Sunteți pe pagina 1din 17

Chapter 4

Sim ulated Design of


X Charts
4.1 Introduction
In the third chapter, theoretical design of the new X chart is discussed and compared
with the theoretical design of Shewhart X chart for various shifts in the sample mean. It
is also important to study the simulated performance of new X chart to verify the design
of new X chart. Comparisons of simulated ARLs with theoretical ARLs, using strategy
CSQ are presented in section 4.3. Performance of new X chart is also compared with
CUSUM and EWMA charts when process parameters vary from sample to sample in
subsequent section. Procedure for generation of random numbers is discussed in the
following section.

4.2 Procedure for generation of random numbers with normal


distribution
As suggested by Box and Muller (1958), a ‘X’ series of numbers generated by the
following relation will have a normal distribution with mean, zero and standard deviation,

one.
X= 2 log ri xCos (2 TC r2) ------------- (4.1)

Where,
ri and r2 are random numbers generated between 0 and 1,
The random number generated from Equation 4.1 can be substituted in Equation
4.2 to give the ‘ Y’ series of random numbers.
Y = am + asd*X ------------ (4.2)
Where,
am = actual mean of the process
asd = actual standard deviation of the process
The variate ‘Y’ will have a normal distribution with mean of ‘am’ and standard deviation
of‘asd’.
Frequency distribution and relevant calculations of random numbers falling in
various class intervals, with mean of 110 and standard deviation of 3 is given in Table IB
(appendix B). % test shows that, with a confidence level of 99%, the numbers generated

109
conform to a normal distribution with the specified mean and standard deviation. Figure
IB (appendix B) shows the histogram of frequency distribution of random numbers
generated from Equations 4.2.
4.3 Comparison of simulated results with theoretical results of new
X chart for Strategy CSQ

A computer program is prepared to simulate the ARLs of new X chart. Simulation


procedure to calculate the ARLs is presented with the help of three flow diagrams shown
in appendix C. Same constraint of false alarm rate (ARLo) of 371 is maintained to

compute to ARLs of new X chart. The simulation run of 10000 is considered and ARL
values (shown in Tables) are the average of five ARL values for each shift in the sample
mean. The ARLo of 371 can be obtained with parameters: L = 3.2, K = 2.2, H = 4, n = 4,
v =16 and U* = 26.3.
Considering these chart parameters, the ARLs have been calculated by
simulation and compared with theoretical ARLs, as presented in Table 4.1.
Table 4.1 Comparison of simulated ARLs with theoretical ARLs for strategy CSQ

Shift (5) ARLs ARLs


(Simulated) (Theoretical)
0.0 373.00 370.32
0.2 308.21 311.56
0.4 201.22 202.25
0.5 152.35 155.52
0.6 115.20 117.32
0.8 64.33 65.06
1.0 35.41 36.32
1.2 20.34 20.22
1.4 10.92 11.30
1.5 8.72 8.56
1.6 6.31 6.52
1.8 3.92 3.97
2.0 2.53 2.66

110
Table 4.1 shows that the theoretical average run lengths (ARLs) match with
simulated ARLs for strategy CSQ. It is found that there is a little difference between
simulated ARLs and theoretical ARLs for all the shifts in the process average.

In this chapter, the simulated results of new X chart with simulated results of
CUSUM and EWMA charts will be compared when process parameters vary from
sample to sample, as discussed in the following section.

4.4 Varying Process parameters from sample to sample


Most popular and efficient charts to catch small shifts in process mean are CUSUM and
EWMA. However they can catch the shift in the process average when there is a single
and sustained shift. In practice, samples to a control chart are drawn in a random manner.
All samples may not necessarily be drawn under identical conditions. For example, in a
machine shop, there may be a number of identical machines and one or more of these
machines might be out of control. Because control charts are drawn by taking random
samples, all the samples may not contain products of these machines which are out of
control. Ideally the control chart should be such that it should generate the signal when
the sample is draw from these machines. Such a process has been simulated and

compared the performance of new X with CUSUM and EWMA charts. The signals of
indication of ‘out of control’ and average run length (ARLs) have been found out. For all
the three schemes, the simulation run length of 5000 is maintained and the process has
been simulated in such a way that shift in process average may occur after any number of
samples. The ARLs and signals of ‘out of control’ have been found out for shift in mean
from 0.5a to 4a. The ARLs have been computed, when there is change in the process
mean only. All the charts have been designed to have in-control ARL (false alarm rate) of
371. All other assumptions are same as stated earlier. Ideally all the schemes should
indicate the out of control signals, at the point where process mean shifts.
For the above situation, 5000 random samples of 4 each have been generated from
simulation by using Equation 4.2, but it is not possible to show such large number of
observations and calculations in the thesis. That’s why a sample study of 100 samples is
presented in the following section.

