Hindawi Publishing Corporation

Journal of Quality and Reliability Engineering

Volume 2013, Article ID 542305, 14 pages
http://dx.doi.org/10.1155/2013/542305

Research Article
Robust Control Charts for Monitoring Process Mean of Phase-I
Multivariate Individual Observations

Asokan Mulayath Variyath and Jayasankar Vattathoor

Department of Mathematics and Statistics, Memorial University of Newfoundland, St. John’s, NL, Canada A1C 5S7

Correspondence should be addressed to Asokan Mulayath Variyath; variyath@mun.ca

Received 22 November 2012; Revised 22 March 2013; Accepted 2 April 2013

Academic Editor: Adiel Teixeira de Almeida

Copyright © 2013 A. M. Variyath and J. Vattathoor. This is an open access article distributed under the Creative Commons
Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is
properly cited.

Hoteling’s 𝑇2 control charts are widely used in industries to monitor multivariate processes. The classical estimators, sample
mean, and the sample covariance used in 𝑇2 control charts are highly sensitive to the outliers in the data. In Phase-I monitoring,
control limits are arrived at using historical data after identifying and removing the multivariate outliers. We propose Hoteling’s
𝑇2 control charts with high-breakdown robust estimators based on the reweighted minimum covariance determinant (RMCD)
and the reweighted minimum volume ellipsoid (RMVE) to monitor multivariate observations in Phase-I data. We assessed the
performance of these robust control charts based on a large number of Monte Carlo simulations by considering different data
scenarios and found that the proposed control charts have better performance compared to existing methods.

1. Introduction control chart for individual observations in Phase-I, for each

observation x𝑗 we calculate
Control charts are widely used in industries to moni-
tor/control processes. Generally, the construction of a control 󸀠
chart is carried out in two phases. The Phase-I data is 𝑇2 (x𝑗 ) = (x𝑗 − x) S−1
1 (x𝑗 − x) ,
(1)
analyzed to determine whether the data indicates a stable (or
in-control) process and to estimate the process parameters where x𝑗 = (𝑥𝑗1 , 𝑥𝑗2 , . . . , 𝑥𝑗𝑝 )󸀠 is the 𝑗th 𝑝-variate observation,
and thereby the construction of control limits. The Phase- (𝑗 = 1, 2, . . . , 𝑚) and the sample mean x, sample covariance
II data analysis consists of monitoring future observations matrix S1 are based on 𝑚 Phase-I observations. In Phase-
based on control limits derived from the Phase-I estimates I monitoring, the 𝑇2 (x𝑗 ) values are compared with the 𝑇2
to determine whether the process continues to be in control control limit derived by assuming that the x𝑗 ’s are multivariate
or not. But trends, step changes, outliers, and other unusual normal so that the 𝑇2 control limits are based on the beta
data points in the Phase-I data can have an adverse effect on distribution with the parameters 𝑝/2 and (𝑚 − 𝑝 − 1)/2.
the estimation of parameters and the resulting control limits. However, the classical estimators, sample mean, and sample
That is, any deviation from the main assumption (in our case, covariance are highly sensitive to the outliers, and hence
identically and independently distributed from normal distri- robust estimation methods are preferred as they have the
bution) may lead to an out-of-control situation. Therefore, it advantage of not being unduly influenced by the outliers.
becomes very important to identify and eliminate these data The use of robust estimation methods is well suited to
points prior to calculating the control limits. In this paper, all detect multivariate outliers because of their high breakdown
these unusual data points are referred to as “outliers.” points which ensure that the control limits are reasonably
Multivariate quality characteristics are often correlated, accurate. Sullivan and Woodall [3] proposed a 𝑇2 chart with
and to monitor the multivariate process mean Hoteling’s 𝑇2 an estimate of the covariance matrix based on the successive
control chart [1, 2] is widely used. To implement Hoteling’s 𝑇2 differences of observations and showed that it is effective in
2 Journal of Quality and Reliability Engineering

