ
Ϭ
Ϭ͘Ϯϱ
Ϭ͘ϱ
Ϭ͘ϳϱ
į
DĞŵŽƌLJͲƚLJƉĞŽŶƚƌŽůŚĂƌƚƐŝŶ^ƚĂƚŝƐƚŝĐĂůWƌŽĐĞƐƐŽŶƚƌŽů
ϭ
ϭ͘ϱ
ZƵŶƐƌƵůĞƐďĂƐĞĚh^hDĂŶĚtD
Figure 2.4: 9:; curves for the simple FOF and modified FOI EWMA schemes, the runs rules
scheme I for CUSUM and the the runs rules scheme II for CUSUM at 9:;= F==
tDƐŝŵƉůĞϮͬϮ;ʄсϬ͘ϭͿ
h^hD/;ŬсϬ͘ϱͿ
tDŵŽĚŝĨŝĞĚϮͬϯ;ʄсϬ͘ϭͿ
h^hD//;ŬсϬ͘ϱͿ
ϵϬ
ϴϬ
ϳϬ
Z>Ɛ
ϲϬ
ϱϬ
ϰϬ
ϯϬ
ϮϬ
ϭϬ
Ϭ
Ϭ͘Ϯϱ
Ϭ͘ϱ
Ϭ͘ϳϱ
ϭ
ϭ͘ϱ
į
Figure 2.5: 9:; curves for the simple FOF and modified FOI EWMA schemes, the classical
EWMA, the FIR EWMA, the weighted CUSUM and the double CUSUM at 9:;= ?==
tDƐŝŵƉůĞϮͬϮ;ʄсϬ͘ϭͿ
tD;ʄсϬ͘ϭͿ
tDŵŽĚŝĨŝĞĚϮͬϯ;ʄсϬ͘ϭͿ
&/ZtD;ʄсϬ͘ϭͿ
ϲϬ
ϱϬ
ϰϬ
Z>Ɛ
ϯϬ
ϮϬ
ϭϬ
Ϭ
Ϭ͘ϱ
ϭ
Ϯ
į
DĞŵŽƌLJͲƚLJƉĞŽŶƚƌŽůŚĂƌƚƐŝŶ^ƚĂƚŝƐƚŝĐĂůWƌŽĐĞƐƐŽŶƚƌŽů
ϯϱ
ϯϲ
Particularly, the
curve of the modified MOP EWMA scheme is on the lower side
compared to all other schemes. This shows the best
performance of the modified MOP
EWMA scheme compared to all others. For the small shifts, the gap between the
curves
of the proposed schemes with those of the other schemes is large, whereas this gap reduces as
the size of the shift increases. This implies that the proposals of the study (particularly
modified MOP EWMA scheme) are generally more beneficial for small shifts.
ʹǤ͵Ǥ͵
This section presents an illustrative example to show how the proposed schemes can
be applied in real situations. For this purpose we have used dataset 2.1 and dataset 2.2 from
subsection 2.2.3. The EWMA statistics are calculated with * >+ and the three schemes
(i.e. the two proposed schemes and the classical scheme with
B) are applied to the
datasets. The graphical display of the control chart with all the three schemes applied to the
datasets 2.1 and 2.2 are given in Figures 2.6 and 2.7 respectively.
From Figure 2.6 we can see that the first 20 points are plotted around the central line
whereas an upward shift in the points can be seen afterwards. The classical scheme is not
signaling any shift whereas the simple MOM scheme is signaling at points # 49 and 50. The
modified MOP scheme is giving 3 out of control signals and these are at points # 45, 49 and
50. This clearly indicated that the modified MOP scheme is not only signaling earlier than the
classical scheme but also is giving more number of signals. The situation is not much
different in Figure 2.7 where the classical and the simple MOM schemes failed to detect any
shift while the modified MOP scheme gives out of control signals at points # 25, 26 and 27.
DĞŵŽƌLJͲƚLJƉĞŽŶƚƌŽůŚĂƌƚƐŝŶ^ƚĂƚŝƐƚŝĐĂůWƌŽĐĞƐƐŽŶƚƌŽů
ZƵŶƐƌƵůĞƐďĂƐĞĚh^hDĂŶĚtD
Figure 2.6: EWMA chart of the classical scheme and the the simple FOF and modified FOI
schemes for the dataset 2.1
ŝ
>>;ĐůĂƐƐŝĐĂůͿ
>>;ƐŝŵƉůĞϮͬϮͿ
>>;ŵŽĚŝĨŝĞĚϮͬϯͿ
Ϭ͘ϴ
Ϭ͘ϲ
Ϭ͘ϰ
Ϭ͘Ϯ
Ϭ
ϭ
ϯ
ϱ
ϳ
ϵ ϭϭ ϭϯ ϭϱ ϭϳ ϭϵ Ϯϭ Ϯϯ Ϯϱ Ϯϳ Ϯϵ ϯϭ ϯϯ ϯϱ ϯϳ ϯϵ ϰϭ ϰϯ ϰϱ ϰϳ ϰϵ
ͲϬ͘Ϯ
ͲϬ͘ϰ
ͲϬ͘ϲ
ͲϬ͘ϴ
Figure 2.7: EWMA chart of the classical scheme and the the simple FOF and modified FOI
schemes for the dataset 2.2
ŝ
>>;ĐůĂƐƐŝĐĂůͿ
>>;ƐŝŵƉůĞϮͬϮͿ
>>;ŵŽĚŝĨŝĞĚϮͬϯͿ
Ϭ͘ϴ
Ϭ͘ϲ
Ϭ͘ϰ
Ϭ͘Ϯ
Ϭ
ϭ Ϯ ϯ ϰ ϱ ϲ ϳ ϴ ϵ ϭϬ ϭϭ ϭϮ ϭϯ ϭϰ ϭϱ ϭϲ ϭϳ ϭϴ ϭϵ ϮϬ Ϯϭ ϮϮ Ϯϯ Ϯϰ Ϯϱ Ϯϲ Ϯϳ Ϯϴ Ϯϵ ϯϬ
ͲϬ͘Ϯ
ͲϬ͘ϰ
ͲϬ͘ϲ
ͲϬ͘ϴ
DĞŵŽƌLJͲƚLJƉĞŽŶƚƌŽůŚĂƌƚƐŝŶ^ƚĂƚŝƐƚŝĐĂůWƌŽĐĞƐƐŽŶƚƌŽů
ϯϳ
ϯϴ
The above example indicates that the modified MOP EWMA scheme is giving the
advantage in terms of run length as well the number of signals for both small and moderate
shifts. The outcomes of these two illustrative examples are completely in accordance with the
findings of subsection 3.3.1.
ʹǤͶ
For small shifts CUSUM charts and EWMA charts are considered most effective. The
efficiency of these charts can also be increased by using different sensitizing rules and runs
rules schemes with their usual design structure. We have proposed two runs rules schemes
each for CUSUM and EWMA charts for the location parameter. By investigating the
performance of the these proposed schemes and by comparing them with some existing
schemes we found that the proposed schemes have the ability to perform better for small and
moderate shifts while reasonably maintaining their efficiency for large shifts as well.
To make the CUSUM and EWMA charts even more efficient, some other sensitizing
rules/runs rules schemes can be used with their respective structures on the similar lines as
followed in this chapter. The proposals and the recommendations of this chapter can also be
extended for the attribute control charts based on CUSUM and EWMA patterns.
DĞŵŽƌLJͲƚLJƉĞŽŶƚƌŽůŚĂƌƚƐŝŶ^ƚĂƚŝƐƚŝĐĂůWƌŽĐĞƐƐŽŶƚƌŽů
ŚĂƉƚĞƌϯ
͵
Ǧ
Shewhart-type control charts are sensitive for large disturbances in the process, while
CUSUM- and EWMA-type control charts are intended to spot small and moderate
disturbances. In this chapter we propose a mixed EWMA-CUSUM control chart for detecting
a shift in the process mean and evaluating its
8. Comparisons of the proposed control
chart are made with some representative control charts including the classical CUSUM,
classical EWMA, Fast Initial Response (FIR) CUSUM, FIR EWMA, adaptive CUSUM with
EWMA based shift estimator, weighted CUSUM and runs rules based CUSUM and EWMA.
The comparisons reveal that the mixing of the two charts makes the proposed scheme even
more sensitive to small shifts in the process mean than the other schemes designed for
detecting small shifts.
Following the mixed EWMA-CUSUM chart for location, we also propose a new
control chart for monitoring the process dispersion. This chart is named the CS-EWMA chart
as its plotting statistic is based on a cumulative sum of the exponentially weighted moving
averages. Comparisons with other memory charts used to monitor the process dispersion are
done by means of the
. An illustration of the proposed technique is done by applying the
CS-EWMA chart on a simulated dataset.
This chapter is based on two papers; one for monitoring the location parameter (cf. Abbas,
Riaz and Does (2012a)) and the other for monitoring the dispersion parameter (cf. Abbas,
Riaz and Does (2012b)).
DĞŵŽƌLJͲƚLJƉĞŽŶƚƌŽůŚĂƌƚƐŝŶ^ƚĂƚŝƐƚŝĐĂůWƌŽĐĞƐƐŽŶƚƌŽů
ϯϵ
ϰϬ
͵Ǥͳ Ǧ
After the development of CUSUM and EWMA charts, several modifications of these
charts have been presented in order to further enhance the performance of these charts. Lucas
(1982) presented the combined Shewhart-CUSUM quality control scheme in which Shewhart
limits and CUSUM limits are used simultaneously. Lucas and Crosier (1982) recommended
the use of the FIR CUSUM which gives a head start to the CUSUM statistic by setting the
initial values of the CUSUM statistic equal to some positive value (non-zero). This feature
gives better
performance but at the cost of a decrease in
.
Yashchin (1989)
presented the weighted CUSUM scheme which gives different weights to the previous
information used in CUSUM statistic. Section 2.2 introduced the runs rules schemes to the
CUSUM charts and shown that the runs rules based CUSUM performs better than the
classical CUSUM for small shifts. Similarly, on the EWMA side, Lucas and Saccucci (1990)
presented the combined Shewhart-EWMA quality control scheme which gives better
performance for both small and large shifts. Steiner (1999) provided the FIR EWMA which
gives a head start to the initial value of the EWMA statistic (like FIR CUSUM) and hence
improves the
performance of the EWMA charts. Section 2.3 discussed the runs rules
schemes to the EWMA charts and showed that the runs rules based EWMA performs better
than the classical EWMA for small shifts. In the next subsection we present a mixed EWMACUSUM quality control scheme for monitoring the mean of a normally distributed process.
The inspiration is to get an improved
performance by combining the features of EWMA
and CUSUM charts in a single control structure.
DĞŵŽƌLJͲƚLJƉĞŽŶƚƌŽůŚĂƌƚƐŝŶ^ƚĂƚŝƐƚŝĐĂůWƌŽĐĞƐƐŽŶƚƌŽů
DŝdžĞĚtDͲh^hDĐŚĂƌƚƐ
͵ǤͳǤͳ
In this subsection we propose an assortment of the classical EWMA and CUSUM
schemes by combining the features of their design structures. The said proposal mainly
depends on two statistics named as 7 and 7 which are defined as:
7
.
7
where
q
q
! 7
q
! 7
(3.1)
is a time varying reference value for the proposed charting structure, the quantities
7 and 7 are known as the upper and lower CUSUM statistics which are initially set to
zero (i.e. 7 7 ) and are based on the EWMA statistic which is defined as:
* ! 5+ * 6
(3.2)
In (3.2), * is the constant like * in (1.3) such that , * - + and the initial value of the
statistic is set equal to the target mean i.e. . Now the mean and variance of statistic
is given as:
'
0
0
\+ 5+ * 6 a
(3.3)
and this will be used later in the calculation of the parameters of the proposed chart.
In (3.1) and (3.2) we are considering the case of individual observations Z +
which may be extended easily for the subgroups. Now the statistics 7 and 7 are plotted
against the control limit, say $q . As long as the values of 7 and 7 are plotted inside the
control limit, the process is said to be in control, otherwise out of control. It is to be noted
here that if the statistic 7 is plotted above $q the process mean is said to be shifted above
the target value and if the statistic 7 is plotted above $q the process is said to be shifted
below the target value. The control limit $q is selected according to a prefixed
.
A large
DĞŵŽƌLJͲƚLJƉĞŽŶƚƌŽůŚĂƌƚƐŝŶ^ƚĂƚŝƐƚŝĐĂůWƌŽĐĞƐƐŽŶƚƌŽů
ϰϭ
ϰϮ
value of the prefixed
q
.
q
and $q are defined as:
,
will give a larger value of $q and vice versa. The two quantities
& K & ' /
0
0
\+ 5+ * 6 a 4
\+ 5+ * 6 a3
0
1
$q ( K ( ' /
0
(3.4)
where & and ( are the constants like & and (, respectively, in the classical set up for the
CUSUM (cf. Section 1.2). The time varying values
and $ are due to the variance of the
EWMA statistic in expression (3.3). For a fixed value of & , we can select the value of (
from the tables (that are given later in this subsection) that fix the
In general
q
at our desired level.
is chosen equal to half of the shift (in units of the standard deviation of ).
Hence, we choose & >B as it makes the CUSUM structure more sensitive to the small
and moderate shifts (cf. Montgomery (2009)), to which memory charts actually target.
To evaluate the
performance of a control scheme, we have used the Monte Carlo
simulation approach in this chapter. An algorithm in R language (provided in Appendix 3.1)
is developed to calculate the run lengths. The algorithm is run 50,000 times to calculate the
average of those 50,000 run lengths. A detailed study on the
performance of the
proposed EWMA-CUSUM control chart to monitor the mean of a normally distributed
process is provided in Tables 3.1 – 3.3 for some selective choices of @, * and ( . For this
purpose
8
other values of
are fixed at 168, 400 and 500 which are the commonly used choices. For
8
one may easily obtain the results on similar lines.
The relative standard errors for the results provided in Tables 3.1 – 3.3 are also calculated
and found to be less than 1.2%. Moreover, we have also replicated the
results of the
classical CUSUM and the classical EWMA using our simulation algorithm and found almost
DĞŵŽƌLJͲƚLJƉĞŽŶƚƌŽůŚĂƌƚƐŝŶ^ƚĂƚŝƐƚŝĐĂůWƌŽĐĞƐƐŽŶƚƌŽů
DŝdžĞĚtDͲh^hDĐŚĂƌƚƐ
similar results as by Hawkins and Olwell (1998) and Lucas and Saccucci (1990),
respectively, ensuring the validity of the simulation algorithm used.
TABLE 3.1: 9:; values for the proposed EWMA-CUSUM scheme with < => ? at
9:;= EgG
C
0
D => E
l FE> I
168.0441
D => F?
l EI> F
168.0652
D => ?
l G> EF
169.8763
D => J?
l ?> HG
171.0422
0.25
52.6449
54.1752
59.7829
68.15245
0.5
24.85945
22.40665
22.54895
24.12865
0.75
17.0208
14.0235
12.85555
12.60565
1
13.3323
10.4832
8.9565
8.2741
1.5
9.743
7.3272
5.78565
4.99665
2
7.90705
5.8231
4.4341
3.7365
C
0
D => E
l II> ?H
402.0894
D => F?
l EG> J
397.404
D => ?
l E=> ?F
398.6486
D => J?
l g> H
0.25
73.31955
78.02035
90.45915
108.0086
0.5
33.06085
29.0845
28.94885
31.44015
0.75
22.39445
17.79425
15.75695
15.66785
1
17.63975
13.2232
10.94695
10.17115
1.5
12.88105
9.113
6.9587
6.03875
2
10.45315
7.2235
5.2808
4.4203
C
0
D => E
l IJ> HF
498.3882
D => F?
l F=> EG
502.018
D => ?
l EE> F
507.9555
D => J?
l J> IF
0.25
80.13585
83.7529
100.2635
121.9883
TABLE 3.2: 9:; values for the proposed EWMA-CUSUM scheme with < => ? at
9:;= H==
400.8962
TABLE 3.3: 9:; values for the proposed EWMA-CUSUM scheme with < => ? at
9:;= ?==
507.5152
0.5
35.524
30.88825
30.7466
33.5054
0.75
24.0522
18.8755
16.6399
16.5139
1
18.8637
13.8816
11.45835
10.6107
1.5
13.79075
9.6036
7.29565
6.3101
2
11.19775
7.59055
5.52345
4.589
DĞŵŽƌLJͲƚLJƉĞŽŶƚƌŽůŚĂƌƚƐŝŶ^ƚĂƚŝƐƚŝĐĂůWƌŽĐĞƐƐŽŶƚƌŽů
ϰϯ
ϰϰ
The main findings about our proposed EWMA-CUSUM quality control scheme for
monitoring the mean of a normally distributed process are given as:
i.
mixing of the EWMA and CUSUM schemes really boosts the
performance of
the resulting combination of the two charts especially for small and moderate shifts in
the process (cf. Tables 3.1 – 3.3);
ii.
for detecting small shifts in the process, the performance of the proposed scheme is
iii.
the proposed scheme is
iv.
better with smaller values of * and vice versa (cf. Tables 3.1 – 3.3);
,
the
decreases with a decrease in the value of @ and vice versa (cf. Tables 3.1 – 3.3);
for a fixed value of @, the
v.
unbiased, i.e. for a fixed value of
of the proposed scheme decreases with a decrease in
(cf. Tables 3.1 – 3.3);
for a fixed value of
,
in * (cf. Tables 3.1 – 3.3).
the control limit coefficient ( decreases with the increase
͵ǤͳǤʹ
In this subsection we present a comprehensive comparison of the proposed mixed
EWMA-CUSUM scheme with some existing representative EWMA and CUSUM control
charts available in the literature. The performance of the control chart is compared in terms of
. The set of the schemes considered for the comparison consist of the classical CUSUM,
the classical EWMA, the FIR CUSUM, the FIR EWMA, the adaptive CUSUM with EWMA
based shift estimator, the weighted CUSUM and the runs rules based CUSUM and EWMA.
Proposed versus the classical CUSUM: The
values for the classical CUSUM control
scheme proposed by Page (1954) are given in Table 2.1. Comparison of the classical
DĞŵŽƌLJͲƚLJƉĞŽŶƚƌŽůŚĂƌƚƐŝŶ^ƚĂƚŝƐƚŝĐĂůWƌŽĐĞƐƐŽŶƚƌŽů
DŝdžĞĚtDͲh^hDĐŚĂƌƚƐ
CUSUM with the proposed schemes reveal that the proposed scheme is performing really
good for all the values of * , particularly for small values of * . We can see that, for all
values of * , the proposed scheme has better
performance as compared to the classical
Proposed versus the classical EWMA: The
values for the classical EWMA with time
CUSUM (cf. Table 2.1 vs. Table 3.1).
varying limits, given by Steiner (1999), are provided in Table 2.2. Comparing the classical
EWMA * >MBh with the proposed scheme we observe that the proposed scheme has
better
8
performance with its respective values of * (cf. Table 2.2 vs. Table 3.3).
Proposed versus the FIR CUSUM: The FIR CUSUM presented by Lucas and Crosier
(1982) provides a head start to the CUSUM statistic. The
8 of the CUSUM with FIR
feature are given in Table 2.10 in which head start is represented by . The FIR feature
decreases the
of the CUSUM chart and more importantly this decreased
very small for the larger values of (for +,
becomes
+XP) which is not recommended
in case of sensitive processes like in health care (cf. Bonetti et al. (2000)). Comparing the
proposed scheme with the FIR CUSUM we see that for smaller values of * the proposed
scheme has a better
not have the fixed
3.1).
performance than the FIR CUSUM, even if the FIR CUSUM does
at +Xf but has smaller
value, i.e. +XP (cf. Table 2.10 vs. Table
Proposed versus the FIR EWMA: FIR EWMA presented by Steiner (1999) is similar to the
FIR CUSUM as it also gives a head start to the EWMA statistic. The control limits for the
FIR based EWMA chart are given as:
' 5+ + }p 6
*
+ + *
M*
DĞŵŽƌLJͲƚLJƉĞŽŶƚƌŽůŚĂƌƚƐŝŶ^ƚĂƚŝƐƚŝĐĂůWƌŽĐĞƐƐŽŶƚƌŽů
ϰϱ
ϰϲ
where MN+ +O+e. The
8 for the FIR EWMA with * >+ and the
proposed chart with * >+ are given in Table 3.4. Comparing the proposed scheme with
the FIR EWMA we see that the proposed scheme is performing better than the FIR EWMA
for smaller shifts i.e. @ , >B. For moderate and larger shifts, FIR EWMA seems superior as
compared to the proposed chart.
TABLE 3.4: 9:; values for the FIR EWMA scheme and the proposed chart
C
0
* >+, P
>A
515.6
* >+, P
>B
613.8
* >+, >B
Pd>eA
516.48
* >+, >B
A>f
0.25
83.1
99.2
81.03
85.79
FIR EWMA
EWMA-CUSUM
613.62
0.5
18.5
22.1
35.76
37.55
0.75
7.3
8.8
24.2
25.41
1
3.8
4.6
19.01
19.94
1.5
1.7
2.1
13.9
14.55
2
1.3
1.4
11.29
11.8
3
1
1
8.48
8.88
4
1
1
6.96
7.29
Proposed versus the adaptive CUSUM with EWMA based shift estimator: Jiang et al.
(2008) proposed the use of adaptive CUSUM with EWMA-based shift estimator. They used
the concept of adaptively updating the reference value of the CUSUM chart using the EWMA
estimator and then using a suitable weighting function. The
values for the adaptive
CUSUM are given in Table 3.5 in which @|`
, *, j and ( are the parameters of the chart.
Comparing the performance of the proposed scheme we notice that the proposed scheme is
outperforming the adaptive CUSUM for small values of @. For moderate and large values of
@, both the proposed scheme and adaptive CUSUM have almost the same
(cf. Table 3.5 vs. Table 3.2).
DĞŵŽƌLJͲƚLJƉĞŽŶƚƌŽůŚĂƌƚƐŝŶ^ƚĂƚŝƐƚŝĐĂůWƌŽĐĞƐƐŽŶƚƌŽů
performance
DŝdžĞĚtDͲh^hDĐŚĂƌƚƐ
TABLE 3.5: 9:; values for adaptive CUSUM with C
=> ? and D => I at 9:;= H==
C
0
j +>B
( B>B
0.25
399.7
ji
jM
jP
jA
j M>B
( A>dP ( A>BB ( A>PeA ( A>PPd ( A>PPA
400.85
400.19
399.29
399.39
399.97
92.82
91.65
88.96
87.02
85.81
85.8
0.5
30.52
30.1
29.32
28.79
28.46
28.45
0.75
14.7
14.5
14.2
14
13.89
13.88
1
9.07
8.96
8.81
8.72
8.67
8.66
1.5
4.89
4.87
4.84
4.83
4.83
4.82
2
3.23
3.25
3.28
3.31
3.35
3.34
There is also an adaptive EWMA chart (cf. Capizzi and Masarotto (2003)), but its
performance is inferior to the adaptive CUSUM, so the results of our proposal are superior to
the adaptive EWMA as well.
Proposed versus the weighted CUSUM: Weighted CUSUM presented by Yashchin (1989)
gives weights to the past information in the CUSUM statistic. The
8 for the weighted
CUSUM are given in Table 2.9 in which the weights given to the past information are
represented by j. The comparison of the proposed scheme with the weighted CUSUM shows
that the proposed scheme is performing better than the weighted CUSUM for the small and
moderate shifts (like @ , +>B). For larger values of @, the weighted CUSUM almost coincide
with the proposed scheme (cf. Table 2.9 vs. Table 3.3).
