Received: 4 January 2024 Revised: 11 May 2024 Accepted: 30 May 2024

DOI: 10.1002/qre.3603

RESEARCH ARTICLE

Adaptive EWMA control chart by using support vector

regression

Muhammad Waqas Kazmi Muhammad Noor-ul-Amin

Department of Statistics, COMSATS

University Islamabad, Lahore Campus, Abstract
Lahore, Pakistan Traditional control charts depend on the process parameters that are used to
Correspondence
monitor the shifts in the process. The adaptive control charts are used to adapt
Muhammad Noor-ul-Amin, COMSATS a process parameter during the online monitoring. This research introduces a
University, Defense Road off Raiwind support vector regression (SVR) based adaptive exponentially weighted moving
Road, Lahore 54000, Pakistan.
Email: [email protected] average control chat to enhance the monitoring of the mean in industrial pro-
cesses. The study systematically investigates the comparative efficiency of linear,
radial basis function (RBF), and polynomial functions within the SVR frame-
work. The proposed SVR-based AEWMA control chart leverages the strengths
of the RBF kernel, providing a robust mechanism for detecting shifts in the pro-
cess mean by adapting the smoothing constant according to the size of the shift.
To validate the efficacy of the proposed methodology, a practical application
is presented by using real-life data. The application showcases the adaptabil-
ity and reliability of the SVR-based adaptive EWMA control chart in effectively
monitoring location shifts.

KEYWORDS
adaptive control chart, machine learning, radial basis function, support vector regression

1 INTRODUCTION

Statistical techniques have attracted practitioners and researchers to keep an eye on the procedure under study. Quality
control charts are one of the common tools among these techniques to signal out of control when a process violates the
standards. One of the main objectives of using such monitoring tools is to find abnormalities in a process of manufacturing
or production. This helps empower the control system to promptly implement appropriate measures before the production
of faulty items. The fundamental elements of statistical process control (SPC) involve recognizing and overseeing special-
cause anomalies in a manufacturing workflow. This helps to maximize process efficiency and manufacturing capacity
while improving the quality of the goods.1
Walter A. Shewhart introduced the notion of a control chart, which subsequently became the cornerstone in mod-
ern SPC. Recent and highly effective statistical process monitoring tools encompass the Cumulative Sum (CUSUM) and
EWMA control charts. Page2 proposed the CUSUM chart, whereas Roberts3 suggested the EWMA chart. Such control
charts and the modified versions of them, find widespread application in process and service industries. They are pri-
marily employed to monitor slight variations in process parameters, as these minor disruptions can potentially lead to
significant financial consequences. However, in comparison to the Shewhart chart, such control charts do not work well
while attempting to spot significant process alterations. Although Control charts based on memory are widely used, their
effectiveness is restricted to scenarios in which shift magnitudes are recognized or in which the quality practitioner is

Qual Reliab Engng Int. 2024;40:3831–3843. wileyonlinelibrary.com/journal/qre © 2024 John Wiley & Sons Ltd. 3831
 10991638, 2024, 7, Downloaded from https://onlinelibrary.wiley.com/doi/10.1002/qre.3603 by Muhammad Noor-ul-Amin - COMSATS Inst of Info Techn - I , Wiley Online Library on [11/11/2024]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
3832 KAZMI and NOOR-UL-AMIN

specifically responsible for creating a shift-specific based monitoring chart. While, in practical scenarios, the circum-
stances often differ; shift sizes may not be predetermined and may need to be estimated. Following this, scholars have
turned their attention to the more recent creation of dual control-charting frameworks and adaptive control-charting
frameworks in an effort to improve safeguarding in response to varying-sized shifts. Typically, an unbiased estimator
is assisted by EWMA statistic which is applied to adaptive control charts in order to estimate process shift. Subse-
quently, the derived shift estimator is used to optimize the control chart.4 Sparks5 developed an adaptive Cumulative
Sum (ACUSUM) chart to track the variations in process location, it demonstrated better performance in detecting shifts
of varying sizes than the conventional CUSUM chart. To improve the sensitivity of the CUSUM chart, this concept entails
estimating shift magnitude using the EWMA statistic after which adding it to the slack value. An adaptive AEWMA chart
was introduced by Capizzi and Masarotto6 as a means to observe the changes in process location. AEWMA chart was
crafted using the Huber score function7 to effectively combine the characteristics of Shewhart and EWMA charts. Sar-
war and Noor-ul-Amin (2022) formulated an adaptive EWMA control chart by incorporating a continuous function. One
notable benefit of this chart is that it allows for uniform adjustments to the smoothing parameter value based on the pro-
cess’s mean shift magnitude. The parameter-free AEWMA control chart that was developed by Noor et al.8 depends on
Bayesian theory under various loss functions, such as the squared error loss function and the linex loss function to adopt
the value of smoothing constant in accordance with the shift size in the process while constructing EWMA statistic.
Extensive scholarly literature attests to the increased effectiveness of employing machine learning-based tools for the
purpose of statistical process monitoring. These tools rooted in the domain of machine learning have demonstrated notable
capabilities in enhancing the monitoring and control of various processes. This affirmation underscores the growing
recognition and adoption of such advanced methodologies, affirming their utility in improving the overall efficiency of
process monitoring practices. For instance, within the realm of statistical process monitoring, various studies conducted
by researchers such as Cuentas et al.,9 He et al.,10 Shao and Hu,11 Apsemidis et al.,12 Chowdhury and Janan,13 and Pieters
et al.14 have extensively leveraged machine learning tools. These notable contributions have demonstrated the integra-
tion and application of advanced machine learning methodologies to enhance and refine statistical process monitoring
practices. The fact that these researchers are using these tools highlights a larger trend in the academic community and
validates the usefulness and potential of machine learning in process monitoring.
Riaz et al.15 introduced an adaptive EWMA control chart utilizing the principal component method to monitor the
process mean vector efficiently. Their method aims to enhance process monitoring by effectively identifying shifts in mean
vector patterns. This approach offers a concise and adaptable solution for real-time process control in diverse industrial
settings. Similarly, Noor et al.16 proposed a strategy for creating EWMA control charts, using residuals from the SVM
regression model. This approach aims to manage the adjustment of risk factors linked to patients. Getting motivation
from the literature, in this work, we proposed an adaptive EWMA control chart that adapts the smoothing constant from
a trained SVR model. To generate the run length profiles for the AEWMA chart, extensive Monte Carlo simulations are
used to calculate the average run length (ARL) and standard deviation of the run length (SDRL). Also, the performance of
the proposed chart is compared to the competitor control charts. Section 2 gives a brief review of SVR and the scheme of
adaptive control charts explained in Section 3. The Section 4 explains the design of the proposed chart and the performance
of the proposed chart is evaluated in Section 5. Section 6 investigates the results of a proposed chart with a considered
control chart. The real-life application is presented in Section 7 and finally, Section 8 concludes the findings.