detecting process shift. However, these charts are not effective above three properties of a good robust estimator, it should
in detecting multiple multivariate outliers because of their be possible to calculate the estimator in a reasonable amount
low breakdown point. of time to make it computationally efficient.
Vargas [4] introduced two robust 𝑇2 control charts based It is difficult to get an affine equivariant and robust
on robust estimators of location and scatter, namely, the estimator as affine equivariance and high breakdown will
minimum covariance determinant (MCD) and minimum not come simultaneously. Lopuhaä and Rousseeuw [8] and
volume ellipsoid (MVE) for identifying the outliers in Phase-I Donoho and Gasko [9] showed that the finite sample
multivariate individual observations. Jensen et al. [5] showed breakdown point of (𝑚 − 𝑝 + 1)/(2𝑚 − 𝑝 + 1) is difficult for
2 2
that 𝑇MCD and 𝑇MVE control charts have better performance an affine equivariant estimator. The largest attainable finite
when outliers are present in the Phase-I data. Chenouri et sample breakdown point of any affine equivariant estimator
al. [6] used reweighted MCD estimators for monitoring the of the location and scatter matrix with a sample size 𝑚 and
Phase-II data, without constructing Phase-I control charts. dimension 𝑝 is (𝑚 − 𝑝 + 1)/2𝑚 [10]. Therefore relaxing the
However, in many situations Phase-I control charts are affine equivariance condition of the estimators to invariance
necessary to assess the performance of the process and under the orthogonal transformation makes it easy to find an
also to identify the outliers. We propose 𝑇2 control charts estimator with the highest breakdown point.
based on the reweighted minimum covariance determinant The classical estimators, sample mean vector, and covari-
(RMCD)/reweighted minimum volume ellipsoid (RMVE) ance matrix of location and scatter parameters are affine
2 2
(𝑇RMCD /𝑇RMVE ) for monitoring Phase-I multivariate indi- equivariant but their sample breakdown point is as low as
vidual observations. RMCD/RMVE estimators are statisti- 1/𝑚. The MCD and MVE estimators have the highest possible
cally more efficient than MCD/MVE estimators and have a finite sample breakdown point (𝑚 − 𝑝 + 1)/2𝑚. However,
manageable asymptotic distribution. We empirically arrive at both of these estimators have very low asymptotic efficiency
2 2 under normality. But the reweighted versions of MCD and
Phase-I control limits for the 𝑇RMCD /𝑇RMVE control chart for
some specific sample sizes and fitted a nonlinear model to MVE estimators have better efficiency without compromising
determine control limits for any sample size for dimensions 2 on the breakdown point and rate of convergence compared
2
to 10. Our simulation studies show that 𝑇RMCD 2
/𝑇RMVE control to MCD and MVE. In the next two subsections, we discuss
2 2 in detail about the MCD and MVE estimators and their
charts are performing well compared to 𝑇MCD /𝑇MVE control
reweighted versions.
charts for monitoring the Phase-I data.
The organization of the remaining part of the paper
is as follows. In Section 2, we discuss the properties of a 2.1. MCD and RMCD Estimators. The MCD estimators of
good robust estimator and we briefly explain the MCD/MVE location and scatter parameters of the distribution are deter-
estimators and their reweighted versions. The proposed mined by a two-step procedure. In step 1, all possible subsets
2 2
𝑇RMCD /𝑇RMVE control charts are given in Section 3 along of observations of size ℎ = (𝑚 ∗ 𝛾), where 0.5 ≤ 𝛾 ≤ 1
with the control limits arrived at based on Monte Carlo sim- are obtained. In step 2, the subset whose covariance matrix
ulations. We assess the performance of the proposed control has the smallest possible determinant is selected. The MCD
charts in Section 4, and the implementation of the proposed location estimator xMCD is defined as the average of this
methods is illustrated in a case example in Section 5. Our selected subset of ℎ points, and the MCD scatter estimator
𝑚
conclusions are given in Section 6. is given by SMCD = 𝑎𝛾,𝑝 ∗ 𝑏𝛾,𝑝 ∗ CMCD , where CMCD is the
covariance matrix of the selected subset, the constant 𝑎𝛾,𝑝 is
𝑚
the multiplication factor for consistency [11], and 𝑏𝛾,𝑝 is the
2. Robust Estimators
finite sample correction factor [12]. Here (1 − 𝛾) represents
The affine equivariance property of the estimator is important the breakdown point of the MCD estimators. The MCD
because it makes the analysis independent of the measure- estimator has its highest possible finite sample breakdown
ment scale of the variables as well as the transformations point when ℎ = (𝑚 + 𝑝 + 1)/2 and has an 𝑚−1/2 rate of
or rotations of the data. The breakdown point concept convergence but has a very low asymptotic efficiency under
introduced by Donoho and Huber [7] is often used to normality. Computing the exact MCD estimators (xMCD ,
assess the robustness. The breakdown point is the smallest SMCD ) is computationally expensive or even impossible for
proportion of the observations which can render an estimator large sample sizes in high dimensions [13], and hence various
meaningless. A higher breakdown point implies a more algorithms have been suggested for approximating the MCD.
robust estimator, and the highest attainable breakdown point Hawkins and Olive [14] and Rousseeuw and van Driessen [15]
is 1/2 in the case of median in the univariate case. For more independently proposed a fast algorithm for approximating
details on affine equivariance and breakdown points one may MCD. The FAST-MCD algorithm of Rousseeuw and van
refer to Chenouri et al. [6] or Jensen et al. [5]. Driessen finds the exact MCD for small datasets and gives a
An estimator is said to be relatively efficient compared to good approximation for larger datasets, which is available in
any other estimator if the mean square error for the estimator the standard statistical software SPLUS, R, SAS, and Matlab.
is the least for at least some values of the parameter compared MCD estimators are highly robust, carry equivariance
to others. A robust estimator is considered to be good if it properties, and can be calculated in a reasonable time using
carries the property of affine equivariance along with a higher the FAST-MCD algorithm; however, they are statistically not
breakdown point and greater efficiency. In addition to the efficient. The reweighted procedure will help to carry both
Journal of Quality and Reliability Engineering 3

2 2
Table 1: Estimates of the model parameters 𝑎1(𝑝,𝛼) , 𝑎2(𝑝,𝛼) , 𝑎3(𝑝,𝛼) for 𝑇RMCD /𝑇RMVE control charts.

