- Email: [email protected]
Fault diagnosis of an industrial plant using a Monte Carlo analysis coupled with systematic troubleshooting
Accepted Manuscript Fault diagnosis of an industrial plant using a Monte Carlo analysis coupled with systematic troubleshooting Irina Boiarkina, Nick Depree, Wei Yu, David I. Wilson, Brent R. Young PII:
S0956-7135(17)30113-5
DOI:
10.1016/j.foodcont.2017.02.061
Reference:
JFCO 5496
To appear in:
Food Control
Received Date: 1 August 2016 Revised Date:
2 February 2017
Accepted Date: 25 February 2017
Please cite this article as: Boiarkina I., Depree N., Yu W., Wilson D.I. & Young B.R., Fault diagnosis of an industrial plant using a Monte Carlo analysis coupled with systematic troubleshooting, Food Control (2017), doi: 10.1016/j.foodcont.2017.02.061. This is a PDF file of an unedited manuscript that has been accepted for publication. As a service to our customers we are providing this early version of the manuscript. The manuscript will undergo copyediting, typesetting, and review of the resulting proof before it is published in its final form. Please note that during the production process errors may be discovered which could affect the content, and all legal disclaimers that apply to the journal pertain.
ACCEPTED MANUSCRIPT
Fault Diagnosis of an Industrial Plant Using a Monte Carlo Analysis Coupled with Systematic Troubleshooting Irina Boiarkinaa , Nick Depreea , Wei Yua , David I. Wilsonb , Brent R. Younga,1 a Industrial
RI PT
Information and Control Centre, Chemical and Materials Engineering, University of Auckland, 2-6 Park Ave, Grafton, Auckland 1023, New Zealand b Auckland University of Technology, School of Engineering, Computing & Mathematical Sciences, Auckland 1142, Auckland
SC
Abstract
Efficiently troubleshooting a fortification issue at an industrial milk powder plant is a complex undertaking given the myriad of possible causes. Multiple causes, even when simple, are not
M AN U
easy to diagnose, however every single cause needs to be addressed in order to consistently meet product quality specifications. This paper uses statistical modelling in the form Monte Carlo simulations to investigate the probable causes for unexpected excessive product variation. This approach alone, refines but does not completely solve, the production issues, so a systematic approach was required to definitively solve other root causes. This two-step fault diagnosis approach ensured that all of the differing causes proposed by plant personnel could be addressed, and sound recommendations for good manufacturing operations could be made and adopted.
1
1. Introduction
TE D
Keywords: fault diagnosis, Monte Carlo, dairy, case study, quality, fortification
Many processing industries have recently seen a shift away from maximising production with
3
process control to a focus on quality, and process analytical technology (PAT) has come to stand
4
for the assessment and control of product quality. The popularity of PAT was partly due to
5
the strong FDA encouragement in the US pharmaceutical industry (FDA, 2004), but since has
6
spread to other manufacturing industries (Munir et al., 2015). The international dairy company
7
Fonterra Co-operative Group Ltd, the world’s largest fluid milk processor, has recently been
8
looking to accelerate the development and use of PAT tools to achieve ‘real time quality’ (RTQ),
AC C
EP
2
Email addresses: [email protected] (Irina Boiarkina), [email protected] (Nick Depree), [email protected] (Wei Yu), [email protected] (David I. Wilson), [email protected] (Brent R. Young) 1 Corresponding author
Preprint submitted to Elsevier
February 2, 2017
ACCEPTED MANUSCRIPT
9
10
combining the benefits of advanced process control (APC) with an explicit focus on quality (Hunter et al., 2012; Munir et al., 2015; Rimpil¨ainen et al., 2015). The attractions of real time quality are obvious. If one is confident that the product cur-
12
rently being manufactured is to specification, then savings can be made on off-line subsequent
13
testing, while simultaneously minimising the possibility of producing significant amounts of off-
14
spec product that must be recycled or rejected. However the development of appropriate tools to
15
achieve this requires that one understands the nature of the underlying quality issue in order to
16
carry out the appropriate corrective action. From an analysis of historical poor quality events,
17
it was decided that this work would concentrate on the timely identification and subsequent
18
correction of faults. Whilst seemingly simple, practical fault diagnosis on large interconnected
19
plants is complicated, and the natural human bias to search for a single phenomenological cause,
20
(as opposed to multiple, single failures) is often unwarranted.
SC
RI PT
11
Standard techniques for industrial fault diagnosis and monitoring can be found in Gertler
22
(1998) and Chiang et al. (2001), those employing simple rule-based methods such as expert
23
systems (Rich and Venkatasubramanian, 1987; Zahedi et al., 2011), or dynamic process modelling
24
(Bertanza et al., 2013), or even data driven multi-variate methods such as principal components
25
analysis (PCA) and multi-variate data analysis (Qin, 2012; Ralston et al., 2001; Singhal and
26
Seborg, 2006; Li et al., 2011; Eslamloueyan, 2011). It may be prudent to distinguish between
27
methodologies applied to a simulated process, such as the benchmark Tennessee Eastman plant
28
(Yin et al., 2012; Lee, 2004; Gertler, 1998; Chiang et al., 2001), those applied at a pilot plant
29
scale (Ruiz et al., 2001), and those applied on an actual industrial plant (Bertanza et al., 2013;
30
Ralston et al., 2001; Zahedi et al., 2011; Ge et al., 2011). The latter often contain subtleties that
31
are important to the overall success of the programme.
34
35
36
37
38
TE D
EP
33
When looking at it from the industrial point of view the method depends on its appropriateness to the end goal and the information available. Possible aims include: 1. Find general problems across the entire plant. This may include standard equipment fail-
AC C
32
M AN U
21
ures or abnormal operation. Data driven methods such as PCA and MVDA are appropriate here as they capture a large quantity of information simultaneously and are non-specific. However these techniques require knowing a ‘normal’ mode of operation, which is not always easily established.
39
2. Troubleshoot a specific problem. The causes may be singular or multiple and varied. Ex-
40
pert systems can be of some help here, as troubleshooting a specific issue may require
2
ACCEPTED MANUSCRIPT
knowledge specific to the process to explain it adequately. A dynamic model of the process
42
for elucidating the exact cause, such as that used by Bertanza et al. (2013) when trou-
43
bleshooting a wastewater treatment plant. General data processing methods can also be
44
applied although they will result in black box models, sucha as PCA, that may be difficult
45
to interpret.
46
1.1. An application of fault diagnosis in the dairy industry
RI PT
41
This work considers an industrial milk powder plant where the product is fortified with
48
specific ingredients in minute and carefully controlled quantities to the customer specification.
49
However one of the added ingredients showed larger than expected concentration variations in
50
the final product, with below-specification results.
SC
47
Initially it was unclear whether the variation was a fault, or a natural consequence from the
52
normal processing. This uncertainty precluded the use of some data driven processing methods
53
that require the identification of a ‘normal’ operating state, such as PCA. However a model of
54
the fortification process could be used to establish whether the variation and below-specification
55
results were probable or not, based on the variation of the inputs. This was combined with a
56
Monte Carlo (MC) strategy of running simulations of the process model repeatedly and compar-
57
ing with the available quality data, which was measured infrequently and difficult to trace back
58
to the process conditions.
M AN U
51
Goldfeld and Dubi (1987) reviewed the use of the Monte Carlo method for reliability engi-
60
neering within the manufacturing industry and an application of the Monte Carlo method for
61
analysing manufacturing failures in electronic components is reported by Accumolli (1996). In
62
this latter case it was used to estimate the percentage of the final product that could be expected
63
fail, as the final product could not be tested. In both cases the failure causes were already known,
64
and neither work looked at using the Monte Carlo method specifically for troubleshooting.
EP
TE D
59
Monte Carlo for uncertainty analysis has also been used for understanding penicillin V pro-
66
duction by Biwer et al. (2005) and for analysing the uncertainty around wastewater treatment
67
plant models by Sin et al. (2009). However, again both of these works look at the uncertainty
68
propagating through the process. For penicillin production it was used for assessing the result-
69
ing variability in the economics and environmental performance of the process, whilst for the
70
wastewater treatment it was used for design modelling. Thus, the aim in both cases was not
71
to troubleshoot and evaluate whether the variation was normal for the process or not. This
72
troubleshooting aim is a novel aspect of this work.
AC C
65
3
ACCEPTED MANUSCRIPT
Where the broad Monte Carlo strategy was found to be ineffective, a case-by-case approach
74
followed, made far more manageable by the reduction in the fault search-space by the broader
75
approach used initially. This ensured that all aspects of the problem could be covered.This
76
requires domain specific knowledge. For example, the troubleshooting of a distillation column re-
77
quires case-by-case evaluation of possible column malfunction causes (Kister et al., 2007; Kister,
78
2014). Similar case-by-case approaches can be taken for other operations such as pneumatic
79
conveying and filter operation (Mills, 2016; Sparks and Chase, 2016). However, analysing each
80
failure individually can be time consuming when each one can have multiple route causes, there-
81
fore having a broad technique to eliminate as many of the potential causes as possible before
82
carrying out a case-by-case analysis is very useful.
SC
RI PT
73
During this fault diagnosis, we noted themes that could be generalised to troubleshooting at
84
any industrial plant. We found that plant personnel had differing pre-conceived notions on the
85
root causes of the problem and this meant that during the fault diagnosis the approach taken
86
had to resolve the differences in opinion in order for any proposed solutions to be adopted at the
87
plant. Using the Monte Carlo uncertainty analysis followed by individual analyses was effective
88
at uncovering both single phenomenological and multiple single simple causes and capturing the
89
quality issue holistically.