Proposed versus the runs rules based CUSUM: Section 2.2 introduced the use of the runs
rules schemes with the design structure of the CUSUM charts. The
8 for the two runs
rules based CUSUMs are given in Tables 2.3 and 2.4 in which WL and AL are representing
the warning limits and action limits, respectively. The comparison of the proposed scheme
with both the runs rules based CUSUM schemes shows that the proposed scheme has the
ability to perform better than the runs rules based CUSUM for all the choices of * (cf.
Tables 2.3 and 2.4 vs. Table 3.1).
DĞŵŽƌLJͲƚLJƉĞŽŶƚƌŽůŚĂƌƚƐŝŶ^ƚĂƚŝƐƚŝĐĂůWƌŽĐĞƐƐŽŶƚƌŽů
ϰϳ
ϰϴ
Proposed versus the runs rules based EWMA: Section 2.3 introduced the use of the runs
rules schemes EWMA structure. The
8 for the two runs rules based EWMAs are given in
Tables 2.15 and 2.16. The comparison of the proposed schemes with both the runs rules
based EWMA schemes shows that the proposed scheme is performing better as long as
* % >+ for the runs rules based EWMA schemes. For * >+, modified MOP EWMA
scheme becomes a bit superior to the proposed chart (cf. Tables 2.15 and 2.16 vs. Table 3.3).
Overall View: In order to provide an overall comparative view of the proposed scheme with
the other existing counterparts we have made some graphical displays, in the form of
curves. Three selective graphs of different charts/schemes are given in Figures 3.1 – 3.3. In
these figures RR CUSUM (EWMA) stands for the runs rules based CUSUM (EWMA)
schemes and the other terms/symbols used are self-explanatory.
Figure 3.1: 9:; curves for the proposed scheme, the Classical CUSUM, the FIR CUSUM and
Runs Rules based CUSUMs at 9:;= EgG
WƌŽƉŽƐĞĚ;ʄсϬ͘ϱͿ
&/Zh^hD;ŽсϭͿ
ZZh^hD//;t>сϯ͘ϱ͕>сϰ͘ϰϰͿ
ϴϬ
ůĂƐƐŝĐĂůh^hD;ŬсϬ͘ϱ͕ŚсϰͿ
ZZh^hD/;t>сϯ͘ϱϯ͕>сϰ͘ϮͿ
ϳϬ
ϲϬ
Z>Ɛ
ϱϬ
ϰϬ
ϯϬ
ϮϬ
ϭϬ
Ϭ
Ϭ͘Ϯϱ
Ϭ͘ϱ
Ϭ͘ϳϱ
į
ϭ
DĞŵŽƌLJͲƚLJƉĞŽŶƚƌŽůŚĂƌƚƐŝŶ^ƚĂƚŝƐƚŝĐĂůWƌŽĐĞƐƐŽŶƚƌŽů
ϭ͘ϱ
Ϯ
DŝdžĞĚtDͲh^hDĐŚĂƌƚƐ
Figure 3.2: 9:; curves for the proposed scheme and adaptive CUSUM at 9:;= H==
WƌŽƉŽƐĞĚ;ʄсϬ͘ϭͿ
WƌŽƉŽƐĞĚ;ʄсϬ͘ϱͿ
ĚĂƉƚŝǀĞh^hD;ɶсϮ͘ϱͿ
ϭϬϬ
ϵϬ
ϴϬ
Z>Ɛ
ϳϬ
ϲϬ
ϱϬ
ϰϬ
ϯϬ
ϮϬ
ϭϬ
Ϭ
Ϭ͘Ϯϱ
Ϭ͘ϱ
Ϭ͘ϳϱ
į
ϭ
ϭ͘ϱ
Ϯ
Figure 3.3: 9:; curves for the proposed scheme, the classical EWMA, the FIR EWMA, the
runs rules based EWMA and the weighted CUSUM at 9:;= ?==
WƌŽƉŽƐĞĚ;ʄсϬ͘ϱͿ
ZZtD/;ʄсϬ͘ϱͿ
ůĂƐƐŝĐĂůtD;ʄсϬ͘ϱͿ
ZZtD//;ʄсϬ͘ϱͿ
tĞŝŐŚƚĞĚh^hD;ɶсϬ͘ϵͿ
ϭϬϬ
ϵϬ
ϴϬ
Z>Ɛ
ϳϬ
ϲϬ
ϱϬ
ϰϬ
ϯϬ
ϮϬ
ϭϬ
Ϭ
Ϭ͘ϱ
ϭ
į
ϭ͘ϱ
Ϯ
DĞŵŽƌLJͲƚLJƉĞŽŶƚƌŽůŚĂƌƚƐŝŶ^ƚĂƚŝƐƚŝĐĂůWƌŽĐĞƐƐŽŶƚƌŽů
ϰϵ
ϱϬ
By examining the graphs of
curves of different schemes under study we see that the
curve of the proposed schemes are on the lower side which shows evidence for the
dominance of the proposed scheme over the other schemes. For small values of @, the
difference between the
of the proposed scheme and the other schemes is larger whereas
for the moderate values of @ this difference almost disappears. For large values of @, the
curve of the proposed chart seems above the
curves of some other charts showing the
poor performance of the proposed chart for large shifts.
To sum up, we may infer that in general the proposed chart is superior for small and
moderate shifts while for larger shifts its performance is inferior to some of the other schemes
under investigation.
͵ǤͳǤ͵
Besides exploring the statistical properties of a method it is always good to provide its
application on some data for illustration purposes. Here we present an illustrative example to
show how the proposed scheme can be applied in the real situation. For this purpose a dataset
is generated containing 40 observations. The first 20 observations are generated from the in
control situation (i.e. [+ so that the target mean is 0) and the remaining 20 observations
are generated from an out of control situation with a small shift introduced in the process (i.e.
[>B+). The classical CUSUM, the classical EWMA and the proposed scheme are applied
to this dataset and the parameters are selected to be & >B and ( B>d for the classical
CUSUM scheme, * >MB and
P for the classical EWMA scheme, * >MB, & >B
and ( M>+f for the proposed scheme to guarantee that
DĞŵŽƌLJͲƚLJƉĞŽŶƚƌŽůŚĂƌƚƐŝŶ^ƚĂƚŝƐƚŝĐĂůWƌŽĐĞƐƐŽŶƚƌŽů
B.
DŝdžĞĚtDͲh^hDĐŚĂƌƚƐ
Figure 3.4: The classical CUSUM chart for the simulated dataset using < => ? and l ?> =
at 9:;= ?==
н
Ͳ
Ś
ϲ
ϱ
ŝ
ϰ
ϯ
Ϯ
ϭ
Ϭ
ϭ
ϯ
ϱ
ϳ
ϵ
ϭϭ ϭϯ ϭϱ ϭϳ ϭϵ Ϯϭ Ϯϯ Ϯϱ Ϯϳ Ϯϵ ϯϭ ϯϯ ϯϱ ϯϳ ϯϵ
^ĂŵƉůĞEƵŵďĞƌ
Figure 3.5: The classical EWMA chart for the simulated dataset using D => F? and
; I at 9:;= ?==
ŝ
ŽŶƚƌŽů>ŝŵŝƚƐ
Ϭ͘ϴ
Ϭ͘ϯ
ŝ
ϱϭ
ͲϬ͘Ϯ
ϭ
ϯ
ϱ
ϳ
ϵ
ϭϭ ϭϯ ϭϱ ϭϳ ϭϵ Ϯϭ Ϯϯ Ϯϱ Ϯϳ Ϯϵ ϯϭ ϯϯ ϯϱ ϯϳ ϯϵ
ͲϬ͘ϳ
Ͳϭ͘Ϯ
^ĂŵƉůĞEƵŵďĞƌ
DĞŵŽƌLJͲƚLJƉĞŽŶƚƌŽůŚĂƌƚƐŝŶ^ƚĂƚŝƐƚŝĐĂůWƌŽĐĞƐƐŽŶƚƌŽů
ϱϮ
Figure 3.6: The proposed scheme for the simulated dataset using D => F?, < => ? and
l F=> EG at 9:;= ?==
Dн
DͲ
ďŝ
ϭϮ
ϭϬ
ŝ
ϴ
ϲ
ϰ
Ϯ
Ϭ
ϭ
ϯ
ϱ
ϳ
ϵ
ϭϭ ϭϯ ϭϱ ϭϳ ϭϵ Ϯϭ Ϯϯ Ϯϱ Ϯϳ Ϯϵ ϯϭ ϯϯ ϯϱ ϯϳ ϯϵ
^ĂŵƉůĞEƵŵďĞƌ
The calculations for the proposed scheme are given in Table 3.6 and the graphical display of
all three control structures are provided in Figures 3.4 – 3.6 with the statistics and
plotted against the control limit $ for the classical CUSUM scheme; ) plotted against the
control limits given in (1.4) for the classical EWMA scheme; and 7 and 7 plotted against
the control limit $ for the proposed scheme.
From Table 3.6 and Figure 3.6 it is obvious that out of control signals are received at
samples # 32, 33, 34, 35, 36, 37, 38, 39 and 40 by the proposed scheme (giving 8 out of
control signals). Figures 3.4 – 3.5 show that the separate applications of the classical
CUSUM and EWMA schemes fail to detect any out of control situation for the given dataset.
This clearly indicates superiority of the proposed scheme over the classical CUSUM and
EWMA schemes and it is exactly in accordance with the findings of subsection 3.1.1.
DĞŵŽƌLJͲƚLJƉĞŽŶƚƌŽůŚĂƌƚƐŝŶ^ƚĂƚŝƐƚŝĐĂůWƌŽĐĞƐƐŽŶƚƌŽů
DŝdžĞĚtDͲh^hDĐŚĂƌƚƐ
TABLE 3.6: Application example of the proposed scheme using D => F?, ¡ => ? and
¢ F=> EG at 9:;= ?==
Sample
No.
q
7
0
7
0
$q
5.045
Sample
No.
q
7
7
$q
1
-0.113
-0.028 0.125
21
0.781
0.452 0.189
3.175
0
7.627
2
-1.906
-0.498 0.156
0
0.341 6.306
22
-0.016
0.335 0.189
3.321
0
7.627
3
-1.891
-0.846 0.171
0
1.016 6.915
23
-0.061
0.236 0.189
3.368
0
7.627
4
0.508
-0.508 0.179
0
1.344 7.235
24
0.332
0.26
0.189
3.439
0
7.627
5
1.374
-0.037 0.184
0
1.198 7.409
25
1.391
0.543 0.189
3.793
0
7.627
6
0.05
-0.015 0.186
0
1.027 7.506
26
1.89
0.879 0.189
4.483
0
7.627
7
0.401
0.089 0.187
0
0.751 7.559
27
0.709
0.837 0.189
5.131
0
7.627
8
0.692
0.239 0.188 0.051 0.323 7.589
28
-0.82
0.423 0.189
5.364
0
7.627
9
0.851
0.392 0.188 0.255
0
7.606
29
1.481
0.687 0.189
5.863
0
7.627
10
0.927
0.526 0.189 0.593
0
7.615
30
0.314
0.594 0.189
6.268
0
7.627
11
2.187
0.941 0.189 1.346
0
7.621
31
2.231
1.003 0.189
7.082
0
7.627
12
0.02
0.711 0.189 1.868
0
7.623
32
0.802
0.953 0.189 7.846*
0
7.627
13
0.12
0.563 0.189 2.242
0
7.625
33
-1.25
0.402 0.189 8.059*
0
7.627
14
2.138
0.957 0.189
3.01
0
7.626
34
0.351
0.389 0.189 8.260*
0
7.627
15
0.183
0.764 0.189 3.585
0
7.627
35
1.362
0.632 0.189 8.703*
0
7.627
16
-2.389
-0.024 0.189 3.371
0
7.627
36
-0.529
0.342 0.189 8.856*
0
7.627
17
-0.269
-0.086 0.189 3.097
0
7.627
37
2.59
0.904 0.189 9.571*
0
7.627
18
0.317
0.015 0.189 2.923
0
7.627
38
0.287
0.75
0.189 10.132*
0
7.627
19
0.055
0.025 0.189 2.759
0
7.627
39
1.676
0.981 0.189 10.924*
0
7.627
20
1.293
0.342 0.189 2.912
0
7.627
40
-0.303
0.66
0
7.627
0.189 11.395*
* indicates proposed scheme giving out of control signal
͵Ǥʹ Ǧ
Page (1963) introduced the CUSUM chart for monitoring the increase in process
dispersion using sample ranges. Following him, Tuprah and Ncube (1987), Chang and Gan
(1995) and Acosta-Mejia et al. (1999) proposed several improved versions of the CUSUM
chart for process dispersion. On the other hand, Wortham and Ringer (1971) suggested an
EWMA control chart for monitoring the process dispersion. Ng and Case (1989) and
Crowder and Hamilton (1992) proposed improved versions of the EWMA chart for
DĞŵŽƌLJͲƚLJƉĞŽŶƚƌŽůŚĂƌƚƐŝŶ^ƚĂƚŝƐƚŝĐĂůWƌŽĐĞƐƐŽŶƚƌŽů
ϱϯ
ϱϰ
monitoring process variance. Castagliola (2005) and Castagliola et al. (2009) proposed
EWMA respectively CUSUM control charts for monitoring the process variance based on a
logarithmic transformation of the sample variance. This article proposes a new memory-type
control chart based on the same transformation, named as CS-EWMA chart, for monitoring
the process dispersion by mixing the effects of EWMA and CUSUM charts.
After presenting the basic structures of the EWMA and CUSUM charts for
monitoring the process dispersion in the next subsection, we present the details of our
proposed CS-EWMA chart for process standard deviation in the subsequent subsection.
͵ǤʹǤͳ
F
Ǧ
Castagliola (2005) proposed an -EWMA control chart for monitoring the process
dispersion. This control structure is based on a three parameter logarithmic transformation
which is given as:
£ ! £ ¤ ! ¥£
(3.5)
where is the sample variance for # pV sample defined as
©
¦ª̈«¬5§¨ § 6
`
, y represents
the wpV observation from the # pV sample of size Z from a normal distribution with mean ,
standard deviation ' and is the average of the # pV sample. The constants £ , £ and ¥£ are
defined as £ £ Z, ¥£ £ Z' and £
£ Z
M£ Z ¤' as in Castagliola
(2005). He derived the distribution of and showed that if the constants £ , £ and ¥£ are
judiciously selected, then the distribution of variable becomes very close to the normal
distribution with mean £ Z and variance '£ Z, i.e. ® [5 £ Z '£ Z6 (cf. Appendix
A in Castagliola (2005)). Table 3.7 reproduces the values of
£ Z,
'£ Z for Z PAB x > +B from Table I in Castagliola (2005).
DĞŵŽƌLJͲƚLJƉĞŽŶƚƌŽůŚĂƌƚƐŝŶ^ƚĂƚŝƐƚŝĐĂůWƌŽĐĞƐƐŽŶƚƌŽů
£ Z, £ Z, £ Z and
DŝdžĞĚtDͲh^hDĐŚĂƌƚƐ
Now using the approximately normally distributed variable from Castagliola (2005), the
plotting statistic for the -EWMA chart is defined as:
) * ! + *)
(3.6)
where * is the sensitivity parameter chosen as , * - + and the initial value of ) is taken
as )
given as:
£ Z
! £ Z ¤5+ ! £ Z6. The control limits for the statistic given in (3.6) are
£ Z /
where
The
0
0
'£ Z,
£ Z,
£ Z ! /
0
0
'£ Z
(3.7)
is the control limit coefficient that determines the distance between and .
Table 3.7: Values of ¯° , ±° , 9° , ²° and n°
3
9°
-0.6627
²°
1.8136
n°
0.6777
¯°
±°
0.02472
0.9165
4
-0.7882
2.1089
0.6261
0.01266
0.9502
5
-0.8969
2.3647
0.5979
0.00748
0.9670
6
-0.9940
2.5941
0.5801
0.00485
0.9765
7
-1.0827
2.8042
0.5678
0.00335
0.9825
8
-1.1647
2.9992
0.5588
0.00243
0.9864
9
-1.2413
3.1820
0.5519
0.00182
0.9892
10
-1.3135
3.3548
0.5465
0.00141
0.9912
11
-1.3820
3.5189
0.5421
0.00112
0.9927
12
-1.4473
3.6757
0.5384
0.00090
0.9938
13
-1.5097
3.8260
0.5354
0.00074
0.9947
14
-1.5697
3.9705
0.5327
0.00062
0.9955
15
-1.6275
4.1100
0.5305
0.00052
0.9960
values of -EWMA are given in Table 3.8 for different values of *, where ³
represents the amount of shift in the standard deviation (i.e. ³
the shifted standard deviation) and the in control
work on this subject. The
'
Y' with ' representing
is fixed at M, as is done in earlier
8 in this section are evaluated through simulation procedures
by running +´ replications. The program is developed in R language.
DĞŵŽƌLJͲƚLJƉĞŽŶƚƌŽůŚĂƌƚƐŝŶ^ƚĂƚŝƐƚŝĐĂůWƌŽĐĞƐƐŽŶƚƌŽů
ϱϱ
ϱϲ
Table 3.8: 9:; values for the µF -EWMA chart with
0.5
D => =?
; F> Fgh
9.257
D => E
; F> H?Fh
0.6
11.679
8.931
0.7
16.101
13.03
0.8
26.108
23.679
29.961
¶
6.866
D => F
; F> ?Fh
? and 9:;= F==
D => I
; F> gIHh
5.448
D => H
; F> gHIh
5.928
D => ?
; F> gIh
7.856
8.481
10.735
16.909
13.064
16.699
25.58
48.763
46.058
80.157
166.467
5.616
7.606
0.9
63.459
70.501
107.839
169.543
274.071
474.331
0.95
133.554
153.817
204.856
264.011
327.699
392.103
1
199.781
200.702
200.756
200.262
200.59
199.224
1.05
78.77
92.938
98.675
98.004
97.342
96.541
1.1
32.542
41.746
47.688
49.736
50.94
51.373
1.2
11.986
15.382
17.449
18.537
19.274
20.022
1.3
7.064
8.766
9.571
9.909
10.226
10.527
1.4
5.054
6.09
6.419
6.517
6.6
6.753
1.5
3.983
4.722
4.835
4.801
4.787
4.82
2
2.133
2.418
2.343
2.225
2.148
2.098
3
1.338
1.456
1.395
1.336
1.294
1.267
Note that the results from Table 3.8 coincide with the results of Table III of Castagliola
(2005).
͵ǤʹǤʹ Ǧ
F
Taking inspiration from the -EWMA chart, Castagliola et al. (2009) proposed a
CUSUM- chart for monitoring the process dispersion which is based on the statistic
given in (3.5). The CUSUM- chart uses two plotting statistics, named as and , and
given as:
5 £ Z6
.
5 £ Z6
where
!
!
(3.8)
is the reference value and the sensitivity parameter of the CUSUM- chart. The
initial values for the plotting statistics given in (3.8) are taken equal to zero, i.e.
DĞŵŽƌLJͲƚLJƉĞŽŶƚƌŽůŚĂƌƚƐŝŶ^ƚĂƚŝƐƚŝĐĂůWƌŽĐĞƐƐŽŶƚƌŽů
DŝdžĞĚtDͲh^hDĐŚĂƌƚƐ
. These plotting statistics are plotted against a control limit $ and an out of control signal is
received if either of the two statistics (i.e. and ) is plotted above $.
and $ are
jointly the two parameters of CUSUM- chart and their standard forms are given as:
&'£ Z,
$ ('£ Z
(3.9)
where k and h are constants which determine the properties of the chart. The
values for
the CUSUM- chart for different choices of its parameters are given in Table 3.9 with
fixed at M.
Table 3.9: 9:; values for the CUSUM-µF chart with
0.5
>+
$ +>BPh
0.6
11.382
8.452
0.7
15.489
12.169
0.8
24.394
21.619
29.699
60.662
170.474
0.9
54.649
63.997
116.766
226.253
475.443
0.95
114.332
143.17
213.662
294.554
379.161
1
199.846
199.241
199.841
200.769
199.625
1.05
102.864
104.01
104.806
102.974
100.707
1.1
52.641
50.675
53.502
54.227
54.99
1.2
25.057
20.867
20.373
20.777
21.353
1.3
16.451
12.763
11.255
11.092
11.18
1.4
12.434
9.256
7.654
7.2
7.115
1.5
10.068
7.341
5.832
5.26
5.126
2
5.509
3.885
2.873
2.41
2.197
3
3.347
2.379
1.686
1.404
1.3
6.516
>B
$ P>fBBh
? and 9:;= F==
³
9.059
>MB
$ X>AdXh
>dB
$ M>XMh
5.071
+
$ +>eXh
7.303
8.264
13.223
12.295
18.314
43.032
5.199
6.121
͵ǤʹǤ͵
In this subsection we propose a memory-type control chart which is based on mixing
the effects of EWMA and CUSUM charts into a single control chart structure. For the
location parameter this idea was explored in Section 3.1. Again let y (i.e. the wpV
DĞŵŽƌLJͲƚLJƉĞŽŶƚƌŽůŚĂƌƚƐŝŶ^ƚĂƚŝƐƚŝĐĂůWƌŽĐĞƐƐŽŶƚƌŽů
ϱϳ
ϱϴ
observation of # pV sample with w +M x > > Z and # +M x x>) be distributed normally with
mean and variance ' under an in control situation. Then the two plotting statistics
(named as 7 and 7 ) for the proposed CS-EWMA chart are given as:
7 5 £ Z6
.
7 5 £ Z6
where
q
q
! 7
q
! 7
(3.10)
is the reference value for the proposed chart, like K in (3.8). The initial value for
both plotting statistics is taken equal to zero, i.e. 7 7 . is the EWMA statistic
which is defined as:
* ! 5+ * 6
(3.11)
where * is the smoothing constant like * in (3.6) and is chosen as , * - +. is the
statistic defined in (3.5). The initial value for the statistic is taken as
£ Z
!
£ Z ¤5+ ! £ Z6. The statistics 7 and 7 are now plotted against the control limit $q
and an out of control signal is received if either of the two plotting statistics given in (3.10) is
plotted above $q . If 7 is plotted above $q that would indicate a positive shift in the process
standard deviation and if the value of 7 gets larger than $q then it would be declared that
the process standard deviation has shifted downwards. The standard forms of
depending upon the variance of (i.e. ·u¸ '£ Z ¹
.
q
& ¹'£ Z/
$q ( ¹'£ Z/
where
0
0
0
0
º
0
/0
4
3
0 1
º $ /
0
0
0
q
and $q
º, cf. (3.7)) are given as:
(3.12)
& '£ Z and $ ( '£ Z. Here in (3.12), we have used the asymptotic
standard deviation of the statistic but the practitioner may use the exact standard deviation
as discussed by Steiner (1999). Note that the CUSUM- is a special case of the proposed
DĞŵŽƌLJͲƚLJƉĞŽŶƚƌŽůŚĂƌƚƐŝŶ^ƚĂƚŝƐƚŝĐĂůWƌŽĐĞƐƐŽŶƚƌŽů
DŝdžĞĚtDͲh^hDĐŚĂƌƚƐ
CS-EWMA chart with * +. Finally, a detailed study on the
performance of the
proposed CS-EWMA chart is given in Tables 3.10 – 3.15, where the
and n = 5. The algorithm developed in R language for computing the
Appendix 3.2.
is fixed at M
8 is given in
Values of $ for sample sizes other than B can be found by the relation $`»´
$`¼´ \
i.
ii.
