You are on page 1of 16

Received: 2 February 2018 Revised: 17 May 2018 Accepted: 29 July 2018

DOI: 10.1002/qre.2377

RESEARCH ARTICLE

Hotelling's T2 control charts with fixed and variable sample


sizes for monitoring short production runs

Nger Ling Chong1 | Michael B. C. Khoo1 | Abdul Haq2 | Philippe Castagliola3

1
School of Mathematical Sciences,
Abstract
Universiti Sains Malaysia, 11800 Minden,
Penang, Malaysia Short production runs are common in enterprises that require a high degree of
2
Department of Statistics, Quaid‐i‐Azam flexibility and variety in manufacturing processes. To date, past research on
University, Islamabad, Pakistan short production runs has little focus on the multivariate control charts. In
3
Department of Quality and Logistics,
view of this, fixed sample size (FSS) and variable sample size (VSS) Hotelling's
LUNAM Université, Université de Nantes
and IRCCyN UMR CNRS 6597, T2 charts are designed to monitor the process mean when the production hori-
Carquefou, France zon is finite. Optimal parameters to minimize the out‐of‐control (1) truncated

Correspondence
average run length (TARL) and (2) expected TARL (ETARL) are provided
Michael B. C. Khoo, School of such that the in‐control TARL is equal to the number of inspections (say I).
Mathematical Sciences, Universiti Sains The numerical study considers the run length performances of the FSS and
Malaysia, 11800 Minden, Penang,
Malaysia.
VSS T2 short‐run charts for both known and unknown shift sizes. The VSS
Email: mkbc@usm.my T2 short‐run chart performs well in swiftly detecting various mean shifts in
Funding information comparison with the FSS T2 short‐run chart. Additionally, the VSS T2 short‐
Universiti Sains Malaysia, Graduate run chart is superior to the FSS T2 short‐run chart, in terms of the truncated
Assistant Scheme and Research Univer-
standard deviation of the run length, expected truncated standard deviation of
sity, Grant/Award Number: 1001.
PMATHS.8011039 the run length, probability that the chart signals an alarm within the I
inspections, ie, P(I) and expected P(I). A case study on the impurity profile
of a crystalline drug substance illustrates the implementation of the VSS T2
short‐run chart.

KEYWORDS
control chart, fixed and variable sample sizes, Hotelling's T2, short production runs, truncated run
length

1 | INTRODUCTION Therefore, adaptive control charts, where the process


parameters (sample size, sampling interval, and control
A control chart is an effective Statistical Process Control limit (CL) coefficient) can vary during production
(SPC) tool that is widely adopted in the process and ser- depending on the previous sample statistics, are con-
vice industries to monitor changes in a process over time structed to enhance the sensitivity of the Shewhart chart
and increase productivity. The main purpose of a control for detecting small and moderate shifts. In comparison to
chart is to identify the occurrence of assignable causes in the standard Shewhart charts, the adaptive control charts
a production system so that corrective actions can be perform better in shift detection because of the consider-
taken well in time to eliminate them before the produc- ation of past process information to determine suitable
tion of nonconforming units. The Shewhart X chart is parameters for future inspections. Tagaras1 highlighted
extensively used due to its easy implementation, but it is the developments of adaptive charts and provided com-
very slow in detecting small and moderate shifts. parisons between the static and adaptive charts. Some

14 © 2018 John Wiley & Sons, Ltd. wileyonlinelibrary.com/journal/qre Qual Reliab Engng Int. 2019;35:14–29.
CHONG ET AL. 15

of the literature on adaptive control charts can be seen in production runs, where the run of a process is short, ie,
Khaw et al,2 Lee and Khoo,3 Lin,4 and Zhou.5 a few hours or days. In Just‐in‐Time manufacturing,
The variable sample size (VSS) scheme, which is an goods are produced to meet demand, while the waste
adaptive strategy, has a widespread application and thus due to overproduction, waiting, and surplus inventory
has been studied by many researchers. Prabhu et al6 and are minimized. As for job shop settings, small lot sizes
Costa7 presented the VSS X chart whose sample size of diverse custom items are manufactured using unique
selection changes on various levels, according to the setups. The importance of short production runs has been
position of the last point on the chart. Recently, Teoh shown by various research publications. Nenes and
et al8 investigated an optimal design of the VSS X chart, Tagaras23 proposed the cumulative sum chart for short
in terms of the median and expected median run lengths. production runs. They also introduced the truncated
Hu et al9 discussed the effect of measurement errors on average run length (TARL) and truncated average time
the performance of the VSS X chart. The performance to signal criteria, which are performance measures used
of the VSS X chart when process parameters are esti- in short runs, as there are limited and fixed total number
mated was studied by Castagliola et al10. Castagliola of samples, to replace their counterparts in long runs, ie,
et al11 evaluated the performance of the VSS chart that average run length (ARL) and average time to signal,
monitors the coefficient of variation (CV) of a process. respectively. In addition, Celano et al24 compared the per-
Additionally, Yeong et al12 presented the VSS chart that formances of the Shewhart and EWMA t charts in a finite
monitors the CV directly instead of the transformed sta- horizon, where t is the Student's t test statistic. Castagliola
tistics employed by Castagliola et al.11 Amiri et al13 et al25 adopted the adaptive approach by proposing the
extended the VSS approach to the exponentially VSS t short‐run chart. In their paper, it was found that
weighted moving average (EWMA) control chart. More the VSS t short‐run chart outperforms the fixed‐parame-
research works on the VSS control charts can be found ter t short‐run chart for detecting moderate to large shifts
in Mahadik,14 Wu and Luo,15 Noorossana and when the production horizon is finite. Nenes et al26 pro-
Maryam,16 Mahadik,17 Cheng and Wang,18 Aslam posed a variable sampling interval (VSI) chart, where
et al,19 Zhou and Lian,20 Nyau et al,21 and Lee.22 the length of the sampling interval varies, in a short pro-
In VSS charts, the sample sizes change as a function duction runs process. Concerning short‐run charts that
of the prior sample position, ie, small or large sample monitor the CV, a one‐sided Shewhart, VSS, and VSI
size. The VSS control chart consists of 3 regions, namely, charts were investigated by Castagliola et al,27 Amdouni
the central region, warning region, and out‐of‐control et al,28 and Amdouni et al,29 respectively.
region, which are divided by the control and warning Recently, there is an increased interest in monitoring
limits. The selected sample size depends on the position several quality characteristics or variables simultaneously
of the previous sample point on the VSS chart. If a using multivariate control charts. In multivariate SPC,
sample point lies within the central region of the VSS the Hotelling's T2 chart is commonly used due to its easy
chart, a small sample size, nS (“S” stands for small), is implementation. For brevity, the Hotelling's T2 chart is
adopted for the next inspection as there is a lower referred to as the T2 chart. To improve the performance
tendency for the process to require adjustment. Mean- of the T2 chart in detecting small process shifts, Aparisi30
while, if a sample point lies within the warning region, applied the VSS procedure to the T2 control chart and
a large sample size, nL (“L” stands for large), is required obtained considerable reduction in the out‐of‐control
for the next inspection to tighten the control as the ARL value, which translates to a lower mean time to
process is inclined to shift to an out‐of‐control state. detect an out‐of‐control condition. As for Aparisi and
When a sample point falls in the out‐of‐control region, Haro,31 the VSI strategy was included in the T2 chart.
the process is declared as out‐of‐control and a search Aparisi and Haro32 further developed the T2 chart by
for assignable causes is then made to detect and remove incorporating the variable sample size and sampling
them. interval (VSSI) scheme. Other past literature on adaptive
Traditional SPC techniques involve large volume T2 charts can be found in Mahadik and Shirke,33 Seif
manufacturing and an infinite production horizon. In et al,34 Mahadik,35 Chen and Hsieh,36 Faraz and
recent years, short production runs, which is character- Moghadam,37 and Aparisi et al38, to name a few. In the
ized by a high‐variety and low‐volume production, have present literature, approaches to monitor multivariate
been adopted by many companies to be more flexible, data in a short production runs setting are limited. In
specialized, and to meet demands of frequent changes fact, most of the past research works on finite production
in the manufactured products. There is an increased horizons are on univariate charts. Hence, this paper fills
emphasis on the Just‐in‐Time philosophy and job shop this gap by proposing both multivariate fixed sample size
settings that require a finite production horizon or short (FSS) and VSS T2 short‐run charts to monitor the mean
16 CHONG ET AL.

