Time Series Control Charts in the

Daniel W. Apley
Presence of Model Uncertainty
Assistant Professor,
Time series control charts are popular methods for statistical process control of autocor-
Mem., ASME
related processes. In order to implement these methods, however, a time series model of
Department of Industrial Engineering,
the process is required. Since time series models must always be estimated from process
Texas A&M University,
data, model estimation errors are unavoidable. In the presence of modeling errors, time
College Station, TX 77843-3131
series control charts that are designed under the assumption of a perfect model may have
e-mail: [email protected]
an actual in-control average run length that is substantially shorter than desired. This
paper presents a method for incorporating model uncertainty information into the design
of time series control charts to provide a level of robustness with respect to modeling
errors. The focus is on exponentially weighted moving average charts and Shewhart
individual charts applied to the time series residuals. 关DOI: 10.1115/1.1510520兴

1 Introduction esen 关7兴, Dooley and Kapoor 关8兴, Dooley, Kapoor, Dessouky, and
DeVor 关9兴, English, Krishnamurthi, and Sastri 关10兴, Lin and Ad-
Statistical process control 共SPC兲 techniques are widely used for
ams 关11兴, Montgomery and Mastrangelo 关12兴, Notohardjono and
monitoring and improving quality in industrial processes. Tradi-
Ermer 关13兴, Runger, Willemain, and Prabhu 关14兴, Superville and
tional SPC is based on the assumption that process data are inde-
Adams 关15兴, Vander Wiel 关16兴, and Wardell, Moskowitz, and
pendent. Advances that have occurred in the areas of measurement
Plante 关17兴兲 has investigated applying Shewhart, CUSUM,
and data collection technology, however, result in the common
EWMA, and other types of control charts to the residuals.
scenario where large amounts of highly autocorrelated in-process
In practice, the ARMA model must always be estimated from
measurement data are available for process monitoring and diag-
nosis 共Montgomery and Woodall 关1兴兲. It is well known that the process data. Let 兵 ␾ˆ i 其 i⫽1
p
, 兵 ␪ˆ i 其 i⫽1
q
, and ␴ˆ 2a be estimates of the
in-control average run length 共ARL兲 of traditional SPC methods ARMA parameters, and define ⌽̂(B) and ⌰ ˆ (B) to be the AR and
like CUSUM, X̄, and EWMA charts may be much shorter than MA polynomials constructed from the estimated parameters. For
intended when data are autocorrelated 共Johnson and Bagshaw 关2兴; simplicity, it is assumed throughout this paper that p and q are
Vasilopoulos and Stamboulis 关3兴兲. As a result, there has been sig- known. The residuals are generated using the estimated model via
nificant research in recent years on designing SPC procedures
suitable for autocorrelated processes 共see, e.g., Montgomery and ⌽̂ 共 B 兲 ⌽̂ 共 B 兲 ⌰ 共 B 兲
e t⫽ x t⫽ at . (3)
Woodall 关1兴, Lu and Reynolds 关4兴, and the references therein兲. ⌰
ˆ 共B兲 ⌰
ˆ 共 B 兲⌽共 B 兲
Some of the most widely investigated methods for SPC of au-
tocorrelated processes are time series control charts, also referred With estimation errors, the residuals follow the ARMA(p⫹q,p
to as residual-based control charts. Typically, the process data x t 共t ⫹q) model 共3兲 and are no longer i.i.d. Residual autocorrelation
is a time index兲 are assumed to follow an autoregressive moving- due to modeling errors can have large impact on the in-control
average 共ARMA兲 model of the form ARL of residual-based charts, as demonstrated empirically in Ad-
ams and Tseng 关18兴, Apley and Shi 关6兴, and Lu and Reynolds 关4兴.
x t ⫺ ␾ 1 x t⫺1 ⫺ ␾ 2 x t⫺2 ⫺¯⫺ ␾ p x t⫺ p To illustrate why, suppose a residual-based EWMA chart is
used to monitor an ARMA process. The residual-based EWMA
⫽a t ⫺ ␪ 1 a t⫺1 ⫺ ␪ 2 a t⫺2 ⫺¯⫺ ␪ q a t⫺q , (1)
statistic is defined recursively via
where the ␾’s and ␪’s are constant parameters, p is the autoregres-
z t ⫽ 共 1⫺␭ 兲 z t⫺1 ⫹␭e t , (4)
sive 共AR兲 order, q is the moving average 共MA兲 order, and a t is
assumed to be an identically independently distributed 共i.i.d.兲 where 0⬍␭⭐1 is the user-specified EWMA parameter 共Mont-
zero-mean random sequence that follows a normal distribution gomery 关19兴兲. For a specified ␭, the typical EWMA design proce-
with variance ␴ 2a . Using the backward shift operator B, Eq. 共1兲 dure when the model is assumed perfect is to set the control limits
can be written as for z t at 共Lu and Reynolds 关4兴兲

x t⫽
⌰共 B 兲
a ,
⌽共 B 兲 t
(2) ⫾L ␴ Z,0⫽⫾L ␴ˆ a 冑 1⫺ ␯
1⫹ ␯
, (5)

where ⌰(B)⫽1⫺ ␪ 1 B⫺ ␪ 2 B 2 ⫺ . . . ⫺ ␪ q B q and ⌽(B)⫽1⫺ ␾ 1 B where the constant L is chosen to provide a desired in-control
⫺ ␾ 2B 2⫺ . . . ⫺ ␾ pB p. ARL, ␯ ⫽1⫺␭, and ␴ z,0 2
⫽ ␴ˆ 2a (1⫺ ␯ )/(1⫹ ␯ ) is the variance of z t
The basic idea behind residual-based charts is to directly moni- 共in the steady-state, after initial transients have died out兲 under the
tor the residuals 共i.e., the one-step-ahead prediction errors兲 of the assumption that there are no modeling errors. Tables provided in
time series model, generated via e t ⫽⌰ ⫺1 (B)⌽(B)x t . From Eq. Lucas and Saccucci 关20兴 give the L values that result in several
共2兲, it follows that after the initial transients have died out, e t is different in-control ARLs for various choices of ␭.
exactly the i.i.d. sequence a t . Thus, the residuals are uncorrelated, Since with modeling errors the residuals will be autocorrelated,
and traditional control charts can be applied with well understood the actual variance of z t may differ considerably from ␴ z,0 2
. Com-
in-control run length properties. Various research 共see e.g., Alwan bining Eq. 共3兲 and Eq. 共4兲, it follows that z t obeys the ARMA(p
and Roberts 关5兴, Apley and Shi 关6兴, Berthouex, Hunter, and Pall- ⫹q⫹1,p⫹q) model

Contributed by the Manufacturing Engineering Division for publication in the 共 1⫺ ␯ 兲 ⌽̂ 共 B 兲 ⌰ 共 B 兲

