9 Cumulative Sum and

Exponentially Weighted
Moving Average Control
Charts

1
 The cumulative sum control chart
Sample, i xi Sample, i xi
Suppose σ = 1 and μ = 10
1 9.45 16 9.37
n=1
2 7.99 17 10.62
UCL = 13
3 9.29 18 10.31
CL = 10
4 11.66 19 8.52
LCL = 7
5 12.16 20 10.84
6 10.18 21 10.90
7 8.04 22 9.33
8 11.46 23 12.29
9 9.20 24 11.50
10 10.34 25 10.60
11 9.03 26 11.08
12 11.47 27 10.38
13 10.51 28 11.62
14 9.40 29 11.31
15 10.08 30 10.52

2
 The cumulative sum control chart
The mean shifted to 11, however, it was not detected.
The shift of 1σ is small and Shewhart chart can not
detect it easily.

P(7 x 13 | 11, 1)
P
P(7 11 z 13 11)
b
P( z 2) P( z 4) 0.97725

ARL
Ip
3
The cumulative sum control chart
Let 0 be a target value.
i
In cusum charts we plot Ci j 1
(xj 0 )
It can be used with individual readings.
It is useful for chemical and process industry where n = 1.

4
 The cumulative sum control chart
i i 1
Ci j 1
( x j 10) ( x j 10) j 1
( x j 10)
( x j 10) Ci 1

i xi xi –10 Ci i xi xi –10 Ci
1 9.45 –0.55 –0.55 16 9.37 –0.63 –0.37
2 7.99 –2.01 –2.56 17 10.62 0.62 0.25
3 9.29 –0.71 –3.27 18 10.31 0.31 0.56 6
4 11.66 1.66 –1.61 19 8.52 –1.48 –0.92
5
6
12.16
10.18
2.16
0.18
0.55
0.73
20
21
10.84
10.90
0.84
0.90
005
–0.08
0.82 I
5
7 8.04 –1.96 –1.23 22 9.33 –0.67 0.15 2 4
7
8 11.46 1.46 0.23 23 12.29 2.29 2.44 y
9 9.20 –0.80 –0.57 24 11.50 1.50 3.94 4 I
10 10.34 0.34 –0.23 25 10.60 0.60 4.54 5 V
11 9.03 –0.97 –1.20 26 11.08 1.08 5.62 E
12 11.47 1.47 0.27 27 10.38 0.38 6.00
13 10.51 0.51 0.78 28 11.62 1.62 7.62
14 9.40 –0.60 0.18 29 11.31 1.31 8.93
15 10.08 0.08 0.26 30 10.52 0.52 9.45
5
The Tabular Cusum for Monitoring the Mean for n = 1
2K the smallest mean shiftthat we would like to detect
Main Assumption: The process starts in statistical control,
hence σ can be estimated
Will monitor deviations from μ0 that are more than K.
Need two statistics to monitor shifts up and down.
Ci max[0, xi ( K ) Ci 1 ]
0
| 1 0 |
K
Ci max[0, ( 0 K ) xi Ci 1 ] 2
where the starting values are 𝐶0+ = 𝐶0− = 0

If the statistics exceed the decision interval H, check for

assignable causes. Recommended H = 5σ

6
 Ci max[0, xi ( K ) Ci 1 ]
l
t
0

Ci max[0, ( K ) xi Ci 1 ]
0
I
xi ( 0 K)

( 0 K ) xi
o
Yi

7
Cusum Status chart

8
 Use of the Cusum chart
• The user sets a value for H, decision interval.
• Usually H = 5σ
• If C+ or C– > H, the process is considered out of
control.
• The nearest period where N = 1, indicates when the
shift probably started.
• After the assignable cause is found reset C to zero
and repeat the calculations.
• Can not use sensitizing rules.

9
 Adjustment to bring the mean back to μ0
i xi xi – (μ0+k) C+ N+ +
𝐶111 +
= 𝑥111 − 𝜇0 + 𝐾 + 𝐶110
105 x105 x105– (μ0+k) 0 0 = 𝑥111 − 𝜇0 + 𝐾
+
+
+𝑥110 − 𝜇0 + 𝐾 +𝐶109
106 x106 x106– (μ0+k) 𝐶106 >0 1 .
107 x107 x107– (μ0+k) +
𝐶107 2 .
+
= 𝑥111 − 𝜇0 + 𝐾
108 x108 x108– (μ0+k) 𝐶108 3 + 𝑥110 − 𝜇0 + 𝐾
109 x109 x109– (μ0+k) +
𝐶109 4 + 𝑥109 − 𝜇0 + 𝐾
+
+ 𝑥108 − 𝜇0 + 𝐾
110 x110 x110– (μ0+k) 𝐶110 5 + 𝑥107 − 𝜇0 + 𝐾
111 x111 x111– (μ0+k) +
𝐶111 6 + 𝑥106 − 𝜇0 + 𝐾

