Chart for Monitoring Univariate Autocorrelated Processes

The Autoregressive T 2 Chart for Monitoring Univariate Autocorrelated Processes DANIEL W APLEY Texas A&M University, College Station, TX 77843-33 FUGEE TSUNG Hong Kong University of Science and Technology, Kowloon, Hong Kong In this paper we investigate the autoregressive T 2 control chart for statistical process control of autocorrelated processes The method involves the monitoring, using Hotelling s T 2 statistic, of a vector formed from a moving window of observations of the univariate autocorrelated process It is shown that the T 2 statistic can be decomposed into the sum of the squares of the residual errors for various order autoregressive time series models fit to the process data Guidelines for designing the autoregressive T 2 chart are presented, and its performance is compared to that of residual-based CUSUM and Shewhart individual control charts The autoregressive T 2 chart has a number of characteristics, including some level of robustness with respect to modeling errors, that make it an attractive alternative to residual-based control charts for autocorrelated processes Introduction IN recent years, statistical process control (SPC) for autocorrelated processes has received a great deal of attention, due in part to the increasing prevalence of autocorrelation in process inspection data With improvements in measurement and data collection technology, processes can be sampled at higher rates, which often leads to data autocorrelation It is well known that the run length properties of common SPC methods like CUSUM and X charts are strongly affected by data autocorrelation, and the in-control average run length (ARL) can be much shorter than intended if the autocorrelation is positive (Johnson and Bagshaw (974) and Vasilopoulos and Stamboulis (978)) The most widely researched methods of SPC for autocorrelated processes are residual-based control charts, which involve fitting some form of autoregressive moving average (ARMA) model to the data Dr Apley is an Assistant Professor in the Department of Industrial Engineering His email address is apley@tamuedu Dr Tsung is an Assistant Professor in the Department of Industrial Engineering and Engineering Management His email address is season@uxmailusthk and monitoring the model residuals (ie, the onestep-ahead prediction errors) If the model is exact, then the model residuals are independent Consequently, standard SPC control charts can be applied to the residuals with well understood in-control run length properties (see eg, Alwan and Roberts (988); Apley and Shi (999); Berthouex, Hunter, and Pallesen (978); Chow, Wu, and Ermer (979); English, Krishnamurthi, and Sastri (99); Lin and Adams (996); Montgomery and Mastrangelo (99); Runger, Willemain, and Prabhu (995); Superville and Adams (994); Vander Wiel (996); and Wardell, Moskowitz, and Plante (994)) Krieger, Champ, and Alwan (992) and Alwan and Alwan (994) proposed a different approach to monitoring autocorrelated processes The basic idea was toform a multivariate vector of a moving window of observations from a univariate autocorrelated process, and then apply multivariate control charts Krieger et al (992) applied a multivariate CUSUM, and Alwan and Alwan (994) applied a T 2 chart to the constructed vectors In this paper, we analyze a slightly modified version of the T 2 approach of Alwan and Alwan (994) One difference between the approaches is that Alwan and Alwan (994) recommended time delays between samples so that the constructed vectors have less statistical dependency, Journal of Quality Technology 80 Vol 34, No, January 2002

THE AUTOREGRESSIVE T 2 CHART FOR MONITORING UNIVARIATE AUTOCORRELATED PROCESSES 8 whereas we do not In addition, Alwan and Alwan (994) focused on the case of a first-order autoregressive processes, whereas we take a more general approach The method, which will be referred to as the autoregressive T 2 control chart, has a close connection to residual-based control charts It is shown in this paper that the autoregressive T 2 statistic can be decomposed into the sum of the squares of the residual errors for various order autoregressive models, similar to well-known decompositions of multivariate T 2 statistics (see, for example, Hawkins (993) and Mason, Tracy, and Young (997)) In spite of the relationship to residual-based control charts, the autoregressive T 2 chart does not explicitly require an ARMA model of the process Only the process autocovariance function up to a pre-specified lag is required Guidelines for designing the autoregressive T 2 chart are presented, and its performance is compared to residual-based CUSUM, recently studied by Lu and Reynolds (200), and residual-based Shewhart individual charts In many situations, the autoregressive T 2 chart compares favorably In addition, it possesses other advantages over residualbased CUSUMs For example, the optimal CUSUM design (ie, selection of reference value and decision threshold) that minimizes the out-of-control ARL for a specified in-control ARL depends strongly on the mean shift size of interest (Montgomery (200)) A single autoregressive T 2 chart design, in contrast, is often suitable for a wide range of mean shift sizes This has practical significance, in that it is generally difficult to select one mean shift size that is of primary interest It is, therefore, desirable that a control chart performs well over a range of mean shift sizes Another disadvantage of residual-based CUSUMs is that an accurate ARMA model of the process is required If the model is inaccurate, the residuals will not be uncorrelated, and the in-control ARL of a residual-based CUSUM may be substantially shorter than what is intended This results in too many nuisance alarms Since ARMA models are always estimated from data, some level of model uncertainty will be present; if the data are limited, then the uncertainty may be large The autoregressive T 2 chart incorporates a mechanism for taking into account model uncertainty (or, analogously, uncertainty in the autocovariance structure of the process) due to limited data from which the process characteristics must be estimated In the remainder of this paper, we present the autoregressive T 2 chart and a performance comparison with residual-based control charts, provide an interpretation in terms of an autoregressive decomposition of the T 2 statistic, give guidelines for designing the chart, and investigate issues related to model uncertainty and limited data The Autoregressive T 2 Control Chart Let x t represent a measurement from an autocorrelated process at sampling instant t Suppose the process is Gaussian with mean µ and autocovariance function γ k = E[(x t µ)(x t+k µ)], where E[ ] denotes the expectation operator Let µ 0 denote the mean when the process is in-control, and assume µ 0 is known Consider the sequence of p-dimensional vectors X t = [x t p+ x t p+2 x t ] formed from observations of the univariate process Assume the process x t is such that X t is multivariate normal This will be the case if, for example, x t follows a stable ARMA model The covariance matrix of X t is γ 0 γ γ p γ Σ = γ 0, () γ γ p γ γ 0 and when the process is in-control the mean vector is µ 0 =[µ 0 µ 0 µ 0 ] IfΣ is known, the T 2 statistic T 2 t =[X t µ 0 ] Σ [X t µ 0 ], (2) which follows a chi-square distribution with p degrees-of-freedom when the process is in-control, could be used to monitor for departures from the in-control state This method will be referred to as the autoregressive T 2 chart One can specify a false alarm probability α and compare Tt 2 to the α percentile of the chi-square distribution with p degreesof-freedom, denoted χ 2 ( α, p) We discuss how to specify α to ensure a desired in-control ARL later in this paper In Alwan and Alwan (994), the intent was that X t be formed from a sample of p successive observations of the process The next vector would be X t+p, formed from a sample of observations that do not overlap those used to form the previous vector It was also suggested that there could be time delays between samples, so that the T 2 statistics are closer to being independent The intent in this paper isslightly different, in that Tt 2 is to be formed for t = p, p +,p+2,p+3,inother words, Tt 2 is cal- Vol 34, No, January 2002 wwwasqorg

