You are on page 1of 11

ARTICLE IN PRESS

WAT E R R E S E A R C H 41 (2007) 2679 – 2689

Available at www.sciencedirect.com

journal homepage: www.elsevier.com/locate/watres

Control charts for the toxicity of finished water—Modeling


the structure of toxicity

Eleni M. Smetia,, Demetrios E. Koronakisb, Spyridon K. Golfinopoulosa


a
Department of Financial & Management Engineering, University of the Aegean, 31 Fostini Street, 82100 Chios, Greece
b
Water Quality Control Division, Water Supply & Sewerage Corporation of Athens, 156 Oropou Street, 11146 Athens, Greece

art i cle info ab st rac t

Article history: The aim of this study was to construct control charts for the data that were obtained from
Received 5 February 2006 the daily toxicological analysis of finished water (that is the water which has already been
Received in revised form treated for human consumption) from the treated-water tanks of the Water Supply &
15 February 2007 Sewerage Corporation of Athens. The basic idea of the control charts is to test the
Accepted 18 February 2007 hypothesis that there are only common causes of variability versus the alternative that
Available online 17 April 2007 there are special causes. The control charts are designed and evaluated under the
Keywords: assumption that the observations from the process are independent and identically
Statistical process control distributed (iid) normally. However, the independence assumption is often violated in
Control charts practice. Time or serial dependence (autocorrelation) may be present in many chemical
Autocorrelation procedures, and may have a significant effect on the properties of the control charts. The
Time series models fact that a serious amount of autocorrelation was present in the data for the toxicity of the
Water toxicity fished water had a serious impact on the typical control charts. The problem of the
autocorrelation was overcome by the usage of a more sophisticated method based on time-
series ARIMA models.
& 2007 Elsevier Ltd. All rights reserved.

1. Introduction the variability. There are two reasons of variability: common


causes and special causes. Common causes concern the
Ecotoxicology deals with potentially harmful effects of natural variability that always exists in every process and
man-made chemicals, released into biosphere, on organisms there is no way to avoid it. Special causes are not an inherent
in the receiving environments. Ecotoxicological moni- characteristic of the process, and therefore, they can be
toring techniques are playing an increasingly important identified and eliminated. SPC aims at the separation of the
role in the evaluation of water quality. Traditionally the two types of variability.
necessary tests involve the use of fish. However, the use of The main tools of SPC are the control charts. The control
fish in toxicity testing is expensive and time-consuming, charts (Fig. 1) attempt to follow-up the process in order to
so alternatives are required. Microbial tests have locate and then eliminate the special causes of variability,
similar complex biochemical functions to those of higher resulting in the improvement of the process. The most
organisms. important types of control charts for variables are: the
To test the quality of drinking water in terms of toxicity, Shewhart control charts, the exponentially weighted moving
there are several methods of statistical process control (SPC). average (EWMA) control charts and the cumulative sum
SPC uses statistical methods to improve the quality of a (CUSUM) control charts. All types of control charts have some
process. This can be achieved by the systematic reduction of common features. The center line corresponds to average

Corresponding author. Tel.: +30 2102823257; fax: +30 2271035499.


E-mail address: e.smeti@fme.aegean.gr (E.M. Smeti).
0043-1354/$ - see front matter & 2007 Elsevier Ltd. All rights reserved.
doi:10.1016/j.watres.2007.02.036
ARTICLE IN PRESS
2680 WAT E R R E S E A R C H 41 (2007) 2679– 2689

Nomenclature PDM the time series of the toxicity data from the
Menidi old treated-water tank
ASP the time series of the toxicity data from the PDMf the time series of the residuals for the Menidi old
Aspropyrgos treated-water tank treated-water tank
ASPf the time series of the residuals for the Aspropyr- q the order of the moving average ARIMA operator
gos treated-water tank s the sample standard deviation
c4 a function of the sample size which is used for zi the exponentially weighted moving average sta-
estimating sigma by the sample standard devia- tistic
tion SHi the superior cumulative sum for the i sample
d the order of the non-seasonal differencing opera- SLi the inferior cumulative sum for the i sample
tor Xi the observation at time i
d2 a control chart function used for estimating sigma X̄ the average (arithmetic mean) of a sample
by the moving ranges
dASP stationary series formed by differencing the ASP Greek symbols
by 1 lag
dNDM stationary series formed by differencing the NDM l a constant in the formula for the exponentially
by 1 lag weighted moving average statistic
dPDM stationary series formed by differencing the PDM s2 variance
by 1 lag s standard deviation
ei the residual or error, i.e. the difference between
the observed value and the predicted value for the
Abbreviations
time period i
h a multiplier of the standard deviation for a
ACF autocorrelation function
cumulative sum control chart
AIC Akaike’s information criterion
H the decision interval for a cumulative sum control
ARL average run length
chart
ARL0 in-control average run length
Ht the inhibitory effect of a sample after the contact
BIC (Schwartz) Bayesian information criterion
time t in %
CCC common cause chart
I0 the initial luminescence intensity of the control
suspension in relative luminescence units
CUSUM cumulative sum control chart
It the luminescence intensity of sample after the
CV coefficient of variation

contact time t in relative luminescence units


EWMA exponentially weighted moving average control
chart
K the reference value for a cumulative sum control
chart
MAE mean absolute error

