1.

04 Statistical Control of Measures and Processes

A. J. Ferrer-Riquelme, Technical University of Valencia, Valencia, Spain
ª 2009 Elsevier B.V. All rights reserved.

1.04.1 Introduction: Basics of Process Monitoring 100

1.04.2 Phases in Process Monitoring 102
1.04.3 Shewhart Control Charts 102
1.04.4 CUSUM Control Charts 104
1.04.5 EWMA Control Charts 105
1.04.6 Performance Measures of Control Charts 106
1.04.6.1 Example 1 106
1.04.7 Control Charts for Autocorrelated Processes 108
1.04.7.1 Example 2 108
1.04.8 Integration of SPC and Engineering Process Control 111
1.04.9 Multivariate Control Charts 112
1.04.9.1 Original Variables-Based MSPC Schemes 112
1.04.9.1.1 Hotelling T 2 control charts for monitoring process mean 113
1.04.9.1.2 Multivariate Shewhart-type control charts for monitoring process variability 114
1.04.9.1.3 MEWMA control charts 114
1.04.9.1.4 Fault diagnosis 115
1.04.9.2 Latent Variables-Based MSPC Schemes 115
1.04.9.2.1 PCA-based MSPC: exploratory data analysis and offline process monitoring (Phase I) 116
1.04.9.2.2 PCA-based MSPC: online process monitoring (Phase II) 119
1.04.9.3 Potential of Latent Variables-Based MSPC in Industry 120
1.04.9.3.1 Example 3 121
1.04.9.4 Monitoring Measurement Systems: Applications of SPC in Analytical Laboratory
Methods 121
1.04.10 Software 123
References 124

Symbols b sample mean of the

A dimension of the SPE sample
principal component BðK=2;ðm – K – 1Þ=2Þ; 100(1–)% percen-
subspace tile of the beta
A3 ,B3 , B4 , and c4 coefficients for the distribution with K/2
calculation of the and (m–K–1)/2
control limits of the degrees of freedom
x – s control chart c correction factor for
A2 ,D3 ,D4 , and d2 coefficients for the DModX
calculation of the Ct CUSUM control
control limits of the statistic at stage t
x – R control chart (definition 1)
at  Nð0;2 Þ series of indepen- number of noncon-
dent and identically formities in samples
normally distributed of size n at stage t
(iind) random distur- (definition 2)
bances with zero ContðDModX;xnew;k Þ contribution of kth
mean and variance variable to the
2

97
98 Statistical Control of Measures and Processes

DModX in a new p population fraction

observation nonconforming
ContðSPE;xnew;k Þ contribution of kth P
m P
m
average sample
¼
p Dt =nm ¼ p̂t =m
variable to the SPE in t¼1 t¼1 fraction
a new observation nonconforming
Contðtnew;a ;xnew;k Þ contribution of kth pak loading of the kth
variable to the ath variable at the
score at a new ath principal com-
observation ponent
D magnitude of the p̂t sample fraction
shift in the process nonconforming at
parameter stage t
 ¼ P Rt =m
m
Dt number of noncon- R average range
forming units in t¼1
samples of size n at rk sample autocorrela-
stage t tion function
E (m  K) residual S (K  K) sample
matrix variance–
enew,k residual correspond- covariance matrix
ing to the kth variable jSj sample generalized
in a new observation variance
Et EWMA control sta- P
m
average standard
s ¼ st =m
tistic at stage t t¼1
qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
ffi deviation
Pn sample standard
ei residual vector (ith st ¼  2
i¼1 ðxti – xt Þ =n – 1
row of matrix E) deviation at stage t
et residual at stage t s0 pooled residual
FðK;ðmn – m – Kþ1ÞÞ; 100(1–)% percen- standard deviation
tile of the F T (m  A) score matrix
distribution with K tnew,a score of a new
and mn–m–Kþ1 observation at the
degrees of freedom ath principal
f(x; ) probability distribu- component
tion of x tr S trace of the covar-
H decision value iance matrix
(threshold) in the T t2 Hotelling T2
CUSUM control control statistic at
chart stage t
K reference value in the tTi ¼ fti1 ;ti2 ;:::;tiA g A-dimensional
CUSUM control score vector (ith
chart row of matrix T)
L distance of the con- P
m P
m
average sample
¼
u Ct =nm ¼ ut =m
trol limits from the t¼1 t¼1 number of noncon-
centerline formities per unit
m number of samples ut sample average
P
m
average of the mov- number of noncon-
MR ¼ MRð2Þt =ðm – 1Þ
t¼2 ing ranges formities per unit at
n sample size stage t
P (K  A) loading v sample variance of
matrix the SPE sample
 Statistical Control of Measures and Processes 99

vt ¼ v(zt) control statistic to be xnew;k from the PCA

plotted in the control model
xTt ¼ fxt1 ; xt2 ;:::; xtn g observations vector
chart
wk square root of the of the key character-
explained sum of istic on n units
squares for the kth sampled randomly
variable from the production
wt MEWMA control sta- or measurement
tistics at stage t process at stage t
x̂tþ1=t one-step ahead
Wt multivariate control
statistic for monitor- forecast of x made
ing process at stage t
variability (a direct Z (m  K) data matrix
zt ¼ – ln½fðxt ; 0 Þ=fðxt ; 1 Þ likelihood ratio
extension of the uni-
variate s2 control statistic
zt ¼ MRð2Þt ¼ jxt – xt – 1 j moving range of
chart)
x key characteristic to two consecutive
be measured observations at
X preprocessed Z stage t
zt ¼ Rt ¼ maxðxt Þ – minðxt Þ sample range
(m  K) data matrix P
 zt ¼ s2t ¼ ni¼1 ðxti – 
xt Þ2 =n – 1 sample variance
X (m  K) prediction Pn
zt ¼ 
xt ¼ i¼1 xti =n sample mean
matrix
xNð; SÞ x follows a multivari- zt ¼ z(xTt ;0 ) suitable statistic
ate normal (or seed) to build
distribution with the chart on.
Y (A  A) covariance
mean vector m and
covariance matrix S matrix of T
jS j generalized
x̄¯ (K1) overall sample
mean vector. variance
x̄t (K  1) vector of  Type I risk
sample means at  number of standard
stage t deviations of the
Xt ¼ fx1t ;x2t ;:::;xKt g (n  K) observations shift in the process
matrix of K charac- parameter
teristics measured  process parameter
on n units sampled at to be monitored
0 in-control (target or
stage t
P
m
reference) value of 
x̄¯ ¼ xt =m grand average
t¼1 ˆt minimum variance
xTkt ¼ fxk1t ; xk2t ;:::; xknt g n-dimensional vector unbiased estimator
of sample values of of  based on the
kth variable (kth col- sample values xTt ¼
umn of matrix X) fxt1 ;xt2 ;:::;xtn g
xTnt ¼ fxn1t ; xn2t ;:::; xnKt g K-dimensional  smoothing constant
observation vector of of the EWMA
nth unit (nth row of control statistic
matrix X) (definition 1)
xnew;k prediction of the kth population average
variable in a new number of
observation
100 Statistical Control of Measures and Processes

nonconformities in  process mean

samples of size n  autoregressive
(definition 2) coefficient of
smoothing constant AR(1) model
of the MEWMA con- 2K; 100(1–)% percen-
trol statistic tile of the
(definition 3) 2 distribution with
a eigenvalue of the ath K degrees of
freedom
principal component

1.04.1 Introduction: Basics of Process Monitoring

Processes, whether chemical, biochemical, or analytical in nature, are subject to certain variability. You
will not always get the same result each time no matter how tightly you operate a process. This is
exactly what any customer perceives from the output (i.e., product, service, or information) of any process.
If this output is to meet or exceed customer expectations, it should be produced by a process that is stable
or repeatable.1
Variation remaining in a stable process reflects ‘common causes’ that are an inherent part of the process and
which cannot be removed easily without fundamental changes in the process itself: for example, the design and
operating settings of a polymerization reactor contribute to the common-cause variation of the viscosity of a
polymer. As long as you keep the process the same, common-cause variation is the same from day to day,
yielding a process consistent over time. A process affected by only common (random) causes of variation is said
to be ‘in statistical control’. This means that the process is predictable, that is, you can predict what the process
will make in the near future, for example, how many batches will meet the customer specifications for a
particular key property.
In addition to the common-cause system, other types of variability may occasionally affect the stability of a
process: for example, operator errors, defective raw material, improperly adjusted or controlled equipment,
sudden change in environmental conditions, and so on. These sources of variability are not part of the common-
cause system and, thus, they are usually referred to as ‘assignable’ or ‘special causes’. They are sporadic in
nature and cannot be predicted. A process also affected by this type of special causes of variation is said to be
‘out of control’ in the sense that it is unstable and, thus, unpredictable.
It is crucial to distinguish between the two sources of variation because sharing the responsibility for
process improvement depends on what type of variation is present: Frontline personnel are responsible for
finding and taking decisions on assignable causes; common-cause variation, on the contrary, is management’s
responsibility.
Statistical process control (SPC) is a powerful collection of problem-solving tools useful in understanding
process variability, achieving process stability, and improving process performance through the reduction of
avoidable variability (i.e., special causes). The goal of any SPC scheme is to monitor the performance of a
process along time to check whether the process behaves as it is expected to do (i.e., the predicted behavior
from the common-cause system), and to detect any unusual (special) event that may occur. By finding
assignable causes for them, significant improvements in the process performance can be achieved by eliminat-
ing these causes (or implementing them if they are beneficial). Control charts are the essential tools for pursuing
this goal.
A control chart is a picture of a process over time that helps to identify the magnitude and the type of
variation present. To implement a control chart, one must register data from a process over time. Decisions are
required about the variables to measure, the sample statistics to be monitorized, the sample size, the time
between samples, the type of control chart, the control limits, and the decision rules. All these choices
 Statistical Control of Measures and Processes 101

determine the cost of the monitoring scheme and its ability to detect certain out-of-control conditions. The
design of a control chart involves the following steps:

1. Identifying the key characteristic x to be measured (e.g., a quality property, a key process variable that
in some way affects the quality of a product, or an analytical measurement of a control or reference
sample), the process parameter  to be monitored (e.g., process mean, variance, proportion of
defectives or nonconforming items, rate of defects or nonconformities), and the probability distribu-
tion of x, f (x; ) (e.g., normal, binomial, or Poisson). In this context, a process is said to be in
statistical control (i.e., in control) if the underlying statistical model representing the key characteristic
is stable over time. In this case, the process is said to be operated under normal operating conditions
(NOCs). If there is some change over time in this statistical model, the process is said to be out of
control.2
2. Selecting a suitable statistic (or seed) zt ¼ zðxT t ; 0 Þ to build the chart, where 0 is the in-control (target or
reference) value of the monitorized process parameter and xT t ¼ fxt 1 ;xt 2 ; . . .; xtn g is the observations vector
of the key characteristic on n units sampled randomly from the production or measurement process at stage t
(t ¼ 1, 2, . . .). In general, seeds of the form zt ¼ ˆt , where ˆt is the minimum variance unbiased estimator of 
based on the sample values xT t ¼ fxt1 ; xt 2 ; . . .; xtn g, are often considered. Another choice is the likelihood
ratio statistic, zt ¼ – ln½f ðxt ; 0 Þ=f ðxt ; 1 Þ, motivated by the Neyman–Pearson theory of testing statistical
hypotheses (H0:  ¼ 0 versus H1:  ¼ 1).3
The observations in xT t can be measurements (e.g., filling volumes, weights, viscosities, particle sizes, and
analyte sample concentrations) or counts (e.g., defective/nondefective item or number of defects) yielding
the so-called control charts for variables or attributes, respectively. When observations are measurements,
sample size n is small with values ranging from 1 to 5; when dealing with counts, sample size n is usually large
(e.g., 50 or even more).
Standard control charts for variables assume that the key characteristic x at stage t is a normally distributed
random variable that can be expressed as xt ¼  þ at , where  is the process mean and at follows a stochastic
‘white noise’ process, that is, a series of independent and identically normally distributed (iind) random
disturbances with zero mean and variance 2 (at  N(0, 2)). Therefore, in control, the observations Pn in xt
are assumed to be iind with mean 0 and variance 20 . The sample mean zt ¼ xt ¼ i¼1 xti =n is the
seed usually chosen to monitor the process mean  ¼ . To monitor the process variance  ¼ 2, several
seeds have been proposed: for example, the sample range zt ¼ Rt ¼ maxðxt Þ – minðxt Þ or the sample variance
P
zt ¼ st2 ¼ ni¼1 ðxti – xt Þ2 =n – 1
Control charts for attributes are based on counts (attributes or discrete data). There are two kinds of
attributes data: (i) yes/no and (ii) number of defects. In the yes/no case, each unit (item or sample)
inspected is either defective (i.e., it does not meet some preset criterion) or not defective (i.e., it meets
the criterion). The key characteristic x is a dichotomous random variable with only two possible
outcomes: It either passes or fails the preset criterion. Examples of this kind of yes/no data include
the following: catalyst has been correctly activated or not; polymer reaction has suffered from fouling
problems or not; batch meets the requirements on one or more product properties or not; and so on. The
statistical principles underlying this attributes control chart are based on the binomial distribution.
Suppose the production process is operating in a stable manner (i.e., in control), such that the probability
that any unit will not meet the preset criterion (i.e., the unit is defective or nonconforming) is p, and that
successive units produced are independent, then each unit produced is a realization of a Bernouilli
random variable with parameter p. If at stage t a random sample of n units is selected and if Dt is the
number of nonconforming units, then Dt has a binomial distribution with parameters n and p
(Dt  binomial (n; p)), with mean Dt ¼ np and variance 2Dt ¼ npð1 – pÞ. In this case,  ¼ p is the only
parameter to be monitorized and the seed usually chosen is the sample fraction nonconforming
zt ¼ p̂t ¼ Dt =n.
A nonconforming unit will contain at least one defect or nonconformity. If it is feasible to count these defects,
the key characteristic x is the number of defects or nonconformities in each inspected unit. The statistical
assumption in this case is that at stage t the occurrence of nonconformities in samples of size n, Ct , is well
modeled by the Poisson distribution with parameter  (Ct  Poisson ()), with mean and variance
102 Statistical Control of Measures and Processes