M AN U
83
The outline of this work is as follows. Section 2.1 describes the specific industrial problem used
91
as a case study and lists the competing solution hypotheses. To resolve which potential problem
92
was indeed the root cause, a Monte Carlo analysis was undertaken in Section 2.2. However the
93
statistical analysis alone could not resolve all the potential causes, so a systematic cases-by case
94
analysis was used in Section 3.5 and both strategies resulted in concrete operational changes
95
outlined in Section 3.6. Finally a generalisation of this specific case study illustrating how it can
96
be applied in other instances is given in Section 4.
97
2. Theory
98
2.1. Fault Description
EP
AC C
99
TE D
90
Milk powder is typically fortified during processing by the addition of several ingredients.
100
These ingredients are normally added to the evaporator flow using a dedicated dosing system,
101
as shown in Figure 1.
102
The dosing system consists of two tanks, one for dosing and one for making up fresh solution
103
of the ingredient to be ready for switch-over as shown in Figure 1. The target (denoted by an
4
SC
RI PT
ACCEPTED MANUSCRIPT
Figure 1: General process flow diagram of a milk dryer showing the location of the ingredient addition point in
M AN U
the process and the dosing system.
104
asterisk superscript) solid mass of ingredient to be diluted for dosing can be calculated using a
105
mass balance based on the mass of solids going through the evaporators in the milk, and the
106
target concentration in the final milk powder (specified by the customer). The final milk powder
107
is composed of the ingredient, the milk solids and a small amount of left over water. Using a
108
mass balance approach, the target solid mass of ingredient to dissolve, m?I (kg), is, C ? V ? qE xs qD (1 − xm )(1 − L)
(1)
TE D
m?I =
where C ? is the target milk powder ingredient concentration (mg/g milk powder), V ? is the
110
target fill volume of the dosing tank used (m3 ), qD and qE are the expected dosing and evaporator
111
flowrates respectively (m3 /hr and t/hr), xs is total solids fraction in the evaporator flow (%), xm
112
is the final moisture content in the powder (%), and L is a factor used to account for ingredient
113
losses (dimensionless), such as by degradation during processing, and is assumed in the region
114
0.1 to 0.15. The estimates for the loss factor were obtained from in-situ historical studies carried
115
out at the plants.
117
118
AC C
116
EP
109
Once the ingredient is made up to the required concentration in the dosing tank, the flow
rate was assumed to be controlled in ratio to the evaporator flow. One particular milk powder plant identified that the concentration of the ingredient showed
119
unexpected, large concentration variations in the milk powder and unusually low concentration
120
results, well below the specification limit. The (competing) hypotheses posed by different plant
121
personnel for these two problems were:
5
ACCEPTED MANUSCRIPT
122
123
H1: The large variation in the final ingredient concentration was due to poor repeatability of the analytical technique used for measuring the ingredient concentration. H2: The large variation was due to inherent variations in the processing conditions (perhaps due
125
to environmental disturbances) that were not sufficiently addressed by the control system.
126
H3: The large variation and unexpected below specification results were due to unpredictable
127
RI PT
124
degradation of the ingredient during processing and subsequent storage.
Without further analysis, the available evidence was insufficient to support or reject these hy-
129
potheses, thus preventing controls from being implemented. In part this was due to the fact
130
that the concentration is measured infrequently, off-line, and with a variable multi-day delay,
131
thus making it difficult to trace, troubleshoot and correlate with processing conditions. Hence a
132
Monte Carlo uncertainty analysis was used to solve this dilemma.
133
2.2. Monte Carlo Analysis to Establish the Source of Variation
M AN U
SC
128
134
A Monte Carlo uncertainty analysis was used to address the three hypotheses stated in
135
Section 2.1, namely to test if the observed variations are due to poor control or excessive variations
136
in the incoming product streams.
The Monte Carlo strategy compares the actual probability density function (PDF) distribu-
138
tion of the output, with that computed by propagating the input distributions through a model
139
of the process (Saltelli et al., 2000) as shown in Figure 2. Such a simple, albeit computationally
140
expensive, strategy can be used to estimate the output distributions through complicated non-
141
linear systems without excessive approximation. The two requirements for implementation are a
142
model, in this case given by the mass balance, and knowledge of the input variations which can
143
be extracted from the plant historian.
EP
Rearranging Equation 1 to solve for the milk concentration in final product milk powder,
AC C
144
TE D
137
C=
m∗ qD (1 − xm )(1 − L) V qE xs
(2)
145
gives the predictive model used to propagate the input distributions through to the required
146
output distribution in Figure 2.
147
Each argument in the right hand side of Equation 2 (inputs in Figure 2) is quantified using
148
historical plant data without having to make any additional assumptions about the underlying
149
distributions. The following assumptions were used:
6
RI PT
ACCEPTED MANUSCRIPT
Figure 2: Diagram of the Monte Carlo uncertainty analysis strategy for quantifying the milk powder ingredient
151
1. For the variables qE , qD , xs and V , the plant historian was used to extract the data directly. Using sufficient data, the PDFs of each input can be regressed.
M AN U
150
SC
concentration variation.
152
2. The final moisture content, xm was assumed constant since it is both very tightly controlled,
153
and given the magnitude relative to the other variables, could not have contributed more
154
than 0.25 % variation to the final concentration.
3. The degradation factor, L, is poorly characterised but historically is assumed to be a 0.12
156
for the specific plant. If the variation is large without accounting for variation in the losses,
157
then unpredictable degradation is not required to explain the large output variation.
TE D
155
158
4. The added mass of the ingredient used for dosing tank make-up, m∗, was constant across a
159
specification with an expected normal three-sigma variation of 0.5% due to weighing error,
160
based on historical work carried out at the plant. If the output PDF as computed from the Monte Carlo simulation differed significantly from
162
the actual distribution, then either the postulated input distributions are a poor approximation
163
(unlikely in this instance), or the model is deficient, or there are disturbance variables unac-
164
counted for.
AC C
165
EP
161
After establishing input PDFs, several scenarios were examined to assess the uncertainty of
166
the final ingredient concentration. The first scenario evaluated the concentration variation due
167
to normal process variation, and this was used as the base case for comparison. Following that
168
‘what-if?’ scenarios were simulated to evaluate the effect of possible solutions.
7
ACCEPTED MANUSCRIPT
169
3. Results and Discussion
170
3.1. Input Probability Density Functions
171
172
Due to commercial sensitivity, all processing variable values and associated units have been deliberately removed and/or normalised as this uses industrially confidential data. The dosing tank fill volume PDF shown in Figure 3a was regressed using five months of
174
historical data. The bi-modal behaviour is due to the fact that actually two separate tanks were
175
used; one being re-filled whilst the other tank was dosing. The two modes are most likely due
176
to a slightly different fill level being used, as each tank is manually filled up to an inscribed
177
mark on the tank. The fill level varies by up to 14%, and is always higher than designed for,
178
assuming that the measured volume is calibrated correctly. The PDFs were converted to cumu-
179
lative distribution functions (CDFs) for analysis. From the CDF in Figure 3a it was found that
180
approximately 50% of the time, the tank is overfilled by 10% or more and that 95% of the time,
181
the tank is overfilled by at least 6%. This is shown by the line drawn on the CDF of the tank
182
volume in Figure 3a. Given that extra ingredient is added to compensate for processing losses,
183
the additional ingredient could be compensating for over-dilution.
M AN U
SC
RI PT
173
184
Figure 3b shows the PDF and CDF functions for the evaporator feed total solids. The
185
evaporator solids vary significantly and this variation was not accounted for during dosing control
186
historically.
The evaporator feed and dosing flow rate PDFs and CDFs are shown in Figures 3c and 3d.
188
Prior to this analysis it was assumed that the dosing flow rate was controlled in a specific ratio
189
to the evaporator flow rate. The dosing flow rate should be exactly proportional to the evap-
190
orator flow rate—when the evaporator flow rate increases by 10 % from the original value, the
191
dosing flow rate should increase by 10 % as well to provide enough ingredient for the extra solids
192
flow. However, the dosing flow rate only increased by 1 %, which would have been insufficient
193
to maintain the required concentration of ingredient in the final milk powder. As a consequence
194
of this investigation, the variation in the dosing flow rate was found to be far too small for it to
195
be in ratio to the evaporator flow. This is shown diagrammatically in Figure 4. This, seemingly
196
trivial oversight, was an unexpected result.
EP
AC C
197
TE D
187
However, the measured dosing flow rate did increase very slightly with the evaporator feed
198
flow, most likely due to a venturi effect at the dosing point. The evaporator and dosing flow
199
rates have a Spearman’s correlation coefficient of 0.84, and this was used in the Iman-Conover
200
method (Iman and Conover, 1982), to generate a correlated sample for the analysis.
8
RI PT
ACCEPTED MANUSCRIPT
(b)
(c)
M AN U
SC
(a)
(d)
Figure 3: Probability and cumulative density functions for a) dosing tank fill level b) total solids entering the
AC C
EP
TE D
evaporator c) flow rate to the evaporator d) dosing flow rate.
Figure 4: A comparison of the current normalised dosing flow rate to the evaporator flow rate overlayed with the design ratio (solid red line) that should be used for control. The correct ratio stipulates that a percentage increase in the evaporator flow should result in an equal percentage increase of the dosing flow rate.