MR moving range
ME mean error

MR the average of the moving ranges


MSE mean squared error

NDM the time series of the toxicity data from the


PACF partial autocorrelation function
Menidi new treated-water tank RMSE root mean squared error
NDMf the time series of the residuals for the Menidi new QC quality control
treated-water tank SCC special cause chart
p the order of the autoregressive ARIMA operator SPC statistical process control

performance, whereas the control limits (the other two lines) ‘‘autocorrelation’’ is used to describe the correlation of one
correspond to the expected range of variation based on the variable at one point in time with observations of the same
process. If all the points plot between the two control limits variable at prior time points. When autocorrelation is present,
and do not exhibit any identifiable pattern the process is said the in-control average run length (ARL), which concerns the
to be in statistical control. expected number of samples that should be received in order
The typical control charts (Shewhart, CUSUM, EWMA) are to point out an erroneous signal (false alarm) for an out-of-
designed and evaluated on the assumption that the observa- control process (while it is actually in control) decreases
tions from the process are independent and identically dramatically. Therefore, there are problems of noticing
distributed (iid) normally. However, the independence as- ‘‘special causes’’ that do not exist and not detecting ‘‘special
sumption is often violated in practice in some manufacturing causes’’ that truly exist, implying a high probability of false
processes, such as chemical processes, where consecutive positives and/or false negatives. As autocorrelation may have
measurements on process or product characteristics are often a significant effect on the properties of the typical SPC control
highly autocorrelated (Montgomery, 2001). ‘‘Dependence’’ in a charts they seem to be inappropriate for monitoring process
time series refers to ‘‘serial dependence’’. So, the term quality.
ARTICLE IN PRESS
WAT E R R E S E A R C H 4 1 (200 7) 267 9 – 268 9 2681

Fig. 1 – Shewhart charts (I Charts) on original data for PDM, NDM and ASP. The estimation of r is based on the average moving
range or the sample standard deviation.

Two techniques have been propounded in the field of by other authors (see, e.g. Wardell et al., 1992; Montgomery
control charts that deal with autocorrelation. The first and Mastrangelo, 1991; Runger et al., 1995; Lu and Reynolds,
approach is based on the typical control charts of the original 1999a, 1999b, 2001).
observations with an adjustment to the control limits to take In this study, we analyze toxicity data of water for human
the autocorrelation into account (see, e.g. Vasilopoulos and consumption from three treated-water tanks of the Water
Stamboulis, 1978; Wardell et al., 1992, 1994; Zhang, 1998; Van Supply and Sewerage Company of Athens. Two of these tanks
Brackle and Reynolds, 1997; Lu and Reynolds, 2001). This are in Menidi (old and new treated-water tanks) and the third
approach is advisable only in case of small values of one is in Aspropyrgos. The finished water in each of these
autocorrelation. The second approach, which is the main tanks comes from the ‘‘Menidi old unit’’, ‘‘Menidi new unit’’
one, has been firstly introduced by Alwan and Roberts (1988). and ‘‘Aspropyrgos unit’’ water treatment plants, respectively.
They suggested the application of typical control charts to the The raw water in the above water treatment plants comes
residuals from the fit of a time-series model to the data. exclusively from the same reservoir, Mornos, which is an
Control charts based on residuals have also been considered artificial lake. Mornos is 250 km away from Athens. Its
ARTICLE IN PRESS
2682 WAT E R R E S E A R C H 41 (2007) 2679– 2689

