Decision Support for Production Control based on Machine Learning

by Simulation-generated Data

Konstantin Muehlbauer

, Lukas Rissmann

and Sebastian Meissner

Technology Center for Production and Logistics Systems, Landshut University of Applied Sciences, Am Lurzenhof 1,

Landshut, Germany

Keywords: Artificial Intelligence, Decision-making, Machine Learning, Order-sequence Optimization, Logistics

Simulation.

Abstract: Data-oriented approaches enable new opportunities to analyze processes and support managers in decision-

making during planning and control tasks. In particular, the application of simulations has been a widely used

tool for many years to evaluate alternative system configurations or to predict future process outcome. Due

to a rapidly changing environment in a cross-linked domain such as production and logistics systems, more

and more decisions have to be made in a shorter time under consideration of multi-factorial influences.

Simulation based approaches often reach limits regarding time constraints assuming limited computing power.

The article describes how data, generated by production and logistics simulation can be used to train a machine

learning model. Thus, the generalized framework presented can be utilized to support decision-making during

planning and control tasks. By applying the framework to a case study on order sequence optimization, it was

possible to verify its feasibility and potential to improve the operational performance of a manufacturing

system.

1 INTRODUCTION AND

PROBLEM STATEMENT

The ongoing technological progress enables new

potentials regarding planning and control of

production and logistics systems (Windt et al., 2008).

One fundamental aim of computational applications

in this field is to support managers in time-consuming

activities or activities with a high degree of

complexity regarding decision-making. In particular,

potential through data-oriented approaches (e.g.,

simulation or machine learning) can be leveraged in

areas where enormous amount of data and its

situational dependency has to be considered. (Hasan

et al., 2016; Koot et al., 2021)

Simulations have been used for many years to

support decision-making during planning of

production and logistics systems (Pfeiffer et al.,

2016). The use of simulations in production control

will also become more important due to the further

https://orcid.org/0000-0003-0986-7009

https://orcid.org/0000-0002-9747-7707

https://orcid.org/0000-0002-5808-9648

implementation of digital twins. A digital twin is a

virtual representation of a physical object or process

(Kauke et al., 2021). It should help to understand the

behavior of an object by a dynamic prediction based

on diverse data (Qi and Tao, 2018). Simulations are

often an essential part of digital twins (Kritzinger et

al., 2018).

The rising complexity of production and logistics

systems also leads to increasingly demanding

requirements for simulation models and necessitate

an growing amount of simulation runs in order to

better represent the reality (Rose, 2007). In particular,

executing different scenarios can make simulation

runs computing and time intensive. Despite increased

computing power, simulating various problems can

take more time than is available (Rose, 2007). In case

of time-critical decisions, this might imply that not all

alternative scenarios can be simulated in time. Thus,

only an insufficiently evaluated decision can be

made. Machine learning (ML) can provide a solution

to this problem. Based on a trained ML model

Muehlbauer, K., Rissmann, L. and Meissner, S.

Decision Support for Production Control based on Machine Lear ning by Simulation-generated Data.

DOI: 10.5220/0011538000003335

In Proceedings of the 14th International Joint Conference on Knowledge Discovery, Knowledge Engineering and Knowledge Management (IC3K 2022) - Volume 3: KMIS, pages 54-62

ISBN: 978-989-758-614-9; ISSN: 2184-3228

enormous amounts of data can be processed and

evaluated faster and, thus, time-critical decisions can

be made on time.

This article presents a framework for application

of simulation models of manufacturing and logistics

processes to generate data in order to train an ML

model. Based on an ML model that has been trained

in advance, the objective is to support decision-

making through predictive analytics for planning and

control tasks in manufacturing and logistics systems.

The approach describes important information flows

and process steps required. The framework is

designed to tackle two challenges. On the one hand,

the approach can help to cope with the issue that

simulating different scenarios takes more time than is

available for time-critical decisions. On the other

hand, it enables the training of ML models in

processes with a quantitatively or qualitatively

limited data basis. Additional simulation-generated

data can be provided as training data, thus, enabling

better control decision. This can increase the

performance of a production and logistics system

significantly. With regard to the current performance

and applicability of ML, as well as the extensive

availability of simulation tools for production and

logistics systems, this article aims to answer the

following research question (RQ):

RQ: How can a generalized framework for machine-

learning-application based on discrete-event

simulation be described in order to be implemented

for decision-support in production and logistics

control system with insufficient data quality and

quantity?

