Robustness of the EWMA control chart to non-normality Connie M Borror; Douglas C Montgomery; George C Runger Journal of Quality Technology; Jul 1999; 31, 3
國立雲林科技大學 工業工程與管理所 Introduction Individual measurements occur frequently in the chemical and process industries. The traditional method of dealing with the case of n=1 is to use the Shewhart individuals control chart to monitor the process mean. The individuals control chart has two widely-cited disadvantages : (1) the chart is not very sensitive to small shifts in the process mean. (2) the performance of the chart can be adversely affected if the observations are not normally distribution. It is certainly true that non-normality of the process data is often not a significant concern if the X-bar control chart is used to monitor the mean.
國立雲林科技大學 工業工程與管理所 Introduction In this paper, we show that the ARL performance of the Shewhart individuals control chart when the process is in control is very sensitive to the assumption of normality. We suggest the EWMA control chart as an alternative to the individuals chart for non-normal data. We show that, in the non-normal case, a properly designed EWMA control chart will have an in-control ARL that is reasonably close to the value of for the individuals chart for normally distributed date. For all cases, the ARL’s were computed using the Markov chain method.
國立雲林科技大學 工業工程與管理所 Background Information-EWMA The EWMA is defined as Where x i is the current observation and λ, smoothing parameter, is a constant for 0 ≦ λ ≦ 1 The control limits for the EWMA control chart are Where L determines the width of the control limits
國立雲林科技大學 工業工程與管理所 Background Information-EWMA For large values of i, the steady-state EWMA control limits are If any point exceeds the control limits, the process is assumed to be out of control.
國立雲林科技大學 工業工程與管理所 Background Information-Skewed and symmetric distribution To study the robustness of the EWMA control chart and the individuals control chart to normality assumption, both skewed and symmetric distribution were examined. Symmetric distribution : t distribution Let k is degree of freedom The Mean is 0 The Variance is k/(k-2)
國立雲林科技大學 工業工程與管理所 Various t distribution and normal distribution with the same mean and variance
國立雲林科技大學 工業工程與管理所 Background Information-Skewed and symmetric distribution Skewed distribution : Gamma distribution Let α=0.5, 1, 2, 3, and 4, while holding β=1
國立雲林科技大學 工業工程與管理所 Various Gamma distribution and normal distribution with the same mean and variance
國立雲林科技大學 工業工程與管理所 Results The normal-theory ARL for individuals control chart with 3σ is known to be For the EWMA, we can determine the values of λ and L to obtain approximately the same ARL of Value of 0.05, 0.1, and 0.2 were chosen for λ, with the corresponding value of 2.492, 2.703, and 2.86, respectively, chosen for L.
國立雲林科技大學 工業工程與管理所 In-Control ARL’s for EWMA-Gamma EWMAShewhart λ L Normal Gam(4,1) Gam(3,1) Gam(2,1) Gam(1,1) Gam(0.5,1) The Best Case
國立雲林科技大學 工業工程與管理所 Out-of-control ARL’s for the EWMA-Gamma Shift ( Number of Standard Deviations ) EWMA λ=0.05 L=2.492 Normal Gam(4,1) Gam(3,1) Gam(2,1) Gam(1,1) Gam(0.5,1) EWMA λ=0.1 L=2.703 Normal Gam(4,1) Gam(3,1) Gam(2,1) Gam(1,1) Gam(0.5,1)
國立雲林科技大學 工業工程與管理所 Out-of-control ARL’s for the EWMA-Gamma Shift ( Number of Standard Deviations ) EWMA λ=0.2 L=2.86 Normal Gam(4,1) Gam(3,1) Gam(2,1) Gam(1,1) Gam(0.5,1) Shewhart Normal Gam(4,1) Gam(3,1) Gam(2,1) Gam(1,1) Gam(0.5,1)
國立雲林科技大學 工業工程與管理所 In-Control ARL’s for EWMA-t EWMAShewhart λ L Normal t t t t t t t t t
國立雲林科技大學 工業工程與管理所 Out-of-control ARL’s for the EWMA-t ShiftEWMA λ=0.05 L=2.492 EWMA λ=0.1 L=2.703 EWMA λ=0.2 L=2.86 Shewhart 0.5 t50~10 ( 26 ) N ( 26.5 ) t8~4 ( 27 ) N 、 t50~40 ( 28.3 ) t30~4 ( 28.4~30 ) N ( 36.2 ) t ( 36~41 ) N ( ) t50~6 ( 137~73 ) t4 ( 63 ) 1 N ( 10.8 ) t ( 11 ) N 、 t ( 9.8 ) N 、 t50~20 ( 9.8 ) t15~4 ( 9.9~10.3 ) N ( 44 ) t50~8 ( 43~39 ) t6~4 ( 38 ) 1.5 N ( 6.8 ) t ( 6.7 ) N 、 t ( 5.8 ) N 、 t ( 5.2 ) N 、 t50~20 ( 15 ) t15~4 ( 16~19 ) 2 N、t(5)N、t(5) N 、 t ( 4.2 ) N 、 t ( 3.6 ) N ( 6.3 ) t50~4 ( 6.4~9 ) 2.5 N、t(4)N、t(4) N 、 t ( 3.3 ) N 、 t ( 2.6 ) N ( 3 ) t50~4 ( 3.3~4 ) 3 N ( 3.4 ) t ( 3.3 ) N 、 t ( 2.8 ) N 、 t ( 2.3 ) N、t(2)N、t(2) EWMA is better than Shewhart
國立雲林科技大學 工業工程與管理所 Comparing three EWMA control chart designs There have been many suggestion in the literature for designing an EWMA control chart. The table compares three EWMA control chart designs. 1st column : λ=0.1 and L=2.7 ( Montgomery, 1996 ) 2rd column : λ=0.1 and L=3 ( computer packages ) 3th column : λ=0.4 and L=3 ( Hunter, 1989 )
國立雲林科技大學 工業工程與管理所 Comparing three EWMA control chart designs λ=0.1 λ=0.4 L=2.7L=3 Normal Gam(4,1) Gam(3,1) Gam(2,1) Gam(1,1) Gam(0.5,1) λ=0.1 λ=0.4 L=2.7L=3 Normal t t t t t t t t t For λ=0.1 and L=3, the ARL’s are too large. For λ=0.4 and L=3, the ARL’s are smaller than the normal- theory value. In-Control
國立雲林科技大學 工業工程與管理所 Conclusions 在 In control 的情況下, λ=0.05 and L=2.492 EWMA 管 制圖在非常態 ARL 值接近常態假設的 ARL 值。不會超出 8% 的差距(沒有低於 )。 在 In control 的情況下,除了極端非常態的分配參數值 ( t6 、 t4 、 Gam1,1 、 Gam0.5,1 ), λ=0.1 and L=2.703 EWMA 管制圖在非常態的 ARL 與常態的 ARL 不會超出 15% 的差距(很少低於 315 )。 在不同的分配參數的情況下, EWMA 偵測製程偏移的能 力並沒有太大的差別。