MAX-EWMA CHART FOR AUTOCORRELATED PROCESSES ( MEWMAP CHART )

This paper proposes an exponentially weighted moving average (EWMA) control chart that is capable of detecting changes in both process mean and standard deviation for autocorrelated data (referred to as the Maximum Exponentially Weighted Moving Average Chart for Autocorrelated Process, or MEWMAP chart). This chart is based on fitting a time series model to the data, and then calculating the residuals. The observations are represented as a first-order autoregressive process plus a random error term. The Average Run Lengths (ARLs) for fixed decision intervals and reference values (h, k) are calculated. The proposed chart is compared with the Max-CUSUM chart for autocorrelated data proposed by Thaga (2003). Comparisons are based on the out-of-control ARLs. The MEWMAP chart detects moderate to large shifts in the mean and/or standard deviation at both low and high levels of autocorrelations more quickly than the Max-CUSUM chart for autocorrelated processes.


INTRODUCTION
Statistical process control (SPC) charts such as the Shewhart control chart (Shewhart [14]), the cumulative sum control chart (Page [10]), and the exponentially weighted moving average control chart (Roberts [12]), are used to monitor product quality and to detect special events that may be indicators of out-of-control situations.These charts are designed on the assumption that a process being monitored will produce measurements that are independent and identically distributed over time, when only the inherent sources of variability are present in the system.However, in some applications the dynamics of the process will produce correlations in observations that are closely spaced in time.If the sampling interval used for process monitoring in these applications is short enough for the process dynamics to produce significant correlation, then this correlation can have a very serious effect on the properties of standard control charts (VanBrackle & Reynolds [16], Lu & Reynolds [7] and [8], Runger & Willemain [13], Atienza, Tang & Ang [2]).
Positive autocorrelation in observations can result in severe negative bias in traditional estimators of the standard deviation.This bias produces control limits that are much tighter than desired.This can result in a much higher average false alarm rate than expected.Furthermore, when observations are positively autocorrelated, when there is a shift in the process mean, only a fraction of the shift will be transferred to the residual mean, and the chart will not quickly detect this shift.It is therefore very important to take autocorrelation among observations into consideration when designing a process-monitoring scheme -in particular, control charts -in order to maximize full benefit from the control charts.
Recently, new control charts have been proposed for dealing with autocorrelated data.Two approaches have been advocated for dealing with this phenomenon.The first approach uses standard control charts on original observations, but adjusts the control limits and methods of estimating parameters to account for the autocorrelation in the observations (VanBackle & Reynolds [16], Lu & Reynolds [7]).This approach is particularly applicable when the level of autocorrelation is not high.
A second approach for dealing with autocorrelation fits a time series model to the process observations.The procedure forecasts observations from previous values and then computes the forecast errors or residuals.These residuals are then plotted on standard control charts, because -when the fitted time series model is the same as the true process model and the parameters are estimated without error -the residuals are independent and identically distributed normal random variables when the process is in control (Alwan & Roberts [1], Montgomery & Mastrangelo [9], Wadell, Moskowitz, & Plante [17], Lu & Reynolds [7], and Runger, Willemain & Prabhu [3]).Yashchin [19] recommends directly charting raw data when the level of autocorrelation is low; at a high level of autocorrelation he recommends some transformation procedures that create residuals.If a shift in the mean and/or standard deviation of the process occurs, this will cause a shift in the mean and/or standard deviation of the residuals.Control charts based on residuals seem to work best when the level of correlation is high.When the level of correlation is low, forecasting is more difficult and residual charts are not very effective at detecting process changes.
The studies mentioned above used several methods, such as simulation, asymptotic approximation, and direct calculation, to evaluate the properties of the control charts.A conclusion that can be drawn from these studies is that correlation between observations has a significant effect on the properties of the control charts.In particular, when the level of autocorrelation is high, control charts run for a long time before detecting shifts in the process parameters from in-control values.
The objective of this paper is to investigate control charts for simultaneously monitoring the process mean and variation using a single chart in the presence of autocorrelation.We propose an exponentially weighted moving average (EWMA) control chart for autocorrelated data that can simultaneously monitor shifts in the mean and standard deviation using a single plotting variable.This investigation is done for the case of processes that can be modeled as a first order autoregressive AR(1) process plus an additional random error, which can correspond to sampling, or measurement error.This model allows relatively accurate numerical techniques to be used to evaluate the properties of the control charts.Lu & Reynolds [7] proposed a simultaneous EWMA control chart for autocorrelated processes, which runs two control charts concurrently.Chen, Cheng & Xie [5] developed a single EWMA chart on the assumption that the process produces measurements that are independent over time when the process is in control.
Our proposed chart monitors the process by monitoring the residual means and standard deviations.The results show that by taking the autocorrelation structure of the process into consideration, the EWMA chart can effectively detect small shifts in the process mean and/or spread.

