Public.

Pollution in manufacturing:
An unavoidable incidence?
A simulation study into cross-contamination effects in a multi-product factory
Delft University of Technology
Paweł Tomasz Kołodziejczyk
P OLLUTION IN MANUFACTURING : A N
UNAVOIDABLE INCIDENCE ?
A SIMULATION STUDY INTO CROSS - CONTAMINATION EFFECTS
IN A MULTI - PRODUCT FACTORY
by
Paweł Tomasz Kołodziejczyk
in partial fulfillment of the requirements for the degree of
Master of Science
in Management of Technology
at the Delft University of Technology,
to be defended publicly on Friday October 30, 2015 at 11:00 AM.
Student number:
4325273
Thesis committee:
Prof. dr. ir. A. Verbraeck
Dr. ir. R. M. Stikkelman
Dr. L. Rook
Ir. R. Bijlsma
Ir. H. Landeweerd
Ing. B. Lekkerkerker
External supervisors:
TU Delft
TU Delft
TU Delft
Systems Navigator
Nutreco
Sloten
An electronic version of this thesis is available at http://repository.tudelft.nl/.
P REFACE
This report concludes my thesis project of the Master in Management of Technology as well as my education
at Delft University of Technology in the Netherlands. It has been fun and challenging time abroad, and during
these just over two years I have evolved considerably. It is high time now to part ways with the academia and
set out to the ‘real world’ taking all the learned lessons with me. This thesis is the prime example of the
abilities I obtained.
I would like to thank all those who contributed to the final version of this thesis in its current form. I
am grateful to have had the opportunity of working closely with not just one, but three companies, solving a
real-world problem, and based on that writing a thesis to graduate at a university. Despite many frustrations
coming from various conflicts of interests among the involved parties as well as my own affinities, I am glad
for the experience that will definitely shape my future. Now, with a relief that is it all over, I would like to
express my gratitude towards those, who especially assisted in making it happen.
Eventually, I conclude this foreword with a remark on the process. In theory, writing a thesis should start
with a research question, and all needs and requirements ought to naturally follow from there and create a
complementary whole. Although this is hardly ever done so, just as in my case, there is little room left to
enjoy the writing itself and the freedom of expression that comes with it. Yet, I did my best to rather follow
the approach of one of my favourite novelists:
I chart a little first-list of names, rough synopsis of chapters, and so
on. But one daren’t overplan; so many things are generated by the
sheer act of writing.
— Anthony Burgess
In the end, this sizeable text is a verbalisation of my understanding of the significance of the performed work
as well as the storytelling skills that I have. I only wish I could make it more like a work of fiction to ease the
read but apparently this would disqualify me in the field of engineering where everything must follow a clear
path, and where sudden twists in the plot are certainly not appreciated. Perhaps some other time. . .
Paweł Tomasz Kołodziejczyk
Delft, October 2015
iii
S UMMARY
Companies can lose a substantial portion of their revenues due to deterioration in value of their products as
a result of processing impediments, which might occur in the entire supply chain. Sometimes, the reasons
for a decline in product quality are not fully understood, and thus there are no appropriate measures taken to
prevent it. Yet, it is vital to comprehend what is actually happening during production and material handling
so that proper, well-informed decisions can be made. This project deals with one of the unexplored problems – quality loss due to product cross-contamination, which is a common occurrence in material handling,
multi-product plants. It is a process of mixing product going through multi-purpose equipment and piping
with leftover residue that is currently there, which takes place when there is no intermediate cleaning in between two consecutive products, though the precise character of it is usually not known. Omitting cleaning
runs to save production time and money is a common practice in the industry.
This research introduces a novel simulation method to quantify and predict the extent of cross-contamination as well as to assess its effects. Many models for material handling in multi-product plant have been
made in the past but very few relate to the issue of cross-contamination, which is of extreme importance
in quality assurance and informed decision-making. As no similar models have been developed previously,
or other research done based on a real industrial setting, it is a truly innovative study, intended to establish
foundation for further research, and to raise awareness of the issue by starting to fill a gap of knowledge about
what can happen during material processing.
Initially, a thorough literature research is done, dealing with issues of trade-offs in manufacturing, crosscontamination in production, as well as how to combine scheduling methodologies with a powder mixing
model in a discrete event simulation (DES). Then modelling boundaries and assumptions are established
together with a conceptual model. Based on that, a general DES building blocks – class models for a bagging
machine, conveyor, mixer, silo as well as product batches and orders are created.
Specific analysis is performed in Sloten, a young animal feed plant in Deventer, the Netherlands, in order
to find a more precise character of cross-contamination over time, as well as to test the model application
in a realistic setting. Contamination measurements using tracer–collector method over material flow with
multiple intermediate sampling points help, together with relevant theoretical models, build mathematical
representation of chaotic changes occurring within material flow. Obtained curves fitted with a sum of two
exponentially decreasing curves correspond well to the measurements but do not explain everything that
happens during the process. Nevertheless they are used to model the cross-contamination phenomenon.
Two unique cross-contamination models suitable for implementation in DES are developed, basing on
principles of segmentation, quantity conservation, product similarity and proportionality, and are fundamentally models of mixing between material flow and residue in the crossed container. More general one,
called partial mass exchange model, allows to customize fraction of material mixed, characteristic to a given
piece of equipment, while the other, mixing model, assumes homogeneous resulting composition.
Finally, a case-specific simulation model, joining material handling system with cross-contamination calculations as well as customizable layout configuration and settable scheduling methodologies, is established.
Experimental setup with four-stage production process consisting of ingredient feeder, a single mixer, limited
intermediate storage buffers and multiple system exits, mostly bagging machines, connected by conveying
equipment, is discussed. All of the machines involved are multi-purpose, flexible but with some limitations,
together with the interconnections and logic among them, basing on a genuine example.
Analysis shows considerable impact of system layout and scheduling rules on product contamination.
In order to limit cross-contamination appropriate product sequencing, minimising the differences between
consecutive products in all stages of the process in needed, while that might impede throughput because of
resulting constraints. There is thus a trade-off relation among various aspects of production efficiency, as well
as amidst that efficiency and flexibility, when number of storage buffers and their capacity is changed.
In the end, research first benefits from measuring the extent of cross-contamination, then succeeding
in giving an original example of how to build a cross-contamination model, and how to combine it with a
stochastic DES, capable of suiting different designs and scheduling methodologies. Performed experiments
show, that cross-contamination effects can be reasonably well simulated with a DES model. It also describes
possible effects of certain interventions on various performance indicators, and demonstrates possible risks
of increasing flexibility in the system. Additional, more thorough trials are needed in the future to improve
the model and generalise it further. Yet, the very important first step to understand the issue has been taken,
which will aid in application to other types of systems as well as help in mitigating product contamination.
v
C ONTENTS
Summary
v
Contents
vi
List of Figures
ix
List of Tables
xi
Abbreviations and Definitions
1 Introduction
1.1 Research Problem . . . . .
1.1.1 Background . . . .
1.1.2 Problem Statement
1.2 Research Objectives . . . .
1.3 Research Questions . . . .
1.4 Methodology . . . . . . .
1.5 Deliverables. . . . . . . .
1.6 Case Company . . . . . .
1.7 Structure . . . . . . . . .
xiv
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
1
1
1
2
3
4
4
5
5
6
2 Literature Analysis
2.1 Production Trade-offs . . . . . . . . . . . .
2.1.1 Plant Layout Design . . . . . . . . .
2.2 Animal Feed Production and Contamination
2.3 Simulation Overview . . . . . . . . . . . .
2.4 Powder Mixing . . . . . . . . . . . . . . .
2.5 Production Scheduling . . . . . . . . . . .
2.6 Literature Relevance . . . . . . . . . . . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
8
8
9
9
10
11
12
13
3 Modelling Specifications and Requirements
3.1 Investigation Boundaries . . . . . . . . . .
3.2 Modelling Choices . . . . . . . . . . . . .
3.3 Recognising Excessive Cross-contamination
3.4 Model Structure . . . . . . . . . . . . . . .
3.4.1 Analysis Criteria . . . . . . . . . . .
3.4.2 Simulation Approach . . . . . . . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
14
14
16
16
17
18
19
4 Production Model
4.1 Purpose . . . . . . . . . . . . . . . .
4.2 Assumptions . . . . . . . . . . . . .
4.3 Model Characteristics . . . . . . . . .
4.3.1 Objects and Sub-models . . . .
4.3.2 Material Flow Logic . . . . . .
4.3.3 Important Features . . . . . .
4.4 Cross-contamination Investigation . .
4.4.1 Modelling Cross-contamination
4.5 Determined Needs . . . . . . . . . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
20
20
20
21
21
22
23
24
24
30
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
5 Experimental Setup
31
5.1 Production Process . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31
5.1.1 Proposed Outline . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 32
vi
C ONTENTS
vii
5.2 Strategic Overview . . . . . . . . . . . . . .
5.3 Modelling Implications . . . . . . . . . . . .
5.3.1 Simulation Model Elements . . . . . .
5.3.2 Stochasticity . . . . . . . . . . . . . .
5.4 Contamination in Sloten . . . . . . . . . . .
5.4.1 Contamination Measurements Results.
5.4.2 Simulated Contamination Curves . . .
5.5 Scheduling . . . . . . . . . . . . . . . . . .
5.5.1 Scheduling Choices . . . . . . . . . .
5.6 Quality of Gathered Data . . . . . . . . . . .
5.7 Verification and Validation . . . . . . . . . .
5.7.1 Conceptual Model Validation . . . . .
5.7.2 Model Verification . . . . . . . . . . .
5.7.3 Simulation Model Validation . . . . . .
5.7.4 Validation Summary . . . . . . . . . .
6 Experimentation
6.1 Scenario Navigator Setup . . . . . . . . .
6.2 Required Warm-up Periods . . . . . . . .
6.3 Base Scenario . . . . . . . . . . . . . . .
6.4 Experiment Design . . . . . . . . . . . .
6.4.1 Initial Sensitivity Analysis . . . . .
6.4.2 Cross-contamination Investigation
6.4.3 Additional Silos Experiments. . . .
6.4.4 Scheduling Rules Experimentation.
6.4.5 Random Component Investigation
6.5 Experimentation Conclusion . . . . . . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
32
34
34
34
35
35
36
38
38
40
41
41
41
41
43
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
45
45
45
46
47
47
49
49
52
54
57
7 Results Analysis
7.1 Results Presentation . . . . . . . . . . . . . . . .
7.2 Initial Sensitivity Analysis Discussion . . . . . . . .
7.3 Impact of Cross-contamination . . . . . . . . . . .
7.3.1 Cross-contamination Model Comparison . .
7.4 Layout Interventions Impact . . . . . . . . . . . .
7.5 Impact of Scheduling Analysis . . . . . . . . . . .
7.6 Random Component Investigation Analysis . . . .
7.6.1 Size of the Random Component Effect . . . .
7.6.2 Flexibility Exploration Analysis. . . . . . . .
7.7 Propagation of Errors Estimation . . . . . . . . . .
7.7.1 Changeover Time Errors . . . . . . . . . . .
7.7.2 Short Stops Inaccuracies . . . . . . . . . . .
7.7.3 Cross-contamination Uncertainty Evaluation
7.8 Variance Reduction . . . . . . . . . . . . . . . . .
7.9 Recommendations . . . . . . . . . . . . . . . . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
58
58
59
59
59
60
61
63
63
64
65
65
65
66
66
67
8 Conclusions
8.1 Discussion . . . . . . . . . . . . .
8.1.1 Research Questions Answers .
8.2 Practical Relevance . . . . . . . . .
8.3 Reflection . . . . . . . . . . . . . .
8.4 Future Research. . . . . . . . . . .
8.4.1 Academic Relevance . . . . .
8.4.2 Practical Applications . . . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
69
69
69
72
72
73
73
74
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
viii
C ONTENTS
Appendices
75
A Model Data
77
B Validation Data
82
C Case study
83
D SN Tool
95
E Model Implementation
102
F Verification & Validation
115
G Experiment Results
118
H Output Analysis
131
References
135
L IST OF F IGURES
1.1 A schematic overview of the research steps . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
1.2 An overview of the document structure overlaid on the research steps . . . . . . . . . . . . . . . .
5
6
3.1 A diagram of a representative process for the chosen problem . . . . . . . . . . . . . . . . . . . .
3.2 A causal diagram of cross-contamination process . . . . . . . . . . . . . . . . . . . . . . . . . . . .
3.3 A simplified class diagram for the model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
15
16
18
4.1 An example resource chart for bagging machines utilisation, grouped by package size . . . . . .
4.2 A comparison of derived cross-contamination models . . . . . . . . . . . . . . . . . . . . . . . . .
23
29
5.1
5.2
5.3
5.4
5.5
A schematic of the investigation boundaries . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Measured contamination lab results for all sampling points . . . . . . . . . . . . . . . . . . . . . .
A comparison between measured, fitted and simulated contamination at the bagging level . . .
Simulated contamination with random component for all measurement points . . . . . . . . . .
A comparison between the average measured (2013) and simulated packaging throughputs in
terms of bags . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
5.6 A comparison between the average measured (2013) and simulated packaging throughputs in
terms of processed quantity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
5.7 An example chart of contamination per manufacturing order . . . . . . . . . . . . . . . . . . . . .
32
36
37
38
6.1 Scheduling parameters chosen for the base case experimentation scenario . . . . . . . . . . . . .
6.2 A box plot presenting average hourly packaging throughput in experiment runs without including cross-contamination calculations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
6.3 A box plot for comparison of orders completed on time . . . . . . . . . . . . . . . . . . . . . . . .
6.4 A box plot for comparison of impact of sequencing on cross-contamination . . . . . . . . . . . .
6.5 Scheduling parameters chosen for silo number and capacity impact investigation . . . . . . . .
6.6 A scatter plot displaying average results for throughput with respect to varying silo capacity and
number . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
6.7 A scatter plot displaying average results for contamination with respect to varying silo capacity
and number . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
6.8 A scatter plot displaying average results for contamination with respect to achieved throughput
for silo parameter analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
6.9 A scatter plot displaying average results for contamination with respect to achieved throughput
for scheduling analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
6.10 A box plot displaying average results for bagging speed per hour for each of the investigated
scenarios . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
46
42
43
44
48
48
50
50
51
52
52
54
54
7.1 All scenarios plot for total contamination as a function of the average throughput . . . . . . . . .
7.2 Scheduling parameters chosen for the base case experimentation scenario . . . . . . . . . . . . .
7.3 A scatter plot displaying average results for contamination with respect to achieved throughput
with error bars for 95% confidence intervals in silo parameter analysis . . . . . . . . . . . . . . .
7.4 A scatter plot displaying average results for contamination with respect to achieved throughput
with error bars for 95% confidence intervals in scheduling analysis . . . . . . . . . . . . . . . . .
7.5 A scatter plot displaying average contamination in polluted orders with respect to the total contamination, including error bars for 95% confidence intervals in scheduling analysis . . . . . . .
7.6 A scatter plot displaying average contamination in polluted orders with respect to the total contamination, including error bars for 95% confidence intervals in random component analysis .
7.7 A scatter plot indicating the average variance for investigated scenarios . . . . . . . . . . . . . . .
58
60
C.1 A schematic of the proposed production organisation, BTH1-3 are bagging lines . . . . . . . . .
84
ix
61
62
63
63
67
x
L IST OF F IGURES
C.2
C.3
C.4
C.5
C.6
A flowchart of the current production planning and scheduling characteristics
A schematic of the investigated route and sampling points . . . . . . . . . . . .
A photograph of a sampling access point for pneumatic conveyor . . . . . . . .
A detailed schematic of proposed production organisation . . . . . . . . . . . .
A diagram of achieved bagging speeds for scenario Scheduling_C . . . . . . . .
D.1 SN input of the main scenario parameters . . . . . . . . . . . . . .
D.2 SN input of the product portfolio . . . . . . . . . . . . . . . . . . . .
D.3 SN input of the recipes for products . . . . . . . . . . . . . . . . . .
D.4 SN input of silo recipe constraints . . . . . . . . . . . . . . . . . . .
D.5 SN input of a manufacturing orders set . . . . . . . . . . . . . . . .
D.6 SN input of the scheduling rules . . . . . . . . . . . . . . . . . . . .
D.7 SN input of the initial conditions . . . . . . . . . . . . . . . . . . . .
D.8 SN run properties screen with possible data collection restrictions
D.9 Main dashboard with scenario high-level results . . . . . . . . . .
D.10 Mixing performance statistics dashboard . . . . . . . . . . . . . . .
D.11 Details of order contamination and statistics . . . . . . . . . . . . .
D.12 Time statistics for the main equipment . . . . . . . . . . . . . . . .
D.13 Silo utilisation statistics dashboard . . . . . . . . . . . . . . . . . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
85
88
89
92
93
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
95
96
96
97
97
98
98
99
99
100
100
101
101
E.1 A flowchart of the order sequence logic . . . . . . . . . . . . . . . . . . . . .
E.2 A flowchart of how sequencing of daily orders is executed . . . . . . . . . .
E.3 A flowchart of how nutrient similarity is defined . . . . . . . . . . . . . . . .
E.4 A flowchart of how final bag contents are gathered . . . . . . . . . . . . . .
E.5 A flowchart of bag contents with nutrient levels comparison . . . . . . . .
E.6 A flowchart of the crucial silo logic . . . . . . . . . . . . . . . . . . . . . . . .
E.7 A flowchart of contamination calculation implementation; Part 1 of 2 . . .
E.8 A flowchart of contamination calculation implementation; Part 2 of 2 . . .
E.9 A flowchart of how possible destinations are assigned to orders . . . . . . .
E.10 An example order silo allocation chart . . . . . . . . . . . . . . . . . . . . . .
E.11 A flowchart of processing logic for mixer after each order dosing finishes .
E.12 A flowchart of how orders to be bagged by BTH3 machine are searched for
E.13 A flowchart of how bagging machine decides to end early for a given day .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
102
103
104
105
106
107
108
109
110
111
112
113
114
F.1
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
Comparison between probability density function of measured and fitted distribution for bagging changeovers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 117
G.1 A box plot for comparison of mixing throughput . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
G.2 A box plot comparing the number of completed orders within the simulation run per scenario .
G.3 A box plot comparing the number of completed orders on time within the simulation run per
scenario . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
G.4 A box plot comparing the average bagging throughput in silo number/size sensitivity analysis .
G.5 A box plot comparing the number of completed orders within the simulation run per scenario .
G.6 A box plot comparing the number of completed orders on time within the simulation run per
scenario . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
G.7 A scatter plot displaying average results of orders completed on time as a function of the total
number of completed orders . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
G.8 A scatter plot displaying average results for contamination with respect to achieved throughput,
for analysis with random component . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
G.9 A scatter plot displaying average number of orders completed on time as a function of total
number of completed orders . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
G.10 A scatter plot displaying average contamination in polluted order against total contamination .
120
124
124
125
126
127
127
130
130
130
L IST OF TABLES
5.1
5.2
5.3
5.4
Short stops and changeovers distributions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Fitted parameters for cross-contamination curves as a function of processed material quantity .
Accuracy of the simulated contamination curves with respect to measured contamination . . .
Statistical t-test comparing measured and simulated throughput values . . . . . . . . . . . . . .
35
36
37
43
6.1
6.2
6.3
6.4
6.5
46
47
47
49
6.6
6.7
6.8
6.9
6.10
6.11
6.12
6.13
6.14
6.15
6.16
Model input parameters set for the base scenario . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Model scheduling parameters set for the base case experimentation scenario . . . . . . . . . . .
Initial experiments inputs without contamination investigation . . . . . . . . . . . . . . . . . . .
Initial cross-contamination experiments inputs . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Altered independent variables for the exploration of the impact of additional silos and their
capacities on production performance . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Scheduling rules chosen for scenarios exploring their impact . . . . . . . . . . . . . . . . . . . . .
Dispatching rules chosen for scenarios exploring impact of scheduling . . . . . . . . . . . . . . .
Random components chosen for analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Experiment results of scenario Random_A . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Experiment results of scenario Random_B . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Experiment results of scenario Mix_B . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Dispatching rules for scenarios Random_H and Random_I with silo reservation investigation .
Scheduling rules for scenarios Random_H and Random_I with silo reservation investigation . .
Chosen combinations of silos reserved for specific recipes in scenario Random_I . . . . . . . . .
Experiment results of scenario Random_H . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Experiment results of scenario Random_I . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
51
53
53
55
55
55
56
56
56
56
56
57
7.1
7.2
7.3
7.4
Statistics on order above limits ratio for different cross-contamination models .
Comparison of impact of silo reservation on the average throughput for 8 silos
Comparison of impact of silo reservation on total contamination . . . . . . . .
Error propagation parameters for short stops . . . . . . . . . . . . . . . . . . . .
.
.
.
.
60
64
64
65
A.1 Short stops data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
A.2 Changeover time statistics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
A.3 Time to repair statistics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
A.4 Contamination results reference points . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
A.5 Parameters of the investigated screw conveyors . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
A.6 Correlation parameters of the investigated screw conveyors . . . . . . . . . . . . . . . . . . . . . .
A.7 Parameters for correlation analysis for all measured equipment (excluding points G and Bag) . .
A.8 Correlation results for all equipment parameters (as in table A.7) . . . . . . . . . . . . . . . . . . .
A.9 Accuracy of the exponential curve fitting expressed with the sum of squared deviations method
A.10 One sample t-test results for the number of late orders for 95% confidence . . . . . . . . . . . . .
A.11 One sample t-test results for average number of processed bags per hour for 95% confidence . .
77
77
77
77
78
79
79
79
79
81
81
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
F.1
F.2
F.3
F.4
List of performed general verification tests . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Accuracy of the simulated contamination curves with respect to the fitted contamination . . .
Goodness of fit (Kolmogorov–Smirnov test) for short stops and changeover time distributions
Statistical data on obtained throughputs in the base case simulation . . . . . . . . . . . . . . .
.
.
.
.
115
117
117
117
G.1
G.2
G.3
G.4
Experiment results of scenario NC_A
Experiment results of scenario NC_B
Experiment results of scenario NC_C
Experiment results of scenario NC_D
.
.
.
.
118
118
118
118
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
xi
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
xii
L IST OF TABLES
G.5 Experiment results of scenario NC_E . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
G.6 Experiment results of scenario NC_F . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
G.7 Experiment results of scenario NC_G . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
G.8 Experiment results of scenario NC_H . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
G.9 Experiment results of scenario NC_I . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
G.10 Experiment results of scenario NC_J . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
G.11 Experiment results of scenario NC_K . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
G.12 Experiment results of scenario NC_L . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
G.13 Experiment results of scenario MassC_A . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
G.14 Experiment results of scenario MassC_B . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
G.15 Experiment results of scenario MassC_C . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
G.16 Experiment results of scenario MassC_D . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
G.17 Experiment results of scenario MassC_E . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
G.18 Experiment results of scenario MixC_A . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
G.19 Experiment results of scenario MassC_F . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
G.20 Experiment results of scenario MassC_G . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
G.21 Experiment results of scenario MassC_H . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
G.22 Experiment results of scenario MassC_I . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
G.23 Experiment results of scenario MassC_J . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
G.24 Experiment results of scenario MassC_K . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
G.25 Experiment results of scenario MassC_L . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
G.26 Experiment results of scenario MassC_M . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
G.27 Experiment results of scenario MassC_N . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
G.28 Experiment results of scenario MassC_O . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
G.29 Experiment results of scenario MassC_P . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
G.30 Experiment results of scenario MassC_R . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
G.31 Experiment results of scenario Scheduling_A . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
G.32 Experiment results of scenario Scheduling_B . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
G.33 Experiment results of scenario Scheduling_C . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
G.34 Experiment results of scenario Scheduling_D . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
G.35 Experiment results of scenario Scheduling_E . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
G.36 Experiment results of scenario Scheduling_F . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
G.37 Random components chosen for analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
G.38 Silo parameters for all experiments with random component for cross-contamination calculations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
G.39 Dispatching rules fixed for scenarios with random component for cross-contamination . . . . .
G.40 Scheduling rules chosen for scenarios including the random component . . . . . . . . . . . . . .
G.41 Experiment results of scenario Random_C . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
G.42 Experiment results of scenario Random_D . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
G.43 Experiment results of scenario Random_E . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
G.44 Experiment results of scenario Random_F . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
G.45 Experiment results of scenario Random_G . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
H.1
H.2
H.3
H.4
H.5
H.6
H.7
Comparison of impact of different discharge speeds on order completion for no additional silos
Comparison of impact of different discharge speeds on order completion for 2 additional silos .
Comparison of impact of different discharge speeds on order completion for 4 additional silos .
Comparison of impact of 0 to 2 additional silos on order completion . . . . . . . . . . . . . . . .
Comparison of impact of 0 to 4 additional silos on order completion . . . . . . . . . . . . . . . .
Comparison of impact of 2 to 4 additional silos on order completion . . . . . . . . . . . . . . . .
Comparison of impact of silo discharge and reservation on bagging throughput with no additional silos . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
H.8 Comparison of impact of silo discharge and reservation on bagging throughput with 2 additional silos . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
H.9 Comparison of impact of silo discharge and reservation on bagging throughput with 4 additional silos . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
H.10Comparison of silo discharge speed and number of extra silos impact on bagging throughputs .
119
119
119
119
119
119
120
120
120
121
121
121
121
121
122
122
122
122
122
123
123
123
123
123
124
124
125
125
125
126
126
126
127
128
128
128
128
129
129
129
129
132
132
132
132
132
132
132
132
133
133
L IST OF TABLES
H.11Comparison of impact of order sequencing and nutrient-based silo allocation on total crosscontamination . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
H.12Comparison of impact of additional silos on order completion . . . . . . . . . . . . . . . . . . . .
H.13Comparison of impact of additional silos on total contamination . . . . . . . . . . . . . . . . . . .
H.14Comparison of impact of additional silos on the average throughput . . . . . . . . . . . . . . . .
H.15Comparison of impact of random component on the average throughput . . . . . . . . . . . . .
H.16Comparison of impact of random component on the total contamination . . . . . . . . . . . . .
xiii
133
133
134
134
134
134
A BBREVIATIONS AND D EFINITIONS
In the document the following abbreviations are used:
APS
CRN
CSLSP
DEM
DES
ERP
EU
FIFO
FMS
GMO
KPI
MAS
MO
MHS
MTBF
OR
SN
TTR
WIP
Advanced planning and scheduling
Common random numbers
Capacitated Stochastic Lot-Sizing Problem
Discrete element method
Discrete event simulation
Enterprise Resource Planning
European Union
First in first out (queuing principle)
Flexible manufacturing system
Genetically Modified Organism
Key performance indicator
Multi-agent systems
Manufacturing orders
Material handling system
Mean time between failures
Operations Research
Scenario Navigator (software)
Time to repair
Work in progress
And the following most important definitions are distinguished:
Carry-over
a separation of portions of mixture from the production batch, left in the manufacturing line or its parts
as a remainder
Cross-contamination
unwanted release of the carry-over by mixing it into a subsequent product
Nutrient
any distinguishable substance that provides nourishment essential for the maintenance of life and for
growth
Product Contamination
amount of foreign, other than specified material in a product, comprised of picked up residue
Release curve
also called “cross-contamination curve”, a graphical representation of ratio of contamination as a function of material flow for a given container
Residue
stationary amount of carry-over in a distinguished processing or transportation segment, comprised of
products priorly occupying it
Tardiness
also “lateness”, the quality or condition of occurring after scheduled time.
xiv
1
I NTRODUCTION
This chapter establishes the problem to be investigated, and gives a description about the work context. It
is followed by a research objectives and questions formulation, together with the utilised methodology and
planned deliverables. Finally, the supporting company, that delivers vital data and allows to place the issue is
more precise context, is briefly introduced.
1.1. R ESEARCH P ROBLEM
In 2004 four edible products were recalled from shelves in New Zealand due to exceedingly high levels of lead.
Investigation quickly traced the source to a particular batch of corn imported from China, then milled into
flour. However, detailed analysis could not identify the exact point of entry to either maize supplier, milling
or further processing, and the origin of the pollution remained unknown for a long time. Eventually, after
extensive and costly search, the source of the contamination was determined to come from a bulk shipment
of corn, where an unclean cargo container was used, and the previous product residue was carried over.1
The above is one of many examples of product cross-contamination, which can occur throughout the
entire supply-chain, being very difficult to pinpoint, and often resulting in commodity quality deterioration.
That is why companies need to put more effort to recognising and dealing with risks of such pollution, if
they want to assure long-term stability in the value of their output. Especially, that their good reputation is at
stake, as the eventual company on a product label is considered responsible by the customers or the public
opinion.
1.1.1. B ACKGROUND
Production processes in factories are often optimised for high capacity, where a lot of high technology solutions are used to increase productivity (Mieghem, 2003). With more added automation comes more complexity, that makes it difficult to understand the production process, especially when allowing for high manufacturing flexibility and a large product base. In such cases the technology making it possible is also the source
of hidden problems, that manifest themselves in deteriorated finished product quality, but are otherwise very
difficult to locate. Managing not fully understood problems has then little chance of success. It is alarming
that the knowledge of what happens during product processing is so limited, and that so many managers
only prioritise performance.
Nevertheless, a lot of companies spend plenty of their resources on quality assurance of their finished
goods, but fail to understand what really happens or what could happen during processing necessary to deliver that particular product. Misguided actions can measure the exact quality of selected finished commodities, but do not specify reasons for it being at a determined level, providing purely factual data on the end
state, which cannot be used for a proper investigation. Moreover, there is a limit to expenditure on quality
assurance, as the enterprises must believe in acquiring profits while these measures have a direct impact of
the prices of goods (Mapes et al., 1997). That is why, more focus needs to be put to understanding why a
specific output is achieved and what causes it, as otherwise it is not possible to take steps to improve the
process. Better recognition of quality deterioration risks can not only help in avoiding them, but also reduce
costs associated with quality assurance.
1 Source:
http://www.foodsafety.govt.nz/elibrary/industry/Source_Lead-Nzfsa_Confident.htm
1
2
1. I NTRODUCTION
This research deals with a possible decline in product quality caused by cross-contamination, which
often occurs in multi-product manufacturing lines. Term “cross-contamination” refers to unwanted additions to a product as a result of picking up residue from the preceding product, left in utilised equipment,
which has not been cleaned (or properly cleaned) in between (Strauch, 2003). It is a very common but
hardly well-understood issue, which might result in a significant decline in quality of some goods. Specialists with considerable industry experience can name plenty of factors that they think have influence over
the process,2 ranging from equipment properties, through product specifications to production scheduling.
Cross-contamination is acknowledged as probably the greatest source of quality related issues and the need
to improve on that is generally expressed. However, there is no knowledge about the precise nature of the
phenomenon or its largest contributors, and this is the named reason why measures taken to mitigate the
problem are not comprehensive.
1.1.2. P ROBLEM S TATEMENT
A lot of factories use versatile and flexible multi-product lines to manufacture various goods, employing the
same machinery with a whole network of interconnections. Such solution allows for being agile, producing
smaller batches, quickly responding to customer demands, and keeping warehousing to minimum. With
more advanced and multi-purpose automated equipment, it has become possible to produce miscellaneous
products easily, utilising the devices as much as possible (Erenguc et al., 1999). Because of that, separate lines
for each product are no longer required, allowing to save money on equipment and personnel, especially
when potential production capacity is much higher than the market demand. The common practice is now
to increase the product base considerably, suiting more precisely to various needs of the identified customer
segments. More narrow client groups, limited product shelf life and required manufacturing agility lead to
small batches and numerous production changeovers.
With production line versatility also come certain complex problems. Flexibility allows companies to have
multiple connections, and use various machines for the same job, combining the sets almost freely, depending on the organisational structure. However, the complexity of scheduling and managing production rises
along with flexibility, leading to situations where it is very difficult to find the best solution for manufacturing planned goods. Moreover, companies tend to prioritise machine utilisation as a measure of production
performance, which in general needs to be high to create more finished products (Smith, 2003). But high
equipment utilisation leads to time pressure, and might result in poor preparation for the following product,
in terms of e.g. improper cleaning, lack of supplies or rushed setup. Determining, what is the relation between production efficiency (the aspects of which are e.g. throughput, quality, tardiness, etc.) and process
flexibility in relation to cross-contamination could help gain better insight on choices to be made, especially
if there are any trade-off characteristics.
Every time a manufacturing line has to be converted to suit a different good, there is a risk of crosscontamination. Equipment that has previously handled certain product is very likely to be polluted with a
carry-over from it, which might transfer into the subsequent one. When a cleaning run in between them is
omitted or one performed does not remove all residue, there is a chance of mixing it with contents of the
following commodity. Naturally, the amount of pollution in the beginning is on average higher and decreases
with every processed unit. However, due to sticking to surfaces or deposition in pockets some of the material
can be released in a chaotic, seemingly random way, that is very difficult to predict. If the transferred material
is unwanted, it might lead to a significant decline in quality, and result in an inferior product, considered by
the manufacturer as out of specification, where certain measurable properties are outside established limits.
Moreover, in cases like food processing or for pharmaceutical use, such occurrence might even be harmful
for the user. But not all cross-contamination is adverse to product quality. Sometimes, large amounts of very
similar in composition additive might not be as unwelcome as small inclusion of another, sensitive material.
Surprisingly, little research has been done to measure and quantify the extent of cross-contamination.
Companies are very reluctant to admit their quality assurance problems or do not realise them fully3 . Moreover, the cross-contamination phenomenon exists along the entire production line and is very specific to
equipment characteristics, handled material properties and cleaning procedures, having many difficult, if
not impossible to predict, aspects. Determining cross-contamination features for one product does not necessarily lead to increasing knowledge about another.
The accurate knowledge on how residue from one product is transferred into the subsequent one is
normally missing. Because of that and a considerable difficulty in procuring it, investigation into cross2 Based on 5 separate interviews with employees of Sloten B.V.
3 See footnote 2
1.2. R ESEARCH O BJECTIVES
3
contamination is normally absent in production models or scheduling solutions. At best, based on the historical issues, companies include such scrutiny for operations as a final ‘nuisance’ to be accounted for. In
such cases cross-contamination is treated as a black-box to which ‘clean’ ingredients enter to be processed,
and exit as finished goods with some pollution. Thus, understanding the phenomenon, and ability to predict
its extent as a multi-stage process, dependent on different variables, would be very beneficial to final quality
assurance.
There are three potential sources of quality norms: internal standards, regulations in law and customer
expectation. Normally companies follow the legal requirements but also specify their own internal limits on
contents or significant aspects of their products, or comply with quality standards of private systems, either
compulsory or voluntary (Trienekens & Zuurbier, 2008). However, clients often do not have resources to measure the quality of products or prove their inferiority, and the producer reputation serves then as assurance.
Because of that, customers are possibly more prone to react to scandals publicized by the media. Sometimes
to comply to regulations or mitigate risks the companies reluctantly admit that there are some possible trace
amount of additives present, which are not included in their recipe. An example for that is presence of trace
amounts of peanuts in many products in the food industry, only because the same line is used to produce
commodities including them. Following products, which have no specified peanut content, despite a thorough cleaning, are still at risk of cross-contamination (Taylor & Baumert, 2010). Due to a large number of
people allergic to peanuts it is safer for companies to admit possible (but not certain) contamination, than to
face potential lawsuits.
Prediction of the extent of cross-contamination has a direct relation to manufacturing costs and risks,
and can be used for devising interventions. First of all, measures different than cleaning can be taken to
reduce cross-contamination hazard. If certain, known portion of the goods in front could be reworked or
discarded, it might prove less costly than to perform cleaning and losing production time. Moreover, more
informed production scheduling can be done to mitigate risks associated with cross-contamination. All in
all, understanding and predicting the extent of quality loss due to cross-contamination is the first step to
devising procedures aiming to limit it.
Nonetheless, there is a perceived trade-off relation between production performance, product quality as
well as system flexibility. The latter, viewed often as having more options to choose from when allocating resources, is an essential source of the cross-contamination problem, as well as its inherent solution. In principle, multi-product manufacturing requires easily adaptable connections among various pieces of machinery
and storage, which is enabled by technology. Was that not possible, the cross-contamination problem would
no longer be a major issue as the production lines would be fixed and product-specific (or product family specific). When that is not the case, broad flexibility poses a greater risk of cross-contamination, when multiple
connections can be contaminated to various degrees, and the ability to predict the outcome decreases.
Finally, investigation into possible effects of cross-contamination is not specific to factory operations but
could come much earlier, in the design stage, to evaluate different layout possibilities and choose one best
suited. Normally, detailed analysis in this matter is not done, and the knowledge about the issue is gathered
from operational experience, when it is too late or too costly to introduce any major changes to the line. Moreover, such analysis can result in alternative strategic choices, like limiting the number of handled products
or specific sensitive ingredients, striving for better monitoring, reducing system flexibility to reduce risks, or
other.
1.2. R ESEARCH O BJECTIVES
The following section lists several intermediate steps to be taken during the study. It is divided into two
groups, the first dealing with the main research itself and the other with testing and experimentation based
on a case study. At first, in the theoretical part, an initial analysis is done and conceptual model derived
with general definitions of possible interventions. Then, in a case study, cross-contamination measurements
lead to a complete formulation of a specific experimental setup, fully defined simulation model, that is later
investigated. Then, data gathered is analysed and conclusions are made. Thus, the intermediate goals are:
• Analysis of factors related to cross-contamination in a multi-product plant, based on interviews;
• Literature study on cross-contamination issues,
• Construction of a conceptual model,
• Derivation of generalised cross-contamination models,
• Definition of possible dispatching and sequencing rules to be investigated,
4
1. I NTRODUCTION
• Case study experimentation:
– Cross-contamination measurements preparation, conduct and analysis,
– Simulation model build, comprising production and cross-contamination models with user input
of various scheduling rules, using Simio software,
– Verification and validation of the simulation model,
– Experimentation with various system layouts and scheduling methodologies,
– Analysis of the experiments results.
• Set of general conclusions arising from the case study,
• Proposal of possible future additions and improvements to the model and research.
The most important research step is to draw general conclusions from the case study and experiments within,
that could be applicable to a wider group of problems.
1.3. R ESEARCH QUESTIONS
The following section formulates questions to be answered during research. The central issue of the analysis
is cross-contamination phenomenon and its impact on production performance. The main study inquiry is
specified to be:
What are the effects of cross-contamination on the amount of out of specification goods in a multi-product
plant with varying buffer capacity and scheduling rules?
Thus, the research is to contribute in quantifying the consequences of cross-contamination process for finished products in factories, with multiple production changeovers, where there are two main areas of interventions: intermediate buffer number and capacity as well as production scheduling. With the main research
question come a few ancillary ones, formulated to specify the study focus:
1. What is cross-contamination, and why is it important for multi-product factories?
2. What are the relationships between factors influencing cross-contamination?
3. What are the effects of cross-contamination, and how are they relevant?
4. When is a product considered out of specification?
5. How to express product-residue mixing in a mathematical way so that it can be used as a general approach for expressing product contamination?
6. Can cross-contamination phenomenon be modelled in a discrete event simulation paradigm?
7. What are the effects of additional buffer capacity in manufacturing on cross-contamination and production capacity?
8. Which scheduling rules are beneficial for reducing the extent of cross-contamination, and what is their
impact on production capacity?
Provided questions are to structure research objectives to in the end provide insight into considered problem,
and assist in answering the main question. They indicate that the research is of quantitative, experimental
nature but often basing on more abstract issues like risk and importance perceptions, quality assurance.
1.4. M ETHODOLOGY
Several different methodologies are used to tackle various problems of the project. Firstly, some imposed
constraints and requirements are mentioned, and a general approach to supply chain as well as production
planning is presented. Then, a simulation approach, describing the requirements and limitations of DES is
given, followed by a more guided scheduling scheme, especially utilising dispatching rules. The research can
be thus divided into 5 main phases: problem formulation, specification, model build & testing, experimentation and analysis with conclusions. At first, knowledge base is built and the problem to investigate defined.
Requirements for further investigation as well as theoretical foundations are prepared in the second stage.
Then, these are translated into a simulation model for a specific case and tested. Consequently, a set of experiments exploring the impact of independent variables on dependent ones (KPIs) is performed, and the
results are analysed in the last part. Finally, conclusions and recommendations are made.
1.5. D ELIVERABLES
5
Figure 1.1: A schematic overview of the research steps
1.5. D ELIVERABLES
The following deliverables are to be submitted as a result of the project:
1. Conceptual model on how to include cross-contamination effects in a multi-product, powder-handling
plant,
2. Discrete event simulation model based on the case of Sloten constructed in Simior software,
3. Scenario Navigatorr tool allowing for scenario setup and simulation result visualisation,
4. Recommendations for limiting the possible effects of cross-contamination regarding system layout and
scheduling methodology.
As it is not the intent of this research to deliver a tool to assist with experimentation and result assessment,
the created SN software is to be of demonstrative nature, handling data and automatically preparing visualisations.
1.6. C ASE C OMPANY
The experimental setup bases on the case of a major young animal feed producer Sloten, a subsidiary of
Nutreco corporation. Its facility in Deventer, the Netherlands, is chosen as an example to provide a real-world
data, and to test the performance as well as validity in a realistic setting.
Because of the handled material structure - a sticky powder, and lower safety standards in the industry,
comparing to human food production, animal feed manufacturing is a prime example of a relatively high
cross-contamination issue, where a lot can be gained if the process is understood and taken advantage of.
Strauch (2003) bringing the result of a survey from the German Research Institute of Feed Technology shows,
that cross-contamination is present in all examined plants and it is increasing, because of constantly introduced new feed formulations, produced on the same line. Moreover, there seems to be a consensus, that
contamination-free manufacturing is not possible, due to scale effects, in comparison to especially pharmaceutical industry, where it is an absolute must. Despite that, and also in consequence of limited knowledge
on the effects contamination has on animals and further on humans, regulatory bodies strive for a “zero tolerance” restrictions on any carry-over. Being pushed by the authorities, their customers as well as trying to provide high quality products, companies are forced to research and mitigate the effects of cross-contamination
in their facilities.
Cleaning the processing equipment, to reduce the impact of cross-contamination between one product
and the following one, is normally possible. It is based on releasing another material in place of the product,
and sending it through the same line to gather the residue. But for handling low-moisture powder, as in the
feed industry, no liquid or any watery substance can be used, because it would affect product quality. Thus,
normally a regular sales item is used to perform the cleaning, which is then disposed of, sold as inferior quality
6
1. I NTRODUCTION
product or reworked. All of these are costly to the manufacturer. When asked about possible impact of closer
scrutiny of cross-contamination, specialists4 list greater complexity of planning, more production time and
product losses, and more constricting manufacturing procedures. As all of them are viewed as negative, the
ignorance about the cross-contamination character is viewed as bliss, because no immediate action needs
to be taken. Thus, organisational reluctance has its tow on possible investigation. All in all, if the extent of
quality loss in unknown, then the potential gains of better quality assurance are even more uncertain.
Moreover, mixing powders, an intermediate step in feed production, is very difficult as usually the added
ingredients have varying physical properties, like size, viscosity or density, which prevent from equal distribution of particles. What is more, a reverse process to mixing, called segregation, occurs strongly in the
mixing of solids, and is not limited to stirring equipment but happens with every powder movement (Poux et
al., 1991). Ramifications of this process for long transportation lines are vital if the material is to be divided
into smaller parts (e.g. packaged), as some of them might contain ratios of certain components that is much
lower or higher than intended. On top of that, measuring the homogeneity is not an easy task as it requires
sampling, in which it is vital to determine location, size and number of taken samples as based on that, the
quality of the entire structure is assessed. Reducing the uncertainty about the mixture contents would yield
quantifiable benefits to the producers.
Cross-contamination has an impact not only on product quality, but also on feed safety issues, which
might lead to human food safety concerns. One of the representative examples is adding copper to piglet
feed, which has positive effect on their growth. If such additive gets mixed up with products targeted to lamb,
the excess of copper might lead to the animal’s poisoning and eventual death (Zervas et al., 1990). For impact
on humans, it is sometimes desired to exclude e.g. medicated feed, especially antibiotics, because of public
health issues, or genetically modified organisms5 (GMOs). Naturally, to avoid lawsuits and damage to the
brand name, companies would like to prevent such occurrences.
Finally, the project is also supported by Systems Navigator, a consultancy firm with a vast experience
in supply-chain simulations. They offered their expertise as well as software tools necessary to handle the
problem. Together with the Sloten’s parent company Nutreco, a global animal nutrition and fish feed supplier,
they are involved to provide counsel and know-how to help gain insight on the issue.
