Anda di halaman 1dari 12

IIE Transactions (2005) 37, 1011–1021

Copyright 
C “IIE”

ISSN: 0740-817X print / 1545-8830 online


DOI: 10.1080/07408170500232214

Economic design of EWMA control charts using regions


of maximum and minimum ARL

J. CARLOS GARCÍA-DÍAZ∗ and FRANCISCO APARISI


Applied Statistics, Operations Research and Quality Department, Polytechnic University of Valencia, Camino de Vera S/n,
Building I-3, 46022 Valencia, Spain
E-mail: juagardi@eio.upv.es

Received and accepted October 2004

Nowadays, it is common to find industries that utilize processes that either have a value of the process capability index Cpk larger than
two or are very difficult to adjust. In these cases, the detection of very small shifts may not be of interest due to the possible extra
variability introduced into the process by the detection process. It would be more interesting in these situations to decide what shift
size is important for detection, and to design a chart capable of quickly detecting this shift whilst having a low probability of false
alarms for the shifts that we do not wish to detect. The Exponentially Weighted Moving Average (EWMA) control chart, although
originally developed to successfully detect small shifts, can be designed to cope with these requirements. This paper presents a method
for the economic-statistical design of EWMA charts for control processes, in which the detection of small shifts is not necessary, and
which is, at the same time, effective in detecting important shifts. A genetic algorithm is used to optimize the design. A sensitivity
analysis of the optimal solution is performed to determine the influence of certain factors on the economic model.

1. Introduction and when zt > UCL or zt < LCL, an out-of-control signal


is obtained.
Exponentially Weighted Moving Average (EWMA) In order to measure a shift in the process mean, the pa-
schemes were introduced by Roberts (1959) as an alterna- rameter δ = |µ0 − µ1 |/σ0 is employed, which is the new
tive to the Shewhart control chart (Shewhart, 1931). The out-of-control process mean, µ1 , µ1 = µ0 ± δσ0 . When the
EWMA chart, developed by Lucas and Saccucci (1990), process is in control δ = 0 and when the process is out of
accumulates data from past-samples and is therefore more control δ > 0.
sensitive to small process shifts. Reynolds and Arnold In the statistical design of control charts, the statistical re-
(2001) and also Jones (2002) have studied the possibility quirements of Type I and Type II error probabilities, as well
of improving the EWMA control chart efficiency to detect as the Average Run Length (ARL) and the Average Time
small process shifts. to Signal (ATS) are considered once the chart parameters
Let xt be a single observation from a random variable X are chosen. The economic design of control charts aims to
with mean µ0 and variance σ02 for the in-control process, determine the parameters of the control chart that mini-
i.e., X ≈ N(µ0 , σ0 ). The statistic EWMA is calculated as mize the cost incurred in the production process. Usually,
follows: the cost model includes the cost of sampling, the cost of de-
zt = r xt + (1 − r )zt−1 , fective products and the cost of a search for out-of-control
signals which are, in fact, false alarms.
where z0 = µ0 and 0 < r ≤ 1. To detect the average process The first economic model was formulated by Duncan
shifts, the Upper Control Limit (UCL) and Lower Control (1956) and was designed for the X̄ chart. Duncan devel-
Limit (LCL) of the EWMA chart are calculated as follows: oped a process to determine the optimal parameters that
 1/2 optimize the expected cost per hour of production. The
r
UCL = µ0 + Lσ0 , parameters to be determined are the sampling size n, the
2−r sampling frequency h and the control limits coefficient k.
 1/2
r This work is the basis of all subsequent research carried out
LCL = µ0 − Lσ0 , on economic models.
2−r
The statistical economic design of a control chart ob-
tains those values of the parameters of the sampling plan
∗ that optimize the expected cost function of the “per time
Corresponding author