2 BRIEF REVIEW OF SVR

A supervised machine learning algorithm called SVR was created to deal with regression issues. Regression analysis is a
useful tool for studying the relationship between an explained variable and one or more explanatory variables. SVR for-
mulates an optimization problem and uses that to obtain a regression function that connects input predictor variables to
observed response values. It is especially useful for handling data with significant dimensions because it strikes a balance
between prediction error and the complexities of models. Notably, SVR extends the Support Vector Machine (SVM) clas-
sification algorithm.17 Following18 the objective of SVR is to identify a function f(x) that minimizes the departure from the
actual target values across all training data while also maintaining a flat structure.
Suppose a given training data (x1 ,y1 , x2 ,y2 ),. . . ,(xi , yi ), ⊂ X × ℜ and i = 1,2,. . . ,n.
Consider the linear function f defined as

𝑓 (𝑥) = ⟨𝜔, 𝑥⟩ + 𝑏, with 𝜔 ∈ 𝑋, 𝑏 ∈ ℜ,

 10991638, 2024, 7, Downloaded from https://onlinelibrary.wiley.com/doi/10.1002/qre.3603 by Muhammad Noor-ul-Amin - COMSATS Inst of Info Techn - I , Wiley Online Library on [11/11/2024]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
KAZMI and NOOR-UL-AMIN 3833

where <⋅⋅> denotes the dot product in x, x is the input space, 𝜔 is the weight vector, and b is the bias term. In the above
equation, flatness implies the pursuit of a small 𝜔. To achieve this minimizing the norm of 𝜔.
Minimize

∑𝑙
1 2 ( )
‖𝜔‖ + 𝐶 𝜉 + 𝜉𝑖∗
2 𝑖=1

Subject to

𝑦𝑖 − ⟨𝜔, 𝑥𝑖 ⟩ − 𝑏 ≤ 𝜀 + 𝜉

⟨𝜔, 𝑥𝑖 ⟩ + 𝑏 − 𝑦𝑖 ≤ 𝜀 + 𝜉𝑖∗

𝜉, 𝜉𝑖∗ ≥ 0

where C > 0, dictates the balance between the flatness of f and the extent to which deviations beyond ε are acceptable.
The regression control chart problems are mostly related to the entire count of variables, assumption of linear-
ity, and precision of fitting techniques. Conventional regression techniques frequently perform worse when dealing with
these issues. The utilization of the SVR method in constructing a regression-based adaptive control chart aimed at achiev-
ing improved results by adapting the value of the smoothing constant. Two key factors contribute to the success and utility
of SVR: the application of mathematical programming techniques and the use of kernel functions. As a result, quadratic
programming provides a competitive solution and a strong mathematical basis for SVR. Additionally, the incorporation
of kernel methods enables SVR to be a suitable technique for handling both linear and nonlinear data.
The predictive efficacy of SVR fluctuates based on the selection of the kernel function and the parameters that go
with it. Thus, it is imperative to initially choose an appropriate kernel function for SVR implementation. Linear kernels,
known for their simplicity, assume a linear relationship between input and output. Polynomial kernels, on the other hand,
offer flexibility by allowing for nonlinear relationships through polynomial transformations. Radial Basis Function (RBF)
kernels are particularly powerful for capturing intricate nonlinear patterns, as they map data into higher-dimensional
spaces. Each kernel type has its strengths and weaknesses, providing researchers and practitioners with versatile tools to
address diverse regression tasks effectively. The selection of kernel functions is highly contingent on the specific problem
and constitutes a crucial factor in SVR applications as illustrated in Figure 1.
These types of kernel functions are, such.