]½ `»´
a
]½ `¼´
for a fixed
for fixed values of Z,
M. From Tables 3.10 – 3.15, we can conclude that:
and
,
the value of $ decreases with an increase in the
for fixed values of Z, * and
,
large values of
value of * and vice versa;
are giving small
values
for detecting a positive shift in the process dispersion;
Table 3.10: 9:; values for the CS-EWMA chart with D => =?,
³
0.5
>+
$ XM>Xh
23.156
>MB
$ Ad>Xh
0.6
26.652
23.825
0.7
32.317
28.983
0.8
43.072
38.98
34.238
20.626
>B
$ Me>Bh
? and 9:;= F==
>dB
$ +f>+Bh
15.133
+
$ +>XMh
20.32
17.749
15.758
24.993
22.144
19.969
31.092
28.876
17.482
13.289
0.9
73.716
68.713
63.844
61.61
61.053
0.95
127.175
123.052
120.583
121.6
123.108
1
199.718
200.262
200.951
200.69
199.752
1.05
94.196
88.525
84.496
82.271
81.696
1.1
52.536
47.005
41.056
37.801
35.939
1.2
31.726
27.331
22.219
18.725
16.32
1.3
24.789
21.241
16.942
13.848
11.534
1.4
21.141
18.078
14.341
11.569
9.429
1.5
18.774
16.091
12.706
10.183
8.233
2
13.481
11.562
9.112
7.224
5.712
3
10.16
8.729
6.877
5.442
4.271
DĞŵŽƌLJͲƚLJƉĞŽŶƚƌŽůŚĂƌƚƐŝŶ^ƚĂƚŝƐƚŝĐĂůWƌŽĐĞƐƐŽŶƚƌŽů
ϱϵ
ϲϬ
Table 3.11: 9:; values for the CS-EWMA chart with D => E,
? and 9:;= F==
³
0.5
>+
$ Af>+h
17.682
>MB
$ PB>Bh
15.366
>B
$ MM>Mdh
12.769
>dB
$ +A>+h
10.995
+
$ f>f+h
0.6
20.542
17.895
14.96
13.009
11.626
0.7
25.299
22.109
18.707
16.487
15.072
0.8
34.747
30.731
26.633
24.372
23.188
0.9
63.816
58.764
55.477
55.787
57.655
0.95
115.831
113.49
115.278
120.601
126.425
9.717
1
199.304
199.885
200.85
199.751
200.113
1.05
102.296
96.919
95.415
95.943
96.935
1.1
55.637
50.376
45.925
44.641
44.409
1.2
30.526
26.334
22.075
19.62
18.345
1.3
22.596
19.244
15.62
13.331
11.874
1.4
18.601
15.786
12.672
10.624
9.178
1.5
16.218
13.729
10.956
9.064
7.73
2
11.075
9.385
7.424
6.039
4.997
3
8.098
6.881
5.438
4.395
3.559
Table 3.12: 9:; values for the CS-EWMA chart with D => F,
? and 9:;= F==
³
0.5
>+
$ PB>Mh
14.017
>MB
$ MA>eXh
11.652
>B
$ +B>Adh
9.421
>dB
$ +>Ph
8.106
+
$ X>BPh
0.6
16.572
13.807
11.243
9.796
8.858
0.7
20.988
17.597
14.579
12.978
12.141
0.8
30.235
25.828
22.383
21.258
21.411
0.9
59.291
54.54
54.423
59.188
65.493
0.95
112.046
111.302
120.099
132.689
145.693
1
200.201
199.572
200.733
200.695
200.392
1.05
104.94
100.801
100.762
103.131
104.378
1.1
56.775
51.104
48.576
48.77
49.734
1.2
29.533
24.981
21.284
19.8
19.186
1.3
20.761
17.182
13.998
12.367
11.485
1.4
16.545
13.588
10.88
9.367
8.401
1.5
14.048
11.505
9.131
7.731
6.791
2
9.041
7.414
5.805
4.797
4.078
3
6.369
5.252
4.117
3.372
2.797
DĞŵŽƌLJͲƚLJƉĞŽŶƚƌŽůŚĂƌƚƐŝŶ^ƚĂƚŝƐƚŝĐĂůWƌŽĐĞƐƐŽŶƚƌŽů
7.222
DŝdžĞĚtDͲh^hDĐŚĂƌƚƐ
Table 3.13: 9:; values for the CS-EWMA chart with D => I,
? and 9:;= F==
³
0.5
>+
$ Mf>+Ph
12.39
>MB
$ +e>Pdh
9.956
>B
$ ++>fdh
7.911
>dB
$ d>dfh
6.822
+
$ B>+Xh
0.6
14.876
11.99
9.621
8.444
7.772
0.7
19.235
15.661
12.875
11.729
11.362
0.8
28.467
23.896
21.086
21.167
22.79
0.9
57.66
53.647
56.972
66.231
78.395
0.95
111.3
112.994
128.065
147.259
164.091
6.134
1
200.262
199.154
199.005
200.568
199.814
1.05
105.433
101.759
102.912
105.905
106.502
1.1
56.456
50.968
49.573
51.114
52.289
1.2
28.726
23.988
20.749
19.841
19.724
1.3
19.793
16.096
13.122
11.908
11.206
1.4
15.467
12.395
9.873
8.664
7.917
1.5
12.948
10.325
8.125
6.969
6.265
2
7.991
6.376
4.933
4.108
3.535
3
5.473
4.409
3.412
2.789
2.366
Table 3.14: 9:; values for the CS-EWMA chart with D => H,
? and 9:;= F==
³
0.5
>+
$ MP>APh
11.425
>MB
$ +B>fPh
8.956
>B
$ e>XMh
7.014
>dB
$ X>Pfh
6.056
+
$ A>MAh
0.6
13.878
10.937
8.707
7.703
7.26
0.7
18.23
14.579
12.041
11.259
11.47
0.8
27.332
22.911
20.802
22.072
25.63
0.9
56.545
53.599
61.015
76.068
94.607
0.95
110.015
115.617
138.526
161.045
184.439
1
199.007
200.336
200.485
199.411
199.888
1.05
104.542
102.206
105.385
106.762
107.729
1.1
55.908
50.87
50.536
52.258
53.902
1.2
28.069
23.363
20.537
19.92
20.051
1.3
19.113
15.296
12.611
11.57
11.075
1.4
14.786
11.625
9.3
8.193
7.625
1.5
12.235
9.581
7.494
6.484
5.892
2
7.32
5.716
4.373
3.647
3.158
3
4.884
3.852
2.965
2.414
2.134
5.526
DĞŵŽƌLJͲƚLJƉĞŽŶƚƌŽůŚĂƌƚƐŝŶ^ƚĂƚŝƐƚŝĐĂůWƌŽĐĞƐƐŽŶƚƌŽů
ϲϭ
ϲϮ
Table 3.15: 9:; values for the CS-EWMA chart with D => ?,
iii.
³
0.5
>+
$ M>+h
10.819
>MB
$ +P>MBh
8.244
>B
$ d>eeh
6.394
>dB
$ B>Mdh
5.563
+
$ P>Bdh
0.6
13.25
10.215
8.096
7.292
7.082
0.7
17.575
13.835
11.575
11.222
12.113
0.8
26.704
22.207
21.05
23.96
30.154
0.9
56.118
53.989
65.611
86.907
113.788
0.95
111.12
118.942
146.808
177.365
205.588
1
200.775
199.405
199.3062
199.499
200.476
1.05
104.617
102.52
105.77
107.044
106.577
1.1
55.436
50.59
51.073
53.56
54.65
1.2
27.554
22.734
20.376
20.218
20.334
1.3
18.626
14.656
12.186
11.379
11.054
1.4
14.272
11.049
8.842
7.888
7.449
1.5
11.76
9.015
7.041
6.138
5.654
2
6.863
5.226
3.98
3.31
2.91
3
4.487
3.464
2.601
2.22
1.87
for fixed values of Z, * and
,
like >MB -
shifts in the negative direction, large values of
for fixed values of Z,
and
,
5.16
negative shifts of small magnitude are detected
efficiently using moderate values of
iv.
? and 9:;= F==
- >B, whereas for large
are recommended;
large values of * are recommended for detecting
large positive shifts and vice versa while for negative shifts, varied behavior is seen;
v.
for fixed values of Z, * and
value of
and vice versa;
,
the value of $ decreases with an increase in the
In this section we have used the statistic to design the control structure of our proposed
chart. Many other transformations of , that result into a statistic which is distributed
approximately normal, can be used. Acosta-Mejia et al. (1999) proposed two such
transformations named as ] and ¾. Castagliola et al. (2010) proposed a four parameter
Johnston transformation and named the resulting variable as . Huwang et al. (2010)
DĞŵŽƌLJͲƚLJƉĞŽŶƚƌŽůŚĂƌƚƐŝŶ^ƚĂƚŝƐƚŝĐĂůWƌŽĐĞƐƐŽŶƚƌŽů
DŝdžĞĚtDͲh^hDĐŚĂƌƚƐ
proposed a logarithmic transformation and showed that ¤ ¿ À Á follows
'
approximately a normal distribution which can be used to design a memory control chart to
monitor the process dispersion. In the next subsection we compare the control charts based on
these transformations with our proposal.
͵ǤʹǤͶ
This subsection contains the comparison of the proposed CS-EWMA chart with the
-EWMA, CUSUM- and some other recently proposed CUSUM and EWMA charts for
monitoring the process dispersion.
CS-EWMA versus µF -EWMA: The
values for the -EWMA chart (discussed in
Section 2.1) are given in Table 3.8. Comparison reveals that the performance of the proposed
chart with
+ is almost the same as compared to the -EWMA for the positive shifts,
but for negative shifts, the performance of the CS-EWMA chart is far more superior than the
-EWMA. Moreover, the performance of -EWMA becomes very poor for moderate and
large values of *, even the
values become larger than the prefixed
for negative
shifts. This is not the case with the proposed CS-EWMA chart, as for large values of * , the
CUSUM factor in the proposed chart still makes it remain better in terms of
3.8 vs. Tables 3.10 – 3.15).
CS-EWMA versus CUSUM-µF : Table 3.9 contains the
8 (cf. Table
values for the CUSUM-
chart proposed by Castagliola et al. (2009). This CUSUM- chart is a special form of our
proposed CS-EWMA chart with * +. The performance of the proposed chart is better than
CUSUM- chart, especially for small values of smoothing constant * . The additional
DĞŵŽƌLJͲƚLJƉĞŽŶƚƌŽůŚĂƌƚƐŝŶ^ƚĂƚŝƐƚŝĐĂůWƌŽĐĞƐƐŽŶƚƌŽů
ϲϯ
ϲϰ
parameter (i.e. * ) in CS-EWMA chart makes sure that the performance of the proposed chart
is not deflated for large values of
of
like it occurs with the CUSUM- chart for large values
(cf. Table 3.9 vs. Tables 3.10 – 3.15).
CS-EWMA versus some other EWMA and CUSUM charts: Acosta-Mejia et al. (1999)
proposed some CUSUM-type charts (named as ¾ CUSUM, ] CUSUM and CUSUM) for
monitoring process dispersion and compared the performance of their proposed charts with
the CUSUM
chart by Page (1963), CUSUM chart by Tuprah and Ncube (1987) and
CUSUM ¤ chart by Chang and Gan (1995). They showed through comparison that their
proposed CUSUM is performing better (in terms of
values) than the other charts
discussed. Similarly, Huwang et al. (2010) proposed two new EWMA type charts (named as
$$R+ CUSUM and $$RM CUSUM) for monitoring the standard deviation of a process.
They also compared the performance of their proposed chart with some other competitors,
like the $ EWMA chart by Crowder and Hamilton (1992) and the  EWMA chart by Shu
and Jiang (2008), and showed that their proposed charts are more sensitive (in detecting
shifts) than the other competitors. The
values of the charts discussed by Acosta-Mejia et
al. (1999) are given in Table 3.16, while Table 3.17 contains the
by Huwang et al. (2010).
8 of the charts discussed
Table 3.16: 9:; values for some one-sided CUSUM-type charts for detecting variance increases
with ? and 9:;= F==
199.93
M>BX
$ A>ffh
201.80
¾
CUSUM
>Pf
$ A>Mfh
200.70
]
CUSUM
>Pf
$ A>Mfh
201.10
CUSUM
+>+PA
$ +>eh
200.60
h
CUSUM
+>+eP
$ +f>ABh
42.94
40.40
41.04
41.04
38.80
34.60
1.2
18.07
17.60
17.17
17.15
16.85
14.14
1.3
10.75
10.82
10.23
10.21
10.36
8.42
1.4
7.63
7.81
7.26
7.24
7.50
5.93
1.5
5.98
6.13
5.66
5.65
5.85
4.58
2
3.18
3.13
2.90
2.98
3.01
2.20
1
CUSUM
¤
>Xf
$ M>XXh
1.1
³
CUSUM
DĞŵŽƌLJͲƚLJƉĞŽŶƚƌŽůŚĂƌƚƐŝŶ^ƚĂƚŝƐƚŝĐĂůWƌŽĐĞƐƐŽŶƚƌŽů
200.76
DŝdžĞĚtDͲh^hDĐŚĂƌƚƐ
ϲϱ
Table 3.17: 9:; values for some one-sided EWMA-type charts for detecting variance increases
with ? and 9:;= F==
³
* >B
* >+
CH-EWMA
( +>BB)
SJ-EWMA
( +>BXf)
200.75
HHW1EWMA
( +>fMf)
200.92
HHW2EWMA
( +>fdM)
1.1
43.24
32.26
28.89
27.28
44.26
35.15
34.32
32.05
1.2
18.09
14.43
11.69
10.78
18.23
14.96
14.1
12.69
1.3
10.77
9.17
6.85
6.2
10.56
9.09
8.2
7.21
1.4
7.63
6.73
4.75
4.24
7.35
6.53
5.65
4.89
1.5
5.98
5.38
3.62
3.22
5.68
5.13
4.28
3.68
2
3.18
2.93
1.8
1.62
2.95
2.72
2.03
1.76
CH-EWMA
( +>B+P)
SJ-EWMA
( M>Md)
199.48
HHW1EWMA
( M>MBP)
199.43
HHW2EWMA
( M>PBB)
CH-EWMA
( +>Bef)
SJ-EWMA
( M>APP)
199.67
HHW1EWMA
( M>PM)
200.22
HHW2EWMA
( M>Add)
1.1
46.63
39.73
41.18
37.87
48.48
43.45
46.14
41.79
1.2
18.79
16.05
16.66
14.7
19.52
17.25
18.65
16.2
1.3
10.54
9.21
9.45
8.16
10.67
9.56
10.35
8.82
1.4
7.16
6.4
6.45
5.49
7.09
6.43
6.9
5.81
1.5
5.41
4.89
4.83
4.07
5.24
4.8
5.11
4.26
2
2.67
2.45
2.24
1.88
2.47
2.3
2.32
1.93
1
³
1
200.33
200.64
* >M
199.57
200.65
CH-EWMA
( +>PP)
SJ-EWMA
( +>eAP)
200.36
HHW1EWMA
( M>de)
199.51
HHW2EWMA
( M>+Pe)
200.02
199.4
Table 3.18: 9:; values for one-sided CS-EWMA chart with r E,
* >P
200.35
199.45
? and 9:;= F==
³
1
* >B
$ B>Peh
200.4035
* >+
$ B>+Ph
200.9845
* >M
$ A>BAh
200.7247
* >P
$ P>eMfh
200.1826
* >A
$ P>AMh
200.621
* >B
$ P>+h
1.1
23.963
31.478
36.48
39.44
41.226
43.263
1.2
11.261
13.798
15.39
16.031
16.599
17.147
1.3
8.006
9.1
9.485
9.553
9.635
9.684
199.6103
1.4
6.53
7.041
7.012
6.837
6.721
6.641
1.5
5.685
5.941
5.689
5.432
5.224
5.112
2
3.975
3.834
3.423
3.08
2.858
2.673
3
3.04
2.732
2.347
2.143
1.894
1.674
All charts presented in Tables 3.16 – 3.17 are one-sided, i.e. designed to detect just positive
shift in the process dispersion. For a valid comparison of the proposed chart with these charts,
we have evaluated the
values of the proposed chart with the one-sided structure for
DĞŵŽƌLJͲƚLJƉĞŽŶƚƌŽůŚĂƌƚƐŝŶ^ƚĂƚŝƐƚŝĐĂůWƌŽĐĞƐƐŽŶƚƌŽů
ϲϲ
8 are given in Table 3.18
monitoring an increase in process standard deviation. These
with
+ and
M.
It can be noticed through the comparison that the proposed chart is outperforming all its
competitors in the form of EWMA and CUSUM type charts (cf. Tables 3.16 – 3.17 vs. Table
3.18).
Finally, before concluding this subsection, we provide the
discussed above. Figure 3.7 contains the
EWMA chart with * >M,
curves of the different charts
curves of the two-sided charts containing: CS-
>MB and $ MA>eX; -EWMA chart with * >M and
M>BeM; CUSUM- chart with
>MB and $ X>AdX. Figure 3.8 shows the
curves of the one-sided charts for positive shifts containing: CS-EWMA chart with *
>B,
+ and $ B>Pe; CUSUM chart with
chart with
+>+eP and $ +f>AB; CUSUM
+>+PA and $ +>e; $ EWMA chart with * >B and
EWMA chart with * >B and
+>fdM.
+>BB; Â
+>BXf; and $$RM EWMA chart with * >B and
Figure 3.7: 9:; curves for two-sided structures of CS-EWMA, µF -EWMA and CUSUM-µF
charts with 9:;= F==
^ͲtD
h^hD
tD
ϮϱϬ
ϮϬϬ
Z>Ɛ
ϭϱϬ
ϭϬϬ
ϱϬ
Ϭ
Ϭ͘ϱ
Ϭ͘ϲ
Ϭ͘ϳ
Ϭ͘ϴ
Ϭ͘ϵ
Ϭ͘ϵϱ
ϭ
Ã
ϭ͘Ϭϱ
ϭ͘ϭ
ϭ͘Ϯ
DĞŵŽƌLJͲƚLJƉĞŽŶƚƌŽůŚĂƌƚƐŝŶ^ƚĂƚŝƐƚŝĐĂůWƌŽĐĞƐƐŽŶƚƌŽů
ϭ͘ϯ
ϭ͘ϰ
ϭ͘ϱ
Ϯ
ϯ
DŝdžĞĚtDͲh^hDĐŚĂƌƚƐ
Figure 3.8: 9:; curves for one-sided structures of CS-EWMA, nÄ CUSUM, CUSUM µ, µÅ
EWMA, ns EWMA and ssQF EWMA charts with 9:;= F==
^ͲtD
Wh^hD
h^hD^
ϱϬ
ϰϬ
Z>Ɛ
ϯϬ
ϮϬ
ϭϬ
Ϭ
ϭ͘ϭ
ϭ͘Ϯ
ϭ͘ϯ
Ã
ϭ͘ϰ
ϭ͘ϱ
Ϯ
Figure 3.7 shows that the performance of all three charts is almost the same for positive shifts
but for negative shifts, the proposed chart is giving a better
Figure 3.8, the
performance. Similarly, in
curve of the proposed chart seems lower than all other curves for small
values of ³ and thus showing a better performance for small positive shifts.
͵ǤʹǤͷ
An application of CS-EWMA chart on a simulated dataset is provided in this
subsection to show the implementation of the proposal. For this purpose, two datasets are
generated having 40 samples of size Z B for both the samples. First 20 samples are
generated from [+A referring to an in control situation with + and ' A. For
dataset 3.1, the remaining 20 observations are generated from [+B showing a positive
DĞŵŽƌLJͲƚLJƉĞŽŶƚƌŽůŚĂƌƚƐŝŶ^ƚĂƚŝƐƚŝĐĂůWƌŽĐĞƐƐŽŶƚƌŽů
ϲϳ
ϲϴ
shift in the process standard deviation with ³ _BÀ +>++f. Similarly, for dataset 3.2,
_A
the remaining 20 observations are generated from [+P showing a negative shift in
dispersion with ³ >fXX. Now the -EWMA chart, CUSUM- chart and their mixture
named as CS-EWMA chart are applied to the given datasets. The chart output of the -
EWMA chart with * >M and
M>BeM is shown in Figure 3.9. Figure 3.10 shows the
graphical display of the CUSUM- with
the proposed chart with * >M,
>B and $ P>fBB. The calculation details for
>B Æ
q
>+Xd and $ +B>Ad Æ $q B>+Bd
are given in Table 3.19, whereas the chart output is given in Figure 3.11.
Figure 3.9: Graphical display of the µF -EWMA chart with D => F and ; F> ?F
ũ;ĚĂƚĂƐĞƚϯ͘ϭͿ
ũ;ĚĂƚĂƐĞƚϯ͘ϮͿ
ĐŽŶƚƌŽůůŝŵŝƚƐ
ϭ
Ϭ͘ϴ
Ϭ͘ϲ
Ϭ͘ϰ
Ϭ͘Ϯ
Ϭ
ͲϬ͘Ϯ
ͲϬ͘ϰ
ͲϬ͘ϲ
ͲϬ͘ϴ
Ͳϭ
ϭ
ϯ
ϱ
ϳ
ϵ
ϭϭ ϭϯ ϭϱ ϭϳ ϭϵ Ϯϭ Ϯϯ Ϯϱ Ϯϳ Ϯϵ ϯϭ ϯϯ ϯϱ ϯϳ ϯϵ
Figures 3.9 – 3.11 clearly indicate that all three charts are giving out of control signals at
samples 39 and 40 for dataset 3.1 (i.e. positive shift in the process dispersion). Moreover, the
proposed CS-EWMA chart detected a negative shift for dataset 3.2 at samples 37, 38, 39 and
DĞŵŽƌLJͲƚLJƉĞŽŶƚƌŽůŚĂƌƚƐŝŶ^ƚĂƚŝƐƚŝĐĂůWƌŽĐĞƐƐŽŶƚƌŽů
DŝdžĞĚtDͲh^hDĐŚĂƌƚƐ
40 but the other two charts did not signal the downward shift in the process standard
deviation.
Figure 3.10: Graphical display of the CUSUM µF chart with r => ? and s I> G??
