Beruflich Dokumente
Kultur Dokumente
DOI: 10.1002/qre.2385
RESEARCH ARTICLE
KEYWORDS
metrics, nonlinear profile data, nonparametric EWMA control chart, support vector regression
Qual Reliab Engng Int. 2018;1–9. wileyonlinelibrary.com/journal/qre © 2018 John Wiley & Sons, Ltd. 1
2 LI ET AL.
of historical data and used to establish control limits for profile autocorrelation of error terms in Phase II monitor-
Phase II monitoring. In Phase II, the data are sequentially ing. First, they proposed U statistic to eliminate autocor-
collected over time to assess whether the parameters of relation. Then, based on the adjusted parameter
the process have changed from the estimated values in estimation, a multivariate T2 control chart was designed
the Phase I study. In this paper, we focus our research to monitor the parameters of the model. Finally, they
on the Phase II study for timely detecting the shifts/ used average run length (ARL) to evaluate the detecting
changes in the process parameters. Due to the complexity performance of their proposed method. Woodall et al7
of parameters estimation involved in nonlinear profile and Woodall8 provided a detailed review on the use of
monitoring, Williams et al1 considered the use of non- control charts for monitoring process and product quality
parametric monitoring methods and the use of metrics profiles. Zi et al9 developed a nonparametric method (a
to measure deviations from a baseline profile. After multivariate sign EWMA control chart) for monitoring
applying spline smoothing techniques to model the verti- linear profile data in case where the normality assump-
cal density profile (VDP) data, they calculated the five tion is doubtful. Recently, Chiang et al10 investigated
metrics and employed an individual Shewhart control another operational and sensitive Phase‐II monitoring
chart based on the moving range (I‐MR chart) to establish method for the linear profile model that exhibits within‐
control limits. However, according to Montgomery2 and profile autocorrelation.
Pan et al,3 exponentially weighted moving average
(EWMA) and cumulative sum control charts are proved
to be more efficient than Shewhart control chart for
2.2 | Nonlinear profile monitoring
detecting the small sustained shift in process mean. Thus,
instead of using I‐MR chart, we propose a nonparametric Nonlinear profile monitoring, on the other hand, was
rank‐based EWMA control chart in Phase II monitoring. considered by other researchers, for instance, Walker
Moreover, a more efficient support vector regression and Wright,11 Williams et al,1 Moguerza et al,12 and
(SVR) method is used to detect and classify the process Hung et al.13 Williams et al1 proposed using three differ-
shifts in nonlinear profiles. ent T2 statistics for phase I analysis to monitor the coeffi-
cients resulting from a parametric nonlinear regression
model that was used to fit profile data. They also consid-
2 | L I T E R A T U R E RE V I E W
ered the use of nonparametric regression method and the
use of metrics to measure deviations from a reference
2.1 | Linear profile monitoring
profile. Hung et al13 proposed using SVR to fit in‐control
Linear profile monitoring was studied by several profiles. Then, they employed the moving block bootstrap
researchers, such as Kang and Albin,4 and Khedmati method to generate correlated samples for each in‐control
and Niaki.5 Kang and Albin4 proposed two control chart profile and obtain a simultaneous confidence region for
methods for Phase II monitoring of simple linear profiles the underlying functional relationship. The obtained con-
process. Their first method used a multivariate T2 control fidence region was used to monitor the real AIDS data
chart to monitor the coefficients of regression. The sec- collected from hospitals in Taiwan. In practice, the func-
ond one used EWMA and R control chart to monitor tional relationship of the profile data rarely occurs in lin-
the mean and variance of the regression residuals. Kim ear form and the real data usually do not follow normal
et al6 proposed a method based on the combination of distribution. Thus, in this paper, the functional relation-
three EWMA charts. They showed that their proposed ship of profile data is described via a nonparametric
method is superior to those proposed by Kang and Albin.4 regression model and a nonparametric rank‐based
Khedmati and Niaki5 considered the effect of between‐ EWMA control chart is developed for detecting the
LI ET AL. 3
process shifts for nonlinear profile data in Phase II denotes the mean absolute deviation, M4 denotes the
monitoring. absolute value of maximum deviation (disregarding the
direction of dissimilarity), and M5 denotes the sum of
squared differences between the predicted values for the
3 | RESEARCH METHODOLOGY
measurement yij and the corresponding baseline profiles.