shifts in short production runs. The proposed charts are λ ¼ nðiÞðμ 1 −μ 0 ÞT Σ−1 0 ðμ 1 − μ 0 Þ ¼ nðiÞδ
2
is the
compared with their existing counterparts in terms of noncentrality parameter. The Mahalanobis distance,
the TARL, truncated standard deviation of the run length which is used to determine the magnitude of the changes
(TSDRL), average sample size (ASS), and the probability in the process mean vector, is denoted by
P(I) of getting a signal within the number of scheduled qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
δ ¼ ðμ 1 −μ 0 ÞT Σ−1 0 ðμ 1 − μ 0 Þ, where μ1 is the mean vector
inspections (I), for known and unknown shift sizes when
the number of variables p = 2. when there is a shift in at least one of the means of the p
The remainder of this paper is organized as follows: quality characteristics.32
Section 2 discusses the properties of the Hotelling's T2 sta- In a short production run, the manufacturing process
tistic and short production runs, whereas Section 3 out- produces a small lot with I number of scheduled inspec-
lines the designs of the FSS and VSS T2 short‐run tions in a finite production horizon H. The interval
charts. Subsequently, the formulae for the performance between two consecutive inspections in a short produc-
measures are derived in Section 4. In Section 5, the tion run, also known as the sampling interval, is obtained
results of the numerical analysis are reported followed with h = H/(I + 1) hours, as there is no inspection sched-
by a discussion of the case study in Section 6. Conclusions uled at the end of the run.25 To illustrate the meaning of
and suggestions for future research works complete the the terms, the following example is presented and shown
paper in Section 7. in Figure 1: Let the production horizon be H = 10 hours
and the number of scheduled inspections be I = 4. Conse-
quently, the sampling interval is h = 10/(4 + 1) = 2 hours.
2 | PROPER T I E S OF T H E The first inspection happens after h = 2 hours from the
H O T E L L IN G ' S T 2 S T A T I S T I C AN D beginning of the short runs, while the last (fourth) inspec-
SHOR T P RODUC TION R UNS tion occurs h = 2 hours before the production horizon
ends. The time interval between the last inspection and
Suppose that a short production run process is monitored the end of the run is h = 2 hours.
by taking data on p correlated quality characteristics
simultaneously, which is assumed to follow a multivari-
ate normal distribution with the mean vector μ0 and the
3 | F SS A ND V SS T 2 SHORT ‐R UN
covariance matrix Σ0, when the underlying process is
CONTROL C HARTS
in‐control. The Hotelling's T2 statistic is as follows30
 T   To implement the FSS T2 short‐run chart, the sample size
T 2i ¼ nðiÞ X i −μ0 Σ−1
0 X i − μ0 ; for i ¼ 1; 2; …; I; (1)
n(i), CL (CL > 0), and h have to be determined. Note that
where n(i) is the sample size at the ith inspection and CL is determined to attain the in‐control TARL (TARL0)
 T when δ = 0. A simple schematic representation of the
X i ¼ X 1i ; X 2i ; …; X pi is the p × 1 vector of sample means.
FSS T2 short‐run chart's operation is given in Figure 2.
nðiÞ
Here, X ki ¼ ∑ X jki =nðiÞ, where Xjki is the jth observation This chart is implemented as follows:
j¼1
on the kth variable for the ith inspection, when k = 1, 2, Step 1. At the ith inspection, for i = 1, 2, …, I, a sample
…, p. Note that, as expected, T 2i > 0. For an in‐control pro- with fixed size, n(i) = n0, is obtained.
cess (μ = μ0), T 2i follows a central chi‐square distribution Step 2. If T 2i < CL, the process is in‐control. Otherwise,
with p degrees of freedom, ie, T 2i ~ χ 2p . On the other hand, the process is out‐of‐control and corrective actions are
for the out‐of‐control process (μ = μ1), T 2i follows a noncen- taken to identify and eliminate the assignable causes,
tral chi‐square distribution with the noncentrality parame- in order to return the process to its in‐control condi-
ter λ and p degrees of freedom, ie, T 2i ∼ χ 2p ðλÞ. Here, tion. Then the control flow returns to Step 1.

FIGURE 1 Example of a short


production run, where H = 10, I = 4, and
h = 10/(4 + 1) = 2
CHONG ET AL. 17

FIGURE 3 Schematic representation of the variable sample size


FIGURE 2 Schematic representation of the fixed sample size T2
T2 chart's operation for short production runs
chart's operation for short production runs

Instead, corrective actions are taken so that the process


2
The VSS T short‐run chart consists of a warning limit returns to the state of statistical control. The transition
W and a CL, where 0 < W ≤ CL. The sample size taken probability matrix (TPM) for a given shift δ is
can be either nS or nL depending on the location of the 0 1
  Pδ1;1 Pδ1;2 1 − Pδ1;1 − Pδ1;2
previous T2 statistic on the chart. Note that if T 2i falls in Q r B C
the central or warning region, the process is declared as Pδ ¼ ¼ @ Pδ2;1 Pδ2;2 1 − Pδ2;1 − Pδ2;2 A; (2)
0T 1
in‐control but a different sample size, ie, the small or 0 0 1
large sample size, respectively, is taken at the next inspec-
tion. It is assumed that there is a perfect initial setup; where Pδk;l denotes the transition probability from state k
thus, the first sample size is small, ie, n(1) = nS. to state l when the mean vector involves a shift δ. Q is a
The VSS strategy works as follows: 25 2 × 2 matrix of transient probabilities, and r = 1 − Q ⋅ 1
(as the row sum of the TPM is unity) is a 2 × 1 vector,
• If T 2i ∈ I 1 ¼ ½0; W  (central region), the process is where 1 ¼ ð1; 1ÞT . The transition probabilities in Q are
declared as in‐control and the sample size adopted defined as
for taking the next sample is n(i + 1) = nS.  
• If T 2i ∈ I 2 ¼ ðW ; CL (warning region), the process is Pδ1;1 ¼ P T 2i ≤ W jnS ; δ ¼ F χ 2 ðW jp; λ1 Þ; (3)
declared as in‐control and the sample size adopted
for taking the next sample is n(i + 1) = nL.  
Pδ1;2 ¼ P W < T 2i ≤ CLjnS ; δ
• If T 2i ∈ I 3 ¼ ½CL; ∞Þ (out‐of‐control region), the pro-
¼ F χ 2 ðCLjp; λ1 Þ − F χ 2 ðW jp; λ1 Þ; (4)
cess is declared as out‐of‐control and corrective
actions are taken to identify and remove the assign-
 
able causes. After the elimination of assignable Pδ2;1 ¼ P T 2i ≤ W jnL ; δ ¼ F χ 2 ðW jp; λ2 Þ; (5)
causes, the process is restarted with a sample size
n(i + 1) = nS.  
Pδ2;2 ¼ P W < T 2i ≤ CLjnL ; δ
A schematic representation of the VSS T2 chart's opera- ¼ F χ 2 ðCLjp; λ2 Þ − F χ 2 ðW jp; λ2 Þ; (6)
tion for short production runs is shown in Figure 3.  
where F χ 2 ⋅jp; λj is the cumulative distribution function
The sample points plotted on the VSS T2 short‐run
chart follow a stochastic process, and the TARL can of the non‐central chi‐square random variable with p
be computed using a 3‐state Markov chain, where the degrees of freedom and the noncentrality parameter λj,
  for j = 1 or 2, where λ1 = nSδ2 and λ2 = nLδ2.
first state T 2i ∈ ½0; W  corresponds to adopting nS in
 
taking sample i + 1, the second state T 2i ∈ ðW ; CL
corresponds to adopting nL in taking
 sample i + 1, 4 | S T A T I S T I C A L ME A S U R E S OF
and the third state T 2i ∈ ðCL; ∞Þ corresponds to the PERFO RM AN CE
out‐of‐control condition. The first two states are the tran-
sient or in‐control states. On the contrary, the third state This section discusses the performance measures of the
is an absorbing or out‐of‐control state, where process FSS and VSS T2 short‐run charts for both known and
monitoring does not proceed with the next inspection. unknown shift sizes. The charts are studied in terms of
18 CHONG ET AL.