0740-817X 
C 2005 “IIE”
1012 Garcı́a-Dı́az and Aparisi

unit of operation”, subject to the constraints of a minimum 1. The in-control region [0, A] or region of maximum ARL
ARL in-control value, and a maximum ARL value when value. This region corresponds to a state equivalent to
the process is out of control. Constraints in terms of the the in-control state and to a shift size between δ = 0 and
ATS, which express the ARL in time units, can also be used. δ = A. In this region, there is no intention to detect any
References on economic models in quality control include. shift. A maximum ARL value is sought. If the chart
Montgomery (1980), Ho and Case (1994), Al-Sultan and shows an out-of-control signal, this signal is considered
Rahim (1997) and Wu et al. (2004). to be a false alarm.
The first aim of the present paper is to develop economic 2. The out-of-control region, [B, ∞) or region of minimum
and economic-statistical models that permit the establish- ARL value, corresponding to the shift value δ = B, from
ment of optimal values for the design parameters of EWMA which the maximum power of detection is required. A
control charts with regions of maximum and minimum minimum ARL value is sought.
ARL (Woodall, 1985), minimizing the expected cost func- 3. The indifferent region, (A, B), between δ = A and δ = B.
tion of the operation per time unit. These parameters are: In this region, it is uncertain as to whether or not there
(i) the sample size n; (ii) the sampling frequency h; (iii) the is a detected process shift.
weighting coefficient for the chart r; and (iv) the coefficient
L. The second aim is to carry out a sensitivity analysis for In Fig. 1 a comparison of ARL curves between the follow-
the optimal solution. The model proposed by Montgomery ing charts can be seen: the X̄ Shewhart chart (with an in-
et al. (1995) and Torng et al. (1995) for the EWMA chart is control ARL equal to 500) and two EWMA charts that have
the starting point from which we will achieve the objectives been designed according to different criteria. The “EWMA-
of the present paper. point” is a chart that has been designed following the “tra-
The paper is structured as follows: Section 2 presents ditional” model, based on the earliest detection of a small
charts with a maximum and minimum ARL value. In deviation in the mean, with an in-control ARL equal to 500.
Section 3 an economic model for cost minimization, as In addition, the “EWMA-regions” is an EWMA chart de-
devised by Montgomery et al. (1995) for the standard signed to have very high ARL values in the in-control region
EWMA chart, is adapted to a design based on regions of (A = 0.25), whilst still maintaining ARL values equivalent
maximum and minimum ARL values. Section 4 presents to those presented in the X̄ chart in the out-of-control re-
a study on the new role of false alarms in these charts, gion (B = 1.2). Thus, the EWMA chart can be designed in
from the point of view of the region in which the detec- the “classical” way, so as to be highly effective in the detec-
tion of shifts is not required, obtaining an “average” prob- tion of small shifts. However, as has been seen, it can also be
ability of false alarms. An example is presented in Sec- designed so as to not detect small shifts, considered as being
tion 5. A sensitivity analysis of the optimal solution is unimportant, while maintaining the efficiency for detecting
carried out in Section 6, and in Section 7 we present our shifts of a greater magnitude, considered to be significant.
conclusions. Woodall’s approach is justified in situations that fre-
quently arise in industry: high capacity processes, difficult
to adjust processes, and processes with a high adjustment
2. Control charts with regions of maximum cost.
and minimum ARL values
1. Manufacturing processes with a high capacity. Processes
that show a high actual capacity are those where, with
From an economic perspective, as later outlined, it is consid-
ered that the average operational cost of the process when
the mean has changed, is proportional to the number of
false alarms and, from a statistical viewpoint, the number
of false alarms can be an extra source of variability for the
process, as well as contributing to a lack of motivation for
the personnel in charge of control (Deming ,1982; Woodall,
1985).
Woodall (1985) studied the statistical design of control
charts and recommended selecting shift sizes whose detec-
tion is important, as a criterion for the design of control
charts. For this, he proposed the definition of three regions:
an in-control region, an indifferent region and an out-of-
control region. These regions are limited by two values, A
and B.
According to the approach of Woodall (1985), the re-
gions of minimum and maximum ARL values are defined
as follows: Fig. 1. Regions of maximum and minimum ARL.
Economic design of EWMA charts 1013

respect to the nominal value, it is of little interest to detect sumes that the process continues during the search for the
small shifts in the mean. assignable cause.
2. Processes difficult to adjust. In some industrial processes Each of the aforementioned cycles, from the in-control
(regardless of economic considerations) for technical or state to the occurrence, detection and elimination of the
physical reasons when the process begins to work outside assignable cause, has a mean time and cost. The assump-
the nominal value its behavior makes it difficult to adjust. tions we have adopted in this work will allow the calculation
Within this group we might find, for example, processes of the average production cost per hour as the ratio of the
where a cutting tool is subject to continuous and gradual cycle’s mean cost and the mean time of the cycle under
wear. Therefore, if a small process shift does not mean consideration.
a significant quality loss in the product, it may be of no We will use the cost model developed by Lorenzen and
interest to detect such small shifts. Vance (1986) and improved upon by Montgomery et al.
3. Processes with a high adjustment cost. In many processes, (1995) in order to develop a model that allows the economic-
any adjustment to enable them to work at the nomi- statistical design of an EWMA chart with regions of max-
nal value based on predetermined specifications requires imum and minimum ARL values.
stopping the process and assessing the condition of the
machine on which the work is carried out. This operation 3.2. Notation
may require a considerable amount of time with a sub-
sequent loss of manufacturing time. Design parameters of the chart

In processes of this kind it may be inappropriate to con- n = the sample size;


sider detecting a difference between the mean value of the h = the interval between samples or sampling
quality characteristic to be controlled and the objective frequency;
nominal value that must be estimated below a threshold r = the exponential weight used;
that we call A. L = the parameter of the EWMA chart.