Let Mj, p, j = 1, 2, …, be the pth metric for the corre-
3.1 | Developing a nonparametric EWMA
sponding jth profile in the Phase II study. The rank of
control chart for phase II monitoring
M with respect to the reference metric
n j, p o
Let yij be the measurement of the ith observations in the M 1;p ; M 2;p ; ; …; M *g−1;p
* *
can be calculated as
jth profile and xij be the vector of the corresponding
R*j;p ¼ 1 þ ∑k¼1 I M j;p > M *k;p , where the indicator
g−1
explanatory variables such that i = 1, 2, …, nj for each
j = 1, 2, …. When the process is in statistical control, function I M j;p > M *k;p ¼ 1 if M j;p > M *k;p ; otherwise,
the underlying model is assumed to be the indicator function = 0, and g − 1 is the sample
size of referencen metric. The detailed o calculation of
yij ¼ f x ij þ εij ; i ¼ 1; 2; …; nj ; j ¼ 1; 2; …; (1) reference metric M *1;p ; M *2;p ; ; …; M *g−1;p is explained in
Section 3.2. Then, the standardized rank of R*j;p can be
where f (x) is a known function with certain degree of
calculated as
smoothness and the random errors εij are generally
assumed to follow some distributions. Without loss of
2 * gþ1
generality, the nj are taken to be equal and the explana- Rj;p ¼ Rj;p − : (2)
g 2
tory variables are assumed to be fixed for different j's.
The predicted values byij of the measurement yij for the Replacing the individual observation by the standard-
corresponding explanatory variables are calculated based ized rank Rj, p, our proposed EWMA statistics can be
on the SVR model. Then, the five metrics for the jth written as Equation (3).
profile suggested by Williams et al1 can be calculated
as follows: EWMAj;p ¼ ð1 − λÞEWMAj−1;p þ λ Rj;p ; j ¼ 1; 2; …; (3)
1 3 1
n the points − 1; − 1; … ; 1 − with mean zero
Mj2 ¼ ∑ byij − e
y i g g g
i¼1 g2 − 1
and variance , the control limits of our proposed
3g2
1 n
nonparametric EWMA control chart are defined as
Mj3 ¼ ∑ b yij − e
y i
n i¼1 sffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
λ g2 − 1
±LR (4)
Mj4 ¼ Mj1 2 − λ 3g2
where e y i is the reference profile for the ith observation. 3.2 | Constructing the reference profile in
The detailed calculation of e y i is explained in Section 3.2. Phase I study
In conducting the Phase II study, we adopt the concept
In this paper, we use SVR to fit profiles for describing the
of residual control chart and employ a nonparametric
functional relationship shown in Equation (1). The SVR
EWMA control chart proposed by Hackel and Ledolter14
is a supervised statistical learning algorithm for regres-
to monitor these five metrics. In other words, we consider
sion problem. In SVR, the decision variables are mapped
the metric for profile j as an individual observation and
onto a feature space, and a linear model in Equation (5) is
use a nonparametric EWMA control chart is to monitor
constructed in the feature space.
the mean change of these five metrics with n observations
over j time period, where M1 denotes the maximum devi- gðx; wÞ ¼ wT ϕðx Þ þ b; (5)
ation, M2 denotes the sum of absolute deviations, M3 where w is normal vector, ϕ(.) is a nonlinear
4 LI ET AL.
8
>
> y − gðx i ; w Þ ≤ ϵ þ ξ i where β0 = 0.5, β1 = 2, and the explanatory variable are
< ij
gðx i ; w Þ − yij ≤ ϵ þ ξ *i (6) fixed as xij = 0.01i, i = 1, …, 300 for each profile j. For ran-
>
> dom error terms εij, we consider the following two scenar-
:
ξ i ; ξ *i ≥ 0 ios: (1) a standard normal distribution N(0, σ) with
parameter σ = 1; (2) an exponential distribution with rate
where ϵ > 0 is a certain threshold, constant C > 0 is a
parameter θ = 0.5. Since a reference metric
penalization that can be viewed as a way to control n o
M *1;p ; M *2;p ; ; …; M *g−1;p obtained from Phase I study
over‐fitting, and ξi and ξ *i are slack variables. The loss
function is given by (see Sections 3.1 and 3.2 for details) is needed to construct
a Phase II control chart, an SVR model for obtaining
(
the reference metric is built based on Phase I profile
0 ; if yij − gðx i ; wÞ < ϵ
L yij ; gðx i ; w Þ ¼ : training data generated from the SVR model shown in
yij − gðx i ; wÞ − ϵ ;
otherwise Equation (6). In the parameter optimization process with
5‐fold cross validation, the optimal values of the parame-
The esp‐regression machine in R package e1071 (Mayer ters C and ϵ in Equation (6) are set as 512 and 0.0001,
et al15) is used to build the above SVR model based on respectively. Then, the reference metrics can be obtained
the training dataset collected from the Phase I study. accordingly. In this paper, the sample size of the refer-
Through the SVR model, the reference profile for the ence metrics is set as g = 21. The simulation results for
g−1
ith observation is calculated as yei ¼ ∑j¼1 y_ ij =ðg − 1Þ, various control parameter LR values under different λ
i = 1, …, n, where y_ ij is the predicted value for the corre- are summarized in Table 1.
sponding observation yij. Thus, the following reference For the random error terms following a standard nor-
metrics can be calculated accordingly: mal distribution or an exponential distribution, the simu-
lation results of ARL1 values with λ = 0.05 and λ = 0.1
n o under different shifts in parameter β0, β1, and standard
M *j;1 ¼ sign max y_ ij − e
y i ; j ¼ 1; …; ðg − 1Þ derivation σ are shown in Figures 2 and 3, respectively.
i¼1; …; n
Note that the shifts in parameters β0 and β1 are defined
as β0 + δ and β1 + δ. The shifts in parameters σ and θ
n
are defined as 1 × δ and 0.5 + δ for a standard normal
M *j;2 ¼ ∑ y_ ij − e
y i ; j ¼ 1; …; ðg − 1Þ
i¼1 and an exponential distribution, respectively.
