A machine learning assisted truncated adaptive EWMA chart for robust process monitoring Authors: Abdullah A. Zaagan, Hassan M. Aljohani, Hamad Zogan, Muhammad Nabi Published: 2025 ABSTRACT The classical memory-type control charts, Adaptive EWMA (AE) and Truncated Adaptive EWMA (TAE) are chart types whose analytical properties are traditionally derived under normality assumptions, although they can be applied to non-normal data and operate on truncating the impact of extreme values. However, truncation-based methods can possibly fail to respond appropriately to genuine changes in the process. The research is based on a new Hybrid Machine Learning-assisted Truncated Adaptive EWMA (ML-TAE) chart which involves the use of anomaly scores associated with the machine learning models that adjustively weight each new observation before it is added to the EWMA statistic. Monte Carlo method is usedto evaluate the suggested chart in range of scenarios of contamination and varying shifts. According to the given results, the ML-TAE chart shows a consistent better performance in the stability of Average Run Length (ARL) as it keeps the rates of false-alarm lower and allows detecting a moderate shift much faster than the classical models. The validity of the practical use of the proposed method is shown by a real life case study as well regarding contaminated process environment.Supplementary InformationThe online version contains supplementary material available at 10.1038/s41598-025-33073-9. Introduction Outliers are data points that significantly deviate from the majority of observations, appearing unusually small or large relative to the rest of the data. While some outliers arise from explainable sources such as instrument malfunction, faulty measurements, or data entry errors and should be corrected or removed, others may represent meaningful information about abnormal system behavior, process instability, or impending failures1,2. Their detection and analysis are therefore critical for improving process reliability and ensuring timely intervention. Outliers are often identified through residuals that are unusually large in absolute value, typically exceeding four or more standard deviations from the mean. Control charts remain one of the most powerful and widely applied tools in Statistical Process Control (SPC). They find application both in manufacturing and service industries to observe process behavior, and to indicate loss of statistical control, thus making it possible to identify possible problems with quality early. Memory-type control charts, e.g. the exponentially weighted moving average (EWMA) chart with Roberts3 and the cumulative sum (CUSUM) chart with Page4, is most effective due to its greater sensitivity to minor and intermediate shifts. Such changes, however minimal, can add up in the course of time to bring about an entire loss of finances or even a collapse in the course of the processes. This has reportedly led memory-type charts to be used in a variety of other fields such as healthcare, fisheries, image and signal processing and public health surveillance5–7. The presence of outliers, however, can severely degrade the performance of traditional control charts by inflating false-alarm rates. To address this challenge, researchers have proposed robust charting methods that employ resistant estimator of process location and scale, particularly when subgroup data are available8–12. Hawkins2, for instance, proposed a winsorized CUSUM chart to reduce the influence of extreme observations on monitoring performance. Nevertheless, there have been relatively lesser studies on strong charts to be used in individual observations even though they are pertinent in most contemporary applications. More recent research has further developed the robust process monitoring field with the development of control charts that work efficiently in the deviation of models, and in the presence of data contamination. To give an example13, developed an efficient and strong CUSUM chart of mean monitoring that has been applied in chemical process control, and14,15 came up with adaptive monitoring methods that do not deteriorate accuracy when there is uncertainty in the model assumptions16 also applied high breakdown point estimators to increase the stability of the process on contaminated healthcare data, which helps to understand the further demand of powerful statistical instruments in the real world. On the same note14,15, designed a nonparametric Hampel score-function-adaptive EWMA chart, which is useful in alleviating non-normality of manufacturing processes. The latter contributions provide emphasis on the growing interest in integrating strength and flexibility in control chart design, which is consistent with the driving force behind the current ML-TAE framework. Adaptive control charts have been adapted to adjust automatically to changes in sensitivity in situations where the degree of change in the process is not known. The grounds of adaptive designs were laid by Lucas and Crosier17 who suggested that standardized observations above and below ± 4.0 may be viewed as possible outliers. Expanding on this concept, Capizzi and Masarotto18 suggested AE. chart, a dynamically-smoothed version that is sensitive to a variety of shift sizes. These concepts have been extended later, with adaptive CUSUM procedures suggested by19,20, and an adaptive-parameter-free EWMA (EAE) chart suggested by21, which is even more efficient at detection. The22 have recently created the truncated normal-based EWMA (TE) chart which is stronger as it employs truncated values rather than extreme ones to rationalize the run-length distribution. Nonetheless, preliminary experiments suggest, that charts, such as the EAE, are sensitive to contamination and the false-alarm level rises tremendously when there are outliers. This promotes development of more powerful strategies which are sensitive but not greatly influenced with radical values. Over the past years, the use of machine learning (ML) in Statistical Process Control (SPC) has increasingly been addressed due to the ability of machine learning to describe nonlinear trends, detect nuanced variations, and deal with high-dimensional or noisy process data. Various ML algorithms have been integrated to improve in different areas of the performance of control charts. As an example, the Support Vector Regression (SVR) has been used to predict changes in means and dynamically adjust control limits23. Similarly, the group techniques like the Random Forests and Gradient Boosting have been applied in the classification of process states and enhanced fault-detection in the complex industrial systems. Autoencoders and Artificial