A Debottlenecking Study of An Industrial

Downloaded from orbit.dtu.
dk on: Nov 01, 2024
A Debottlenecking Study of an Industrial Pharmaceutical Batch Plant
Baehner, Franz D.; Huusom, Jakob K.
Published in:
Industrial & Engineering Chemistry Research
Link to article, DOI:

10.1021/acs.iecr.9b03134
Publication date:
2019
Document Version
Peer reviewed version
Link back to DTU Orbit
Citation (APA):
Baehner, F. D., & Huusom, J. K. (2019). A Debottlenecking Study of an Industrial Pharmaceutical Batch Plant.
Industrial & Engineering Chemistry Research, 58(43), 20003-20013. https://doi.org/10.1021/acs.iecr.9b03134
General rights
Copyright and moral rights for the publications made accessible in the public portal are retained by the authors and/or other copyright
owners and it is a condition of accessing publications that users recognise and abide by the legal requirements associated with these rights.
 Users may download and print one copy of any publication from the public portal for the purpose of private study or research.
 You may not further distribute the material or use it for any profit-making activity or commercial gain
 You may freely distribute the URL identifying the publication in the public portal
If you believe that this document breaches copyright please contact us providing details, and we will remove access to the work immediately
and investigate your claim.
Subscriber access provided by DTU Library
Process Systems Engineering

A Debottlenecking Study of an Industrial Pharmaceutical Batch Plant
Franz David Baehner, and Jakob Kjøbsted Huusom
Ind. Eng. Chem. Res., Just Accepted Manuscript • DOI: 10.1021/acs.iecr.9b03134 • Publication Date (Web): 07 Oct 2019
Downloaded from pubs.acs.org on October 14, 2019
Just Accepted
“Just Accepted” manuscripts have been peer-reviewed and accepted for publication. They are posted
online prior to technical editing, formatting for publication and author proofing. The American Chemical
Society provides “Just Accepted” as a service to the research community to expedite the dissemination
of scientific material as soon as possible after acceptance. “Just Accepted” manuscripts appear in
full in PDF format accompanied by an HTML abstract. “Just Accepted” manuscripts have been fully
peer reviewed, but should not be considered the official version of record. They are citable by the
Digital Object Identifier (DOI®). “Just Accepted” is an optional service offered to authors. Therefore,
the “Just Accepted” Web site may not include all articles that will be published in the journal. After
a manuscript is technically edited and formatted, it will be removed from the “Just Accepted” Web
site and published as an ASAP article. Note that technical editing may introduce minor changes
to the manuscript text and/or graphics which could affect content, and all legal disclaimers and
ethical guidelines that apply to the journal pertain. ACS cannot be held responsible for errors or
consequences arising from the use of information contained in these “Just Accepted” manuscripts.
is published by the American Chemical Society. 1155 Sixteenth Street N.W.,

Washington, DC 20036
Published by American Chemical Society. Copyright © American Chemical Society.
However, no copyright claim is made to original U.S. Government works, or works
produced by employees of any Commonwealth realm Crown government in the course
of their duties.
Page 1 of 38 Industrial & Engineering Chemistry Research
1
2
3
4
5
A Debottlenecking Study of an Industrial