maximum capacity is 700 million m3 and the surrounding 83.5 with a total standard deviation (‘‘within’’ and ‘‘between’’)
area is free from any activities. The water treatment of 5.9. This is an estimate of the bioluminescence test
procedure involves the following steps: variability over time. The standard deviation within duplicate
Prechlorination–Flocculation–Sedimentation–Filtration and measurements of the water samples was 1.7.
Postchlorination.
The typical SPC techniques and the time-series method 2.2. Theoretical background of the statistical analysis
were applied to the data. This application showed the serious
effects of autocorrelation when the typical SPC control charts 2.2.1. Typical control charts of SPC
are applied to autocorrelated observations. The problem of The first and simplest control chart for individual observa-
the autocorrelation was overcome by the usage of the Alwan tions, which is used for the control of the process mean, is the
and Roberts (1988) approach. Shewhart chart for individuals (I Chart). In this chart, the
measurement data are plotted according to the time order.
The center line is set in the mean of the process or in its
2. Materials and methods estimation and the control limits are usually in a distance of
three standard deviations from both sides of the mean. The
2.1. Analytical toxicological control of treated water standard deviation of the process is often practically un-
known and is estimated from the ‘‘moving ranges’’ of size 2 as
The toxicity data from the Menidi old treated-water tank MR=d2 or from the sample standard deviation of the data as
(variable PDM), the Menidi new treated-water tank (variable s=c4 , where d2 (which is exactly 1.128) can be read from special
NDM) and the Aspropyrgos treated-water tank (variable ASP) tables (see Ryan, 1989), and c4 is a function of the sample size
have been registered using a bioluminescence test, which is and approximates the value 1 as the sample size increases.
based on the correlation between toxicity of the water sample These charts are appropriate for fast detection of large shifts
and its effects on the light intensity of marine bacteria Vibrio in the process, while they are insensible to small shifts. For
fischeri (former Photobacterium phosphoreum), measured by the the increase of the effectiveness of the Shewhart charts there
bioluminometer Microtox (Strategic Diagnostics Microtox have been recommended criteria which indicate that a
1010, 2003). Toxic substances causing a disturbance to the process shift or trend has begun. These criteria are known
normal metabolism of the bacterium result in a reduction of as run rules and are based on runs of consecutive points
light output. The degree of toxicity is proportional to the increasing/decreasing or oscillating above and below the
measured light loss. center line.
Samples were adjusted to contain 2% NaCl, which provides The EWMA and the CUSUM charts are better at detecting
osmotic protection for the marine bioassay organism. small changes in the process mean than Shewhart charts
Expression of toxicity is the inhibitory effect of a test (Montgomery, 2001; Champ and Woodall, 1987).
sample after the contact time of 30 min (H30 Þ. This expression The EWMA chart was first introduced by Roberts (1959). The
is given by the below formula: EWMA statistic is defined as zi ¼ lxi þ ð1  lÞzi1 and it is a
weighted average of all past and current observations with
Ht ¼ ½ðI0  It Þ=It   100,
weights that decrease geometrically. The constant l can take
where Ht is the inhibitory effect of a sample after the contact values in the interval 0olp1. The most common choices for l
time of 30 min in %, It is the luminescence intensity of sample are in the interval ½0:05; 0:25. The starting value z0 is the
after the contact time of 30 min in relative luminescence units process target value m0 or the average of preliminary data or, if
and I0 is the initial luminescence intensity of the control it is not available, the sample mean. The EWMA control chart
suspension in relative luminescence units. is constructed by plotting zi versus the time order i. Additional
If the difference of light intensity is more than 20% the information about the EWMA chart can be found elsewhere
sample is considered to be toxic. This is an upper specifica- (e.g. Lucas and Saccucci, 1990).
tion limit for the toxicity of finished water. The CUSUM charts were first introduced by Page (1954).
The data available concern 163 daily samples with size There are two approaches for CUSUM charts: the tabular
n ¼ 1 for each one of the Menidi tanks and 91 for the CUSUM and the V-mask CUSUM. Montgomery (2001) strongly
Aspropyrgos tank. All the tests were carried out according to advises not to use the V-mask procedure due to disadvan-
standard protocols described in the Microtox Acute Toxicity tages and problems that are associated with this scheme. The
Users’ Guide. The duplicate basic test procedure was used. In tabular CUSUM is constructed as follows. For each observa-
duplicate testing, data quality is improved through cross tion two cumulative sums SHi and SLi are calculated for the
comparison. The duplicate test results are unacceptable if the detection of positive and negative shifts of the mean,
light measurements are different by more than 20% of either respectively. The cumulative sums are named superior and
value. The average value from the two readings is used in inferior CUSUMS, respectively. It is
subsequent calculations. All the tests were carried out by the
SHi ¼ max½0; Xi  ðX̄ þ KÞ þ SHi1  and
same person under the same operating conditions and on the
SLi ¼ max½0; ðX̄  KÞ  Xi þ SLi1 .
same day on which each water sample was received. With
respect to the Microtox test repeatability (within-laboratory The initial prices of these sums are zero ðSH0 ¼ SL0 ¼ 0Þ. The
precision), zinc sulphate was used (as the reference toxicant) price of K is selected to be the half of the mean change that is
to produce internal quality control (QC) data. The overall wished to be detected, usually 0:5s, and H ¼ hs (with h usually
mean of the reference toxicant duplicate measurements was 4 or 5) is the decision interval. If either SHi or SLi exceeds the
ARTICLE IN PRESS
WAT E R R E S E A R C H 4 1 (200 7) 267 9 – 268 9 2683