2 RESEARCH ADVANCES

In following section, current research advances in

simulation and ML as well as applications within

manufacturing and logistics systems are described.

2.1 Simulation of Production and

Logistics Processes

A simulation is a method of reproducing a system in

an experimentable model, which can be used to

observe and analyze the temporal behavior of

complex systems (VDI 3633, 2018). Simulations can

help companies develop, implement, and execute

plans and strategies, giving them a significant

competitive advantage. They have proven their

potential by predicting performance, utilization,

bottlenecks, as well as analyzing interactions of

different components of a system. Results of

simulations can significantly improve decisions in

terms of planning and control. A key advantage

compared to other operations research approaches is

the ability to perform experiments with different

elements of a business system (Agalianos et al.,

2020). Fowler and Rose mention further advantages

such as time compression, component integration,

and risk avoidance. Simulation models are already

often used for applications in high-tech production

systems such as semiconductor or automotive

industries (Fowler and Rose, 2004).

Application scenarios of simulation models

regarding short-term decision-making within

production and logistics systems are described below.

Korth et al. developed a simulation model within a

digital twin for a critical real-time use case in logistics.

Objective of the application is to support shift planning

of employees and time window planning within a

warehouse (Korth et al., 2018). Kauke et al. describe a

digital twin for order picking systems by using a

simulation. It is emphasized that due to a high system

complexity, simulation is often the only way to check

different parameters of a picking system. Simulations

should help to support decisions like the size of picking

orders, use of employees, or order-release strategies

(Kauke et al., 2021). Further applications of

discrete-event simulation within production and

logistics systems are shown by Agalianos et al. in their

literature review.

However, the literature indicates that in particular

real-time simulations are still in early development

phase. According to the current state, the application

of simulation for time-critical decisions is only

possible with: (1) use of a simulation model that runs

continuously and is synchronized with the factory, (2)

automated modeling of a simulation models based on

the factory data basis, or (3) by simplifying the

simulation (Fowler and Rose, 2004; Rose, 2007).

2.2 Simulation-based Machine

Learning in Production and

Logistics Systems

In literature the combination of ML and simulation

models are described in different applications.

Vernickel et al. introduce a ML approach for

parameterizing and synchronizing a material flow

simulation model. This approach shows how ML can

be used to identify relevant process information from

a dataset and integrate this information into a

simulation model. This enables a better determination

of resource processing time compared to a normal

simulation model (Vernickel et al., 2020). Nagahara

Decision Support for Production Control based on Machine Learning by Simulation-generated Data

et al. pursue a job sequencing rule identification

method by using ML to generate an automatic

modeling of operational control rules for a

simulation. Another approach is presented by Müller

et al. using a material flow simulation to control

automated guided vehicles which communicates with

other digital twins, e.g., in manufacturing cells. Other

author attempts to validate an ML model for

predicting disruptive effects in production logistics

by simulation models (Vojdani and Erichsen, 2018).

The generated data of a simulation model represent

real production data. This could be important as some

companies do not have the necessary data basis to use

ML. Data from simulation models are often the only

way to test an algorithm's applicability in advance and

to transfer them to a real production or logistics

system. Pfeiffer et al. describe an approach for multi-

model-based prediction of lead times within a

manufacturing system. The method is tested on data

generated by a simulation model.

The literature analysis shows that the application

of simulation in combination with ML in production

and logistics systems increases. The lack of sufficient

real production and logistics data encourages the

usage of simulation models to generate data for ML.

In summary, three possible applications for the

collaborative use of ML and simulations can be

identified. First, ML can support simulation runs by

optimizing the parameterization of simulation

models. Secondly, ML could make complete

predictions on its own and replace the entire

simulation model (also called surrogate modelling)

(Bárkányi et al., 2021). The third application is to use

simulations to generate data of production and

logistics systems for training and validating ML

models.

3 MACHINE LEARNING

FRAMEWORK BASED ON

SIMULATION DATA

Section 3.1 describes the developed framework with

all components required for the implementation and

section 3.2 presents an application example by a

specific case study.

3.1 Components of the Framework

The framework consists of different components

which are shown in Figure 1. The following

components are required: (1) problem statement

level, (2) input data to perform a simulation run, (3) a

validated simulation model, (4) output data of a

simulation run, (5) a data preparation utility, and (6)

a selected ML model.