𝛼 = 0.05 𝛼 = 0.01 𝛼 = 0.001

𝑝
𝑎̂1 𝑎̂2 𝑎̂3 𝑎̂1 𝑎̂2 𝑎̂3 𝑎̂1 𝑎̂2 𝑎̂3
2
𝑇RMCD
2 17.223 41102 2.647 21.134 38170 2.329 27.051 192909 2.508
3 20.134 35844 2.209 24.287 128924 2.344 31.350 1144947 2.718
4 23.152 269357 2.548 28.181 1272773 2.773 35.575 5989325 2.973
5 24.685 467949 2.524 28.437 1417059 2.632 31.013 2666196 2.593
6 26.962 1762051 2.746 29.654 3061216 2.711 31.662 5414248 2.669
7 24.892 1099128 2.493 22.882 1585224 2.416 19.058 3465278 2.444
8 27.236 2908821 2.667 27.245 4922576 2.644 28.326 12134778 2.710
9 23.974 2447649 2.534 21.420 4726835 2.554 18.772 14096595 2.676
10 31.894 12572909 2.914 37.085 34375654 3.033 56.573 172176786 3.301
2
𝑇RMVE
2 17.442 29553 2.494 21.365 31571 2.244 27.594 148747 2.434
3 20.286 22497 2.066 24.387 59096 2.13 31.326 338665 2.402
4 23.095 108855 2.286 27.549 291064 2.372 35.109 1255429 2.576
5 24.796 238966 2.334 28.302 508097 2.367 32.008 1063783 2.377
6 27.585 1041090 2.606 31.126 1882888 2.601 37.136 4714353 2.671
7 28.151 1541634 2.598 30.936 3183762 2.635 39.357 12199414 2.827
8 34.917 14798692 3.127 45.767 75616029 3.419 70.875 840512379 3.904
9 39.191 59094377 3.415 50.271 275604839 3.679 72.768 1960966919 4.039
10 50.733 950607720 4.099 68.154 4696452032 4.379 110.587 56398461817 4.881

robustness and efficiency. That is, first a highly robust but where the robust distance RD(x𝑗 ) =
perhaps an inefficient estimator is computed, which is used as
√(x𝑗 − xMCD )󸀠 S−1
MCD (x𝑗
− xMCD ) and 𝑞𝛼 is (1 − 𝛼)100%
a starting point to find a local solution for detecting outliers
and computing the sample mean and covariance of the quantile of the chi-square distribution with 𝑝 degrees of
cleaned data set as in Rousseeuw and van Zomeren [16]. This freedom.
consists of discarding those observations whose Mahalanobis This reweighting technique improves the efficiency of
distances exceed a certain fixed threshold value. MCD is the the initial MCD estimator while retaining (most of) its
current best choice for the initial estimator of a two-step robustness. Hence the RMCD estimator inherits the affine
procedure as it contains the robustness, equivariance, and equivariance, robustness, and asymptotic normality proper-
computational efficiency properties along with its 𝑚−1/2 rate ties of the MCD estimators with an improved efficiency.
of convergence. Hence RMCD estimators are the weighted
mean vector 2.2. MVE and RMVE Estimators. Determining the MVE
estimators of location and scatter parameters of the distri-
(∑𝑚
𝑗=1 𝑤𝑗 x𝑗 ) bution is almost in line with that of the MCD estimator. As
xRMCD = , (2) in the case of MCD, all the possible subsets of data points
(∑𝑚
𝑗=1 𝑤𝑗 ) with size ℎ = (𝑚 ∗ 𝛾) (where 0.5 ≤ 𝛾 ≤ 1) is obtained
first. Then the ellipsoid of minimum volume that covers
and the weighted covariance matrix the subsets are obtained to determine the MVE estimators.
𝑚,𝑝 The MVE location estimator is the geometrical center of
SRMCD = 𝑐𝛼,𝑝 ∗ 𝑑𝛾,𝛼 the ellipsoid, and the MVE scatter estimator is the matrix
󸀠 defining the ellipsoid itself, multiplied by an appropriate
∑𝑚
𝑗=1 𝑤𝑗 (x𝑗 − xRMCD ) (x𝑗 − xRMCD )
(3)
constant to ensure consistency [13, 16]. Thus MVE estimator
∗ ,
∑𝑚
𝑗=1 𝑤𝑗
does not correspond to the sample mean vector and the
sample covariance matrix as in the case of the MCD estimator.
where 𝑐𝛼,𝑝 is the multiplication factors for consistency [11], Here (1 − 𝛾) represents the breakdown point of the MVE
𝑚,𝑝 estimators, as in the case of MCD, and it has the highest
𝑑𝛾,𝛼 is the finite sample correction factor [12], and the weights
possible finite sample breakdown point when ℎ = (𝑚 +
𝑤𝑗 are defined as
𝑝 + 1)/2𝑚 [8, 17]. The MVE estimator has an 𝑚−1/3 rate of
convergence and a nonnormal asymptotic distribution [17].
1 if RD(x𝑗 ) ≤ √𝑞𝛼 , As in the case for MCD estimators, MVE estimators are
𝑤𝑗 = { (4)
0 otherwise, also not efficient. Hence, a reweighted version similar to that
4 Journal of Quality and Reliability Engineering

RMCD, 𝑝 = 2, 𝛼 = 0.05 RMCD, 𝑝 = 2, 𝛼 = 0.01

35 35

30 30

Limits
Limits

𝑅2 = 0.9904
𝑅2 = 0.9806
25 25

20 20

50 100 150 200 50 100 150 200

Sample size Sample size
(a) (b)
RMVE, 𝑝 = 2, 𝛼 = 0.05 RMVE, 𝑝 = 2, 𝛼 = 0.01

35 35

30 30
Limits
Limits

𝑅2 = 0.9326 𝑅2 = 0.962
25 25

20 20

50 100 150 200 50 100 150 200

Sample size Sample size
Actual Actual
Fitted Fitted
(c) (d)

2 2
Figure 1: Scatter plot of 𝑇RMCD /𝑇RMVE control limits and the fitted curve for 𝑝 = 2.

