A new Bayesian multivariate exponentially weighted moving average control chart for phase II monitoring of multivariate multiple linear profiles

Recently, statistical profile monitoring methods have become efficient tools for monitoring the quality of a product (or a production process) using control charts. The key idea is to describe the relationship between a response variable and a set of explanatory variables in the form of a statistical regression model, which called profile. Traditionally, those control charts are constructed with standard “frequentistic” regression models. Recently, it has been proposed to apply Bayesian regression models instead, and it has been empirically demonstrated that Bayesian regression models have the potential to perform significantly better. In this paper, we introduce a novel Bayesian multivariate exponentially weighted moving average control chart for monitoring multivariate multiple linear profiles in phase II. The key idea is to use the data from historical data sets to generate informative prior distributions for the regression models in phase II. The results of our empirical simulation studies show that the Bayesian multivariate multiple linear regression model is superior to its classical “frequentistic” counterpart in terms of the average run length. Our empirical findings are in agreement with findings reported in recently published articles. To shed more light onto the merit of the proposed Bayesian method, we carry out a sensitivity analysis, in which we investigate how the amount of phase I data influences the results. We also demonstrate the applicability and superiority of the proposed Bayesian method by a real‐world application.     

Gaussian process regression for multivariate spectroscopic calibration

Traditionally multivariate calibration models have been developed using regression based techniques including principal component regression and partial least squares and their non-linear counterparts. This paper proposes the application of Gaussian process regression as an alternative method for the development of a calibration model. By formulating the regression problem in a probabilistic framework, a Gaussian process is derived from the perspective of Bayesian non-parametric regression, prior to describing its implementation using Markov chain Monte Carlo methods. The flexibility of a Gaussian process, in terms of the parameterization of the covariance function, results in its good performance in terms of the development of a calibration model for both linear and non-linear data sets. To handle the high dimensionality of spectral data, principal component analysis is initially performed on the data, followed by the application of Gaussian process regression to the scores of the extracted principal components. In this sense, the proposed method is a non-linear variant of principal component regression. The effectiveness of the Gaussian process approach for the development of a calibration model is demonstrated through its application to two spectroscopic data sets. A statistical hypothesis test procedure, the paired t-test, is used to undertake an empirical comparison of the Gaussian process approach with conventional calibration techniques, and it is concluded that the Gaussian process exhibits enhanced behaviour.     

An adaptive EWMA control chart for monitoring the process mean in Bayesian theory under different loss functions

The Shewhart control chart is used for detecting the large shift and an exponentially weighted moving average (EWMA) control chart is used for detecting the small/moderate shift in the process mean. A scheme that combines both the Shewhart control chart and the EWMA control chart in a smooth way is called the adaptive EWMA (AEWMA) control chart. In this paper, we proposed a new AEWMA control chart for monitoring the process mean in Bayesian theory under different loss functions (LFs). We used informative (conjugate prior) under two different LFs: (1) squared error loss function and (2) linex loss function for posterior and posterior predictive distributions. We used the average run length and standard deviation of run length to measure the performance of the AEWMA control chart in the Bayesian theory. A comparative study is conducted for comparing the proposed AEWMA control chart in Bayesian theory with the existing Bayesian EWMA control chart. We conducted a Monte Carlo simulation study to evaluate the proposed AEWMA control chart. For the implementation purposes, we presented a real-data example.     

Process monitoring based on mode identification for multi-mode process with transitions

Some industrial processes frequently change due to various factors, such as alterations of feedstocks and compositions, different manufacturing strategies, fluctuations in the external environment and various product specifications. Most multivariate statistical techniques are under the assumption that the process has one nominal operation region. The performance of it is not good when they are used to monitor the process with multiple operation regions. In this paper, we developed an effective approach for monitoring multi-mode continuous processes with the following improvements. 1). Offline mode identification algorithm is proposed to identify (i) stable modes, (ii) transitional modes between two stable modes, and (iii) noise. 2). According to the data distribution, proper multivariate statistical algorithm is selected automatically to realize fault detection for each mode. 3). When online monitoring, the right model is chosen based on Mode Transformation Probability (MTP), which makes full use of the empirical knowledge hidden in offline data. This method can enhance real-time performance of online mode identification for continuous process and timely monitoring can be further realized. The proposed method is illustrated by application in furnace temperature system of continuous annealing line. The effectiveness of mode identification and fault detection is demonstrated in the results.     