Based on a key performance indicator (KPI)

system and the deviation between target and actual

values, an identified potential for optimization in the

respective production and logistics system serves as

the starting point for the application. Consequently, a

target can be determined. This target has to be

reflected in the real system by one or more KPIs (e.g.,

lead time, throughput, failures). The factors that

influence the target (process parameters and process

constants) or other causes have to be determined from

the real production and logistics systems. Due to the

fact that not all influencing factors can be adjusted,

control variables have to be defined. Various process

analysis methods as well as expert knowledge have to

be used for this. This can be done manually (e.g., by

Value Stream Mapping) or with data-oriented

approaches (e.g., by rule-based or ML approaches).

The identified control variables as well as the target

KPIs will be used for the ML model.

After the problem statement and the analysis of

the process, simulation input data has to be prepared

in order to generate sufficient simulation output data.

Also, different control variables have to be defined,

so that the simulation model can be parameterized

depending on the application. This simulation input is

used for a validated simulation model of the

production and logistics system. The model should

reflect the real process in as much detail as is

reasonable based on the defined target KPIs and

influencing factors. It is necessary to ensure that

results of this model have been checked in advance

and produce comparable results to the real process.

For this purpose, it is important to use the same data

structure between input data of the simulation and the

real system. This is critical to validate the results of

the simulation. It has to be mentioned that by using

simulations as well as simulations in combination

with ML multiple factors for inaccuracies can exist.

Further research is required on this issue. These

effects are not considered in this article.

The simulation input (e.g., production or transport

orders, resources, etc.) required for implementing a

simulation model can be taken from different systems

such as Enterprise Resource Planning (ERP),

Manufacturing Execution System (MES), or

Warehouse Management System (WMS). The

aforementioned input should be used for the

simulation model as well as the ML model. Based on

the defined simulation input (2) and a validated

simulation model (3), the required output data (4) for

determining the target KPIs can be generated.

KMIS 2022 - 14th International Conference on Knowledge Management and Information Systems

Through the use of a simulation model it is

possible to generate multiple years’ worth of data

where only the input variables of the system have

changed. Thus, the configuration as well as the

restrictions of the production and logistics systems

are the same. This allows occurrences that happen

very rarely to be reflected in the data and provide

comprehensive data for training an ML model.

Furthermore, existing datasets or datasets with

insufficient quality and quantity can be enriched with

additional data. For determining the number of entries

in the dataset required, it has to be considered that the

duration of a single simulation run can be a regulating

variable. It is not possible to specify the quantity of

required entries in a dataset. This is due to different

factors such as the complexity of the simulation as

well as the number of process parameters. The

complexity of the problem to be solved or the ML task

can also influence the number of entries.

Furthermore, the input and generated output data

of each simulation run (=simulation results) must be

stored together. This can be done using a database

system e.g., SQLite, etc. Next, the dataset can be split

by a random training and test splitting function for

cross validation. The training is performed with the

available features (=simulation input) and labels

(=simulation output). Once the generated data has

been splitted, it is analyzed in the next step using

various data preparation methods. Where, incorrect or

missing data are sorted out. By using a simulation

model only failed simulation runs or incorrect models

can create erroneous data. In case of using a validated

simulation model, no erroneous simulation runs

should occur at this step. Nevertheless, the results

should be checked as failed simulation runs can create

incorrect or incomplete data. Furthermore, data

argumentation as well as techniques for

dimensionality reduction can applied.

In complex problems, ML clustering (e.g., K-

Means, etc.) may be used to automatically find

patterns or correlations in the data. This allows to

create classes, which substitute the original label.

Techniques such as scaling and principal component

analysis (PCA) can used to improve model

performance. This may help to cope with imbalanced

data and improve the ML result. Subsequently, it

must be decided which ML learning type, ML class,

and finally ML model are roughly suitable. As labels

are available, supervised learning is selected.

Supervised learning, involves learning with different

features of a dataset, annotated with a label. The goal

is to map input to output values by minimizing the

discrepancy between real and predicted values within

the dataset (Goodfellow et al., 2016). Through the

description of the application scenario, an ML class

can be specified e.g., classification or regression.