decision interval H, then the process is considered to be out of BIC tends to choose models that are more parsimonious than
control. those chosen by AIC.
The moving-range control chart MR, where MRi ¼ jXiþ1  Xi j
and Xi is the observation at time i, may be used to control
variability. However, some authors (e.g. Sullivan and Woodall, 3. Results–discussion
1996a) have pointed out that this chart cannot really
contribute significantly to the identification of a shift in 3.1. Typical SPC control charting for the toxicity data
process variability.
Table 1 shows the descriptive statistics of the toxicity data for
the three treated-water tanks. The normality tests (Kolmo-
2.2.2. The time-series model based approach for gorov Smirnov–Lilliefore Correction) indicated that the null
autocorrelated data (Alwan and Roberts approach) hypothesis, according to which the data are distributed
The main approach for autocorrelated data was proposed by normally, cannot be rejected in any of the treated-water
Alwan and Roberts (1988) and uses time-series modeling to tanks at 5% significant level (p-values: more than 0.150 for
help the detection of the existence of systematic variation both PDM and ASP, and 0.055 for NDM).
and to gain a more accurate sorting out of special causes. The typical Shewhart charts on original data (Fig. 1), using
Alwan and Roberts suggest the implementation of two charts: moving ranges (MR=d2 ) to estimate standard deviation s,
The common cause chart (CCC) and mainly the special cause detected several out-of-control points in every treated-water
chart (SCC). The CCC is a time plot of the fitted values tank (13 in the Menidi old tank, 9 in the Menidi new tank and
obtained when the autocorrelated process is modeled with an 12 in the Aspropyrgos tank). On the other hand, the typical
ARIMA model. The CCC is not a control chart because of the Shewhart charts on original data, using the sample standard
fact that it does not have control limits but it basically deviation ðs=c4 Þ to estimate standard deviation s, gave only
accounts for the systematic variation that exists in the one point out of the control limits in every water tank (Fig. 1).
process. The SCC is a typical SPC control chart applied to In the first case there are many false positive signals and in
the residuals of the time-series ARIMA model. the second one, some false negatives might occur. In case the
data are not autocorrelated, there is no substantial difference
between the two approaches of the estimation of s (Alwan,
2.2.3. ARIMA models
1992). In presence of autocorrelation, the estimator of s that is
The Box–Jenkins methodology (Box et al., 1994) is followed to
based upon moving ranges underestimates s (Ryan, 1989)
find an appropriate ARIMAðp; d; qÞ model. This methodology
giving as a result many false out of control indications,
involves three steps:
whereas the estimator of s that is based on the sample
standard deviation overestimates s resulting into non-sensi-
a. Identification of the model (specification of the number of tive control charts even though we have large shifts on the
differences ðdÞ, of the original series that is needed to process (Wardell et al., 1994). Therefore, if the data are time
produce the stationary process and identification of the depended, the typical SPC control charts are inadequate.
orders p and q for the autoregressive and moving average). Herein, the form of the data in the charts indicates that
b. Estimation of the parameters. there exists a serious amount of autocorrelation in the data.
c. Diagnostic checking. The inquest of special causes becomes even more compli-
cated if run rules are applied to these control charts (Alwan,
An appropriate model is selected by using criteria such as 1992). Therefore, the Shewhart control charts on these data
mean squared error (MSE), Akaike’s information criterion are inadequate to distinguish common causes from special
(AIC), Schwartz Bayesian information criterion (BIC). The MSE causes of variation.
is the average or mean value of the squared errors (residuals).
When two or more models are compared, the one with the 3.2. Time-series modeling
smaller MSE is the one with the better fit. The Box–Jenkins
methodology premises that the selected model has as few The first 80% of the total measurements of each variable were
parameters as possible in terms of parsimony (simplicity). used to identify and estimate the parameters of the time-
The AIC (Akaike, 1974) and BIC (Schwartz, 1978) take into series models (estimation period). The rest of the data (20%)
account the number of terms in the models. The penalty term were withheld to validate the models (validation period).
of BIC is more stringent than the penalty term of AIC. Thus, Holding data out for validation purposes is a very important

Table 1 – Descriptive statistics of the toxicity data for PDM, NDM and ASP

Toxicity Number of samples Mean Standard deviation Minimum Median Maximum Range

PDM 163 4.4 12.64 49 4 20 69


NDM 163 4.5 12.23 43 4 19 62
ASP 91 0.8 9.98 29 2 19 48
ARTICLE IN PRESS
2684 WAT E R R E S E A R C H 41 (2007) 2679– 2689