3. Cost model Process parameters


3.1. Model assumptions λ = the occurrence level of the assignable cause per
time unit;
Several researchers have studied the economic design of l/λ = the mean time in the in-control state between two
EWMA control charts including for example Montgomery assignable causes;
et al. (1995) and Park et al. (2004). Montgomery et al. (1995) T0 = the expected time spent searching for a false
considered the economic design of EWMA charts by ex- alarm;
tending the Lorenzen–Vance flexible cost model (Lorenzen T1 = the expected time required to discover the
and Vance, 1986). We will use the following assumptions. assignable cause (true alarm);
The process starts under an in-control state and therefore T2 = the expected time required to adjust the process
only common variability causes affect the process. The qual- (elimination of the assignable cause);
ity characteristic to be controlled has a probability distri- d1 = 1 if production continues during searches, 0
bution of mean µ0 and a standard deviation σ0 . It is further otherwise;
assumed that only one cause can randomly occur, result- d2 = 1 if production continues during the repair of the
ing in a shift in the mean from µ0 to µ0 + δσ0 where δ is process, 0 otherwise;
the known size of the shift, and σ0 is a constant. The inter- E = the expected time to sample and chart one item;
val between occurrences has an exponential distribution of τ = the mean time of occurrence of an assignable
parameter λ (1/λ is the mean time in the in-control state). cause, if it occurred between the jth and (j + 1)
The control limits of the EWMA charts are given by UCL samples (Duncan 1956):
and LCL, and random samples of size n are taken at fixed
sampling intervals every h hours. When a point exceeds a τ = {1 − (1 + λh) exp(−λh)}/{λ − λ exp(−λh)};
control limit, the process is considered to be out of control,
and a search for the assignable cause is initiated. A cor- h−τ = the expected time between a shift in the process
rective action must be carried out in order to reinstate the (assignable cause) and the next sample;
process to the in-control state. s = the number of samples while in control
After a corrective action, the process returns to the initial (Lorenzen and Vance, 1986):
state of statistical control with the variable at its nominal
value. Two different operational policies are applied dur- s = exp(−λh)/(1 − exp(−λh));
ing the process (Panagos et al., 1985; Lorenzen and Vance,
1986). Policy 1 assumes that the process is stopped during α = the probability of a Type I error, that is, the prob-
the search for the assignable cause, whereas policy 2 pre- ability that the chart signals a false alarm;
1014 Garcı́a-Dı́az and Aparisi

β = the probability of a Type II error, that is, the C0
probability that the chart does not signal that C(n, h, r, L) = + C1 [−τ + nE + h(ARLB )
λ
there has been a shift; sY
δ = the size of the shift in the process mean, measured + d1 T1 + d2 T2 ] + +W
  ARL0
in σ0 units;
ARLo = the expected number of points charted until the a + bn 1
+ − τ + nE + h(ARLB )
control chart signals that the process is out of h λ
 
control, given that the process is in control; 1 (1 − δ1 )sT0
+ d1 T1 + d2 T2 ÷ +
ARLB = the average number of points charted until the λ ARL0
control chart signals, when there is a shift of mag- 
nitude δ = B. − τ + nE + h(ARLB ) + T1 + T2 . (3)

It should be pointed out that both the ARL0 and the ARLB
Cost parameters depend on n, r and L, and that s and τ depend on h.
C0 = the cost of producing non-conforming units If d1 = d2 = 1, then production continues during the
when the process is in control; search for and repair of an assignable cause and if d1 =
C1 = the cost of producing non-conforming units d2 = 0, then, production is stopped during the search and
when the process is out of control; repair. In the case where production only continues during
Y = the cost of investigating false alarms; the search, we obtain d1 = 1 and d2 = 0. The case d1 = 0
W = the cost of locating and repairing an assignable and d2 = 1 does not seem reasonable. The expression given
cause; by Equation (3) depends on 10 cost and time parameters
a = the fixed sampling cost; which describe the process, two state variables which indi-
b = the variable sampling cost per item sampled and cate whether production continues during the search and
evaluated; the elimination of assignable variability causes, a parameter
C = the average cost to be optimized. which measures the size of the shift to be detected δ, and
the four design parameters of the chart n, h, r and L of the
For a production cycle based on the proposed model, the EWMA chart.
average length of the production cycle is given by Equa- The aim is to establish the value of the design parameters
tion (1). of the EWMA control chart, using regions of maximum
and minimum ARL that optimize the mean cost function
1 (1 − d1 )sT0
E(L) = + h(ARLB ) − τ + + T1 + T2 + nE. per hour of production given by Equation (3), for a shift of
λ ARL0 magnitude δ in the mean. However, we can observe that C
(1) depends on the in-control ARL, ARL0 , and on the out-of-
3.3. Cost per production cycle control ARL, ARLB . In this paper, where in-control and
Incurred costs per cycle are due to the production of out-of-control regions are introduced, a detailed analysis
non-conforming units when the process is out of control, of these ARL values will be carried out in the next section.
false alarms, the location and elimination of an assignable
cause, and sampling and inspection. For a production
cycle, the total expected cost can be calculated using 4. Interpreting and calculating the probability
Equation (2) of a false alarm

C0 The probability of a false alarm is “classically” defined as


E(C) = + C1 [−τ + nE + h(ARLB ) + d1 T1 + d2 T2 ]
λ  
the probability that a point falls outside the control limits
sY a + bn when the process is in an in-control state, and is known as
+ +W + α, the probability of a Type I error.
ARL0 h
  Tang et al. (1999), Xie et al. (2001) and Cai et al. (2002)
1
× − τ + nE + h(ARLB ) + d1 T1 + d2 T2 . analyze trended productive processes. This sort of behavior
λ in which the process mean changes involves regular grad-
(2) ual adjustments. In each adjustment, the process mean is
restored to its nominal value, and I is the interval between
3.3.1. Average cost per hour two consecutive adjustments, that is, the number of sam-
Dividing Equation (2) by Equation (1), the average cost per ples taken between two process adjustments. In these sort
production hour is obtained, as given by Equation (3): of processes, since the process mean varies, the probability
  that a value falls out of the control limits changes according
E C to the number of samples, and, therefore, the probability of
C(n, h, r, L) =   ,
E L a false alarm is not constant in the design value.
Economic design of EWMA charts 1015