+
𝐶111 /6 =(𝑥111 + 𝑥110 + 𝑥109 + 𝑥108 + 𝑥107 + 𝑥106 )/6 − 𝜇0 + 𝐾
𝜇Ƹ
ˆ 0 K Ci / N
10
 Adjustment to bring the mean back to μ0

If C+ > H If C– > H
• Assume the mean shifted • Assume the mean shifted
N+ time periods back. N– time periods back.
• The average of • The average of
xi – (μ0+k) = C+/N+ (μ0 – k) – xi = C–/N–
• An estimate of the mean • An estimate of the mean
now is now is
ˆ 0 K Ci / N ˆ 0 K Ci / N

11
 Example

ˆ 0 K Ci / N
+
• Suppose 𝐶30 = 5.4 > H = 5 O
• The corresponding N+ = 12
• Average difference = 5.4/12 = 0.45
• An estimate of the mean now is 10+0.5+0.45 =10.95
• Take action to reduce the current mean by 0.95

12
 Choice of K and H
• Choose K and H that result in small ARL.
• Define K = kσ and H = hσ.
• h = 4 or 5 and k = 0.5 give small ARL for a shift of 1σ
Shift in mean
(multiple of σ) h=4 h=5
0 168 465
0.25 74.2 139
0.50 26.6 38.0
0.75 13.3 17.0
1.00 8.38 10.4
1.50 4.75 5.75
2.00 3.34 4.01
2.50 2.62 3.11
3.00 2.19 2.57
4.00 1.71 2.01
13
 Choice of K and H

Values of k and the corresponding values of h that give

ARL0 = 370 for the two sided tabular Cusum
Similar to chasing 1 7 2 0 0027 ARLO 770

k 0.25 0.5 0.75 1 1.25 1.5

h 8.01 4.77 3.34 2.52 1.99 1.61

14
 The Standardized Cusum

Ci max[0, xi ( 0 K ) Ci 1 ]
xi
yi 0
Nco 1

xi ( 0 k ) ( yi k)

The Standardized Two–Sided Cusum

Ci max 0, yi k Ci 1

Ci max 0, k yi Ci 1

15
Improving cusum performance for large shifts
Combine with Shewhart control chart
Use L = 3.5
Shift in mean Basic Cusum-
(multiple of σ) Cusum Shewart
0 465 391 ARLO b
0.25
0.50
0.75
139
38.0
17.0
130.9
37.2
16.8
Ii
k = ½, h = 5 1.00 10.4 10.2
1.50
2.00
2.50
5.75
4.01
3.11
5.58
3.77
2.77
fan
3.00 2.57 2.1
4.00 2.01 1.34

16
 The Fast Initial Response (FIR) Cusum
µ+K
K=3, µ = 100, H =12
Ci max[0, x1 103 C0 ] 5
Set the starting value of C0 C0 H /2 6 Ci max[0,97 x1 C0 ] 1
µ-K
i xi xi–103 C+ N+ 97–xi C– N–
1 102 –1 5 1 –5 1 1
2 97 –6 0 0 0 1 2
3 104 1 1 1 –7 0 0
4 93 –6 0 0 4 4 1
5 100 –3 0 0 –3 1 2
6 105 2 1 1 –8 0 0
7 96 –7 0 0 1 1 1
8 98 –5 0 0 –1 0 0
9 105 2 1 1 –8 0 0
10 99 –4 0 0 –2 0 0 17
 Suppose the mean shifted to µ = 105
i xi xi–103 C+ N+ 97–xi C– N–
1 107 4 10 1 –10 0 0
2 102 –1 9 2 –5 0 0
3 109 6 15 3 –12 0 0
4 98 –5 10 4 –1 0 0
5 105 2 12 5 –8 0 0
6 110 7 19 6 –13 0 0
7 101 –2 17 7 –4 0 0
8 103 0 17 8 –6 0 0
9 110 7 24 9 –13 0 0
10 104 1 25 10 –7 0 0
• H = 12 implies that the cusum signals at sample 3
• Without the headstart, it would not signal until sample 6