Autocorrelation Function for PDM Partial Autocorrelation Function for PDM


1 1

Partial Autocorrelations
Autocorrelations
0.6 0.6
0.2 0.2
-0.2 -0.2
-0.6 -0.6
-1 -1
0 5 10 15 20 25 0 5 10 15 20 25
lag lag
Autocorrelation Function for NDM Partial Autocorrelation Function for NDM

Partial Autocorrelations
1 1
Autocorrelations

0.6 0.6

0.2 0.2

-0.2 -0.2

-0.6 -0.6

-1 -1
0 5 10 15 20 25 0 5 10 15 20 25
lag lag
Autocorrelation Function for ASP Partial Autocorrelation Function for ASP
Partial Autocorrelations

1 1
Autocorrelations

0.6 0.6
0.2 0.2

-0.2 -0.2

-0.6 -0.6

-1 -1
0 5 10 15 20 25 0 5 10 15 20 25
lag lag

Fig. 2 – Sample autocorrelation functions and sample partial autocorrelation functions for PDM, NDM and ASP.

diagnostic test of a model. It gives the best indication of the different from zero after lag 1. The PACFs decay quickly.
accuracy that can be expected when forecasting the future. A Therefore, an ARIMAð0; 1; 1Þ would be an advisable model
good model has similar error measures in the validation in order to explain the process. However, several alternative
period to those in the estimation period (although they are models (stationary and non-stationary) were examined
often at least slightly larger (Nau, 2005). and the final selection of the appropriate ARIMA models
The sample autocorrelation function (ACF) and the sample was based on BIC and on the performance of the models in
partial autocorrelation function (PACF) with 95% probability the validation set. Low BIC value in comparison to other
limits (Fig. 2) confirm that the measurements on water models is indicative of a good model. Table 2 summarizes the
toxicity are moderate to highly positively autocorrelated in performance of the selected models for both the estimation
every treated-water tank. The graphs of the ACF and PACF and validation periods. It displays the root squared error
show that the procedure is not stationary for PDM, NDM and (RMSE), the mean absolute error (MAE) and the mean error
ASP. Stationarity is achieved by taking the first differences of (ME). The selected models gave similar results both for the
the series (differencing by one lag). That is, the creation of a estimation and the validation periods. Therefore, the models
new time series of successive differences (Xi  Xi1 , where Xi are likely to perform as well as expected in forecasting the
is the observation at time i and Xi1 is the observation of the future.
previous time period). The new time series that the differen- The estimated models are ARIMAð0; 1; 1Þ for each one of the
cing produces from PDM, NDM and ASP are named dPDM, three variables.
dNDM and dASP, respectively. The ACFs and PACFs for
the new variables (Fig. 3) are useful identification tools of
the orders p and q for the autoregressive and moving PDM: Xi ¼ Xi1  0:6294ei1 þ ei ,
average ARIMA operators. The ACFs of the differenced series NDM: Xi ¼ Xi1  0:6954ei1 þ ei ,
display a sharp cut-off at lag 1 and is not significantly ASP: Xi ¼ Xi1  0:5883ei1 þ ei ,
ARTICLE IN PRESS
WAT E R R E S E A R C H 4 1 (200 7) 267 9 – 268 9 2685

Autocorrelation Function for dPDM Partial Autocorrelation Function for dPDM


1

Partial Autocorrelations
1

Autocorrelations
0.6 0.6

0.2 0.2

-0.2 -0.2

-0.6 -0.6

-1 -1
0 5 10 15 20 25 0 5 10 15 20 25
lag lag

Autocorrelation Function for dNDM Partial Autocorrelation Function for dNDM


1

Partial Autocorrelations
1
Autocorrelations

0.6 0.6

0.2 0.2

-0.2 -0.2

-0.6 -0.6

-1 -1
0 5 10 15 20 25 0 5 10 15 20 25
lag lag

Autocorrelation Function for dASP Partial Autocorrelation Function for dASP


Partial Autocorrelations

1 1
Autocorrelations

0.6 0.6

0.2 0.2

-0.2 -0.2

-0.6 -0.6

-1 -1
0 5 10 15 20 25 0 5 10 15 20 25
lag lag

Fig. 3 – Sample autocorrelation functions and sample partial autocorrelation functions for dPDM, dNDM and dASP.