To counter this, Xie et al. (2001) propose the use of the δ ∈ [0, A], when it is, in fact, outside it, δ ∈ [0, A]. In this
Average Alarm Ratio (AAR). For instance, for the X̄ con- case, it would be useful to obtain a minimum ARL value.
trol chart, we take a sample at instant i, where the average If the values of δ are random values, fluctuating from one
value of the process is the nominal value, and as such, the instant to another in the production process while it remains
in-control state. The probability of a point falling out of in control, we can consider that for each instant, the proba-
the control limits, i.e., the probability of a false alarm, can bility that a value falling outside the control limits, changes
be calculated as αi = P(yi > UCL) + P(yi < LCL), where with the instant considered, and, therefore, the probability
yi is the sampling value of the quality characteristic to of a false alarm is not constant in the design value, since
be controlled. The assumption is made that it is a chart it depends on δ, but instead fluctuates randomly. Thus, we
of single measurements, as in the case of modern, highly can define the random variable S(δ) as the “probability that
automated industrial environments. It should be pointed a point falls outside the control limits when the process re-
out that αi , known here as the punctual alarm ratio, de- mains in control in a region of amplitude A”. The following
pends on the sample number i. Taking the average value step is to determine the mean value of this random variable.
for a cycle of length I, we can calculate the ratio of average The mean value of S(δ) is:
alarms as:
A
E[S(δ)] = s(δ)fs (δ)dδ, (5)

I
0
AAR(I) = αj /I. (4)
j=1 where s(δ) is a value of variable S(δ) and fs (δ) represents the
density probability function of S(δ).
A model for the economic design of control charts for From the viewpoint of a “standard” X̄ chart, to obtain
trended processes is presented by Cai et al. (2002). The aim s(δ), it must be noted that the probability of a false alarm,
of this model is to obtain the optimal adjustment inter- α, refers to the case in which δ = 0, and therefore no shift
val of the process I, minimizing the cost function. In this has occurred. From the viewpoint of a chart with regions
paper, the in-control and out-of-control states are defined. of maximum and minimum ARL, this concept extends to
The probability of a false alarm is estimated using modi- values of shifts where detection is not required, δ ∈ [0, A]. In
fied Type I and Type II errors, as proposed by Tang et al. the standard X̄ chart, assuming normality, the probability
(1999), for processes with trend adjustment, and, finally, of a false alarm is given by:
the average probability of a false alarm, αi , is the average

probability of a false alarm in each sample taken within the α=2 φ(z)dz, (6)
cycle or adjustment interval of the process, equivalent to k

the concept of AAR as presented in Equation (4). where φ(z) is the density probability function of the stan-
dard normal distribution and k = Zα/2 = 3 with a three-
sigma standard design.
4.1. The average probability for an alarm in control charts When we wish to detect a shift in the process mean de-
with regions of maximum and minimum ARL. The case fined by a magnitude δ > 0 with the X̄ chart, we use the
of the X̄ control chart concept of power or efficacy of the standard chart given by
the expression Power = 1 − β where β is the probability of
In accordance with previous considerations, we will, first of
accepting that no shift has taken place when, in actual fact,
all, study, with regard to the X̄ chart, how a probability of an
a shift has occurred, Pa (δ). This probability is determined
“average false alarm” can be determined for an in-control
by the expression:
state corresponding to an in-control region of amplitude √ √
A. Later we will apply this work to the EWMA chart. For Pa (δ) = β = (Zα/2 − δ n) − (−Zα/2 − δ n), (7)
this, let us suppose that the process fluctuates around the
mean, µ0 , randomly producing causes that modify the mean where (·) is the accumulated probability distribution func-
value of the control object characteristic, with a mean that tion of the standard normal distribution and n is the sam-
oscillates from µ0 to µ1 = µ0 ± δσ0 , where δ is the random ple size used. Finally, the probability of detecting that an
shift value in the range δ ∈ [0, A] for A > 0. assignable cause has occurred in the following sample is
We can consider the probability of a false alarm as being determined by Equation (8):
the probability of accepting that a shift has occurred when
−k−δ√n

the process was in control or in the region δ ∈ [0, A] where Power = 1 − β = φ(z)dz + √
(z)dz. (8)
the detection of shifts is not necessary. In this case, it would −∞ k−δ n
be useful for this probability to be as low as possible, and, Thus, to calculate s(δ) we can use the following
as such, produce a maximum ARL value. expressions:
The probability of committing a Type II error is the
probability of accepting that the shift produced is within if δ = 0 ⇒ s(δ) = α,
the region in which shifts are not required to be detected, if δ = 0 ⇒ s(δ) = 1 − β.
1016 Garcı́a-Dı́az and Aparisi

Taking into account Equations (5) to (8), we can calculate In the case of a uniform distribution for S(δ), the mean
for the case δ ∈ (0, A] the mean value of S(δ) as follows: value of S(δ) is determined by the equation:

−k−δ√n

1 A 1
δ ∈ (0, A] ⇒ s(δ) = 1 − β = φ(z)dz E[S(δ)] = dδ. (14)
−∞ A 0 ARLB


+ φ(z)dz, The problem found in this case is that we cannot analytically

k−δ n find an equation to determine E[S(δ)] due to the fact that

A
A 

−k−δ n it depends on the ARLB . Since the values plotted on the
E[S(δ)] = s(δ) · fs (δ) · dδ = φ(z)dz chart are not independent, in order to determine the ARL
−∞