THE AR(1) PROCESS WITH AN ADDITIONAL RANDOM ERROR
This model has been used previously in a number of contexts, and has the advantage that it will account for the correlation between observations that are close together in time, for variability in the process mean over time, and for additional variability due to sampling or measurement error.
To model observations from an autocorrelated process, we use a model that has been discussed previously in quality control by authors such as Lu & Reynolds [7 and 8] and VanBackle & Reynolds [16].For this model, X t can be represented as .
Where X t represents an observation taken from the process at sampling time; t, t μ is the random process mean at sampling time; t and , t ε s are independent normal random errors with mean 0 and variance .
over time, and for additional variability due to sampling or measurement error.It is assumed that t μ can be described as an AR(1) process, defined as Where ξ is the overall process mean, , t α s are independent normal random variables with mean 0 and variance 2 α σ , and φ is the autoregressive parameter satisfying |φ | <1 for a process to be stationary.We assume that the starting value 0 μ follows a normal distribution with mean ξ and variance The distribution of X t is therefore constant with mean ξ and variance given by Var(X t ) = Var( t μ ) + Var( t ε ).This variance is given as In this case 2 μ σ represents long-term variability, and 2 ε σ represents a combination of short-term variability and the variability associated with measurement error.When assessing processes following models (1) and ( 2), it is often convenient to consider the proportion of total process variability that is due to variation in t μ and the proportion due to error variability.The proportion of the process variability due to variation in t μ is defined as .
The AR(1) process with an additional random error is equivalent to a first order autoregressive moving average, ARMA(1,1) process (Box,Jenkins & Reinsel [3]), which can be written as Where , t γ s are independent normal random variables with mean 0 and variance , 2 γ σ θ is the moving average parameter, φ is the autoregressive parameter defined in equation (2), and B is a backshift operator such that BX t = X t-1 .If φ >0, Koons & Foutz [6] estimate θ and 2 γ σ as http://sajie.journals.ac.za The standard time series estimation techniques can be used to estimate the parameters in the ARMA(1,1) model.In some production processes, a large volume of items is produced in a single lot.In this situation, more than one observation is sampled each time.Let X ti be the i th observation at sampling time t.We assume that X ti can be represented as Where the ti ε ′ s are independent and identically distributed normal random variables with mean 0 and variance 2 ε σ ′ , and t μ follows model (2).Lu & Reynolds [7] indicated that the sample means from this process will follow models (1) and ( 2) In this article, we monitor the process mean and standard deviation by monitoring the residuals from a forecast.To do this, we first determine the distribution of the residuals when the process is in control.When the process is in control, the residual at observation t from the minimum mean square error forecast made at observation t- Where φ and θ are parameters in the ARMA(1,1) model given in equation (4), and 0 ξ is the in-control process mean -that is, the residual at time t is the difference between X t and the prediction of X t based on the previous data.
If the fitted time series model is the same as the true process model and the parameters are estimated without error, then the residuals are independent and identically distributed normal random variables when the process is in control.We can then monitor the process by using standard control charts for independent observations using these residuals.If there is a step change in the process mean from the in-control value 0 ξ to 1 ξ between time t = 1 − τ and t = , τ the expectations of the residuals for various times are (Lu & Reynolds [7]): The asymptotic mean of these residuals is given as These residuals are independent and normally distributed with variance 2 γ σ .The expectation of the residuals after the shift occurs is a decreasing function of time.
Also, as φ increases, a small fraction of shift in the process mean will be transferred to the mean of the residuals.As a result, the ability of the chart for residuals to detect the mean shift is reduced.On the other hand, the residuals chart is theoretically very appealing because it takes the serial correlation into account, and reduces the problem to the well-known case of a shift in the process mean for independent observations.
A change in the process variance can be attributed to a change in the autoregressive parameter φ , the individual observation random shock variance 2 ε σ , and/or change in variability of the random shocks associated with the mean . Where is a sequence of independent normal random variables with a mean 0 and variance , 2 0 where is a sequence of independent normal random variables with mean 0 and variance , 2 0 ε respectively.Therefore, using the model in equation (1) and equations ( 8), ( 9) and (10), it can be shown by induction that http://sajie.journals.ac.za and The asymptotic variance of these residuals after the shift will increase to the limit The residuals after these shifts are correlated normal random variables with an asymptotic mean in equation (10) and asymptotic variance given in equation (15).
From equation (15), we can see that changes in We can then fit the AR(1) plus random error model in equations ( 1) and ( 2), which is the model considered in this paper.We consider the case of positive autocorrelation, which is more prevalent than negative autocorrelation in control chart applications.