∙ Linear Kernel: K(x,y) = x.y.

∙ Polynomial Kernel: K(x,y) = (x.y + 1)d , where d > 0 is a constant that defines the kernel order.
∙ Gaussian RBF Kernel: 𝐾(𝑥, 𝑦) = 𝑒𝑥𝑝(−‖𝑥 − 𝑦‖2 ∕2𝜎2 ), where σ > 0 is a parameter that defines the kernel width.

3 ADAPTIVE EWMA CONTROL CHARTS

When a control chart responds to variations in process data by altering at least one of its design parameters, it is referred
to as an adaptive control chart. Within an adaptive EWMA control chart designed to seamlessly integrate aspects of both
Shewhart and EWMA charts, the core concept involves choosing the significance of the earlier observation based on the
error magnitude. This method seeks to mitigate the inertia problem while identifying shifts of various sizes.19 Sarwar and
Noor-ul-Amin (2022) introduced a new design of Adaptation of the smoothing constant being used for EWMA control
charts. With a location parameter of 𝜇 and a scale parameter of 𝜎2 at time 𝑡 ≥ 1, imagine {X𝑡} as a series of independent,
identically distributed random variables that follow a normal distribution. Consider 𝛿̂𝑡 as an estimate for a mean shift at
time t. Now, the estimator for the shift (𝛿) by following the formulation Jiang et al.20 is expressed as:

𝛿̂𝑡∗ = 𝜑 (𝑋𝑡 ) + (1 − 𝜑) 𝛿̂𝑡−1

∗
 10991638, 2024, 7, Downloaded from https://onlinelibrary.wiley.com/doi/10.1002/qre.3603 by Muhammad Noor-ul-Amin - COMSATS Inst of Info Techn - I , Wiley Online Library on [11/11/2024]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
3834 KAZMI and NOOR-UL-AMIN

FIGURE 1 SVR with various kernels.

where smoothing constant 𝜑 ∈ (0, 1) and 𝛿̂0∗ = 0. The following estimator is used by the authors such as Sarwar and
Noor-ul-Amin (2022) for the estimation of the shift

𝛿̂𝑡∗
𝛿̂𝑡∗∗ = 𝑡
.
1 − (1 − 𝜑)

It is suggested to use 𝛿̂𝑡 = |𝛿̂𝑡∗∗ | , for estimating δ.

Consequently, the following equation represents the plotting statistics of the suggested AEWMA control chart for
observing process mean by using a sequence {Xt }

𝐹𝑡 = 𝑔(𝛿̂𝑡 )𝑋𝑡 + (1 − 𝑔(𝛿̂𝑡 ))𝐹𝑡−1

where Fo = 0, and 𝑔(𝛿̂𝑡 ) ∈ (0, 1) such that

⎧ [ 1 −𝑐 ] 𝑖𝑓 0 < 𝛿̂𝑡 ≤ 2.7

( ) ⎪ 𝑎 1+(𝛿̂𝑡 )
𝑔 𝛿̂𝑡 = ⎨ (1)
⎪
⎩1 𝑖𝑓 𝛿̂𝑡 > 2.7

The provided function has recommended to set the parameter a as 7, and for the constant c, values of 2 and 1 are
suggested when 𝛿̂𝑡 ≤ 1 𝑎𝑛𝑑 1 < 𝛿̂𝑡 ≤ 2.7, respectively.

4 DESIGN OF THE PROPOSED CONTROL CHART

This section suggests a novel AEWMA control chart for tracking shifts in a normally distributed process mean. Let Xt be an
independent and identically random observations sequence with mean μo and standard deviation σo from an in-control
process. The idea is to first train the SVR model given below with the specified kernel and other parameters of SVR. To
 10991638, 2024, 7, Downloaded from https://onlinelibrary.wiley.com/doi/10.1002/qre.3603 by Muhammad Noor-ul-Amin - COMSATS Inst of Info Techn - I , Wiley Online Library on [11/11/2024]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
KAZMI and NOOR-UL-AMIN 3835

FIGURE 2 Flow chart of proposed control chart.

do this use the standard normal variable Xt and the values of the smoothing constant (A supplementary file is provided,
containing the values of the smoothing constant alongside the shift values to train the SVR model) against specified values
Xt for training the SVR model given below.

𝑡
∑
𝑓 (𝑥) = 𝛼𝑖 𝐾 (𝑥, 𝑥𝑖 ) + 𝑏 (2)
𝑖=1

αi is the Langrage multiplier for each support vector determined during the training process, K(x, xi ) is the kernel func-
tion and b is the bias term. The various types of kernels available for use have been previously outlined in Section 2. Now
take the value of Xt of the desired quality characteristic and convert it to the standard normal form like 𝑍𝑡 = (𝑋𝑡 − 𝑥)∕𝑠̄
and then predict the value of smoothing constant λ through the trained SVR model given in Equation (2) for charting
statistic SAEWMAt .