9
ACCEPTED MANUSCRIPT
201
3.2. A Comparison of Actual Variation of Ingredient Concentration The first case investigated simulated the expected variation due to the normal processing
203
conditions encountered at the plant. Figure 5 shows a comparison of the expected variation due
204
to the natural input variations and subsequent processing, with and without analytical variation
205
(red and orange curves respectively), and the actual variation found in the milk powder ingredient
206
concentration (blue curve). For this simulation, a population of 50,000 data random points were
207
used for each input. The realisation of the random variates were generated by passing uniformly
208
distributed random numbers through the inverse cumulative distribution functions (iCDFs) of
209
each input.
RI PT
202
The Kolmogorov-Smirnov test was used to compare the similarity between the actual and
211
simulated distributions (Wilks, 1995). Even with the inclusion of the addition of the analytical
212
variation to the processing variation, the distributions are statistically different with a p-value
213
of 5.7×10−12 at the 95 % confidence level. The simulated distributions was found to be narrower
214
than actual, which shows that the Monte Carlo simulation did not account for all possible sources
215
of variation. However, both the simulated and actual distributions were right skewed with the
216
mean above the mid-specification value (normalised to zero in the figure). Both the actual and
217
simulated distributions also show that the concentration is more likely to be out of specification
218
on the upper limit, as opposed to the lower limit. However, something other than the normal
219
processing and analytical variation is contributing to the out-of-specification results, and thus a
220
case-by-case analysis was used.
AC C
EP
TE D
M AN U
SC
210
Figure 5: Comparison of the distribution of the predicted output concentration using an MC analysis with the actual measured distribution of the data.
221
The repeatability of the analytical technique had been quantified previously from historical 10
ACCEPTED MANUSCRIPT
work carried out at the plants, and the coefficient of variation was found to be 3.6 %. This is
223
of the same order of magnitude as the standard deviation of the simulation which accounts for
224
processing variation only, of 4.7 %. Figure 5 shows the expected output concentration if this
225
analytical variation is added on top of the processing variation, assuming that it comes from nor-
226
mally distributed data. As expected, it increases the total variation seen in the simulated results.
227
The difference between the distributions, with and without analytical variation, is statistically
228
significant at a 95 % confidence level with a p-value of 9.1×10−79 when using the two sample
229
Kolmogorov-Smirnov test. This means that a reduction in the variation due to either processing,
230
or the analytical technique, will decrease the final measured concentration variation in the milk
231
powder.
SC
RI PT
222
Consequently both hypotheses H1 and H2 by plant personnel contributed to the variation
233
found in the milk powder, and that scenarios could be tested to make recommendations for
234
improving the system operation. These are discussed in Sections 3.3 and 3.4. On the other
235
hand if the variation due to the analytical technique was significantly larger than that due to
236
processing, then any improvements in processing would end up being masked and therefore
237
fruitless.
M AN U
232
However, neither the normal variation, nor the variation in the analytical technique explained
239
the unexpected below-specification results that were defined as part of the problem. Furthermore,
240
normal processing variation was unlikely to lead to any below-specification results, since Figure
241
5 clearly indicates that off-specification results are far more probable on the high end given the
242
normal processing variation. Thus in order to adequately address the concerns of the plant
243
personnel, and find a clear explanation, the below-specification results were investigated case-
244
by-case. This is discussed in Section 3.5.
245
3.3. The Effect of Dosing Tank Fill Volume to the Concentration Variation
EP
TE D
238
Section 3.1 showed that because the dosing tanks are filled manually to a pre-specified mark,
247
there is actually significant variation in the dosing tank fill volume, with the volume usually
248
being larger than that designed. There was some concern that this manual operation was to
249
blame, and it would be desirable to quantify the economic implications in order to balance the
250
costs of implementing an automated filling procedure. To do this, a Monte Carlo simulation
251
was performed where the tank fill volume input variable is replaced with the constant setpoint
252
thereby removing the influence of the uncontrolled variations.
253
AC C
246
Figure 6 shows the effect of filling the dosing tanks to 100% of the intended design volume.
11
M AN U
SC
RI PT
ACCEPTED MANUSCRIPT
Figure 6: A comparison between the predicted distributions of the ingredient concentration in the milk powder with variable dosing tank fill volume (upper) and if the tank was filled exactly to the designed volume (lower).
254
Interestingly if the tank is filled to the design volume, then this means that a significant portion
255
of the results would be expected to violate the upper limit, approximately 16%. This implies that the plant processing losses (assumed to be 12%) may not be as high as
257
previously estimated by the plant personnel, and that the current over-dilution compensates
258
for this. Clearly some of the losses are not encountered during the processing so much as by
259
unaccounted for dilution. Therefore, either the tank volume fill procedure should be better
260
controlled, or the dosing flow rate should be adjusted based on the actual volume of water used
261
to dissolve the ingredient.
EP
TE D
256
Maintaining a constant tank volume would not significantly reduce the concentration variation
263
in the milk powder on its own. Therefore the reinstatement of ratio control was also investigated,
264
discussed in the following section.
265
3.4. The Effect of Ratio Control on Ingredient Concentration Variation
266
AC C
262
Figure 7 shows a comparison of different processing changes that could be made and what
267
impact they would have on the variation of the ingredient concentration. Proper ratio control
268
could be instated, with the dosing flow rate changing in proportion to the evaporator flow rate,
269
to account for the increase is solids being sent to the dryer. With the use of proper ratio control
12
ACCEPTED MANUSCRIPT
(Figure 7b), the distribution becomes significantly narrower (compared with Figure 7a), with the
271
standard deviation reducing from 4.7 to 3.7 %. Furthermore, if the tank is filled to a consistent
272
volume each time, then the variability decreases further, with the standard deviation reducing to
273
3.1 % (see Figure 7c). Having the dosing flow rate ratio controlled and a constant tank volume
274
would allow the plant to be operated closer to the upper limit.
RI PT
270
The distribution of the ingredient in the milk powder in Figure 7c now mirrors the total
276
solids distribution (Figure 3b), and therefore if ratio control is implemented for the evaporator
277
flow rate, it should be in ratio to mass solids flow, rather than the volumetric flowrate, as the
278
water that contributes to the volumetric flow rate is removed during the drying process, leaving
279
effectively only the solids behind.
280
3.5. Case-by-Case Analyses
SC
275
The work in part was initiated due to unexpected below specification results. The broad fault
282
diagnosis strategy found that off-specification results were more likely to be on the high limit
283
due to normal process variation. However, the Monte Carlo analysis did not reveal a possible
284
cause of the unexpectedly low below-specification results.
M AN U
281
Therefore, a case-by-case analysis was employed to investigate these intermittent occurrences,
286
using the the cause-and-effect diagram shown in Figure 8. Given that the Monte Carlo analysis
287
from Section 2.2 eliminated normal operation as being a possible cause, we looked for abnormal
288
operating conditions. Diagrams and data of the operating variables at the time of the event were
289
tracked and investigated, and the possible causes, shown in Figure 8, were eliminated one by one.
290
The degradation hypothesis assumed that all relevant plant systems are functioning correctly,
291
however, this was never verified for the low concentration results. Therefore, the dosing system,
292
and all relevant plant systems, were explored during these specific runs to see if any abnormal
293
behaviour was present at the time.
EP
TE D
285
During the case-by-case verification of normal system performance abnormal operations were
295
found for very short durations that coincided with the extremely low test results. Two repeating
296
themes were found for these results that were likely have caused them:
299
AC C
294
300
2. Unintended mixing of different specification powders. This was also not picked up on using
297
298
301
1. Temporary dosing suspension. This was not picked up on with the Monte Carlo analysis as the data range used did not cover the periods where this occurred, as this is extremely
rare and occurs only for very short durations.
the Monte Carlo analysis, as this data was not included in the model used. 13
SC
M AN U
(a)
RI PT
ACCEPTED MANUSCRIPT
(c)
AC C
EP
TE D
(b)
Figure 7: Comparison of effect of different processing scenarios on the final concentration in the powder a) current dosing practice b) effect of the implementation of ingredient dosing ratio control to the evaporator flow rate on the ingredient concentration c) effect of the implementation of ingredient dosing ratio control to the evaporator flow rate with a constant tank volume.
14
RI PT
ACCEPTED MANUSCRIPT
SC
Figure 8: Fish bone diagram showing the different aspects of the process that were assessed systematically for performance during off-specification events.
Due to the two-tank dosing system, a fresh tank has to be made up prior to the dosing tank
303
running dry, preferably with sufficient time for adequate mixing. However if the fresh solution is
304
not made up in time, then dosing is suspended until fresh solution becomes available. This was
305
found to have occurred on days when below-specification results were detected roughly during
306
the sampling time period.
M AN U
302
The other common cause found was due to the management of the milk powder storage
308
hoppers. Given that not all powder is fortified and large hoppers are used when packing powder
309
for transport, if a small amount of unfortified powder remains at the bottom of the storage
310
hoppers when a new, fortified specification powder is packed on top, then this will to be picked
311
up as a below-specification result during testing. During these occurrences the hopper level was
312
not quite zero, and although the quantity remaining was tiny in reference to the total amount of
313
powder produced, it is more than sufficient to be easily picked up during sampling.
316
317
318
319
EP
315
These causes were not picked up on for a number of reasons that need to be considered when carrying out any fault diagnosis:
AC C
314
TE D
307
1. Neither flow stoppages, nor the mixing of fortified and unfortified powder were hypothesised as possible causes for the rare below-specification results. It was assumed that the system operated always as intended, and thus these possibilities were not investigated. In retrospect all possibilities should be eliminated using supporting data.
320
2. The concentration is measured with a significant delay. This lack of immediate feedback
321
means that possible causes are easily overlooked or forgotten amongst the other information
15
ACCEPTED MANUSCRIPT
322
needed to run a complex plant. This means that the fault search space needs to look beyond
323
pre-conceived ideas.