82 DANIEL W APLEY AND FUGEE TSUNG culated and charted for every value of t, asopposed to t = p, 2p, 3p, The matrix Σ could be considered known if a sufficiently large data set were available to estimate the autocovariance function The situation where the data is insufficient to assume that the T 2 statistic in Equation (2) follows a chi-square distribution is treated later in this paper and is closely related to the issue of model uncertainty in residual-based control charts An alternative to directly estimating the autocovariance function is to fit an ARMA model to the data, and calculate the autocovariance function theoretically from the model parameters Define y t = x t µ, and suppose y t follows the ARMA(n, m) model (Box et al, 994) Φ(B)y t =Θ(B)a t, (3) where a t is an independent Gaussian process with mean 0 and variance σa, 2 and B is the backshift operator Here Φ(B) and Θ(B) are polynomials in B of degree n and m, respectively, and are parameterized as Φ(B) = φ B φ 2 B 2 φ n B n and Θ(B) = θ B θ 2 B 2 θ m B m For the special case of a first order model, when n = m =,aclosed-form expression for the autocovariance function is (Pandit and Wu (990)) σ 2 a (φ θ)( θφ) ( φ 2 γ k = ) (φ ) k k σ 2 a ( 2θφ+θ2 ) ( φ 2 ) k =0 (4) Although no closed-form expression exists for the more general situation, γ k can easily be calculated recursively from the ARMA parameters, as described in Appendix A The matrix Σ would then be formed directly from γ k via Equation () In either method, σa 2 and the ARMA parameters must be known or estimated using standard time series modeling procedures (eg, as described in Box et al, 994) If the process is not well represented by an ARMA model, then γ k must be estimated directly from the data in order to form the expressions in Equations () and (2) Unless otherwise noted, it will be assumed that the sample of data from which the ARMA parameters or γ k are estimated is large enough that the effects of modeling errors can be ignored In a later section of the paper we address the issue of model uncertainty Performance Comparison Monte Carlo simulation results comparing the autoregressive T 2 control chart with residual-based CUSUM and Shewhart individual charts for a variety of ARMA(,) processes are shown in Tables through 3 The residual-based charts involve applying standard (two-sided) CUSUM and Shewhart charts to the ARMA model residuals (see eg, Alwan and Roberts (988); Runger, Willemain, and Prabhu (995); and Superville and Adams (994)) We used 0,000 Monte Carlo trials in all simulations The data were generated as an ARMA(,) process, according to Equation (3), with σa 2 =and µ 0 =0 The true model was assumed known A range of values for φ and θ were used Only positive values for φ were considered, however, since negative values of φ generally result in processes that have negative autocorrelation, a situation that is not commonly encountered in industry A range of out-ofcontrol mean shift magnitudes, denoted µ through µ 4 in Tables through 3, were also considered The mean shifts are expressed in units of σ a (=) for each model The particular values of µ through µ 4 vary with the model parameters, and were chosen to span a range from what can be considered difficult to detect to what can be considered easy to detect An alternative would have been to use a single set of mean shift magnitudes (scaled in terms of the process standard deviation σ x ) consistent for each model It was felt that the comparisons would have been less informative had this been done For example, a mean shift of 25σ x (=288) for the AR() process with φ =05yields an interesting comparison between the various tests In contrast, a mean shift of 25σ x (=255) for the AR() process with φ =098 would have been detected by all tests on the initial observation with a probability of almost one Table 4 expresses the mean shifts considered for each model in units of σ x For each model and each mean shift magnitude, four different autoregressive T 2 charts (with p = 2, 5, 0, and 20), four different CUSUM charts (each with different reference value K and decision threshold H), and a Shewhart chart were compared All charts were designed so that the in-control ARL was 500 Results for other in-control ARLs exhibited the same general trends and are not shown A subsequent section of this paper discusses guidelines for selecting p and specifying the α that is needed to achieve a desired in-control ARL for the autoregressive T 2 chart The Shewhart chart signaled when an individual residual fell outside the control limits, which coincided with the upper and lower 000 percentiles Journal of Quality Technology Vol 34, No, January 2002

THE AUTOREGRESSIVE T 2 CHART FOR MONITORING UNIVARIATE AUTOCORRELATED PROCESSES 83 TABLE Comparison of Out-of-Control ARLs for Various Magnitude Mean Shifts and Various ARMA(,) Model Parameters (All charts have an in-control ARL of 500) Mean Shift T 2 ARL CUSUM ARL Shewhart ARL φ θ µ µ 2 p α µ µ 2 K H µ µ 2 µ µ 2 098 0 3 4 2 00037 279 679 030 756 3966 3256 2640 888 098 0 3 4 5 00049 2538 942 050 507 423 3273 098 0 3 4 0 00073 2786 226 075 354 3886 2330 098 0 3 4 20 0022 3027 505 00 267 3447 60 09 09 2 3 2 00034 35 2264 00 480 233 445 3005 763 09 09 2 3 5 00048 3428 487 020 000 2670 659 09 09 2 3 0 00073 272 560 030 756 2989 943 09 09 2 3 20 002 202 94 050 507 354 2423 09 05 3 4 2 00033 442 503 00 480 066 74 233 605 09 05 3 4 5 00046 398 332 020 000 8 725 09 05 3 4 0 0007 600 422 030 756 48 834 09 05 3 4 20 009 73 540 050 507 920 36 09 0 2 3 2 0003 89 65 00 480 069 606 3565 807 09 0 2 3 5 00046 2227 883 020 000 85 587 09 0 2 3 0 0007 2538 082 030 756 422 653 09 0 2 3 20 009 2624 28 050 507 932 98 09 05 2 3 2 0003 527 35 020 000 346 48 288 732 09 05 2 3 5 00046 67 42 030 756 354 7 09 05 2 3 0 0007 806 72 050 507 447 97 09 05 2 3 20 002 907 209 075 354 668 05 09 2 3 2 0003 975 36 020 000 283 59 35 226 05 09 2 3 5 00049 827 205 030 756 274 37 05 09 2 3 0 00073 74 29 050 507 325 26 05 09 2 3 20 0023 570 77 075 354 464 33 05 0 2 2 0003 20 80 020 000 3 22 989 48 05 0 2 5 00046 72 74 030 756 309 04 05 0 2 0 0007 00 65 050 507 386 92 05 0 2 20 002 993 63 075 354 562 00 05 05 2 3 2 0003 547 38 020 000 205 8 6 297 05 05 2 3 5 00047 500 8 030 756 89 99 05 05 2 3 0 00072 465 4 050 507 202 86 05 05 2 3 20 002 437 8 075 354 272 89 of the standard normal distribution The CUSUM chart signaled when either the upper or lower onesided CUSUM statistic exceeded H The upper and lower CUSUM statistics are calculated recursively via S t + = max{0,s t + + e t K} and S t = max{0,st e t K}, respectively, where e t is the residual for the observation at time t In the Monte Carlo simulations, the process, the residuals, and the test statistics were first allowed to reach steady-state before the mean shift was introduced The run length was taken to be the time from when the shift was introduced to when the chart first signaled Consequently, the ARL can be viewed as a steady-state ARL The steady-state ARL is an appropriate measure of performance (Lu and Reynolds (999)) if one is interested in measuring signaling properties for mean shifts that occur after the chart has been running for some time If one were more Vol 34, No, January 2002 wwwasqorg