Ct ¼ 2Ct ¼ . This requires that the number of opportunities for defects be infinitely large, the probability of
occurrence of a nonconformity be small and constant, and that successive units produced are independent. This
process depends only on parameter  ¼  and usually the seed used is the sample average number of
nonconformities per unit zt ¼ ut ¼ Ct =n.
3. Selecting the control statistic vt ¼ v(zt) most appropriate to detect certain out-of-control situations (e.g., changes
in the parameter  or other kind of abnormalities), the control limits of the chart, and the decision rules.
The role of a control chart is to provide a procedure for a timely detection of a change in the value of .
Depending on the particular goal, to detect a pulse, exponential decrease or increase, rectangular bump
or sustained change in the parameter , a Shewhart, EWMA (exponentially weighted moving average),
MA (moving average), or CUSUM (cumulative sum) type of control chart may be constructed on
the selected seed zt, respectively.4 These various types of charts differ in the way present and past seeds zt
are weighted, leading to different control statistics vt to be plotted in a time sequence on the control chart.

1.04.2 Phases in Process Monitoring

Standard control chart usage involves two distinct phases, with two different goals: Phase I (model building)
and Phase II (model exploitation). In Phase I, a set of process data is gathered and analyzed all at once in a
retrospective way, estimating the parameters of the control statistics used and calculating trial control
limits to determine if the process has been in control over the period of time where the data were
collected. This is an iterative process where at each step the out-of-control observations are detected, carefully
analyzed, and root causes are investigated. Those out-of-control observations considered as real anomalies
are omitted from the data set and new trial control limits are recalculated from the resulting data. This
phase ends when there is enough evidence that the data gathered come from an in-control process and hence
the control limits can be used for Phase II. This means that the performance of the process has been understood
and modeled, and the assumptions of its behavior and process stability are checked. No matter whether the
collected data come from a historical database or they have been recently sampled, a wise use of Phase I control
charts can increase process understanding and process improvement, assisting operating personnel and process
managers in bringing the process into a state of statistical control.
In Phase II, real online process monitoring is carried out by plotting the control statistics from new samples on
the control charts defined in Phase I. As in Phase I, once an out-of-control sample is detected, this is carefully
studied trying to find out the nature of the assignable causes affecting the process and implementing counter-
measures to eliminate them if they deteriorate the process performance (or incorporate them in case they
improve the process performance). Process knowledge and experience from operators and process managers are
critical inputs for a successful diagnostic process. In simplex processes (e.g., univariate), the out-of-control action
plans (OCAPs) as discussed by Montgomery1 can also be very useful in expediting this diagnostic process. The
OCAPs are based on the analysis of known failure modes and provide the manual for identification and
elimination (or incorporation if beneficial) of the assignable causes in an expert system-like fashion. In complex
processes (e.g., multivariate), other diagnostic tools are needed as explained later on. As a result of this diagnostic
process, process understanding and process improvement can also be obtained.

1.04.3 Shewhart Control Charts

This type of monitoring schemes refers to any control chart developed according to the statistical principles
proposed in the 1920s by Shewhart.5 In this case, the control statistic to be charted along stages t is the
selected seed:

vt ¼ zt ¼ zðxT
t ; 0 Þ ð1Þ
 Statistical Control of Measures and Processes 103

The chart contains a centerline that represents the in-control average value of the selected seed zt and
two horizontal lines, called upper control limit (UCL) and lower control limit (LCL). These control limits
are chosen so that if the process is in control, the expected fraction of seeds beyond the control limits takes
a prespecified value . Choosing the control limits is equivalent to setting up the critical region for testing
the hypothesis H0:  ¼ 0 versus H1:  6¼ 0,  being the Type I risk of the test (i.e., the chance of wrongly
rejecting the null hypothesis when it is true; for more information on hypothesis testing, see Chapter 1.02).
Under the assumption that zt ðiindÞ N ðz ;2z Þ, the control limits are set at z  Lz , where L is the
distance of the control limits from the centerline, expressed in standard deviation units, and this depends
on . Usually,  ¼ 0.0027, leading to the classical ‘three-sigma’ control limits (L ¼ 3).
One of the most common Shewhart charts Pn for variables is the xR control chart. This is made of two charts: The x
chart uses the sample mean ðzt ¼ xt ¼ i¼1 xti =nÞ to monitor the process mean  ¼ , whereas the R chart uses
the sample range ðzt ¼ Rt ¼ maxðxt Þ – minðxt ÞÞ to monitor the process variance  ¼ 2. Control limits on both
charts are estimated from preliminary samples taken when the process is thought to be in control. To obtain good
estimates, at least m ¼ 20–25 samples are required. The control limits on the x chart are set at x  A2 R,  where
Pm
x ¼ t ¼2 xt =m is the grand average. 
The control limits on the R chart are plotted at D3 R (LCL) and D4 R (UCL), 
P
where the average range R ¼ mt ¼1Rt =m is the centerline. Assuming the process is in control, the process standard
deviation  can be estimated by R=d 2 . The values for A2, D3, D4, and d2 are tabulated for various sample sizes n.1
The sample range R, although easy to calculate by hand and easy to understand for most people, is not the best
pP ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
n
ffi
way to monitor process variability. Sample standard deviations st ¼ i¼1 ðxti – x t Þ2 =n – 1 give better estimates
of the process standard deviation, particularly for moderate to large sample sizes. The current availability of
handheld calculators and personal computers for online implementation of control charts has eliminated any
computational difficulty. If st replaces Rt as the seed, this leads to the x – s control chart. The control limits Pmon the x
 The centerline of the s chart is plotted at the average standard deviation s ¼ t¼1
chart are set at x  A3 R. st =m,
and the control limits at B3 s (LCL) and B4 s (UCL). Assuming the process is in control, the process standard
deviation  can be estimated by s =c 4 . The values for A3, B3, B4, and c4 are tabulated for various sample sizes n.1
There are many situations in which the sample size used for process monitoring is 1, that is, n ¼ 1. Some
examples are (i) every unit manufactured is analyzed (e.g., 100% inspection in highly automated parts industry
or batch processes), (ii) the process generates data on a very limited frequency (e.g., the test method used to
analyze the process is very expensive to run or takes a long time, or when dealing with accumulated quantities
such as daily energy consumption), and (iii) repeated measurements on the process differ only because of
laboratory or analysis error, as in many chemical continuous processes.
In such situations, a control chart for individual measurements is useful to monitor the process. The most
common individuals control chart is the x – MRð2Þ individuals control chart (moving range (MR)). This actually
includes two charts: The x chart displays the value of the key characteristic at stage t ðzt ¼ xt Þ to
monitor the process mean  ¼ , whereas the MR(2) chart uses the moving range of two consecutive observations
ðzt ¼ MRð2Þt ¼ jxt – xt –1 jÞ to monitorPthe process variance  ¼ 2. The MR(2) chart has centerline at the
m
average of the moving ranges MR ¼ t ¼2 MRð2Þt =ðm – 1Þ and control limits Pm at D3 MR (LCL) and D4 MR
(UCL). The control limits on the x chart are set at x  3MR=d2 , where x ¼ t ¼1 xt =m is the sample average of
all the observations. Assuming the process is in control, the process standard deviation  can be estimated by
MR=d2 . The values for D3, D4, and d2 are tabulated for various sample sizes n.1 Caution should be taken when
normality assumption cannot be assumed. In this case, the control limits may be based on resampling methods.6
Regarding control charts for attributes, the most frequently used are the p chart and the u chart. The p chart
displays the sample fraction nonconforming zt ¼ p̂t ¼ Dt =n to monitor the population fraction nonconforming
p, where Dt is the number of nonconforming units in the sample. Assuming that the binomial probability model
pffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
is adequate and can be approximated by a normal distribution, the controlPlimits are set at P
p  3 
p ð1 – pÞ=n,
m m
where the centerline is the average sample fraction nonconforming p ¼ t ¼1 Dt =nm ¼ t ¼1 p̂t =m. In those
cases where the normality assumption is not appropriate (i.e., when n pð1 – pÞ < 9), the control limits have to be
set based on /2 and 1/2 percentiles from the binomial probability model.7 On the contrary, the u chart
displays the sample average number of nonconformities per unit zt ¼ ut ¼ Ct =n to monitor the population
average number of nonconformities per unit /n, where Ct is the occurrence of nonconformities in samples of
size n. Assuming Ct is well modeled by the Poisson distribution with parameter  and can be approximated by a
104 Statistical Control of Measures and Processes

Table 1 Formulas for Shewhart control charts

Chart Centerline Control limits

x (using R) x x  A2 R
R R LCL ¼ D3 R;  UCL ¼ D4 R
x (using s) x x  A3 R
s s LCL ¼ B3 s ; UCL ¼ B4 s
x (using MR) x x  3MR=d2
MR MR LCL ¼ D3 MR; UCL ¼ D4 MR
pffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
p p p  3 pð1 – pÞ=n
pffiffiffiffiffiffiffi
u u u  3 u=n

pffiffiffiffiffiffiffi
 3 u=n, where the centerline is the average sample number
Pm are set at u P
normal distribution, the control limits
m
of nonconformities per unit u ¼ t¼1 Ct =nm ¼ t ¼1 ut =m. In those cases where the normality assumption is
not appropriate (i.e., when  < 9), the control limits have to be set based on /2 and 1/2 percentiles from the
Poisson probability model.7
Table 1 summarizes the formulas to calculate the centerline and control limits for the different Shewhart
control charts introduced.
To determine whether a process is out of control, several rules have to be defined. These rules refer to
extremely low probability events when the process is in control. So, when they occur, you are fairly certain that
something has changed in the process and it is worth looking for an assignable cause. The main rule used on
Shewhart control charts is one point exceeding the control limits. Using three-sigma limits for the control charts,
this rule has an in-control probability of 0.0027. This is called the false alarm probability for this out-of-control
rule. Other rules, called supplementary rules, have been suggested for detecting any nonrandom pattern of
behavior. Some examples are as follows: a run of eight (or seven) consecutive points on one side of the centerline;
eight (or seven) points in a row trending up (or down); and two out of three consecutive points beyond the two-
sigma warning limits ðz  2z Þ. Although the application of these supplementary rules increases the sensitivity
of the Shewhart control charts to the detection of small process shifts, care should be exercised when using
several decision rules simultaneously because of the increase in the overall false alarm probability.8
Shewhart control charts are extremely useful in Phase I implementation of SPC (model building), where
the process is likely to be out of control and experiencing assignable causes resulting in large shifts in the
monitored parameters. They are also very useful in the diagnostic aspects of bringing a ‘wild’ process into
statistical control, because the patterns on these charts often provide guidance regarding the nature of the
assignable cause.1 This advantage comes from the fact that they are plots of the actual data, providing a picture
of what the process is actually doing, which makes the interpretation easy. Because they are not specialized,
these charts act as global radars, being potentially capable of drawing attention to unusual kinds of behavior and
hence to possible signals and causes previously unsuspected.4
The major disadvantage of Shewhart control charts is that they are relatively insensitive to small process shifts,
quite often in Phase II monitoring (model exploitation). This drawback is due to the fact that at each sampling
time t it uses only the contemporary information, ignoring any potential information contained in past samples
(t1, t2, . . .). Time weighted control charts can be designed for this purpose (i.e., detecting small shifts). Out of
the different proposals, the CUSUM control charts and the EWMA control charts are the best choices.