0
 0
of EWMA control charts, three methods can be used: (i)
+ φ(z)dz fs (δ) dδ. (9) simulation; (ii) integral equations (Crowder, 1987); and (iii)

k−δ n Markov chains (Lucas and Saccucci, 1990; Saccucci and
Lucas, 1990). In this paper, we have decided to calculate the
Therefore, the mean value of S(δ), given by Equation (9), ARL using the Markov chain approximation. Conversely,
depends on the distribution of the possible shift values δ ∈ in the model given by Equation (3), we can observe that
[0, A]. the probability of a false alarm is specified through the in-
For example, we can analyze the case of a uniform dis- control ARL, ARLo , and the chart power by the out-of-
tribution of the possible shift values δ ∈ [0, A]. Thus, the control ARL, ARLB .
probability density function of a uniform distribution de- Since we cannot directly use Equation (13), we will use
fined for interval [0, A] is given by the expression: an approximate method to determine the value of the out-
1 of-control ARL within the area in which the detection of
if δ ∈ [0, A],
fs (δ) = A (10) shifts is not required, or within the in-control region. In
0 if δ ∈ [0, A]. our case, according to the AAR concept given by Equa-
tion (4), we will consider the following as the mean ARLo
Taking into account Equations (8) and (10), we can calcu- value:
late that in the case of δ ∈ (0, A], the mean value of S(δ) is
as follows: 1
ARLo = (w1 ARL0 + w2 ARLA/2 + w3 ARLA )


−k−δ√n
∞  3
1 A when A < 0.5,
E[S(δ)] = φ(z)dz + √
φ(z)dz dδ.
A 0 −∞ k−δ n 1
ARLo = (w1 ARL0 + w2 ARLA/3 + w3 ARLA2/3
(11) 4
+ w4 ARLA ) when A ≥ 0.5,
Therefore, the mean value of S(δ), in the case of a uniform (15)
distribution, can be calculated using Equation (11) in terms
of the size of the sample to be used n, the value of α and where A is the magnitude of the shifts that define the limit
the size of the region where the detection of shifts is not of the in-control region (0 < δ < A), ARLo is the ARL for
required, A. δ = 0, that is, the quality characteristic is maintained under
statistical control. The rest of the ARL values are for the
4.2. Average probability of an alarm for EWMA charts shifts δ = A, A/2, A/3 and A2/3.
with regions of maximum and minimum ARL The wi coefficients are weighting factors to assign greater
importance to the different intervening ARL values, in ac-
In Section 2 the regions of minimum and maximum ARL cordance with the assumed variable distribution S(δ). Ob-
in the EWMA chart were defined as those regions that are viously, it is necessary that wi = 1. For our case, we have
equivalent to the in-control and out-of-control states. Using taken into consideration four possible distributions: (i) a
the same methodology as that used for the X̄ chart, we will uniform distribution; (ii) a triangular distribution; (iii) a
determine the mean value of S(δ) for the case of the EWMA triangularly decreasing distribution; and (iv) a triangularly
control chart. In this case, we can write: increasing distribution (see distribution shapes on the user
1 interface of the in-house developed software, Fig. 2).
δ ∈ (0, A] ⇒ s(δ) = , (12) A uniform distribution assumes that the probability den-
ARLB sity at interval δ ∈ [0, A] is constant. A triangular distri-
for the EWMA chart, we can calculate that in the case of bution is used when a higher probability density is found
δ ∈ (0, A], the mean value of S(δ), considering Equation (5) around shifts with a magnitude of δ = A/2. A triangularly
is as follows: decreasing distribution is used in situations in which there

A is a high probability density around a shift in δ close to zero.
1 A triangularly increasing distribution is used when the large
E[S(δ)] = fs (δ)dδ. (13)
0 ARL B probability density is around a shift of δ close to A.
Economic design of EWMA charts 1017

Fig. 2. User interface of the in-house developed software.

4.3. Use of a genetic algorithm to establish the parameters where C = C(n, h, r, L) is the mean cost equation given
of the sampling plan by Equation (3) and m is a constant that takes the fol-
lowing values in order to observe the following statistical
A Genetic Algorithm (GA) has been used as the optimiza-
constraints:
tion tool to establish the parameters of the sampling plan
that minimize the cost function. The programme and its 
code can be downloaded at http://ttt.upv.es/∼faparisi. GAs  0 if ARL0 > ARL0 min and ARLB < ARLB max,


are optimization algorithms based on the natural evolution 1 if ARL > ARL min and ARLB > ARLB max,
0 0
of species (Holland, 1975; Goldberg 1989; Michalewicz; m=
1 if ARL0 < ARL0
 min and ARLB < ARLB max,
1996). The use of GAs as an optimization tool for statis- 

tical quality control can be found in He et al. (2002) and 2 if ARL0 < ARL0 min and ARLB > ARLB max,
Aparisi and Garcı́a-Dı́az (2004). (17)
The quality of a solution is measured using an evaluation
function, known as the fitness function. In the case devel- where ARL0 min is the statistical constraint that the in-
oped in this paper, the in-house developed software can control ARL must satisfy and ARLB max corresponds to
optimize an economic design, and also an economic design the minimum power required for a shift δ = B. This fitness
with some statistical ARL constrains, know as an economic- function is the objective function to be minimized.
statistical design. The fitness function, implemented in the Therefore, we utilize the following factors:
GA, is:
 1. the process parameters: λ, T2 , T1 , d1 , d2 , E, r, s, and δ;