1.7. S TRUCTURE
This report comprises of eight main chapters, complemented with an extensive appendix. The structure of
the main part is overlaid on the methodology graph and presented in Figure 1.2. Following this introduc-
Figure 1.2: An overview of the document structure overlaid on the research steps
tory chapter is a thorough analysis of current state of knowledge about the topic (chapter 2). Specifications
and requirements for modelling are included in the first part of chapter 3, along with boundaries, simulation approach and modelling choices made. Chapter 4 describes the production model - a base for capacity
analysis and model components. It is followed by cross-contamination model derived in section 4.4, where
4 Independent reaction of three employees of Sloten, working in operations.
5 See:
http://www.cdc.gov/narms/animals.html
1.7. S TRUCTURE
7
the method and path to obtain both mixing and partial mass exchange models is depicted. This is followed
by description of implemented scheduling methodology and dispatching rules, that concludes the design
phase.
Experimental setup, based on the factory layout and organisation of Sloten, is defined in chapter 5, after which all derived models are verified and validated to the best extent of possessed data in section 5.7.
Chapters 6 and 7 deal respectively with conducted experiments and their analysis, where the conclusions on
the scientific relevance are made. These are followed by recommendations arisen from the results scrutiny.
Conclusions about the work can be found in chapter 8 with a rumination on the project and the report is
completed with a set of possible future expansion of the investigation.
2
L ITERATURE A NALYSIS
In order to undertake any research, a solid theoretical foundations are essential beforehand. This chapter
presents a review of fundamental and contemporary knowledge on the topic at hand.
As the main scope of the research is to determine the effects of cross-contamination in production, these
need to be defined first before the simulation model exploring production scheduling with cross-contamination calculations is built and investigated. Then, answering how does the analysis fit in a landscape of
Operations Research (OR) and in particular in production line manufacturing of similar endeavours. Especially, how to define production efficiency, versatility as well as flexibility and what are the trade-offs between
them, if there are any. Furthermore, how to construct and validate a production line DES model and complement it with cross-contamination model as well as scheduling possibilities. Moreover, a short review on the
animal feed manufacturing framework and an extensive production scheduling review is included, referring
to topics like dispatching and sequencing.
2.1. P RODUCTION T RADE - OFFS
Production is an intermediate step in supply chain logistics. It deals with transforming raw materials into
products, by giving them value that is recognized by customers (Min & Zhou, 2002). In this context, it is
highly focused on a single production line, taken from the entire landscape of supply chain (mostly procurement and distribution), and dealt with on its own. Yet it is acknowledged, as Min & Zhou (2002) or Beamon
(1998) emphasize, that the entire supply chain needs to be treated as an integrated system. Furthermore,
uncertainties need to be incorporated into the system to achieve better results (Mula et al., 2006). However,
Erenguc et al. (1999) recognises that some uncertainties, like poor coordination, discrimination against internal customers or linking problems between inbound and outbound logistics are not easily quantifiable, and
thus often not included in such analysis.
There are four main categories of supply-chain goals, each looking at a different part of the business. First
deals with customer service and distinguishes variables like product availability or response time. Second,
monetary value, is characterized by cost behaviour, return on investment and asset utilisation. Furthermore,
there could be a goal to gain information or knowledge by the company, and finally risk elements in terms of
quality and imperfect information sharing (Min & Zhou, 2002). All of those are different aspects of production
efficiency, and are typically interconnected with trade-off relations, where one has to choose which variable
is more important to satisfy. This is of great importance to companies, having real economic significance.
Gupta & Goyal (1992) give an example of a large product base that might reduce the plant’s ability to produce
large volumes. Yet Erenguc et al. (1999) claim that these should not be each other trade-offs but rather need
to be simultaneously prioritised, looking from a strategic point of view. Thus products need to be delivered
quickly and cheaply, keeping production agile, their quality impeccable and inventories low. Mapes et al.
(1997) name six most common decision variables in these terms: quality consistency, quality specification,
lead time, delivery reliability, cost, flexibility and innovativeness. Min & Zhou (2002) expand it and classify to
determine three major group of restrictions on possible decisions managers can make and the feasibility of
which, they need to asses. These are:
• capacity, including financial, production, supply, technical, workforce or IT adoption, among others,
8
2.2. A NIMAL F EED P RODUCTION AND C ONTAMINATION
9
• service compliance, for meeting customer requirements, or e.g. quality,
• the extent of demand, so that the capacity of supply is balanced with the extent of consumption.
Adler et al. (1999) define flexibility as the ability of a manufacturing system to cope with changing circumstances, and calls the trade of between it and efficiency one of the firmer grounded notions in the entire
OR. The idea is that efficiency requires bureaucracy, standardisation, formalisation, specialisation as well as
hierarchy, and these all impede adjustments and smoothness.
An interesting shift in understanding and dealing with production trade-offs can be found in literature.
Gupta & Goyal (1992) and Mapes et al. (1997) understand trade-offs literally, as sacrificing one performance
indicator for another, possibly finding a minimum impact point of their combination. Since then, Adler et al.
(1999) change this view to postulate for parallel improvement of all, supported by an appropriate organisational approach. Also Silveira & Slack (2001) call trade-offs a ‘myth’, which holds back companies while other
manage to outperform them in many areas simultaneously. As trade-offs are believed to exist, they create a
significant bias on how managers approach process improvement.
For example Sandborn & Vertal (1998) perform inquiry into trade-offs in production due to different packaging options. They conclude that such analysis needs to be done for entire product families life-cycles,
sometimes sacrificing individual performance for the entire portfolio. Mapes et al. (1997) notice that producing wide variety of products with different quality specifications increases process complexity that leads
to errors in scheduling, unplanned delays and poorer quality in general. On the other hand, Silveira & Slack
(2001) regard the trade-off concept as more problematic to academics than it is for practising managers.
2.1.1. P LANT L AYOUT D ESIGN
System design is making decisions on long-term features of the system, such as layout, configurations and
capacities. It is vital to commit proper resources to that, to assure the best possible structure, as it might not
change for a long time. The analysis and adjustments can be much more easily done with support of simulation (Smith, 2003). In this case, a problem of production scheduling in multi-stage, multi-item capacitated
systems is described. Moreover, analysis of production trade-offs needs to be incorporated with functional
verification and layout design so that more optimum solution can be found (Sandborn & Vertal, 1998).
As companies are under constant pressure to offer a wider group of products, the complexity in managing
that rises as well. Mapes et al. (1997) acknowledge that such impact might be limited by the organisational
structure, but it is unlikely to be non-negative. That is why organisations tend to offer great variety of products but try to limit it within particular plants. Still, for a single facility there are numerous products, whose
efficiency of production needs to be supported by the plant layout (Benjaafar et al., 2002, Djassemi, 2007).
This is a serious problem as often factories are upgraded from their previous schemes, which was not built
for multi-product purposes, and the changes are hardly ever composite, involving only a part of the process.
Typical design criteria for plant layouts also do not capture the relationship between layout flexibility (in
terms of operational interconnections) and performance. Thus, analysis into trade-off relations of different design possibilities are vital for finding superior solution (Benjaafar et al., 2002). Most measures evaluating layout suitability deal with material handling efficiency (like throughputs) to evaluate candidate scenarios (Djassemi, 2007). However, these often fail to capture other characteristics like quality (or crosscontamination as in this research), punctuality or organisational structure. It is very important to expand
such analysis to suit wider needs and increase prediction validity.
2.2. A NIMAL F EED P RODUCTION AND C ONTAMINATION
Because animal agriculture has been rapidly growing for the past decades, there is an increasing interest in
choosing the right feeds to allow for faster and more efficient growth of livestock. Young animals are especially susceptible to given nutrients, so the industry puts special effort to providing them with best suited and
balanced diet. This often includes dried dairy products and derivatives (Crawshaw, 2012). Thus, the industry
expanded and now every species has a specially developed for them feed or multiple feeds. But with variety comes also greater manufacturing cost and production process complexity. All that while striving for low
production costs and retail price as well as increased manufacturing techniques and quality (Behnke, 1996).
Growing number of products manufactured on a single line entails higher difficulty of keeping their critical
components separate and larger risk of cross-contamination via their carry-over (Strauch, 2003).
According to Fink-Gremmels (2012) animal feed production industry is often disregarded and only mentioned when some emergency arises causing harm to people via livestock. Such occurrences might include
10
2. L ITERATURE A NALYSIS
pathogens (e.g. prions), dioxins or adverse substances like melamine in milk. But the importance of proper
nutrition of animals, especially the young ones, should not be underestimated, as it becomes a highly innovative business with specialised targets, high quality resources and considerable socio-economic impact. Moreover, the number of regulation and contamination limitations increases steadily, especially in the EU, where
since 2002 animal feed is recognised as a part of the human food chain (Crawshaw, 2012, Fink-Gremmels,
2012).
However, typical contamination considerations involve mostly potentially harmful additives, comprising
only marginal portion of the feed, especially microbiological or chemical agents that can pollute the material
not only during production, but also be a part of incoming raw materials or enter during storage or transportation (Fink-Gremmels, 2012). This report concentrates on cross-contamination, which is, according to
Leloup et al. (2011), Strauch (2003), a process of mixing of carry-over of additives during material handling,
which is a common issue for animal feed industry that can be extended to many powder-based industries.
In consequence, sometimes finished products contain other than specified ingredients, which in general are
unwanted but have no adverse effects on the animals. However, there also might be severe repercussions,
for example when lamb product gets contaminated with high levels of copper (typical for piglet feed), which
causes harsh poisoning to the animal (Zervas et al., 1990). To understand more about the issue Leloup et al.
(2011) prepared a test-bench for material elevator using several different feeds, and a tracer with collector
method. Although differences among feeds are noticeable (which is explained by different cohesiveness due
to varying fat content), the behaviour for single powders is consistent with limited variability. Fitzpatrick et al.
(2004) perform analysis on the flow characteristics of milk powders with different fat contents. They conclude
that these differ significantly in terms of cohesiveness as well as friction angles thus varying considerably in
flow properties, which demands distinguishing in modelling.
Moreover, Leloup et al. (2011) emphasise that trials during production are difficult to carry-out, especially because of lack of access points to equipment and freedom to watch the particle movements, while
experimental setup differs from the real system, and is often simplified. Therefore careful trials on the actual equipment with intermediate steps are necessary to derive a model for cross-contamination. Toso et al.
(2009) investigate a production scheduling problem for an animal feed plant, which can be a starting point
for the research at hand. They claim that proper sequencing of products can take advantage of a ‘cleaning’
property of two consecutive products if the lots are big enough, which means that the previous residue is collected and no intermediate cleaning is required. But that usually works only with very similar products and
other might pick up too much residue to fit quality norms.
Furthermore, manipulating product batches by means od lot-sizing (see Lee et al., 1997), can join or split
orders and take advantage of the lower residue and changeover times that may not follow the triangular rule.
That means that separate set-up change from first product A to B and then B to C might be in total shorter than
straight change from A to C. Also, Toso et al. (2009) point out that while normally lot-sizing and sequencing
is dealt with separately, combining those might lead to greater flexibility and overall performance, which is
also confirmed by (e.g Drexl & Kimms, 1997). Finally, animal feed demand follows a seasonal pattern which
impacts the scheduling greatly. Normally, companies deal with it by working overtime and making sure they
have enough ingredients in stock to avoid a backlog (Toso et al., 2009, Behnke, 1996). Clark et al. (2014)
propose a more general solution to scheduling with contamination and non-triangular sequence-dependent
setup times. They point out that a capability to absorb contamination presents opportunity that cannot be
ignored as it might result in reducing the number of intermediate cleaning runs as well as the processing
time.
2.3. S IMULATION O VERVIEW
According to Min & Zhou (2002) supply chain models can be divided into four categories: deterministic,
stochastic, IT-driven and hybrid. This analysis concentrates on the latter, as it includes simulation (especially
DES), which combines all of these methods and can potentially result in superior outcomes. Also Mula et
al. (2006) recognise that models which do not include uncertainty produce inferior results. By further dividing uncertainties into several categories they discover that simulation models, including Multi-agent systems
(MAS), have been utilised in the past for each of distinguished categories (Mula et al., 2006, see Table 2, pp.
272). Terzi & Cavalieri (2004) claim that the use of simulation methods has been increasing more rapidly than
a general use of IT, and in particular DES is considered a vital key-success factor for company survival. Also
Jeong & Kim (1998), Gupta & Goyal (1992) claim that simulation can be effectively used in design, implementation and operations of FMS.
2.4. P OWDER M IXING
11
Negahban & Smith (2014) postulate that simulation is an appropriate method to evaluate design of manufacturing systems, and to attempt to improve it by exploring different layout alternatives. Most important
part of the simulation is a material handling system (MHS), that is concerned with the use of raw materials
and their transformation. Jeong & Kim (1998) present an extension approach to production called a flexible manufacturing system (FMS), which consists of a collection of numerically controlled machines, that are
further regulated by an automatic MHS. Setup is then divided into four decision-making stages: design, planning, scheduling and control, that work closely together. For this case, it is vital to understand how planning
and scheduling exchange information and cooperate.
In a discrete event simulation the state of the model can change only at discrete time points, called events.
Between those events the time is continuous, and the system trajectories can be thus considered piecewise
constant, i.e. the state variables remain fixed between events. There can be any number of state changes,
called transitions during an event, for the being of which time does not progress. A simulation model specifies
how the events are scheduled and what state transitions are caused. Then a given simulator tool handles the
execution of the model (Schriber et al., 2012, Zeigler et al., 2000). Simior is a popular DES commercial tool1
that follows object-oriented world-view, in which all objects are derived from the same base object. Execution
is based on Processes, that can be expanded and customized by users to create own objects (Schriber et
al., 2012). Software includes stochasticity considerations, allows for experimentation and use of additional
packages. For further details see Kelton et al. (2013).
Furthermore, it is crucial that any developed production simulation model could be used as a part of a
larger supply-chain business simulation, as interconnectivity plays a major role and needs to be taken care
of, by choosing appropriate techniques in the early stages of modelling (Terzi & Cavalieri, 2004). Support of
distributed real-time platforms might be necessary for proper information exchange, and is definitely useful
for possible future additions. As such, any model cannot be considered completely standalone, but should
allow for coupling it with other models or systems. Simulation can play an important role as a support for
decision making on design of system layouts. Providing the model is complete and transparent, built with
full involvement of the analyst skills and process experts, it can be useful to both high-level alternative comparison as well as visual interactive communication means (van der Zee & van der Vorst, 2005). They also
point out that the model control structure needs to be clear and centralised, in order to allow the decision
makers to efficiently control the system, without arbitrary hard-coded parameters. Only with clearly and
realistically defined model elements and relationships, adequately reflected model dynamics, separate user
interface and easy setup of various scenarios, can the demands for suitable decision support tool be fulfilled.
Consequently, combining different performance indicators with a means of comparing them directly is a vital
part of a possible multi-criteria decision-making problem (Gupta & Goyal, 1992).
2.4. P OWDER M IXING
Powder mixing is an important process in many industries. It is a mechanism of bringing together solid
particles to produce a mixture of regular consistency, which occurs due to different velocities and directions of
the agitated material (Sen & Ramachandran, 2013). Being able to mathematically express and predict the level
of homogeneity in mixing is an important step in order to control that, and thus many models for that have
been developed in the past. Obtaining homogeneity might be problematic due to mixed product diversity,
expressed by parameters such as size, shape, moisture, surface nature as well as mixing apparatus differences.
Moreover, mixtures also undergo a reverse process to mixing simultaneously, called segregation, which often
prevents from achieving homogeneity (Poux et al., 1991, Sen & Ramachandran, 2013).
Sen & Ramachandran (2013) distinguish several modelling approaches to the problem: Monte-Carlo
methods, continuum and constitutive models, data-driven statistical models, compartment models, residence time distribution methods, hybrid models and discrete element method (DEM) based models. Additionally, they create a population balance model which describes the development of separate entities, where
particles can combine and break up over time. Portillo et al. (2006) compartmentalize investigated V-blender
in order to decrease computational intensity and allow for a relatively large number of particles. At discrete
time steps particles can randomly (and depending on the particle flux) enter or leave given compartment,
that is assumed to be uniformly mixed, leading to a large-scale mixing process. Most accurate models for
granular mixing are based on DEM, especially useful for processes requiring high accuracy e.g. pharmaceutical, where proper dosage in pills is of utmost importance. Bertrand et al. (2005) reviews several different
approaches and solutions to this problem, also discussing limitations, especially in processing power. Al1 Software website:
http://www.simio.com/index.php
12
2. L ITERATURE A NALYSIS
though this has improved significantly over the past years, it still remains one of the biggest drawbacks of this
method (Sen & Ramachandran, 2013).
However, all aforementioned models describe processes of intentional mixing, where rotating equipment
stirs the particles. Models for coincidental mixing or involving a mixer and adjacent particle transportation
system are very limited in the literature, especially when including a whole production system and not just
its part. This gap must be filled in order to suitably support the industry as these effect will occur whether
recognised or not, and the only right way to limit them is by analysis first. Still, many investigated the possible
mixing effect in silos. E.g. Wu et al. (2009), Cleary & Sawley (2002) perform detailed analysis on silo flows,
and conclude that the incoming material to silo amasses roughly in a heap shape, while discharge forms a
funnel, resulting in shifting of material. As such, if the inflow material has some changing characteristic (e.g.
contamination level), its outflow behaviour will be altered in a way, that is very difficult to predict unless the
material and silo properties are very well known. This makes it extremely difficult to make general predictions
on material quality.
2.5. P RODUCTION S CHEDULING
Scheduling is the process of arranging, controlling and optimizing work and workloads in a production process. Scheduling is used to allocate plant and machinery resources, plan human resources, plan production
processes and purchase materials. In this case, it is the determination of a production sequence, manufacturing runs and times, and the route to be followed. At the same time, scheduling takes into account the
constraints on production activities and resources in such a way that the goals, e.g. no deadline violation,
are realized (Drexl & Kimms, 1997, Herrmann, 2006). Supply-chain logistics is characterized by a large number of interdependent variables. For its transformation network models Erenguc et al. (1999) list: products,
stages, minimum and desired stocks, time periods as the most important ones at this stage, given the predecessor relationships among all supply-chain stages. Proper scheduling has real economic significance for all
manufacturers, and they are extremely interested in taking advantage of it to suit their needs.
Given problem can be described as Capacitated Stochastic Lot-Sizing Problem (CSLSP) with intermediate
storage, that deals with scheduling for a stochastic production system with multiple items, for which set-up
times and setup costs occur when a production is switched from one product to another (Kämpf & Köchel,
2006, Drexl & Kimms, 1997). The issue is to find a sequencing and lot-sizing rules to maximise performance
(in terms for e.g. maximizing profit). The solving of CSLSP usually leads to mixed integer programming formulations, which can only be solved efficiently by the means of heuristics, i.e. either there is no analytical
solution or it is too expensive and does not handle stochasticity well. According to Mula et al. (2006) scheduling problems in production are optimisation problems and are very complex, requiring new approaches to
manage uncertainty. There are a few criteria that drive the choice of production scheduling, depending on
the level of aggregation. In this case Min & Zhou (2002) names six decision variables. These are: allocation of
resources, number of stages, service sequence, volume, inventory level, workforce and extent of outsourcing.
On the other hand, Erenguc et al. (1999) reduces all this variables to monetary value beforehand and optimises costs as a whole. This approach reduces complexity for optimisation itself but increases uncertainty of
transitions to single dimension.
Scheduling can be also incorporated with an enterprise resource planning (ERP) system, as an attempt,
as stated before, to integrate various branches of the supply chain. Such an attempt is made by Józefowska
& Zimniak (2008). They first point out that it is important to take into account the aspects of capacity constraints, backlogging, set-up costs, sequence dependence, number of production stages and possible number of products. Only then one can endeavour to solve the problems, which depending on above parameters
are called differently, and the approach to tackle them should change. Developing heuristics to find nearoptimum results fast is suggested, to help solve multi-objective combinatorial optimisation problem at hand.
Despite not incorporating stochasticity Genetic Algorithms approach is used, mostly because the model has a
much broader spectrum and includes not only production but also costs, in and out stocks as well as multiple
products and machines (Józefowska & Zimniak, 2008).
Another approach, that does not require optimisation and fits well with simulation, is scheduling via
dispatching rules. When lot-sizes are fixed, and there is no desire for costly optimisation, in many cases
dispatching rules can be a good enough substitution (Jeong & Kim, 1998, Pinedo, 2008). Blackstone et al.
(1982) describe dispatching rules as logic, according to which the next job is chosen from a set of jobs awaiting
service. While this logic can be simple or complex the drawback is that these rules are applied independently
among workstations and only within the logic prescribed to them. For complex systems, introducing highly
2.6. L ITERATURE R ELEVANCE
13
sophisticated dispatching rules, taking into account the state of other, relevant stations might not be feasible.
Moreover, as Blackstone et al. (1982) mentions, no single rule can be identified as best in all circumstances,
and lists a comprehensive set of dispatching rules for job shop operations in the paper. Interestingly, these
rules can be also intentionally violated, using dispatching heuristics, to e.g. insert idle time to a station to
await for a high-priority job. Another example of such application is pre-emption, i.e. interrupting a task if a
more suitable situation can be achieved.
Job selection can be also solved via job sequencing, which is ordering all awaiting jobs in a queue, while
dispatching rules only choose a single next job. Job sequencing is useful especially when the set of awaiting
jobs is fixed, without new arrivals, which are then processed by a single server (Blackstone et al., 1982). Lee et
al. (1997) distinguish two methodologies for that: constructive heuristics, which builds up a non-reversible
queue of jobs, and improvement heuristics, that tries to optimize the set either via local search or guided
search of the sampling space. The constructive heuristic is thus similar to dispatching rules for a fixed set of
jobs. Selection of dispatching rules might be in the end left to an experienced planner, who knows the system
and uses such scheduling tool for support (Herrmann, 2006).
2.6. L ITERATURE R ELEVANCE
Desk research done in this chapter, together with performed interviews, allows to conceptualise how to constrain and model the given problem, and elaborate more on the current state of knowledge about the issue.
Clearly, there are a few gaps that need to be filled to fully understand the subject, especially involving the
precise nature of cross-contamination and its impact, but also the effect facility layout can have on pollution.
Then, understanding how material residue is released during processing is missing, and thus it is not yet possible to express it with a mathematical model. As no DES model for cross-contamination could be found, a
question arises whether this method can be suitably well used for its representation. Moreover, it is unknown
whether there are any trade-offs involved in the process because of cross-contamination, or what could they
be exactly. In the end, literature is used to help develop the cross-contamination models in section 4.4 and to
relate to the experimentation results in chapter 7.
3
M ODELLING S PECIFICATIONS AND
R EQUIREMENTS
This chapter describes the implications arising from the introduced research problem and choices made, in
order to come up with tools to help answer the research questions. In the first part, modelling boundaries
and constraints are listed, and then the theoretical model is constructed.
3.1. I NVESTIGATION B OUNDARIES
As the problem environment – multi-product manufacturing is put in general terms, there is a need to constrain it. Cross-contamination is an issue for all multi-functional production lines, but for some it is considerably greater, depending of multitude of factors like equipment properties: internal area, volume, line length,
angle of inclination, internal shape, surface texture, processing speed and others. Moreover, it is highly dependent on the type of handled products, whether these are liquid, bulk particulates, discrete solid parts or
assemblies, and on the technique of production - batch or continuous. Product properties like material type
and quantity, particle size, shape and number, cohesiveness, friction angles, are also vital components. It is
not possible to investigate all of them in a single research project, and thus a representative of the larger class
of problems is chosen. Investigation into continuously handling loose, dry material in form of powder can
serve as a good enough example for the group with relatively high pollution effects (Strauch, 2003). Then,
cross-contamination happens not only in particular machines but without break along the entire route with
varying intensity, especially including transportation and storage equipment. In addition, only short term
repercussions of cross-contamination are studied, where material mixes in predictable way, excluding rare
events of long-term residue release.
Production process might be quite complex with alternative routes and possible multiple, varying number of stages per product. Additionally, there could be repeated processing on the same machine, rework,
exclusions and other aspects, making it a very compound problem from scheduling perspective, and blurring the insight into cross-contamination in between (Mapes et al., 1997). Therefore, in order to understand
the complexity of cross-contamination, the process should be relatively simple but allowing for scheduling
and layout interventions. Figure 3.1 depicts the chosen example of process to investigate further. It starts
with material dosing, where automatic equipment applies several ingredients from storage, that is not within
the scope of the inquiry, into a single, main mixer. Mixing machine distributes the components evenly and
creates a homogeneous powder, which is then sent to an intermediate storage buffer – a silo. There is a limited number of the silos with varying capacity and connections and the mixed product can be sent to more
than one. Subsequently, there are several bagging machines that package the ready products, to which they
are transported with a network of automated conveying equipment. Carry-over occurs then in every interval of the process and so can cross-contamination, depending on the route chosen for a given product and
previous occupants of this path.
Additional constraints, specific to equipment, products or job organisation can be applied to increase
the realism of the process. Moreover, pollution gathered before the mixing can be incorporated into the
analysis and accounted for. The resulting production scheme can be described as scheduling and sequencing
problem with limited intermediate storage (silos) in which all the products go through the main mixer to be
14
3.1. I NVESTIGATION B OUNDARIES
15
Figure 3.1: A diagram of a representative process for the chosen problem
dispatched to one of the silos and then await the capacity for bagging (or bulk loading).
Every multi-product manufacturing process comprises of orders or jobs that need to be scheduled and
then handled (Herrmann, 2006). This highest level list contains information of the product (normally there
is only a single type of product per order), its quantity, identifier and due date. Every job has to have at least
one product assigned to it, and usually more that physically moves through the system. For bulk processing
items can be discretized into small batches and processed as individual units of homogeneous composition,
small enough to withstand that assumption for large processing amounts. Furthermore, every product has
a recipe, or unique specified component composition, that can identify counterparts that are packaged into
different size of bag (or released as bulk), or possibly have different contamination rules. These are based on
nutrients and sensitive components where every product for every nutrient has particularised specification
level as well as lower and upper acceptable thresholds. The number of such rules can differ per system, and
they can be easily compared to their equivalents in another products.
Normally, production planning is done on three levels: strategic, tactical and operational (Erenguc et al.,
1999). This investigation concentrates mostly on the tactical one with some considerations about the strategy,
and names possible operational uses with further development. In order to constrain the investigation and
focus on the most important aspects of the problem several boundaries are set:
•
•
•
•
•
•
•
•
•
•
Model begins at the end of the feeding system to the mixer,
There are four types of equipment involved: mixer, bagging machine, silo and conveyor;
Dynamic object classes include orders, product batches and bags;
Possible contamination before the mixer (e.g. in the elevator) is not taken into account directly, but is
included together with one occurring in the mixer;
Raw materials (ingredients) are always available and the finished products are removed and stored independently,
Total amount of residue in the system is roughly constant so the contamination is an equal exchange of
the manufactured product with the equipment residue,
Cross-contamination is independent of the type of transported material. It varies with equipment/pipe
type and contents of materials participating in the exchange,
Products with different recipes cannot be in the same place,
Backlogs or rush orders are not taken into account,
The model does not investigate any further than the bagging lines and bulk stations.
These model boundaries are set to concentrate on investigating the research problem, and to exclude nonvital system parts from scrutiny. Based on them, modelling assumptions are made that are used for model
implementation.
16
3. M ODELLING S PECIFICATIONS AND R EQUIREMENTS
3.2. M ODELLING C HOICES
Several modelling choices are made to describe the system in more detail and prepare for implementation.
First of all, it is vital to split mixing and bagging to allow for taking advantage of the system’s flexibility. Since
due dates only apply to the product leaving the system and shelf life is not an issue, product can be mixed
any time beforehand and be ‘suspended’ in silos until further processed. Then, orders can leave the system
no earlier than the set due date from planning.
System can be therefore imagined as a push-pull construct. The main mixer prepares the recipes and
tries to allocate them in a silo. Mixing process can only start when there is sufficient silo capacity available,
and any time the mixing and storage capacity are available, there should be material ‘pushed’ to the silos. As
such, the amount of work in progress (WIP) is of no concern, providing there is sufficient material available
to be bagged. Once the material reaches the silo it can be ‘pulled’ by an appropriate system exit with available
capacity, providing the transportation line is not blocked. If there is always material available in the storage,
it is up to the exit points to decide which order to process next, that is to be included in the model logic and
scheduling rules.
Another important choice is to assign every material batch to its order and move it through the system
in a FIFO manner with no passing by, and cross-contamination is also to work in this way. It is to assure
proper sequence of orders, though can be limiting, e.g. in case where there are two or more products of
the same recipe in a silo, and only one in a middle could be processed at the moment, due to availability of
equipment. On the other hand, it ensures that the first product collects most of the (possible) contamination,
and is useful when splitting orders among multiple silos. Given a possible large number of entities in the
system, all materials in containers i.e. mixer, silos and buffers are to be batched to a single entity but retain
their references and contamination data. Calculation of cross-contamination is then done after each interval
with only single possible route, distinguished in a model by separate objects (one object - one contamination
calculation per product batch).
3.3. R ECOGNISING E XCESSIVE C ROSS - CONTAMINATION
Classifying whether given occurrence is cross-contamination, and determining if the outcome product is too
largely affected by it, is a vital step to take measures in order to limit it. Assuming no cross-contamination
Figure 3.2: A causal diagram of cross-contamination process
3.4. M ODEL S TRUCTURE
17
is allowed should be treated as an unrealistic idealism, impossible to implement for larger quantities, even
despite multiple cleaning treatments. Although pharmaceutical industry manages to limit cross-contamination, which is an absolute must for drug production, other industries, such as food processing, are not as
thorough with cleanliness norms. For example trace amounts of peanuts can be found in other products,
even after considerable following material flow, which can cause a serious threat to allergy sufferers (Taylor &
Baumert, 2010).
Assuming all products are subjected to unintended mixing along the route, the outcome composition depends on what is being mixed and how. Most commonly, it is between different portions the same product,
and then it does not matter how much product was involved and how it happened, as it has no consequence
to the resulting structure. But when there is another product involved, the quantity and path are vital. Differentiating among neutral and significant processes for a model is difficult, though being able to do so could
save a lot of processing power. Therefore, all calculations done for cross-contamination are performed regardless of material and residue structure. Moreover, the idealised situation of mixing just two products is
hardly ever going to happen. Diluted residues of products remain in the system for a long time, until their
amount in a given place can be considered so small, that can be omitted (Fink-Gremmels, 2012).
A causal diagram for cross-contamination is presented in Figure 3.2. There are two main factors, except
for equipment properties, influencing the resulting product contamination in the outgoing products: initial
material residue and contamination in incoming product. In general, it is expected that the more product is
processed, the more of its residue remains in the system and the less of other products (Leloup et al., 2011).
When a finished product is contaminated, it does not necessarily mean it is out of specification, though
there is a higher chance for that. It is the ingredients, especially the sensitive ones, that are vital to determine
whether it is within requirements (Fink-Gremmels, 2012). The number of product ingredients is generally
vast, and it is difficult to handle them all. Therefore the approach taken is to include necessary nutrients,
combining them with relevant other, usually trace components, to which the product can be susceptible.
By giving specification limits to all, and comparing with the actual content, final determination of quality in
terms of cross-contamination can be made. This way, minor addition of some might result in rejection of a
given item, while large intake of another, similar product might be neutral to the outcome.
3.4. M ODEL S TRUCTURE
A simplified Unified Modeling Language (UML) class diagram is presented in Figure 3.3. It contains the most
important interactions and variables, based on the aforementioned boundaries and assumptions.
The most important relation in the system is one of product batch – place, that defines material handling
logic and applies cross-contamination calculations. Every product batch has a parent object - its order, defining type of product, quantity and a due date. Product batch also has a recipe, limiting the maximum dosing
speed and identified by its number, and several different items may have the same recipe. Furthermore, products should abide contamination rules, which specify the particularised nutrient level and its minimum and
maximum values for the finished product. They can be also empty if such levels are not specified.
Product batch always occupies a certain place, one of four distinguished: a mixer, silo, bagging machine or
conveyor. Every place has a certain capacity limiting the number of product batches occupying it and the total
residue parameter used for cross-contamination calculations. Moreover, every place keeps information on
the current product residues in it and their quantities. Clean product batches, with original recipe content of
100%, might become contaminated along their route by the residue in places they go through (by the process
of residue-product mixing), keeping the information in the contamination table. As such product batches
contaminate each other via an indirect interaction in a certain place. Material content from products of the
same item number is not considered contamination, while any other additives are recognised as pollutants.
Contamination is only considered extensive if the nutrient levels are breached for the sum of product batches
in a bag leaving the system. Depending on the content of a place and a batch, the total contamination can
either increase, stay the same or decrease.
Conveyors are places that connect places with each other (so a conveyor can connect two conveyors as
well). They transport material in FIFO manner according to their fixed speed along fixed conveyor length.
Conveyor types, screw or pneumatic, are distinguished to calculate their capacity and conveying speed. An
air filtering station, placed at the end on pneumatic conveying system and before a bagging buffer, can be
also viewed as a conveyor with zero length that only performs material exchange for cross-contamination.
A silo is an object that holds product batches for an indefinite amount of time, i.e. from the moment they
enter until they can be processed further, which does not depend on the silo. Contingent upon product batch
18
3. M ODELLING S PECIFICATIONS AND R EQUIREMENTS
destination the discharge speed from a silo may vary. Discharge is a process of sending material form a silo
to its destination, subjected to various model constraints. Another important function performed by a silo
is checking whether material from a single order is split among multiple silo, and attempting to merge its
discharge when the silo runs empty of a given order child objects.
Figure 3.3: A simplified class diagram for the model
A mixer is a similar piece of equipment to a silo. It also holds multiple product batches and delays them
before sending to their assigned destination (a silo). But the discharge logic is subjected to the chosen mixer
output as well as material level monitors, that can hold off discharge if breached.
Finally, bagging is responsible for grouping product batches together, and putting them into bags, which
can hold several of them. Product batches are stored in buffers, from where they are put together in accordance with the machine processing speed and product batch package size. In the end, whole contamination
in a bag is added up, and recalculated into nutrients to be compared with the nutrient minimum and maximum levels to eventually determine whether the bag is within specified limits or not.
The class diagram from Figure 3.3 is thus a vital specification element to be used further in software implementation.
3.4.1. A NALYSIS C RITERIA
Cost analysis is not within the scope of investigation, therefore it is the line performance, mostly in terms of
capacity, flow times and contamination, that is to be examined. All derived useful performance indicators
3.4. M ODEL S TRUCTURE
19
are listed in section C.3. The following set of high level Key Performance Indicators (KPIs) is used to compare
different scenarios with each other, often joining them to express in more general term (e.g. throughput):
Important equipment:
• average daily throughput
• average daily throughput
Manufacturing orders:
◦
◦
◦
◦
number completed
number completed on time
average contaminated order ratio above limit
ratio of material above limits to total processed
[bags]
[kg]
[-]
[-]
[-]
[-]
Thus, all comparisons among scenarios defined in chapter 6 are done using (some of ) the above six parameters.
3.4.2. S IMULATION A PPROACH
One of the most important issues to point out, is that the production in question can be best described as
continuous, rather than batch, because of handled material properties and its flow through the system. This
is so because the mixing process, after an initial setup, gives out a steady flow of mixed product. And that
continuous process is to be represented by a discrete event simulation. According to Kelton et al. (2013) the
modelling steps can be divided into seven parts:
• Problem formulation,
• System and simulation specification,
• Model formulation and construction,
• Verification and validation,
• Experimentation and analysis,
• Presenting and preserving the results,
• Reflection.
These are to be followed in search for a well-founded approach to model the problem and represent it in
the Simio package. As the research problem is stated and methodology chosen in chapter 1, this chapter
continues with system specification. At first, the specification of the modelling problem in more detail is
made, followed by an identification of constraints and boundaries as foundations for the further steps. These
are vital to represent the issue at the right level and only including relevant features, as in general such systems
are too complex to be modelled fully, and unnecessary features only obscure its understanding. The following
stages continue according to the list, where a model formulation is done in the next chapter. Then, the final
simulation model is defined and validated, as the essential tool to discover possible consequences of certain
interventions (chapter 5), and following experimentation is described in chapter 6, its analysis in 7, and is
finally concluded in chapter 8. Detailed tables with experimentation results can be found in Appendix G.
4
P RODUCTION M ODEL
Production model is the basis of the project that provides the fundamental functionalities of the simulation. This chapter explores the translation of a real-world manufacturing plant problem into a discrete-event
simulation model, as a necessary means to explore impact of selected specific interventions on production
performance. In the following sections, the model is complemented with cross-contamination calculations
and dispatching logic.
4.1. P URPOSE
In order to explore possible performance of different scheduling alternatives as well as to determine the extent of contamination in the products, the production system within investigation boundaries needs to be
represented in a DES paradigm. For this purpose an object oriented simulation tool Simior from Simio LLC
is chosen as a powerful and commercially well-established program. Production model as such is meant here
as an ‘engine’ that creates and directs the flow of material, and steers the execution logic in a manner similar to
the real-world system. With parametrised properties and hierarchical structure, the final model is to consist
of several parts that harmoniously perform necessary production steps, i.e. create product batches (entities)
in specified intervals, and send them to their destinations with relevant delays for mixing, storage, waiting
and transportation, to end up in one of the system exit points as according to entity properties, be handled
there and leave the system as finished product. Eventually, the production model contains four sub-models:
Mixer model to represent mixing delay, control discharge logic and material destination,
Silo model to store material in a designated space and discharge it to a specified system exit when available,
Bagging machine model to combine material batches into bags according to particularized characteristics,
Conveyor model to act as transportation delay of specified capacity.
4.2. A SSUMPTIONS
On top of the system boundaries and assumptions specified earlier in section 3.1, there are several additional
assumptions made to help model the investigated problem at the chosen level, disregarding unwanted details. Thus the following assumptions are made:
• Dosing and mixing are always exact and according to specifications,
• Mixing is completely uniform, and fits exactly to the specifications (excluding contamination),
• Ingredient mixing is not to be modelled – products enter the system with specification contents to be
only delayed in the mixer,
• All material flow considerations are expressed in terms of mass flow rather than volume,
• All material is represented in batches (entities) of equal size,
• Workforce availability is modelled as a schedule with possible overtime,
• Model is to have modular construction to allow for future expansions.
20
4.3. M ODEL C HARACTERISTICS
21
An important choice is not to include considerations on the uniformity of mixing, and thus not to model the
mixing itself. Although it could be very interesting to look at the cross-contamination issue when products
might deviate from their specification values, by either non-homogeneous mixing, imprecise ingredient dosing or segregation, at this stage of knowledge about the phenomenon, it could only blur the outcome. In the
presented model any deviation from the norm is a result of cross-contamination and thus is easily recognised.
Moreover, to assure consistency and ease of the modelling of cross-contamination, all material flow is to
be divided into batches of equal size (mass), in order for the discrete simulator engine to process them. Discretising entities is vital to record specific statistics as well as to allow for cross-contamination investigation.
The size of an order to be mixed next is divided by the size of a single product batch to determine the number
of active entities to insert into the system. They enter via a dosing feeder in intervals with restrictions of the
automated system (recipe specific), and this way arrival of new products is solved.
Equipment involved, that needs to be manned in order to operate, is given a capacity schedule. Thus,
without modelling the specific workers, the limitations of working day availability can be incorporated. Additionally, it is possible to set a working overtime to be used when not all orders for a given day are finished.
Finally, to support the generic nature of the structure, the model needs to be modular, also allowing for
future expansions. Creating a full model of a system is then done by including chosen equipment, joining
them with conveyors, and assigning specific parameters and references. Then, on the highest level, material
flow logic between the main object instances needs to be incorporated.
4.3. M ODEL C HARACTERISTICS
In the model two major parts can be distinguished: object definitions and flow logic description. Both are
shortly described in the following sections. Additionally, characteristics of more specific model features are
given next.
4.3.1. O BJECTS AND S UB - MODELS
The model is divided into several important object classes and components that execute the logic and steer
the flow of material through the system as defined in Figure 3.3. Important object characteristics from the
point of view of production capacity model are described below.
E NTITIES
There are three types of entities (dynamic objects) in the system: orders, product batches and batch entities.
First represent the actual production orders, containing essential information about it and gathering important performance indicators. They appear in the system in groups of orders for a given day, are rearranged
according to scheduling rules and wait their turn for mixing. Once that happens, child entities – product
batches, are being created to express the flow of material through the system. As such, order entities are just
logical objects, that have no physical real-life equivalent, performing function of handling and translating
unordered data from table into schedule conforming to the set rules. They also perform an important role
in distributing their child products through the system, as they hold information about designated system
exit and according to that are assigned a list of possible intermediate storage silos. Product batches leave the
system as soon as they are bagged, while the order entities only after the last child entity has been packaged.
Batch entities are containers used to group products together (e.g. in bags or silos) and hold information on
their total contents. A comprehensive flowchart of order entity logic can be found in Appendix, in Figure E.1
on page 102.
D OSING AND THE MAIN MIXER
Dosing is the source of product batches in the system for the main mixer. It creates them in intervals in
accordance with specified dosing speeds that differ among recipes, and conforms to model logic imposed by
the order entities and the main mixer. Dosing can be only started if there is sufficient intermediate storage
space available for the entire order. Thus a situation that an order was started and there is no equipment
to receive its product cannot occur. Product batches are directly getting to the main mixer that delays them
accordingly controlled by the level switch logic. When leaving the mixer an entity is assigned its intermediate
storage destination according to possible ones held by the parent order and the current fill level of the silos.
The main mixer can be thus viewed as a push system that sends out material to possible silos, trying to
get rid of it and to be able to start the next order as soon as possible.
22
4. P RODUCTION M ODEL
C ONVEYORS
There are two types of conveyors in the system: screw and pneumatic, represented with a single sub-model
based on a time-path construct, and a switching value is used to assign specific parameters and limit amount
of repeated data. They transport motionless product to assigned destination in FIFO manner, without passing-by. Both are accumulating and have fixed, theoretical speed.
Screw conveyor uses rotational movement of an inner helical screw to transport material within a tube
in a roughly linear trajectory, and is a typical mechanism used in transportation of bulk products, especially
upwards. It is often used for high capacity lines, when large volumes need to be moved. Assuming there is no
slip, the resulting material speed is a product of the rotational speed and the screw pitch, and the maximum
volume of material inside is close to the volume of the conveyor. When the angle of inclination to the ground
is bigger than zero the performance of such conveyor decreases, so in order to account for that different
mechanism properties or higher rotational speed is normally chosen.
Pneumatic conveyor, on the other hand, takes advantage of compressed air for material propulsion. A
separate inflow pushes in air that snatches powder, and quickly moves it to the destination where it is filtered
before further processing. Particle, once it is accelerated, has a velocity similar to the air, but the piping has
a typically much lower diameter than screw conveyors, and thus can hold less material. Also, the piping can
hold much less material, as with excessive quantity it might become clogged. Moreover, too high air velocity
can damage powder particles and thus cannot be used. Pneumatic conveyors end with an air filter to separate
powder from the air, which cone be modelled also as one with zero length and a different contamination
characteristic.
S ILOS
Silos are storage units that hold entities, batching them together and always utilising the First In First Out
(FIFO) principle. Having a fixed volume, they can always hold the same amount of uniform density product.
Inflow to silos is unrestricted, providing there is enough space and the outflow rate depends on the destination. Outflow valve can be closed at any time if the amount of discharged material in the bagging buffers
exceeds predefined limits. Additionally, silos hold information about the products that are currently in, and
perform a very important function in the model logic, as they are involved in allocating material from the
mixer as well as discharging it to system exit points. The logic preventing different products to occupy the
same conveyors (links) is embedded within silo processes, so that a discharge will not commence while there
is a parallel material flow. A flowchart with description of the silo logic can be found in Appendix in Figure
E.6.
B AGGING
Bagging bases on a pull principle to choose next material to package, and it only accounts for the current
system state to make such decision (no prediction). At the beginning of a shift, when there is a new inflow
to a silo, and when a machine finished its changeover, model logic performs a search on which product to
process further. It bases on system state and boundaries, product constraints and scheduling rules (described
in section 5.5). A flowchart presenting how a given machine picks orders is shown in Figure E.12 on page 113.
A packaging machine takes material from a bagging buffer, located just before it, of a quantity that is corresponding to the package size. These entities (at least 2) are batched into a parent entity – a bag, specially
created to hold them and move further as a single entity, after a processing delay. Bagging speed is modelled deterministically, which is a close enough representation of a numerically controlled machine without
failures. To reduce this speed and provide a more realistic performance of the equipment short stops are
introduced, that limit the speed indirectly by suspending some of the processed items for a short time.