JOURNAL OF MANUFACTURING SCIENCE AND ENGINEERING. Manuscript received z t⫽ at . (6)
September 2000; Revised January 2002. Associate Editor: S. J. Hu. 共 1⫺ ␯ B 兲 ⌽̂ 共 B 兲 ⌽ 共 B 兲

Journal of Manufacturing Science and Engineering NOVEMBER 2002, Vol. 124 Õ 891
Copyright © 2002 by ASME
For a given set of parameters and parameter estimates, the vari- 1⫺ ␪ 21
ance of z t may be calculated using the impulse response method ⌺ ␥⫽ : for MA共 1 兲 (9)
共Box et al. 关21兴兲 discussed in Section 3. If the parameter estimates n
are such that the residuals are positively autocorrelated, the vari-
ance of z t will be larger than ␴ z,0
2
and the resulting in-control ARL
will be shorter than desired, possibly by a substantial amount
⌺ ␥⫽ 冋
共 1⫺ ␾ 1 ␪ 1 兲 共 1⫺ ␾ 1 兲共 1⫺ ␾ 1 ␪ 1 兲
2

n 共 ␾ 1 ⫺ ␪ 1 兲 2 共 1⫺ ␾ 21 兲共 1⫺ ␪ 21 兲
共 1⫺ ␾ 21 兲共 1⫺ ␪ 21 兲
共 1⫺ ␪ 21 兲共 1⫺ ␾ 1 ␪ 1 兲
册 :

共Adams and Tseng 关18兴兲. for ARMA共 1,1,兲 (10)

Viewing the parameter estimates as random, the EWMA vari-
ance is itself a random variable. Let ␴ z2 denote the expected value These expressions for ⌺ ␥ provide the model uncertainty represen-
of the EWMA variance, where the expectation is with respect to tation that will be used when determining the EWMA variance ␴ z2
the distribution of the random parameter estimates. It will be in Section 4.
shown that ␴ z2 is a function of the covariance matrix of 兵 ␾ ˆ i 其 i⫽1
p

and 兵 ␪ i 其 i⫽1 , an estimate of which is typically available as a mea-

ˆ q
3 The Effects of Model Uncertainty on the EWMA
sure of model uncertainty when ARMA parameters are estimated
共Box et al. 关21兴兲. In analogy with 共5兲, the basic idea of this paper
Variance for General ARMA Processes
is to use From Eq. 共6兲, the variance of z t clearly depend on ␥ˆ . Denote
the conditional variance of z t , given ␥ˆ , as ␴ z2兩 ␥ˆ ⫽E 关 z t2 兩 ␥ˆ 兴 , where
⫾L ␴ z , (7)
E 关 • 兴 is the expectation operator. The unconditional variance of z t
as control limits for the residual-based EWMA. As will be shown, can be written as
␴ z2 is always larger than ␴ z,0
2
because of the uncertainty involved ␴ z2 ⫽E 关 z t2 兴 ⫽E 关 E 关 z t2 兩 ␥ˆ 兴兴 ⫽E 关 ␴ z2兩 ␥ˆ 兴
in estimating ARMA models. Consequently, the control limits 共7兲
are wider than the traditional control limits 共5兲 that are used when The strategy for finding ␴ z2兩 ␥ˆ and ␴ z2 is to linearize the model 共6兲
models are assumed perfect. In this sense, the proposed method about ⌽̂(B)⫽⌽(B) and ⌰ ˆ (B)⫽⌰(B). Linearizing gives
safeguards against an undesirably short in-control ARL due to
ARMA modeling errors. Although the sensitivity of residual- 共 1⫺ ␯ 兲 ⌽̂ 共 B 兲 ⌰ 共 B 兲
based charts to modeling errors is well known, the proposed
method is the first that incorporates model uncertainty information 共 1⫺ ␯ B 兲 ⌰
ˆ 共 B 兲⌽共 B 兲

冋 册冋 册
into the control chart design to provide such a safeguard.
The majority of the technical content of the paper is devoted to 共 1⫺ ␯ 兲 ⌰ 共 B 兲 ⌽ 共 B 兲 ⫹⌽̃ 共 B 兲
developing an expression for ␴ z2 , to be used in 共7兲. The focus is on ⫽
共 1⫺ ␯ B 兲 ⌽ 共 B 兲 ⌰ 共 B 兲 ⫹⌰
˜ 共B兲
residual-based EWMA charts for first-order ARMA processes, al-

冋 册冋 册
though results for arbitrary-order ARMA processes are presented
in Section 3. The results also apply to the Shewhart individual 共 1⫺ ␯ 兲 ⌰ 共 B 兲 ⌽共 B 兲 ⌽̃ 共 B 兲 ⌽共 B 兲⌰
˜ 共B兲
⬵ ⫹ ⫺
chart, which is a special case of the EWMA chart with ␭⫽1.0. 共 1⫺ ␯ B 兲 ⌽ 共 B 兲 ⌰共 B 兲 ⌰共 B 兲 ⌰ 2共 B 兲

2 Representing Model Uncertainty

⫽ 共 1⫺ ␯ 兲 冋 1
1⫺ ␯ B
⫹
⌽̃ 共 B 兲
共 1⫺ ␯ B 兲 ⌽ 共 B 兲
⫺
⌰
˜ 共B兲
共 1⫺ ␯ B 兲 ⌰ 共 B 兲
册 ,

Suppose x t follows a stable, invertible ARMA(p,q) model 共2兲, (11)

p and q are known, and the model parameters are estimated from
n observations of x t . Define the parameter vector as ␥ where ⌽̃(B)⫽⌽̂(B)⫺⌽(B)⫽⫺ ␾ ˜ 1 B⫺ ␾
˜ 2B 2⫺ . . . ⫺ ␾
˜ p B p , and
⌰
˜ (B)⫽⌰ ˆ (B)⫺⌰(B)⫽⫺˜␪ 1 B⫺˜␪ 2 B 2 ⫺ . . . ⫺˜␪ q B q . ␾ ˜ i⫽ ␾ˆ i⫺ ␾ i
⫽ 关 ␾ 1 ␾ 2 . . . ␾ p ␪ 1 ␪ 2 . . . ␪ q 兴 T . Similarly, define ␥ˆ to be the esti-
mated parameter vector and ˜␥⫽ ␥ˆ ⫺ ␥ to be the error. Most of the and ␪ i ⫽ ␪ˆ i ⫺ ␪ i are defined as the parameter estimation errors.
˜
widely used estimation methods, such as nonlinear least squares Note that the third line of 共11兲 involves the linearization of the
and exact or approximate maximum likelihood methods, produce operator ⌰ ˆ ⫺1 (B)⌽̂(B) that maps the space of infinite random
similar results and are largely based on minimizing the sum of the sequences into itself. This operator can also be viewed as a func-
squares of the model residuals 共Box et al. 关21兴兲. For these meth- tion of the ARMA parameter estimates that is continuously differ-
ods, an approximate expression 共that is asymptotically exact兲 for entiable 共where continuity is with respect to any suitable operator
the covariance matrix of ␥ˆ is 共Box et al. 关21兴兲 norm兲 at ␾ ˆ i ⫽ ␾ i and ␪ˆ i ⫽ ␪ i via the assumption of a stable, invert-
ible ARMA model. Differentiating the first line of 共11兲 with re-
␴ 2a ⫺1 spect to each of the ARMA parameter estimates and then substi-
⌺ ␥⫽ ⌺ ,
n W tuting the definition of ⌽̃(B) and ⌰ ˜ (B) into the result yields the
third line of 共11兲.
where ⌺ W is the covariance matrix of the random vector wt , Let g j , j⫽0, 1, 2, . . . denote the coefficients of the impulse
defined as wt ⫽ 关 u t u t⫺1 . . . u t⫺ p⫹1 v t v t⫺1 . . . v t⫺q⫹1 兴 T . The ran- response function 共Box et al. 关21兴兲 of the term in brackets in Eq.
dom processes u t and v t are defined via 共11兲, i.e.,

