SPC ALGORITHM APPLICATION WITH FUZZY TO MONITORING PLANT BY ONLINE Rina Krisnayana, Katherin Indriawati, Totok Ruki Biyanto Department of Physical Engineering, Faculty of Industrial Technology Institute Technology of Sepuluh November Surabaya email :
[email protected]
ABSTRACT In this research will be made an algorithm that can be used as support tool to make a decision about status of a process that can be apply on monitoring phase (online). The methods that used to make that algorithm is using fuzzy as a decision maker, so the result chart is a process condition status (status chart), which is normal, warning and action. Input for fuzzy is individual data deviation and moving range. The probability of false alarm can be minimized by widen the control limit. Using the widen control limit must be followed by CUSUM value on each deviation (X or R) so the speed to detect the real error is not slow. In this research, the SPC algorithm with fuzzy is applied to monitor process on Final Tail Gas Scrubber by on-line. This plant is built by HYSIS software. From the result on this research, SPC Algorithm with Fuzzy can be used to monitoring plant Final Tail Gas Scrubber’s vessel. SPC algorithm with fuzzy have a better performance than individual-MR chart on detect the small mean change, false alarm probability and false detection probability. Keywords : control chart, fuzzy, status chart, online.
1
INTRODUCTION
Statistical Process Control (SPC) is a statistic method to monitor, analyze, control, and improve performance of the process. The most SPC that common to used is control chart. For industry with continuous process, like chemical industry, control charts used commonly are individual – moving range (MR) [1], that is one of kind of Shewhart control chart. The use of Shewhart control chart has been investigated by SPC researchers and shown that by using statistical performance, for example, that cumulative sum (CUSUM) and exponentially weighted moving average (EWMA) charts are much more effective than Shewhart charts in detecting small and moderate-sized sustained shifts in the parameters of the probability distribution of a quality [2]. Nevertheless, this is not meant Shewhart control chat can be completely replaced by those two methods, especially in detecting a wider assortment of effects due to assignable causes. It is frequently recommended that Shewhart limits be used in conjunction
with a CUSUM or EWMA chart [3]. Unfortunately, in practice, this combination make chart more complex in view and need interpretation more carefully to decide about the process status. El-Shal and Morris in 1999 had used a fuzzy rulebased algorithm to improve the performance of SPC in quality systems for grouped data [4]. Fuzzy inference system (FIS) was used to make decision about the process status. According to Indriawati (2005), false alarm, false detection and time to first signal are the performances that have to be observed on SPC algorithm with fuzzy in offline phase [5]. In this paper, we propose to use an algorithm for individual data that can be used as support tool to make a decision about status of a process that can be apply on monitoring phase (online). As a case study, we choose the process on plant final tail gas scrubber and measure the performance of the algorithm through false alarm, false detection, and time to first signal.
2 2.1
DESIGN AND IMPLEMENTATION The Status Chart Design
In this research, the SPC algorithm with fuzzy is used to phase II, that is phase monitoring. The criteria function that muse be minimized for SPC performance is Criteria = sum of false detection that can’t detect + sum of false alarm that rises + time to detect error (1) The number of false detection can be maximized with fused two control chart that is shewhart and CUSUM. The using of limit control from control chart is widened to minimized the false alarm. Limit control for X-deviation is 4σx and limit control for R deviation is 4,2 MR . Widening limit control can detect false detection longer than usual. To make the speed to detect false detection not slow, must be followed by cusum value on each deviation. The cusum value for X deviation is kx = x and hx = 4x, then cusum value for R deviation is kR = R dan hR = 4,2 MR . The value of limit control, k and h are given for fuzzy to make the status. In fuzzy there is 2 configurations that is used: DCSC and RCSC. DCSC and RCSC
configurations are adopt from the Indriawati’s research [5]. To determine limit control from real plant data in controlled statistically condition is by using software SPCPI. Figure 1 is individual chart and moving range chart for real plant data in controlled condition statistically. Table number 1 is limit control for cusum from real plant data in controlled statistically condition. Limit control for phase II (monitoring phase) is constant. The step of decision making for process status : First step is detection if there is moving on mean value. DCSC is used for this situation Second step is detection if there is a change on variability process which represent by range changes. RCSC is used for this situation.