the TARL, TSDRL, ASS, and P(I) criteria when the shifts f RL ðiÞ if i ¼ 1; 2; …; I
are known and expected TARL, TSDRL, ASS, and P(I) f TRL ðiÞ ¼ ; (9)
1 − F RL ðI Þ if i ¼ I þ 1
criteria when the shifts are unknown.
and
4.1 | TARL, TSDRL, ASS, and P(I) 
F RL ðiÞ if i ¼ 1; 2; …; I
When a process operates indefinitely, a typical perfor- F TRL ðiÞ ¼ ; (10)
1 if i ¼ I þ 1
mance measure used is the ARL, which measures the
expected number of sample points plotted on a control where the probability mass function f RL(i) and cumula-
chart until the occurrence of the first out‐of‐control sig- tive distribution function FRL(i) of the usual run length
nal. However, for short production runs, there is a possi- (RL) are
bility that the process ends without any out‐of‐control
signal being issued by the chart. To address this phenom- f RL ðiÞ ¼ qT Qi−1 r; (11)
enon, Nenes and Tagaras23 introduced the TARL crite-
rion as a performance measure for charts with finite and
production horizons. The speed of signaling an out‐of‐
control condition is measured by the TARL, which is F RL ðiÞ ¼ 1 − qT Qi 1; (12)
defined as the expected number of samples inspected
until an out‐of‐control condition is signaled or until the respectively. The TARL1(δ) of the VSS T2 short‐run chart
production process is completed, whichever occurs first, can be obtained from the expected value of the TRL, ie,
following a process shift.24 As the production run is short, E(TRL), which can be shown to be23
there is a possibility that the run ends before the chart
 
signals an out‐of‐control condition. If there is no signal I
when the run ends within the I inspections, the truncated TARL1 ðδÞ ¼ q T
∑ Q 1;i
(13)
i¼0
run length (TRL) has a value of I + 1.24 Conversely, the
TRL can be assigned values from 1 to I when there is
where q in Equation (13) is a 2 × 1 vector of initial
an out‐of‐control signal within the I inspections. The per-
probabilities for the transient states. As it is assumed in
formance measures are evaluated by considering that an
this paper that the initial sample size is small, ie,
out‐of‐control condition occurs immediately after the
n(1) = nS then q = (1, 0)T. Otherwise, if n(1) = nL, then
start‐up of the short run. As the possible number of
q = (0, 1)T.
inspections (I = 10, 30, 50) is small during short produc-
Another measure of performance, the out‐of‐control
tion runs, no steady‐state condition where an out‐of‐con-
TSDRL computed for shift δ, which is denoted as
trol condition occurs at some time in the future is
TSDRL1(δ), is used to gain insight about the spread of
achieved.24
the TRL for a finite horizon, instead of SDRL1(δ), which
The probability mass function of the TRL for the FSS
2 is used in the case of infinite horizons. To derive the
T short‐run control chart is
TSDRL1(δ) for the FSS T2 short‐run chart, the second
( moment of TRL (denoted by TRL2), ie, E(TRL2), can be
ð1 − βÞβi−1 if i ¼ 1; 2; …; I computed as follows:27
f TRL ðiÞ ¼ ; (7)
βI if i ¼ I þ 1
2IβIþ2 − 2IβIþ1 þ βIþ2 − 3βIþ1 þ β þ 1
TRL2 ¼ : (14)
where β ¼ F χ 2 ðCLjp; λÞ is the Type‐II error probability. ð1−βÞ2
Consequently, the out‐of‐control TARL (or E(TRL)) for
qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
the shift δ is computed as23
Then TSDRL1 ðδÞ ¼ EðTRL2 Þ − TARL1 ðδÞ2 can be

Iþ1
shown to be27
TARL1 ðδÞ ¼ ∑ i f TRL ðiÞ
i¼1 qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
  ffi
I 1 − βIþ1 β 1 − β2Iþ1 − ð1 − βÞβIþ1 ð1 þ 2I Þ
¼ ð1 − βÞ ∑ iβi−1 þ ðI þ 1ÞβI ¼ : (8) TSDRL1 ðδÞ ¼ : (15)
i¼1 1−β 1−β

As for the VSS T2 short‐run chart, f TRL(i) and F TRL(i) As for the VSS T2 short‐run chart, the second moment
of the TRL criterion can be obtained as follows:28 of the TRL is28
CHONG ET AL. 19

I 
PðI Þ ¼ 1 − qT QIþ1 1; (23)
TRL2 ¼ qT ∑ ð2i þ 1ÞQi 1: (16)
i¼0
where qTQI+11 is the probability that there is no signal
Consequently, within the I inspections.
qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
TSDRL1 ðδÞ ¼ TRL2 − ½EðTRLÞ2 4.2 | Expected TARL, TSDRL, ASS, and P(I)
qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
¼ TRL2 − TARL1 ðδÞ2 : (17) When a specific shift size, δ, is known, the performance
of a chart can be studied in terms of TARL1(δ). However,
As stated by Castagliola et al,25 the out‐of‐control ASS
in practice, the exact value of δ is unknown to the SPC
for a shift δ denoted by ASS1(δ) is a ratio between the
practitioners before the shift occurs and the chart will
expected total number of observations taken in a finite
have poor statistical performance if the actual shift size
production horizon and I, as shown in Equation (18).
differs from the value specified at the chart's designing
I  stage. Thus, the out‐of‐control ETARL, denoted as
1 1 I
ASS1 ðδÞ ¼ E ∑ nðiÞ ¼ ∑ EðnðiÞÞ: (18) ETARL1(δmin, δmax), calculated for shifts in the interval
I i¼1 I i¼1
(δmin, δmax) is adopted instead of TARL1(δ) as a perfor-
The TPM Pδ in Equation (2) is transformed to P e δ in mance measure. Note that δmin and δmax denote the min-
order to evaluate ASS1(δ), with the assumption that when imum and maximum shifts in the mean vector (in the
there is an out‐of‐control condition, corrective actions are interval (δmin, δmax)), in terms of the Mahalanobis dis-
taken swiftly to remove the assignable cause(s) before the tance, respectively.
next sample n(i + 1) = n(1) is taken. As n(1) = nS is The ETARL1(δmin, δmax) is determined as
adopted in this paper, the third row of P e δ is a repetition
δmax
of the first row that corresponds to the small sample size. ETARL1 ðδmin ; δmax Þ ¼ ∫δmin TARL1 ðδÞf ðδÞdδ; (24)
All states are now accessible, and the modified TPM P e δ is
shown in Equation (19): where TARL1(δ) is computed using either Equation (8)
(for the FSS T2 short‐run chart) or Equation (13) (for
0 1
Pδ1;1 Pδ1;2 1 − Pδ1;1 − Pδ1;2 the VSS T2 short‐run chart), while f (δ) is the probability
B C density function of δ. According to Sparks,39 δ can be
e δ ¼ B Pδ
P Pδ2;2 1 − Pδ2;1 − Pδ2;2 C
@ 2;1 A: (19)
assumed to follow a uniform distribution U(δmin, δmax)
Pδ1;1 Pδ1;2 1− Pδ1;1 − Pδ1;2 as there is a general assumption that an arbitrary shift
size δ in the interval δmin ≤ δ ≤ δmax has an equal proba-
e δ , E(n(i)) in Equation (18) becomes25
By using P bility of occurrence. Celano et al40 noted that practi-
tioners generally have a preference for a range of
EðnðiÞÞ ¼ e e ðnS ; nL ; nS ÞT ;
qT P
i−1
(20) shifts that should be detected with reasonable sensitivity
δ
by the chart. Due to the uniform distribution,
where qeT ¼ ð1; 0; 0Þ is a vector representing the initial Equation (24) becomes
probabilities of the modified Markov chain. From Equa-
tions (18) and (20), ASS1(δ) can be computed as25 1 δmax
ETARL1 ðδmin ; δmax Þ ¼ ∫ TARL1 ðδÞ dδ:
  δmax − δmin δmin
1 T I e i−1 (25)
ASS1 ðδÞ ¼ qe ∑ Pδ ðnS ; nL ; nS ÞT (21)
I i¼1
The Gauss‐Legendre quadrature can be used to
The in‐control ASS (ASS0) is computed using Equa- approximate the integral in Equation (25) as it is difficult
tion (21) by considering δ = 0 in the transition probabili- to evaluate it analytically. The same strategy as
ties in Equations (3)–(6). ETARL1(δmin, δmax) is used to compute the expected
For the purpose of measuring the sensitivity of the TSDRL1(δ), which is represented as ETSDRL1(δmin, δmax)
FSS T2 short‐run chart, the probability that the chart sig- when δ is unknown and it is given as follows:
nals an alarm within the I inspections is24
1 δmax
PðI Þ ¼ 1 − βI ; (22) ETSDRL1 ðδmin ; δmax Þ ¼ ∫δmin TSDRL1 ðδÞ dδ:
δmax − δmin
(26)
where βI is the probability that there is no signal within
the I inspections. The probability of getting a signal For the VSS T2 short‐run chart, the expected out‐of‐
within the I inspections for the VSS T2 short‐run chart is control ASS, EASS1(δmin, δmax) is also used to evaluate
20 CHONG ET AL.