Neural Networks (ANNs) have demonstrated special potential in the study of nonlinear structures and the reconstruction of normal patterns of the processes, thus allowing detecting abnormal deviations at an early stage24. Moreover, the unsupervised models such as the Isolation Forest and the One-Class SVM have been found useful in anomaly detection and have provided adaptive measures to limit the effect of data contamination to monitoring performance. Even though these hybrid SPC-ML models have achieved significant reductions in false alarms and improvement in responsiveness to medium-speed process changes, the vast majority of them are relying on fixed thresholds, or offline training. This reduces their flexibility to change as conditions of the processes change with time. Also, though models using deep learning (e.g. Autoencoder) can be more informative, they are less interpretable and are often not reliable without using large datasets. Thus it is evident that there is a necessity to come up with a solution that will be able to combine strength against outliers, adaptive responsiveness and computational efficiency to online monitoring. As a solution to these difficulties, the current study suggests a control chart of Machine Learning-Assisted Truncated Adaptive EWMA (ML-TAE). The approach uses scores of anomaly which are the results of the unsupervised ML algorithms to keep changing the weight attributed to new process observations. By this combination, the proposed chart achieves a good balance between stability and sensitivity that allows it to have a reliable detection performance when data is contaminated.Based on such developments the current study proposes a ML –TAE control chart. The chart incorporates ML obtained anomaly scores which continuously tune the weights of the observations before inclusion in the EWMA statistic. This approach aims to find an excellent compromise between robustness and responsiveness by minimizing the contribution of extreme changes and maintaining the ability to discern the true process changes. The efficiency of the offered ML-TAE chart is strictly checked by means of the Monte Carlo simulation studies and one practical example of the IDA, which implies the demonstration of its practical interest. The remainder of this paper is structured as follows. Sect. “Truncated adaptive EWMA control chart” reviews the background of AE and truncated EWMA charts, highlighting their strengths and limitations in the presence of outliers. Sect. “Methodology of ML-TAE control chart” presents the design of the proposed ML-TAE control chart, including its statistical formulation and weighting mechanism based on anomaly scores. Sect. “Simulation design and validation” describes the Monte Carlo simulation study used to evaluate the in-control and out-of-control performance of the proposed chart under various contamination levels and shift scenarios. Sect. “Run length computation and evaluation” gives a practical application of the ML-TAE chart using real data. Lastly, Sect. “Comparative analysis” will tend to summarize the primary findings, address the practical implications, and outline the possible future research directions. Truncated adaptive EWMA control chart The methodology is built upon the framework proposed by Haq (2024) that creates a truncated adaptive EWMA (TAE) control chart to track the mean of a normally distributed process that is polluted by infrequent outliers.Let \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$X_{t}^{{}}$$\end{document} denote an individual observation at timet, assumed to follow a normal distribution.Observations are standardized as \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$X_{t}^{*}$$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$Y_{t}^{{}}$$\end{document} is defined as\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$Y_{t} = \left\{ \begin{gathered} sign\left( {X_{t}^{*} } \right)h,\,\,\,\,\,\,\,if\left| {X_{t}^{*} } \right|> u_{t} \hfill \\ X_{t}^{*} ,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,otherwise \hfill \\ \end{gathered} \right.$$\end{document}where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$u_{t}$$\end{document} is a user-specified outlier cut-off. The TAE statistic updates according to Update the monitoring statistic as given by.\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$E_{t} = \zeta (Y)Y_{t} + \left( {1 - \zeta (Y)} \right)E_{t - 1} ,E_{0} = 0$$\end{document} Where, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\zeta ( \cdot )$$\end{document} is a stepwise score function controlling the smoothing level. Following Haq and Khoo21, four smoothing levels (0.05, 0.25, 0.50, 1.00) are used. An out-of-control (OOC) signal is triggered whenever: \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\left| {E_{t} } \right|> h,$$\end{document} the h is the predefined threshold value for a specified ARL0. Methodology of ML-TAE control chart This section presents the formulation and operational structure of the proposed Machine Learning–Assisted Truncated Adaptive EWMA (ML-TAE) control chart. We propose a ML- TAE chart for monitoring the mean of a normal process in the presence of outliers. The method enhances the robustness of memory-type charts in contaminated environments by integrating anomaly-score weighting, adaptive truncation, and adaptive smoothing. The goal is to protect the monitoring statistic from the influence of extreme values while preserving sensitivity to genuine process shifts.Although this truncation method does offer certain defense against contamination, it can have two important shortcomings: (i) failing to respond to the actual process changes where the corresponding observations are truncated and (ii) responding to harmless fluctuations that do not reflect actual process change. Conceptual motivation Classical types of EWMA charts are usually formulated under normality when obtaining control limits, although they do not assume normal data, hence, are very sensitive to extreme outlier values. As contamination develops, a single outlier will inflate the EWMA statistic at a high rate leading to excessive false alarms. The classic truncation-based methods like the TAE chart22 and Haq, 2024 constrain the size of the extreme values, but might erroneously smother the process changes in case the truncation is overly fierce. The ML-TAE model overcomes these weaknesses through the application of machine-learned anomaly scores, which are used to determine the validity of every incoming observation. The weighting of each new value based on the probability of it being an outlier is used in place of magnitude based truncation. The resulting monitoring statistic obtained through this soft