100
the cost parameters: C0 , C1 , Y, W, a and b;
 no statistical constraints, 2.
C
Fitness = (16) 3. the mean cost equation, Equation (3);

 100 4. the possible statistical constraints in the form of a
 (1 − 0.2m) statistical constraints,
C minimum ARL0 and a maximum ARLB .
1018 Garcı́a-Dı́az and Aparisi
Table 2. Multiple range tests for mean cost. Method: 95.0% LSD

Factor Count LS mean LS sigma Contrast Difference

Y
50 24 23.0951 0.010 3361
500 24 23.5214 0.010 3361
50–500 −0.426 346∗
λ
0.01 24 15.8872 0.010 3361
0.05 24 30.7292 0.010 3361
0.01–0.05 −14.842∗
A
0.2 16 23.2556 0.012 6591
0.5 16 23.2965 0.012 6591
Fig. 3. The curve of the ARL values of the optimal chart. 0.8 16 23.3756 0.012 6591
0.2–0.5 —∗
0.2–0.8 0.040 9375∗
The problem we wish to solve is to find the optimal values 0.5–0.8 —∗
of the design parameters for the EWMA chart: r, L, n and 0.117 0440
h that minimize the fitness function given by Equation (16). 0.076 1062
Distribution
5. Example of application 1 12 23.3026 0.014 6174
2 12 23.3246 0.014 6174
3 12 23.2710 0.014 6174
Statistical control charts have been successfully used in 4 12 23.3347 0.014 6174
industry to monitor stable processes. However, processes 1–2 —
with uncontrollable but acceptable trends are commonly 1–3 0.022 0167
observed. A typical example is wear processes in cutting 1–4 0.031 6417
tools. Many production processes are subject to the tool 2–3 —∗
wear problem (Duncan, 1986; Quesenberry, 1988). Such 2–4 0.032 1333
processes are often plagued by deterioration factors such as 3–4 0.053 6583∗
tool wear and increased material consumption and hence —
need regular adjustments. 0.010 1167
Consider a machine turning the Outer Diameter (OD) —
0.063 7750
of a piston. As the tool wears, the OD size of succes-
sively manufactured pistons increases. When the OD size ∗
Denotes a statistically significant difference.
approaches the upper specification limit, the tool is changed 1: uniform; 2: triangular; 3: decreasing; 4: increasing.

Table 1. ANOVA for mean cost


and the OD sizes become small again. The manufactur-
Sum of Mean ing process is centered at the target with Cpk ≥ 2.0. This is
Source squares Df square F-ratio P-value a process with “six-sigma” capability. In a six-sigma pro-
cess it is assumed that the mean may drift by as much as
Main effects 1.5 standard deviations off-target without causing serious
Y 2.181 25 1 2.181 25 850.71 0.0000
problems (Montgomery, 2001). Let us consider a produc-
λ 2643.43 1 2643 43 1030 966.19 0.0000
A 0.112 892 2 0.056 4461 22.01 0.0000 tion process with these characteristics, where shifts δ in
D 0.028 7019 3 0.009 567 29 3.73 0.0255 the mean below 0.25σ (A = 0.25) do not imply any per-
Interaction ceivable loss of product quality. For this reason, it is de-
Yλ 0.382 72 1 0.382 72 151.04 0.0000 sirable that the control chart shows a high ARL value
YA 0.004 415 37 2 0.002 207 68 0.86 0.4359 for this region. In our case, at δ = 0, a minimum ARL
YD 0.011 2201 3 0.003 740 02 1.46 0.2519 value of 500 is required. On the other hand, it is well
λA 0.024 7461 2 0.012 373 4.83 0.0178 known that shifts in the mean above 1.5σ produce unac-
λD 0.005 0325 3 0.001 6775 0.65 0.5885 ceptable losses in quality (out-of-control region, δ > 1.5σ ),
AD 0.019 0556 6 0.003 175 93 1.24 0.3235 and therefore the process is regularly adjusted. The con-
Residual 0.058 9727 23 trol chart to be designed must present a maximum ARL
Total 2646.26 47
value of four at B = 1.5. A historical study of the process
(corrected)
reveals that the deviation values in the in-control region dis-
D = Distribution. play a triangularly decreasing distribution. The following
Economic design of EWMA charts 1019
Table 3. ANOVA for n, r , L and h

n r L h

Source DF F-ratio P-value F-ratio P-value F-ratio P-value F-ratio P-value

Main effects
Y 1 161.96 0.0000 3.63 0.0693 1358. 66 0. 0000 12.55 0.0017
λ 1 86.49 0.0000 11.87 0.0022 37.59 0.0000 491.75 0.0000
A 2 21.74 0.0000 0.12 0.8832 74.88 0.0000 0.68 0.5162
D 3 3.67 0.0269 1.25 0.3159 12.50 0.0000 2.54 0.0816
Interaction
Yλ 1 2.16 0.1555 1.03 0.3214 0.59 0.4494 4.89 0.0372
YA 2 0.78 0.4708 0.97 0.3930 4.40 0.0241 3.22 0.0584
YD 3 0.48 0.6999 2.18 0.1173 6.76 0.0020 0.74 0.5414
λA 2 0.06 0.9420 0.64 0.5376 0.52 0.6003 0.14 0.8712
λD 3 2.64 0.0739 0.51 0.6822 0.20 0.8979 1.69 0.1962
AD 6 1.46 0.2365 1.62 0.1869 2.38 0.0614 0.98 0.4622
Residual 23
Total 47
D = Distribution.