𝑆𝐴𝐸𝑊𝑀𝐴𝑡 = 𝜆𝑋𝑡 + (1 − 𝜆) 𝑆𝐴𝐸𝑊𝑀𝐴𝑡 (3)

If |SAEWMAt | > h, then the process triggered out of control signal. Where ℏ, is always expected to be greater than zero,
serves as a threshold. The in-control ARL is fixed at a given level, say ARL0 , by determining the value of h. Every kernel
function that is specified has its threshold value h determined independently. The feature of the proposed SAEWMA
control chart is that it does not depend on some pre-specified parameter(s) like other AEWMA controls. Figure 2 briefly
describes the design of the proposal.
 10991638, 2024, 7, Downloaded from https://onlinelibrary.wiley.com/doi/10.1002/qre.3603 by Muhammad Noor-ul-Amin - COMSATS Inst of Info Techn - I , Wiley Online Library on [11/11/2024]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
3836 KAZMI and NOOR-UL-AMIN

TA B L E 1 ARL (SDRL) of proposed chart w.r.t. the training data sets.

Training data sets
Shift size (δ) 100 500 1000 5000
0.1 211.81(196.09) 211.37 (195.8) 206.64 (188.58) 204.88 (188.82)
0.25 73.56(57.57) 75.1 (59.3) 73 (57.41) 71.32 (55.87)
0.5 27.07(15.42) 27.43 (15.92) 27.22 (15.61) 26.06 (15.00)
0.75 15.64(7.22) 16.23 (7.43) 15.76 (7.14) 15.13 (7.35)
1.00 11.1(4.26) 11.32 (4.35) 11.08 (4.27) 10.97 (4.03)
1.5 7.01(2.07) 7.03 (2.12) 6.98 (2.09) 6.82 (2.11)
2.00 5.19(1.22) 5.15 (1.24) 5.09 (1.24) 5.03 (1.23)
3.00 3.66(0.63) 3.66 (0.63) 3.56 (0.62) 3.56 (0.63)

5 PERFORMANCE EVALUATION

An analysis using simulation is done to determine how well the suggested control chart works. The run length profile
is used to investigate the effectiveness of the proposed chart in this study, which is a commonly used evaluation met-
ric to evaluate how well out-of-control processes are monitored. This section explains how the run length profiles of
the suggested control chart are computed using the Monte Carlo simulation approach for quality characteristics follow-
ing a normal distribution. To get the value of the smoothing constant, the SVM model must first be trained with the
consideration of sample observation.
Setting up the threshold value

(i) Decide the in-control ARL0 (say 370) and select the value of h (threshold).
(ii) Select the sample of size n from a standard normal distribution and predict the value of the smoothing constant
according to the mean magnitude of the desired quality characteristic through the trained SVM model.
(iii) Find the value of the proposed SAEWMA statistic given in Equation (3).
(iv) Follow the procedure of the control chart, if the process is declared as out-of-control record the number of subgroups
as the in-control run length.
(v) Repeat steps from (i) to (iv) 50,000 times to calculate the in-control ARL and SDRL, and modify the value of ℎ if
requisite.

Evaluate the out-of-control ARL

(i) Generate the random sample from a shifted process and calculate the value of the proposed SAEWMA statistic given
in Equation (3)
(ii) Follow this procedure, if an out-of-control process is declared record the subgroup number as run length.
(iii) To compute the ARLs and 𝑆𝐷𝑅𝐿𝑠, repeat the whole procedure 50,000 times.

Hence, the RL profiles stated in Tables 1–4 have been widely produced for the performance assessment of this study.

5.1 Performance analysis to the number of training data sets

An important factor influencing the performance of the monitoring algorithms is the quantity of training sets. The effi-
ciency of the SVR can only be maximized when it is sufficiently trained. The accuracy to detect the false alarm rate can be
reduced if improper training is exercised. Datasets of varying sizes with parameters C = 0.1, epsilon = 0.1, and gamma = 1
(Optimal choice of parameters with RBF kernel) at ARL0 = 370, are created to assess the SAEWMA control chart. The
efficacy of the SAEWMA is examined as the total number of training examples increases over time. Several inferences can
be made from these findings: (1) Up to a certain degree of accuracy, the SAEWMA chart’s performance can be enhanced
by increasing the number of training examples (early detection of false alarm rate). This might be accounted for by the
increased likelihood of accurate representation of a problem space that comes with sufficient large training sets. However,
 10991638, 2024, 7, Downloaded from https://onlinelibrary.wiley.com/doi/10.1002/qre.3603 by Muhammad Noor-ul-Amin - COMSATS Inst of Info Techn - I , Wiley Online Library on [11/11/2024]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
KAZMI and NOOR-UL-AMIN 3837

TA B L E 2 Run length profile of proposed chart w.r.t different kernels.