weighting, then adaptive truncation and adaptive smoothing is robustly and responsively sensitive. The list of abbreviation is given in the Table B1 of Appendix B. Anomaly scoring using machine learning Let \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$X_{t}$$\end{document} denote the observation at time t. Prior to online monitoring, an unsupervised anomaly detection algorithm is trained using Phase-I standardized in-control data. In this study, an Isolation Forest (IF) model is used due to its ability to isolate extreme observations through recursive partitioning.The IF model contains three main hyperparameters, chosen based on standard guidelines and sensitivity analysis and described in Table A2 of Appendix A. Given a trained model, each incoming observation \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$X_{t}$$\end{document} generates an anomaly score.1\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$s_{t} = IF\left( {X_{t} } \right)$$\end{document}where larger values of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$s_{t}$$\end{document} indicate a higher likelihood of contamination. To convert the score into a probabilistic measure, we apply a logistic transformation:2\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$w_{t} = \frac{1}{{1 + \exp \left\{ {k\left( {s_{t} - \tau } \right)} \right\}}},$$\end{document}wherek > 0 controls the steepness of the logistic function and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\tau$$\end{document} represents the anomaly threshold (e.g., the 95th percentile of Phase-I scores). The resulting weight \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$w_{t}$$\end{document} ∈ (0,1] modulates how much influence the new value should have in the EWMA recursion.Instead of directly using the raw value \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$X_{t}$$\end{document}, the ML-TAE chart forms a robust substitute, i.e.,3\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$X_{t}^{ * } = w_{t} X_{t}$$\end{document} The weighted values may still remain extreme when contamination is severe. To further protect the statistic, the truncation mechanism used in the ML-TAE chart is inspired by the winsorization strategy of2, the truncated EWMA design of22, and classical robust M-estimators. When the weighted observation \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$X_{t}^{*}$$\end{document} exceeds a data-adaptive truncation limit \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$u_{t}$$\end{document}, the value is replaced by \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$sign\left( {X_{t}^{*} } \right)h$$\end{document}. This rule ensures that extremely large values which typically arise from contamination cannot dominate the EWMA recursion, while the direction of the deviation is still preserved to maintain sensitivity to genuine process shifts.4\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$Y_{t} = \left\{ \begin{gathered} sign\left( {X_{t}^{*} } \right)h,\,\,\,\,\,\,\,if\left| {X_{t}^{*} } \right|> u_{t} \hfill \\ X_{t}^{*} ,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,otherwise \hfill \\ \end{gathered} \right.$$\end{document}where, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$u_{t} = u_{0} \left( {1 + c\left( {1 - w_{t} } \right)} \right)$$\end{document}. Here, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$u_{0}$$\end{document} = 5is the baseline truncation threshold, and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$c> 0$$\end{document} that controls how the limit expands when \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$w_{t}$$\end{document} is small (i.e., when the observation appears abnormal). Replacing extreme values with ± h preserves the directional information of the shift while preventing extreme magnitude from destabilizing the monitoring statistic. This is in line with the philosophy of22 who demonstrated that truncation enhances stability of charts in the presence of contamination. Adaptive smoothing parameter and plotting statistic The ML-TAE chart continuously varies the smoothing parameter to become more responsive when the incoming data point looks normal and less sensitive when it looks anomalous. Set a time-varying smoothing parameter:5\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\lambda_{t} = \lambda_{\min } + \left( {\lambda_{\max } - \lambda_{\min } } \right)\left( {1 - w_{t} } \right).$$\end{document} The adaptive smoothing parameter \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\lambda_{t}$$\end{document} is designed to respond dynamically to the reliability of each incoming observation through the anomaly-score weigh. It is constructed according to 0 < \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\lambda_{\min }$$\end{document} < \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\lambda_{\max }$$\end{document} < 1, which means that the statistic is stable, and there is permissible adaptively.When an observation is considered normal, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$w_{t}$$\end{document}≈1,satisfies its weight, therefore, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\lambda_{t}$$\end{document}≈\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\lambda_{\min }$$\end{document}.The behavior of the chart, in this instance, is like a classical EWMA with a small value of the smoothing parameter making it a slow reacting statistic which effectively filters the noise and not over-reacts to small fluctuations. On the other hand, as the weight \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$w_{t}$$\end{document} goes down which means higher probability of change in the process or contamination, the smoothing parameter approaches \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\lambda_{\max }$$\end{document}.This causes the statistic to be more sensitive and responsive to be able to respond to meaningful changes, and it also protects the statistic against extreme values because of the truncation mechanism. This adaptive behavior balances robustness and sensitivity, enabling the ML-TAE chart to perform effectively under both clean and contaminated environments. Update the monitoring statistic as given by.6\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$C_{t} = \lambda_{t} Y_{t} + \left( {1 - \lambda_{t} } \right)C_{t - 1} ,C_{0} = 0$$\end{document} An out-of-control (OOC) signal is triggered whenever:7\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\left| {C_{t} } \right|> h,$$\end{document}where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$h> 0$$\end{document} is the control limit chosen to achieve a desired in-control \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$ARL_{0}$$\end{document}. Since