process and cost parameters of the economic model used in results are listed in Tables 1, 2 and 3. The factors studied
this application example are the ones used by Montgomery are: (i) the size of the region in which the detection of shifts
et al. (1995): λ = 0.01, a = 0.5, b = 0.1, Y = 50, W = is not required, A; (ii) the intensity of the occurrences of an
25, C0 = 10, C1 = 100, E = 0.05, and T0 = 0, T1 = T2 = assignable cause, λ; (iii) the cost of a search for false alarm
2, d1 = d2 = 1 Y ; and (iv) the type of distribution of shifts in the region
Entry conditions are A = 0.25, B = 1.5, distribu- where shifts are not required to be detected. Five DOEs
tion of δ is triangularly decreasing, ARL0 min = were carried out, where the response variables are: (i) the
500 and ARLB max = 4. The EWMA control chart, mean cost, E(C); (ii) the control limit, L; (iii) the smoothing
obtained through the economic-statistical design and parameter, r; (iv) the sample size, n; and (v) the sampling
the in-house developed software is specified through interval, h.
parameters L = 3.26, r = 0.93, n = 8, h = 1.59, ARL0 = Values for the rest of the variables were kept con-
897.9 and ARLB = 1.19. stant during the study, using the values found in
A control scheme has been obtained that, in addition to the reference studied (Montgomery et al. 1995) and
good economic conditions, also presents excellent statistical are: E = 0.05, T0 = 0, T1 = 2.0, T2 = 2, C0 = 10, C1 =
conditions, as can be seen in Fig. 3. 100, W = 25, a = 0.5, b = 0.1, d1 = d2 = 1. The levels of
the factors studied appear in Table 4.
Table 1 is the Analysis of VAriance (ANOVA) corre-
6. Sensitivity analysis sponding to the study of the influence of the factors on
the mean cost. As can be observed, all the main effects and
A sensitivity analysis of the optimal solution, employing interactions between Y with λ and λ with A are signifi-
Design of Experiments (DOE) was carried out to deter- cant. These interactions of Y and A with λ, increase the
mine the influence of some factors that have an effect on
the economic model developed in the present paper. The
factorial experimental design used for the performance and
sensitivity study is a 22 × 31 × 41 design. A total of 48 runs
are required to conduct the experiment. The experimental

Table 4. Levels of the factors studied

Factor Levels

A 0.2, 0.5, 0.8


λ 0.01, 0.05
Y 50, 500
Distribution Uniform, triangular, decreasing, increasing
Fig. 4. Mean and LSD intervals for A as a function of r.
1020 Garcı́a-Dı́az and Aparisi

timal solution’s response to variations in the model vari-


ables has been carried out. As a summary of this analysis,
the size of the region in which the detection of shifts is
not required, limited by A, has an influence on the opti-
mal solution. When A increases, the control limit L val-
ues are larger, whilst the smoothing parameter, r , remains
constant.