18
The Fast Initial Response (FIR) Cusum

19
The Fast Initial Response (FIR) Cusum

k = ½, h = 5, L =3.5
Cusum-
Shift in mean Basic Cusum- Cusum
Shewart
(multiple of σ) Cusum Shewart with FIR
with FIR
0 465 391 430 360
0.25 139 130.9 122 113.9
0.50 38.0 37.2 28.7 28.1
0.75 17.0 16.8 11.2 11.2
1.00 10.4 10.2 6.35 6.32
5
1.50 5.75 5.58 3.37 3.37
2.00 4.01 3.77 2.36 2.36
2.50 3.11 2.77 1.86 1.86
3.00 2.57 2.1 1.54 1.54
4.00 2.01 1.34 1.16 1.16

20
Exponentially weighted moving average control chart
EW MA

• Used to detect small variations in the mean

• Approximately equivalent to the cusum chart.
• Easier than the cusum chart in operation.
• We consider the use with individual
observations.

21
Exponentially weighted moving average control chart

It l d I

I C oil

Chart the quantities:

𝑧𝑖 = 𝜆𝑥𝑖 + (1 − 𝜆)𝑧𝑖−1
where 𝑧0 = 𝜇0 and 0 < λ < 1

larger I mean less memory forget faster

22
 General expression of zi
𝑧𝑖 = 𝜆𝑥𝑖 + (1 − 𝜆)𝑧𝑖−1

• z0 = μ0
• z1 = λx1 + (1– λ)z0
• z2 = λx2 + (1– λ)z1
= λx2 + (1– λ)[λx1 + (1– λ)z0]
= λx2 + λ(1– λ)x1 + (1– λ)2z0
• z3 = λx3+ (1– λ)z2
= λx3 + λ(1– λ)x2 + λ(1– λ)2x1 + (1– λ)3z0
i 1 j i
zi j 0
1 xi j 1 z0

23
 Mean of zi

i 1 j i
zi j 0
1 xi j 1 z0 N M ri

i 1 j i
E ( zi ) j 0
1 E ( xi j ) 1 E ( z0 )
i 1 j i
0 j 0
1 1

24
 Variance of zi

i 1 j i
zi j 0
1 xi j 1 z0

i 1 j i
V ( zi ) V j 0
1 xi j 1 z0
2 i 1 2j
j 0
1 V ( xi j )
I C Coil
2 2 i 1 2j
j 0
1 Gl D CCon
2
1 1
2i
at O
2
2
If i → ∞ V ( zi )
2
25
 EWMA Control Chart

𝑧𝑖 = 𝜆𝑥𝑖 + (1 − 𝜆)𝑧𝑖−1
UCL 0 L [1 (1 ) 2i ]
2
Center line 0 I

UCL 0 L [1 (1 ) 2i ]
2
Steady–state control limits: if i → ∞

UCL 0 L
2

UCL 0 L
2

26
 Example
λ = 0.1, L = 2.7, µ0 = 10, σ = 1
x1 = 9.45, z1 = 0.1*9.45 + 0.9*10 = 9.945

Period 1 Control Limits: Steady–State Control Limits:

UCL 0 L 1 (1 )2i UCL 0 L

2 2
0.1 0.1
10 2.7 *1 1 (1 0.1) 2(1) 10 2.7 *1
2 0.1 2 0.1
10.27 10.62

LCL L
UCL 0 L 1 (1 )2i 0
2
2
0.1
0.1 10 2.7 *1
10 2.7 *1 1 (1 0.1) 2(1) 2 0.1
2 0.1
9.38
9.73
27
28
 Design of the EWMA

If

If the mean increases by a value of σ/2, what are the

values of L and λ that result in ARL1 ≤ 35?
29
 Choice of λ and L

• 0.05 ≤ λ ≤ 0.25 works well in practice.

• Use small λ for small shift
• If λ ≤ 0.1, use L ε [2.6, 2.8]
• If λ is large use L = 3

30
Robustness of EWMA to Non–normal Data
Comparison with skewed distributions

ARL

Ri
Go
Lo

JAR

31
Comparison with symmetric distributions

In Shewhart charts, small

departure of normality affects
ARL dramatically

32
Extensions of the EWMA
• Combine with Shewhart charts
• Fast initial response feature
• Monitoring variability
• Monitoring count data

33
 Zi d ki t Il h ki i

Rst RstNa xi xi 1 xi w 1
Mi
W 3 Mf 3 w
E Mi 0 We 5
2 t Nitze
V Mi Mz
w 3

UCL 0 3
w
UCL 0 3
w

34
35