THE NEW CONTROL CHART
We propose a new exponentially weighted moving average control chart for residuals in this section.Let X i = X i1 , X i2 , K , X in , i = 1, 2, K , denote a sequence of samples of size n taken on a quality characteristic X.It is assumed that, for each i, X i1 , X i2 , K , X in follows a normal distribution and is autocorrelated and can be expressed as in equation (7).We monitor the process by first fitting the time series model into the process observations, and then computing the residuals.Let 0 otherwise, the process has changed due to some assignable causes.Then a represents the shift in the process mean and b represents the shift in the process standard deviation. Let be the mean and variance for the i th sample residuals respectively.These statistics are independently distributed, as are the sample residual values when the process is in-control.These two statistics follow different distributions.The EWMA charts for the mean and standard deviation are based on i ξ and MSE i respectively.
To develop an EWMA chart for the process mean and process standard deviation using residuals, we carry out the following transformations: http://sajie.journals.ac.za .
is the inverse function of ), (⋅ Φ the cumulative distribution function of N(0, 1) and H(w; p) = P(W ≤ w|p) for W ~ , Z i and Y i are independent and when a = 0 and b = 1, they follow the standard normal distribution.The EWMA statistics based on Z i and Y i are defined as: respectively, where U 0 and V 0 are starting values of the chart and . 1 0 <

< λ
The parameter λ is called the smoothing parameter of the EWMA chart.Because Z i and Y i follow the same distribution, a new statistic for a new single control chart is defined as If the process is operating out of control, the M i 's will be plotted outside the control limits, otherwise the M i values are within the limits.Since M i >0, we plot only the upper limit for this chart, and consider the process to be out of control if an M i value is plotted above the upper control limit.
In statistical process control (SPC), we often use the average run length (ARL) of the chart to assess the performance of the scheme.This is the expected number of samples (or observations, if we take a single observation each time) required by the chart to signal an out-of-control situation.For a change in variability, we consider the effects of changes in ε σ and α σ separately to calculate the ARL.This is because the two parameters have different impacts on the level of variability of the process as shown in equation (11).The shifts in these parameters are considered for different values of . φ As shown previously, when the process is in control, the residuals are independent and identically distributed normal random variables with mean If there is a change in the mean and/or standard deviation, the residuals are correlated normal random variables with mean given in equation (10) and variance given in equation (15).In this article we consider a situation of an increase in the mean, standard deviation, or both mean and standard deviation of the process.

DESIGN OF A MAX-EWMA CHART FOR AUTOCORRELATED PROCESS (MEWMAP CHART)
When the process is in control and U 0 = V 0 = 0, U i and V i can be written as: It can be shown that Where again λ is called the smoothing parameter of the EWMA chart.
Chen, Cheng & Xie [5] showed that the in-control cumulative distribution function of M i is given as: The corresponding probability density function of M i is the derivative of Using cubic-spline numerical integration, Xie [18] computed the mean and variance of M i and obtained The upper control limit (UCL) is obtained as follows: where L is the width of the control limits.As i gets large To calculate the ARL of the new chart, we use the modified Markov chain procedure (VanBrackle & Reynolds [16]).For the AR(1) plus random error model investigated in this article for shifts in mean and/or standard deviation, we use the asymptotic mean and variance given in equation (10) and equation ( 15) respectively.For a given in-control ARL and a shift of the mean and/or standard deviation intended to be detected by the chart, we find the combination of the chart design parameters ) , ( L λ that gives the desired in-control ARL and also minimizes the out-of-control ARL.This guideline takes into consideration the autocorrelation structure between the variables.We calculate the out-of-control ARL for the effect of changes in the standard deviation that is due to changes in ε σ and http://sajie.journals.ac.za α σ respectively.The smallest value of an out-of-control ARL is calculated with respect to a pair of specified shifts in both mean and standard deviation, using the optimal in-control ARL EWMA chart parameters.We assume that the process starts in an in-control state, and thus the initial value of the EWMA statistic is set at zero.For example, if one wants to have an in-control ARL of 250 and to guard against a parameter values are λ = 0.9853 and L = 2.9996.These shifts can be detected on the second sample inspection, i.e., the ARL is approximately two, as seen in Table 1.
Table 2 gives the optimal combinations of λ and L for an in-control ARL fixed at 250 and the autoregressive parameter , 75 .0 = φ with 80% of process variability due to variation in , t μ and the correlation between adjacent observations equal to 0.6.
We use the same procedure to calculate the ARL for this table as for Table 1.
Comparing these tables, it can be seen that at a low level of autocorrelation, the charts detect small shifts in the parameters more quickly than at a high level of autocorrelation.The scheme is slightly more sensitive to shifts in the standard deviation due to shifts in ε σ than it is to shifts in the process standard deviation resulting from shifts in .α σ This is due to the fact that an increase in α σ increases the level of correlation between adjacent observations.This is because the variance of t μ increases, and thus the proportion of total process variability due to variation in the autocorrelated mean increases.
An increase in ε σ decreases the level of correlation between observations.This is particularly evident at higher levels of autocorrelations.This improves the performance of the MEWMAP chart.As would be expected, the chart detects small shifts with small values of λ , and large shifts are quickly detected when using large values of .
λ When , 1 = λ the EWMA chart is equivalent to the Shewhart chart, where all the weight is given to the current observation and L = 3.2359 which gives an UCL = 3.0899, which is the same as the UCL for the Shewhart-type Max-chart proposed by Chen & Cheng [3].