1.04.4 CUSUM Control Charts

The CUSUM control chart was first designed by Page during the 1950s.9,10 In this case, the control statistic to
be charted at stage t is the CUSUM of past and present deviations of the selected seed zt ¼ zðxTt ;0 Þ from a
target (in-control) value 0:
X
t
vt ¼ Ct ¼ ðzk – 0 Þ ð2Þ
k¼1
 Statistical Control of Measures and Processes 105

This feature makes this scheme very powerful for detecting persistent changes in the process parameter ,
particularly a shift from the in-control value 0 to another (out-of-control) value 1.
CUSUMs can be represented in two ways: the tabular (or algorithmic) CUSUM and the V-mask CUSUM.
The tabular CUSUM, which is the preferred one, involves two statistics, Ctþ and Ct– . Ctþ is the sum of deviations
above the target and is referred to as the one-sided upper CUSUM. Ct– is the sum of deviations below the target and
is referred to as the one-sided lower CUSUM:
 
Ctþ ¼ max 0;zt – ð0 þ K Þ þ Ctþ– 1 ð3Þ
 
Ct– ¼ max 0;ð0 – K Þ – zt þ Ct–– 1 ð4Þ

where the starting values are Ctþ ¼ Ct– ¼ 0. The constant K is referred to as the reference value, and it is often
chosen as K ¼ j1 – 0 j=2, that is, halfway between the target value 0 and the out-of-control value 1 we are
aiming to detect quickly. If the shift is expressed in standard deviation units as 1 ¼ 0 þ , then K ¼ ð=2Þ,
that is, one-half of the magnitude of the shift. However, other values of K can be used.
The two-sided CUSUM control chart plots the values of Ctþ and Ct– for each sample. If either statistics lies
beyond a stated decision value (or threshold) H, the process is considered to be out of control. The choice of H is
not arbitrary and should be chosen after careful consideration. A reasonable value is H ¼ 5. Actually, the
proper selection of both parameters K and H is crucial for the good performance of the CUSUM control chart.3
CUSUMs have been developed for a large variety of sample statistics (monitoring seeds) such as individual
observations, sample averages, ranges, and standard deviations of rational subgroups, sample fractions non-
conforming, and average numbers of nonconformities per inspection unit.

1.04.5 EWMA Control Charts

The EWMA control chart is also a good alternative to the Shewhart control chart for detecting small shifts. The
EWMA control chart was first introduced by Roberts.11 In this case, the control statistic to be charted is an
EWMA of present and past values of the selected seed zt :
vt ¼ Et ¼ zt þ ð1 – ÞEt – 1 ð5Þ
where  is a smoothing constant (0 <   1). The EWMA can also be expressed as
X
t
Et ¼ ð1 – Þt E0 þ  ð1 – Þt – k zk ð6Þ
k¼1

where the starting value E0 is the process target 0. From this expression, it is clear how the weights ð1 – Þt –k
decrease geometrically with the age of the sample statistic zt. The parameter  is a tuning parameter. If  is
small, the EWMA performs like a CUSUM. On the contrary, if  is close to one, the EWMA is quite similar to
the Shewhart chart.
Several EWMA charts have been developed for different goals using a large variety of sample statistics
(monitoring seeds). The most used are the EWMA with individual observations ðzt ¼ xt Þ or sample averages
qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
 ffi
ðzt ¼ xt Þ for monitoring the process mean. The control limits are set at 0  Lz ð=ð2 – ÞÞ 1 – ð1 – Þ2t ,
where L is the width of the control limits and z is the standard deviation of the sample statistic z. Once the
chart is started, after several time periods these control limits will approach steady-state values given by
pffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
0  Lz =ð2 – Þ. The process is considered out of control if one or more zt values exceed the control limits.
For a given smoothing constant , the value L is chosen to give the desired performance monitoring of the
EWMA chart. Usual values for both parameters are 0.05    0.25 and 2.6  L  3 (see Hunter12 for details). In
practice, the EWMA chart is used more than the CUSUM chart because it is easier to set up and operate.
For monitoring the variability of normally distributed processes, several EWMA-based statistics have been
proposed. One example is the exponentially weighted mean square error (EWMS) chart. This chart uses the
sample statistic zt ¼ ðxt – 0 Þ2 . To avoid the sensitivity of this chart to shifts in process mean, the exponentially
106 Statistical Control of Measures and Processes

weighted moving variance (EWMV) chart has been proposed by replacing 0 in the sample statistic of EWMS
with an estimate of the process mean obtained from the classical EWMA statistic for the process mean. EWMA
charts can also be used for monitoring the average numbers of nonconformities per inspection unit. This is
called the Poisson EWMA, and uses the sample average number of nonconformities per unit zt ¼ ut ¼ Ct =n as
the sample statistic to be weighted. For technical details on control limits of these charts see Montgomery.1

1.04.6 Performance Measures of Control Charts

The performance of a control chart is usually evaluated based on two criteria: how often the control chart
signals for an in-control process (false alarm rate) and how slow the control chart is in detecting a specific change
in the process (excessive delay in detection). Unfortunately, both criteria cannot be optimized at the same time. For
example, to reduce the false alarm rate, the control limits can be moved farther from the centerline. However,
widening the control limits will increase the risk of a point falling between the control limits when the process is
really out of control. Therefore, it will take more time to detect a change in the process, hence causing excessive
delay in detection. This concept is closely related to Type I error (rejecting the null hypothesis when it is true)
and Type II error (accepting the null hypothesis when it is false) in hypothesis testing. Two different
performance measures, in-control average run length (ARL0) and out-of-control average run length (ARLD),
are defined for evaluating the false alarm rate and the excessive delay in detection, respectively. ARLD depends
on the magnitude of the shift in the process parameter to be monitored, D ¼ j1 – 0 j ¼ .
The average run length (ARL) is defined as the expected (average) number of samples that must be plotted
before an out-of-control signal is observed. For an in-control process ARL0 should be as high as possible. For
example, if the process observations are uncorrelated, then for any Shewhart control chart with no supple-
mentary out-of-control rules ARL0 ¼ 1/, where  is the probability that any point exceeds the control limits
when the process is in control. Under normality and three-sigma control limits, ARL0 ¼ 1/0.0027 ¼ 370. This
means that, even if the process remains in control, a point exceeding the control limits will appear every 370
samples, on the average. For a control chart with given control limits, ARL0 is unique. On the contrary, if the
process is out of control owing to a shift of magnitude D, ARLD is a function of D and it should be as small as
possible, meaning that the control chart signals the change as soon as possible right after the shift occurs. The
performance of a particular control chart can be described by plotting the ARLD values against different
suspected shifts D (note that ARLD¼0 ¼ ARL0). This is also a useful tool to compare the performance of
different control charts.
When comparing the performance of different control charts, ARL0 should be equal for all the schemes. This
can be done by adjusting the control limits of the control charts: For example, in a CUSUM control chart, for a
given particular K value (chosen to efficiently detect a particular shift in the process parameter D ¼ j1 – 0 j),
the value of threshold H is selected to give the desired ARL0. On the contrary, in an EWMA control chart, for a
given smoothing constant , the value L is chosen to give the desired ARL0.

1.04.6.1 Example 1
Let us consider the following example elaborated from a data set available in Montgomery,1 page 387: the first
20 observations were drawn at random from a normal distribution z0  N ð0 ¼ 10; 20 ¼ 1Þ; the last 10
observations were drawn at random from a normal distribution z1  N ð1 ¼ 11; 20 ¼ 1Þ. This data set could
be considered, for example, as reaction run times for 30 batches; the last 10 as corresponding to batches
produced with a different catalyst. Figures 1–3 display the x individuals Shewhart, CUSUM, and EWMA
control charts, respectively. All the charts have been adjusted to yield the same in-control ARL, ARL0 ¼ 370, by
setting L ¼ 3 in the Shewhart chart, H ¼ 4.773 and K ¼ 0.5 in the CUSUM chart, and L ¼ 2.8 and  ¼ 0.15 in
the EWMA chart. The control limits of the charts have been worked out based on the parameters of the normal
distribution z0.
The CUSUM and EWMA charts signal an out-of-control point: Sample 29 is exceeding the UCL. See the
cross points in Figures 2 and 3. Both charts also show a stable upward trend in the last 10 observations,
indicating that a shift in the process mean likely occurred around samples 21 and 22. This matches the time
 Statistical Control of Measures and Processes 107

14
13
12
11
10

x
9
8
7
6
0 5 10 15 20 25 30
Observation
Figure 1 x individuals Shewhart control chart for the reaction run time data (L ¼ 3). Cross points indicate out-of-control
observations. Dotted lines indicate control limits.

3
CUSUM

–1

–3

–5
0 5 10 15 20 25 30
Observation
Figure 2 CUSUM individuals control chart for the simulated reaction run time data (H ¼ 4.773; K ¼ 0.5). Cross points
indicate out-of-control observations. Dotted lines indicate control limits.

11

10.5
EWMA

10

9.5

9
0 5 10 15 20 25 30
Observation
Figure 3 EWMA individuals control chart for the simulated reaction run time data (L ¼ 2.8;  ¼ 0.15). Cross points indicate
out-of-control observations. Dotted lines indicate control limits.

when a new catalyst was used. The slope of the trend estimates the magnitude of the shift, D ¼ j1 – 0 j. In this
case, D is approximately 1, corresponding to one standard deviation of the run time, D ¼  ¼ 1. Figure 1
shows that the Shewhart chart apparently looks in control (there is no point exceeding the control limits). Only
if the supplementary rules are considered, a run of seven consecutive points above the centerline occurs at
sample 29, suggesting the process is out of control. This example illustrates the benefits of using supplementary
rules: increasing the sensitivity of the Shewhart chart to detect small shifts in the process. However, the shift in
the process is much more evident from CUSUM and EWMA control charts.
108 Statistical Control of Measures and Processes

1.04.7 Control Charts for Autocorrelated Processes

As commented in Section 1.04.1, the standard model assumed in most SPC applications is that when the process
is in control the key characteristic x to be measured at stage t can be expressed as
xt ¼  þ at ð7Þ
where at is normally and independently distributed with mean zero and standard deviation  (i.e., at
follows a white noise stochastic process). In many cases, the monitoring schemes are robust to slight
deviations from normality, and then the statistical properties of the control charts are not compromised.
The independence assumption, however, is much more critical and has a profound impact on the
performance of the control charts if indeed violated. For example, if the data are positively correlated
(i.e., values above the mean tend to be followed by other values above the mean and vice versa), control
charts will give misleading results in the form of too many false alarms and they will not signal real
out-of-control situations.
Lack of independence between successive observations (i.e., data correlation over time) shows up whenever
the interval between samples becomes small relative to the process dynamics (i.e., inertial elements as raw
materials flow, storage tanks, reactors’ residence times, or environmental conditions, defining the settling time of
the process). This being more the rule than the exception in modern process environments because of the
information technologies that allow registering data from every part produced (e.g., in computer integrated
manufacturing (CIM) environments) or at high sampling rates (e.g., in continuous and batch process industries).
Most of the present-day analytical systems contain mechanical, electrical, and/or optical parts, in which
the correlation over time can be of great importance. Some examples are lamp aging of the photometric detector,
a decrease in the separation capacity of the chromatographic column, changes in the detector response with use,
the influence of points of contamination on subsequent responses in continuous flow analytical systems,
sensor deterioration, or changes in the environmental conditions (temperature, humidity, etc.), causing drifts in
the results.13

1.04.7.1 Example 2
As an example, consider the data in Figure 4. This graph plots the x–MR(2) individuals control chart from the
results of 56 consecutive sulfate determinations in a control sample using a sequential injection analysis (SIA)
system with UV/visible detection for determining sulfates in water.14
Note that mainly the x individuals control chart signals many out-of-control measurements (cross points in
Figure 4): for example, measurement 8 completes a run of seven consecutive points below the centerline; or
measurements 11–18 are above the UCL. The sulfate determinations of the reference sample are drifting or
wandering over time, indicating that the analytical process is apparently out of control. Before taking any
decision, it would be of great interest to know whether the out-of-control signals are indicating real problems or
whether they are false alarms due to autocorrelation (correlation over time).