84 DANIEL W APLEY AND FUGEE TSUNG TABLE Continued Mean Shift T 2 ARL CUSUM ARL Shewhart ARL φ θ µ 3 µ 4 p α µ 3 µ 4 K H µ 3 µ 4 µ 3 µ 4 098 0 5 6 2 00037 02 46 030 756 259 759 43 84 098 0 5 6 5 00049 96 22 050 507 955 7473 098 0 5 6 0 00073 33 432 075 354 89 553 098 0 5 6 20 0022 47 9 00 267 375 464 09 09 4 5 2 00034 67 448 00 480 975 73 44 05 09 09 4 5 5 00048 308 272 020 000 060 7070 09 09 4 5 0 00073 36 20 030 756 23 7692 09 09 4 5 20 002 9 29 050 507 445 659 09 05 5 6 2 00033 6 200 00 480 55 3876 8 32 09 05 5 6 5 00046 46 2 020 000 476 3240 09 05 5 6 0 0007 60 22 030 756 508 383 09 05 5 6 20 009 96 56 050 507 536 744 09 0 4 5 2 0003 47 256 00 480 392 275 486 666 09 0 4 5 5 00046 233 388 020 000 330 2035 09 0 4 5 0 0007 3 600 030 756 332 730 09 0 4 5 20 009 395 862 050 507 385 356 09 05 4 5 2 0003 33 27 020 000 76 446 25 7 09 05 4 5 5 00046 27 6 030 756 50 289 09 05 4 5 0 0007 34 28 050 507 3 84 09 05 4 5 20 002 45 56 075 354 24 26 05 09 4 5 2 0003 96 322 020 000 06 767 22 06 05 09 4 5 5 00049 45 46 030 756 85 576 05 09 4 5 0 00073 25 3 050 507 64 346 05 09 4 5 20 0023 24 45 075 354 45 75 05 0 3 4 2 0003 43 6 020 000 73 50 06 232 05 0 3 4 5 00046 44 76 030 756 58 393 05 0 3 4 0 0007 49 22 050 507 44 274 05 0 3 4 20 002 59 266 075 354 37 99 05 05 4 5 2 0003 38 4 020 000 8 602 54 32 05 05 4 5 5 00047 3 28 030 756 65 46 05 05 4 5 0 00072 34 47 050 507 48 292 05 05 4 5 20 002 43 87 075 354 36 68 interested in mean shifts that are present at the time the control chart is started, zero-state ARLs would be more appropriate In this situation, the T 2 chart ARL would be at least as large as p, since the T 2 statistic is not formed until at least p observations are available One could use a modified procedure in which the T 2 statistic is formed beginning from the initial observation, however, if {x 0,x,,x p+2 } are taken to be zero This is similar to what is done in residual-based charts, if the residual is calculated on the initial observation with {x 0,x,x 2,} taken to be zero This is not recommended unless the control limits over the initial period are modified appropriately so as to avoid an excessively large probability of a false alarm before the test statistic reaches its steady-state distribution All ARL results presented below are for the steady-state scenario In Table we show the out-of-control ARLs for the different charts with the various size mean shifts Generally speaking, the autoregressive T 2 chart (with appropriate p) outperforms the residualbased CUSUM and Shewhart charts when φ is large (eg 09 or 098) and θ is negative For example, when φ =09 and θ = 09, the autoregressive T 2 chart with p =20has substantially lower out-of- Journal of Quality Technology Vol 34, No, January 2002

THE AUTOREGRESSIVE T 2 CHART FOR MONITORING UNIVARIATE AUTOCORRELATED PROCESSES 85 TABLE 2 Comparison of P Values for Various Magnitude Mean Shifts and Various ARMA(,) Model Parameters (All charts have an in-control ARL of 500) Mean Shift T 2 P CUSUM P Shewhart P φ θ µ µ 2 p α µ µ 2 K H µ µ 2 µ µ 2 098 0 3 4 2 00037 043 079 030 756 000 000 046 082 098 0 3 4 5 00049 03 066 050 507 000 006 098 0 3 4 0 00073 023 054 075 354 00 038 098 0 3 4 20 0022 08 044 00 267 025 063 09 09 2 3 2 00034 005 08 00 480 000 000 04 046 09 09 2 3 5 00048 007 024 020 000 000 000 09 09 2 3 0 00073 006 022 030 756 000 000 09 09 2 3 20 002 006 08 050 507 000 000 09 05 3 4 2 00033 03 064 00 480 000 000 046 082 09 05 3 4 5 00046 029 065 020 000 000 000 09 05 3 4 0 0007 023 054 030 756 000 000 09 05 3 4 20 009 08 043 050 507 000 006 09 0 2 3 2 0003 02 040 00 480 000 000 04 046 09 0 2 3 5 00046 008 030 020 000 000 000 09 0 2 3 0 0007 006 023 030 756 000 000 09 0 2 3 20 009 006 08 050 507 000 000 09 05 2 3 2 0003 00 035 020 000 000 000 04 046 09 05 2 3 5 00046 008 029 030 756 000 000 09 05 2 3 0 0007 006 023 050 507 000 000 09 05 2 3 20 002 006 08 075 354 00 00 05 09 2 3 2 0003 006 020 020 000 000 000 04 046 05 09 2 3 5 00049 007 024 030 756 000 000 05 09 2 3 0 00073 006 02 050 507 000 000 05 09 2 3 20 0023 005 06 075 354 00 00 05 0 2 2 0003 002 0 020 000 000 000 002 04 05 0 2 5 00046 00 008 030 756 000 000 05 0 2 0 0007 00 006 050 507 000 000 05 0 2 20 002 002 005 075 354 000 00 05 05 2 3 2 0003 009 033 020 000 000 000 04 046 05 05 2 3 5 00047 008 029 030 756 000 000 05 05 2 3 0 00072 006 022 050 507 000 000 05 05 2 3 20 002 005 06 075 354 00 00 control ARLs than any of the CUSUMs or the Shewhart chart The exception is that when µ = 5,the Shewhart chart has a slightly lower ARL Another general trend is that the performance of the autoregressive T 2 chart relative to the CUSUM improves as the mean shift magnitude increases For example, consider the case where φ =09and θ =0 For mean shifts of 2 and 3, the CUSUMs with K = 0 and 02 have the lowest ARL, whereas for larger mean shifts of 4 and 5, the autoregressive T 2 chart with p = 2has substantially better ARL values than any of the CUSUMs When φ =05 and θ =0 or 05 (ie, when the autocorrelation is moderate) CUSUMs with appropriate K values perform substantially better for small to moderate mean shifts, and the autoregressive T 2 chart with p =2or5performs the best for large mean shifts One primary advantage of the autoregressive T 2 chart over the CUSUM chart is that, for a given ARMA model, the optimal choice of p does not seem to depend strongly on the magnitude of the mean Vol 34, No, January 2002 wwwasqorg

