Conference PaperPDF Available

Towards Digital Twins for Optimizing the Factory of the Future


Abstract and Figures

Logistics are essential regarding the efficiency of factories, and therefore their optimization increases productivity. This paper presents an approach and an initial implementation for optimizing a fleet of automated transport vehicles, which transports products between machines in the factory of the future. The approach exploits a digital twin derived from a model of the factory representing the artifacts and information flow required to build a valid digital twin. It can be executed faster than real-time in order to assess different configurations, before the best-fitting choice is applied to the real factory. The paper also gives an outlook on how the digital twin will be extended in order to use it for additional optimization aspects and to improve resilience of the transport fleet against anomalies.
Content may be subject to copyright.
Patrick Eschemann, Philipp Borchers,
Linda Feeken and Ingo Stierand
OFFIS - Institute for Computer Science
Escherweg 2
26121 Oldenburg, Germany
E-mail: {patrick.eschemann, philipp.borchers,
linda.feeken, ingo.stierand}
Jan Stefan Zernickel and Martin Neumann
InSystems Automation
egeny-Straße 16
12489 Berlin, Germany
E-mail: {janstefan.zernickel,
Computer Integrated Manufacturing (CIM), Production, Trans-
portation, Optimization, Decision support systems.
Logistics are essential regarding the efficiency of factories,
and therefore their optimization increases productivity. This
paper presents an approach and an initial implementation
for optimizing a fleet of automated transport vehicles, which
transports products between machines in the factory of the
future. The approach exploits a digital twin derived from a
model of the factory representing the artifacts and information
flow required to build a valid digital twin. It can be executed
faster than real-time in order to assess different configurations,
before the best-fitting choice is applied to the real factory.
The paper also gives an outlook on how the digital twin
will be extended in order to use it for additional optimization
aspects and to improve resilience of the transport fleet against
The factory of the future (FoF) does not consist of individual
machines or plants, but of a network of machines, plants
and factories, constituting a system of systems (SoS). The
interconnection of machines can be used as infrastructure for
enabling machines to manufacture products intelligently, to
request required products, and to communicate their status
in order to troubleshoot issues. Manufacturing cycles become
more flexible, and enable creating various different products
down to production size one.
The operation of a FoF calls for appropriate approaches to op-
timize manufacturing cycles, and to maintain resilience against
production issues. This paper focuses on the transportation
of products. Suppose a fleet of automated guided vehicles
(AGV) delivering products between machines, and to and
from warehouses. The machines are able to issue delivery
requests to the manufacturing execution system (MES) if they
require input products, and pickup requests for completed
This work received funding from the BMBF through the ITEA CyberFac-
tory#1 project under grant agreement No 01IS18061C, and through the BMBF
CrESt project under grant agreement No 01IS16043L.
output products. The MES searches for a matching source (or
target) of the requested product, and sends a corresponding
transport request to the AGV fleet. The AGV fleet in turn
schedules the incoming transport requests among the AGVs,
and processes them accordingly.
FoF optimization and resilience is considered in this setting
as follows. Concerning optimization, the AGV fleet should
schedule and process transport requests in a way that mini-
mizes downtimes of machines due to missing input products.
This includes situations with frequently changing requirements
due to flexible manufacturing cycles, and other production
dynamics such as regular supply deliveries. Concerning re-
silience, the fleet should distribute transport requests equally
among the AGVs in order to disperse wear and tear equally,
so that maintenance operations can be better planned and
performed. This paper does not address resilience against
anomalies such as AGV outages and road blocks preventing
or delaying transport tasks. This is considered as future work.
Optimizing production flows and improving resilience in an
FoF raises the challenge of how it can be achieved. It re-
quires testing different strategies and configurations in order
to identify the best fitting result. In our setting, it means to
parameterize the AGV fleet strategy to optimize product flows,
and to apply counter-measures in the fleet behavior in case of
detected anomalies to re-establish fluid transportation service.
It is however often not desired or even impossible to perform
these test operations online in the actual factory.
Instead we exploit a digital twin (DT) of the FoF. The DT is
fed with the actual configuration of the FoF and test parameters
of investigated components, in our case the AGV fleet. The
DT is designed as a faster than real-time simulation, and thus
can be used to test a large number of different configurations.
The best fitting result is then played into the real AGVs.
Such approach indeed requires the DT to be valid. This means
that the behavior observed with the DT is comparable to the
behavior of the actual FoF.
The contribution of the paper is an initial implementation of
the above approach. After placing the approach into the state
of the art, we present in Section FACTORY MODEL a factory
model containing all artifacts and information flows of the real
FoF necessary to derive a valid DT. Section SIMULATION
discusses a real-time simulation that serves as a substitute for
the real factory, as this one is not at hand. The simulation can
be instantiated according to a given configuration of the model.
In Section DIGITAL TWIN we develop a simple DT, which
can also be configured according to the developed model,
and executes faster than real-time. The DT is evaluated in
Section EVALUATION in two steps. First, the DT is applied
to an exemplary use case to solve a simple optimization task
by assessing different scheduling strategies for the AGV fleet
for a given factory configuration. It is then shown by the use
case how the DT could be validated in order to ensure that
the results obtained by the DT correspond to the real-factory
(simulation). To this end, a realistic similarity metric is defined
allowing to compare behavior. In Section DISCUSSION AND
FUTURE WORK we give an outlook on how the developed DT
will be extended to support additional optimization problems,
and to improve resilience against anomalies. The last section
concludes the paper.
The term “Factory of the Future” is used for decades and
reinvents itself reconsidering the challenges and opportunities
in context of the so called fourth industrial revolution, which is
a digitalization revolution. Whereas highly automated factories
where solely a vision of the FoF in 1987, the FoF nowa-
days transforms from the digital factory to the smart factory
(Meredith 1987), (Salierno et al. 2019). This emphasizes that
the FoF is about utilizing computer based developments to
alter manufacturing processes with the goal to adapt quickly,
spare resources and deliver a high sustainability. A smart
factory adds decision-making capabilities to the shop floor,
typically by collecting data from it to build realistic models
for accurate simulation results (Salierno et al. 2019).