240 24

230 20
16
220
MR(2)

12
x

210
8
200 4
190 0
0 10 20 30 40 50 60 0 10 20 30 40 50 60
Observation Observation
Figure 4 x–MR(2) individuals control chart for sulfate determinations in a control sample. Cross points indicate
out-of-control measurements. Dotted lines indicate control limits.
 Statistical Control of Measures and Processes 109

The correlation over a series of n time-oriented observations can be estimated by the sample autocorrelation
function (ACF):
Pn – k
ðxt – xÞðx t þk – xÞ
rk ¼ t ¼1Pn ð8Þ
Þ2
t ¼1 ðxt – x

This assumes that the mean and the variance of the process are constant and that correlation between successive
observations depends only on the lag k (number of time periods between observations).
The sample ACF for the sulfate determinations is shown in Figure 5. Note that there is a statistically
significant (p-value < 0.05) strong positive correlation till lag k ¼ 3. This means that measurements that
are one, two, and three periods apart are positively correlated. This level of autocorrelation is sufficient
to distort the performance of a Shewhart control chart and the interpretation rules. For example,
supplementary run rules should never be used because they can be generated from autocorrelation
(i.e., they are part of the common-cause system) and not from assignable causes of variability. On the
contrary, estimating the variability of a positive autocorrelated process from moving ranges (as done in
Figure 4) may lead to a considerably high underestimation of the real process variability, yielding
control limits too much narrower than should be and, hence, increasing the frequency of measurements
exceeding them.
It is important to highlight that autocorrelation must be considered as part of the common-cause system affecting
the process, and any attempt to get rid of it (e.g., by reducing the sampling rate) will lead to a very inefficient use of
available data, degrading the performance of the resulting monitoring scheme. Therefore, we should model the
autocorrelation. For this purpose, autoregressive integrated moving average (ARIMA) models15 can be used.
In addition to the sample ACF, the so-called partial ACF15 is a useful tool to give more information about the
correlation structure of the data. The partial ACF measures the correlation between measurements that are
shifted k lags without the influence of the intermediate values.
Figure 6 shows the partial ACF for the sulfate determinations. In this case, only the partial ACF at lag 1 is
statistically significant (p-value < 0.05).
Figures 5 and 6 suggest a first-order autoregressive model, AR(1), to describe this analytical process

xt –  ¼ ðxt –1 – Þ þ at ð9Þ

where  is the process mean,  is the autoregressive coefficient ðjj < 1Þ, and at follows a white noise stochastic
process, at  N(0, ). The standard deviation of xt is given by


x ¼ pffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi ð10Þ
ð1 – 2 Þ

0.6

0.2
r(k)

–0.2

–0.6

–1
0 3 6 9 12 15
Lag
Figure 5 Sample ACF for the sulfate determinations in a control sample. Dotted lines indicate limits for statistical
significance ( ¼ 0.05).
110 Statistical Control of Measures and Processes

0.6

0.2

r(k)
–0.2

–0.6

–1
0 3 6 9 12 15
Lag
Figure 6 Partial ACF for the sulfate determinations in a control sample. Dotted lines indicate limits for statistical significance
( ¼ 0.05).

The estimated AR(1) model for the sulfate measurements is

xt – 215:70 ¼ 0:85ðxt –1 – 215:70Þ þ et ð11Þ

where the residuals et ¼ xt – x̂t are approximately normally and independently distributed with mean zero and
constant variance (et  N(0, s ¼ 4.85)). This is validated by computing the simple and partial ACF of the
residuals (not shown) where no coefficient are statistically significant (p-value > 0.05).
There are several approaches to monitor this analytical process. One is to estimate the process variability
qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
from Equation (10) and set the control limits on the x individuals chart at ˆ  3s= ð1 – Þ ˆ 2 . No supplementary
rules should be used when interpreting the chart. Note that in the sulfate data the process standard deviation
estimated from the average moving range ðMR=d2 ¼ 2:99Þ is approximately one-third of the one estimated
(i.e., 9.21) from Equation (10). This explains the large number of measurements exceeding the control limits in
Figure 4.
Another approach is to apply classical control charts to the residuals. Points out of control or unusual patterns
on such charts would indicate that the parameters , , or  had changed, implying that the analytical process
was out of control. Figure 7 is an x–MR(2) individuals control chart of the residuals from the fitted AR(1)
model. Most of the out-of-control points in the x–MR(2) individuals control chart of the original data
(Figure 4) have disappeared. This confirms the hypothesis that they were real false alarms due to autocorrela-
tion. As shown in Figure 7, there is only one out-of-control point (measurement 36) in the x chart that
generates two out-of-control points (measurements 36 and 37) in the MR(2) chart. By looking at the original
data (Figure 4) at measurement 35, there is a drastic jump to measurement 36 that is signaled by the MR(2)
chart. This momentary dysfunction of the analytical systems can be caused by bubbles that temporarily produce
some anomalous data but, once they have disappeared, the systems begin to function correctly again.16

27 24
20
17
16
MR(2)

7 12
x

8
–3
4
–13 0
0 10 20 30 40 50 60 0 10 20 30 40 50 60
Observation Observation
Figure 7 x–MR(2) individuals control chart for the residuals of AR(1) model from sulfate determinations in a control sample.
Cross points indicate out-of-control observations. Dotted lines indicate control limits.
 Statistical Control of Measures and Processes 111

240

230

Sulfate (mg l–1)

220

210

200

190
0 20 40 60
Observation
Figure 8 One-step ahead forecast sulfate determinations from AR(1) model.

To exploit the inertial properties of the process, in addition to the residual control charts, one-step ahead
forecasts, x̂t þ1=t (i.e., prediction for the measurement at stage t þ 1 made at stage t), can also be plotted in a run
chart:

x̂t þ1=t ¼  þ ðx t – Þ ð12Þ

Figure 8 displays the run chart of the one-step ahead forecasts, x̂tþ1=t , for the sulfate measurements. Note the
big gap (error) between the real measurement 36 and its prediction when measurement 35 was obtained, x̂36=35 .
If this monitoring scheme was implemented online, this kind of anomalies could be detected in real time. Note
that for estimating this model, outliers should be eliminated. Several methods can be used for this purpose. See
for example Box et al.15
Finally, it is important to note that in correlated processes, if there is any easily manipulated variable related
to the key characteristic of interest, process dynamics can be exploited to design control rules to force the key
characteristic to track a desired trajectory. By integrating control theory with SPC ideas, process improvement
can be achieved as commented in the following section.

1.04.8 Integration of SPC and Engineering Process Control

The traditional control charting discipline of ‘monitor, then adjust when out of control’ is not the best strategy
for process improvement in processes with some kind of dynamics. This is especially true in processes having a
tendency to drift or wander away from the target. In these contexts, another approach called stochastic control,
also known as engineering process control (EPC), is needed. This approach is based on process compensation
and regulation (feedback and/or feedforward control), in which some easily manipulable process variables are
adjusted following some control rules with the goal of keeping the key characteristic (controlled variable) close
to the desired target. This requires good knowledge of the relationship between the controlled variable and the
manipulated variables, as well as an understanding of process dynamics. If the cost of making adjustments
(control actions) in the manipulable process variables is negligible, then the variability of the controlled variable
is minimized by taking control actions in every sample. This is in sharp contrast to SPC.1
SPC and EPC originated in different industries – the parts industry and the process industry – and have been
developed independently with some controversy in the past because of the different meaning of ‘control’ in
both approaches: process monitoring (SPC) versus process regulation (EPC). It is true that a dynamic process
should not be adjusted using SPC, but it is also true that assignable causes cannot be detected with EPC. In fact,
regulation schemes react to process upsets; they do not make any effort to remove the root causes. For example,
let us consider a control loop that regulates the inside reactor temperature through a thermostat (feedback
controller) by making adjustments in the heating flow. If there appears a leaking problem in the heating jacket,
the reactor temperature decreases and the thermostat tries to compensate for the deviation by increasing the
heating flow. By continuous adjustments (provided that the heating system can keep up with the temperature
112 Statistical Control of Measures and Processes

differential), the reactor temperature is kept around the target but the assignable cause (leaking) goes
undetected. The EPC scheme masks the presence of the assignable cause keeping the process at the desired
target at all costs. In the above example, this will result in high energy and repair costs (at the very least the
heater will eventually breakdown). By monitoring, for example, the adjustments in the heating flow, the leaking
problem would have been detected, saving energy and avoiding the breakdown of the heating system.
Therefore, EPC and SPC should be considered as two complementary (not alternative) strategies for
quality improvement. SPC monitoring procedures seek to reduce output variability by detecting and
eliminating assignable causes of variation. On the contrary, EPC tries to minimize output variability by
making regular adjustments exploiting the process dynamics (common-cause system). Hence, ideas from
both fields can be used together in an integrated EPC/SPC system called to secure both optimization and
improvement.
The notion of superimposing SPC on a closed-loop system has opened new lines of research in the area of
quality improvement. This new strategy called engineering statistical process control (ESPC), also known as
algorithmic statistical process control (ASPC),17 is a proactive approach to quality improvement that reduces
predictable variations in quality characteristics using feedback and feedforward techniques and then monitors
the entire system (by plotting deviations from the target, prediction errors, adjustments, etc.) to detect and help
remove unpredictable process upsets. ESPC integrates two complementary approaches: process adjustment and
process monitoring. As such, ESPC is a marriage of control theory and SPC that aims to reduce both short- and
long-term variability by replacing the traditional control charting discipline of ‘monitor, then adjust when out
of control’ with ‘adjust optimally and monitor’.18 Successful applications and thorough research of this
methodology can be found in the literature.4,19–24

1.04.9 Multivariate Control Charts

Most SPC schemes currently in practice are based on charting one or a small number of product quality
variables in a univariate way. These approaches, although effective in the past where data were scarce, are
totally inadequate for modern continuous and batch processes and manufacturing industries where massive
amounts of highly correlated variables are being collected. As commented by Professor John F. MacGregor ‘‘the
presence of variable interactions in experimental designs leads to the same difficulties in interpreting the results
of one factor at a time experimentation as does the presence of correlation among variables in interpreting
univariate SPC charts’’.25 Applying univariate SPC charts to each individual variable separately will force the
operator to inspect a large number of control charts. When special events occur in a process, they affect not only
the magnitude of the variables but also their relationship to each other. These events are often difficult (or even
impossible) to detect by charting one variable at a time in the same way as the human eye cannot perceive
perspective by looking at an object with one eye at a time.26 Multivariate statistical process control (MSPC)
schemes that treat all the variables simultaneously are required in these new data-rich environments. Several
approaches have been proposed. In this chapter, they are going to be classified as original variables-based
MSPC and latent variables-based MSPC.

1.04.9.1 Original Variables-Based MSPC Schemes

Let us consider Xt ¼ fx1t ; x2t ; . . .; xKt g, the n  K observations matrix of K characteristics measured on n units
sampled randomly from the production or measurement process at stage t (t ¼ 1, 2, . . .), where xT kt ¼ fxk1t ; xk2t ;
. . .; xknt g is the n-dimensional vector of sample values of the kth variable (the kth column of matrix X) and
xTnt ¼ fxn1t ; xn2t ; . . .; xnKt g is the K-dimensional observation vector of the nth unit (the nth row of matrix X).
Original variables-based MSPC schemes are based on multivariate statistics built from measured variables Xk
(k ¼ 1, . . ., K). They can be thought of as the multivariate counterpart to the Shewhart, CUSUM, and EWMA
control charts for monitoring the process mean and variability. The Hotelling T 2 and multivariate EWMA
(MEWMA) control charts are the most used in practice.
 Statistical Control of Measures and Processes 113

1.04.9.1.1 Hotelling T 2 control charts for monitoring process mean

The Hotelling T 2 control chart is the multivariate extension of the univariate Shewhart x control chart for
monitoring the process mean. This approach assumes that the (K  1) vector of measured variables x is not time
dependent and follows a K-dimensional normal distribution, x NK ðm; SÞ. The chart checks if the (K  1) mean
vector of the process m remains stable, assuming a constant (K  K) covariance matrix S. This can be used for
either individual (n ¼ 1) or subgroup data (n > 1).