86 DANIEL W APLEY AND FUGEE TSUNG TABLE 2 Continued Mean Shift T 2 P CUSUM P Shewhart P φ θ µ 3 µ 4 p α µ 3 µ 4 K H µ 3 µ 4 µ 3 µ 4 098 0 5 6 2 00037 096 00 030 756 000 003 097 00 098 0 5 6 5 00049 09 099 050 507 028 067 098 0 5 6 0 00073 084 097 075 354 076 096 098 0 5 6 20 0022 075 093 00 267 09 099 09 09 4 5 2 00034 042 070 00 480 000 000 082 097 09 09 4 5 5 00048 056 086 020 000 000 000 09 09 4 5 0 00073 052 082 030 756 000 000 09 09 4 5 20 002 043 074 050 507 006 028 09 05 5 6 2 00033 089 098 00 480 000 000 097 00 09 05 5 6 5 00046 09 099 020 000 000 000 09 05 5 6 0 0007 084 097 030 756 000 003 09 05 5 6 20 009 074 093 050 507 028 067 09 0 4 5 2 0003 077 096 00 480 000 000 082 097 09 0 4 5 5 00046 065 09 020 000 000 000 09 0 4 5 0 0007 054 084 030 756 000 000 09 0 4 5 20 009 043 074 050 507 006 028 09 05 4 5 2 0003 070 093 020 000 000 000 082 097 09 05 4 5 5 00046 065 09 030 756 000 000 09 05 4 5 0 0007 054 084 050 507 006 028 09 05 4 5 20 002 043 074 075 354 038 076 05 09 4 5 2 0003 048 076 020 000 000 000 082 097 05 09 4 5 5 00049 057 086 030 756 000 000 05 09 4 5 0 00073 05 082 050 507 006 028 05 09 4 5 20 0023 042 074 075 354 038 076 05 0 3 4 2 0003 040 076 020 000 000 000 046 082 05 0 3 4 5 00046 029 065 030 756 000 000 05 0 3 4 0 0007 022 054 050 507 000 006 05 0 3 4 20 002 06 043 075 354 00 038 05 05 4 5 2 0003 068 092 020 000 000 000 082 097 05 05 4 5 5 00047 065 09 030 756 000 000 05 05 4 5 0 00072 054 084 050 507 006 028 05 05 4 5 20 002 042 074 075 354 038 076 shift whereas the optimal choice of K for the CUSUM does As evidence, consider the case when φ =09 and θ = 05 The autoregressive T 2 chart with p =5provides the lowest ARL for mean shifts of magnitude 3, 4, 5, or 6 In contrast, the optimal choices of K for the CUSUM range from approximately 0 to something substantially larger when the mean shift magnitude varies from 3 to 6 Of the Kvalues considered, K = 05 provided the lowest ARL for µ =6 The Shewhart chart, however, had a much lower ARL Since a Shewhart chart is the limiting form of a CUSUM as K becomes large, it is clear that the optimal CUSUM has K much larger than 05 The optimal K for µ =3performs poorly when µ = 6, and vice versa Similar observations apply to almost all of the cases considered For φ =05 and θ = 09, the optimal p value for the autoregressive T 2 chart was approximately 20 for all mean shifts ranging from µ =2to 5 (for µ =5,p =0provided a slightly lower ARL, however) In contrast, the optimal K for the CUSUM varied from 03 for µ =2 to very large (ie, the Shewhart limiting form) for µ =5 Invariance, with respect to the mean shift magnitude, of the optimal value of p for the autoregressive Journal of Quality Technology Vol 34, No, January 2002

THE AUTOREGRESSIVE T 2 CHART FOR MONITORING UNIVARIATE AUTOCORRELATED PROCESSES 87 TABLE 3 Comparison of P 5 Values for Various Magnitude Mean Shifts and Various ARMA(,) Model Parameters (All charts have an in-control ARL of 500) Mean Shift T 2 P 5 CUSUM P 5 Shewhart P 5 φ θ µ µ 2 p α µ µ 2 K H µ µ 2 µ µ 2 098 0 3 4 2 00037 044 079 030 756 00 003 047 082 098 0 3 4 5 00049 039 073 050 507 005 07 098 0 3 4 0 00073 03 063 075 354 07 048 098 0 3 4 20 0022 024 05 00 267 030 067 09 09 2 3 2 00034 006 09 00 480 000 000 03 078 09 09 2 3 5 00048 07 053 020 000 000 000 09 09 2 3 0 00073 03 080 030 756 000 00 09 09 2 3 20 002 032 082 050 507 00 005 09 05 3 4 2 00033 032 065 00 480 000 000 049 084 09 05 3 4 5 00046 046 08 020 000 000 000 09 05 3 4 0 0007 04 076 030 756 000 00 09 05 3 4 20 009 032 065 050 507 003 0 09 0 2 3 2 0003 03 042 00 480 000 000 05 047 09 0 2 3 5 00046 03 039 020 000 000 000 09 0 2 3 0 0007 0 032 030 756 000 002 09 0 2 3 20 009 009 025 050 507 003 00 09 05 2 3 2 0003 08 052 020 000 000 007 08 054 09 05 2 3 5 00046 022 06 030 756 004 028 09 05 2 3 0 0007 09 055 050 507 07 060 09 05 2 3 20 002 05 044 075 354 026 072 05 09 2 3 2 0003 009 027 020 000 000 00 027 07 05 09 2 3 5 00049 06 048 030 756 00 007 05 09 2 3 0 00073 023 067 050 507 006 024 05 09 2 3 20 0023 022 067 075 354 0 038 05 0 2 2 0003 005 029 020 000 000 00 004 020 05 0 2 5 00046 004 027 030 756 000 009 05 0 2 0 0007 004 022 050 507 003 026 05 0 2 20 002 004 07 075 354 005 034 05 05 2 3 2 0003 05 044 020 000 000 002 07 05 05 05 2 3 5 00047 06 05 030 756 002 02 05 05 2 3 0 00072 04 045 050 507 008 032 05 05 2 3 20 002 0 034 075 354 03 043 T 2 chart has important practical significance, since it is generally difficult to identify one single mean shift magnitude of interest It is desirable that a test be effective at detecting a variety of mean shift magnitudes Guidelines for selecting the optimal value of p for the autoregressive T 2 chart are provided in a following section Another advantage of the autoregressive T 2 chart over the CUSUM chart is that the probability of detecting the mean shift immediately after it occurs is usually much larger for the T 2 chart Tables 2 and 3 show the probability of detecting the mean shift on the first observation (denoted by P ) and on or before the fifth observation (denoted by P 5 ) after the occurrence of the shift For many of the cases, P for the autoregressive T 2 chart is almost as large as that for the Shewhart chart In this sense, the autoregressive T 2 chart possesses desirable properties of both the Shewhart chart (fast detection of large shifts) and the CUSUM (good protection against small shifts) For many of the models, a single T 2 chart with an Vol 34, No, January 2002 wwwasqorg