325
326
3.6. Recommended Operational Changes Both the Monte Carlo approach and the case-by-case analysis resulted in distinct recommendations being suggested for addressing the system performance.
RI PT
324
The Monte Carlo analysis showed that correcting the ratio control gain of the dosing flow
328
rate, preferably to the mass of solids passing through the dryer, would reduced the variation
329
of the ingredient in the milk powder. Furthermore, the dilution of the ingredient in the dosing
330
tank should also be accounted for, whether through better control of the fill volume, or through
331
adjustment of the dosing rate based on the actual dilution. The consequences of these can be
332
easily economically quantified. The MC analysis could also quantify the relative importance of
333
decreasing processing variance as opposed to decreasing the variance of the incoming material
334
streams.
M AN U
SC
327
The unexpected below-specification results can be eliminated by giving operators adequate
336
warning when the dosing tank is going to run empty to prevent dosing stoppages, and by ensuring
337
that unfortified and fortified powders do not get mixed inadvertently during packing. This could
338
be programmed as a warning into the human machine interface (HMI), or another plant systems.
339
4. Conclusions
TE D
335
This work looked at the fault diagnosis of a specific issue at a large, complex industrial milk
341
powder plant. Two strategies were employed to investigate why a fortification ingredient showed
342
a large variation in the final milk powder and unexpected below-specification results: a broad
343
Monte Carlo analysis, supplemented by a case-by-case study. This resulted in a wider framing
344
of the problem, and such a dual-pronged approach is necessary when it is likely that more than
345
one issue is actually active at any one time.
AC C
346
EP
340
In this industrial case study, it was found that the large variation came from normal processing
347
which was quantified by the MC analysis, and the below-specification results came from simple
348
causes such as flow stoppages and inadvertent mixing of fortified and unfortified milk powder.
349
Both issues have been addressed with specific process changes to work towards real-time quality.
350
In general, the fault diagnosis followed a number of different hypothesised causes, and we
351
found that all had to be reconciled for any recommendations to be adopted at the plant. A
352
single tool is unlikely to be completely successful given that at any one time on a complex 16
ACCEPTED MANUSCRIPT
industrial plant, the high probability of multiple seemingly independent, possibly simple, issues.
354
However simple problems in complex environments are not so simple to diagnose due to the
355
interconnections, noise and amount of data that has to be processed in a large system. Hence
356
the suggestion to use a two-step approach consisting of using statistical models to reduce the
357
search space, after which one can then undertake a more detailed, directed efficient case-by-case
358
study.
359
5. Acknowledgements
RI PT
353
The authors would like to acknowledge the Primary Growth Partnership program from the
361
New Zealand Ministry of Primary Industries for funding the project and would also like to thank
362
Fonterra staff, specifically James Winchester, Richard Croy, Hong Chen, Brian Woods, Steve
363
Holroyd, Nigel Russell and Tristan Hunter for providing resources and support throughout the
364
project.
365
References
366
Accumolli, A. (1996). Manufacturing process problem analysis using Monte Carlo simulation.
M AN U
367
SC
360
Quality and Reliability Engineering International, 12:39–41.
Bertanza, G., Pedrazzani, R., Manili, L., and Menoni, L. (2013). Bio-p release in the final
369
clarifiers of a large WWTP with co-precipitation: Key factors and troubleshooting. Chemical
370
Engineering Journal, 230:195–201.
373
374
using Monte Carlo simulation. Biotechnology and Bioengineering, 90(2):167–179.
EP
372
Biwer, A., Griffith, S., and Cooney, C. (2005). Uncertainty analysis of penicillin V production
Chiang, L., Russell, E., and Braatz, R. (2001). Fault Detection and Diagnosis in Industrial Systems. Springer-Verlag London.
AC C
371
TE D
368
375
Eslamloueyan, R. (2011). Designing a hierarchical neural network based on fuzzy clustering for
376
fault diagnosis of the Tennessee Eastman process. Applied Soft Computing, 11(1):1407–1415.
377
FDA (2004). Guidance for industry PAT —- A framework for innovative pharmaceutical de-
378
velopment, manufacturing, and quality assurance. Technical report, U.S. Department of
379
Health and Human Services, Food and Drug Administration, 5600 Fishers Lane, Rockville,
380
MD 20857, USA. 17
ACCEPTED MANUSCRIPT
382
383
384
385
386
Ge, Z., Chen, T., and Song, Z. (2011). Quality prediction for polypropylene production process based on clgpr model. Control Engineering Practice, 19:423–432. Gertler, J. (1998). Fault detection and diagnosis in engineering systems. Marcel Dekker, Inc, New York. Goldfeld, A. and Dubi, A. (1987). Monte Carlo methods in reliability engineering. Quality and Reliability Engineering International, 3:83–91.
RI PT
381
Hunter, T. J., Russell, N., Wilson, D. I., Young, B. R., Munir, T., and Depree, N. (2012).
388
Developing a real-time quality programme for dairy. In Chemeca, 23-36 September 2012,
389
Wellington, New Zealand.
SC
387
Iman, R. L. and Conover, W. J. (1982). A distribution free approach to induce rank correlation
391
among input variables. Communication in Statistics-Simulation and Computation, 11:311–
392
334.
394
395
396
Kister, H., Stupin, W., Lenferink, J. O., and Stupin, S. (2007). Troubleshooting a packing maldistribution upset. Chemical Engineering Research and Design, 85(1):136 – 143. Kister, H. Z. (2014). Chapter 2 - common technques for distillation troubleshooting. In Garak, A. and Schoenmakers, H., editors, Distillation, pages 37 – 101. Academic Press, Boston.
TE D
393
M AN U
390
397
Lee, G. (2004). Multiple-fault diagnosis of the tennessee eastman process based on system decom-
398
position and dynamic PLS. Industrial and Engineering Chemistry Research, 43(25):8037–
399
8048.
Li, G., Alcala, C., Qin, S. J., and Zhou, D. (2011). Generalized reconstruction-based contribu-
401
tions for output-relevant fault diagnosis with application to the Tennessee Eastman process.
402
Control Systems Technology, IEEE Transactions, 19(5):1114–1127.
EP
400
Mills, D. (2016). Chapter 24 - troubleshooting and material flow problems. In Mills, D., ed-
404
itor, Pneumatic Conveying Design Guide (Third Edition), pages 555 – 576. Butterworth-
405
AC C
403
Heinemann, third edition edition.
406
Munir, M. T., Yu., W., Young, B. R., and Wilson, D. I. (2015). The current status of pro-
407
cess analytical technologies in the dairy industry. Trends in Food Science & Technology,
408
43(2):205–218.
18
ACCEPTED MANUSCRIPT
410
411
412
413
414
Qin, S. J. (2012). Survey on data-driven industrial process monitoring and diagnosis. Annual Reviews in Control, 36:220–234. Ralston, P., DePuy, G., and Graham, J. H. (2001). Complant-based monitoring and fault diagnosis: a chemical process case study. ISA Transactions, 40:85–98. Rich, S. H. and Venkatasubramanian, V. (1987). Model-based reasoning in diagnostic expert
RI PT
409
ssystem for chemical process plants. Computer Chemical Engineering, 11(2):111–122.
Rimpil¨ ainen, V., Kaipio, J. P., Depree, N., Young, B. R., and Wilson, D. I. (2015). Predicting
416
functional properties of milk powder based on manufacturing data in an industrial-scale
417
powder plant. Journal of Food Engineering, 153:12–19.
419
420
421
Ruiz, D., maria Nougues, J., and Puigjaner, L. (2001). Fault diagnosis support system for complex chemical plants. Computers and Chemical Engineering, 25:151–160.
M AN U
418
SC
415
Saltelli, A., Chan, K., and Scott, E. M. (2000). Sensitivity Analysis. John Wiley & Sons Ltd, West Sussex.
422
Sin, G., Gernaey, K. V., Neumann, M. B., van Loosdrecht, M. C., and Gujer, W. (2009). Un-
423
certainty analysis in {WWTP} model applications: A critical discussion using an example
424
from design. Water Research, 43(11):2894 – 2906.
426
Singhal, A. and Seborg, D. E. (2006). Evaluation of a pattern matching method for the Tennessee
TE D
425
Eastman challenge process. Journal of Process Control, 16(6):601–613. Sparks, T. and Chase, G. (2016). Section 7 - filter selection, process design, testing, optimization
428
and troubleshooting guidelines. In Sparks, T., , and Chase, G., editors, Filters and Filtration
429
Handbook (Sixth Edition), pages 383 – 413. Butterworth-Heinemann, Oxford, sixth edition
430
edition.
432
Wilks, D. S. (1995). Chapter 5 hypothesis testing. In Statistical Methods in the Atmospheric
AC C
431
EP
427
Sciences, volume 59 of International Geophysics, pages 114 – 158. Academic Press.
433
Yin, S., Ding, S. X., Haghani, A., Hao, H., and Zhang, P. (2012). A comparison study of basic
434
data-driven fault diagnosis and process monitoring methods on the benchmark Tennessee
435
436
437
Eastman process. Journal of Process Control, 22:1567–1581.