6
7
8
9
10
11
12
Pharmaceutical Batch Plant
13
14
∗
15 Franz D. Bähner and Jakob K. Huusom
1
16
17
18 Process and Systems Engineering Center (PROSYS), Department of Chemical and
19
20 Biochemical Engineering, Technical University of Denmark, 2800 Lyngby, Denmark
21
22
23 E-mail: [email protected]
24
25
26 2 Abstract
27
28 3 Operational complexity in batch-operated plants is large and companies struggle
29
30 4 to produce at high equipment utilisation. Due to high xed costs, capacity utilisation
31
32 5 is an important operational directive. Many current plants are not in a state that
33
34 6 allows the application of rigorous process systems engineering tools due to modelling
35
36 7 challenges. This work proposes identifying the right engineering projects based on
37
38
8 statistic evidence. In the case of incomplete process monitoring strategies in semi-
39 9 automated facilities, these analyses are specically challenging. The power of modern
40
41 10 data processing tools in this context is shown at hand of a case study in an industrial
42
43 11 pharmaceutical production. This includes the development of a recursive monitoring
44
45 12 algorithm as well as plant performance evaluation based on established heuristics.
46
47
48
49
50
13 1 Introduction
51
52
14 The operational paradigm of batch process plants producing high-value products at com-
53
54
15 parably low volumes diers from that of bulk chemical processes. Such batch plants are
55
56
16 frequently encountered in speciality chemical-, pharmaceutical-, and bio-based industries.
57
58
59 1
60 ACS Paragon Plus Environment
Industrial & Engineering Chemistry Research Page 2 of 38
1
2
3
17 Due to operational complexity especially in multi-product scenarios, it is dicult to operate
4
5
18 these plants at maximum throughput. Therefore, incremental debottlenecking is generally
6
7
19 among the primary production objectives. If high xed costs arise from quality control and
8
9 20 quality assurance in the context of good manufacturing practice (GMP) compliance,
1,2
the
10
11 21 importance of high capacity utilisation is increased even more. Furthermore, reductions in
12
13 22 public healthcare spending by means of price caps or endorsed production of generic (non-
14
15 23 branded) drugs is expected.
3
This forces pharmaceutical enterprises to streamline existing
16
17 24 production facilities and supply chains.
4
The situation is complicated further for research-
18
19 25 driven pharmaceutical enterprises due to the complex interrelations between capacity plan-
20
21 26 ning, risk management (i.e. from failed new drug development processes), and operational
22
23 27 issues.
5
24
25 28 Incremental debottlenecking of continuous plants by installing additional recycles or by other
26
27 29 means of deecting workload from the bottleneck to less pressed units is generally possible.
6
28
29 30 Yet, exibility in such plants is limited compared to designs with parallelised, intensied
30
31 31 units
7
(scale-out vs. scale-up). Accordingly, systems of parallel standard units facilitate
32
33 32 organic growth, enabling plants to track markets more exibly. This requires continual (ex-
34
35 33 pensive) engineering expertise and therefore eective project execution. Not least, parallel
36
37 34 standard units facilitate performance monitoring and predictive maintenance as trends from
38
39 35 equipment wear & tear are more easily distinguishable from trends induced by uid proper-
40
41 36 ties.
42
43 37 Several frameworks exist for retrot debottlenecking of batch process systems, the most re-
44
45 38 cent by Amaran et al., and a large base of works has been presented by and around Petrides
46
47 39 et al.. All of these frameworks advocate the use of computational tools, which ts well
48
49 40 the desire to more consciously apply process systems engineering methods in the bio-based
50
51 41 and i.e. pharmaceutical industries.
10
In an industrial context it is however not so that
52
53 42 model-based work is apt for all projects. Particular challenges arise in bio-based produc-
54
55 43 tion, and model-building as well as verication are likely to be expensive or even impossible
56
57
58
59 2
1
2
3
44 due to unmeasured, uncertain, and unknown variables. This can stem from the complexity
4
5 11 12
45 of solid-liquid suspension properties, large natural process variability, but also due to
6
7
46 non-deterministic manual control and incomplete process monitoring systems. This work
8
9 47 discusses these and other challenges, and proposes a solution approach. To this end, rstly
10
11 48 the general notion of batch processing, economical considerations, and arising operational
12
13 49 challenges are introduced in section 2. Regard is paid to the special case of bio-based pro-
14
15 50 cesses. Identication and handling of bottlenecks are discussed in section 3. Important
16
17 51 heuristics are extracted from batch process plant debottlenecking frameworks. An algorithm
18
19 52 for the reconstruction of machine states based on time-series values of commonly measured
20
21 53 variables is introduced in section 4. This is necessary to identify scheduling bottlenecks in
22
23 54 process plants without conclusive monitoring schemes, as is also the case in the regarded
24
25 55 production plant. This industrial case study is described in section 5, functionality of the
26
27 56 algorithm shown, and debottlenecking candidate projects derived accordingly. Operational
28
29 57 challenges and generalisations are discussed in section 6, hereafter the work is concluded.
30
31
32
33
34 58 2 Background
35
36
37 59 This section discusses batch processes at a high level of abstraction as well as disadvantages
38
39 60 and challenges immanent in them. This is embedded in the context of bio-based processes
40
41 61 before introducing retrot process re-designs aimed at debottlenecking.
42
43
44
45 62 2.1 Economics of Batch Process Plants
46
47 63 In a single-product scenario, continuous plants outperform batch processes. An estimate for
48
49 64 possible savings is given for instance by Calabrese and Pissavini. Bauer and Craig state
50
51 65 that there are less batch processes than continuous in chemical, petrochemical, and rening
52
53 66 industries. This prevalence of continuous processes over batch production in large sectors of
54
55 67 the chemical industry is a good indicator for their economic superiority, as batch operation
56
57
58
59 3
1
2
3 15
68 constituted the rst paradigm. One explanation for this lies in lower equipment ecien-
4
5 16 17
69 cies, which may be as low as 30% in pharmaceutical plants.
6
7
70 Furthermore, reductions in equipment scale and consequently also temperature/concentration
8
9 71 gradients due to better heat and mass transfer are expected for continuous reactors.
18
Co-
10
11 72 standy et al. oer an optimisation-based framework for the comparison of batch- and contin-
12
13 73 uous reactors based on rst-principles. A more practical approach to evaluating the benets
14
15 74 of continuous processing including downstream lines is presented by Teoh et al.. Continuous
16
17 75 processes can also be preferable for safety reasons due to reductions of hazardous material
18
19 76 hold-ups,
21
and have been classied as more easily automatable.
19
Finally, a batch sched-
20
21 77 ule impedes process integration as the points in time of heating and cooling may change
22
23 78 relative to one-another.
22
This requires either advanced scheduling, which is expectedly con-
24
25 79 nected with some cost, or extra equipment in the form of an intermediate thermal storage
26
27 80 to decouple the integration events in time.
28
29
30
31 81 2.1.1 Batch Production of Specialty Chemicals
32
33 82 To reap the benets of economies of scale, suppliers of volume-limited markets need to
34
35 83 operate exible multi-purpose plants.
23
Due to changeover and shut-down as well as start-
36
37 84 up costs, continuous multi-purpose plants can be unattractive. Furthermore, batch duration
38
39 85 as an extra degree-of-freedom renders systems highly exible. Fifteen years ago, Edgar
40
41 86 claim that, in total, there were more batch operated plants than continuous plants in the
42
43 87 United States. More recently, Kano and Ogawa or Amaran et al. mention the growing
44
45 88 importance of specialised products and consequently batch processes. An alternative lies
46
47 89 in the development of exible multi-purpose continuous plants, which is however connected
48
49 90 with substantial technological challenges.
26
50
51
52
53
54
55
56
57
58
59 4
1
2
3
4
91 2.1.2 Bio-Based and Pharmaceutical Batch Production
5
6 92 In bio-based production, but especially in the case of pharmaceuticals, the plantwide pro-
7
8 93 duction regime is traditionally batch, likely for three reasons:
9
10
11 94 Fermentation was initially a batch process (and frequently still is), this operational
12
95 regime is propagated through the downstream line.
13
14
15 96 Recurrent cleaning and sterilisation of units is necessary to prevent cross-contamination
16
17 97 between batches and accumulation of biological cell matter in general.
18
19
20 98 Limited batch durations reduce the risk of mutation in the deployed organisms.
21
22
99 The latter also implies that, aside from the complications in heat-integrating processes, ma-
23
24
100 terial recycles may be entirely forbidden due to cross-contamination risk policies calling for
25
26
101 rigorous cleaning-in-place (CIP) or sterilisation-in-place (SIP) barriers.
27
28
102 Due to the historical prevalence of batch processes, regulatory entities (U.S. Food and Drug
29
30
103 Administration, European Medicines Agency, etc.) as well as industrial stakeholders have
31
32
104 accumulated a lot of experience from frequent qualication of these processes. This is also
33
34
105 mentioned by Wu et al., who acknowledge that batch processes are likely to have lower ef-
35
36
106 ciencies, but bear the convenience of the "three-batch process validation approach". This
37
38
107 adds a lot of inertia in a transition to continuous process regimes, as especially in rst-to-
39
40
108 market scenarios commissioning delays are likely to outweigh suboptimally high operational
41
42 27
109 costs - even if they persist for numerous years. Nevertheless, there is a push for the pharma-
43
44 2830
110 ceutical industry to rely more on continuous and semi-continuous processing technology.
45
46
111 This is enabled by the fact that product batches which undergo nal testing and are ulti-
47
48 31,32
112 mately sold need not strictly stand in relation to a single upstream batch.
49
50
113 Beyond the process-oriented economic arguments, there are upstream processes that favour
51
52 33
114 a continuous production regime due to biological properties. Schaber et al. compared costs
53
54
115 for continuous and batch processing for a large-scale pharmaceutical production and con-
55
56
116 clude signicant cost savings in the case of continuous processing technology. While most big
57
58