2.2
Simulating Algorithm by On-line
In this research, the purposed algorithm is applied to monitor process on Final Tail Gas Scrubber by on-line. Final tail gas scrubber is a complex system that have several part, i.e. input of sulphate acid, water process, gas of ammonia and flour that will be washed, liquid acid or base, output gas ammonia with flour and output of catharsis result. To simulate monitoring system by on-line, the plant is built by HYSIS software and also is simulated on HYSIS. Then it communicated with the DDE protocol with Matlab as DDE server that design the algorithm and HYSIS as DDE client that design the plant.. Several run have been made to estimate the probability of false alarm and false detection, and to compute time to first signal. Run data is come from the final tail gas scrubber plant simulation that is sent from HYSIS and received by MATLAB.
Individuals
Individuals (X) and Moving-Range Charts Data1_A
Start UCL=5.303
5.30 5.25 5.20 5.15 5.10 5.05 5.00 4.95 4.90 4.85 4.80 4.75 4.70
Is there Data?
y
CL=5
Make FIS woth DCSC and RCSC configuration
LCL=4.697 0
20
40
60
80
Read data from observed parameter
100 UCL=0.3717
0.35
Deviation normalization from CUSUM
0.30
Moving Range
N
0.25
Input normalization value to FIS
0.20 0.15 CL=0.1138
0.10
Y
Or the DCSC output>=limit action?
0.05
N
0.00 0
20
40
60
80
Y
100
Subgroup No.
Or the RCSC output>=limit action?
. N
Y
Figure 1. Individual chart and moving range chart real plant data
Or the DCSC output>=limitwarning? N
Table 1. Value of CUSUM parameter
Value of CUSUM parameter
Y
Variable Individual
Warning
Moving Range
Or the RCSC output>=limit warning? N Normal
Action
Is there another data
Y
k h
0,10074 0,40296
N
0,07414 0,4778424
If the output from DCSC configuration is bigger that limit action (0.85 is chosen for this case), so the process status is action. If output from RCSC configuration is bigger than 0.85, the process status is action. If the output from DCSC and RCSC configuration is bigger than limit warning (0.5 is chosen for this case), so the process status is warning. If output from DCSC and RCSC less than 0.5, the process status is normal. The flowchart for SPC algorithm with fuzzy by on-line is showed in figure 2.
STOP
Figure 2. Flowchart of the status chart
3 3.1
RESULT The False Alarm Probability Estimation
From the simulation by using 60 run data, it is shown that there is 2 run that give false alarm decision on status chart. Individual-MR control chart and status chart from two run show on figure 5 and 6. The first error model is happen on 10th run. On figure 3, shows the cause of false alarm is 2 points from 3 sample nearby points, from sample number 75 until 77, located outside 2 limit on X chart. It is shown that the
purposed algorithm is sensitive on one of the nonrandom pattern rules, i.e. 2 from 3 points in series located outside 2 limit. So that, false alarm that happen is caused by nonrandom pattern on data run which is shown the moving from aim but actually it’s not happen in reality.
between 1 and 3, from sample number 65 until 69. It’s shown that the purposed algorithm is sensitive on one of the nonrandom pattern rule, i.e. 4 from 5 points that nearby is located on area between 1 and 3. So that, false alarm that happen is caused by nonrandom pattern on run data that show the moving from aim but actually it’s not happen in reality. According to simulation result on this stage, so the probability of false alarm by using SPC algorithm with fuzzy is 3.3%. By using same run data, there is 30 false alarms that produce by individual-MR control chart. So that, false alarm from individual-MR control chart for simulation is 50%. So that false alarm probability from status chart is smaller than false alarm probability for individual-MR control chart. According to criteria function that written on eq.1 and based on simulation result, show the number of false alarm that have been rise by SPC algorithm with fuzzy is minimized then the number of false alarm that produce from individual –MR control chart. Therefore, the effort to minimized criteria function on rise false alarm section is done successfully.