Table 2 – Error statistics for the selected models of the toxicity data for PDM, NDM and ASP

Statistic PDM: ARIMA(0,1,1) with no NDM: ARIMA(0,1,1) with no ASP: ARIMA(0,1,1) with no
constant constant constant

Estimation Validation Estimation Validation Estimation Validation


period period period period period period

RMSE 9.33073 9.05085 9.28666 9.79153 6.41239 5.80558


MAE 6.98312 7.62017 7.32305 7.78385 5.03069 4.78155
ME 0.022535 0.483222 0.435758 0.573015 0.465342 1.75819

where Xt is the observation at time t, Xt1 is its previous time- for all lags up to the 48th lag, equal zero is not rejected. The
period value, et are the residuals and et1 are the previous ACF and PACF of the residuals (Fig. 4) corroborate the
time-period residuals. conclusion that the residuals are not autocorrelated. We
The series of the residuals from the models that were can, therefore, conclude that the autocorrelation is elimi-
applied to PDM, NDM and ASP are named PDMf, NDMf and nated. The normal probability plots as well as the KS–Lillifore
ASPf, respectively. The residuals fulfilled all the diagnostic correction tests for normality indicate that the residuals are
checks. The p-value of the Ljung–Box statistic equals to following a normal distribution. The Breusch and Pagan test
0:934; 0:836 and 0.175 for PDM, NDM and ASP, respectively and for difference in variance of the residuals between the first
signifies that the null hypothesis that the autocorrelations, half and the second half was not significant. We can,
ARTICLE IN PRESS
2686 WAT E R R E S E A R C H 41 (2007) 2679– 2689

Autocorrelation Function for PDMf Partial Autocorrelation Function for PDMf


1 1

Partial Autocorrelations
Autocorrelations
0.6 0.6

0.2 0.2

-0.2 -0.2

-0.6 -0.6

-1 -1
0 5 10 15 20 25 0 5 10 15 20 25
lag lag
Autocorrelation Function for NDMf Partial Autocorrelation Function for NDMf
1 1

Partial Autocorrelations
Autocorrelations

0.6 0.6

0.2 0.2

-0.2 -0.2

-0.6 -0.6

-1 -1
0 5 10 15 20 25 0 5 10 15 20 25
lag lag
Autocorrelation Function for ASPf Partial Autocorrelation Function for ASPf
1 1
Partial Autocorrelations

0.6
Autocorrelations

0.6
0.2 0.2
-0.2 -0.2

-0.6 -0.6

-1 -1
0 5 10 15 20 25 0 5 10 15 20 25
lag lag

Fig. 4 – Sample autocorrelation functions and sample partial autocorrelation functions for PDMf, NDMf and ASPf.

therefore, infer that the residuals can be considered as ‘‘white Wardell et al. (1994) showed that the Shewhart chart on
noise’’ and the estimated models are good models for the residuals has a small sensitivity in the small shifts when the
time series. Now, typical SPC control charts can be applied on process is positively autocorrelated and they consider that
the residuals. EWMA or CUSUM on the residuals can give better results in
Fig. 5 shows the CCC, which are charts of the fitted values detecting small shifts, as in the case of traditional SPC control
from the ARIMA models for the toxicity in each one of the charts. In order to improve the sensitivity of the Shewhart
three treated-water tanks and they give the current level of chart on residuals Runger et al. (1995) suggested the use of the
the conditional mean of the processes. On the same chart, the CUSUM chart on the residuals.
actual observations were plotted, although the CCC does not Fig. 7 presents the EWMA charts on the residuals PDMf,
require it. NDMf and ASPf. In order to achieve the same in-control ARL
A SCC is a Shewhart chart for individuals applied to ðARL0 ¼ 370:4Þ as in the case of 3-sigma Shewhart charts, we
the residuals. Fig. 6 shows the SCC charts of the residuals use l ¼ 0:2 and 2.859-sigma control limits (Lu and Reynolds,
(PDMf, NDMf and ASPf for the PDM, NDM and ASP treated- 1999b). EWMA charts on the residuals give 2 points out of the
water tanks, respectively), using moving ranges to estimate control limits for PDMf and none for NDMf and ASPf.
the standard deviation. There are 3 points out of the control Fig. 8 presents the CUSUM charts on the residuals of the
limits for PDMf and none for NDMf and ASPf. Even though variables PDMf, NDMf and ASPf. These charts are designed
we use the sample standard deviation s to estimate the using K ¼ 0:5s. Therefore, they are appropriate for detecting
standard deviation s, there is no substantial difference in shifts of magnitude one standard deviation. The decision
these graphs. Therefore, the differences between the two interval is H ¼ 4:77s in order to achieve the same in-control
ways to estimate standard deviations that were noticed ARL ðARL0 ¼ 370:4Þ as in the case of 3-sigma Shewhart charts
when we applied Shewhart charts on the original data are (Hawkins, 1993). The two-sided CUSUM charts on the
eliminated. residuals give 5 points out of the control limits for PDMf
ARTICLE IN PRESS
WAT E R R E S E A R C H 4 1 (200 7) 267 9 – 268 9 2687

Time Sequence Plot for PDM Time Sequence Plot for NDM
ARIMA(0,1,1) ARIMA(0,1,1)
30 37
actual actual
10 fitted 17 forecast
PDM

NDM
-10 -3

-30 -23

-50 -43
0 30 60 90 120 150 180 0 30 60 90 120 150 180
Time Sequence Plot for ASP
ARIMA(0,1,1)
20
actual
10
forecast
0
ASP

-10

-20

-30
0 20 40 60 80 100

Fig. 5 – Time sequence plots (CCC) for PDM, NDM and ASP.