59 5
1
2
3
117 players in pharmaceutical production engage in related research projects, the industry per-
4
5
118 spective is not without restraint, especially concerning the replacement of existing (batch)
6
7 35
119 infrastructure. Croughan et al. point out that production scales of pharmaceuticals are
8
9 120 in general not comparable to bulk chemical- and petrochemical industries. Therefore, they
10
11 121 conclude [...] should we close existing batch operations? The answer, at least for now is
12
13 122 no, but the factory of the future embodied in new facility design is likely to evolve around
14
15 123 integrated continuous bioprocessing". All in all, there is strong evidence that batch plants
16
17 124 in today's chemical and especially biochemical industries are of economic relevance.
18
19
20
21 125 2.2 Operational Objectives in High-Value Bio-Based Production
22
23
24 126 In the above it has been mentioned that bio-based plants tend to operate in batch regimes.
25
26 127 In these plants, operational complexity is substantial due to the additional scheduling tasks
27
28 128 on unit operation layer. Low- and high-level automation may is complicated by the absence
29
30 129 of reliable property models, presence of uncertainties and delays, and not lastly entirely
31
37
32 130 unmeasurable states. Furthermore, the comparably small scale of the facilities and the
33
34 131 indicated disproportion between cost of utilities and xed costs (i.e. quality control) may
35
36 132 complicate the business cases for automation projects. Finally, due to fault-tolerant processes
37
38 133 and equipment breakdowns, exible control systems are needed. This exibility is added
39
40 134 in most easily by including educated operators into the control structure, which implies
41
38
42 135 operating mistakes due to human error. All of the above may lead to suboptimal operation
43
44 136 conditions including suboptimal capacity utilisation, which can easily go unnoticed in a
45
46 137 complex environment. Therefore, if xed costs (for quality control, quality assurance, GMP
47
48 138 facility depreciation, and labour) outweigh the cost of utilities, full capacity utilisation of
49
50 139 installed equipment is likely to be the primary operational objective as long as the market
51
52 140 is not saturated.
53
54
55
56
57
58
59 6
1
2
3
4
141 2.2.1 Retrot Process Design
5
6 142 Most enterprises have to react to market evolution in a recurrent manner. This may be due
7
8 143 to changes in demand, product line-up, raw material supply, energy prices, and emerging
9
10 144 or disappearing competitors. In the operated plants this implies an ongoing sequence of
11
12 39
145 engineering projects, which aligns well with the plantwide control task, also iterative.
13
14 146 Re-optimisations are especially important in newly-built plants after experience has been
15
16 147 gathered, or after signicant process or market changes. In these projects, economic plant
17
18 148 performance should be evaluated based on accumulated sales, process, and product data,
19
20 149 but also extraordinary sampling campaigns may be sensible. This usually requires manual
21
40
22 150 processing as, unlike in the case of control system performance monitoring, data sets
23
24 151 are relatively heterogeneous and unstructured, and on top of that case-specic - making it
25
26 152 dicult to apply a set of standard tools. It is important for an enterprise to enable ecient
27
28 153 and robust execution of these projects; similarly process designs that allow incremental
29
30 154 capacity increases may be a conscious choice.
31
32
33
34
35
155 3 Debottlenecking Methodology
36
37
156 Several frameworks and articles incorporating best practices specically for or including de-
38
39 8,9,23,4144
157 bottlenecking of batch process plants have been presented. Many of these include
40
41 4550
158 application examples, and further cases have been studied. Most of these frameworks
42
43
159 advocate the use of models to augment decision-making. Undoubtedly, the dynamic be-
44
45
160 haviour of complex batch process systems is hard to conceptualise for the human mind,
46
47
161 calling for computational support. However, not every production facility or enterprise is
48
49
162 apt for model-based optimisation. This is discussed for instance by Guimarães et al. for the
50
51 52
163 case of discrete-event models. Here, with a reference to a basis for a maturity measure
52
53
164 for an organisation's aptitude for process modelling is introduced. The latter depends both
54
55 165 on the production process itself, which has to be structured and operate in stable routines,
56
57
58
59 7
1
2
3
166 furthermore automatic data collection needs to be in place. Finally, the right skill-set needs
4
5
167 to be available within the enterprise. These measures are subject to some ambiguity, not
6
7
168 lastly stemming from a lack of documented industrial cases.
8
9 169
10
11
12
13
14
170 3.1 Identication of Bottlenecks
15
53
16 171 Any unit not capable of handling a throughput increase is per denition a bottleneck.
17
18 172 In stable, continuous processes, personnel is usually able to pin-point this unit or these
19
20 173 units with high precision. This information can also be concluded from statistical analysis
21
54
22 174 of process data. Alternatively, the plant can be tried in a maximum throughput trial.
23
24 175 However, the shortness of a trial bears the risk of missing eects that would become relevant
25
26 176 in the long-term. Secondly, it is usually so that operators are more attentive during out-
27
28 177 of-the-ordinary temporary scenarios. Capacity estimates can also be basted on for instance
29
30 178 model-based extrapolation of the current state - this may be necessary if capacity is elevated
31
32 179 past several bottlenecking stages at once. However, if uid properties change frequently and
33
34 180 transient periods make up a substantial part of uptime, the task takes on a more complicated
35
36 181 character. Here, a more advanced analysis including statistical elements is required.
37
38
39 182 3.1.1 Identication of Bottlenecks in Batch Plants
40
41
42 183 In a batch plant all of these complications can arise in the same way. Recalling section
43
44 184 2.1.2, complex scenarios are actually more probable, as batch plants are more likely to pro-
45
46 185 duce multiple products. Furthermore, operation may be less regular due to operators being
47
48 186 actively in-the-loop. Also, one needs to rene the denition of a bottleneck in a batch pro-
49
50 187 cessing context. According to Calandranis and Petrides, bottlenecks can be either size or
51
52 188 time inicted.
53
54 189 The size bottleneck limits the amount of coherent material ('a batch') passing through the
55
56 190 system. It is evidently so that the size bottleneck's volume should be utilised to the maxi-
57
58
59 8
1
2
3
191 mum in each batch. Note that there can be multiple size bottlenecks in processes where the
4
5
192 product changes streams (for example elution processes) or processes with purge streams.
6
7
193 Time bottlenecks are those with the largest stage cycle time, or those where it takes longest
8
9 194 to process a batch until starting the next. Ideally, time bottlenecks operate at 100% equip-
10
11 195 ment uptime, it is then also trivial to understand that throughput on these units cannot
12
13 196 be increased further. However, due to operational complexity, it may not be possible to
14
15 197 challenge a plant in a maximum throughput scenario, thus some form of extrapolation from
16
17 198 nominal production rate may be of need.
18
19 199 In general, batch plants ordinarily employ both 'native' batch and semi-batch units, but also
20
21 200 semi-continuous units that only follow the batch regime of the incorporating plant. For a
22
23 201 detailed overview of common nomenclature, the reader is referred to Barrera.
24
25
26
27 202 3.1.2 Batch and Semi-Batch vs. Semi-Continuous Units
28
29 203 Batch processes are characterised by their size or volume, have specied lling, processing,
30
31 204 draining steps, and no material may be added to or removed during the processing phase. In
32
33 205 fermentations this requirement is usually relaxed as addition of acids/bases for pH control
34
35 206 purposes is allowed.
56
There is always at least one transient variable and a constraint value
36
37 207 for it. This variable can be cycle time, in which case the transient is reduced to a timer. A
38
39 208 batch process is depicted schematically in gure 1 (left), where it can also be seen that for
40
41 209 batch units with predominantly 'holding' operations (size bottlenecks), debottlenecking by
42
43 210 means of volume maximisation is eective (for example fermentation).
44
45 211 Native batch units are contrasted with continuous units operated under a batch operating
46
47 212 regime. A 'semi-continuous' unit can be inactive due to planned downtime (as are all non-
48
49 213 bottleneck units) or CIP / SIP / maintenance activities. However, start-up and shut-down
50
51 214 procedures are generally negligible and the unit operates continuously during uptime. This
52
53 215 is represented by gure 1 (right), and it is also intuitive to understand that scheduling
54
55 216 bottlenecks are more likely to arise upon batch-size increases on 'ow' units (for example
56
57
58
59 9
1
2
3
217 decantation).
4
5
218 The restriction of no addition or removal of material is removed in semi-batch steps. A dead-
6
7
219 end ltration process with uncertain cake properties is a semi-batch process with, depending
8
9 220 on the control strategy (ow rate or pressure controlled regime) an unknown duration or
10
11 221 even unknown ow rate and duration.
57
These processes lie somewhat between batch and
12
13 semi-continuous units in terms of how to identify and treat bottlenecks.
14
15
16
17
Batch Operation Continuous Operation
18
19
20
21
22 a a
Volume
Volume
23
24
25
26
27
28
29
30
31
32
33
Time b Time b
34
35
Figure 1: Native batch vs. semi-continuous units and the eect of throughput increases 'a'
36
37 on cycle time 'b'.
38 222
39
40
41 223 3.1.3 Engineering Projects
42
43
44 224 Implementation of debottlenecking strategies even in a continuous improvement cycle is
45
46 225 usually project-based. Calandranis and Petrides list three primary debottlenecking strategies
47
48 226 (increase number of cycles per batch for the limiting procedure, rearrange the equipment
49
50 227 assignment, introduce new equipment). Amaran et al. suggest the following taxonomy of
51
52 228 possible retrot debottlenecking project turnouts which are applicable to a broader class of
53
54 229 problems:
55
56
230 "Type-1: operational improvements by removing slack and uncovering hidden capacity
57
58
59 10
1
2
3
231 with current infrastructure".
4
5
6 232 "Type-2: upgrades to current equipment, planning and scheduling improvements and
7
8 233 improvements in operating discipline".
9
10
11 234 "Type-3: fundamental improvements in operating policies".