Nieto et al. conducted a wide research regarding different
approaches towards the FoF. They identified three research
streams (Nieto et al. 2017): (1) Computer Aided Manufactur-
ing (CIM), (2) the role of human workforce, and (3) other,
which includes
continuous improvement, e.g., lean and total quality man-
holistic factory, e.g., cyber physical systems and cloud
processes for sustainable production, e.g. zero-waste,
decision support system.
In this paper we focus on the last point and present an approach
to improve decision support. For this purpose knowledge of
the current state and how processes in the factory develops
are quite important. Problems with the current setup can be
detected before occurring in the reality. This requirement can
be caught by a concept called DT, originated 2002 at the
University of Michigan (Grieves 2014).
With increasing digitalization of the FoF the DT concept gets
intensified research focus regarding various applications in the
production domain. A DT can be defined as “a formal digital
representation of a real physical system that captures attributes
and behaviors” (Malakuti et al. 2020). It “is an integrated
multi-physics, multi-scale, and probabilistic simulation of a
complex product and uses the best available physical models,
sensor updates, etc., to mirror the life of its corresponding
twin” (Tao et al. 2018).
There are two main aspects to motivate the DT. First, all the
applications that arises with the FoF require and produce a lot
of information about the assets in factories. The data, coming
from different parties, is probably duplicated, inconsistent or
incomplete (Malakuti et al. 2020). The DT can collect the
data centrally to provide it for different applications. Therefore
the DT must include a data-model abstracting the physical
dimension to keep meaning and context of data (Lu et al.
2019). Otherwise the collected data just forms a incoherent
data lake, which is not sufficient for a high quality data
Another important motivation for the DT is the simulation
aspect. Boschert and Rosen set the DT as the next generation
step in the trend of industrial simulation (Boschert and Rosen
2016). Because of accurate models fitted with actual and
history data as well as artificial intelligent techniques, the DT
is rather a digital and independent imitation of corresponding
assets (Rosen et al. 2015). Also operational dynamics are
representable, which enhances the DT to more than just a
simulation (Lu et al. 2019).
With those aspects the DT supports the overall production
process and life cycle of a product. Designers can test and
validate products even without a prototype fabrication. Im-
provements of economic and financial product plans can be
checked, and consequences of decisions regarding product
features and configuration options can be predicted (Tao et al.
2018). The DT can also anticipate consequences of production
steps, which helps to decide between alternative strategies
e.g. for production scheduling. Furthermore the DT can detect
whether misbehavior of machines or malfunction of products
arise in the future, and reveals the effect of countermeasures
(Rosen et al. 2015).
The development of DTs require a data model to interpret
information as well as a behavioral model for accurate sim-
ulation. DTs for the FoF distinguish DTs for product design
and DTs for the manufacturing process, for both exist certain
industrial standards that come up with different models (Lu
et al. 2019). As stated in B´
ecue et al., the FoF DT comes
up with the challenge to handle the amount of models of the
underlying subjects and simulate them coherently (B´
ecue et al.
A probable solution is the co-simulation technique, which
can be defined as “the coordinated execution of two or more
models in their run-time environments” (Steinbrink 2016). It
enables coupling of two simulation environments via interfaces
(Alzaga et al. 2018) and the life mirroring of corresponding
twins in a complex system (Negri et al. 2019). Scheifele et al.
presented a real-time co-simulation for virtual commissioning
(Scheifele et al. 2019). They managed to carry out commis-
sioning beforehand by using a DT. The standardized functional
mock-up interface (FMI) for co-simulation is used in this
approach for high-resolution mapping of specific effects. In
the FMI technique different simulators are wrapped up in
separated functional mock-up units (FMUs) resulting in equal
interfaces which are connected to a global simulation.
Another approach is presented by Korth et al. (Korth et al.
2018). They present a DT architecture based on a discrete
event time model to simulate the future material flow. The
DT is split into a dynamic model (events of past, present and
future) and a static model (process structure, resources, orders,
business objects). Therefore they are able to run a simulation
based on the static model an exchangeable dynamic model.
The decision support uses the simulation to predict distur-
bances and to test counter measures like real-time resource
allocation or order shifting.
To conclude, none of the examined works considers dif-
ferent job distribution strategies of an AGV-based logistic
system. The approach presented in this paper employs a DT
to optimize the flow of goods and therefore counts to the
decision support approaches. Our DT is based on discrete
event simulation and does not require co-simulation.
As a DT should reflect all aspects of the real factory that
are relevant for the optimization task, we developed a corre-
sponding factory model that captures those aspects. The model
thus serves as an intermediate representation of the factory for
the DT. It can be instantiated according to the current factory
configuration, and then be fed to the DT. Since there is no
real factory at hand whose data output can be transferred into
the DT, a simulation is used to generate reasonable data.
The model focuses on two components, the manufacturing
execution system (MES) and the Transport System (TS). It
also encompasses typical elements that occur in factories like
warehouses, machines, and their dependencies. This part is
explained in Section Modeling the Factory from the MES Point
of View. Additional parameters capture production dynamics
close to the real behavior, see Section Modeling Dynamics of
the Factory. The TS consists of a fleet of AGVs that takes
care of the logistic processes in the factory. The Section The
Transport System gives a detailed view.
Modeling the Factory from the MES Point of View
The factory model consists of stocks and machines, which are
usually found in most factories and are used to resemble the
factorial layout. A stock can either be a source for goods or a
sink of goods. Machines are modeled as black boxes, meaning
the actual manufacturing process is not further detailed.
The model also contains buffers and recipes to capture pro-
duction flows between machines (and stocks). Buffers are used
to store goods at the input and output of a machine. Recipes
define required quantities to start a manufacturing process, and
what is created.