I Chart of PDMf I Chart of NDMf


30 UCL=28.64
40

30 20
UCL=25.43
Individual Value

Individual Value

20
10
10 _
_ X=0.46
0 0
X=-0.08

-10 -10
-20
LCL=-25.59 -20
-30
LCL=-27.72
-40 -30
1 16 32 48 64 80 96 112 128 144 160 1 16 32 48 64 80 96 112 128 144 160
Observation Observation

I Chart of ASPf
20
UCL=17.08

10
Individual Value

_
0 X=-0.72

-10

LCL=-18.52
-20
1 9 18 27 36 45 54 63 72 81 90
Observation

Fig. 6 – Special cause charts—I Charts for PDMf, NDMf and ASPf.

and none for NDMf and ASP. CUSUM charts show more points of future quality (Wardell et al., 1994). Similarly, the SCC can
out of control than the EWMA charts do. These additional be used to detect any assignable cause, including changes in
out-of-control points were close to the control limits of the the structure of the time series.
EWMA charts. Therefore, if we compare the CUSUM charts The differences between the estimated models for the three
with the EWMA charts we can see that the CUSUM charts are treated-water tanks can be explained by the fact that,
more effective in detecting small shifts than the EWMA although the initial raw water comes from the same reservoir,
charts. there are slight differences in the way of treatment between
The Alwan and Roberts (1988) approach takes advantage of the three treatment water plants and in the time that the
the fact that the process is autocorrelated to make forecasts water remains in each tank.
ARTICLE IN PRESS
2688 WAT E R R E S E A R C H 41 (2007) 2679– 2689

EWMA Chart of PDMf EWMA Chart of NDMf


10 10
+2.9SL=9.42
+2.9SL=8.02

5 5
EWMA

EWMA
_
_ _
0 X=-0.08 0 X=0.46

-5 -5

-2.9SL=-8.18 -2.9SL=-8.49
-10 -10
1 16 32 48 64 80 96 112 128 144 160 1 16 32 48 64 80 96 112 128 144 160
Sample Sample

EWMA Chart of ASPf


5.0 +2.9SL=4.93

2.5
EWvMA

0.0 _
_
X=-0.72

-2.5

-5.0
-2.9SL=-6.38
-7.5
1 9 18 27 36 45 54 63 72 81 90
Sample

Fig. 7 – EWMA charts for PDMf, NDMf and ASPf.

CUSUM Chart of PDMf CUSUM Chart of NDMf


50
50
UCL=44.8
UCL=40.6

25
Cumulative Sum
Cumulative Sum

25

0 0 0 0

-25 -25

LCL=-40.6
LCL=-44.8
-50 -50
1 16 32 48 64 80 96 112 128 144 160 1 16 32 48 64 80 96 112 128 144 160
Sample Sample

CUSUM Chart of ASPf


30
UCL=28.30
Cumulative Sum

20

10

0 0

-10

-20

LCL=-28.30
-30
1 9 18 27 36 45 54 63 72 81 90
Sample

Fig. 8 – CUSUM charts for PDMf, NDMf and ASPf.

There is no reason for concern about the out-of-control cient to ensure good quality characteristics (Deming, 1986). A
indications that were located by the control charts used, as all modern definition of quality is ‘‘quality is inversely propor-
the original measurements of the toxicity in every one of the tional to variability’’ (Montgomery, 2001). In this sense, quality
treated-water tanks were under the specification limit. improvement means the reduction of the unwanted varia-
However, meeting only the specification limits is not suffi- bility. The control charts aim exactly at detecting the
ARTICLE IN PRESS
WAT E R R E S E A R C H 4 1 (200 7) 267 9 – 268 9 2689