u t⫽
1
a , and
⌽共 B 兲 t 冋 1
⫹
⌽̃ 共 B 兲
⫺
⌰˜ 共B兲
1⫺ ␯ B 共 1⫺ ␯ B 兲 ⌽ 共 B 兲 共 1⫺ ␯ B 兲 ⌰ 共 B 兲
⫽
⬁

j⫽0
g jB j. 册 兺
1 (12)
v t ⫽⫺ a .
⌰共 B 兲 t The linearized approximation of z t becomes
It can be shown 共Box et al. 关21兴兲 that for the special cases of
AR共1兲, MA共1兲, and ARMA共1,1兲 processes, the covariance matri-
ces are
z t ⬵ 共 1⫺ ␯ 兲 冋兺 册⬁

j⫽0
g j B j a t ⫽ 共 1⫺ ␯ 兲 兺ga
j⫽0
⬁

j t⫺ j ,

where the impulse response coefficients are random variables that

1⫺ ␾ 21 depend on the random ˜␥, in addition to ␥ and ␯. The conditional
⌺ ␥⫽ : for AR共 1 兲 (8)
n and unconditional variance of z t become

892 Õ Vol. 124, NOVEMBER 2002 Transactions of the ASME

 ⬁ Table 1 EWMA control limits with and without consideration
␴ z2兩 ␥ˆ ⬵ 共 1⫺ ␯ 兲 2 ␴ 2a 兺
j⫽0
g 2j , and of model uncertainty, for various ␭, n ␾
ˆ , and ␪ˆ

␴ z2 ⬵E 关 ␴ z2兩 ␥ˆ 兴 ⫽ 共 1⫺ ␯ 兲 2 ␴ 2a 兺 E关g 兴.
j⫽0
2
j (13)

Since ⌽̃(B) and ⌰ ˜ (B) only appear in the numerators of the

terms in Eq. 共12兲, each coefficient g j will be a linear function of
˜␥. Consequently, each g 2j will be a quadratic function of ˜␥, so that
Eq. 共13兲 could be expressed solely in terms of ␯, ␴ 2a , ␥, and ⌺ ␥ .
Appendix B describes a numerical procedure for calculating ␴ z2
for arbitrary order ARMA models. Much of the remainder of this
paper focuses on the special cases of AR共1兲, MA共1兲, and
ARMA共1, 1兲 processes, for which relatively simple closed-form
expressions for ␴ z2 can be derived.

4 Results for First-Order ARMA Processes

For ARMA共1, 1兲 processes, Eq. 共12兲 becomes

冋 1
⫺
␾
˜B
⫹
˜␪ B
1⫺ ␯ B 共 1⫺ ␯ B 兲共 1⫺ ␾ B 兲 共 1⫺ ␯ B 兲共 1⫺ ␪ B 兲
⫽
⬁

j⫽0
g jB j, 册 兺
(14)
where the subscript on the ARMA parameters has been dropped
for convenience. In Appendix A it is shown that after finding the
impulse response coefficients for Eq. 共14兲, the EWMA variance
for ARMA共1, 1兲 processes becomes

␴ z2 ⬵ ␴ 2a 冉 冊冋
1⫺ ␯
1⫹ ␯
1⫹
⌺␾
冉
1⫹ ␯ ␾
1⫺ ␾ 2 1⫺ ␯ ␾
⫹ 冊
⌺ ␪ 1⫹ ␯ ␪
1⫺ ␪ 2 1⫺ ␯ ␪ 冉 冊
⫺ 冉
2⌺ ␾␪ 1⫺ ␯ 2 ␾␪
1⫺ ␾␪ 共 1⫺ ␯ ␾ 兲共 1⫺ ␯ ␪ 兲 冊册 . (15) ␴ z2 ⬵ ␴ 2a 冉 冊冋
1⫺ ␯
1⫹ ␯
1⫹
1⫹ ␯ ␪
n 共 1⫺ ␯ ␪ 兲册. (18)

⌺ ␾ , ⌺ ␪ , and ⌺ ␾␪ denote the variance of ␾ˆ , the variance of ␪ˆ , Since the true parameters are unknown, these expressions for
and the covariance between ␾ ˆ and ␪ˆ , respectively, and are the ␴ z2 cannot be directly used in the control limits ⫾L ␴ z from 共7兲.
elements of the matrix ⌺ ␥ . The recommended procedure is to substitute the parameter esti-
Substituting ⌺ ␥ from Eq. 共10兲 into Eq. 共15兲 gives mates for their true values in Eq. 共16兲, 共17兲, or 共18兲. L can be

冉 冊冋
selected using existing methods 共e.g., Table 4 of Lucas and Sac-
1⫺ ␯ 共 1⫺ ␾␪ 兲 2 共 1⫹ ␯ ␾ 兲 共 1⫺ ␾␪ 兲 2 共 1⫹ ␯ ␪ 兲 cucci 关20兴 or the Markov chain approach described in Lu and
␴ z2 ⬵ ␴ 2a 1⫹ ⫹
1⫹ ␯ n 共 ␾ ⫺ ␪ 兲 2 共 1⫺ ␯ ␾ 兲 n 共 ␾ ⫺ ␪ 兲 2 共 1⫺ ␯ ␪ 兲 Reynolds 关4兴兲, based on ␭ and the desired in-control ARL. By

册
inspection of Eq. 共16兲 through 共18兲, it is clear that ␴ z2 is strictly
2 共 1⫺ ␾ 2 兲共 1⫺ ␪ 2 兲共 1⫺ ␯ 2 ␾␪ 兲
⫺ . larger than the EWMA variance ␴ z,0 2
⫽ ␴ 2a (1⫺ ␯ )/(1⫹ ␯ ) that re-
n 共 ␾ ⫺ ␪ 兲 2 共 1⫺ ␯ ␾ 兲共 1⫺ ␯ ␪ 兲 sults when the model is assumed perfect. Eq. 共B.2兲 in Appendix B
Since 2(1⫺ ␯ 2 ␾␪ )(1⫺ ␯ ␾ ) ⫺1 (1⫺ ␯ ␪ ) ⫺1 ⫽(1⫹ ␯ ␾ )(1⫺ ␯ ␾ ) ⫺1 demonstrates that this is also true for higher-order ARMA models.
⫹(1⫹ ␯ ␪ )(1⫺ ␯ ␪ ) ⫺1 , the term in brackets simplifies to Consequently, the end result of considering model uncertainty in