As one can see from Figure 2, the ARL1 decreases as δ
increases and ARL1 increases as λ increases. The
1 n
M *j;3 ¼ ∑ y_ ij − e
y i ; j ¼ 1; …; ðg − 1Þ
n i¼1
TABLE 1 Various control parameter LR values under different λ
with g = 21. for achieving ARL0=370
M *j;4 ¼ M *j;1 ; j ¼ 1; …; ðg − 1Þ λ 0.05 0.1 0.2 0.3 0.4
LR 1.37 1.77 2.10 2.18 2.17
λ
n 2 0.5 0.6 0.7 0.8 0.9
M *j;5 ¼ ∑ y_ ij −e
y i ; j ¼ 1; …; ðg − 1Þ: LR 2.13 2.05 1.96 1.87 1.77
i¼1
LI ET AL. 5
simulation results show that the detecting ability of the are almost the same under different shifts in various
proposed nonparametric EWMA chart increases when parameters when the random error terms follow a
the magnitudes of shift in parameter δ increase or the standard normal distribution.
smoothing parameter λ decreases. In most cases, the As one can see from Figure 3, the ARL1 decreases as
ARL1 of the nonparametric EWMA control chart with δ increases and ARL1 increases as λ increases. The
the metric M1 and M4 for detecting the shifts in various simulation results show that the detecting ability of
parameters are larger than those of M2, M3, and M5, the proposed nonparametric EWMA chart increases
ie, the detecting performances of the nonparametric when the magnitudes of shift in parameter δ increase
EWMA control chart with metrics M1 and M4 are worse or the smoothing parameter λ decreases. Note that the
than those of M2, M3, and M5. Note that the detecting detecting performances of the nonparametric EWMA
performances of the nonparametric EWMA control chart control chart with metrics M2 and M3 are the same;
with metrics M2 and M3 are the same; thus, only four thus, only four curves are shown in Figure 3. In other
curves are shown in Figure 2. The results are expected words, the detecting performances of the control chart
since the statistical forms of the two metrics M2 and M3 with metrics M2, M3, and M5 are almost the same under
differ only by a constant. Moreover, Figure 2 shows different shifts in various parameters when the random
that the detecting performances of the nonparametric error terms follow an exponential distribution. More-
EWMA control chart with metrics M2, M3, and M5 over, in most cases, the ARL1 of the EWMA control
FIGURE 2 ARL1 values under different shifts in parameter: A, β0; B, β1; or C, σ for the given λ = 0.05 (left panel) and λ = 0.1 (right panel)
when the random error terms follow a standard normal distribution
6 LI ET AL.
chart with the metric M1 and M4 for detecting shifts in density property of finished boards is the quality charac-
various parameters except β0 are larger than metrics M2, teristic required to be closely monitored. Each resulting
M3, and M5. Therefore, one can conclude that the profile consists of 314 density measurements, and the dis-
detecting performances of the nonparametric EWMA tance between two consecutive measures is 0.002 inch.
control chart with metrics M1 and M4 are worse than The original VDP data with 24 profiles are illustrated in
those of M2, M3, and M5. These results are expected Figure 4, where the depth of thickness of the particle
since the metric M1 and M4 did not use all the informa- boards ranges from 0 to 0.626 inch.
tion in the data set. To construct a Phase II control chart, we first employ
the SVR technique as mentioned in Section 3 to model
the VDP data. Note that the parameter C and ϵ are set
5 | A NUMERICAL EXAMPLE as 8 and 0.001, respectively, in Equation (6). Then, the
mean squared error (MSE) and mean absolute percent
In order to demonstrate the practical application of our error (MAPE) for VDP data can be obtained and summa-
proposed control chart, the numerical example with rized in Table 2.
VDP data given in Walker and Wright11 is used for illus- According to Table 2, the average MAPE can be
tration purpose. In manufacturing the particle boards, the approximated to 0.49%, which indicates its percent
FIGURE 3 ARL1 values under different shifts in parameter: A, β0; B, β1; or C, θ for given the λ = 0.05 (left panel) and λ = 0.1 (right panel)
when the random error terms follow an exponential distribution
LI ET AL. 7
FIGURE 4 The original VDP data with 24 particle boards FIGURE 5 The 24 reference profiles after using SVR model to fit
[Colour figure can be viewed at wileyonlinelibrary.com] VDP data [Colour figure can be viewed at wileyonlinelibrary.com]