88 DANIEL W APLEY AND FUGEE TSUNG TABLE 3 Continued Mean Shift T 2 P 5 CUSUM P 5 Shewhart P 5 φ θ µ 3 µ 4 p α µ 3 µ 4 K H µ 3 µ 4 µ 3 µ 4 098 0 5 6 2 00037 096 00 030 756 008 020 097 00 098 0 5 6 5 00049 094 099 050 507 045 077 098 0 5 6 0 00073 089 098 075 354 08 097 098 0 5 6 20 0022 080 095 00 267 092 099 09 09 4 5 2 00034 043 070 00 480 000 000 098 00 09 09 4 5 5 00048 088 099 020 000 000 00 09 09 4 5 0 00073 099 00 030 756 002 006 09 09 4 5 20 002 099 00 050 507 07 048 09 05 5 6 2 00033 089 098 00 480 000 000 098 00 09 05 5 6 5 00046 097 00 020 000 000 00 09 05 5 6 0 0007 096 00 030 756 004 0 09 05 5 6 20 009 090 099 050 507 035 07 09 0 4 5 2 0003 078 096 00 480 000 000 082 097 09 0 4 5 5 00046 074 094 020 000 00 004 09 0 4 5 0 0007 065 090 030 756 007 02 09 0 4 5 20 009 053 082 050 507 030 06 09 05 4 5 2 0003 085 098 020 000 037 080 088 099 09 05 4 5 5 00046 092 099 030 756 073 096 09 05 4 5 0 0007 090 099 050 507 092 00 09 05 4 5 20 002 08 098 075 354 096 00 05 09 4 5 2 0003 057 083 020 000 007 027 097 00 05 09 4 5 5 00049 083 098 030 756 028 062 05 09 4 5 0 00073 096 00 050 507 058 088 05 09 4 5 20 0023 096 00 075 354 077 097 05 0 3 4 2 0003 073 097 020 000 08 068 057 090 05 0 3 4 5 00046 074 097 030 756 050 09 05 0 3 4 0 0007 065 095 050 507 076 098 05 0 3 4 20 002 053 089 075 354 083 099 05 05 4 5 2 0003 079 097 020 000 03 043 086 098 05 05 4 5 5 00047 085 098 030 756 040 077 05 05 4 5 0 00072 082 098 050 507 070 093 05 05 4 5 20 002 07 094 075 354 080 097 TABLE 4 Mean Shift Magnitudes Used in the Simulations, Expressed in Units of σ x (The values for µ, µ 2, µ 3, and µ 4 are given in Tables 3 for each of the eight ARMA(,) models) φ θ σ x µ /σ x µ 2 /σ x µ 3 /σ x µ 4 /σ x 098 0 503 060 080 099 9 09 09 425 047 07 094 8 09 05 336 089 9 49 78 09 0 229 087 3 74 28 09 05 36 47 22 295 368 05 09 90 05 58 20 263 05 0 5 087 73 260 346 05 05 53 3 96 262 327 Journal of Quality Technology Vol 34, No, January 2002

THE AUTOREGRESSIVE T 2 CHART FOR MONITORING UNIVARIATE AUTOCORRELATED PROCESSES 89 appropriate choice of p performed comparably to the optimal CUSUM for small mean shifts and the Shewhart chart for large mean shifts An Autoregressive Decomposition of the T 2 Statistic Consider an nth order autoregressive model fit to y t = x t µ, even though y t may not truly be an ARMA process The AR(n) model form is y t = β,n y t + β 2,n y t 2 + + β n,n y t n + e t,n, (5) where β,n,β 2,n,,β n,n are the AR(n) parameters which are optimal in the sense of minimizing the variance of e t,n, the residual error for the AR(n) model Let σn 2 denote the variance of e t,n In Appendix B we show that Σ can be factored as Σ = B DB, (6) where β, β 2,2 β p,p β,2 β p 2,p B = 0 β,p is an upper triangular matrix containing the parameters for the set of AR(n) models with n =, 2,,p, and D = σ 2 0 σ 2 σ 2 p (7) is a diagonal matrix containing the variances of the residual errors for the various order AR(n) models Here, e t,0 is defined as y t, and σ0 2 is defined as γ 0, the variance of x t The factorization in Equation (6) is a scaled version of the Cholesky factorization of Σ Substituting Equation (6) into Equation (2), it follows that T 2 t =[X t µ 0 ] [ B DB ] [Xt µ 0 ] =[X t µ 0 ] [ BD B ] [X t µ 0 ] = p n=0 e 2 t p++n,n σn 2 (8) From Equation (8), the autoregressive T 2 control chart has a close relationship to residual-based control charts The value Tt 2 is the sum of the squares of the residual errors for various order AR models, scaled by the residual variances (the residual for the n th order model is delayed by p n observations) This includes the residual for the AR(0) model, which is just the original process y t Note that for this interpretation to be valid it is not necessary that y t be an ARMA process Similar T 2 decompositions for true multivariate processes are well known (see eg, Hawkins (993) and Mason, Tracy, and Young (997)) When the process is in-control, each of the p terms in the summation in Equation (8) follows a chi-square distribution with one degree-of-freedom This follows from the fact that each e i,n,being a linear combination of Gaussian random variables, is itself Gaussian The square of a zero-mean Gaussian random variable, divided by its variance, is a chi-square random variable with one degree-of-freedom Furthermore, we show in Appendix B that the covariance matrix of the e i,n terms in Equation (8) is the diagonal matrix in Equation (7) Hence, they are uncorrelated Since they are Gaussian, they are also independent, and so are their squared values Thus, all p terms in Equation (8) are chi-square random variables with one degree-of-freedom and are independent of each other Thus, Tt 2 is chi-square distributed with p degrees-offreedom The relationship to residual-based control charts becomes even more apparent when x t is an AR() process x t µ = φ (x t µ)+a t Since the optimal AR(n) model for n =, 2,,p isjust the true AR() model, B and D in Equation (6) have the structure φ 0 0 φ B = 0 0 φ and D = σ 2 x σ 2 a (9) σa 2 where σ 2 x = γ 0 is the variance of x t Consequently, the decomposition in Equation (8) becomes T 2 t = (x t p+ µ 0 ) 2 σ 2 x p 2 + i=0 e 2 t i σa 2, Vol 34, No, January 2002 wwwasqorg