冋 冉 冊 册
the design procedure is to use wider control limits than in the
1⫹ ␯ ␾ 1⫹ ␯ ␪ 共 1⫺ ␾␪ 兲 2 ⫺ 共 1⫺ ␾ 2 兲共 1⫺ ␪ 2 兲 standard approach.
1⫹ ⫹
1⫺ ␯ ␾ 1⫺ ␯ ␪ n共 ␾⫺␪ 兲2 The extent to which the control limits are widened depends on

冋 冉 冊册
the level of model uncertainty. Table 1 compares the standard
1 1⫹ ␯ ␾ 1⫹ ␯ ␪ (⫾L ␴ z,0) and modified (⫾L ␴ z ) control limits for various ␭, n, ␾ ˆ,
⫽ 1⫹ ⫹
n 1⫺ ␯ ␾ 1⫺ ␯ ␪ and ␪ˆ for ARMA共1, 1兲 processes. Only positive values of the
ARMA parameters were considered, since negative ␾ values re-
Thus, the EWMA variance for ARMA共1, 1兲 processes becomes
sult in negative lag one autocorrelation, which is not commonly

␴ z2 ⬵ ␴ 2a 冉 冊冋
1⫺ ␯
1⫹ ␯
1⫹
1⫹ ␯ ␾
⫹
1⫹ ␯ ␪
n 共 1⫺ ␯ ␾ 兲 n 共 1⫺ ␯ ␪ 兲 册 (16)
encountered in industrial processes 共Box, et al. 关21兴兲. For ␭
⫽0.05 and ␭⫽0.1, the L values that provide a desired in-control
ARL of 500 are L⫽2.616 and L⫽2.814, respectively 共Lucas and
Note that Eq. 共10兲, and thus Eq. 共16兲, is not valid for ␾ ⫽ ␪ . Saccucci 关20兴兲. As sample size n decreases, model uncertainty
This is a trivial case, however, since an ARMA共1, 1兲 model with increases, and the control limits are widened to a greater extent.
␾ ⫽ ␪ is equivalent to x t ⫽a t . This is more apparent from Fig. 1, which shows the relative in-
For AR共1兲 and MA共1兲 processes, the results are similar. Substi- crease in the control limit width as a function of n and ␭, for four
tuting Eq. 共8兲 into Eq. 共15兲 with ⌺ ␪ ,⫽⌺ ␾␪ ⫽ ␪ ⫽0, the EWMA different combinations of ␾ ˆ and ␪ˆ . The relative increase r is de-
variance for AR共1兲 processes becomes fined as r⫽( ␴ z ⫺ ␴ z,0)/ ␴ z,0 . The relative increase depends not

␴ z2 ⬵ ␴ 2a 冉 冊冋
1⫺ ␯
1⫹ ␯
1⫹
1⫹ ␯ ␾
n 共 1⫺ ␯ ␾ 兲
. 册 (17)
only on n, but also on ␭ and the ARMA parameters. For n
⫽100, ␾ ˆ ⫽0.95, and ␪ˆ ⫽0.70, for example, the relative increase is
15.5% when ␭⫽0.02, but only 3.9% when ␭⫽0.30. For the same
Substituting Eq. 共9兲 into Eq. 共15兲 with ⌺ ␾ , ⫽⌺ ␾␪ ⫽ ␾ ⫽0, the n⫽100 but with ␾ ˆ ⫽0.80 and ␪ˆ ⫽0.40, the relative increase be-
EWMA variance for MA共1兲 processes becomes comes 5.1% when ␭⫽0.02 and 2.6% when ␭⫽0.30.

Journal of Manufacturing Science and Engineering NOVEMBER 2002, Vol. 124 Õ 893
 Fig. 1 Relative increase r in the EWMA control limits when model uncertainty is considered,
for various ␭, n , ␾
ˆ , and ␪ˆ

One question of interest is how large must n be before the be 1.0 for convenience. 75 additional observations from the pro-
relative increase can be considered insignificant, the answer to cess were simulated and used to obtain the parameter estimates
which depends on ␭ and the ARMA parameters. For ␭⫽0.02, ␾ ˆ ␾ˆ ⫽0.909, ␪ˆ ⫽0.652, and ␴ˆ 2a ⫽1.007.
⫽0.95, and ␪ˆ ⫽0.7, the relative increase does not drop below If ␭⫽0.05 ( ␯ ⫽0.95) and a desired in-control ARL of 500 are
共say兲 5% until n⬎325. In contrast, for ␭⫽0.30 and the same chosen, the appropriate L value is 2.616 共from Table 4 of Lucas
ARMA parameters, the relative increase drops below 5% when and Saccucci 关20兴兲. From Eq. 共16兲 with the unknown parameters
n⬎77. For the limiting case of a Shewhart individual chart (␭ replaced by their estimates, the EWMA variance is