Figure 1 shows an example of a factory layout and on how the
elements depend to each other. The example models a simple
magazine production plant. It consists of two machines, a
Printer and a Binder, and a warehouse with four stocks. Three
of them are (infinite) sources of goods. Source q1 provides
paper, q2 ink, and q3 glue. One stock (s1) receives completed
magazines, and thus is a sink.
Figure 1: Factory Model Example
The ware flow is modeled by unidirectional lines that always
connect output buffers to input buffers. They also represent
transport routes of the AGVs. Each line is annotated with a
number representing the amount of goods transported with
each transport job. For example, the annotation of the line
between q1 and Printer means that an AGV transports 5000
sheets of paper each time.
Production recipes result from the numbers annotated to
buffers. For example, the Printer requires 1.000 sheets of
paper and 25 ink units for a production cycle. In each cycle,
the Printer produces also 1.000 sheets of paper, now being
printed. The manufactured items are buffered at the output.
The Binder takes 2.500 pages and creates 50 magazines, 50
pages each. It needs 15 glue packages for the binding process.
Source and sink buffers are also following the recipe principle.
They however “produce” and “consume” the defined amount
of products out of nothing and into nothing, respectively.
A machine starts a manufacturing cycle when the required
filling level of the input buffers are reached, and the output
buffers have the capacity to receive the produced goods.
The maximum buffer sizes are not shown in Figure 1, but
modeled together with a threshold parameter. A machine
sends one delivery request to the MES if the fill level of
an input buffer is below this threshold. Correspondingly, it
sends a pickup request if an output buffer reaches the defined
threshold. Also the time needed for a production cycle can be
configured in the model.
The upper part of Figure 1 shows the MES, which is connected
to the TS. All machines are connected to the MES and report
their input and output filling level as explained above. With
these information the MES matches full output buffer with
empty input buffer and creates corresponding transport jobs,
which are forwarded to the TS component. For example, a
transport job will be created when the MES receives a pickup
request from the Printer and a delivery request from the
Binder. The strategy by which the MES matches requests and
creates transport jobs depends on the implemented strategy,
and is not in the focus of this paper.
Note that the number of machines, stocks, buffers and recipes
can be configured as required. In this manner a wide variety
of layouts can be created.
Modeling Dynamics of the Factory
To this point the model reflects a static 24/7 production
process. Since this does not apply for real world factories,
the model allows changes of the duration of manufacturing
processes in predefined boundaries. Those changes are not
allowed to be purely random, since we expect the dynamic
behavior of a real factory to follow certain patterns as well.
Instead, the model defines a configurable time scheme in terms
of working days and working hours. Hereby different shift
models become testable. A so-called efficiency factor in the
range [0,1]for manufacturing operations can be defined for
each working hour of the week. This can either be done
globally for all processes within the factory, or individually
for each machine.
The machine-related recipe introduced in the previous section
defines a minimum and maximum duration for the manufactur-
ing process. With an efficiency of 1, the manufacturing process
requires the minimum time, and with an efficiency of 0 the
maximum duration. Values between these boundaries result in
linear interpolated durations. A possible application could be
a factory with a workstation operated by humans, where the
manufacturing duration at the human station varies over the
working hours.
The Transport System
As shown in Figure 1, the TS gets transport requests from the
MES, and is responsible for processing them by scheduling
and performing corresponding transport jobs. As for the MES,
the TS component is implementation-dependent. In our case, it
consists of a fleet of a configurable number of AGVs. Trans-
port requests are published by the MES to the whole fleet.
Each AGV implements a bidding protocol over a common
communication medium, allowing the fleet to negotiate which
AGV is taking responsibility for the request. To this end, each
AGV calculates a bid value for the request, and broadcasts
it to the fleet. The AGV with the lowest bid value gets the
transport job. Being implementation-dependent, configuration
of the bidding is detailed in the following section.
As stated before, there was no access to a real factory, nor to
sufficiently detailed production data. Hence, we use a factory
simulation as a substitute. The starting point for this was
a simulation of the AGV fleet implemented by InSystems
Automation GmbH. It has been designed to be as realistic
as possible, and also provides for hardware in the loop
simulation scenarios with mixed simulated and real AGVs.
It thus includes various functionalities, such as the simulation
of battery consumption, path finding and collision detection,
and also allows to configure the bidding strategy.
The TS component is embedded in a modular simulation
system consisting of individual (docker) containers, which
enables implementing and integrating various extensions into
the simulation. In order to resemble a whole factory, a MES
component has been implemented in accordance with the
model discussed in Section Modeling the Factory from the
MES Point of View. The simulation system is configured by
a simple factorial layout that defines the existing elements
such as the machines and buffers, together with their location.
Also machine recipes, buffer sizes and initial fill levels are
configurable. Finally, working days and hours together with
machine efficiencies can be configured in order to implement
the production dynamics (cf. Section Modeling Dynamics of
the Factory).
The MES implementation offers the possibility to prioritize
several possible matches of pickup and delivery requests ac-
cording to four parameters (random, distance, capacity, time).
For example, if the parameters distance and time were both set
to 50 percent, the MES would prioritize the requests with the
shortest distance and those that were received at the earliest
point in time equally. For the purpose of this paper we have
set the random parameter to 100 percent and the others to
0 percent. This results in an equal distribution among the
possible matches. We have chosen this configuration in order
to avoid the generation of always identical transport jobs. The
longer the simulation time, the greater the probability that all
possible transfer orders will be executed.
The TS component is configured by the number of AGVs in
the fleet, and their bidding strategy. The calculation of the bid
values is a key element for the performance and resilience of
the TS component. Depending on the chosen strategy it may
involve various parameters, such as the estimated completion
time, battery status, and the distances already driven. The
strategy, i.e., the selection and weighting of the bid parameters,
largely determines the performance of the AGV fleet. On
one hand, the fleet shall perform transportation service in
a way that minimizes machine down-times due to missing
products and exhausted output buffers. On the other hand, the
AGVs shall be kept in moderate operational state in order
to minimize maintenance effort. This includes preservation of
battery lifetime and uniform wear and tear among the fleet.