A spatial rank–based multivariate EWMA chart for monitoring process shape matrices

In this article, we propose a nonparametric EWMA control chart for monitoring the shape matrix of a multivariate process based on a spatial rank test and the exponentially weighted moving average scheme. The proposed control chart is essentially developed using an estimated spatial rank covariance matrix to test the shape matrix of the covariance matrix of multivariate distributions with heavy tails. Based on our simulation studies, the proposed control chart outperforms the only existing nonparametric control chart in many practical out‐of‐control scenarios for monitoring the shape matrix of the covariance matrix of many multivariate processes. Further, we point out the weaknesses of both the nonparametric EWMA control charts for monitoring the shape matrix of multivariate processes in real applications and propose one possible method to overcome these weaknesses. We also use an example from a white wine production process to demonstrate the applicability and implementation of the proposed control chart.     

Multivariate statistical process control charts: an overview

In this paper we discuss the basic procedures for the implementation of multivariate statistical process control via control charting. Furthermore, we review multivariate extensions for all kinds of univariate control charts, such as multivariate Shewhart‐type control charts, multivariate CUSUM control charts and multivariate EWMA control charts. In addition, we review unique procedures for the construction of multivariate control charts, based on multivariate statistical techniques such as principal components analysis (PCA) and partial least squares (PLS). Finally, we describe the most significant methods for the interpretation of an out‐of‐control signal. Copyright © 2006 John Wiley & Sons, Ltd.     

Engineering model-based Bayesian monitoring of ramp-up phase of multistage manufacturing process

Process monitoring of full mass production phase of multistage manufacturing processes (MMPs) has been successfully implemented in many applications; however, monitoring of ramp-up phase of MMPs is often more difficult to conduct due to the limited information to establish valid process control parameters (such as mean and variance). This paper focuses on the estimation of the process control parameters used for monitoring scheme design of ramp-up phase of MMPs. An engineering model of variation propagation of an MMP is developed and reconstructed to a linear model, establishing a relationship between the error sources and the variation of product characteristics. Based on the developed linear model, a two-step Bayesian method is proposed to estimate the process control parameters. The performance of the proposed Bayesian method is validated with simulation data and real-world data, and the results demonstrate that the proposed method can effectively estimate process parameters during ramp-up phase of MMP.     

Markov chain Monte Carlo in Bayesian models for testing gamma and lognormal S-type process qualities

The process capability index C_pu is widely used to measure S-type process quality. Many researchers have presented adaptive techniques for assessing the true C_pu assuming normality. However, the quality characteristic is often abnormal, and the derived techniques based on the normality assumption could mislead the manager into making uninformed decisions. Therefore, this study provides an alternative method for assessing C_pu of non-normal processes. The Markov chain Monte Carlo, an emerging popular statistical tool, is integrated into Bayesian models to seek the empirical posterior distributions of specific gamma and lognormal parameters. Afterwards, the lower credible interval bound of C_pu can be derived for testing the non-normal process quality. Simulations show that the proposed method is adaptive and has good performance in terms of coverage probability.     

Should observations be grouped for effective monitoring of multivariate process variability?

A multivariate dispersion control chart monitors changes in the process variability of multiple correlated quality characteristics. In this article, we investigate and compare the performance of charts designed to monitor variability on the basis of individual and grouped multivariate observations. We compare one of the most well-known methods for monitoring individual observations—a multivariate exponentially weighted mean squared deviation (MEWMS) chart—with various charts based on grouped observations. In addition, we compare charts based on monitoring with overlapping and nonoverlapping subgroups. We recommend using charts based on overlapping subgroups when monitoring with subgroup data. The effect of subgroup size is also investigated. Steady-state average time to signal is used as the performance measure. We show that monitoring methods based on individual observations are the quickest in detecting sustained shifts in the process variability. We use a simulation study to obtain our results and illustrated these with a case study.     