1.04.9.1.1(i) Subgroup data Assume that there are m samples of K measured variables each of size n > 1
available from the process. The Hotelling T 2 control chart is based on monitoring at each time t the Hotelling
T 2 statistic:1,27–30
T
t – xÞ S – 1 ðx
Tt2 ¼ nðx t – xÞ ð13Þ
where xt is the (K  1) vector of sample means at stage t, x the (K  1) overall sample mean vector (estimate of
the in-control (K  1) mean vector m), and S is the (K  K) pooled sample variance–covariance matrix (estimate
of the in-control (K  K) covariance matrix S) from the m samples. The Tt2 statistic represents the estimated
weighted distance (Mahalanobis distance) of any observation from the target m. Therefore, only UCLs are
defined for this chart. The Phase I and II control limits at significance level (Type I risk)  are given by
K ðm – 1Þðn – 1Þ
UCLðT 2 ÞPhase I; ¼ FðK ;ðmn – m – K þ1ÞÞ; ð14Þ
mn – m – K þ 1
K ðm þ 1Þðn – 1Þ
UCLðT 2 ÞPhase II; ¼ FðK ;ðmn – m – K þ1ÞÞ; ð15Þ
mn – m – K þ 1
where FðK ;ðmn – m – K þ1ÞÞ; is the 100(1)% percentile of the corresponding F distribution. Thus, if the value of
the Tt2 statistic plots above the UCL, the chart signals a potential out-of-control process. The difference in both
control limits comes from the fact that in Phase II, when the chart is used for monitoring new observations, x t is
independent of x and S sample estimators. This is not the case in Phase I. However, when the number of
samples m is large (m > 50), both limits match and the Tt2 statistic can be approximated by a 2 distribution with
K degrees of freedom. Therefore, UCLðT 2 Þ ¼ 2K ; , where 2K ; is the 100(1)% percentile of the
corresponding 2 distribution.

1.04.9.1.1(ii) Individual observations As commented in Section 1.04.3, there are situations where the
subgroup size is naturally n ¼ 1. This occurs frequently in chemical and process industries and in measurement
processes. In this case, the Hotelling T 2 statistic at stage t is given by

ÞT S – 1 ðxt – x
Tt2 ¼ ðxt – x Þ ð16Þ
where xt is the (K  1) vector of measured variables at stage t, xT ¼ fx1 ; xP xK g is the sample mean vector
2 ; . . .; 
(estimate of the in-control (K  1) mean vector m), and S ¼ ðm – 1Þ – 1 mi¼1 ðxi – x ÞT is the (K  K)
Þðxi – x
sample variance–covariance matrix (estimate of the in-control (K  K) covariance matrix S) from the m samples.
Following Tracy et al.,31 in Phase I the Tt2 =ðm – 1Þ2 m – 1 statistic follows a beta distribution with K/2 and
(mK1)/2 degrees of freedom. Therefore, UCLs at significance level (Type I risk)  can be obtained for
Phase I as

ðm – 1Þ2
UCLðT 2 ÞPhase I; ¼ BðK =2;ðm – K – 1Þ=2Þ; ð17Þ
m
where BðK =2;ðm – K – 1Þ=2Þ; is the 100(1)% percentile of the corresponding beta distribution. This can be
computed from the 100(1)% percentile of the corresponding F distribution by using the relationship

ðK =ðm – K – 1ÞÞFðK ;m – K – 1Þ;

BðK =2;ðm – K – 1Þ=2Þ; ¼ ð18Þ
1 þ ðK =ðm – K – 1ÞÞFðK ;m – K – 1Þ;
114 Statistical Control of Measures and Processes

In Phase II, the Tt2 =K ðm – 1Þ2 ½mðm – K Þ – 1 statistic follows an F distribution with K and (mK) degrees of
freedom. Thus, the corresponding UCL at significance level (Type I risk)  is given by
K ðm2 – 1Þ
UCLðT 2 ÞPhase II; ¼ FðK ;ðm – K ÞÞ; ð19Þ
mðm – K Þ
If the multivariate normality assumption is not reasonable, control limits may be obtained from resampling
methods such as bootstrap.6

1.04.9.1.2 Multivariate Shewhart-type control charts for monitoring process variability

The Hotelling T 2 control charts assume that process variance–covariance structure, summarized in the (K  K)
covariance matrix S, remains constant. This assumption is not necessarily true and must be validated in
practice. There are two indices for measuring the overall variability of a set of multivariate data: (i) the
generalized variance jSj, that is, the determinant of the covariance matrix, and (ii) the trace of the covariance
matrix trS, that is, the sum of the variances of the K variables. F. B. Alt32 presents two control charts for
monitoring process variability based on the two previous indices.
Assume that the vector x follows an NK ðm; SÞ and that there are m samples of size n > 1 available. The first
procedure is a direct extension of the univariate s2 control chart. The procedure is equivalent to repeated test of
significance of the null hypothesis that the process covariance matrix is S. The control statistic to be charted at
stage t is
Wt ¼ – Kn þ Kn lnðnÞ – n lnðjAt j=jSjÞ þ trðS – 1 At Þ ð20Þ
where At ¼ (n1)St, St is the sample covariance matrix of the tth subgroup, and tr is the trace operator. The Wt
statistic follows an asymptotic 2 distribution with K(K þ 1)/2 degrees of freedom. Hence, the UCL at
significance level (Type I risk)  is UCLðW Þ ¼ 2ðK ðK þ1Þ=2Þ; .
The second approach is based on the sample generalized variance jSj (the square root of this quantity is
proportional to the area or volume generated by a set of data). By using the mean
pffiffiffiffiffiffiffiffiffiffiffiffi
ffi and variance of jSj, that is,
E(jSj) and V(jSj) respectively, the control limits are set at EðjSjÞ  3 VðjSjÞ.
Although jSj is a widely used measure of multivariate dispersion, it is a relative simplistic scalar representa-
tion of a complex multivariate structure. Therefore, its use can be misleading in the sense that different
correlation structures can yield identical generalized variance. To face this problem, the use of univariate
dispersion charts in addition to the control chart for jSj is proposed.32
Other references to multivariate SPC methods for controlling the process variability can be found in Mason
et al.33

1.04.9.1.3 MEWMA control charts

The Hotelling T 2 control chart is a Shewhart-type control chart, that is, it uses information only from the
current sample. As a result, it is relatively insensitive to small and moderate shift in the mean vector m. The
MEWMA, introduced by Lowry et al.,34 is the multivariate extension of the univariate EWMA to provide more
sensitivity to small shifts in the mean vector m. This is particularly useful in Phase II studies. Assuming
individual observations (n ¼ 1), the MEWMA statistics is defined as
wt ¼ xt þ ð1 – Þwt – 1 ð21Þ

where  is a smoothing constant (0 <   1) and xt is the tth, K-dimensional observation sampled from a K-
dimensional normal distribution, x NK ðm; SÞ. The control statistic to be charted at stage t is
–1
MEWMA2t ¼ wT
t wt wt ð22Þ

where the covariance matrix is

  
Swt ¼ 1 – ð1 – Þ2t S ð23Þ
2–
 Statistical Control of Measures and Processes 115

which is analogous to the variance of the univariate EWMA. The control limits and tuning parameters
necessary for implementing the MEWMA control chart are dependent on the shift in the mean vector to
detect the in-control ARL and the number of characteristics K to be monitored.35 MEWMA control charts, like
their univariate counterparts, are robust to the assumption of normality, if properly designed.
Bersimis et al.36 provide an extensive review of several types of MEWMA control charts for different
purposes, for example, to account for multivariate time-dependent observations and to deal with monitoring
changes in process variability.

1.04.9.1.4 Fault diagnosis

Once the multivariate control chart signals an out-of-control alarm, it is required to diagnose an assignable
cause for it. This involves two steps: first (diagnostic) find which measured variable(s) contributes to the out-of-
control signal and second (corrective) determine what happens in the process that upsets the behavior of these
variables. For pursuing the first step (isolation of variables responsible for the out-of-control signal), several
approaches have been reported in the literature. An extensive review and references of diagnostic procedures in
original variables-based MSPC schemes can be found in Mason et al.33 Bersimis et al.36 and Kourti and
MacGregor.37 Regarding the second step (identifying root causes of the problem), management and operator
actions based on technical process knowledge will be required.

1.04.9.2 Latent Variables-Based MSPC Schemes

Although original variables-based MSPC is well sounded from a statistical point of view, it suffers from lack of
applicability in data-rich environments, typical of modern processes where, as commented in Section 1.04.3, the
subgroup size is naturally n ¼ 1. This serious drawback comes from the fact that the multivariate control
statistics to be charted need the inversion of a covariance matrix (see Equations (13), (16), (20), and (22)). To
avoid problems with this inversion, the number of multivariate observations or samples (m) has to be larger than
the number of variables (K), and covariance matrix has to be well conditioned (slightly correlated variables). In
addition, complete data (no missing values) are required to work out the Hotelling T 2 or MEWMA2 statistics
for any particular sample (see Equations (13), (16), and (22)). Nevertheless, these requirements are not met in
highly automated processes where not only a few product quality variables but hundreds (or even thousands) of
process variables are measured at a higher frequency rate.
For treating large and ill-conditioned data sets that are not full statistical rank (m < K), we advocate the use of
latent variables-based MSPC in the way they were proposed by Kourti and MacGregor.37 Latent variable
methodology exploits the correlation structure of the original variables by revealing the few independent
underlying sources of variation (latent variables) that are driving the process at any time. Multivariate statistical
projection methods such as principal component analysis (PCA)38 are used to reduce the dimensionality of the
monitoring space by projecting the information in the original variables down onto low-dimensional subspaces
defined by a few latent variables. The process is then monitored in these latent subspaces by using a few
multivariate control charts built from multivariate statistics that can be thought of as process performance indices or
process wellness indices.39 These charts retain all the simplicity of presentation and interpretation of conventional
univariate SPC charts. However, by using the information contained in all the measured (process and quality)
variables simultaneously, they are much more powerful for detecting out-of-control conditions.
As commented by T. Kourti,40 by monitoring only the quality variables (as frequently done in original
variables-based MSPC) we are in fact performing statistical quality control (SQC). For true SPC, one must look
at all the process data as well. Monitoring the process variables is expected to provide much more information
on the state of the process and to supply it more frequently. Furthermore, any abnormal events that occur will
also have their fingerprints in the process data. Thus, once an abnormal situation is detected, it is easier and
faster to diagnose the source of the problem, as we are dealing directly with the process variables. On the
contrary, control charts on the quality variables will only signal when the product properties are no longer
consistent with expected performance, but they will not point to the process variables responsible for the
problem, making the fault diagnostic process more difficult.
Another advantage of monitoring process data is that quality data may not be available at certain stages of the
process. Sometimes, product quality is determined only by the performance of the product later, during another
116 Statistical Control of Measures and Processes

process. For example, if a catalyst is conditioned in a batch process before being used for polymer production, the
quality of the catalyst (success of conditioning) is assessed by its performance in the subsequent polymer production.
It would be useful to know if the catalyst will perform well before using it; monitoring the batch process variables
would detect abnormal situations and would provide an early indication of poor catalyst performance.
In some cases, the scarce properties measured on a product are not enough to define entirely the product
quality and performance for different applications. For example, if only the viscosity of a polymer is measured
and kept within specifications, any variation in end-use application that arises owing to variation in chemical
structure (branching, composition, and end-group concentration) will not be captured. In these cases, the
process data may contain more information about events with special causes that may affect the product
structure and thus its performance in different applications. Finally, by monitoring process variables, other
abnormal operating conditions may be detected such as a pending equipment failure or a process failure.40
Finally, by using latent variables-based MSPC, missing and noisy data are easily handled, and predictive
models based on projection to latent structures, such as partial least squares (PLS) or principal component
regression (PCR), can also be used.41
Given the practical relevance of this approach, in the following an overview of implementation of MSPC
schemes based on principal component analysis (PCA-based MSPC), describing both Phase I and II, is
presented.