Thus, possible models can be delimited. A regression

model attempts to predict continuous values based on

given data (Han et al., 2012). On the other hand, a

classification model aims to predict a correct class

from several classes of data (Han et al., 2012).

As shown in Figure 1, it is important to define the

output requirements in order to be able to evaluate the

final ML results. Nevertheless, a specific requirement

cannot be described due to the different application

scenarios within production and logistics systems.

However, the proposed solution must be better than

the current approach. Thus, the performance or

reliability of the system should be increased, such as

lower throughput times, better adherence to

schedules, or an increased throughput. These KPIs

can be specified in percentages or absolute numbers.

After training, the model is validated with a

comparison between a prediction of previously

unknown data against the labeled data. Thus, the ML-

based results are referred to as predicted, while the

simulation-based results are referred to as real. It is

analyzed whether the prediction accuracy achieved

by the ML model is sufficient to meet the defined

output (performance) requirements. Based on the

chosen ML model, metrics are used to quantify the

ML results. For regression tasks ‘mean absolute

error’ and ‘root mean squared error’ can be applied.

In classification, ‘accuracy’, ‘recall’, ‘precision’, and

‘f1-score’ can be used.

Depending on the complexity and quality of the

ML results, this step leads to further iteration loops as

shown in Figure 1. If further data generation is

chosen, the amount of data is increased step by step.

In performance enhancement the settings of the ML

model (e.g., further data preparation) or the complete

ML model itself (e.g., other algorithm classes or

algorithm) is adjusted. If the results of the ML model

correspond to the previously defined requirements,

the model can then be tested with real (historical)

data. However, it should be mentioned that this step

can only be performed if data of the real system is

available. Otherwise, this step must be performed

with the future real system data. Depending on the

result, this step leads to the execution of the

corresponding feedback loop. If these results match

the test data results in terms of accuracy, the ML

model can be used in the real system as a decision

support tool. Here, the ML model helps to decide

whether a replanning is necessary or only a few

adjustments (control variables) are required. In the

following, the key applications of the framework are

demonstrated by means of a case study.

Decision Support for Production Control based on Machine Learning by Simulation-generated Data

Figure 1: Framework for training an ML model with simulation-generated data of a production and logistics system.

Real Production or Logistics System Problem Statement ML Training

Real System Initial Situation Machine Learning (ML)

Real production or logistics

system

Planning &

control data

Real system

output

ML input

data

Programming enviroment

Result

req.

validation

optimization

YES

Validated simulation model

Simulation output

Labels

Define results requirements

ML Performance enhancement

ML Validation

ML input

data

Programming enviroment

Simulation

Data Generation by Simulation Model

Test ML

model with

real data

Simulation input

Generating

data based

on planning

& control

information

Parameter-

ization of

control

variables

Features

ML Application

Decision support tool

ML results

Data

selection

(filter)

Result

req.

Real

application

optimization

YES

Trained

model

Trained

model

Real System

Simulation

results

KPI

calculation

Compared with test data results

Target

achieved

Change

control

variables in

the real

system

YES

Simulation data

process flow

Real data

process flow

Data

Activities

Decision for

implementation

ML Data

process flow

Knowledge

transfer flow

ML feedback

loop

Problem to target

process flow

Systems

Transfer AI

model

Real system

decision

Training

ML model

Data

preparation

ML clustering

(optional)

Training and

test data

splitting

Test ML

model with

simulation

data

Identifying

problem

based on

KPIs

Defining

farget &

target KPIs

Determing

influencing

factors

Determing

control

variables

Modeling

Planning &

control data

KPI-System

Comparing

target vs.

actual KPIs

Identified from

problem statement

Identified from

real system

Further data generation

Programming enviroment

Validation

Common problem definition

KMIS 2022 - 14th International Conference on Knowledge Management and Information Systems

3.2 Case Study

The case study represents a U-cell assembly line of a

medium-sized company and is exemplarily set up in

the Technology Center for Production and Logistics

Systems (TZ PULS) of the University of Applied

Science Landshut (Blöchl and Schneider, 2016). A

simulation model was built up in Plant Simulation

based on the real system and validated against

multiple KPIs (e.g., throughput in units, cycle time

etc.). As the U-cell assembly line is used for

educational purposes, the data from the educational

production runs were used for validation. As

displayed in Figure 2 the whole value stream from

goods reception over storage to assembly and finally

goods issue is simulated. In the considered system,

floor rollers in six different variants are assembled in

seven steps. The input data are the production orders

of the assembly line. Goods reception, storage, and

goods issue are only influenced indirectly through

requests within the U-cell.