Ill
4.4.1 Sample study

The total run length is considered as 100 and assumed that assignable causes may occur
after any number of samples in all the three schemes. Same values of XI, X2, X3, X4, and

X have been considered for three schemes. Following data are considered for analyzing
the results.
Sample size of each sample (n) = 4, History (H) = 4, K= 2.2 and L = 3.2,

(Jo = 110 and Go= 3.0


For sample size of 4, the combination of K = 2.2 and L =3.2 is taken because this

combination gives in-control ARL (false alarm rate) of 371, which is equal to in-control

ARL of Shewhart chart when there is no shift in process mean or standard deviation.

Other combinations have also been tried for same in-control ARL but those combinations

yield higher ARLs when sample mean shifts from its target value. Other parameters and

calculations are given as under.

Upper warning limit (UWL) = po + Ko or po + K oo/Vn = 113.3

Lower warning limit (LWL) = po - Ka or po - K ao/Vn = 106.7

Upper control limit (UCL) = po + Lo or po + L ao/Vn = 114.8

Lower control limit (LCL) = po - La or po - L Go/Vn = 105.2

Chart parameters for CUSUM chart are: k = 0.5 and h = 4.8

Chart parameters for EWMA chart are: X, = 0.28

Table 4.2 gives 100 simulated samples of 4 each, out of which 90 have been

generated from a normal universe with a mean of 110 and standard deviation of 3. Ten

out of these, shown by bold faces are generated by a subroutine that gives random

numbers with normal distribution with mean of 113 and standard deviation of 3. Number

of correct and false signals by new X, CUSUM and EWMA charts are presented in

Table 4.3.

112
Table 4.2 Comparisons of signals of proposed scheme with CUSUM and EWMA
schemes
Mean
S. no XI X2 X3 X4 Remarks
(X)

1 109.557 111.169 114.76 107.926 110.853


2 108.472 110.367 109.747 110.477 109.766
3 110.722 108.005 113.875 108.859 110.365
4 110.164 112.433 110.206 111.042 110.961
5 111.286 115.22 110.544 104.786 110.459
6 110.479 107.094 111.125 111.885 110.146
7 112.349 109.428 108.934 111.564 110.569
8 114.332 111.556 111.918 102.531 110.084
9 115.874 108.286 109.877 108.225 110.566
10 112.846 119.489 115.134 113.811 115.32 SIGNAL BY NEW CHART ONLY
11 109.044 111.071 105.251 105.372 107.685
12 104.529 110.265 112.864 108.418 109.019
13 110.04 105.712 109.113 112.969 109.459
14 104.471 110.972 107.424 108.496 107.841
15 106.641 113.692 104.54 103.342 107.054
16 109.337 114.581 106.415 113.679 111.003
17, 109.792 110.802 107.883 115.683 111.04
18 108.281 115.7 110.618 110.393 111.248
19 114.469 108.524 105.9 111.117 110.003
20 118.042 116.033 109.251 118.821 115.537 SIGNAL BY NEW CHART ONLY
21 113.174 112.005 110.692 108.705 111.144
22 113.858 104.882 113.407 110.726 110.718
23 106.353 110.387 110.95 111.091 109.695
24 109.651 111.419 110.257 118.525 112.463
25 113.091 110.277 108.098 115.795 111.815
26 112.193 109.598 105.869 113.032 110.173
27 112.446 116.436 106.867 111.925 111.919 SIGNAL BY CUSUM CHART ONLY
28 106.101 105.432 110.875 110.264 108.168
29 107.207 109.989 112.109 108.089 109.349
SIGNAL BY BOTH NEW CHART
30 115.712 114.456 116.461 116.886 115.879 AND CUSUM CHART