Normal operation of a bagging machine during shift time involves processing an order (including short
stops), then a changeover, and a following product. Sometimes in between products, after a changeover,
there is a break in product supply, resulting in machine starving time. An example resource chart for bagging
is shows in Figure 4.1.
4.3.2. M ATERIAL F LOW L OGIC
Material flow, or specifically movements of differentiated substance batches, needs to follow a clear, customisable progress logic. In this case, material enters the system via a single source, and never returns to the
place, that it has already been to before. Moreover, for a given product batch, it can only be placed in a silo
once. Then, flow logic needs to assure that in a given place only product batches with the same recipes can
reside.
4.3. M ODEL C HARACTERISTICS
23
Figure 4.1: An example resource chart for bagging machines utilisation, grouped by package size
Negahban & Smith (2014) name DES as an appropriate tool to analyse and understand the dynamics of a
modern manufacturing system, the vital part of which is often a material handling system (MHS). It is used
to solve problems regarding movement, storage, control and protection of goods throughout the process. It
can be thus understood as the principle logic steering the progression of material in the entire system, which
in this case is mostly:
• Choosing when to start and dose a new order,
• Delaying the material appropriately in the mixer,
• Controlling discharge from the main mixer to a previously chosen silo,
• Keeping track of the material in the silos,
• Allocating the silo discharge to an appropriate, available system exit,
• Assigning product progression speed as a resultant of the product and equipment properties,
• Recording important data about the performance.
In the end, as described by Schriber et al. (2012), the execution of a DES model is characterised mostly by
assigning entities into appropriate states, and handling the consecutive event list as well as delay lists. Thus,
one can imagine material progress as a set of queues and delays on its way in between starting the simulation
run and leaving the system.
However, having fully described product and system properties as well as established order set, does not
complete the execution logic. In any manufacturing circumstance, orders need to be scheduled and appropriate resources allocated to allow them to progress. This can be done at random or any other way, but usually
follows a certain approach, aimed to be beneficial in some aspect to the production process. This way has to
be clearly defined, and preferably customisable to determine, whether (or how big) scheduling interventions
have effect on production performance.
4.3.3. I MPORTANT F EATURES
Next to the main objects and flow logic there are a few more interesting and important characteristics of the
built model that are defined below.
Work schedules
The most important equipment is given a customizable work schedule feature – an ability to define specific
working hours separately. The standard operation assumes 5-day week with 16-hour shift and no breaks but
it can be changed at will, for instance taking some machine off-line or reducing (extending) its shift. The
intention is to mimic availability of the workforce while not specifically modelling workers. Thus, a work
schedule is defined as a resource that has to be usable in order to proceed with work. Dosing and bagging
machines are equipped with this feature, and it is assumed that supervision over the entire system is performed by the same workers as for the dosing. Logic used to make a decision on ending shift for an example
bagging machine is presented in Figure E.13 on page 114.
24
4. P RODUCTION M ODEL
Order possible destination search
Before mixing can start, the next order in line has to have guaranteed available space in the intermediate storage to fit the entire order. But the way which of the available intermediate storage silos are sent material to,
should be efficient and subjected to user defined logic i.e. scheduling rules. More about these can be found
in chapter 5.5. In practice, every order needs to undergo a process of checking this availability, picking the
best silo(s) according to defined rules and then start dosing. A flow chart of the procedure is presented in
Figure E.9 on page 110, and its core involves two logical queues - possible destination queue and arranged
order allocation queue. First contains all silos that are present in current configuration of the system, excluding those that are switched off by user and those which were reserved for specific recipes (another user input
feature is possibility to constrain material in chosen silo to particular recipe). From the possible destination
queue in the chosen sequence by scheduling rules silos are moved to the order allocation queue, and once
that is done there is another attempt to reorder this queue, if more appropriate arrangement can be found in
case of split orders among multiple silos.
Splitting and merging material
In some cases material can be sent to multiple silos. This happens when an order is too big to fit to a single silo,
or when it is preferred to join identical recipes and the remaining capacity of the silo is not sufficient. When
silo high level mark is reached the material from the main mixer is sent to the next silo in the queue. After
the second is full, the list is always searched from the beginning (in case some material was removed from
the silo in the meantime), and the chosen destination follows the queue till the order is finished. Material
that is then placed in multiple intermediate storages should be then bagged as a single order to avoid lengthy
changeovers. Thus, a search function is executed any time an order finishes in one silo, but only when globally
not all of the material has been processed. Then the flows are smoothly merged, unless the route is blocked or
is not the first one in line. If it is impossible to merge the order, bagging machines does not wait but executes
a changeover and proceeds with another order, to finish the split one later.
Data collection
An important feature of any simulation model is gathering suitable and comprehensive data about the performance, to be later processed and appropriately presented (van der Zee & van der Vorst, 2005). In order to
allow for communication between Simio and SN text files are created to store the data. Then Simio inbuilt
write to file step is used to save important parameters after significant events or periodically in accordance to
timers. KPIs defined in section C.3 are thus obtained in a manner that is suitable for further data processing.
4.4. C ROSS - CONTAMINATION I NVESTIGATION
Product cross-contamination is a serious issue for quality when the same equipment is used for more than
a single product. It happens when a given product enters a piece of equipment, sweeps some of a leftover
residue released by the previous product, and drops some of its own behind. In the following sections an
investigation into powder cross-contamination is performed, where a generalised mathematical model for
such process is derived and discussed. The analysis and implementation is then conducted based on the
case of Sloten in section 5.4.
4.4.1. M ODELLING C ROSS - CONTAMINATION
Every product leaves some of its contents behind, and picks up whatever is in the system. This means there
is a considerable extent of accidental static mixing in the system, as the flow of material proceeds through
the piping. Parts left by a product are called residue, and only become contamination once picked by the
following one, and only if it is not a specified product content. It is assumed that the portfolio of products
sent through the same network is rather similar. Known differences among products are considered negligible
and any product is assumed to have exactly the same behaviour.
Investigated powder has a very fine grade, leading to enormous number of involved particles, the simulation of which is beyond the capabilities of modern computing. To assure efficiency but not hinder the
quality of analysis, moving material is grouped into fixed-sized batches (segments), that are assumed to have
homogeneous composition and always mix with the equipment residue, also divided into compartments,
proportionally to their contents. The approach is thus similar to one proposed by Portillo et al. (2006). The
material starts uncontaminated at the dosing machine, i.e. having only its original, intended composition,
and exchanges some of its contents for different products via mixing with the residue of equipment it is pass-
4.4. C ROSS - CONTAMINATION I NVESTIGATION
25
ing through. The size of a batch remains constant but the composition, in terms of comprising products,
changes along the path.
Once the product is bagged and ready to send to the customer, the analysis on its nutrient specifications
and limits needs to be performed. Product batches making up the bag contents are grouped together and
assumed homogeneously spread in the bag, so that sampling is not an issue. Because some of the products
picked up by the batch are more similar than others, a means of comparison between them is necessary,
in order to determine whether the cross-contamination is acceptable or not. As normally the number of
relevant ingredients comprising products is considerable, expressing the final bag contents in terms of them
might be counter-productive, leading to even greater complexity. Moreover, comparing similar ingredients
(e.g. different concentrates of the same material) would still be impossible. Thus an approach to compare
different products with each other is limited to nutrients with addition of relevant, recognised contaminants,
collectively referred to as nutrients. This way, a list of important, company-specific elements can be identified
and assigned to each product. By adding lower and upper limits for each of the nutrients, the desired levels
of these can be checked and controlled. Then, after bagging the product batches, the model is to add up
all nutrients from comprising contents and compare each of them with specified limits. The number of offspecification bags is counted and recorded, to allow comparison with other products but especially with
another scenarios, having different scheduling rules.
In literature, product contamination is usually connected to trace elements such as microbiological or
chemical agents, that can be potentially harmful after consumption. Some adverse cases of animal diseases,
such as Bovine spongiform encephalopathy (mad cow disease), originate from such contamination, by prions in this case (Fink-Gremmels, 2012). These occurrences are not in question in this analysis as they should
be a part of thorough risk analysis and laboratory trials, because they cannot be quantified and expressed
as a regular part of the production process. However, unwanted pollution of one product by another (crosscontamination), as a result of production method, where the total amount of material introduced into the
system is known, can be quantified in search for the answer of how long and how much specific residue
remains in the system. In general there are three sources of norms that dictate what is considered as unwanted pollution: internal company norms, legislation and customer demands. These norms can be strict or
conveyed as a preference, without major impact. In any case, all added nutrients are grouped together and
expressed with their own units of measurement and only compared to each other. For various plants and
industries different ingredients will be considered significant in such analysis. These are either main ingredients creating the nutritional value or trace amounts of other, sensitive because of aforementioned reasons.
Small bits of other additives can be considered irrelevant and excluded from the analysis.
D EFINING CROSS - CONTAMINATION
Cross-contamination character is specific to the investigated system. Although detailed simulation analysis
of particle movement using DEM, based on powder and equipment properties, could be possible, it would
not be feasible to include it as a part other analysis (e.g. throughput, scheduling), mostly because of computational limitations (Bertrand et al., 2005). Thus, the powder behaviour is to be investigated based on field
trials, its characteristic plotted and generalised for any situation. This section presents one of the methods of
such research, well established in the industry, including two products with a tracer material.
A typical process consists of two consecutive stages: first polluting the system with sufficient amount of
material, including high amounts of the tracer, then sending collector material (without the tracer particles)
through the same route, and taking samples in multiple points over time, so that the amount of tracer material
as a function of time can be expressed. Assuming the tracer to be homogeneously mixed in drawn samples,
one can proportionally establish the amount of the initial material and thus the level of cross-contamination
in the samples with reference to some base level. According to literature, the behaviour of contamination
over time is exponentially decreasing (Leloup et al., 2011), and by integrating the established curve over time
one can determine the total (initial) amount of residue material in the measured intervals. The change in
this amount constitutes for how much is swept by incoming product with respect to the constant base level.
However, expressing incrementally from point to point how contamination changes is not straightforward,
as reference levels change continuously and the exponential character of the curves is lost. In such a case,
it is expected that without additional mixing effects the level of contamination in respective samples can
only increase for consecutive measurement points. To illustrate it, if contamination level measured after 100
kg flow is 5% at point A, then for point B being further along the route, the corresponding contamination
after 100 kg flow should be no less than 5% and possibly more. However, due to trial conduct, additional
(unaccounted for) mixing effects and method uncertainty, this amount could actually be lower. Therefore
26
4. P RODUCTION M ODEL
it is more convenient to express relation to the base level, even though the change in contamination and
its calculations need to be incremental in a model. The description of the actual trials performed in Sloten,
Deventer can be found in Appendix C.4.
The character of cross-contamination and resulting exponential curve can be explained as a random
chance release curve. More specifically, it can be connected to axial dispersion curve in chemical reactions
in fluids, defined as a dynamic but reproducible blending of sample area with a reagent and/or carrier, as
a result of turbulent flow patterns, created by fluid going through a narrow-bore tubing1 . Models for axial
dispersion are common in chemical engineering, (see e.g. Liao & Shiau, 2000).
Another part of the analysis is extending the tracer-collector scrutiny to multiple products, that could be
involved with cross-contamination. At this point, the requirement for homogeneity and proportionality of
mixing is vital, especially that no trials with more than two products can be found in literature. Even though
these assumptions are not exactly exact, they are certainly sensible. Multi-product contamination possibility is necessary because of some trace elements, which are important from the product safety perspective
and which need to decrease by a large factor. Moreover, as the consecutive products are often allocated into
different silos, taking alternative routes, the primary contaminants in various places differ, so the final batch
composition must comprise of multiple products. Proportionality of mixing is maintained with regards to the
current residue composition of a certain piece of equipment, and the contents of the product batch coming
through it. This means that any amount that is mixed, which can be only partial, takes equal amounts of
material from both, distributes homogeneously and attaches the mixture back to them, changing their composition and conserving the mass balance. As the products move in FIFO manner, the calculations can be
discretized and performed when the batch leaves the equipment.
C ONTAMINATION ASSUMPTIONS
Based on the analysis of the trials the following assumptions are made for contamination investigation and
modelling:
• Material always moves on First In First Out (FIFO) basis,
• There is no loss of mass in the system,
• Material in equipment and product batches is always homogeneously mixed,
• Total amount of residue in equipment is constant,
• Cross-contamination is a process of equal exchange of residue between the equipment and the product
batch coming through it,
• Contamination in a product batch is always expressed as a mass content of different than nominal
product in it,
• Cross-contamination behaviour is independent of the type of transported material and only bases on
equipment properties,
• Resultant material composition after cross-contamination mixing is proportional to the contents before the process,
• Depending on the residue in a container the level of contamination in a product batch can increase,
remain the same or even decrease,
• After the product is packaged, contamination is calculated for bags as the average of its contents,
• Total bag nutritional elements are in the end compared with product admissible limits, determining its
quality.
M ODEL DERIVATION
In order for the measured curves to be useful for modelling, they need to be expressed with numerical equations. Because the shapes highly resemble exponential lines, which should also be the case based on the
previous experience, only exponential curve fitting is considered. In general, single measurement points are
too uncertain to be used directly, and an equation for a curve best fitting all values for a given measurement
point needs to be found, to account for random errors and allow generating values in between measured intervals. For all curve fitting done for contamination trials, the minimisation of the sum of squared deviations
method is used, employing MS Excel solver to find the minimum. Investigated functions are fully constrained
1 Definition origin:
http://www.globalfia.com/tutorials/lesson-4-dispersion
4.4. C ROSS - CONTAMINATION I NVESTIGATION
27
i.e. there are maximum and minimum values given for each parameter and the result in order to obtain values of contamination in between 0 and 1. Both non-linear and evolutionary solvers are used to find the best
fit.
To achieve better fitting results a sum of two exponentially decreasing curves is utilised. Using two basic
functions implies that there is more than a single mixing effect in the equipment. As cross-contamination
is a process of partial mixing of the residue with product coming through, there might be more than one
cumulative effect, which cannot be distinguished in the tracer–collector measurements. For such, a thorough
statistical analysis needs to be conducted to find correlations between equipment characteristics and fitted
curves. Such limited analysis is done further in this document. The general equation for the fitting curve is
thus:
R(x) = a 1 · e a2 x + a 3 · e a4 x
(4.1)
Where x is the amount of material that has been put through and a 1 to a 4 are constants. The first exponential
element in the equation relates thus to residue that is flushed quickly, and the second to a long-term pollution. As such, both the steep descent part of the curve as well as long tail in the end can be accounted for.
Total picked up residue can be calculated for each of the obtained curves by integrating them with respect to
weight:
Z∞
T R = R(x)dx
(4.2)
0
Then, by subtracting the consecutive total residues the amount of added carry-over can be determined for
each of the measured intervals.
For the model construction the following designations are chosen:
R i [i t em, quant i t y]
C j [i t em, quant i t y]
i, j ∈ N+
M
C 0 [or i g i nal i t em, M ]
E AP ∈ (0, M )
EA
TR
σ
Pi , L j
n
Equipment residue matrix containing item number and its quantity
Product batch contamination containing item number and its quantity
row indexes
Mass of a product batch
Initial, original content of the product batch
Exchange amount parameter, specific for a segment
Exchanged amount of material in a given transaction
Total material residue in a given piece of equipment
Random deviation parameter
Picked and left material, auxiliary variables
Number of product batches in a segment
Partial mass exchange model
The cross-contamination process is actually a mixing process, after which the amount of contamination in a
product batch can either increase, remain the same or decrease, depending on the contents of the equipment
Ri and the batch Cj beforehand. In order to retain conservation of mass the mixing has to be a proportional
mass exchange between the batch and the equipment. This exchange can apply to the entire material involved or just a portion of it, but it has to be smaller to both batch size M and equipment residue T R. Therefore, assuming this amount to be equipment property, due to similarity of all products, and calling it E AP , a
mathematical method for this exchange can be derived. For each transaction, involving a single product in
an exchange, there are certain amounts of picked contamination P i and left residue L j . The equations can
be thus formulated:
R i [quant i t y]
(4.3)
Pi = E A
TR
Lj = E A
C j [quant i t y]
M
(4.4)
The resulting exchange is:
0
R i [quant i t y] = R i [quant i t y] + P i − L j
0
C j [quant i t y] = C j [quant i t y] − P i + L j
(4.5)
(4.6)
Providing that R i [i t em] = C j [i t em] and the calculations are done for all i , j where R i [i t em]! = NULL and
C j [i t em]! = NULL. If one of the participants does not have given product, the corresponding transaction
28
4. P RODUCTION M ODEL
quantity is zero. Exchanged amount is normally equal to the exchange amount parameter (E A = E AP ) but it
can be varied by adding to it a random component, then:
E A = min[max(0, random.normal(E AP, σ · E AP )), min(T R, M )]
(4.7)
This way one can assign different exchange values from the distribution, trying to account for e.g. measurement error, while still being within set boundaries. Note that when σ = 0, the value of E A equals the parameter E AP . Also, the exchanged amount cannot be lower than zero or higher that the lower of equipment residue
or batch mass, as it would result in negative values. Because of numerical errors when E A = mi n(T R, M ), and
another method exploring full mixing, this amount is constrained to:
E A = min[max(0, random.normal(E AP, σ · E AP )), 0.95 · min(T R, M )]
(4.8)
Also, as there is no possibility of having dynamic vectors in Simio the size of vectors R i ,C j is set to 20. To
show that for both equipment and product batch eventually exchange the amount E A, the following sums of
individual transactions are made:
∞
X
∞
P
Pi = E A
i =1
R i [quant i t y]
TR
i =1
∞
P
∞
X
Lj = E A
TR
=EA
TR
(4.9)
M
=EA
M
(4.10)
C j [quant i t y]
j =0
M
j =0
=EA
=EA
Meaning that both participants always exchange the whole amount E A, but the difference between their
previous state and one after exchange depends on the contents, and is usually lower because material of the
same items is often present and the actual difference is between the absolute difference of P i , L j .
Also, the calculations are general for a piece of equipment and independent of time or material flow. In
the implementation it is possible to set the random component σ > 0 until a certain quantity of product is
through, and then assume σ = 0, in order to investigate resemblance to peaks and valleys from the contamination trials, shown in Figure 5.2.
Mixing model
Another, specific possibility of material exchange between a product batch and a piece of equipment is full
mixing. It can be imagined as homogeneously mixing both first, and then detaching part of the material from
the mixture, with the size of the batch, and sending it further. Following the nomenclature from the previous
paragraph and adjusting to the methodology, the exchanged amount is:
EA=
M
M +TR
(4.11)
While the whole material for given item R i [i t em] = C j [i t em] is summed together for the exchange:
P i = E A · (R i [quant i t y] +C j [quant i t y])
(4.12)
L j = (1 − E A) · (R i [quant i t y] +C j [quant i t y])
(4.13)
Then, the final contents are not dependent on the previous state:
0
R i [quant i t y] = P i
0
C j [quant i t y] = L j
(4.14)
(4.15)
In this case, the cross-contamination process is quicker and has greater extent with steeper curves, but the
total amount of contamination picked up by the flow of material is not bigger than T R.
4.4. C ROSS - CONTAMINATION I NVESTIGATION
29
Model comparison
The main differences between the proposed models is shown in Figure 4.2. In the partial mass exchange
model a certain settable, equal portions of the equipment residue and product content are taken and mixed
uniformly, to be later proportionally redistributed to the participants. By changing the size of the exchange,
different outcome characteristics can be obtained. In the mixing model, on the other hand, all material contents are mixed together and then product batch is detached from this homogeneous blend and the rest
remains in the equipment.
Figure 4.2: A comparison of derived cross-contamination models
It is expected, that in the mixing model the foreign residue is swept much quicker than in the partial
mass exchange model, as all of the contents always participate in the exchange and the equipment residue is
normally bigger that the product batch size.
Main mixer
Final special case of cross-contamination calculations is the main mixer, the only piece of equipment that
purposely stirs the material within it to stimulate homogeneity. Thus, the material picked up by products
before the mixer and the residue in it, are proportionally spread among everything that is in the mixer and
not only with the first batch, as in case of the previous models. Then the product batches can be treated as
‘clean’ when entering the main mixer, meaning they contain only their original product:
C 0 [or i g i nal i t em, M ]
∞
X
C j [quant i t y] = 0
(4.16)
(4.17)
j =1
These batches exchange their original contents with whatever is in the mixer, and the exchanged amount can
be written as:
M
EA=
(4.18)
n ∗M +TR
30
4. P RODUCTION M ODEL
This way the final cross-contamination models were established and implemented in software. For detailed
algorithm see Figures E.7 and E.8 on page 109 in the Appendix. The user has thus a choice which of the general cross-contamination models to include: partial mass exchange, mixing or none, and whether to include
contamination occurring before and in the mixer. For the partial mass exchange model random component
can be added to vary the swap of material in a chaotic way, which can be stopped for each new inflow after a
certain amount of material has passed.
4.5. D ETERMINED N EEDS
Having introduced the specifications in chapter 3, this stage describes the purpose, assumptions and derived
features of the simulation model, including specific classes and steering logic. Especially, it concentrates on
mathematical representation of cross-contamination process and how to implement it in a DES model.
In order to fully define the cross-contamination character in a model, the equipment properties T R and
E AP , as well as product mass M , and possibly random deviation parameter σ are needed (as defined in
section 4.4.1). Since there is no appropriate knowledge in literature, the best approach is to perform measurements in an actual factory. Then, not only an example data for determining the impact of certain interventions can be obtained, but also a means of validation of the model, which is vital to the generalisation of
the possible findings.
Moreover, placing the research in a realistic setting, where an actual production data is used, is needed to
test the developed methods. Access to specialist expertise, plant resources as well as statistics and knowledge
base, obtained over an extended period of time, is useful not only for evaluation and validity assessment of
the performed research, but also to determine the usability for the industry.
Then, a scheduling solution is required for the production environment to fully define manufacturing
logic, and make room for an analysis of the impact of certain changes to the model execution logic. Full
scheduling optimisation, by means of e.g. linear programming, looking for the best solution to maximise
chosen KPIs, is not within the scope of the research. It would be beneficial though to derive a method for
scheduling, where cross-contamination is involved in detail, and not just as a cost function for certain sequences as in Toso et al. (2009). However, control over the choices made by the simulation model in terms
of scheduling, or at least a clear formalisation of the method, is necessary for a complete solution. One without a means to intervene would be, and rightfully so, rejected by the problem owner as insufficient. That is
why, for the simulation model, scheduling using dispatching rules, incremental choices for a set of available
jobs, is a good enough substitute for scheduling optimisation (Pinedo, 2008, pp. 377). Furthermore, it allows
to approach scheduling as a set of consecutive choices, that can be altered in another experiment and not
a pre-defined, fixed collection. These selection rules can be then defined based on a specific case, investigated thoroughly, and in the end be quite complex and close to the optimal solution. Finally, the issue of
stochasticity is not as difficult to handle in dispatching rules, as it is for optimisation. The former is even
more transparent, as various simulation runs (with different values of the stochastic variables) can be directly
compared with each other.
The above three problems can be tackled as a whole, if a case study is used to combine them. This is
done and presented in the following chapter (5). The company is thus used as a realistic case to procure data,
perform measurements and test the developed approach in a genuine production environment. Moreover,
such setting, and insiders knowledge of company specialists, are important aspects in judging the validity of
the model.
5
E XPERIMENTAL S ETUP
This chapter introduces the experimental setup, based on the case company Sloten, to test the foundations
of the model. At first, the production process is briefly described and system boundaries with assumptions
given. Then, the DES production setting is fully defined and complemented with cross-contamination model,
established on the performed measurements. Finally, scheduling solution for the case is presented, and the
complete model is verified and validated, to the best extent of possessed data.
5.1. P RODUCTION P ROCESS
Sloten’s factory in Deventer is used to mix spray-dried milk fat concentrate with other raw ingredients, e.g.
whey, soy or skimmed milk powders, minerals, vitamins. Then, most of the products are packaged into small
bags, and then palletized. The rest is put either to big bags or moved as bulk material. As a result the final
product, young animal feed, is prepared, and can be transported to a customer.
Production begins with raw ingredients that are shipped from another Sloten factory in Friesland province
in the Netherlands or procured from third parties. These are then stored in designated raw component silos
or storage containers when the usage is low. Sometimes before the main mixing, minor elements need to be
premixed separately to assure uniformity in a mixture. They are then weighted and manually added to one
of the available premix mixers. From then on, the process is controlled numerically. Weight feeding system
is used to automatically draw material from ingredient silos and premix mixers and convey it via a single
possible route to the main mixer.
From the main mixer, the product can be transported either directly to the bagging machines, or sent to
an intermediate storage silo. Currently, most material is bagged immediately, causing a downtime interlink
between engaged equipment, but in the future it is desired to solve this issue by using intermediate storage
for all products. There are 7 big and 4 small silos that can be used for such purpose and, if need be, there can
be a few additional medium-sized silos added. For details on their placement and product route constraints
see Figure 5.1. For transportation either screw conveyors are used, mostly for high capacity connections
between the main mixer and silos, or pneumatic conveying lines starting from existing intermediate silos. In
accordance to the product type, material has to end in one of the five system exits and often can take many
routes to get there, depending on the current system state and product properties.
There are three main product groups in Sloten. First and foremost, there are small bag products that are
put into paper packaging of 10, 20 or 25 kilograms. Under normal operations these comprise over 90% of the
entire production, and are considered most important. Then, there are big bag products packaged separately
into large polyester bags of 950 to 1100 kilograms. These require considerable input of manual labour to
fill and replace, while small bags are sent by conveyor belts for palletizing. Because there is a single shared
palletizing machine for bagging machines BTH1 and BTH2, they are treated as a single line. The last way for
product to leave the system is via bulk station, when discharged directly to a truck as a loose product. This
option is used rarely, mostly for clients in the Netherlands with relatively high usage, and appropriate storage
possibility.
With a large number of produced goods in Sloten, there is a considerable complexity in planning. Customer orders average around 13 tonnes, which results in roughly 11 different mixing orders per day after planning consolidation and usually more for the most-used bagging lines, leading to considerable changeover
31
32
5. E XPERIMENTAL S ETUP
losses. The number of sales products is anticipated to steadily increase and so is the number of customer
orders. Moreover, in the investigated period of January 2015 there were 226 mixing orders (and 352 bagging
orders), totalling over 8000 tonnes for 21 working days, averaging 381 tonnes per day.
Preparing production schedule is a constant burden for planners who need to take into consideration
several variables and make a decision. The operational planning in Sloten is a daily routine, for two days
ahead. Every day, the planning department makes a list of the bagging and bulk loading orders that need to
be dispatched and beforehand produced, on a specific date. Then the required mixing orders are defined and
consolidated as much as possible. For more detailed information on planning and scheduling process see
Figure C.2 on page 85.
5.1.1. P ROPOSED O UTLINE
The initial planned production outline, designed by the company specialists is presented in Figure 5.1. It
only includes in-scope parts, and thus starts with dosing (green), having 5 different system exits: bagging
machines BTH1-3, big bag filling line and bulk stations (red).
Figure 5.1: A schematic of the investigation boundaries
Silos with dashed line borders are proposed ones, that may but do not have to be included, and their
desired number is to be determined using simulation. As it is a network system and different parts are contaminated by various products, a means of keeping track what pollution is in a given interval is necessary.
5.2. S TRATEGIC O VERVIEW
Sloten as a subsidiary of Nutreco is slowly finding its place in organisational culture of the corporation. Being
a specialised producer of feed for infant animals, the changes to adjust to the corporate strategy seem to
have been limited. The apparent biggest aspiration is to increase quality control, and unify it with Nutreco
standards, for which e.g. OEE system was introduced. As the project might be expanded in the future, it is
important that the general approach to the problem is taken, and that the Sloten-specific circumstances are
used to set the project in a realistic setting with a possibility to compare with the current structure.
As previously mentioned the number of products offered by Sloten is expected only to increase as well as
the total amount of goods to produce. Therefore the company is very much interested in bigger production
flexibility as well as higher overall capacity, both of which might be in conflict. Anticipating a certain trade-off
between them, the company is willing to investigate the possible solutions to the problem. By establishing
5.2. S TRATEGIC O VERVIEW
33
a good scheduling methodology one might take advantage of the lack of cleaning runs, while maintaining
acceptable contamination. Although cleaning is done with a sales product of similar content to the next one
in line, its quality is poorer and cost increase. A solution solving two problems at once, or at least limiting
their extent, would be most advantageous.
Normally, regarding production efficiency, the most important factor for Sloten managers is definitely
throughput, with rapidly increasing recognition of high quality standards, including especially cross-contamination. Planners are also keen on ensuring timely completion of products and their delivery. It has also
been noted that the cost assessment should be the ultimate decision aspect, but the relations between its
components are not trivial. Versatility is the ability to process different products by the same equipment,
which is definitely the case for the main mixer. Also bagging lines can process multiple products with varying
bag sizes, packaging options and material properties. But the flexibility is not complete, as only a single
machine is equipped with possibility of bagging into 10 kg packages or dealing with coloured products. In
this sense, BTH3 machine is the most versatile one in the system, capable of processing the biggest number
of products. Thus, it is also the most flexible one, as it should be considered for handling them all. Production
flexibility mostly arises from interconnections and their multi-purpose character. The main mixer blends
all products and is connected to all intermediate storage silos, which are normally not reserved for certain
recipes. Moreover, the silos can be used as buffers for indefinite period, allowing for flexibility in choosing
the orders to be processed. Also when unforeseen circumstances occur, like machine breakage or rushed
jobs, a flexible system is much more likely to handle them well. But with flexibility arises complexity, which,
if not managed properly, might lead to inferior planning and lower production efficiency.
Currently, there is little mitigation of trade-offs between the production efficiency and flexibility, as the
layout of the system differs from the investigated one. Bad experience with downtime interlinking, lack of
freedom and low equipment utilisation, is the reason for exploring a possible performance of a system with
intermediate storage buffers. There is a certain dose of planning flexibility, when contingencies are made
if not all orders are completed for a given day (e.g. including overtime), but it affects the entire job floor,
which is a big drawback. All in all, the level of flexibility needs to increase but so will the complexity and the
company needs to be prepared for that.
P RODUCT P ORTFOLIO
There are around 600 sales products that could be produced by Sloten, based on around 230 recipes i.e.
unique ingredient structures. Thus, often several sales products have the same recipe but differ in terms
of bag size or packaging type (sealing, label language etc.), and might have different contamination rules.
The number of products and ingredients is only believed to grow1 as more specialised feeds are created, to
better suit specific animal breeds. Also more and more regulations arise, depending on the destination country, further specifications of quality standards, and increasing knowledge of their dietary impact. Detailed
company specific description of contamination issues is included in section C.4 on page 88.
Due to limitations in transportation of materials from raw ingredient silos, each recipe has a maximum
speed of delivery to the system, that is often smaller than the theoretical mixing speed, limiting it. Products
and recipes are identified by their unique identification numbers, and a product can only have one recipe,
while a single recipe can be assigned to multiple products.
D ATA AND F EATURES
Many parameter values used in the models are taken from the analysis of the Overall Equipment Effectiveness
(OEE) measurement system data that is gathered in the facility. Most of all, the data is used to determine
the distributions of random variables described in the next section but also to help with validation od the
simulation results. Traditionally OEE systems are used as a process improvement tools to determine possible
problems and thus help solving them. Only some of the performance indicators are utilised in the simulation
model, as defined in Appendix C.3. Representing the same performance indicators is believed to be helpful
to readers and useful for validation.
There are two sets of data available, one detailing stops in production, and interval query, measuring performance in given time periods. The latter divides the day into one hour intervals and adds extra periods once
a product is changed. This way performance characteristics in relation to processed products are obtained
for the mixer and all bagging machines separately. OEE data is used in chapter 5.7 to validate simulation
results.
1 View of Sloten employees responsible for product formulations
34
5. E XPERIMENTAL S ETUP
5.3. M ODELLING I MPLICATIONS
There are several implications arising from the chosen case that need to be incorporated into the simulation
model.
Product batches have fixed density of 500 kg/m3 and mass of 5 kg, as the highest common divisor from
the package sizes they need to fit in. Although that will result in a number of active entities in the system
of magnitude of 30000–70000 (see Appendix A.3), and significantly reduce computing speed, it is a preferred
solution for possible future expansions of the model as well as one more favoured by the client. Moreover,
cross-contamination calculations described in section 4.4 can be generalised only for a uniform entity size.
5.3.1. S IMULATION M ODEL E LEMENTS
Some of the used objects are special cases of the classes depicted in the conceptual model from Figure 3.3 on
page 18. Below paragraphs shortly describe the specifics.
Bulk stations
Bulk dispensers are similar to silos but smaller, that imitate system exit points for loose product discharged to
trucks. It is important that the product is mixed and transported before the due date and await for the truck.
Currently, it is common practice to mix and discharge material via silo to bulk stations on the day before,
as the number of bulk orders is low (typically one truck a week). The model explores a worse case scenario,
when the product can be mixed to silo beforehand but is dispensed further on the same date, typically in the
morning. This way for some time the transportation capacity is seized which, as in case of big bag station,
can lower capacity of other components by delaying discharge.
BTH machines
Three small bagging machines called shortly BTH1-3 (common name originates from their producer) are
the most utilised system exits, recognised as system bottlenecks in terms of capacity. Bagging consists of a
buffer (small silo), batching into a bag and a fixed delay while the flow logic for buffer is based on threshold
levels. BTH machines process small bags of sizes 10 kg (only BTH3), 20 kg and 25 kg with a fixed deterministic
processing speed per bag. Random components are included for short stops, that effectively slow down the
bagging process, and for changeovers. For more details about stochasticity see section 5.3.2
Big bag station
Big bags are just another type (size) of bags, utilising the same bagging sub-model, with a different buffer
discharge logic to fill the bags. Because big bag filling station is not used as frequently as BTH machines,
and because the rate of filling depends highly on the workers abilities, there is no randomisation included.
Although the simulation results for big bags are not that interesting, this component performs an important
role of using mixing capacity, blocking a silo and discharge line, and can in some cases limit the throughout
of the other system exits.
Air filters
Air filters are a special case of conveyors, located at the end of pneumatic conveying line. The path inside
them, that product batches traverse, is limited and assumed to be zero but the residue inside, due to the
equipment properties, is much higher than the average conveyor. But in fact, they base on the same principle
and can be considered zero-length conveyors. For the ease of representation in the interactive mode in Simio,
as well as to avoid infinities in calculations, a node construct is used to suit the modelling needs.
5.3.2. S TOCHASTICITY
The OEE system in Sloten also gathers and classifies information on stops in production. Upon filtering this
data, knowledge on short stops and changeovers is obtained. Short stops for each equipment are fitted with a
single distribution and contain for mixer categories of short stop (all events below 60 seconds), supply faults
and weighing interference and for bagging machines: corresponding short stop events, bag closing/sewing
problems and lack of possible discharge to palletizing. These events are then fitted to a distribution with a
statistical software tool, and the following results are obtained and shown in table 5.1.
Similar analysis is performed for changeovers, i.e. delays needed to change production from one product
to another. Although it is known that for bagging machines these times are sequential (depending on the
5.4. C ONTAMINATION IN S LOTEN
35
type of products), no classification for that exists in the OEE system and the differences are believed by the
managers to be small. Thus, a single distribution is given to all BTH machines, and the results are put into
table 5.1 below:
Table 5.1: Short stops and changeovers distributions
Equipment
Uptime between failures [s]
Time to repair [s]
Changeover time [s]
Main mixer
BTH1
BTH2
BTH3
Exponential(1426)
Exponential(185.5)
Exponential(380.1)
Exponential(651.2)
Log-normal(3.12, 0.93)
Log-normal(1.91, 1.20)
Log-normal(2.26, 1.22)
Log-normal(2.93, 1.17)
Pert(120, 210, 480)
Weibull(2.21, 289.56)
Weibull(2.21, 289.56)
Weibull(2.21, 289.56)
The numbers from table 5.1 are calculated in several different steps. First of all, overall time between
failures in not recorded by the OEE system when short stops are measured, thus mean time between failures
is determined from total operating time, total downtime and number of occurrences, and assumed to be
exponentially distributed. Results are put into table A.1 on page 77, based on the equation (5.1), and the
rounded values are used for simulation.
MT BF =
Total_Staffed_Time − Total_Downtime
Number_of_Occurences
(5.1)
Moreover, there are no precise measurements of the main mixer changeover times, as there is current interlinking between bagging and mixing, which causes the changeovers to be longer. Because of limited information and after consulting with a company specialist, Pert distribution is chosen with most likely occurrence
of 210 seconds.
Finally, the remaining values are fitted in search for their statistical distribution, using 95% confidence
level and the Kolmogorov–Smirnov test. Results are put into table F.3 on page 117 and are significant only for
BTH1-3 changeover, which is the only well-fit distribution (see also Figure F.1). However, thorough investigation into other statistical distributions showed that the log-normal one is best suited for the OEE time to
repair (TTR) data, and after rounding is used in the simulation. Plotting histograms for TTR values showed
clearly that they are not distributed exponentially and similar method to the mean time between failures
could not be used. Although with so many values available one could draw directly a random value from the
set, without fitting them to a distribution, this approach is considered suboptimal for clarity, data analysis
and possible model alterations. Another solution would be to use a sum of multiple distributions or trying to
split them into different short stop categories, and fitting them separately.
5.4. C ONTAMINATION IN S LOTEN
Factory in Deventer is facing a significant increase of cross-contamination, because of the considerable rise
in material transportation length in between the main mixer and system exits, if the new plant outline is to be
implemented. The material, which is always in a form of powder, mostly consists of fat, milk, various protein
sources and other ingredients necessary to provide for nutritional needs of the specified animals. In this
terms the product portfolio is quite similar and is treated as such. For Sloten there are 17 specific nutrients
defined (16 of which are used in simulation due to lack of data on moisture content). These are:
Animal Fat in Fat
Ash
Colourant
Copper
Fat
GMO
Probiotic
GMO in GMO Protein
GMO Soy Flour
Iron
Lactose
Protein
Protimax
Soyabean Protein
Vitamin A
Vital Wheat Gluten Dry
Moisture
It needs to be noted that ash as a nutrient refers to any inorganic content, such as minerals, present in the
feed, rather than leftover from the combustion process. Also, since much of the derived nutrient specification and limit levels has not yet been established by the company, such cases are treated as not defined and
excluded from comparison.
5.4.1. C ONTAMINATION M EASUREMENTS R ESULTS
A general curve expressing all measured contamination is presented in Figure 5.2, shown as a ratio of foreign
product content with respect to the collector material in function of mass flow. Due to a common frame
36
5. E XPERIMENTAL S ETUP
Contamination based on Fe analysis
1.0
Position
A
Contamination
B
BAG
C
D
E
0.5
F
G
0.0
0
1000
2000
3000
4000
5000
Quantity [kg]
Figure 5.2: Measured contamination lab results for all sampling points
of reference, one would expect that the contamination ever increases from point to point, and cannot be
lower than for the previous point. This is unfortunately not the case and there are several identified reasons
for that. First of all, there are additional mixing effects in the equipment, especially silos, meaning that the
material flow through it is not on FIFO basis. While the composition of the inflow material contamination has
a decreasing exponential character, it amasses unevenly, and then discharges in a different manner (Wu et al.,
2009, Cleary & Sawley, 2002). This results in observed peaks in Figure 5.2, especially from point E onwards.
Moreover, the uncertainty of manual sample drawing is considerable and impossible to determine, and the
iron and protein content analysis is subjected to some error as well. The laboratory testing the samples have
not provided the possible margin of error for the method.
The acquired curves are exponentially fitted as described before and put into table 5.2, using sum of
squared deviations method. For goodness of fit see table A.9 on page 79.
Table 5.2: Fitted parameters for cross-contamination curves as a function of processed material quantity
A
B
C
D
E
F
G
BAG
a1
0.384
0.95
0.94
0.94
0.935
0.935
0.92
0.9
a2
-0.0178
-0.021
-0.017
-0.016
-0.008
-0.008
-0.007
-0.0065
a3
0.04
0.05
0.06
0.06
0.065
0.065
0.08
0.1
a4
-0.004
-0.0035
-0.002
-0.001
-0.0007
-0.0006
-0.00055
-0.0005
Carry-over total [kg]
31.573
59.524
85.294
118.750
209.732
225.208
276.883
338.462
Carry-over added [kg]
31.573
27.951
25.770
33.456
90.982
15.476
51.675
61.579
5.4.2. S IMULATED C ONTAMINATION C URVES
Contamination in product batches leaving specified intervals is simulated by recreating the investigated route
as much as possible. There is a slight difference in length of taken pneumatic conveyors (in total simulated
route is shorter), and the bagging machine is assumed to have the same characteristics, although one of a
different type was used. The contamination is expressed as percentage content of different than specified
5.4. C ONTAMINATION IN S LOTEN
37
product within the batch to the same base level, which in this case is 0, by not taking into account contamination arisen from the main mixer and before. Sum of squared deviations method is used to determine the
goodness of fit and the parameters are shown in table 5.3.
Most of the curves are very-well fitted with exception of points E, F and G, where the initial measurements
showed lower contamination than in the previous sampling points, which would violate made assumptions.
However, the simulated curves and the fitted ones are very similar to each other, as shown in table F.2 on
page 117 in Appendix.
Comparision between fitted and simulated contamination curves for bagging
1.00
0.75
Type
Contamination
Fitted
Measured
Mixing
0.50
MixingWithMixer
PartialExchange
PartialExchangeWithMixer
0.25
0.00
0
1000
2000
3000
4000
5000
Quantity [kg]
Figure 5.3: A comparison between measured, fitted and simulated contamination at the bagging level
An example fitting at bag level is shown in Figure 5.3, comparing it to the simulated and measured curves.
Partial mass exchange and mixing models differ significantly in shape, though having exactly the same area
underneath the curve. Mixing model collects a lot of residue in the first part, leading to an almost clean
material after 1000 kg flow. This is not consistent with Sloten’s experience and measurements, leading to
greater confidence in the partial exchange model. However, by adding contamination in the mixer, which
was not plotted in the trials, the results are more alike.
Table 5.3: Accuracy of the simulated contamination curves with respect to measured contamination
Point
SS Total
SS Regression
SS Residual
R Squared
StDev
A
B
C
D
E
F
G
BAG
0.1728
0.1623
0.0106
0.9389
0.0249
1.7298
1.5860
0.1438
0.9168
0.0920
0.9043
0.8733
0.0310
0.9657
0.0427
0.9833
0.8946
0.0887
0.9098
0.0898
0.8724
0.6366
0.2358
0.7297
0.1214
0.3888
-0.2250
0.6138
-0.5787
0.1959
0.6963
0.4390
0.2573
0.6304
0.1268
1.4541
1.4045
0.0497
0.9658
0.0541
Moreover, simulated results for all investigated points, including a random component and cut-off point,
are shown in Figure 5.4, in attempt to recreate peaks and valleys of the measured curves for deviation parameter σ = 0.4. The cut-off points are set at 2000 kg, except for points A and B, where they were set to 0.
38
5. E XPERIMENTAL S ETUP
1.00
0.75
Point
A
Contamination
B
BAG
C
0.50
D
E
F
G
0.25
0.00
0
2000
4000
6000
Quantity [kg]
Figure 5.4: Simulated contamination with random component for all measurement points
As the mixing involved regards solid particles, some dose of ‘ripples’ on the function plot are expected,
based on literature (e.g. Wu et al., 2009)) and experience of Sloten’s specialists. Although Figure 5.4 does not
resemble Figure 5.2 well, it is believed that with more measurements they would be more alike. Perhaps by
introducing trends in the exchanged amount, and not treating material exchange independently the curves
would be smoother and fit better to measurements. Either way, the risks (extent of contamination) represented by both curves should be similar, even they do not resemble one another exactly. That is, because
the total area beneath the curve is the same in both cases and due to multiple (at least 10) application of the
cross-contamination calculations that smoothers the outcome.
5.5. S CHEDULING
The following section explores the definition of various scheduling rules, the model is capable to use. By
scheduling in this case is meant sequencing daily set of orders before mixing and making decisions on silo
allocation. Then, using dispatching rules to make a choice of which product to bag next.
Scheduling is a vital part of the process, as a proper one can significantly improve throughputs and limit
the amount of cross-contamination. It is expected, that there is no universally good solution to include both,
and certain trade-offs have to be made (Blackstone et al., 1982). However, with simulation one can try to
explore the best settings for each, in search of the reasons behind. In the end, a proper scheduling can help
deliver better quality products to the customers in more timely manner or in bigger quantities.