closed-form variance is not available due to adaptively, h is determined via Monte Carlo Simulation under in-control conditions. Simulation design and validation The effectiveness of the proposed ML-TAE chart was assessed through Monte Carlo simulations. The control limit, h = 0.41730, was chosen to achieve an in-control ARL₀ of roughly 500 under nominal conditions, based on preliminary trials. For every combination of process mean shift (δ) and contamination proportion (P), 10,000 independent simulation runs were performed to ensure reliable estimates. The stability of the ARL results was confirmed through convergence checks, and fixed random seeds were applied to allow reproducibility. Out-of-control ARLs (ARL₁) were calculated for δ values ranging from 0.1 to 6.0 across different contamination levels, and the corresponding standard deviations of the run lengths (SDRLs) were computed to measure variability.The anomaly-score component of the ML-TAE chart was generated using an unsupervised Isolation Forest model. The anomaly score used in our study was the Isolation Forest anomaly score generated by the isotree package (Liu et al., 2008). This score ranges between 0 and 1, with higher values indicating more anomalous observations. It was chosen because Isolation Forest is widely recognized for its ability to isolate outliers efficiently in high-noise environments, and its score integrates naturally with the weighting framework of the ML-TAE chart. The model was trained using the Phase-I in-control standardized data. Our nominal 5% false-alarm threshold of 100 trees, sub-sampling of 256 observations per tree, and a contamination parameter range of [0.01 0.05] was used to arrive at a nominal 5% false-alarm.The anomaly level was adopted as the 95 th percentile of the in-control score. All anomaly scores were converted to probabilities through an empirical distribution estimated from Phase-I data. The probability was then transformed into a time-varying weight that directly affected the EWMA update. After defining the ML-TAE statistic, ARL and SDRL values were computed using Monte Carlo simulation over 10,000 independent replications until a signal occurred.All simulated observations were generated sequentially using the following rules: Phase-I (training data): Drawn from a standard normal distribution representing an in-control process without contamination. Phase-II (monitoring data): i.Regular observations followed a normal distribution with the same standard deviation but with a shifted mean to represent the specified out-of-control condition.ii.With a small predefined probability, an observation was replaced by an outlier. Outliers were defined as extreme values generated uniformly from a specified interval (“outlier range”), representing large unexpected deviations that contaminate the process.The outlier probability P specifies the chance that any individual observation is contaminated. In our simulations, for instance, if we set P=0.1, meaning that on average, 10% of observations are outliers. Because the monitoring sequence stops once the chart signals, the actual number of outliers per run is random and may differ from 10%. i.Regular observations followed a normal distribution with the same standard deviation but with a shifted mean to represent the specified out-of-control condition. Regular observations followed a normal distribution with the same standard deviation but with a shifted mean to represent the specified out-of-control condition. ii.With a small predefined probability, an observation was replaced by an outlier. Outliers were defined as extreme values generated uniformly from a specified interval (“outlier range”), representing large unexpected deviations that contaminate the process.The outlier probability P specifies the chance that any individual observation is contaminated. In our simulations, for instance, if we set P=0.1, meaning that on average, 10% of observations are outliers. Because the monitoring sequence stops once the chart signals, the actual number of outliers per run is random and may differ from 10%. With a small predefined probability, an observation was replaced by an outlier. Outliers were defined as extreme values generated uniformly from a specified interval (“outlier range”), representing large unexpected deviations that contaminate the process.The outlier probability P specifies the chance that any individual observation is contaminated. In our simulations, for instance, if we set P=0.1, meaning that on average, 10% of observations are outliers. Because the monitoring sequence stops once the chart signals, the actual number of outliers per run is random and may differ from 10%. These details ensure transparency in how both normal observations and outliers are generated in the simulation. The following steps describe how the Monte Carlo simulations were conducted to estimate the performance of the proposed ML-TAE chart combined with an Isolation Forestbased anomaly detection mechanism.Step 1: Training the isolation forest model Step 1: Training the isolation forest model Step 1: Training the isolation forest model An Isolation Forest model was fitted using the Phase-I data.The model learned the typical structure of the process and provided an anomaly score for any new observation.This score ranged from low (normal observation) to high (potential outlier).Step 2: Anomaly scoring using the isolation forest Step 2: Anomaly scoring using the isolation forest Step 2: Anomaly scoring using the isolation forest Each new observation was passed through the trained Isolation Forest model as discussed in (1).The model returned an anomaly score indicating the likelihood of the observation being abnormal.The hyperparametersof IF model are given in Table A3 of Appendix A.Step 3: Conversion of anomaly score into a weight Step 3: Conversion of anomaly score into a weight Step 3: Conversion of anomaly score into a weight The anomaly score was transformed into a weight between zero and one using a logistic-type mapping as given in (2).A weight near one indicated a normal observation, while a weight near zero indicated a highly suspicious one.Step 4: Robust substitution of the observation Step 4: Robust substitution of the observation Step 4: Robust substitution of the observation The raw observation was adjusted using the computed weight.Highly suspicious observations were down-weighted more strongly, reducing their influence on the monitoring statistic by using the (3).Step 5: Adaptive truncation mechanism Step 5: Adaptive truncation mechanism Step 5: Adaptive truncation mechanism A dynamic threshold was computed based on the weight.If the adjusted observation exceeded