In this paper, two parameters have been implemented for both
the simulation and the DT, as detailed in Section DIGITAL
Unfortunately, the accurate AGV fleet simulation requires
large computational power in order to achieve acceptable
runtime. Hence, it has been decided for the time being to
replace the accurate AGV simulation by a simplified version,
which concentrates on commissioning strategies. The main
abstraction is that AGV movements have been abstracted by
randomized velocities as arbitrary outcome along a uniform
distribution within configurable bounds, and fixed distances
between machines. This excludes for example evasive maneu-
ver as well as adjustments on buffer locations.
In order to minimize the effect of such simplification, the
selected example has a factory layout where AGVs can move
freely, with a minimal danger of necessary avoidance maneu-
ver that would result in larger deviations of transport comple-
tions. This indeed affects the validation of the DT discussed in
the Section EVALUATION. Removing this simplification hence
constitutes a high-priority item on the future work list.
Our main objective for using a DT is to provide decision sup-
port for the fleet of AGVs for optimizing the transport flow in
the factory. In particular, the time for transport task completion
and the difference in the wear and tear of different AGVs
shall both be minimized. The proposed approach is sketched in
Figure 2: Important properties of the real factory are captured
in a factory model (cf. Section FACTORY MODEL). Important
data like the factory map and the number of transport robots
is fed to the DT for enabling the instantiation of the DT. Since
we do not have access to a real factory for playing around, we
additionally set up a factory simulation as substitution of the
real factory (cf. Section SIMULATION). For running the DT,
the transport job list produced by the simulation is used. The
DT is then used in a faster than real-time mode for predicting
the fleet performance for the given transport tasks for different
behavioral alternatives of the fleet. The resulting KPIs on
fleet behavior can be used for picking the alternative with the
best performance, which is then given to the real robots as
proposed new configuration. Running the factory simulation
with the same fleet configuration as in the DT, fleet KPIs of
the simulated reality and the DT are compared for checking
validity of the DT.
Figure 2: Digital Twin Usage and Approach
In this paper, we focus on the distribution of transport tasks
over the fleet, i.e. on the choice of the bidding algorithm of the
AGVs, and the consequence on the two optimization goals.
For making the optimization goals measurable, the following
key performance indicators (KPIs) are chosen and tracked in
the DT:
1) Average time and variance needed for completing a
transport task.
2) Average number of fulfilled transport tasks per hour
(efficiency of the fleet).
3) Maximal difference in the driven distances of the AGVs
when all transport tasks are completed.
4) Maximal difference in the number of completed tasks
over all AGVs when all transport tasks are completed.
We implemented an initial version of a DT for the fleet of
AGVs with the free software GNU Octave. For instantiating
the DT, some information on the factory and on properties
of the fleet of AGVs are needed. In most factory environ-
ments, access to data is limited for protecting confidential
information. For example, we cannot expect that a logistic
system like a fleet of AGVs that is provided by third-party
gets access to machine recipes or the desired number of final
goods. Consequently, we can only use data in the DT that
is also accessible for the AGVs in real applications. Hence,
we limited the required information to data that are directly
accessible for the AGVs, or that can be derived from known
data and from observations from previous AGV runs. The
following information are taken by the DT as input:
1) Identifiers of machine buffers, charging stations and
2) Distances between buffers.
3) Number of AGVs in the fleet.
4) Minimal and maximal velocities for each AGV.
The information is taken from the simulation model described
in Section FACTORY MODEL. Existing buffers are extracted
from a factory map that each robot has access to, and dis-
tances between buffers are calculated automatically using the
Manhattan distance as metric. The number of AGVs is also
set based on the chosen setting in the simulation. For the
velocities, the same distribution as chosen in the simulation
is used. In addition to the above information, the DT gets
an initial situation (location of each AGV and already driven
distances), and a list of transport tasks (including information
on when the task is sent to the fleet, and between which buffers
or warehouses the transportation needs to be done).
In order to obtain the list of transport tasks, the simulation logs
all transport requests to the TS in a list, which can directly be
fed to the DT.
Last but not least, the DT requires information on which
configuration (or: strategy) of the AGVs should be used.
Currently implemented is a family of the wear and tear
strategies: Each AGV calculates a bid for a newly requested
transport task based on its already driven distance and on the
number of not yet finalized transport tasks, i.e. the length of the
local queue of the AGV. To this end, the AGV is calculating
the sum of these two values, weighted according to the selected
strategy parameters. The AGV with the lowest bid is taking
over the task. The importance of the two parameters driven
distance and number of unfinished tasks for the bid calculation
can be varied by using different weights for the parameters.
All AGVs are using the same weights. The lower the weight
for one of the parameters, the less influence has the parameter
on the task distribution. For example, a weight of zero for the
driven distance results in assigning a new task to the robot
with the lowest number of unfinished tasks, independent on
the traveled distance.
Taking the input data, the DT can run faster than real-time for
calculating the KPIs listed above. Because of build-in non-
determinism in the velocity of each AGV, we let it run several
times in order to get statistically relevant data. In each run, the
velocity of the robots are chosen non-deterministically within
the ranges given for the DT instantiation, possibly resulting in
different task distributions over the fleet.
As output, the DT delivers in each run detailed information
on each transport task: (i) an identifier of the AGV that took
over the task, (ii) point in time in which the AGV started the
task, (iii) point in time in which the AGV reached the source
of the task, (iv) task completion time, and (v) the distance the
AGV has been driven in order to fulfil the task. Additionally,
information on the distance evolution of each AGV is given.