As the current scheduling methodology is entirely based on planner’s experience, without a proper procedure and having only restrictions on sequences, an entirely new approach for Sloten is taken, by proposing
relevant rules from literature, company specialists suggestions and own expertise. The latter arises from investigating model execution in the interactive mode and thinking up solution for model logic, which could
be beneficial.
5.5.1. S CHEDULING C HOICES
There are three main aspects of scheduling to be investigated in the system. First of all, it is sequencing of
daily orders for mixing, in order to achieve good performance, mostly in terms of contamination. As the num-
5.5. S CHEDULING
39
ber of orders is fixed and the changeover times between are independent of the mixed product, there is little
to be done to increase throughput. However, order sizes have significant impact on the possible allocated
silo, providing that there are at least a few available to choose from for any given order, or ability to finish it
within the current shift.
Silo allocation is important to find a good fit between order size, or size of all consecutive orders with the
same recipe, and silo capacity so that a good one is chosen, a need for splitting orders is limited and there is
a higher chance that the next order in line has space to be allocated to.
Finally, bagging order is vital especially for throughput, as it is the bottleneck of the system, and limiting
product changeover would be beneficial for increasing efficiency. On the other hand, all products need to
be bagged either way, and there need to be a changeovers between them, so the impact rather comes from
choosing the right bagging line (BTH1-2 or BTH3), as long as there is material available to bag and no machine
starving. Thus, there is a preference that there are multiple silos to choose from available, meaning that
mixing leads bagging by some time. Moreover, at this point to limit contamination, one can take advantage
of the mixing sequence and similarity of products, or try to minimise the tardiness of products.
Implemented scheduling rules are defined in Appendix in section C.5 on page 89, and the screen capture
from SN for user input of the rules is shown in Figure D.6 on page 98. In the following section the reasons for
defining these rules are given.
S EQUENCING
Sequencing is altering the order of items within specified set, and can be understood as queuing them for
a certain task. In this case, the mixing sequence is the one to be influenced by rearranging the sequence
from the daily planning. In general there are two steps to sequencing in this case: order sorting and recipe or
nutrient incremental adjoining.
At first, the order is sorted in accordance to a pre-specified rule, such as smallest order size first or biggest
bag first. Then, products with the same recipes can be shifted together with the same rule, or more complex nutrient similarity approach can be used to prepare the sequence. Naturally, the orders might remain
unchanged, if desired.
Adjoining nutrients is an attempt to find a sequence that would result in lowest outcome nutrient difference among consecutive products and therefore the smallest number of off-specification bags. It is noted,
that the best would be a directed search, or possibly even full factorial analysis as the number of orders per
day is limited, through the entire daily plan in order to minimise the overall relative nutrient difference. However, as the utilised software is not capable of doing so, and programming an actual sequencing algorithm
would be too time-consuming, an incremental method is used. This is also due to a lack of data on most of
the product nutrients, which makes the full sequencing algorithm premature. The incremental method differs from the search as it chooses the best option and makes it fixed, cumulatively ending up with a sequence.
However, this sequence is most likely not the best one there is. A description of the implemented method is
shown in Figure E.3 on page 104 with a short description next to it.
D ISPATCHING RULES
Dispatching rules allow to choose the most preferred option from the available ones, based mostly on the
current state of the system with possible simple predictions. These rules are utilised for both silo allocation
and bagging order.
The implemented method of choosing a silo is pictured in a flowchart E.9 on page 110 in Appendix. Empty
silos and not-full silos with the same recipe are subjected to allocation dispatching rules. In general, silos
with the same recipe are preferred to any other silos, providing the scheduling rules are met, and if not, the
particular silo is rejected. This can be as a result of a different order type or not being able to fit entire order
in that silo. Preference is then included as a multiplier to a minimisation search for the best one according to
the silo matching rule, which can be based on size fit, shortest distance or disregarded. The search function
implemented in Simio then is thus:
min[(1 + 100 · Candidate.Silo_Model.NutrientDissimilarity) · SiloAllocationMatch[1].SiloMatch · (1+
+(Candidate.Silo_Model.ContaminatingProduct! = RecipeTable.RecipeNumber) + 1000·
(5.2)
·(Candidate.Silo_Model.HighLevelMark < (ManufacturingOrders.OrderQuantity + CapacityNeeded)))]
As shown in equation (5.2), hard-coded fixed rules with assigned fixed weights (determined empirically) are
multiplied by the matching rule SiloAllocationMatch[1].SiloMatch, determined by the user and nutrient similarity function. Nutrient similarity is set to 0 if it is not calculated, as according to the chosen rule. Because
40
5. E XPERIMENTAL S ETUP
that can be 0, 1 is added not to bring down the equation result to 0. Finally, function strongly penalises different contaminating recipes, not currently present in the silo, and prefers silos that have higher capacity than
the order.
Before bagging can be started, an appropriate silo needs to be chosen first and then discharge evaluated.
Scheduling rules can force some checks to be performed earlier than others. This is done in the order: late
products, BTH3 only products (coloured or for 10 kg bags) and then silo groups (designated to machine,
common or all silos). For all of these dispatching rules are executed, similarly to the case of silo allocation
with minimisation function for a search. See also Figure E.12 on page 113. Equation (5.3) is shown as an
example search for BTH3 machine:
BTH3_SiloMatching[1].BTH3Rule · (Math.If(Candidate.Silo_Model.TimerDischargeRate.Enabled, 10000, 1)+
+(Candidate.Silo_Model.DischargeSize > 2)) · Math.If(SchedulingAlternatives[7].Include,
(5.3)
Math.If(BTH3.ContaminatingRecipe == Candidate.Silo_Model.RecipeInsideNumber, 0.5, 1), 1)
First of all, the search function is multiplied by the matching rule, which can be e.g. smallest order first,
biggest silo content or earliest mixing time. Then, there is a strong preference not to choose a silo which is
currently discharging, although in some cases that might be necessary, e.g. for parallel machines or joining
to bag the last product for the day. Then, BTH3 either way prefers smaller bags and, if chosen, will prefer
matching recipes with the one previously bagged.
For a reference on all implemented scheduling rules see section C.5 on page 89 in Appendix.
5.6. QUALITY OF G ATHERED D ATA
An important aspect of the investigation, is assessing the quality of data used as an input to the simulation
model. Although it is very difficult to determine the range of possible errors made when using these values,
this section presents the total extent of knowledge about the used data. There are two sources of quantifiable
information about the system – historical statistics gathered by the company, and measurements performed
specially for the project.
The measurements conducted in the factory were carefully prepared beforehand, and the crew was informed about their tasks. Nevertheless, some identified mistakes were made and possibly also other. As the
access points to the equipment were located throughout the facility, no proper supervision could be maintained over the entire conduct. Employees, in pairs for each of the sampling points, had to, using stopwatches
to determine intervals, manually draw samples from the material flow. With a considerable difficulty in timing, the acquired samples might have been taken at a wrong instance, leading to increased uncertainty. Unfortunately, the extent of this error cannot be measured. Moreover, as for measurement point D, due to human error, samples were taken in entirely wrong intervals, and there were fewer of them. Further sources of
uncertainties in these measurements come from: other human errors, stops in discharge, lack of homogeneity in samples and laboratory conduct. Then, whether gathered samples are representative to the mixture
in that location and time is also unknown. Finally, no precision of the laboratory analysis of the collected
samples is given. Acquired data about the contamination is then fitted with exponential curves, which also
have their precision (as described in table A.9). The simulated curves in a model have a different relation to
the measured ones (see table: 5.3), which even further increases uncertainties about cross-contamination.
Nutrient content data is incomplete and full information is given for roughly 20% of the product portfolio.
Remaining products have some of their information filled but not all. Only for Probiotic, colourant, Protimax,
Vital wheat gluten and copper, all of the specification data fields are given but not all the limits. With full
information about the nutrients, it is expected that the nutrient-based choices would be entirely different.
OEE data, although considered most certain as gathered by automated system and not prone to human
errors, also exhibits some issues in certain aspects. When attempted to extract average production data many
discrepancies were seen, e.g. repeated quality of products above 100%, unrealistically low equipment utilisation or recorded production with unstaffed line. Because of that, the OEE data is used only to obtain information about the short stops and changeover duration. With multiple data points and consistent structure (see
tables A.2 and A.3), the quality is considered high. It is possible though, that there is a systemic bias in recording, yet nothing could be identified. The other way to extract specific production data, is using historical
summaries, measured manually on the job floor and recorded day by day. Although this data is definitely not
perfect, having the same working time for all the machines, which is most likely not true, and some significant
outliers, including ones that are impossible to achieve in reality (e.g. bagging speed over 400 bags/hour), this
is the only remaining source.
5.7. V ERIFICATION AND VALIDATION
41
Finally, data regarding equipment properties, including for instance piping length, sizes or silo capacity, is
approximated by company specialists. No technical as-built drawings are available for equipment properties,
save for general piping and instrumentation drawings without great detail. As such, especially lengths, in
particular for designed parts, are inexact. Also, the impact of equipment angles of inclination to the ground
is not taken into account, while there are a lot of vertical, sloped or horizontal connections.
5.7. V ERIFICATION AND VALIDATION
This section describes the efforts in verification and validation of the simulation model, basing on the approaches presented in Kelton et al. (2013), Sargent (2007). As the model is developed to suit the purpose
described in chapter 1, its validity has to be determined with respect to that purpose. However, thorough
model validation for the full domain of intended applicability is often too expensive, and thus selective tests
are performed until sufficient confidence is achieved (Sargent, 2007).
The general approach to model validity was establishing trust by involving the company experts from the
start in conceptualization and implementation of the model, and relying on their remarks about how the
problem needs to be tackled. This is important for conceptual model validation and usefulness perception,
in the case where it is impossible to have a third-party validation or another model for comparison.
5.7.1. C ONCEPTUAL M ODEL VALIDATION
The first step to confirm the validity of a model is assuring that employed theories and made assumptions are
correct, and that the problem is adequately expressed by the conceptual model’s structure, logic and relationships (Sargent, 2007). These are introduced in section 3 and especially the class diagram from Figure 3.3 on
page 18. At this stage, the experts are asked to perform face validation of proposed structures, boundaries and
logic for each of the sub-models separately and then assess the consolidated model. Over the course of several meetings, involving different experts, proposed solutions are discussed and agreed upon. Data analysis
of historical performance and other quantitative investigations, to e.g. extent of contamination or field tests
of conveyor speeds, are also presented, and compared with the proposed mathematical functions modelling
individual model features. This is how the initial versions of responses shown in Figure 5.3 are created and
improved.
5.7.2. M ODEL V ERIFICATION
Model verification requires checking whether the simulation model executes in the designed way, i.e. if it
accurately translates the design into the implementation. This stage can be further divided into two steps:
general verification and verification tests, where only the latter is formalised. The first involves running the
simulation in the interactive mode (with animation) and monitoring the flow of entities, values of model variables and time-dependent charts. Also checking model for specific events and monitoring whether the model
executes as desired. After the initial verification several more formal tests are performed and their inputs and
results recorded. These include degenerate analysis and traces basing on a dynamic testing approach and
their details are included in Appendix F.1 on page 115.
5.7.3. S IMULATION M ODEL VALIDATION
Model validation is the most important step in determining the usefulness of the model, as any further analysis of experiments should be based on the premise that these indeed sufficiently well describe the possible
model outputs, given the experiment inputs. Here, as the company specialists have been involved in model
development from the start, their perceptions about accurateness of capture of the problem are vital. Furthermore, validation includes parameter variability tests and component/feature individual comparison with
field tests or historical data. As the factory layout has not been changed yet, a comparison of the actual production performance and the simulated output cannot be made. The following sections describe the key
activities during the validation phase. Details, including a quantitative analysis can be found in Appendix F.2.
E XPERT VALIDATION
Expert validation has been undertaken with company specialists involved in the project. It consisted of two
main parts: face validation and experiment result discussion. It concentrated on getting acceptance to modelling approach, and establishing trust in the future results.
The process began with a presentation of the simulation execution in the interactive mode in Simio. At
first, the model components were explained so that the people involved could connect the model layout to
42
5. E XPERIMENTAL S ETUP
the factory. Then, the model was started and material flow initiated. A set of labels, colours and displays was
specially added to support understanding, and provide reason behind what was happening. Consequently,
almost an entire day of production was simulated to show how material is transported through the system,
delayed in between to leave through one of the exits.
The second part involved a discussion of preliminary results and comparison with a general, high-level
historical data. An early version of the result presentation set-up, described in section 6.1, was used to support
the process. General throughputs achieved by the simulation were within acceptable limits by the company
experts, even though the are some differences between the current and the future production organisation,
resulting in a different performance.
After the meeting, the company specialists were provided an opportunity to explore and get familiar with
the simulation model on their own, which they did. Upon inquiry, their thoughts on execution correctness
and agreed upon representation of reality were positive. The general consensus allowed to regard the model
as valid.
C OMPONENT VALIDATION WITH H ISTORICAL D ATA
This section contains validation data based on the case od Sloten. Since the layout of the system is changed
considerably, expected performance validation of the entire system cannot be made. However, some measures can be compared and the differences between them explained.
P RODUCTION M ODEL
The most important performance indicator is speed at the bottlenecks, i.e. bagging. Data on the performance
of all bagging machines BTH combined, is obtained from the client, and concerns years 2013 and 2014 gathered in daily intervals. Data collected in the base case scenario on achieved throughputs is put into table F.4
on page 117. The following Figures 5.5 and 5.6 show box plots comparing these results.
In general, performance obtained in the simulation is much better than the measured one, especially
when looking at the number of processed bags. Measured data has plenty of outliers, including some when
performance is better than the maximum possible, which constitutes for recording error. Some are also very
low, meaning there could possibly be an extended downtime, caused by e.g. machine breakage. Moreover,
the average measured value is lower due to the single reference shift time for all machines, unrealistic for such
a long period, and downtimes due to interlinking with the main mixer.
BTH2
BTH3
100
100
100
200
200
200
300
Average processed bags
400
300
Average processed bags
300
Average processed bags
400
400
500
500
500
BTH1
Measured
Simulated
Measured
Simulated
Measured
Simulated
Figure 5.5: A comparison between the average measured (2013) and simulated packaging throughputs in terms of bags
When expressed in terms of processed kilograms of material, the differences are smaller, especially for
machine BTH3, but only because in the simulated setup this is the place to process 10 kg bags, while historically this equipment has been used only for 20 and 25 kg packages. The BTH3 processed quantity is the
5.7. V ERIFICATION AND VALIDATION
43
only of the six presented comparisons where the simulated value is not bigger than measured, with 95% confidence, see table 5.4. Statistical tests performed were Welch two-sample t-test with one sided hypothesis.
BTH2
BTH3
2000
12000
10000
8000
2000
2000
4000
4000
4000
6000
Average processed quantity
10000
8000
Average processed quantity
6000
8000
6000
Average processed quantity
10000
12000
12000
14000
BTH1
Measured
Simulated
Measured
Simulated
Measured
Simulated
Figure 5.6: A comparison between the average measured (2013) and simulated packaging throughputs in terms of processed quantity
Company specialists, when confronted with the expected base-case scenario performance and its relation to historical data, expressed satisfaction with presented results and hope that it would be so when implemented. According to their expectations as well as evidence from the simulation, the performance is likely
to increase considerably. However, the extent of the increase is uncertain, and thus no validation can be made
to support achieved results.
Table 5.4: Statistical t-test comparing measured and simulated throughput values
Alternative hypothesis
t-value
df
p-value
95% conf. interval
H 1 :BTH1_Bags_Simulated > BTH1_Bags_Measured
H 1 :BTH3_Bags_Simulated > BTH2_Bags_Measured
H 1 :BTH3_Bags_Simulated > BTH3_Bags_Measured
H 1 :BTH1_Quantity_Simulated > BTH1_Quantity_Measured
H 1 :BTH2_Quantity_Simulated > BTH3_Quantity_Measured
H 1 :BTH3_Quantity_Simulated > BTH3_Quantity_Measured
16.504
15.973
10.686
15.561
14.873
-1.045
239.97
278.29
297.40
310.50
296.43
462.00
< 2.2e-16
< 2.2e-16
< 2.2e-16
< 2.2e-16
< 2.2e-16
0.8517
48.45214: Inf
48.49943: Inf
63.43501: Inf
1147.779: Inf
1140.839: Inf
-264.1085: Inf
C ROSS - CONTAMINATION M ODEL
Compared simulation results from trials presented in Figure 5.3 use slightly different route than originally. As
there was just a single trial performed, this section cannot offer more than it is shown in table 5.3 on page 37,
for the accuracy of simulated contamination curves to the measured ones. Pollution in orders is inspected
based on charts like 5.7, where per each order the average amount, of all replications, of material over limits
(red) is compared with the accepted quantity (green).
5.7.4. VALIDATION S UMMARY
All in all, the level of confidence in the model is high, and it is considered valid by the involved company
experts. Performed tests prove that it suits the purpose and is able to assist with the answers to relevant
research questions from chapter 1, and as defined in chapter 3.
Nonetheless, as for any model, additional tests could be performed to further increase this confidence.
There are several options to choose from, while still assuming that involving a third party or building another model of the problem is too expensive for implementation. An interesting alternative is e.g. so called
44
5. E XPERIMENTAL S ETUP
Figure 5.7: An example chart of contamination per manufacturing order
Turing test, testing company specialists whether they can distinguish model outputs from actual system performance. Since the factory layout is not reorganised yet, it is too early for such analysis. Obviously, statistical
tests for the actual system output and model results, would be most useful but are unachievable at this stage.
They would be a vital step for more generalised purpose of the model and application to other facilities.
This section concludes the design phase, where the final simulation model is define, based on the case
company example. Moreover, cross-contamination measurements are discussed, scheduling rules designed
and data quality examined. Finally, the model is verified and validated. It is further used in chapter 6, to
explore possible impact of the devised scenarios, i.e. certain arrangements of the input variables, on the
dependent variables. The results are discussed in chapter 7, where also a connection to the literature is made.
6
E XPERIMENTATION
This chapter explores what is the impact of the chosen independent variables on the dependent variables,
called KPIs. Experiments are grouped in five sets, in which there are multiple scenarios defined, i.e. unique
combinations of the input variables. For each group specific values of the parameters are given, along with
reasons for choosing them. Due to a large number of possible combinations, of almost 500 billion for full
factorial analysis, just for different scheduling options, and long simulation execution time, the investigation
is limited to a certain number of viable scenarios. Replication runner is capable of running 2–4 concurrent
replications and the standard investigated production period is 3–4 weeks. The analysis is based on an actual
set of Sloten bagging orders from January 2015, which have assigned fixed due dates.
A funnel approach is taken to design groups of scenarios, which within a group have varied only a part of
the investigated input parameters (independent variables). Thus, at first a base scenario is defined (see section 6.3), which has a fixed set of independent variables. Then, two sensitivity analysis scenario collections
are performed: one (section 6.4.1) looking at performance without including cross-contamination calculations, with varying silo parameters and discharge speeds, and another investigating the disparity in different
sequencing as well as silo allocation approaches on resultant contamination (section 6.4.2). These are followed by another group, analysing the impact of varying additional silo number and size, which is succeeded
by an analysis of various scheduling rules definitions. Finally, a refined set of very specific inquiries into e.g.
particular scheduling rules, and including random component for cross-contamination, is investigated. The
choice to approach the problem in a given way is aimed to reduce the number of different scenarios to run,
which is important not to drag on the analysis as well as to limit extensive computing time.
All of the scenarios are compared between each other based on pre-defined general KPIs, as defined in
section 3.4.1. All results are put into tables, where for each KPI the average value, standard deviation (SD), median and number of measurements are recorded. For details on specific results see Appendix G on page 118.
These values, as well as figures placing various scenarios dependent variables in comparison to each other,
are presented for each set of experiments in this chapters. The analysis, i.e. drawing conclusions from the
experiment results is done in chapter 7.
6.1. S CENARIO N AVIGATOR S ETUP
A proprietary software of Systems Navigator, called Scenario Navigator is used to define an environment capable of cooperation with Simio, in terms of defining system inputs, managing run of scenarios and displaying the results. Being able to store data from multiple scenarios, it is useful for analysis and exploring possible
options. Tables stored in the program are linked to the simulation engine via *.csv files and the text files produced by the simulations are read after the completion of each scenario and kept in a database. Obtained
data is then used to create graphs, gauges and data tables for the user, either directly or via SQL queries to
filter more specific entries or higher level indicators. For further details, see Appendix D.1. Some of the figures
used in this chapter are screen captures from the tool.
6.2. R EQUIRED WARM - UP P ERIODS
Warm-up period is a time before the simulation reaches its steady state. As the model execution always begins
with an empty system, it is vital to reach this point first, and then start gathering relevant statistics (Kelton et
45
46
6. E XPERIMENTATION
al., 2013). In this case warm-up period is essential to fill intermediate silos, so that there are products to be
bagged immediately, and also to get rid of the initial unknown residue in the equipment.
When running without cross-contamination analysis, only the first part is important and thus a single
day of warm-up period suffices. As the mixer has much higher capacity than bagging at the end of the first
day there are several silos filled. For runs with contamination, it is vital that there are multiple products
going through all equipment, especially conveyors. Some of them are used rarely, like e.g. bulk stations, and
the time needed to sweep the initial contamination is considerable. Thus the warm-up period is set to 3
working days and although the initial contamination may not gone completely, it is small enough to consider
negligible.
For considerations of the replication number as well as the chosen contamination negligence point see
sections A.3.2 on page 80 onwards in the Appendix.
6.3. B ASE S CENARIO
This section presents the initial point of the investigation, a so called base scenario, from which the whole
analysis diverges. The following settings are used for all performed experiments, unless otherwise specified.
Initial model parameters, i.e. independent variables, are put into table 6.1, and mostly are an input from
the company specialists or current equipment settings with no error margin. Stochastic input variables, the
same for all scenarios, are directly taken as in table 5.1. Moreover, independence among all input parameters
is assumed, though especially for scheduling options some are interrelated variables.
Table 6.1: Model input parameters set for the base scenario
Parameter
Unit
Value
Silo to BTH Discharge Rate
Silo to Bulk Stations Average Discharge Rate
Silo to BigBag Discharge Rate
Polems to BigBag/BTH3 Discharge Rate
BTH1 nominal bagging speed
BTH2 nominal bagging speed
BTH3 nominal bagging speed
Big bag nominal bagging speed
Maximum allowed daily overtime (predicted)
Minimum order size left for pre-emption
Contamination cut-off point
kg/h
kg/h
kg/h
kg/h
bags/hour
bags/hour
bags/hour
bags/hour
h
kg
kg
20000
15000
12000
15000
400
400
400
5.5
1
15000
0.01
Figure 6.1: Scheduling parameters chosen for the base case experimentation scenario
Base case scheduling parameters are put into Figure 6.1 and table 6.2, depicting which rules are initially
thought, after consultations with relevant planning personnel, to be advantageous to the production efficiency, especially for a high throughput.
6.4. E XPERIMENT D ESIGN
47
Table 6.2: Model scheduling parameters set for the base case experimentation scenario
Scheduling rule
Value
Mixing order
Silo match
BTH1_2_SiloGroupPreference
BTH3_SiloGroupPreference
BTH1_2_DispatchRule
BTH3_DispatchRule
Smallest Orders First
Best Size Fit
Common Silos First
All silos
Biggest Order First
Smallest Order First
There are other input parameters, not presented in this section, due to their considerable number, which
are kept fixed for all experiments. They include other discharge or transmission speeds, secondary equipment
properties, piping lengths etc. The exact values are kept constant in the SN tool or are hard-coded in the Simio
simulation model.
6.4. E XPERIMENT D ESIGN
The following sections contain funnel-based groups of designed experiments as well as explanations why are
certain values of the independent variables chosen. For each set, a certain number of input parameters are
fixed, often as a result of previous experiments outcome, while some others are varied.
6.4.1. I NITIAL S ENSITIVITY A NALYSIS
The aim of the initial experiments is to perform a high level analysis on production throughput, without including cross-contamination calculations. Changing discharge rates from silos is done to determine if these
are bottlenecks and should be increased, as the powder transportation should not limit equipment throughput. Moreover, an initial investigation into the performance with additional storage silos is to be made as
a sensitivity analysis, to focus further inquiry and reduce number of variables. The following independent
variables are altered in the first phase of the experiments:
•
•
•
•
•
Silos to BTH machines discharge rate
Silos to big bag line discharge rate
Number of additional silos
Capacity of additional silos
Silo recipe restrictions
E XPERIMENTS D ESIGN
Recipe restriction reflects the current operations in Sloten, where two silos, number 21 and 22, are designated
for specific recipes, in this case for recipe 3290 and 3211 respectively, and then no other product can be placed
in them. Production operators believe it to be convenient, so it is desired to assess the impact of their habits
if this situation continued in the new plant layout. Then, reserving silos for specific recipes reduces system
Table 6.3: Initial experiments inputs without contamination investigation
Scenario
NC_A
NC_B
NC_C
NC_D
NC_E
NC_F
NC_G
NC_H
NC_I
NC_J
NC_K
NC_L
Discharge Rate
BTH [kg/h]
20000
18500
22000
20000
20000
18500
22000
20000
20000
18500
22000
20000
Discharge Rate
Bigbag [kg/h]
12000
12000
15000
12000
12000
12000
15000
12000
12000
12000
15000
12000
Additional
Silos Number
0
0
0
0
2
2
2
2
4
4
4
4
Additional Silos
Capacity [kg]
52000 / 25000
52000 / 25000
52000 / 25000
52000 / 25000
52000 (2) / 25000 (2)
52000 (2) / 25000 (2)
52000 (2) / 25000 (2)
52000 (2) / 25000 (2)
Reserved Silos
NO
NO
NO
Silos 21 & 22
NO
NO
NO
Silos 21 & 22
NO
NO
NO
Silos 21 & 22
48
6. E XPERIMENTATION
flexibility, diminishing choices for silo allocation, and impacting the number of available jobs for bagging
machines as a result. The full set of independent variables values is presented in table 6.3.
E XPERIMENTS R ESULTS
Acquired data is put into tables in Appendix from G.1 on page 118 to G.12. Additionally, bagging throughput
in terms of average processed bags per hour is displayed in a box plot in Figure 6.2. Black vertical line inside the distinguished rectangular represents the median of the results. Lower and upper borders of the box
signify respectively 25th and 75th percentile, further dashed lines expected range of the data, and possible
circles farther along the line show outliers1 . Bagging throughputs for Figure 6.2 and the following ones, are
Figure 6.2: A box plot presenting average hourly packaging throughput in experiment runs without including cross-contamination calculations
calculated as an average of a day for given machine and then added up, just as in equation (6.1):
BT H 1_D ai l y_B ag s
BT H 2_D ai l y_B ag s
BT H 3_D ai l y_B ag s
+
+
BT H 1_T i me_W or ki ng BT H 2_T i me_W or ki ng BT H 3_T i me_W or ki ng
250
245
240
235
225
230
Number of Orders completed on time
255
Aver ag e_T hr oug hput =
A
B
C
D
E
F
G
H
I
J
K
L
Scenario NC_
Figure 6.3: A box plot for comparison of orders completed on time
1 For details see R package documentation http://www.rdocumentation.org/packages/graphics/functions/boxplot
(6.1)
6.4. E XPERIMENT D ESIGN
49
Then, timely order completion comparison, using another box plot is presented in Figure 6.3. There are
261 orders to be completed, most of which are expected to finish within the run time-frame, but there is no
single run where all orders are completed on time.
6.4.2. C ROSS - CONTAMINATION I NVESTIGATION
The second part of the initial sensitivity analysis is determining the impact of mixing sequencing, defined
in section 5.5.1, and nutrient-based silo allocation, as in equation (5.2), on the total contamination. When
the parameter ‘Nutrient-based silo allocation’ is set to ‘YES’, it is calculated based on schematic displayed in
Figure E.3 on page 104, and if it is ’NO’ the value of Candidate.Silo_Model.NutrientDissimilarity = 0. Recipe
based sequencing (combining recipes) is at first sorting the daily orders according to a specified ‘Mixing order’
rule, also done for nutrients, and then grouping all products with the same recipe together, starting from the
first one for a given day. Order sequencing with option ’combine nutrients’ bases on the same principle as
described in section 5.5.1.
It is expected that mixer sequencing has a considerable impact on the total amount of excess contamination, as any choices done for this set have direct repercussions in mixer allocation and bagging order.
Moreover, most of the storage space is in the main silos numbered 14 to 23, and to reach any of them there is
a long common route for the product, up until the end of the main elevator, an possibly even further. There
are large screw conveyors there, which are measured to have a relatively high amount of material residue,
comprising roughly half of the measured total residue for the investigated route (described in section C.4).
E XPERIMENTS D ESIGN
The base scenario from section 6.3 is used to explore the impact of nutrient or recipe based allocation. Special
controls are added to the model to disconnect nutrient based sequencing and silo allocation. All scenarios are
run including (pre)mixer residue of 166 kg, and the set of experiment differentiated values for independent
variables is listed in table 6.4.
Table 6.4: Initial cross-contamination experiments inputs
Scenario
Order sorting
Order sequence
Nutrient-based silo allocation
MassC_A
MassC_B
MassC_C
MassC_D
MassC_E
MixC_A
Largest bags first
Largest bags first
Largest bags first
Largest bags first
Largest bags first
Largest bags first
Combine nutrients
Combine nutrients
Combine recipes
Combine recipes
None
Combine nutrients
YES
NO
NO
YES
NO
YES
By comparing the results of the experiments presented in table 6.4, a sensitivity analysis to order sequencing and silo allocation can be made, assessing what factor has a bigger impact on the total amount of offspecification product. Additionally, experiments from MassC_A to MassC_E are run with the partial mass
exchange cross-contamination model, while MixC_A is using the mixing one, and the same remaining input parameters as MassC_A. Thus, assessment on how much worse, for the total amount of off-specification
product, is a slower release of the residue product, can be made. According to the company specialists, the
newest state-of-the-art software, used in another Nutreco factory, utilises a full mixing based material behaviour in the process. As such, comparing it with a measured behaviour can indicate the extent of error
made with this assumption. Being able to compare these two models can provide argumentation whether
closer scrutiny needs to be paid to this phenomenon.
E XPERIMENTS R ESULTS
Determining the impact of sequencing on overall contamination is vital to understand the implications and
extent of it, as well as to show how much savings can be achieved with a simple ordering scheme. A box plot
for total contamination in all packaged product is shown in Figure 6.4. Specific results for each scenario are
put into tables from G.13 on page 120 to G.18.
6.4.3. A DDITIONAL S ILOS E XPERIMENTS
A vital part of the investigation is determining the impact of different plant layouts on production performance. In this case, design interventions are limited to the number and size of additional intermediate storage silos, based on the proposed factory layout of Sloten, shown in Figure 5.1. There are 11 existing silos,
6. E XPERIMENTATION
15
10
5
Total contamination above limits [‰]
20
50
A
B
C
D
E
Mix
Scenario
Figure 6.4: A box plot for comparison of impact of sequencing on cross-contamination
varying in size and connections that are included in the analysis as fixed input. Then, up to eight additional
silos are added, with predefined constraints. Thus possible silos 51, 52, 61 and 62 are parallel to each other
and connected only to packaging line 1, comprising bagging machines BTH1 and BTH2. Remaining potential
silos 53, 54, 63 and 64, can only discharge to BTH3 machine.
This set of experiments can also be viewed as a sensitivity analysis of the varying number of additional
silos and their size, because the factory layout interventions only regard a single dimension.
E XPERIMENTS D ESIGN
Since the joint capacity of machines BTH1 and BTH2 is bigger than of BTH3, it is likely that the first ones
might benefit more from increased silo capacity, and thus several experiments vary silo sizes with shift to that
side. There are some chosen alterations to the scheduling rules of the base case scenario from section 6.3.
Figure 6.5 shows a screen capture of used scheduling alternatives. Moreover, BTH3 bagging rule also deviates
from the base scenario and is set to “Smallest Remaining Material to Package”. These complete the final set
of the independent variables.
Figure 6.5: Scheduling parameters chosen for silo number and capacity impact investigation
Table 6.5 contains varied silo capacities, expressed in kilograms, of the additional investigated silos, the
only alterations done within this set. If the value is set to 0, then the silo is not included. Moreover, the silo
6.4. E XPERIMENT D ESIGN
51
high level marks were set 1000–5000 kg lower. For capacities of 25000 kg, the high level marks are set to 24000
kg, for remaining silos below 80000 kg they are 2000 kg lower and for the biggest silos 5000 kg lower.
Table 6.5: Altered independent variables for the exploration of the impact of additional silos and their capacities on production performance
Scenario
Silo_51
Silo_52
Silo_53
Silo_54
Silo_61
Silo_62
Silo_63
Silo_64
MassC_F
MassC_G
MassC_H
MassC_I
MassC_J
MassC_K
MassC_L
MassC_M
MassC_N
MassC_O
MassC_P
MassC_R
0
52000
0
22000
42000
52000
80000
130000
52000
52000
32000
32000
0
52000
0
22000
32000
52000
62000
105000
52000
52000
32000
32000
0
0
32000
22000
32000
32000
42000
52000
32000
32000
32000
32000
0
0
22000
22000
22000
22000
32000
32000
22000
22000
32000
32000
0
0
0
0
0
0
0
0
25000
25000
32000
0
0
0
0
0
0
0
0
0
25000
25000
32000
0
0
0
0
0
0
0
0
0
25000
0
0
0
0
0
0
0
0
0
0
0
25000
0
0
0
The investigation explores only a small part of the solution space that is expected to be beneficial to production performance. Silo sizes from table 6.5 are chosen based on early company specialists prediction and
then further varied in size, mostly increased. The biggest proposed silo from scenario MassC_M is specially
selected for the set of production orders used with the experiments, so that the largest order would fit in it
entirely without splitting.
E XPERIMENTS R ESULTS
Obtained results for all scenarios are put into tables in Appendix from G.19 to G.30 in the same manner as before. Additionally, scatter plots are provided in Figures 6.6 and 6.7 to visualise the average dependent variable
values for the total contamination as well as average hourly bagging throughputs as functions of the total silo
capacity. Additionally, the dots representing the results are coloured to show how many silos are used.
1085
Average throughput in bags per hour
1080
Silos
8
1075
6
4
2
0
1070
1065
0
100
200
300
Total Silo Capacity
Figure 6.6: A scatter plot displaying average results for throughput with respect to varying silo capacity and number
Total contamination from Figure 6.7 is shown in parts per thousand and represents the ratio of the total
amount of material considered off-specification to the total packaged quantity.
52
6. E XPERIMENTATION
10.0
Total contamination above limits [‰]
9.5
Silos
8
6
9.0
4
2
0
8.5
8.0
7.5
0
100
200
300
Total Silo Capacity
Figure 6.7: A scatter plot displaying average results for contamination with respect to varying silo capacity and number
Furthermore, the two plots from Figures 6.7 and 6.6 are combined in one to show the total contamination
as a function of bagging throughput in Figure 6.8.
10.0
Total contamination above limits [‰]
9.5
SiloCapacity
300
9.0
200
100
0
8.5
8.0
7.5
1065
1070
1075
1080
1085
Average throughput in bags per hour
Figure 6.8: A scatter plot displaying average results for contamination with respect to achieved throughput for silo parameter analysis
Finally, box plots for comparison of completed orders and ones finished on time are included in the Appendix in Figures G.2 and G.3 respectively.
6.4.4. S CHEDULING RULES E XPERIMENTATION
The following group of scenarios concentrates on assessing the impact of different scheduling logic on the
production performance. They utilise specific silo parameters from one of the scenarios from section 6.4.3,
namely MassC_K, having four fixed silos with capacities of 52, 52, 32 and 22 tonnes. It is expected that there
6.4. E XPERIMENT D ESIGN
53
is a considerable impact of different scheduling rules on all the KPIs and that the reasons for the differences
among them can be identified.
Yet again, experimentation can be viewed as a sensitivity analysis, this time exploring the expected tradeoffs between aspects of production performance. Thus the main reason for choosing the following sets of
independent variables is identifying the extent of impact of scheduling on the average throughput as well as
corresponding amount of the off-specification product.
E XPERIMENTS D ESIGN
There are a number of independent variables that are varied within this set of investigated scenarios, all of
which are scheduling rules. Although the number of scheduling possibilities just for different combinations
of the rules is large, of almost 500 billion, there are only six experiments performed. Table 6.6 contains numbered scheduling rules as in Figure 6.1 and proposed scenarios. If the rule is included, the cell contains ‘YES’,
otherwise ‘NO’.
Table 6.6: Scheduling rules chosen for scenarios exploring their impact
Scenario
1
2
3
4
5
6
7
8
9
10
11
Scheduling_A
Scheduling_B
Scheduling_C
Scheduling_D
Scheduling_E
Scheduling_F
YES
YES
YES
YES
YES
YES
YES
YES
YES
YES
NO
YES
YES
YES
NO
YES
YES
YES
NO
NO
YES
YES
YES
YES
NO
NO
NO
NO
YES
NO
NO
NO
YES
YES
NO
YES
YES
YES
NO
NO
NO
NO
NO
NO
YES
NO
YES
YES
NO
NO
NO
NO
NO
NO
YES
NO
YES
NO
NO
YES
NO
NO
NO
NO
YES
YES
Chosen dispatching rules, that are also varied, are put into table 6.7. The mixing order for all is ‘Smallest orders first’, and the nutrient based allocation is also always included, because it, based on results from
section 6.4.2, reduces the total contamination significantly.
First two scenarios in the set are aimed at achieving low contamination, by taking advantage of the mixer
sequence, where dispatching rules for bagging machines set to ‘earliest mixing time’. The following two are
set to increase the average bagging throughput, by taking advantage of the different capacities of lines 1 &
2, and the final two provide reference for the first two by having a combination of independently reasonable
choices, that together do not fit too well.
All of the values are based on the input from company specialists, literature study from section 2.5, as well as
own exploration of the model in the interactive mode.
Table 6.7: Dispatching rules chosen for scenarios exploring impact of scheduling
Scenario
Silomatch
BTH1-2silopref.
BTH3silopref.
BTH1-2dispatch
BTH3dispatch
Scheduling_A
Scheduling_B
Scheduling_C
Scheduling_D
Scheduling_E
Scheduling_F
Bestsizefit
Bestsizefit
Bestsizefit
Shortestroute
Shortestroute
Bestsizefit
All
Designated
All
All
All
Common
All
Designated
All
All
Common
All
Earliestmixingtime
Earliestmixingtime
Biggestsilocontent
Biggestorder
Biggestorder
Biggestorder
Earliestmixingtime
Earliestmixingtime
Smallestsilocontent
Smallestpackage
Smallestquantityleft
Earliestmixingtime
As such, within this set there are fifteen varied input variables. Although independence among them is
assumed, there are quite possibly notable interconnections. These are not explored.
E XPERIMENTS R ESULTS
The results of the experiments are put into tables in Appendix from G.31 on page 125 to G.36. Additionally,
a similar graph to 6.8 is prepared for scheduling investigation and shown in Figure 6.9, but grouped by scenario name (ScA is short for Scheduling_A etc.). Different colours in the scatter plot signify scenarios whose
independent variable values are defined in tables 6.6 and 6.7.
54
6. E XPERIMENTATION
Total contamination above limits [‰]
9.0
Scenario
ScA
ScB
ScC
8.5
ScD
ScE
ScF
8.0
275
280
285
290
Orders completed on time
Figure 6.9: A scatter plot displaying average results for contamination with respect to achieved throughput for scheduling analysis
1000
900
800
700
600
500
Average number of processed bags per hour
1100
Moreover, a box plot showing the total average packaging throughput per hour for the investigated group
is shown in Figure 6.10.
A
B
C
D
E
F
Scenario
Figure 6.10: A box plot displaying average results for bagging speed per hour for each of the investigated scenarios
Similarly to the previous section, box plots for comparison of completed orders and ones completed on
time are included in the Appendix in Figures G.5 and G.6 respectively.
6.4.5. R ANDOM C OMPONENT I NVESTIGATION
The final set of scenarios includes the parametrised random component for cross-contamination calculations. The leading assumption, on top of ones made previously, is that the exchanged amount, as described
in equation (4.8), differs per product batch according to the normal distribution, where the expected value is
E AP , with a standard deviation of σ · E AP , but altogether constrained not to result in negative values. Doing
so creates ripples in the cross-contamination curve, similar, but lower and much more frequent as measured
6.4. E XPERIMENT D ESIGN
55
(shown in Figure 5.4 on page 38). The following experiments explore the effects of this randomisation on the
total amount of off-specification material.
R ANDOM COMPONENT SIZE
Scenario MassC_J (see section 6.4.3) is taken as the source for the set of independent variables, because it has
relatively good results for both throughput and contamination. This is thus a starting point for comparison
with σ = 0. Then, by increasing the value of the random component σ the ripples appear and become higher,
which might have an effect on total contamination. The following two experiments are given the same input variables as scenario MassC_J and only differ in the random component value, as defined in table 6.8.
This way a comparison can be made to results without the random component, and two scenarios including
it. Additionally, the mixing cross-contamination model is also explored to provide a frame of reference for
method comparison. For scenario MixC_B the random component is not applicable.
Table 6.8: Random components chosen for analysis
Scenario
σ
Cross-contamination model
Random_A
Random_B
MixC_B
0.2
0.5
-
Partial mass exchange
Partial mass exchange
Mixing
To sum up the design of this set, scenario Random_A differs from MassC_J by introducing a random component for cross-contamination σ = 0.2. Then, experiment Random_B has it increased to σ = 0.5. Finally,
scenario MixC_B has the same input parameters as Random_A with exception of the cross-contamination
model, which is set for the mixing one, as the name suggests.
R ANDOM COMPONENT SIZE EXPERIMENT RESULTS
The results from scenarios Random_A, Random_B and Mix_B are put into tables 6.9, 6.10 and 6.11 respectively. Additionally, the results, including further experimentation with the random component is put into
plots, presented in the Appendix G.5 on page 127.
Table 6.9: Experiment results of scenario Random_A
Scenario
KPI
Avgthroughputbags
Avgthroughputquantity
Orderscompleted
Orderscompletedontime
Avgcontaminatedorderratioabovelimit
Totalcontaminationabovelimits
Random_A
Average
1077.61
24821.37
304.05
292.40
0.08903
0.0080
SD
38.42
1351.91
1.16
5.73
0.00837
0.0006
median
1085.99
24939.92
304.50
292.50
0.08814
0.0078
no. measurements
320
320
20
20
20
20
Unit
bags/h
kg/h
-
Random_B
Average
1077.61
24821.37
304.05
292.40
0.08753
0.0080
SD
38.42
1351.91
1.16
5.73
0.00849
0.0006
median
1085.99
24939.92
304.50
292.50
0.08563
0.0078
no. measurements
320
320
20
20
20
20
Unit
bags/h
kg/h
-
Table 6.10: Experiment results of scenario Random_B
Scenario
KPI
Avgthroughputbags
Avgthroughputquantity
Orderscompleted
Orderscompletedontime
Avgcontaminatedorderratioabovelimit
Totalcontaminationabovelimits
There is a number of additional scenarios including the effects of the random component together with
various scheduling rules. As they are not directly devised to answer the research questions, they are put into
the Appendix G.5 on page 127, with short analysis. Their results though, often appear in figures providing
comparison among multiple scenarios.
56
6. E XPERIMENTATION
Table 6.11: Experiment results of scenario Mix_B
Scenario
KPI
Avgthroughputbags
Avgthroughputquantity
Orderscompleted
Orderscompletedontime
Avgcontaminatedorderratioabovelimit
Totalcontaminationabovelimits
MixC_B
Average
1079.11
24873.79
304.25
293.65
0.04018
0.0036
SD
30.47
1278.97
1.26
3.64
0.00171
0.0001
median
1084.61
24913.92
305.00
294.00
0.04056
0.0036
no. measurements
320
320
20
20
20
20
Unit
bags/h
kg/h
-
F LEXIBILITY EXPLORATION
In general, it is expected that by introducing additional silos the flexibility of the system would increase, as
there would be more freedom to make choices. But then, reserving certain silos for specific recipes should
decrease this flexibility to reduce the total contamination. The following experiments try to provide quantifiable information to assess whether this expectation is correct, by introducing first scenario with 8 additional
silos and measuring the values of the dependent variables. These are then analysed and six specific recipes,
which are connected to products which are likely to be contaminated, are picked and silos are reserved for
them.
Silo sizes are the same as in scenario MassC_N, but scheduling rules differ, as these are based on scenario
Scheduling_A to have low contamination. The exact values of chosen independent variables for scheduling
are put into tables 6.12 and 6.13.