Kernel Linear function Radial basis function Polynomial function
ARL
h 200 370 500 200 370 500 200 370 500
δ 0.3952 0.4382 0.4632 0.6205 0.6910 0.7279 0.8005 0.9651 1.0215
0.1 140.95 219.87 315.26 123.9 204.88 257.21 130.25 233.94 298.63
(129.62) (206.91) (293.23) (112.49) (188.82) (241.28) (113.8) (225.08) (279.4)
0.25 51.87 72.32 91.25 55.15 71.32 83.72 61.58 100.7 129.17
(40.82) (58.6) (77.8) (42.76) (55.87) (65.01) (51.24) (94.61) (116.72)
0.50 21.92 27.23 31.47 22.67 26.06 29.62 27.62 40.29 45.8
(13.19) (16.54) (19.46) (13.17) (15.00) (16.96) (19.29) (29.38) (34.83)
0.75 13.53 15.95 16.58 13.75 15.13 16.94 15.56 21.02 23.08
(6.42) (7.49) (7.81) (6.44) (7.35) (7.61) (9.62) (13.37) (15.07)
1.00 9.27 10.24 11.3 9.65 10.97 11.8 10.37 12.81 14.4
(3.73) (4.01) (4.37) (3.91) (4.03) (4.5) (6.05) (7.55) (8.39)
1.50 5.66 6.63 6.7 6.1 6.82 7.28 5.55 6.44 6.85
(1.83) (1.98) (2.07) (1.96) (2.11) (2.19) (3.1) (3.59) (3.81)
2.00 4.34 4.61 5.21 4.48 5.03 5.36 3.35 3.85 4.16
(1.15) (1.22) (1.29) (1.17) (1.23) (1.27) (1.84) (2.14) (228)
3.00 2.93 3.25 3.36 3.21 3.56 3.79 1.7 1.84 1.9
(0.63) (0.66) (0.67) (0.52) (0.63) (0.64) (0.84) (0.94) (0.98)

TA B L E 3 SVR employing a Gaussian RBF with varying values of G.

ARLo = 370
δ G=1 G = 1.5 G=2 G = 2.5
0.1 204.88 (188.82) 233.58 (215.24) 209.91 (191.71) 210.27 (191.71)
0.2 96.11 (78.92) 99.32 (83.03) 99.46 (81.33) 92.54 (77.49)
0.25 71.32 (55.87) 73.7 (57.35) 69.98 (53.56) 71.19 (57.84)
0.50 26.06 (15) 27.29 (15.33) 27.89 (15.89) 26.81 (15.69)
0.75 15.13 (7.35) 15.99 (7.05) 15.53 (6.9) 15.4 (7.04)
1.00 10.97 (4.03) 11.42 (4.27) 11.05 (4.12) 10.84 (4.21)
1.5 6.82 (2.11) 7.09 (2.06) 7.01 (2.01) 6.88 (2.04)
2 5.03 (1.23) 5.36 (1.23) 5.2 (1.19) 5.09 (1.21)
2.5 4.11 (0.82) 4.41 (0.81) 4.44 (0.79) 4.28 (0.8)

the SAEWMA chart’s performance would not be significantly enhanced by increasing the training size after hitting these
limits. (2) Larger training datasets are used to provide a more detailed representation of the density distribution in situa-
tions where out-of-control signals have minor magnitudes. (3) Out-of-control signals with significant shifts are easier to
identify with smaller training examples. This is because signals with large shifts have distinct pattern features that make
them stand out more clearly from other random in-control patterns as shown in Table 1.

5.2 Performance analysis with different kernels

The performance of the proposed SAEWMA control chart investigated with different kernels and results are given in
Table 2. It includes ARL values calculated for various shifts in the mean of monitored quality characteristics. The calcula-
tions were carried out for different kernel functions with n = 1, considering ARL0 values of 200, 370, and 500. The quality
characteristic is thought to be normally distributed in the situations, both during the training of the SVR model and when
utilizing the trained model for making predictions based on input values. This indicates that the mean shifted from μo = 0
to μ1 = 0 + 0.5 (1) = 0.5 if the tested (shift) value was 0.5. Table 2 reveals that with a shift size (δ) of 0, the ARL closely
aligns with the initial value (ARL0 ) for every kernel type. However, as the shift size increases, the ARL experiences a rapid
 10991638, 2024, 7, Downloaded from https://onlinelibrary.wiley.com/doi/10.1002/qre.3603 by Muhammad Noor-ul-Amin - COMSATS Inst of Info Techn - I , Wiley Online Library on [11/11/2024]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
3838 KAZMI and NOOR-UL-AMIN

TA B L E 4 Comparison with AEWMA—I and AEWMA—II control charts.