90 DANIEL W APLEY AND FUGEE TSUNG TABLE 5 Relationship Between the Optimal p (Denoted p ) and the AR Model Order (Denoted n )Above Which the Parameter Magnitudes Drop Below 0 (The values for p are approximate, from Table ) φ 098 090 090 090 090 050 050 050 θ 000 090 050 000 050 090 000 050 p 2 20 5 2 2 5 20 2 5 n + 2 8 5 2 4 8 2 5 β,9 098 80 40 090 040 40 050 00 β 2,9 000 6 070 000 020 25 000 050 β 3,9 000 44 035 000 00 2 000 025 β 4,9 000 29 07 000 005 00 000 02 β 5,9 000 5 009 000 003 089 000 006 β 6,9 000 02 004 000 00 079 000 003 β 7,9 000 090 002 000 00 070 000 002 β 8,9 000 080 00 000 000 062 000 00 β 9,9 000 070 00 000 000 055 000 000 β 0,9 000 06 000 000 000 048 000 000 β,9 000 053 000 000 000 042 000 000 β 2,9 000 046 000 000 000 036 000 000 β 3,9 000 039 000 000 000 030 000 000 β 4,9 000 032 000 000 000 025 000 000 β 5,9 000 026 000 000 000 020 000 000 β 6,9 000 020 000 000 000 06 000 000 β 7,9 000 04 000 000 000 0 000 000 β 8,9 000 008 000 000 000 007 000 000 β 9,9 000 003 000 000 000 003 000 000 where e j is the residual error for the true model at observation number j and is given by e j = (x j µ 0 ) φ (x j µ 0 ) Consequently, when the process is AR(), Tt 2 is a moving average of the squares of the residuals added to the square of (x t p+ µ 0 )/σ x Autoregressive T 2 Chart Design To implement the autoregressive T 2 control chart, the user must specify p and α Wefirst discuss guidelines for selecting p Asdiscussed previously, the simulation results indicate that the value of p that provides the lowest out-of-control ARL does not depend strongly on the size of the mean shift In other words, p can be selected based only on the autocovariance structure of the process (equivalently, on the ARMA parameters if x t can be represented as an ARMA process) Table 5 shows what were somewhat subjectively determined to be the optimal p values (from the four values of p (2, 5, 0, and 20) that were considered) for the eight different ARMA(,) models used in the simulation results of Tables through 3 The optimal values for p are denoted by p For some models, a single value of p was optimal for all mean shifts considered For others, different values of p provided a lower out-of-control ARL, depending on the size of the mean shift It was generally true, however, that for a given model a single value of p was close to optimal for all mean shifts considered For example, for φ =05and θ = 05, the optimal values of p for mean shifts of size 2, 3, 4, and 5 were 20, 0, 5, and 5, respectively For the smaller mean shifts of size 2 and 3, however, the ARL for p =5 was only slightly higher than for p =0and 20 In this case, p =5was designated the optimal value Foragiven model, there appears to be a simple relationship between p and the model characteristics Table 5 also shows the parameters, defined in Equation (5), of a high order AR(n)model (ie, n = 9) fit to the autocovariance function of each ARMA model More specifically, the AR(9) model parameters are the solution to the Yule-Walker equations of order 9 (Box et al, 994) The parameters can be easily de- Journal of Quality Technology Vol 34, No, January 2002

THE AUTOREGRESSIVE T 2 CHART FOR MONITORING UNIVARIATE AUTOCORRELATED PROCESSES 9 termined by finding B in the Cholesky factorization in Equation (6) with p = 20 The AR(9) parameters are then contained in the last column of B If the model is invertible, the parameters of a high order AR model will decay to zero as the lag increases (Box et al, 994) Let n be the lag after which the magnitude of the parameters drops below some small value, say 0 In other words, n is defined as the smallest integer such that β j, < 0 for all j>n Thus, n is essentially the AR model order that is needed to capture the dynamics of the process From the discussion on the autoregressive T 2 decomposition of the previous section, if p is set as n +, then Σ will also capture the dynamics of the process Setting p = n +appears to be an effective rule-of-thumb for selecting the optimal value of p This follows from Table 5, which shows that for all eight models considered, p is close to n + As discussed above, n can be found by forming Σ with a relatively large p (eg, p =20or30), taking its Cholesky factorization, and inspecting the last column of B Note that if the process truly is AR(n), then p = n +For example, p =2for each of the three AR() models In the preceding method for finding n we assume that the true process autocovariance function (or, equivalently, the true ARMA model describing the process) is accurately known, and then attempt to approximate the process characteristics using a sufficiently high-order AR model In situations where the autocovariance or ARMA model is estimated from limited data and cannot be assumed to be accurate, one could still apply this method using the estimated ARMA model In this situation, however, the following alternative method for finding n may be preferable The alternative method is to fit various order AR models to the data, until a further increase in the model order (beyond n )isnolonger statistically significant; p would again be set as n + After selecting p, one potential means of selecting α is to fix the false alarm probability If the threshold for Tt 2 is set as the α percentile of the chi-square distribution with p degrees-of-freedom, then clearly the false alarm probability will be α for any given isolated time There is no simple relationship, however, between α and the in-control ARL The sequence Tt 2 can have high autocorrelation, in particular if p is large, and the in-control ARL may be substantially larger than /α This is evident from Table, where the in-control ARL (denoted by ARL 0 )was500 for all cases Empirically, we have observed that for a given ARMA model (ie, a given autocovariance structure) and p, the relationship between ARL 0 and α is very close to log-linear In other words, log(arl 0 ) = c 0 c log(α), (0) where c 0 and c are constants that depend on p and the ARMA model parameters Tables 6 and 7 give TABLE 6 Values for c 0 for Various p and ARMA(,) Parameters (To Be Used in Equation (0)) (φ, θ) (098,0) (09, 09) (09, 05) (09,0) (09,05) (05, 09) (05, 05) (05,0) 2 069 0864 0833 0709 0706 0624 0572 052 3 202 009 0977 0900 0900 0835 088 0792 4 36 33 093 06 052 0996 029 0998 5 40 250 29 27 20 67 225 203 6 505 385 349 357 323 309 37 343 7 68 509 482 486 462 460 490 464 p 8 695 63 600 606 554 577 597 570 9 774 738 726 723 659 702 695 673 0 854 834 830 832 746 807 798 765 2 983 975 968 978 90 962 928 900 4 2097 2093 2077 2090 2028 2086 2043 209 6 298 222 270 29 240 289 249 236 8 2296 2320 2266 2299 2273 2293 2242 2239 20 238 2407 234 2377 2364 2378 2325 2323 Vol 34, No, January 2002 wwwasqorg