113
Table 4.2 Comparisons of signals of proposed scheme with CUSUM and EWMA
schemes (Cont...)
31 109.623 109.138 109.999 110.82 109.895
32 109.739 109.316 109.148 111.412 109.904
33 108.506 104.423 105.814 109.69 107.108
34 115.474 111.148 109.702 110.067 111.598
35 107.22 110.678 107.794 108.012 108.426
36 109.54 115.686 110.392 115.111 112.682
37 113.624 110.191 113.014 114.258 112.772
38 107.594 109.308 106.063 116.026 109.748
39 107.083 110.929 108.269 112.024 109.576
40 117.826 109.515 113.869 112.129 113.335 NO SIGNAL BY ANY CHART
41 107.317 103.968 109.895 111.878 108.265
42 109.794 107.919 109.596 110.06 109.342
43 111.516 109.688 106.316 112.538 110.015
44 110.43 113.726 115.293 114.184 113.408
45 115.037 107.379 107.947 110.01 110.093
46 107.194 104.059 110.413 105.123 106.697
47 112.483 113.307 110.955 106.712 110.864
48 109.397 117.473 111.967 111.012 112.462
49 107.473 109.669 110.844 109.581 109.392
50 113.395 113.881 116.554 113.757 114397 NO SIGNAL BY ANY CHART
51 109.385 109.972 109.482 112.176 110.254
52 110.39 106.943 108.788 112.602 109.681
53 109.917 110.704 111.127 112.969 111.179
54 111.516 110.803 108.839 109.189 110.087
55 110.92 109.402 105.757 111.262 109.335
56 106.183 104.3 110.107 115.113 108.926
57 110.94 110.591 106.003 111.162 109.674
58 115.13 108.015 108.709 112.593 111.112
59 112.823 110.68 105.804 108.085 109.348
60 114.768 114.988 115.721 114.914 115.098 SIGNAL BY NEW CHART ONLY
61 110.87 111.37 110.453 107.489 110.046
62 110.549 109.918 108.629 107.157 109.063
63 104.724 104.88 105.72 107.81 105.784
64 111.996 105.517 107.324 109.183 108.505

114
Table 4.2 Comparisons of signals of proposed scheme with CUSUM and EWMA
schemes (Cont...)
65 106.456 112.464 105.747 106.504 107.793
66 107.119 111.686 109.848 111.933 110.147
67 111.533 105.908 108.327 115.5 110.317
68 111.949 109.795 107.458 107.454 109.164
69 116.421 108.01 114.118 114.385 113.234
70 114.228 114.731 115.275 117.781 115.504 SIGNAL BY NEW CHART ONLY
71 104.796 116.968 105.496 112.563 109.956 SIGNAL BY EWMA CHART ONLY
72 103.041 111.318 107.255 108.228 107.461
73 110.929 111.258 111.645 108.366 110.55
74 108.011 108.929 113.884 110.914 110.435
75 113.792 112.093 103.849 103.131 108.216
76 107.347 111.484 112.718 114.518 111.517
77 107.181 106.572 106.048 109.911 107.428
78 115.06 102.839 109.477 103.966 107.836
79 105.748 104.795 111.227 105.684 106.864
80 114.796 106.886 114.422 119.158 113.816 SIGNAL BY NEW CHART ONLY
81 109.645 109.428 104.277 114.907 109.564
82 107.834 108.753 107.653 111.372 108.903
83 110.106 109.018 113.106 107.105 109.834
84 113.554 108.386 107.741 108.874 109.639
85 109.124 110.751 109.408 113.333 110.654
86 114.433 106.987 111.437 115.347 112.051
87 104.215 112.86 111.047 105.542 108.416
88 108.349 112.732 103.069 106.348 107.625
89 112.76 109.973 112.406 109.89 111.257
90 115.193 111.306 116.169 115.137 114.451 NO SIGNAL BY ANY CHART
91 119.104 112.333 107.27 112.042 112,687
92 113.753 106.991 104.116 102.463 106.831
93 112.44 115.662 103.645 110.397 110.536
94 110.537 110.877 106.988 115.224 110.907
95 109.008 109.463 113.435 113.961 111.467
96 105.814 107.578 112.008 111.486 109.222
97 108.177 113.65 112.076 109.835 110.935
98 112.184 111.369 109.175 110.566 110.824

115
Table 4.2 Comparisons of signals of proposed scheme with CUSUM and EWMA
schemes (Cont...)
99 108.697 108.518 112.607 114.082 110.976
100 111.548 103.663 113.667 112.496 110344 NO SIGNAL BY ANY CHART

Table 4.3 Number of correct and false signals by new*, CUSUM and EWMA
charts

No. of points with shifted mean, for all the three charts = 10
No. of points with target mean, for all the three charts = 90