12
13
14 235 "Type-4: investment in new pieces of equipment/trains".
15
16
17 236 The framework by Amaran et al. takes a process systems engineering approach and these
18
19 237 projects are turnouts of model-based simulation studies with an optimisation sub-step. This
20
21 238 allows balancing mathematical optimality and degree of subjectivity, and can therefore cap-
22
23 239 italise exibly on skills and experience of the process optimisation team. Amaran et al.
24
25 240 propose to validate the discrete-event model by means of comparing simulated inter-batch
26
27 241 start times, overall production times, queue lengths, and individual processing and wait times
28
242 to sets of historical data that are most descriptive of future operations. The cycle times of
29
30
243 individual unit operation steps in turn are assumed to be captured by the process monitoring
31
32
244 system. With a reference to the fact that validation can be a very time-consuming step, one
33
34
245 could nd that comparably little attention is paid to it. Further, Amaran et al. remark that
35
36
37 246 especially implementation of Type-3 projects may be dicult as also theses require opera-
38
39 247 tional discipline. The above classication are adapted in the following and projects will be
40
248 dened in these terms.
41
42
43
44
45 249 4 Machine State Reconstruction Algorithm
46
47
48 The absence of control system state information due to manual operation makes it dicult
49
50 to impossible to reliably identify scheduling bottlenecks, furthermore it is not possible to
51
52 validate models merely at hand of the collected time-series data. Sequential control systems
53
54 are characterised by a set of discrete states S; examples are given in table 1. These states
55
56
57
58
59 11
1
2
3 Table 1: Exemplary machine states in batch operation regime.
4
5
6 Type of Common machine states
7 unit (aside from idle)
8
Stabilise
9 Batch Reactor Heat
10 Fill Inoculate Ferment (chemical, Drain CIP
(Fermentation) Treatment
11 cooling)
12
13 Holding Tank Fill Wait for Drain
CIP
14 (Storage ) (from) resource (to)
15
Tank Fill Perform Wait for Drain
16 CIP
17 (Flocculation, ...) (from) operation resource (to)
18
Flow Unit
19
20 (Centrifugation, Active Reinitialise CIP
21 Filtration, ...)
22
23
24 dene the control system outputs ω, and thereby nally the evolution of the set of continu-
25
26 ous states X of the plant as a consequence of actuation. The continuous states furthermore
27
28 experience signicant process noise, for instance from slurry properties as a consequence of
29
30 batch-to-batch variability in fermentations. The control system outputs in this sequential
31
32 context are often binary or categorical (for instance open/closed valve position), but also
33
34 continuous process data (ow, temperature, pH) are measured. Some of the continuous vari-
35
36 ables may be controlled, usually by means of proportional-integral control.
37
38 This knowledge needs to be harnessed into the state reconstruction algorithm. Further infor-
39
40 mation available comes in the form of standard operating procedures, piping and instrumen-
41
42 tation diagrams, and not lastly experience of operators and engineers. A non-deterministic
43
44 recurrent state machine RSM is chosen as a modelling basis:
45
46
47
48 RSM = hM1 , ..., Mk i (1)
49
50 Mi = (Si , s0i , ΣSF Ci , ΣT Si , Fi ) (2)
51
52
53
54 250 For each component state machine Mi , S i describes the set of discrete states and s0i the
55
56 251 initial state (which for all machines is idle). The input alphabet Σ will be delineated in
57
58
59 12
1
2
3
252 greater detail in the following. Fi denotes the set of nal states for the component machines.
4
5
253 In the reconstruction algorithm, the outputs ω of the sequential batch control system become
6
7
254 inputs to the state machine reconstruction algorithm:
8
9
10
11 ΣSF C = f (ω, S) (3)
12
13
14
15 255 This is intuitive, as every time a valve (∈ ω) opens or closes, or a ow magnitude changes
16
17 256 persistently, this must have been dictated by a state transition in S. Note that in this,
18
19 257 operators are control system elements and therefore introduce non-determinism. They make
20
21 258 mistakes and react delayed. In general there may be some leeway in the operating procedures
22
259 that are not strictly product-safety critical. Further, operators may have to react exibly to
23
24
260 changing uid properties: sometimes a routine will have to be called multiple times instead
25
26
261 of once until the desired eect is achieved. Finally, units malfunction, which also leads to
27
28
262 extraordinary state sequences.
29
30
263 Thus, this input alphabet is not enough to reconstruct the machine states due to countless
31
32
264 multiplicities: a valve opening / closing may result in dierent transitions depending on the
33
34
265 current machine step - which may be unknown yet due to preceding multiplicities or irregular
35
36
266 events. Therefore, characteristic points in the time-series data need to be translated into
37
38
267 inputs
39
40
41 ΣT S = f (X, S) , (4)
42
43
268 which is described in more detail in section 4.2. Secondly, it may be necessary to call a sub-
44
45
269 set of unique component state machines which will be supplied with the subsequent input
46
47
270 sequence (both ΣSF C and ΣT S ) to see which - if any - of the machines converges to a nal
48
49
271 state. Upon unequivocal convergence, the algorithm, recursive in its nature, will trigger a
50
51
272 state transition at the point-in-time of initial execution of the subroutines. Otherwise it is
52
53
273 concluded that an erroneous procedure must have occurred during processing of the active
54
55
274 batch which is also logged.
56
57
58
59 13
1
2
3
4
275 4.1 Identication of ΣSF C
5
6 276 As indicated, opening and closing of valves as well as switching motors on or o are trivial
7
8 277 and usable indicators. However, in some cases and especially in the case of ow control,
9
10 278 more scrutiny is of need. Changes in absolute values of rates (for example aeration) can
11
12 279 allow pin-pointing specic steps. Secondly, in the case of strongly sequential operations (for
13
14 280 example a sequence of ve-fold elution or washing steps), a component state machine is
15
16 281 needed that keeps track of the sequence on the one hand, and potentially duration of the
17
18 282 steps on the other. These checks (range-, spread-, and timers) have proven to be powerful
19
20 283 elements for sequential owchart reconstruction.
21
22
23
24 284 4.2 Identication of ΣT S
25
26
285 There are countless characteristic points in the time-series data and it is neither realistic nor
27
28
286 necessary to delineate each one of them as they are highly case-specic. The most trivial
29
30
287 characteristic points are a ow stopping persistently on a 'ow' unit, or a tank running dry
31
32
33
288 for a 'holding' unit. The respective unit then must be in Idle/Empty. Note that for holding
34
35
289 units, Empty is an unambiguous state, whereas Full may be more dicult to pin-point in
36
37
290 the case of altering ll-levels. (Also for Empty, a tolerance may be of need due to sensor
38
291 noise.) As the nalisation of a draining procedure on one tank normally means that a ll
39
40
292 procedure on another tank has been nished, this information is of two-fold value.
41
42
293 Beyond these simple indicators, more advanced routines can easily be included into the
43
44
294 algorithm. Moving-horizon linear ts, smoothers, maximum norms, or variance estimators
45
46
295 can be used eciently to detect and classify changes in uncertain environments. Averaged
47
48
296 threshold values can usually be inferred from observations. These vector operations are
49
50
297 however computationally expensive and should only be applied when scalar criteria fail.
51
52
298 Finally, the duration of a procedure can be estimated early on and gives a good indication
53
54 299 about its nature (standard operation, CIP routine, irregular event), enabling an a-priori
55
56 300 selection of component state machines that will be triggered subsequently. It should be noted
57
58
59 14
1
2
3
301 that, at times, knowledge of future information may be needed in order to unambiguously
4
5
302 classify an event. In these, the algorithm cannot be used for on-line monitoring purposes.
6
7
303 However, the real-time requirements in production are rather soft as there is signicant value
8
9 304 in having knowledge about the past 12-24 hours of production after a shift changeover even
10
11 305 if there is some delay.
12
13
14
15
16
306 4.3 Superstructure, Equipment Assignments
17
18 307 Due to the predominantly forward-oriented ow of product in batch plants without recycle, a
19
20 308 separation into subroutines is possible, which allows step-wise implementation and validation.
21
22 309 Therefore, generally a link between unit state machines only has to be drawn when material is
23
24 310 transferred, which may require identication of a resource (ow unit) in the case of parallel
25
26 311 units, and identication of a target holding unit. By tracking material accumulation /
27
28 312 depletion and by estimating the ow rates into and out of the respective tanks, this link can
29
30 313 then be established with high certainty.
31
32
33
34 314 4.4 Illustrative Example
35
36 315 At hand of a small excerpt of the case study plant (introduced in the next section), these
37
38 316 principles will be demonstrated. The Gantt chart of the example plant in gure 2 is limited
39
40 317 to a single fermenter F E03, from where material is transferred to holding tanks T 11 and
41
42 318 T 12. At hand of the pseudo-code for algorithm 1 it is shown how machine states for the
43
44 319 fermenters are identied from ΣSF C and ΣT S . Secondly, the posterior propagation of batch
45
46 320 numbers to the holding tanks is indicated algorithm 2. On plant scale this is necessary
47
48 321 to create batch traceability, and for complex transitions. Furthermore, sometimes it is not
49
50 322 possible to identify all necessary ΣT S & ΣSF C based on time-series values from one unit
51
52 323 operation, and steps on preceding / succeeding units need to be included. In these cases, the
53
54 324 batch ID needs to be propagated simultaneously with the RSM calculations. (For instance,
55
56 325 this is dierent for the special case of the recycle section (Tanks 31/32, Unit 3, Tanks 41/42
57
58
59 15
1
2
3
4
5
6
Algorithm 1 RSM Principle
7 1: T is vector of time-points
8 2: ω is matrix of inputs (valves initially closed), X is matrix of continuous states
9 3: SF E03 is unknown fermenter machine state
10 4: procedure trackFermStates(X, ω, T )
11
5: SF E03 = idle; tk = T (1)
12
13 6: while tk < T (end) do
14 7: if SF E03 is idle then
15 8: CheckTransitionsFromIdle(... X.F E03.Level, ω.F E03.V alveIn (tk ) , T, tk )
16
17
9: else if SF E03 is f illing then
18 10: CheckTransitionsFromFilling(...)
19 11: else if SF E03 is ... then