the performance of the chart as the exact value of δ is subject to the


usually unknown in real‐life settings. Unlike ASS1(δ),
which requires the exact value of δ, practitioners only
constraint: TARL0 ¼ I: (30b)
need a possible range of δ values to compute
EASS1(δmin, δmax), where
Note that TARL0 in constraints (29b) and (30b)
1 δmax
EASS1 ðδmin ; δmax Þ ¼ ∫δmin ASS1 ðδÞdδ: (27) denotes the in‐control TARL.
δmax − δmin It is worth noting that ASS0 of the VSS T2 short‐run
chart is set to be equal to n0 of the FSS T2 short‐run chart
On the other hand, the expected probability of getting
to ensure a fair comparison between the two charts. The
a signal within the I inspections, E(P(I)) is computed as
optimal sample sizes (nS, nL), satisfying the constraint
1 δmax 2 ≤ nS < ASS0 < nL ≤ 31, as considered by Amdouni
EðPðI ÞÞ ¼ ∫δmin PðI; δÞdδ: (28) et al,19 are also adopted for the VSS T2 short‐run chart.
δmax − δmin
The optimal parameters and results of the performance
measures are computed using the optimization programs
written in the ScicosLab software. The accuracy of the
5 | N U M E R I C A L A N A LY S I S
results has been verified using simulation in the Statisti-
cal Analysis System (SAS) 9.4 software.
In this section, the performance of the FSS and VSS T2
In Tables 1–3, the statistical performances of the
short‐run charts are evaluated and compared. The statis-
FSS and VSS T2 short‐run charts, for I ∈ {10, 30, 50},
tical performance is studied in terms of the performance
ASS0 ∈ f5; 7; 9g, number of variables p = 2, and the
measures TARL1(δ) and TSDRL1(δ), for known shifts.
mean shifts δ ∈ {0.1, 0.2, 0.3, 0.4, 0.5, 0.6, 0.7, 0.8, 0.9,
The FSS and VSS T2 short‐run charts are TARL unbiased
1.0, 2.0}, are presented. Note that the TARL0 is set equal
as TARL0 > TARL1(δ), for all shift sizes. Additionally,
to I. It can be seen from Tables 1–3 that the VSS T2
in the case of the VSS T2 short‐run chart, where the
short‐run chart is generally superior to the FSS T2 short‐
sample size deviates from one sample to another, it is of
run chart for small and moderate shifts, in terms of the
interest to analyze the performance of the chart in terms
TARL criterion, and the outperformance generally
of ASS1(δ). Furthermore, P(I) is computed to study the
increases with ASS0 and I. On the other hand, the FSS
sensitivity of the charts. As practitioners usually do
T2 short‐run chart surpasses the VSS T2 short‐run chart
not know the actual shift size in the mean vector, an
when the shift is large (δ = 2). Note that when TARL1(δ)
interval of shift size δ, say (δmin, δmax), is selected in the
is smaller, the chart is more effective as an out‐of‐control
next step of the analysis, and the performance measures,
condition is signaled earlier. To illustrate, consider
including ETARL1(δmin, δmax), ETSDRL1(δmin, δmax),
Table 1 when I = 10, δ = 0.5 and ASS0 ¼ 5, TARL1(δ)
EASS1(δmin, δmax), and E(P(I)), are adopted.
for the VSS and FSS T2 short‐run charts are 6.19 and
The parameters of the VSS T2 short‐run chart are nS,
7.50, respectively (VSS signals quicker than FSS by 1.21
nL, W, and CL. On the other hand, the parameters of
times), whereas in Table 3, when I = 50, δ = 0.5, and
the FSS T2 short‐run chart are n0 and CL. When an exact
ASS0 ¼ 9, TARL1(δ) for the VSS and FSS T2 short‐run
shift size can be specified, the optimal design of the FSS
charts are 9.26 and 32.68, respectively (VSS signals
and VSS T2 short‐run charts can be formulated by the
quicker than FSS by 3.53 times). In view of a larger
model shown below:
ASS0 , larger values of nS can be taken and nL usually
takes values close to the upper bound of the constraint
Objective function: Minimize TARL1 ðδÞ (29a)
2 ≤ nS < ASS0 < nL ≤ 31. Consequently, increases in
the sample sizes for both nS and nL improve the sensitiv-
subject to the ity of the VSS T2 short‐run chart to detect an assignable
cause. On the other hand, a larger value of I enhances
constraint: TARL0 ¼ I: (29b) the effect of the adaptive strategy as the possibility of
the samples to switch to a large sample size (nL)
As for a shift whose exact size cannot be specified, the
increases.16 Meanwhile, the larger the shift size δ, the
optimization procedure of the two charts is shown by the
smaller is the corresponding TARL1(δ) value indicating
following model:
that the charts have a better ability to detect a process
change when the shift is larger. This is justifiable as large
Objective function: Minimize ETARL1 ðδmin ; δmax Þ shifts result in considerable loss of quality and require
(30a) swift detections.
CHONG ET AL. 21

TABLE 1 TARL1(δ) minimization with charts' parameters and results of performance measures, for I = 10, TARL0 = I, ASS0 ∈ f5; 7; 9g,
p = 2, and δ ∈ {0.1, 0.2, 0.3, 0.4, 0.5, 0.6, 0.7, 0.8, 0.9, 1.0, 2.0}

FSS VSS
δ ns nL CL W TARL1 TSDRL1 P(I) TARL1 TSDRL1 ASS1 P(I)

ASS0 = 5
0.1 2 31 7.9003 4.0169 9.91 2.57 0.19 9.91 2.56 5.13 0.21
0.2 2 31 7.9003 4.0169 9.62 2.83 0.24 9.55 2.84 5.50 0.28
0.3 2 31 7.9003 4.0169 9.12 3.17 0.32 8.76 3.24 6.05 0.43
0.4 2 31 7.9003 4.0169 8.41 3.50 0.43 7.51 3.49 6.52 0.64
0.5 2 28 7.9003 3.8284 7.50 3.73 0.57 6.19 3.37 6.70 0.82
0.6 2 21 7.9003 3.2727 6.45 3.76 0.70 5.01 2.95 6.66 0.93
0.7 3 18 7.9003 3.5747 5.37 3.57 0.82 4.03 2.50 6.77 0.98
0.8 3 16 7.9003 3.3194 4.34 3.18 0.91 3.29 1.95 6.63 1.00
0.9 4 14 7.9003 4.0749 3.47 2.67 0.96 2.74 1.65 6.47 1.00
1.0 4 13 7.9003 3.8939 2.79 2.15 0.99 2.31 1.28 6.32 1.00
2.0 4 8 7.9003 2.4278 1.04 0.20 1.00 1.09 0.29 4.29 1.00
ASS0 = 7
0.1 3 31 7.9003 3.4521 9.87 2.61 0.20 9.87 2.60 7.20 0.22
0.2 2 31 7.9003 3.1135 9.46 2.95 0.27 9.35 2.96 7.87 0.32
0.3 2 31 7.9003 3.1135 8.76 3.36 0.38 8.22 3.38 8.70 0.53
0.4 2 31 7.9003 3.1135 7.76 3.68 0.53 6.58 3.40 9.17 0.78
0.5 3 31 7.9003 3.4521 6.55 3.77 0.69 5.10 3.01 9.41 0.92
0.6 4 25 7.9003 3.4521 5.26 3.54 0.83 3.94 2.42 9.44 0.98
0.7 5 21 7.9003 3.6885 4.08 3.04 0.93 3.11 1.89 9.25 1.00
0.8 5 19 7.9003 3.4521 3.13 2.43 0.98 2.53 1.39 8.98 1.00
0.9 6 16 7.9003 4.0749 2.44 1.84 1.00 2.10 1.16 8.49 1.00
1.0 6 15 7.9003 3.8939 1.96 1.37 1.00 1.80 0.90 8.20 1.00
2.0 6 10 7.9003 2.4278 1.00 0.07 1.00 1.01 0.11 6.05 1.00
ASS0 = 9
0.1 5 31 7.9003 3.3194 9.83 2.64 0.21 9.83 2.63 9.25 0.23
0.2 2 31 7.9003 2.4934 9.30 3.06 0.29 9.16 3.06 10.20 0.36
0.3 2 31 7.9003 2.4934 8.39 3.51 0.44 7.74 3.45 11.18 0.61
0.4 3 31 7.9003 2.7147 7.13 3.77 0.62 5.88 3.26 11.70 0.86
0.5 5 31 7.9003 3.3194 5.68 3.65 0.79 4.30 2.64 12.07 0.97
0.6 6 28 7.9003 3.5349 4.30 3.16 0.91 3.25 1.96 11.96 1.00
0.7 7 24 7.9003 3.7946 3.18 2.47 0.98 2.55 1.47 11.53 1.00
0.8 8 20 7.9003 4.3813 2.40 1.80 1.00 2.08 1.18 10.77 1.00
0.9 8 19 7.9003 4.2362 1.89 1.29 1.00 1.74 0.88 10.43 1.00
1.0 8 17 7.9003 3.8939 1.56 0.93 1.00 1.52 0.69 9.90 1.00
2.0 8 13 7.9003 2.8420 1.00 0.02 1.00 1.00 0.04 8.01 1.00

From the results in Tables 1–3, it is observed that the implying that the latter has a larger spread of the TRL dis-
TSDRL1(δ) value of the VSS T2 short‐run chart is tribution; hence, it is inferior to the former. Owing to a
generally lower than that of the FSS T2 short‐run chart, more intensive sampling to implement the VSS T2
22 CHONG ET AL.