for MCD has been proposed by Rousseeuw and van Zomeren for monitoring the process mean of Phase-I multivariate
[16]. Note that it has been shown more recently that the individual observations. RMCD/RMVE estimators inherit
RMVE estimators do not improve on the convergence rate the nice properties of initial MCD estimators such as affine
(and thus the 0% asymptotic efficiency) of the initial MVE equivariance, robustness, and asymptotic normality while
estimator [8, 12]. Therefore, as an alternative, a one-step M- achieving a higher efficiency. We now define a robust 𝑇2
estimator can be calculated with the MVE estimators as the control chart with RMCD and RMVE estimators for 𝑖th
initial solution [13, 18] which results in an estimator with multivariate observation as
the standard 𝑚−1/2 convergence rate to a normal asymptotic
2 󸀠
distribution. For more details on MCD/MVE estimators 𝑇RMCD (x𝑖 ) = (x𝑖 − xRMCD ) S−1
RMCD (x𝑖 − xRMCD ) ,
one may refer to Chenouri et al. [6] or Jensen et al. [5]. (5)
2 󸀠
The algorithm to determine the MVE/RMVE estimators 𝑇RMVE (x𝑖 ) = (x𝑖 − xRMVE ) S−1
RMVE (x𝑖 − xRMVE ) ,
is available in the statistical software SPLUS, R, SAS, and
Matlab. where xRMCD , xRMVE are the mean vectors and SRMCD ,
3. Robust Control Charts SRMVE are the dispersion matrices under the RMCD/RMVE
methods based on 𝑚 multivariate observations.
We propose to use 𝑇2 charts with robust estimators of 2
The exact distribution of 𝑇RMCD 2
/𝑇RMVE estimators not
location and dispersion parameters based on RMCD/RMVE available, hence the control limits for Phase-I data are
Journal of Quality and Reliability Engineering 5

RMCD, 𝑝 = 6, 𝛼 = 0.05 RMCD, 𝑝 = 6, 𝛼 = 0.01

350 350

250 250
Limits

Limits
𝑅2 = 0.9816
𝑅2 = 0.9815
150 150

50 50

50 100 150 200 50 100 150 200

Sample size Sample size
(a) (b)
RMVE, 𝑝 = 6, 𝛼 = 0.05 RMVE, 𝑝 = 6, 𝛼 = 0.01
350 350

250 250
Limits
Limits

𝑅2 = 0.9878 𝑅2 = 0.984
150 150

50 50

50 100 150 200 50 100 150 200

Sample size Sample size
Actual Actual
Fitted Fitted
(c) (d)

2 2
Figure 2: Scatter plot of 𝑇RMCD /𝑇RMVE control limits and the fitted curve for 𝑝 = 6.

2 2
obtained empirically. In the next subsection we apply Monte of 𝛾 = 0.50, 𝑇RMCD /𝑇RMVE statistics for each observation
Carlo simulation to estimate quantiles of the distribution of in the data set were calculated using (5), and the maximum
2 2
𝑇RMCD and 𝑇RMVE for several combinations of sample sizes value attained for each data set of size 𝑚 was recorded. The
2 2
and dimensions. For each dimension, we further introduce a empirical distribution of maximum of 𝑇RMCD and 𝑇RMVE was
method to fit a smooth nonlinear model to arrive, the control inverted to determine the (1 − 𝛼)100% quantiles. We used
limits for any given sample size. the R-function “CovMcd()” in the “rrcov” package written by
Torodov [19] to ascertain the RMCD/RMVE estimators.
We have constructed the empirical distribution of
3.1. Computation of Control Limits. We performed a large
𝑇𝑅2 MCD/𝑇𝑅2 MVE as above for 𝑚 = [30(1)50, 55(5)100,
number of Monte Carlo simulations to obtain the control
110(10)200], 𝑝 = (2, 3, . . . 10) when 𝛾 = 0.50 and arrived
limits. We generated 𝑛 = 200, 000 samples of size 𝑚 from
at the control limits for 𝛼 = (0.05, 0.01, and 0.001). The
a standard multivariate normal distribution MVN(0, 𝐼𝑝 )
2
scatter plots of the quantiles and sample sizes for different
with dimension 𝑝. Due to the invariance of the 𝑇RMCD and dimensions suggest a family of nonlinear models of the form
2
𝑇RMVE statistics, these limits will be applicable for any values
of 𝜇 and Σ. Using the reweighted MCD/MVE estimators 𝑎2,(𝑝,𝛼,𝛾)
xRMCD , SRMCD , xRMVE , and SRMVE with a breakdown value 𝑓𝑝,𝛼,𝛾,𝑚 = 𝑎1,(𝑝,𝛼,𝛾) + , (6)
𝑚𝑎3,(𝑝,𝛼,𝛾)
6 Journal of Quality and Reliability Engineering

RMCD, 𝑝 = 10, 𝛼 = 0.05 RMCD, 𝑝 = 10, 𝛼 = 0.01

1500 1500

Limits
1000
Limits

1000 𝑅2 = 0.978

𝑅2 = 0.9805

500 500

0 0
50 100 150 200 50 100 150 200
Sample size Sample size
(a) (b)
RMVE, 𝑝 = 10, 𝛼 = 0.05 RMVE, 𝑝 = 10, 𝛼 = 0.01

1500 1500

𝑅2 = 0.9808
Limits

1000
Limits

1000
𝑅2 = 0.9785

500 500

0 0
50 100 150 200
50 100 150 200
Sample size
Sample size
Actual Actual
Fitted Fitted
(c) (d)

2 2
Figure 3: Scatter plot of 𝑇RMCD /𝑇RMVE control limits and the fitted curve for 𝑝 = 10.