this threshold, it was replaced by a fixed, bounded value as discussed in (4).Step 6: Adaptive smoothing parameter Step 6: Adaptive smoothing parameter Step 6: Adaptive smoothing parameter The smoothing constant used in the updating step was adjusted adaptively.Normal observations used a smaller smoothing value, while suspicious observations used a larger one.This adaptive structure enabled the statistic to react appropriately depending on data quality by using the (5).Step 7: Updating statistic and run length profile Step 7: Updating statistic and run length profile Step 7: Updating statistic and run length profile The monitoring statistic was updated by using the (6). After each update, the statistic was compared with the control limit.If the limit was exceeded, the run length for that simulation was recorded and the run was terminated.If the limit was not exceeded, the simulation continued until the maximum allowed length.The entire sequence described above was repeated for a large number of replications. After all replications, the average run length (ARL) and standard deviation of the run length (SDRL) were computed.These metrics were used to evaluate the performance of the proposed ML-TAE chart under the specified conditions. Note that the Steps 1–7 are repeated 10,000 times to obtained the vlaue of h for a specified value of ARL0 and the process is done again with a changed value of h if the desired ARL0 is not met. Run length computation and evaluation The performance of the proposed TAE chart under varying levels of outliers has been examined through zero-state average run length (ZS ARL) profiles. An in-control ARL of 500 (false-alarm rate ≈ 0.002) was chosen because this level is widely used in recent robust and adaptive SPC studies10. These studies demonstrate that higher in-control ARL values are especially appropriate in contaminated or heavy-tailed environments, where conventional EWMA or CUSUM charts may generate excessive false alarms. The Tables 1, 2, 3 and 4 summarize the results for different standardized mean shifts \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$(\delta )$$\end{document} when outliers are generated across four ranges: (5, 7), (7, 9), (9, 11), and (11, 13), under varying proportions of outliers (P). Several important observations can be drawn from these findings.Table 1ARL of the proposed chart of the various mean shifts and various level of contamination at Outliers range = 5,7.h = 0.41730δP = 0.005P = 0.01P = 0.02P = 0.03P = 0.05P = 0.08P = 0.10.0ARL500.83503.87519.47535.04573.58612.62657.21SDRL485.17498.03498.76527.27555.11597.41638.910.1ARL274.01274.81283.80293.41312.44336.86350.48SDRL256.27250.88268.57270.82295.96319.28328.960.2ARL119.08118.24123.76126.33133.94142.48149.08SDRL102.19102.64105.62108.29119.74123.12130.540.5ARL28.8929.0529.7130.2131.2032.9333.91SDRL16.7517.0417.0617.5818.6919.9921.180.75ARL16.4216.5916.8517.0417.4018.4318.95SDRL7.417.567.507.978.028.849.081.00ARL11.5011.5511.6011.7912.1112.6712.97SDRL4.444.304.504.534.705.015.171.50ARL7.107.137.317.357.457.807.99SDRL2.052.062.222.242.282.442.572.5ARL4.214.254.304.324.424.624.68SDRL0.930.941.011.0381.111.251.263ARL3.643.653.703.743.823.984.02SDRL0.750.7680.800.850.931.031.083.5ARL3.343.383.403.4603.523.663.73SDRL0.730.730.770.800.870.971.004ARL3.3883.433.463.513.613.7103.78SDRL0.910.950.981.001.091.161.225ARL4.684.754.7164.844.895.0435.23SDRL2.12.092.092.192.252.362.48Table 2ARL of the proposed chart of the various mean shifts and various level of contamination at Outliers range = 7,9.h = 0.41730δCCP = 0.005P = 0.01P = 0.02P = 0.03P = 0.05P = 0.08P = 0.10.0ARL503.42518.12524.73538.10558.74625.95651.85SDRL494.92501.53519.33524.10539.30629.92624.730.1ARL275.27283.29282.80294.50307.98336.84354.42SDRL257.44256.95264.98277.80291.34326.47338.080.2ARL116.05121.73125.07124.89133.66142.47151.40SDRL98.41104.34109.89108.04116.50124.03131.010.5ARL28.5428.8229.9030.2731.1533.5533.82SDRL16.6216.7517.4617.9518.3620.6820.380.75ARL16.3516.4516.9716.9417.5818.2218.97SDRL7.167.257.787.658.008.309.291.00ARL11.3711.5211.7211.8012.0612.6513.05SDRL4.254.324.384.554.675.035.231.50ARL7.1407.187.227.337.607.807.98SDRL2.102.092.162.222.332.492.632.5ARL4.214.244.294.344.474.614.71SDRL0.930.930.981.031.111.231.293ARL3.623.643.703.723.823.944.02SDRL0.740.770.810.840.921.021.093.5ARL3.363.383.433.473.543.673.73SDRL0.7270.740.780.820.881.001.054ARL3.413.403.443.473.573.683.74SDRL0.930.940.960.981.061.141.235ARL4.644.734.694.834.925.105.21SDRL2.102.152.152.262.292.442.503Table 3ARL of the proposed chart of the various mean shifts and various level of contamination at Outliers range = 9,11.h = 0.41730δCCP = 0.005P = 0.01P = 0.02P = 0.03P = 0.05P = 0.08P = 0.10.0ARL505.54519.14521.67527.88561.24606.02649.47SDRL504.38503.34499.58527.43542.80592.15645.360.1ARL271.14279.00284.44294.78309.16330.97354.10SDRL251.80265.12274.30272.34287.33311.26331.440.2ARL117.81120.29122.84127.64131.24145.09151.19SDRL101.82100.83106.02107.33114.19123.37132.090.5ARL28.8429.4729.6430.5231.1932.9434.26SDRL16.6317.3417.0918.4318.5419.8821.260.75ARL16.5916.3116.7817.1217.4918.5518.83SDRL7.637.277.717.878.018.848.971.00ARL11.4211.5511.6011.7612.2112.6713.00SDRL4.214.354.374.434.725.115.271.50ARL7.137.157.297.327.557.807.96SDRL2.122.072.212.212.332.442.552.5ARL4.2054.264.324.334.434.594.70SDRL0.9170.940.991.031.1091.231.283ARL3.643.653.693.733.823.944.04SDRL0.740.750.800.850.931.011.093.5ARL3.363.373.413.463.523.663.74SDRL0.720.750.750.800.860.971.054ARL3.423.413.483.513.553.693.78SDRL0.930.920.981.021.051.151.205ARL4.644.664.764.814.895.125.23SDRL2.082.122.182.202.272.452.53Table 4ARL of the proposed chart of the various mean shifts and various level of contamination at Outliers range = 11,13.h = 0.41730δCCP = 0.005P = 0.01P = 0.02P = 0.03P = 0.05P = 0.08P = 0.10.0ARL498.51512.13524.76528.79567.86624.36669.83SDRL498.76496.75512.07523.74552.58615.74643.550.1ARL277.59280.68281.80298.67318.51332.17355.33SDRL262.48261.44264.19285.41295.78314.04339.440.2ARL118.33120.05122.16125.62132.90143.78155.43SDRL100.32101.58107.04108.77117.85127.57132.980.5ARL28.6729.1429.8730.3031.4032.9133.92SDRL16.5417.1217.2818.0518.6719.6320.790.75ARL16.4416.7116.8616.9017.3218.3618.79SDRL7.527.727.557.747.908.608.751.00ARL11.4211.5911.6811.7712.1212.5212.98SDRL4.194.314.454.474.635.055.151.50ARL7.147.187.267.327.527.798.01SDRL2.102.152.162.212.302.482.642.5ARL4.224.234.294.344.434.624.70SDRL0.920.930.991.021.091.221.313ARL3.633.643.663.743.813.944.04SDRL0.750.770.800.840.911.031.093.5ARL3.373.393.413.453.543.663.74SDRL0.730.740.770.820.890.971.074ARL3.403.423.453.493.573.693.75SDRL0.940.920.970.991.071.141.245ARL4.694.684.784.814.905.065.14SDRL2.132.132.182.222.302.392.42 Table 1ARL of the proposed