TABLE 2 TARL1(δ) minimization with charts' parameters and results of performance measures, for I = 30, TARL0 = I, ASS0 ∈ f5; 7; 9g,
p = 2, and δ ∈ {0.1, 0.2, 0.3, 0.4, 0.5, 0.6, 0.7, 0.8, 0.9, 1.0, 2.0}

FSS VSS
δ ns nL CL W TARL1 TSDRL1 P(I) TARL1 TSDRL1 ASS1 P(I)

ASS0 = 5
0.1 2 31 12.2418 4.4289 29.85 4.72 0.07 29.83 4.74 5.16 0.08
0.2 2 31 12.2418 4.4289 29.35 5.58 0.11 29.02 6.02 5.72 0.13
0.3 2 31 12.2418 4.4289 28.39 6.86 0.16 26.27 8.58 6.77 0.31
0.4 2 31 12.2418 4.4289 26.79 8.36 0.26 19.73 10.49 8.17 0.67
0.5 2 31 12.2418 4.4289 24.41 9.80 0.39 11.97 8.48 9.03 0.94
0.6 2 31 12.2418 4.4289 21.19 10.78 0.56 7.69 5.57 8.86 1.00
0.7 2 25 12.2418 3.9737 17.35 10.91 0.73 5.70 3.80 8.60 1.00
0.8 3 22 12.2418 4.3948 13.37 10.00 0.87 4.43 2.74 8.78 1.00
0.9 3 20 12.2418 4.1766 9.83 8.29 0.95 3.60 2.01 8.53 1.00
1.0 3 18 12.2418 3.9304 7.09 6.32 1.00 3.05 1.53 8.23 1.00
2.0 4 10 12.2418 3.4904 1.16 0.43 1.00 1.27 0.45 5.18 1.00
ASS0 = 7
0.1 2 31 12.2418 3.4234 29.78 4.84 0.08 29.76 4.87 7.30 0.08
0.2 2 31 12.2418 3.4234 29.06 6.01 0.12 28.53 6.62 8.27 0.16
0.3 2 31 12.2418 3.4234 27.60 7.68 0.21 24.38 9.56 9.99 0.42
0.4 2 31 12.2418 3.4234 25.13 9.44 0.35 15.96 10.03 11.87 0.82
0.5 2 31 12.2418 3.4234 21.49 10.73 0.55 8.74 6.26 12.38 1.00
0.6 3 31 12.2418 3.7945 16.95 10.87 0.75 5.60 3.66 12.17 1.00
0.7 4 30 12.2418 4.2147 12.30 9.58 0.90 4.15 2.48 12.12 1.00
0.8 5 26 12.2418 4.5908 8.45 7.37 0.98 3.31 1.81 11.92 1.00
0.9 5 23 12.2418 4.2888 5.76 5.16 1.00 2.74 1.32 11.41 1.00
1.0 6 21 12.2418 5.2856 4.06 3.52 1.00 2.35 1.12 10.92 1.00
2.0 6 12 12.2418 3.4904 1.03 0.17 1.00 1.06 0.25 6.35 1.00
ASS0 = 9
0.1 2 31 12.2418 2.7575 29.72 4.96 0.08 29.69 5.00 9.43 0.09
0.2 2 31 12.2418 2.7575 28.75 6.42 0.14 28.06 7.13 10.79 0.19
0.3 2 31 12.2418 2.7575 26.74 8.40 0.26 22.76 10.10 13.00 0.51
0.4 2 31 12.2418 2.7575 23.32 10.24 0.45 13.52 9.21 14.95 0.90
0.5 3 31 12.2418 2.9934 18.51 10.98 0.68 7.17 5.00 15.34 1.00
0.6 5 31 12.2418 3.6484 13.18 9.93 0.88 4.55 2.75 15.29 1.00
0.7 6 31 12.2418 4.1376 8.65 7.52 0.97 3.36 1.80 15.02 1.00
0.8 7 30 12.2418 4.7686 5.61 5.02 1.00 2.71 1.34 14.78 1.00
0.9 7 27 12.2418 4.4953 3.80 3.26 1.00 2.28 1.00 14.08 1.00
1.0 8 23 12.2418 5.2856 2.73 2.17 1.00 1.96 0.86 12.87 1.00
2.0 8 14 12.2418 3.4904 1.00 0.07 1.00 1.01 0.11 8.07 1.00

short‐run chart, a higher cost may be incurred. In fact, chart, n0 (cf., Tables 1–3). However, there is only a
most of the ASS1 when the process is out‐of‐control is slight increase in the ASS compared with that of the
larger than the sample size of the FSS T2 short‐run FSS T2 short‐run chart, and the slight increase in cost
CHONG ET AL. 23

TABLE 3 TARL1(δ) minimization with charts' parameters and results of performance measures, for I = 50, TARL0 = I, ASS0 ∈ f5; 7; 9g,
p = 2, and δ ∈ {0.1, 0.2, 0.3, 0.4, 0.5, 0.6, 0.7, 0.8, 0.9, 1.0, 2.0}

FSS VSS
δ ns nL CL W TARL1 TSDRL1 P(I) TARL1 TSDRL1 ASS1 P(I)

ASS0 = 5
0.1 2 31 14.2755 4.4820 49.82 6.21 0.05 49.80 6.26 5.17 0.04
0.2 2 31 14.2755 4.4820 49.21 7.58 0.07 48.69 8.50 5.76 0.09
0.3 2 31 14.2755 4.4820 47.98 9.67 0.12 44.23 13.52 6.95 0.26
0.4 2 31 14.2755 4.4820 45.81 12.29 0.20 31.54 17.63 8.78 0.67
0.5 2 31 14.2755 4.4820 42.31 15.04 0.32 15.90 12.37 10.26 0.97
0.6 2 31 14.2755 4.4820 37.21 17.29 0.49 8.74 6.45 10.17 1.00
0.7 2 29 14.2755 4.3401 30.65 18.18 0.68 6.24 4.22 9.49 1.00
0.8 3 26 14.2755 4.8265 23.43 17.10 0.84 4.79 3.01 9.70 1.00
0.9 3 22 14.2755 4.4474 16.75 14.22 0.95 3.87 2.16 9.36 1.00
1.0 3 20 14.2755 4.2265 11.57 10.61 0.99 3.26 1.64 9.03 1.00
2.0 4 10 14.2755 3.5338 1.26 0.57 1.00 1.37 0.49 5.54 1.00
ASS0 = 7
0.1 2 31 14.2755 3.4663 49.74 6.40 0.04 49.71 6.48 7.31 0.05
0.2 2 31 14.2755 3.4663 48.85 8.27 0.08 48.01 9.56 8.34 0.11
0.3 2 31 14.2755 3.4663 46.93 11.06 0.15 41.16 15.44 10.31 0.37
0.4 2 31 14.2755 3.4663 43.40 14.32 0.28 24.95 16.70 12.88 0.83
0.5 2 31 14.2755 3.4663 37.70 17.14 0.47 11.26 8.66 14.12 0.99
0.6 3 31 14.2755 3.8405 29.95 18.17 0.69 6.42 4.23 13.77 1.00
0.7 4 31 14.2755 4.3401 21.43 16.43 0.87 4.52 2.67 13.34 1.00
0.8 5 30 14.2755 4.9917 14.13 12.55 0.97 3.56 1.95 13.22 1.00
0.9 5 27 14.2755 4.7383 9.10 8.47 0.99 2.94 1.41 12.79 1.00
1.0 5 24 14.2755 4.4474 6.05 5.52 1.00 2.53 1.07 12.20 1.00
2.0 6 13 14.2755 3.8405 1.05 0.24 1.00 1.11 0.31 6.66 1.00
ASS0 = 9
0.1 2 31 14.2755 2.7959 49.67 6.58 0.05 49.62 6.68 9.45 0.05
0.2 2 31 14.2755 2.7959 48.45 8.94 0.09 47.35 10.45 10.89 0.14
0.3 2 31 14.2755 2.7959 45.74 12.36 0.19 38.53 16.54 13.46 0.46
0.4 2 31 14.2755 2.7959 40.63 15.97 0.37 20.90 15.26 16.27 0.90
0.5 2 31 14.2755 2.7959 32.68 18.11 0.62 9.26 6.84 16.90 0.99
0.6 4 31 14.2755 3.3240 23.07 16.99 0.84 5.31 3.24 16.59 1.00
0.7 6 31 14.2755 4.1872 14.52 12.82 0.96 3.71 1.98 16.30 1.00
0.8 6 31 14.2755 4.1872 8.83 8.21 0.99 2.92 1.35 15.73 1.00
0.9 7 30 14.2755 4.8265 5.59 5.07 1.00 2.45 1.04 15.60 1.00
1.0 8 26 14.2755 5.7125 3.77 3.23 1.00 2.13 0.89 14.41 1.00
2.0 8 15 14.2755 3.8405 1.01 0.10 1.00 1.02 0.15 8.16 1.00

is negligible compared with the significant improvement of the charts is P(I). The TARL0 value is similar for both
in the sensitivity because of the adaptive procedure. FSS and VSS T2 short‐run charts, for all values of I,
Another performance measure to study the sensitivity which provides a common ground for comparison.
24 CHONG ET AL.