COMPARISON WITH OTHER CHARTS
In this section we compare the MEWMAP chart with a single CUSUM chart for autocorrelate processes (Thaga [15]).These charts' parameters are adjusted so that their in-control ARLs are equal to 250.These charts are compared in Tables 3 and 4.
We compare these charts for the autocorrelated processes when the values of φ -the correlation between t μ and 1 − t μ -are 0.25 and 0.75, with the proportion of variation in the process attributed to variation in t μ fixed at 0.8.The conclusion that can be made is that the Max-CUSUM chart performs better than the Max-EWMA chart for small shifts in the process mean and/or standard deviation at both low and high levels of autocorrelations, while the Max-EWMA outperforms the Max-CUSUM for moderate to large shifts in the process mean and/or standard deviation.Both schemes are more sensitive when change in process variability is due to change in ε σ than they are to changes in process variability due to changes in  This indicates an increase in the mean and a decrease in the standard deviation of the process.
This indicates a decrease in the mean and an increase in the standard deviation of the process.
This shows a decrease in both the mean and the standard deviation of the process.
8. Once an out of control signal is given by the chart plotting outside the control limit, the process should be stopped, an investigation of the cause(s) of shift for each out-of-control point in the chart should be carried out, and remedial measure(s) needed to bring the process back into an in-control state should be implemented.

APPLICATION IN INDUSTRY
Observations from processes in the chemical and pharmaceutical industries are frequently autocorrelated.The standard Shewhart chart, the cumulative sum chart, and the exponentially weighted moving average chart are effective in detecting shifts in the process when measurements are independent and identically distributed over time.The proposed chart is an improved version of the traditional EWMA chart.It has an added advantage of being able to detect shifts in the process when the measurements are not independent over time.
The new chart can be used to monitor observations from processes in the chemical and pharmaceutical industries where observation comes in batches.In this case a special cause might produce an increase in within-batch variability, in between-batch variability, or in both.