2 2
where 𝑎1(𝑝,𝛼,𝛾) , 𝑎2(𝑝,𝛼,𝛾) , and 𝑎3(𝑝,𝛼,𝛾) are the model parame- control limits for 𝑇RMCD and 𝑇RMVE can be found using (6)
ters. For clarity, the scatter plot of the actual and the fitted for any sample size.
2 2 For the implementation of a robust control chart, first
values of the quantiles of 𝑇RMCD and 𝑇RMVE for 𝑝 = 2, 6, and
10 are given in Figures 1, 2, and 3; other plots are omitted to collect a sample of 𝑚 multivariate individual observations
save space. with dimension 𝑝. Compute robust estimates of mean and
From Figures 1, 2, and 3, we can see that the nonlinear fit covariance matrix using R or any other software with 𝛾 =
2 2
is very well supported by the high 𝑅2 values, which help us to 0.50, and determine 𝑇RMCD /𝑇RMVE . Outliers can be deter-
2 2 2 2
determine the 𝑇RMCD and 𝑇RMVE control limits for any given mined by comparing the 𝑇RMCD /𝑇RMVE values with control
sample size. The least square estimates of the parameters limits obtained using (6) for specific values of 𝛼, 𝑚, 𝑝, and the
𝑎1(𝑝,𝛼) , 𝑎2(𝑝,𝛼) , and 𝑎3(𝑝,𝛼) when 𝛾 = 0.50 for dimensions 𝑝 = constants given in Table 1. The outlier free data can be used
2 2
(2, 3, . . . , 10) and 𝛼 = (0.05, 0.01 and 0.001) for 𝑇RMCD /𝑇RMCD to construct the standard 𝑇2 control chart for monitoring the
control charts are given in Table 1. Using these estimates, the Phase-II observations.
Journal of Quality and Reliability Engineering 7

𝛼 = 0.05, 𝜋 = 10% 𝛼 = 0.01, 𝜋 = 10%

1 1

0.8 0.8

0.6 0.6

0.4 0.4

0.2 0.2

0 0
0 5 10 15 20 25 30 0 5 10 15 20 25 30
Noncentrality parameter Noncentrality parameter
(a) (b)
𝛼 = 0.05, 𝜋 = 20% 𝛼 = 0.01, 𝜋 = 20%
1 1

0.8 0.8

0.6 0.6

0.4 0.4

0.2 0.2

0 0
0 5 10 15 20 25 30 0 5 10 15 20 25 30
Noncentrality parameter Noncentrality parameter
Std MVE Std MVE
MCD RMVE MCD RMVE
RMCD RMCD
(c) (d)

Figure 4: Probability of signal for 𝑇2 control chart with different estimation methods for 𝑝 = 2, 𝑚 = 50.

4. Performance Analysis outliers. We set 𝜋 = 0.10 and 0.20 to ensure that the sample
We assess the performance of the proposed charts when contains few outliers. Without loss of generality, we consider
outliers are present due to the shift in the process mean. In the in-control distribution as 𝑁(0, 𝐼𝑝 ). The out-of-control
their study, Jensen et al. [5] concluded that the 𝑇MCD2 2
/𝑇MVE distribution is a multivariate normal with a small shift in the
control charts had better performance in terms of probability mean vector with same covariance matrix. The amount of
of signal. Hence, we compare the performance of our pro- mean shift is defined through a noncentrality parameter (𝛿),
2 2 which is given by
posed method with 𝑇MCD /𝑇MVE charts as well as the standard
2
𝑇 charts based on classical estimators. Our study compares 󸀠
𝛿 = (𝜇1 − 𝜇) Σ−1 (𝜇1 − 𝜇) , (7)
more combinations of dimension 𝑝, sample size 𝑚, and 𝜋.
For a particular combination of 𝑝, 𝑚, and 𝜋, a number of where (𝜇1 − 𝜇) is the shift in the mean vector. The larger
datasets are generated. Out of the 𝑚 observations generated, the value of 𝛿 is, the more extreme the outliers are. The
2 2
𝑚 ∗ 𝜋 of them are random data points generated from the proportion of datasets that had at least one 𝑇RMCD or 𝑇RMVE
out-of-control distribution, and the remaining 𝑚 ∗ (1 − 𝜋) statistic greater than the control limit was calculated, and this
observations are generated from the in-control distribution proportion becomes the estimated probability of signal. We
so that the sample of 𝑚 data points may contain some compared the performance of these charts with standard 𝑇2
8 Journal of Quality and Reliability Engineering

𝛼 = 0.05, 𝜋 = 10% 𝛼 = 0.01, 𝜋 = 10%

1 1

0.8 0.8

0.6 0.6

0.4 0.4

0.2 0.2

0 0
0 5 10 15 20 25 30 0 5 10 15 20 25 30
Noncentrality parameter Noncentrality parameter
(a) (b)
𝛼 = 0.05, 𝜋 = 20% 𝛼 = 0.01, 𝜋 = 20%
1 1

0.8 0.8

0.6 0.6

0.4 0.4

0.2 0.2

0 0
0 5 10 15 20 25 30 0 5 10 15 20 25 30
Noncentrality parameter Noncentrality parameter
Std MVE Std MVE
MCD RMVE MCD RMVE
RMCD RMCD
(c) (d)

Figure 5: Probability of signal for 𝑇2 control chart with different estimation methods for 𝑝 = 2, 𝑚 = 100.