As the VSS T2 short‐run chart has a higher P(I) for all VSS T2 short‐run chart surpasses the FSS T2 short‐run
shift sizes, it is superior to the FSS T2 short‐run chart as the former has smaller ETSDRL1(δmin, δmax)
chart in detecting mean shifts. Furthermore, P(I) values than the latter. Similar to the results when the shifts
increases with ASS0 because of an increase in the can be specified, the outperformance of the VSS T2 short‐
sensitivities of the charts when the sample sizes are run chart requires a larger EASS1(δmin, δmax) (>n0) value
larger. when the process is out‐of‐control, for most range of shift
To account for the case where the exact shift size can- sizes. Nonetheless, the slight increase in the
not be specified, a range of shift sizes such that EASS1(δmin, δmax) value is insignificant compared with
δ~U(δmin, δmax) is adopted to study the performances of the improved sensitivity attributed to the adaptive proce-
the FSS and VSS T2 short‐run charts, for p = 2, I ∈ {10, dure. The VSS T2 short‐run chart has a higher E(P(I)) value
30, 50} and ASS0 ∈ {5, 7, 9}, as shown in Tables 4–6. The for all ranges of shifts compared with the FSS T2 short‐run
VSS T2 short‐run chart outperforms its FSS T2counterpart, chart, indicating that the former surpasses the latter in the
in terms of the mean detection speed, as the former has sensitivity of detecting mean shifts (cf., Tables 4–6). Addi-
lower ETARL1(δmin, δmax) values, except for the range of tionally, the charts become more sensitive toward shifts
large shifts when (δmin, δmax) = (1.5, 2). In addition, the in the mean vector as ASS0 increases.

TABLE 4 ETARL1(δmin, δmax) minimization with charts' parameters and results of performance measures, for I = 10, TARL0 = I,
ASS0 ∈ f5; 7; 9g, p = 2, and δ~U(δmin, δmax)

FSS VSS
δmin δmax ns nL CL W ETARL1 ETSDRL1 E(P(I)) ETARL1 ETSDRL1 EASS1 E(P(I))

ASS0 = 5
0.1 0.3 2 31 7.9003 4.0169 9.64 2.86 0.24 9.53 2.88 5.56 0.29
0.1 0.5 2 31 7.9003 4.0169 9.03 3.18 0.34 8.53 3.16 6.03 0.46
0.1 0.7 2 26 7.9003 3.6885 8.18 3.37 0.46 7.43 3.11 6.23 0.61
0.1 1 3 20 7.9003 3.7946 6.78 3.22 0.62 6.08 2.72 6.39 0.72
0.5 0.8 3 20 7.9003 3.7946 5.95 3.63 0.76 4.67 2.80 6.82 0.93
0.5 1 3 18 7.9003 3.5747 4.98 3.25 0.84 3.96 2.33 6.76 0.96
1 1.5 4 11 7.9003 3.4521 1.87 1.25 1.00 1.77 0.84 5.80 1.00
1.5 2 4 9 7.9003 2.8420 1.14 0.38 1.00 1.22 0.42 4.72 1.00
ASS0 = 7
0.1 0.3 2 31 7.9003 3.1135 9.46 2.97 0.27 9.30 2.99 7.95 0.34
0.1 0.5 2 31 7.9003 3.1135 8.61 3.32 0.40 7.97 3.16 8.53 0.55
0.1 0.7 3 31 7.9003 3.4521 7.51 3.38 0.54 6.72 2.95 8.83 0.68
0.1 1 4 26 7.9003 3.5349 5.96 2.97 0.69 5.37 2.44 9.02 0.78
0.5 0.8 4 23 7.9003 3.2727 4.74 3.26 0.87 3.67 2.17 9.33 0.98
0.5 1 5 22 7.9003 3.7946 3.84 2.70 0.92 3.10 1.83 9.30 0.98
1 1.5 6 14 7.9003 3.6885 1.41 0.74 1.00 1.42 0.59 7.51 1.00
1.5 2 6 11 7.9003 2.8420 1.04 0.18 1.00 1.07 0.23 6.28 1.00
ASS0 = 9
0.1 0.3 2 31 7.9003 2.4934 9.29 3.08 0.30 9.08 3.07 10.28 0.38
0.1 0.5 2 31 7.9003 2.4934 8.21 3.40 0.46 7.53 3.11 10.82 0.61
0.1 0.7 5 31 7.9003 3.3194 6.93 3.31 0.61 6.18 2.78 11.31 0.72
0.1 1 6 31 7.9003 3.7601 5.36 2.74 0.74 4.85 2.24 11.57 0.81
0.5 0.8 6 27 7.9003 3.4521 3.84 2.81 0.93 3.03 1.76 11.87 0.99
0.5 1 7 25 7.9003 3.8939 3.08 2.21 0.96 2.56 1.47 11.52 0.99
1 1.5 8 16 7.9003 3.6885 1.21 0.47 1.00 1.23 0.44 9.05 1.00
1.5 2 8 14 7.9003 3.1750 1.01 0.08 1.00 1.02 0.12 8.13 1.00
CHONG ET AL. 25

TABLE 5 ETARL1(δmin, δmax) minimization with charts' parameters and results of performance measures, for I = 30, TARL0 = I,
ASS0 ∈ f5; 7; 9g, p = 2, and δ~U(δmin, δmax)

FSS VSS
δmin δmax ns nL CL W ETARL1 ETSDRL1 E(P(I)) ETARL1 ETSDRL1 EASS1 E(P(I))

ASS0 = 5
0.1 0.3 2 31 12.2418 4.4289 29.43 5.68 0.11 28.86 6.28 5.83 0.15
0.1 0.5 2 31 12.2418 4.4289 28.11 7.04 0.18 24.22 8.05 6.98 0.40
0.1 0.7 2 31 12.2418 4.4289 25.80 8.26 0.31 18.86 7.31 7.60 0.59
0.1 1 2 31 12.2418 4.4289 21.14 8.52 0.51 14.14 5.90 7.90 0.73
0.5 0.8 2 31 12.2418 4.4289 19.27 10.66 0.64 7.21 5.13 8.70 0.99
0.5 1 2 28 12.2418 4.2147 15.57 9.71 0.76 6.02 4.10 8.60 0.99
1 1.5 4 15 12.2418 4.6818 3.79 3.22 1.00 2.27 1.02 7.83 1.00
1.5 2 4 11 12.2418 3.7945 1.47 0.82 1.00 1.48 0.55 5.98 1.00
ASS0 = 7
0.1 0.3 2 31 12.2418 3.4234 29.09 6.13 0.13 28.20 6.88 8.44 0.19
0.1 0.5 2 31 12.2418 3.4234 27.08 7.77 0.24 22.22 8.08 10.07 0.48
0.1 0.7 2 31 12.2418 3.4234 23.72 8.74 0.41 16.84 6.71 10.55 0.65
0.1 1 2 31 12.2418 3.4234 18.29 7.94 0.60 12.48 5.18 10.47 0.77
0.5 0.8 3 31 12.2418 3.7945 14.80 9.97 0.81 5.27 3.40 11.95 1.00
0.5 1 4 31 12.2418 4.2888 11.26 8.10 0.88 4.36 2.70 12.38 1.00
1 1.5 6 18 12.2418 4.8517 2.29 1.70 1.00 1.79 0.74 9.79 1.00
1.5 2 6 14 12.2418 4.0574 1.15 0.39 1.00 1.20 0.39 7.21 1.00
ASS0 = 9
0.1 0.3 2 31 12.2418 2.7575 28.73 6.55 0.15 27.59 7.34 10.99 0.23
0.1 0.5 2 31 12.2418 2.7575 25.97 8.33 0.30 20.86 7.98 12.83 0.54
0.1 0.7 2 31 12.2418 2.7575 21.77 8.79 0.49 15.61 6.34 13.04 0.69
0.1 1 3 31 12.2418 2.9934 16.18 7.30 0.66 11.45 4.69 13.15 0.79
0.5 0.8 5 31 12.2418 3.6484 11.28 8.59 0.90 4.26 2.53 15.06 1.00
0.5 1 6 31 12.2418 4.1376 8.35 6.51 0.94 3.55 2.00 15.11 1.00
1 1.5 8 21 12.2418 5.0076 1.67 1.03 1.00 1.51 0.59 11.35 1.00
1.5 2 8 16 12.2418 4.0574 1.05 0.20 1.00 1.08 0.24 8.54 1.00

6 | A RE A L A P P L I C A T I O N run with a production horizon, H = 9 hours. Due to eco-


nomic constraints, the quality engineer makes only I = 8
This example illustrates the implementation of the VSS inspections with ASS0 = 4 when the process is in‐con-
T2 short‐run chart to monitor the impurity profile of a trol. As the quality engineer is unable to specify the
crystalline drug substance, which is multivariate in exact shift size δ for which a quick detection is required,
nature. The case study from the chemical and pharma- the optimal parameters of the chart are acquired by
ceutical industry is obtained from Gonzalez and minimizing ETARL1(δmin, δmax), where (δmin, δmax) = (0.1,
Rodriguez.41 The quality characteristics monitored in 2) is considered. Here, it is assumed that the shift size
this example are the levels of the two organic impurities; follows a uniform distribution over the interval
ie, the number of variables is p = 2, denoted by B and C (δmin, δmax), ie, δ ~ U(0.1, 2). Consequently, nS = 3,
in the case study. It is assumed that the crystalline drug nL = 12, CL = 7.0296, and W = 3.7064 are obtained.
substance is manufactured during a short production The target values, μ0 and Σ0, when the process is in‐
26 CHONG ET AL.