For each point in time a new transport task is received by the
fleet of AGVs, it is recorded how much distance each AGV
has driven at this time.
In order to obtain the KPIs for a particular configuration, they
are first calculated for each individual simulation run from the
recorded data. The results are then averaged over all simulation
We applied the DT for simulating the AGV performance in a
factory over 24 hours with different strategies for the AGVs.
The used factory topology is the same as shown in Figure
1: The factory floor consists of two machines (printer and
binder) and a warehouse. The Manhattan distance between the
buffers is between 5m and 63m. Transport tasks are generated
according to the machine model in Section Modeling the
Factory from the MES Point of View, which leads in the chosen
machine and MES configuration to 225 transport tasks in 24
hours, or approximately one request every 6.4 minutes. This
list of tasks is given as input to the DT with three AGVs. The
minimal and maximal velocity of AGVs is set to 0.2m
s, and
srespectively. Initially, the AGVs are set to some waiting
positions (not shown in Figure 1), have an empty local queue,
and no driven distance.
Three different strategies were applied for distributing the
transport tasks: With Strategy S1, the AGVs distribute the tasks
solely based on the number of unfinished tasks per AGV, i.e.
the AGV with the smallest number of unfinished tasks at the
generation time of a new task is taking over the new task. With
Strategy S2, the AGV with the least driven distance until the
generation time of a new task is taking over the task. The last
chosen strategy, S3, is a mix of the other two strategies: the
AGVs decide on task distribution based on covered distance
and the number of unfinished tasks per AGV. For all three
strategies holds: are two or more AGVs equally equipped to
handle the next task (e.g. when the number of unfinished tasks
per AGV is the same), then the AGV with the smaller ID is
taking over the task.
The results of applying the three strategies in the DT in terms
of KPIs are summarized in Table 1. For each strategy, the KPIs
are averaged over 20 runs of the DT, each run varying in the
velocities of the AGVs, leading to potential variations in task
distribution. The KPIs show that all strategies lead to equal
efficiency of the fleet of AGVs (around 11 completed tasks
per hour). This is no surprising result as mean time for task
completion is rather low. Strategy S1 has by far the largest
differences in the covered distance, diff(D) (in meters), and
in the number of completed tasks, diff(C), per AGV. This
is because the driven distance is not used to calculate bid
values with this strategy. As the transportation times are rather
low, the effect of the ut parameter is somehow restricted,
leading to an unbalanced distribution of tasks among the
AGVs. Conversely, strategy S2 leads to the highest average
completion time for tasks T T C (in seconds), with variance
Var(T TC ) more than twice as high as for S1 and S3. S3
results in the lowest average time for completing tasks, lowest
variance, lowest difference in the covered distance and in the
number of completed tasks per AGV.
Table 1: Strategies and Resulting KPIs in the DT
Strategy T TC Var(T TC ) efficiency diff(D) diff(C)
in sin s2in tasks
hin min tasks
S1dd =0 202 23480 11 7507 125
ut =1
S2dd =1 227 57371 11 54 6
ut =0
S3dd =1 194 22843 11 41 3
ut =1
(dd = weight of driven distances parameter,
ut = weight of unfinished tasks parameter)
Interpreting the KPIs, we conclude that strategy S3 performs
the best in the chosen setting, and hence this strategy should
be fed back from DT to its physical counterpart for real
application. The fact that both T TC and Var(T T C) are slightly
lower in S3 than in S1 is considered as an effect of the
randomization involved in the DT.
Additionally, the KPIs can be used for the identification of
potential improvements of all strategies: The large difference
in the number of completed tasks in S1, assigning tasks to the
AGV with the shortest local queue, indicates that all AGVs
are equally occupied most of the time, such that the AGV
with the lowest ID is taking over much more tasks than
the other AGVs. A more in-depth analysis of the DT runs
supports this hypothesis. Modification of the AGV selection
in cases of equal bids of two or more AGVs to e.g. random
task assignment for those AGVs could improve the AGV
Feeding the result of an evaluation of operational parameters
by a DT back to the real factory requires a certain trust that
those parameters will indeed have the predicted impact, i.e.,
that the DT is valid. Concerning validity, one should keep
in mind that all data for the DT have been obtained from a
simulation. Hence, we can in fact assess validity of the DT
only with respect to the simulation. Although the validity of
the simulation with respect to the real factory is out of scope
of the paper, an as accurate as possible simulation is desired.
However, the validation procedure for a real factory would be
A usual way to assess the validity of the DT is to define a set
of KPIs, and to compare a sufficient large set of data obtained
from the simulation and from the DT. The KPIs used above (cf.
Table 1) may serve as a starting point. As this paper represents
an early development stage, there is not yet much simulation
data available. The simulation speed-up highly depends on the
used simulation hardware.
The experiment from the previous section represents roughly
24 hours of simulated time. The mean TTC in this simulation
run is 241s, and the variance is 27893. The driven distances
and which AGV has completed the task have not yet been
logged, and thus are not accessible in order to calculate the
other KPIs.
As an initial observation one could state that the value obtained
in the simulation and those predicted by the DT are at least in
a similar range. Secondly, one could argue that the simulated
AGVs need more time on average to complete transport tasks
than predicted by the DT. On the other hand, this might also
be a statistical artifact. Moreover, at least the mean TTC of
the different strategies predicted by the DT are quite similar,
indicating that this KPI alone is not the most suited one for
arguing about similarity.
This is actually a unsatisfactory situation concerning the
validity of the DT. More data is required from simulation runs,
and particularly with different strategies applied to the AGVs.
This is, as stated, work in progress.
The paper presents only a first step towards an industrially
applicable approach for optimizing transportation service of
products by a fleet of AGVs. Further improvements and
extensions are required to this end. Our road map contains
the following tasks.