New X Number of correct signals 6


chart given by the chart
Number of false signals given 0
by the chart
Percentage of correct signals 60%
Percentage of false signals 0.0%
ARL 1.67
CUSUM Number of correct signals 1
chart given by the chart
Number of false signals given 1
by the chart
Percentage of correct signals 10%
Percentage of false signals 1.11%
ARL 10
EWMA Number of correct signals 0.0%
chart given by the chart
Number of wrong signals 1
given by the chart
Percentage of correct signals 0
Percentage of false signals 1.11% '
ARL -

116
It can be noted that new X chart indicates 6 correct signals out of total 10

signals, while CUSUM chart gives only one correct signal out of 10 signals. The EWMA

chart doesn’t show any correct signal out of 10 signals.

Considering all the parameters of sample study, experimental (simulated) results

for 5000 samples for new X, CUSUM and EWMA charts are presented in section 4.4.2.

4.4.2 Comparison of new X chart with CUSUM and EWMA charts for
5000 samples

In the above situation, the assignable cause/causes may occur after any number of

samples. The new X chart (strategy CSQ), CUSUM and EWMA charts have been
compared for 5000 observations. The ARLs have been computed, when the process mean
shifts only.
For the above situation, 5000 random samples of 4 each have been generated from
simulation by using Equation 4.2. Out of which, 4500 samples have been generated at
‘am’= 110 and ‘asd’= 3, while remaining 500 samples have been generated at
‘am’ = 110+8a, where 8 > 0. The Other details are described as under:

Total number of observations (N) = 5000


Number of points falling out of control after shifting the mean (Ni) = 500
Number of in-control points (N2) = 4500

Number of correct catches by new X chart out of 5000 samples = Np


Number of correct catches by CUSUM chart out of 5000 samples = Nc
Number of correct catches by EWMA chart out of 5000 samples = Ne

Number of wrong signals (false alarms) given by new X chart out of 5000 samples = Fp
Number of wrong signals (false alarms) given by CUSUM chart out of 5000 samples= Fc
Number of wrong signals (false alarms) given by EWMA chart out of 5000 samples = Fe

Probability of catching the shift by new X chart (Pp) = (Np/ Ni)

117
Probability of catching the shift by CUSUM chart (Pc) = (Nc/ Ni)
Probability of catching the shift by EWMA chart (Pe) = (Ne/ N i)

Probability of occurrence of false alarms (wrong signals) in new X chart (Pfp) = (Fp/ N2)
Probability of occurrence of false alarms (wrong signals) in CUSUM chart (Pfc)= (Fc/ N2)
Probability of occurrence of false alarms (wrong signals) in EWMA chart (Pfe) = (Fe/ N2)

Average run length (ARLp) of new X chart = (1/ Pp)


Average run length (ARLc) of CUSUM chart = (1/ Pc)
Average run length (ARLe) of EWMA chart = (1/ Pe)
The simulated results of new X, CUSUM, and EWMA charts are shown in Table
4.4.

Table 4.4 Experimental (simulated) results of new X, CUSUM and EWMA charts

Shift No. of Correct signals generated by False signals generated by


(in terms points New CUSUM EWMA New CUSUM EWMA
of a) with shifted X chart chart chart X chart chart
mean chart
0.5 500 4 3 1 9 13 10
1.0 500 11 6 2 11 26 21
1.5 500 26 17 1 15 40 35
2.0 500 58 28 1 16 49 44
2.5 500 127 71 1 13 58 74
3.0 500 218 88 2 13 63 146
3.5 500 299 143 1 14 65 175
4.0 500 380 197 1 17 68 241

The performance of individual charts is shown in Table 4.5, 4.6, and 4.7 and
graphical comparisons are shown in Figure 4.1, 4.2 and 4.3.