occurrence of special cause of variability. An out-of-control Deming, W.E., 1986. Out of Crisis. Massachusetts Institute of
indication shows that something unusual has happened in Technology, Center for Advanced Engineering Study, Cam-
the process. After the investigation of the process, one can bridge, MA.
Hawkins, D.M., 1993. Cumulative sum control charting: an
trace the cause of trouble. Hence, corrective actions may be
underutilized SPC tool. Qual. Eng. 5, 463–467.
undertaken before problems accumulate. Lu, C.W., Reynolds Jr, M.R., 1999a. EWMA control charts for
monitoring the mean of autocorrelated processes. J. Qual.
Technol. 31, 166–188.
4. Conclusions Lu, C.W., Reynolds Jr, M.R., 1999b. Control charts for monitoring
the mean and variance of autocorrelated processes. J. Qual.
The typical SPC control charts are ineffective, because of the Technol. 31, 259–274.
existence of autocorrelation in the toxicity data for the three Lu, C.W., Reynolds Jr, M.R., 2001. CUSUM charts for monitoring
treated-water tanks. The serious amount of autocorrelation an autocorrelated process. J. Qual. Technol. 33,
316–334.
that was present in the data dramatically reduced the
Lucas, J.M., Saccucci, M.S., 1990. Exponentially weighted moving
performance of the charts, giving a large number of false average control schemes: properties and enhancements.
alarms when s was estimated via the average moving range Technometrics 32, 1–12.
MR and covering real alarms when s was estimated via the Montgomery, D.C., 2001. Introduction to Statistical Quality Con-
sample standard deviation s. trol, fourth ed. Wiley, New York.
The more sophisticated Alwan and Roberts (1988) approach Montgomery, D.C., Mastrangelo, C.M., 1991. Some statistical
eliminated the autocorrelation of the data. The SCC, which is process control methods for autocorrelated data. J. Qual.
Technol. 23, 179–193.
a Shewhart chart of individual observations applied to the
Nau, R.F., 2005. Three types of forecasts: estimation period,
residuals which were obtained after fitting the process with validation period, and ‘‘the future’’. Statistical Forecasting
an ARIMA model, gave 3 out-of-control clues on PDMf and (MBA Course), h http://www.duke.edu/rnau/three.htmi.
none on NDMf and ASPf. For fast localization of small shifts, Page, E.S., 1954. Continuous inspection Schemes. Biometrika 41,
EWMA and CUSUM charts applied to the residuals are more 100–115.
suitable. Roberts, S.W., 1959. Control chart tests based on geometric
moving averages. Technometrics 1, 239–250.
The appropriate models to explain the toxicity procedure
Runger, G.C., Willemain, T.R., Prabhu, S., 1995. Average run
were ARIMAð0; 1; 1Þ in any case of the three examined treated-
lengths for CUSUM control charts applied to residuals.
water tanks. The time-series based method captures the Commun. Stat. Theory Methods 24, 273–283.
dynamic structure of the data and gives reasonable control Ryan, T.P., 1989. Statistical Methods for Quality Improvement.
charts that can be useful tools for the improvement of the Wiley, New York.
water quality. Schwartz, G., 1978. Estimating the dimension of a model. Ann.
Stat. 5 (2), 461–464.
Strategic Diagnostics Microtox 1010 (2003). Method for Measuring
Acknowledgment the Acute Toxicity of Wastewater and Receiving Water with
Vibrio fischeri (NRRL B-11177) Microtoxs Test System. Strategic
Diagnostics, Inc.
We would like to thank Dr. N. Thanasoulias for the useful
Sullivan, J.H., Woodall, W.H., 1996a. A control chart for prelimin-
discussion and exchange of ideas. ary analysis of individual observations. J. Qual. Technol. 28,
265–278.
R E F E R E N C E S Van Brackle, L.N., Reynolds, M.R., 1997. EWMA and CUSUM
control charts in the presence of autocorrelation. Commun.
Stat. Simulation Comput. 26, 979–1008.
Vasilopoulos, A.V., Stamboulis, A.P., 1978. Modification of control
Akaike, H., 1974. A new look at the statistical model identification.
chart limits in the presence of data correlation. J. Qual.
IEEE Trans. Autom. Control AC-19, 716–723.
Technol. 10, 20–30.
Alwan, L.C., 1992. Effects of autocorrelation on control chart
Wardell, D.G., Moskowitz, H., Plante, R.D., 1992. Control charts
performance. Commun. Stat. Theory Methods 21, 1025–1049.
in the presence of data correlation. Manage. Sci. 38,
Alwan, L.C., Roberts, H.V., 1988. Time-series modeling for
1084–1105.
statistical process control. J. Bus. Econ. Stat. 6, 87–95.
Wardell, D.G., Moskowitz, H., Plante, R.D., 1994. Run-length
Box, G.E.P., Jenkins, G.M., Reinsel, G.C., 1994. Time Series Analysis,
distributions of special-cause control charts for correlated
Forecasting and Control. Prentice-Hall, Inc., Englewood Cliffs, NJ.
processes. Technometrics 36, 3–27.
Champ, C.W., Woodall, W.H., 1987. Exact results for Shewhart
Zhang, N.F., 1998. A statistical control chart for stationary process
control charts with supplementary runs rules. Technometrics
data. Technometrics 40, 24–38.
29, 393–399.

You might also like