92 DANIEL W APLEY AND FUGEE TSUNG TABLE 7 Values for c for Various p and ARMA(,) Parameters (To Be Used in Equation (0)) (φ, θ) (098,0) (09, 09) (09, 05) (09,0) (09,05) (05, 09) (05, 05) (05,0) 2 098 094 0940 095 0950 097 0972 0975 3 093 0938 0938 0944 0944 0964 0955 0957 4 0909 0935 0937 0938 0937 0957 0942 0945 5 0905 093 0934 0930 093 0948 0929 0933 6 090 0925 0927 0922 0926 0939 0920 0924 7 0897 096 097 092 0920 0926 093 097 p 8 0893 0907 0905 0903 095 094 0906 090 9 0889 0897 0892 0893 0909 0903 0899 0906 0 0887 0888 0884 0886 0904 0895 0895 090 2 0883 0879 088 0879 0896 0887 0890 0895 4 0879 0872 0878 0874 0888 0882 0887 0890 6 0876 0867 0877 0870 088 0878 0884 0887 8 0873 0863 0875 0868 0875 0874 0883 0883 20 0870 0860 0874 0865 0869 087 088 0880 the constants c 0 and c, respectively, for each of the ARMA models and for various values of p ranging from 2 to 20 The constants were identified empirically by running Monte Carlo simulations for the eight ARMA(,) models considered in Table, with ARL 0 ranging from 200 to 000 For larger values of p, c 0 and c are relatively insensitive to the ARMA parameters and depend predominantly on p For example, for p = 0 and 20, the approximations p =0: p =20: log(arl 0 ) = 83 0892 log(α) log(arl 0 ) = 2364 087 log(α) were always within 29% and 23%, respectively, of the true ARL 0 The suggested method for selecting the threshold for T 2 t is to specify a desired ARL 0, determine the approximate c 0 and c from Tables 6 and 7, determine α from Equation (0), and then set the threshold as the α percentile of the chi-square distribution with p degrees-of-freedom We point out that the procedures for selecting the optimal p and the thresholds are based on a number of empirical approximations and limited simulation for a specific set of ARMA parameters As such, they can only be expected to provide approximate in-control ARLs Robustness to Model Uncertainty and Unknown Σ Up to this point, only the case where Σ is assumed known has been considered Essentially, this means that enough data are available so that an estimate, denoted by Σ, issufficiently close to Σ It is often the case that the available data is insufficient to warrant this assumption In this event, Tt 2 of Equation (2) with Σ replaced by Σ does not follow a chi-square distribution The primary problem with this is that the actual in-control ARL may then be substantially shorter than what is desired, resulting in too many false alarms This is no less a problem in residual-based control charts, and is analogous to having insufficient data to accurately estimate the ARMA model parameters, the model order, and σa 2 If the estimated ARMA model is inaccurate, then the residuals will no longer be uncorrelated, their variance will differ from σ a, 2 and the incontrol ARL of (for example) a CUSUM on the residuals will not be as desired Adams and Tseng (998), Lu and Reynolds (999), and Apley and Shi (999) have demonstrated the effects of model estimation errors on the ARLs for various residual-based charts Adams and Tseng (998) provide guidelines for determining the sample size (from which the model parameters are estimated) such that the effects are negligible Although the adverse effects of model estimation errors on the ARLs of residual-based control charts are known, no approaches for modifying the control limits to account for the model uncertainty have been proposed A natural extension of the autoregressive T 2 control chart, when Σ must be estimated from Journal of Quality Technology Vol 34, No, January 2002

THE AUTOREGRESSIVE T 2 CHART FOR MONITORING UNIVARIATE AUTOCORRELATED PROCESSES 93 limited data, provides one means of accommodating model uncertainty Suppose N observations {x,x 2,,x N } from Phase I operation (Sullivan and Woodall (996)) are available, from which Σ is to be estimated Consider, as an estimate of Σ, Σ = N p + N [X t µ 0 ][X t µ 0 ], t=p where µ 0 =[ µ 0 µ 0 µ 0 ], with µ 0 the usual sample average of {x,x 2,,x N } The suitability of the above estimate of Σ requires that the process is in-control when the Phase I data are collected When this is not the case, some form of rational subgrouping is recommended when estimating Σ Sullivan and Woodall (996), to which the reader is referred for details, provides an excellent discussion of strategies for more robustly estimating Σ during Phase I operation In related work, Boyles (2000) discusses estimating ARMA model parameters when assignable causes are present during Phase I Assuming that our estimate of Σ is used and is suitable, the T 2 statistic would be T 2 t =[X t µ 0 ] Σ [X t µ 0 ] () The distribution of this statistic is better approximated as an F -distribution (scaled by an appropriate constant), which takes into account the uncertainty in Σ, than as a chi-square distribution The F -distribution is still an approximation, however, since {X p, X p+,,x N } are not independent of each other, and Σ will not exactly follow a Wishart distribution Approximating by an F -distribution, the threshold for Tt 2 would be (N p +)p F (p, N 2p +2, α), (2) N 2p +2 where F (j, k, ν) denotes the ν percentile of the F - distribution with j numerator degrees-of-freedom and k denominator degrees-of-freedom As N gets larger, the expression in Equation (2) approaches the α percentile of the chi-square distribution with p degrees-of-freedom This is reasonable, since as N gets large, Σ approaches Σ As N becomes smaller, the value in Equation (2) increases monotonically, and the threshold for Tt 2 gets larger This serves to prevent an in-control ARL that is substantially shorter than desired There exists an analogous decomposition of Equation () for the case that Σ is estimated, similar to the decomposition in Equation (8) that was derived for Σ known The form of the decomposition is identical to Equation (8), except that the e j,n terms are the residual errors for AR(n) models whose coefficients (the β j,n s from Equation (5)) are estimated from the Phase I data using linear least squares Likewise, each σn 2 term is the sample variance of the AR(n) model residuals over the Phase I data from which the parameters are estimated The proof is a straightforward extension of that in Appendix B, where the correlation inner-product associated with minimum mean square error predictors is replaced by the sample correlation inner-product associated with linear least squares regression Mason et al (997) and Mason and Young (999) present similar decompositions for multivariate processes Table 8 shows Monte Carlo simulation results that demonstrate how control limits based on Equation (2) help in preventing an undesirably short incontrol ARL The actual in-control ARL for various residual-based CUSUM and autoregressive T 2 charts are shown when the necessary quantities (the ARMA model for the CUSUM chart and Σ for the T 2 chart) are estimated from limited data with N = 50and N = 00 For each case, 20,000 Monte Carlo trials were run For each trial, N observations were generated from the true ARMA model (with parameters given in Table 8), the model was estimated from these data, and the test was implemented on TABLE 8 True ARL 0 for Residual-Based CUSUM and T 2 Charts When the Model is Estimated from N =50and N = 00 Observations (The desired ARL 0 in all cases is 500) true parameters CUSUM ARL 0 T 2 ARL 0 φ θ N K=02 σ a K=05 σ a p =2 p =5 p =0 09 0 50 340 265 499 585 70 09 0 00 400 338 467 472 454 05 05 50 372 275 707 704 98 05 05 00 402 352 578 595 555 Vol 34, No, January 2002 wwwasqorg