118
Table 4.5 Performance of new X chart

Shift Np Ni Pp ARLp Fp n2 Pfp

(in terms
of o)
0.5 4 500 0.008 125.00 9.0 4500 0.002
1.0 11 500 0.022 45.50 11.0 4500 0.002
1.5 26 500 0.052 19.20 15.0 4500 0.003
2.0 58 500 0.116 8.60 16.0 4500 0.003
2.5 127 500 0.254 3.90 13.0 4500 0.003
3 .0 218 500 0.436 2.30 13.0 4500 0.003
3.5 299 500 0.598 1.70 14.0 4500 0.003
4.0 380 500 0.760 1.30 17.0 4500 0.004

Table 4.6 Performance of CUSUM chart (k= 0. 5, h = 4.8)

Shift (in terms Nc N, Pc ARLc Fc n2 Pfc

of n)
0.5 3 500 0.006 166.70 13 4500 0.003
1.0 6 500 0.012 83.30 26 4500 0.006
1.5 17 500 0.034 29.40 40 4500 0.011
2.0 28 500 0.056 17.90 49 4500 0.011
2.5 71 500 0.142 7.00 58 4500 0.013
3 .0 88 500 0.176 5.70 63 4500 0.014
3.5 143 500 0.286 3.50 65 4500 0.014
4.0 197 500 0.394 2.50 68 4500 0.015

Where ‘k’ is reference value and ‘h’ is decision interval for CUSUM chart

119
Table 4.7 Performance of EWMA chart ( X = 0.28)

Shift (in Ne N, Pe ARLe Fe n2 Pfe

terms of a)

0.5 1 500 0.002 500 10 4500 0.002

1.0 2 500 0.004 250 21 4500 0.005

1.5 1 500 0.002 500 35 4500 0.008

2 .0 1 500 0.002 500 44 4500 0.010

2.5 1 500 0.002 500 74 4500 0.016

3 .0 2 500 0.004 250 146 4500 0.032

3.5 1 500 0.002 500 175 4500 0.039

4.0 1 500 0.002 250 241 4500 0.054

Where X = Weight factor for data of EWMA chart

The parameter X determines the rate at which older data enter into the
calculation of the EWMA statistic. X = 0.28 generates the in-control ARL (false alarm
rate) of 371 for EWMA chart. We have compared the signals amongst new X chart,
CUSUM chart and EWMA chart at the same false alarm rate.

120
Prob. of catching the shift
oo
by new chart

Prob. of catching the shift


by CUSUM chart
b>
Prob. of catching the shift

Prob. of catching the shift


> 0 0 0 0
by EWMA chart
cn
let
o>
to

1.5 2 2.5 3.5

Shift in mean

Figure 4.1 Comparison of probability of catching the process shifts

121
0.06
Prob. of false alarms
in new chart

Prob. of false alarms in


CUSUM chart

Prob. of false alarms in


EWMA chart j

Figure 4.2 Comparison of Probability of false alarms (wrong signals)

122
]

------ARL ( new chart)

ARL (CUSUM chart)

Figure 4.3 Comparison of average run length (ARLs)

4.4.2.1 Analysis of the situation

Tables 4.5, 4.6, and 4.7 show that the percentage of correct signals is very high for new

scheme compared to those provided by CUSUM and EWMA schemes. For 2o shift in the

sample mean, the new X chart indicates 11.6% correct signals, while CUSUM and

EWMA charts indicate only 5.6% and 0.2% respectively. Following conclusions can be

drawn from Table 4.5,4.6, and 4.7.

123
(i) New X chart is most efficient chart as it provides the highest percentage of

correct signals compared to CUSUM and EWMA schemes for all the shifts in

sample mean.

(ii) Maximum percentage of correct signals provided by new scheme is 76%, while

maximum percentage of correct signals provided by CUSUM and EWMA

schemes are 39.4% and 0.4% respectively.

(iii) For all the process shifts, the EWMA scheme is most ineffective out of the three

schemes as its maximum capability to catch the process shift is 0.4%.

(iv) The average run length of proposed scheme is much lower than both CUSUM

and EWMA schemes for any shift in the process average.

(v) The new scheme indicates the minimum percentage of false alarms compared to

CUSUM and EWMA schemes for all the shifts in sample mean.

(vi) The maximum percentage of false signals generated by new chart is 0.37% while

it is 1.51% for CUSUM chart and 5.36% for EWMA chart.

4.5 Discussion
It is clear from the comparisons of simulated and theoretical ARLs that the simulated
average run lengths (ARLs) match with theoretical ARLs for both all the shifts in the
sample mean. When process parameters vary from sample to sample, the performance of

new X chart has been compared with CUSUM and EWMA charts. It can be concluded

that new X chart is most efficient as it provides the highest percentage of correct signals
compared to CUSUM and EWMA schemes for all shifts in the process average

considered. The percentage of false signals generated by new X chart are also lowest
compared to CUSUM and EWMA charts.

124

S-ar putea să vă placă și