ARL 370 500
CC SAEWMA AEWMA—I AEWMA—II SAEWMA AEWMA—I AEWMA—II
h 0.6910 0.1996 0.4036 0.7279 0.20991 0.4250
δ ARL SDRL ARL SDRL ARL SDRL ARL SDRL ARL SDRL ARL SDRL
0 370.9 355.72 369.87 358.36 370.83 357.57 499.34 488.8 502.71 490.36 502.57 493.17
0.1 204.88 188.82 217.49 204.02 221.35 207.42 257.21 241.28 278.21 260.94 279.24 263.70
0.15 136.31 119.58 145.18 129.34 148.61 133.02 167.06 148.86 175.46 156.40 181.03 164.68
0.2 96.11 78.92 101.48 85.19 102.57 85.96 114.46 95.69 119.83 103.52 122.29 104.71
0.25 71.32 55.87 73.57 58.28 75.02 59.95 83.72 65.01 84.65 67.91 86.50 69.70
0.50 26.06 15.00 26.48 15.24 26.89 15.53 29.62 16.96 28.89 16.76 29.38 17.08
0.75 15.13 7.35 15.41 6.92 15.53 7.14 16.94 7.61 16.42 7.30 16.62 7.47
1 10.97 4.03 10.70 3.99 10.73 4.11 11.8 4.5 11.42 4.22 11.43 4.35
1.5 6.82 2.11 6.77 1.97 6.70 2.16 7.28 2.19 7.13 2.07 7.04 2.29
2 5.03 1.23 4.97 1.24 4.77 1.56 5.36 1.27 5.24 1.27 4.99 1.65
2.5 4.11 0.82 3.98 0.87 3.55 1.35 4.37 0.84 4.17 0.89 3.71 1.43
3 3.56 0.63 3.34 0.67 2.69 1.21 3.79 0.64 3.50 0.68 2.77 1.29
4 2.93 0.39 2.57 0.53 1.54 0.77 3.05 0.41 2.70 0.52 1.56 0.80
5 2.43 0.5 2.09 0.31 1.10 0.34 2.68 0.5 2.15 0.38 1.11 0.34

FIGURE 3 (A) Kernel performance in terms of ARLs. (B) Kernel performance in terms of parameter prediction.

decline. Indeed, the rationale for employing SVM charts with various kernel functions is to enhance the effectiveness of
process monitoring by addressing certain limitations associated with conventional SPC methods.

1. With a minor shift, the RBF and polynomial kernels exhibit early detection of out-of-control points as compared to the
linear function. For instance, when ARL0 = 200, at a shift of 0.1, the linear kernel yields an ARL of 140.95, while its
counterparts RBF and polynomial kernels achieve ARL values of 123.9 and 130.25, respectively.
2. The RBF kernel demonstrates superior performance compared to the linear and polynomial kernels, particularly when
the initial ARL0 is substantial (e.g., 370). This is attributed to the fact that at higher ARL0 values, the SVR has a greater
capacity to effectively distinguish between in-control and out-of-control points, as demonstrated in Figure 3A.
 10991638, 2024, 7, Downloaded from https://onlinelibrary.wiley.com/doi/10.1002/qre.3603 by Muhammad Noor-ul-Amin - COMSATS Inst of Info Techn - I , Wiley Online Library on [11/11/2024]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
KAZMI and NOOR-UL-AMIN 3839

(A) (B)

(C) (D)

FIGURE 4 SVR employing a Gaussian RBF with varying values of Gamma.

3. It is evident that, given our focus on determining the smoothing parameter value for constructing the SEWMA
statistic, the SVR model using the RBF Kernel provides more precise predicted values for the smoothing
constant compared to other kernels across all mean shift values as supported by the findings presented in
Figure 3B.
4. When considering substantial shifts (δ ≥ 1), all kernel functions exhibit comparable performances. For instance, at
a shift of δ = 1, with an initial ARL0 set to 200, the corresponding ARL values for the linear, RBF, and polynomial
kernels are observed to be 9.27, 9.65, and 10.37, respectively. This convergence in performance highlights the consistent
behavior of the different kernel functions under larger shift conditions.

5.3 Performance analysis with different values of Gamma in RBF kernel

It is observed that the predictive capacity of the SVR model with the RBF kernel is significantly improved when the
parameter Gamma (G) = 1 is compared to other G values. The shift (increase) in the mean causes the smoothing constant
(λ) to increase respectively, resulting in a decrease in average run lengths (ARLs). For G = 1, when there is a mean shift
of 0.1, 0.2, 0.25, 0.50, 0.75, 1.00, 1.5, 2.00, and 2.5, the corresponding ARLs are 204.88, 96.11, 71.32, 26.06, 15.13, 10.97, 6.82,
5.03, and 4.11 as it is obvious from Table 3 and Figure 4. When the value of G is set to 1, the control chart exhibits increased
sensitivity to minor shifts. However, for substantial shifts, the control chart maintains consistent efficiency across all
Gamma values.
 10991638, 2024, 7, Downloaded from https://onlinelibrary.wiley.com/doi/10.1002/qre.3603 by Muhammad Noor-ul-Amin - COMSATS Inst of Info Techn - I , Wiley Online Library on [11/11/2024]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
3840 KAZMI and NOOR-UL-AMIN

6 COMPARATIVE STUDY

The focus of this section is to contrast the adaptive control chart that is suggested using SVR with the AEWMA—I control
chart suggested by Noor et al.8 and the AEWMA—II control chart designed by Haq and Khoo.21 Noor et al.8 is a recent
paper on an adaptive control chart in the context of a parameter-free version. ARLs and SDRLs for previously suggested
control charts as well as the current ones are presented in Table 4. The proposed control chart using radial basis kernel is
more efficient than its competitor control chart. The efficiency of the newly developed control chart surpasses that of its
competitor for small shifts (less than 0.50) and performs equally well for shifts greater than or equal to 0.50. For instance,
when the shift δ is less than 0.50, ARLs for the proposed control chart are 370.9, 204.88, 136.31, 96.11, and 71.32, while
those for the existing control chart AEWMA—I are 369.87, 217.49, 145.18, 101.48, and 73.57 and for AEWMA—II are 370.83,
221.35, 148.61, 102.57, and 75.02, respectively. Moreover, for shifts δ ranging from 0.50 to 4, the ARLs of the proposed control
chart are 26.06, 15.13, 10.97, 6.82, 5.03, 3.56, and 2.93, compared to the ARLs of the existing control chart, which are 26.48,
15.41, 10.70, 6.77, 4.97, 3.34, and 2.57.