TABLE 6 ETARL1(δmin, δmax) minimization with charts' parameters and results of performance measures, for I = 50, TARL0 = I,
ASS0 ∈ f5; 7; 9g, p = 2, and δ~U(δmin, δmax)

FSS VSS
δmin δmax ns nL CL W ETARL1 ETSDRL1 E(P(I)) ETARL1 ETSDRL1 EASS1 E(P(I))

ASS0 = 5
0.1 0.3 2 31 14.2755 4.4820 49.37 7.74 0.07 48.41 9.04 5.89 0.11
0.1 0.5 2 31 14.2755 4.4820 47.60 10.06 0.13 39.75 12.50 7.33 0.38
0.1 0.7 2 31 14.2755 4.4820 44.11 12.43 0.25 29.69 10.69 8.25 0.59
0.1 1 2 31 14.2755 4.4820 36.22 13.41 0.46 21.45 8.19 8.51 0.72
0.5 0.8 2 31 14.2755 4.4820 33.84 17.40 0.58 8.25 6.04 9.78 1.00
0.5 1 2 31 14.2755 4.4820 27.14 16.10 0.72 6.73 4.76 9.46 1.00
1 1.5 4 17 14.2755 5.0694 5.63 5.05 1.00 2.44 1.06 8.80 1.00
1.5 2 4 12 14.2755 4.1061 1.75 1.13 1.00 1.61 0.57 6.68 1.00
ASS0 = 7
0.1 0.3 2 31 14.2755 3.4663 48.94 8.47 0.09 47.42 10.13 8.55 0.15
0.1 0.5 2 31 14.2755 3.4663 46.10 11.40 0.19 36.47 12.55 10.64 0.46
0.1 0.7 2 31 14.2755 3.4663 40.71 13.53 0.35 26.68 9.94 11.45 0.64
0.1 1 2 31 14.2755 3.4663 31.21 12.56 0.56 19.11 7.38 11.27 0.76
0.5 0.8 3 31 14.2755 3.8405 25.89 16.75 0.77 6.04 3.99 13.28 1.00
0.5 1 4 31 14.2755 4.3401 19.33 13.53 0.86 4.92 3.10 13.43 1.00
1 1.5 6 21 14.2755 5.3525 3.07 2.50 1.00 1.94 0.76 11.24 1.00
1.5 2 6 15 14.2755 4.3401 1.25 0.53 1.00 1.29 0.44 7.84 1.00
ASS0 = 9
0.1 0.3 2 31 14.2755 2.7959 48.46 12.49 0.10 46.51 10.97 11.15 0.18
0.1 0.5 2 31 14.2755 2.7959 44.41 12.49 0.25 34.33 12.47 13.56 0.51
0.1 0.7 2 31 14.2755 2.7959 37.39 13.83 0.44 24.89 9.55 14.10 0.67
0.1 1 2 31 14.2755 2.7959 27.51 11.55 0.62 17.74 6.94 13.48 0.78
0.5 0.8 5 31 14.2755 3.6931 19.42 14.53 0.88 4.98 3.08 16.50 1.00
0.5 1 5 31 14.2755 3.6931 14.01 10.85 0.93 4.03 2.27 15.95 1.00
1 1.5 8 24 14.2755 5.4800 2.07 1.46 1.00 1.64 0.63 12.90 1.00
1.5 2 8 18 14.2755 4.5493 1.09 0.28 1.00 1.12 0.30 9.02 1.00

control, are obtained from the historical Phase‐I dataset, current T2 statistic lies in the warning region, the next
given by sample with a large sample size is taken. On the
contrary, if the current T2 statistic lies in the central
   
74 4314:483 2775:862 region, the next sample with a small sample size is
μ0 ¼ and ∑0 ¼ :
341:667 2775:862 16883:333 taken. It is clearly seen that in this short production
runs case study, an out‐of‐control condition is signaled
Table 7 shows the sample size n(i), sample mean by the VSS T2 short‐run chart at the 7th sample
vector X i , and the T2 statistic for the crystalline drug (see the boldfaced T 2i value in Table 7 and the
substance. At each inspection, either nS = 3 or nL = 12 point plotted above the CL in Figure 4). Following this
bivariate observations are sampled depending on the out‐of‐control signal, the quality engineer initiates a
location in which the previous T2 statistic plots on the search for assignable causes and takes appropriate cor-
chart. Process monitoring begins by taking a rective actions in order to bring the process back to
sample with the small sample size, ie, n(1) = nS. If the the in‐control state.
CHONG ET AL. 27

TABLE 7 Phase II dataset from a crystalline drug substance and I. The study of the statistical performance of the
process charts is also extended to the case when the shift size
i n(i) Xi Ti2 Region
cannot be specified, where the selected range of the
  shifts is modeled by a uniform distribution. In general,
1 3 160 7.019 Warning the mean detection speed of shifts of the VSS T2 short‐
680 run chart is higher than that of the FSS T2 short‐run
 
2 12 134:167 4.695 Warning chart, except for the range of large shifts,
621:667 (δmin, δmax) = (1.5, 2.0). Additionally, when the shift size
  can be specified and when it cannot be specified, the
3 12 84:167 0.432 Central
426:667 VSS T2 short‐run chart is superior to the FSS T2 short‐
  run chart in terms of the TSDRL and ETSDRL criteria.
4 3 80 1.589 Central
500
As the VSS T2 short‐run chart requires a more intensive
  sampling, there is a slight increase in the ASS when the
5 3 66:667 3.508 Central
process is out‐of‐control, which leads to a slight increase
566:667
  in the cost. However, the minimal increase in cost due
6 3 106:667 1.558 Central to a larger ASS is negligible as compared with the signif-
503:333 icant improvement in the sensitivity when the adaptive
 
7 3 243:333 10.031 Out‐of‐control strategy is applied.
676:667 Future research can involve presenting the VSI,
  VSSI, and variable parameters (VP) T2 charts for a
8 3 13:333 1.037 Central
250 finite production horizon. In addition, the multivariate
cumulative sum (MCUSUM) and multivariate EWMA
(MEWMA) schemes can also be studied for short produc-
tion runs to reduce the detection time in signaling small
to moderate shifts.

ACKNOWLEDGEMENTS
This research is supported by the Graduate Assistant
Scheme awarded by Universiti Sains Malaysia and is
funded by Universiti Sains Malaysia, Research University
Grant, number 1001.PMATHS.8011039.

FIGURE 4 The variable sample size T2 chart for short ORCID


production runs corresponding to the crystalline drug substance
process Michael B. C. Khoo http://orcid.org/0000-0002-3245-1127
Abdul Haq http://orcid.org/0000-0002-4467-9719

RE FER EN CES
7 | CONCLUSIONS
1. Tagaras G. A survey of recent developments in the design of
In this paper, we have developed the FSS and VSS ver- adaptive control charts. J Qual Technol. 1998;30(3):212‐231.
sions of the T2 chart to monitor a finite production hori-
2. Khaw KW, Khoo MBC, Yeong WC, Wu Z. Monitoring the coef-
zon. The charts are studied in terms of the TARL,
ficient of variation using a variable sample size and sampling
TSDRL, ASS, and P(I) when the exact shift size can be interval control chart. Commun Stat Simul Comput.
specified and in terms of the ETARL, ETSDRL, EASS, 2017;46(7):5772‐5794.
and E(P(I)) when the shift size cannot be specified.
The number of variables p = 2 is considered. Based on 3. Lee MH, Khoo MBC. Multivariate synthetic |S| control chart
with variable sampling interval. Commun Stat Simul Comput.
the comparison between the FSS and VSS T2 short‐run
2015;44(4):924‐942.
charts, it is found that the VSS T2 short‐run chart sur-
passes its FSS counterpart for most shift sizes (except 4. Lin YC. The variable parameters X control charts for monitoring
δ = 0.1 and 2, where the two charts perform equally autocorrelated processes. Commun Stat Simul Comput.
well) and the outperformance increases with the ASS0 2009;38(4):729‐749.
28 CHONG ET AL.