chart of the various mean shifts and various level of contamination at Outliers range = 5,7.h = 0.41730δP = 0.005P = 0.01P = 0.02P = 0.03P = 0.05P = 0.08P = 0.10.0ARL500.83503.87519.47535.04573.58612.62657.21SDRL485.17498.03498.76527.27555.11597.41638.910.1ARL274.01274.81283.80293.41312.44336.86350.48SDRL256.27250.88268.57270.82295.96319.28328.960.2ARL119.08118.24123.76126.33133.94142.48149.08SDRL102.19102.64105.62108.29119.74123.12130.540.5ARL28.8929.0529.7130.2131.2032.9333.91SDRL16.7517.0417.0617.5818.6919.9921.180.75ARL16.4216.5916.8517.0417.4018.4318.95SDRL7.417.567.507.978.028.849.081.00ARL11.5011.5511.6011.7912.1112.6712.97SDRL4.444.304.504.534.705.015.171.50ARL7.107.137.317.357.457.807.99SDRL2.052.062.222.242.282.442.572.5ARL4.214.254.304.324.424.624.68SDRL0.930.941.011.0381.111.251.263ARL3.643.653.703.743.823.984.02SDRL0.750.7680.800.850.931.031.083.5ARL3.343.383.403.4603.523.663.73SDRL0.730.730.770.800.870.971.004ARL3.3883.433.463.513.613.7103.78SDRL0.910.950.981.001.091.161.225ARL4.684.754.7164.844.895.0435.23SDRL2.12.092.092.192.252.362.48 ARL of the proposed chart of the various mean shifts and various level of contamination at Outliers range = 5,7. Table 2ARL of the proposed chart of the various mean shifts and various level of contamination at Outliers range = 7,9.h = 0.41730δCCP = 0.005P = 0.01P = 0.02P = 0.03P = 0.05P = 0.08P = 0.10.0ARL503.42518.12524.73538.10558.74625.95651.85SDRL494.92501.53519.33524.10539.30629.92624.730.1ARL275.27283.29282.80294.50307.98336.84354.42SDRL257.44256.95264.98277.80291.34326.47338.080.2ARL116.05121.73125.07124.89133.66142.47151.40SDRL98.41104.34109.89108.04116.50124.03131.010.5ARL28.5428.8229.9030.2731.1533.5533.82SDRL16.6216.7517.4617.9518.3620.6820.380.75ARL16.3516.4516.9716.9417.5818.2218.97SDRL7.167.257.787.658.008.309.291.00ARL11.3711.5211.7211.8012.0612.6513.05SDRL4.254.324.384.554.675.035.231.50ARL7.1407.187.227.337.607.807.98SDRL2.102.092.162.222.332.492.632.5ARL4.214.244.294.344.474.614.71SDRL0.930.930.981.031.111.231.293ARL3.623.643.703.723.823.944.02SDRL0.740.770.810.840.921.021.093.5ARL3.363.383.433.473.543.673.73SDRL0.7270.740.780.820.881.001.054ARL3.413.403.443.473.573.683.74SDRL0.930.940.960.981.061.141.235ARL4.644.734.694.834.925.105.21SDRL2.102.152.152.262.292.442.503 ARL of the proposed chart of the various mean shifts and various level of contamination at Outliers range = 7,9. Table 3ARL of the proposed chart of the various mean shifts and various level of contamination at Outliers range = 9,11.h = 0.41730δCCP = 0.005P = 0.01P = 0.02P = 0.03P = 0.05P = 0.08P = 0.10.0ARL505.54519.14521.67527.88561.24606.02649.47SDRL504.38503.34499.58527.43542.80592.15645.360.1ARL271.14279.00284.44294.78309.16330.97354.10SDRL251.80265.12274.30272.34287.33311.26331.440.2ARL117.81120.29122.84127.64131.24145.09151.19SDRL101.82100.83106.02107.33114.19123.37132.090.5ARL28.8429.4729.6430.5231.1932.9434.26SDRL16.6317.3417.0918.4318.5419.8821.260.75ARL16.5916.3116.7817.1217.4918.5518.83SDRL7.637.277.717.878.018.848.971.00ARL11.4211.5511.6011.7612.2112.6713.00SDRL4.214.354.374.434.725.115.271.50ARL7.137.157.297.327.557.807.96SDRL2.122.072.212.212.332.442.552.5ARL4.2054.264.324.334.434.594.70SDRL0.9170.940.991.031.1091.231.283ARL3.643.653.693.733.823.944.04SDRL0.740.750.800.850.931.011.093.5ARL3.363.373.413.463.523.663.74SDRL0.720.750.750.800.860.971.054ARL3.423.413.483.513.553.693.78SDRL0.930.920.981.021.051.151.205ARL4.644.664.764.814.895.125.23SDRL2.082.122.182.202.272.452.53 ARL of the proposed chart of the various mean shifts and various level of contamination at Outliers range = 9,11. Table 4ARL of the proposed chart of the various mean shifts and various level of contamination at Outliers range = 11,13.h = 0.41730δCCP = 0.005P = 0.01P = 0.02P = 0.03P = 0.05P = 0.08P = 0.10.0ARL498.51512.13524.76528.79567.86624.36669.83SDRL498.76496.75512.07523.74552.58615.74643.550.1ARL277.59280.68281.80298.67318.51332.17355.33SDRL262.48261.44264.19285.41295.78314.04339.440.2ARL118.33120.05122.16125.62132.90143.78155.43SDRL100.32101.58107.04108.77117.85127.57132.980.5ARL28.6729.1429.8730.3031.4032.9133.92SDRL16.5417.1217.2818.0518.6719.6320.790.75ARL16.4416.7116.8616.9017.3218.3618.79SDRL7.527.727.557.747.908.608.751.00ARL11.4211.5911.6811.7712.1212.5212.98SDRL4.194.314.454.474.635.055.151.50ARL7.147.187.267.327.527.798.01SDRL2.102.152.162.212.302.482.642.5ARL4.224.234.294.344.434.624.70SDRL0.920.930.991.021.091.221.313ARL3.633.643.663.743.813.944.04SDRL0.750.770.800.840.911.031.093.5ARL3.373.393.413.453.543.663.74SDRL0.730.740.770.820.890.971.074ARL3.403.423.453.493.573.693.75SDRL0.940.920.970.991.071.141.245ARL4.694.684.784.814.905.065.14SDRL2.132.132.182.222.302.392.42 ARL of the proposed chart of the various mean shifts and various level of contamination at Outliers range = 11,13. First, when the process is in control \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$(\delta = 0.0)$$\end{document}, the ARL values are close to the nominal target of 500 in all cases. However, as the proportion of outlier’s increases, the in-control ARL gradually rises. For example, in Table 1, the ARL increases from 500.83 at P = 0.005 to 657.21 at P = 0.10. A similar trend is observed across Tables 2, 3 and 4, with in-control ARLs exceeding 600 when the contamination proportion is high. This indicates that the proposed chart maintains robustness by preventing an excessive increase in false alarms, even when the magnitude of outliers is relatively large.Second, the sensitivity of the chart to small shifts \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$(\delta = 0.1,0.2)$$\end{document} demonstrates stable behavior across different outlier ranges. Although the ARL values increase slightly with higher contamination, the change is modest and consistent. For instance, in Table 3 at δ = 0.2, the ARL rises from 117.81 at P = 0.005 to 151.19 at P = 0.10, which shows that the proposed chart retains its detection ability despite the presence of extreme data points.Third, as the shift magnitude increases \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$(\delta \ge 0.4)$$\end{document}, the ARL values decline rapidly toward the lower bound, showing the expected property of effective detection of moderate and large shifts. Across all outlier ranges, the ARL converges to values between 3 and 7 for \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$(\delta \ge 1.5)$$\end{document}, which demonstrate that the proposed chart remains highly responsive to meaningful process deviations. The standard deviation of the run length (SDRL) also decreases substantially as \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\delta$$\end{document} increases, which confirms that the detection becomes not only faster