Simulation improvements: As the simulation provides our
“ground truth”, it should be as realistic as possible. Currently, a
number of simplifications and restrictions have been made. For
example, only a limited number of strategies of the MES for
determining transport requests have been implemented. Also
the behavior of the AGV fleet needs to be improved, such
as the inclusion of effects of the battery status like charging
the AGVs, and pickup and delivery latencies at the buffers.
Another line of work considers the availability of data. Many
important information is currently not accessible, such as some
AGV status information. This would be required to assess
important effects, like the actual transport job queues of the
individual AGVs.
Improving the DT: The current implementation is intended
as proof of concept. We plan to improve the DT in two
dimensions: The first dimension is on improving the digital
replicas of the AGVs. In the current DT implementation,
battery consumption of robots and the need for recharging
is not yet reflected. Since a good fleet performance strongly
depends on the robots being operational with charged batteries,
we plan to equip each digital AGV with a battery model. This
will enable us to test not only alternatives for task distribution,
but also alternatives for charging times. Additionally, we want
to weaken the assumption that an AGV is always driving
the same distance between two locations in the factory by
adding some non-determinism in the driven distances between
The second dimension tackles the interaction of the AGVs with
the environment, and the synchronization with the physical
counterpart. For example, we plan to add dynamics in the
factory maps of the (digital) AGVs: If the real AGV fleet
learns detected blocked paths in the factory, the DT should be
able to adapt to the new situation in order to still reflect the
behavior of the real AGVs.
As discussed in Section FACTORY MODEL, the factory model
is an intermediate model to get production process informa-
tion for the DT. Considering that the robots have limited
information about the production process, further development
of the model is a important approach to improve a DT.
Moreover, with a model-integrated DT, robots could interpret
their sparsely available data in the model context, thus using
their data more efficiently and in turn improve the prediction
quality of production dynamics.
Extension of use cases: Beyond these improvements, further
work envisions the application of the approach to additional
use cases.
(i) Anticipation of production dynamics: The DT is currently
being fed with a list of transport tasks from the factory
simulation. Hence, the DT can currently only be used for
analyzing how each AGV configuration would have performed
for the already completed tasks. We plan to add a transport
task prediction component that can feed the DT with upcoming
transport tasks instead of previous tasks. This component will
receive extensive information on recorded transport tasks from
historic data. With such a prediction component, the DT can
be used for anticipating future fleet performance for different
(ii) Anomaly detection: The approach can be applied for
improving the resilience of the FoF by a component that
detects anomalies in the product flows. Given that certain
patterns can be identified with a particular configuration of
the factory, deviations from this pattern and possible causes
would be identified. These findings could be again fed into
the DT in order to assess counter-measures. The development
of such a component is part of our research agenda.
(iii) Component for automated decisions for configurations: A
combination of the current DT implementation and anticipated
production dynamics allows to test the effect of different
strategies of the AGVs in advance. As a next step, we want
to investigate how this can be used for automated choosing of
a strategy for the AGVs that promises good fleet behavior
in terms of KPIs. Such a component could be triggered
for example by detected anomalies, or by changes in the
production dynamics, for considering to change the strategy.
The chosen strategy is then automatically fed to the fleet and
triggers a respective reconfiguration of the AGVs.
This paper presents an initial implementation of the approach
to exploit DTs for optimizing manufacturing and improving
resilience against production flaws in the FoF. Focusing on the
transport service of a fleet of autonomous transport vehicles,
a DT is designed to achieve faster than real-time simulation,
thus enabling testing a large number of different factory and
AGV fleet configurations, before the best fitting result then is
applied to the real fleet.
The paper also gives an outlook on upcoming improvements
and extensions of this initial work, which will largely increase
the applicability of the approach for further kinds of optimiza-
tions and resilience improvements.
Alzaga, Aitor et al. 2018. “Machine Tool: From the Digital
Twin to the Cyber-Physical Systems”. In: Twin-Control.
ecue, Adrien et al. 2020. “A New Concept of Digital Twin
Supporting Optimization and Resilience of Factories of the
Future”. In: Applied Sciences 10.13, p. 4482.
Boschert, Stefan and Roland Rosen. 2016. “Digital Twin
- The Simulation Aspect. Challenges and solutions for
Mechatronic Systems and their Designers”. In: Mechatronic
Grieves, Michael. 2014. “Digital twin: manufacturing excel-
lence through virtual factory replication”. In: White paper
1, pp. 1–7.
Korth, Benjamin et al. 2018. “Simulation-ready digital twin for
realtime management of logistics systems”. In: 2018 IEEE
International Conference on Big Data (Big Data). IEEE,
pp. 4194–4201.
Lu, Yuqian et al. 2019. “Digital Twin-driven smart manu-
facturing: Connotation, reference model, applications and
research issues”. In: Robotics and Computer-Integrated
Malakuti, S. et al. 2020. “Digital Twins for Industrial Applica-
tions”. In: An Industrial Internet Consortium White Paper.
URL: https : / / www . iiconsortium . org / press - room / 02 - 18 -
Meredith, Jack R. 1987. “The Strategic Advantages of the
Factory of the Future”. In: California Management Review
29.3, pp. 27–41. DO I: 10.2307/41165250.
Negri, Elisa et al. 2019. “FMU-supported simulation for CPS
Digital Twin”. In: Procedia manufacturing 28, pp. 201–206.
Nieto, Flor de Asis Marti et al. 2017. “From Factory of the
Future to Future of the Factory: Integration Approaches”.
In: IFAC-PapersOnLine 50.1, pp. 11695–11700.
Rosen, Roland et al. 2015. “About the importance of autonomy
and digital twins for the future of manufacturing”. In: IFAC-
PapersOnLine 48.3, pp. 567–572.
Salierno, Giulio et al. 2019. “Different perspectives of a
factory of the future: An overview”. In: International
Conference on Advanced Information Systems Engineering.
Springer, pp. 107–119.
Scheifele, Christian et al. 2019. “Real-time co-simulation
for the virtual commissioning of production systems”. In:
Procedia CIRP 79, pp. 397–402.