7 REAL-LIFE APPLICATION

In this section, our focus will center on defining the contextual background of authentic data and explaining the procedural
methodology employed in implementing the proposed scheme.

7.1 Background of data

The pistons also known as reciprocating engines find extensive applications in compressors, cars, pumps, motors, and
automobiles, utilizing various natural fuels such as biogas, natural gas, and biodiesel. Functioning as moving components
within cylinders, these engines compress gas by tightly sealing it through the piston ring. A piston engine’s main goal is
to use a piston rod to transfer force from the gas that is expanding in the cylinder bore to the crankshaft. Every piston has
a bore, also known as a cylinder bore, whose diameter is used to determine its size. A number of variables, including the
cylinder bore, stroke, and the space between the piston and the cylinder, affect an engine’s displacement, which collectively
plays a critical role in engine performance. Hence, accurate measurement of the internal diameter of the cylinder bore is
crucial. To serve the objectives of our study, data on the cylinder bore diameters were extracted from research conducted
by Zaman et al.22 A total of 40 samples of size 1 were gathered, with measurements conducted at the precision of 1/10,000
of an inch, exemplified by values like 3.5205, 3.5202, and so forth, signifying the engine block’s cylinder bores’ internal
diameter. Only the final three digits of each measurement were kept in order to make comparisons easier. For instance,
from the measurement 3.5205, only 205 was recorded, as the first two digits (3.5) are consistent across all data points. It
is important to note that the original data underwent standardization prior to its utilization for the intended purposes in
the study.

7.2 Control chart application

√ bore diameters, where i = 1,2.3,. . . ,40.

a) Start with Xi values from the real data of cylinder
̄
b) Transform the raw data into 𝑍𝑖 = (𝑥𝑖 − 𝑥)∕(𝜎 𝑥 ∕ 𝑛) score because we have trained the SVR model for the standardized
data.
c) Predict the value of the smoothing constant for the proposed charting statistic SAEWMAt by using the trained SVR
model as suggested in Section 4.
d) Find the statistic as described in Equation (3). Plot the statistics of the control chart by following the proposed design.

Figures 5 and 6 clearly show that during the first 30 samples, the control chart stays stable, indicating that the process
is in an in-control state. The control charts demonstrate that in the next 10 samples, the process mean shifts upward.
Notably, out-of-control conditions are detected by the proposed SAEWMA control chart earlier than the AEWMA—II
control chart proposed by Haq and Khoo,21 that is, SAEWMA and AEWMA—II charts trigger out-of-control signals at
the 32nd and 38th samples, respectively. The data indicates that, on average, the suggested chart demonstrates greater
sensitivity compared to the current chart.
 10991638, 2024, 7, Downloaded from https://onlinelibrary.wiley.com/doi/10.1002/qre.3603 by Muhammad Noor-ul-Amin - COMSATS Inst of Info Techn - I , Wiley Online Library on [11/11/2024]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
KAZMI and NOOR-UL-AMIN 3841

FIGURE 5 Parameter-free SAEWMA control chart for cylinder bore diameters.

FIGURE 6 AEWMA—II control chart for cylinder bore diameters.

8 CONCLUSIONS AND FUTURE WORK

This paper presents a new method for efficiently tracking changes in the process mean: the SVR-based EWMA control
chart. The charting statistics are derived by employing the smoothing constant, predicted through the training of the SVR
model. A thorough examination is conducted on the SVR model, with an emphasis on how various kernels affect the
suggested control chart’s performance in particular. The effectiveness of the proposed SAEWMA chart is assessed, and
numerical results (ARLs) are compared with the alternatives that are currently available. The findings show that in terms
of ARL performance, the SAEWMA chart performs better than the AEWMA chart. Furthermore, a practical application
involving cylinder bore diameter measurements is presented to illustrate the real-life application of the SAEWMA chart.
The outcomes from this application show the effectiveness of the proposed methodology in a practical industrial setting.
Future endeavors could explore enhancing the performance of the SAEWMA chart through the integration of alterna-
tive machine-learning methods. Additionally, exploring different kernel functions, not covered in this study, presents a
fascinating path for further research, to be addressed in subsequent works.

AU T H O R CO N T R I B U T I O N S
Muhammad Waqas Kazmi: Research Design; methodology; analysis; writing—original draft preparation. Muham-
mad Noor-ul-Amin: Conceptualization; supervision; methodology review; validation; writing—reviewing and editing.
 10991638, 2024, 7, Downloaded from https://onlinelibrary.wiley.com/doi/10.1002/qre.3603 by Muhammad Noor-ul-Amin - COMSATS Inst of Info Techn - I , Wiley Online Library on [11/11/2024]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
3842 KAZMI and NOOR-UL-AMIN

AC K N OW L E D G M E N T
This study was conducted without external funding.