AN EXAMPLE
To provide a visual picture of how the MEWMAP chart responds to various kinds of process changes, a set of simulated data is used.Specific process changes are introduced into the data, and the chart is plotted to monitor these changes in the parameters.The data set was generated using the first order autoregressive models in equations ( 1) and (2).
For a fixed sequence of , t α s and , t ε s, a shift in α σ can be simulated by multiplying t α in equation ( 2) by a constant.A change in ε σ can be simulated by multiplying t ε in equation (1) by a constant, and a change in the mean is simulated by adding a constant to the generated observations.This approach is discussed in the literature (Lu & Reynolds [7]).This procedure allows different types of process changes to be investigated on the same basic sequence of , t α s and , t ε s.In this example, we assume the autoregressive parameter φ remains constant. http://sajie.journals.ac.za We simulated 100 observations with the following parameters: ξ = 0, φ = 0.75, α σ = 0.59 and ε σ =0.5.This gives X σ = 1.02 and ψ = 0.76.This implies that 76% of variability in the process is due to variation in t μ and that the correlation between the adjacent observations is ρ = φ ψ = 0.57.Using equations ( 5) and ( 6), the corresponding parameters in the ARMA(1,1) model in equation ( 4) are θ = 0.27 and  σ from 1.02 to 1.56.This corresponds to a 52% increase in the process standard deviation.This also leads to an increase in ψ from 0.76 to 0.90, and an increase in the correlation between adjacent observations from 0.57 to 0.675.
Therefore 90% of variation in the process is due to variation in α by the factor 0.97/0.59= 1.644.Adding 1 to the last 40 observations of the new process observation X t accomplishes the increase in the process mean for the last 40 observations.These shifts are signaled for the first time on the 67th observation.σ is shown in Figure 3.We consider an increase in mean from 0 to 1 and an increase in ε σ from 0.5 to 1. Assume that due to some special causes, these shifts occur immediately after the 60 th observation and remain in effect for the rest of the process.This increase in ε σ will result in an increase in the process standard deviation from 1.02 to 1.34, and this represents a 30% increase in the process standard deviation.Unlike the increase in , α σ the increase in ε σ results in a decrease in the correlation between adjacent observations from 0.57 to 0.33, and the value of the proportion of total process variability that is due to the variability in t μ also decreases from 56% to 44%.
The increase in ε σ for the last 40 observations was accomplished by multiplying the last 40 observations of the simulated t ε values by 1/0.5 = 2.00, while the increase in the mean is accomplished by adding 1 to the last 40 observations of the new process value X t .The chart signals a shift for the first time on the 62 nd observation for an increase in both parameters.
Comparing the results of the simulated shifts, we realize that this chart quickly detects a simultaneous shift in the process mean and standard deviation with a shift in the process standard deviation due to shift in .ε σ This is due to the fact that an increase in ε σ results in a decrease in the correlation between adjacent observations, while an increase in α σ results in an increase in this correlation structure.
http://sajie.journals.ac.zaAlthough it is very difficult to draw general conclusions based on one set of data corresponding to one set of process parameters, the ARL results given in this article, together with charts plotted in Figures 2 and 3, allow some conclusions to be drawn.
The results reported here have shown that correlation among observations from a process can have significant effect on the performance of a EWMA control chart.
Computer simulation of individual data from a first order autoregressive plus a random error model was used to show a pictorial display of the MEWMAP chart.
The monitoring problem in this model is very complicated, as it requires more parameters than is the case when the observations are independent.We have shown how a change in one of the two components of residual variances and the process mean impacts on the overall process performance.
However, in many applications, a change in the process may be because of a combination of changes in these parameters.Therefore it becomes very difficult to diagnose the variance component that has caused the process variance to change.It might be necessary to estimate the residual variance at the point of the shift to see which component has shifted.

X
The proportion of the variance due to t ε is then 1 -ψ .The covariance between two observations that are i units apart is 2 μ σ φ i , and the correlation between two adjacent observations is α for .l i ≤ Therefore the effect of a shift in the variance is to induce correlation in the residuals(Lu & Reynolds [7]).Therefore the variance of the residuals after the shift is in equation(15) will vanish, and therefore we will use the variance of the residual as Var(e t ) = /sajie.journals.ac.za

2 α σ and 2 εσσ
have different impacts on the variability of the residuals.Given the parameters in the ARMA(1,1) model, can be obtained from (Reynolds, Arnold & Baik[11]): nominal process mean and standard deviation of the residuals for this fitted model.Assume that the process residual parameters ξ and γ σ can be expressed as ξ = a = 0 and b = 1 when the process is in-control;

2 pχ
the chi-square distribution with p degrees of freedom.

)
approaches unity and the UCL approaches the steady-state value, given as: use the statistic M i to construct a new control chart.Because M i is the maximum of the two statistics, we call this new chart the Maximum Exponentially Weighted Moving Average for Autocorrelated Process chart (MEWMAP chart).
the optimal combinations of λ and L for an in-control ARL fixed at 250 ., a = 3 and b = 2, the optimal in-control chart for these simulated observations is drawn in Figure 1.All points fall within the acceptance region; thus the simulated process is in control.The chart's parameters ) , ( L λ are for an in-control ARL of 250 runs.The chart is designed to detect a σ 1 increase in the process mean and σ 2 increase in the process standard deviation.

Figure 1 :σ
Figure 1: The MEWMAP chart for the in-control simulated values

.tμ
The increase in α σ for the last 40 observations was accomplished by multiplying the last 40 values of t

Figure 2 :σ
Figure 2: The MEWMAP chart for shift in the process mean and variability due to shift in .α σ

Figure 3 :
Figure 3: The MEWMAP chart for shift in the process mean and variability due to shift in .ε σ

Table 4 : Comparison of the MEWMAP chart with the MCAP chart with
i V