Figure 2: Representation of the real system and simulation

model in the TZ PULS.

The identified potential for optimization is that the

assembly line in question has a fluctuating throughput

per working day, resulting in a lower average

throughput than planned. Hence, the considered

target KPI is throughput in units. Since this KPI

depends on many different influencing factors, it was

necessary to narrow down the scope with regard to

the problem to be solved. The following restrictions

have been placed on this: the feasibility of the

solution should not involve any physical changes to

the material flow and should be implementable in a

short time without additional costs. Consequently, the

adjustment of the production order sequence was

identified as a changeable and monitorable control

variable. Thus, the goal was to predict the throughput

in units based on the production order sequence using

ML. The production orders as well as the throughputs

generated by the simulation model are combined to

form the input data for the ML model.

To verify the functionality of the ML model, a set

with 1,000 random production orders sequences

(numpys.random.choice

)

, are prepared. The

distribution of variants within a production order is

60 % for high-runner, 30 % for middle, and 10 % for

least demanded variants. Considering their

distribution within the production orders, the six

different variants were sequenced randomly. The

number of items per production order has been

limited to 751 units, since this number is the

maximum output quantity of the assembly line for

one working day. Each simulation run corresponds to

the processing of a production order per working day

with two shifts and sixteen hours of working time.

The output data of the simulation (=throughput in

units) is used as label.

The first approach was to predict the throughput

in units without clustering. Classic non-linear

regression algorithms (logistic regression, elastic net)

struggle to identify patterns in the data probably

because of the large number of features (751

features). A production order consists of 751 different

products that can be distinctly sequenced according to

its distribution (60 % - 30 % - 10 %). This results in

an extremely large amount of possible production

order sequences (≈10

289

). Hence, regression models

seemed to predict only floating averages. These

predictions showed a ‘mean absolute error’ of 162

units. Consequently, the approach was discarded and

following classification model was selected.

To increase the prediction accuracy the simulation

model results are first clustered into five classes.

K-Means (sklearn.cluster.KMeans) was used to

identify five different clusters. The result of the

clustering replaces the label throughput in units for

the upcoming classification task. Furthermore, the

four clusters with lower yields are grouped together.

Further data preparation contains scaling

(sklearn.preprocessing.StandardScaler) and finally a

PCA (sklearn.decomposition.PCA) step keeping

96 % of the components.

Hence, the class of the ML model is defined as

classification of the KPI throughput in units

represented by previously mentioned clusters (high

and low yield). The objective is to classify whether a

given production order sequence will produce a high

yield or not. Due to the fact that input data can be

categorized as features and output data of the

simulation model as labels, supervised learning can

be applied as ML type. Since this is a classification,

regression algorithms can be named for delimitation

of the algorithm class. For the case study a multilayer

perceptron (MLP) model (sklearn.neural_network.

MLPClassifier) is used for classification. MLP is one

Decision Support for Production Control based on Machine Learning by Simulation-generated Data

of a widely used algorithm which consists of a fully

connected input and output layer with multiple hidden

layers and is only feedforward (Goodfellow et al.,

2016). They form the basis of all ANN and are

suitable for unknown structures in the data.

Hence, the classification task has to separate

between the two classes high and low yield. High

yield should contain production order sequences with

a high throughput in units and vice versa. In Figure 3

training and test data as well as the mean (=708 units)

of the throughput in units - before applying the ML

model - are displayed.

Figure 3: 200 production orders with random sequences

with corresponding throughput in units.

The requirements for the validation of the ML

result were defined as an increase of the average

throughput in units by 1% (≈7 units) of each working

day. To split the dataset into training and test data a

random data split function (sklearn.model_selection.

train_test_split) is used. The results of the

classification are presented below.

4 RESULTS

In order to verify the approach, the predicted results

are compared with the simulated throughputs in units.

The maximum accuracy reached is 68 % with an f1-

score of 73 % for the high yield class (Table 1). In

addition, 60 % of the given production orders are

correctly classified into low yield class.

Table 1: Results of the ML model (MLP).