冉 冊冋 册
⫽1.0), Eq. 共16兲 implies the relative increase is r⫽(1⫹2/n) 1/2
⫺1⬵1/n for all combinations of ␾ ˆ and ␪ˆ . Consequently, for the 1⫺ ␯ 1⫹ ␯ ␾ˆ 1⫹ ␯ ␪ˆ
␴ z2 ⬵ ␴ˆ 2a 1⫹ ⫹ ⫽0.0320,
Shewhart individual chart the relative increase drops below 5% 1⫹ ␯ n 共 1⫺ ␯ ␾
ˆ 兲 n 共 1⫺ ␯ ␪ˆ 兲
when n⬎20 and drops below 1% when n⬎100.
Caution should be used when deciding whether the relative in- and the modified control limits become ⫾L ␴ z ⫽⫾0.468. In con-
crease is insignificant, since a 5% change in the control limit trast, since ␴ z,0
2
⫽ ␴ˆ 2a (1⫺ ␯ )/(1⫹ ␯ )⫽0.0258, the standard control
width can have a large effect on the ARL. Consider a Shewhart limits are ⫾L ␴ z,0⫽⫾0.420. Thus, to account for model uncer-
individual chart on i.i.d. data x t ⫽a t 共with no ARMA modeling or tainty the control limits are widened by 11.4%. This has signifi-
modeling errors兲. Control limits of ⫾3.09␴ a result in an in- cant impact on the in-control ARL. Monte Carlo simulation with
control ARL of 500 共Montgomery 关19兴兲. Control limits of 10,000 replicates reveals that the actual in-control ARL is only
⫾2.94␴ a , which are 5% narrower, result in a substantially lower 237 using the traditional ⫾0.420 control limits. In contrast, when
in-control ARL of 305. Control limits of ⫾3.06␴ a , which are 1% the modified ⫾0.468 control limits are used, the actual in-control
narrower, result in an in-control ARL of 452. ARL is 445, much closer to the desired ARL of 500.
As mentioned above, widening the control limits will also in-
crease the out-of-control ARL when a mean shift does occur. To
5 Examples and Discussion
investigate the extent to which the out-of-control ARL is in-
The obvious drawback of widening the control limits to safe- creased, mean shifts of various magnitude ␮ were added to the
guard against an undesirably low in-control ARL is that the out- process x t on the initial observation, and Monte Carlo simulation
of-control ARL for any size mean shift will inevitably be in- was conducted to estimate the out-of-control ARLs using the stan-
creased. Since the purpose of control charts is to detect faults as dard and modified control limits. 10,000 replicates were con-
quickly as possible, increases in the out-of-control ARL are unde- ducted for each case. The results are summarized in Table 2. The
sirable. The following example illustrates this tradeoff, as well as ARL values for ␮ ⫽0 correspond to the in-control ARL. Although
the proposed design procedure for widening the control limits. the in-control ARL is almost doubled when the modified control
Consider the concentration readings from the chemical production limits are used, the out of control ARL is only slightly larger for
process represented by series A from Box, et al. 关21兴. Based on moderate to large mean shifts. For example, when ␮ ⫽3.0 the
these data, the estimated process model is 共see Box et al. 关21兴 for out-of-control ARLs using the modified and standard control lim-
details兲 x t ⫺0.87x t⫺1 ⫽a t ⫺0.48a t⫺1 , where the in-control mean its are 8.59 and 6.85, respectively.
has already been subtracted out of the data. For illustrative pur- As discussed in Section 4, the Shewhart individual chart (␭
poses, it will be assumed that this is the true model. ␴ 2a is taken to ⫽1.0) is much less affected by model uncertainty, to the point that

894 Õ Vol. 124, NOVEMBER 2002 Transactions of the ASME

Table 2 ARL comparison of the standard EWMA, modified vidual chart with standard control limits (⫾L ␴ˆ a with L⫽3.09),
EWMA, and Shewhart individual chart for the ARMA„1,1… ex- indicates that this is not the case. Even with widened control
ample in Section 5. The in-control ARL corresponds to a mean limits, the EWMA still detects small to moderate size mean shifts
shift magnitude ␮ Ä0. Control limits for the charts are shown in
much faster that the Shewhart chart. The Shewhart chart only
parentheses.
outperforms the EWMA when ␮ is roughly 4.5 or larger. If one
were primarily interested in mean shifts this large, then a Sh-
ewhart individual chart would be the best option. For smaller size
mean shifts, the EWMA with modified control limits is recom-
mended. Note that the in-control ARL for the Shewhart chart
共450兲 is slightly less than the desired 500. This is because the
Shewhart ARL is still affected by modeling errors, albeit by a
much lesser extent than the EWMA.
Figure 2 illustrates the differences between the standard
EWMA, the modified EWMA, and the Shewhart chart with a
typical set of observations from the Monte Carlo simulation. A
mean shift of magnitude ␮ ⫽2.5 was added to the original process
x t on observation number 11. Note that the resulting mean shift in
the residuals, which are obtained by filtering x t via 共3兲, will not
simply be a constant mean shift of magnitude ␮. Rather, the mean
shift in the residuals will be the time-varying function
⌽̂(B)⌰ ˆ ⫺1 (B) ␮ t , where ␮ t denotes a step function of magnitude
␮ occurring at time step 11. The time-varying mean of the residu-
als was referred to as the fault signature in Apley and Shi 关6兴 and
the change pattern in Hu and Roan 关22兴, to which the reader is
widened control limits may not be necessary. Consequently, since referred for more details. The mean of the residuals gradually
the drawback of widening the EWMA control limits is that the decays to the steady-state value ␮ (1⫺ ␾ ˆ )/(1⫺ ␪ˆ )⫽0.65. This can
out-of-control ARL is increased somewhat, one may speculate that be seen in the top panel of Fig. 2, which shows the residuals,
a Shewhart individual control chart with standard control limits along with their mean function and the control limits for a Sh-
would be preferable to an EWMA chart with widened control ewhart individual chart. In this case, none of the first 50 residuals
limits. Table 2, which also shows the ARLs for a Shewhart indi- exceeded the Shewhart control limits. The bottom panel of Fig. 2

Fig. 2 Residuals e t and EWMA statistic z t for the ARMA„1,1… example in Section 5 with a
mean shift of magnitude ␮ Ä2.5 occurring at observation number 11

Journal of Manufacturing Science and Engineering NOVEMBER 2002, Vol. 124 Õ 895
 Fig. 3 Relative increase r in the EWMA control limits for p cascaded AR„1… processes, for
ˆ . Sample size is n Ä100 for all cases
various ␭, p , and ␾

shows the EWMA statistic z t , together with the standard and ries process models are estimated with uncertainty. The end result
modified EWMA control limits. The EWMA statistic hovers is that the control limits are widened by an amount that depends
around zero until the mean shift occurs at observation 11, at which on the level of model uncertainty, as well as on the choice of
point it gradually climbs towards the upper control limit. The EWMA parameter ␭. The design method is a relatively straight-
standard EWMA control limit is crossed at observation 21, forward modification of existing methods, in which time series
whereas it takes one additional observation for z t to cross the models are assumed perfect. The focus has been on EWMA charts
modified EWMA control limit. applied to first-order ARMA processes, although results for arbi-
The preceding discussions centered around first-order ARMA trary order ARMA processes have also been presented. For the
processes. It is reasonable to speculate that since higher-order pro- case of first-order ARMA processes, the only information that is
cesses have more parameters that must be estimated, parameter needed are the parameter estimates and the number of observa-
uncertainty will generally be larger and have larger effect on the tions from which the estimates were obtained. For the case of
EWMA variance. This is illustrated in the following example, in higher-order ARMA processes the parameter covariance matrix is
which the process is AR(p) and results from a series of p cas- needed, and numerical methods must be used to determine the
caded AR共1兲 subprocesses. In other words, the output from the i th control limit width.
AR共1兲 subprocess acts as the input to the (i⫹1) st subprocess. The The extent to which the control limits are widened depends on
input to the first subprocess is assumed to be white noise, and the a number of factors. When sample size and/or the EWMA param-
output of the p th process is the autocorrelated process x t to be eter are large, the relative increase in the control limit width may
monitored. This type of cascaded process is common in the be quite small. In this case, the standard EWMA control limits
chemical and other continuous flow process industries, and is dis- should provide an acceptable in-control ARL and it would not be
cussed in more detail in English et al. 关10兴. For illustrative sim- necessary to use the modified control limits. In situations when
plicity, we consider the scenario where each of the p subprocesses the sample size is small 共e.g., n⬍100) and the EWMA parameter
has the same 共estimated兲 AR共1兲 parameter ␾ ˆ , so that the cascaded is small 共e.g., ␭⬍0.05), however, the relative increase in the con-
AR(p) process is x t ⫽(1⫺ ␾ˆ B) ⫺ p a t . Figure 3 shows the relative trol limit width may be substantial. In these situations, using the
increase r in the EWMA control limit width as a function of ␭, for widened control limits provides some level of protection against
an unacceptably short ARL due to modeling errors.
various p and ␾ ˆ . The sample size was n⫽100 for all cases, and
The obvious drawback of the proposed method is that widening
the method described in Appendix B was used to calculate ␴ z for
the control limits also increases the out-of-control ARLs, resulting
the control limits. From Fig. 3 it is clear that higher-order cas-
in slower detection of mean shifts. In the example considered in
caded processes require a larger relative increase in the control
Section 5 the control limits were widened by 11%, resulting in a
limit width. For ␾ ˆ ⫽0.95 and ␭⫽0.05, for example, the relative
substantial increase in the in-control ARL 共from 237 to 445兲. The
increase is 9.3% for p⫽1 and 40.6% for p⫽5. The relative in- resulting increase in the out-of-control ARL for moderate to large
crease also depends strongly on ␾ ˆ . For ␾ˆ ⫽0.5 and ␭⫽0.05, the
mean shifts was much less dramatic 共from 6.85 to 8.59 for a mean
relative increase drops to 1.4% for p⫽1 and 6.8% for p⫽5. shift of size 3 ␴ a ). In this situation, the benefits of using widened
control limits would most likely outweigh the disadvantages. Sup-
6. Conclusions pose, on the other hand, there had been much larger model uncer-
A method has been presented for designing residual-based tainty and the design procedure had called for even wider control
EWMA control charts for autocorrelated processes when time se- limits. The use of excessively widened control limits should be