н;ĚĂƚĂƐĞƚϯ͘ϭͿ
Ͳ ;ĚĂƚĂƐĞƚϯ͘ϮͿ
,
ϱ
ϰ͘ϱ
ϰ
ϯ͘ϱ
ϯ
Ϯ͘ϱ
Ϯ
ϭ͘ϱ
ϭ
Ϭ͘ϱ
Ϭ
ϭ
ϯ
ϱ
ϳ
ϵ
ϭϭ
ϭϯ ϭϱ
ϭϳ
ϭϵ
Ϯϭ Ϯϯ
Ϯϱ
Ϯϳ
Ϯϵ ϯϭ
ϯϯ
ϯϱ
ϯϳ ϯϵ
^ĂŵƉůĞEƵŵďĞƌ
Figure 3.11: Graphical display of the CS-EWMA chart with D => F, r => ? and s
E?> HJ
Dн;ĚĂƚĂƐĞƚϯ͘ϭͿ
DͲ ;ĚĂƚĂƐĞƚϯ͘ϮͿ
,ƋΖ
ϴ
ϳ
ϲ
ϱ
ϰ
ϯ
Ϯ
ϭ
Ϭ
ϭ
ϯ
ϱ
ϳ
ϵ
ϭϭ
ϭϯ
ϭϱ
ϭϳ
ϭϵ
Ϯϭ
Ϯϯ
Ϯϱ
Ϯϳ
Ϯϵ
ϯϭ
ϯϯ
ϯϱ
ϯϳ
ϯϵ
^ĂŵƉůĞEƵŵďĞƌ
DĞŵŽƌLJͲƚLJƉĞŽŶƚƌŽůŚĂƌƚƐŝŶ^ƚĂƚŝƐƚŝĐĂůWƌŽĐĞƐƐŽŶƚƌŽů
ϲϵ
ϳϬ
Table 3.19: Calculation details of the proposed CS-EWMA chart for dataset 3.1
Sample
No.
1
5.61
0.74
0.32
7
0.14
7
0
Sample
No.
21
0.81
-1.42
-0.44
7
7
0
0.32
2
4.48
0.38
0.33
0.3
0
22
11.47
2.04
0.06
0
0.1
3
2.58
-0.38
0.19
0.31
0
23
4.69
0.45
0.14
0
0
4
1.7
-0.84
-0.02
0.12
0
24
3.21
-0.1
0.09
0
0
5
7.04
1.13
0.21
0.16
0
25
6.42
0.97
0.26
0.09
0
6
5.96
0.84
0.34
0.32
0
26
3.84
0.15
0.24
0.16
0
7
3.84
0.15
0.3
0.45
0
27
4.94
0.54
0.3
0.29
0
8
3.29
-0.07
0.23
0.5
0
28
9.35
1.65
0.57
0.68
0
9
8.62
1.5
0.48
0.81
0
29
4.21
0.29
0.51
1.02
0
10
10.33
1.84
0.75
1.39
0
30
10.5
1.87
0.79
1.64
0
11
3.33
-0.05
0.59
1.81
0
31
1.45
-0.99
0.43
1.89
0
12
0.65
-1.54
0.17
1.8
0
32
10.1
1.8
0.7
2.42
0
13
3.99
0.21
0.17
1.8
0
33
3.16
-0.12
0.54
2.79
0
14
2.21
-0.57
0.03
1.65
0
34
5.16
0.61
0.55
3.16
0
15
0.88
-1.38
-0.25
1.22
0.1
35
7.84
1.32
0.71
3.7
0
16
9.86
1.75
0.15
1.19
0
36
3.14
-0.13
0.54
4.06
0
17
5.48
0.71
0.26
1.28
0
37
3.44
-0.01
0.43
4.32
0
18
1.1
-1.22
-0.04
1.07
0
38
6.96
1.11
0.57
4.71
0
19
3.48
0.01
-0.03
0.86
0
39
11.04
1.97
0.85
5.38*
0
20
1.67
-0.86
-0.19
0.5
0.03
40
6.47
0.98
0.87
6.08*
0
* indicates an out of control signal by CS-EWMA chart
͵Ǥ͵
CUSUM control charts and EWMA control charts are the two most commonly used
memory control charts in the literature. These control schemes do 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 chapter we have combined the CUSUM and EWMA
control schemes into a single control structure and proposed a mixed EWMA-CUSUM
control scheme for monitoring the process location. Performance of the proposed scheme is
compared with other CUSUM and EWMA type control charts which are meant to detect
small and moderate shifts in the process. The comparisons revealed that the proposed scheme
DĞŵŽƌLJͲƚLJƉĞŽŶƚƌŽůŚĂƌƚƐŝŶ^ƚĂƚŝƐƚŝĐĂůWƌŽĐĞƐƐŽŶƚƌŽů
DŝdžĞĚtDͲh^hDĐŚĂƌƚƐ
is really good at detecting the smaller shifts in the process as compared to the other schemes
under study.
Following our proposal for location, this chapter also proposes a new memory-type
control chart, named as the CS-EWMA chart, for monitoring the process standard deviation.
The design of the proposed chart is based on mixing the effects of EWMA and CUSUM
control charts into a single control structure. The performance of the proposed CS-EWMA
chart is evaluated using
as indicator. In terms of
values, the proposed chart is
compared with existing CUSUM and EWMA control charts and it is noticed that the
proposed chart has better performance for both positive as well as negative shifts in the
process dispersion. Finally, an illustrative example is provided to show the application of the
proposed chart.
DĞŵŽƌLJͲƚLJƉĞŽŶƚƌŽůŚĂƌƚƐŝŶ^ƚĂƚŝƐƚŝĐĂůWƌŽĐĞƐƐŽŶƚƌŽů
ϳϭ
ϳϮ
͵Ǥͳ
x=c(); Q=c(); sdp=c(); K=c(); H=c(); Mp=c(); Mn=c(); rl=c(); k=0.5;
h=37.42; ld=0.1
mu=0; sig=1
for(j in 1:50000)
{
for(i in 1:1000000)
{
x[i]=rnorm(1,mu,sig)
if(i==1)
{Q[i]=ld*x[i]+(1-ld)*mu;}
else{Q[i]=ld*x[i]+(1-ld)*Q[i-1];}
sdp[i]=sqrt((ld/(2-ld))*(1-(1-ld)^(2*i)))
K[i]=k*sdp[i]
H[i]=h*sdp[i]
if(i==1)
{Mp[i]=max(0,Q[i]-K[i]);}
else{Mp[i]=max(0,Q[i]-K[i]+Mp[i-1]);}
if(i==1)
{Mn[i]=max(0,-Q[i]-K[i]);}
else{Mn[i]=max(0,-Q[i]-K[i]+Mn[i-1]);}
if(Mp[i]>H[i] | Mn[i]>H[i])
{rl[j]=i; break;}
else{rl[j]=0;}
}
}
mean(rl)
͵Ǥʹ
Q=c(); Mp=c(); Mn=c(); rl=c(); Tk=c()
n=5; a=-0.8969; b=2.3647; c=0.5979; ETi=0.00748; STi=0.9670
Z0=a+b*log(1+c,exp(1))
ld=0.05; K=0.1*sqrt(ld/(2-ld)); H=62.6*sqrt(ld/(2-ld))
for(j in 1:50000)
{
for(i in 1:1000000)
{
x=rnorm(n,0,1)
T[i]=a+b*log(var(x)+c,exp(1))
if(i==1)
{Q[i]=ld*T[i]+(1-ld)*Z0;}
else{Q[i]=ld*T[i]+(1-ld)*Q[i-1];}
if(i==1)
DĞŵŽƌLJͲƚLJƉĞŽŶƚƌŽůŚĂƌƚƐŝŶ^ƚĂƚŝƐƚŝĐĂůWƌŽĐĞƐƐŽŶƚƌŽů
DŝdžĞĚtDͲh^hDĐŚĂƌƚƐ
{Mp[i]=max(0,Q[i]-ETi-K);}
else{Mp[i]=max(0,Q[i]-ETi-K+Mp[i-1]);}
if(i==1)
{Mn[i]=max(0,-Q[i]+ETi-K);}
else{Mn[i]=max(0,-Q[i]+ETi-K+Mn[i-1]);}
if(Mp[i] > H | Mn[i] > H)
{rl[j]=i;break;}
else{rl[j]=0;}
}
}
mean(rl)
DĞŵŽƌLJͲƚLJƉĞŽŶƚƌŽůŚĂƌƚƐŝŶ^ƚĂƚŝƐƚŝĐĂůWƌŽĐĞƐƐŽŶƚƌŽů
ϳϯ
ϳϰ
DĞŵŽƌLJͲƚLJƉĞŽŶƚƌŽůŚĂƌƚƐŝŶ^ƚĂƚŝƐƚŝĐĂůWƌŽĐĞƐƐŽŶƚƌŽů
ŚĂƉƚĞƌϰ
Ͷ
Two popular categories of control charts are CUSUM and EWMA charts, which are good at
quickly detecting the presence of small and moderate disturbances. Targeting on small and
moderate shifts in the process mean, this chapter proposes EWMA- and CUSUM-type
control charts which utilize the information of auxiliary variable(s). The regression estimation
technique for the mean is used in defining the control structure of the proposed charts. It is
shown that the proposed charts are performing better than their competitors which are also
designed for detecting small shifts. This chapter is based on an article by Abbas, Riaz and
Does (2012c), consisting of the EWMA-type control charts based on auxiliary information.
ͶǤͳ
Information accessible at the stage of estimation other than that in the sample is called
auxiliary information. The concept of using auxiliary information is frequently used in the
field of survey sampling and estimation techniques. Auxiliary information can be used at
either or both the design and estimation stage. Sampling techniques like probability sampling
(cf. Fuller (2009)) and rank set sampling (cf. McIntyre (1952)) are examples of the utilization
of auxiliary information at the sample selection stage. Ratio, product and regression-type
estimators are examples of utilization of auxiliary information at the estimation stage (cf.
Cochran (1977) and Fuller (2002)). These estimators are designed in such a way that they not
DĞŵŽƌLJͲƚLJƉĞŽŶƚƌŽůŚĂƌƚƐŝŶ^ƚĂƚŝƐƚŝĐĂůWƌŽĐĞƐƐŽŶƚƌŽů
ϳϱ
ϳϲ
only utilize the sample information but also make use of the information available other than
that. This makes these estimators more efficient than the conventional ones.
This auxiliary information is also used in control charting techniques in order to
enhance their performance. Examples are the regression control chart proposed by Mandel
(1969) and cause-selecting control charts proposed by Zhang (1985). The control structure of
these control charts is based on regressing the study variable on the auxiliary variable. The
residuals obtained from that regression are used for monitoring the process (see also Wade
and Woodall (1993)). Riaz (2008a) introduced the concept of using auxiliary information at
the time of estimating the plotting statistic of a control chart. He proposed a control chart
which uses a regression-type estimator as the plotting statistic to monitor the variability of the
process and showed the dominance (in terms of power) of his proposed control chart over the
well-known Shewhart-type control charts for the same purpose (i.e. , and charts). Riaz
(2008b) proposed a regression-type estimator to monitor the location of the process. He not
only showed the superiority of his proposal over the Shewhart’s chart but also over the
regression charts (cf. Mandel (1969)) and the cause-selecting charts (cf. Zhang (1985)). Later
Riaz and Does (2009) proposed another variability chart based on a ratio-type estimator and
proved the dominance of their proposed chart over the one based on regression-type
estimator. Following these authors, this chapter proposes the use of auxiliary information
with the control structure of EWMA and CUSUM charts. The regression estimation
technique is used to exploit the information from the auxiliary variable and without loss of
generality the case of individual observations have been considered.
DĞŵŽƌLJͲƚLJƉĞŽŶƚƌŽůŚĂƌƚƐŝŶ^ƚĂƚŝƐƚŝĐĂůWƌŽĐĞƐƐŽŶƚƌŽů
ƵdžŝůŝĂƌLJŝŶĨŽƌŵĂƚŝŽŶďĂƐĞĚh^hDĂŶĚtDĐŽŶƚƌŽůĐŚĂƌƚƐ
ͶǤʹ
Let an auxiliary variable be correlated with the variable of interest and let us
denote the correlation between these two variables by Ç . The observations of and are
obtained in the paired form for each sample and the population mean and variance of (i.e.
and ' respectively) are assumed to be known. Also we assume bivariate normality of
and , i.e.
'
\ a È[ É\ a Ê
Ç ' '
Ç ' '
ËÌ
'
(4.1)
where [ represents the bivariate normal distribution. The regression estimate of the
population mean (cf. Cochran (1977)) is given as:
7§ ! Í
(4.2)
where Í is the change in due to one unit change in and is Í Ç \ Î a. The mean
and variance of the statistic 7 in (4.2) are given as:
Ð7 , 7 'Ñ ' + Ç
]
]Ï
(4.3)
Equation (4.3) implies that 7 is also an unbiased estimator of and 'Ñ , ' as long as
Ç
% . Based on the regression estimator in (4.2), the plotting statistic for the proposed
EWMA chart based on a single auxiliary variable (named as MXEWMA chart) is defined as:
q
)q *q 7§ ! + *q )
(4.4)
where *q is the smoothing constant for the proposed statistic and 7§ is the value of statistic
q
7 for the # pV sample. )
represents the past information (like ) ) and its initial value
DĞŵŽƌLJͲƚLJƉĞŽŶƚƌŽůŚĂƌƚƐŝŶ^ƚĂƚŝƐƚŝĐĂůWƌŽĐĞƐƐŽŶƚƌŽů
ϳϳ
ϳϴ
(i.e. )q ) is also taken equal to the target mean , i.e. the in control mean of . Now based on
(4.3) the time varying control limits for the proposed chart are:
.
q 'Ñ /
0Ò
0Ò
+ + *q 4
2
3
0Ò
2
q
q
! 'Ñ / Ò + + *
0
1
where
q
(4.5)
determines the width of the control limits for the proposed MXEWMA chart. The
values for the proposed MXEWMA chart with time varying limits are given in Tables
4.1 – 4.5 for some selective choices of Ç in which @ represents the amount of shift in the
study variable, i.e. @
ÓÔ¬ Ô^ Ó
]Î
where is the out of control mean of . The program
(developed in R language) for evaluating the
8 is given in Appendix 4.1 which is
replicated 50,000 for each simulated value. Note that remains constant.
Table 4.1: 9:; values for the proposed MXEWMA chart with time varying limits, ÕÖ× => =?
and 9:;= ?==
@
0
*q >P
q
M>AfP
500.3011
*q >B
q
M>XPe
500.8313
*q >+
M>fMA
499.7023
*q >MB
q
P
0.25
66.3103
77.5273
103.1312
168.8081
254.324
321.6227
0.5
21.2101
23.6211
28.7434
47.2173
88.1445
139.8713
0.75
10.7125
11.8433
13.592
19.2345
35.401
62.2502
1
6.632
7.2905
8.2072
10.3617
17.1148
30.3995
1.5
3.4443
3.7573
4.1644
4.7647
6.2628
9.7633
2
2.244
2.4205
2.6563
2.9313
3.3701
4.4521
2.5
1.6555
1.7719
1.9147
2.0831
2.2565
2.6079
3
1.3401
1.4104
1.5112
1.6109
1.6904
1.8151
4
1.0642
1.0882
1.1224
1.1609
1.184
1.1946
5
1.0049
1.0085
1.0131
1.0221
1.0265
1.0281
q
499.6045
DĞŵŽƌLJͲƚLJƉĞŽŶƚƌŽůŚĂƌƚƐŝŶ^ƚĂƚŝƐƚŝĐĂůWƌŽĐĞƐƐŽŶƚƌŽů
*q >B
P>dM
q
500.9026
*q >dB
q
P>ff
499.9704
ƵdžŝůŝĂƌLJŝŶĨŽƌŵĂƚŝŽŶďĂƐĞĚh^hDĂŶĚtDĐŽŶƚƌŽůĐŚĂƌƚƐ
Table 4.2: 9:; values for the proposed MXEWMA chart with time varying limits, ÕÖ× => F?
and 9:;= ?==
@
500.0592
*q >B
q
M>XPe
500.5358
*q >+
M>fMA
499.1371
*q >MB
q
P
q
0
*q >P
q
M>AfP
499.8049
*q >B
P>dM
499.951
*q >dB
q
P>ff
0.25
63.1554
73.8038
97.5858
161.9697
245.9632
313.2722
0.5
20.1268
22.4316
27.1393
44.1797
83.0098
132.8202
0.75
10.1982
11.2488
12.8277
18.0008
32.9028
57.9059
1
6.3049
6.9254
7.7548
9.7748
15.8281
28.0123
1.5
3.2871
3.5816
3.9532
4.5005
5.832
8.9444
2
2.1384
2.3119
2.5261
2.7962
3.1754
4.0979
2.5
1.5945
1.6997
1.8406
1.9877
2.1497
2.4354
3
1.2952
1.363
1.4519
1.5523
1.6178
1.7158
4
1.0495
1.0687
1.0955
1.132
1.1476
1.1565
5
1.0044
1.0067
1.0108
1.0149
1.0178
1.0183
q
500.5678
Table 4.3: 9:; values for the proposed MXEWMA chart with time varying limits, ÕÖ× => ?
and 9:;= ?==
@
0
*q >P
q
M>AfP
500.7792
*q >B
q
M>XPe
499.5635
*q >+
M>fMA
499.8114
*q >MB
q
P
499.692
*q >B
P>dM
500.7859
*q >dB
q
P>ff
0.25
52.6523
61.1011
80.6591
135.7683
216.0974
285.4555
0.5
16.6817
18.5391
22.044
34.5785
65.3989
108.1412
0.75
8.4242
9.3261
10.5432
14.1099
24.7989
44.3004
1
5.2477
5.7592
6.4306
7.8001
11.9099
20.6193
1.5
2.7663
3.0129
3.3183
3.7182
4.5573
6.5619
2
1.8392
1.9673
2.1455
2.3486
2.5852
3.136
2.5
1.3979
1.48
1.5834
1.7044
1.7964
1.9476
3
1.17
1.2168
1.2786
1.3501
1.3947
1.4325
4
1.0155
1.0253
1.038
1.0513
1.0606
1.064
5
1.0007
1.0012
1.0019
1.002
1.0028
1.0056
q
q
500.9686
DĞŵŽƌLJͲƚLJƉĞŽŶƚƌŽůŚĂƌƚƐŝŶ^ƚĂƚŝƐƚŝĐĂůWƌŽĐĞƐƐŽŶƚƌŽů
ϳϵ
ϴϬ
Table 4.4: 9:; values for the proposed MXEWMA chart with time varying limits, ÕÖ× => J?
and 9:;= ?==
@
0
*q >P
q
M>AfP
500.0659
*q >B
q
M>XPe
499.5868
*q >+
M>fMA
500.7509
*q >MB
q
P
500.4051
*q >B
P>dM
499.2859
*q >dB
q
P>ff
0.25
33.8532
38.4587
48.8401
84.1265
146.8905
211.6496
0.5
10.6126
11.7078
13.4213
18.9673
34.9019
61.2737
0.75
5.398
5.9343
6.6291
8.0816
12.4585
21.7022
1
3.4068
3.7177
4.1045
4.7175
6.1522
9.5862
1.5
1.8835
2.0268
2.2095
2.4126
2.6787
3.2731
2
1.3299
1.4043
1.4969
1.5949
1.6784
1.7904
2.5
1.1006
1.1309
1.1772
1.2269
1.2577
1.2721
3
1.0199
1.0296
1.0429
1.0608
1.0733
1.0742
4
1.0001
1.0004
1.0006
1.0011
1.0014
1.0017
5
1
1
1
1
1
1
q
q
499.7069
Table 4.5: 9:; values for the proposed MXEWMA chart with time varying limits, ÕÖ× => ?
and 9:;= ?==
@
0
*q >P
q
M>AfP
500.1483
*q >B
q
M>XPe
499.3424
*q >+
M>fMA
500.5895
*q >MB
q
P
500.2327
*q >B
P>dM
500.0067
*q >dB
q
P>ff
0.25
9.6039
10.6139
12.1289
16.7024
30.2287
53.736
0.5
3.1274
3.4017
3.7577
4.2576
5.389
8.1474
0.75
1.75
1.8739
2.0315
2.2142
2.4181
2.8603
q
q
500.7743
1
1.2547
1.3136
1.3951
1.4846
1.5468
1.622
1.5
1.0102
1.0162
1.0216
1.034
1.0418
1.0432
2
1
1.0003
1.003
1.0005
1.0006
1.0013
2.5
1
1
1
1
1
1
3
1
1
1
1
1
1
4
1
1
1
1
1
1
5
1
1
1
1
1
1
The
in Tables 4.1 – 4.5 is fixed at B which will enable us to make
comparison of the proposed control chart with some other charts/schemes. Tables 4.1 – 4.5
refer to a situation where the information about the population correlation coefficient is
DĞŵŽƌLJͲƚLJƉĞŽŶƚƌŽůŚĂƌƚƐŝŶ^ƚĂƚŝƐƚŝĐĂůWƌŽĐĞƐƐŽŶƚƌŽů
ƵdžŝůŝĂƌLJŝŶĨŽƌŵĂƚŝŽŶďĂƐĞĚh^hDĂŶĚtDĐŽŶƚƌŽůĐŚĂƌƚƐ
assumed to be known, because the information about the population correlation coefficient
Ç is known in many practical situations (cf. Garcia and Cebrian (1996)). However, this
may not be the case in every situation. Then we have to estimate the value of Ç from
preliminary samples (L). The estimators used to estimate Ç and Í are:
ÇØ
¦Ù
§«¬§ §
©
© Ù
/¦Ù
§«¬§ ¦§«¬§
¦ § §
ÍÚ §«¬
¦Ù
©
Ù
,
§«¬
§
(4.6)
From Tables 4.1 – 4.5, the main findings about our proposed MXEWMA control chart
for monitoring the location of a process are given as:
i.
the use of auxiliary information in the form of a regression estimator boosts the
ii.
for fixed values of Ç and @, the performance of the proposed chart with time
iii.
iv.
performance of EWMA control chart, especially for large values of Ç ;
varying limits is better for small values of *q ;
for fixed values of *q ,
large values of Ç ;
for all choices of *q ,
exceeds
v.
q
q
and @, the performance of the proposed chart is better for the
and Ç the proposed chart is
for any value of @;
for small values of Ç , the
unbiased, i.e.
never
for the proposed scheme decreases gradually with an
increase in the value of @ but for the large values of Ç , the
scheme decreases rapidly with an increase in the value of @.
for the proposed
Note that if we apply the same set up with samples sizes Z % + instead of Z +, the results
will be the same with the obvious adjustments in the control limits in (4.2) and (4.3).
DĞŵŽƌLJͲƚLJƉĞŽŶƚƌŽůŚĂƌƚƐŝŶ^ƚĂƚŝƐƚŝĐĂůWƌŽĐĞƐƐŽŶƚƌŽů
ϴϭ
ϴϮ
ͶǤʹǤͳ
In this subsection we provide a broad comparison of our proposed MXEWMA chart
with the classical CUSUM, the classical EWMA and some of their recent modifications.
Below, we present the one by one comparison of the proposed scheme with its counterparts:
MXEWMA versus classical EWMA:
8 for the classical EWMA chart with time varying
are given in Table 2.2. Comparing the MXEWMA chart (with *q >MB) with the classical
EWMA chart we can see that for all the values of Ç the
performance of the
MXEWMA chart is better than the classical EWMA for a fixed value of @ (cf. Table 2.2 vs.
Tables 4.1 – 4.5). Moreover, an important point here is that the classical EWMA is a special
case of the MXEWMA chart, i.e. applying MXEWMA chart to a process where Ç is
equivalent to applying the classical EWMA. From Table 4.1 we see that for Ç >B the
results almost coincides with the results of Table 2.2 as was to be expected.
It is to be noted here that the results of proposed MXEWMA chart with time varying
limits are on the same pattern as compared to the classical EWMA with time varying limits
while the proposed chart with asymptotic limits (computational results are not provided here)
is mainly following the pattern of classical EWMA with asymptotic limits.