The precision value shows the true and false

positive rate of all positive values. As seen in Table 1

a majority of prediction is correct. Through recall the

true positive and false negative ratio is described. F1-

score is defined as the harmonic mean of the precision

and recall. Last but not least the accuracy of both

classes shows the correct prediction of the total

number of predictions for the two classes high and

low yield.

The results prove that the approach is able to

identify the high yield class with a high probability.

The classification of the low yield classes is not as

good as that of the high yield classes, i.e., it is harder

to classify production order sequences with lower

throughput in units than vice versa. By applying the

approach presented in this paper, with a 68 %

accuracy of the trained ML model, the mean value of

the throughput in units can be increased by 10 units

from a mean value of 708 units (Figure 3) to 718 units

(Figure 4). Further, the variability of the throughput

has also been reduced. This fulfilled the target of

increasing the average throughput in units by 1 %.

These results confirm the successful application of

the framework by using a simulation model to

generate input data for a ML model in this specific

case study.

Figure 4: Throughput in units of the 127 pre-validated

production orders.

5 DISCUSSON

In the following section the results are discussed. The

implementation of the framework within a case study

shows exemplarily that simulation models can be

used to generate data to train ML models. The results

of classifying the throughput in units based on

production order sequences, shows an accuracy of

68 %. If the accuracy satisfies the requirements of the

process, the ML model can be used as a decision

support tool for planning and control task in

precision recall f1-score

high yield (1) 0.69 0.78 0.73

low yield (2) 0.67 0.55 0.60

accuracy - - 0.68

KMIS 2022 - 14th International Conference on Knowledge Management and Information Systems

production and logistics systems. Furthermore, it is

possible to classify the production order sequence

faster than with a simulation model. For the case

study a complete simulation run took 20 seconds

while a single classification required 0.01 seconds.

Although these figures only apply in this specific case

study. Further research on time-savings is required.

This enables faster decision-making as compared to a

simulation model in the presented case study. Also, it

is shown that the framework applied is suitable for

extend an insufficient data basis (quantity and

quality) of processes from production and logistics

systems with additional data in order to train an ML

model. With these results and the provided

limitations, the RQ can be answered: Key elements of

the framework are a well described problem

statement based on target KPIs and control variables,

generated simulation input data based on the

identified control variables, a validated simulation

model for data generation as well as suitable data

preparation step for an appropriate ML model.

The framework can also be applied to other

control processes within production and logistics

systems. Nevertheless, there are still some

limitations. First of all, it should be mentioned that,

there is still room for improvement regarding the ML

model. The determination of suitable AI models with

regard to this specific problem of production order

sequencing has already been studied by (Rissmann et

al., 2022). It can be stated that the application of more

specific ML models, such as deep neural networks,

could provide even better results. Further

investigation is expected to demonstrate how the

application works on other random problems (e.g.,

failures, downtimes etc.) within production and

logistics systems. Furthermore, only the classification

of throughput in units was tested. For other KPIs,

such as the prediction of the production time of

individual units or lead time, the simulation-based

data may have to be enriched.

6 SUMMARY AND OUTLOOK

In this paper, we present a framework that supports

the implementation and training of ML models based

on generated datasets from production and logistics

simulations. To achieve this, the input and output data

of a simulation model are used for training. Thus, ML

models can be developed even in processes with

limited data or insufficient data quality, which can

then be used for decision support. By applying the

approach within an exemplary case study, the ML

model was able to increase the average throughput.

In future research activities, the existing

simulation model is to be supplemented by further

influencing factors such as downtimes and failures.

This will allow the simulation model to reflect a real

production and logistics system even more

accurately. The next research steps will be the

implementation of a data-oriented problem

identification and optimization approach based on

KPIs as well as another verification of the approach

in a real production and logistics system.

FUNDING

This research was supported by KIProLog project

funded by the Bavarian State Ministry of Science and

Art (FKZ: H.2-F1116.LN33/3).

REFERENCES

Agalianos, K., Ponis, S. T., Aretoulaki, E., Plakas, G., and

Efthymiou, O. (2020). Discrete Event Simulation and

Digital Twins: Review and Challenges for Logistics. In:

Procedia Manufacturing 51, pp. 1636–1641.

Bárkányi, Á., Chován, T., Németh, S., and Abonyi, J.

(2021). Modelling for Digital Twins—Potential Role of

Surrogate Models. In: Processes 9 (3), pp. 476.