896 Õ Vol. 124, NOVEMBER 2002 Transactions of the ASME

considered with caution, as the adverse impact on the out-of- ⬁
⌺␾
control ARLs may be unacceptable. In this situation, a more ef-
fective approach may be to collect more data to improve the pa- 共 ␯ ⫺ ␾ 兲2 兺 共␯
j⫽0
2j
⫺2 ␯ j ␾ j ⫹ ␾ 2 j 兲

冋 册
rameter estimates and reduce model uncertainty. Equation 共16兲
could be used to provide approximate guidelines for how large a ⌺␾ 2 共 1⫺ ␯ 2 兲 1⫺ ␯ 2
⫽ 1⫺ ⫹ ,
sample size is required to have control limits that are widened by 共 ␯ ⫺ ␾ 兲 共 1⫺ ␯ 兲
2 2
1⫺ ␯ ␾ 1⫺ ␾ 2
not more than 共say兲 5%. (A5)
As a final comment, we point out that the impact of modeling
⬁
errors is not unique to residual-based charts. For example, if a ⌺␪
CUSUM or X̄ chart is applied to the original autocorrelated pro- 共 ␯⫺␪ 兲2 兺 共␯
j⫽0
2j
⫺2 ␯ j ␪ j ⫹ ␪ 2 j 兲
cess x t , the ARL approximations of Johnson and Bagshaw 关2兴 or
Vasilopoulos and Stamboulis 关3兴 could be used to suitably widen
the control limits. This, however, requires that the parameters of
an ARMA process model are known. If the parameters are inac-
⫽
⌺␪
共 ␯ ⫺ ␪ 兲 共 1⫺ ␯ 兲
2 2 1⫺
1⫺ ␯ ␪
⫹ 冋
2 共 1⫺ ␯ 2 兲 1⫺ ␯ 2
1⫺ ␪ 2
, and 册
curate, the resulting modified control limits will also fail to pro- (A6)
vide the desired ARL. Similar arguments hold if the method of ⬁
Zhang 关23兴 is used to design an EWMA on x t . Although an 2⌺ ␾␪
ARMA process model is not explicitly required in Zhang 关23兴, 共 ␯ ⫺ ␾ 兲共 ␯ ⫺ ␪ 兲 兺 共␯
j⫽0
2j
⫺␯ j␾ j⫺␯ j␪ j⫹␾ j␪ j 兲
accurate knowledge of the autocorrelation function of x t is.
⫽
2⌺ ␾␪
共 ␯ ⫺ ␾ 兲共 ␯ ⫺ ␪ 兲共 1⫺ ␯ 兲
2 1⫺
1⫺ ␯ 2 1⫺ ␯ 2
⫺ ⫹
1⫺ ␯ 2
1⫺ ␯ ␾ 1⫺ ␯ ␪ 1⫺ ␾␪
. 冋 册
Appendix A (A7)
After algebraic manipulation, the terms in brackets in Eq. 共A5兲
This appendix derives the approximate expression for ␴ z2 in Eq.
through Eq. 共A7兲 reduce to
共15兲 for ARMA共1,1兲 processes. The impulse response g j in Eq.
共14兲 is the sum of the impulse responses for the three transfer
functions in Eq. 共14兲. That is, g j ⫽g 0,j ⫺ ␾
⫺␾
˜ Bg ␾ , j ⫹˜␪ Bg ␪ , j ⫽g 0,j
˜ g ␾ , j⫺1 ⫹˜␪ g ␪ , j⫺1 , where g 0,j , g ␾ , j , and g ␪ , j are defined via
冋 1⫺
2 共 1⫺ ␯ 2 兲 1⫺ ␯ 2
1⫺ ␯ ␾
⫹
1⫺ ␾ 2 ⫽
共 1⫹ ␯ ␾ 兲共 ␯ ⫺ ␾ 兲 2
共 1⫺ ␯ ␾ 兲共 1⫺ ␾ 2 兲
册
, (A8)

⬁
1
⬁ 冋 1⫺
2 共 1⫺ ␯ 2 兲 1⫺ ␯ 2
1⫺ ␯ ␪
⫹
1⫺ ␪ 2 ⫽ 册
共 1⫹ ␯ ␪ 兲共 ␯ ⫺ ␪ 兲 2
共 1⫺ ␯ ␪ 兲共 1⫺ ␪ 2 兲
, and (A9)
兺 g 0,j B j ⫽ ⫽ 兺
␯ jB j,
冋 册
(A1)
j⫽0 1⫺ ␯ B j⫽0 1⫺ ␯ 2 1⫺ ␯ 2 1⫺ ␯ 2 共 1⫺ ␯ 2 ␾␪ 兲共 ␯ ⫺ ␾ 兲共 ␯ ⫺ ␪ 兲
1⫺ ⫺ ⫹ ⫽ .
⬁ 1⫺ ␯ ␾ 1⫺ ␯ ␪ 1⫺ ␾␪ 共 1⫺ ␯ ␾ 兲共 1⫺ ␯ ␪ 兲共 1⫺ ␾␪ 兲
1
兺
(A10)
g ␾, jB j⫽
j⫽0 共 1⫺ ␯ B 兲共 1⫺ ␾ B 兲 Substituting Eq. 共A8兲 through Eq. 共A10兲 into Eq. 共A5兲 through Eq.
⬁ 共A7兲, it follows that