5. Zhou M. Variable sample size and variable sampling interval 23. Nenes G, Tagaras G. Evaluation of CUSUM charts for finite‐
Shewhart control chart with estimated parameters. Oper Res. horizon processes. Commun Stat Simul Comput.
2017;17(1):17‐37. 2010;39(3):578‐597.
6. Prabhu SS, Runger GC, Keats JB. X chart with adaptive sample 24. Celano G, Castagliola P, Trovato E, Fichera S. Shewhart and
sizes. Int J Prod Res. 1993;31(12):2895‐2909. EWMA t control charts for short production runs. Qual Reliab
7. Costa AFB. X charts with variable sample size. J Qual Technol. Eng Int. 2011;27(3):313‐326.
1994;26(3):155‐163. 25. Castagliola P, Celano G, Fichera S, Nenes G. The variable sam-
8. Teoh WL, Chong JK, Khoo MBC, Castagliola P, Yeong WC. ple size t control chart for monitoring short production runs. Int
Optimal designs of the variable sample size X chart based on J Adv Manuf Technol. 2013;66(9):1353‐1366.
median run length and expected median run length. Qual Reliab 26. Nenes G, Castagliola P, Celano G, Panagiotidou S. The variable
Eng Int. 2017;33(1):121‐134. sampling interval control chart for finite‐horizon processes. IIE
9. Hu X, Castagliola P, Sun J, Khoo MBC. The performance of var- Trans. 2014;46(10):1050‐1065.
iable sample size X chart with measurement errors. Qual Reliab 27. Castagliola P, Amdouni A, Taleb H, Celano G. One‐sided
Eng Int. 2016;32(3):969‐983. Shewhart‐type charts for monitoring the coefficient of variation
10. Castagliola P, Zhang Y, Costa A, Maravelakis P. The variable in short production runs. Qual Technol Quan Manage.
sample size X chart with estimated parameters. Qual Reliab 2015;12(1):53‐67.
Eng Int. 2012;28(7):687‐699. 28. Amdouni A, Castagliola P, Taleb H, Celano G. Monitoring the
11. Castagliola P, Achouri A, Taleb H, Celano G, Psarakis S. coefficient of variation using a variable sample size control chart
Monitoring the coefficient of variation using a variable in short production runs. Int J Adv Manuf Technol.
sample size control chart. Int J Adv Manuf Technol. 2015;80(9): 2015;81(1):1‐14.
1561‐1576. 29. Amdouni A, Castagliola P, Taleb H, Celano G. A variable sam-
12. Yeong WC, Khoo MBC, Lim SL, Lee MH. A direct procedure for pling interval Shewhart control chart for monitoring the
monitoring the coefficient of variation using a variable sample coefficient of variation in short production runs. Int J Prod
size scheme. Commun Stat Simul Comput. 2017;46(6):4210‐4225. Res. 2017;55(19):5521‐5536.
13. Amiri A, Nedaie A, Alikhani M. A new adaptive variable sample 30. Aparisi F. Hotelling's T2 control chart with adaptive sample
size approach in EWMA control chart. Commun Stat Simul sizes. Int J Prod Res. 1996;34(10):2853‐2862.
Comput. 2014;43(4):804‐812.
31. Aparisi F, Haro CL. Hotelling's T2 control chart with variable
14. Mahadik SB. X charts with variable sample size, sampling sampling intervals. Int J Prod Res. 2001;39(14):3127‐3140.
interval, and warning limits. Qual Reliab Eng Int. 2013;29(4):
32. Aparasi F, Haro CL. A comparison of T2 control charts with var-
535‐544.
iable sampling schemes as opposed to MEWMA chart. Int J Prod
15. Wu Z, Luo H. Optimal design of the adaptive sample size and Res. 2003;41(10):2169‐2182.
sampling interval np control chart. Qual Reliab Eng Int.
33. Mahadik SB. Shirke DT. A special variable sample size and sam-
2004;20(6):553‐570.
pling interval Hotelling's T2 chart. Int J Adv Manuf Technol.
16. Noorossana R, Maryam SA. Monitoring two dependent process 2011;53(1‐4):379‐384.
steps using special variable sample sizes and sampling intervals
cause‐selecting control charts. Qual Reliab Eng Int. 2012;28(4): 34. Seif A, Faraz A, Heuchenne C, Saniga E, Moghadam MB. A
437‐453. modified economic‐statistical design of the T2 control chart with
variable sample sizes and control limits. J Appl Stat.
17. Mahadik SB. Variable sample size and sampling interval X 2011;38(11):2459‐2469.
charts with runs rules for switching between sample sizes and
sampling interval lengths. Qual Reliab Eng Int. 2013;29(1):63‐76. 35. Mahadik SB. Variable sampling interval Hotelling's T2 charts
with runs rules for switching between sampling interval lengths.
18. Cheng XB, Wang FK. VSSI median control chart with estimated
Qual Reliab Eng Int. 2012;28(2):131‐140.
parameters and measurement errors. Qual Reliab Eng Int.
2018;1‐15. 36. Chen YK, Hsieh KL. Hotelling's T2 charts with variable
sample size and control limit. Eur J Oper Res. 2007;182(3):
19. Aslam M, Arif OH, Jun CH. A new variable sample size control
1251‐1262.
chart using MDS sampling. J Stat Comput Simul. 2016;86(18):
3620‐3628. 37. Faraz A, Moghadam MB. Hotelling's T2 control chart with two
adaptive sample sizes. Qual Quant. 2009;43(6):903‐912.
20. Zhou W, Lian Z. Optimum design of a new VSS‐NP chart
with adjusting sampling inspection. Int J Prod Econ. 38. Aparisi F, Epprecht E, Carrión A, Ruiz O. The variable sample
2011;129(1):8‐13. size variable dimension T2 control chart. Int J Prod Res.
2014;52(2):368‐383.
21. Nyau SY, Lee MH, Wong DML. Optimal statistical design of var-
iable sample size multivariate exponentially weighted moving 39. Sparks RS. CUSUM charts for signalling varying location shifts.
average control chart based on median run‐length. Qual Technol J Qual Technol. 2000;32(2):157‐171.
Quan Manage. 2017;14(4):478‐495. 40. Celano G, Castagliola P, Fichera S, Nenes G. Performance of t
22. Lee MH. Multivariate EWMA control chart with adaptive sam- control charts in short runs with unknown shift sizes. Comput
ple sizes. Commun Stat Simul Comput. 2010;39(8):1548‐1561. Ind Eng. 2013;64(1):56‐68.
CHONG ET AL. 29

41. González M, Rodriguez P. Application of the multivariate T2


control chart and the Mason–Tracy–Young decomposition pro-
Islamabad, Pakistan, in 2010. He obtained his PhD
cedure to the study of the consistency of impurity profiles of from the School of Mathematics and Statistics, Uni-
drug substances. Qual Eng. 2003;16(1):127‐142. versity of Canterbury, Christchurch, New Zealand, in
2014. He is a lecturer at the Department of Statistics,
QAU, Islamabad, Pakistan. His research interests are
Nger Ling Chong is currently a full‐time PhD stu- quality control charts and their applications, ranked
dent in the School of Mathematical Sciences, set sampling, and Bayesian inference.
Universiti Sains Malaysia and she is doing her PhD
in the field of Statistical Quality Control. She obtained Philippe Castagliola is a professor in Université de
her Bachelor of Applied Science (Statistics) degree Nantes, Institut Universitaire de Technologie de
with honors and Master of Science (Statistics) from Nantes, France. He is also a member of the IRCCyN
Universiti Sains Malaysia. (Institut de Recherche en Communications et
Cybernétique de Nantes), UMR CNRS 6597. His
Michael B. C. Khoo is a professor of Statistics in the research activity involves development of new SPC
School of Mathematical Sciences, Universiti Sains techniques (non‐normal control charts, optimized
Malaysia. His research area is Statistical Quality Con- EWMA type control charts, control charts with esti-
trol. He publishes extensively in numerous interna- mated parameters, multivariate capability indices,
tional journals and is a member of the American monitoring of batch processes, etc).
Society for Quality (ASQ) and the Malaysian Mathe-
matical Sciences Society.
How to cite this article: Chong NL, Khoo MBC,
Abdul Haq gained his MS in Statistics from the Haq A, Castagliola P. Hotelling's T2 control charts
Department of Statistics, Quaid‐i‐Azam University with fixed and variable sample sizes for monitoring
(QAU), Islamabad, Pakistan, in 2007, and MPhil in short production runs. Qual Reliab Engng Int.
Statistics from the Department of Statistics, QAU, 2019;35:14–29. https://doi.org/10.1002/qre.2377

You might also like