1.04.9.2.1 PCA-based MSPC: exploratory data analysis and offline process monitoring
(Phase I)
As commented in Section 1.04.2, the main goal in Phase I is to model the in-control process performance based
on a set of historical in-control (reference) data. This data set is one in which the process has been operating
consistently (stable over time) in an acceptable manner, and in which only good quality products have been
obtained. Occasionally, this historical in-control data set is not directly available, but has to be extracted from
historical databases in an iterative fashion as commented below. This explorative analysis of historical databases
is a useful technique for improving process understanding and detecting past faults in the process (out-of-
control samples). By correctly diagnosing their root causes, some countermeasures can be implemented,
optimizing the future performance of the process.
Consider that the historical database consists of a set of m multivariate observations (objects or samples of size
n ¼ 1) on K variables (online process measurements, dimensional variables, or product quality data) arranged in
an (m  K) data matrix Z. Variables in matrix Z are often preprocessed by mean centering and scaling to unit
variance. With mean centering, the average value of each variable is calculated and then subtracted from the
data. This usually improves the interpretability of the model because all preprocessed variables will have mean
value zero. By scaling to unit variance, each original variable is divided by its standard deviation and will have
unit variance. Given that projection methods are sensitive to scaling, this is particularly useful when the
variables are measured in different units. After preprocessing, matrix Z is transformed into matrix X.
PCA is used to reduce the dimensionality of the process by compressing the high-dimensional original data
matrix X into a low-dimensional subspace of dimension A (A  rank(X)), in which most of the data variability is
explained by a fewer number of latent variables, which are orthogonal and linear combinations of the original
ones. This is done by decomposing X into a set of A rank 1 matrices

X
A X
rankðXÞ

X¼ t a pT
a þ ta p T T
a ¼ TP þ E ¼ X þ E ð24Þ
a¼1 a¼Aþ1

P (K  A) is the loading matrix containing the loading vectors pa, which are the eigenvectors, corresponding to the
A largest eigenvalues of the covariance matrix of the original pretreated data set X, and define the directions of
highest variability of the new latent A-dimensional subspace. T (m  A) is the score matrix containing the
location of the orthogonal projection of the original observations onto the latent subspace. The columns ta of the
score matrix T (ta ¼ Xpa) represent the new latent variables with variances given by their respective eigenva-
lues (a). These new latent variables summarize the most important information of the original K variables, and
thus can predict (reconstruct) X, by means X ¼ TPT, with minimum mean square error. Matrix E (m  K)
contains the residuals (statistical noise), that is, the information that is not explained by the PCA model.
 Statistical Control of Measures and Processes 117

The dimension of the latent variable subspace is often quite small compared with the dimension of the
original variable space (i.e., A << rank(X)). Several algorithms can be used to extract the principal components.
For large ill-conditioned data sets, it is recommended to compute the principal components sequentially via the
NIPALS (noniterative partial least squares) algorithm42 and to stop based on cross-validation criteria or some
other procedure.38,43 Another advantage of NIPALS algorithm is that it easily handles missing data (i.e.,
observation vectors from which some variable measurements are missing). The quality of the fitted PCA
model can be evaluated by computing several parameters, such as R2, which measures the goodness of fit, or Q2,
which indicates the predictive capability of the model.44
Equation (24) shows that the PCA model transforms each K-dimensional original observation vector xi (the
ith row of matrix X) into an A-dimensional score vector tT i ¼ fti1 ;ti2 ; . . .; tiA g(the ith row of matrix T) and a
residual vector ei (the ith row of matrix E). From the scores and the residuals (prediction errors) associated with
each observation, two complementary (orthogonal or independent) statistics are derived: the Hotelling’s TA2 and
the SPE (sum of squared prediction errors). The TA2 statistic for the ith observation is defined as

X
A
t2
–1
TA2 ¼ tT
i Q ti ¼
a
ð25Þ
a¼1
a

where Q (A  A) is the covariance matrix of T (diagonal matrix of the highest A eigenvalues f1 ; . . .; A g). This
statistic is the Hotelling T 2 statistic when a reduced subspace with A components is used instead of the original
variables space, and it represents the estimated Mahalanobis distance from the center of the latent subspace to
the projection of an observation onto this subspace. Under the assumption that the scores follow a multivariate
normal distribution (they are linear combinations of random variables), it holds31 that in Phase I, TA2 (times a
constant) has a beta distribution

ðm – 1Þ2
TA2  BA=2;ðm – A – 1Þ=2 ð26Þ
m
whereas in Phase II, TA2 (times a constant) follows an F distribution

Aðm2 – 1Þ
TA2  FA;ðm – AÞ ð27Þ
mðm – AÞ

The difference in both distributions comes from the fact that in Phase I, the same observation vectors xi
collected in the reference data set are used for two purposes: (i) to build the PCA model and work out the
control limits of the charts and (ii) to check whether they fall within these control limits. Therefore, observa-
tions in the reference data set are not independent of PCA model parameters used to derive the statistics to be
monitored. In contrast, in Phase II, new observations (not used for model building) are checked against the
control limits calculated from the in-control data and, therefore, independence is guaranteed. However, if a
large reference data set is available, Equation (27) can also be used for approximating the distribution of the TA2
statistic in Phase I.
On the contrary, the SPE statistic for the ith observation xi is given by

 T 
SPE ¼ eT
i ei ¼ ðxi – xi Þ ðxi – xi Þ ð28Þ

where ei is the residual vector of the ith observation and xi is the prediction of the observation vector xi from
the PCA model. The SPE statistic represents the squared euclidean (perpendicular) distance of an observation
from this subspace, and gives a measure of how close the observation is from the A-dimensional subspace.
Assuming that residuals follow a multivariate normal distribution, several approximate distributions for such
quadratic forms have been proposed.44–46
From these two statistics, in PCA-based MSPC, two complementary multivariate control charts are con-
structed. Shewhart-type control charts for individual observations are often used in practice. Nevertheless,
other types of multivariate charts such as MEWMA charts (see Section 1.04.9.1.3) may be used. The latter may
be specially suited for autocorrelated processes.
118 Statistical Control of Measures and Processes

The control limits of the multivariate control charts are calculated following the traditional SPC philosophy.
In Phase I, an appropriate historical or reference set of data (collected from one or different periods of plant
operation or analytical process when performance was good) is chosen, which defines the normal or in-control
operating conditions for a particular process corresponding to common-cause variation. The in-control PCA
model is then built on these data. Any periods containing variations arising from special events that one would
like to detect in the future are omitted at this stage. The choice of the reference (in-control) data set is critical to
the successful application of the procedure.37 Control limits for good operation on the control charts are defined
based on this reference data set. In Phase II, values of future measurements are compared against these limits.
UCL for the Shewhart TA2 chart at significance level (Type I risk)  can be obtained for Phase I from
Equation (26):

ðm – 1Þ2
UCLðTA2 Þ ¼ BðA=2;ðm – A – 1Þ=2Þ; ð29Þ
m
where BðA=2;ðm – A – 1Þ=2Þ; is the 100(1)% percentile of the corresponding beta distribution that can be
computed from the 100(1)% percentile of the corresponding F distribution by using the following
relationship:31

ðA=ðm – A – 1ÞÞFðA;m – A – 1Þ;

BðA=2;ðm – A – 1Þ=2Þ; ¼ ð30Þ
1 þ ðA=ðm – A – 1ÞÞFðA;m – A – 1Þ;

For Phase II, the corresponding UCL from Equation (27) is given by
Aðm2 – 1Þ
UCLðTA2 Þ ¼ FðA;ðm – AÞÞ; ð31Þ
mðm – AÞ
Regarding the UCL for the Shewhart SPE chart, several procedures can be used. Jackson and Mudholkar46
showed that an approximate SPE critical value at significance level  is given by
" pffiffiffiffiffiffiffiffiffiffiffi #1=h0
z 22 h02 2 h0 ðh0 – 1Þ
UCLðSPEÞ ¼ 1 þ1þ ð32Þ
1 21
PrankðXÞ
where k ¼ j ¼Aþ1 ðj Þk , h0 ¼ 1 – ð21 3 =322 Þ, j are the eigenvalues of the PCA residual covariance matrix
ETE/(N1), and z is the 100(1)% standardized normal percentile.
Alternatively, one can use an approximation based on the weighted 2 distribution ðg 2h Þ proposed by Box.45
Nomikos and MacGregor47 suggested a simple and fast way to estimate the parameters g and h, which is based
on matching moments between a g 2h distribution and the sample distribution of SPE. The mean ð ¼ ghÞ and
variance ð2 ¼ 2g 2 hÞ of the g 2h distribution are equated with the sample mean (b) and variance (v) of the SPE
sample. Hence, the upper SPE control limit at significance level  is given by
v 2
UCLðSPEÞ ¼ 2 ð33Þ
2b ð2b =vÞ;
where 2ð2b2 =vÞ; is the 100(1)% percentile of the corresponding 2 distribution.
Another method based on the statistical test of equality of variances from normal distributions is proposed by
Eriksson et al.44 Based on the SPE, they define the absolute distance to the model (DModX) of an observation as
its (corrected) residual standard deviation
sffiffiffiffiffiffiffiffiffiffiffiffiffiffi
SPE
DModX ¼ c ð34Þ
ðK – AÞ

where c is a correction factor (function of the number of observations and the number of components) to be used
in Phase I. This correction factor takes into account that the DModX is expected to be slightly smaller for an
observation in the reference set because it has influenced the model. This correction matters only if the number
of observations in the reference set is small. In Phase II, c ¼ 1.
 Statistical Control of Measures and Processes 119

They also define the normalized distance to the model (DModXnorm) as

DModX
DModXnorm ¼ ð35Þ
s0
qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
Pm PK 2
where s0 ¼ i¼1 k¼1 eik =ðm – A – 1ÞðK – AÞ is the pooled residual standard deviation. This is an estimation
of the residual variability taking into account all the observations used to build the model (reference data set).
Assuming that the statistic (DModXnorm)2 has an approximate F distribution with KA and (mA1)(KA)
degrees of freedom for the in-control observations, the UCL for the Shewhart SPE chart at significance level 
is expressed as

K –A 2
UCLðSPEÞ ¼ s FðK – A;ðm – A – 1ÞðK – AÞÞ; ð36Þ
c2 0

where FðK – A;ðm – A – 1ÞðK – AÞÞ; is the 100(1)% percentile of the corresponding F distribution.
The normality assumption on which these calculations are based is usually quite reasonable in practice.
Nevertheless, control limits for multivariate charts can be obtained from distribution free methods by repeated
sampling. The only requirement is to have a large in-control data set from which the external reference
distribution48 for any statistic can be obtained.
The two multivariate control charts (TA2 and SPE) differ in their conceptual meaning. They are two
complementary indices that provide a picture of the wellness of the process at a glance.39 The TA2 chart checks
if the projection of an observation on the hyperplane defined by the latent subspace is within the limits
determined by the reference (in-control) data. Thus, a value of this statistic exceeding the control limits
indicates that the corresponding observation presents abnormal extreme values in some (or all) of its original K
variables, even though it maintains the correlation structure between the variables in the model. This
observation can be tagged as an abnormal outlier inside the PCA model (an extremist or severe outlier).41 On
the contrary, the SPE chart checks if the distance (noise variation) of an observation to the latent hyperplane is
inside the control limits. The SPE chart values exceeding the control limits are related to observations that do
not behave in the same way as the ones used to create the model (in-control data), in the sense that there is a
breakage of the correlation structure of the model. This chart will detect the occurrence of any new event that
causes the process to move away from the hyperplane defined by the reference model. This kind of observations
can be tagged as outliers outside the model (an alien or moderate outlier).41
Severe outliers are influential observations with high leverage on the model, that is, with strong power to pull
the principal directions toward themselves, creating fictitious components and misleading the PCA model.44
Therefore, model validation is extremely needed in the Phase I stage to remove from the data matrix these
dangerous outlier (out-of-control) observations and, afterward, recalculate the PCA model. However, before
removing any observation from the data matrix, some diagnostics using contribution plots (discussed below)
and process insight should be used to sort out false alarms from the real ones. This process of model building and
validation is done iteratively until no multivariate control chart signals any real outlier. As a side effect of this
debugging procedure, the root causes of the out-of-control observations can be discovered, improving process
knowledge and future process performance.

1.04.9.2.2 PCA-based MSPC: online process monitoring (Phase II)

Once the reference PCA model and the control limits for the multivariate control charts are obtained, new
process observations can be monitored online. When a new observation vector zi is available, after preproces-
sing it is projected onto the PCA model yielding the scores and the residuals, from which the value of the
Hotelling TA2 and the value of the SPE are calculated. This way, the information contained in the original K
variables is summarized in these two indices, which are plotted in the corresponding multivariate TA2 and SPE
control charts. No matter what the number of the original variables K is, only two points have to be plotted on
the charts and checked against the control limits. The SPE chart should be checked first. If the points remain
below the control limits in both charts, the process is considered to be in control. If a point is detected to be
beyond the limits of one of the charts, then a diagnostic approach to isolate the original variables responsible for
120 Statistical Control of Measures and Processes

the out-of-control signal is needed. In PCA-based MSPC, contribution plots37 are commonly used for this
purpose.
Contribution plots can be derived for abnormal points in both charts. If the SPE chart signals a new out-of-
control observation, the contribution of each original kth variable to the SPE at this new abnormal observation
is given by its corresponding squared residual:
2
ContðSPE; xnew;k Þ ¼ enew;k 
¼ ðxnew;k – xnew;k Þ2 ð37Þ

where enew,k is the residual corresponding to the kth variable in the new observation and is the prediction
xnew;k
of the kth variable xnew;k from the PCA model.
In case of using the DModX statistic, the contribution of each original kth variable to the DModX is
given by44

ContðDModX; xnew;k Þ ¼ wk enew;k ð38Þ

where wk is the square root of the explained sum of squares for the kth variable. Variables with high
contributions in this plot should be investigated.
If the abnormal observation is detected by the TA2 chart, the diagnosis procedure is carried out in two steps:
(i) a bar plot of the normalized scores for that observation (tnew,a/a)2 is plotted and the ath score with the
highest normalized value is selected; (ii) the contribution of each original kth variable to this ath score at this
new abnormal observation is given by

Contðtnew;a ; xnew;k Þ ¼ pak xnew;k ð39Þ

where pak is the loading of the kth variable at the ath component. A plot of these contributions is created.
Variables on this plot with high contributions but with the same sign as the score should be investigated
(contributions of the opposite sign will only make the score smaller). When there are some scores with high
normalized values, an overall average contribution per variable can be calculated over all the selected scores.39
Contribution plots are a powerful tool for fault diagnosis. They provide a list of process variables that
contribute numerically to the out-of-control condition (i.e., they are no longer consistent with NOCs), but they
do not reveal the actual cause of the fault. Those variables and any variables highly correlated with them should
be investigated. Incorporation of technical process knowledge is crucial to diagnose the problem and discover
the root causes of the fault.
Apart from the TA2 and SPE control charts, other charts such as the univariate time-series plots of the scores
or scatter score plots can be useful (both in Phase I and II) for detecting and diagnosing out-of-control situations
and also for improving process understanding.