20 12: Check transitions from...
21 13: end if
22 14: Increment tk in T
23
24 15: end while
25 16: end procedure
26 17: procedure CheckTransitionsFromIdle(Level, V alveIn, T, tk )
27 18: if V alveIn.opens then
28
19: ti tk ;
=
29
30 20: while Level (ti ) > 0 do
31 21: Increment ti in T (tk : end)
32 22: end while
33 23: ∆TF erm and LF erm are thresholds
34
24: Condition 1: ti -tk > ∆TF erm , Condition 2: avg (Level (tk : ti )) > LF erm
35
36 25: if Condition 1 and Condition 2 then
37 26: store(Ferm. nr. begins at tk )
38 27: store(Ferm. nr. ends at ti )
39 28: SF E03 = f illing . (Transition)
40
41
29: else
42 30: # Irregular event has occured
43 31: Store time-points

44 32: tk = ti . Jump period
45 33: end if
46
47
34: end if
49 36: procedure CheckTransitionsFromFilling(V alveIn)
50 37: if V alveIn.closes then . Simple ΣSF C
51 38: SF E03 = waiting
52
53
39: end if
55
56
57
58
59 16
1
2
3 Algorithm 2 Posterior Propagation of Batch ID
4
5 1: TS,F E03 contains points in time of all fermenter state transitions from algorithm 1 as well
6 as batch IDs
7 2: TS,T 11 and TS,T 12 contain holding tank state transitions
8 3: procedure propagateBatchID(TS,F E03 ,TS,T 11 ,... TS,T 12 )
9
4: # Times of all tank lling events:
10
11 5: TF ill,T 11 = TS,T 11 .allBatches.periodOf(lling)
12 6: TF ill,T 12 = TS,T 12 .allBatches.periodOf(lling)
13 7: for each batchID in TS,F E03 do
14 8: tdrain = batchID.periodOf(draining)
15
16
9: if tdrain = any(TF ill,T 11 ) then
17 10: Propagate batchID to batch in TS,T 11
18 11: else if tdrain = any(TF ill,T 12 ) then