Zahedi, G., Saba, S., al Otaibi, M., and Mohd-Yusof, K. (2011). Troubleshooting of crude oil desalination plant using fuzzy expert system. Desalination, 266:162–170. 19
ACCEPTED MANUSCRIPT
•
AC C
EP
TE D
M AN U
SC
•
Significant variation in fortified milk powder quality Large, complex industrial powder plant made issues difficult to track Monte Carlo analysis used to understand the processing sources of the large quality variation Case-by-case analysis used for off-spec events guided by Monte Carlo results Fault diagnosis required to ensure all causes were found to produce consistent quality powder
RI PT
• • •
S0956-7135(17)30113-5
DOI:
10.1016/j.foodcont.2017.02.061
Reference:
JFCO 5496
To appear in:
Food Control
Received Date: 1 August 2016 Revised Date:
2 February 2017
Accepted Date: 25 February 2017
Please cite this article as: Boiarkina I., Depree N., Yu W., Wilson D.I. & Young B.R., Fault diagnosis of an industrial plant using a Monte Carlo analysis coupled with systematic troubleshooting, Food Control (2017), doi: 10.1016/j.foodcont.2017.02.061. This is a PDF file of an unedited manuscript that has been accepted for publication. As a service to our customers we are providing this early version of the manuscript. The manuscript will undergo copyediting, typesetting, and review of the resulting proof before it is published in its final form. Please note that during the production process errors may be discovered which could affect the content, and all legal disclaimers that apply to the journal pertain.
ACCEPTED MANUSCRIPT
Fault Diagnosis of an Industrial Plant Using a Monte Carlo Analysis Coupled with Systematic Troubleshooting Irina Boiarkinaa , Nick Depreea , Wei Yua , David I. Wilsonb , Brent R. Younga,1 a Industrial
RI PT
Information and Control Centre, Chemical and Materials Engineering, University of Auckland, 2-6 Park Ave, Grafton, Auckland 1023, New Zealand b Auckland University of Technology, School of Engineering, Computing & Mathematical Sciences, Auckland 1142, Auckland
SC
Abstract
Efficiently troubleshooting a fortification issue at an industrial milk powder plant is a complex undertaking given the myriad of possible causes. Multiple causes, even when simple, are not
M AN U
easy to diagnose, however every single cause needs to be addressed in order to consistently meet product quality specifications. This paper uses statistical modelling in the form Monte Carlo simulations to investigate the probable causes for unexpected excessive product variation. This approach alone, refines but does not completely solve, the production issues, so a systematic approach was required to definitively solve other root causes. This two-step fault diagnosis approach ensured that all of the differing causes proposed by plant personnel could be addressed, and sound recommendations for good manufacturing operations could be made and adopted.
1
1. Introduction
TE D
Keywords: fault diagnosis, Monte Carlo, dairy, case study, quality, fortification
Many processing industries have recently seen a shift away from maximising production with
3
process control to a focus on quality, and process analytical technology (PAT) has come to stand
4
for the assessment and control of product quality. The popularity of PAT was partly due to
5
the strong FDA encouragement in the US pharmaceutical industry (FDA, 2004), but since has
6
spread to other manufacturing industries (Munir et al., 2015). The international dairy company
7
Fonterra Co-operative Group Ltd, the world’s largest fluid milk processor, has recently been
8
looking to accelerate the development and use of PAT tools to achieve ‘real time quality’ (RTQ),
AC C
EP
2
Email addresses: [email protected] (Irina Boiarkina), [email protected] (Nick Depree), [email protected] (Wei Yu), [email protected] (David I. Wilson), [email protected] (Brent R. Young) 1 Corresponding author
Preprint submitted to Elsevier
February 2, 2017
ACCEPTED MANUSCRIPT
9
10
combining the benefits of advanced process control (APC) with an explicit focus on quality (Hunter et al., 2012; Munir et al., 2015; Rimpil¨ainen et al., 2015). The attractions of real time quality are obvious. If one is confident that the product cur-
12
rently being manufactured is to specification, then savings can be made on off-line subsequent
13
testing, while simultaneously minimising the possibility of producing significant amounts of off-
14
spec product that must be recycled or rejected. However the development of appropriate tools to
15
achieve this requires that one understands the nature of the underlying quality issue in order to
16
carry out the appropriate corrective action. From an analysis of historical poor quality events,
17
it was decided that this work would concentrate on the timely identification and subsequent
18
correction of faults. Whilst seemingly simple, practical fault diagnosis on large interconnected
19
plants is complicated, and the natural human bias to search for a single phenomenological cause,
20
(as opposed to multiple, single failures) is often unwarranted.
SC
RI PT
11
Standard techniques for industrial fault diagnosis and monitoring can be found in Gertler
22
(1998) and Chiang et al. (2001), those employing simple rule-based methods such as expert
23
systems (Rich and Venkatasubramanian, 1987; Zahedi et al., 2011), or dynamic process modelling
24
(Bertanza et al., 2013), or even data driven multi-variate methods such as principal components
25
analysis (PCA) and multi-variate data analysis (Qin, 2012; Ralston et al., 2001; Singhal and
26
Seborg, 2006; Li et al., 2011; Eslamloueyan, 2011). It may be prudent to distinguish between
27
methodologies applied to a simulated process, such as the benchmark Tennessee Eastman plant
28
(Yin et al., 2012; Lee, 2004; Gertler, 1998; Chiang et al., 2001), those applied at a pilot plant
29
scale (Ruiz et al., 2001), and those applied on an actual industrial plant (Bertanza et al., 2013;
30
Ralston et al., 2001; Zahedi et al., 2011; Ge et al., 2011). The latter often contain subtleties that
31
are important to the overall success of the programme.
34
35
36
37
38
TE D
EP
33
When looking at it from the industrial point of view the method depends on its appropriateness to the end goal and the information available. Possible aims include: 1. Find general problems across the entire plant. This may include standard equipment fail-
AC C
32
M AN U
21
ures or abnormal operation. Data driven methods such as PCA and MVDA are appropriate here as they capture a large quantity of information simultaneously and are non-specific. However these techniques require knowing a ‘normal’ mode of operation, which is not always easily established.
39
2. Troubleshoot a specific problem. The causes may be singular or multiple and varied. Ex-
40
pert systems can be of some help here, as troubleshooting a specific issue may require
2
ACCEPTED MANUSCRIPT
knowledge specific to the process to explain it adequately. A dynamic model of the process
42
for elucidating the exact cause, such as that used by Bertanza et al. (2013) when trou-
43
bleshooting a wastewater treatment plant. General data processing methods can also be
44
applied although they will result in black box models, sucha as PCA, that may be difficult
45
to interpret.
46
1.1. An application of fault diagnosis in the dairy industry
RI PT
41
This work considers an industrial milk powder plant where the product is fortified with
48
specific ingredients in minute and carefully controlled quantities to the customer specification.
49
However one of the added ingredients showed larger than expected concentration variations in
50
the final product, with below-specification results.
SC
47
Initially it was unclear whether the variation was a fault, or a natural consequence from the
52
normal processing. This uncertainty precluded the use of some data driven processing methods
53
that require the identification of a ‘normal’ operating state, such as PCA. However a model of
54
the fortification process could be used to establish whether the variation and below-specification
55
results were probable or not, based on the variation of the inputs. This was combined with a
56
Monte Carlo (MC) strategy of running simulations of the process model repeatedly and compar-
57
ing with the available quality data, which was measured infrequently and difficult to trace back
58
to the process conditions.
M AN U
51
Goldfeld and Dubi (1987) reviewed the use of the Monte Carlo method for reliability engi-
60
neering within the manufacturing industry and an application of the Monte Carlo method for
61
analysing manufacturing failures in electronic components is reported by Accumolli (1996). In
62
this latter case it was used to estimate the percentage of the final product that could be expected
63
fail, as the final product could not be tested. In both cases the failure causes were already known,
64
and neither work looked at using the Monte Carlo method specifically for troubleshooting.
EP
TE D
59
Monte Carlo for uncertainty analysis has also been used for understanding penicillin V pro-
66
duction by Biwer et al. (2005) and for analysing the uncertainty around wastewater treatment
67
plant models by Sin et al. (2009). However, again both of these works look at the uncertainty
68
propagating through the process. For penicillin production it was used for assessing the result-
69
ing variability in the economics and environmental performance of the process, whilst for the
70
wastewater treatment it was used for design modelling. Thus, the aim in both cases was not
71
to troubleshoot and evaluate whether the variation was normal for the process or not. This
72
troubleshooting aim is a novel aspect of this work.
AC C
65
3
ACCEPTED MANUSCRIPT
Where the broad Monte Carlo strategy was found to be ineffective, a case-by-case approach
74
followed, made far more manageable by the reduction in the fault search-space by the broader
75
approach used initially. This ensured that all aspects of the problem could be covered.This
76
requires domain specific knowledge. For example, the troubleshooting of a distillation column re-
77
quires case-by-case evaluation of possible column malfunction causes (Kister et al., 2007; Kister,
78
2014). Similar case-by-case approaches can be taken for other operations such as pneumatic
79
conveying and filter operation (Mills, 2016; Sparks and Chase, 2016). However, analysing each
80
failure individually can be time consuming when each one can have multiple route causes, there-
81
fore having a broad technique to eliminate as many of the potential causes as possible before
82
carrying out a case-by-case analysis is very useful.
SC
RI PT
73
During this fault diagnosis, we noted themes that could be generalised to troubleshooting at
84
any industrial plant. We found that plant personnel had differing pre-conceived notions on the
85
root causes of the problem and this meant that during the fault diagnosis the approach taken
86
had to resolve the differences in opinion in order for any proposed solutions to be adopted at the
87
plant. Using the Monte Carlo uncertainty analysis followed by individual analyses was effective
88
at uncovering both single phenomenological and multiple single simple causes and capturing the
89
quality issue holistically.