1.04.9.3 Potential of Latent Variables-Based MSPC in Industry

As commented by Kourti,39 multivariate monitoring schemes based on latent variables methods have been
receiving increasing attention by industrial practitioners in the last 20 years. Several companies have enthu-
siastically adopted the methods and have reported many success stories. Latent variables-based MSPC has been
used for process monitoring, fault detection, and diagnosis in continuous and batch industrial processes.
Kourti39 also provides a long list of publications reporting successful industrial applications in different fields
(e.g., petrochemical, polymer, and chemical).
In pharmaceutical industry, latent variables-based MSPC scheme is playing a critical role in the successful
implementation of process analytical technology (PAT) supported by the United States Food and Drug
Administration (FDA). PAT has been defined by the FDA as systems for the analysis and control of manufacturing
processes based on timely measurements during processes of critical quality parameters and performance attributes of raw and
in-process materials and processes to assure acceptable end product quality at the completion of the process.49 With this
initiative, the FDA tries to motivate the pharmaceutical industry to improve process control strategies for high-
quality, cost-effective pharmaceutical products. Kourti50 discusses the critical role of latent variables-based
MSPC methods for process understanding, abnormal situation detection, and fault diagnosis, as linked to PAT.
 Statistical Control of Measures and Processes 121

Latent variables-based MSPC techniques can also be used to extract subtle information from digital images
related to product quality and use such information for prediction, monitoring, and control. The methodology,
based on multivariate image analysis (MIA),51 can be used to monitor solids and other heterogeneous materials
(lumber, steel sheets, pulp and paper products, polymer films, multiphase streams, etc.). Image analysis provides
informative, inexpensive, and robust online sensors for the solids industry. Therefore, it opens new ways for the
successful monitoring and control of processes, which was traditionally difficult owing to lack of sensors.
Several applications have been reported in the literature: online monitoring of time-varying images52 and
lumber defects;53 monitoring and control of the amount of coating applied to the base food product and the
distribution of the coating among the individual product pieces in snack food industry;54,55 defect detection and
classification in ceramic tiles, artificial stone countertops, and orange fruits.56
Maintaining and controlling data quality is a key problem in large-scale microarray studies. In particular,
systematic changes in experimental conditions across multiple chips can seriously affect quality and even lead
to false biological conclusions. Model et al.57 introduce a novel approach for microarray process control based
on MSPC.

1.04.9.3.1 Example 3
An example of successful implementation of latent variables-based MSPC in wastewater treatment plants is
presented in Aguado et al.58 Data were collected from a sequencing batch reactor (SBR) operated for enhanced
biological phosphorus removal from wastewater. The reactor was equipped with a mechanical mixer, an air
diffuser, and several electronic sensors: temperature (Temp), pH, electric conductivity (Cond), oxidation–
reduction potential (ORP), and dissolved oxygen (DO). Data from the five electronic sensors were recorded
approximately every 1 min along the duration of the batch yielding 340 values per sensor and a total of K ¼ 1700
values per batch. The historical database available corresponding to 4 months of operation with the SBR was
used to build the in-control model in Phase I. A PCA-based MSPC scheme was designed after several cycles of
model validation and outlier detection (out-of-control batches). Finally, m ¼ 75 batches were selected to build
the in-control model. Technical knowledge of the process was also incorporated to check if the real faulty
batches in the database were detected and the responsible variables correctly identified (diagnosis ability). Note
that in this case the structure of the matrix X (75  1700) makes unfeasible the use of original variables-based
MSPC introduced in Section 1.04.9.1.
Figure 9 illustrates the proposed PCA-based MSPC scheme. The upper part shows the DModX control
chart (from Equation (34)) (Figure 9, top left) and TA2 control chart (from Equation (25)) (Figure 9, top right) at
significance level  ¼ 0.01. The first 30 batches represent in-control batches from Phase I. Batches 31 and 32 are
new batches monitored in Phase II. Batch 31 looks in control, whereas batch 32 exceeds the UCL in both charts
(values out of range), suggesting that this is an out-of-control batch. To isolate the variables responsible for this
out-of-control situation, a contribution plot in DModX of out-of-control batch 32 is displayed (Figure 9,
bottom left). The plot signals DO as the process variable related to the fault. The negative values of the
contributions indicate that DO takes lower values than required, may be due to an aeration system fault. This is
confirmed in Figure 9 (bottom right) by comparing the trajectory of DO for batch 32 (out-of-control) to that of
an in-control batch. Around the time point 124, the DO dramatically drops in batch 32, confirming a fault in the
aeration system control that should be fixed before a new batch starts.
The potential of latent variables-based MSPC is really impressive nowadays and will be a key tool for
success in the future high-tech environments.

1.04.9.4 Monitoring Measurement Systems: Applications of SPC in Analytical Laboratory

Methods
Every analytical or measurement process that works continuously (e.g., in routine analysis, or a method used for
a long period of time) can be considered as a process with inputs (the samples) and output (the results of the
analysis: measurements). Hence, the basics of SPC can be applied to monitor them leading to the so-called
analytical process monitoring. In this case, it is the quality of the measurement that matters. Control charts can
be used to attain and maintain statistical control of measuring systems. When the control chart signals an
122 Statistical Control of Measures and Processes

40
1.7
30
DmodX [7]

T 2 crit (99%)

T 2 [7]
D crit (99%)
1.3
20

0.9 10

0.5 0
0 5 10 15 20 25 30 0 5 10 15 20 25 30
Batch Batch

4 Anaerobic Aerobic Settle

6
Faulty batch
Contribution in DmodX

2 5 Normal batch
0

(mg l–1)
DO 4
–2
3
–4

DO
–6 2
pH ORP Cond Temp 1
–8
–10 0
0 340 680 1020 1360 1700 0 50 100 150 200 250 300
Variable number Time instant
Figure 9 PCA-based MSPC scheme for SBR data. DModX control chart (top left) and TA2 control chart (top right): the solid
red line represents control limits. Contribution plot in DModX of out-of-control batch 32 (bottom left) signaling DO as the
process variable related to the fault. Trajectory of DO for batch 32 (out of control due to an aeration system fault) and for an in-
control batch (bottom right).

out-of-control measurement, the analyst must search for the assignable cause and fix the problem before
reanalyzing some of the previous samples and continuing with other analyses.59
Instrumental methods of analysis are usually comparative in nature and the analytical instruments require
periodic recalibration to compensate for the effects of uncontrollable environmental factors and gradual
changes in instrument performance. To be confident of the test results, the analyst should have systems for
monitoring and controlling the accuracy and precision of the test method.
Standard or check samples (samples with known properties) are used to test analytical methods periodically
to ensure that they are in control. For example, latex suspensions with known particle size are used to test the
calibration in particle sizing methods, and samples with known concentration of an analyte are used to test gas
chromatography. The results for these samples can be plotted on control charts to test that the value measured
exhibits only common-cause variation. When the response obtained from the standard is not a single number
(i.e., instead of composition of one analyte, we get compositions of several analytes), one should consider
multivariate charts.40
In chromatography, process variables such as flow rate and pressure can be of significant importance for the
performance of the chromatographic system, whereas, for example, column aging or a contaminated detector can
take the system out of control. Nijhuis et al.60 present a PCA-based MSPC scheme of a chromatographic system
for the quantification of the total content of trans fatty acids in vegetable oils by capillary gas chromatography
using a check sample containing five analytes to test column performance. A TA2 chart and an SPE chart are used
to monitor analyte peak area percentage of five analytes of the check sample along time. The multivariate
monitoring scheme clearly outperforms the univariate approach in the sense of reducing the number of false
alarms and allowing the detection of real out-of-control observations due to a change in correlation.
Zhu et al.61 use online high-performance liquid chromatography (HPLC) to monitor a steady-state reaction
over 35.2 h, with 197 chromatograms recorded as the reaction progresses. For each chromatogram, peaks are
detected, baseline corrected, aligned, and integrated to provide a peak table consisting of intensities of 19 peaks,
 Statistical Control of Measures and Processes 123

2 corresponding to the reactants, 1 to the product, and 1 to the solvent, the remaining being impurities, by-
products, or intermediates. A PCA-based MSPC scheme successfully detects out-of-control samples and
provides diagnostic insight (using contribution plots) into why these samples are problematic. The TA2 charts
appear to be better at monitoring the reactants and products, and could be employed, for example, to determine
if there were problems with mixing or difficulties with instrument operation, whereas the SPE charts are best at
detecting impurities during the reaction.
When developing multivariate calibration models (as in SIA), it is crucial to ensure that the analytical system
works under statistical control during the modeling process, and that it remains under control while the unknown
samples are analyzed. To do so, a representative control sample is analyzed repeatedly before and during the
multivariate model development to check the system’s stability. Rius et al.14 illustrate the use of a PCA-based
MSPC approach in the development of an SIA method for determining sulfates in natural water using spectral data.
As shown in the previous examples, many analytical methods are based on liquid chromatography and
commonly the only measure of system stability is standards, injected repeatedly throughout the sequence.
Fransson et al.62 present a liquid chromatographic process control (LCPC) system where the analytical run
(chromatographic analysis) is treated as a process with the chromatographic data as the product. To ensure
high quality of the analytical data, focus is on not only already existing postrun checks of retention time,
plate numbers, resolution, and peak areas, but also the process monitoring itself. Several key parameters,
such as the pressure at the column, the injection valve, the temperature, and the conductivity of the mobile
phase, are monitored. LCPC provides better control of the analysis, enhancing the quality of the data
obtained. This allows a reduction in the number of standards and replicates. Furthermore, troubleshooting is
facilitated: Using contribution plots, system disturbance can easily be tracked back to the original para-
meters that caused the incident. The authors state that the approach can also be adapted to other flow
systems, such as gas chromatography, capillary electrophoresis, capillary electrochromatography, and flow
injection systems.
Apart from monitoring spectral signals in analytical processes, latent-based MSPC techniques can also be used
to diagnose the possible bias in analytical methods as a means of internal validation in analytical laboratories.
Ortiz-Estarelles et al.63 illustrate the use of MSPC tools for automatic detection of possible errors in the method
used for routine multiparametric analysis to design an internal multivariate analytical quality control (iMAQC)
program. Such tools have the ability to signal possible failures in the analytical methods without resorting to any
external reference, as they use their own analytical results as a source for the diagnosis of the method’s quality.
The potential of the proposed approach is illustrated in the routine analysis of drinking water.

1.04.10 Software

There exist many commercial software packages containing univariate SPC tools. A brief list includes
Statgraphics Centurion from StatPoint Inc., Minitab, STATISTICA from StatSoft, SAS, and SPSS.
Microsoft Excel also provides a basic univariate SPC module. Some of these packages (Statgraphics
Centurion, Minitab, and STATISTICA) also include multivariate SPC tools in the original variables.
Nevertheless, latent-based MSPC tools are not very common in commercial software; out of the already
mentioned, only STATISTICA offers these tools. Another multivariate specialized software package is SIMCA
from Umetrics. Some websites offer MATLAB freeware with specialized MSPC tools (e.g., the Multivariate
Statistical Engineering Research Group website: http://mseg.webs.upv.es).