19 12: Propagate batchID to batch in TS,T 12
20 13: else
21
14: Throw Exception . Indicates error in algorithm as each batch must be
22
processed
23
24 15: end if
25 16: end for
27
28
29
326 in gure 3) of the case study plant, which requires a more complex approach, namely one
30
31
327 state machine model describing the simultaneous evolution on ve units.)
32
33
34
35
36
328 4.5 Logging of Important Time-Series Values
37
38 329 While this information is not crucial for the identication and quantication of scheduling
39
40 330 bottlenecks, values beyond points-in-time can be of relevance to the analyst. Volumes as
41
42 331 indicators for size bottlenecks are of evident interest. Still, rened feature engineering is easy
43
44 332 once the overall frame has been implemented, and tracking for instance a relative average
45
46 333 pH change between two steps is trivial. On the other hand, once the time-steps for machine
47
48 334 state changes have been identied, post-hoc extraction of features is similarly easy and it is
49
50 335 a matter of taste how to generate the specic data sets.
51
52
53
54
55
56
57
58
59 17
1
2
3
4
5
6
7 2 3 45 2 3 45 2 3 4 5
8 1 0 1 0 1 0
Unit Operation
9
03
10
FE
11
12
13
1
14
T1
15
16
17
18
2
T1
19
20
21
26-08 29-08 01-09 04-09
22 Datetime
23
24
Figure 2: Gantt chart excerpt with propagation of the batch ID (color). Grey lines indicate
25
26 level in the tanks. (Fermenter states: 0:idle, 1:lling substrate, 2:autoclaving & waiting for
27 inoculation, 3:fermentation, 4:waiting for transfer, 5:emptying )
28
29
30 336 5 Pharmaceutical Production Plant Case Study
31
32
33 337 At hand of an industrial production of fermentation-derived antibiotic, the challenges men-
34
35 338 tioned above are delineated. Furthermore, it is shown that in the case of a strongly capacity-
36
37 339 leveraged process, local improvements can be identied eectively using visual statistics and
38
39 340 domain knowledge. The study is to be of descriptive value, outlining the challenges that
40
41 341 are likely to exist in many semi-automatic production facilities currently in operation, while
42
43 342 proposing some ways of how to handle them.
44
45
46
47 343 5.1 Process System Description
48
49
344 Figure 3 shows the process in terms of semi-continuous and batch units as well as those
50
51
345 entailing strong elements of both. A more detailed description was not compliant with the
52
53
346 intellectual property policy of the industrial partner, but this does not harm the analysis
54
55
347 in the following. The CIP barrier running through a certain fraction of the downstream
56
57
58
59 18
1
2
3
348 section is also indicated. In the regarded plant, the batch number is logged exclusively for
4
5
349 fermentation units. For all other machines, only time-series data (i.e. ows, pressures, levels,
6
7
pH) are available on the historian.
8
9
10 CIP
…
11
12 Recycle
13
14 Fermentation
Tank Tank Unit Tank Tank
15 11 21 21 31 41
16 Fermentation Unit 1 Unit 3 …
17 Tank Tank Unit Tank Tank
12 22 22 32 42
18 Fermentation
19
20
21
22
CIP
23 …
24 Tank
25 Tank
61
… Unit 4 Unit 4 End of analysis (no data)
26 5
27 Tank
62
28
29
30 Figure 3: Indication of process owsheet in terms of batch (predominantly holding: 'tank')
31 and ow 'units'. Units 21 and 22 (light orange underlay) with strong aspects of both. Note
32 the structured CIP that passes through a certain fraction of the downstream line sequentially.
33
34 350
35
36
37 351 5.1.1 Application of State Reconstruction Algorithm
38
39
352 The proposed recursive state machine based algorithm is applied, which allows reconstruc-
40
41
353 tion of the machine states for an entire production campaign of almost 70 batches with high
42
43
354 certainty. This enables taking advantage of the advanced plotting capabilities available in
44
45
355 open-source software (here Python). Figure 4 shows a Gantt chart of the production cam-
46
47
356 paign with the units from the process owsheet (gure 3) on the y-axis over date-time on
48
49
357 the x-axis. The colour code allows facile visual tracking of a batch through the system, fur-
50
51
358 thermore the scaled time-series variable that best characterises the respective process steps
52
53
359 on the machines is displayed. The boxes on U1 have been omitted to show the discontinuous
54
55
360 nature of this ow unit which may make reliable state identication dicult. As the infor-
56
57
58
59 19
1
2
3
361 mation about unit-activity is implicitly dened both by the level evolution of upstream and
4
5
362 downstream tanks, this induces no loss in informative value. Furthermore, a second indicator
6
7
363 (the level in the wastewater treatment plant) has been added as a dotted line to units T21
8
9 364 and T22. High material inventory in the water treatment plant can delay a step transition
10
11 365 on T21 and T22. However, as the water treatment plant as a resource is shared by several
12
13 366 plants, there is substantial noise and no unambiguous classication of delays was possible.
14
15 367 Visual analysis of some operational patterns with engineers and operators was found to be
16
17 insightful, but is not to be discussed in greater detail in this work.
18
19
20
U1 T12 T11 FE03 FE02 FE01
21
22
23
24
25
26
27
U3 T32 T31 U22 U21 T22 T21
28
29
Unit Operation
30
31
32
33
34
U4 T42 T41
35
36
37
38
T5
39
U5
40
2 1
T6 T6
41
03-08 06-08 09-08 12-08 15-08 18-08 21-08 24-08 27-08 30-08 02-09 05-09 08-09 11-09 14-09 17-09 20-09 23-09 26-09 29-09 02-10 05-10 08-10 11-10 14-10 17-10 20-10 23-10
42 Datetime
43
44 Figure 4: Reconstructed Gantt chart of an entire production campaign with unique colour-
45 code for batches and scaled time-series data (usually ows or volumes) as indicator of process
46 step. Cleaning-in-Place with black underlay for units T31/T32/U3. Last batch of campaign
47 has been excluded.
48
49 368
50
51
52
53
54
55
56
57
58
59 20
1
2
3
4
369 5.2 Bottleneck Identication
5
6 370 By mere visual assessment of gure 4, the density of the colour-bars on units T21/T22 and
7
8 371 U21/U22 indicates a bottleneck. Especially downstream of these units, long idle periods
9
10 372 indicate free capacities. Upstream, the fermenters may exhibit long cycle times. However,
11
12 373 this front-end process is easily plannable and long cycle times occur as material is prepared in
13
14 374 advance to be ready for downstream processing. In the end this is possible due to overcapacity
15
16 375 also on the fermenters.
17
18
19
20
376 5.2.1 Bottleneck Identication (Frequentist Statistics)
21
22 377 There are multiple ways to assess capacity at current planning and operating skill. Firstly,
23
24 378 it is possible to draw an average over the entire campaign (76.5 days) in which 69 batches
25
26 379 have passed through the largest part of the downstream line which gives a capacity of 0.9
27
28 380 batches per day.
29
30 381 Bar-plots showing the eective (normalised w.r.t. number of units) cycle times are a common
31
32 382 visual tool for comparing equipment capacities. Figure 5 indicates a strongly leveraged
33
34 383 process with a bottleneck around T21/T22 and U21/U22. Note that here waiting time
35
36 384 between batches is not counted, and exclusively processing or waiting with material-in-the-
37
38 385 loop is tracked. An equivalent visualisation is shown in gure 6, however expressed in batches
39
40 386 per day which was preferred by the industry stakeholders due to its more expressive nature.
41
42 387 The average stage cycle time (including the time units wait for material from upstream)
43
44 388 amounts to 53.3 hours. Converting to the eective average stage cycle time, a capacity
45
46 389 estimate based on the bottleneck can be made and adds up to
47
48
49
50 C = (24h/d)/(53.3h/2b) = 0.9b/d , (5)
51
52
53 390 b denoting batches. This is a good indication that the Frequentist analysis is meaningful.
54
55 391 A reduction of waiting for upstream material on the bottlenecking units would increase this
56
57
58
59 21
1
2
3
4
5
6
7
8
9 25
10
11
12 20
13
Time (h)
14
15
15
16
17 10
18
19
20
5
21
22 0
23 3xFE 2xT1 U1 2xT2 2xU2 2xT3 2xT4 T5 2xT6
24
25 Figure 5:
Unit Operations
Eective cycle times (normalised by number of installed machines), excluding
26
periods of waiting for material-transfer from upstream.
27
28
29
30
31
32
33
34
35
36
37 8
38
Batches / day (b/d)
39
40 6
41
42
43
44
4
45
46
47 2
48
49
50 0
51 3xFE 2xT1 U1 2xT2 2xU2 2xT3 2xT4 T5 2xT6
52
53 Unit Operations
Figure 6: Eective capacity expressed in batches per day.
54
55
56
57
58
59 22
1
2
3
392 to 1 batch per day. However, this requires substantial operating discipline and is thus both a
4
5
393 dicult and risky project. Again, it is not possible to dierentiate waiting periods further as
6
7
394 there are many process steps with naturally varying durations due to changing slurry/uid
8
9 395 properties, and the re-constructed machine steps do not contain enough information to un-
10
11 396 ambiguously characterise the waiting periods. It is known that CIP-procedures throw the
12
13 397 downstream-line o of the planned fermentation schedule on a regular basis. Careful exami-
14
15 398 nation of the Gantt chart reveals that CIPs of units T31/T32 as well as U3 prohibit passing
16
17 399 material on from U21/U22. This delays the already long processing step and furthermore
18
19 400 propagates upstream to units T21/T22, which have previously also been identied to have
20
21 401 naturally long cycle times. Figure 7 shows the results of an algorithm that classies cycles on
22
23 402 units U21/U22 into CIP-aected and non-CIP-aected (and post-CIP) populations. Despite
24
25 403 of large variability, CIP-aected cycles can be clearly distinguished and are substantially
26
27 delayed, the cycle immediately after a CIP is already unaected by this.
28
29
30
31
32
6 CIP-affected
33
Indirect CIP affected
Frequency
34
35
36 4 non-CIP-affected
37
38
39 2
40
41
42 0
43 0 5 10 15 20 25 30 35 40 45 50 55 60 65 70
44
45 Hours
46
47
48 Figure 7: Eect of Cleaning-in-Place on cycle durations. Indirectly aected denotes the cycle
49 after the CIP-aected cycle has been completed.
50
404
51
52
53
54
55
56
57
58
59 23
1
2
3
4
405 5.2.2 Bottlenecks in a Transient Scenario
5
6 406 Due to the inability to dierentiate inter-unit waiting and include this into the bar plot
7
8 407 (gures 5 and 6), the scenario needs to be assessed in the transient regime. The time-
9
10 408 evolutions of the machine capacities are overlaid in gure 8. (Here fermentation has been
11
12 409 omitted as it is known that ample capacity is available.) Units T21/T22 and U21/U22 exhibit
13
14 410 the lowest capacities over the entire campaign. It is known that they are strongly correlated,
15
16 411 furthermore it appears that both can be rate-limiting at dierent times. Therefore, in the
17
18 412 following a selection of possible debottlenecking projects is identied from process knowledge
19
20 substantiated with the statistical analysis.
21
22
12 2xT1
23 U1
24 10 2xT2
25 2xU2
2xT3
Batches / day
26 8
27 2xT4
28 6 T5
29 2xT6
30 4
31
2
32
33
34 2018-08-14 2018-08-28 2018-09-11 2018-09-25 2018-10-09 2018-10-23
35
Date-time
36 Figure 8: Eective capacities over time for all units but fermentation. Proximity in the
37
process and similar capacities indicate a shared bottleneck on units T21/T22 and U21/U22;
38
all other cycle times lie far below these two throughout the campaign.
39
40 413
41
42
43
44 414 5.3 Retrot Engineering Projects
45
46
415 It is not a possibility to increase batch size further (avert 'size bottlenecks'). Fermenter
47
48
416 volume has been enhanced previously by increasing the height of the tanks, but another
49
50
417 increase at this point is not realistic. However, a project aimed at substantially increasing the
51
52
418 product concentration in the slurry by adjusting feed composition as well as feeding strategy
53
54
419 is ongoing. This constitutes debottlenecking on the biochemical level, which suggests an
55
56
420 extension of the classication size/scheduling bottleneck by a quality bottleneck, as it not
57
58