Multivariate process monitoring of an experimental blast furnace

Process monitoring by use of multivariate projection methods has received increasing attention as it can reduce the monitoring problem for richly instrumented industrial processes with many correlated variables. This article discusses the monitoring and control of a continuously operating experimental blast furnace (EBF). A case study outlines the need for monitoring and control of the EBF and the use of principal components (PCs) to monitor the thermal state of the process. The case study addresses design, testing and online application of PC models for process monitoring. The results show how the monitoring problem can be reduced to following just a few PCs instead of many original variables. The case study highlights the problem of multivariate monitoring of a process with frequently shifting operating modes and process drifts and stresses the choice of a good reference data set of ‘normal’ process behavior. Possible solutions for adaptations of the multivariate models to process changes are also discussed. Copyright © 2009 John Wiley & Sons, Ltd.     

Memory-type multivariate charts with fixed and variable sampling intervals for process mean when covariance matrix is unknown

Memory-type multivariate charts have been widely recognized as a potentially powerful process monitoring tool because of their excellent speed in detecting small-to-moderate shifts in the mean vector of a multivariate normally distributed process, namely, the multivariate EWMA (MEWMA), double MEWMA, Crosier multivariate CUSUM (MCUSUM), and Pignatiello and Runger MCUSUM charts. These multivariate charts are based on the assumption that the covariance matrix is known in advance; but, it may not be known in practice. It is thus not possible to use these multivariate charts unless a large Phase I dataset is available from an in-control process. In this paper, we propose multivariate charts with fixed and variable sampling intervals for the process mean vector when the covariance matrix is estimated from sample. Using the Monte Carlo simulation method, the run length characteristics of the multivariate charts are computed. It is shown that the in-control and out-of-control run length performances of the proposed multivariate charts are robust to the changes in the process covariance matrix, while the existing multivariate charts are not. A real dataset is taken to explain the implementation of the proposed multivariate charts.     

Multivariate process capability using principal component analysis

Wang and Chen (Qual. Eng. 1998; 11:21–27) have defined process capability indices (PCIs) for multivariate normal processes data using principal component analysis (PCA). Veevers (Statistical Process Monitoring and Optimization. Marcel Dekker: New York, NY, 1999; 241–256) has suggested a multivariate capability index based on the first principal component (PC). In this paper we demonstrate the problem in the definition of PCIs given by Wang and Chen (Qual. Eng. 1998; 11:21–27) and the non‐suitability of PCI given by Veevers (Statistical Process Monitoring and Optimization. Marcel Dekker: New York, NY, 1999; 241–256) through some examples. We also suggest an alternative method for assessing multivariate process capability based on the empirical probability distribution of PCs. This method has been performed on industrial and simulated data. Copyright © 2008 John Wiley & Sons, Ltd.     

Memory‐type multivariate control charts with auxiliary information for process mean

In statistical process control, it is a common practice to increase the sensitivity of a control chart with the help of an efficient estimator of the underlying process parameter. In this paper, we consider an efficient estimator that requires information on several study variables along with one or more auxiliary variables when estimating the mean of a multivariate normally distributed process. Using this auxiliary‐information‐based (AIB) process mean estimator, we propose new multivariate EWMA (MEWMA), double MEWMA (DMEWMA), and multivariate CUSUM (MCUSUM) charts for monitoring the process mean, denoted by the AIB‐MEWMA, AIB‐DMEWMA, and AIB‐MCUSUM charts, respectively. The run length characteristics of the proposed multivariate charts are computed using Monte Carlo simulations. The proposed charts are compared with their existing counterparts in terms of the run length characteristics. It turns out that the AIB‐MEWMA, AIB‐DMEWMA, and AIB‐MCUSUM charts are uniformly and substantially better than the MEWMA, DMEWMA, and MCUSUM charts, respectively, when detecting different shifts in the process mean. A real dataset is considered to explain the implementation of the proposed and existing multivariate control charts.     