Blöchl, S. J., and Schneider, M. (2016). Simulation Game

for Intelligent Production Logistics – The PuLL®

Learning Factory. In: Procedia CIRP 54, pp. 130–135.

Fowler, J. W., and Rose, O. (2004). Grand Challenges in

Modeling and Simulation of Complex Manufacturing

Systems. In: SIMULATION 80 (9), pp. 469–476.

Goodfellow, I., Bengio, Y., and Courville, A. (2016). Deep

learning, The MIT Press, Cambridge, Massachusetts,

London, England.

Han, J., Kamber, M., and Pei, J. (2012). Data Mining

Concepts and Techniques, Elsevier.

Hasan, M. S., Ebrahim, Z., Wan Mahmood, W. H., and Ab

Rahman, M. N. (2016). Decision support system

classification and its application in manufacturing

sectore: a review. In: Jurnal Teknologi 79 (1).

Kauke, D., Galka, S., and Fottner, J. (2021). Digital Twins

in Order Picking Systems for Operational Decision

Support. In: Hawaii International Conference on

System Sciences, pp. 1655–1664.

Koot, M., Mes, M. R., and Iacob, M. E. (2021). A

systematic literature review of supply chain decision

making supported by the Internet of Things and Big

Data Analytics. In: Computers & Industrial

Engineering 154, pp. 107076.

Korth, B., Schwede, C., and Zajac, M. (2018). Simulation-

ready digital twin for realtime management of logistics

systems. In: 2018 IEEE International Conference on

Big Data, pp. 4194–4201.

Decision Support for Production Control based on Machine Learning by Simulation-generated Data

Kritzinger, W., Karner, M., Traar, G., Henjes, J., and Sihn,

W. (2018). Digital Twin in manufacturing: A

categorical literature review and classification. In:

IFAC-PapersOnLine 51 (11), pp. 1016–1022.

Müller, M., Mielke, J., Pavlovskyi, Y., Pape, A., Masik, S.,

Reggelin, T., and Häberer, S. (2021). Real-time

combination of material flow simulation, digital twins

of manufacturing cells, an AGV and a mixed-reality

application. In: Procedia CIRP 104, pp. 1607–1612.

Nagahara, S., Sprock, T. A., and Helu, M. M. (2019).

Toward data-driven production simulation modeling:

dispatching rule identification by machine learning

techniques. In: Procedia CIRP 81, pp. 222–227.

Pfeiffer, A., Gyulai, D., Kádár, B., and Monostori, L.

(2016). Manufacturing Lead Time Estimation with the

Combination of Simulation and Statistical Learning

Methods. In: Procedia CIRP 41, pp. 75–80.

Qi, Q., and Tao, F. (2018). Digital Twin and Big Data

Towards Smart Manufacturing and Industry 4.0: 360

Degree Comparison. In: IEEE Access 6, pp. 3585–3593.

Rissmann, L., Muehlbauer, K., and Meissner, S. (2022).

Application of Deep Learning based on a Simulation

Model to classify Production Orders. In: Proceedings

of the 34rd European Modeling & Simulation

Symposium, (accepted).

Rose, O. (2007). Improved simple simulation models for

semiconductor wafer factories. In: 2007 Winter

Simulation Conference. Washington, DC, USA, IEEE,

pp. 1708–1712.

Verein Deutscher Ingenieure (2018). Simulation von

Logistik-, Materialfluss- und Produktionssystemen, In:

Verein Deutscher Ingenieure (Ed.), 2018 (3633).

Vernickel, K., Brunner, L., Hoellthaler, G., Sansivieri, G.,

Härdtlein, C., Trauner, L. et al. (2020). Machine-

Learning-Based Approach for Parameterizing Material

Flow Simulation Models. In: Procedia CIRP 93, pp.

407–412.

Vojdani, N., and Erichsen, B. (2018). Ein Ansatz für ein

Predictive-Monitoring-System zur Identifikation von

Störungswirkungen in der Produktionslogistik mittels

künstlichen neuronalen Netzen. In: Logistics journal.

Windt, K., Philipp, T., and Böse, F. (2008). Complexity

cube for the characterization of complex production

systems. In: International Journal of Computer

Integrated Manufacturing 21 (2), pp. 195–200.

KMIS 2022 - 14th International Conference on Knowledge Management and Information Systems