2 2
charts, 𝑇MCD , and 𝑇MVE charts. The standard 𝑇2 chart was values of 𝛿. We have presented only a selected set of plots to
included in our performance study as a reference because of save space. The plots of probability of signal for 𝛼 = 0.05 and
its common usage. 0.01, 𝑝 = 2 and 6, and 𝑚 = 50 and 100 are given in Figures 4, 5,
The probability of a signal for different values of 𝛿 = 6, and 7 for easier understanding. For dimension 𝑝 = 10, we
(0, 5, 10, 15, 20, 25, 30) and for some of the values of 𝑚 = used 𝑚 = 100 and 150, and the plots of probability of signal
(30, 50, 100, 150), 𝑝 = (2, 6, 10) and 𝜋 = (10%, 20%) was are given in Figures 8 and 9.
considered in our study. Fifty thousand datasets of size 𝑚 From Figures 4–9, we can see that when the value of
were generated for each combination of 𝑝, 𝜋, and 𝛿, and the the noncentrality parameter is zero or close to zero, the
probability of signal was estimated for 𝛼 = 0.05, 0.01, and probability of signal is close to 𝛼 which is expected for an in-
0.001. We considered various combinations of 𝜇1 , 𝜇2 , and 𝜌 control process. As the value of the noncentrality parameter
which determine 𝛿 as per (7) and found that the probability of increases the probability of signals also increases. Using
signal is the same irrespective of the combination of 𝜇1 , 𝜇2 and this criterion, we select the best method for identifying the
𝜌. Hence we have considered 𝜇1 = 𝜇2 and 𝜌 = 0 for various outliers. If the probability of signal does not increase for
Journal of Quality and Reliability Engineering 9

𝛼 = 0.05, 𝜋 = 10% 𝛼 = 0.01, 𝜋 = 10%

0.4 0.4

0.3 0.3

0.2 0.2

0.1 0.1

0 0
0 5 10 15 20 25 30 0 5 10 15 20 25 30
Noncentrality parameter Noncentrality parameter
(a) (b)
𝛼 = 0.05, 𝜋 = 20% 𝛼 = 0.01, 𝜋 = 20%
0.4 0.4

0.3 0.3

0.2 0.2

0.1 0.1

0 0
0 5 10 15 20 25 30 0 5 10 15 20 25 30
Noncentrality parameter Noncentrality parameter
Std MVE Std MVE
MCD RMVE MCD RMVE
RMCD RMCD
(c) (d)

Figure 6: Probability of signal for 𝑇2 control chart with different estimation methods for 𝑝 = 6, 𝑚 = 50.

2
increase in noncentrality parameter, then it is clear that the possesses little ability to detect the outliers and the 𝑇MVE , and
2 2 2
estimator has broken down and is not capable of detecting 𝑇MVE stands below the 𝑇RMCD /𝑇RMVE charts throughout all
the outliers. the values of 𝛿.
A careful examination of these plots of probability of As 𝑝 increases for a fixed value of 𝑚, the breakdown
signals corresponding to various values of 𝑝, 𝑚, and 𝜋 points of RMCD and RMVE get smaller as the breakdown
2
indicates that for small values of 𝑝 and 𝑚, 𝑇RMVE performs value is given by (𝑚 − 𝑝 + 1)/2𝑚. This suggests that the
2
well. As 𝑚 and 𝑝 increase, 𝑇RMCD chart is superior. For larger 𝑝 is, the larger 𝑚 will need to be in order to maintain
2
example, from Figures 4 and 5 we see that 𝑇RMVE has slight the breakdown point, which is very well demonstrated in
2 2 2
advantage over 𝑇RMCD . But compared to 𝑇MCD /𝑇MVE charts, Figures 8 and 9. In general, there was always one estimator,
2 2
𝑇RMCD /𝑇RMVE charts are performing well which is evident RMCD or RMVE, that was found to be superior across all
from all the plots presented here. When 𝑝 is large (see Figures the values of the noncentrality parameter as long as the
2 2
8 and 9), the 𝑇RMCD has clear advantage compared to 𝑇RMVE . proportion of outliers was not so big as to cause the estimators
2
From these figures, we see that standard 𝑇 control chart to break down. This greatly simplifies the conclusions that
10 Journal of Quality and Reliability Engineering

𝛼 = 0.05, 𝜋 = 10% 𝛼 = 0.01, 𝜋 = 10%

0.8 0.8

0.6 0.6

0.4 0.4

0.2 0.2

0 0
0 5 10 15 20 25 30 0 5 10 15 20 25 30
Noncentrality parameter Noncentrality parameter
(a) (b)
𝛼 = 0.05, 𝜋 = 20% 𝛼 = 0.01, 𝜋 = 20%
0.8 0.8

0.6 0.6

0.4 0.4

0.2 0.2

0 0

0 5 10 15 20 25 30 0 5 10 15 20 25 30
Noncentrality parameter Noncentrality parameter
Std MVE Std MVE
MCD RMVE MCD RMVE
RMCD RMCD
(c) (d)

Figure 7: Probability of signal for for 𝑇2 control chart with different estimation methods for 𝑝 = 6, 𝑚 = 100.

can be made about when the RMCD or RMVE estimators are for Phase-I applications where the number of outliers is
2
preferred to the MCD and MVE estimators. unknown, 𝑇RMVE should be used only for smaller sample
2 2
Nevertheless, 𝑇RMCD and 𝑇RMCD charts are preferred for 2
sizes, and it is also computationally feasible. 𝑇RMCD should be
the various combinations of 𝑚, 𝑝, and 𝜋, and some broad used for larger sample sizes or when it is believed that there
recommendations can be made on the selection among these is a large number of outliers. When the dimension is large,
2
two charts. When 𝑚 < 100, the 𝑇RMVE will be the best for larger sample sizes are needed to ensure that the estimator
2
small dimension. When 𝑚 ≥ 100, the 𝑇RMCD is preferred. does not break down and lose its ability to detect outliers.
As 𝑝 increases, then the percentage of outliers that can be 2
Hence for larger dimension cases, 𝑇RMCD is preferred with
2
detected by the 𝑇RMVE chart decreases. It is true for both large sample sizes. For very small samples (𝑚 < 30), one may
the charts that when 𝑝 is higher, the number of outliers opt for higher values of 𝛾, for which control limits need to be
that can be detected decreases for smaller sample sizes. Thus developed.
Journal of Quality and Reliability Engineering 11