再 冋 册 冋 册
1
⫽ 兺 共␯ j⫹1
⫺ ␾ j⫹1 兲 B j , and (A2) ⬁
1 ⌺ ␾ 1⫹ ␯ ␾ ⌺ ␪ 1⫹ ␯ ␪
␯⫺␾ j⫽0
兺 E 关 g 兴 ⫽ 1⫺ ␯
j⫽0
2
j 2 1⫹
1⫺ ␾ 2 1⫺ ␯ ␾
⫹
1⫺ ␪ 2 1⫺ ␯ ␪

冋 册冎
⬁ ⬁
1 1 共 1⫺ ␯ 2 ␾␪ 兲
兺g
j⫽0
␪, jB
j
⫽ ⫽
共 1⫺ ␯ B 兲共 1⫺ ␪ B 兲 ␯ ⫺ ␪ 兺 共␯
j⫽0
j⫹1
⫺ ␪ j⫹1 兲 B j . ⫺
2⌺ ␾␪
1⫺ ␾␪ 共 1⫺ ␯ ␾ 兲共 1⫺ ␯ ␪ 兲
. (A11)
(A3)
Substituting Eq. 共A11兲 into Eq. 共13兲 gives the result Eq. 共15兲.
Equation 共A2兲 follows from equation 共3.1.18兲 of Pandit and Wu
关24兴, by noting that (1⫺ ␯ B) ⫺1 (1⫺ ␾ B) ⫺1 is the transfer function Appendix B
for an AR共2兲 process. Equation 共A3兲 follows similarly. Equating
coefficients on the left and right sides of Eq. 共A1兲 through Eq. This appendix describes a numerical procedure for calculating
共A3兲 to obtain g 0,j , g ␾ , j , and g ␪ , j , and substituting these into ␴ z2 for arbitrary order ARMA models. Similarly to what was done
g j ⫽g 0,j ⫺ ␾
˜ g ␾ , j⫺1 ⫹˜␪ g ␪ , j⫺1 , gives in Appendix A, the impulse response g j in Eq. 共13兲 can be written
as
␾
˜ 共 ␯ j ⫺ ␾ j 兲 ˜␪ 共 ␯ j ⫺ ␪ j 兲
g j ⫽g 0,j ⫹⌽̃ 共 B 兲 g ␾ , j ⫺⌰
˜ 共 B 兲g ␪, j
g j⫽␯ j⫺ ⫹ , and
␯⫺␾ ␯⫺␪
⫽g 0,j ⫺ 共 ␾
˜ 1 B⫹ ␾
˜ 2B 2⫹ . . . ⫹ ␾
˜ p B p 兲 g ␾ , j ⫹ 共 ˜␪ 1 B⫹˜␪ 2 B 2
⌺␾
E 关 g 2j 兴 ⫽ ␯ 2 j ⫹ 共 ␯ 2 j ⫺2 ␯ j ␾ j ⫹ ␾ 2 j 兲 ⫹ . . . ⫹˜␪ q B q 兲 g ␪ , j ,
共 ␯⫺␾ 兲2
⌺␪ ⫽g 0,j ⫺ ␾
˜ 1 g ␾ , j⫺1 ⫺ ␾
˜ 2 g ␾ , j⫺2 ⫺ . . . ⫺ ␾
˜ p g ␾ , j⫺ p ⫹˜␪ 1 g ␪ , j⫺1
⫹ 共 ␯ 2 j ⫺2 ␯ j ␪ j ⫹ ␪ 2 j 兲
共 ␯ ⫺ ␪ 兲2 ⫹˜␪ 2 g ␪ , j⫺2 ⫹ . . . ⫹˜␪ q g ␪ , j⫺q ,
2⌺ ␾␪ ⫽g 0,j ⫹gTj ˜␥, (B1)
⫺ 共 ␯ 2 j ⫺ ␯ j ␾ j ⫺ ␯ j ␪ j ⫹ ␾ j ␪ j 兲 . (A4)
共 ␯ ⫺ ␾ 兲共 ␯ ⫺ ␪ 兲 where g 0,j , g ␾ , j , and g ␪ , j are defined as the impulse response
Summing Eq. 共A4兲 term by term for j⫽0,1,2, . . . gives coefficients of (1⫺ ␯ B) ⫺1 , (1⫺ ␯ B) ⫺1 ⌽ ⫺1 (B), and (1
⫺ ␯ B) ⫺1 ⌰ ⫺1 (B), respectively. The p⫹q length vector g j
⬁ is defined as g j ⫽ 关 ⫺g ␾ , j⫺1 ⫺g ␾ , j⫺2 . . .
1
兺␯
j⫽0
2j
⫽
1⫺ ␯ 2
, ⫺g ␾ , j⫺ p g ␪ , j⫺1 g ␪ , j⫺2 . . . g ␪ , j⫺q 兴 T , where it is understood that
g ␾ , j ⫽g ␪ , j ⫽0 for j⬍0. Thus, from Eqs. 共13兲 and 共B1兲,

Journal of Manufacturing Science and Engineering NOVEMBER 2002, Vol. 124 Õ 897
 ⬁ ⌺ ␪ ⫽ variance MA parameter estimate
␴ z2 ⬵ 共 1⫺ ␯ 兲 2 ␴ 2a 兺
j⫽0
E 关 g 2j 兴 ⌺ ␾␪ ⫽ covariance between AR and MA parameter estimates
n ⫽ number of observations used to estimate ARMA
⬁ model
⫽ 共 1⫺ ␯ 兲 2 ␴ 2a 兺 E 关共 g
j⫽0
␥兲 2 兴
j,0⫹g j ˜
T ␴ z2兩 ␥ˆ ⫽ conditional EWMA variance, given ␥ˆ
g j ⫽ impulse response function of EWMA model
⬁

⫽ 共 1⫺ ␯ 兲 2 ␴ 2a 兺 E关g
j⫽0
␥⫹gTj ˜␥˜␥T g j 兴
j,0⫹2g j ˜
2 T References
关1兴 Montgomery, D. C., and Woodall, W. H., 1997, ‘‘A Discussion on Statistically-
Based Process Monitoring and Control,’’ J. Quality Tech. 29共2兲, pp. 121–162.
⬁ 关2兴 Johnson, R. A., and Bagshaw, M., 1974, ‘‘The Effect of Serial Correlation on
⫽ 共 1⫺ ␯ 兲 2 ␴ 2a 兺 共␯
j⫽0
2j
⫹gTj ⌺ ␥ g j 兲 the Performance of CUSUM tests,’’ Technometrics, 16共1兲, pp. 103–112.
关3兴 Vasilopoulos, A. V., and Stamboulis, A. P., 1978, ‘‘Modification of Control
Chart Limits in the Presence of Data Correlation,’’ J. Quality Tech., 10共1兲, pp.