MXEWMA versus classical CUSUM: The classical CUSUM proposed by Page (1954) is
discussed briefly in Section 1.2. A comprehensive study on the CUSUM charts is given by
Hawkins and Olwell (1998). The
fixed at B.
8 for the CUSUM chart are given in Table 4.6 with
Comparing the MXEWMA chart with the classical CUSUM chart we observe that the
newly proposed chart is outperforming the classical CUSUM even for small values of Ç
(cf. Table 4.6 vs. Tables 4.1 – 4.5).
DĞŵŽƌLJͲƚLJƉĞŽŶƚƌŽůŚĂƌƚƐŝŶ^ƚĂƚŝƐƚŝĐĂůWƌŽĐĞƐƐŽŶƚƌŽů
ƵdžŝůŝĂƌLJŝŶĨŽƌŵĂƚŝŽŶďĂƐĞĚh^hDĂŶĚtDĐŽŶƚƌŽůĐŚĂƌƚƐ
TABLE 4.6: 9:; values for the classical CUSUM scheme
&
(
0.25
@
0
0.25
0.5
0.75
1
1.5
2
2.5
3
4
5
8.585
500
94.8
31.08
17.54
12.17
7.58
5.55
4.41
3.69
2.84
2.26
0.5
5.071
500
145.5
38.87
17.32
10.52
5.82
4.06
3.15
2.6
2.03
1.72
0.75
3.539
500
200.7
57.07
22.13
11.6
5.45
3.55
2.67
2.18
1.63
1.24
1
2.665
500
249.5
81.44
30.9
14.67
5.75
3.41
2.45
1.94
1.38
1.09
MXEWMA versus runs rules based CUSUM and EWMA: The
8 for the runs rules
based CUSUM are provided in Tables 2.7 – 2.8 and for runs rules based EWMA these are
given in Tables 2.15 – 2.16. Comparing the MXEWMA chart with runs rules based CUSUM
and EWMA we can see that the proposed chart is uniformly surpassing both the CUSUM
schemes I & II and the EWMA simple MOM scheme. The EWMA modified MOP scheme is
performing better than the proposed chart as long as Ç , >B, but once we have Ç Û >B,
the proposed chart outperforms the EWMA modified MOP scheme as well (cf. Tables 2.7 –
2.8 & 2.15 – 2.16 vs. Tables 4.1 – 4.5).
MXEWMA versus MEWMA: Lowry et al. (1992) introduced a multivariate extension of the
EWMA chart named as MEWMA chart. For the bivariate case, the MEWMA statistic is
)+
)+
¹ º K ¹ º ! + K ¹ º
)M
)M
and the chart gives an out of control signal if )+
(Ü . Here
Ý
© 5Þ© 6
Ý5Ý©§ 6]Ω ]Ï
ÎÏ
)M Ê
'
'
' )+
˹ º %
)M
'
. Lowry et al. (1992) assumed, that the mean of both the
variables are equal (i.e. ). According to this assumption the
MEWMA only depend upon the shift parameter @ /
ÞÎÏ
values of
. We have evaluated the
values of this bivariate EWMA chart through Monte Carlo simulation by running +´
DĞŵŽƌLJͲƚLJƉĞŽŶƚƌŽůŚĂƌƚƐŝŶ^ƚĂƚŝƐƚŝĐĂůWƌŽĐĞƐƐŽŶƚƌŽů
ϴϯ
ϴϰ
replications. To validate our simulation code, we have replicated Table 1 of Lowry et al.
(1992) article and found the same results.
Now using our simulation code, we find the
values of the bivariate EWMA with
the mean of fixed and defining the shift in the mean of as we have defined in Tables 4.1
8 are given in Table 4.7 where K >+ and (Ü +>fPP with
– 4.5. These
B.
TABLE 4.7: 9:; values for MEWMA charts with ß => E and lH E=> GII at 9:;= ?==
@
0
Ç >B
500.57
Ç >MB
501.6147
Ç >B
501.0223
Ç >dB
498.4386
Ç >eB
0.5
36.81699
34.50276
27.63295
16.425
4.41316
1
9.8781
9.31899
7.66392
4.85723
1.55707
1.5
4.91344
4.65957
3.88276
2.54956
1.05029
2
3.08706
2.94745
2.47956
1.6849
1.00066
2.5
2.19454
2.10018
1.79616
1.27854
1.00001
3
1.6936
1.6269
1.41069
1.08457
1
500.4125
Comparing the performance of the bivariate EWMA with the proposed chart, it can be
noticed that the proposed chart has smaller
values as compared to the bivariate EWMA
chart for all the corresponding values of Ç (cf. Table 4.7 vs. Tables 4.1 – 4.5).
ͶǤʹǤʹ
Let us consider three variables, named as , and R, following a trivariate normal
distribution ([à ) out of which is the study variable, while and R are two auxiliary
variables, i.e.
'
É Ì È[à áÉ Ì ¿ '
â
R
'â
'
'
'â
'â
'â Áã
'â
DĞŵŽƌLJͲƚLJƉĞŽŶƚƌŽůŚĂƌƚƐŝŶ^ƚĂƚŝƐƚŝĐĂůWƌŽĐĞƐƐŽŶƚƌŽů
(4.7)
ƵdžŝůŝĂƌLJŝŶĨŽƌŵĂƚŝŽŶďĂƐĞĚh^hDĂŶĚtDĐŽŶƚƌŽůĐŚĂƌƚƐ
Based on the above trivariate normal distribution, Kadilar and Cingi (2005) provided a
regression estimator for estimating the population mean of the study variable which are
given as:
7£§ ! Í ! Íâ â R
where Í
]ÎÏ
©
]Ï
andÍâ
]Îä
©
]ä
(4.8)
. The mean and variance of 7£ is given as:
Ð7£ , 7£ '£ ' + Ç
Çâ
! MÇ Çâ Çâ
(4.9)
where Ç is the Pearson’s correlation coefficient between and , and similarly Çâ and
Çâ are defined. Now the plotting statistic and the control limits of the EWMA chart based
on two auxiliary variables is given as:
å
)å *å 7£§ ! 5+ *å 6)
.
å
'Ñ /
0å
0å
(4.10)
+ + *å 4
2
3
2
0å
å
å
/
+ + *
! 'Ñ
1
0å
(4.11)
å
where *å is the smoothing constant for the proposed statistic. )
represents the past
information and its initial value (i.e. )å ) is also taken equal to the target mean i.e. the in
control mean of .
å
determines the width of the control limits for the proposed EWMA
chart based on two auxiliary variables.
Finally, after a simulation study like we have done with one auxiliary variable, the key
conclusions about the EWMA chart based on two auxiliary variables are:
i.
values for the proposed chart decrease with an increase in values of either or
both Ç and Çâ but they increase as the value of Çâ increases;
DĞŵŽƌLJͲƚLJƉĞŽŶƚƌŽůŚĂƌƚƐŝŶ^ƚĂƚŝƐƚŝĐĂůWƌŽĐĞƐƐŽŶƚƌŽů
ϴϱ
ϴϲ
ii.
iii.
iv.
v.
the combinations of Ç , Çâ and Çâ (e.g. Ç Çâ Çâ >B>dB and
Ç Çâ Çâ >dB>B) for which the values of plotting statistic 7£ and its
variance '£ are the same, yield the same
for fixed values of *å and
,
8;
the value of
combination of Ç , Çâ and Çâ ;
å
remains the same for any possible
the performance of the proposed chart is usually better for the smaller values of *å ;
auxiliary information really improves the performance (which is not a surprise!).
ͶǤʹǤ͵
In this subsection, we give an illustrative example for which we generate a dataset
(named as dataset 4.1) containing 30 observations in total. The first 20 observations are
generated from a trivariate normal distribution given as:
+
+ >B >B
É Ì È[à ¿É B Ì É>B +
ÌÁ
R
B
>B
+
Hence the in control mean of the study variable equals +. The remaining 10 observations
are generated from an out of control situation with a shift of one sigma in the mean of the
study variable (i.e. @ +) given as:
++
+ >B >B
É Ì È[à ¿É B Ì É>B +
ÌÁ
R
B
>B
+
Now the variable is used to setup the classical EWMA control chart (with plotting statistic
represented by ) and parameters * >MB and
P), variables and are used to setup
the MXEWMA control chart (with plotting statistic represented by )q and parameters
*q >MB and
q
P) and variables , and R are used to setup the MTEWMA control
chart (with plotting statistic represented by )å and parameters *å >MB and
DĞŵŽƌLJͲƚLJƉĞŽŶƚƌŽůŚĂƌƚƐŝŶ^ƚĂƚŝƐƚŝĐĂůWƌŽĐĞƐƐŽŶƚƌŽů
å
P). The
ƵdžŝůŝĂƌLJŝŶĨŽƌŵĂƚŝŽŶďĂƐĞĚh^hDĂŶĚtDĐŽŶƚƌŽůĐŚĂƌƚƐ
calculation steps for applying the MXEWMA chart are given in Table 4.7, whereas the
graphical display of the three charts is given in Figure 4.1.
Table 4.8: Simulated dataset 4.1 and calculation steps of MXEWMA chart
Sample
No.
1
10.07
2
11.21
3
8.32
4
9.63
5
7.4
6
R
5.22
7
)q
4.34
9.97
9.99
7.15
4.15
10.13
10.03
4.91
4.46
8.37
9.61
2.95
4.32
10.65
9.87
4.08
4.26
7.86
9.37
10.56
4.16
5.86
10.98
7
9.52
4.55
4.69
8
8.85
5.23
4.02
9
9.58
5.45
10
10.98
11
10.11
12
Sample
No.
10.65
16
11.06
10.81
17
10.15
10.89
18
9.12
10.93
19
10.5
10.95
20
8.67
9.77
10.97
21
9.74
9.76
10.97
8.74
9.51
10.98
4.62
9.36
9.47
5.04
4.48
10.96
5.52
5.68
9.85
9.6
4.06
4.76
13
10.7
5.25
14
8.35
4.33
15
11.09
5.88
5.86
R
7
)q
6.67
10.63
10.12
10.98
5.48
3.77
9.92
10.07
10.98
5.18
4.58
9.03
9.81
10.98
4.97
6.67
10.51
9.99
10.98
5.5
3.55
8.42
9.59
10.98
10.41
5.29
4.05
10.26
9.76
10.98
22
10.11
3.33
3.91
10.95
10.06
10.98
23
12.59
5.2
7.1
12.49
10.67
10.98
10.98
24
11.59
4.98
4.65
11.6
10.9
10.98
9.84
10.98
25
11.24
6.42
4.27
10.53
10.81
10.98
9.84
10.98
26
10.65
4.81
4.66
10.74
10.79
10.98
10.07
9.9
10.98
27
11.09
5.29
3.56
10.94
10.83
10.98
6.63
10.57
10.07
10.98
28
10.48
5.25
4.36
10.36
10.71
10.98
4.24
8.68
9.72
10.98
29
11.92
5.04
4.14
11.9
11.01
10.98
5.02
10.65
9.95
10.98
30
11.17
4.32
5.53
11.51
11.13
10.98
Figure 4.1: Graphical display of MTEWMA, MXEWMA and classical EWMA charts for the
simulated dataset 4.1
DdtD
DdžtD
>ŝŵŝƚƐ;DdtDͿ
>ŝŵŝƚƐ;DdžtDͿ
ϭϭ͘Ϯ
ϭϬ͘ϴ
ϭϬ͘ϰ
ϭϬ
ϵ͘ϲ
ϵ͘Ϯ
ϴ͘ϴ
ϭ
ϯ
ϱ
ϳ
ϵ
ϭϭ
ϭϯ
ϭϱ
ϭϳ
ϭϵ
Ϯϭ
Ϯϯ
Ϯϱ
Ϯϳ
Ϯϵ
^ĂŵƉůĞEƵŵďĞƌ
DĞŵŽƌLJͲƚLJƉĞŽŶƚƌŽůŚĂƌƚƐŝŶ^ƚĂƚŝƐƚŝĐĂůWƌŽĐĞƐƐŽŶƚƌŽů
ϴϳ
ϴϴ
Figure 4.1 indicates that the classical EWMA control chart failed to detect any shift in the
mean of the study variable . MXEWMA detected the shift at samples # 29 and 30 (cf. Table
4.8), whereas MTEWMA detected the shift at samples # 24, 25, 26, 27, 28, 29 and 30.
ͶǤ͵
In this section we utilize the efficiency of regression estimator to design a CUSUMtype structure and try to study the effect of this efficient estimator on the
performance of
CUSUM chart. Now the plotting statistics for the proposed MXCUSUM chart (which is based
on the estimator given in (4.2)) is given as:
[ 57§ 6
q
! [
,
[ 57§ 6
q
! [
(4.12)
Initial values for the statistics given in (4.12) are taken equal to zero i.e. [ [ . The
decision rule for the proposed chart is given as: the statistics [ and [ are plotted against
the control limit $q . For any value of #, if the value of [ exceeds the value of $q than the
process mean is declared to be shifted upwards and if the value of [ exceeds the value of $q
than the process mean is said to be moved downwards.
q
& q 'Ñ ,
$q (q 'Ñ
q
and $q are defined as:
(4.13)
The values of & q and (q need to be selected carefully because the
properties of
MXCUSUM chart mainly depend on these two constants (along with the value of Ç ). The
specifications of MXCUSUM chart makes it the generalized form of the classical CUSUM
chart given in Section 1.2, i.e. for Ç , the proposed MXCUSUM chart becomes
equivalent to the classical CUSUM in terms of plotting statistic as well
For the proposed chart Tables 4.9 – 4.11 contain the
8 with
DĞŵŽƌLJͲƚLJƉĞŽŶƚƌŽůŚĂƌƚƐŝŶ^ƚĂƚŝƐƚŝĐĂůWƌŽĐĞƐƐŽŶƚƌŽů
performance.
fixed at B.
ƵdžŝůŝĂƌLJŝŶĨŽƌŵĂƚŝŽŶďĂƐĞĚh^hDĂŶĚtDĐŽŶƚƌŽůĐŚĂƌƚƐ
Table 4.9: 9:; values for the proposed MXCUSUM chart with <q => F? and lq G> ?G? at
9:;= ?==
0.05
0.25
Ç
0.5
0.75
0.95
0
499.3466
499.1864
499.2012
500.9817
500.6632
0.25
94.6109
89.987
74.984
48.1801
16.1035
0.5
31.0236
29.6155
25.1541
17.3604
7.0516
0.75
17.5129
16.8127
14.5609
10.4637
4.5919
1
12.1546
11.7041
10.2408
7.5157
3.4645
1.5
7.5715
7.3144
6.4712
4.8689
2.3575
2
5.5412
5.3636
4.7791
3.6584
1.9967
2.5
4.4047
4.2701
3.8258
2.9861
1.7963
3
3.6817
3.5743
3.2238
2.5107
1.2199
@
Table 4.10: 9:; values for the proposed MXCUSUM chart with <q => ? and lq ?> =JE at
9:;= ?==
0.05
0.25
Ç
0.5
0.75
0.95
0
499.015
500.0021
500.8843
499.0574
500.1008
0.25
145.2716
138.2263
114.5163
68.9901
15.3663
0.5
38.777
36.3617
28.9044
17.0683
5.3405
0.75
17.2793
16.3181
13.3748
8.6464
3.2904
1
10.4965
9.9897
8.4132
5.7578
2.4388
1.5
5.8086
5.5756
4.8327
3.5082
1.7847
2
4.0503
3.9052
3.4373
2.5806
1.2021
2.5
3.1451
3.0414
2.7049
2.1126
1.0074
3
2.5978
2.5187
2.2684
1.8667
1
@
The algorithm used for calculating the
where 50,000 simulations are used.
8 in Tables 4.9 – 4.11 is given in Appendix 4.2
DĞŵŽƌLJͲƚLJƉĞŽŶƚƌŽůŚĂƌƚƐŝŶ^ƚĂƚŝƐƚŝĐĂůWƌŽĐĞƐƐŽŶƚƌŽů
ϴϵ
ϵϬ
Table 4.11: 9:; values for the proposed MXCUSUM chart with <q E and lq F> gg? at
9:;= ?==
0.05
0.25
Ç
0.5
0.75
0.95
0
499.0466
499.1701
500.5972
499.4241
500.9488
0.25
249.1542
240.639
209.9628
139.8514
26.0941
0.5
81.2212
75.9419
58.9892
30.2808
5.0512
0.75
30.8019
28.4313
21.2702
10.7289
2.5918
1
14.6235
13.5187
10.2682
5.6554
1.795
1.5
5.7314
5.3872
4.3635
2.814
1.1276
2
3.4063
3.2422
2.7406
1.9231
1.0031
2.5
2.4483
2.3484
2.0355
1.4745
1
3
1.9387
1.8674
1.6356
1.1929
1
@
Before concluding this section, we present the main findings about our proposed MXCUSUM
control chart.
i.
the use of an auxiliary variable with the control structure of CUSUM chart is really
advantageous in terms of the
Tables 4.9 – 4.11);
ii.
iii.
iv.
values if the value of Ç is reasonably large (cf.
to attain a fixed value of
,
the value of (q has to remain fixed for all the value of
to attain a fixed value of
,
the value of (q decrease with an increase in the value
Ç (cf. Tables 4.9 – 4.11);
of & q and vice versa (cf. Tables 4.9 – 4.11);
for a fixed value of
,
the
values decrease rapidly with a decrease in the
values of either or both Ç and @ (cf. Tables 4.9 – 4.11).
Like MXCUSUM, a CUSUM chart based on the information of two auxiliary variables
(named as MTCUSUM) can also be easily designed by following the procedure given in
subsection 4.2.2.
DĞŵŽƌLJͲƚLJƉĞŽŶƚƌŽůŚĂƌƚƐŝŶ^ƚĂƚŝƐƚŝĐĂůWƌŽĐĞƐƐŽŶƚƌŽů
ƵdžŝůŝĂƌLJŝŶĨŽƌŵĂƚŝŽŶďĂƐĞĚh^hDĂŶĚtDĐŽŶƚƌŽůĐŚĂƌƚƐ
ͶǤ͵Ǥͳ
Now we provide a comparison of MXCUSUM chart with some of the revisions and
extensions of CUSUM and EWMA-type control charts.
MXCUSUM versus the classical CUSUM:
values of & with
8 for two-sided CUSUM chart for different
fixed at B are given in Table 4.6. Collectively Tables 4.9 – 4.11 and
Table 4.6 validate that the classical CUSUM is a special case of the proposed MXCUSUM
chart. The
performance of the proposed chart with Ç >B (i.e. close to zero) is
almost same as compared to the
value of Ç for the proposed chart, its
performance of classical CUSUM. As we increase the
values decrease.
MXCUSUM versus the classical EWMA: Table 2.2 contains the
where * and
8 of classical EWMA
are the parameters of EWMA control chart. The comparison of MXCUSUM
chart with the classical EWMA show that the proposed chart is performing better in general
for Ç Û >MB (cf. Table 2.2 vs. Tables 4.9 – 4.11). For large values of Ç (like Ç Û
>dB) the proposed chart (with any value of & q ) outperfoms the classical EWMA (with any
value of *).
MXCUSUM versus MXEWMA: Comparing the performance of MXCUSUM chart with
MXEWMA chart for a specific value of Ç (e.g. Ç >B) it can be observed that
MXEWMA (with *q >+) has a slightly better
performance as compared to the
proposed chart but once *q is greater than >+ for the MXEWMA chart, the proposed chart
becomes better than MXEWMA (cf. Table 4.1 – 4.5 vs. Tables 4.9 – 4.11).
The MTCUSUM control chart, which is based on two auxiliary variables, has even better
properties as was to be expected (results not given here).
DĞŵŽƌLJͲƚLJƉĞŽŶƚƌŽůŚĂƌƚƐŝŶ^ƚĂƚŝƐƚŝĐĂůWƌŽĐĞƐƐŽŶƚƌŽů
ϵϭ
ϵϮ
ͶǤ͵Ǥʹ
An example to illustrate the application of the proposed chart with a real dataset is
provided in this subsection. The three CUSUM-type control charts are applied to the
simulated dataset 4.1 with & >B and ( B>d+ for the classical CUSUM; & q >B and
(q B>d+ for the MXCUSUM; and & å >B and (å B>d+ for MTCUSUM chart with
4.2.
fixed at B for all three charts. The chart output for all three charts is given in Figure
Figure 4.2: Graphical display of MTCUSUM, MXCUSUM and classical CUSUM charts for
simulated dataset 4.1
Ddh^hD
,;Ddžh^hDͿ
ϭϬ
,;Ddh^hDͿ
h^hD
Ddžh^hD
,;h^hDͿ
ϴ
ϲ
ϰ
Ϯ
Ϭ
ϭ
ϯ
ϱ
ϳ
ϵ
ϭϭ
ϭϯ
ϭϱ
ϭϳ
ϭϵ
Ϯϭ
Ϯϯ
Ϯϱ
Ϯϳ
Ϯϵ
^ĂŵƉůĞEƵŵďĞƌ
Figure 4.2 shows that MTCUSUM chart is detecting the positive shift at the 24th sample and
onwards, whereas the classical CUSUM and MXCUSUM are detecting the shift at the 27th
and the 29th sample, respectively. This verifies the observations in subsection 4.3.1.
DĞŵŽƌLJͲƚLJƉĞŽŶƚƌŽůŚĂƌƚƐŝŶ^ƚĂƚŝƐƚŝĐĂůWƌŽĐĞƐƐŽŶƚƌŽů
ƵdžŝůŝĂƌLJŝŶĨŽƌŵĂƚŝŽŶďĂƐĞĚh^hDĂŶĚtDĐŽŶƚƌŽůĐŚĂƌƚƐ
ͶǤͶ
This chapter proposes the use of auxiliary information with the control structures of
the EWMA and CUSUM charts. The regression estimation technique is used to exploit the
auxiliary information. Note that the proposed MXEWMA and MXCUSUM charts are the
extended forms of the classical EWMA and CUSUM charts respectively, i.e. the proposed
charts are equal to the classical ones when the correlation between the study variable and
auxiliary variable(s) is equal to . The performance of the proposed charts is evaluated in
terms of
for different values of the correlation between the study variable and the
auxiliary variable(s). A comparison of the proposed charts with the classical CUSUM, the
classical EWMA and some of their recent modifications is also made. The comparisons
showed that the proposed charts are good at detecting small to moderate shifts in the process
location, while their ability to detect large shifts is not bad either. Finally, results are also
supported by the illustrative examples.