3.2
Figure 3. The first false alarm error model that is produced by status chart
Figure 4. The second false alarm error model that is produced by status chart
While the 2nd error model is happen on run number 42. On figure 4, show that the cause of false alarm is come from 4 from 5 points that nearby which located on area
The Probability Estimation
False
Detection
From the simulation result by using 72 run data, false detection of the status chart algorithm can receive 100%. It means all runs error can be detected by the algorithm. Meanwhile, with the same data run, false detection probability on individual-MR control chart for this simulation is 84.7%. It because individual-MR control chart is unable to detect the change of process condition on 11 run: run 1 on pH sensor error 0,9 σ run 2 on pH sensor error 0,9 σ run 4 on pH sensor error 0,9 σ run 2 on pH sensor error 1 σ run 5 on pH sensor 1 σ run 1 on pH sensor 1,1 σ run 6 on pH sensor error 1,1 σ run 1 on pH sensor 1,2 σ run 4 on pH sensor 1,2 σ run 6 on pH sensor error 1,2 σ run 3 on pH sensor error 1,3 σ According to criteria function on eq. 1 and based on simulation result on this section, show the number of real error that is undetected by status chart has been minimized to 0% than the number or real error that is undetected by individual-MR control chart (15.3%). So, the effort to maximize the criteria function on real error that can not detected is done successfully.
3.3
Time To First Signal
There is the comparison between time to first signal for individual-MR control chart and status fuzzy chart, from false detection simulation (table 2). Based on this result, show that status chart also detect the mean change faster than individual-MR control chart do. Therefore, the effort to increase criteria function on time that is needed to detect the error according to eq.1 is done successfully.
Sensor Error
Run Number
IndividualMR Chart
Status Chart
1,8σ
1 2 3 4 5 6 1 2 3 4 5 6 1 2 3 4 5 6
108 106 113 117 103 104 104 102 103 104 128 143 106 117 111 104 105 101
107 104 106 104 102 103 103 102 102 103 115 104 105 105 107 103 105 101
1,9σ
Table 2. Time to first signal value Sensor Error
Run Number
IndividualMR Chart
Status Chart
0,9σ
1 2 3 4 5 6 1 2 3 4 5 6 1 2 3 4 5 6 1 2 3 4 5 6 1 2 3 4 5 6 1 2 3 4 5 6 1 2 3 4 5 6 1 2 3 4 5 6 1 2 3 4 5 6
109 111 139 108
105 134 107 117 111 126 105 105 120 111 140 102 144 125 110 113 116 137 117 130 116 105 102 122 105 107 110 102 105 108 110 105 106 119 122 112 104 107 105 106 116 109 104 106 104 107 104 103 104 104 106 106 103 107
1σ
1,1σ
1,2σ
1,3σ
1,4σ
1,5σ
1,6σ
1,7σ
146 117 107 125 115 131 116 130 116 102 105 129 123 118 112 142 146 117 113 139 134 113 149 122 109 129 118 109 126 110 118 106 132 114 105 107 108 110 120
2σ
4
CONCLUSION
False alarm probability from the status chart algorithm with is smaller (3.33%) than individual-MR control chart (50%) False detection probability from the status chart algorithm is bigger (100%) than individual-MR control chart (84.7%) SPC algorithm with fuzzy have a better performance than individual-MR control chart to detect the small and medium mean change (0,9 until 2)
REFERENCE - C.L., Introduction to statistical process [1] Mamzic, control, in Statistical Process Control, Bab 1, Mamzic, C.L, Editor, Instrument Society of America, 1995, 1 – 58.
[2] Montgomery, D.C., Introduction to Statistical Quality Control, 3rd ed., John Wiley & Sons, New York, 1996. [3] Woodall, W. H., Controversies and contradictions in statistical process control, Journal of Quality Technology, 2000, 32, no.4, 341 – 350. [4] El-Shal, S.M., Morris, A.S., A fuzzy rule-based algorithm to improve the performance of SPC in quality systems, Proceedings of IEEE International Conference on Systems, Man., and Cybernetics, Tokyo, Japan, 1999: 284 – 289. [5] Indriawati, K, Studi Peningkatan Kemampuan Interpretasi Data Pada SPC Berbantukan Sistem Inferensi Fuzzy: Kasus Evaluasi Kinerja Electrostatic Precipitator, Tesis S2, Teknik Fisika, ITB, 2005