Table 6.12: Dispatching rules for scenarios Random_H and Random_I with silo reservation investigation
Mixing order
Silo match
BTH1-2 silo pref.
BTH3 silo pref.
BTH1-2 dispatch
BTH3 dispatch
Smallest order
Best size fit
All
All
Earliest Mixing
Earliest Mixing
Table 6.13: Scheduling rules for scenarios Random_H and Random_I with silo reservation investigation
Rule
1
2
3
4
5
6
7
8
9
10
11
Value
YES
YES
YES
NO
NO
NO
YES
NO
NO
YES
NO
Moreover, a random component of σ = 0.2 is included. The specific silos and recipes that are reserved in
scenario Random_I (6 silos in total) are shown in table 6.14.
Table 6.14: Chosen combinations of silos reserved for specific recipes in scenario Random_I
Silo Number
Reserved Recipe
15
3100
20
3113
22
3130
23
3149
52
3301
54
3412
Then the final two scenarios are fully defined.
F LEXIBILITY EXPLORATION RESULTS
The results from the investigation are put into tables 6.15 and 6.16
Table 6.15: Experiment results of scenario Random_H
Scenario
KPI
Avgthroughputbags
Avgthroughputquantity
Orderscompleted
Orderscompletedontime
Totalcontaminationabovelimits
Ratioofmaterialabovelimitstototalprocessed
Random_H
Average
1074.17
24769.10
303.60
291.75
0.08179
0.0073
SD
28.57
1426.32
1.36
3.14
0.00669
0.0004
median
1079.21
24977.36
304.00
291.50
0.07926
0.0073
no. measurements
320
320
20
20
20
20
Unit
bags/h
kg/h
-
6.5. E XPERIMENTATION C ONCLUSION
57
Table 6.16: Experiment results of scenario Random_I
Scenario
KPI
Avgthroughputbags
Avgthroughputquantity
Orderscompleted
Orderscompletedontime
Totalcontaminationabovelimits
Ratioofmaterialabovelimitstototalprocessed
Random_I
Average
1052.19
24178.45
301.00
284.25
0.08079
0.0068
SD
56.29
1816.54
2.86
7.29
0.00668
0.0005
median
1073.53
24472.04
301.00
284.00
0.08057
0.0069
no. measurements
320
320
20
20
20
20
Unit
bags/h
kg/h
-
Additionally, the Appendix contains scatter plots for all experiments including the non-zero random component. Figure G.8 on page 130, to depict the relation between the total contamination and the average
throughput, in Figure G.9 to show the order completion, and finally in Figure G.10 to show the relation of the
total contamination to the average contaminated order ratio above limits.
6.5. E XPERIMENTATION C ONCLUSION
This section concludes the experimentation phase. In total 46 experiments are performed in five scenario
groups in chapter 6, following a funnel approach. The reasons for the choice of independent variables, in
short to determine the impact of plant layout and scheduling interventions on production performance, are
discussed, and the values of dependent variables (KPIs) given in Appendix G. Moreover, figures displaying
selected charts of different KPIs in a scenario groups are included, to visualise the obtained data.
The following chapter provides an analysis of those results as well as recommendations arising from the
experimentation.
7
R ESULTS A NALYSIS
This stage deals with an interpretation and assessment of the results from the experiments performed in
chapter 6. As described by van der Zee & van der Vorst (2005), a proper and universal way to compare various
scenarios is needed to fairly judge their performance. The analysis is based on KPIs defined in section 3.4.1.
Statistical tests performed in this section are done with 95% confidence. T-tests, unless otherwise specified, are Welch two-sample ones, where equality of variances is not assumed. For variables where there are 20
values available additional F-tests are calculated to determine equality of variances (see column ‘var equal’ in
result tables in Appendix G). When equality of variances is confirmed the pooled variance is used to estimate
the variance, otherwise the Welch approximation to the degrees of freedom is calculated.
7.1. R ESULTS P RESENTATION
Just as in case of experiment inputs, an extensive set of processed data is included in the SN tool for results
visualisation. This comprises of six screens and multiple charts, gauges and table to provide the user with
relevant information, allowing to make design (and in the future operational) decisions. Typically, when possible, information displayed is based on the average of multiple replications, and if needed, tables with raw
data are available for preview. One of the important features are resource charts that can display machine
utilisation in time intervals per different replications. Thus, important data is communicated in easy to understand, visual form. Several screen captures of the aforementioned SN tabs are presented in section D.2
and further.
Total contamination above limits [‰]
15
12
Set
Contamination
Random
Scheduling
Silos
9
6
3
1060
1070
1080
Average throughput in bags per hour
Figure 7.1: All scenarios plot for total contamination as a function of the average throughput
58
7.2. I NITIAL S ENSITIVITY A NALYSIS D ISCUSSION
59
Moreover, Figure 7.1 shows a scatter plot of all scenarios including cross-contamination analysis for the
average total contamination, shown against the average packaging throughput as a means to compare results
among different scenario groups.
7.2. I NITIAL S ENSITIVITY A NALYSIS D ISCUSSION
Initial experiments designed in section 6.4.1 are focused on finding the starting point for a more oriented
search. The aim is to perform limited sensitivity analysis on adding extra silos, varying discharge speed, and
reserving some silos for specific recipes. All done without including cross-contamination in the investigation.
For more detailed description see section H.1, and for statistics see tables from H.1 on page 132 to H.10 on
page 133.
There is some evidence to suggest that the more silos the more orders finished on time, and the higher
average throughput. However, in total it is rather inconclusive. The only piece of convincing results refers to
the difference in the average throughputs, when the silo discharge speed is lowered from 20 tonnes per hour
to 18. In all three corresponding cases, with different silo number, the average throughput is significantly
smaller. Likewise, when increasing the discharge to 22 tonnes per hour, the average throughput in all three
cases is not significantly higher than for 20. This can be clearly seen in Figure 6.2. When comparing the
same scenarios with the only difference of reserving silos, two out of three have a significant difference in
means when comparing orders finished on time. Moreover, one in three is showing similar significance in
the average achieved bagging throughput.
In the end, from the initial analysis, it can be concluded that the model is not highly sensitive to the varied
parameters.
7.3. I MPACT OF C ROSS - CONTAMINATION
Experiments aiming to determine the impact of sequencing based on nutrient similarity on production efficiency are shown in table 6.4 on page 49. The goal is to determine the contamination characteristic with
different methods of mixer order sequencing and nutrient based silo allocation.
Investigation into the total contamination in bagged products is shown in Figure 6.4, which clearly indicates, together with performed t-test shown in table H.11 on page 133, that mixer sequencing based on
nutrient similarity is significantly better (lower contamination) than simple recipe joining or no intervention,
except for order sorting. Surprisingly, recipe joining is not significantly better than no specific sequencing.
This can be because of limited knowledge about the nutrients and their limits, based on which the final contamination is assessed, or even an outcome due to the used order set. What is vital, nutrient based sequencing
is a much better method to keep contamination low.
Also as expected, the total contamination when using the mixing cross-contamination model is much
lower than for partial mass exchange one. This is due to the steepness of the release curve, as a result of
which, the foreign residue is concentrated in the early batches after the change of product. Of course, the
same amount of material is carried, but in the mixing model the early batches are more contaminated than
in the partial mass exchange model, resulting in quicker disposal of the residue.
Figure 7.2 shows a scatter plot of the total contamination in the bagged material as a function of the
average contamination in polluted orders, which have at least one off-specification bag. Additionally, error
bars are added to signify 95% confidence intervals. Clearly, there is a strong connection between the average
pollution and the total contamination, almost linear based on Figure 7.2, and the lower the contamination,
the smaller the uncertainties.
Finally, no evidence that allocating silos based on nutrient similarity has limiting effect on contamination
could be found, when comparing scenarios MassC_A and MassC_B as well as MassC_C with MassC_D. It is
possible, that there is no such correspondence for the given system, or it is due to the limited choice of silos
when starting to mix a new order. With increasing number of silos this effect could be noticeable. Moreover,
since most of the residue is concentrated in equipment before the silo, it is possible that further carry-over
has little effect on final contamination.
7.3.1. C ROSS - CONTAMINATION M ODEL C OMPARISON
To compare the developed models and asses the possible error made by using a wrong one, data on offspecification order ratio is extracted and put into table 7.1.
Nutreco specialists claim, that the best state-of-the-art automatic production support tools, available in
the corporation and commercially on the market, measure inflow material and use a mixing model for pre-
60
7. R ESULTS A NALYSIS
17.5
Total contamination above limits [‰]
15.0
Scenario
A
12.5
B
C
D
E
10.0
MixA
7.5
5.0
0.05
0.10
0.15
0.20
Average contamination in polluted orders
Figure 7.2: Scheduling parameters chosen for the base case experimentation scenario
diction of the flow progress. Assuming that the developed partial mass exchange model provides an exact
prediction of the process, the error made when using the mixing model instead can be estimated.
Table 7.1: Statistics on order above limits ratio for different cross-contamination models
n
mean
stdev
median
MassC_A
696
0.10955
0.09948
0.08750
MixC_A
714
0.04790
0.02972
0.04063
Thus, the expected value of the difference between all possible sample means is equal to the difference
between population means:
E (x 1 − x 2 ) = µd = µ1 − µ2
(7.1)
The standard deviation of the difference between sample means σd is approximately equal to:
s
σd =
σ21
n1
+
σ22
n2
(7.2)
When calculated the values are:
µd = 0.06165
(7.3)
σd = 0.00393
(7.4)
The 95% confidence interval for the difference is thus from 0.06144 to 0.06186. The discrepancy accounts thus
for over 6% of the contaminated order size, that is misjudged as within limits, while assessed off-specification
with a more accurate prediction method. Relatively to the order ratio, this is an almost 130% increase, which
deems such mixing models unreliable for estimation.
7.4. L AYOUT I NTERVENTIONS I MPACT
Experiments designed in section 6.4.3 deal with assessing the impact of introducing additional intermediate
storage in between mixing and bagging. Chosen independent variables are listed in table 6.5, and the results
in tables from G.19 on page 122 to G.30 on page 124. Statistical tests for order completion, total contamination
and average throughput are put in tables H.12, H.13 and H.14 respectively.
7.5. I MPACT OF S CHEDULING A NALYSIS
61
Based on Figure 6.6 there seems to be a relationship that when silo capacity increases, so does the average
throughput. It is stronger for the overall capacity than for silo number. The plotted relationship shows that
with further increase of silo capacity, the relative increase of the average throughput decreases, and seems
to approach asymptote at over 300 tonnes. Fewer bigger silos have a higher positive effect on the average
throughput than more smaller ones.
Similarly, Figure 6.7 shows the relationship of the total contamination and additional silo size. Here, the
relationship does not seem to be so straightforward. In the beginning, with an increasing additional silo
capacity the total contamination decreases, but after it has reached around 200 tonnes, it starts increasing
rapidly. But when looking at the silo number, the link is ever decreasing – the more additional silos, the lower
the total contamination.
Both of these relationships can be explained. For capacity, the total contamination initially decreases because of increased flexibility in choosing the intermediate storage, as well as lower chance of splitting orders
among multiple silos, which could result in higher contamination. But with increasing silo capacity, the total
material residue in the system increases as well. At some point, the benefits from better fitting the orders
outweigh the drawbacks of putting a few relatively small orders into too big silos. As such, there is a increased
risk of higher contamination with very large silos.
Then, Figure 6.8 shows a scatter plot of the total contamination as a function of the average throughput.
Another version of this chart, including error bars for 95% confidence intervals is shown in Figure 7.3, to
depict the uncertainties to the means shown in the plot, arising from the simulation.
Total contamination above limits [‰]
10
SiloCapacity
300
200
9
100
0
8
1060
1070
1080
Average throughput in bags per hour
Figure 7.3: A scatter plot displaying average results for contamination with respect to achieved throughput with error bars for 95%
confidence intervals in silo parameter analysis
Based on Figure 7.3 there is a region, with relatively low contamination and high average throughput, not
very sensitive to changes in silo parameters. Even large variations in additional silo capacity have little effect
on the average throughput, which is relatively less than 2% between the lowest and the highest achieved. On
the other hand, the relative difference in the total contamination amounts up to 35% relative difference.
Finally, statistical tests, the results of which are put into tables H.12, H.13 and H.14, are performed with
assumption that the bigger the total silo capacity, the higher the average throughput, more orders completed
on time and less total off-specification material. This is the case only in the beginning, when increasing
additional silo capacity from 0 to about 100 tonnes (2 or 4 extra silos). Further on, the sensitivity of response
decreases.
7.5. I MPACT OF S CHEDULING A NALYSIS
An analysis of the impact of scheduling rules on the production efficiency is performed to assess possibility o
positively influencing certain parameters and determining coupled trade-offs. The results are put into tables
from G.31 to G.36 on page 126.
62
7. R ESULTS A NALYSIS
Figure 7.4 depicts a scatter plot with error bars, where the average total contamination is plotted against
the average bagging throughput. Since the scheduling interventions are based on scenario MassC_K, it is
also included in the plot. Most importantly, it is shown, that a proper choice of scheduling rules can, up to
some point, increase multiple factors of production efficiency. Scenarios Scheduling_D and Scheduling_E
have both higher contamination and lower throughput than scenarios Scheduling_A or Scheduling_B. But,
at some point, with this method of scheduling logic there seems to be a trade-off relation between lower
contamination and higher throughput. Additional experiments are needed to explore this relation, but basing
on this preliminary analysis, there might not be a set of scheduling rules achieving best results in all aspects.
9.5
Total contamination above limits [‰]
9.0
Scenario
K
ScA
ScB
8.5
ScC
ScD
ScE
ScF
8.0
7.5
1050
1060
1070
1080
Average throughput in bags per hour
Figure 7.4: A scatter plot displaying average results for contamination with respect to achieved throughput with error bars for 95%
confidence intervals in scheduling analysis
To add to the insight, a plot similar to 7.2, exploring the same relationship between total contamination
and the average contamination in polluted orders, is shown in Figure 7.5. This time, the corresponding relation is not close to linear. On the contrary, no visible correspondence between these two factors can be
distinguished. Nevertheless, from the limited performed analysis, which does not include scheduling optimisation, there are several conclusions that can be made.
Taking advantage of the mixing sequence is beneficial to low contamination. Chosen dispatching rule for
bagging ‘Earliest mixing time’ sufficiently well performs this function, with significant lead of mixing before
bagging. Thus increasing chances that cross-contamination occurs between products of similar nominal
composition and as a result the participating products are not too heavily affected. Of course, as all scheduling choices, it is not optimised, and comparison with an optimum schedule cannot be made.
Attempts to increase the throughput are, in this case, based on the specific factory layout, and difference
in capacities between packaging lines. By diverging larger orders to line 1 with higher capacity, the number
of needed production changeovers decreases, as there are two machines on line 1 in contrast to line 2, where
there is only one. However, when this methodology is utilised, the number of available orders to bag in the
intermediate silos is an important factor, and the low contamination logic can no longer be utilised. It can
be thus concluded, that there is a trade-off between these two factors, and it should be up to the scheduling
decision maker, to choose one preferred or provide a function weighing these (e.g. total perceived monetary
cost). Only then, can there be judgement made on which rule is the best.
However, there is an important limitation in the analysis – it is performed based on a single set of manufacturing orders. Although these are taken directly from the facility to increase realism, it is possible that
for a different set the results could be entirely different. In the end, this inquiry is to give predictions and
increase insight into the process, and not give the best scheduling solution, which is the outcome not only
of scheduling logic and manufacturing orders, but also operational circumstances, which are not explored in
this investigation.
7.6. R ANDOM C OMPONENT I NVESTIGATION A NALYSIS
63
9.5
Total contamination above limits [‰]
9.0
Scenario
K
ScA
ScB
8.5
ScC
ScD
ScE
ScF
8.0
7.5
0.085
0.090
0.095
Average contamination in polluted orders
Figure 7.5: A scatter plot displaying average contamination in polluted orders with respect to the total contamination, including error
bars for 95% confidence intervals in scheduling analysis
7.6. R ANDOM C OMPONENT I NVESTIGATION A NALYSIS
Random component for cross-contamination calculations is introduced in attempt to recreate similar peaks
and valleys in the contamination curve as measured, and assess the impact of uncertainties in single calculations on the total amount of cross-contamination.
7.6.1. S IZE OF THE R ANDOM C OMPONENT E FFECT
Figure 7.6 shows a scatter plot of results from chosen scenarios investigating the random component size
effect on the average throughput and total contamination.
8
Total contamination above limits [‰]
7
Scenario
J
6
MixB
RaA
RaB
5
4
1072.5
1075.0
1077.5
1080.0
1082.5
Average throughput in bags per hour
Figure 7.6: A scatter plot displaying average contamination in polluted orders with respect to the total contamination, including error
bars for 95% confidence intervals in random component analysis
Random component has a slight influence on the total contamination. Scenarios MassC_J and MixC_B,
which do no include the random component have the exact same achieved throughput, but different total
64
7. R ESULTS A NALYSIS
contamination because of utilised cross-contamination models. If that was not the case, the validity of the
model should be questioned, because a technique called common random numbers (CRN) is utilised i.e.
different scenarios run with the same random value stream, from which the values are drawn. Similarly,
scenarios Random_A and Random_B have the exact same throughput but slightly different contamination.
As all of the scenarios are run with the same set of independent variables for throughput and scheduling, it
should be expected that, due to the CRN the achieved throughput is the same. This is not the case as random
value streams are not differentiated for stochastic variables and scenarios including the random component
draw from this stream much more frequently, and the correlation to runs without the random component
is not maintained. In the future, utilised random value streams for different stochastic variables should be
separated.
When the random component is included, the contamination changes. However, based on the t-test
shown in table H.16 there is no significant difference in means among the results achieved without the random component and when it is set to σ = 0.2 or σ = 0.5. Similarly, no difference for throughput can be
determined for any of the scenarios in this set.
The reason for little difference between using a non-zero random component and not including it, might
arise due to investigation into total effects of contamination and not specific changes in the release curves.
As the total amount of residue swept is the same for both models, the total number of off-specification bags
are not affected by slight changes in the individual contaminations.
That is why a more in-depth inquiry is performed looking at specific orders. Scenario Random_B has 668
recorded polluted orders, while there are 659 such orders for Random_A, for all 20 replications. Comparison
shows, that out of them in 281 cases the order contamination is simulated to be the same, 158 times contamination in orders from Random_B is greater that equivalent one for Random_A and 229 is the other way
around.
It can be concluded that additional efforts in modelling are required to better recreate the contamination
curve. Devised method, even assuming that additional assumptions are valid, is not a means to adequately
model the issue. However, it gives a few important indications, mostly that the assumptions made about
constant value of specific equipment total residue is in all likelihood not valid. Cross-contamination has a
very complex character, which is only in the early stages of analysis and additional data is needed to more
accurately model the process.
7.6.2. F LEXIBILITY E XPLORATION A NALYSIS
The final analysis concentrates on the impact of reserving silos on contamination. It is established in section
7.3, that the biggest impact on total contamination comes from initial order sequencing, which should be
done based on consecutive product nutrient contents.
Tables 7.2 and 7.3 provide one-sided t-tests on the difference of the average throughput and total contamination when six arbitrarily chosen silos are reserved for recipes, which compose some of the contaminated
products from scenario Random_H.
Table 7.2: Comparison of impact of silo reservation on the average throughput for 8 silos
Hypothesis
H 1 :Ra_H>Ra_I
n
mean
SD
t
df
p
95%conf. interval
640
1063.18
46.01
6.2206
473.137
5.452e-10
16.1614: Inf
Table 7.3: Comparison of impact of silo reservation on total contamination
Hypothesis
n
mean
SD
varequal
t
df
p
95%conf. interval
H 1 :Ra_H>Ra_I
40
0.00708
0.00052
TRUE
3.049
38
0.002083
0.00730: 0.00685
It can be thus concluded, that although the effect is small, there is a statistically significant, with 95% confidence, decrease in contamination when silos are reserved for some recipes to prevent cross-contamination
in them. In parallel, because of that action, there is a significant decrease in the average achieved throughput
as a cost of lower number of off-specification products. Indeed, there is a trade-off relation between flexibility
in terms of free choice of intermediate storage silo and production efficiency factors of contamination and
throughput.
7.7. P ROPAGATION OF E RRORS E STIMATION
65
7.7. P ROPAGATION OF E RRORS E STIMATION
Below considerations assume that all investigated variables are independent, and that their errors have a
normal probability distribution. For the numerous observations, according to the central limit theorem, this
is an acceptable assumption.
7.7.1. C HANGEOVER T IME E RRORS
Changeovers C H are independent delays that happen after the end of processing of a product at a machine,
as defined in section 5.3.2. Then, the uncertainty of function basing on these delays is a linear combination
of the exact number of observations for a given machine n, and given by equation:
fC H = n · C H
(7.5)
Then the standard deviation of the total error is :
σ fC H = |n| · σC H
(7.6)
For example, in scenario Scheduling_C there are on average 111 changeovers for BTH3 machine. Then, for
σC H = 123.75 seconds, the total deviation is:
σ fC H ≈ 3.82 [hour s]
(7.7)
While the total average changeover time for the period is:
fC H ≈ 9.92 [hour s]
(7.8)
7.7.2. S HORT S TOPS I NACCURACIES
Short stops are cyclic combinations of two factors: time to failure and time to repair, each with their own
independent uncertainties. Then, the expected time of the cycle is:
S = MT BF +T T R
(7.9)
f S = n M T B F · M T B F + nT T R · T T R
(7.10)
And the total function can be expressed:
Having the same number of occurrences of failures and repairs the standard deviation of the resulting function is:
q
(7.11)
σ f S = n σ2M T B F + σ2T T R
Then, knowing that the standard deviation of an exponential function is equal to the mean, the final formula
can be written:
q
σ f S = n M T B F 2 + σ2T T R
(7.12)
As the time to repair is approximated by a log-normal distribution in table 5.1, the non-logarithmized values
need to be calculated. Denoting distribution parameters µl og and σl og the mean and standard deviation of
the variable’s natural logarithm respectively, the sought for values are:
q
σT T R = (exp(σ2l og ) − 1) · exp(2µl og + σ2l og )
(7.13)
T T R = exp(µl og + 0.5σ2l og )
(7.14)
Then, based on table 5.1, parameter summary for error propagation is constructed in table 7.4.
Table 7.4: Error propagation parameters for short stops
µl og [s]
σl og [s]
M T B F [s]
T T R [s]
σT T R [s]
S [s]
σS [s]
Mixer
3.12
0.93
1426.00
34.90
40.92
1460.80
1426.49
BTH1
1.91
1.20
185.50
13.87
24.90
199.37
187.16
BTH2
2.26
1.22
380.10
20.17
37.36
400.27
381.93
BTH3
2.93
1.17
651.20
37.13
63.57
688.33
654.30
66
7. R ESULTS A NALYSIS
The total number of short stops is not recorded in the simulation, and thus only estimation for a single
cycle is given.
7.7.3. C ROSS - CONTAMINATION U NCERTAINTY E VALUATION
There is no function to describe cross-contamination of a product batch given. This section tries to derive a
generalised way of expressing it, and discuss a method of error approximation. An expansion of the method
presented below, not for a single product batch but for an entire order, would be vital to further use in mixed
integer linear programming methods for optimisation.
Contamination in a product batch C B is then a sum of contamination picked up in the mixer and swept
further along the route until bagged. The following consideration assumes using the partial mass exchange
model, without the random component and all designations as in section 4.4.1. Then, assuming exactly
known route for the product, the function can be given as:
C B = C Mi xi ng +C Rout e
(7.15)
The mixing contamination, given that the batch enters it uncontaminated, is then a function of the residue,
its composition, total material quantity in the mixer and batch mass, see equation (4.18).
C Mi xi ng = f (T R, R i [i t em, quant i t y], n, M )
(7.16)
And the route contamination is a sum of es crossed equipment segments S, denoted by function g S :
C Rout e =
es
X
g S (T R, R i [i t em, quant i t y], E AP, M )
(7.17)
S=1
It is also possible that contamination in a given segment or even entire route is negative, thus decreasing the
total contamination. Then, assuming that material quantity in the mixer and mass of a batch are exact, without uncertainties, and that the current residue composition in a piece of equipment R i [i t em, quant i t y] is a
derivative of previous exchanges with exactly known starting value, the only uncertain variables are T R and
E AP . The contamination functions, presumed differentiable with respect to uncertain variables, are in fact
sets of non-linear combinations of the variables that need to be linearised by e.g. Taylor series expansion.
Furthermore, neglecting correlations between variables, and adding denotion S to route segments parameters, the standard deviation of the formula yields:
v
uµ
¶2
µ
¶
¸
es ·µ ∂C
u ∂C Mi xi ng ¶2
X
∂C Rout e 2 2
Rout e
σ2T R
+
σ2T R +
σE AP
(7.18)
σC B = t
Mi xi ng
∂T R Mi xi ng
∂T R
∂E AP
S
S=1
As in (Ku, 1966, see equation 2.10)
7.8. VARIANCE R EDUCTION
Most figures presented in this chapter are drawn with 95% confidence intervals, implying that the true mean
lies somewhere between the minimum and maximum value, not necessarily being as calculated. This is due
to the stochastic nature of some of the input variables and different outcomes of the scheduling rules. Ideally,
this interval should be as low as possible to signify high certainty of the achieved output, providing there is no
bias in data. In the presented case the intervals might be viewed as large. However, the figures are scaled so
that the differences between achieved outputs are clearly visible. Figure 7.7, a version of Figure 7.1, shows the
dependent variable ranges from zero, and does not allow for distinguishing much. Plotted 95% confidence
intervals indicate, that the relative difference in achieved values is not high, often hardly distinguishable on
the plot.
Yet, it is still possible to reduce the variance by increasing the number of performed replications. All experiments, the results of which are presented in Figure 7.7 are run with 20 replications, while increasing this
number even further would decrease the intervals (see also section on how number of replications is chosen
in Appendix A.3.3). By default Simio uses the same stream of random numbers for a given replication, so that
wherever possible the advantage of common random numbers (CRN) technique can be taken.1 Differentiating random number streams in the simulation, currently running with a single stream, would have positive
effect on variance reduction.
1 See Simio Reference Guide for details, topic "Simulation Replications" in version 7
7.9. R ECOMMENDATIONS
67
Total contamination above limits [‰]
15
Set
Contamination
10
Random
Scheduling
Silos
5
0
0
300
600
900
Average throughput in bags per hour
Figure 7.7: A scatter plot indicating the average variance for investigated scenarios
On the other hand, due to differences in manufacturing order list, and consequences of machine starving time, changeovers, shift duration for a given day etc., the gain cannot be very high. A single change in
sequence or silo allocation of a single order can result in it taking a polluted path and skewing the results for
total contamination, while in another replication with a different random value stream such situation might
not happen. Box plots for throughput from chapter 6 depict a significant number of outliers, which are not
going to disappear with simple increase of replications.
7.9. R ECOMMENDATIONS
The experiment analysis chapter is concluded with a set of general recommendations, that are relevant to
similar cases, including cross-contamination calculations. For case specific recommendations see section
C.7, and for thoughts on future research direction see section 8.4.
First of all, the research shows that it can be extremely useful to include investigation into cross-contamination effects in production environments. By managing to explore product contamination in this way, an
additional aspect of the production performance can be analysed, which is a supplementary layer to typical
consideration providing a lot of useful information, that could help decision makers in their responsibilities.
Thus for most, or possibly all industries with cross-contamination problem, it is recommended to include a
form of such analysis.
Cross-contamination character is directly connected to the equipment properties, especially the total
amount of residue remaining in the system. The more short-term residue, the more carry-over between
products and the larger the cross-contamination. Reducing this amount would have a direct and significant
influence on lowering the amount of off-specification product. This can be done by e.g. choosing a different
piece of equipment, one which has a better cross-contamination characteristic. During the measurements
it was determined, that a pneumatic conveying system has much less residue than screw conveyors. Just as
for Sloten, replacing all the screw conveyors with pneumatic ones of sufficient transportation capacity could
reduce cross-contamination problem, for any other system choosing machinery with consideration of this
factor could help avoiding extensive pollution. Conveyor manufacturers often claim that there is a very low
residue in their products2 , but do not provide any measurements of that characteristic. Including comprehensive analysis for many bulk products would be very expensive, or even impossible, but some indication
2 Based on several internet folders scrutiny
68
7. R ESULTS A NALYSIS
of the extent would be useful in assessing which manufacturer to choose.
Discrete event simulation is a useful way to attempt predicting performance of a future production system, which is not yet fully defined. It allows for relatively easy addition of extra features and exploring impact
of uncertain values, by means of sensitivity analysis. Moreover, it is possible to include stochastic variables in
DES and assessing their joint influence on dependent variables. It is thus recommended to be used as one of
the methods of validating a system design.
Acquiring such deep understanding of the cross-contamination issue, will most certainly help making
more informed decisions on both system design as well as daily operations. Taking it further to create a
decision support scheduling tool, that is integrated into the planning landscape, and that is used to assist
planners in making scheduling decisions, would be very beneficial. Such appliance, able to quickly assess
different scenarios, could help find a scheduling arrangement most suited for a particular period, given specific preferences that are difficult to formalize (like breaks, convenience, preference, experience, workforce
skill impact etc.). Cross-contamination analysis undoubtedly should be included for cases of similar systems.
Finally, investigating a problem, whose composing system has a large number of input variables, because
the system does not exist in reality, is not ideal. To deliver more valid results, it would be better to perform
the analysis on an existing setup, and not combine two different analysis goals in one. But then, assessing
the effects of layout interventions would probably not be the goal, and the analysis would have much more
operational character.
8
C ONCLUSIONS
This chapter first gives the conclusion of the project findings, then author’s reflection on the research, and finally identifies several important directions for future research. First of all, a discussion divided into scientific
part, including research questions answers, and practical implications is given.
8.1. D ISCUSSION
This research presents a unique and novel method, which describes how to model cross-contamination in
a discrete event simulation. Typical models analysing production systems do not include such feature. Although its creation is connected to multitude of assumptions and the use has its limitations, by including
such considerations an improved insight into the problem of cross-contamination in high technology production environments can be created. In this way, organisations can better explore and understand how
technology can be used, to deliver products of better quality, in the end contributing to improving outcomes,
such as customer satisfaction, corporate productivity, profitability and competitiveness.
Performed cross-contamination measurements to determine the magnitude of the problem tell about the
general character of the release curve, and confirm its similarity to axial dispersion problem (Liao & Shiau,
2000). As Leloup et al. (2011) stated, the tracer-collector method is an adequate means of measuring the
extent of the issue, and by adding intermediate sampling points a lot more information can be determined.
Two cross-contamination models built based on this investigation are generic, and could be utilised in other
models or models of different systems.
The achieved results show, that such scrutiny should be included, and more effort needs to be put to solve
the problem of quality decline as a consequence of cross-contamination. Analysing the effects of various interventions in the chosen case shows, that both system layout and production scheduling have a considerable
impact on production efficiency.
8.1.1. R ESEARCH QUESTIONS A NSWERS
Eight auxiliary research questions are formulated in section 1.3, to focus the study efforts. The answers for
them are gathered throughout the research, allowing in the end to phrase them:
1. What is cross-contamination, and why is it important for multi-product factories?
Cross-contamination is a process of mixing material residue, left by preceding products, into the next processed batch. Short term residue is leftover product that sticks to surfaces it is transported through. Because
of the flow and friction, material adhered to them gets loose and is picked up with a stream. Its place is then
taken by some of the powder coming through (Leloup et al., 2011).
It is assumed, because of lack of detailed analysis, that the amount of residue in a particular location is roughly
constant once it is built up, providing the properties of handled material are similar, mostly in terms of particle size, viscosity and friction parameters. Some aspects are difficult to model as there is some amount of
chaotic mixing involved. Though it is expected that such assumptions are in the end not valid.
In multi-product factories cross-contamination might result in a decline in product quality, which can further lead to issues concerning product safety (Fink-Gremmels, 2012). Handling a large number of products
on the same manufacturing line poses a higher risk of cross-contamination, forces them to put bigger effort
69
70
8. C ONCLUSIONS
to production scheduling (Toso et al., 2009) and often requires additional treatments, like cleaning, rework,
or possibly even reject affected portion altogether.
2. What are the relationships between factors influencing cross-contamination?
Developed model distinguishes two major components affecting the magnitude of cross-contamination: the
total amount of residue located in the equipment, and the way this material is released to the incoming product. Although this is not a complete list, the model is capable of sufficiently well recreating behaviour determined during measurements in the case factory.
Then, the more residue in the system, the bigger the total extent of cross-contamination, i.e. more products
are affected. But the precise character of sweeping material residue with the flow is determined by the shape
of the release curve, which is recreated by one of two cross-contamination models in the developed simulation. Comparison made between mixing and partial mass exchange cross-contamination models shows,
that the way residue material is incorporated into the next product has a tremendous effect on final material
composition. In essence, it can be taken very quickly and amassed in front of a batch, or more slowly released
and affecting larger portions of material.
As Fitzpatrick et al. (2004) determine, distinguishing investigated powders based on cohesiveness and friction angles, would lead to better prediction of the possible behaviour. Further inquiry into the total amount
of residue, as well as to shape of the release curve, is undoubtedly needed.
3. What are the effects of cross-contamination, and how are they relevant?
The effects of cross-contamination can be divided into three categories: neutral, unwanted but acceptable
and forbidden. Neutral is when mixing happens between the same or very similar materials, and the resulting
mixture upholds all relevant quality standards. In this case, cross-contamination occurs, but its effects are
limited, if at all distinguishable. In some cases, transferred carry-over contains material that is not specific to
the manufactured product or results in concentrations that are different than intended, but still acceptable.
In such cases, the total product quality declines but there is no reason to reject the product altogether. It
might have to change its product category, destination (client) or require rework of some magnitude, but is
going to be still regarded as sales product.
Finally, there is a chance that some forbidden material is transferred cannot be there, either because of quality
standards or safety reasons. Such product must be discarded and creates significant losses for the company.
That is, if it is actually discovered as the quality control of related issues is often lacking. In the investigated
case, cross-contamination of animal feeds poses risk of even poisoning target animals (Zervas et al., 1990).
4. When is a product considered out of specification?
A final product is considered off specification when it does not meet quality standards that it was initially
intended to. In the case of Sloten and other similar businesses, there are several company-specific categories
of final mixture contents, based on which the products are evaluated. There is a certain level of specification
contents of each of the components ,or nutrients in this case, that are added to create the product, and some
level of acceptable deviation from this specification, called lower and higher threshold levels. If either of this
thresholds is breached, just for one out of many investigated components, the product is considered out of
specification.
5. How to express product-residue mixing in a mathematical way so that it can be used as a general approach
for expressing product contamination?
A means of quantifying cross-contamination is devised and modelled. Two alternative methods, basing on
principles of segmentation, quantity conservation, product similarity and proportionality, are devised. Succeeding in sufficiently well recreating measured release curve shows, that it is possible to create calculation
method for one-by-one indirect reassignment of some material contents between consecutive discretised
product batches. One that is mathematically sound, can be used to investigate an actual production system,
and in the end help in better understanding of the involved issues.
The devised general approach is a series of independent material exchanges, in essence mixing, happening
whenever a material batch leaves distinguished segment of the system. The mixing model always results in
homogeneous composition of a product batch and residue in the equipment, while the partial mass exchange
model has an extra parameter saying how big portion of both product batch and equipment residue is mixed
together. In the end, it is a necessary intermediate step for further investigation (Sen & Ramachandran, 2013).
6. Can cross-contamination phenomenon be modelled in a discrete event simulation paradigm?
Cross-contamination can definitely be included in a DES model, complementing typical investigation into
8.1. D ISCUSSION
71
production with a directed analysis into the issue. By expressing batch contamination and equipment residue
in tables where specific components and their quantities are saved and handled, a generic way of keeping
track of the current components and proportionally mixing them is devised. The biggest limitation is that
it requires a fixed size, small material batches, which results in considerable computing time. Moreover, so
far the ability to recreate peaks and valleys observed in the measurements is rather narrow. Further work is
needed to improve on the method and increase its usefulness, as the ability to include and assess uncertainties can provide a better insight to the process (Mula et al., 2006).
7. What are the effects of additional buffer capacity in manufacturing on cross-contamination and production capacity?
Production efficiency comprises many factors relevant for companies. In general it is concerned with achieving the greatest output with the lowest input, but due to its multiple aspects, and also because of changing
goals, it is often difficult to explicitly present. For Sloten in this research, there were two most important
aspects: line capacity (in terms of finished product throughput) and resulting product contamination. Flexibility manifests as the ability to use multiple pieces of equipment for different products (multi-purpose allocation) and appropriate conveying connections to suit it. Moreover, relatively small lot sizes, agility, quick
response to customer and WIP quantity, are all aspects of this adaptability. Normally companies try to limit
the trade-offs between performance and flexibility by understanding, measuring and constantly improving
their process (Sandborn & Vertal, 1998). Sometimes these are understood literally when a minimum in a
perceived cost function is sought for, especially in daily operations. But when a chance arises, for example
before bigger organisational changes, some seize the opportunity to introduce more systemic improvements
in pursuit for simultaneous increase of several performance indicators.
Additional buffer capacity introduces more flexibility to the process but also more complexity in scheduling.
Adding more silos can help in increasing line capacities, as there is then more work in progress and better
scheduling choices can be made. For the investigated case where a bottleneck is at the end, capacity increase is minimal, only due to a better distribution of products among the machines and slight reduction of
changeovers. However, this occurrence is very specific to the investigated system and might not happen in
another, which has to be analysed separately.
Increased buffer capacity can have a positive effect on the total contamination, where it can be used to reduce the negative effects of cross-contamination by choosing a different intermediate storage, which was
previously used for the same or similar material. Providing the silo size is similar to the order, which is not
split among multiple storage places, the resulting contamination can be limited. However, large increase of
storage capacity poses a risk of much higher contamination, as it increases the total amount of residue in the
system. Fitting relatively small orders in too big silos might result in excessive impurities and in general poses
a risk of higher contamination.
8. Which scheduling rules are beneficial for reducing the extent of cross-contamination, and what is their impact on production capacity?
Most of all, proper sequencing of products can reduce the negative effects of cross-contamination. By making sure that consecutive products in given intervals of the production lines are similar, such mitigation can
be attempted. In the investigated system, scheduling choices of first incrementally sequencing most similar
orders, then fitting them tightly in a single silo with a neutral material residue, and finally packaging them in
an order as close to the mixing sequence as possible, are best.
When certain choices to reduce the overall contamination are made, they have their tow on achieved capacity,
i.e. they limit the available remaining choices and might constrain achieved capacity. Yet, for the investigated
system the negative effect, although statistically significant, was low, and possible gains from less contaminated materials, in the view of company specialists, outweigh the slight decrease in throughput. However, a
quantitative analysis into the total connected costs needs to be made to accurately depict the difference, and
possibly confirm this assessment.
To sum up and answer the main research question, there are many effects of cross-contamination on the
amount of off-specification goods in a multi-product factory when varying buffer capacity and scheduling
rules, most of which are adverse to product quality. The investigation concentrates on the impact of scheduling interventions as well as varying buffer capacity, and determines that both these sets have a significant
ramifications on off-specification goods.
In essence, cross-contamination is unavoidable in a multi-product factory, but its negative effects can be
mitigated. Biggest impact comes from production sequencing, a specific scheduling aspect, which, if done
72
8. C ONCLUSIONS
properly, can limit the negative repercussions of cross-contamination. In the investigated system most of the
residue is located in the beginning of the process and thus the initial sequence is so important. But in fact, all
choices need to consider the effects of cross-contamination.
To a large extent facility layout and available space are relevant as well, mostly because of the total amount
of residue in the system, that has a direct connection to the extent of cross-contamination, and as a result
number of off-specification products. When designing or altering a production system an analysis into specific contamination issues is needed, because once it is created there operational (scheduling or procedural)
interventions to contain it are more limited.
Yet, as Erenguc et al. (1999) claims, all aspects of production efficiency should not be seen as each other
trade-offs but be simultaneously prioritised in search for the best solution. Although a comparison with an
optimum schedule cannot be made, scheduling via dispatching rules seems as a good enough and relatively
simple method for scheduling (Jeong & Kim, 1998).
8.2. P RACTICAL R ELEVANCE
There are operational actions to be performed that help limiting the adverse effects of cross-contamination.
Since it is impossible to avoid it, scheduling similar products next to each other have very strong positive effect on this matter. Then, including intermediate storage silos increases flexibility in the system and can positively influence both throughput and total contamination. However, introducing large intermediate storage
silos poses an increased risk of higher cross-contamination and larger number of off-specification products,
due to higher total material residue in the system. Flexibility increase might be as a result done at a cost of
increasing risk of contamination. It is expected, that several varied in size silos and proper scheduling can
outperform silos of equal size, allowing to fit production orders more tightly in them. Furthermore, reserving
silos for certain product can decrease contamination, but at a cost of decline in flexibility, and possible limited deterioration of the production throughput. The validity of the results is too limited to generalize further,
as the analysis is made based on a single system and a single set of manufacturing orders. Still, it provides a
valuable insight into the uncharted territory of the cross-contamination phenomenon.
A simulation study can help establish boundaries for non-existing production layouts and create insight
into its design. Thus making it possible to make more informed decisions in the design phase, for aspects
like included number and size of intermediate storage silos. By analysing potential performance of a future
system with a relatively cheap method, conceptual changes can be introduced, based on the acquired data
to support the design process, and deliver a better solution in the end. Depending on the case of Sloten, it
can be concluded that an informed design of a facility can most definitely help in limiting trade-offs between
production efficiency and flexibility. Utilising simulation to help determine the impact of different layouts
and scheduling rules with special regard to cross-contamination helps to create insight into the process, and
to give the decision makers quantitative data on proposed solutions. Although the changes in layout are not
fundamental, they are easily implemented in the simulation and can create a difference in performance.
In many cases, also partially for investigated Sloten’s facility in Deventer, cross-contamination poses a
serious threat to food safety issues, and a thorough investigation into it is vital for delivering high quality, trustworthy products. This method has a potential to mitigate some of the risks that are characteristic to multi-product manufacturing lines. Surprisingly, very little data is available on cross-contamination,
and highly theoretical concepts, from e.g. fluid dynamics, have little practical use. The behaviour of crosscontamination is not transparent and a lot more needs to be done to precisely predict its actual character. It is
crucial, when dealing with food, to provide safe, dependable product and a means of its automatic prediction
would be valuable to the industry. Ideally, during cross-contamination it is the nutrients and not products
that should take part in the exchange and be recorded. Then, the number of different fields to record for
every product batch as well as equipment residue would be known and equal to the number of distinguished
nutrients. There would also be no need for recalculation into different frame of reference or to neglect some
products when their content drops below certain threshold. However, the requirement for that is for all nutrients in a product to be registered, expressed as percentage content and sum up to 100%. This was not
possible for Sloten to achieve for the project, and the data was acquired on nutrients was near the end of the
design phase, so the product representation was in the end chosen as a good enough substitute.
8.3. R EFLECTION
This section aims to provide a few personal thoughts on the completed project and give perspective on the
way things progressed.
8.4. F UTURE R ESEARCH
73
Working on the verge of academia and industry is very alluring and challenging. It requires looking at a
problem from different perspectives, and creating a more complete picture of the problem, while each of the
viewpoints is mostly concerned with only a part of it. Having two ‘masters’ implies additional work to satisfy
both parties, and to present the same facts differently, in a way that is better understood and preferred by
them. For a long run it could get frustrating, but I was more amused by, curious of and even a bit worried
about the gap that separates these mindsets.
In this research simulation is used to perform quantitative investigation into the problem of production
efficiency trade-off for versatility, in an animal feed manufacturing industry. Showing that till certain level all
parameters can be improved simultaneously, and after that there is a clear trade-off relation, is from perspective better outcome than anticipated when the project commenced. Simulation as a methodology is a very
powerful means of determining possible performance in systems that are too expensive to investigate or not
existing. Discrete event simulation in particular is a robust and precise method that should be used much
more often in research as well as for process improvement.
Lack of proper literature on the topic made it more difficult to commence with the project, especially with
the design of cross-contamination models. This gap of knowledge should be filled to help to understand and
counteract these as well as enable future process improvements. I hope that I have contributed to this notion,
as solving real problems and drawing general conclusions that could help further, is probably the most stimulating and rewarding occupation one can have. I am thus hoping to be one day able to go a step further and
apply the principle in the food industry. Yet, it is my belief that with more solid theoretical background for
cross-contamination phenomenon and more knowledge about the tracer–collector measurement methodology beforehand, it would be possible to influence the factory contamination trials conduct to gather data of
higher quality.