Steinbrink, C. 2016. “A nonintrusive uncertainty quantification
system for modular smart grid co-simulation”. PhD thesis.
At University of Oldenburg.
Tao, Fei et al. 2018. “Digital twin-driven product design,
manufacturing and service with big data”. In: The Interna-
tional Journal of Advanced Manufacturing Technology 94.9,
pp. 3563–3576. URL:
... This component uses the predictions from the DT to check the effect of planned reconfigurations. While our previous work investigates on the left and the middle layer of this architecture (Eschemann et al. 2020), the present paper elaborates on the "Production Dynamic Learning" component. Section RELATED WORK provides an overview of related work on this topic. ...
... We assume factories that can be characterized by a factory model as in (Eschemann et al. 2020). Those factories consist of machines and warehouses with an individual set of input and output buffers to receive and send products (cf. Figure 2). ...
Conference Paper
Full-text available
The performance of factory-internal logistic systems plays a central role for the overall productivity of the factory of the future. A key element is the optimization of logistic systems based on predictive analytics of transport tasks in order to anticipate and to adapt to changes of the production flows in the factory. Although this information might be derivable from production plans and machine configurations, the data is often not available for transport systems like automated guided vehicles provided by third parties due to data protection concerns. Here, prediction needs to be based on observable information. In this paper, we investigate methods for the prediction of future transport tasks between machines based on recorded task histories. For this, we modify several prediction methods from the literature for making them applicable to predict timestamps and destinations of future tasks. The results are compared with evaluation data taken from a factory simulation via multiple metrics, which are revealing suitable predictors based on achieved prediction accuracy.
... Transparency Reduced Lead Time Service Level Reliability Process Resilience Increased Process Efficiency Yao et al. (2018) [75]; Eschemann et al.(2020)[76] ...
Full-text available
Background: Digital supply chain twins (DSCT) are gaining increased attention in academia and practice as they emerge as one of the most important trends in logistics and supply chain management (LSCM). Still, there seems to be no common understanding of the term in scientific literature. Moreover, the broad field of LSCM allows for a multitude of feasible application areas and use cases, yet there exists no conclusive list of them as to date. Methods: This study builds upon a systematic literature review of 66 DSCT articles to identify application areas of DSCT in LSCM as well as specific use cases and their respective intended benefits. Results: To start with, the study derives a unified definition of DSCTs, including possible scopes of applications. Afterwards, five application areas of DSCT in LSCM are synthesized as well as 14 individual use cases and their respective intended benefits. Conclusions: The study leads towards a conceptual clarification of DSCT that is of importance for research and practice alike. For managers it additionally provides up-to-date use cases to guide DSCT applications in practice.
... This demonstrates the importance of short and optimized transport routes as well as the avoidance of unnessecary movements. In (Eschemann et al. 2020) the robot behavior was optimized, this work delivers an optimized environment for the AGVs. Figure 1 shows a simplified layout of the actual factory. Machine "0" has five feeds "0a -0e" which can receive, hold and extract magazines. ...
Conference Paper
Full-text available
In this paper, the facility layout problem is studied using a real existing factory, with unequal-sized facilities and unequal-sized factory floor. The overall goal is to find optimized layouts to increase the productivity of the manufacturing process. To find optimal layouts a genetic algorithm is used in combination with a digital image that serves as visual interface for a supervisor. Via a 2D and 3D environment, the supervisor can apply manual mutations into the current genetic algorithm solution. The fitness of the layouts is determined by summing up material handling costs. The supervisor concept allows an unlimited number of constraints to be considered during layout generation. The solutions are compared to the actual factory layout, which was constantly improved over a period of 20 years. Through the combination of genetic algorithms and human knowledge the creation of application-oriented solutions can be reached quicker and with less effort than conventionally.
... [87] X X X X X 10 [88] X X X X X X 11 [89] X X X X X 12 [90] X X X X 13 ...
Full-text available
In the Industry 4.0 era, the Digital Twin has become one of the most promising enabling technologies supporting material flow. Although the literature on the Digital Twin is becoming relatively well explored, including a certain number of review papers, the context of the Digital Twins application in internal transport systems has not been investigated so far. This paper thoroughly reviews the research on the Digital Twins applied in internal transport systems concerning major research trends within this research area and identification of future research directions. It provides clarification of various definitions related to the Digital Twin concept, including misconceptions such as a digital shadow, a digital model, and a digital mirror. Additionally, the relationships between terms such as material handling, material flow, and intralogistics in the context of internal transport systems coupled with the Digital Twin are explained. This paper’s contribution to the current state of the art of the Digital Twins is three-fold: (1) recognition of the most influential and high-impact journals, papers, and researchers; (2) identification of the major research trends related to the Digital Twins applications in internal transport systems, and (3) presentation of future research agendas in investigating Digital Twins applied for internal transport systems.
... • для оптимизации работы внутреннего автопарка продукции [57]; ...
Full-text available
Рассматриваются практические аспекты создания цифровых двойников (ЦД), прежде всего, применяемых в промышленности. Описываются прикладные платформы, реализующие ключевые элементы киберфизических систем: промышленный Internet вещей, анализ больших данных, коммуникационную и вычислительную инфраструктуру, собственно ЦД. Приводятся многочисленные примеры практического внедрения ЦД в промышленности. Характеризуются текущее состояние и тренды рынка ЦД в мире и в России. Обсуждаются социально-технические вызовы ширящегося распространения ЦД и задачи участников этого революционного этапа трансформации экономики – промышленности, науки, правительственных и финансовых институтов, общества. / The practical aspects of Digital Twins’ (DT) development, primarily used in industry, are considered. The article describes the application platforms that implement the key elements of cyber-physical systems: Industrial Internet of Things, Big data analysis, communicating and computing infrastructure, and the DT itself. Numerous examples of practical DT implementation in industry are given. The current state and trends of DT market in the world and in Russia are characterized. The socio-technical challenges of DT expanding are discussed as well as the tasks of the participants of this economy transformation – industry, science, government, financial institutions, and society.