DĞŵŽƌLJͲƚLJƉĞŽŶƚƌŽůŚĂƌƚƐŝŶ^ƚĂƚŝƐƚŝĐĂůWƌŽĐĞƐƐŽŶƚƌŽů
ϵϯ
ϵϰ
ͶǤͳ
library(MASS)
Mx=c(); Zp=c(); ucl=c(); lcl=c(); rl=c()
muy=0; mux=0; sigyy=1; sigxx=1; sigyx=0.5
betayx=sigyx/sigxx; rho=sigyx/sqrt(sigyy*sigxx)
meanv=c(muy,mux)
sigmam=matrix(c(sigyy,sigyx,sigyx,sigxx),2,2)
muM=muy; sigM=sqrt(sigyy*(1-rho^2))
ld=0.25; L=3
for(j in 1:10000)
{
for(i in 1:1000000)
{
w=mvrnorm(1,meanv,sigmam)
y=w[1]; x=w[2]
Mx[i]=y-betayx*x
if(i==1)
{Zp[i]=ld*Mx[i]+ (1-ld)*muM;}
else{Zp[i]=ld*Mx[i]+(1-ld)*Zp[i-1];}
ucl[i]=muM+L*sigM*sqrt((ld/(2-ld))*(1-(1-ld)^(2*i)))
lcl[i]=muM-L*sigM*sqrt((ld/(2-ld))*(1-(1-ld)^(2*i)))
if(Zp[i]>ucl[i]|Zp[i]<lcl[i])
{rl[j]=i;break;}
else{rl[j]=0;}
}
}
mean(rl)
ͶǤʹ
library(MASS)
Mx=c(); Np=c(); Nn=c(); rl=c()
muy=0; mux=0; sigyy=1; sigxx=1; sigyx=0.5;
betayx=sigyx/sigxx; rho=sigyx/sqrt(sigyy*sigxx)
meanv=c(muy,mux)
sigmam=matrix(c(sigyy,sigyx,sigyx,sigxx),2,2)
muM=muy; sigM=sqrt(sigyy*(1-rho^2))
K=0.5*sigM; H=5.071*sigM
for(j in 1:100000)
{
for(i in 1:1000000)
{
w=mvrnorm(1,meanv,sigmam)
DĞŵŽƌLJͲƚLJƉĞŽŶƚƌŽůŚĂƌƚƐŝŶ^ƚĂƚŝƐƚŝĐĂůWƌŽĐĞƐƐŽŶƚƌŽů
ƵdžŝůŝĂƌLJŝŶĨŽƌŵĂƚŝŽŶďĂƐĞĚh^hDĂŶĚtDĐŽŶƚƌŽůĐŚĂƌƚƐ
y=w[1]; x=w[2]
Mx[i]=y-betayx*x
if(i==1)
{Np[i]=max(0,(Mx[i]-muM)-K);}
else{Np[i]=max(0,(Mx[i]-muM)-K+Np[i-1]);}
if(i==1)
{Nn[i]=max(0,-(Mx[i]-muM)-K);}
else{Nn[i]=max(0,-(Mx[i]-muM)-K+Nn[i-1]);}
if(Np[i]>H|Nn[i]>H)
{rl[j]=i;break;}
else{rl[j]=0;}
}
}
mean(rl)
DĞŵŽƌLJͲƚLJƉĞŽŶƚƌŽůŚĂƌƚƐŝŶ^ƚĂƚŝƐƚŝĐĂůWƌŽĐĞƐƐŽŶƚƌŽů
ϵϱ
ϵϲ
DĞŵŽƌLJͲƚLJƉĞŽŶƚƌŽůŚĂƌƚƐŝŶ^ƚĂƚŝƐƚŝĐĂůWƌŽĐĞƐƐŽŶƚƌŽů
ŚĂƉƚĞƌϱ
ͷ
This chapter proposes a control chart for monitoring the process location, named as
Progressive Mean (7) control chart, in which a progressive mean is used as a plotting
statistic. Taking the inspiration from the 7 chart, two new memory control charts for
monitoring the process dispersion, named as floating and floating control
charts, are also proposed. The proposed charts are designed such that they utilize not only the
current information but also the past information. Therefore, the proposed charts are natural
competitors for the classical CUSUM, the classical EWMA and some recent modifications of
these two charts.
This chapter is based on two papers; one for monitoring the location parameter (cf. Abbas,
Zafar, Riaz and Does (2012)) and the other for monitoring the dispersion parameter (cf.
Abbas, Riaz and Does (2012d)).
ͷǤͳ
Let be a quality characteristic of which the mean will be monitored. It is assumed
that we use individual observations from a normal distribution. If , # + M P x x> is the
sequence of independent and identically distributed observations from the process under
investigation, then progressive mean 7 is defined as the cumulative average over time.
Mathematically, we may define the 7 statistic as:
DĞŵŽƌLJͲƚLJƉĞŽŶƚƌŽůŚĂƌƚƐŝŶ^ƚĂƚŝƐƚŝĐĂůWƌŽĐĞƐƐŽŶƚƌŽů
ϵϳ
ϵϴ
7
¦§¨«¬ ¨
(5.1)
7 is an unbiased estimator of population mean and its variance is given by
]^©
.
According to the typical three sigma control limits, the control charting structure based on the
7 statistic and its variance may be defined as:
P ^ h
]
_
h
! P
]^
_
(5.2)
where all the parameters used in (5.2) are as explained earlier. It is obvious that the control
limits given in (5.2) are time varying and exploit the past and present information using equal
weights. Note that the design structure of the proposed chart is relative simple and easy to
execute compared with the CUSUM and EWMA control charts.
A problem with the control structure in (5.2) is that the control limits remain too wide for the
large values of # (wide relative to the plotting statistic). It turns out that it is almost impossible
for the plotting statistic in (5.1) to cross the control limits in (5.2), in case of shifted mean.
We have solved this issue by imposing a penalty on the control limits such that the control
limits are a bit narrower for large values of #. We choose this penalty function equal to
# # æ . Hence, the penalized limits for the proposed 7 chart are given by:
q
]^
^>çèé
hh
h
! q
where q is a constant that is used to control the run lengths.
]^
^>çèé
(5.3)
In (5.3) we have used different possibilities of ê and we have searched for a suitable
constant q for each possibility to fix the in control process properties in terms of
optimum out of control
properties in terms of
suitable choice in terms of optimizing the
.
and
We identified ê >M as the most
properties. For this optimum choice of # (i.e.
# > ) we have worked out the values of constants q by fixing
DĞŵŽƌLJͲƚLJƉĞŽŶƚƌŽůŚĂƌƚƐŝŶ^ƚĂƚŝƐƚŝĐĂůWƌŽĐĞƐƐŽŶƚƌŽů
.
These constants for some
WƌŽŐƌĞƐƐŝǀĞĐŽŶƚƌŽůĐŚĂƌƚŝŶŐ
commonly used
values are provided in Table 5.1. For these constants and
ϵϵ
values
given in Table 5.1 we have carried out a RL study and the resulting properties in terms of
and
are provided in Table 5.2. Appendix 5.1 contains the details regarding the
8. The standard errors
algorithm, that is made in R language, used for the computation of
for the results of Tables 5.1 – 5.2 remain less than 1.3%.
TABLE 5.1: Values of the control limit constant nq for different choices of 9:;= for the
proposed Äë control chart
q
168
200
3
3.129
400
500
3.568
3.639
3.846
TABLE 5.2: 9:; values for the proposed chart with different shifts
Pre-fixed
370
0
0.25
0.5
0.75
1
500
498.14
47.2353
19.032
11.1564
400
400.938
44.474
17.872
370
369.00
17.452
200
200.82
168
170.33
@
1.5
2
3
4
5
7.5504
4.5204
3.1455
1.9803
1.4248
1.1186
10.313
7.200
4.261
2.985
1.877
1.360
1.0847
10.166
10.166
7.0941
4.184
2.931
1.721
1,236
1.043
34.7366
14.6914
8.6191
5.9814
3.5996
2.5415
1.6124
1.1981
1.0338
33.0695
12.1479
8.1779
5.659
3.4086
2.4437
1.5555
1.1598
1.0263
We conclude from Tables 5.1 – 5.2:
i.
The proposed 7 control chart is really good in detecting small and moderate shifts
and is still good in detecting large shifts (cf. Table 5.2);
ii.
iii.
decreases fastly with an increase in @ (cf. Tables 5.1 and 5.2);
The control structure of the proposed chart is very easy compared to the CUSUM and
EWMA charts.
Note that if we apply the same set up with samples sizes Z % + instead of Z +, the results
will be the same with the obvious adjustments in the control limits in (5.3).
DĞŵŽƌLJͲƚLJƉĞŽŶƚƌŽůŚĂƌƚƐŝŶ^ƚĂƚŝƐƚŝĐĂůWƌŽĐĞƐƐŽŶƚƌŽů
ϭϬϬ
ͷǤͳǤͳ
To detect small and moderate shifts, EWMA and CUSUM charts and some of their
modifications are available. We have introduced in this chapter a rather simple alternative to
these charts, namely the 7 control chart, and in this section we compare the performance of
our proposed chart with some of its counterparts in terms of
. We compare the
performance of the proposed chart with the classical EWMA, the classical CUSUM, the fast
initial response (FIR) CUSUM, the FIR EWMA, the runs rules based CUSUM, the runs rules
based EWMA and the adaptive EWMA. We use for the
the values +Xf, M, Pd, A
and B, so that valid comparisons with each chart can be made. Below, we present one by
one the comparison of the proposed chart with its competitor.
Proposed versus the classical CUSUM:
values for the classical CUSUM are given by
in Table 2.1. Comparison of the classical CUSUM with the proposed 7 chart clearly shows
that the proposed chart almost outperforms the classical CUSUM for all the values of @ (cf.
Table 2.1 vs. Table 5.2).
Proposed versus the classical EWMA: Table 2.2 contains the
values for the classical
EWMA chart. Comparison of Table 2.2 with Table 5.2 shows the uniform superiority of the
proposed 7 chart over the classical EWMA chart.
Proposed versus FIR CUSUM and FIR EWMA:
values for the FIR CUSUM
presented by Lucas and Crosier (1982) and FIR EWMA by Lucas and Saccucci (1990) are
given in Table 2.10 and Table 2.17, respectively. The comparison of FIR CUSUM and the
proposed 7 chart indicates that the performance of proposed control chart is better for small
and moderate shifts even when the
for the FIR CUSUM is not fixed at +Xf (cf. Table
2.10 vs. Table 5.2). We may conclude that the proposed chart performs really well for small
DĞŵŽƌLJͲƚLJƉĞŽŶƚƌŽůŚĂƌƚƐŝŶ^ƚĂƚŝƐƚŝĐĂůWƌŽĐĞƐƐŽŶƚƌŽů
WƌŽŐƌĞƐƐŝǀĞĐŽŶƚƌŽůĐŚĂƌƚŝŶŐ
shifts. The same is the case for the FIR EWMA, i.e. for a 25% head start the performance of
the proposed chart is superior to the FIR EWMA for every choice of *, but as we increase the
head start to 50%, the performance of the FIR EWMA becomes better than the proposed
chart for the larger shifts with * >+ (cf. Table 2.17 vs. Table 5.2).
Proposed versus runs rules based CUSUM and EWMA: In chapter 2 we introduced the
8 for the
runs rules based CUSUM and the runs rules based EWMA, respectively. The
runs rules based CUSUM are provided in Tables 2.7 – 2.8 and for runs rules based EWMA
these are given in Tables 2.15 – 2.16. The comparison of the runs rules based CUSUM with
the proposed chart shows that the proposed chart performs better than the runs rules based
CUSUM uniformly for both schemes (cf. Tables 2.7 – 2.8 vs. Table 5.2). Similarly, the
comparison of the runs rules based EWMA with the proposed chart also reveals that the
proposed chart is superior to the runs rules based EWMA for most of the values of @ (cf.
Tables 2.15 – 2.16 vs. Table 5.2).
Proposed versus adaptive EWMA: The adaptive EWMA of Capizzi and Masarotto (2003)
is designed so that it performs better for small and large shifts at the same time by giving
weights to past information using a suitable function of the current error. Three functions of
error represented by ìVí > , ìîo > and ìïíî > are used in their article.
values for the
adaptive EWMA with these 3 functions of errors are given in Table 5.3, where the value of @
is targeted between >MB A.
TABLE 5.3: 9:; values for the adaptive EWMA at 9:;= ?==
Error function
ìVí >
ìîo >
ìïíî >
0.25
0.5
0.75
1
98.51
40.94
25.04
135.01
42.72
97.03
41.54
@
1.5
2
3
4
17.59
10.11
6.08
2.29
1.26
21.99
13.91
7.12
4.25
2.01
1.28
25.68
18.16
10.52
6.36
2.35
1.27
DĞŵŽƌLJͲƚLJƉĞŽŶƚƌŽůŚĂƌƚƐŝŶ^ƚĂƚŝƐƚŝĐĂůWƌŽĐĞƐƐŽŶƚƌŽů
ϭϬϭ
ϭϬϮ
The comparison of the proposed chart with the adaptive EWMA shows that the proposed
chart also outperforms the adaptive EWMA for all the values of @ (cf. Table 5.3 vs. Table
5.2).
Proposed versus adaptive CUSUM: Jiang et al. (2008) proposed the use of an adaptive
CUSUM with EWMA based shift estimator.
8 for the adaptive CUSUM with @|`
>B
and * >P are given in Table 3.5. Comparison of Table 5.2 and Table 3.5 shows that the
proposed 7 chart is uniformly superior compared to the adaptive CUSUM with EWMA
based shift estimator.
curves of the proposed 7 chart against
To summarize the results, we have made some
its existing counterparts. These are given in Figures 5.1 – 5.3.
curves for the proposed
chart, the classical EWMA (with * >+), the adaptive EWMA (with error function ìVí > ),
the runs rules based CUSUM scheme II (with R A>f and
B>++) and the runs rules
based EWMA modified MOP scheme (with * >+) are presented in Figure 5.1.
FIGURE 5.1: 9:; Curves for proposed chart, classical EWMA, adaptive EWMA, runs rules
based CUSUM and runs rules based EWMA at 9:;= ?==
WƌŽƉŽƐĞĚ
tD
Ϭ͘Ϯϱ
Ϭ͘ϱ
ĚĂƉƚŝǀĞtD
ZZh^hD//
ZZtD//
ϭϱϬ
ϭϮϬ
Z>
ϵϬ
ϲϬ
ϯϬ
Ϭ
Ϭ͘ϳϱ
ϭ
į
DĞŵŽƌLJͲƚLJƉĞŽŶƚƌŽůŚĂƌƚƐŝŶ^ƚĂƚŝƐƚŝĐĂůWƌŽĐĞƐƐŽŶƚƌŽů
ϭ͘ϱ
Ϯ
WƌŽŐƌĞƐƐŝǀĞĐŽŶƚƌŽůĐŚĂƌƚŝŶŐ
ϭϬϯ
FIGURE 5.2: 9:; curves for proposed chart and adaptive CUSUM at 9:;= H==
WƌŽƉŽƐĞĚ
ĚĂƉƚŝǀĞh^hD;ɶсϰ͕Śсϰ͘ϯϰϴͿ
ĚĂƉƚŝǀĞh^hD;ɶсϮ͕Śсϱ͘ϭϬϱͿ
ϴϬ
Z>
ϲϬ
ϰϬ
ϮϬ
Ϭ
Ϭ͘Ϯϱ
Ϭ͘ϱ
Ϭ͘ϳϱ
ϭ
į
ϭ͘ϱ
Ϯ
ϯ
ϰ
FIGURE 5.3: 9:; curves for proposed chart, classical CUSUM and FIR CUSUM at
9:;= EgG
WƌŽƉŽƐĞĚ
h^hD
&/Zh^hD;ŽсϭͿ
ϴϬ
Z>
ϲϬ
ϰϬ
ϮϬ
Ϭ
Ϭ͘Ϯϱ
Ϭ͘ϱ
Ϭ͘ϳϱ
į
ϭ
ϭ͘ϱ
Ϯ
DĞŵŽƌLJͲƚLJƉĞŽŶƚƌŽůŚĂƌƚƐŝŶ^ƚĂƚŝƐƚŝĐĂůWƌŽĐĞƐƐŽŶƚƌŽů
ϭϬϰ
In Figure 5.2 comparison of the proposed chart and the adaptive CUSUM are provided,
whereas Figure 5.3 contains the
curves for the proposed chart, the classical CUSUM and
the FIR CUSUM (with head start +). Figures 5.1 – 5.3 clearly show the best
performance of the proposed chart against the other methods.
ͷǤͳǤʹ
In this section we demonstrate the application of the proposed 7 chart. Also the
classical EWMA and the classical CUSUM charts are included in the example to validate the
superiority of the proposal. For this purpose we have generated two datasets of 40 and 30
observations, respectively, such that in dataset 5.1 the first 20 observations are generated
from [+ (i.e. the in control situation) and the second set of the 20 observations from
[>B+ (i.e. the out of control situation having a shift of >B' (small shift)). Similarly, in
dataset 5.2 the first 20 observations are generated from [+ and the second set of 10
observations from [+>B+ (i.e. an out of control situation having a shift of +>B' (moderate
shift)). The 7 statistic for the proposed chart, the EWMA statistic with * >MB and the
CUSUM statistic with & >B are calculated. To fix the
+>MXd for the 7 chart,
at B we have used q
P for the classical EWMA and ( B>d for the classical
CUSUM. The graphical displays of the proposed 7, EWMA and CUSUM charts are
presented in the Figures 5.4, 5.5 and 5.6 for dataset 5.1 and in Figures 5.7, 5.8 and 5.9 for
dataset 5.2, respectively.
From Figure 5.4 we can see that the proposed chart gives out of control signals at
samples # 35, 36, 37, 38, 39 and 40, thus giving a total of 6 out of control signals. Figure 5.5
shows that the classical EWMA control chart gives one out of control signal at sample # 38
and Figure 5.6 depicts that the classical CUSUM control chart gives out of control signals at
DĞŵŽƌLJͲƚLJƉĞŽŶƚƌŽůŚĂƌƚƐŝŶ^ƚĂƚŝƐƚŝĐĂůWƌŽĐĞƐƐŽŶƚƌŽů
WƌŽŐƌĞƐƐŝǀĞĐŽŶƚƌŽůĐŚĂƌƚŝŶŐ
samples # 38, 39 and 40, thus giving 3 signals. An upward shift occurred after sample # 20,
which is detected by the proposed chart more quickly than the EWMA and the CUSUM. It
illustrates the ability of the proposed chart to quickly detect small shifts in the process.
FIGURE 5.4: Graphical display of the proposed Äë chart for dataset 5.1
WDƐƚĂƚŝƐƚŝĐƐ
ϰ
h>
>>
ϯ
Ϯ
ϭ
Ϭ
Ͳϭ
ͲϮ
Ͳϯ
Ͳϰ
ϭ
ϯ
ϱ
ϳ
ϵ
ϭϭ
ϭϯ
ϭϱ
ϭϳ
ϭϵ
Ϯϭ
Ϯϯ
Ϯϱ
Ϯϳ
Ϯϵ
ϯϭ
ϯϯ
ϯϱ
ϯϳ
ϯϵ
^ĂŵƉůĞEƵŵďĞƌ
FIGURE 5.5: Graphical display of the classical EWMA chart for dataset 5.1
tDƐƚĂƚŝƐƚŝĐ
ϭ͘ϱ
h>
>>
ϭ
Ϭ͘ϱ
Ϭ
ͲϬ͘ϱ
Ͳϭ
Ͳϭ͘ϱ
ϭ
ϯ
ϱ
ϳ
ϵ
ϭϭ ϭϯ ϭϱ ϭϳ ϭϵ Ϯϭ Ϯϯ Ϯϱ Ϯϳ Ϯϵ ϯϭ ϯϯ ϯϱ ϯϳ ϯϵ
^ĂŵƉůĞEƵŵďĞƌ
DĞŵŽƌLJͲƚLJƉĞŽŶƚƌŽůŚĂƌƚƐŝŶ^ƚĂƚŝƐƚŝĐĂůWƌŽĐĞƐƐŽŶƚƌŽů
ϭϬϱ
ϭϬϲ
FIGURE 5.6: Graphical display of the classical CUSUM for dataset 5.1
h^hDƐƚĂƚŝƐƚŝĐ
Ś
ϲ
ϱ
ϰ
ϯ
Ϯ
ϭ
Ϭ
ϭ
ϯ
ϱ
ϳ
ϵ
ϭϭ
ϭϯ
ϭϱ
ϭϳ
ϭϵ
Ϯϭ
Ϯϯ
Ϯϱ
Ϯϳ
Ϯϵ
ϯϭ
ϯϯ
ϯϱ
ϯϳ
^ĂŵƉůĞEƵŵďĞƌ
FIGURE 5.7: Graphical display of the proposed Äë chart for dataset 5.2
WDƐƚĂƚŝƐƚŝĐ
ϰ
h>
>>
ϯ
Ϯ
ϭ
Ϭ
Ͳϭ
ͲϮ
Ͳϯ
Ͳϰ
ϭ
ϯ
ϱ
ϳ
ϵ
ϭϭ
ϭϯ
ϭϱ
ϭϳ
ϭϵ
Ϯϭ
^ĂŵƉůĞEƵŵďĞƌ
DĞŵŽƌLJͲƚLJƉĞŽŶƚƌŽůŚĂƌƚƐŝŶ^ƚĂƚŝƐƚŝĐĂůWƌŽĐĞƐƐŽŶƚƌŽů
Ϯϯ
Ϯϱ
Ϯϳ
Ϯϵ
ϯϵ
WƌŽŐƌĞƐƐŝǀĞĐŽŶƚƌŽůĐŚĂƌƚŝŶŐ
The situation is not much different in dataset 5.2, where the 7 chart detects the shift at
samples # 23, 24, 25, 26, 27, 28, 29 and 30 (cf. Figure 5.7). The classical EWMA and the
CUSUM detect the shift at samples # 27, 28, 29 and 30 (cf. Figures 5.8 and 5.9).
FIGURE 5.8: Graphical display of the classical EWMA chart for dataset 5.2
Ϯ
tDƐƚĂƚŝƐƚŝĐ
h>
>>
ϭ͘ϱ
ϭ
Ϭ͘ϱ
Ϭ
ͲϬ͘ϱ
Ͳϭ
Ͳϭ͘ϱ
ϭ
ϯ
ϱ
ϳ
ϵ
ϭϭ
ϭϯ
ϭϱ
ϭϳ
ϭϵ
Ϯϭ
Ϯϯ
Ϯϱ
Ϯϳ
Ϯϵ
^ĂŵƉůĞEƵŵďĞƌ
FIGURE 5.9: Graphical display of the classical CUSUM for dataset 5.2
h^hDƐƚĂƚŝƐƚŝĐ
Ś
ϭϮ
ϭϬ
ϴ
ϲ
ϰ
Ϯ
Ϭ
ϭ
ϯ
ϱ
ϳ
ϵ
ϭϭ
ϭϯ
ϭϱ
ϭϳ
ϭϵ
Ϯϭ
Ϯϯ
Ϯϱ
Ϯϳ
Ϯϵ
^ĂŵƉůĞEƵŵďĞƌ
DĞŵŽƌLJͲƚLJƉĞŽŶƚƌŽůŚĂƌƚƐŝŶ^ƚĂƚŝƐƚŝĐĂůWƌŽĐĞƐƐŽŶƚƌŽů
ϭϬϳ
ϭϬϴ
In both small and moderate shifts we see that the proposed chart detect the shift more quickly
than the others. The number of signals given by the proposed chart is also greater than the
classic ones. These outcomes are exactly in accordance with the findings of subsection 5.1.2.
ͷǤʹ
There is a lot of literature available on CUSUM and EWMA-type control charts for
monitoring the process dispersion, e.g. see Page (1963), Hawkins (1981), Acosta-Mejia et al.
(1999) and Chang and Gan (1995) for CUSUM-type charts and Ng and Case (1989),
Crowder and Hamilton (1992) and Huwang et al. (2010) for EWMA-type charts.
Additionally, Wu and Tian (2005) and Zhang and Chang (2008) also provided the CUSUM
and EWMA charts, respectively, for monitoring the process mean and variance
simultaneously.