M AN U
83
The outline of this work is as follows. Section 2.1 describes the specific industrial problem used
91
as a case study and lists the competing solution hypotheses. To resolve which potential problem
92
was indeed the root cause, a Monte Carlo analysis was undertaken in Section 2.2. However the
93
statistical analysis alone could not resolve all the potential causes, so a systematic cases-by case
94
analysis was used in Section 3.5 and both strategies resulted in concrete operational changes
95
outlined in Section 3.6. Finally a generalisation of this specific case study illustrating how it can
96
be applied in other instances is given in Section 4.
97
2. Theory
98
2.1. Fault Description
EP
AC C
99
TE D
90
Milk powder is typically fortified during processing by the addition of several ingredients.
100
These ingredients are normally added to the evaporator flow using a dedicated dosing system,
101
as shown in Figure 1.
102
The dosing system consists of two tanks, one for dosing and one for making up fresh solution
103
of the ingredient to be ready for switch-over as shown in Figure 1. The target (denoted by an
4
SC
RI PT
ACCEPTED MANUSCRIPT
Figure 1: General process flow diagram of a milk dryer showing the location of the ingredient addition point in
M AN U
the process and the dosing system.
104
asterisk superscript) solid mass of ingredient to be diluted for dosing can be calculated using a
105
mass balance based on the mass of solids going through the evaporators in the milk, and the
106
target concentration in the final milk powder (specified by the customer). The final milk powder
107
is composed of the ingredient, the milk solids and a small amount of left over water. Using a
108
mass balance approach, the target solid mass of ingredient to dissolve, m?I (kg), is, C ? V ? qE xs qD (1 − xm )(1 − L)
(1)
TE D
m?I =
where C ? is the target milk powder ingredient concentration (mg/g milk powder), V ? is the
110
target fill volume of the dosing tank used (m3 ), qD and qE are the expected dosing and evaporator
111
flowrates respectively (m3 /hr and t/hr), xs is total solids fraction in the evaporator flow (%), xm
112
is the final moisture content in the powder (%), and L is a factor used to account for ingredient
113
losses (dimensionless), such as by degradation during processing, and is assumed in the region
114
0.1 to 0.15. The estimates for the loss factor were obtained from in-situ historical studies carried
115
out at the plants.
117
118
AC C
116
EP
109
Once the ingredient is made up to the required concentration in the dosing tank, the flow
rate was assumed to be controlled in ratio to the evaporator flow. One particular milk powder plant identified that the concentration of the ingredient showed
119
unexpected, large concentration variations in the milk powder and unusually low concentration
120
results, well below the specification limit. The (competing) hypotheses posed by different plant
121
personnel for these two problems were:
5
ACCEPTED MANUSCRIPT
122
123
H1: The large variation in the final ingredient concentration was due to poor repeatability of the analytical technique used for measuring the ingredient concentration. H2: The large variation was due to inherent variations in the processing conditions (perhaps due
125
to environmental disturbances) that were not sufficiently addressed by the control system.
126
H3: The large variation and unexpected below specification results were due to unpredictable
127
RI PT
124
degradation of the ingredient during processing and subsequent storage.
Without further analysis, the available evidence was insufficient to support or reject these hy-
129
potheses, thus preventing controls from being implemented. In part this was due to the fact
130
that the concentration is measured infrequently, off-line, and with a variable multi-day delay,
131
thus making it difficult to trace, troubleshoot and correlate with processing conditions. Hence a
132
Monte Carlo uncertainty analysis was used to solve this dilemma.
133
2.2. Monte Carlo Analysis to Establish the Source of Variation
M AN U
SC
128
134
A Monte Carlo uncertainty analysis was used to address the three hypotheses stated in
135
Section 2.1, namely to test if the observed variations are due to poor control or excessive variations
136
in the incoming product streams.
The Monte Carlo strategy compares the actual probability density function (PDF) distribu-
138
tion of the output, with that computed by propagating the input distributions through a model
139
of the process (Saltelli et al., 2000) as shown in Figure 2. Such a simple, albeit computationally
140
expensive, strategy can be used to estimate the output distributions through complicated non-
141
linear systems without excessive approximation. The two requirements for implementation are a
142
model, in this case given by the mass balance, and knowledge of the input variations which can
143
be extracted from the plant historian.
EP
Rearranging Equation 1 to solve for the milk concentration in final product milk powder,
AC C
144
TE D
137
C=
m∗ qD (1 − xm )(1 − L) V qE xs
(2)
145
gives the predictive model used to propagate the input distributions through to the required
146
output distribution in Figure 2.
147
Each argument in the right hand side of Equation 2 (inputs in Figure 2) is quantified using
148
historical plant data without having to make any additional assumptions about the underlying
149
distributions. The following assumptions were used:
6
RI PT
ACCEPTED MANUSCRIPT
Figure 2: Diagram of the Monte Carlo uncertainty analysis strategy for quantifying the milk powder ingredient
151
1. For the variables qE , qD , xs and V , the plant historian was used to extract the data directly. Using sufficient data, the PDFs of each input can be regressed.
M AN U
150
SC
concentration variation.
152
2. The final moisture content, xm was assumed constant since it is both very tightly controlled,
153
and given the magnitude relative to the other variables, could not have contributed more
154
than 0.25 % variation to the final concentration.
3. The degradation factor, L, is poorly characterised but historically is assumed to be a 0.12
156
for the specific plant. If the variation is large without accounting for variation in the losses,
157
then unpredictable degradation is not required to explain the large output variation.
TE D
155
158
4. The added mass of the ingredient used for dosing tank make-up, m∗, was constant across a
159
specification with an expected normal three-sigma variation of 0.5% due to weighing error,
160
based on historical work carried out at the plant. If the output PDF as computed from the Monte Carlo simulation differed significantly from
162
the actual distribution, then either the postulated input distributions are a poor approximation
163
(unlikely in this instance), or the model is deficient, or there are disturbance variables unac-
164
counted for.
AC C
165
EP
161
After establishing input PDFs, several scenarios were examined to assess the uncertainty of
166
the final ingredient concentration. The first scenario evaluated the concentration variation due
167
to normal process variation, and this was used as the base case for comparison. Following that
168
‘what-if?’ scenarios were simulated to evaluate the effect of possible solutions.
7
ACCEPTED MANUSCRIPT
169
3. Results and Discussion
170
3.1. Input Probability Density Functions
171
172
Due to commercial sensitivity, all processing variable values and associated units have been deliberately removed and/or normalised as this uses industrially confidential data. The dosing tank fill volume PDF shown in Figure 3a was regressed using five months of
174
historical data. The bi-modal behaviour is due to the fact that actually two separate tanks were
175
used; one being re-filled whilst the other tank was dosing. The two modes are most likely due
176
to a slightly different fill level being used, as each tank is manually filled up to an inscribed
177
mark on the tank. The fill level varies by up to 14%, and is always higher than designed for,
178
assuming that the measured volume is calibrated correctly. The PDFs were converted to cumu-
179
lative distribution functions (CDFs) for analysis. From the CDF in Figure 3a it was found that
180
approximately 50% of the time, the tank is overfilled by 10% or more and that 95% of the time,
181
the tank is overfilled by at least 6%. This is shown by the line drawn on the CDF of the tank
182
volume in Figure 3a. Given that extra ingredient is added to compensate for processing losses,
183
the additional ingredient could be compensating for over-dilution.
M AN U
SC
RI PT
173
184
Figure 3b shows the PDF and CDF functions for the evaporator feed total solids. The
185
evaporator solids vary significantly and this variation was not accounted for during dosing control
186
historically.
The evaporator feed and dosing flow rate PDFs and CDFs are shown in Figures 3c and 3d.
188
Prior to this analysis it was assumed that the dosing flow rate was controlled in a specific ratio
189
to the evaporator flow rate. The dosing flow rate should be exactly proportional to the evap-
190
orator flow rate—when the evaporator flow rate increases by 10 % from the original value, the
191
dosing flow rate should increase by 10 % as well to provide enough ingredient for the extra solids
192
flow. However, the dosing flow rate only increased by 1 %, which would have been insufficient
193
to maintain the required concentration of ingredient in the final milk powder. As a consequence
194
of this investigation, the variation in the dosing flow rate was found to be far too small for it to
195
be in ratio to the evaporator flow. This is shown diagrammatically in Figure 4. This, seemingly
196
trivial oversight, was an unexpected result.
EP
AC C
197
TE D
187
However, the measured dosing flow rate did increase very slightly with the evaporator feed
198
flow, most likely due to a venturi effect at the dosing point. The evaporator and dosing flow
199
rates have a Spearman’s correlation coefficient of 0.84, and this was used in the Iman-Conover
200
method (Iman and Conover, 1982), to generate a correlated sample for the analysis.
8
RI PT
ACCEPTED MANUSCRIPT
(b)
(c)
M AN U
SC
(a)
(d)
Figure 3: Probability and cumulative density functions for a) dosing tank fill level b) total solids entering the
AC C
EP
TE D
evaporator c) flow rate to the evaporator d) dosing flow rate.
Figure 4: A comparison of the current normalised dosing flow rate to the evaporator flow rate overlayed with the design ratio (solid red line) that should be used for control. The correct ratio stipulates that a percentage increase in the evaporator flow should result in an equal percentage increase of the dosing flow rate.
9
ACCEPTED MANUSCRIPT
201
3.2. A Comparison of Actual Variation of Ingredient Concentration The first case investigated simulated the expected variation due to the normal processing
203
conditions encountered at the plant. Figure 5 shows a comparison of the expected variation due
204
to the natural input variations and subsequent processing, with and without analytical variation
205
(red and orange curves respectively), and the actual variation found in the milk powder ingredient
206
concentration (blue curve). For this simulation, a population of 50,000 data random points were
207
used for each input. The realisation of the random variates were generated by passing uniformly
208
distributed random numbers through the inverse cumulative distribution functions (iCDFs) of
209
each input.