Acknowledgements

We would like to acknowledge the helpful comments of


Fig. 5. Mean and LSD intervals for A as a function of L. the anonymous reviewers. The authors acknowledge the
effect of factor λ on the mean cost. In Table 2, the shift in financial support of the Ministry of Education and Science
the mean cost as a function of the factors is shown, using of Spain, Research Project Reference DPI2002-03537 and
the multiple test based on Fisher’s Least Significant proce- European FEDER funding. In addition, the authors would
dure (LSD). The results show that when the cost of false like to thank the Foreign Language Co-ordination Office
alarms increases, the value of λ and the value of A produce at the Polytechnic University of Valencia for their help in
a greater operational mean cost. When the distribution of translating this paper.
δ in the process is decreasing, a reduction in the mean cost
will be produced.
Table 3 is a summary of the ANOVA results for the other References
considered variables (n, r, L and h). It can be observed that
factor Y has a significant influence on n, L and h. λ has a Al-Sultan, K.S. and Rahim, M.A. (1997) Optimization in Quality Control
significant influence on the four parameters of the sampling Kluwer, Boston, MA.
scheme. The size of the in-control region A and the type of Aparisi, F. and Garcı́a-Dı́az, J.C. (2004) Optimization of univariate and
distribution have a significant influence on n and L. multivariate exponentially weighted moving-average control charts
With regard to the interactions, it can be said that, in using genetic algorithms. Computers & Operations Research, 31,
1437–1454.
general, they are not significant, except in the following Cai, D.Q., Xie, M., Goh, T.N. and Tang, X.Y. (2002) Economic design
cases: in the case of response variable L, when Y interacts of a control chart for trended processes. International Journal of
with A and the type of distribution. In the case of response Production Economics, 79, 85–92.
variable h, the cost of the false alarm interacts with λ and Crowder, S.V. (1987) A simple method for studying run-length distribu-
with A. tions of exponentially weighted moving average charts. Technomet-
rics, 29, 401–407.
It is interesting to observe the influence of the size of the Deming, W.E. (1982) Quality, Productivity and Competitive Position, MIT
in-control region on the design parameters of the chart, r Press: Cambridge, MA.
and L. Figures 4 and 5 show this effect. As can be observed, Duncan, A.J. (1956) The economic design of X̄ charts used to maintain
variations in A only affect the value of the control limit, current control of a process. Journal of the American Statistical As-
increasing its value as the value of A increases. sociation, 51, 228–241.
Duncan, A.J. (1986) Quality Control and Industrial Statistics, Irwin,
Homewood, IL.
Goldberg, D.E. (1989) Genetic Algorithms in Search. Optimization and
7. Conclusions Machine Learning, Addison Wesley: Reading, MA.
He, H., Grigoryan, A. and Sigh, M. (2002) Design of double- and triple-
From this paper, the following conclusions can be drawn. sampling X-bar control charts using genetic algorithms. Interna-
tional Journal of Production Research, 40(6), 1387–1404.
An economic model has been developed for EWMA con- Ho, C. and Case, K.E. (1994) Economic design of control charts: a litera-
trol charts for regions of maximum and minimum ARL ture review for 1981–1991. Journal of Quality Technology, 26, 1–78.
values in order to obtain optimal sampling plans, from the Holland, J. (1975) Adaptation in natural and artificial systems. University
perspective of obtaining minimum mean costs of opera- of Michigan Press, Ann Arbor, MI.
tion per time unit to satisfy statistical constraints, imposed Jones, L.A. (2002) The statistical design of EWMA control charts with
estimated parameters. Journal of Quality Technology, 34, 277–288.
as in-control and out-of-control regions. A methodology Lorenzen, T.J. and Vance, L.C. (1986) The economic design of control
has been developed to estimate the probability of average charts: a unified approach. Technometrics, 28, 3–10.
false alarms in the region in which the detection of shifts Lucas, J.M. and Saccucci, M.S. (1990) Exponentially weighted moving
is not required, δ ∈ [0, A]. This is necessary for the appli- average control schemes : properties and enhancements. Technomet-
cation of the economic model developed. Also software rics, 32, 1–12.
Michalewicz, Z. (1996) Genetic Algorithms + Data Structures = Evolution
has been developed to determine the optimal solution of Programs, Springer, Berlin, Germany.
the proposed economic model, which implements a GA Montgomery, D.C. (1980) The economic design of control charts: a review
as an optimization tool. A sensitivity analysis of the op- and literature survey. Journal of Quality Technology, 12, 75–87.
Economic design of EWMA charts 1021
Montgomery, D.C. (2001) Introduction to Statistical Quality Control. 4th Torng, J.C.-C., Cochran, J.K., Montgomery, D.C. and Lawrence, F.P.
edn., Wiley, New York, NY. (1995) Implementing statistically constrained economic EWMA
Montgomery, D.C., Torng, J.C., Cochran, J.K. and Lawrence, F.P. (1995) control charts. Journal of Quality Technology, 27(3), 257–264.
Statistically constrained economic design of the EWMA control Woodall, W.H. (1985) The statistical design of quality control charts. The
chart. Journal of Quality Technology, 27, 250–256. Statistician, 34, 155–160.
Panagos, M.R., Heikes, R.G. and Montgomery, D.C. (1985) Economic Wu, Z., Lam, Y.C., Zhang, S., and Shamsuzzaman, M. (2004) Optimiza-
design of X̄ control charts for two manufacturing process models. tion design of control chart systems. IEE Transactions, 36(5), 447–
Naval Research Logistics Quarterly, 32, 631–646. 455.
Park, C., Lee, J. and Kim, Y. (2004) Economic design of a vari- Xie, M., Del Castillo, E., Goh, T.N. and Cai, D.Q. (2001) On the mon-
able sampling rate EWMA chart. IEE Transactions, 36(5), 387– itoring of trended and regularly adjusted processes. International
399. Journal of Production Research, 39(16), 3641–3650.
Quesenberry, C.P. (1988) An SPC approach to compensating a tool-wear
process. Journal of Quality Technology, 20, 220–229.
Reynolds, M.R., Jr. and Arnold, J.C. (2001) EWMA control charts with Biographies
variable sample sizes and variable sampling intervals. IEE Transac-
tions, 33, 511–530. J. Carlos Garcı́a-Dı́az received his B.S. in 1991, M.S. in 1996 and his
Roberts, S.W. (1959) Control chart tests based on geometric moving av- Ph.D. in Statistics in 2003 from the Polytechnic University of Valencia,
erages. Technometrics, 1, 239–250. Spain. He is currently a senior quality of consultant and is an Associate
Saccucci, M.S. and Lucas, J.M. (1990), Average run lengths for exponen- Professor of statistical quality control at the Polytechnic University of
tially weighted moving average control schemes using the Markov Valencia.
chain approach. Journal of Quality Technology, 22, 154–162.
Shewhart, W.A. (1931), Economic Control of Quality of Manufactured Francisco Aparisi was awarded his BSc in 1991 from the Polytechnic Uni-
Product, Van Nostrand, New York, NY. versity of Valencia, Spain, his M.Sc. from Cranfield University, Cranfield,
Tang, X.Y., Cai, D.Q., Xie, M. and Goh, T.N. (1999) Statistical process UK in 1992, and his Ph.D. in Statistics in 1995 from the Polytechnic Uni-
control with intermediate adjustment for trend. Proceedings of the versity of Valencia. He has worked for IBM in Valencia and also as a
25th International Conference on Computers & Industrial Engineer- quality consultant and, at present, is an Associate Professor at the Poly-
ing, New Orleans, LA, pp. 33–36. technic University of Valencia.

Anda mungkin juga menyukai