Managing to devise a model of cross-contamination nevertheless was gratifying, especially that its final
form is very simple, customizable and does not base on complex theories from chemical engineering field or
fluid dynamics. Method seems quite neat, powerful and allowing for further development.
Surprisingly, the biggest challenge to finalise the thesis came from my favourite part – writing. Being normally able to endlessly spew fairly cohesive and well written text on virtually any topic, I had to constrain
myself to the rigidity of a scientific report, and to writing in a plain, easily understandable manner. What
a nightmare it is, not to be able to use humour, digressions, vague allusions, or to lesser extent, other languages (quidquid Latine dictum sit, altum videtur). Constant revisions of text so that it is more clear, better
understood and cohesive are, despite their obvious need and value, the biggest discouragement from trying it
again. If ever facing a binary choice between a scientific text and a work of fiction, I will most certainly choose
the latter. The reason is, perhaps, that my thesis is written for, at least in theory, scientific community, and
my own scribblings are primarily directed to my selfish ego. . .
8.4. F UTURE R ESEARCH
As the research presented in this report is describing area, which was never investigated thoroughly before
there are a lot of possible directions to take for further analysis. Most likely, animal feed industry is going to
expand, increasing their complexity of operations while trying to limit costs and increase production. The
following proposal of future research possibilities is divided into two parts: method improvement (scientific)
and method application (practical).
8.4.1. A CADEMIC R ELEVANCE
Presented model and research are the first step in understanding and predicting what happens or could happen during production in terms of cross-contamination, and what mechanisms describe this phenomenon.
There are therefore, several distinct areas of inquiry to be listed. First of all, a further analysis of actual and
perceived trade-offs arising from cross-contamination effects should be made. To enhance the approach
taken for this research and allow for better analysis of the trade-offs between production efficiency and versatility, several improvements could be made. Firstly, an inquiry into the trade-off perceptions of managers
and their view of the importance of cross-contamination of several facilities would be extremely useful in
determining the willingness to face them to find a mutually beneficial solution, if possible. Moreover, the
trade-offs expressed during the analysis do not have clear weights and have to be interpreted subjectively by
decision makers. While this not necessarily has to be a bad thing, a more objective performance indicator,
such as reducing trade-offs to monetary value and interpreting costs.
Moreover, a clear single objective could prove useful in the future for optimisation. If cross-contamination
74
8. C ONCLUSIONS
carries hidden quality depreciation, as analysed in this report, then attempts can be made to minimise this
cost or overall cost in system design, or operational scheduling.
In order to optimise, a generic scheduling approach, capturing the essentials of cross-contamination effects, needs to be developed. Current solutions involving non-triangular setup times for cleaning (Toso et
al., 2009) are insufficient, and need to be improved to include acceptable limits of carry-over, possibly distinguished for sensitive ingredients. Such method would be much more difficult to define than the usual mixed
integer formulations in similar problems, and most likely much more demanding computationally.
Furthermore, improving the modelling method to capture the relevance of physical properties of the material, its flow and the equipment, to create a universal description of the problem, if that is even possible,
should be attempted. Representing exact chaotic movement of particles might be extremely difficult, and
the benefits coming from it limited. In substitution, a smooth curve with a safety factor could be used to
sufficiently well model the phenomenon. Then, to include a differentiation among product properties, the
assumption made of constant residue in the system needs to be set aside, and the mass balance kept for the
system in total, not for each of the material contents exchange.
Finally, as cross-contamination occurs especially with edibles, a more thorough risk analysis into its effects on health should be made. Current methods involve reactive measures to determined liabilities, while
with a proper investigation method, a hazards could be identified and eliminated before they occur, with
significant benefit to public safety.
8.4.2. P RACTICAL A PPLICATIONS
In terms of application a natural advancement seems to expand analysis to other powder-based systems that
encounter similar problems. First of all, for other feed production facilities, in which characteristics and
complexity of issues is different, to establish a more general method, applicable to the entire industry and
tested on multiple cases.
Furthermore, research can be expanded for human food with much stricter regulations and possible
safety repercussions than in the animal feed case. As the industry is much bigger, the possible use is also
greater. Being able to reduce lot sizes and gain a lot of flexibility, while limiting efficiency loss, would definitely be interesting to plant managers, that could more quickly respond to market needs without bearing
considerable costs.
Moreover, the method could be applicable for non-edibles like metal powders, chemicals, plastics & polymers or minerals. There are many industries that use common conveying equipment for various products
without realising or caring for resulting pollution. An analysis into its extent could help design better systems
or use them in a more efficient manner.
There are also possible extensions to the simulation model, that could better capture characteristics of
the real system and provide improved insight into it. Built simulation model could be further expanded by
adding or improving features listed below:
• Dynamic silo allocation with additional checks during mixing (e.g. when current silo full),
• Perform nutrient similarity sequencing based on a directed search or full factorial analysis,
• Taking into account specific product properties (e.g. viscosity, friction) to distinguish products from
each other to determine impact of that on cross-contamination extent,
• Including nutrient analysis for all products,
• Include nutrient similarity check for bagging dispatching rules,
• Adding prediction possibilities for bagging dispatching rules, including forcing idle time in anticipation
of a better sequence.
Obviously, increasing validity of the model, by measuring an actual production system and not just its
concept, would prove useful for confidence in results, in terms of not only throughput but further measurements of contamination. Designed model is validated based on partial data and specialists assessment, but
quantitative analysis of the entire system is invaluable.
Finally, this analysis with the simulation model is a first step to an on-line scheduling tool for operations,
that could include learning capabilities to capture the knowledge of planners. Being able to expand usefulness of the method, making it a composite solution, could only encourage potential users to undertake it.
All in all, possibilities for expansions are considerable and should be made in the future to provide more
value to both science and industry.
Appendices
75
A
M ODEL D ATA
A.1. M ODEL I NPUT D ATA
Table A.1: Short stops data
BTH1
BTH2
BTH3
Mixer
Failure number
6539
4062
1990
1060
Total Downtime [h]
26.81
22.30
16.81
8.87
Total Staffed Time [h]
363.82
451.19
376.78
428.73
MTBF [h]
0.052
0.106
0.181
0.396
MTBF [s]
185.5
380.1
651.2
1425.9
Table A.2: Changeover time statistics
BTH1
BTH2
BTH3
Mixer
Observations
287
295
310
733
Mean [s]
277.27
256.93
235.00
266.91
SD [s]
128.08
116.45
123.53
248.12
Median [s]
271
246
213
246
Min value [s]
62
61
62
61
Max value [s]
658
603
625
1048
Mean [s]
16.37
22.34
21.34
35.66
SD [s]
34.66
41.53
60.15
51.39
Median [s]
6
8
20
25
Min value [s]
1
1
1
1
Max value [s]
580
599
628
661
Table A.3: Time to repair statistics
BTH1
BTH2
BTH3
Mixer
Observations
5895
3593
1634
896
A.2. C ONTAMINATION M EASUREMENTS A NALYSIS
Several weeks after the trials the laboratory results came in and first conclusions could be made. The most
important reference values were measured to be:
Table A.4: Contamination results reference points
Fe ppm
432
20
8
Protein %
12.22
19.46
19.93
Material
First material (contaminant)
Reference value with contamination
Second material (without contamination)
77
78
A. M ODEL D ATA
For each of the points A to F normally 16 – 17 samples were taken over time, which allowed to plot contamination curves for this point. Measured points, in relation to the reference sample, are expressed as a
ratio, and plotted with straight-line connections between points. Cross-contamination measurements for all
the points are shown in Figure 5.2 on page 36. These lines show exponential behaviour and all are based on
the reference value of 20 ppm Fe, as such displaying picked contamination from the start to the measurement
point and not incrementally from point to point.
M IXER C ONTAMINATION A NALYSIS
Below calculations show how much residue was picked up by the second material before the mixer and inside
the mixer. Assuming that the route was entirely polluted, there was some amount T R Mi xer of the first material
with content of 432 ppm Fe at a point in the mixer, together with 5700 kg of the second material with 8 ppm Fe,
that was added during the trial. Resulting uniform mixture had iron content of 20 ppm and thus an equation
can be formulated as the total amount of iron needs to remain constant:
5700 kg · 8 ppm + T R Mi xer kg · 432 ppm = (5700 + T R Mi xer ) kg · 20 ppm
(A.1)
Solving the equation the total mass of picked residue is calculated:
T R Mi xer = 166.0194 [kg]
(A.2)
Similar calculations for protein yield over 370 kg of contamination, which is deemed way to excessive and
highly improbable by the operations expert, and thus is rejected. During the project period, the main mixer
was specially cleaned and weighed before and after that process. Operators reported that the difference
amounted to almost 80 kg. The rest of the R M value would be picked up before the main mixer. Either
way, this shows that there is a considerable amount of material residue in the system and since this is the
only possible route this cannot be helped by scheduling. Redesign of that part or special cleaning in between
is not considered in this research.
C ONVEYORS C ORRELATIONS
One of the aims of contamination modelling is ability to generalise the obtained curves to fit other, not measured equipment. That would mean finding correlations among significant parameters of the given equipment, and finding a general expressions to be applied to any other ones in the system. At first these significant
parameters need to be defined and found. An example of such parameters for investigated screw conveyors
is shown in Table A.5. Interval C –> D is not investigated due to fewer and inconsistent measurements.
Table A.5: Parameters of the investigated screw conveyors
screw
length [mm]
diameter [mm]
compartments
rotational speed [rpm]
pitch [mm]
angle of inclination [deg]
particle speed [m/s]
internal area [m2 ]
volume [m3 ]
Residue [kg]
Mixer –> A
A –> B
B –> C
3000
500
1
64
320
0
0.3413
7.9168
0.3770
32.7778
14000
400
2
294
250
90
1.225
29.6968
0.6333
23.1667
17090
380
3
306
280
45
1.428
33.2932
0.9785
30.1667
First parameters from Table A.5 relate to physical characteristics of the conveyors (or the sum if there were
more conveyors in between measurement points, see the number of compartments). Further come from the
measurements, and are fitted for the incremental curves. Finally, a total residue is the conveyor is estimated
by integrating the curves. In search for cross-correlations among the values, an Excel in-built correlation
function is used. Because of the limited number of measurements, the findings can hardly be expected to be
perfect. The results are put into Table A.6 and are inconclusive.
A.2. C ONTAMINATION M EASUREMENTS A NALYSIS
79
Table A.6: Correlation parameters of the investigated screw conveyors
diameter [mm]
rotational speed [rpm]
pitch [mm]
angle of inclination [deg]
particle speed [m/s]
internal area [m2 ]
volume [m3 ]
Residue [kg]
length
[mm]
diameter
[mm]
-0.99854
0.98622
-0.79519
0.74266
0.99943
0.99689
0.92165
-0.98542
1
-0.99371
0.82675
-0.77771
-0.99979
-0.99969
-0.89937
0.97480
rotational
speed
[rpm]
pitch
[mm]
1
-0.88452
0.84319
0.99122
0.99618
0.84479
-0.94371
angle of
inclination
[deg]
particle
speed
[m/s]
internal
area
[m2 ]
1
0.76471
0.79308
0.42464
-0.61790
1
0.99897
0.90811
-0.97915
1
0.88824
-0.96896
1
-0.99661
-0.81509
-0.84047
-0.49763
0.68044
volume
[m3 ]
1
-0.97423
Table A.7: Parameters for correlation analysis for all measured equipment (excluding points G and Bag)
screw
Mixer –> A
A –> B
B –> C
C–>D
D–>E
E–>F
F–>G
G–>Bag
length [mm]
diameter [mm]
internal area [m^2]
volume [m^3]
Residue [kg]
3000
500
7.91681
0.37699
31.573
14000
400
29.69684
0.63334
27.9508
17090
380
33.29318
0.97849
25.7703
11000
380
21.29772054
0.60670
33.4559
12700
3600
154.62320
130
90.982
60000
150
28.273
1.06028
15.476
51.675
61.579
Table A.8: Correlation results for all equipment parameters (as in table A.7)
length [mm]
diameter [mm]
internal area [m^2]
volume [m^3]
Residue [kg]
length [mm]
1
-0.24992
-0.09062
-0.16325
-0.38089
diameter [mm]
internal area [m^2]
volume [m^3]
Residue [kg]
1
0.97461
0.99588
0.98663
1
0.98673
0.93875
1
0.97141
1
Although there are promising values for relation between the total residue and volume (0.97423) or internal area (-0.96896), no expected relationship could be made, because of the limited number of samples.
Additional analysis needs to be made before any results are accepted and generalisation made. In the end,
the non-measured intervals of screw conveyors are assigned total residue T R that is proportional to their volume and corresponding measured interval’s volume. Thus, when the volume changes in the simulation, for
e.g. included extra silos, the total residue changes as well.
P NEUMATIC C ONVEYORS
This type of conveyors contains least amount of residue out of all the measured equipment, according to
the trials. Because there was only a single long section of pneumatic conveyor investigated no correlations
can be found, and the generalisation bases on the findings from screw conveyors. Assuming that residue is
also in this case correlated with pipe volume then with the same diameter in all pneumatic conveyors, the
amount of residue can be expressed as material per running meter. The investigated portion of the conveyor
has approximately 60 m, and had 15.476 kg residue total. That yields:
· ¸
15.479
kg
R(l ) = l ·
= 0.25793 · l
(A.3)
60
m
Where l is length of a given interval.
Table A.9 shows the parameters obtained by using sum of squared deviations method to find exponential
curves for contamination.
Table A.9: Accuracy of the exponential curve fitting expressed with the sum of squared deviations method
Point
SSTotal
SSRegression
SSResidual
RSquared
StDev
A
B
C
D
E
F
G
BAG
0.172808
0.171882
0.000926
0.994640
0.007381
1.729847
1.597465
0.132382
0.923471
0.090960
0.904290
0.901128
0.003161
0.996503
0.014057
0.983347
0.967468
0.015878
0.983852
0.037993
0.872396
0.611294
0.261101
0.700707
0.127745
0.388773
-0.154848
0.543621
-0.398299
0.184326
0.696348
0.458486
0.237861
0.658416
0.121927
1.454125
1.409095
0.045030
0.969032
0.051466
80
A. M ODEL D ATA
A.3. M ODELLING AND E XPERIMENTATION C HOICES
This section contains considerations on specific choices to be made for proper model execution, so that
necessary input variables are chosen after appropriate deliberation. At first, modelling consideration about
entity size is done, then when to neglect certain contamination and finally what number of replications to
choose for reliable experimentation results.
A.3.1. E NTITY S IZE AND N UMBER
Choice for entity (dynamic object) size, a discretisation of continuous material flow into small individual
product batches is done, based on the production specification introduced in chapter 5. General model as
shown in Figure 3.3 assumes there is a uniform size of all product batches but does not specify its value, which
needs to be explicit for the chosen system.
In the end, most of the products need to be packaged into bags having size of 10, 20 or 25 kilograms.
Natural, easiest choice suggest taking their highest common divisor – 5 kg, from which contents for all these
bags can be assembled. However, such selection poses a risk of having a lot of entities in a system at any given
moment, which would impede execution speed. Because of cross-contamination, the entities need to store
their product contents, and the record needs to be available during every exchange calculation. For that it
is easier to store these values in each entity individually, but then operations of merging them into a single
entity in storage would require additional logic and value storage.
Assuming that most material in the system is stored in the intermediate silos, the maximum capacity of
which (in the basic case) is 535 tonnes, then by having them all filled to the maximum, they would contain
107 thousand entities of 5 kg. This is an unrealistic, ‘worst case’ scenario and the actual number of them in
a system should be lower, as the silos are most likely not going to be full all the time. Still, considerations are
done on this number, to assure smooth execution even in unlikely situations. Having this many active entities
is, based on field experts experience, not possible in Simio, which would probably crash during execution. It
is thus required to put the entities in storage into some form, that is less expensive for model execution or
perform some actions to lessen that impact.
Additional tests with simplified Simio models are performed to determine a sufficient method for modelling. Roughly 100 thousand entities are put into several Simio constructs called a station1 and then material
flow movement to and from the stations representing silos is approximated. The speed is relatively low, but
is deemed acceptable. However, additional layer is added to increase the speed, which involves process of
batching entities entering a station into a single entity (called batch entity). Organisation of having a silo
represented by a station, in which there is constantly a single batch entity, which has then a collection of the
actual product batches that enter and leave it in a FIFO manner, is empirically deemed faster and thus utilised
for the final production model.
A.3.2. C UT- OFF P OINT
There is an amount of material in a given place which can be considered negligible for the outcome i.e. even
if a given product is sensitive to it releasing its entirety into the product batch in subsequent stages would
not change the outcome on its own. There is a chance, that composition of factors might eventually have
a higher impact, but due to modelling limitations the limit when a given product is considered negligible
should be established. In general higher precision comes at a cost of extended computing time and thus the
cut-off limit is a variable in the model. The following considerations derive a limit deemed acceptable for the
vast majority of cases. Moreover, is the limit is too high, the biggest repercussion is considering up to a few
finished product bags as acceptable, whereas they should be counted as outside of specification, most likely
together with earlier identified ones.
The biggest difference in between one product’s specification level and the higher level of another determined for Sloten is 2880 times, which could happen only in a very extreme case. This is a difference of
bacteria content between 1.44 · 108 cfu (colony forming units) and limit of 5 · 104 cfu (with specification of 0).
Supposing there is total amount of X kilograms of material in a finished product batch of 5 kg. Then, basing
on only this component and these two product it should within:
1 See Simio Reference Guide for details
X · 1.44 · 108 < 5 · 5 · 104
(A.4)
X < 1.736 · 10−3
(A.5)
[kg]
A.3. M ODELLING AND E XPERIMENTATION C HOICES
81
The biggest number of intervals in between the main mixer and bagging is 17 and thus assuming that at
each of them the product batch is picking all of the contents, leaving no behind the maximum equal picked
material x at each stage would be:
x = 1.02 · 10−4 [kg]
(A.6)
This is highly unlikely as picking up entire residue without leaving anything behind is against cross-contamination modelling specification and the cut-off value should be higher. In fact, whenever given product
contamination is higher that equipment residue it strives to equality with every exchange and thus the inequality from (A.5) holds for every exchange and not only totality. Still the amount of just above one gram is
very conservative and is in practice relaxed to the next order of magnitude to 0.01 kg. Practical screening of
results with chosen threshold of 0.01 kg and 0.0001 kg revealed no apparent differences.
There is a chance, that there are multiple trace amount of products which together constitute for significant amount of a given, sensitive nutrient and could have a negative impact on the bag acceptance determination, but this is unlikely. Although higher precision is desired, the chosen negligence threshold of 10 grams
is deemed sufficient for great majority of cases.
A.3.3. R EPLICATION N UMBER
Each defined scenario must be run several times with different pseudo-random number generator seeds to
account for stochasticity in the model an give an accurate prediction of the expected value (or output possible interval based on some confidence level). Although in the designed model the number of stochastic
variables is low and include only changeovers, short stops and possibly random exchange component for
cross-contamination calculation, their impact might be profound on the performance. Mixing sequence remains fixed for a given scenario but the chosen silos and bagging order do vary from replication to replication,
creating a ripple effect since it first occurs.
Therefore the necessary number of replications is investigated first manually and then using statistical
tools to find an optimum number. In general, the more replications are run the more certain can one be
about the results but that comes at a cost of computing power (Kelton et al., 2013). In this case for used
Intelr CoreTM i5-4300U CPU and 64 bit Simio with 4 parallel replications the run time is around 70 minutes
without contamination calculations and exceeds 200 minutes including cross-contamination for gathered
statistics of 3 weeks of production time. Therefore it is vital to limit the number of replications per scenario
to a comfortable level in order to explore more than just few options.
A comparison of the results for one sample t-test for 95% confidence interval for different number of
replications is given in tables A.10 and A.11 for three weeks production and a single day of warm-up period
(excluded).
Table A.10: One sample t-test results for the number of late orders for 95% confidence
t-value
d. freedom
p-value
interval low
interval high
mean
n=10
9.8466
9
4.069e-06
8.01071
12.789287
10.4
Late orders
n=20
13.369
19
4.093e-11
8.60311
11.79689
10.5
n=40
20.466
39
< 2.2e-16
9.732637
11.867363
10.8
Table A.11: One sample t-test results for average number of processed bags per hour for 95% confidence
t-value
d. freedom
p-value
interval low
interval high
mean
Avg Bags per hour, n = 10
BTH1
BTH2
BTH3
260.24
218.72
277.53
149
149
149
<2.2e-16
<2.2e-16
<2.2e-16
351.7474
352.0303
354.5835
357.1299
358.4490
359.6689
354.4387
355.2396
357.1262
Avg Bags per hour, n = 20
BTH1
BTH2
BTH3
407.83
342.34
410.42
299
299
299
<2.2e-16
<2.2e-16
<2.2e-16
354.2724
355.1371
356.1618
357.7080
359.2437
359.5938
355.9902
357.1904
357.8778
Avg Bags per hour, n = 40
BTH1
BTH2
BTH3
539.39
471.18
575.29
599
599
599
<2.2e-16
<2.2e-16
<2.2e-16
354.2870
355.4477
356.7238
356.8763
358.4232
359.1677
355.5817
356.935
357.9458
It is concluded that having 20 replications giving well enough results of experiments with acceptable confidence interval for 5% significance level as well as manageable runtime. Thus, all performed experiments are
to have 20 replications, unless otherwise specified.
B
VALIDATION D ATA
This chapter contains historical information considered sensitive by the company, and is left blank in the
public version.
82
C
C ASE STUDY
This chapter provides information on the company, their goals, production process and more case-specific
data hat can be addition to general summary presented in the main text.
C.1. P ROJECT B ACKGROUND
Sloten B.V., a subsidiary of Nutreco corporation, is a major developer and supplier of young animal feeds, taking care of these sensitive animals by meeting their specific nutritional needs with high-quality products. To
optimize the capacity and efficiency in the Sloten facility in Deventer, a plan is devised for a thorough plantupgrade, which is to be evaluated. It is desired to further analyse the capacity of the proposed plant-upgrade
based on real customer order, product and production data. As it is very likely that the new manufacturing
layout will require different scheduling rules or even a completely new planning approach, an in-depth inquiry into current state of the art knowledge is to follow. A discrete event simulation (DES) including stochastic effects is requested to analyse the performance, and be the base for determining additional new structural
requirements and scheduling.
The company started the project as it is mostly interested in gaining insight about the future plant operations, specifically about the possible capacity of the proposed upgrade and above all – about the impact
of cross-contamination on production and product quality. By having an outside party come to the factory
and ask questions about their process, they believe to be stimulated to investigate the right issues and improve the manufacturing performance. One of the most important requirements in their production process
is keeping strict cross-contamination procedures to prevent unwanted or excessive impurity of a product,
due to collection of residue from a previously produced one and sent through the same piping system. Thus
there is a desire to keep a single production for as long as possible, by e.g. joining orders and making sure that
the next product in line is not in violation with the defined contamination rules. However, the consecutive
products can also be sequenced in an order that mitigates the impact of cross-contamination, due to feed
ingredient similarity, and some products cannot be put in sequence without a cleaning run, that takes time.
Specifically, contamination has two main aspects. Firstly, due to the quality standards, the amount of different than specified ingredients cannot exceed certain threshold. Additionally, some products are susceptible
to particular ingredients (which potentially may even be harmful to the animals) and are allowed to contain
only much stricter, trace amounts of those. It is also desired to perform cleaning with sales products, that
are not sensitive to a given type of contamination. If that is not possible, certain amount of product can be
discarded, or reworked, which generates losses.
Moreover, it is requested to investigate and determine the number of needed additional intermediate storage silos and overall line capacities. In the end, basing on real manufacturing orders, a feasibility of the upgrade is to be assessed. After such, the company management will be presented with the results and will determine whether to undertake it. A schematic of the proposed production scheme is shown in Figure C.1. The
new approach is to introduce more flexibility into the system and more routes. With more choices and complexity, the handling of scheduling is going to be more challenging, and the extent of cross-contamination
due to longer routes greater, limiting production efficiency. Thus a thorough comparison between possibilities of the new approach and limitations of the old one is necessary to convince the management whether
this is a viable improvement. Before the production can be rearranged as indicated in Figure C.1, a feasibility
83
84
C. C ASE STUDY
Figure C.1: A schematic of the proposed production organisation, BTH1-3 are bagging lines
study determining its performance and requirements needs to be carried out. Most importantly, assessing
the level of cross-contamination that is to be expected, and preparing a suitable scheduling approach. The
management identified several issues that should be addressed and improved, these are:
• Mixing speed limited by bagging lines,
• Downtime interlinking for dosing, mixing and packaging,
• Missing route options and limitations on line combinations (with negative efficiency impact),
• Low equipment utilization,
• Limited use of finished product silo.
In the end the Sloten management is to be provided with an answer on the feasibility of production process
transformation, its requirements in terms of resources and scheduling as well as an assessment of possible
improvements (e.g. in terms of capacity) that could result from it.
C.2. P LANNING AND S CHEDULING IN S LOTEN
Planning procedures are normally a result of organisational choices and differ among the companies. In
Sloten, planning is largely detached from scheduling (sequencing of manufacturing orders within a day) with
some overlapping. The process is illustrated in Figure C.2, and starts with either customer order or low stock
of some, frequently sold items. The main tasks of planners is to fit these orders to daily plans in accordance
to their needed lead times and lot sizes. Two days ahead the daily plan for manufacturing is released and
sent to production department. Manufacturing has normally freedom to sequence the orders in the manner
they see fit, providing that contamination rules are abided. As the current process is much simpler from
the investigated one, due to direct connection of mixer and bagging, in the new outline there requires more
choices to be made, because next to mixing sequence silo capacity for the order needs to be allocated and
then the sequence of bagging from the silos. There are two main sources of manufacturing orders: customer
orders and low inventory of some make-to-stock items. Planners are then responsible to consolidate orders,
fit them into daily schedules, and to determine final lot sizes. Since customer orders might come with a
limited advance, there is a constant need to rearrange the plan until it is released two working days ahead of
the manufacturing date. All the planning is done manually within an ERP system, without specialised tools
to help.
Daily production plan is then sequenced by the manufacturing crew (main mixer operators), who compose it manually according to their expertise, preference and relevant contamination rules. The contamination rules are in form of written-down guidelines, often prohibiting certain sequences and giving minimum
material throughput in between them. Nonetheless, there are no strict procedures on how to prepare the
daily schedule and the process is very subjective. Moreover, until recently there was no centralised contamination rules register and various responsible parties were only given seemingly relevant pieces of information
with no central site for the full knowledge.
C.2. P LANNING AND S CHEDULING IN S LOTEN
Sales Orders
85
ERP System
Low stock of
item
ERP System
In case of make-toorder item
Different lead times
In case of make-tostock item
Characteristics
Generates
request for
planning
Generates
alert for new
order
Size determined by
planner, often fitted
to available capacity
Characteristics
Needed extra time
for quality-check
Also used to fill
capacity in case of
smaller number of
orders
Recipe consolidation
Constant
shifting and
reordering
Pre-ordered
list for each
day
Daily capcity constraints
Characteristics
Separate orders for
bagging and mixing
Facilitates rush orders
Released
daily plan
Characteristics
Allows for some flexibility
for part of orders
Two working
days ahead
PLANNING
Based on product
availability in
intermediate storage
Approved by both planning
and production
Often starts in the
morning from material
mixed the previous day
Different priority
Characteristics
Daily
production
plan
Bigbag
schedule
Capacity dependant
on workforce
Roughly once/twice a
week
Mixer
schedule
Normally mixed the
previous day
Based on specific
contamination rules
Characteristics
Manual silo allocation
Limited contingencies
Stations can be filled
in advance
Must have
transportation
capacity
Characteristics
Bulk orders
Small bags
Department 2
Standard location for
10kg bags and
colourants
Characteristics
Only bagging
disconnected from
mixer
Direct interlinking
with the mixer
Characteristics
Manual silo allocation
Rough schedule based on
average capacities
Small bags
KLM
Biggest production,
most changeovers
Low capacity and old
machine
Scheduling
Figure C.2: A flowchart of the current production planning and scheduling characteristics
Therefore for the researched problem, the total production plan is to based on the actual historical data
split into fixed due dates. Orders within a daily production plan can be rearranged freely in terms of mixing
86
C. C ASE STUDY
sequencing, silo allocation and bagging order, and are to be subjected to the simulation model logic and
its scheduling rules. As the simulation model is to give indication on the trade-off between manufacturing
versatility, flexibility and efficiency loss as well as validating a plant layout design, further investigation into
planning is not necessary at this point. Nonetheless, integrating short-term planning with daily scheduling
in a single APS solution could bring further benefits in terms of better production schedules.
C.3. P RODUCTION KPI DEFINITION
There are two main categories of key performance indicators (KPIs). First is connected to overall average
system performance (mixing and bagging capacities, silo utilisation or working time), and the other is order
specific, meaning it creates a separate set of statistics for each order (contamination, throughput). General
statistics are gathered in set intervals.
G ENERAL KPI S
General KPIs are mostly gathered in intervals, averaged over the measured period. It is assumed that during
working hours the equipment is staffed at all times.
Mixer
Mixer works continuously during working hours. If all orders for that day are finished it continues working,
providing it estimates there is sufficient time to finish the order during the shift. Records the time last order
is finished and compares it with the shift ending time (can be a little late if there are many stops).
Daily statistics:
•
•
•
•
•
•
•
Time available [hours]
Under- or overtime to the schedule [hours]
Time working [hours]
Total downtime [hours]
Total changeover time [hours]
Time starving [hours]
Total throughput during working time [kg]
Bagging
For each of the bagging machines (BTH1-3) the following set of statistics is recorded.
Daily:
•
•
•
•
•
•
•
•
•
Time available [hours]
Under- or overtime to the schedule [hours]
Time working [hours]
Time working bagging late products [hours]
Total downtime [hours]
Total changeover time [hours]
Time starving [hours]
Total throughput during working time [bags]
Total throughput during working time [kg]
Hourly:
• Throughput [bags]
• Throughput [kg]
Silos
Each of the silos is measured to provide the following indicators.
Daily:
• Average weight [kg]
• Time empty [hours]
• Ratio of average contents weight to silo capacity during a non-empty time (utilisation) [-]
Every 15 minutes:
C.3. P RODUCTION KPI DEFINITION
87
• Material weight in the silo [kg]
• Ratio of contents weight to silo capacity [-]
Before mixing start three availability categories (with sub-categories):
• Total number of silos where given order type could be stored, based on system output destination
• Number of available silos to discharge to:
– With the same material
– Empty
• Total usable space in available silos [kg]
• Number of blocked silos:
–
–
–
–
Due to contamination rules
Due to scheduling constraints
With the same material but full
Non-empty with different materials
In order to start mixing, there needs to be at least one available silo for allocation and the total available space
in silos needs to be equal or greater to the order size.
O RDER KPI S
For each of the processed orders the following statistics are recorded.
Performance statistics for every order:
•
•
•
•
•
•
•
Time started mixing [hours]
Time ended mixing [hours]
Time started bagging [hours]
Time ended bagging [hours]
Order quantity [kg]
Order package size [kg]
Whether order was finished on time [-]
Contamination:
The statistics for residue within equipment are not directly shown. Contamination is thus only measured separately for each bag that is leaving the system and which contents comprise of the sum of entity composition
within it. The bags are then classified into two contamination categories: within or out of set contamination
limits. Regarding a bag as above the limit does not necessarily mean that is should be rejected due to its quality. Once the final criteria on how to evaluate the acceptance or rejection of a given bag are made, it is desired
to provide the following statistics for every order:
• Count of bags above limit
• Amount of material above limit (bag count multiplied by bag size) [kg]
• Ratio of order of material above limit (amount of material above limit divided by the order size)
OTHER S TATISTICS
There are also other performance indicators recorded, that cannot be assigned to any previously mentioned
category.
Report of order arrangement:
As a reporting feature, an outcome set of schedules
• A list of consecutive orders in the mixer with starting and ending times
• A list of utilised silos for each order
• A list of consecutive products at bagging machines with starting, ending times and source silo
All aforementioned statistics are shown to the user in specially designed for the purpose environment in
Scenario Navigator software.
88
C. C ASE STUDY
Run ending statistics:
When a simulation run is ending the following information is to be stored:
• Number of completed orders
• Number of completed orders on time
• Percentage of the simulation time when different numbers of silos were empty
C.4. C ONTAMINATION T RIALS
The contamination trials were conducted in the facility on 05.03.2015. They consisted of two main steps.
At first, the investigated route (shown in Figure C.3) is contaminated entirely by a product with high iron
content, with nominal value of 432 parts per million (ppm). After that, a new product with a low iron content
and a reference value of 8 ppm is sent through. Because the mixing process in Sloten is continuous, the trials
aimed to measure in different places in the system (see Figure C.3) the amount of iron after a certain amount
of material was sent through. Due to difficulties in frequent sampling, some measurements were made in
quite big intervals, making a good result analysis more challenging. Contaminating with a high iron content
SILO 19
SILO 20
SILO 21
SILO 22
SILO 23
SILO 14
SILO 15
LINT MENGER
1
Buffer
afzak
BTH1
Buffer
afzak
BTH2
Buffer
afzak
2
3
4
5
6
7
8
9
10
Buffer
afzak
BTH3
Matam
SAMPLE F
SAMPLE G
Buffer
afzak
1
2
Arodo
BIGBAG
BIGBAG
BIGBAG
MONSTER BAG
ZAK
SAMPLE
BIGBAG
Figure C.3: A schematic of the investigated route and sampling points
material is way to leave a tracer marker within the system, that is then picked up by the incoming product.
Iron content measurement is generally accepted to be quite precise and not very sensitive to measurement
errors. An investigation into iron content of the consecutive product batches leaving the system allows to
draw a curve of how contamination changes over time. In fact, iron was not the only tracer measured for the
trial. Additionally, protein content was investigated for the same samples as drawn. Yet, because of much
lower protein content difference between products and more uncertain laboratory measurement method,
resulting values are considered to be less precise from the iron content analysis.
First Steps
At first, the system is completely polluted by the high iron content product. The amount of product to be sent
through needed to entirely contaminate the system is taken from the experience of operations staff and based
on the previous trials. The following product is then put to mixer, filling it completely and thoroughly mixed.
Because there was some pollution before the mixer and within it, a reference sample is taken from the mixer
C.5. S CHEDULING RULES
89
itself to establish a base for the further measurements. In order not to dilute that, there is no material dosing
once the mixer start discharging, which deviates from the normal operating procedures but is required for
trustworthy results.
Screw Conveyor Measurements
Then, the discharge is started and all the material is directed to the same storage (Silo 23), just as the high
iron content pollutant, at a steady pace. Along the screw conveyors, four access points (A to D) were made,
for sample taking. Sampling for each point started when the material reached it, and then after previously
specified time intervals, same for all the points. This way was aimed to draw samples after a concurrent mass
flow for each point. The phase was ended once all the material from the mixer was sent to the designated silo.
Silo and Pneumatic Conveyor Measurements
Product from silo was directly discharged to a bagging line. Because of a higher discharge capacity than
bagging speed, this process was interrupted several times, based on the material level in the bagging line
buffer. To draw samples from lines with high pressure, special access points were made to divert some part
of the flow away to a sampling container. Such a device is shown in Figure C.4. The samples were then taken
manually from points D to G and by a machine just before bagging. Labelled containers with samples from
each access point were then taken to laboratory for chemical analysis to determine iron content.
Figure C.4: A photograph of a sampling access point for pneumatic conveyor
C.5. S CHEDULING RULES
Below, all possible defined and implemented rules are listed. There are four categories of scheduling rules,
connected to sequence of mixing orders, allocation of mixed products in silos and the choice of which product to package next. First category lists several alternatives that can be either applied or not and the following
ones base on preferences (expressions in model) used to make choices.
Scheduling alternatives
The following list contains an independent set of alternative choices that can be used in any combination
with each other. Rule 1 regards order manipulation, rules 2–5 silo allocation and the rest bagging.
90
C. C ASE STUDY
1. Combine matching recipes in a given day for mixing
To position all products with the same recipe next to each other;
2. Allocate silos based on all consecutive orders of the same recipe
To consider finding suitable space for more than a single order ahead;
3. Avoid splitting small orders among multiple silos
To prevent splitting small orders (less than 90 tonnes) into multiple silos. Might delay mixing;
4. Prefer splitting large orders into parallel silos to combining recipes
To send a large order (bigger than 90 tonnes), that by default would almost certainly be split, rather into
parallel empty silos than to a partially full with the same recipe and non-parallel one(s) elsewhere;
5. Put orders of the same recipe but different type into separate silos
To prevent from combining different order types of the same recipe in the same silo. Distinguished 4 types
of orders: bulk, big bag, 10 kg bag and coloured (for BTH3 only) and rest of small bag products;
6. Allow pre-emptions
BTH3 machine in cases when BTH3 and BTH1 or BTH2 are bagging from the same silo and a new, better
product is available for bagging is stopped. Product is redirected to BTH1-2 and after BTH3 changeover
another product is chosen. Requires to specify the lowest remaining order size to consider it beneficial;
7. Strongly prefer matching recipes for consecutive bagging
Forces bagging machines to foremost look for the same recipe as was bagged before and check other only
if not found;
8. Prefer for BTH3 to bag coloured or 10 kg size packages first
To prioritize BTH3-specific products;
9. Always combine recipes in silos. Allow splitting orders into parallel silos only
Overrules points 3–5 and silo matching choice (for ’best size fit’). Forces combining of the recipes and
allocates product to the chosen silo and its parallels only. Might cause significant starving time for large
orders;
10. Try redirecting Mixer to Polem discharge if Big bag filling is interfering
Tries to reduce waiting time in case mixer is to discharge to Polem silos while there is a simultaneous
discharge from the main silos (14–23) to big bag;
11. Prefer bagging late products first
Increases priority of late products. If not applied late products conform to the same rules as the rest.
Order manipulation
For each scheduled day the simulation logic can rearrange the order of consecutive products to be mixed.
Firstly, it organizes them based on the list below and then applies chosen alternatives from scheduling alternatives. Rearranging the consecutive orders to mix for the given day is based on:
•
•
•
•
•
Smallest bags first (bulk)
Smallest orders first
No sorting
Biggest bags first
Biggest orders first
Silo matching
When an empty silo is chosen as a potential candidate for filling with the current product, the silo matching
rule is applied to rank it priority. As such, e.g. the ’best size fit’ rule picks a silo that has the smallest positive
difference between the silo capacity and the order size. By default, the model logic prefers silos that already
have products of the same recipe inside (or used to have it) but that can be overruled by scheduling alternative
choices. Possible silo matching rules are:
• None
• Best size fit
• Shortest route
Bagging rules
A bagging machine chooses a new product to package from among all that are reachable at a given moment,
If there are more available it has to choose which one to pick first. Rules from scheduling alternatives are
applied prior to preferences listed below. In general the aim is to utilize bagging lines for as much time as
C.6. P RODUCTION O UTLINE
91
possible therefore waiting for silos to become available for discharge (that are blocked by parallel silo discharge) are not investigated. BTH3 is the only one to process 10kg bags and coloured products and there
is a strong preference that BTH3 would bag other product than parallel BTH1-2. Also, there is an in-built
preference to start discharging material to big bag or bulk stations prior to small package bagging machines.
For each of the small bagging lines there are two choices to be made - which silos to consider first and what
preference to use. There are 3 choices of the silo groups:
• All silos with no group preference
• Common silos first that can reach all bagging lines
• Designated silos first that can only discharge to the particular bagging line
The preferences below are then applied first to the group from above and then to the rest if there are no silos
ready to discharge within the group. The preferences then are:
•
•
•
•
•
•
•
•
•
•
•
Biggest order first
Biggest package first
Biggest silo content first
Biggest silo content ratio to capacity first
Earliest mixing time
None
Smallest order first
Smallest remaining quantity to bag first
Smallest package first
Smallest silo content first
Smallest silo content ratio to capacity first
C.6. P RODUCTION O UTLINE
A detailed schematic of production outline, including exact transportation connections is shown in Figure
C.5. Marked in green is material entry point and system exits are red. Parts that are conceptualised and not yet
built have dashed lines, whereas existing conveyors are distinguished between screw based and pneumatic
(blue).
Every existing screw conveyor and equipment is considered separately for cross-contamination calculations. For pneumatic conveyors the calculations are done after each of the independent intervals with only a
single direction.
C.7. C ASE S PECIFIC R ESULTS AND R ECOMMENDATIONS
There are a number of specific outcomes of the study, relevant to Sloten. This section discusses briefly the
results in scope of the company and gives case specific recommendations on further actions.
C.7.1. A DDITIONAL S ILO E FFECTS
The results have shown that installing additional silos can, up to a point, increase line capacities, because
there are:
• More jobs to choose from,
• More work-in-progress, bigger lead of mixer,
• Less chance of being blocked by parallel discharge.
Moreover, by installing bigger silos there is in general less order splitting among multiple silos, which leads to
fewer changeovers and higher throughput. More silos have also effects on product contamination, in general:
• The more silos, the fewer off specification products,
• There is a point where additional silos have little effect on contamination (8 silos not considerably
better than 4),
• Silo size has bigger impact on contamination than their number,
• Increasing silo sizes introduces higher risk of contamination.
92
C. C ASE STUDY
Figure C.5: A detailed schematic of proposed production organisation
C.7.2. S CHEDULING
Informed scheduling is vital for high production efficiency and should be the goal not only for Sloten, but for
any manufacturing company. Any schedule is an outcome of the order set, scheduling logic and operational
circumstances. Best schedule devised by a planning tool might not be superior if the algorithm does not have
the full information, which it rarely has. As the planning in Sloten is done off-line in advance, the information
on scheduled breaks, workforce availability etc. is not full. Moreover, scheduling is not updated once a change
in production occurs. Thus, for a company an operational tool to assist with making decisions would be
extremely helpful, especially when it comes to contamination.
There are certain limitations to the investigation. It was done with a single set of manufacturing orders,
and by changing this set the results could differ considerably. Moreover, scheduling by dispatching rules
is not a single choice, but a set of dependent selections, though for statistical analysis assumed independent. For the limited number of designed scheduling rules, there are almost 500 billion combinations of the
scheduling logic alone, which is by no means all of the choices made. Moreover, there is probably no fixed
set of scheduling rules that will perform best for any situation. However, the investigation gave a few good
indications about what is important and/or has the biggest impact:
•
•
•
•
Mixing sequencing is the most important aspect,
Allocate silos based on all orders of the same recipe,
Splitting when order type is different might be considered too, as it decrease contamination,
Unconditional combination of recipes in a silo and never splitting orders among multiple silos is unrealistic without forced waiting time for mixer,
• Orders need to fit tightly in their designated silo, therefore varying capacity of silos is suggested,
C.7. C ASE S PECIFIC R ESULTS AND R ECOMMENDATIONS
93
• For low contamination choose to bag earliest mixed orders (take advantage of nutrient sequencing),
• For high throughput:
– BTH1 & BTH2 should bag material from silos with largest amount in, or choose the largest available order,
– BTH3 should bag coloured and 10 kg bag products first,
– BTH3 then should go for smallest orders or smallest content in silos;
• There should be no discrimination of silo groups (designated/common),
• Pre-emptions have positive effect on throughput,
• When bagging late products is a priority, it has a slight negative effect on throughput.
C.7.3. L INE C APACITIES
Figure C.6 shows average achieved throughput in scenario Scheduling_C (see table G.33, which is a scenario
with a fairly high throughput.
Figure C.6: A diagram of achieved bagging speeds for scenario Scheduling_C
Yearly extrapolation
To give an indication about the possible yearly performance, some calculations are made to extrapolate that.
Assuming there are 240 working days in a year and the factory operates for 16 hours a day one can calculate
the expected yearly value as:
240 · 16 · (17.57 + 7.31) ≈ 95.5 [thousand tonnes]
(C.1)
Of course, for this time-frame, adding extended machine downtime, scheduled workforce downtime, lack of
orders or workforce etc. needs to be included as well, as the model investigates the issue only for a short
period, and does not include several aspects that could be important in the long run. Thus, the expected
extrapolated throughput should in the end decrease.
C.7.4. R ECOMMENDATIONS
Based on the performed analysis with described assumptions and limitations, it is concluded that additional
four silos with capacities of 30 to 40 tonnes would be beneficial for production performance. To reduce contamination varying the sizes of these silos should be considered, as the orders need to fit well in them.