Full-text available
На основе анализа 320 публикаций за период 2020-22 гг. приводится обзор текущего состояния и новшеств в разработке и практическом внедрении цифровых двойников (ЦД) в задачах автоматизации промышленности и других отраслей народного хозяйства. Обсуждается сложившееся в профессиональном сообществе понимание состава, назначения и границ концепции цифровых двойников. Характеризуются базовые технологии ЦД, уточняются ближайшие и долговременные перспективы их развития. Дается сводка последних ключевых публикаций по исследованию концепции и технологий ЦД, разработке отраслевых решений, проработке смежных концепций и интегрированных подходов, например, таких как городские ЦД. "Автоматизация в промышленности", 2022, № 7, 3-14.
Full-text available
In the context of Industry 4.0, a growing use is being made of simulation-based decision-support tools commonly named Digital Twins. Digital Twins are replicas of the physical manufacturing assets, providing means for the monitoring and control of individual assets. Although extensive research on Digital Twins and their applications has been carried out, the majority of existing approaches are asset specific. Little consideration is made of human factors and interdependencies between different production assets are commonly ignored. In this paper, we address those limitations and propose innovations for cognitive modeling and co-simulation which may unleash novel uses of Digital Twins in Factories of the Future. We introduce a holistic Digital Twin approach, in which the factory is not represented by a set of separated Digital Twins but by a comprehensive modeling and simulation capacity embracing the full manufacturing process including external network dependencies. Furthermore, we introduce novel approaches for integrating models of human behavior and capacities for security testing with Digital Twins and show how the holistic Digital Twin can enable new services for the optimization and resilience of Factories of the Future. To illustrate this approach, we introduce a specific use-case implemented in field of Aerospace System Manufacturing.
Technical Report
Full-text available
This whitepaper provides practical guidance on digital twin, including the definition, benefits, architectures and the necessary building blocks to implement one. We illustrate the relationships between an Industrial Internet of Things (IIoT) system and its twin with use cases.
Full-text available
During virtual commissioning, the real control technology is put into operation in advance using a model-based digital twin of the production system. The growing demands on the significance of digital twins result in the need for higher model depths and model scopes of the behavior models used. However, new approaches to enhance the usable computing power in a real-time simulation are required. This paper presents a real-time co-simulation platform for virtual commissioning. The platform includes the integration of powerful technology-specific simulation solutions based on integration interfaces and a real-time co-simulation architecture based on partitioning, parallelization, synchronization and data exchange mechanisms.
Full-text available
Manufacturing companies are experiencing the fourth industrial revolution characterised by the introduction of new technologies into production equipment, such as the Cyber Physical Systems and the Digital Twin simulations. Companies are then challenged by the digitization of products and production systems information, which leads to new potentials for digital continuity – i.e. information available and continuously updated for the decision makers – along the lifecycles. A semantic data model, that structures and stores physical and operational data from the field, can support the digital continuity to be used in production system simulations in a Digital Twin paradigm. This work proposes to model specific aspects and behaviours of the production system separately from the core simulation, in order to flexibly decide whether to activate the replica of the specific behaviours only when needed. The modules interact with the main simulation run through standard interfaces, allowing an easy reusability of the single modules also in different simulation environments.
Full-text available
Nowadays, manufacturing systems transform themselves to become Factories of the Future (FoF) that is to say highly flexible, rapidly adaptable to external changes and aiming for a high degree of sustainability. This trend has generated several research streams that we analyse in this paper through a deep bibliographical review. This review leads us to conceptualize 3 main approaches: Computer Integrated Manufacturing, role of human work force and other integration approaches. A potential issue identified within the research is the interplay between computer-based and organization based approaches to manage the dynamic interactions among product, processes and production systems. The conclusions and orientations of future works are thus outlined to support the development of this co-evolution perspective for the successful transformation of Factories.
Full-text available
Nowadays, along with the application of new-generation information technologies in industry and manufacturing, the big data-driven manufacturing era is coming. However, although various big data in the entire product lifecycle, including product design, manufacturing, and service, can be obtained, it can be found that the current research on product lifecycle data mainly focuses on physical products rather than virtual models. Besides, due to the lack of convergence between product physical and virtual space, the data in product lifecycle is isolated, fragmented, and stagnant, which is useless for manufacturing enterprises. These problems lead to low level of efficiency, intelligence, sustainability in product design, manufacturing, and service phases. However, physical product data, virtual product data, and connected data that tie physical and virtual product are needed to support product design, manufacturing, and service. Therefore, how to generate and use converged cyber-physical data to better serve product lifecycle, so as to drive product design, manufacturing, and service to be more efficient, smart, and sustainable, is emphasized and investigated based on our previous study on big data in product lifecycle management. In this paper, a new method for product design, manufacturing, and service driven by digital twin is proposed. The detailed application methods and frameworks of digital twin-driven product design, manufacturing, and service are investigated. Furthermore, three cases are given to illustrate the future applications of digital twin in the three phases of a product respectively.
Information and communication technology (ICT) applications are increasingly being applied in industry. Indeed, in the last years, initiatives like Industry 4.0, Factories of the Future and Industrial Internet of things, pushed by governments and industrial leaders, refer to ICT as the future of manufacturing. To meet this trend, machine tool industry needs to include all these advances. This chapter provides a state of the art of ICT applied in machine tool industry and serves as a reference of all the developments of twin-control project.
Industrie 4.0 - the “brand” name of the German initiative driving the future of manufacturing - is one of several initiatives around the globe emphasizing the importance of industrial manufacturing for economy and society. Besides the socio-economical if not political question which has to be answered - including the question about the future of labor - there are a couple of substantial technical and technological questions that have to be taken care of as well.