Detecting the process changes for multivariate nonlinear profile data

In profile monitoring for a multivariate manufacturing process, the functional relationship of the multivariate profiles rarely occurs in linear form, and the real data usually do not follow a multivariate normal distribution. Thus, in this paper, the functional relationship of multivariate nonlinear profile data is described via a nonparametric regression model. We first fit the multivariate nonlinear profile data and obtain the reference profiles through support vector regression (SVR) model. The differences between the observed multivariate nonlinear profiles and the reference profiles are used to calculate the vector of metrics. Then, a nonparametric revised spatial rank exponential weighted moving average (RSREWMA) control chart is proposed in the phase II monitoring. Moreover, a simulation study is conducted to evaluate the detecting performance of our proposed nonparametric RSREWMA control chart under various process shifts using out‐of‐control average run length (ARL₁ ). The simulation results indicate that the SREWMA control chart coupled with the metric of mean absolute deviation (MAD) can be used to monitor the multivariate nonlinear profile data when a common fixed design (CFD) is not applicable in the phase II study. Finally, a realistic multivariate nonlinear profile example is used to demonstrate the usefulness of our proposed RSREWMA control chart and its monitoring schemes.     

Dual multivariate CUSUM charts with auxiliary information for process mean

It is customary to increase the sensitivity of a control chart using an efficient estimator of the underlying process parameter which is being monitored. In this paper, using an auxiliary information-based (AIB) mean estimator, we propose dual multivariate CUSUM (DMCUSUM) and mixed DMCUSUM (MDMCUSUM) charts, called the AIB-DMCUSUM and AIB-MDMCUSUM charts, with and without fast initial response features for monitoring the mean vector of a multivariate normally distributed process. The DMCUSUM chart combines two similar-type multivariate CUSUM (MCUSUM) charts while the MDMCUSUM chart combines two different-type MCUSUM charts, into a single chart. The objective of two multivariate subcharts in the DMCUSUM/MDMCUSUM chart is to simultaneously detect small-to-moderate and moderate-to-large shifts in the process mean vector. Monte Carlo simulations are used to compute the run length characteristics, including the average run length (ARL), extra quadratic loss, and integral of the relative ARL. Based on detailed run length comparisons, it turns out that the AIB-DMCUSUM and AIB-MDMCUSUM charts uniformly and substantially outperform the DMCUSUM and MDMCUSUM charts when detecting different sizes of shift in the process mean vector. A real dataset is used to explain the implementation of proposed AIB multivariate charts.     

Wavelet-based multiscale statistical process monitoring: A literature review

Data that represent complex and multivariate processes are well known to be multiscale due to the variety of changes that could occur in a process with different localizations in time and frequency. Examples of changes may include mean shift, spikes, drifts and variance shifts all of which could occur in a process at different times and at different frequencies. Acoustic emission signals arising from machining, images representing MRI scans and musical audio signals are some examples that contain these changes and are not suited for single scale analysis. The recent literature contains several wavelet-decomposition-based multiscale process monitoring approaches including many real life process monitoring applications. These approaches are shown to be effective in handling different data types and, in concept, are likely to perform better than existing single scale approaches. There also exists a vast literature on the theory of wavelet decomposition and other statistical elements of multiscale monitoring methods, such as principal components analysis, denoising and charting. To our knowledge, no comprehensive review of the work relevant to multiscale monitoring of both univariate and multivariate processes has been presented to the literature. In this paper, over 150 both published and unpublished papers are cited for this important subject, and some extensions of the current research are also discussed.     