Most of these charts are based on transforming the sample variance such that the new
transformed form may be closely approximated by a normally distributed variable and hence
applying the usual CUSUM and EWMA structures (recommended by Page (1954) and
Roberts (1959), respectively) on it. In a similar direction, Castagliola (2005) proposed a new
EWMA chart for monitoring the process dispersion. He used a logarithmic three
parameter transformation to obtain a normal approximation for sample variance. A similar
transformation is used by Castagliola et al. (2009) to setup a CUSUM chart for monitoring
process dispersion. The details regarding the EWMA and CUSUM charts based on the
logarithmic transformation are given in Section 3.2.
Castagliola et al. (2010) proposed another similar type of transformation based on a
four parameter Johnson ð transformation. They claimed that this four parameter
transformation gives a better approximation to the normal distribution as compared to the
DĞŵŽƌLJͲƚLJƉĞŽŶƚƌŽůŚĂƌƚƐŝŶ^ƚĂƚŝƐƚŝĐĂůWƌŽĐĞƐƐŽŶƚƌŽů
WƌŽŐƌĞƐƐŝǀĞĐŽŶƚƌŽůĐŚĂƌƚŝŶŐ
three parameter logarithmic transformation. With the notation of Castagliola et al. (2010), it
follows that
ñ ! ñ ¤ ¹
where ñ
ò§© ïó
ôó ïó ò§©
ñ Z,
º
(5.4)
ñ ñ Z, ¥ñ ñ Z' and õñ cñ Z' . Variable in (5.4)
follows approximately a normal distribution with mean ñ Z and variance 'ñ Z where the
values of ñ Z, 'ñ Z,
in Table 5.4.
ñ Z,
ñ Z, ñ Z and cñ Z for all Z P A B x +B are given
Table 5.4: Values of ¯ö , ±ö , 9ö , ²ö and nö
3
¯ö
0.0184
±ö
0.9475
9ö
3.1936
²ö
1.1952
nö
-0.2588
÷ö
4
0.0078
0.9739
3.3657
1.3983
-0.2438
12.591
5
0.0039
0.9852
3.5402
1.5727
-0.2352
11.312
6
0.0022
0.9908
3.7111
1.7281
-0.2295
10.530
7
0.0014
0.994
3.8768
1.8698
-0.2254
10.000
8
0.0009
0.9958
4.0369
2.0010
-0.2224
9.618
15.077
9
0.0006
0.9970
4.1918
2.1238
-0.2200
9.328
10
0.0004
0.9978
4.3417
2.2396
-0.2181
9.100
11
0.0003
0.9983
4.4869
2.3495
-0.2166
8.917
12
0.0002
0.9987
4.6279
2.4544
-0.2152
8.766
13
0.0002
0.9989
4.7648
2.5549
-0.2141
8.640
14
0.0001
0.9991
4.8981
2.6515
-0.2132
8.532
15
0.0001
0.9993
5.0279
2.7446
-0.2123
8.440
Note that in case of õñ ! ¥ñ - , the transformation given in (5.4) is not possible, but
Castagliola et al. (2010) showed that the probability of occurrence of this event is so close to
zero that it can be neglected. From the values of ¥ñ and õñ it can be noticed that õñ ! ¥ñ
- implies a very large value of as compared to the value of ' so it can be taken as
an out of control situation with a large positive shift. Details about the distributional
properties of can be found in Castagliola et al. (2010).
DĞŵŽƌLJͲƚLJƉĞŽŶƚƌŽůŚĂƌƚƐŝŶ^ƚĂƚŝƐƚŝĐĂůWƌŽĐĞƐƐŽŶƚƌŽů
ϭϬϵ
ϭϭϬ
Furthermore, it should be noted that any change in the process standard deviation will
change the mean of the normalized variables given in (3.5) and (5.4). So based on the these
two (approximately) normalized statistics, we are now able to define our new control
structures, named as floating and floating charts, respectively. These charts
monitor basically the mean of the transformed statistics in (3.5) and (5.4) and hence control
the process dispersion.
F
ͷǤʹǤͳ
The first proposed chart, named as floating chart, is based on the three
parameter logarithmic transformation given in (3.5). The plotting statistic is given as:
¦§ø«¬ £ø
(5.5)
The statistic in (5.5) is a cumulative average of the three parameter logarithmic
transformation given in (3.5). According to the probability distribution theory we have that, if
follows (approximately) a normal distribution with mean £ Z and variance '£ Z that
¦U¼ U N# will have mean £ Z and variance
]½© `
. This implies that the control
limits (including the upper control limit ( ), center line ( ) and lower control limit
( )) for the floating statistic given in (5.5) can be defined as:
£ Z
£
]½ `
_
£ Z,
,
£ Z !
where the width of the control limits is determined by
adjusting this constant (
£)
as the
£.
£
The
]½ `
_
(5.6)
can be controlled by
8 for a control chart with wider limits are larger and
vice versa. The same problem (like the control limits in (5.2)) also persists with the control
structure in (5.6), i.e. the control limits remain too wide for the larger values of # (wide
relative to the plotting statistic). Therefore, following (5.3), the penalized control limits for
DĞŵŽƌLJͲƚLJƉĞŽŶƚƌŽůŚĂƌƚƐŝŶ^ƚĂƚŝƐƚŝĐĂůWƌŽĐĞƐƐŽŶƚƌŽů
WƌŽŐƌĞƐƐŝǀĞĐŽŶƚƌŽůĐŚĂƌƚŝŶŐ
the proposed floating chart are given as:
£
where
£ Z
q
£
q ]½ `
£ y éè^>ç ,
£
£ Z,
£
£ Z !
q ]½ `
£ y éè^>ç
(5.7)
is the adjusted control limit coefficient.
Table 5.5: 9:; values for the floating ° µF chart with rq° I> EF and 9:;= F==
Ch
0.5
Ih
6.041
?h
3.779
Jh
2.890
h
2.417
0.6
8.069
5.003
3.810
3.164
0.7
11.874
7.316
5.546
4.585
0.8
20.468
12.625
9.533
7.835
0.9
49.392
31.652
23.918
19.851
0.95
98.625
70.337
56.347
47.163
1.05
92.869
67.545
54.129
46.352
1.1
47.906
31.469
24.452
20.236
1.2
21.249
13.379
10.268
8.496
1.3
12.892
8.130
6.259
5.213
1.4
9.113
5.778
4.461
3.740
1.5
7.040
4.482
3.485
2.941
2
3.316
2.205
1.774
1.527
3
1.834
1.323
1.148
1.067
Table 5.6: 9:; values for the floating ° µF chart with rq° I> ?gG and 9:;= IJ=
Ch
Ih
?h
Jh
h
0.5
7.157
4.425
3.391
0.6
9.629
5.892
4.478
3.696
0.7
14.123
8.674
6.542
5.371
0.8
24.638
15.057
11.310
9.280
0.9
61.342
38.450
28.858
23.761
0.95
135.241
91.241
71.136
59.045
1.05
128.999
89.714
69.639
58.704
1.1
60.472
38.941
29.562
24.367
1.2
25.692
16.054
12.208
10.039
1.3
15.356
9.646
7.356
6.110
1.4
10.757
6.792
5.220
4.333
1.5
8.234
5.229
4.049
3.377
2
3.813
2.506
2.001
1.703
3
2.030
1.434
1.206
1.103
2.807
DĞŵŽƌLJͲƚLJƉĞŽŶƚƌŽůŚĂƌƚƐŝŶ^ƚĂƚŝƐƚŝĐĂůWƌŽĐĞƐƐŽŶƚƌŽů
ϭϭϭ
ϭϭϮ
Table 5.7: 9:; values for the floating ° µF chart with rq° I> GHg and 9:;= ?==
0.5
Ch
7.896
Ih
4.870
?h
3.712
Jh
3.085
h
0.6
10.599
6.491
4.912
4.054
0.7
15.639
9.538
7.207
5.913
0.8
27.327
16.588
12.501
10.242
0.9
69.082
42.830
32.216
26.335
0.95
158.678
104.768
80.317
66.427
1.05
153.847
102.791
79.859
66.697
1.1
69.135
43.487
33.006
27.250
1.2
28.649
17.766
13.468
11.097
1.3
17.037
10.578
8.085
6.666
1.4
11.847
7.443
5.696
4.728
1.5
9.008
5.711
4.405
3.684
2
4.133
2.719
2.148
1.838
3
2.179
1.519
1.259
1.133
Note thaththe control limits given in (5.6) are a special case of the limits in (5.7) with ê .
We have tested several values of ê and ê >M was found to be optimal, i.e. ê >M gives
smaller
values for a fixed
proposed floating chart. The
.
Tables 5.5 – 5.7 contain the
values for the
8 for the proposed chart are evaluated by running
+´ simulations. The simulation program is developed in R language.
ͷǤʹǤʹ
F
The plotting statistic for the second proposed chart (based on a four parameter
Johnson ð transformation) to monitor the process dispersion is given as:
¦§ø«¬ ñø
Like in (5.5), here also have mean ñ Z and variance
(5.8)
© `
]ó
. Therefore the control
limits for this second proposed chart, named as floating chart, are given as:
DĞŵŽƌLJͲƚLJƉĞŽŶƚƌŽůŚĂƌƚƐŝŶ^ƚĂƚŝƐƚŝĐĂůWƌŽĐĞƐƐŽŶƚƌŽů
WƌŽŐƌĞƐƐŝǀĞĐŽŶƚƌŽůĐŚĂƌƚŝŶŐ
ñ
where
ñ Z
q
ñ
q ]ó `
ñ y éè^>ç ,
ñ
ñ Z,
ñ
ñ Z !
q ]ó `
ñ y éè^>ç
is the control limit coefficient for this second proposed chart. The
the floating chart are given in Tables 5.8 – 5.10.
(5.9)
values for
Table 5.8: 9:; values for the floating ö µF chart with rqö I> EF and 9:;= F==
0.5
Ch
5.737
Ih
3.557
?h
2.737
Jh
2.305
h
0.6
7.877
4.830
3.676
3.056
0.7
11.766
7.201
5.447
4.472
0.8
20.684
12.632
9.476
7.789
0.9
49.990
31.946
24.242
19.811
0.95
101.245
71.182
56.534
47.38
1.05
94.654
67.969
54.991
46.672
1.1
48.593
32.012
24.519
20.390
1.2
21.784
13.621
10.427
8.551
1.3
13.266
8.241
6.315
5.227
1.4
9.410
5.883
4.505
3.737
1.5
7.176
4.549
3.496
2.924
2
3.363
2.206
1.754
1.520
3
1.846
1.316
1.141
1.063
From Tables 5.5 – 5.10 we may conclude that:
i.
both floating charts are performing good, not only for positive shifts but also for
negative shifts in the process standard deviation;
ii.
iii.
for a fixed
,
the proposed floating chart is performing better for small
shifts, like @ Û >e and @ - +>P, whereas the performance of floating chart is
better for large shifts, like @ - >f and @ Û +>A;
for the fixed values of ê and
,
the values of the control limit coefficients are the
same for both proposed charts;
iv.
for large values of Z, the
values for both charts are more symmetric with respect
to @ as the distribution of both and becomes very close to normal as Z increases.
DĞŵŽƌLJͲƚLJƉĞŽŶƚƌŽůŚĂƌƚƐŝŶ^ƚĂƚŝƐƚŝĐĂůWƌŽĐĞƐƐŽŶƚƌŽů
ϭϭϯ
ϭϭϰ
Table 5.9: 9:; values for the floating ö µF chart with rqö I> ?gG and 9:;= IJ=
0.5
Ch
6.809
Ih
4.176
?h
3.183
Jh
2.649
h
0.6
9.339
5.692
4.301
3.562
0.7
14.052
8.522
6.413
5.266
0.8
24.861
14.998
11.250
9.176
0.9
62.300
38.592
29.075
23.830
0.95
138.501
92.615
71.719
59.133
1.05
131.474
89.931
70.533
59.158
1.1
62.083
39.362
29.939
24.634
1.2
26.376
16.284
12.254
10.120
1.3
15.833
9.753
7.411
6.090
1.4
11.044
6.888
5.247
4.337
1.5
8.401
5.280
4.027
3.364
2
3.821
2.473
1.946
1.667
3
2.015
1.402
1.187
1.093
Table 5.10: 9:; values for the floating ö µF chart with rqö I> GHg and 9:;= ?==
Ch
0.5
Ih
?h
Jh
h
7.511
4.583
3.480
2.884
0.6
10.318
6.257
4.716
3.889
0.7
15.587
9.379
7.044
5.769
0.8
27.370
16.606
12.389
10.122
0.9
70.351
43.080
32.349
26.301
0.95
161.226
105.404
80.668
66.694
1.05
155.272
103.961
80.665
66.948
1.1
70.928
44.268
33.470
27.389
1.2
29.405
18.016
13.538
11.106
1.3
17.526
10.787
8.137
6.688
1.4
12.167
7.525
5.735
4.723
1.5
9.239
5.746
4.397
3.652
2
4.125
2.654
2.085
1.771
3
2.119
1.460
1.226
1.110
DĞŵŽƌLJͲƚLJƉĞŽŶƚƌŽůŚĂƌƚƐŝŶ^ƚĂƚŝƐƚŝĐĂůWƌŽĐĞƐƐŽŶƚƌŽů
WƌŽŐƌĞƐƐŝǀĞĐŽŶƚƌŽůĐŚĂƌƚŝŶŐ
ͷǤʹǤ͵
In this section, we compare the performance of the proposed floating charts with some
recently proposed CUSUM and EWMA-type control charts for monitoring the process
dispersion. The control charts selected for the comparison purpose include the -EWMA by
Castagliola (2005), EWMA Â by Castagliola et al. (2010) and CUSUM by
Castagliola et al. (2009) directly, while we have also compared the performance of our
proposed charts with the Shewhart
chart, a CUSUM chart for process dispersion
proposed by Page (1963) and an EWMA chart proposed by Crowder and Hamilton (1992),
indirectly.
Proposed versus µF -EWMA and EWMA Å µF : Castagliola (2005) proposed an EWMA
chart for monitoring the process dispersion based on the same logarithmic transformation as
in (3.5), named as -EWMA. Following him, Castagliola et al. (2010) proposed another
EWMA chart based on the same four parameter Johnson ð transformation as in (5.4), named
as EWMA Â for controlling the process standard deviation. The two parameters of these
charts are the smoothing parameter * and the control limit coefficient
these two charts for the optimal choices of * and
. The
values of
are given in Table 5.11. Comparing the
performance of proposed charts with these EWMA-type charts, we can notice that both
proposed charts have smaller
values for a fixed
Pd. Moreover, the proposed
charts are showing more dominance for small shifts as compared to large values of @ (cf.
Table 5.11vs. Tables 5.6 & 5.9).
Castagliola (2005) showed in his article that the -EWMA control chart performs better than
the Shewhart
chart for small shifts like @ - M. He also proved the dominance of his
proposed chart over the CUSUM chart proposed by Page (1963) and the EWMA charts
DĞŵŽƌLJͲƚLJƉĞŽŶƚƌŽůŚĂƌƚƐŝŶ^ƚĂƚŝƐƚŝĐĂůWƌŽĐĞƐƐŽŶƚƌŽů
ϭϭϱ
ϭϭϲ
proposed by Crowder and Hamilton (1992). Therefore, we can state that the performance of
our proposed charts is better than these charts too.
Table 5.11: 9:; values for the µF -EWMA and EWMA Å µF charts with 9:;= IJ=
C
µF -EWMA
Ih
EWMA Å µF
0.5
10
5.6
?h
Jh
4
3.1
h
9.4
Ih
5.2
?h
3.7
Jh
2.9
h
0.6
13.9
8
5.7
4.5
13.6
7.7
5.5
4.3
0.7
21.3
12.6
9.1
7.1
21.2
12.4
8.9
7
0.8
40.8
23.1
17
13.5
40.8
23
16.8
13.4
0.9
130.3
68.9
48.5
38.2
126.5
68.3
48.4
38.1
0.95
289.7
184.8
137.6
110.3
274.3
179.9
135.8
109.4
1.05
173.2
142.3
115.1
96.8
195.9
148
118.1
98.8
1.1
91.9
59.8
45.2
36.9
97
61.6
46.1
37.5
1.2
36.5
22.8
17
13.8
38.8
23.5
17.4
14
1.3
20.3
12.4
9.3
7.5
21.3
13
9.6
7.7
1.4
13.4
8.1
6.1
4.9
13.8
8.4
6.3
5.1
1.5
9.8
5.8
4.4
3.6
9.8
6
4.5
3.7
2
3.9
2.3
1.8
1.5
3.8
2.3
1.8
1.5
Proposed vs. CUSUM µF : Castagliola et al. (2009) proposed a CUSUM chart based
on the three parameter logarithmic transformation as in (3.5). The
chart with optimal parameter choices are given in Table 5.12.
8 of the CUSUM
Table 5.12: 9:; values for the CUSUM µF chart with 9:;= IJ=
C
0.5
0.6
0.7
0.8
0.9
0.95
1.05
1.1
1.2
1.3
1.4
1.5
2
3
10.8
15.4
24.1
44
108.9
216.9
183.3
98.6
39.5
21.7
14.1
10.2
3.8
5
5.6
8.3
13.4
25.4
68.4
154.8
145.7
64.6
24.3
13.1
8.3
5.9
2.3
7
3.8
5.7
9.4
18.2
51.1
122.9
117.5
49.2
18
9.6
6.1
4.3
1.8
9
2.9
4.4
7.3
14.3
41.3
103.1
99.5
40.3
14.5
7.7
4.9
3.5
1.5
DĞŵŽƌLJͲƚLJƉĞŽŶƚƌŽůŚĂƌƚƐŝŶ^ƚĂƚŝƐƚŝĐĂůWƌŽĐĞƐƐŽŶƚƌŽů
WƌŽŐƌĞƐƐŝǀĞĐŽŶƚƌŽůĐŚĂƌƚŝŶŐ
The performance of this CUSUM control chart is more or less similar to that of
-EWMA and EWMA Â charts. Comparing the performance of our proposed charts
with the CUSUM , we may conclude that the proposed charts are performing better than
the CUSUM chart for almost all values of @ (cf. Table 5.12 vs. Tables 5.6 & 5.9).
Apart from the tabular comparison, Figures 5.10 and 5.11 provide the
curves of the
charts discussed in this subsection for an increase and a decrease, respectively, in the process
dispersion. It is clear from Figures 5.10 – 5.11 that the
curves of both proposed charts
are on the lower side of other curves. This shows that the proposed charts have smaller
values for a fixed
Pd. In addition, both proposed charts are showing almost the
same performance as their
curves coincide in both figures.
Figure 5.10: 9:; curves for floating ° µF , floating ö µF , µF -EWMA, EWMA Å µF and
CUSUM µF charts for increase in the process dispersion
&ůŽĂƚŝŶŐdͲ^Ϯ
&ůŽĂƚŝŶŐhͲ^Ϯ
^ϮͲtD
tD:Ͳ^Ϯ
h^hD^Ϯ
ϭϲϬ
ϭϰϬ
ϭϮϬ
ϭϬϬ
ϴϬ
ϲϬ
ϰϬ
ϮϬ
Ϭ
ϭ͘Ϭϱ
ϭ͘ϭ
ϭ͘Ϯ
ϭ͘ϯ
ϭ͘ϰ
ϭ͘ϱ
Ϯ
DĞŵŽƌLJͲƚLJƉĞŽŶƚƌŽůŚĂƌƚƐŝŶ^ƚĂƚŝƐƚŝĐĂůWƌŽĐĞƐƐŽŶƚƌŽů
ϭϭϳ
ϭϭϴ
Figure 5.11: 9:; curves for floating ° µF , floating ö µF , µF -EWMA, EWMA Å µF and
CUSUM µF charts for decrease in the process dispersion
&ůŽĂƚŝŶŐdͲ^Ϯ
&ůŽĂƚŝŶŐhͲ^Ϯ
^ϮͲtD
tD:Ͳ^Ϯ
h^hD^Ϯ
ϮϬϬ
ϭϴϬ
ϭϲϬ
ϭϰϬ
ϭϮϬ
ϭϬϬ
ϴϬ
ϲϬ
ϰϬ
ϮϬ
Ϭ
Ϭ͘ϵϱ
Ϭ͘ϵ
Ϭ͘ϴ
Ϭ͘ϳ
Ϭ͘ϲ
Ϭ͘ϱ
ͷǤʹǤͶ
For illustrating the application of the proposed charts, we generate two datasets
(namely dataset 5.3 and dataset 5.4) of 25 subgroups each of size Z B, i.e. one for an
increase and the other for a decrease in the process standard deviation. For dataset 5.3, the
first 15 subgroups are generated from [+ showing an in control standard deviation while
the remaining 10 subgroups are generated from [+>M referring to an out of control
standard deviation with @ +>M. Similarly, for dataset 5.4 the first 15 subgroups are the same
as for dataset 5.3, whereas the remaining 10 subgroups are taken from [>f showing an
negative shift in the process dispersion with @ >f. Both proposed charts are applied to the
datasets with parameters; £ Z >dAf, '£ Z >eXd,
M>PXAd, £ Z >BeXe, ê >M and
>Pe, 'ñ Z >efBM,
ñ Z
q
£
£ Z
>feXe, £ Z
P>BXf for the floating chart; ñ Z
P>BAM, ñ Z +>BdMd, ñ Z >MPBM, cñ Z
DĞŵŽƌLJͲƚLJƉĞŽŶƚƌŽůŚĂƌƚƐŝŶ^ƚĂƚŝƐƚŝĐĂůWƌŽĐĞƐƐŽŶƚƌŽů
WƌŽŐƌĞƐƐŝǀĞĐŽŶƚƌŽůĐŚĂƌƚŝŶŐ
++>P+M, ê >M and
q
ñ
P>BXf for the floating chart. The calculations for both
proposed charts with dataset 5.3 are given in Table 5.13. Figure 5.12 shows the chart output
of the floating chart for both datasets while the chart output of the floating
chart is given in Figure 5.13.