59 24
1
2
3
421 evident that all downstream unit operations can handle this increase (thinking for instance
4
5 44
422 about resin capacity in chromatographic units ).
6
7
423 Having identied the scheduling bottleneck consisting of an interlinkage of units T21/T22
8
9 424 and U21/U22, it is time to capitalise on the process knowledge of the experienced engineers
10
11 425 for the identication of possible engineering projects.
12
13 426 Firstly, it is desirable to reduce the eect of downstream CIPs. These not only delay the
14
15 427 bottleneck, but also induce substantial operational complexity due to the irregularity arising
16
17 428 in the schedule. Currently, CIPs of U3 block both tanks T31/T32 (gure 9) and therefore also
18
19 the transmission of material from bottleneck U21/U22 for ca. 12 hours. A reconguration of
20
21
22
23
24
25
26
27
1
28
U2
29
30
Unit Operation
31
2
32
U2
33
34
35
1
T3
36
37
38
39
2
T3
40
41
42
43
U3
44
45 01-10 04-10
46
47
Datetime
48 Figure 9: Strongly delayed material transfer from bottleneck (U21, olive-green batch) due
49 to T31/T32 blockage by U3 CIP.
50
429
51
52 430 the CIP piping allows running a CIP on unit U3 with only one of the tanks (there will still
53
54 431 be two shorter tank CIPs). This implies that the other tank is free to receive the upstream
55
56 432 material-in-waiting, freeing up roughly 12 hours of bottleneck operation once every 6-8 days
57
58
59 25
1
2
3
433 as dictated by the CIP schedule. This is expected to yield a capacity increase of 0.5 batches
4
5
434 per week for a comparably simple Type-2 project (3.1.3). Again, it must be pointed out that
6
7
435 this reduction in delay is likely to reduce the propagation of delays into the upstream process,
8
9 436 reducing operational complexity for operators due to increased regularity, facilitating future
10
11 437 optimisation projects.
12
13 438 Secondly, aside from these irregular delays, it is desirable to elevate the baseline capacity on
14
15 439 the bottlenecking units. Figure 10 shows some rened machine step statistics (anonymized).
16
17 440 Steps 4-6 denote the transfer of material from T21/T22 to U21/U22 and therefore take up
18
19 441 an equally long amount of processing time on both units. The distinction is drawn as step 4
20
21 442 is largely automatic, whereas step 5 requires substantial manual operation. Delays here arise
22
23 443 especially due to challenging uid properties as a result of variability in the fermentations.
24
25 444 In the case of highly viscous slurries, the transfer of material requires a lot of manual re-
26
27 445 tuning of set points. This can lead to delays if the operator is not disposable and can create
28
29 446 problems in other parts of the process as the operator needs to keep an eye on the procedure.
30
31 447 Installation of a new pumping system with a higher degree of automation as a Type-4 project
32
33 448 is likely to shorten this step signicantly. Furthermore, variability of the step duration will
34
35 449 be reduced and the operator be given more capacities to engage in other parts of the process.
36
37 450 Roughly, each 5 hours of shortening the procedure will increase capacity by 0.7 batches per
38
39 451 week. It is not clear exactly how much time will be saved, but as indicated by gure 11 there
40
41 452 is high potential due to extreme delays. (The large tails in steps 2-4 in gure 10 are i.e. due
42
43 453 to CIPs.)
44
45
46
47
48 454 6 Discussion
49
50
51 455 Firstly, the case study has conrmed that it is not always possible or justiable to apply a for-
52
53 456 mal model-based systems engineering approach to optimisation of industrial processes. Lim-
54
55 457 itations in model-building and verication, as well as uncertainty regarding implementability
56
57
58
59 26
1
2
3
4
5
Frequency Total duration: Step 1 until Step 6
6
7
8
25 T21
9 T22
10
11 0
12 0 8 16 24 32 40 48 56 64 72
13
14 Step 1 until Step 2
Frequency
15
16
17
25
18
19
20 0
21
22
0 8 16 24 32 40 48 56 64 72
23
24
Step 2 until Step 3
Frequency
25
26 25
27
28
29
0
30
31 0 8 16 24 32 40 48 56 64 72
32
33 Step 3 until Step 4
Frequency
34
35
36
25
37
38
39 0
40 0 8 16 24 32 40 48 56 64 72
Step 4 until Step 5
41
42
43
Frequency
44
45 25
46
47
48 0
49
50 0 8 16 24 32 40 48 56 64 72
51
52
Step 5 until Step 6
Frequency
53
54
55
25
56
57
58 0
59
60
0 8 16 ACS24Paragon32Plus Environment
27 40 48 56 64 72
Figure 10: Granularity of the sub-steps with kernel-smoothed density estimates to facilitate
1
2
3
4
5
6
15
7
8
9
Frequency
10
11
10
12
13
14
15
16
5
17
18
0
19
20
21 5 10 15 20 25 30
22
23 Cycle Duration (h)
24
25 Figure 11: Distribution of cumulative transfer times (steps 4-6).
26
27
28 458 especially of projects that require operational discipline are main inhibitors in this. On the
29
30 459 other hand it is shown that modern data processing environments can be a viable tool in
31
32 460 the incremental optimisation of bio-based processes, especially if the increments have only
33
34 461 moderate impact on operations. Statistical means create a more explicit evidence-base than
35
36 462 mental models. Furthermore, it has been shown that it is possible to reconstruct necessary
37
38 463 cycle time information even in plants without a rigorous monitoring system, as is the case in
39
40 464 many semi-automated facilities due to the diculties that arise in monitoring manual pro-
41
42 465 cess steps. The proposed algorithm is closely linked to process and plant structures, which
43
44 466 renders it intuitive and easy to maintain. Furthermore, it allows identifying batches that did
45
46 467 not follow a standard pattern with high certainty.
47
48 468 In the regarded scenario, the recursive state machine functioned eectively despite of the
49
50 469 uncertain process regime as many input sequences themselves were very reproducible (due
51
52 470 to being automated or standardised operating procedures). Augmented by the time-series
53
54 471 data from common sensors and domain knowledge, this enabled precise reconstruction of the
55
56 472 schedule - it was however connected with a substantial programming eort. Accelerating this
57
58
59 28
1
2
3
473 development process should greatly contribute to usability of the proposed methodology. The
4
5
474 same applies to an extension of the method which better accommodates or even automat-
6
7 58
475 ically handles irregular sequences or time-series proles. Bayesian changepoint detection
8
9 476 seems to oer a theoretic foundation which can help with both ambitions. Here, thought
10
11 477 should be given to balancing programming eort with the workload which may arise in pos-
12
13 478 terior validation of the algorithm, which expectedly grows with its exibility/independence.
14
15 479 Utilising a exible algorithm for detailed analyses within the superstates (which can be re-
16
17 480 liably generated by a robust embedding algorithm such as the one presented in this work)
18
19 481 might constitute a good trade-o.
20
21 482 In terms of process optimisation, Amaran et al.'s model-based approach is justied by set-
22
23 483 ting an ambitious future production target. Due to the implied complexity, for instance
24
25 484 de-bottlenecking of multiple stages at once calls for a notable level of formality. Still, small
26
27 485 projects along the way must not be neglected, especially if they pave the way for these large
28
29 486 projects by facilitating or enabling modelling, validation and implementation.
30
31 487 In these small projects, a model-based approach may not always be the correct choice. Again,
32
33 488 besides unknowns on the physiochemical level, modelling may be dicult in the bio-based
34
35 489 industries as
36
37
38 490 Machine steps are not tracked in semi-automated facilities (incomplete process moni-
39
40 491 toring)
41
42
43 492 Production processes are unsteady/irregular which makes validation of the models
44
45 493 complicated and costly or even impossible
46
47
48 494 Model-uncertainties and complex manual operations can create substantial uncertainty
49
50 495 about the goodness of the results after implementation
51
52
496 On the other hand, the amount of products and number of parallel trains are usually quite
53
54
497 manageable.
55
56
498 It could be argued that Amaran et al. presuppose that this type of statistical analysis is a
57
58
59 29
1
2
3
499 standard operation, but especially in smaller-scale and semi-automated facilities this is likely
4
5
500 not so. Furthermore, in industry especially Type-1 and Type-2 problems will often not jus-
6
7
501 tify the dedicated development and validation of models, as it may be both too expensive
8
9 502 and too time-consuming, calling for simpler approaches. The situation may be dierent if a
10
11 503 digital twin of the plant is available and rigorously maintained. On the other hand, this is
12
13 504 still an emerging technology and the economics behind it are in need of further illumination.
14
15 505 In general, the value of creating a database as an enabler of rigorous process systems engi-
16
17 506 neering should be acknowledge by industrial stakeholders. This calls for investments in new
18
19 507 data acquisition and better data contextualisation, but also in integrated manufacturing ex-
20
21 508 ecution & control systems, which facilitate the often cumbersome data-collection process.
59
22
23 509 In the above analysis, no performance indicators beyond current practices
8,23,41
are intro-
24
25 510 duced. However, a statistical analysis should have its place in a framework directed at
26
27 511 especially bio-based batch plant debottlenecking. Modern data processing environments
28
29 512 such as Python or R are attractive (cheap, fast, advanced, exible) as they enable highly ef-
30
31 513 cient analyses as well as visualisations. Thus, the case study should hopefully be of notable
32
33 514 practical interest to industrial readers.
34
35
36
37
38 515 7 Conclusion
39
40
41 516 Batch processing is a relevant production paradigm especially in the bio-based industries,
42
43 517 and a number of generalisable operational challenges have been shown at hand of an indus-
44
45 518 trial case study. While several methodologies for debottlenecking of batch process plants
46
47 519 exist in literature, it is likely that applicability in large parts of the bio-based industries is
48
49 520 limited. The reasons for that lie in the irregularities that arise as a cause of manual process
50
51 521 control, but also from missing monitoring systems (both of biochemical properties as well
52
53 522 as machine states). An algorithm has been introduced that is capable of reconstructing
54
55 523 the machine states necessary to identify and avert scheduling bottlenecks. Based on visual
56
57
58
59 30
1
2
3
524 statistical and heuristics, a powerful combination that maybe deserves a mentioning in a
4
5
525 exible framework, promising candidate debottlenecking projects have been identied in the
6
7
526 case study plant. The chosen projects are robust and have a positive eect on reproducibility
8
9 527 (process stability), and thus bring the plant closer to a state that allows rigorous model-based
10
11 528 optimisation. Transforming industrial process plants into predictable entities should be a
12
13 529 general ambition both for academic and industrial stakeholders in order to reap the benets
14
15 530 of modern systems engineering tools. While it is a shared endeavour, it seems likely that the
16
17 531 academy has to play an orchestrating role in this.
18
19
20
21
22 532 Acknowledgements
23
24
25 533 The project received nancial support from Innovation Fund Denmark.
26
27
28
29
30
534 References
31
32 535 (1) FDA, Facts About the Current Good Manufacturing Practices (CGMPs). 2015.
33
34
35 536 (2) Meyer, H. P.; Minas, W.; Schmidhalter, D. In Industrial Biotechnology: Products and
36
37 537 Processes ; Wittman, C., Liao, J. C., Eds.; Wiley-VCH, 2016.
38
39
40 538 (3) Birg, L. Pharmaceutical Regulation, Mandatory Substitution, and Generic Competi-
41
42 539 tion. SSRN Electronic Journal 2015, 241 .