D A T A AVA I L A B I L I T Y S T A T E M E N T
The data that support the findings of this study are available as supplementary material accompanying this article.

ORCID
Muhammad Noor-ul-Amin https://orcid.org/0000-0003-2882-221X

REFERENCES
1. Montgomery DC. Statistical Quality Control. Vol 7. Wiley; 2009.
2. Page ES. Continuous inspection schemes. Biometrika. 1954;41(1/2):100-115.
3. Roberts SW. Control chart tests based on geometric moving averages. Technometrics. 1959;42(1):97-101.
4. Haq A, Gulzar R, Khoo MB. An efficient adaptive EWMA control chart for monitoring the process mean. Qual Reliab Eng Int.
2018;34(4):563-571.
5. Sparks RS. CUSUM charts for signalling varying location shifts. J Qual Technol. 2000;32(2):157-171.
6. Capizzi G, Masarotto G. An adaptive exponentially weighted moving average control chart. Technometrics. 2003;45(3):199-207.
7. Huber P. Robust Statistics. John Wiley and Sons; 1981.
8. Noor-ul-Amin M, Noor S, Feroze N. A parameter free Bayesian adaptive EWMA mean chart under different loss functions. Qual Reliab
Eng Int. 2022;38(4):2156-2169.
9. Cuentas S, Peñabaena-Niebles R, Garcia E. Support vector machine in statistical process monitoring: a methodological and analytical
review. Int J Adv Manuf Technol. 2017;91:485-500.
10. He S, Jiang W, Deng H. A distance-based control chart for monitoring multivariate processes using support vector machines. Ann Oper
Res. 2018;263:191-207.
11. Shao YE, Hu YT. Using machine learning classifiers to recognize the mixture control chart patterns for a multiple-input multiple-output
process. Mathematics. 2020;8(1):102.
12. Apsemidis A, Psarakis S, Moguerza JM. A review of machine learning kernel methods in statistical process monitoring. Comput Ind Eng.
2020;142:106376.
13. Chowdhury NR, Janan F. Control chart pattern recognition: a comparison between statistical correlation measure and support vector
machine (svm). In Proceedings of the 5th NA International Conference on Industrial Engineering and Operations Management Detroit,
pp. 10-14, 2020.
14. Pieters L, Malela-Majika JC, Human S, Castagliola P. A new multivariate extended homogeneously weighted moving average monitoring
scheme incorporated with a support vector machine. Qual Reliab Eng Int. 2023;39(6):2454-2475.
15. Riaz M, Zaman B, Raji IA, Omar MH, Mehmood R, Abbas N. An adaptive EWMA control chart based on principal component method to
monitor process mean vector. Mathematics. 2022;10(12):2025.
16. Noor-ul-Amin M, Khan I, Alzahrani ARR, Ayari-Akkari A, Ahmad B. Risk-adjusted EWMA control chart based on support vector machine
with application to cardiac surgery data. Sci Rep. 2024;14(1):9633.
17. Zhang F, O’Donnell LJ. Support vector regression. Machine Learning. Academic Press; 2020:123-140.
18. Gani W, Taleb H, Limam M. Support vector regression based residual control charts. J Appl Statist. 2010;37(2):309-324.
19. Sarwar MA, Noor-ul-Amin M. Design of a new adaptive EWMA control chart. Qual Reliab Eng Int. 2022;38(7):3422-3436.
20. Jiang W, Shu L, Apley DW. Adaptive CUSUM procedures with EWMA-based shift estimators. IIE Trans. 2008;40(10):992-1003.
21. Haq A, Khoo MBCMichael. A parameter-free adaptive EWMA mean chart. Qual. Technol. Quant. Manag. 2020,17(5):528-543.
22. Zaman B, Lee MH, Riaz M, Abujiya MR. An adaptive approach to EWMA dispersion chart using Huber and Tukey functions. Qual Reliab
Eng Int. 2019;35(6):1542-1581.

S U P P O RT I N G I N F O R M AT I O N
Additional supporting information can be found online in the Supporting Information section at the end of this article.

How to cite this article: Kazmi MW, Noor-ul-Amin M. Adaptive EWMA control chart by using support vector
regression. Qual Reliab Eng Int. 2024;40:3831–3843. https://doi.org/10.1002/qre.3603
 10991638, 2024, 7, Downloaded from https://onlinelibrary.wiley.com/doi/10.1002/qre.3603 by Muhammad Noor-ul-Amin - COMSATS Inst of Info Techn - I , Wiley Online Library on [11/11/2024]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
KAZMI and NOOR-UL-AMIN 3843

AU T H O R B I O G R A P H I E S

Muhammad Waqas Kazmi is pursuing a PhD degree in Statistics at COMSATS University Islam-
abad, Lahore Campus. He holds a Master’s in Statistics from the Department of Statistics, Faculty
of Sciences, University of Sargodha. His research primarily focuses on statistical process control,
artificial neural networks, machine learning, and deep learning.

Muhammad Noor-ul-Amin received his PhD degree from NCBA and E, Lahore, Pakistan. He is
currently working as an Associate Professor at COMSATS University Islamabad-Lahore Campus.
His research interests include Sampling methods and Statistical Quality Control.