94 DANIEL W APLEY AND FUGEE TSUNG additional data generated from the true model until the test signaled a false alarm For the T 2 charts, only Σ was estimated For the residual-based CUSUMs, the model order, the parameters, and σa 2 were estimated For simplicity, µ 0 was assumed to be zero without loss of generality To estimate the model parameters, the Matlab ARMAX routine, which is based on nonlinear least squares, was used Various ARMA(n, m) models were fit to the data, and Akaike s Final Prediction Error criterion (Ljung (987)) was used to estimate model order The set of models that were fitted was AR(), AR(2), AR(3), AR(4), ARMA(,), ARMA(2,), ARMA(3,2), and ARMA(4,3) In all cases, the tests were designed with a desired ARL 0 of 500 in mind The CUSUMs were designed as if the estimated model equaled the true model For reference values of K = 02 σ a and 05 σ a, the decision thresholds were H = 00 σ a and 507 σ a, respectively For the T 2 charts with p =2, 5, and 0, the values of α in Table were used The thresholds for Tt 2 were then set according to Equation (2) For N = 50, the CUSUM ARL 0 s were substantially lower than 500, in particular when K = 05 Even for N = 00, the CUSUM ARL 0 s were much lower than 500 Although the T 2 ARL 0 s were at times substantially larger than 500, they were never substantially lower Consequently, when there is model uncertainty due to limited data, the autoregressive T 2 chart with threshold given by Equation (2) can be considered more robust with respect to excessive false alarms Use of Equation (2) should be viewed only as an approximate means of widening the control limits to account for model uncertainty, since the assumed F -distribution of () is only an approximation For arbitrary p, the covariance matrix would therefore be of the form φ φ 2 φ p φ φ φ p 2 Σ = σ2 a φ 2 φ 2 φ φ p 3 φ p φ p 2 φ p 3 with φ =0847 Taking the Cholesky factor of Σ or using Equation (9) directly gives 0847 0 0 0847 B = 0 0 0847 By inspection of the last column of B, itisclear that β j, =0for j>n = Using the guidelines presented earlier, p = n + =2is recommended The covariance matrix for p = 2reduces to σ2 a Σ = φ 2 ( φ φ ) = ( 354 300 300 354 ) From Tables 6 and 7, c 0 =0709 and c =095 for a similar AR() model with φ =09 Using these in Equation (0) gives α = 0003 for an approximate ARL 0 of 500 The test threshold is χ 2 ( α, p) = 55 An autoregressive T 2 chart with p =2and α = 0003 was applied to the simulated AR() process, Illustrative Example Consider the chemical process described in Montgomery and Mastrangelo (99), which is represented as an AR() process with φ = 0847 For simplicity, assume that σa 2 =0, and that the data is centered so that the in-control mean is zero Suppose one wishes to apply an autoregressive T 2 chart with ARL 0 = 500 We assume the model is accurate enough that the χ 2 control limit can be used Two design parameters, p and α, must be selected Applying Equation (4) with θ =0gives σ2 a γ k = φ 2 φ k FIGURE Autoregressive T 2 Chart Example Journal of Quality Technology Vol 34, No, January 2002

THE AUTOREGRESSIVE T 2 CHART FOR MONITORING UNIVARIATE AUTOCORRELATED PROCESSES 95 where a mean shift of magnitude 45 was introduced at observation number 2 Since the process standard deviation is σ x = γ /2 0 = 88, the mean shift is approximately 24 standard deviation units The simulated process and the T 2 statistics are shown in Figure In this case, the T 2 chart signaled on the first observation following the mean shift Conclusions A method for monitoring autocorrelated processes, termed the autoregressive T 2 chart, has been presented in this paper The terminology results from the fact that the T 2 statistic can be decomposed into the sum of the squares of the residual errors for various order autoregressive models fit to the process Hence, there is a close relationship between residual-based control charts and the autoregressive T 2 chart The performance of the autoregressive T 2 chart, in terms of the ARL, has been compared to residualbased CUSUM and Shewhart charts for a number of ARMA(,) processes For certain ranges of the ARMA parameters, the autoregressive T 2 chart performs substantially better than either the CUSUM or Shewhart charts In general, for moderate to large mean shifts, the autoregressive T 2 chart is superior for most of the processes considered For small mean shifts the CUSUM chart is often superior, depending on the process parameters Even in situations when the CUSUM has a lower out-of-control ARL, the autoregressive T 2 chart usually has a much higher probability of detecting the mean shift within the first few observations following the shift, and is similar to a residual-based Shewhart chart in this respect One primary advantage of the autoregressive T 2 chart over a residual-based CUSUM is that a single T 2 chart design is often nearly optimal for a wide range of mean shift sizes In contrast, the optimal CUSUM design almost always depends on the mean shift size of interest A CUSUM chart that is optimal for small mean shifts may perform poorly for large mean shifts, and vice-versa Guidelines for designing the autoregressive T 2 chart have been provided An additional advantage is that the autoregressive T 2 chart provides some robustness with respect to an excessive number of false alarms when there is large uncertainty in the process model Specifically, when there are limited data available for estimating an ARMA process model, a residual-based CUSUM may have substantially lower in-control ARL than what is intended The autoregressive T 2 chart possesses a natural mechanism for taking into account model uncertainty due to limited data In the presence of model uncertainty the actual in-control ARL for the autoregressive T 2 chart is never substantially lower than the intended ARL in the examples considered In some cases it is, however, substantially higher There may exist better approximations to the distribution of Tt 2 in Equation () that will allow more accurate specification of the in-control ARL when data are limited Although SPC for autocorrelated processes and residual-based control charts have been widely studied, how to modify the control chart to account for model uncertainty has not been studied and deserves more attention In this paper we have investigated the autoregressive T 2 chart performance only for the case of process mean shifts Typically, it is desirable to detect variance changes as well It is likely that the autoregressive T 2 chart would also be effective at detecting variance increases This speculation follows from the decomposition of Tt 2 as the sum of the squares of the autoregressive residual errors Appendix A This appendix provides an algorithm for calculating Σ, when y t follows the ARMA(n, m) model of Equation (3) Let g j, j =0,, 2,, denote the Green s function for the ARMA model Green s function is essentially the output of the ARMA model when the input sequence a t is a single pulse of unit magnitude at time t = 0(see, for example, Pandit and Wu (990)) It can be calculated recursively via n φ i g j i θ j j m i= g j =, n φ i g j i m<j i= with initial conditions g 0 =and g j =0for j<0 It can be shown (Pandit and Wu, (990)) that the autocovariance function of x t is given by γ k = σa 2 g j g j+k j=0 Since g j decays exponentially for stable ARMA models, the infinite summation can be truncated Appendix B This appendix provides a simple proof of Equation (7) Define Y t =[y t p+ y t p+2 y t ] and E t = [e t p+,0 e t p+2, e t,p ] with e i,j as in Equation Vol 34, No, January 2002 wwwasqorg