Acknowledgments

The author gratefully acknowledges D. Aguado and the CALAGUA research group for providing the experi-
mental data of the wastewater treatment process, and Professor M. P. Callao for providing the SIA sulfate
determinations data set.
124 Statistical Control of Measures and Processes

References

1. Montgomery, D. C. Introduction to Statistical Quality Control, 5th ed.; Wiley: New York, 2005.
2. Woodall, W. H. Controversies and Contradictions in Statistical Process Control. J. Qual. Technol. 2000, 32 (4), 341–350.
3. Hawkins, D. M.; Olwell, D. H. Cumulative Sum Charts and Charting for Quality Improvement; Springer: New York, 1988.
4. Box, G. E. P.; Luceño, A. Statistical Control by Monitoring and Feedback Adjustment; Wiley: New York, 1997.
5. Shewhart, W. A. Economic Control of Quality of Manufactured Product; Van Nostrand: New York, 1931. Reissued by the
American Society for Quality: Milwaukee, WI, 1980.
6. Liu, R. Y.; Tang, J. Control Charts for Dependent and Independent Measurements Based on Bootstrap Methods. J. Am. Stat.
Assoc. 1996, 91, 1694–1700.
7. McNeese, W. H.; Klein, R. A. Statistical Methods for the Process Industries (Quality & Reliability 28); Marcel Dekker: New York,
1991.
8. Champ, C. W.; Woodall, W. H. Exact Results for Shewhart Control Charts with Supplementary Run Rules. Technometrics 1987,
29 (4), 393–399.
9. Page, E. S. Continuous Inspection Schemes. Biometrika 1954, 41, 110–115.
10. Page, E. S. Cumulative Sum Control Charts. Technometrics 1961, 3 (1), 1–9.
11. Roberts, S. Control Chart Tests Based on Geometric Moving Averages. Technometrics 1959, 42 (1), 97–102.
12. Hunter, J. S. Exponentially Weighted Moving Average. J. Qual. Technol. 1986, 18, 203–210.
13. Rius, A.; Ruisánchez, I.; Callao, M. P.; Rius, F. X. Reliability of Analytical Systems: Use of Control Charts, Time Series Models and
Recurrent Neural Networks (RNN). Chemom. Intell. Lab. Syst. 1998, 40, 1–18.
14. Rius, A.; Callao, M. P.; Rius, F. X. Multivariate Statistical Process Control Applied to Sulfate Determination by Sequential Injection
Analysis. Analyst 1997, 122, 737–741.
15. Box, G. E. P.; Jenkins, G. M.; Reinsel, G. C. Time Series Analysis: Forecasting and Control, 3rd ed.; Prentice Hall: New Jersey,
1994.
16. Callao, M. P.; Rius, A. Time Series: A Complementary Technique to Control Charts for Monitoring Analytical Systems. Chemom.
Intell. Lab. Syst. 2003, 66, 79–87.
17. Vander Wiel, S. A.; Tucker, W. T.; Faltin, F. W.; Doganaksoy, N. Algorithmic Statistical Process Control: Concepts and an
Application. Technometrics 1992, 34, 286–297.
18. Faltin, F. W.; Hahn, G. J.; Tucker, W. T.; Vander Wiel, S. A. Algorithmic Statistical Process Control: Some Practical Observations.
Int. Stat. Rev. 1993, 61, 67–80.
19. Box, G. E. P.; Kramer, T. Statistical Process Monitoring and Feedback Adjustment – A Discussion. Technometrics 1992, 34,
251–285.
20. Capilla, C.; Ferrer, A.; Romero, R.; Hualda, A. Integration of Statistical and Engineering Process Control in a Continuous
Polymerization Process. Technometrics 1999, 41, 14–28.
21. Ferrer, A.; Barceló, S.; Hermenegildo, F. Engineering Statistical Process Control (ESPC) in a Tile Industry. Proceedings of the
Industrial Statistics in Action 2000 International Conference. University of Newcastle upon Tyne, Newcastle, 2000; pp 145–158.
22. Janakiram, M.; Keats, J. B. Combining SPC and EPC in a Hybrid Industry. J. Qual. Technol. 1998, 30, 189–200.
23. MacGregor, J. F. On-Line Statistical Process Control. Chem. Eng. Prog. 1988, 84, 21–31.
24. Montgomery, D. C.; Keats, B. J.; Runger, G. C.; Messina, W. S. Integrating Statistical Process Control and Engineering Process
Control. J. Qual. Technol. 1994, 26, 79–87.
25. MacGregor, J. F. Using On-Line Process Data to Improve Quality. Is There a Role for Statisticians? Are They Up for the
Challenge? ASQC Stat. Div. Newsl. 1996, 16 (2), 6–13.
26. Ferrer, A. Multivariate Statistical Process Control Based on Principal Component Analysis (MSPC-PCA): Some Reflections and a
Case Study in an Autobody Assembly Process. Qual. Eng. 2007, 19, 311–325.
27. Hotelling, H. Multivariate Quality Control. In Techniques of Statistical Analysis; Eisenhart, C., Hastay, M., Wallis, W. A., Eds.;
MacGraw-Hill: New York, 1947; pp 111–184.
28. Jackson, J. E. Multivariate Quality Control. Commun. Stat. Theor. Meth. 1985, 14 (11), 2657–2688.
29. Wierda, S. J. Multivariate Statistical Process Control – Recent Results and Directions for Future Research. Stat. Neerl. 1994, 48
(2), 147–168.
30. Fuchs, C.; Kenett, R. S. Multivariate Quality Control: Theory and Applications (Quality & Reliability 54); Marcel Dekker: New York,
1998.
31. Tracy, N. D.; Young, J. C.; Mason, R. L. Multivariate Control Charts for Individual Observations. J. Qual. Technol. 1992, 24 (2),
88–95.
32. Alt, F. B. Multivariate Quality Control. In The Encyclopedia of Statistical Sciences; Kotz, S., Johnson, N. L., Read, C. R., Eds.;
Wiley: New York, 1985; pp 110–122.
33. Mason, R. L.; Champ, C. W.; Tracy, N. D.; Wierda, S. J.; Young, J. C. Assessment of Multivariate Process Control Techniques.
J. Qual. Technol. 1997, 29 (2), 140–143.
34. Lowry, C. A.; Woodall, W. H.; Champ, C. W.; Rigdon, S. E. A Multivariate Exponentially Weighted Moving Average Control Chart.
Technometrics 1992, 34, 46–53.
35. Kulahci, M.; Borror, C. Advanced Statistical Process Control. In Statistical Practice in Business and Industry; Coleman, S.,
Greenfield, T., Stewardson, D., Montgomery, D. C., Eds.; Wiley: West Sussex, UK, 2008; Chapter 13, pp 211–237.
36. Bersimis, S.; Psarakis, S.; Panaretos, J. Multivariate Statistical Process Control Charts: An Overview. Qual. Reliab. Eng. Int. 2007,
23, 517–543.
37. Kourti, T.; MacGregor, J. F. Multivariate SPC Methods for Process and Product Monitoring. J. Qual. Technol. 1996, 28 (4),
409–428.
38. Jackson, J. E. A User’s Guide to Principal Components; Wiley: New York, 2003.
39. Kourti, T. Application of Latent Variable Methods to Process Control and Multivariate Statistical Process Control in Industry. Int. J.
Adapt. Control Signal Process. 2005, 19, 213–246.
 Statistical Control of Measures and Processes 125

40. Kourti, T. Process Analysis and Abnormal Situation Detection: From Theory to Practice. IEEE Control Syst. 2002, 22, 10–25.
41. Martens, H.; Naes, T. Multivariate Calibration; Wiley: New York, 1989.
42. Wold, H. Nonlinear Estimation by Iterative Least Squares Procedures. In Research Papers in Statistics; David, F., Ed.; Wiley: New
York, 1966; pp 411–444.
43. Wold, S. Cross-Validatory Estimation of the Number of Components in Factor and Principal Component Models. Technometrics
1978, 20 (4), 397–405.
44. Eriksson, L.; Johansson, E.; Kettaneh-Wold, N.; Wold, S. Multi- and Megavariate Data Analysis: Principles and Applications;
Umetrics AB, 2001.
45. Box, G. E. P. Some Theorems on Quadratic Forms Applied in the Study of Analysis of Variance Problems: Effect of Inequality of
Variance in One-Way Classification. Ann. Math. Stat. 1954, 25, 290–302.
46. Jackson, J. E.; Mudholkar, G. S. Control Procedures for Residuals Associated with Principal Component Analysis.
Technometrics 1979, 21 (3), 341–349.
47. Nomikos, P.; MacGregor, J. F. Multivariate SPC Charts for Monitoring Batch Processes. Technometrics 1995, 37 (1), 41–59.
48. Box, G. E. P.; Hunter, J. S.; Hunter, W. G. Statistics for Experimenters: Design, Innovation and Discovery, 2nd ed.; Wiley:
Hoboken, NJ, 2005.
49. Guidance for Industry, PAT – A Framework for Innovative Pharmaceutical Development, Manufacturing, and Quality Assurance;
U.S. Department of Health and Human Services, Food and Drug Administration, Center for Drug Evaluation and Research
(CDER), Center for Veterinary Medicine (CVM), Office of Regulatory Affairs (ORA), Pharmaceutical CGMPs, September 2004.
http://www.fda.gov/cder/guidance/6419fnl.pdf
50. Kourti, T. Process Analytical Technology beyond Real-Time Analyzers: The Role of Multivariate Analysis. Crit. Rev. Anal. Chem.
2006, 36, 257–278.
51. Geladi, P.; Grahn, H. Multivariate Image Analysis; Wiley: Chichester, UK, 1996.
52. Bharati, M. H.; MacGregor, J. F. Multivariate Image Analysis for Real Time Process Monitoring and Control. Ind. Eng. Chem. Res.
1998, 37, 4715–4724.
53. Bharati, M.; MacGregor, J. F. Softwood Lumber Grading through On Line Multivariate Image Analysis. Ind. Eng. Chem. Res.
2003, 42, 5345–5353.
54. Yu, H.; MacGregor, J. F. Multivariate Image Analysis and Regression for Prediction of Coating and Distribution in the Production
of Snack Foods. Chemom. Intell. Lab. Syst. 2003, 67, 125–144.
55. Yu, H.; MacGregor, J. F.; Haarsma, G.; Bourg, W. Digital Imaging for On-Line Monitoring and Control of Industrial Snack Food
Processes. Ind. Chem. Eng. Res. 2003, 42, 3036–3044.
56. Prats-Montalbán, J. M.; Ferrer, A. Integration of Colour and Textural Information in Multivariate Image Analysis: Defect Detection
and Classification Issues. J. Chemom. 2007, 21, 10–23.
57. Model, F.; Konig, T.; Piepenbrock, C.; Adorjan, P. Statistical Process Control for Large Scale Microarray Experiments.
Bioinformatics 2002, 18 (Suppl 1), S155–S163.
58. Aguado, D.; Ferrer, A.; Ferrer, J.; Seco, A. Multivariate SPC of a Sequencing Batch Reactor for Wastewater Treatment. Chemom.
Intell. Lab. Syst. 2007, 85, 82–93.
59. Paul, W. L.; Barnett, N. S. Control Charting Instrumental Analyses. Lab. Autom. Inform. Manag. 1995, 31, 141–148.
60. Nijhuis, A.; de Jong, S.; Vandeginste, B. G. M. Multivariate Statistical Process Control in Chromatography. Chemom. Intell. Lab.
Syst. 1997, 38, 51–62.
61. Zhu, L.; Brereton, R. G.; Thompson, D. R.; Hopkins, P. L.; Escott, R. E. A. On-Line HPLC Combined with Multivariate Statistical
Process Control for the Monitoring of Reactions. Anal. Chim. Acta 2007, 584, 370–378.
62. Fransson, M.; Sparén, A.; Lagerholm, B.; Karlsson, L. On-Line Process Control of Liquid Chromatography. Anal. Chem. 2001, 73
(7), 1502–1508.
63. Ortiz-Estarelles, O.; Martı́n-Biosca, Y.; Medina-Hernández, M. J.; Sagrado, S.; Bonet-Domingo, E. On the Internal Multivariate
Quality Control of Analytical Laboratories. A Case Study: The Quality of Drinking Water. Chemom. Intell. Lab. Syst. 2001, 56,
93–103.
126 Statistical Control of Measures and Processes

Biographical Sketch

Alberto Ferrer is Professor of Statistics in the Department of Applied Statistics, Operation

Research and Quality, and Head of the Multivariate Statistical Engineering Research Group
of the Technical University of Valencia, Spain. He holds an M.Sc. in Agricultural
Engineering and a Ph.D. in Statistics. His research focuses on statistical techniques for quality
and productivity improvement, especially those related to multivariate statistical projection
methods. Professor Ferrer is currently an associate editor of Technometrics, a member of the
editorial board of Quality Engineering, a member of the Council of the International Society
for Business and Industrial Statistics (ISBIS), and a member of the European Network for
Business and Industrial Statistics (ENBIS) and the Spanish Chemometrics Network. He is
also active as industrial consultant on Six Sigma and Process Analytical Technology (PAT).