but also more consistent for larger shifts.Another noteworthy observation is the effect of the outlier magnitude itself. Comparing across the four tables, the ARL patterns remain fairly stable regardless of whether the outliers lie in (5, 7) or as extreme as (11, 13). For in-control conditions and small shifts, the rise in ARL with increasing P is slightly more pronounced when the outliers are larger, yet the differences remain minor. This implies that the proposed TAE chart is not only resistant to proportion of outliers but also in response to the level of extreme values. The ML-TAE is overall more robust and sensitive and demonstrates better performance in terms of contamination and shifts of different magnitudes. Overall, the results discussion shows that the proposed chart will have three strengths:i.The ARL is performing well in-control with its level almost at the desirable limit. Pollution, restricting false alarms.ii.Only marginally influenced in detection is consistent sensitivity to small and moderate shifts by outlier proportion or magnitude.iii.Consistency in behavior during outlier intervals and improved detection with the help of ML combination—ML-TAE version is also superior to TAE especially in the case of moderate and large shifts without compromising on its ability to withstand extreme values. i.The ARL is performing well in-control with its level almost at the desirable limit. Pollution, restricting false alarms. The ARL is performing well in-control with its level almost at the desirable limit. Pollution, restricting false alarms. ii.Only marginally influenced in detection is consistent sensitivity to small and moderate shifts by outlier proportion or magnitude. Only marginally influenced in detection is consistent sensitivity to small and moderate shifts by outlier proportion or magnitude. iii.Consistency in behavior during outlier intervals and improved detection with the help of ML combination—ML-TAE version is also superior to TAE especially in the case of moderate and large shifts without compromising on its ability to withstand extreme values. Consistency in behavior during outlier intervals and improved detection with the help of ML combination—ML-TAE version is also superior to TAE especially in the case of moderate and large shifts without compromising on its ability to withstand extreme values. These results prove the suggested ML-TAE chart as an efficient and sophisticated monitoring instrument in the settings in which the processes are likely to be contaminated by outliers. The combination of its strength against the frequency and magnitude of abnormal observations, and enhanced sensitivity related to the integration of machine learning makes it specifically applicable to real life settings where the irregularities of data are frequent. Comparative analysis To rigorously assess the performance of the proposed ML-TAE chart, Table 5 reports its ARL values alongside the traditional TAE chart for two outlier ranges, (5,7) and (7,9) under varying contamination levels P and a broad spectrum of mean shifts δ. Several important observations emerge from these results, further supported by the accompanying visual analyses.A comparison between the proposed ML-TAE chart and the classical TAE chart is presented using the results in Fig. 1 and the ARL values reported in Table 5. These results allow us to evaluate both charts under different shift magnitudes and levels of contamination. The patterns observed across the tables and the graphical trends in Fig. 1 point to three key findings.Table 5Comparison Table of Proposed ML-TAE Chart and TAE Charts for different mean shifts (δ), and outliers are generated in the range (5, 7) and (7, 9) under varying proportions of outliers (P).δProposed ML-TAE with h = (5,7)TAE Chart with h = (5,7)Proposed ML-TAE with h = (7,9)TAE Chart with h = (7,9)P = 0.005P = 0.01P = 0.05P = 0.1P = 0.005P = 0.01P = 0.05P = 0.1P = 0.005P = 0.01P = 0.05P = 0.1P = 0.005P = 0.01P = 0.05P = 0.10.0500.8503.8573.5657.2502.23506.0566.3647.2503.4518.1558.7651.8503.9505.8565.0645.90.528.8929.0531.233.9129.4429.6531.834.7828.5428.8231.1533.8229.4129.8631.834.790.7516.4216.5917.418.9516.6616.7617.819.1816.3516.4517.5818.9716.6516.7817.7419.181.011.511.5512.1112.9711.5411.5912.1913.0611.3711.5212.0613.0511.5411.5912.2313.11.57.17.137.457.997.117.137.477.967.147.187.67.987.097.147.487.962.05.235.35.525.815.065.075.315.635.265.265.525.885.055.085.315.642.54.214.254.424.683.83.823.994.234.214.244.474.713.83.834.04.233.03.643.653.824.022.932.953.083.253.623.643.824.022.942.953.083.253.53.343.383.523.732.352.362.462.63.363.383.543.732.352.362.472.64.03.3883.433.613.782.042.062.152.273.413.43.573.742.052.052.162.275.04.684.754.895.232.492.52.612.754.644.734.925.212.482.52.62.756.012.1212.2512.9513.356.766.827.117.5112.2212.1312.7913.576.86.837.077.51Fig. 1Absolute ARL differences at δ = 0 for different outlier range. Table 5Comparison Table of Proposed ML-TAE Chart and TAE Charts for different mean shifts (δ), and outliers are generated in the range (5, 7) and (7, 9) under varying proportions of outliers (P).δProposed ML-TAE with h = (5,7)TAE Chart with h = (5,7)Proposed ML-TAE with h = (7,9)TAE Chart with h = (7,9)P = 0.005P = 0.01P = 0.05P = 0.1P = 0.005P = 0.01P = 0.05P = 0.1P = 0.005P = 0.01P = 0.05P = 0.1P = 0.005P = 0.01P = 0.05P = 0.10.0500.8503.8573.5657.2502.23506.0566.3647.2503.4518.1558.7651.8503.9505.8565.0645.90.528.8929.0531.233.9129.4429.6531.834.7828.5428.8231.1533.8229.4129.8631.834.790.7516.4216.5917.418.9516.6616.7617.819.1816.3516.4517.5818.9716.6516.7817.7419.181.011.511.5512.1112.9711.5411.5912.1913.0611.3711.5212.0613.0511.5411.5912.2313.11.57.17.137.457.997.117.137.477.967.147.187.67.987.097.147.487.962.05.235.35.525.815.065.075.315.635.265.265.525.885.055.085.315.642.54.214.254.424.683.83.823.994.234.214.244.474.713.83.834.04.233.03.643.653.824.022.932.953.083.253.623.643.824.022.942.953.083.253.53.343.383.523.732.352.362.462.63.363.383.543.732.352.362.472.64.03.3883.433.613.782.042.062.152.273.413.43.573.742.052.052.162.275.04.684.754.895.232.492.52.612.754.644.734.925.212.482.52.62.756.012.1212.2512.9513.356.766.827.117.5112.2212.1312.7913.576.86.837.077.51 Comparison Table of Proposed ML-TAE Chart and TAE Charts for different mean shifts (δ), and outliers are generated in the range (5, 7) and (7, 9) under varying proportions of outliers (P). Absolute ARL differences at δ = 0 for different outlier range. First, for small and moderate shifts (δ ≤ 2.0), both charts behave almost identically. The ARL values in Tables 3, 4 and 5 show only minor differences between ML-TAE and TAE, and the curves in Fig. 1 also follow a very similar trajectory. This reveals that the anomaly-based weighting