RI PT
202
The Kolmogorov-Smirnov test was used to compare the similarity between the actual and
211
simulated distributions (Wilks, 1995). Even with the inclusion of the addition of the analytical
212
variation to the processing variation, the distributions are statistically different with a p-value
213
of 5.7×10−12 at the 95 % confidence level. The simulated distributions was found to be narrower
214
than actual, which shows that the Monte Carlo simulation did not account for all possible sources
215
of variation. However, both the simulated and actual distributions were right skewed with the
216
mean above the mid-specification value (normalised to zero in the figure). Both the actual and
217
simulated distributions also show that the concentration is more likely to be out of specification
218
on the upper limit, as opposed to the lower limit. However, something other than the normal
219
processing and analytical variation is contributing to the out-of-specification results, and thus a
220
case-by-case analysis was used.
AC C
EP
TE D
M AN U
SC
210
Figure 5: Comparison of the distribution of the predicted output concentration using an MC analysis with the actual measured distribution of the data.
221
The repeatability of the analytical technique had been quantified previously from historical 10
ACCEPTED MANUSCRIPT
work carried out at the plants, and the coefficient of variation was found to be 3.6 %. This is
223
of the same order of magnitude as the standard deviation of the simulation which accounts for
224
processing variation only, of 4.7 %. Figure 5 shows the expected output concentration if this
225
analytical variation is added on top of the processing variation, assuming that it comes from nor-
226
mally distributed data. As expected, it increases the total variation seen in the simulated results.
227
The difference between the distributions, with and without analytical variation, is statistically
228
significant at a 95 % confidence level with a p-value of 9.1×10−79 when using the two sample
229
Kolmogorov-Smirnov test. This means that a reduction in the variation due to either processing,
230
or the analytical technique, will decrease the final measured concentration variation in the milk
231
powder.
SC
RI PT
222
Consequently both hypotheses H1 and H2 by plant personnel contributed to the variation
233
found in the milk powder, and that scenarios could be tested to make recommendations for
234
improving the system operation. These are discussed in Sections 3.3 and 3.4. On the other
235
hand if the variation due to the analytical technique was significantly larger than that due to
236
processing, then any improvements in processing would end up being masked and therefore
237
fruitless.
M AN U
232
However, neither the normal variation, nor the variation in the analytical technique explained
239
the unexpected below-specification results that were defined as part of the problem. Furthermore,
240
normal processing variation was unlikely to lead to any below-specification results, since Figure
241
5 clearly indicates that off-specification results are far more probable on the high end given the
242
normal processing variation. Thus in order to adequately address the concerns of the plant
243
personnel, and find a clear explanation, the below-specification results were investigated case-
244
by-case. This is discussed in Section 3.5.
245
3.3. The Effect of Dosing Tank Fill Volume to the Concentration Variation
EP
TE D
238
Section 3.1 showed that because the dosing tanks are filled manually to a pre-specified mark,
247
there is actually significant variation in the dosing tank fill volume, with the volume usually
248
being larger than that designed. There was some concern that this manual operation was to
249
blame, and it would be desirable to quantify the economic implications in order to balance the
250
costs of implementing an automated filling procedure. To do this, a Monte Carlo simulation
251
was performed where the tank fill volume input variable is replaced with the constant setpoint
252
thereby removing the influence of the uncontrolled variations.
253
AC C
246
Figure 6 shows the effect of filling the dosing tanks to 100% of the intended design volume.
11
M AN U
SC
RI PT
ACCEPTED MANUSCRIPT
Figure 6: A comparison between the predicted distributions of the ingredient concentration in the milk powder with variable dosing tank fill volume (upper) and if the tank was filled exactly to the designed volume (lower).
254
Interestingly if the tank is filled to the design volume, then this means that a significant portion
255
of the results would be expected to violate the upper limit, approximately 16%. This implies that the plant processing losses (assumed to be 12%) may not be as high as
257
previously estimated by the plant personnel, and that the current over-dilution compensates
258
for this. Clearly some of the losses are not encountered during the processing so much as by
259
unaccounted for dilution. Therefore, either the tank volume fill procedure should be better
260
controlled, or the dosing flow rate should be adjusted based on the actual volume of water used
261
to dissolve the ingredient.
EP
TE D
256
Maintaining a constant tank volume would not significantly reduce the concentration variation
263
in the milk powder on its own. Therefore the reinstatement of ratio control was also investigated,
264
discussed in the following section.
265
3.4. The Effect of Ratio Control on Ingredient Concentration Variation
266
AC C
262
Figure 7 shows a comparison of different processing changes that could be made and what
267
impact they would have on the variation of the ingredient concentration. Proper ratio control
268
could be instated, with the dosing flow rate changing in proportion to the evaporator flow rate,
269
to account for the increase is solids being sent to the dryer. With the use of proper ratio control
12
ACCEPTED MANUSCRIPT
(Figure 7b), the distribution becomes significantly narrower (compared with Figure 7a), with the
271
standard deviation reducing from 4.7 to 3.7 %. Furthermore, if the tank is filled to a consistent
272
volume each time, then the variability decreases further, with the standard deviation reducing to
273
3.1 % (see Figure 7c). Having the dosing flow rate ratio controlled and a constant tank volume
274
would allow the plant to be operated closer to the upper limit.
RI PT
270
The distribution of the ingredient in the milk powder in Figure 7c now mirrors the total
276
solids distribution (Figure 3b), and therefore if ratio control is implemented for the evaporator
277
flow rate, it should be in ratio to mass solids flow, rather than the volumetric flowrate, as the
278
water that contributes to the volumetric flow rate is removed during the drying process, leaving
279
effectively only the solids behind.
280
3.5. Case-by-Case Analyses
SC
275
The work in part was initiated due to unexpected below specification results. The broad fault
282
diagnosis strategy found that off-specification results were more likely to be on the high limit
283
due to normal process variation. However, the Monte Carlo analysis did not reveal a possible
284
cause of the unexpectedly low below-specification results.
M AN U
281
Therefore, a case-by-case analysis was employed to investigate these intermittent occurrences,
286
using the the cause-and-effect diagram shown in Figure 8. Given that the Monte Carlo analysis
287
from Section 2.2 eliminated normal operation as being a possible cause, we looked for abnormal
288
operating conditions. Diagrams and data of the operating variables at the time of the event were
289
tracked and investigated, and the possible causes, shown in Figure 8, were eliminated one by one.
290
The degradation hypothesis assumed that all relevant plant systems are functioning correctly,
291
however, this was never verified for the low concentration results. Therefore, the dosing system,
292
and all relevant plant systems, were explored during these specific runs to see if any abnormal
293
behaviour was present at the time.
EP
TE D
285
During the case-by-case verification of normal system performance abnormal operations were
295
found for very short durations that coincided with the extremely low test results. Two repeating
296
themes were found for these results that were likely have caused them:
299
AC C
294
300
2. Unintended mixing of different specification powders. This was also not picked up on using
297
298
301
1. Temporary dosing suspension. This was not picked up on with the Monte Carlo analysis as the data range used did not cover the periods where this occurred, as this is extremely
rare and occurs only for very short durations.
the Monte Carlo analysis, as this data was not included in the model used. 13
SC
M AN U
(a)
RI PT
ACCEPTED MANUSCRIPT
(c)
AC C
EP
TE D
(b)
Figure 7: Comparison of effect of different processing scenarios on the final concentration in the powder a) current dosing practice b) effect of the implementation of ingredient dosing ratio control to the evaporator flow rate on the ingredient concentration c) effect of the implementation of ingredient dosing ratio control to the evaporator flow rate with a constant tank volume.
14
RI PT
ACCEPTED MANUSCRIPT
SC
Figure 8: Fish bone diagram showing the different aspects of the process that were assessed systematically for performance during off-specification events.
Due to the two-tank dosing system, a fresh tank has to be made up prior to the dosing tank
303
running dry, preferably with sufficient time for adequate mixing. However if the fresh solution is
304
not made up in time, then dosing is suspended until fresh solution becomes available. This was
305
found to have occurred on days when below-specification results were detected roughly during
306
the sampling time period.
M AN U
302
The other common cause found was due to the management of the milk powder storage
308
hoppers. Given that not all powder is fortified and large hoppers are used when packing powder
309
for transport, if a small amount of unfortified powder remains at the bottom of the storage
310
hoppers when a new, fortified specification powder is packed on top, then this will to be picked
311
up as a below-specification result during testing. During these occurrences the hopper level was
312
not quite zero, and although the quantity remaining was tiny in reference to the total amount of
313
powder produced, it is more than sufficient to be easily picked up during sampling.
316
317
318
319
EP
315
These causes were not picked up on for a number of reasons that need to be considered when carrying out any fault diagnosis:
AC C
314
TE D
307
1. Neither flow stoppages, nor the mixing of fortified and unfortified powder were hypothesised as possible causes for the rare below-specification results. It was assumed that the system operated always as intended, and thus these possibilities were not investigated. In retrospect all possibilities should be eliminated using supporting data.
320
2. The concentration is measured with a significant delay. This lack of immediate feedback
321
means that possible causes are easily overlooked or forgotten amongst the other information
15
ACCEPTED MANUSCRIPT
322
needed to run a complex plant. This means that the fault search space needs to look beyond
323
pre-conceived ideas.
325
326
3.6. Recommended Operational Changes Both the Monte Carlo approach and the case-by-case analysis resulted in distinct recommendations being suggested for addressing the system performance.