Multivariate monitoring of batch process startup

A batch process is finite in duration and can be separated into two stages: startup and production. We develop a methodology to monitor a batch process during the startup stage to reduce the length of the startup stage. We focus on processes that are characterized by multiple process parameters and product characteristics. Because of the complex interdependencies characterizing the process parameters and product characteristics, it is more effective to evaluate them simultaneously. To address the multivariate nature of the process we use a multivariate statistical model: PLS (Projection to Latent Structures). PLS has been applied to several applications in statistical process monitoring. We present a new application of PLS to the startup stage of a batch process. Iterative adjustments made during startup in search of an acceptable production zone consume considerable amounts of material, labor and equipment time. We develop a monitoring procedure to reduce the time as well as the number of iterations and adjustments needed for startup. A PLS model is constructed, using baseline data, to characterize the relationship among process parameters during good production. The startup stage is monitored using the PLS characterization to determine if the process is consistent with good production. We illustrate the improved startup operations with an example from a batch process in filament extrusion, the application that motivates this work. Copyright © 2001 John Wiley & Sons, Ltd.     

A predictive Bayesian approach to sequential time-between-events monitoring

A fundamental problem with all process monitoring techniques is the requirement of a large Phase-I data set to establish control limits and overcome estimation error. This assumption of having a large Phase-I data set is very restrictive and often problematic, especially when the sampling is expensive or not available, eg, time-between-events (TBE) settings. Moreover, with the advancement in technology, quality practitioners are now more interested in online process monitoring. Therefore, the Bayesian methodology not only provides a natural solution for sequential and adaptive learning but also addresses the problem of a large Phase-I data set for setting up a monitoring structure. In this study, we propose Bayesian control charts for TBE assuming homogenous Poisson process. In particular, a predictive approach is adopted to introduce predictive limit control charts. Beside the Bayesian predictive Shewhart charts with dynamic control limits, a comparison of the frequentist sequential charts, designed by using unbiased and biased estimator of the process parameter, is also a part of the present study. To assess the predictive TBE chart performance in the presence of practitioner-to-practitioner variability, we use the average of the average run length (AARL) and the standard deviation of the in-control run length (SDARL).     

Fault detection based on Gaussian process latent variable models

Gaussian processes, GPs, can be used to approximate complex non-linear functions with relative simplicity. Their regression performance is, at least, comparable to that achieved via artificial neural networks (ANN) and, in fact, both methods are intrinsically related. They are both non-parametric and, as Neal (1994) [1] has shown, when the number of nodes in the hidden layer of a neural network tends to infinity the ANN converge to a Gaussian process.In most of the cases, the GP will map a multivariate input into a univariate response. In this paper, however, we present an approach to process monitoring that combines several GPs so that multivariate responses can be appropriately modeled. We review a similar approach recently proposed in the literature and highlight some concerns related to it that needs to be taken into consideration. Additionally, we propose an alternative procedure to the way in which new observations are mapped into the non-linear model. A simulation study is provided that will help understand the method flexibility. Furthermore, results from a real example are also discussed.     

An EWMA chart for monitoring the covariance matrix of a multivariate process based on dissimilarity index

In this article, we propose an exponentially weighted moving average (EWMA) control chart for monitoring the covariance matrix of a multivariate process based on the dissimilarity index of 2 matrices. The proposed control chart essentially monitors the covariance matrix by comparing the individual eigenvalues of the estimated EWMA covariance matrix with those of the estimated covariance matrix from the in‐control (IC) phase I data. It is different from the conventional EWMA charts for monitoring the covariance matrix, which are either based on comparing the sum or product or both of the eigenvalues of the estimated EWMA covariance matrix with those of the IC covariance matrix. We compare the performance of the proposed chart with that of the best existing chart under the multivariate normal process. Furthermore, to prevent the control limit of the proposed EWMA chart developed using the limited IC phase I data from having extensively excessive false alarms, we use a bootstrap resampling method to adjust the control limit to guarantee that the proposed chart has the actual IC ARL(average run length) not less than the nominal level with a certain probability. Finally, we use an example to demonstrate the applicability and implementation of the proposed EWMA chart.