94
C. C ASE STUDY
First of all, in order to understand better as well as draw proper conclusions, the cross-contamination trials should be way more extensive, and possibly include more than a tracer-collector relationship. Basing on
a single set of measurements can only give an approximate idea for the character of cross-contamination. Investigating several routes repeatedly with different materials with known properties is the only way to make
statistically solid judgements for system properties. That could be preceded by an analysis on the material flow within conveyors and silos, possibly supported by a more thorough literature search on this topics.
Moreover, cleaning the system from time to time and part by part is a way to determine the extent of amassed
residue. Doing so several times, in a long enough interval, could confirm or contradict the assumption that
the amount of material in the system remains roughly constant. Then, the possibility of using another type of
equipment, with a different residue specification, is not discussed in the thesis. Manufacturers should have
information on the possible level of leftovers in their equipment. Making it a priority to choose ones with a
low specification could be a simple solution to reduce overall contamination. Definitely finding a way to exchange screw conveyors for pneumatic ones, without deteriorating capacity, would be an example, but perhaps also installing silos with smaller residue characteristic. Furthermore, cross-contamination trials were
performed on a single route with just a single measurement for each mass flow for every sampling point.
This is not enough for a proper scientific study and should be expanded in the future. Doing more crosscontamination measurements, including other equipment and better preparations would help in increasing
prediction certainty of the results. Including factors like specific product properties and differentiating them
(in terms of e.g. viscosity, friction angles, grain size) would allow extending the analysis even further and
concentrate on actual product to product relations.
All scheduling decisions should be made with considerations of nutrient similarity for consecutive products, to reduce contamination and make more informed decisions. Method of matching product similarity
is much more transparent and theoretically sound than the current practice of introducing ‘blocking rules‘,
preventing certain combinations of products, and leaving the remaining scheduling up to the responsible
persons based on their preference and experience. Especially, that there is much to gain (or lose) in terms
of contamination, or specifically the number of off-specification bags of product. It is recommended to approach cross-contamination with closer scrutiny, and prioritize low contamination over higher throughput
when there is a trade-off relation. Interviews with company specialists suggest that rework or rejection (scrapping) of material would be more costly than slight increase in throughput. Closer scrutiny would require
assigning cost to working hours and rejected material, but is out of scope for this investigation.
It was not the goal of this research to give a single answer to which scheduling rules and particular combination of trade-offs for production efficiency are best. Instead, providing answers to the possible extent
of these trade-offs and effects associated with certain scheduling choices so that the users themselves can
pick the most suitable ones for them, either by choosing directly or introducing a cost function to weight
them. Needless to mention, from academic viewpoint the ability to reduce cross-contamination is the most
interesting one, and finding further ways to decrease it, is most strongly recommended. This could be a result of aforementioned trials but possibly also product base consolidation, or using ingredients of the same
nutritional value but different particle properties.
Moreover, it is believed that with some extra effort, the current operational scheduling procedure could
be utilised for the corresponding action for the new factory layout, especially because the mixer sequence
is deemed the most important factor for total material contamination. No advanced APS solution is thus
necessary to operate the future system. However, sloppy scheduling will pose a high risk of contamination,
much bigger than it is currently, and without a scheduling tool, assessment whether a chosen schedule is
close to optimum cannot be made. Thus, either a tool based on linear programming, or simulation-based
exploratory software could be extremely useful to determine the possible outcome to finally help in making
an informed decision for planner.
D
SN T OOL
This chapter contains information on the Scenario Navigator tool developed to support experimentation and
result analysis as well as for possible future use.
D.1. S CENARIO I NPUTS
The tool possesses 8 initial tabs, where vital data can be defined and scenarios prepared for execution. Below,
each of them is shortly described.
Main inputs
This tab is used to input equipment and contamination parameters. The user can choose whether to include
cross-contamination calculations in the run and according to which model. Also total residue in the mixer
(amount of material that remains in the mixer and before it) as well as cut-off point (amount of any product when contamination in given equipment can be neglected) can be set. Relevant equipment properties
comprise of bagging and discharge speeds and container level marks, among others. Additionally, maximum
predicted daily overtime can be adjusted. As advanced options the user can move to altering product and
recipe definitions.
Figure D.1: SN input of the main scenario parameters
Product definitions
Tab contains a table where the whole finished product portfolio is defined. In order to add a new product,
a unique number has to be given, name set, and connected to an existing recipe. Package size parameter
95
96
D. SN T OOL
defines the size of bags for the product to be put into, which reduces the possible system exits that can be
used. Last check-box is an additional parameter, company specific, that provides an extra restriction on the
possible bagging machine usage (in this case products with package size 20 or 25 and checked box will be
sent only to BTH3 bagging machine).
Figure D.2: SN input of the product portfolio
Recipe definitions
Recipes table connects the products that base on the same set of ingredients. As well as the products, they
are identified by a unique code and have their standard names. Additionally, because of the limitations of the
raw material feeding system, maximum dosing speeds to the main mixer are defined per recipe.
Figure D.3: SN input of the recipes for products
Silo constraints
This tab allows for change in the definitions of intermediate storage silos. For those, that do not exist in the
system, the user can choose to include them in the run and what their size should be. Moreover, for any silo
restrictions on accepted recipe can be established. This means, that only products with the same recipe as
chosen can be sent to the given silo.
D.1. S CENARIO I NPUTS
97
Figure D.4: SN input of silo recipe constraints
Order setting
Manufacturing orders are created by choosing a product from a drop-down list and then entering its quantity
and due date. Unique identification number, called MO number, and used to distinguish between orders, is
given automatically. These orders are then used by the simulation an handled in accordance to their assigned
due date.
Figure D.5: SN input of a manufacturing orders set
Scheduling rules
One of the most important features of the model is assigning scheduling rules to equipment, and measuring
their performance in search of rules of thumb that could help in daily operations. Thus the user has a choice
on all possibilities defined in C.5 by entering a boolean value for certain alternatives and picking a rule from
a drop-down list for others.
98
D. SN T OOL
Figure D.6: SN input of the scheduling rules
Initial conditions
An additional feature of the model is an ability to define initial silo contents, that have been previously mixed,
and are waiting to be processed further. This component is treated entirely as an extension and a possible use
in operations. The table contains a list of orders and their locations, where mixed products are awaiting that
is not connected to the actual order list. By using initial silo contents the warm-up period could be reduced.
Figure D.7: SN input of the initial conditions
Run settings
Finally, the scenario run settings need to be defined. These include starting date and duration of the run,
warm-up period, number of replications as well as the simulation model to be used, if those differ. On top of
that the user has a possibility to exclude data gathering for some of the parameters if desired. The following
dashboards in the tool relate to results visualisation and analysis.
D.2. S CENARIO R ESULTS
99
Figure D.8: SN run properties screen with possible data collection restrictions
D.2. S CENARIO R ESULTS
Prepared tool in SN contains 6 tabs for results presentation on a high level, mostly including charts for convenient visual comparison between scenarios, but also having SQL queries to filter acquired data for more
interesting values. It starts with a main results dashboard that has hierarchical links to 5 tabs with more detailed parameters on: mixing, orders & contamination, daily, silo and bagging statistics. Relevant dashboards
contain buttons allowing to export data for processing to another tools.
The following figures included to visualise dashboard descriptions should not be connected to any scenario or each other, unless otherwise specified. There are there for demonstrative purposes only
Main results
This dashboard contains the most important information of the entire scenario divided in sections. By pressing an appropriate button more detailed window about it is displayed. Vital indicators are general throughputs and time statistics for equipment: mixer and BTH machines. Moreover, information on empty silos and
silo utilisation as well as a display of timely completed orders. Finally, contamination with respect to orders
and total material is given.
Figure D.9: Main dashboard with scenario high-level results
100
D. SN T OOL
Mixing performance
A resource chart shows information on the starting and ending times of each order mixing, distinguishing
three item categories: small bag, big bag and bulk products. It allows to see per order how long it takes to
process it in the mixer and when starving times occur. Additionally item names can be displayed on the
product bars to identify products of interest.
Figure D.10: Mixing performance statistics dashboard
Order & contamination parameters
This dashboard contains statistics with regards to orders as well as some average indicators on the total extent
of cross-contamination. A column bar chart displays per each consecutive order how much of it is considered
within (green) and how much above limits (red). Moreover a few gauges present information on the total
material bagged and how much of it was over limit (average from all replications), average order ratio above
limits, subset of that containing only orders with at least one bag above limits and finally total rejected bags
and weight over the run period.
Figure D.11: Details of order contamination and statistics
Also an important feature is silo allocation chart, that is constructed when an order searches for possible
destinations (flowchart presented in Figure E.9). Every time an order is prepared information on available,
blocked and rejected silos is saved to provide some reasons for the silo allocation.
D.2. S CENARIO R ESULTS
101
Daily time statistics
This dashboard presents mostly four stacked column charts for mixer and BTH1-3 machines, displaying the
composition of running, starving, changeover and short stop times (as well as running late products for BTH).
Added up, they show how much time is the given piece of equipment working and thus how much over- or
under-time there is that day as an average of all the replications.
Figure D.12: Time statistics for the main equipment
On the top there are several gauges displaying the total amount of time in given category of working,
overtime and starving is there in the whole run.
Silo statistics
Silo statistics aim to show the usage of silos chosen for the scenario. In general silo utilisation is defined
as the average silo content weigh ratio to its capacity during a non-empty period. The figure displays such
utilisation for the whole run period.
Figure D.13: Silo utilisation statistics dashboard
Further details are available for silo utilisation per given day and then silo contents charts measured in 15
minute intervals showing either the actual mass of material in the silo or its ratio to the silo capacity.
E
M ODEL I MPLEMENTATION
This chapter presents several flowcharts prepared to indicate important model logic.
E.1. O RDER PATH
Manufacturing orders are parent entities that steer the flow of product batches, their child entities. This
process is shown in Figure E.1. Orders for the first day are initialised at the start of simulation run and after
Figure E.1: A flowchart of the order sequence logic
102
E.1. O RDER PATH
103
their rearrangement in accordance to the chosen scheduling rules are kept in a queue. Whenever this queue
runs empty the next production day is initialised. Then, one by one orders search for possible destinations
and once find sufficient space move to dosing, which commences immediately for the same recipes and after
mixer changeover if the recipe is different. When the dosing is finished, another order is taken from the order
daily queue.
Figure E.2: A flowchart of how sequencing of daily orders is executed
Mixing takes time depending on the dosing speed and chosen destination and the cross-contamination
calculation start when leaving the mixer and continue for each independent interval. As the first product
reaches the chosen silo, the discharge might commence but usually the product needs to wait until a relevant
system exit becomes available (or time function reaches its due date). When the mixing finishes the order
entity is moved from the dosing station to the storing station awaiting next steps. From the silo product
batches are sent to either bagging or bulk stations to be processed further. In case the order is split there
might be order merger or the rest of the material continues later. When the whole order has been processed
the order gathers final statistics and is sent to the finished orders station, where it remains till the end of the
run.
The most important function in the possible daily order sequencing (flowchart shown in Figure E.2 is
incrementally basing consecutive products on the least different product. This procedure starts with a ref-
104
E. M ODEL I MPLEMENTATION
erence product, which is the last product from the previous day (the last product in the mixer) or the first
one from the queue in case of the model start-up. All remaining products in the queue then calculate their
difference with respect to the reference product as shown in Figure E.3. All nutrients are equally important in
this check and the relative difference between specified level of the reference product and the lower of higher
threshold value of the investigated product is calculated.
Figure E.3: A flowchart of how nutrient similarity is defined
The order with the lowest relative difference is then taken and placed next in the order mixing sequence.
In case there is more than one with the same, lowest value the first one in accordance with mixing order
scheduling rule (sequencing based on order parameters such as size or package size). The reference is set
then to the taken product and the procedure is repeated for the set with one fewer order. The process is
finished when all orders are allocated in the mixing sequence. Similar approach is taken for order allocation
in silos when these are empty candidate silos. Then a nutrient similarity function is a multiplier of a silo
objective function that needs to be minimised based on the current state. This way in case of multiple empty
silos, depending on their last contents a preference is given based on nutrient levels on which one to choose.
E.2. A SSESSING B AG C ONTENTS
In order to determine whether a packaged bag complies with specification limits for a given product a two
step analysis is performed. At first, all bag contents in terms of products are gathered together, as previously
the batch entities comprising it were just put together. The logic for it is shown in Figure E.4. The following
step is to assess the nutrient content of a given mixture, the process of which is shown in Figure E.5. The
method assumes equal spread (homogeneity) of contents in the bag as well as no sampling bias. A flowchart
in Figure E.4 shows performed steps to gather together all bag contents from comprising product batches so
that there is no missing or repeated data. Next to that, the process performs an important function of counting finished bags for their parent orders and signifies the end of order if all of them have been processed. To
simplify the process, one by one, each of the bag child entities (batch members) are taken and its comprising
product contents are added to, empty in the beginning, respective bag fields. At first original, intended recipe
content is added and then, again one by one, all comprising product contaminants. For product batches,
the table of contaminants is incremental and there is no removal of entries so once an empty field in the table is reached, the next product batch can be taken. The main logic is thus within two embedded for loops,
where the higher level one is of the size of the number of entities and the lower level has each time the size
of the number of contaminants within examined product batch. There is a significance of not ever removing
product batch contents, not only to easier handling of their tables, but also to express that once a particular
product was contaminated by a given other product. With quite uncertain contamination curve, the means of
empirically establishing whether a particular contamination took place is added, though with a large number
of entities processed one needs to know exactly what to look for in order to take advantage out of that.
Recalculation into nutrients is a sub-process for bag contents assessment, and is shown in Figure E.5. It
E.3. S ILO L OGIC
105
Figure E.4: A flowchart of how final bag contents are gathered
can be viewed as first a for loop for each of the 16 identified nutrients adding them up for the original recipe
content, and then repeating this process for each nutrient for each of the comprising product contaminants.
Thus, if the product has 12 contaminating components then including the original recipe there is a need for
16 · (12 + 1) = 208 calculations, just for a single bag, which is a lot of calculations.
An important feature limiting search problem is searching for specification values not in a table with all
contaminants but within storage station containing all previously processed orders. Because the simulation
time was limited, the build-up of orders was also constrained to manageable number of several hundred. The
full table, on the other hand, has over 10 thousand entries. Once all the nutrients are added up a comparison
with specification limits can be made, and is done separately for lower and higher limits for each nutrient.
Once any value is considered out of specification, the bag is labelled as rejected.
E.3. S ILO L OGIC
Silos are intermediary objects between the main mixer and system exits and thus take part in both silo allocation and final discharge. Their main purpose is to hold (delay) material before processing further. Since there
might be material from different products in a single silo and an order is not necessarily put into a single silo
the knowledge how many different products are there in and when does one finish and the next one starts is
vital.
In general silo behaviour follows one big logical loop as presented in Figure E.6 and the most time is spent
on waiting for discharge or material. Normally silos are initialised empty and then wait for incoming material
from the mixer, which is then connected to a batch entity to save on memory and computing power. Once
there is some material in the silo the discharge might commence, providing there is capacity of a relevant
system exit. If not, the silo waits until it is queried for it, which is a check for possible parallel discharge
106
E. M ODEL I MPLEMENTATION
Figure E.5: A flowchart of bag contents with nutrient levels comparison
and fitting material parameters to the query. IF everything is matching the discharge commences to the
appropriate destination, which might be interrupted if the buffer levels breach their high level. Discharge
can be simultaneous to up to 2 small bagging machine and a single bulk station or big bag filling station.
Once the whole order in a given silo is discharged the outflow stops and references are cleared in case
it runs empty or updated if there is another material in it. There might be an additional check for the same
order in another silo if the order is split. Merging discharge is a process of redirecting flow from another silo
to the same system exit without doing a changeover, if the route is not blocked by a parallel powder flow. Silo
loop starts with a replication run and ends at the finish time.
Cross-contamination function is the central and most expensive calculation made in the model, executed
for each product batch between 10 and 27 times, depending on the route taken and bases on the calculation
made in section 4.4. Flowcharts for implemented functions are split into two and presented in Figures E.7 (in-
E.3. S ILO L OGIC
107
Figure E.6: A flowchart of the crucial silo logic
cluding used designations) and E.8 (also with often repeated sub-process). Simply put the process compares
and shifts some of the values (quantity) in one table (product batch) to and from another table (equipment
residue table) and it might be divided into 4 stages: establishing exchanged amount, updating batch original
recipe content, comparing batch contamination table with equipment residue table and contaminating the
batch by products in the equipment that were not previously present in the batch.
108
E. M ODEL I MPLEMENTATION
E.4. C ROSS - CONTAMINATION I MPLEMENTATION
Figure E.7: A flowchart of contamination calculation implementation; Part 1 of 2
An important assumption is that once any product residue in certain equipment reaches a level below
user-settable threshold level it is removed from the table and treated as 0. Once a product batch is contaminated, its presence cannot be entirely removed, so that the contamination is incremental and there are no
gaps in the static table. Establishing the exchanged amount is the first step. Normally it is equal to the exchange amount parameter EAP unless it is a run with a random component, when the EA is drawn from a
normal distribution of EAP with set deviation and restricted when breaches certain thresholds. Moreover,
E.5. P OSSIBLE S ILO A LLOCATION F UNCTION
109
Figure E.8: A flowchart of contamination calculation implementation; Part 2 of 2
extra variables to keep track of the number of contaminants present in both equipment and product batch
are introduced to save on the number of calculations. This way contaminants present in the product batch
are always compared with the residue in the equipment and the search in the other way is performed only if
the number of product residues in the equipment is bigger than the one in the product batch, after the initial
comparison. This function (as in Figure E.8) is thus calculated infrequently, being very expensive by going
through each product residue in the equipment and searching for respective product in the batch.
E.5. P OSSIBLE S ILO A LLOCATION F UNCTION
Another important feature of the model is allocating possible destination silos for orders before mixing, presented in Figure E.9. When dosing finishes the next order in line is taken and an attempt is made to allocate
possible silos to it. First all silos that are available in a simulation run and are not reserved for another recipe
are queued in a so called possible allocation queue and counted. From that queue silos with the same recipes
are searched. All found are removed and either inserted into actual allocation queue is the silo fill level is below high or counted as blocked with the same recipe when the level is not below high. Moreover, silo with the
same recipe might be rejected based on scheduling rules (e.g. preventing products with the same recipe but
different type from being in the same storage). Afterwards, all silos that are currently holding other recipes
are removed from the possible allocation queue and only empty ones remain.
The remaining empty silos are then ranked in accordance to chosen scheduling rules. In case of a run with
contamination silo residue (last product in) is compared with the order being allocated based on nutrients.
110
E. M ODEL I MPLEMENTATION
Figure E.9: A flowchart of how possible destinations are assigned to orders
Then, final order in term of scheduling rules is established and silos are one by one removed from possible
allocation queue and moved to the actual allocation queue providing they meet scheduling constraints. Once
all silos are processed the statistics are saved and remaining space in the silos is compared with necessary one
for allocation. If that is sufficient the order allocation queue might be rearranged to find a better match in
case the order is to be split (i.e. preferring discharge to parallel silos). Then the order is sent to dosing where
E.6. M
L
OGIC
Title: SiloIXER
Allocation
User: Administrator
Printdate: 2015-10-14
111
Scenario(s): MassC_4_extra_3_Random_F (1)
16
AvailableSameRecipe
AvailableEmpty
BlockedContamination
BlockedSameRecipeFull
BlockedOtherRecipe
RejectedConstraints
14
12
10
8
6
4
119 (M…
125 (M…
126 (M…
124 (M…
121 (M…
123 (M…
114 (M…
112 (M…
113 (M…
117 (M…
116 (M…
115 (M…
118 (M…
122 (M…
120 (M…
111 (M…
106 (M…
109 (M…
108 (M…
101 (M…
99 (Ma…
99 (Ma…
99 (Ma…
99 (Ma…
0
99 (Ma…
2
Figure E.10: An example order silo allocation chart
it awaits mixer availability and follows the procedure as described earlier in Figure E.1. When there is not
enough space in the chosen silos the allocation queue is cleared and the order waits for a silo to run empty to
repeat the search.
After each execution of the possible silo allocation function a data entry is made to categorise assessed
silos for given order, an example of which is shows in Figure E.10. A manufacturing orders can have multiple
entries if there was not enough space available to commence mixing.
E.6. M IXER L OGIC
The main mixer logic is closely connected to its state and especially orders currently allocated to it. Every time
a new order has finished dosing, there is a process determining what to do next, more specifically whether
there is enough time and capacity to process the next order or to finish for a given day. If the next order in
queue is to be finished on a given day then once the mixing has ended, the statistics are saved, and new order
can be processed. Logic for stopping due to time constraints for a given day is executed elsewhere.
Here, the algorithm assesses, whether there are enough resources to mix orders, that are not necessary
for a given day but could be put into silos. At first, predicted ending time is checked, and if that is sufficient,
an order searches for possible destinations (see Figure E.9) to determine available capacity in silos. If there
is enough time and capacity the process ends without consequence. When there is enough time but not silo
capacity the loop continues, waiting for capacity to become available or the end of shift. In the case when
there is not enough time to finish the order on time, the shift is ended early, statistics recorded, and the next
order in queue is prevented from entering dosing.
E.7. C HOOSING S ILO TO D ISCHARGE
Packaging material is a process between a silo and bagging machines, where any two machines can be chosen
to bag material from a given silo. Therefore, a match has to be found between those two, while the packaging
is leading by ‘pulling’ material from the silo. Figure E.12 shows how the most versatile machine BTH3 chooses
a potential silo to bag from, based on scheduling rules and its internal logic. The process commences when a
new product enters an empty silo, at the beginning of a shift, when a bagging machine becomes available, or
when a potential discharge candidate is rejected.
The whole procedure bases on repeated searches for suitable silos (or in fact their first orders for discharge) according to scheduling rules and then checking whether discharge from them is possible and finally
allocating suitable bagging machines. Checking availability comprises of parallel discharge assessment, size
and restrictions for the order control as well as final machine availability check, not to commence more than
112
E. M ODEL I MPLEMENTATION
Figure E.11: A flowchart of processing logic for mixer after each order dosing finishes
one discharge to a bagging machine.
E.8. B AGGING M ACHINE S HIFT E ND
Every time a bagging machine finishes a changeover it executes a decision process, whether there are still
orders to be processed by it, and if that is possible. As the decision is based on the current state of the mixer,
orders in silos and their due dates, the search is a complex one. Other processes are responsible for time
E.8. B AGGING M ACHINE S HIFT E ND
113
Figure E.12: A flowchart of how orders to be bagged by BTH3 machine are searched for
management. Flowchart of the process logic is shown in Figure E.13.
Here, all orders that have already been mixed and could be processed by a given machine, providing the
mixer has finished mixing orders for a given day, are added to a potential queue. From these, orders that are
allocated in reachable silos i.e. with a direct connection to the machine are moved to another queue, called
reachable queue. Then, depending on the number of orders in the reachable queue a decision whether to
continue is made.
With this method, orders that are split among multiple silos or multiple orders in a single silo are recognised and handled, and a simple predictive decision is made. Only when there is a potential reachable order
that can be bagged in the future will the bagging machine stay on shift, even when the order is not immediately to be started. If it is blocked by a parallel discharge it will result in machine starving time.
The method is not optimal, it might result in unnecessary starving time of a single machine if there is one
order blocked for discharge and all three available machines can process it.
114
Figure E.13: A flowchart of how bagging machine decides to end early for a given day
E. M ODEL I MPLEMENTATION
F
V ERIFICATION & VALIDATION
F.1. V ERIFICATION T ESTS
Below several performed verification tests are listed starting from degenerate and general logic tests in table
F.1, followed by a trace test.
Carried out tests
Some degenerate tests performed are listed in table F.1.
Table F.1: List of performed general verification tests
Group
Dosing
Order
sequencing
Initial
conditions
Order
composition
Silos
parameters
Input
Result
Dosing speed > mixer discharge
Dosing speed = mixer discharge
Dosing speed < mixer discharge
Smallest orders first
Largest packages first and join recipes
Nutrient sequencing
Random product with distant due date in silo
All silos blocked for non-existing recipes
All but one silos blocked
No orders in queue
Bulk orders only
Big bag orders only
Orders for colored products/10kg bags only
Large orders only (>100 tonnes)
Silos reserved for some recipes
Silo discharge speed set to 1 kg/h
Silo discharge speed set to 20000 tonnes/h
Parallel silo list set to all silos
Material in mixer increases, cyclic dosing breaks when full
Roughly constant material levl in mixer
Cyclic discharge breaks due to low material level
Order quantity increasing in queue
Package size decreasing in queue and same recipes together
Orders shuffled, low dissimilarity and same recipes together
Size and references correct, no discharge till due date
Mixing never starts
Material allocated to only available silo
No production
Only silos 14-23 used, bulk stations periodic discharge when full
Only silos 14-23 used, quick increase of WIP material
Only silos reaching BTH3 used, machines BTH1-2 idle
Orders split into multiple silos, outflow merged whene possible
Reserved silos not considered for other products
Extremely slow discharge
Extremely fast discharge until buffers full, piping not empty
Only single silo discharge in the system
Trace test
Trace test is based on a movement and cross-contamination of the first product batch of a random order,
including logic for silo allocation and its discharge. A shortened model trace is shown below:
Item Name: Sprayfo geel NL 25kg
Item Number: 10041915
Recipe Number: 10040895
Recipe Name Recept 3100
Order Quantity: 24000kg
Package Size: 25kg
Due Date: 08.01.2015
Order Number: 66
Dosing speed: 50000 kg/h
Coloured product: False
Time statistics:
TimeStartedMixing 154.06402465065776 Hours
TimeEndedMixing 154.74579131731289 Hours
TimeStartedBagging 176.00702175666007 Hours
TimeEndedBagging 177.28228900885603 Hours
Silo allocation statistics:
154.06402465065776 Hours
Name Value
SilosTotal 15
SilosAvailableSameRecipe 0
SilosAvailableEmpty 1
SilosBlockedSameRecipeFull 0
SilosBlockedOtherRecipe 10
SilosRejectedConstraints 4
SilosAvailableCapacityForAllocation 65000 Kilograms
Starting dosing.
115
116
Silo AllocationQueue 1: Silo_19
Silo chosen has big enough capacity.
No queue reordering as there is only a single silo picked.
Nutrient specification:
Contaminant Specification Lower Limit Higher Limit
Animal Fat in Fat 0 NaN 4
Ash 9 NaN 10.1
Colorant 0 NaN NaN
Copper 10 8 12
Fat 16.5 14.4 18.6
GMO 0 NaN NaN
GMO in GMO Protein 0 NaN NaN
GMO Soy Flour 0 NaN 1
Iron 100 80 120
Lactose 46.5 43.5 49.5
Probiotic 1000000 NaN NaN
Protein 21.5 18.8 24.2
Protimax 0 NaN NaN
Soyabean Protein 0 0 NaN
Vital Wheat Gluten Dry 0 NaN 1
Vitamin A 40000 36000 44000
AcceptedBagsOrder 858
AcceptedWeightOrder 21450
AcceptedOrderRatio 0.89375
RejectedBagsNutrients 102
RejectedWeightNutrients 2550
RejectedOrderRatioNutrients 0.10625
Order on time: True
Verification values:
Batch interarrival time: 0.0001 Hours
5kg / 50000 kg/h = 0.001h
Minimum dosing time: 4800 * 0.001h = 0.48h
Minimum discharge time after dosing:
3000kg / 40000kg/h = 0.075h
Dosing-->Mixer transport time: 0.001(6) hours
Mixing time: 0.41499166668038 Hours
Minimum time < Mixing time
Available silos: 15 (11 initial + 4 new)
Investigated batch: Product 346415
1. Initial
Originalrecipecontent: 5 [kg]
2. After Mixing Contamination
OriginalRecipeContent 4.8020569762106762
Contaminants:
Item Number Mass [kg]
10041865 0.0001175069627893614
10040205 0.19782551682653426
3. After the first screw conveyor
OriginalRecipeContent 2.7659848182973494
Contaminants:
Item Number Mass [kg]
10041865 0.00132106141106035
10040205 2.23269412029159
4. After the second screw conveyor
OriginalRecipeContent 0.27659848182973468
Contaminants:
Item Number Mass [kg]
10041865 0.00132106141106035
10040205 4.7206089411318839
5. After the third screw conveyor
OriginalRecipeContent 0.027659848182973484
Contaminants:
Item Number Mass [kg]
10040205 4.9694004232159124
10041865 0.00132106141106035
6. After the sixth screw conveyor
OriginalRecipeContent 0.016885977061830239
Contaminants:
Item Number Mass [kg]
Name Value
10040205 4.9801679257314717
10041865 0.00132106141106035
7. After the eigth screw conveyor
(entering silo 19)
F. V ERIFICATION & VALIDATION
OriginalRecipeContent 0.015562116460182749
Contaminants:
Item Number Mass [kg]
10041865 0.00132106141106035
10040205 4.5897227603541237
10040135 0.0013445461412777936
10040355 0.1986456206184527
10754455 0.0011237284007612216
10276785 0.19886749435773685
8. After the silo
OriginalRecipeContent 0.0140059048141645
Contaminants:
Item Number Mass [kg]
10041865 0.00132106141106035
10040205 4.13075048431871
10040135 0.0013445461412777936
10040355 0.178781058556607
10754455 0.00383359245582761
10276785 0.676134048043703
9. After the panumatic conveyor system
OriginalRecipeContent 0.0043660052721958519
Contaminants:
Item Number Mass [kg]
10041865 0.00132106141106035
10040205 1.2876624989213121
10040135 0.011673490401418091
10040355 1.3021951100181148
10754455 0.003833592455827609
10276785 0.34553561165688396
10038805 0.12495744039917706
10042695 0.12792395710989624
10043355 1.7989290253094348
10. After the air filter
OriginalRecipeContent 0.0031190741664567168
Contaminants:
Item Number Mass [kg]
10041865 0.00132106141106035
10040205 0.91990608922938533
10040135 0.011673490401418091
10040355 0.93028818659694135
10754455 0.003833592455827609
10276785 0.24685064096767789
10038805 0.094967654703374565
10042695 0.091576807732927257
10043355 2.7122954730286364
11. After the bagging buffer
OriginalRecipeContent 0.0029007389748047468
Contaminants:
Item Number Mass [kg]
10041865 0.00132106141106035
10040205 0.8555126629833284
10040135 0.011673490401418091
10040355 0.86516801353515549
10754455 0.003833592455827609
10276785 0.22957109609994045
10038805 0.094967654703374565
10042695 0.085389483279515108
10043355 2.8722112075185415
12. Bag contamiantion
OriginalRecipeContent 0.66430670424439575
Contaminants:
Item Number Mass [kg]
10041865 0.0060438882287918334
10040205 11.80937146779862
10040135 0.016578099923367454
10040355 3.1374302418541409
10754455 0.018826663246733256
10276785 2.4248923189217026
10038805 0.24158115955761184
10042695 0.21869821168682493
10043355 6.4882633848704536
Nutrients:
NutrientAnimalFatInFat 0
NutrientAsh 0.23915041352798244
NutrientColorant 0
NutrientCopper 21.360238935959895
NutrientFat 0.43844242480130119
NutrientGMO 0
F.2. S TATISTICAL T ESTS
117
NutrientGMOinGMOProtein 0
NutrientGMOSoyFlour 0
NutrientIron 2.657226816977583
NutrientLactose 1.235610469894576
NutrientProbiotic 520649.03791019565
NutrientProtein 0.57130376565018026
NutrientProtimax 0.0012649203173318664
NutrientSoyabeanProtein 0
NutrientVitalWheatGluten 0
NutrientVitaminA 1062.8907267910331
Bag rejected: TRUE
F.2. S TATISTICAL T ESTS
Table F.2: Accuracy of the simulated contamination curves with respect to the fitted contamination
Point
SSTotal
SSRegression
SSResidual
RSquared
StDev
A
B
C
D
E
F
G
BAG
1.2000
1.0809
0.1191
0.9008
0.0109
5.5072
5.3972
0.1101
0.9800
0.0105
6.7699
6.3656
0.4043
0.9403
0.0201
7.1377
6.1235
1.0141
0.8579
0.0318
13.0768
12.9185
0.1583
0.9879
0.0126
12.9861
12.7940
0.1921
0.9852
0.0139
14.7372
14.4615
0.2756
0.9813
0.0166
3.5497
3.4512
0.0985
0.9723
0.0222
Table F.3: Goodness of fit (Kolmogorov–Smirnov test) for short stops and changeover time distributions
DN
p-value
result
BTH1
0.185746
0.0
reject
Timetorepair(log-normal)
BTH2
BTH3
Mixer
0.146947
0.0565757
0.06596
0.0
0.0000573061 0.000822343
reject
reject
reject
Changeover(Weibull)
BTH1-3
0.0316606
0.335282
failtoreject
0.003
0.002
density
dataset
Measured
Simulated
0.001
0.000
0
200
400
600
800
Time [s]
Figure F.1: Comparison between probability density function of measured and fitted distribution for bagging changeovers
Table F.4: Statistical data on obtained throughputs in the base case simulation
Machine
BTH1[bags]
BTH2[bags]
BTH3[bags]
BTH1[kg]
BTH2[kg]
BTH3[kg]
Average
SD
Median
Nomeasurements
ValueMin
ValueMax
355.32
17.95
359.57
300
223.41
379.05
360.11
15.92
363.40
300
222.48
382.77
357.72
18.86
362.84
300
258.62
378.36
8700.82
512.05
8875.22
300
5585.15
9476.19
8813.73
484.56
8940.43
300
5562.08
9569.13
7014.55
1007.60
6911.82
300
3201.25
9198.34
G
E XPERIMENT R ESULTS
G.1. I NITIAL E XPLORATION R ESULTS
G.1.1. R ESULT TABLES
Table G.1: Experiment results of scenario NC_A
Scenario
KPI
Avgthroughputbags
Avgthroughputquantity
Orderscompleted
Orderscompletedontime
NC_A
Average
1073.15
24529.10
261.00
248.85
SD
41.57
1400.92
0.00
2.76
median
1080.69
24686.10
261.00
249.00
no. measurements
300
300
20
20
Unit
bags/h
kg/h
-
NC_B
Average
1073.16
24529.1
260.50
247.15
SD
41.57
1400.92
0.81
3.28
median
2080.70
24686.10
261.00
246.50
no. measurements
300
300
20
20
Unit
bags/h
kg/h
-
NC_C
Average
1071.31
24467.82
260.95
248.65
SD
40.65
1439.16
0.22
2.62
median
1080.54
24605.21
261.00
249.00
no. measurements
300
300
20
20
Unit
bags/h
kg/h
-
NC_D
Average
1060.78
24209.60
260.5
249.45
SD
56.30
1774.11
1.07
3.98
median
7078.38
24653.46
261.00
251.00
no. measurements
300
300
20
20
Unit
bags/h
kg/h
-
Table G.2: Experiment results of scenario NC_B
Scenario
KPI
Avgthroughputbags
Avgthroughputquantity
Orderscompleted
Orderscompletedontime
Table G.3: Experiment results of scenario NC_C
Scenario
KPI
Avgthroughputbags
Avgthroughputquantity
Orderscompleted
Orderscompletedontime
Table G.4: Experiment results of scenario NC_D
Scenario
KPI
Avgthroughputbags
Avgthroughputquantity
Orderscompleted
Orderscompletedontime
118
G.1. I NITIAL E XPLORATION R ESULTS
119
Table G.5: Experiment results of scenario NC_E
Scenario
KPI
Avgthroughputbags
Avgthroughputquantity
Orderscompleted
Orderscompletedontime
NC_E
Average
1071.62
24442.90
260.85
247.85
SD
44.79
1490.34
0.48
4.90
median
1081.41
24587.42
261.00
248.00
no. measurements
300
300
20
20
Unit
bags/h
kg/h
-
NC_F
Average
1045.68
23864.88
260.40
248.00
SD
59.84
1705.87
2.18
6.17
median
1064.53
24242.70
261.00
249.50
no. measurements
300
300
20
20
Unit
bags/h
kg/h
-
NC_G
Average
1069.78
24397.41
260.70
250.45
SD
43.29
1465.65
0.56
4.74
median
1080.11
24647.20
261.00
252.00
no. measurements
300
300
20
20
Unit
bags/h
kg/h
-
NC_H
Average
1070.92
24433.75
261.00
250.75
SD
43.27
1323.44
0.00
3.51
median
1080.63
24563.2
261.00
250.50
no. measurements
300
300
20
20
Unit
bags/h
kg/h
-
NC_I
Average
1077.39
24560.26
260.95
252.50
SD
28.80
1323.20
0.22
3.27
median
1081.84
24599.79
261.00
253.50
no. measurements
300
300
20
20
Unit
bags/h
kg/h
-
NC_J
Average
1060.29
24164.30
260.90
250.45
SD
27.75
1306.65
0.30
3.00
median
1063.08
24256.63
261.00
250.50
no. measurements
300
300
20
20
Unit
bags/h
kg/h
-
Table G.6: Experiment results of scenario NC_F
Scenario
KPI
Avgthroughputbags
Avgthroughputquantity
Orderscompleted
Orderscompletedontime
Table G.7: Experiment results of scenario NC_G
Scenario
KPI
Avgthroughputbags
Avgthroughputquantity
Orderscompleted
Orderscompletedontime
Table G.8: Experiment results of scenario NC_H
Scenario
KPI
Avgthroughputbags
Avgthroughputquantity
Orderscompleted
Orderscompletedontime
Table G.9: Experiment results of scenario NC_I
Scenario
KPI
Avgthroughputbags
Avgthroughputquantity
Orderscompleted
Orderscompletedontime
Table G.10: Experiment results of scenario NC_J
Scenario
KPI
Avgthroughputbags
Avgthroughputquantity
Orderscompleted
Orderscompletedontime
120
G. E XPERIMENT R ESULTS
Table G.11: Experiment results of scenario NC_K
Scenario
KPI
Avgthroughputbags
Avgthroughputquantity
Orderscompleted
Orderscompletedontime
NC_K
Average
1076.46
24551.67
261.00
251.20
SD
32.79
1309.24
0.00
3.33
median
1080.83
24645.49
261.00
252.00
no. measurements
300
300
20
20
Unit
bags/h
kg/h
-
SD
38.12
1373.98
0.00
2.55
median
1084.20
24622.98
261.00
250.50
no. measurements
300
300
20
20
Unit
bags/h
kg/h
-
Table G.12: Experiment results of scenario NC_L
Scenario
KPI
Avgthroughputbags
Avgthroughputquantity
Orderscompleted
Orderscompletedontime
NC_L
Average
1073.53
24471.98
261.00
250.25
25000
20000
15000
10000
5000
Average mixer throughput [kg/hour]
30000
35000
G.1.2. P LOTS AND F IGURES
A
B
C
D
E
F
G
H
I
J
K
L
Scenario NC
Figure G.1: A box plot for comparison of mixing throughput
G.2. C ONTAMINATION I NVESTIGATION R ESULTS
Table G.13: Experiment results of scenario MassC_A
Scenario
KPI
Avgthroughputbags
Avgthroughputquantity
Orderscompleted
Orderscompletedontime
Avgcontaminatedorderratioabovelimit
Totalcontaminationabovelimits
MassC_A
Average
1071.79
24699.81
303.95
289.60
0.1097
0.01122
SD
36.61
1441.48
0.80
4.83
0.0079
0.00069
median
1080.72
24974.09
304.00
290.50
0.1079
0.01127
no. measurements
320
320
20
20
20
20
Unit
bags/h
kg/h
-
G.2. C ONTAMINATION I NVESTIGATION R ESULTS
121
Table G.14: Experiment results of scenario MassC_B
Scenario
KPI
Avgthroughputbags
Avgthroughputquantity
Orderscompleted
Orderscompletedontime
Avgcontaminatedorderratioabovelimit
Totalcontaminationabovelimits
MassC_B
Average
1068.89
24655.03
303.75
289.15
0.1062
0.01102
SD
37.10
1417.61
1.22
5.53
0.0114
0.00076
median
1078.86
24817.58
303.00
290.50
0.1073
0.01086
no. measurements
320
320
20
20
20
20
Unit
bags/h
kg/h
-
MassC_C
Average
1064.18
24531.18
303.40
290.55
0.1908
0.01685
SD
59.44
1779.36
2.48
4.09
0.0135
0.00062
median
1082.05
24882.07
304.00
291.00
0.1904
0.01687
no. measurements
320
320
20
20
20
20
Unit
bags/h
kg/h
-
MassC_D
Average
1055.26
24398.01
300.00
282.50
0.1880
0.01684
SD
102.63
2623.48
12.84
21.76
0.0092
0.00090
median
1081.36
24917.00
304.00
290.00
0.1881
0.01684
no. measurements
320
320
20
20
20
20
Unit
bags/h
kg/h
-
MassC_E
Average
1069.26
24650.07
303.90
291.05
0.1856
0.01667
SD
49.08
1671.28
2.19
3.57
0.0138
0.00100
median
1082.72
24976.83
305.00
291.50
0.1827
0.01669
no. measurements
320
320
20
20
20
20
Unit
bags/h
kg/h
-
MixC_A
Average
1067.01
24616.74
303.05
289.35
0.0479
0.00549
SD
39.68
1547.02
2.38
4.67
0.0022
0.00024
median
1078.43
24737.83
304.00
290.50
0.0485
0.00554
no. measurements
320
320
20
20
20
20
Unit
bags/h
kg/h
-
Table G.15: Experiment results of scenario MassC_C
Scenario
KPI
Avgthroughputbags
Avgthroughputquantity
Orderscompleted
Orderscompletedontime
Avgcontaminatedorderratioabovelimit
Totalcontaminationabovelimits
Table G.16: Experiment results of scenario MassC_D
Scenario
KPI
Avgthroughputbags
Avgthroughputquantity
Orderscompleted
Orderscompletedontime
Avgcontaminatedorderratioabovelimit
Totalcontaminationabovelimits
Table G.17: Experiment results of scenario MassC_E
Scenario
KPI
Avgthroughputbags
Avgthroughputquantity
Orderscompleted
Orderscompletedontime
Avgcontaminatedorderratioabovelimit
Totalcontaminationabovelimits
Table G.18: Experiment results of scenario MixC_A
Scenario
KPI
Avgthroughputbags
Avgthroughputquantity
Orderscompleted
Orderscompletedontime
Avgcontaminatedorderratioabovelimit
Totalcontaminationabovelimits
122
G. E XPERIMENT R ESULTS
G.3. A DDITIONAL S ILOS
Table G.19: Experiment results of scenario MassC_F
Scenario
KPI
Avgthroughputbags
Avgthroughputquantity
Orderscompleted
Orderscompletedontime
Avgcontaminatedorderratioabovelimit
Totalcontaminationabovelimits
MassC_F
Average
1063.53
24506.67
303.75
289.75
0.1190
0.01023
SD
40.26
1428.42
1.18
3.46
0.0107
0.00065
median
1075.30
24679.23
304.00
290.00
0.1209
0.01027
no. measurements
320
320
20
20
20
20
Unit
bags/h
kg/h
-
MassC_G
Average
1073.21
24681.34
304.90
294.15
0.1056
0.00926
SD
33.76
1357.57
0.44
4.34
0.0104
0.00073
median
1080.03
24755.60
305.00
295.50
0.1046
0.00918
no. measurements
320
320
20
20
20
20
Unit
bags/h
kg/h
-
MassC_H
Average
1072.62
24724.48
304.30
292.85
0.0997
0.00899
SD
40.51
1307.10
1.27
3.09
0.0089
0.00061
median
1085.16
24938.18
305.00
293.00
0.0980
0.00897
no. measurements
320
320
20
20
20
20
Unit
bags/h
kg/h
-
MassC_I
Average
1075.13
24761.09
303.75
292.90
0.0867
0.00867
SD
34.46
1449.19
1.37
3.99
0.0090
0.00062
median
1083.68
24914.85
304.00
293.50
0.0851
0.00861
no. measurements
320
320
20
20
20
20
Unit
bags/h
kg/h
-
MassC_J
Average
1079.11
24873.79
304.25
293.65
0.0868
0.00789
SD
30.47
1278.97
1.26
3.64
0.0063
0.00049
median
1084.61
24913.92
305.00
294.00
0.0864
0.00779
no. measurements
320
320
20
20
20
20
Unit
bags/h
kg/h
-
Table G.20: Experiment results of scenario MassC_G
Scenario
KPI
Avgthroughputbags
Avgthroughputquantity
Orderscompleted
Orderscompletedontime
Avgcontaminatedorderratioabovelimit
Totalcontaminationabovelimits
Table G.21: Experiment results of scenario MassC_H
Scenario
KPI
Avgthroughputbags
Avgthroughputquantity
Orderscompleted
Orderscompletedontime
Avgcontaminatedorderratioabovelimit
Totalcontaminationabovelimits
Table G.22: Experiment results of scenario MassC_I
Scenario
KPI
Avgthroughputbags
Avgthroughputquantity
Orderscompleted
Orderscompletedontime
Avgcontaminatedorderratioabovelimit
Totalcontaminationabovelimits
Table G.23: Experiment results of scenario MassC_J
Scenario
KPI
Avgthroughputbags
Avgthroughputquantity
Orderscompleted
Orderscompletedontime
Avgcontaminatedorderratioabovelimit
Totalcontaminationabovelimits
G.3. A DDITIONAL S ILOS
123
Table G.24: Experiment results of scenario MassC_K
Scenario
KPI
Avgthroughputbags
Avgthroughputquantity
Orderscompleted
Orderscompletedontime
Avgcontaminatedorderratioabovelimit
Totalcontaminationabovelimits
MassC_K
Average
1076.93
24806.59
304.35
293.45
0.0920
0.00817
SD
37.23
1336.02
0.96
2.50
0.0085
0.00049
median
1085.96
24871.01
305.00
293.50
0.0909
0.00825
no. measurements
320
320
20
20
20
20
Unit
bags/h
kg/h
-
MassC_L
Average
1083.68
24984.12
304.40
295.75
0.1000
0.00872
SD
27.52
1298.89
1.24
2.72
0.0077
0.00057
median
1087.86
25099.30
305.00
0.50
0.0986
0.00881
no. measurements
320
320
20
20
20
20
Unit
bags/h
kg/h
-
MassC_M
Average
1084.26
25016.20
304.55
295.60
0.1091
0.00930
SD
24.38
1289.66
0.50
2.33
0.0085
0.00062
median
1087.53
25042.21
305.00
296.00
0.1081
0.00916
no. measurements
320
320
20
20
20
20
Unit
bags/h
kg/h
-
MassC_N
Average
1073.13
24689.41
302.40
291.90
0.0997
0.00762
SD
37.31
1475.05
2.75
2.83
0.0089
0.00052
median
1082.88
24776.34
303.50
291.50
0.0980
0.00761
no. measurements
320
320
20
20
20
20
Unit
bags/h
kg/h
-
MassC_O
Average
1076.73
24752.60
303.90
291.75
0.0875
0.00793
SD
36.27
1275.93
1.04
4.76
0.0100
0.00057
median
1082.85
24734.79
304.00
292.50
0.0852
0.00794
no. measurements
320
320
20
20
20
20
Unit
bags/h
kg/h
-
Table G.25: Experiment results of scenario MassC_L
Scenario
KPI
Avgthroughputbags
Avgthroughputquantity
Orderscompleted
Orderscompletedontime
Avgcontaminatedorderratioabovelimit
Totalcontaminationabovelimits
Table G.26: Experiment results of scenario MassC_M
Scenario
KPI
Avgthroughputbags
Avgthroughputquantity
Orderscompleted
Orderscompletedontime
Avgcontaminatedorderratioabovelimit
Totalcontaminationabovelimits
Table G.27: Experiment results of scenario MassC_N
Scenario
KPI
Avgthroughputbags
Avgthroughputquantity
Orderscompleted
Orderscompletedontime
Avgcontaminatedorderratioabovelimit
Totalcontaminationabovelimits
Table G.28: Experiment results of scenario MassC_O
Scenario
KPI
Avgthroughputbags
Avgthroughputquantity
Orderscompleted
Orderscompletedontime
Avgcontaminatedorderratioabovelimit
Totalcontaminationabovelimits
124
G. E XPERIMENT R ESULTS
Table G.29: Experiment results of scenario MassC_P
Scenario
KPI
Avgthroughputbags
Avgthroughputquantity
Orderscompleted
Orderscompletedontime
Avgcontaminatedorderratioabovelimit
Totalcontaminationabovelimits
MassC_P
Average
1077.18
24778.73
303.45
292.70
0.0842
0.00784
SD
30.04
1287.48
1.50
3.74
0.0085
0.00049
median
1082.63
24873.11
304.00
294.00
0.0811
0.00768
no. measurements
320
320
20
20
20
20
Unit
bags/h
kg/h
-
Table G.30: Experiment results of scenario MassC_R
MassC_R
Average
1076.67
24830.89
304.10
293.75
0.0834
0.00806
SD
35.86
1355.78
1.37
4.77
0.0041
0.00032
median
1085.14
24985.24
305.00
294.00
0.0831
0.00809
no. measurements
320
320
20
20
20
20
298
300
Number
302
304
Scenario
KPI
Avgthroughputbags
Avgthroughputquantity
Orderscompleted
Orderscompletedontime
Avgcontaminatedorderratioabovelimit
Totalcontaminationabovelimits
F
G
H
I
J
K
L
M
N
O
P
R
Scenario
290
280
285
Number
295
300
Figure G.2: A box plot comparing the number of completed orders within the simulation run per scenario
F
G
H
I
J
K
L
M
N
O
P
R
Scenario
Figure G.3: A box plot comparing the number of completed orders on time within the simulation run per scenario
Unit
bags
kg
-
125
1100
1050
1000
950
900
850
800
Average number of processed bags per hour
1150
G.4. S CHEDULING RULES I NVESTIGATION
F
G
H
I
J
K
L
M
N
O
P
R
Scenario
Figure G.4: A box plot comparing the average bagging throughput in silo number/size sensitivity analysis
G.4. S CHEDULING RULES I NVESTIGATION
Table G.31: Experiment results of scenario Scheduling_A
Scenario
KPI
Avgthroughputbags
Avgthroughputquantity
Orderscompleted
Orderscompletedontime
Avgcontaminatedorderratioabovelimit
Totalcontaminationabovelimits
Scheduling_A
Average
1066.60
24528.63
301.40
286.60
0.0888
0.00774
SD
40.71
1498.91
1.53
3.35
0.0070
0.00067
median
1078.26
24678.77
301.00
286.50
0.0901
0.00762
no. measurements
320
320
20
20
20
20
Unit
bags/h
kg/h
-
SD
42.73
1469.78
1.39
3.02
0.0097
0.00072
median
1076.60
24565.54
301.00
288.00
0.0935
0.00782
no. measurements
320
320
20
20
20
20
Unit
bags/h
kg/h
-
SD
29.31
1461.94
0.83
3.38
0.0068
0.00052
median
1083.41
25060.82
304.00
295.00
0.0912
0.00841
no. measurements
320
320
20
20
20
20
Unit
bags/h
kg/h
-
Table G.32: Experiment results of scenario Scheduling_B
Scenario
KPI
Avgthroughputbags
Avgthroughputquantity
Orderscompleted
Orderscompletedontime
Avgcontaminatedorderratioabovelimit
Totalcontaminationabovelimits
Scheduling_B
Average
1064.49
24483.95
301.60
287.60
0.0930
0.00787
Table G.33: Experiment results of scenario Scheduling_C
Scenario
KPI
Avgthroughputbags
Avgthroughputquantity
Orderscompleted
Orderscompletedontime
Avgcontaminatedorderratioabovelimit
Totalcontaminationabovelimits
Scheduling_C
Average
1078.70
24882.86
304.25
293.60
0.0915
0.00850
126
G. E XPERIMENT R ESULTS
Table G.34: Experiment results of scenario Scheduling_D
Scenario
KPI
Avgthroughputbags
Avgthroughputquantity
Orderscompleted
Orderscompletedontime
Avgcontaminatedorderratioabovelimit
Totalcontaminationabovelimits
Scheduling_D
Average
1053.06
24235.70
300.45
277.00
0.0919
0.00846
SD
88.08
2436.75
5.15
15.22
0.0116
0.00077
median
1081.49
24873.95
303.00
282.50
0.0919
0.00835
no. measurements
320
320
20
20
20
20
Unit
bags/h
kg/h
-
SD
78.49
2242.66
3.40
11.53
0.0079
0.00049
median
1083.57
25031.50
300.00
276.50
0.0869
0.00932
no. measurements
320
320
20
20
20
20
Unit
bags/h
kg/h
-
SD
43.65
1542.45
0.94
5.74
0.0068
0.00063
median
1081.77
24797.85
304.00
292.00
0.0906
0.00856
no. measurements
320
320
20
20
20
20
Unit
bags/h
kg/h
-
Table G.35: Experiment results of scenario Scheduling_E
Scenario
KPI
Avgthroughputbags
Avgthroughputquantity
Orderscompleted
Orderscompletedontime
Avgcontaminatedorderratioabovelimit
Totalcontaminationabovelimits
Scheduling_E
Average
1056.60
24402.81
300.15
272.90
0.0865
0.00932
Table G.36: Experiment results of scenario Scheduling_F
Scheduling_F
Average
1069.88
24593.86
304.10
290.30
0.0917
0.00873
Number
290
295
300
305
Scenario
KPI
Avgthroughputbags
Avgthroughputquantity
Orderscompleted
Orderscompletedontime
Avgcontaminatedorderratioabovelimit
Totalcontaminationabovelimits
A
B
C
D
E
F
Scenario
Figure G.5: A box plot comparing the number of completed orders within the simulation run per scenario
127
270
240
250
260
Number
280
290
300
G.5. R ANDOM C OMPONENT I NVESTIGATION
A
B
C
D
E
F
Scenario
Figure G.6: A box plot comparing the number of completed orders on time within the simulation run per scenario
Average number of orders completed on time
290
Scenario
K
285
ScA
ScB
ScC
ScD
ScE
ScF
280
275
300
301
302
303
304
Average number of completed orders
Figure G.7: A scatter plot displaying average results of orders completed on time as a function of the total number of completed orders
G.5. R ANDOM C OMPONENT I NVESTIGATION
This section contains additional experiments performed for the investigation, which are not directly relevant
to answering the research questions from section 1.3 on page 4 but are nevertheless explored to increase the
knowledge about some possible effects of certain interventions.