冉 冊
⬁ 20–30.
1⫺ ␯ 2
⫽ ␴ ⫹ 共 1⫺ ␯ 兲 2 ␴ 2a
1⫹ ␯ a j⫽0
gTj ⌺ ␥ g j ,兺 (B2) 关4兴 Lu, C. W., and Reynolds, M. R., 1999, ‘‘EWMA Control Charts for Monitor-
ing the Mean of Autocorrelated Processes,’’ J. Quality Tech., 31共2兲, pp. 166 –
188.
where the second to last equality follows from the facts that g 0,j 关5兴 Alwan, L. C., and Roberts, H. V., 1988, ‘‘Time-Series Modeling for Statistical
Process Control,’’ Journal of Business and Economic Statistics, 6共1兲, pp. 87–
⫽ ␯ j and ˜␥ is zero-mean. 95.
Equation 共B2兲 yields a numerical procedure for calculating ␴ z2 , 关6兴 Apley, D. W., and Shi, J., 1999, ‘‘The GLRT for Statistical Process Control of
given the ARMA parameter estimates and their covariance matrix Autocorrelated Processes,’’ IIE Transactions, 31共12兲, pp. 1123–1134.
关7兴 Berthouex, P. M., Hunter, W. G., and Pallesen, L., 1978, ‘‘Monitoring Sewage
⌺ ␥ . To calculate the g j vectors, we must substitute the parameter Treatment Plants: Some Quality Control Aspects,’’ J. Quality Tech., 10, pp.
estimates for their unknown true values when calculating the im- 139–149.
pulse response coefficients g ␾ , j and g ␪ , j ( j⫽0,1,2, . . . ). Numeri- 关8兴 Dooley, K. J., and Kapoor, S. G., 1990, ‘‘An Enhanced Quality Evaluation
cal procedures for calculating the impulse response coefficients System for Continuous Manufacturing Processes,’’ Parts 1 and 2, ASME J.
Eng. Ind., 112, pp. 57– 68.
can be found in Pandit and Wu 关24兴 and Box et al. 关21兴. Since the 关9兴 Dooley, K. J., Kapoor, S. G., Dessouky, M. I., and DeVor, R. E., 1986, ‘‘An
impulse response coefficients decay exponentially to zero for Integrated Quality Systems Approach to Quality and Productivity Improve-
stable, invertible ARMA models, the summation in 共B1兲 can be ment in Continuous Manufacturing Processes,’’ ASME J. Eng. Ind., 108, pp.
truncated to a finite number of terms. ⌺ ␥ is readily available with 322–327.
关10兴 English, J. R., Krishnamurthi, M., and Sastri, T., 1991, ‘‘Quality Monitoring of
most ARMA modeling software. Alternatively, the method dis- Continuous Flow Processes,’’ Computers and Industrial Engineering, 20共2兲,
cussed in Box et al. 关21兴 and Section 2 could be used to calculate pp. 251–260.
⌺␥ . 关11兴 Lin, W. S., and Adams, B. M., 1996, ‘‘Combined Control Charts for Forecast-
Based Monitoring Schemes,’’ J. Quality Tech., 28共3兲, pp. 289–301.
关12兴 Montgomery, D. C., and Mastrangelo, C. M., 1991, ‘‘Some Statistical Process
Nomenclature Control Methods for Autocorrelated Data,’’ J. Quality Tech., 23共3兲, pp. 179–
xt ⫽ process measurement data, t⫽time index 193.
关13兴 Notohardjono, B. D., and Ermer, D. S., 1986, ‘‘Time Series Control Charts for
at ⫽ i.i.d. random noise Correlated and Contaminated Data,’’ ASME J. Eng. Ind., 108, pp. 219–226.
et ⫽ time series residuals 关14兴 Runger, G. C., Willemain, T. R., and Prabhu, S., 1995, ‘‘Average Run Lengths
zt ⫽ EWMA statistic for Cusum Control Charts Applied to Residuals,’’ Commun. Stat: Theory
␾i ⫽ AR parameters Meth., 24共1兲, pp. 273–282.
关15兴 Superville, C. R., and Adams, B. M., 1994, ‘‘An Evaluation of Forecast-Based
␪i ⫽ MA parameters Quality Control Schemes,’’ Commun. Stat: Sim. Comp, 23共3兲, pp. 645– 661.
p ⫽ AR order 关16兴 Vander Wiel, S. A., 1996, ‘‘Monitoring Processes That Wander Using Inte-
q ⫽ MA order grated Moving Average Models,’’ Technometrics, 38共2兲, pp. 139–151.
关17兴 Wardell, D. G., Moskowitz, H., and Plante, R. D., 1994, ‘‘Run-Length Distri-
␴ 2a ⫽ variance of a t butions of Special-Cause Control Charts for Correlated Processes,’’ Techno-
␴ z2 ⫽ variance of z t metrics, 36共1兲, pp. 3–17.
⌽(B) ⫽ AR polynomial 关18兴 Adams, B. M., and Tseng, I. T., 1998, ‘‘Robustness of Forecast-Based Moni-
toring Schemes,’’ J. Quality Tech., 30共4兲, pp. 328 –339.
⌰(B) ⫽ MA polynomial 关19兴 Montgomery, D. C., 1996, Introduction to Statistical Quality Control, 3rd ed.,
B ⫽ time series backshift operator Wiley, New York.
ˆ ⫽ overscore that denotes an estimate of a quantity 关20兴 Lucas, J. M., and Saccucci, M. S., 1990, ‘‘Exponentially Weighted Moving
⬃ ⫽ overscore that denotes an estimation error of a quan- Average Control Schemes: Properties and Enhancements,’’ Technometrics,
32共1兲, pp. 1–12.
tity 关21兴 Box, G., Jenkins, G., and Reinsel, G., 1994, Time Series Analysis, Forecasting,
␭ ⫽ EWMA parameter and Control, 3rd ed., Prentice-Hall, Englewood Cliffs, NJ.
␯ ⫽ 1⫺␭ 关22兴 Hu, S. J., and Roan, C., 1996, ‘‘Change Patterns of Time Series-Based Control
L ⫽ constant factor in EWMA control limits Charts,’’ J. Quality Tech., 28共3兲, pp. 302–312.
关23兴 Zhang, N. F., 1998, ‘‘A Statistical Control Chart for Stationary Process Data,’’
␥ ⫽ ARMA parameter vector Technometrics, 40共1兲, pp. 24 –38.
⌺␥ ⫽ covariance matrix of ARMA parameter estimates 关24兴 Pandit, S. M. and Wu, S. M., 1990, ‘‘Time Series and System Analysis with
⌺␾ ⫽ variance of AR parameter estimate Applications,’’ Krieger, Malabar, FL.

898 Õ Vol. 124, NOVEMBER 2002 Transactions of the ASME