𝛼 = 0.05, 𝜋 = 10% 𝛼 = 0.01, 𝜋 = 10%

0.8 0.8

0.6 0.6

0.4 0.4

0.2 0.2

0 0
0 5 10 15 20 25 30 0 5 10 15 20 25 30
Noncentrality parameter Noncentrality parameter
(a) (b)
𝛼 = 0.05, 𝜋 = 20% 𝛼 = 0.01, 𝜋 = 20%
0.8 0.8

0.6 0.6

0.4 0.4

0.2 0.2

0 0
0 5 10 15 20 25 30 0 5 10 15 20 25 30
Noncentrality parameter Noncentrality parameter
Std MVE Std MVE
MCD RMVE MCD RMVE
RMCD RMCD
(c) (d)

Figure 8: Probability of signal for 𝑇2 control chart with different estimation methods for 𝑝 = 10, 𝑚 = 100.

5. Case Example
To illustrate the applicability of the proposed control chart 5.7125
method, we discuss a real case example taken from an 𝑋RMCD = (6.2643) ,
electronic industry. The data gives 105 measurements of 3 6.0837
axial components of acceleration measured by accelerometer 2.8549 0.1901 −0.2926
on a e-compass unit fixed on the objects. The mean vector and 𝑆RMCD = ( 0.1901 1.1175 0.4433 ) ,
covariance matrix under the classical, RMCD, and RMVE −0.2926 0.4433 2.3115
methods of the sample data considered are given by
5.7790
6.3143 𝑋RMVE = (6.1379) ,
𝑋 = (5.7339) , 5.9894
5.7527
2.7837 −0.2642 −0.5745
3.4022 −1.1524 −1.0746 𝑆RMVE = (−0.2642 1.2778 0.6080 ) .
𝑆 = (−1.1524 1.9249 1.1209 ) , −0.5745 0.6080 2.2909
−1.0746 1.1209 2.2004 (8)
12 Journal of Quality and Reliability Engineering

𝛼 = 0.05, 𝜋 = 10% 𝛼 = 0.01, 𝜋 = 10%

1 1

0.8 0.8

0.6 0.6

0.4 0.4

0.2 0.2

0 0
0 5 10 15 20 25 30 0 5 10 15 20 25 30
Noncentrality parameter Noncentrality parameter
(a) (b)
𝛼 = 0.05, 𝜋 = 20% 𝛼 = 0.01, 𝜋 = 20%
1 1

0.8 0.8

0.6 0.6

0.4
0.4

0.2
0.2

0
0
0 5 10 15 20 25 30 0 5 10 15 20 25 30
Noncentrality parameter Noncentrality parameter
Std MVE Std MVE
MCD RMVE MCD RMVE
RMCD RMCD
(c) (d)

Figure 9: Probability of signal for 𝑇2 control chart with different estimation methods for 𝑝 = 10, 𝑚 = 150.

A simple comparison of these estimators indicates that there 6. Conclusions

are outliers in the Phase-I data. The plots of 𝑇2 , 𝑇RMCD
2
, and
2 Use of robust control chart in Phase-I monitoring is very
𝑇RMVE values along with the respective control limits at 99%
important to assess the performance of the process as well
confidence level for the sample data are given in Figure 10. 2 2
as detecting outliers. We propose 𝑇RMCD /𝑇RMVE control
The control limits for 𝑇2 are arrived at based on beta charts for Phase-I monitoring of multivariate individual
2 2
distribution, and 𝑇RMCD /𝑇RMVE are calculated using (6) for observations. The control limits for these charts are arrived
𝑝 = 3 and 𝑚 = 105. From Figure 10, it is very clear that both empirically and a non-linear regression model is used for
2 2
𝑇RMCD and 𝑇RMVE control chart alarms signal for 3 outliers arriving control limits for any sample size. The performance
whereas the standard 𝑇2 control chart alarm signals for none of the proposed charts were compared under various data
even though all the charts are having the same pattern. This scenarios using large number of Monte Carlo simulations.
2
indicates the effectiveness of the proposed robust control Our simulation studies indicate that 𝑇RMVE control charts
charts in identifying the outliers. are performing well for smaller sample sizes and smaller
Journal of Quality and Reliability Engineering 13

𝑇2 -standard
40
30

𝑇2
20
(UCL)
10
0
0 20 40 60 80 100
Sample number
𝑇2 -RMCD
40
30
20 (UCL)
𝑇2 10
0
0 20 40 60 80 100
Sample number
𝑇2 -RMVE
40
30
20 (UCL)
𝑇2

10
0
0 20 40 60 80 100
Sample number

Figure 10: 𝑇2 , 𝑇RMCD

2 2
, and 𝑇RMVE control charts for the sample data.