43
44
45 540 (4) Gernaey, K. V.; Cervera-padrell, A. E.; Woodley, J. M. A perspective on PSE in phar-
46
47 541 maceutical process development and innovation. Computers & Chemical Engineering
48
49 542 2012, 42, 1529.
50
51
543 (5) Marques, C. M.; Moniz, S.; de Sousa, J. P. Strategic decision-making in the phar-
52
53
54 544 maceutical industry: A unied decision-making framework. Computers and Chemical
55
56 545 Engineering 2018, 119, 171189.
57
58
59 31
1
2
3
546 (6) Litzen, D. B.; Bravo, J. L. Uncover Low-Cost Debottlenecking Opportunities. Chemical
4
5
547 Engineering Progress 1999, March Ed.
6
7
8 548 (7) Baldea, M. From process integration to process intensication. Computers and Chemical
9
10 549 Engineering 2015, 81, 104114.
11
12
13 550 (8) Amaran, S. .; Sharda, B.; Bury, S. J. Targeted Incremental Debottlenecking of Batch
14
15 551 Process Plants. Proceedings of the 2016 Winter Simulation Conference. 2016; pp 2924
16
17 552 2934.
18
19
20 553 (9) Petrides, D.; Carmichael, D.; Siletti, C.; Koulouris, A. Biopharmaceutical Process Op-
21
22 554 timization with Simulation and Scheduling Tools. Bioengineering 2014, 1, 154187.
23
24
25 555 (10) Wu, H.; Khan, M.; Hussain, A. Process control perspective for process analytical tech-
26
27 556 nology: Integration of chemical engineering practice into semiconductor and pharma-
28
29 557 ceutical industries. Chemical Engineering Communications 2007, 194, 760779.
30
31
558 (11) Klatt, K. U.; Marquardt, W. Perspectives for process systems engineering-Personal
32
33
34
559 views from academia and industry. Computers and Chemical Engineering 2009, 33,
35
560 536550.
36
37
38 561 (12) Alford, J. S. Bioprocess control: Advances and challenges. Computers and Chemical
39
40 562 Engineering 2006, 30, 14641475.
41
42
43 563 (13) Calabrese, G. S.; Pissavini, S. From Batch to Continuous Flow Processing in Chemicals
44
45 564 Manufacturing. AIChE Journal 2011, 57 .
46
47
48 565 (14) Bauer, M.; Craig, I. K. Economic assessment of advanced process control - A survey
49
50 566 and framework. Journal of Process Control 2008, 18, 218.
51
52
53 567 (15) Williams, T. J. In Springer Handbook of Automation ; Nof, S. Y., Ed.; Springer Berlin
54
55 568 Heidelberg: Berlin, Heidelberg, 2009; pp 511.
56
57
58
59 32
1
2
3
569 (16) Lee, C. K.; Khoo, H. H.; Tan, R. B. Life Cyle Assessment Based Environmen-
4
5
570 tal Performance Comparison of Batch and Continuous Processing: A Case of 4-D-
6
7
571 Erythronolactone Synthesis. Organic Process Research and Development 2016, 20,
8
9 572 19371948.
10
11
12 573 (17) Vervaet, C.; Remon, J. P. Continuous granulation in the pharmaceutical industry.
13
14 574 Chemical Engineering Science 2005, 60, 39493957.
15
16
17 575 (18) Gutmann, B.; Cantillo, D.; Kappe, C. O. Continuous-ow technology - A tool for
18
19 576 the safe manufacturing of active pharmaceutical ingredients. Angewandte Chemie -
20
21 577 International Edition 2015, 54, 66886728.
22
23
578 (19) Costandy, J. G.; Edgar, T. F.; Baldea, M. Switching from Batch to Continuous Reactors
24
25
26
579 Is a Trajectory Optimization Problem. Industrial & Engineering Chemistry Research
27
28
580 2019, 58, 1371813736.
29
30 581 (20) Teoh, S. K.; Rathi, C.; Sharratt, P. Practical Assessment Methodology for Converting
31
32 582 Fine Chemicals Processes from Batch to Continuous. Organic Process Research and
33
34 583 Development 2016, 20, 414431.
35
36
37 584 (21) Chen, S.; Koirala, Y.; Mannan, M. S. Comparison of batch versus continuous process
38
39 585 in pharmaceutical industry based on safety consideration. 32nd Center for Chemical
40
41 586 Process Safety International Conference 2017, CCPS 2017 - Topical Conference at the
42
43 587 2017 AIChE Spring Meeting and 13th Global Congress on Process Safety. 2017.
44
45
46 588 (22) Fernández, I.; Renedo, C. J.; Pérez, S. F.; Ortiz, A.; Mañana, M. A review: Energy
47
48 589 recovery in batch processes. Renewable and Sustainable Energy Reviews 2012, 16, 2260
49
50 590 2277.
51
52
591 (23) Simon, L. L.; Osterwalder, N.; Fischer, U.; Hungerbühler, K. Systematic retrot method
53
54
592 for chemical batch processes using indicators, heuristics, and process models. Industrial
55
56
593 and Engineering Chemistry Research 2008, 47, 6680.
57
58
59 33
1
2
3
594 (24) Edgar, T. F. Control and operations: When does controllability equal protability?
4
5
595 Computers and Chemical Engineering 2004, 29, 4149.
6
7
8 596 (25) Kano, M.; Ogawa, M. The state of the art in chemical process control in Japan: Good
9
10 597 practice and questionnaire survey. Journal of Process Control 2010, 20, 969982.
11
12
13 598 (26) Roberge, D. M.; Ducry, L.; Bieler, N.; Cretton, P.; Zimmermann, B. Microreactor
14
15 599 technology: A revolution for the ne chemical and pharmaceutical industries? Chemical
16
17 600 Engineering and Technology 2005, 28, 318323.
18
19
601 (27) Hollis, A. The importance of being rst: Evidence from Canadian generic pharmaceu-
20
21
22
602 ticals. Health Economics 2002, 11, 72334.
23
24 603 (28) Zydney, A. L. Continuous downstream processing for high value biological products: A
25
26 604 Review. Biotechnology and Bioengineering 2016, 113, 465475.
27
28
29 605 (29) Lee, S. L.; O'Connor, T. F.; Yang, X.; Cruz, C. N.; Chatterjee, S.; Madurawe, R. D.;
30
31 606 Moore, C. M.; Yu, L. X.; Woodcock, J. Modernizing Pharmaceutical Manufacturing:
32
33 607 from Batch to Continuous Production. Journal of Pharmaceutical Innovation 2015,
34
35 608 10, 191199.
36
37
38 609 (30) Jungbauer, A. Continuous downstream processing of biopharmaceuticals. Trends in
39
40 610 Biotechnology 2013, 31, 479492.
41
42 611 (31) Chatterjee, S. FDA Perspective on Continuous Manufacturing. IFPAC Annual Meeting.
43
44 612 Baltimore, MD, 2012.
45
46
47 613 (32) FDA, Code of Federal Regulations Title 21. 2017; https://www.accessdata.fda.gov/
48
49 614 scripts/cdrh/cfdocs/cfcfr/CFRSearch.cfm?CFRPart=211.
50
51
52 615 (33) Doran, P. M. Bioprocess Development. Bioprocess Engineering Principles 2013, 311.
53
54
616 (34) Schaber, S. D.; Gerogiorgis, D. I.; Ramachandran, R.; Evans, J. M.; Barton, P. I.;
55
56
617 Trout, B. L. Economic analysis of integrated continuous and batch pharmaceutical
57
58
59 34
1
2
3
618 manufacturing: A case study. Industrial and Engineering Chemistry Research 2011,
4
5
619 50, 1008310092.
6
7
8 620 (35) Poechlauer, P.; Manley, J.; Broxterman, R.; Gregertsen, B.; Ridemark, M. Continuous
9
10 621 processing in the manufacture of active pharmaceutical ingredients and nished dosage
11
12 622 forms: An industry perspective. Organic Process Research and Development 2012, 16,
13
14 623 15861590.
15
16
17 624 (36) Croughan, M. S.; Konstantinov, K. B.; Cooney, C. The future of industrial bioprocess-
18
19 625 ing: Batch or continuous? Biotechnology and Bioengineering 2015, 112, 648651.
20
21
22 626 (37) Mandenius, C.-F. Recent developments in the monitoring, modeling and control of
23
24 627 biological production systems. Bioprocess and biosystems engineering 2004, 26, 347
25
26 628 351.
27
28
29 629 (38) Mannan, S. Lees' Loss Prevention in the Process Industries: Hazard Identication,
30
31 630 Assessment and Control: Fourth Edition ; 2012.
32
33
34
631 (39) Downs, J. J. In Plantwide Control: Recent Developments and Applications ; Ranga-
35
632 iah, G. P., Ed.; 2012.
36
37
38 633 (40) Bauer, M.; Horch, A.; Xie, L.; Jelali, M.; Thornhill, N. The current state of control
39
40 634 loop performance monitoring - A survey of application in industry. Journal of Process
41
42 635 Control 2016, 38, 110.
43
44
45 636 (41) Calandranis, J.; Petrides, D. P. Analysis and debottlenecking chemical processes of
46
47 637 integrated batch. Computers & Chemical Engineering 2000, 24, 1387 1394.
48
49
50 638 (42) Petrides, D.; Koulouris, A.; Siletti, C. Throughput analysis and debottlenecking of
51
52 639 biomanufacturing facilities. Chimica oggi 2002, 5, 2228.
53
54
55
56
57
58
59 35
1
2
3
640 (43) Jully, T.; Foo, D.; Yee, C.; Kumaresan, S.; Aziz, R. A. Computer-aided process De-
4
5
641 sign Tools for Debottlenecking a Batch Pharmaceutical Production. The Institution of
6
7
642 Engineers 2007, 68, 3037.
8
9
10 643 (44) Yang, Y.; Farid, S. S.; Thornhill, N. F. Data mining for rapid prediction of facility
11
12 644 t and debottlenecking of biomanufacturing facilities. Journal of Biotechnology 2014,
13
14 645 179, 1725.
15
16
17 646 (45) Tan, J.; Foo, D. C. Y.; Kumaresan, S.; Aziz, R. A. Debottlenecking of a batch phar-
18
19 647 maceutical cream production. Pharmaceutical Engineering 2006, 26 .
20
21
22 648 (46) Sharda, B.; Bury, S. J. Bottleneck analysis of a chemical plant using discrete event
23
24 649 simulation. Proceedings - Winter Simulation Conference. 2010.
25
26
27 650 (47) Monroy, D. F. Z.; Vallejo, C. C. R. Use Cases of Discrete Event Simulation: Appliance
28
29 651 and Research ; 2013.
30
31
652 (48) Schmidtke, D. C. Value Stream Mapping and Discrete Event Simulation for Optimiza-
32
33
653 tion of Chemical Production Processes. Ph.D. thesis, Fakultät für Chemie der Technis-
34
35
654 chen Universität München, 2015.
36
37
38 655 (49) Alshekhli, O.; Foo, D. C.; Hii, C. L.; Law, C. L. Process simulation and debottlenecking
39
40 656 for an industrial cocoa manufacturing process. Food and Bioproducts Processing 2010,
41
42 657 89, 528536.
43
44
45 658 (50) Tan, M. M.; Andiappan, V.; Wan, Y. Process Debottlenecking and Retrot of Palm Oil
46
47 659 Milling Process via Inoperability Input-Output Modelling. MATEC Web of Conferences
48
49 660 2018, 152 .
50
51
52 661 (51) Guimarães, A. M. C.; Leal, J. E.; Mendes, P. Discrete-event simulation software se-
53
54 662 lection for manufacturing based on the maturity model. Computers in Industry 2018,
55
56 663 103, 1427.
57
58
59 36
1
2
3
664 (52) Pennsylvania, P. M. I. N. S. Organizational Project Management Maturity Model
4
5
665 (OPM3) : Knowledge Foundation. 2003.
6
7
8 666 (53) Goldratt, E. M.; Cox, J. The Goal: A Process of Ongoing Improvement ; 2004.
9
10
11 667 (54) Pegels, C. C.; Watrous, C. Application of the theory of constraints to a bottleneck
12
13 668 operation in a manufacturing plant. Journal of Manufacturing Technology Management
14
15 669 2005, 16, 30211.
16
17
18 670 (55) Barrera, M. D. Optimal design and operation of batch processes. Ph.D. thesis, Mas-
19
20 671 sachusetts Institute of Technology, 1989.
21
22
23 672 (56) Spann, R.; Eliasson Lantz, A.; Gernaey, K. V.; Sin, G. Reference Module in Chemistry,
24
25 673 Molecular Sciences and Chemical Engineering ; Elsevier Inc., 2018; pp 117.
26
27
674 (57) Bähner, F. D.; Santacoloma, P. A.; Huusom, J. K. Optimal operation of parallel dead-
28
29
30
675 end lters in a continuous bio-based process. Food and Bioproducts Processing 2019,
31
32
676 114, 26775.
33
34 677 (58) Adams, R. P.; MacKay, D. J. C. Bayesian Online Changepoint Detection. 2007; http:
35
36 678 //arxiv.org/abs/0710.3742.
37
38
39 679 (59) De Carlo, F.; Arleo, M. A.; Tucci, M. OEE evaluation of a paced assembly line through
40
41 680 dierent calculation and simulation methods: A case study in the pharmaceutical en-
42
43 681 vironment. International Journal of Engineering Business Management 2014, 6, 111.
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59 37
1
2
3
4
682 Graphical TOC Entry
5
6
7
8
9
10 683
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59 38

A Debottlenecking Study of An Industrial

Uploaded by

Copyright:

Available Formats

A Debottlenecking Study of An Industrial

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

A Debottlenecking Study of An Industrial

Uploaded by

Copyright:

Available Formats

Downloaded from orbit.dtu.

dk on: Nov 01, 2024

A Debottlenecking Study of an Industrial Pharmaceutical Batch Plant

Baehner, Franz D.; Huusom, Jakob K.

Link to article, DOI:

Link back to DTU Orbit

Process Systems Engineering

is published by the American Chemical Society. 1155 Sixteenth Street N.W.,

A Debottlenecking Study of an Industrial

19 11: else if SF E03 is ... then

43 31: Store time-points

18 11: else if tdrain = any(TF ill,T 12 ) then

You might also like