RI PT
324
The Monte Carlo analysis showed that correcting the ratio control gain of the dosing flow
328
rate, preferably to the mass of solids passing through the dryer, would reduced the variation
329
of the ingredient in the milk powder. Furthermore, the dilution of the ingredient in the dosing
330
tank should also be accounted for, whether through better control of the fill volume, or through
331
adjustment of the dosing rate based on the actual dilution. The consequences of these can be
332
easily economically quantified. The MC analysis could also quantify the relative importance of
333
decreasing processing variance as opposed to decreasing the variance of the incoming material
334
streams.
M AN U
SC
327
The unexpected below-specification results can be eliminated by giving operators adequate
336
warning when the dosing tank is going to run empty to prevent dosing stoppages, and by ensuring
337
that unfortified and fortified powders do not get mixed inadvertently during packing. This could
338
be programmed as a warning into the human machine interface (HMI), or another plant systems.
339
4. Conclusions
TE D
335
This work looked at the fault diagnosis of a specific issue at a large, complex industrial milk
341
powder plant. Two strategies were employed to investigate why a fortification ingredient showed
342
a large variation in the final milk powder and unexpected below-specification results: a broad
343
Monte Carlo analysis, supplemented by a case-by-case study. This resulted in a wider framing
344
of the problem, and such a dual-pronged approach is necessary when it is likely that more than
345
one issue is actually active at any one time.
AC C
346
EP
340
In this industrial case study, it was found that the large variation came from normal processing
347
which was quantified by the MC analysis, and the below-specification results came from simple
348
causes such as flow stoppages and inadvertent mixing of fortified and unfortified milk powder.
349
Both issues have been addressed with specific process changes to work towards real-time quality.
350
In general, the fault diagnosis followed a number of different hypothesised causes, and we
351
found that all had to be reconciled for any recommendations to be adopted at the plant. A
352
single tool is unlikely to be completely successful given that at any one time on a complex 16
ACCEPTED MANUSCRIPT
industrial plant, the high probability of multiple seemingly independent, possibly simple, issues.
354
However simple problems in complex environments are not so simple to diagnose due to the
355
interconnections, noise and amount of data that has to be processed in a large system. Hence
356
the suggestion to use a two-step approach consisting of using statistical models to reduce the
357
search space, after which one can then undertake a more detailed, directed efficient case-by-case
358
study.
359
5. Acknowledgements
RI PT
353
The authors would like to acknowledge the Primary Growth Partnership program from the
361
New Zealand Ministry of Primary Industries for funding the project and would also like to thank
362
Fonterra staff, specifically James Winchester, Richard Croy, Hong Chen, Brian Woods, Steve
363
Holroyd, Nigel Russell and Tristan Hunter for providing resources and support throughout the
364
project.
365
References
366
Accumolli, A. (1996). Manufacturing process problem analysis using Monte Carlo simulation.
M AN U
367
SC
360
Quality and Reliability Engineering International, 12:39–41.
Bertanza, G., Pedrazzani, R., Manili, L., and Menoni, L. (2013). Bio-p release in the final
369
clarifiers of a large WWTP with co-precipitation: Key factors and troubleshooting. Chemical
370
Engineering Journal, 230:195–201.
373
374
using Monte Carlo simulation. Biotechnology and Bioengineering, 90(2):167–179.
EP
372
Biwer, A., Griffith, S., and Cooney, C. (2005). Uncertainty analysis of penicillin V production
Chiang, L., Russell, E., and Braatz, R. (2001). Fault Detection and Diagnosis in Industrial Systems. Springer-Verlag London.
AC C
371
TE D
368
375
Eslamloueyan, R. (2011). Designing a hierarchical neural network based on fuzzy clustering for
376
fault diagnosis of the Tennessee Eastman process. Applied Soft Computing, 11(1):1407–1415.
377
FDA (2004). Guidance for industry PAT —- A framework for innovative pharmaceutical de-
378
velopment, manufacturing, and quality assurance. Technical report, U.S. Department of
379
Health and Human Services, Food and Drug Administration, 5600 Fishers Lane, Rockville,
380
MD 20857, USA. 17
ACCEPTED MANUSCRIPT
382
383
384
385
386
Ge, Z., Chen, T., and Song, Z. (2011). Quality prediction for polypropylene production process based on clgpr model. Control Engineering Practice, 19:423–432. Gertler, J. (1998). Fault detection and diagnosis in engineering systems. Marcel Dekker, Inc, New York. Goldfeld, A. and Dubi, A. (1987). Monte Carlo methods in reliability engineering. Quality and Reliability Engineering International, 3:83–91.
RI PT
381
Hunter, T. J., Russell, N., Wilson, D. I., Young, B. R., Munir, T., and Depree, N. (2012).
388
Developing a real-time quality programme for dairy. In Chemeca, 23-36 September 2012,
389
Wellington, New Zealand.
SC
387
Iman, R. L. and Conover, W. J. (1982). A distribution free approach to induce rank correlation
391
among input variables. Communication in Statistics-Simulation and Computation, 11:311–
392
334.
394
395
396
Kister, H., Stupin, W., Lenferink, J. O., and Stupin, S. (2007). Troubleshooting a packing maldistribution upset. Chemical Engineering Research and Design, 85(1):136 – 143. Kister, H. Z. (2014). Chapter 2 - common technques for distillation troubleshooting. In Garak, A. and Schoenmakers, H., editors, Distillation, pages 37 – 101. Academic Press, Boston.
TE D
393
M AN U
390
397
Lee, G. (2004). Multiple-fault diagnosis of the tennessee eastman process based on system decom-
398
position and dynamic PLS. Industrial and Engineering Chemistry Research, 43(25):8037–
399
8048.
Li, G., Alcala, C., Qin, S. J., and Zhou, D. (2011). Generalized reconstruction-based contribu-
401
tions for output-relevant fault diagnosis with application to the Tennessee Eastman process.
402
Control Systems Technology, IEEE Transactions, 19(5):1114–1127.
EP
400
Mills, D. (2016). Chapter 24 - troubleshooting and material flow problems. In Mills, D., ed-
404
itor, Pneumatic Conveying Design Guide (Third Edition), pages 555 – 576. Butterworth-
405
AC C
403
Heinemann, third edition edition.
406
Munir, M. T., Yu., W., Young, B. R., and Wilson, D. I. (2015). The current status of pro-
407
cess analytical technologies in the dairy industry. Trends in Food Science & Technology,
408
43(2):205–218.
18
ACCEPTED MANUSCRIPT
410
411
412
413
414
Qin, S. J. (2012). Survey on data-driven industrial process monitoring and diagnosis. Annual Reviews in Control, 36:220–234. Ralston, P., DePuy, G., and Graham, J. H. (2001). Complant-based monitoring and fault diagnosis: a chemical process case study. ISA Transactions, 40:85–98. Rich, S. H. and Venkatasubramanian, V. (1987). Model-based reasoning in diagnostic expert
RI PT
409
ssystem for chemical process plants. Computer Chemical Engineering, 11(2):111–122.
Rimpil¨ ainen, V., Kaipio, J. P., Depree, N., Young, B. R., and Wilson, D. I. (2015). Predicting
416
functional properties of milk powder based on manufacturing data in an industrial-scale
417
powder plant. Journal of Food Engineering, 153:12–19.
419
420
421
Ruiz, D., maria Nougues, J., and Puigjaner, L. (2001). Fault diagnosis support system for complex chemical plants. Computers and Chemical Engineering, 25:151–160.
M AN U
418
SC
415
Saltelli, A., Chan, K., and Scott, E. M. (2000). Sensitivity Analysis. John Wiley & Sons Ltd, West Sussex.
422
Sin, G., Gernaey, K. V., Neumann, M. B., van Loosdrecht, M. C., and Gujer, W. (2009). Un-
423
certainty analysis in {WWTP} model applications: A critical discussion using an example
424
from design. Water Research, 43(11):2894 – 2906.
426
Singhal, A. and Seborg, D. E. (2006). Evaluation of a pattern matching method for the Tennessee
TE D
425
Eastman challenge process. Journal of Process Control, 16(6):601–613. Sparks, T. and Chase, G. (2016). Section 7 - filter selection, process design, testing, optimization
428
and troubleshooting guidelines. In Sparks, T., , and Chase, G., editors, Filters and Filtration
429
Handbook (Sixth Edition), pages 383 – 413. Butterworth-Heinemann, Oxford, sixth edition
430
edition.
432
Wilks, D. S. (1995). Chapter 5 hypothesis testing. In Statistical Methods in the Atmospheric
AC C
431
EP
427
Sciences, volume 59 of International Geophysics, pages 114 – 158. Academic Press.
433
Yin, S., Ding, S. X., Haghani, A., Hao, H., and Zhang, P. (2012). A comparison study of basic
434
data-driven fault diagnosis and process monitoring methods on the benchmark Tennessee
435
436
437
Eastman process. Journal of Process Control, 22:1567–1581.
Zahedi, G., Saba, S., al Otaibi, M., and Mohd-Yusof, K. (2011). Troubleshooting of crude oil desalination plant using fuzzy expert system. Desalination, 266:162–170. 19
ACCEPTED MANUSCRIPT
•
AC C
EP
TE D
M AN U
SC
•
Significant variation in fortified milk powder quality Large, complex industrial powder plant made issues difficult to track Monte Carlo analysis used to understand the processing sources of the large quality variation Case-by-case analysis used for off-spec events guided by Monte Carlo results Fault diagnosis required to ensure all causes were found to produce consistent quality powder
RI PT
• • •