2
dimension where as 𝑇RMCD control charts are performing [7] D. L. Donoho and P. J. Huber, “The notion of breakdown point,”
well for larger sample sizes and larger dimensions. We in A Festschrift for Erich Lehmann, P. Bickel, K. Doksum, and
illustrated our proposed robust control chart methodology J. Hodges, Eds., pp. 157–184, Wadsworth, Belmont, Calif, USA,
using a case study from the electronic industry. 1983.
[8] H. P. Lopuhaä and P. J. Rousseeuw, “Breakdown points of affine
equivariant estimators of multivariate location and covariance
Acknowledgments matrices,” The Annals of Statistics, vol. 19, pp. 229–248, 1991.
[9] D. L. Donoho and M. Gasko, “Breakdown properties of location
The authors would like to thank the editor and two anony-
estimates based on halfspace depth and projected outlyingness,”
mous referees for their valuable comments and suggestions The Annals of Statistics, vol. 20, pp. 1803–1827, 1992.
that substantially improved the overall quality of an ear-
[10] P. L. Davies, “Asymptotic behavior of S-estimates of multivariate
lier version of this paper. The research is supported by a
location parameters and dispersion matrices,” The Annals of
grant from the Natural Science and Engineering Council of Statistics, vol. 15, pp. 1269–1292, 1987.
Canada.
[11] C. Croux and G. Haesbroeck, “Influence function and efficiency
of the minimum covariance determinant scatter matrix estima-
References tor,” Journal of Multivariate Analysis, vol. 71, no. 2, pp. 161–190,
1999.
[1] H. Hotelling, “Multivariate quality control,” in Techniques of [12] G. Pison, S. van Aelst, and G. Willems, “Small sample correc-
Statistical Analysis, C. Eisenhart, H. Hastay, and W. A. Wallis, tions for LTS and MCD,” Metrika, vol. 55, no. 1-2, pp. 111–123,
Eds., pp. 111–184, McGrawHill, New York, NY, USA, 1947. 2002.
[2] N. D. Tracy, J. C. Young, and R. L. Mason, “Multivariate [13] D. L. Woodru and D. M. Rocke, “Computable robust estimation
control charts for individual observations,” Journal of Quality of multivariate location and shape in high dimension using
Technology, vol. 24, pp. 88–95, 1992. compound estimators,” Journal of American Statistical Associ-
[3] J. H. Sullivan and W. H. Woodall, “A comparison of multivariate ation, vol. 89, pp. 888–896, 1994.
control charts for individual observations,” Journal of Quality [14] D. M. Hawkins and D. J. Olive, “Improved feasible solution
Technology, vol. 28, no. 4, pp. 398–408, 1996. algorithms for high breakdown estimation,” Computational
[4] J. A. N. Vargas, “Robust estimation in multivariate control Statistics and Data Analysis, vol. 30, no. 1, pp. 1–11, 1999.
charts for individual observations,” Journal of Quality Technol- [15] P. J. Rousseeuw and K. van Driessen, “A fast algorithm for the
ogy, vol. 35, no. 4, pp. 367–376, 2003. minimum covariance determinant estimator,” Technometrics,
[5] W. A. Jensen, J. B. Birch, and W. H. Woodall, “High breakdown vol. 41, no. 3, pp. 212–223, 1999.
estimation methods for phase I multivariate control charts,” [16] P. J. Rousseeuw and B. C. van Zomeren, “Unmasking multivari-
Quality and Reliability Engineering International, vol. 23, no. 5, ate outliers and leverage points,” Journal of American Statistical
pp. 615–629, 2007. Association, vol. 85, pp. 633–639, 1990.
[6] S. Chenouri, S. H. Steiner, and A. M. Variyath, “A multivariate [17] P. L. Davies, “The asymptotics of Rousseeuw’s minimum volume
robust control chart for individual observations,” Journal of ellipsoid estimator,” The Annals of Statistics, vol. 20, no. 4, pp.
Quality Technology, vol. 41, no. 3, pp. 259–271, 2009. 1828–1843, 1992.
14 Journal of Quality and Reliability Engineering

[18] C. Croux and G. Haesbroeck, “An easy way to increase the finite-
sample efficiency of the resampled minimum volume ellipsoid
estimator,” Computational Statistics and Data Analysis, vol. 25,
no. 2, pp. 125–141, 1997.
[19] V. Torodov, “rrcov: Scalable Robust Estimators with High
Breakdown point,” R package version 0.5-03, 2009, http://
cran.r-project.org/web/packages/rrcov/.
 International Journal of

Rotating
Machinery

International Journal of
The Scientific
Engineering Distributed
Journal of
Journal of

Hindawi Publishing Corporation

World Journal
Hindawi Publishing Corporation Hindawi Publishing Corporation
Sensors
Hindawi Publishing Corporation
Sensor Networks
Hindawi Publishing Corporation
http://www.hindawi.com Volume 2014 http://www.hindawi.com Volume 2014 http://www.hindawi.com Volume 2014 http://www.hindawi.com Volume 2014 http://www.hindawi.com Volume 2014

Journal of

Control Science
and Engineering

Advances in
Civil Engineering
Hindawi Publishing Corporation Hindawi Publishing Corporation
http://www.hindawi.com Volume 2014 http://www.hindawi.com Volume 2014

Submit your manuscripts at

http://www.hindawi.com

Journal of
Journal of Electrical and Computer
Robotics
Hindawi Publishing Corporation
Engineering
Hindawi Publishing Corporation
http://www.hindawi.com Volume 2014 http://www.hindawi.com Volume 2014

VLSI Design
Advances in
OptoElectronics
International Journal of

International Journal of
Modelling &
Simulation
Aerospace
Hindawi Publishing Corporation Volume 2014
Navigation and
Observation
Hindawi Publishing Corporation
http://www.hindawi.com Volume 2014
in Engineering
Hindawi Publishing Corporation
http://www.hindawi.com Volume 2014
Engineering
Hindawi Publishing Corporation
http://www.hindawi.com Volume 2014
Hindawi Publishing Corporation
http://www.hindawi.com
http://www.hindawi.com Volume 2014

International Journal of
International Journal of Antennas and Active and Passive Advances in
Chemical Engineering Propagation Electronic Components Shock and Vibration Acoustics and Vibration
Hindawi Publishing Corporation Hindawi Publishing Corporation Hindawi Publishing Corporation Hindawi Publishing Corporation Hindawi Publishing Corporation
http://www.hindawi.com Volume 2014 http://www.hindawi.com Volume 2014 http://www.hindawi.com Volume 2014 http://www.hindawi.com Volume 2014 http://www.hindawi.com Volume 2014