Table 5.13: Calculation details of the proposed charts for dataset 5.3
Subgroup
Number
µF
°
ù°
;n;°
ön;°
ö
3.458
-0.045
2.131
-0.998
ùö
-0.045
;n;ö
-3.511
ön;ö
-0.521
-2.16
2.168
1.633
1
0.815
-0.079
-0.079
-3.443
2
0.363
-0.992
-0.536
-2.116
3.519
3
1.567
0.929
-0.047
-1.592
1.607
0.924
-0.04
-1.625
4
0.594
-0.482
-0.156
-1.3
1.315
-0.45
-0.142
-1.328
1.336
5
1.565
0.927
0.061
-1.111
1.126
0.922
0.071
-1.135
1.143
6
2.384
1.687
0.332
-0.977
0.992
1.654
0.335
-0.999
1.007
7
0.659
-0.356
0.234
-0.876
0.891
-0.321
0.241
-0.896
0.904
8
0.069
-1.857
-0.028
-0.797
0.812
-2.106
-0.052
-0.816
0.824
9
0.709
-0.264
-0.054
-0.734
0.749
-0.228
-0.072
-0.751
0.759
10
0.564
-0.543
-0.103
-0.681
0.696
-0.513
-0.116
-0.697
0.705
11
0.529
-0.614
-0.149
-0.637
0.651
-0.588
-0.159
-0.652
0.66
12
0.696
-0.287
-0.161
-0.598
0.613
-0.251
-0.167
-0.613
0.621
13
0.875
0.018
-0.147
-0.565
0.58
0.051
-0.15
-0.58
0.588
14
2.485
1.765
-0.01
-0.536
0.551
1.731
-0.015
-0.55
0.558
15
0.431
-0.829
-0.065
-0.511
0.526
-0.818
-0.069
-0.524
0.532
16
2.123
1.47
0.031
-0.488
0.503
1.442
0.025
-0.501
0.509
17
3.219
2.27
0.163
-0.467
0.482
2.247
0.156
-0.48
0.488
18
0.332
-1.068
0.094
-0.449
0.464
-1.085
0.087
-0.461
0.469
19
2.552
1.816
0.185
-0.432
0.447
1.782
0.176
-0.444
0.451
20
0.148
-1.592
0.096
-0.416
0.431
-1.731
0.081
-0.428
0.436
21
3.935
2.677
0.219
-0.402
0.417
2.694
0.205
-0.413
0.421
22
1.592
0.957
0.253
-0.389
0.404
0.95
0.239
-0.4
0.408
23
3.744
2.575
0.354
-0.377
0.392
2.579
0.341
-0.388
0.395
24
1.834
1.205
0.389*
-0.366
0.38
1.187
0.376
-0.376
0.384
25
1.215
0.51
0.394*
-0.355
0.37
0.525
0.382**
-0.365
0.373
* indicates an out of control signal by floating chart
** indicates an out of control signal by floating chart
DĞŵŽƌLJͲƚLJƉĞŽŶƚƌŽůŚĂƌƚƐŝŶ^ƚĂƚŝƐƚŝĐĂůWƌŽĐĞƐƐŽŶƚƌŽů
ϭϭϵ
ϭϮϬ
It can be seen from Figures 5.12 – 5.13 that the proposed charts are effectively detecting both
positive and negative shifts. This can also be confirmed from Table 5.13, where the floating
chart is signaling at subgroups # 24 and 25, while the floating chart is
detecting the positive shift at subgroup # 25.
Figure 5.12: Chart output of floating ° µF chart for datasets 5.3 & 5.4
ĚĂƚĂƐĞƚϱ͘ϯ
ĚĂƚĂƐĞƚϱ͘ϰ
ŽŶƚƌŽů>ŝŵŝƚƐ
ϰ
ϯ
Ϯ
ϭ
Ϭ
Ͳϭ
ͲϮ
Ͳϯ
Ͳϰ
ϭ
ϯ
ϱ
ϳ
ϵ
ϭϭ
ϭϯ
ϭϱ
ϭϳ
ϭϵ
Ϯϭ
Ϯϯ
Ϯϱ
^ĂŵƉůĞEƵŵďĞƌ
Figure 5.13: Chart output of floating ö µF chart for datasets 5.3 & 5.4
ĚĂƚĂƐĞƚϱ͘ϯ
ĚĂƚĂƐĞƚϱ͘ϰ
ŽŶƚƌŽů>ŝŵŝƚƐ
ϰ
ϯ
Ϯ
ϭ
Ϭ
Ͳϭ
ͲϮ
Ͳϯ
Ͳϰ
ϭ
ϯ
ϱ
ϳ
ϵ
ϭϭ
ϭϯ
ϭϱ
ϭϳ
^ĂŵƉůĞEƵŵďĞƌ
DĞŵŽƌLJͲƚLJƉĞŽŶƚƌŽůŚĂƌƚƐŝŶ^ƚĂƚŝƐƚŝĐĂůWƌŽĐĞƐƐŽŶƚƌŽů
ϭϵ
Ϯϭ
Ϯϯ
Ϯϱ
WƌŽŐƌĞƐƐŝǀĞĐŽŶƚƌŽůĐŚĂƌƚŝŶŐ
ͷǤ͵
To monitor the location of a process two main types of charts, named as Shewharttype control charts and memory control charts (like EWMA and CUSUM) are available.
Former are recommended for large shifts P - @ - B while latter are good at detecting small
and moderate shifts (like >MB - @ - +>B). In this chapter, we have proposed another type of
memory control chart, the so called 7 control chart. The performance of the new chart is
evaluated in terms of
8 and we have compared the performance of proposed control chart
with different existing memory control charts. Comparisons revealed that the newly proposed
chart performs very good and outperforms its competitors for small and moderate shifts but
also shows good performance for large shifts.
Following the 7 chart for location, two memory-type control charts for process
dispersion, named as the floating control chart (based on a three parameter
logarithmic transformation) and the floating control chart (based on a four parameter
Johnson ð transformation) are proposed. The performance evaluation of the proposed charts
is done by calculating the
values using simulation procedures. These
8 are
compared with some EWMA- and CUSUM-type control charts for monitoring the process
standard deviation. The comparisons show that the proposed charts are dominating the other
charts in terms of
values. Moreover, an inter-proposed charts comparison shows that the
floating chart is better for small shifts, whereas the floating chart is superior
for large shifts in the process dispersion. At the end, an illustrative example is provided
which shows the application of the proposed charts on simulated datasets.
DĞŵŽƌLJͲƚLJƉĞŽŶƚƌŽůŚĂƌƚƐŝŶ^ƚĂƚŝƐƚŝĐĂůWƌŽĐĞƐƐŽŶƚƌŽů
ϭϮϭ
ϭϮϮ
ͷǤͳ
y=c(); PM=c(); ucl=c(); lcl=c(); rl=c()
mu=0; sig=1
p=0.2; K=3.129
for(j in 1:100000)
{
for(i in 1:1000000)
{
y[i]=rnorm(1,mu,sig)
PM[i]=mean(y[1:i])
ucl[i]=mu+K*(sig/i^(0.5+p))
lcl[i]=mu-K*(sig/i^(0.5+p))
if(PM[i]>ucl[i]|PM[i]<lcl[i])
{rl[j]=i;break;}
else{rl[j]=0;}
}
}
mean(rl)
ͷǤʹ
rl=c(); Tj=c(); FT=c(); ucl=c(); lcl=c()
n=5; a=-0.8969; b=2.3647; c=0.5979; ETj=0.00748; STj=0.9670
p=0.2; K=3.129
mu=0; sig=1
for(j in 1:100000)
{
for(i in 1:1000000)
{
x=rnorm(n,mu,sig)
Tj[i]=a+b*log(var(x)+c,exp(1))
FT[i]=mean(Tj[1:i])
ucl[i]=ETj+K*(STj/i^(0.5+p))
lcl[i]=ETj-K*(STj/i^(0.5+p))
if(FT[i]>ucl[i]|FT[i]<lcl[i])
{rl[j]=i; break;}
else{rl[j]=0;}
}
}
mean(rl)
DĞŵŽƌLJͲƚLJƉĞŽŶƚƌŽůŚĂƌƚƐŝŶ^ƚĂƚŝƐƚŝĐĂůWƌŽĐĞƐƐŽŶƚƌŽů
ZĞĨĞƌĞŶĐĞƐ
Abbas, N., Riaz, M. and Does, R.J.M.M. (2011). Enhancing the Performance of EWMA
Charts. Quality and Reliability Engineering International, 27(6), 821 – 833.
Abbas, N., Riaz, M. and Does, R.J.M.M. (2012a). Mixed Exponentially Weighted Moving
Average – Cumulative Sum charts for Process Monitoring. Quality and Reliability
Engineering International, DOI: 10.1002/qre.1385.
Abbas, N., Riaz, M. and Does, R.J.M.M. (2012b). CS-EWMA Chart for Monitoring Process
Dispersion. Quality and Reliability Engineering International, DOI: 10.1002/qre.1414.
Abbas, N., Riaz, M. and Does, R.J.M.M. (2012c). An EWMA-type Control Chart for
Monitoring the Process Mean Using Auxiliary Information. Accepted for publication in
Communications in Statistics – Theory and Methods.
Abbas, N., Riaz, M. and Does, R.J.M.M. (2012d). New Memory-type Control Charts for
Monitoring the Process Dispersion. Submitted to International Journal of Production
Research.
Abbas, N., Zafar, R.F., Riaz, M. and Hussain, Z. (2012). Progressive Mean Control Chart for
Monitoring Location Parameter. Quality and Reliability Engineering International, DOI:
10.1002/qre.1386.
Acosta-Mejia, C., Pigniatiello, J. and Rao, B. (1999). A comparison of Control Charting
Procedures for Monitoring Process Dispersion. IIE Transactions, 31(6), 569-579.
Antzoulakos, D.L. and Rakitzis, A.C. (2008). The Modified r out of m Control Chart.
Communications in Statistics – Theory and Methods, 37(2), 396-408.
Bonetti, P.O., Waeckerlin, A., Schuepfer, G. and Frutiger, A. (2000). Improving TimeSensitive Processes in the Intensive Care Unit: the Example of ‘Door-to-Needle Time’ in
Acute Myocardial Infarction. International Journal for Quality in Health Care, 12(4), 311317.
Brook, D. and Evans, D.A. (1972). An Approach to the Probability Distribution of CUSUM
Run Length. Biometrika, 59(3), 539 - 549.
Capizzi, G.and Masarotto, G. (2003). An Adaptive Exponentially Weighted Moving Average
Control Chart. Technometrics, 45(3), 199-207.
Castagliola, P. (2005). A New S2-EWMA Control Chart for Monitoring Process Variance.
Quality and Reliability Engineering International, 21(8), 781-794.
Castagliola, P., Celano, G. and Fichera, S. (2009). A New CUSUM-S2 Control Chart for
Monitoring the Process Variance. Journal of Quality in Maintenance Engineering, 15(4),
344-357.
DĞŵŽƌLJͲƚLJƉĞŽŶƚƌŽůŚĂƌƚƐŝŶ^ƚĂƚŝƐƚŝĐĂůWƌŽĐĞƐƐŽŶƚƌŽů
ϭϮϯ
ϭϮϰ
Castagliola, P., Celano, G. and Fichera, S. (2010). A Johnson's Type Transformation EWMAS2 Control Chart. International Journal of Quality Engineering and Technology, 1(3), 253275.
Chang, T.C. and Gan, F.F. (1995). A Cumulative Sum Control Chart for Monitoring Process
Variance. Journal of Quality Technology, 27(2), 109-119.
Chatterjee, S. and Qiu, P. (2009). Distribution-Free Cumulative Sum Control Charts Using
Bootstrap-Based Control Limits. Annals of Applied Statistics, 3(1), 349-369.
Cochran, W.G. (1977). Sampling Techniques. 3rd ed. New York: John Wiley and Sons.
Crowder, S.V. and Hamilton, M.D. (1992). An EWMA for Monitoring a Process Standard
Deviation. Journal of Quality Technology, 24(1), 12-21.
Ewan, W.D. and Kemp, K.W. (1960). Sampling inspection of continuous processes with no
autocorrelation between successive results. Biometrika, 47(3-4), 363-380.
Fuller, W.A. (2002). Regression Estimation for Survey Samples. Survey Methodology, 28(1),
5-23.
Fuller, W.A. (2009). Sampling Statistics. John Wiley and Sons Inc., Hoboken, New Jersey.
Garcia, M.R. and Cebrian, A.A. (1996). Repeated Substitution Method: The Ratio Estimator
for the Population Variance. Metrika, 43(1), 101-105.
Hawkins, D.M. (1981). A CUSUM for a Scale Parameter. Journal of Quality Technology,
13(4), 228-231.
Hawkins, D.M. and Olwell, D.H. (1998). Cumulative Sum Charts and Charting
Improvement. Springer, New York.
Huwang, L., Huang, C. J. and Wang, Y.H.T. (2010). New EWMA Control Charts for
Monitoring Process Dispersion. Computational Statistics and Data Analysis, 54(10), 23282342.
Jiang, W., Shu, L. and Aplet, D.W. (2008). Adaptive CUSUM Procedures with EWMABased Shift Estimators. IIE Transactions, 40(10), 992-1003.
Kadilar, C. and Cingi, H. (2005). A New Estimator Using Two Auxiliary Variables. Applied
Mathematics and Computation, 162(2), 901–908.
Khoo, M.B.C. (2004). Design of Runs Rules Schemes. Quality Engineering, 16(1), 27–43.
Klein, M. (2000). Two Alternatives to the Shewhart X Control Chart. Journal of Quality
Technology, 32(4), 427–431.
Koutras, M.V., Bersimis, S. and Maravelakis, P.E. (2007). Statistical Process Control using
Shewhart Control Charts with Supplementary Runs Rules. Methodology and Computing in
Applied Probability, 9(2), 207–224.
Lowry, C.A., Woodall, W.H., Champ, C.W. and Rigdon, S.E. (1992). A Multivariate
Exponentially Weighted Moving Average Control Chart. Technometrics, 34(1), 46 – 53.
DĞŵŽƌLJͲƚLJƉĞŽŶƚƌŽůŚĂƌƚƐŝŶ^ƚĂƚŝƐƚŝĐĂůWƌŽĐĞƐƐŽŶƚƌŽů
ZĞĨĞƌĞŶĐĞƐ
Lucas, J.M. (1982). Combined Shewhart-CUSUM Quality Control Schemes. Journal of
Quality Technology, 14(2), 51-59.
Lucas, J.M. and Crosier, R.B. (1982). Fast Initial Response for CUSUM Quality-Control
Scheme. Technometrics, 24(3), 199-205.
Lucas J.M. and Saccucci M.S. (1990). Exponentially Weighted Moving Average Control
Schemes: Properties and Enhancements. Technometrics, 32(1), 1–12.
Mandel, B.J. (1969). The Regression Control Chart. Journal of Quality Technology, 1(1), 1-9.
McIntyre, G.A. (1952). A Method for Unbiased Selective Sampling, Using Ranked Sets.
Australian Journal of Agricultural Research, 3(4), 385-390.
Montgomery, D.C. (2009). Introduction to Statistical Quality Control. 6th ed. New York:
John Wiley & Sons.
Nelson, L.S., (1984), The Shewhart Control Chart – Tests for Special Causes, Journal of
Quality Technology, 16(4), 237-239
Ng, C.H. and Case, K.E. (1989). Development and Evaluation of Control Charts Using
Exponentially Weighted Moving Averages. Journal of Quality Technology, 21(4), 242-250.
Page, E.S. (1954). Continuous Inspection Schemes. Biometrika, 41(1-2), 100-115.
Page, E.S. (1963). Controlling the Standard Deviation and Warning Lines by CUSUM.
Technometrics, 5(3), 307-309.
Palm, A.C. (1990). Tables of Run Length Percentiles for Determining the Sensitivity of
Shewhart Control Charts for Averages with Supplementary Runs Rules. Journal of Quality
Technology, 22(4), 289–298.
Riaz, M. (2008a). Monitoring Process
Computational Statistics, 23(2), 253-276.
Variability
using
Auxiliary
Information.
Riaz, M. (2008b). Monitoring Process Mean Level using Auxiliary Information. Statistica
Neerlandica, 62(4), 458-481.
Riaz, M., Abbas, N. and Does, R.J.M.M. (2011). Improving the Performance of CUSUM
Charts. Quality and Reliability Engineering International, 27(4), 415 – 424.
Riaz, M. and R.J.M.M. Does (2009). A process variability control chart. Computational
Statistics, 24(2), 345–368.
Roberts, S.W. (1959), Control Chart Tests Based on Geometric Moving Averages,
Technometrics, 1(3), 239-250.
Shewhart, W.A. (1931). Economic Control of Quality of Manufactured Product, New York:
1931. Reprinted by ASQC, Milwaukee, 1980.
Shmueli, G. and Cohen, A. (2003). Run-Length Distribution for Control Charts with Runs
and Scans Rules. Communications in Statistics – Theory and Methods, 32(2), 475–495.
Shu, L.J. and Jiang, W. (2008). A New EWMA Chart for Monitoring Process Dispersion.
Journal of Quality Technology, 40(3), 319-331.
DĞŵŽƌLJͲƚLJƉĞŽŶƚƌŽůŚĂƌƚƐŝŶ^ƚĂƚŝƐƚŝĐĂůWƌŽĐĞƐƐŽŶƚƌŽů
ϭϮϱ
ϭϮϲ
Siegmund, D.O. (1985). Sequential Analysis: Test and Confidence Intervals. New York:
Springer-Verlag.
Steiner, S.H. (1999). EWMA control charts with time varying control limits and fast initial
response. Journal of Quality Technology, 31(1), 75 – 86.
Tuprah, K. and Ncube, M. (1987). A Comparison of Dispersion Quality Control charts.
Sequential Analysis, 6(2), 155-163.
Wade, M.R. and Woodall, W.H. (1993). A Review and Analysis of Cause-Selecting Control
Charts. Journal of Quality Technology, 25(3), 161-169.
Waldmann, K.H. (1995). Design of double CUSUM quality control schemes. European
Journal of Operational Research, 95(3), 641-648.
Westgard, J.O., Groth, T., Aronsson, T. and De Verdier. C.H. (1977). Combined ShewhartCUSUM Control Chart for Improved Quality Control in Clinical Chemistry. Clinical
Chemistry, 23(10), 1881-1887.
Wortham, A.W. and Ringer, L.J. (1971). Control via Exponentially Smoothing. The Logistics
and Transportation Review, 7(1), 33-39.
Wu, Z. and Tian, Y. (2005). Weighted-loss-Function CUSUM chart for Monitoring Mean
and Variance of a Production Process. International Journal of Production Research, 43(14),
3027–3044.
Yashchin, E. (1989). Weighted Cumulative Sum Technique. Technometrics, 31(1), 321-338.
Zhang, G.X. (1985). Cause-Selecting Control Charts-A New Type of Quality Control Charts.
The QR Journal, 12, 221-225.
Zhang, G. and Chang, S.I. (2008). Multivariate EWMA Control Charts Using Individual
Observations for Process Mean and Variance Monitoring and Diagnosis. International
Journal of Production Research, 46(24), 6855–6881.
DĞŵŽƌLJͲƚLJƉĞŽŶƚƌŽůŚĂƌƚƐŝŶ^ƚĂƚŝƐƚŝĐĂůWƌŽĐĞƐƐŽŶƚƌŽů
^ĂŵĞŶǀĂƚƚŝŶŐ
Het doel van statistische procesbeheersing is om een procesvoering te krijgen die
geschoeid is op een kwantitatieve leest. De regelkaart is het belangrijkste gereedschap van
statistische procesbeheersing en werd in 1924 door Shewhart geïntroduceerd in het
bedrijfsleven. Het is een grafiek van metingen van een kwaliteitskarakteristiek van het proces
op de verticale as uitgezet tegen de tijd op de horizontale as. De grafiek wordt aangevuld met
regelgrenzen die de procesinherente variatie markeren. Zodra een meting buiten de
regelgrenzen valt dan noemen we het proces niet beheerst. In de loop der jaren is hier veel
onderzoek naar gedaan.
In begin vijftiger jaren van de vorige eeuw ontwikkelde Page een nieuw type
regelkaart: de gecumuleerde som regelkaart, beter bekend geworden als de CUSUM
regelkaart. Kort daarna, in 1959, presenteerde Roberts de exponentieel gewogen
voortschrijdend gemiddelde regelkaart, die bekend werd onder de naam van EWMA
regelkaart. Beide regelkaarten gebruiken naast de laatst gemeten waarneming ook de
voorafgaande waarnemingen om te oordelen of het proces beheerst is. In die zin hebben deze
regelkaarten een geheugen. In hoofdstuk 1 worden beide regelkaarten geïntroduceerd.
De standaard signaleringsregel is dat als een waarneming buiten de regelgrenzen valt,
het proces niet beheerst is. Door vervanging van de regelgrenzen door waarschuwings- en/of
actiegrenzen kunnen ook patronen van opeenvolgende waarnemingen als signaleringsregels
gebruikt worden. In hoofdstuk 2 worden deze aanpassingen gebruikt om de CUSUM - en
EWMA regelkaarten tot snellere signalering te laten komen bij met name kleine
verschuivingen in het gemiddelde. Hiervoor wordt als criterium genomen de gemiddelde run
DĞŵŽƌLJͲƚLJƉĞŽŶƚƌŽůŚĂƌƚƐŝŶ^ƚĂƚŝƐƚŝĐĂůWƌŽĐĞƐƐŽŶƚƌŽů
ϭϮϳ
ϭϮϴ
lengte, waarbij als definitie van de run lengte gebruikt wordt: het aantal opeenvolgende
waarnemingen dat nodig is om tot een signaal te komen. Het blijkt dat de aanpassingen in
hoofdstuk 2 tot substantiële verbetering leiden voor het sneller signaleren van kleine
verschuivingen in het gemiddelde.
In hoofdstuk 3 worden de CUSUM – en EWMA regelkaarten gecombineerd tot een
gemeenschappelijke regelkaart. Deze gecombineerde regelkaart wordt ontwikkeld voor zowel
het monitoren van de locatie als van de spreiding. Vergelijkingen op basis van de gemiddelde
run lengten worden gemaakt met bestaande regelkaarten. Het blijkt dat de gecombineerde
regelkaart tot betere resultaten leidt.
Een betere beheersing van de procesparameters kan ook verkregen worden door
gebruik te maken van aanvullende informatie uit de data. Door het benutten van de correlatie
tussen de kwaliteitskarakteristiek en de aanvullende kenmerken, zijn nieuwe CUSUM – en
EWMA regelkaarten ontwikkeld. In hoofdstuk 4 worden deze nieuwe regelkaarten
vergeleken met bestaande regelkaarten die gebruikt worden voor hetzelfde doel. Het blijkt dat
aanvullende informatie de prestaties van de regelkaarten substantieel verbetert.
Tot slot wordt in het laatste hoofdstuk een nieuw type regelkaart voorgesteld die
wordt gebaseerd op het progressief voortschrijdend gemiddelde. Deze regelkaart wordt zowel
voor de locatie parameter als voor de spreidingsparameter ontwikkeld. De prestaties,
wederom in termen van de gemiddelde run lengte, blijken in de praktijk uitstekend te zijn.
DĞŵŽƌLJͲƚLJƉĞŽŶƚƌŽůŚĂƌƚƐŝŶ^ƚĂƚŝƐƚŝĐĂůWƌŽĐĞƐƐŽŶƚƌŽů
ƵƌƌŝĐƵůƵŵsŝƚĂĞ
Nasir Abbas was born in Rawalpindi (Pakistan) on 21st June, 1987. He completed his
schooling in 2003 from F.G. Secondary School, Mehfooz Road, Rawalpindi, Pakistan. He
earned his F.Sc. and B.Sc in 2005 and 2007, respectively, from the F.G. Sir Syed College,
The Mall, Rawalpindi, Pakistan with the major subjects as Mathematics, Statistics and
Physics at F.Sc. level, and Mathematics, Statistics and Economics as B.Sc. level. He got his
M.Sc. in Statistics from the Department of Statistics Quaid-i-Azam University Islamabad
Pakistan in 2009 where he secured the top position in his department and M.Phil in Statistics
from the Department of Statistics Quaid-i-Azam University Islamabad Pakistan in 2011.
Additionally, he is also serving as Assistant Census Commissioner in Pakistan Bureau
of Statistics from July 2011-present. His current research interests include Statistical Quality
Control particularly control charting based on the classical, Bayesian as well as nonparametric methodologies.
DĞŵŽƌLJͲƚLJƉĞŽŶƚƌŽůŚĂƌƚƐŝŶ^ƚĂƚŝƐƚŝĐĂůWƌŽĐĞƐƐŽŶƚƌŽů
View publication stats
ϭϮϵ