in ML-TAE does not impair its ability to track gradual departures from the target mean. Second, for larger shifts (δ ≥ 2.5), the classical TAE chart clearly detects shifts more quickly. Across all contamination levels, the ARL values for TAE are consistently smaller than those of ML-TAE, as shown in Table 5. This behaviour is also visible in Fig. 1 where the TAE curves drop more rapidly as the shift magnitude increases. This difference arises because ML-TAE down-weights extreme observations, some of which correspond to genuine process changes when the shift is large. As a result, ML-TAE becomes more conservative while the classical TAE chart responds more promptly. Third, despite being slightly slower for detecting large shifts, the ML-TAE chart demonstrates stronger robustness when contamination is present. As the proportion of outliers increases, the ARL patterns in Tables 5 show that ML-TAE maintains more stable performance compared to TAE. This stability is also visible in Fig. 1 where the ML-TAE curves fluctuate less across different contamination levels. This confirms that the weighting mechanism effectively mitigates the influence of isolated extreme values. In summary, Fig. 1 and Table 5 collectively show that ML-TAE and TAE perform similarly for small and moderate shifts, while the classical TAE chart is more efficient for detecting larger shifts. The ML-TAE chart, however, provides enhanced robustness under contamination and produces more consistent ARL behaviour when data irregularities are present. Thus, ML-TAE is preferable in applications where data contamination is expected, whereas TAE remains advantageous when rapid detection of substantial shifts is the primary objective. Application To demonstrate the practical performance of the proposed ML-TAE control chart, we applied it to the real dataset originally analyzed by9. The Phase I dataset (provided in the Table A1 of Appendix A) was first standardized to follow a standard normal distribution to eliminate location and scale effects. Using this standardized data, Phase I estimates of the process mean and dispersion were obtained, and the corresponding control limits were constructed for the proposed chart. Figs. 2, 3, 4, 5 and 6 display the Phase II behavior of the proposed chart under different process conditions. Fig. 3 and 4 show the chart’s performance under a small mean shift of δ = 0.25 at contamination probabilities of P = 0.05 and P = 0.10, respectively. Fig. 5 and 6 illustrate the response under a moderate shift of δ = 0.5 with the same contamination settings. Under P = 0.05, the chart signals promptly after the mean shift, especially in Fig. 5 where the out-of-control signal appears shortly after the shift point, confirming its ability to detect moderate changes efficiently. Even for the smaller shift of δ = 0.25 shown in Fig. 3, the chart still detects the change with a reasonable delay, which is expected due to the weaker shift magnitude. When the contamination level increases to P = 0.10 (Figs. 4 and 6), the chart maintains almost the same detection pattern, indicating that its performance is not degraded by the presence of outliers. This consistency across different contamination levels highlights the robustness of the proposed ML-TAE chart, showing that it effectively down weights outlier influence while preserving its sensitivity to genuine process changes. Fig. 2 is a flow chart explaining how to get the plots in this example. The detection behavior shown in Figs. 3 and 4 confirms the ARL patterns from the simulation results summarized in Tables 1, 2, 3 and 4. For the small shift of δ = 0.25 with contamination levels of P = 0.05 and P = 0.10, the ML-TAE chart signals after a moderate delay, consistent with the simulation result that small shifts yield ARLs around 120–150 for such contamination proportions. From the tables, increasing P from 0.05 to 0.10 increases ARL only moderately, which indicates stability of the chart when the frequency of outliers rises. Similarly, Figs. 3 and 4 demonstrate the same stability: the time of signal under P = 0.10 is only slightly later than under P = 0.05. The same applies to Figs. 5 and 6. The plots of this case study confirm the same pattern as Tables 1, 2, 3 and 4: the ML-TAE chart maintains robustness to contamination while preserving adequate sensitivity for detecting small process shifts.Fig. 2Flow chart to conduct the practical application.Fig. 3Proposed control chart with shift = 0.5 and P = 0.05.Fig. 4Proposed control chart with shift = 0.25 and P = 0.05.Fig. 5Proposed control chart with shift = 0.5 and P = 0.1.Fig. 6Proposed control chart with shift = 0.25 and P = 0.1. Flow chart to conduct the practical application. Proposed control chart with shift = 0.5 and P = 0.05. Proposed control chart with shift = 0.25 and P = 0.05. Proposed control chart with shift = 0.5 and P = 0.1. Proposed control chart with shift = 0.25 and P = 0.1. Conclusion This study introduces the ML-TAE control chart for the normal data, a robust and adaptive monitoring tool that combines the strengths of truncated EWMA charts and machine learning–based anomaly detection. By effectively incorporating anomaly scores as adaptive weights, the proposed chart minimizes the impact of outliers and maintains sensitivity to meaningful process shifts. The results of simulations confirmed that the ML-TAE chart outperforms the classical AE and TAE charts. Offering superior ARL properties for a wide range of shift magnitudes and contamination levels, the real-data application further demonstrated the practical relevance and effectiveness of the approach within real-world industrial contexts. Extensions may be developed based on the proposed approach in the following directions: multivariate process monitoring, online learning for dynamic anomaly detection, or applications in high-dimensional data streams. Such developments hold further potential to strengthen the role of data-driven quality control within modern smart manufacturing systems. The computational cost of the proposed ML-TAE chart is moderate. In general, both anomaly detection models, such as Isolation Forest or Autoencoder, are trained only once in Phase I data and are updated in real time using lightweight scoring operations. Thus, the approach is feasible for online process monitoring or for implementation in an industrial IoT environment. In future studies, the framework can also be generalized to a multivariate or non-Gaussian setting by incorporating multivariate anomaly detectors or robust nonparametric transformations, widening its applicability to complex manufacturing and service processes. Supplementary Information Supplementary Information. Supplementary Information.