Table G.37: Random components chosen for analysis
Scenario
σ
Nutrient-basedsiloallocation
Random_C
Random_D
Random_E
Random_F
Random_G
0.5
0.5
0.5
0.5
0.5
YES
NO
YES
YES
NO
Chosen values σ for the random component as well as whether nutrient-based silo allocation was performed
128
G. E XPERIMENT R ESULTS
are shown in table G.37. It is firstly desired to assess whether the random component or its size has any impact on the total contamination, and then, whether nutrient-based silo allocation has considerable influence
when there are more silos as in investigation from section 6.4.2. It is an attempt to determine if a recreation
of peaks and valleys as observed for the results of the contamination measurements (Figure 5.2 on page 36)
has a noticeable effect on production performance. Silo input parameters are placed into table G.38.
Table G.38: Silo parameters for all experiments with random component for cross-contamination calculations
SiloNumber
Capacity[kg]
51
42000
52
32000
53
32000
54
22000
61
0
62
0
63
0
64
0
The dispatching rules set for this part of the investigation are put into table G.39
Table G.39: Dispatching rules fixed for scenarios with random component for cross-contamination
Mixingorder
Silomatch
BTH1-2silopref.
BTH3silopref.
BTH1-2dispatch
BTH3dispatch
Smallestorder
Bestsizefit
All
All
BiggestOrder
Smallestremainingmaterial
Then, there are variations made to the values of the scheduling alternatives as shown in table G.40, creating
essentially four groups of scheduling rules.
Table G.40: Scheduling rules chosen for scenarios including the random component
Scenario
1
2
3
4
5
6
7
8
9
10
11
Random_A
Random_B
Random_C
Random_D
Random_E
Random_F
Random_G
MixC_B
YES
YES
YES
YES
YES
YES
YES
YES
YES
YES
YES
YES
YES
YES
YES
YES
YES
YES
NO
NO
NO
YES
YES
YES
YES
YES
NO
NO
NO
NO
NO
YES
YES
YES
NO
NO
NO
NO
NO
YES
YES
YES
YES
YES
NO
YES
YES
YES
NO
NO
NO
NO
NO
NO
NO
NO
YES
YES
YES
YES
YES
YES
YES
YES
NO
NO
NO
NO
NO
NO
NO
NO
YES
YES
YES
YES
YES
YES
YES
YES
NO
NO
NO
NO
NO
NO
NO
NO
To sum up, all the performed additional scenarios with a random component have σ = 0.5 fixed. Scenario
Random_C has fewer constraints on scheduling in comparison to Random_A, where avoidance of splitting
is not forced as well as division of orders with the same recipe but a different type (scheduling alternatives
3–5). Scenario Random_D has the exact same set of input parameters as Random_C, with an exception of
nutrient-based silo allocation, which is suppressed. Similarly, scenario Random_E looks at the impact of preemptions, the exclusion of which is the only difference with experiment Random_C.
Then, scenario Random_F differs from Random_C again by a single parameter - avoidance of small order
splitting among multiple silos (scheduling alternative 3). This, is then also checked against the impact of
nutrient-based silo allocation, which concludes the prepared set.
G.5.1. E XPERIMENT R ESULTS
Table G.41: Experiment results of scenario Random_C
Scenario
KPI
Avgthroughputbags
Avgthroughputquantity
Orderscompleted
Orderscompletedontime
Avgcontaminatedorderratioabovelimit
Totalcontaminationabovelimits
Random_C
Average
1080.44
24917.05
304.40
293.35
0.08719
0.0080
SD
30.01
1402.32
0.66
3.32
0.00675
0.0006
median
1085.94
25067.18
304.50
294.50
0.08515
0.0079
no. measurements
320
320
20
20
20
20
Unit
bags/h
kg/h
-
G.5. R ANDOM C OMPONENT I NVESTIGATION
129
Table G.42: Experiment results of scenario Random_D
Scenario
KPI
Avgthroughputbags
Avgthroughputquantity
Orderscompleted
Orderscompletedontime
Avgcontaminatedorderratioabovelimit
Totalcontaminationabovelimits
Random_D
Average
1081.08
24924.59
304.50
295.10
0.08595
0.0079
SD
23.39
1358.07
0.74
4.62
0.00795
0.0006
median
1083.40
25120.96
305.00
296.50
0.08660
0.0078
no. measurements
320
320
20
20
20
20
Unit
bags/h
kg/h
-
Random_E
Average
1075.62
24830.48
303.75
290.00
0.08533
0.0080
SD
38.68
1502.97
1.70
5.04
0.00681
0.0006
median
1083.98
25069.75
304.00
289.00
0.08496
0.0082
no. measurements
320
320
20
20
20
20
Unit
bags/h
kg/h
-
Random_F
Average
1075.96
24787.01
304.05
293.20
0.08544
0.0078
SD
35.66
1458.89
1.07
3.52
0.00679
0.0007
median
1084.05
24943.80
304.00
293.00
0.08656
0.0077
no. measurements
320
320
20
20
20
20
Unit
bags/h
kg/h
-
Random_G
Average
1073.67
24723.48
304.45
295.05
0.08706
0.0081
SD
36.52
1522.69
0.74
3.17
0.00775
0.0008
median
1084.01
24873.41
305.00
295.50
0.08427
0.0081
no. measurements
320
320
20
20
20
20
Unit
bags/h
kg/h
-
Table G.43: Experiment results of scenario Random_E
Scenario
KPI
Avgthroughputbags
Avgthroughputquantity
Orderscompleted
Orderscompletedontime
Avgcontaminatedorderratioabovelimit
Totalcontaminationabovelimits
Table G.44: Experiment results of scenario Random_F
Scenario
KPI
Avgthroughputbags
Avgthroughputquantity
Orderscompleted
Orderscompletedontime
Avgcontaminatedorderratioabovelimit
Totalcontaminationabovelimits
Table G.45: Experiment results of scenario Random_G
Scenario
KPI
Avgthroughputbags
Avgthroughputquantity
Orderscompleted
Orderscompletedontime
Avgcontaminatedorderratioabovelimit
Totalcontaminationabovelimits
Figure G.8 contains a scatter plot of total contamination as a function of the average bagging throughput,
distinguished by scenarios (RaA is an equivalent of Random_A etc.).
Additionally, a scatter plot displaying relation of orders completed in set simulation time to orders finished on their due date is shown in Figure G.9. Scenarios Random_A and Random_B overlay each other.
Finally, another scatter plot, in Figure G.10, displaying total contamination as a function of the average
contamination in polluted orders. Polluted orders are those, which have at least one bag considered out of
specifications.
130
G. E XPERIMENT R ESULTS
8
Total contamination above limits [‰]
7
Scenario
MixB
RaA
RaB
RaC
6
RaD
RaE
RaF
RaG
RaH
5
RaI
4
1060
1070
1080
Average throughput in bags per hour
Figure G.8: A scatter plot displaying average results for contamination with respect to achieved throughput, for analysis with random
component
Scenario
Number of completed orders on time
292
MixB
RaA
RaB
RaC
RaD
RaE
RaF
RaG
288
RaH
RaI
284
301
302
303
304
Number of completed orders
Figure G.9: A scatter plot displaying average number of orders completed on time as a function of total number of completed orders
8
Total contamination above limits [‰]
7
Scenario
MixB
RaA
RaB
RaC
6
RaD
RaE
RaF
RaG
RaH
5
RaI
4
0.04
0.05
0.06
0.07
0.08
0.09
Average contamination in polluted orders
Figure G.10: A scatter plot displaying average contamination in polluted order against total contamination
H
O UTPUT A NALYSIS
This chapter contains statistical data analysing the experiment results included in chapter G.
H.1. I NITIAL S ENSITIVITY A NALYSIS
First of all the impact of changing discharge rates on order completion is investigated. As not all runs manage
to complete all orders the performance indicator is orders completed on time, being the difference between
orders ended and orders ended late. Thus Scenario NC_A is compared with NC_B and NC_C, then NC_E
with equivalent and so on. For reserved silos two-sided analysis is made and when varying discharge speed
one-sided as there is expectation that higher discharge speed results in more orders completes on time. Only
alternative hypothesis H1 is specified in the result tables (see tables H.1, H.2 and H.3 on the following page).
In two cases it is expected, with 95% confidence, that lower than nominal silo discharge results in fewer completed orders on time. Moreover, also in two cases reserving silos the means significantly differ and when
comparing scenarios NC_E and NC_G it might be concluded that higher discharge speed than nominal has a
positive effect of the number of orders completed on time.
Then, the influence of the increasing number of silos involved on order completion is investigated. Results of t-tests are put into tables H.4, H.5 and H.6 on the next page). They show, that there is no apparent
difference in timely order completion between 0 and 2 additional silos. But having 4 extra silos is better than
having 0 (p-value = 0.0002423) or having 2 (p-value = 0.0005541) for standard discharge rate. Having 4 additional silos and higher discharge speed is also significantly better than adding none or two.
Moreover, similar analysis is done for mixer and bagging throughputs. As only the bagging is expected
to vary more significantly at first a box plot comparing average mixing throughputs is shown in Figure G.1.
Values are indeed very similar and no further tests of means are made. Also, no following tests for different
scenario set of mixing throughput are made. Distinction between bagging machines is done further in the
report when assessing different scheduling rules and not general throughputs. As the intent is to determine
whether adding additional silos and changing silo discharge speeds has an effect on the average throughput.
T-test results are put into tables H.7, H.8, H.9 and H.10 on page 133. First of all, they all consistently
show (with 95% confidence) that lower discharge speed results in smaller average bagging throughput. Consequently, increased discharge speed has no effect on bagging speed and thus it can be concluded that the
nominal discharge speed of 20 t/h is set correctly in order not to hinder performance.
Moreover, having two additional silos has no positive effect on bagging throughput, but having four might.
Although there is no significant increase of throughput between 0 and 4 additional silos (p-value of 0.07403 ),
the one sided test between NC_E and NC_I rejects the null hypothesis, concluding with 95% confidence that
there is an increase in throughput going from 2 to 4 silos. Nevertheless, scenario NC_I has the biggest average bagging throughput and lower standard deviation. From analysis of production performance, without
including cross-contamination calculations can be concluded that silo discharge speed of 20 t/h is needed
and increasing it further to 22 t/h (and boosting discharge to big bag station to 15 t/h) will not increase performance.
131
132
H. O UTPUT A NALYSIS
Table H.1: Comparison of impact of different discharge speeds on order completion for no additional silos
Hypothesis
n
mean
SD
varequal
t
df
p
95%confidenceinterval
H 1 :NC_A>NC_B
H 1 :NC_A<NC_C
H 1 :NC_A!=NC_D
40
40
40
248.00
248.75
249.15
3.11
2.66
3.39
TRUE
TRUE
TRUE
1.7736
0.23506
-0.55414
38
38
38
0.04207
0.5923
0.5827
0.08404611: Inf
-Inf: 1.634496
-2.791932: 1.591932
Table H.2: Comparison of impact of different discharge speeds on order completion for 2 additional silos
Hypothesis
n
mean
SD
varequal
t
df
p
95%confidenceinterval
H 1 :NC_E>NC_F
H 1 :NC_E<NC_G
H 1 :NC_E!=NC_H
40
40
40
247.93
249.15
249.30
5.50
4.94
4.46
TRUE
TRUE
TRUE
-0.085102
-1.7051
-2.1516
38
38
38
0.5337
0.04817
0.03784
-3.12164: Inf
-Inf: -0.02926495
-5.628489: -0.171511
Table H.3: Comparison of impact of different discharge speeds on order completion for 4 additional silos
Hypothesis
n
mean
SD
varequal
t
df
p
95%confidenceinterval
H 1 :NC_I>NC_J
H 1 :NC_I<NC_K
H 1 :NC_I!=NC_L
40
40
40
251.47
251.85
251.38
3.27
3.32
3.11
TRUE
TRUE
TRUE
2.0665
1.2452
2.4264
38
38
38
0.02282
0.8897
0.0201
0.3775205: Inf
-Inf: 3.060188
0.3727977: 4.1272023
Table H.4: Comparison of impact of 0 to 2 additional silos on order completion
Hypothesis
n
mean
SD
varequal
t
df
p
95%confidenceinterval
H 1 :NC_A<NC_E
H 1 :NC_A<NC_G
40
40
248.35
249.65
3.96
3.91
FALSE
FALSE
0.79509
-1.3047
29.935
30.542
0.7836
0.1009
-Inf: 3.134814
-Inf: 0.4802682
Table H.5: Comparison of impact of 0 to 4 additional silos on order completion
Hypothesis
n
mean
SD
varequal
t
df
p
95%confidenceinterval
H 1 :NC_A<NC_I
H 1 :NC_A<NC_K
40
40
250.68
250.03
3.51
3.25
TRUE
TRUE
-3.8166
-2.4288
38
38
0.0002423
0.009995
-Inf: -2.037638
-Inf: -0.7187276
Table H.6: Comparison of impact of 2 to 4 additional silos on order completion
Hypothesis
n
mean
SD
varequal
t
df
p
95%confidenceinterval
H 1 :NC_E<NC_I
H 1 :NC_E<NC_K
40
40
250.18
249.53
4.74
4.47
TRUE
TRUE
-3.5296
-2.5271
38
38
0.0005541
0.007891
-Inf: -2.42885
-Inf: -1.115085
Table H.7: Comparison of impact of silo discharge and reservation on bagging throughput with no additional silos
Hypothesis
n
mean
SD
t
df
p
95%confidenceinterval
H 1 :NC_A>NC_B
H 1 :NC_A<NC_C
H 1 :NC_A!=NC_D
600
600
600
1062.63
1072.23
1066.97
42.38
41.16
49.91
6.2773
0.54843
3.0572
597.56
597.7
550.32
3.312e-10
0.7082
0.002342
15.53185: Inf
-Inf: 7.383496
4.423463: 20.323430
Table H.8: Comparison of impact of silo discharge and reservation on bagging throughput with 2 additional silos
Hypothesis
n
mean
SD
t
df
p
95%confidenceinterval
H 1 :NC_E>NC_F
H 1 :NC_E<NC_G
H 1 :NC_E!=NC_H
600
600
600
1058.65
1070.7
1071.28
54.47
44.09
44.08
6.0044
0.51549
0.19671
554
597.31
597.29
1.739e-09
0.6968
0.8441
18.83294: Inf
-Inf: 7.79169
-6.364977: 7.781979
H.2. C ROSS - CONTAMINATION I NVESTIGATION
133
Table H.9: Comparison of impact of silo discharge and reservation on bagging throughput with 4 additional silos
Hypothesis
n
mean
SD
t
df
p
95%confidenceinterval
H 1 :NC_I>NC_J
H 1 :NC_I<NC_K
H 1 :NC_I!=NC_L
600
600
600
1068.84
1076.93
1075.46
29.57
30.89
33.87
7.3952
0.36823
1.3966
597.17
588.22
556.46
2.4e-13
0.6436
0.1631
13.29432: Inf
-Inf: 5.087692
-1.568617: 9.287001
Table H.10: Comparison of silo discharge speed and number of extra silos impact on bagging throughputs
Hypothesis
n
mean
SD
t
df
p
95%confidenceinterval
H 1 :NC_A<NC_E
H 1 :NC_A<NC_I
H 1 :NC_A<NC_K
H 1 :NC_E<NC_I
600
600
600
600
1072.39
1075.27
1074.81
1074.51
43.25
35.85
37.51
37.80
0.43074
-1.4486
-1.0801
-1.8699
594.7
532.33
567.24
510.18
0.6666
0.07403
0.1403
0.03103
-Inf: 7.344177
-Inf: 0.5825061
-Inf: 1.737458
-Inf: -0.6840012
H.2. C ROSS - CONTAMINATION I NVESTIGATION
Table H.11: Comparison of impact of order sequencing and nutrient-based silo allocation on total cross-contamination
Hypothesis
n
mean
SD
varequal
t
df
p
95%conf. interval
H 1 :MassC_A<MassC_B
H 1 :MassC_A<MassC_C
H 1 :MassC_A<MassC_E
H 1 :MassC_A>Mix_A
H 1 :MassC_C<MassC_D
H 1 :MassC_C<MassC_E
H 1 :MassC_C>Mix_A
H 1 :MassC_E>Mix_A
40
40
40
40
40
40
40
40
0.01112
0.01403
0.01394
0.00835
0.01684
0.01676
0.01117
0.01108
0.00074
0.00293
0.00290
0.00295
0.01394
0.00835
0.00078
0.00570
TRUE
TRUE
TRUE
FALSE
TRUE
FALSE
FALSE
FALSE
0.8324
-26.45
-21.57
34.142
0.0281
0.6643
74.672
47.208
38
38
38
23.4
38
31.6
24.5
21.1
0.7948
<2.2e-16
<2.2e-16
<2.2e-16
0.5111
0.7443
<2.2e-16
<2.2e-16
-Inf: 0.00059
-Inf: -0.00527
-Inf: -0.00518
0.00544: Inf
-Inf: 0.00043
-Inf: 0.00064
0.01110: Inf
0.01077: Inf
H.3. L AYOUT I NTERVENTIONS I NVESTIGATION
Table H.12: Comparison of impact of additional silos on order completion
Hypothesis
n
mean
SD
varequal
t
df
p
95%conf. interval
H 1 :MassC_G>MassC_F
H 1 :MassC_H>MassC_F
H 1 :MassC_I>MassC_F
H 1 :MassC_J>MassC_I
H 1 :MassC_K>MassC_J
H 1 :MassC_L>MassC_K
H 1 :MassC_M>MassC_L
H 1 :MassC_N>MassC_L
H 1 :MassC_M>MassC_N
H 1 :MassC_O>MassC_K
H 1 :MassC_L>MassC_O
H 1 :MassC_P>MassC_K
H 1 :MassC_O>MassC_P
H 1 :MassC_R>MassC_I
H 1 :MassC_K>MassC_R
40
40
40
40
40
40
40
40
40
40
40
40
40
40
40
291.95
291.30
291.32
293.27
293.55
294.60
295.68
293.82
293.75
292.60
293.75
293.07
292.23
293.32
293.60
4.56
3.67
4.10
3.88
3.16
2.89
2.57
3.42
3.22
3.95
4.42
3.25
4.36
4.47
3.86
TRUE
TRUE
TRUE
TRUE
TRUE
TRUE
TRUE
TRUE
TRUE
FALSE
FALSE
TRUE
TRUE
TRUE
FALSE
3.455
2.9132
2.6005
0.6058
-0.1975
2.715
-0.1826
-4.2795
4.4009
-1.3776
3.1793
-0.7263
-0.6836
0.5958
-0.2427
38
38
38
38
38
38
38
38
38
28.7
30.2
38
38
38
28.7
0.0006841
0.002981
0.006594
0.2741
0.5778
0.004956
0.5719
0.9999
4.226e-05
0.9105
0.0017
0.764
0.7508
0.2774
0.595
2.252911: Inf
1.305929: Inf
1.107811: Inf
-1.337114: Inf
-1.906899: Inf
0.8717751: Inf
-1.535288: Inf
-5.366743: Inf
2.282552: Inf
-3.797477: Inf
1.865045: Inf
-2.490852: Inf
-3.293081: Inf
-1.555461: Inf
-2.401188: Inf
134
H. O UTPUT A NALYSIS
Table H.13: Comparison of impact of additional silos on total contamination
Hypothesis
n
mean
SD
varequal
t
df
p
95%conf. interval
H 1 :MassC_G<MassC_F
H 1 :MassC_H<MassC_F
H 1 :MassC_I<MassC_F
H 1 :MassC_J<MassC_I
H 1 :MassC_K<MassC_J
H 1 :MassC_L<MassC_K
H 1 :MassC_M<MassC_L
H 1 :MassC_N<MassC_L
H 1 :MassC_M<MassC_N
H 1 :MassC_O<MassC_K
H 1 :MassC_L<MassC_O
H 1 :MassC_P<MassC_K
H 1 :MassC_O<MassC_P
H 1 :MassC_R<MassC_I
H 1 :MassC_K<MassC_R
40
40
40
40
40
40
40
40
40
40
40
40
40
40
40
0.00975
0.00961
0.00945
0.00828
0.00803
0.00845
0.00901
0.00817
0.00846
0.00805
0.00833
0.00801
0.00789
0.00837
0.00812
0.00085
0.00089
0.00102
0.00069
0.00052
0.00061
0.00067
0.00079
0.00103
0.00055
0.00071
0.00052
0.00054
0.00059
0.00042
TRUE
TRUE
TRUE
TRUE
TRUE
TRUE
TRUE
TRUE
TRUE
TRUE
TRUE
TRUE
TRUE
FALSE
TRUE
-4.325
-6.071
-7.585
-4.329
1.764
3.175
2.958
-6.173
9.002
-1.398
4.260
-2.051
0.491
-3.820
0.801
38
38
38
38
38
38
38
38
38
38
38
38
38
28.6
38
5.329e-05
2.275e-07
2.014e-09
5.251e-05
0.9572
0.9985
0.9973
1.647e-07
1
0.08511
0.9999
0.0236
0.6867
0.00033
0.786
-Inf: -0.00059
-Inf: -0.00089
-Inf: -0.00121
-Inf: -0.00048
-Inf: 0.00055
-Inf: 0.00084
-Inf: 0.00090
-Inf: -0.00080
-Inf: 0.00198
-Inf: 4.99e-05
-Inf: 0.00111
-Inf: -5.82e-05
-Inf: 0.00038
-Inf: -0.00034
-Inf: 0.00034
Table H.14: Comparison of impact of additional silos on the average throughput
Hypothesis
H 1 :MassC_G>MassC_F
H 1 :MassC_H>MassC_F
H 1 :MassC_I>MassC_F
H 1 :MassC_J>MassC_I
H 1 :MassC_K>MassC_J
H 1 :MassC_L>MassC_K
H 1 :MassC_M>MassC_L
H 1 :MassC_N>MassC_L
H 1 :MassC_M>MassC_N
H 1 :MassC_O>MassC_K
H 1 :MassC_L>MassC_O
H 1 :MassC_P>MassC_K
H 1 :MassC_O>MassC_P
H 1 :MassC_R>MassC_I
H 1 :MassC_K>MassC_R
n
mean
SD
t
df
p
95%conf. interval
640
640
640
640
640
640
640
640
640
640
640
640
640
640
640
1068.37
1068.08
1069.33
1076.03
1076.93
1080.3
1083.97
1078.4
1078.7
1076.83
1080.2
1078.02
1077.92
1075.9
1076.8
37.49
40.68
37.95
35.91
37.26
32.93
26.02
33.22
32.03
36.78
32.4
34.06
33.54
35.2
36.58
3.2889
2.8426
3.9084
0.6341
0
2.6019
0.2844
-4.0613
4.4589
-0.0692
2.725
0.8073
-0.8959
0.5526
0.0907
619.152
637.975
623.146
634.22
638
587.403
628.911
586.819
549.507
637.562
594.847
613.98
619.572
636.995
637.099
0.0005315
0.002309
5.155e-05
0.26317
0.5
0.004752
0.3881
1
4.997e-06
0.5276
0.003309
0.2099
0.8147
0.2904
0.4639
4.828893: Inf
3.822827:Inf
6.708968: Inf
-2.877708: Inf
-4.855933: Inf
2.473952: Inf
-2.805336: Inf
-14.8164: Inf
7.014838: Inf
-4.994947: Inf
2.746652: Inf
-2.262811: Inf
-6.744498: Inf
-3.047902: Inf
-4.504707: Inf
H.4. R ANDOM COMPONENT
Table H.15: Comparison of impact of random component on the average throughput
Hypothesis
H 1 :Ra_A<MassC_J
H 1 :Ra_B<MassC_J
H 1 :MixC_B!=MassC_J
n
mean
SD
t
df
p
95%conf. interval
640
640
640
1077.27
1077.27
1078.02
37.86
37.86
34.06
0.2261
0.2261
0.8073
637.366
637.366
613.98
0.5894
0.5894
0.4198
-Inf: 5.611756
-Inf: 5.611756
-3.115306: 7.464032
Table H.16: Comparison of impact of random component on the total contamination
Hypothesis
n
mean
SD
varequal
t
df
p
95%conf. interval
H 1 :Ra_A!=MassC_J
H 1 :Ra_B!=MassC_J
H 1 :Ra_A!=Ra_B
H 1 :MixC_B!=MassC_J
40
40
40
40
0.00794
0.00793
0.00798
0.00573
0.00057
0.00056
0.00062
0.00222
TRUE
TRUE
TRUE
FALSE
0.553
0.4082
0.1359
-37.379
38
38
38
21.1
0.5835
0.685
0.8926
<2.2e-16
-0.00027: 0.00047
-0.00029: 0.00044
-0.00038: 0.00043
-0.00456: -0.00408
R EFERENCES
Adler, P. S., Goldoftas, B., & Levine, D. I. (1999). Flexibility versus efficiency? A case study of model
changeovers in the toyota production system. Organization Science, 10, 43–68. doi: 10.1287/orsc.10.1.43
Beamon, B. M. (1998). Supply chain design and analysis: Models and methods. International Journal of
Production Economics, 55, 281–294. doi: 10.1016/S0925-5273(98)00079-6
Behnke, K. C. (1996). Feed manufacturing technology: current issues and challenges. Animal Feed Science
and Technology, 62(1), 49–57. doi: 10.1016/S0377-8401(96)01005-X
Benjaafar, S., Heragu, S. S., & Irani, S. A. (2002). Next generation factory layouts: Research challenges and
recent progress. Interfaces, 32(6), 58–76. doi: 10.1287/inte.32.6.58.6473
Bertrand, F., Leclaire, L.-A., & Levecque, G. (2005). DEM-based models for the mixing of granular materials.
Chemical Engineering Science, 60(8–9), 2517–2531. (5th International Symposium on Mixing in Industrial
Processes (ISMIP5)) doi: 10.1016/j.ces.2004.11.048
Blackstone, J. H., Philips, D. T., & Hogg, G. L. (1982). A state-of-the-art survey of dispatching rules for
manufacturing job shop operations. International Journal of Production Research, 20(1), 27–45. doi:
10.1080/00207548208947745
Clark, A., Mahdieh, M., & Rangel, S. (2014). Production lot sizing and scheduling with non-triangular
sequence-dependent setup times. International Journal of Production Research, 52(8), 2490–2503. doi:
10.1080/00207543.2014.885662
Cleary, P. W., & Sawley, M. L. (2002). DEM modelling of industrial granular flows: 3d case studies and
the effect of particle shape on hopper discharge. Applied Mathematical Modelling, 26(2), 89–111. doi:
10.1016/S0307-904X(01)00050-6
Crawshaw, R. (2012). 2 - animal feeds, feeding practices and opportunities for feed contamination: an introduction. In J. Fink-Gremmels (Ed.), Animal feed contamination (pp. 11–32). Woodhead Publishing. doi:
10.1533/9780857093615.11
Djassemi, M. (2007). Improving factory layout under a mixed floor and overhead material handling condition.
Journal of Manufacturing Technology Management, 18(3), 281–291. doi: 10.1108/17410380710730611
Drexl, A., & Kimms, A. (1997). Lot sizing and scheduling — survey and extensions. European Journal of
Operational Research, 99(2), 221–235. doi: 10.1016/S0377-2217(97)00030-1
Erenguc, S., Simpson, N., & Vakharia, A. (1999). Integrated production/distribution planning in supply
chains: An invited review. European Journal of Operational Research, 115(2), 219–236. doi: 10.1016/S03772217(98)90299-5
Fink-Gremmels, J. (2012). 1 - introduction to animal feed contamination. In J. Fink-Gremmels (Ed.), Animal
feed contamination (pp. 1–10). Woodhead Publishing. doi: 10.1533/9780857093615.1
Fitzpatrick, J., Iqbal, T., Delaney, C., Twomey, T., & Keogh, M. (2004). Effect of powder properties and storage
conditions on the flowability of milk powders with different fat contents. Journal of Food Engineering,
64(4), 435–444. doi: 10.1016/j.jfoodeng.2003.11.011
Gupta, Y. P., & Goyal, S. (1992). Flexibility trade-offs in a random flexible manufacturing system: A simulation
study. International Journal of Production Research, 30(3), 527–557. doi: 10.1080/00207549208942910
Herrmann, J. (2006). Handbook of production scheduling. Springer. doi: 10.1007/0-387-33117-4
135
136
R EFERENCES
Jeong, K.-C., & Kim, Y.-D. (1998). A real-time scheduling mechanism for a flexible manufacturing system:
Using simulation and dispatching rules. International Journal of Production Research, 36(9), 2609–2626.
doi: 10.1080/002075498192733
Józefowska, J., & Zimniak, A. (2008). Optimization tool for short-term production planning and scheduling.
International Journal of Production Economics, 112(1), 109–120. doi: 10.1016/j.ijpe.2006.08.026
Kelton, D. W., Smith, J. S., & Sturrock, D. T. (2013). Simio and simulation: Modeling, analysis, applications
(3rd ed.). Simio LLC.
Kämpf, M., & Köchel, P. (2006). Simulation-based sequencing and lot size optimisation for a production-andinventory system with multiple items. International Journal of Production Economics, 104(1), 191–200. doi:
10.1016/j.ijpe.2006.02.008
Ku, H. (1966). Notes on the use of propagation of error formulas. Journal of Research of the National Bureau
of Standards, 70C, 263–273. doi: 10.6028/jres.070C.025
Lee, I., Sikora, R., & Shaw, M. (1997). A genetic algorithm-based approach to flexible flow-line scheduling with
variable lot sizes. Systems, Man, and Cybernetics, Part B: Cybernetics, IEEE Transactions on, 27(1), 36–54.
doi: 10.1109/3477.552184
Leloup, M., Solliec, C., & Putter, F. (2011). How process operations affect cross contamination in animal feed
industry. In International Congress On Engineering And Food Process Engineering in a Changing World (pp.
22–26). Athens, Greece.
Liao, H.-T., & Shiau, C.-Y. (2000). Analytical solution to an axial dispersion model for the fixed-bed adsorber.
AIChE Journal, 46(6), 1168–1176. doi: 10.1002/aic.690460609
Mapes, J., New, C., & Szwejczewski, M. (1997). Performance trade-offs in manufacturing plants. International
Journal of Operations & Production Management, 17(10), 1020–1033. doi: 10.1108/01443579710177031
Mieghem, J. A. V. (2003). Commissioned paper: Capacity management, investment, and hedging: Review and recent developments. Manufacturing & Service Operations Management, 5(4), 269–302. doi:
10.1287/msom.5.4.269.24882
Min, H., & Zhou, G. (2002). Supply chain modeling: past, present and future. Computers & Industrial Engineering, 43(1-2), 231–249. doi: 10.1016/S0360-8352(02)00066-9
Mula, J., Poler, R., García-Sabater, G.-S., & Lario, F. (2006). Models for production planning under uncertainty:
A review. International Journal of Production Economics, 103(1), 271–285. doi: 10.1016/j.ijpe.2005.09.001
Negahban, A., & Smith, J. S. (2014). Simulation for manufacturing system design and operation: Literature
review and analysis. Journal of Manufacturing Systems, 33(2), 241–261. doi: 10.1016/j.jmsy.2013.12.007
Pinedo, M. L. (2008). Scheduling: Theory, algorithms, and systems (4th ed.). Springer Publishing Company,
Incorporated. doi: 10.1007/978-0-387-78935-4
Portillo, P., Muzzio, F., & Ierapetritou, M. (2006). Characterizing powder mixing processes utilizing compartment models. International Journal of Pharmaceutics, 320(1–2), 14–22. doi: 10.1016/j.ijpharm.2006.03.051
Poux, M., Fayolle, P., Bertrand, J., Bridoux, D., & Bousquet, J. (1991). Powder mixing: Some practical rules
applied to agitated systems. Powder Technology, 68(3), 213–234. doi: 10.1016/0032-5910(91)80047-M
Sandborn, P. A., & Vertal, M. (1998). Analyzing packaging trade-offs during system design. IEEE Design & Test
of Computers, 15(3), 10–19. doi: 10.1109/54.706028
Sargent, R. (2007, Dec). Verification and validation of simulation models. In Proceedings of the Winter Simulation Conference (pp. 124–137). doi: 10.1109/WSC.2007.4419595
Schriber, T. J., Brunner, D. T., & Smith, J. S. (2012). How discrete-event simulation software works and why it
matters. In Proceedings of the Winter Simulation Conference (pp. 1–15). doi: 10.1109/WSC.2012.6465274
R EFERENCES
137
Sen, M., & Ramachandran, R. (2013). A multi-dimensional population balance model approach to continuous
powder mixing processes. Advanced Powder Technology, 24(1), 51–59. doi: 10.1016/j.apt.2012.02.001
Silveira, G. D., & Slack, N. (2001). Exploring the trade-off concept. International Journal of Operations &
Production Management, 21(7), 949–964. doi: 10.1108/01443570110393432
Smith, J. S. (2003). Survey on the use of simulation for manufacturing system design and operation. Journal
of Manufacturing Systems, 22(2), 157–171. doi: 10.1016/S0278-6125(03)90013-6
Strauch, W. (2003). Is contamination-free feed production realistic? Feed Tech, 7(7), 23–25.
Taylor, S., & Baumert, J. (2010). Cross-contamination of foods and implications for food allergic patients.
Current Allergy and Asthma Reports, 10(4), 265–270. doi: 10.1007/s11882-010-0112-4
Terzi, S., & Cavalieri, S. (2004). Simulation in the supply chain context: a survey. Computers in Industry, 53(1),
3–16. doi: 10.1016/S0166-3615(03)00104-0
Toso, E. A., Morabito, R., & Clark, A. R. (2009). Lot sizing and sequencing optimisation at an animal-feed
plant. Computers & Industrial Engineering, 57(3), 813–821. doi: 10.1016/j.cie.2009.02.011
Trienekens, J., & Zuurbier, P. (2008). Quality and safety standards in the food industry, developments and challenges. International Journal of Production Economics, 113(1), 107–122. doi: 10.1016/j.ijpe.2007.02.050
van der Zee, D.-J., & van der Vorst, J. G. A. J. (2005). A modeling framework for supply chain simulation: Opportunities for improved decision making. Decision Sciences, 36(1), 65–95. doi: 10.1111/j.15405915.2005.00066.x
Wu, J., Binbo, J., Chen, J., & Yang, Y. (2009). Multi-scale study of particle flow in silos. Advanced Powder
Technology, 20(1), 62–73. doi: 10.1016/j.apt.2008.02.003
Zeigler, B., Praehofer, H., & Kim, T. (2000). Theory of modeling and simulation: Integrating discrete event and
continuous complex dynamic systems. Academic Press. doi: 10.1002/rnc.610
Zervas, G., Nikolaou, E., & Mantzios, A. (1990, 6). Comparative study of chronic copper poisoning in lambs
and young goats. Animal Production, 50, 497–506. doi: 10.1017/S0003356100004980
Download PDF